Import Mesa 19.2.8

16 files changed, 7037 insertions, 20325 deletions

diff --git a/lib/mesa/src/amd/vulkan/radv_nir_lower_ycbcr_textures.c b/lib/mesa/src/amd/vulkan/radv_nir_lower_ycbcr_textures.c
index 7d15a78b3..5d771c2fc 100644
--- a/lib/mesa/src/amd/vulkan/radv_nir_lower_ycbcr_textures.c
+++ b/lib/mesa/src/amd/vulkan/radv_nir_lower_ycbcr_textures.c
@@ -21,301 +21,432 @@
  * IN THE SOFTWARE.
  */
 
-#include "nir/nir.h"
-#include "nir/nir_builder.h"
-#include "nir/nir_vulkan.h"
 #include "radv_private.h"
 #include "radv_shader.h"
 #include "vk_format.h"
+#include "nir/nir.h"
+#include "nir/nir_builder.h"
 
 struct ycbcr_state {
-   nir_builder *builder;
-   nir_ssa_def *image_size;
-   nir_tex_instr *origin_tex;
-   nir_deref_instr *tex_deref;
-   const struct radv_sampler_ycbcr_conversion_state *conversion;
-   bool unnormalized_coordinates;
+	nir_builder *builder;
+	nir_ssa_def *image_size;
+	nir_tex_instr *origin_tex;
+	nir_deref_instr *tex_deref;
+	const struct radv_sampler_ycbcr_conversion *conversion;
 };
 
 static nir_ssa_def *
-get_texture_size(struct ycbcr_state *state, nir_deref_instr *texture)
+y_range(nir_builder *b,
+        nir_ssa_def *y_channel,
+        int bpc,
+        VkSamplerYcbcrRange range)
 {
-   nir_builder *b = state->builder;
-   const struct glsl_type *type = texture->type;
-   nir_tex_instr *tex = nir_tex_instr_create(b->shader, 1);
-
-   tex->op = nir_texop_txs;
-   tex->sampler_dim = glsl_get_sampler_dim(type);
-   tex->is_array = glsl_sampler_type_is_array(type);
-   tex->is_shadow = glsl_sampler_type_is_shadow(type);
-   tex->dest_type = nir_type_int32;
+	switch (range) {
+	case VK_SAMPLER_YCBCR_RANGE_ITU_FULL:
+		return y_channel;
+	case VK_SAMPLER_YCBCR_RANGE_ITU_NARROW:
+		return nir_fmul(b,
+		                nir_fadd(b,
+		                         nir_fmul(b, y_channel,
+		                                  nir_imm_float(b, pow(2, bpc) - 1)),
+		                         nir_imm_float(b, -16.0f * pow(2, bpc - 8))),
+		                nir_frcp(b, nir_imm_float(b, 219.0f * pow(2, bpc - 8))));
+	default:
+		unreachable("missing Ycbcr range");
+		return NULL;
+	}
+}
 
-   tex->src[0].src_type = nir_tex_src_texture_deref;
-   tex->src[0].src = nir_src_for_ssa(&texture->dest.ssa);
+static nir_ssa_def *
+chroma_range(nir_builder *b,
+             nir_ssa_def *chroma_channel,
+             int bpc,
+             VkSamplerYcbcrRange range)
+{
+	switch (range) {
+	case VK_SAMPLER_YCBCR_RANGE_ITU_FULL:
+		return nir_fadd(b, chroma_channel,
+		                nir_imm_float(b, -pow(2, bpc - 1) / (pow(2, bpc) - 1.0f)));
+	case VK_SAMPLER_YCBCR_RANGE_ITU_NARROW:
+		return nir_fmul(b,
+		                nir_fadd(b,
+		                         nir_fmul(b, chroma_channel,
+		                                  nir_imm_float(b, pow(2, bpc) - 1)),
+		                         nir_imm_float(b, -128.0f * pow(2, bpc - 8))),
+		                nir_frcp(b, nir_imm_float(b, 224.0f * pow(2, bpc - 8))));
+	default:
+		unreachable("missing Ycbcr range");
+		return NULL;
+	}
+}
 
-   nir_ssa_dest_init(&tex->instr, &tex->dest, nir_tex_instr_dest_size(tex), 32, NULL);
-   nir_builder_instr_insert(b, &tex->instr);
+typedef struct nir_const_value_3_4 {
+	nir_const_value v[3][4];
+} nir_const_value_3_4;
 
-   state->builder->shader->info.uses_resource_info_query = true;
+static const nir_const_value_3_4 *
+ycbcr_model_to_rgb_matrix(VkSamplerYcbcrModelConversion model)
+{
+	switch (model) {
+	case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_601: {
+		static const nir_const_value_3_4 bt601 = { {
+			{ { .f32 =  1.402f             }, { .f32 = 1.0f }, { .f32 =  0.0f               }, { .f32 = 0.0f } },
+			{ { .f32 = -0.714136286201022f }, { .f32 = 1.0f }, { .f32 = -0.344136286201022f }, { .f32 = 0.0f } },
+			{ { .f32 =  0.0f               }, { .f32 = 1.0f }, { .f32 =  1.772f             }, { .f32 = 0.0f } },
+		} };
+
+		return &bt601;
+	}
+	case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_709: {
+		static const nir_const_value_3_4 bt709 = { {
+			{ { .f32 =  1.5748031496063f   }, { .f32 = 1.0f }, { .f32 =  0.0f               }, { .f32 = 0.0f } },
+			{ { .f32 = -0.468125209181067f }, { .f32 = 1.0f }, { .f32 = -0.187327487470334f }, { .f32 = 0.0f } },
+			{ { .f32 =  0.0f               }, { .f32 = 1.0f }, { .f32 =  1.85563184264242f  }, { .f32 = 0.0f } },
+		} };
+
+		return &bt709;
+	}
+	case VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_2020: {
+		static const nir_const_value_3_4 bt2020 = { {
+			{ { .f32 =  1.4746f            }, { .f32 = 1.0f }, { .f32 =  0.0f               }, { .f32 = 0.0f } },
+			{ { .f32 = -0.571353126843658f }, { .f32 = 1.0f }, { .f32 = -0.164553126843658f }, { .f32 = 0.0f } },
+			{ { .f32 =  0.0f               }, { .f32 = 1.0f }, { .f32 =  1.8814f            }, { .f32 = 0.0f } },
+		} };
+
+		return &bt2020;
+	}
+	default:
+		unreachable("missing Ycbcr model");
+		return NULL;
+	}
+}
 
-   return nir_i2f32(b, &tex->dest.ssa);
+static nir_ssa_def *
+convert_ycbcr(struct ycbcr_state *state,
+              nir_ssa_def *raw_channels,
+              uint8_t bits)
+{
+	nir_builder *b = state->builder;
+	const struct radv_sampler_ycbcr_conversion *conversion = state->conversion;
+
+	nir_ssa_def *expanded_channels =
+		nir_vec4(b,
+		         chroma_range(b, nir_channel(b, raw_channels, 0),
+		                      bits, conversion->ycbcr_range),
+		         y_range(b, nir_channel(b, raw_channels, 1),
+		                 bits, conversion->ycbcr_range),
+		         chroma_range(b, nir_channel(b, raw_channels, 2),
+		                      bits, conversion->ycbcr_range),
+		         nir_imm_float(b, 1.0f));
+
+	if (conversion->ycbcr_model == VK_SAMPLER_YCBCR_MODEL_CONVERSION_YCBCR_IDENTITY)
+		return expanded_channels;
+
+	const nir_const_value_3_4 *conversion_matrix =
+		ycbcr_model_to_rgb_matrix(conversion->ycbcr_model);
+
+	nir_ssa_def *converted_channels[] = {
+		nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix->v[0])),
+		nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix->v[1])),
+		nir_fdot4(b, expanded_channels, nir_build_imm(b, 4, 32, conversion_matrix->v[2]))
+	};
+
+	return nir_vec4(b,
+	                converted_channels[0], converted_channels[1],
+	                converted_channels[2], nir_imm_float(b, 1.0f));
 }
 
 static nir_ssa_def *
-implicit_downsampled_coord(nir_builder *b, nir_ssa_def *value, nir_ssa_def *max_value,
-                           int div_scale)
+get_texture_size(struct ycbcr_state *state, nir_deref_instr *texture)
 {
-   return nir_fadd(
-      b, value,
-      nir_fdiv(b, nir_imm_float(b, 1.0f), nir_fmul(b, nir_imm_float(b, div_scale), max_value)));
+	nir_builder *b = state->builder;
+	const struct glsl_type *type = texture->type;
+	nir_tex_instr *tex = nir_tex_instr_create(b->shader, 1);
+
+	tex->op = nir_texop_txs;
+	tex->sampler_dim = glsl_get_sampler_dim(type);
+	tex->is_array = glsl_sampler_type_is_array(type);
+	tex->is_shadow = glsl_sampler_type_is_shadow(type);
+	tex->dest_type = nir_type_int;
+
+	tex->src[0].src_type = nir_tex_src_texture_deref;
+	tex->src[0].src = nir_src_for_ssa(&texture->dest.ssa);
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest,
+	                  nir_tex_instr_dest_size(tex), 32, NULL);
+	nir_builder_instr_insert(b, &tex->instr);
+
+	return nir_i2f32(b, &tex->dest.ssa);
 }
 
 static nir_ssa_def *
-implicit_downsampled_coord_unnormalized(nir_builder *b, nir_ssa_def *value, int div_scale)
+implicit_downsampled_coord(nir_builder *b,
+                           nir_ssa_def *value,
+                           nir_ssa_def *max_value,
+                           int div_scale)
 {
-   return nir_fadd(
-      b, value,
-      nir_imm_float(b, 1.0f / (float)div_scale));
+	return nir_fadd(b,
+	                value,
+	                nir_fdiv(b,
+	                         nir_imm_float(b, 1.0f),
+	                         nir_fmul(b,
+	                                  nir_imm_float(b, div_scale),
+	                                  max_value)));
 }
 
 static nir_ssa_def *
-implicit_downsampled_coords(struct ycbcr_state *state, nir_ssa_def *old_coords)
+implicit_downsampled_coords(struct ycbcr_state *state,
+                            nir_ssa_def *old_coords)
 {
-   nir_builder *b = state->builder;
-   const struct radv_sampler_ycbcr_conversion_state *conversion = state->conversion;
-   nir_ssa_def *image_size = NULL;
-   nir_ssa_def *comp[4] = {
-      NULL,
-   };
-   enum pipe_video_chroma_format chroma_format =
-      pipe_format_to_chroma_format(vk_format_to_pipe_format(state->conversion->format));
-   const unsigned divisors[2] = {chroma_format <= PIPE_VIDEO_CHROMA_FORMAT_422 ? 2 : 1,
-                                 chroma_format <= PIPE_VIDEO_CHROMA_FORMAT_420 ? 2 : 1};
-
-   for (int c = 0; c < old_coords->num_components; c++) {
-      comp[c] = nir_channel(b, old_coords, c);
-
-      if (c < ARRAY_SIZE(divisors) && divisors[c] > 1) {
-         if (state->unnormalized_coordinates)
-            comp[c] = nir_fdiv(b, comp[c], nir_imm_float(b, divisors[c]));
-
-         if (conversion->chroma_offsets[c] == VK_CHROMA_LOCATION_COSITED_EVEN) {
-            if (state->unnormalized_coordinates) {
-               comp[c] = implicit_downsampled_coord_unnormalized(b, comp[c], divisors[c]);
-            } else {
-               if (!image_size)
-                  image_size = get_texture_size(state, state->tex_deref);
-
-               comp[c] = implicit_downsampled_coord(b, comp[c], nir_channel(b, image_size, c), divisors[c]);
-            }
-         }
-      }
-   }
-
-   return nir_vec(b, comp, old_coords->num_components);
+	nir_builder *b = state->builder;
+	const struct radv_sampler_ycbcr_conversion *conversion = state->conversion;
+	nir_ssa_def *image_size = NULL;
+	nir_ssa_def *comp[4] = { NULL, };
+	const struct vk_format_description *fmt_desc = vk_format_description(state->conversion->format);
+	const unsigned divisors[2] = {fmt_desc->width_divisor, fmt_desc->height_divisor};
+
+	for (int c = 0; c < old_coords->num_components; c++) {
+		if (c < ARRAY_SIZE(divisors) && divisors[c] > 1 &&
+		    conversion->chroma_offsets[c] == VK_CHROMA_LOCATION_COSITED_EVEN) {
+			if (!image_size)
+				image_size = get_texture_size(state, state->tex_deref);
+
+			comp[c] = implicit_downsampled_coord(b,
+			                                     nir_channel(b, old_coords, c),
+			                                     nir_channel(b, image_size, c),
+			                                     divisors[c]);
+		} else {
+			comp[c] = nir_channel(b, old_coords, c);
+		}
+	}
+
+	return nir_vec(b, comp, old_coords->num_components);
 }
 
 static nir_ssa_def *
-create_plane_tex_instr_implicit(struct ycbcr_state *state, uint32_t plane)
+create_plane_tex_instr_implicit(struct ycbcr_state *state,
+                                uint32_t plane)
 {
-   nir_builder *b = state->builder;
-   nir_tex_instr *old_tex = state->origin_tex;
-   nir_tex_instr *tex = nir_tex_instr_create(b->shader, old_tex->num_srcs + 1);
-   for (uint32_t i = 0; i < old_tex->num_srcs; i++) {
-      tex->src[i].src_type = old_tex->src[i].src_type;
-
-      switch (old_tex->src[i].src_type) {
-      case nir_tex_src_coord:
-         if (plane && true /*state->conversion->chroma_reconstruction*/) {
-            assert(old_tex->src[i].src.is_ssa);
-            tex->src[i].src =
-               nir_src_for_ssa(implicit_downsampled_coords(state, old_tex->src[i].src.ssa));
-            break;
-         }
-      FALLTHROUGH;
-      default:
-         nir_src_copy(&tex->src[i].src, &old_tex->src[i].src, &tex->instr);
-         break;
-      }
-   }
-
-   tex->src[tex->num_srcs - 1].src = nir_src_for_ssa(nir_imm_int(b, plane));
-   tex->src[tex->num_srcs - 1].src_type = nir_tex_src_plane;
-
-   tex->sampler_dim = old_tex->sampler_dim;
-   tex->dest_type = old_tex->dest_type;
-   tex->is_array = old_tex->is_array;
-
-   tex->op = old_tex->op;
-   tex->coord_components = old_tex->coord_components;
-   tex->is_new_style_shadow = old_tex->is_new_style_shadow;
-   tex->component = old_tex->component;
-
-   tex->texture_index = old_tex->texture_index;
-   tex->sampler_index = old_tex->sampler_index;
-
-   nir_ssa_dest_init(&tex->instr, &tex->dest, old_tex->dest.ssa.num_components,
-                     nir_dest_bit_size(old_tex->dest), NULL);
-   nir_builder_instr_insert(b, &tex->instr);
-
-   return &tex->dest.ssa;
+	nir_builder *b = state->builder;
+	nir_tex_instr *old_tex = state->origin_tex;
+	nir_tex_instr *tex = nir_tex_instr_create(b->shader, old_tex->num_srcs+ 1);
+	for (uint32_t i = 0; i < old_tex->num_srcs; i++) {
+		tex->src[i].src_type = old_tex->src[i].src_type;
+
+		switch (old_tex->src[i].src_type) {
+		case nir_tex_src_coord:
+			if (plane && true/*state->conversion->chroma_reconstruction*/) {
+				assert(old_tex->src[i].src.is_ssa);
+				tex->src[i].src =
+					nir_src_for_ssa(implicit_downsampled_coords(state,
+					                                            old_tex->src[i].src.ssa));
+				break;
+			}
+		/* fall through */
+		default:
+			nir_src_copy(&tex->src[i].src, &old_tex->src[i].src, tex);
+			break;
+		}
+	}
+
+	tex->src[tex->num_srcs - 1].src = nir_src_for_ssa(nir_imm_int(b, plane));
+	tex->src[tex->num_srcs - 1].src_type = nir_tex_src_plane;
+
+	tex->sampler_dim = old_tex->sampler_dim;
+	tex->dest_type = old_tex->dest_type;
+	tex->is_array = old_tex->is_array;
+
+	tex->op = old_tex->op;
+	tex->coord_components = old_tex->coord_components;
+	tex->is_new_style_shadow = old_tex->is_new_style_shadow;
+	tex->component = old_tex->component;
+
+	tex->texture_index = old_tex->texture_index;
+	tex->texture_array_size = old_tex->texture_array_size;
+	tex->sampler_index = old_tex->sampler_index;
+
+	nir_ssa_dest_init(&tex->instr, &tex->dest,
+	                  old_tex->dest.ssa.num_components,
+	                  nir_dest_bit_size(old_tex->dest), NULL);
+	nir_builder_instr_insert(b, &tex->instr);
+
+	return &tex->dest.ssa;
 }
 
 struct swizzle_info {
-   unsigned plane[4];
-   unsigned swizzle[4];
+	unsigned plane[4];
+	unsigned swizzle[4];
 };
 
 static struct swizzle_info
 get_plane_swizzles(VkFormat format)
 {
-   int planes = vk_format_get_plane_count(format);
-   switch (planes) {
-   case 3:
-      return (struct swizzle_info){{2, 0, 1, 0}, {0, 0, 0, 3}};
-   case 2:
-      return (struct swizzle_info){{1, 0, 1, 0}, {1, 0, 0, 3}};
-   case 1:
-      return (struct swizzle_info){{0, 0, 0, 0}, {0, 1, 2, 3}};
-   default:
-      unreachable("unhandled plane count for ycbcr swizzling");
-   }
+	int planes = vk_format_get_plane_count(format);
+	switch (planes) {
+	case 3:
+		return (struct swizzle_info) {
+			{2, 0, 1, 0},
+			{0, 0, 0, 3}
+		};
+	case 2:
+		return (struct swizzle_info) {
+			{1, 0, 1, 0},
+			{1, 0, 0, 3}
+		};
+	case 1:
+		return (struct swizzle_info) {
+			{0, 0, 0, 0},
+			{0, 1, 2, 3}
+		};
+	default:
+		unreachable("unhandled plane count for ycbcr swizzling");
+	}
 }
 
+
 static nir_ssa_def *
-build_swizzled_components(nir_builder *builder, VkFormat format, VkComponentMapping mapping,
+build_swizzled_components(nir_builder *builder,
+                          VkFormat format,
+                          VkComponentMapping mapping,
                           nir_ssa_def **plane_values)
 {
-   struct swizzle_info plane_swizzle = get_plane_swizzles(format);
-   enum pipe_swizzle swizzles[4];
-   nir_ssa_def *values[4];
-
-   vk_format_compose_swizzles(&mapping, (const unsigned char[4]){0, 1, 2, 3}, swizzles);
-
-   nir_ssa_def *zero = nir_imm_float(builder, 0.0f);
-   nir_ssa_def *one = nir_imm_float(builder, 1.0f);
-
-   for (unsigned i = 0; i < 4; ++i) {
-      switch (swizzles[i]) {
-      case PIPE_SWIZZLE_X:
-      case PIPE_SWIZZLE_Y:
-      case PIPE_SWIZZLE_Z:
-      case PIPE_SWIZZLE_W: {
-         unsigned channel = swizzles[i] - PIPE_SWIZZLE_X;
-         values[i] = nir_channel(builder, plane_values[plane_swizzle.plane[channel]],
-                                 plane_swizzle.swizzle[channel]);
-         break;
-      }
-      case PIPE_SWIZZLE_0:
-         values[i] = zero;
-         break;
-      case PIPE_SWIZZLE_1:
-         values[i] = one;
-         break;
-      default:
-         unreachable("unhandled swizzle");
-      }
-   }
-   return nir_vec(builder, values, 4);
+	struct swizzle_info plane_swizzle = get_plane_swizzles(format);
+	enum vk_swizzle swizzles[4];
+	nir_ssa_def *values[4];
+
+	vk_format_compose_swizzles(&mapping, (const unsigned char[4]){0,1,2,3}, swizzles);
+
+	nir_ssa_def *zero = nir_imm_float(builder, 0.0f);
+	nir_ssa_def *one = nir_imm_float(builder, 1.0f);
+
+	for (unsigned i = 0; i < 4; ++i) {
+		switch(swizzles[i]) {
+		case VK_SWIZZLE_X:
+		case VK_SWIZZLE_Y:
+		case VK_SWIZZLE_Z:
+		case VK_SWIZZLE_W: {
+			unsigned channel = swizzles[i] - VK_SWIZZLE_X;
+			values[i] = nir_channel(builder,
+			                        plane_values[plane_swizzle.plane[channel]],
+			                        plane_swizzle.swizzle[channel]);
+			break;
+		}
+		case VK_SWIZZLE_0:
+			values[i] = zero;
+			break;
+		case VK_SWIZZLE_1:
+			values[i] = one;
+			break;
+		default:
+			unreachable("unhandled swizzle");
+		}
+	}
+	return nir_vec(builder, values, 4);
 }
 
 static bool
-try_lower_tex_ycbcr(const struct radv_pipeline_layout *layout, nir_builder *builder,
+try_lower_tex_ycbcr(const struct radv_pipeline_layout *layout,
+                    nir_builder *builder,
                     nir_tex_instr *tex)
 {
-   int deref_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
-   assert(deref_src_idx >= 0);
-   nir_deref_instr *deref = nir_src_as_deref(tex->src[deref_src_idx].src);
-
-   nir_variable *var = nir_deref_instr_get_variable(deref);
-   const struct radv_descriptor_set_layout *set_layout =
-      layout->set[var->data.descriptor_set].layout;
-   const struct radv_descriptor_set_binding_layout *binding =
-      &set_layout->binding[var->data.binding];
-   const struct radv_sampler_ycbcr_conversion_state *ycbcr_samplers =
-      radv_immutable_ycbcr_samplers(set_layout, var->data.binding);
-
-   if (!ycbcr_samplers)
-      return false;
-
-   assert(binding->immutable_samplers_offset);
-   const uint32_t *immutable_samplers =
-      radv_immutable_samplers(set_layout, binding);
-
-   /* For the following instructions, we don't apply any change and let the
-    * instruction apply to the first plane.
-    */
-   if (tex->op == nir_texop_txs || tex->op == nir_texop_query_levels || tex->op == nir_texop_lod)
-      return false;
-
-   assert(tex->texture_index == 0);
-   unsigned array_index = 0;
-   if (deref->deref_type != nir_deref_type_var) {
-      assert(deref->deref_type == nir_deref_type_array);
-      if (!nir_src_is_const(deref->arr.index))
-         return false;
-      array_index = nir_src_as_uint(deref->arr.index);
-      array_index = MIN2(array_index, binding->array_size - 1);
-   }
-   const struct radv_sampler_ycbcr_conversion_state *ycbcr_sampler = ycbcr_samplers + array_index;
-
-   if (ycbcr_sampler->format == VK_FORMAT_UNDEFINED)
-      return false;
-
-   bool unnormalized_coordinates = immutable_samplers[4 * array_index + 0] & S_008F30_FORCE_UNNORMALIZED(1);
-
-   struct ycbcr_state state = {
-      .builder = builder,
-      .origin_tex = tex,
-      .tex_deref = deref,
-      .conversion = ycbcr_sampler,
-      .unnormalized_coordinates = unnormalized_coordinates,
-   };
-
-   builder->cursor = nir_before_instr(&tex->instr);
-
-   VkFormat format = state.conversion->format;
-   const int plane_count = vk_format_get_plane_count(format);
-   nir_ssa_def *plane_values[3];
-
-   for (int p = 0; p < plane_count; ++p) {
-      plane_values[p] = create_plane_tex_instr_implicit(&state, p);
-   }
-
-   nir_ssa_def *result =
-      build_swizzled_components(builder, format, ycbcr_sampler->components, plane_values);
-   if (state.conversion->ycbcr_model != VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY) {
-      VkFormat first_format = vk_format_get_plane_format(format, 0);
-      uint32_t bits =
-         vk_format_get_component_bits(first_format, UTIL_FORMAT_COLORSPACE_RGB, PIPE_SWIZZLE_X);
-      /* TODO: swizzle and bpcs */
-      uint32_t bpcs[3] = {bits, bits, bits};
-      result = nir_convert_ycbcr_to_rgb(builder, state.conversion->ycbcr_model,
-                                        state.conversion->ycbcr_range, result, bpcs);
-   }
-
-   nir_ssa_def_rewrite_uses(&tex->dest.ssa, result);
-   nir_instr_remove(&tex->instr);
-
-   return true;
+	int deref_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
+	assert(deref_src_idx >= 0);
+	nir_deref_instr *deref = nir_src_as_deref(tex->src[deref_src_idx].src);
+
+	nir_variable *var = nir_deref_instr_get_variable(deref);
+	const struct radv_descriptor_set_layout *set_layout =
+		layout->set[var->data.descriptor_set].layout;
+	const struct radv_descriptor_set_binding_layout *binding =
+		&set_layout->binding[var->data.binding];
+	const struct radv_sampler_ycbcr_conversion *ycbcr_samplers =
+		radv_immutable_ycbcr_samplers(set_layout, var->data.binding);
+
+	if (!ycbcr_samplers)
+		return false;
+
+	/* For the following instructions, we don't apply any change and let the
+	 * instruction apply to the first plane.
+	 */
+	if (tex->op == nir_texop_txs ||
+	    tex->op == nir_texop_query_levels ||
+	    tex->op == nir_texop_lod)
+		return false;
+
+	assert(tex->texture_index == 0);
+	unsigned array_index = 0;
+	if (deref->deref_type != nir_deref_type_var) {
+		assert(deref->deref_type == nir_deref_type_array);
+		if (!nir_src_is_const(deref->arr.index))
+			return false;
+		array_index = nir_src_as_uint(deref->arr.index);
+		array_index = MIN2(array_index, binding->array_size - 1);
+	}
+	const struct radv_sampler_ycbcr_conversion *ycbcr_sampler = ycbcr_samplers + array_index;
+
+	if (ycbcr_sampler->format == VK_FORMAT_UNDEFINED)
+		return false;
+
+	struct ycbcr_state state = {
+		.builder = builder,
+		.origin_tex = tex,
+		.tex_deref = deref,
+		.conversion = ycbcr_sampler,
+	};
+
+	builder->cursor = nir_before_instr(&tex->instr);
+
+	VkFormat format = state.conversion->format;
+	const int plane_count = vk_format_get_plane_count(format);
+	nir_ssa_def *plane_values[3];
+
+	for (int p = 0; p < plane_count; ++p) {
+		plane_values[p] = create_plane_tex_instr_implicit(&state, p);
+	}
+
+	nir_ssa_def *result = build_swizzled_components(builder, format, ycbcr_sampler->components, plane_values);
+	if (state.conversion->ycbcr_model != VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY) {
+		VkFormat first_format = vk_format_get_plane_format(format, 0);
+		result = convert_ycbcr(&state, result, vk_format_get_component_bits(first_format, VK_FORMAT_COLORSPACE_RGB, VK_SWIZZLE_X));
+	}
+
+	nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(result));
+	nir_instr_remove(&tex->instr);
+
+	return true;
 }
 
-static bool
-radv_nir_lower_ycbcr_textures_instr(nir_builder *b, nir_instr *instr, void *layout)
+bool
+radv_nir_lower_ycbcr_textures(nir_shader *shader,
+                             const struct radv_pipeline_layout *layout)
 {
-   if (instr->type != nir_instr_type_tex)
-      return false;
+	bool progress = false;
 
-   nir_tex_instr *tex = nir_instr_as_tex(instr);
-   return try_lower_tex_ycbcr(layout, b, tex);
-}
+	nir_foreach_function(function, shader) {
+		if (!function->impl)
+			continue;
 
-bool
-radv_nir_lower_ycbcr_textures(nir_shader *shader, const struct radv_pipeline_layout *layout)
-{
-   return nir_shader_instructions_pass(shader,
-                                       radv_nir_lower_ycbcr_textures_instr,
-                                       nir_metadata_block_index |
-                                       nir_metadata_dominance,
-                                       (void *)layout);
+		bool function_progress = false;
+		nir_builder builder;
+		nir_builder_init(&builder, function->impl);
+
+		nir_foreach_block(block, function->impl) {
+			nir_foreach_instr_safe(instr, block) {
+				if (instr->type != nir_instr_type_tex)
+					continue;
+
+				nir_tex_instr *tex = nir_instr_as_tex(instr);
+				function_progress |= try_lower_tex_ycbcr(layout, &builder, tex);
+			}
+		}
+
+		if (function_progress) {
+			nir_metadata_preserve(function->impl,
+			                      nir_metadata_block_index |
+			                      nir_metadata_dominance);
+		}
+
+		progress |= function_progress;
+	}
+
+	return progress;
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_android.c b/lib/mesa/src/freedreno/vulkan/tu_android.c
index d1f6bb3ab..1ebc9e726 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_android.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_android.c
@@ -1,26 +1,35 @@
 /*
  * Copyright © 2017, Google Inc.
- * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_android.h"
+#include "tu_private.h"
 
 #include <hardware/gralloc.h>
-
-#if ANDROID_API_LEVEL >= 26
-#include <hardware/gralloc1.h>
-#endif
-
 #include <hardware/hardware.h>
 #include <hardware/hwvulkan.h>
+#include <libsync.h>
 
-#include "drm-uapi/drm_fourcc.h"
-
-#include "util/libsync.h"
-#include "util/os_file.h"
-
-#include "tu_device.h"
-#include "tu_image.h"
+#include <vulkan/vk_android_native_buffer.h>
+#include <vulkan/vk_icd.h>
 
 static int
 tu_hal_open(const struct hw_module_t *mod,
@@ -42,7 +51,7 @@ PUBLIC struct hwvulkan_module_t HAL_MODULE_INFO_SYM = {
        .module_api_version = HWVULKAN_MODULE_API_VERSION_0_1,
        .hal_api_version = HARDWARE_MAKE_API_VERSION(1, 0),
        .id = HWVULKAN_HARDWARE_MODULE_ID,
-       .name = "Turnip Vulkan HAL",
+       .name = "AMD Vulkan HAL",
        .author = "Google",
        .methods =
          &(hw_module_methods_t){
@@ -97,161 +106,41 @@ tu_hal_close(struct hw_device_t *dev)
    return -1;
 }
 
-/* get dma-buf and modifier from gralloc info */
-static VkResult
-tu_gralloc_info_other(struct tu_device *device,
+VkResult
+tu_image_from_gralloc(VkDevice device_h,
+                      const VkImageCreateInfo *base_info,
                       const VkNativeBufferANDROID *gralloc_info,
-                      int *dma_buf,
-                      uint64_t *modifier)
+                      const VkAllocationCallbacks *alloc,
+                      VkImage *out_image_h)
 
 {
-   const uint32_t *handle_fds = (uint32_t *)gralloc_info->handle->data;
-   const uint32_t *handle_data = &handle_fds[gralloc_info->handle->numFds];
-   bool ubwc = false;
-
-   if (gralloc_info->handle->numFds == 1) {
-      /* gbm_gralloc.  TODO: modifiers support */
-      *dma_buf = handle_fds[0];
-   } else if (gralloc_info->handle->numFds == 2) {
-      /* Qualcomm gralloc, find it at:
-       *
-       * https://android.googlesource.com/platform/hardware/qcom/display/.
-       *
-       * The gralloc_info->handle is a pointer to a struct private_handle_t
-       * from your platform's gralloc.  On msm8996 (a5xx) and newer grallocs
-       * that's libgralloc1/gr_priv_handle.h, while previously it was
-       * libgralloc/gralloc_priv.h.
-       */
-
-      if (gralloc_info->handle->numInts < 2) {
-         return vk_errorf(device, VK_ERROR_INVALID_EXTERNAL_HANDLE,
-                          "VkNativeBufferANDROID::handle::numInts is %d, "
-                          "expected at least 2 for qcom gralloc",
-                          gralloc_info->handle->numFds);
-      }
+   TU_FROM_HANDLE(tu_device, device, device_h);
+   VkImage image_h = VK_NULL_HANDLE;
+   struct tu_image *image = NULL;
+   struct tu_bo *bo = NULL;
+   VkResult result;
 
-      uint32_t gmsm = ('g' << 24) | ('m' << 16) | ('s' << 8) | 'm';
-      if (handle_data[0] != gmsm) {
-         return vk_errorf(device, VK_ERROR_INVALID_EXTERNAL_HANDLE,
-                          "private_handle_t::magic is %x, expected %x",
-                          handle_data[0], gmsm);
-      }
+   result = tu_image_create(
+      device_h,
+      &(struct tu_image_create_info) {
+         .vk_info = base_info, .scanout = true, .no_metadata_planes = true },
+      alloc, &image_h);
 
-      /* This UBWC flag was introduced in a5xx. */
-      ubwc = handle_data[1] & 0x08000000;
+   if (result != VK_SUCCESS)
+      return result;
 
-      /* QCOM gralloc has two fds passed in: the actual GPU buffer, and a buffer
-       * of CPU-side metadata.  I haven't found any need for the metadata buffer
-       * yet.  See qdMetaData.h for what's in the metadata fd.
-       */
-      *dma_buf = handle_fds[0];
-   } else {
-      return vk_errorf(device, VK_ERROR_INVALID_EXTERNAL_HANDLE,
+   if (gralloc_info->handle->numFds != 1) {
+      return vk_errorf(device->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE,
                        "VkNativeBufferANDROID::handle::numFds is %d, "
-                       "expected 1 (gbm_gralloc) or 2 (qcom gralloc)",
+                       "expected 1",
                        gralloc_info->handle->numFds);
    }
 
-   *modifier = ubwc ? DRM_FORMAT_MOD_QCOM_COMPRESSED : DRM_FORMAT_MOD_LINEAR;
-   return VK_SUCCESS;
-}
-
-static const char cros_gralloc_module_name[] = "CrOS Gralloc";
-
-#define CROS_GRALLOC_DRM_GET_BUFFER_INFO 4
-#define CROS_GRALLOC_DRM_GET_USAGE 5
-#define CROS_GRALLOC_DRM_GET_USAGE_FRONT_RENDERING_BIT 0x1
-
-struct cros_gralloc0_buffer_info {
-   uint32_t drm_fourcc;
-   int num_fds;
-   int fds[4];
-   uint64_t modifier;
-   int offset[4];
-   int stride[4];
-};
-
-static VkResult
-tu_gralloc_info_cros(struct tu_device *device,
-                     const VkNativeBufferANDROID *gralloc_info,
-                     int *dma_buf,
-                     uint64_t *modifier)
-
-{
-   const gralloc_module_t *gralloc = device->gralloc;
-   struct cros_gralloc0_buffer_info info;
-   int ret;
-
-   ret = gralloc->perform(gralloc, CROS_GRALLOC_DRM_GET_BUFFER_INFO,
-                          gralloc_info->handle, &info);
-   if (ret)
-      return VK_ERROR_INVALID_EXTERNAL_HANDLE;
-
-   *dma_buf = info.fds[0];
-   *modifier = info.modifier;
-
-   return VK_SUCCESS;
-}
-
-VkResult
-tu_gralloc_info(struct tu_device *device,
-                const VkNativeBufferANDROID *gralloc_info,
-                int *dma_buf,
-                uint64_t *modifier)
-
-{
-   if (!device->gralloc) {
-      /* get gralloc module for gralloc buffer info query */
-      int ret = hw_get_module(GRALLOC_HARDWARE_MODULE_ID,
-                              (const hw_module_t **)&device->gralloc);
-
-      if (ret) {
-         /* This is *slightly* awkward, but if we are asked to import
-          * a gralloc handle, and there is no gralloc, it is some sort
-          * of invalid handle.
-          */
-         return vk_startup_errorf(device->instance,
-                                  VK_ERROR_INVALID_EXTERNAL_HANDLE,
-                                  "Could not open gralloc\n");
-      }
-
-      const gralloc_module_t *gralloc = device->gralloc;
-
-      mesa_logi("opened gralloc module name: %s", gralloc->common.name);
-
-      /* TODO not sure qcom gralloc module name, but we should check
-       * for it here and move the special gmsm handling out of
-       * tu_gralloc_info_other()
-       */
-      if (!strcmp(gralloc->common.name, cros_gralloc_module_name) && gralloc->perform) {
-         device->gralloc_type = TU_GRALLOC_CROS;
-      } else {
-         device->gralloc_type = TU_GRALLOC_OTHER;
-      }
-   }
-
-   if (device->gralloc_type == TU_GRALLOC_CROS) {
-      return tu_gralloc_info_cros(device, gralloc_info, dma_buf, modifier);
-   } else {
-      return tu_gralloc_info_other(device, gralloc_info, dma_buf, modifier);
-   }
-}
-
-/**
- * Creates the VkImage using the gralloc handle in *gralloc_info.
- *
- * We support two different grallocs here, gbm_gralloc, and the qcom gralloc
- * used on Android phones.
- */
-VkResult
-tu_import_memory_from_gralloc_handle(VkDevice device_h,
-                                     int dma_buf,
-                                     const VkAllocationCallbacks *alloc,
-                                     VkImage image_h)
-
-{
-   struct tu_image *image = NULL;
-   VkResult result;
+   /* Do not close the gralloc handle's dma_buf. The lifetime of the dma_buf
+    * must exceed that of the gralloc handle, and we do not own the gralloc
+    * handle.
+    */
+   int dma_buf = gralloc_info->handle->data[0];
 
    image = tu_image_from_handle(image_h);
 
@@ -264,52 +153,70 @@ tu_import_memory_from_gralloc_handle(VkDevice device_h,
       .image = image_h
    };
 
-   const VkImportMemoryFdInfoKHR import_info = {
-      .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO_KHR,
+   const VkImportMemoryFdInfo import_info = {
+      .sType = VK_STRUCTURE_TYPE_IMPORT_MEMORY_FD_INFO,
       .pNext = &ded_alloc,
       .handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT,
-      .fd = os_dupfd_cloexec(dma_buf),
+      .fd = dup(dma_buf),
    };
+   /* Find the first VRAM memory type, or GART for PRIME images. */
+   int memory_type_index = -1;
+   for (int i = 0;
+        i < device->physical_device->memory_properties.memoryTypeCount; ++i) {
+      bool is_local =
+         !!(device->physical_device->memory_properties.memoryTypes[i]
+               .propertyFlags &
+            VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
+      if (is_local) {
+         memory_type_index = i;
+         break;
+      }
+   }
+
+   /* fallback */
+   if (memory_type_index == -1)
+      memory_type_index = 0;
 
    result =
       tu_AllocateMemory(device_h,
                         &(VkMemoryAllocateInfo) {
                            .sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
                            .pNext = &import_info,
-                           .allocationSize = image->total_size,
-                           .memoryTypeIndex = 0,
+                           .allocationSize = image->size,
+                           .memoryTypeIndex = memory_type_index,
                         },
                         alloc, &memory_h);
    if (result != VK_SUCCESS)
       goto fail_create_image;
 
-   VkBindImageMemoryInfo bind_info = {
-      .sType = VK_STRUCTURE_TYPE_BIND_IMAGE_MEMORY_INFO,
-      .image = image_h,
-      .memory = memory_h,
-      .memoryOffset = 0,
-   };
-   tu_BindImageMemory2(device_h, 1, &bind_info);
+   tu_BindImageMemory(device_h, image_h, memory_h, 0);
 
    image->owned_memory = memory_h;
+   /* Don't clobber the out-parameter until success is certain. */
+   *out_image_h = image_h;
 
    return VK_SUCCESS;
 
 fail_create_image:
+fail_size:
    tu_DestroyImage(device_h, image_h, alloc);
 
    return result;
 }
 
-static VkResult
-format_supported_with_usage(VkDevice device_h, VkFormat format,
-                            VkImageUsageFlags imageUsage)
+VkResult
+tu_GetSwapchainGrallocUsageANDROID(VkDevice device_h,
+                                   VkFormat format,
+                                   VkImageUsageFlags imageUsage,
+                                   int *grallocUsage)
 {
    TU_FROM_HANDLE(tu_device, device, device_h);
    struct tu_physical_device *phys_dev = device->physical_device;
    VkPhysicalDevice phys_dev_h = tu_physical_device_to_handle(phys_dev);
    VkResult result;
 
+   *grallocUsage = 0;
+
    /* WARNING: Android Nougat's libvulkan.so hardcodes the VkImageUsageFlags
     * returned to applications via
     * VkSurfaceCapabilitiesKHR::supportedUsageFlags.
@@ -340,19 +247,12 @@ format_supported_with_usage(VkDevice device_h, VkFormat format,
    result = tu_GetPhysicalDeviceImageFormatProperties2(
       phys_dev_h, &image_format_info, &image_format_props);
    if (result != VK_SUCCESS) {
-      return vk_errorf(device, result,
+      return vk_errorf(device->instance, result,
                        "tu_GetPhysicalDeviceImageFormatProperties2 failed "
                        "inside %s",
                        __func__);
    }
 
-   return VK_SUCCESS;
-}
-
-static VkResult
-setup_gralloc0_usage(struct tu_device *device, VkFormat format,
-                     VkImageUsageFlags imageUsage, int *grallocUsage)
-{
    if (unmask32(&imageUsage, VK_IMAGE_USAGE_TRANSFER_DST_BIT |
                                 VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT))
       *grallocUsage |= GRALLOC_USAGE_HW_RENDER;
@@ -367,7 +267,7 @@ setup_gralloc0_usage(struct tu_device *device, VkFormat format,
     * gralloc swapchains.
     */
    if (imageUsage != 0) {
-      return vk_errorf(device, VK_ERROR_FORMAT_NOT_SUPPORTED,
+      return vk_errorf(device->instance, VK_ERROR_FORMAT_NOT_SUPPORTED,
                        "unsupported VkImageUsageFlags(0x%x) for gralloc "
                        "swapchain",
                        imageUsage);
@@ -390,66 +290,93 @@ setup_gralloc0_usage(struct tu_device *device, VkFormat format,
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_GetSwapchainGrallocUsageANDROID(VkDevice device_h,
-                                   VkFormat format,
-                                   VkImageUsageFlags imageUsage,
-                                   int *grallocUsage)
+VkResult
+tu_AcquireImageANDROID(VkDevice device,
+                       VkImage image_h,
+                       int nativeFenceFd,
+                       VkSemaphore semaphore,
+                       VkFence fence)
 {
-   TU_FROM_HANDLE(tu_device, device, device_h);
-   VkResult result;
+   VkResult semaphore_result = VK_SUCCESS, fence_result = VK_SUCCESS;
+
+   if (semaphore != VK_NULL_HANDLE) {
+      int semaphore_fd =
+         nativeFenceFd >= 0 ? dup(nativeFenceFd) : nativeFenceFd;
+      semaphore_result = tu_ImportSemaphoreFdKHR(
+         device, &(VkImportSemaphoreFdInfoKHR) {
+                    .sType = VK_STRUCTURE_TYPE_IMPORT_SEMAPHORE_FD_INFO_KHR,
+                    .flags = VK_SEMAPHORE_IMPORT_TEMPORARY_BIT,
+                    .fd = semaphore_fd,
+                    .semaphore = semaphore,
+                 });
+   }
 
-   result = format_supported_with_usage(device_h, format, imageUsage);
-   if (result != VK_SUCCESS)
-      return result;
+   if (fence != VK_NULL_HANDLE) {
+      int fence_fd = nativeFenceFd >= 0 ? dup(nativeFenceFd) : nativeFenceFd;
+      fence_result = tu_ImportFenceFdKHR(
+         device, &(VkImportFenceFdInfoKHR) {
+                    .sType = VK_STRUCTURE_TYPE_IMPORT_FENCE_FD_INFO_KHR,
+                    .flags = VK_FENCE_IMPORT_TEMPORARY_BIT,
+                    .fd = fence_fd,
+                    .fence = fence,
+                 });
+   }
 
-   *grallocUsage = 0;
-   return setup_gralloc0_usage(device, format, imageUsage, grallocUsage);
+   close(nativeFenceFd);
+
+   if (semaphore_result != VK_SUCCESS)
+      return semaphore_result;
+   return fence_result;
 }
 
-#if ANDROID_API_LEVEL >= 26
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_GetSwapchainGrallocUsage2ANDROID(VkDevice device_h,
-                                    VkFormat format,
-                                    VkImageUsageFlags imageUsage,
-                                    VkSwapchainImageUsageFlagsANDROID swapchainImageUsage,
-                                    uint64_t *grallocConsumerUsage,
-                                    uint64_t *grallocProducerUsage)
+VkResult
+tu_QueueSignalReleaseImageANDROID(VkQueue _queue,
+                                  uint32_t waitSemaphoreCount,
+                                  const VkSemaphore *pWaitSemaphores,
+                                  VkImage image,
+                                  int *pNativeFenceFd)
 {
-   TU_FROM_HANDLE(tu_device, device, device_h);
-   VkResult result;
-
-   *grallocConsumerUsage = 0;
-   *grallocProducerUsage = 0;
-   mesa_logd("%s: format=%d, usage=0x%x", __func__, format, imageUsage);
+   TU_FROM_HANDLE(tu_queue, queue, _queue);
+   VkResult result = VK_SUCCESS;
 
-   result = format_supported_with_usage(device_h, format, imageUsage);
-   if (result != VK_SUCCESS)
-      return result;
-
-   int32_t grallocUsage = 0;
-   result = setup_gralloc0_usage(device, format, imageUsage, &grallocUsage);
-   if (result != VK_SUCCESS)
-      return result;
+   if (waitSemaphoreCount == 0) {
+      if (pNativeFenceFd)
+         *pNativeFenceFd = -1;
+      return VK_SUCCESS;
+   }
 
-   /* Setup gralloc1 usage flags from gralloc0 flags. */
+   int fd = -1;
 
-   if (grallocUsage & GRALLOC_USAGE_HW_RENDER) {
-      *grallocProducerUsage |= GRALLOC1_PRODUCER_USAGE_GPU_RENDER_TARGET;
-      *grallocConsumerUsage |= GRALLOC1_CONSUMER_USAGE_CLIENT_TARGET;
-   }
+   for (uint32_t i = 0; i < waitSemaphoreCount; ++i) {
+      int tmp_fd;
+      result = tu_GetSemaphoreFdKHR(
+         tu_device_to_handle(queue->device),
+         &(VkSemaphoreGetFdInfoKHR) {
+            .sType = VK_STRUCTURE_TYPE_SEMAPHORE_GET_FD_INFO_KHR,
+            .handleType = VK_EXTERNAL_SEMAPHORE_HANDLE_TYPE_SYNC_FD_BIT,
+            .semaphore = pWaitSemaphores[i],
+         },
+         &tmp_fd);
+      if (result != VK_SUCCESS) {
+         if (fd >= 0)
+            close(fd);
+         return result;
+      }
 
-   if (grallocUsage & GRALLOC_USAGE_HW_TEXTURE) {
-      *grallocConsumerUsage |= GRALLOC1_CONSUMER_USAGE_GPU_TEXTURE;
+      if (fd < 0)
+         fd = tmp_fd;
+      else if (tmp_fd >= 0) {
+         sync_accumulate("tu", &fd, tmp_fd);
+         close(tmp_fd);
+      }
    }
 
-   if (grallocUsage & (GRALLOC_USAGE_HW_FB |
-                       GRALLOC_USAGE_HW_COMPOSER |
-                       GRALLOC_USAGE_EXTERNAL_DISP)) {
-      *grallocProducerUsage |= GRALLOC1_PRODUCER_USAGE_GPU_RENDER_TARGET;
-      *grallocConsumerUsage |= GRALLOC1_CONSUMER_USAGE_HWCOMPOSER;
+   if (pNativeFenceFd) {
+      *pNativeFenceFd = fd;
+   } else if (fd >= 0) {
+      close(fd);
+      /* We still need to do the exports, to reset the semaphores, but
+       * otherwise we don't wait on them. */
    }
-
    return VK_SUCCESS;
 }
-#endif
diff --git a/lib/mesa/src/freedreno/vulkan/tu_cmd_buffer.c b/lib/mesa/src/freedreno/vulkan/tu_cmd_buffer.c
index 0acb45d71..fe436e595 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_cmd_buffer.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_cmd_buffer.c
@@ -1,879 +1,807 @@
 /*
  * Copyright © 2016 Red Hat.
  * Copyright © 2016 Bas Nieuwenhuizen
- * SPDX-License-Identifier: MIT
  *
  * based in part on anv driver which is:
  * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_cmd_buffer.h"
+#include "tu_private.h"
+
+#include "registers/adreno_pm4.xml.h"
+#include "registers/adreno_common.xml.h"
+#include "registers/a6xx.xml.h"
 
-#include "vk_render_pass.h"
-#include "vk_util.h"
-#include "vk_common_entrypoints.h"
+#include "vk_format.h"
 
-#include "tu_clear_blit.h"
 #include "tu_cs.h"
-#include "tu_image.h"
-#include "tu_tracepoints.h"
 
-static void
-tu_clone_trace_range(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-                     struct u_trace_iterator begin, struct u_trace_iterator end)
+void
+tu_bo_list_init(struct tu_bo_list *list)
 {
-   if (u_trace_iterator_equal(begin, end))
-      return;
-
-   tu_cs_emit_wfi(cs);
-   tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
-   u_trace_clone_append(begin, end, &cmd->trace, cs,
-         tu_copy_timestamp_buffer);
+   list->count = list->capacity = 0;
+   list->bo_infos = NULL;
 }
 
-static void
-tu_clone_trace(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-               struct u_trace *trace)
+void
+tu_bo_list_destroy(struct tu_bo_list *list)
 {
-   tu_clone_trace_range(cmd, cs, u_trace_begin_iterator(trace),
-         u_trace_end_iterator(trace));
+   free(list->bo_infos);
 }
 
 void
-tu6_emit_event_write(struct tu_cmd_buffer *cmd,
-                     struct tu_cs *cs,
-                     enum vgt_event_type event)
+tu_bo_list_reset(struct tu_bo_list *list)
 {
-   bool need_seqno = false;
-   switch (event) {
-   case CACHE_FLUSH_TS:
-   case WT_DONE_TS:
-   case RB_DONE_TS:
-   case PC_CCU_FLUSH_DEPTH_TS:
-   case PC_CCU_FLUSH_COLOR_TS:
-   case PC_CCU_RESOLVE_TS:
-      need_seqno = true;
-      break;
-   default:
-      break;
+   list->count = 0;
+}
+
+/**
+ * \a flags consists of MSM_SUBMIT_BO_FLAGS.
+ */
+static uint32_t
+tu_bo_list_add_info(struct tu_bo_list *list,
+                    const struct drm_msm_gem_submit_bo *bo_info)
+{
+   for (uint32_t i = 0; i < list->count; ++i) {
+      if (list->bo_infos[i].handle == bo_info->handle) {
+         assert(list->bo_infos[i].presumed == bo_info->presumed);
+         list->bo_infos[i].flags |= bo_info->flags;
+         return i;
+      }
    }
 
-   tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, need_seqno ? 4 : 1);
-   tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(event));
-   if (need_seqno) {
-      tu_cs_emit_qw(cs, global_iova(cmd, seqno_dummy));
-      tu_cs_emit(cs, 0);
+   /* grow list->bo_infos if needed */
+   if (list->count == list->capacity) {
+      uint32_t new_capacity = MAX2(2 * list->count, 16);
+      struct drm_msm_gem_submit_bo *new_bo_infos = realloc(
+         list->bo_infos, new_capacity * sizeof(struct drm_msm_gem_submit_bo));
+      if (!new_bo_infos)
+         return TU_BO_LIST_FAILED;
+      list->bo_infos = new_bo_infos;
+      list->capacity = new_capacity;
    }
+
+   list->bo_infos[list->count] = *bo_info;
+   return list->count++;
 }
 
-/* Emits the tessfactor address to the top-level CS if it hasn't been already.
- * Updating this register requires a WFI if outstanding drawing is using it, but
- * tu6_init_hardware() will have WFIed before we started and no other draws
- * could be using the tessfactor address yet since we only emit one per cmdbuf.
- */
-static void
-tu6_lazy_emit_tessfactor_addr(struct tu_cmd_buffer *cmd)
+uint32_t
+tu_bo_list_add(struct tu_bo_list *list,
+               const struct tu_bo *bo,
+               uint32_t flags)
 {
-   if (cmd->state.tessfactor_addr_set)
-      return;
-
-   tu_cs_emit_regs(&cmd->cs, A6XX_PC_TESSFACTOR_ADDR(.qword = cmd->device->tess_bo->iova));
-   /* Updating PC_TESSFACTOR_ADDR could race with the next draw which uses it. */
-   cmd->state.cache.flush_bits |= TU_CMD_FLAG_WAIT_FOR_IDLE;
-   cmd->state.tessfactor_addr_set = true;
+   return tu_bo_list_add_info(list, &(struct drm_msm_gem_submit_bo) {
+                                       .flags = flags,
+                                       .handle = bo->gem_handle,
+                                       .presumed = bo->iova,
+                                    });
 }
 
-static void
-tu6_lazy_emit_vsc(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+VkResult
+tu_bo_list_merge(struct tu_bo_list *list, const struct tu_bo_list *other)
 {
-   struct tu_device *dev = cmd->device;
-
-   /* VSC buffers:
-    * use vsc pitches from the largest values used so far with this device
-    * if there hasn't been overflow, there will already be a scratch bo
-    * allocated for these sizes
-    *
-    * if overflow is detected, the stream size is increased by 2x
-    */
-   mtx_lock(&dev->mutex);
+   for (uint32_t i = 0; i < other->count; i++) {
+      if (tu_bo_list_add_info(list, other->bo_infos + i) == TU_BO_LIST_FAILED)
+         return VK_ERROR_OUT_OF_HOST_MEMORY;
+   }
 
-   struct tu6_global *global = dev->global_bo->map;
+   return VK_SUCCESS;
+}
 
-   uint32_t vsc_draw_overflow = global->vsc_draw_overflow;
-   uint32_t vsc_prim_overflow = global->vsc_prim_overflow;
+static VkResult
+tu_tiling_config_update_gmem_layout(struct tu_tiling_config *tiling,
+                                    const struct tu_device *dev)
+{
+   const uint32_t gmem_size = dev->physical_device->gmem_size;
+   uint32_t offset = 0;
 
-   if (vsc_draw_overflow >= dev->vsc_draw_strm_pitch)
-      dev->vsc_draw_strm_pitch = (dev->vsc_draw_strm_pitch - VSC_PAD) * 2 + VSC_PAD;
+   for (uint32_t i = 0; i < tiling->buffer_count; i++) {
+      /* 16KB-aligned */
+      offset = align(offset, 0x4000);
 
-   if (vsc_prim_overflow >= dev->vsc_prim_strm_pitch)
-      dev->vsc_prim_strm_pitch = (dev->vsc_prim_strm_pitch - VSC_PAD) * 2 + VSC_PAD;
+      tiling->gmem_offsets[i] = offset;
+      offset += tiling->tile0.extent.width * tiling->tile0.extent.height *
+                tiling->buffer_cpp[i];
+   }
 
-   cmd->vsc_prim_strm_pitch = dev->vsc_prim_strm_pitch;
-   cmd->vsc_draw_strm_pitch = dev->vsc_draw_strm_pitch;
+   return offset <= gmem_size ? VK_SUCCESS : VK_ERROR_OUT_OF_DEVICE_MEMORY;
+}
 
-   mtx_unlock(&dev->mutex);
+static void
+tu_tiling_config_update_tile_layout(struct tu_tiling_config *tiling,
+                                    const struct tu_device *dev)
+{
+   const uint32_t tile_align_w = dev->physical_device->tile_align_w;
+   const uint32_t tile_align_h = dev->physical_device->tile_align_h;
+   const uint32_t max_tile_width = 1024; /* A6xx */
+
+   tiling->tile0.offset = (VkOffset2D) {
+      .x = tiling->render_area.offset.x & ~(tile_align_w - 1),
+      .y = tiling->render_area.offset.y & ~(tile_align_h - 1),
+   };
+
+   const uint32_t ra_width =
+      tiling->render_area.extent.width +
+      (tiling->render_area.offset.x - tiling->tile0.offset.x);
+   const uint32_t ra_height =
+      tiling->render_area.extent.height +
+      (tiling->render_area.offset.y - tiling->tile0.offset.y);
+
+   /* start from 1 tile */
+   tiling->tile_count = (VkExtent2D) {
+      .width = 1,
+      .height = 1,
+   };
+   tiling->tile0.extent = (VkExtent2D) {
+      .width = align(ra_width, tile_align_w),
+      .height = align(ra_height, tile_align_h),
+   };
+
+   /* do not exceed max tile width */
+   while (tiling->tile0.extent.width > max_tile_width) {
+      tiling->tile_count.width++;
+      tiling->tile0.extent.width =
+         align(ra_width / tiling->tile_count.width, tile_align_w);
+   }
+
+   /* do not exceed gmem size */
+   while (tu_tiling_config_update_gmem_layout(tiling, dev) != VK_SUCCESS) {
+      if (tiling->tile0.extent.width > tiling->tile0.extent.height) {
+         tiling->tile_count.width++;
+         tiling->tile0.extent.width =
+            align(ra_width / tiling->tile_count.width, tile_align_w);
+      } else {
+         tiling->tile_count.height++;
+         tiling->tile0.extent.height =
+            align(ra_height / tiling->tile_count.height, tile_align_h);
+      }
+   }
+}
 
-   struct tu_bo *vsc_bo;
-   uint32_t size0 = cmd->vsc_prim_strm_pitch * MAX_VSC_PIPES +
-                    cmd->vsc_draw_strm_pitch * MAX_VSC_PIPES;
+static void
+tu_tiling_config_update_pipe_layout(struct tu_tiling_config *tiling,
+                                    const struct tu_device *dev)
+{
+   const uint32_t max_pipe_count = 32; /* A6xx */
 
-   tu_get_scratch_bo(dev, size0 + MAX_VSC_PIPES * 4, &vsc_bo);
+   /* start from 1 tile per pipe */
+   tiling->pipe0 = (VkExtent2D) {
+      .width = 1,
+      .height = 1,
+   };
+   tiling->pipe_count = tiling->tile_count;
 
-   tu_cs_emit_regs(cs,
-                   A6XX_VSC_DRAW_STRM_SIZE_ADDRESS(.bo = vsc_bo, .bo_offset = size0));
-   tu_cs_emit_regs(cs,
-                   A6XX_VSC_PRIM_STRM_ADDRESS(.bo = vsc_bo));
-   tu_cs_emit_regs(cs,
-                   A6XX_VSC_DRAW_STRM_ADDRESS(.bo = vsc_bo,
-                                              .bo_offset = cmd->vsc_prim_strm_pitch * MAX_VSC_PIPES));
+   /* do not exceed max pipe count vertically */
+   while (tiling->pipe_count.height > max_pipe_count) {
+      tiling->pipe0.height += 2;
+      tiling->pipe_count.height =
+         (tiling->tile_count.height + tiling->pipe0.height - 1) /
+         tiling->pipe0.height;
+   }
 
-   cmd->vsc_initialized = true;
+   /* do not exceed max pipe count */
+   while (tiling->pipe_count.width * tiling->pipe_count.height >
+          max_pipe_count) {
+      tiling->pipe0.width += 1;
+      tiling->pipe_count.width =
+         (tiling->tile_count.width + tiling->pipe0.width - 1) /
+         tiling->pipe0.width;
+   }
 }
 
 static void
-tu6_emit_flushes(struct tu_cmd_buffer *cmd_buffer,
-                 struct tu_cs *cs,
-                 enum tu_cmd_flush_bits flushes)
+tu_tiling_config_update_pipes(struct tu_tiling_config *tiling,
+                              const struct tu_device *dev)
 {
-   if (unlikely(cmd_buffer->device->physical_device->instance->debug_flags & TU_DEBUG_FLUSHALL))
-      flushes |= TU_CMD_FLAG_ALL_FLUSH | TU_CMD_FLAG_ALL_INVALIDATE;
-
-   if (unlikely(cmd_buffer->device->physical_device->instance->debug_flags & TU_DEBUG_SYNCDRAW))
-      flushes |= TU_CMD_FLAG_WAIT_MEM_WRITES |
-                 TU_CMD_FLAG_WAIT_FOR_IDLE |
-                 TU_CMD_FLAG_WAIT_FOR_ME;
-
-   /* Experiments show that invalidating CCU while it still has data in it
-    * doesn't work, so make sure to always flush before invalidating in case
-    * any data remains that hasn't yet been made available through a barrier.
-    * However it does seem to work for UCHE.
-    */
-   if (flushes & (TU_CMD_FLAG_CCU_FLUSH_COLOR |
-                  TU_CMD_FLAG_CCU_INVALIDATE_COLOR))
-      tu6_emit_event_write(cmd_buffer, cs, PC_CCU_FLUSH_COLOR_TS);
-   if (flushes & (TU_CMD_FLAG_CCU_FLUSH_DEPTH |
-                  TU_CMD_FLAG_CCU_INVALIDATE_DEPTH))
-      tu6_emit_event_write(cmd_buffer, cs, PC_CCU_FLUSH_DEPTH_TS);
-   if (flushes & TU_CMD_FLAG_CCU_INVALIDATE_COLOR)
-      tu6_emit_event_write(cmd_buffer, cs, PC_CCU_INVALIDATE_COLOR);
-   if (flushes & TU_CMD_FLAG_CCU_INVALIDATE_DEPTH)
-      tu6_emit_event_write(cmd_buffer, cs, PC_CCU_INVALIDATE_DEPTH);
-   if (flushes & TU_CMD_FLAG_CACHE_FLUSH)
-      tu6_emit_event_write(cmd_buffer, cs, CACHE_FLUSH_TS);
-   if (flushes & TU_CMD_FLAG_CACHE_INVALIDATE)
-      tu6_emit_event_write(cmd_buffer, cs, CACHE_INVALIDATE);
-   if (flushes & TU_CMD_FLAG_WAIT_MEM_WRITES)
-      tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-   if ((flushes & TU_CMD_FLAG_WAIT_FOR_IDLE) ||
-       (cmd_buffer->device->physical_device->info->a6xx.has_ccu_flush_bug &&
-        (flushes & (TU_CMD_FLAG_CCU_FLUSH_COLOR | TU_CMD_FLAG_CCU_FLUSH_DEPTH))))
-      tu_cs_emit_wfi(cs);
-   if (flushes & TU_CMD_FLAG_WAIT_FOR_ME)
-      tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
-}
+   const uint32_t max_pipe_count = 32; /* A6xx */
+   const uint32_t used_pipe_count =
+      tiling->pipe_count.width * tiling->pipe_count.height;
+   const VkExtent2D last_pipe = {
+      .width = tiling->tile_count.width % tiling->pipe0.width,
+      .height = tiling->tile_count.height % tiling->pipe0.height,
+   };
+
+   assert(used_pipe_count <= max_pipe_count);
+   assert(max_pipe_count <= ARRAY_SIZE(tiling->pipe_config));
+
+   for (uint32_t y = 0; y < tiling->pipe_count.height; y++) {
+      for (uint32_t x = 0; x < tiling->pipe_count.width; x++) {
+         const uint32_t pipe_x = tiling->pipe0.width * x;
+         const uint32_t pipe_y = tiling->pipe0.height * y;
+         const uint32_t pipe_w = (x == tiling->pipe_count.width - 1)
+                                    ? last_pipe.width
+                                    : tiling->pipe0.width;
+         const uint32_t pipe_h = (y == tiling->pipe_count.height - 1)
+                                    ? last_pipe.height
+                                    : tiling->pipe0.height;
+         const uint32_t n = tiling->pipe_count.width * y + x;
+
+         tiling->pipe_config[n] = A6XX_VSC_PIPE_CONFIG_REG_X(pipe_x) |
+                                  A6XX_VSC_PIPE_CONFIG_REG_Y(pipe_y) |
+                                  A6XX_VSC_PIPE_CONFIG_REG_W(pipe_w) |
+                                  A6XX_VSC_PIPE_CONFIG_REG_H(pipe_h);
+         tiling->pipe_sizes[n] = CP_SET_BIN_DATA5_0_VSC_SIZE(pipe_w * pipe_h);
+      }
+   }
 
-/* "Normal" cache flushes, that don't require any special handling */
+   memset(tiling->pipe_config + used_pipe_count, 0,
+          sizeof(uint32_t) * (max_pipe_count - used_pipe_count));
+}
 
 static void
-tu_emit_cache_flush(struct tu_cmd_buffer *cmd_buffer,
-                    struct tu_cs *cs)
-{
-   tu6_emit_flushes(cmd_buffer, cs, cmd_buffer->state.cache.flush_bits);
-   cmd_buffer->state.cache.flush_bits = 0;
+tu_tiling_config_update(struct tu_tiling_config *tiling,
+                        const struct tu_device *dev,
+                        const uint32_t *buffer_cpp,
+                        uint32_t buffer_count,
+                        const VkRect2D *render_area)
+{
+   /* see if there is any real change */
+   const bool ra_changed =
+      render_area &&
+      memcmp(&tiling->render_area, render_area, sizeof(*render_area));
+   const bool buf_changed = tiling->buffer_count != buffer_count ||
+                            memcmp(tiling->buffer_cpp, buffer_cpp,
+                                   sizeof(*buffer_cpp) * buffer_count);
+   if (!ra_changed && !buf_changed)
+      return;
+
+   if (ra_changed)
+      tiling->render_area = *render_area;
+
+   if (buf_changed) {
+      memcpy(tiling->buffer_cpp, buffer_cpp,
+             sizeof(*buffer_cpp) * buffer_count);
+      tiling->buffer_count = buffer_count;
+   }
+
+   tu_tiling_config_update_tile_layout(tiling, dev);
+   tu_tiling_config_update_pipe_layout(tiling, dev);
+   tu_tiling_config_update_pipes(tiling, dev);
 }
 
-/* Renderpass cache flushes */
+static void
+tu_tiling_config_get_tile(const struct tu_tiling_config *tiling,
+                          const struct tu_device *dev,
+                          uint32_t tx,
+                          uint32_t ty,
+                          struct tu_tile *tile)
+{
+   /* find the pipe and the slot for tile (tx, ty) */
+   const uint32_t px = tx / tiling->pipe0.width;
+   const uint32_t py = ty / tiling->pipe0.height;
+   const uint32_t sx = tx - tiling->pipe0.width * px;
+   const uint32_t sy = ty - tiling->pipe0.height * py;
+
+   assert(tx < tiling->tile_count.width && ty < tiling->tile_count.height);
+   assert(px < tiling->pipe_count.width && py < tiling->pipe_count.height);
+   assert(sx < tiling->pipe0.width && sy < tiling->pipe0.height);
+
+   /* convert to 1D indices */
+   tile->pipe = tiling->pipe_count.width * py + px;
+   tile->slot = tiling->pipe0.width * sy + sx;
+
+   /* get the blit area for the tile */
+   tile->begin = (VkOffset2D) {
+      .x = tiling->tile0.offset.x + tiling->tile0.extent.width * tx,
+      .y = tiling->tile0.offset.y + tiling->tile0.extent.height * ty,
+   };
+   tile->end.x =
+      (tx == tiling->tile_count.width - 1)
+         ? tiling->render_area.offset.x + tiling->render_area.extent.width
+         : tile->begin.x + tiling->tile0.extent.width;
+   tile->end.y =
+      (ty == tiling->tile_count.height - 1)
+         ? tiling->render_area.offset.y + tiling->render_area.extent.height
+         : tile->begin.y + tiling->tile0.extent.height;
+}
+
+static enum a3xx_msaa_samples
+tu6_msaa_samples(uint32_t samples)
+{
+   switch (samples) {
+   case 1:
+      return MSAA_ONE;
+   case 2:
+      return MSAA_TWO;
+   case 4:
+      return MSAA_FOUR;
+   case 8:
+      return MSAA_EIGHT;
+   default:
+      assert(!"invalid sample count");
+      return MSAA_ONE;
+   }
+}
 
-void
-tu_emit_cache_flush_renderpass(struct tu_cmd_buffer *cmd_buffer,
-                               struct tu_cs *cs)
+static enum a4xx_index_size
+tu6_index_size(VkIndexType type)
 {
-   if (!cmd_buffer->state.renderpass_cache.flush_bits &&
-       likely(!cmd_buffer->device->physical_device->instance->debug_flags))
-      return;
-   tu6_emit_flushes(cmd_buffer, cs, cmd_buffer->state.renderpass_cache.flush_bits);
-   cmd_buffer->state.renderpass_cache.flush_bits = 0;
+   switch (type) {
+   case VK_INDEX_TYPE_UINT16:
+      return INDEX4_SIZE_16_BIT;
+   case VK_INDEX_TYPE_UINT32:
+      return INDEX4_SIZE_32_BIT;
+   default:
+      unreachable("invalid VkIndexType");
+      return INDEX4_SIZE_8_BIT;
+   }
 }
 
-/* Cache flushes for things that use the color/depth read/write path (i.e.
- * blits and draws). This deals with changing CCU state as well as the usual
- * cache flushing.
- */
+static void
+tu6_emit_marker(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+   tu_cs_emit_write_reg(cs, cmd->marker_reg, ++cmd->marker_seqno);
+}
 
 void
-tu_emit_cache_flush_ccu(struct tu_cmd_buffer *cmd_buffer,
-                        struct tu_cs *cs,
-                        enum tu_cmd_ccu_state ccu_state)
+tu6_emit_event_write(struct tu_cmd_buffer *cmd,
+                     struct tu_cs *cs,
+                     enum vgt_event_type event,
+                     bool need_seqno)
 {
-   enum tu_cmd_flush_bits flushes = cmd_buffer->state.cache.flush_bits;
+   tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, need_seqno ? 4 : 1);
+   tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(event));
+   if (need_seqno) {
+      tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
+      tu_cs_emit(cs, ++cmd->scratch_seqno);
+   }
+}
 
-   assert(ccu_state != TU_CMD_CCU_UNKNOWN);
-   /* It's unsafe to flush inside condition because we clear flush_bits */
-   assert(!cs->cond_stack_depth);
+static void
+tu6_emit_cache_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+   tu6_emit_event_write(cmd, cs, 0x31, false);
+}
 
-   /* Changing CCU state must involve invalidating the CCU. In sysmem mode,
-    * the CCU may also contain data that we haven't flushed out yet, so we
-    * also need to flush. Also, in order to program RB_CCU_CNTL, we need to
-    * emit a WFI as it isn't pipelined.
-    */
-   if (ccu_state != cmd_buffer->state.ccu_state) {
-      if (cmd_buffer->state.ccu_state != TU_CMD_CCU_GMEM) {
-         flushes |=
-            TU_CMD_FLAG_CCU_FLUSH_COLOR |
-            TU_CMD_FLAG_CCU_FLUSH_DEPTH;
-         cmd_buffer->state.cache.pending_flush_bits &= ~(
-            TU_CMD_FLAG_CCU_FLUSH_COLOR |
-            TU_CMD_FLAG_CCU_FLUSH_DEPTH);
-      }
-      flushes |=
-         TU_CMD_FLAG_CCU_INVALIDATE_COLOR |
-         TU_CMD_FLAG_CCU_INVALIDATE_DEPTH |
-         TU_CMD_FLAG_WAIT_FOR_IDLE;
-      cmd_buffer->state.cache.pending_flush_bits &= ~(
-         TU_CMD_FLAG_CCU_INVALIDATE_COLOR |
-         TU_CMD_FLAG_CCU_INVALIDATE_DEPTH |
-         TU_CMD_FLAG_WAIT_FOR_IDLE);
-   }
+static void
+tu6_emit_lrz_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+   tu6_emit_event_write(cmd, cs, LRZ_FLUSH, false);
+}
 
-   tu6_emit_flushes(cmd_buffer, cs, flushes);
-   cmd_buffer->state.cache.flush_bits = 0;
-
-   if (ccu_state != cmd_buffer->state.ccu_state) {
-      struct tu_physical_device *phys_dev = cmd_buffer->device->physical_device;
-      tu_cs_emit_regs(cs,
-                      A6XX_RB_CCU_CNTL(.color_offset =
-                                          ccu_state == TU_CMD_CCU_GMEM ?
-                                          phys_dev->ccu_offset_gmem :
-                                          phys_dev->ccu_offset_bypass,
-                                       .gmem = ccu_state == TU_CMD_CCU_GMEM));
-      cmd_buffer->state.ccu_state = ccu_state;
+static void
+tu6_emit_wfi(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+   if (cmd->wait_for_idle) {
+      tu_cs_emit_wfi(cs);
+      cmd->wait_for_idle = false;
    }
 }
 
 static void
-tu6_emit_zs(struct tu_cmd_buffer *cmd,
-            const struct tu_subpass *subpass,
-            struct tu_cs *cs)
+tu6_emit_zs(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
+   const struct tu_subpass *subpass = cmd->state.subpass;
+
    const uint32_t a = subpass->depth_stencil_attachment.attachment;
    if (a == VK_ATTACHMENT_UNUSED) {
-      tu_cs_emit_regs(cs,
-                      A6XX_RB_DEPTH_BUFFER_INFO(.depth_format = DEPTH6_NONE),
-                      A6XX_RB_DEPTH_BUFFER_PITCH(0),
-                      A6XX_RB_DEPTH_BUFFER_ARRAY_PITCH(0),
-                      A6XX_RB_DEPTH_BUFFER_BASE(0),
-                      A6XX_RB_DEPTH_BUFFER_BASE_GMEM(0));
+      tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_BUFFER_INFO, 6);
+      tu_cs_emit(cs, A6XX_RB_DEPTH_BUFFER_INFO_DEPTH_FORMAT(DEPTH6_NONE));
+      tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_BUFFER_PITCH */
+      tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_BUFFER_ARRAY_PITCH */
+      tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_BUFFER_BASE_LO */
+      tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_BUFFER_BASE_HI */
+      tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_BUFFER_BASE_GMEM */
+
+      tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_DEPTH_BUFFER_INFO, 1);
+      tu_cs_emit(cs,
+                 A6XX_GRAS_SU_DEPTH_BUFFER_INFO_DEPTH_FORMAT(DEPTH6_NONE));
 
-      tu_cs_emit_regs(cs,
-                      A6XX_GRAS_SU_DEPTH_BUFFER_INFO(.depth_format = DEPTH6_NONE));
+      tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_LRZ_BUFFER_BASE_LO, 5);
+      tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_FLAG_BUFFER_BASE_LO */
+      tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_FLAG_BUFFER_BASE_HI */
+      tu_cs_emit(cs, 0x00000000); /* GRAS_LRZ_BUFFER_PITCH */
+      tu_cs_emit(cs, 0x00000000); /* GRAS_LRZ_FAST_CLEAR_BUFFER_BASE_LO */
+      tu_cs_emit(cs, 0x00000000); /* GRAS_LRZ_FAST_CLEAR_BUFFER_BASE_HI */
 
-      tu_cs_emit_regs(cs, A6XX_RB_STENCIL_INFO(0));
+      tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCIL_INFO, 1);
+      tu_cs_emit(cs, 0x00000000); /* RB_STENCIL_INFO */
 
       return;
    }
 
-   const struct tu_image_view *iview = cmd->state.attachments[a];
-   const struct tu_render_pass_attachment *attachment =
-      &cmd->state.pass->attachments[a];
-   enum a6xx_depth_format fmt = tu6_pipe2depth(attachment->format);
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_BUFFER_INFO, 6);
-   tu_cs_emit(cs, A6XX_RB_DEPTH_BUFFER_INFO(.depth_format = fmt).value);
-   if (attachment->format == VK_FORMAT_D32_SFLOAT_S8_UINT)
-      tu_cs_image_depth_ref(cs, iview, 0);
-   else
-      tu_cs_image_ref(cs, &iview->view, 0);
-   tu_cs_emit(cs, tu_attachment_gmem_offset(cmd, attachment));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_GRAS_SU_DEPTH_BUFFER_INFO(.depth_format = fmt));
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_FLAG_BUFFER_BASE, 3);
-   tu_cs_image_flag_ref(cs, &iview->view, 0);
-
-   if (attachment->format == VK_FORMAT_D32_SFLOAT_S8_UINT ||
-       attachment->format == VK_FORMAT_S8_UINT) {
-
-      tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCIL_INFO, 6);
-      tu_cs_emit(cs, A6XX_RB_STENCIL_INFO(.separate_stencil = true).value);
-      if (attachment->format == VK_FORMAT_D32_SFLOAT_S8_UINT) {
-         tu_cs_image_stencil_ref(cs, iview, 0);
-         tu_cs_emit(cs, tu_attachment_gmem_offset_stencil(cmd, attachment));
-      } else {
-         tu_cs_image_ref(cs, &iview->view, 0);
-         tu_cs_emit(cs, tu_attachment_gmem_offset(cmd, attachment));
-      }
-   } else {
-      tu_cs_emit_regs(cs,
-                     A6XX_RB_STENCIL_INFO(0));
-   }
+   /* enable zs? */
 }
 
 static void
-tu6_emit_mrt(struct tu_cmd_buffer *cmd,
-             const struct tu_subpass *subpass,
-             struct tu_cs *cs)
+tu6_emit_mrt(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
    const struct tu_framebuffer *fb = cmd->state.framebuffer;
+   const struct tu_subpass *subpass = cmd->state.subpass;
+   const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+   unsigned char mrt_comp[MAX_RTS] = { 0 };
+   unsigned srgb_cntl = 0;
 
-   enum a6xx_format mrt0_format = 0;
-
+   uint32_t gmem_index = 0;
    for (uint32_t i = 0; i < subpass->color_count; ++i) {
       uint32_t a = subpass->color_attachments[i].attachment;
-      if (a == VK_ATTACHMENT_UNUSED) {
-         /* From the VkPipelineRenderingCreateInfo definition:
-          *
-          *    Valid formats indicate that an attachment can be used - but it
-          *    is still valid to set the attachment to NULL when beginning
-          *    rendering.
-          *
-          * This means that with dynamic rendering, pipelines may write to
-          * some attachments that are UNUSED here. Setting the format to 0
-          * here should prevent them from writing to anything. This also seems
-          * to also be required for alpha-to-coverage which can use the alpha
-          * value for an otherwise-unused attachment.
-          */
-         tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_BUF_INFO(i), 6);
-         for (unsigned i = 0; i < 6; i++)
-            tu_cs_emit(cs, 0);
-
-         tu_cs_emit_regs(cs,
-                         A6XX_SP_FS_MRT_REG(i, .dword = 0));
+      if (a == VK_ATTACHMENT_UNUSED)
          continue;
-      }
 
-      const struct tu_image_view *iview = cmd->state.attachments[a];
+      const struct tu_image_view *iview = fb->attachments[a].attachment;
+      const struct tu_image_level *slice =
+         &iview->image->levels[iview->base_mip];
+      const enum a6xx_tile_mode tile_mode = TILE6_LINEAR;
+      uint32_t stride = 0;
+      uint32_t offset = 0;
 
-      tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_BUF_INFO(i), 6);
-      tu_cs_emit(cs, iview->view.RB_MRT_BUF_INFO);
-      tu_cs_image_ref(cs, &iview->view, 0);
-      tu_cs_emit(cs, tu_attachment_gmem_offset(cmd, &cmd->state.pass->attachments[a]));
+      mrt_comp[i] = 0xf;
 
-      tu_cs_emit_regs(cs,
-                      A6XX_SP_FS_MRT_REG(i, .dword = iview->view.SP_FS_MRT_REG));
+      if (vk_format_is_srgb(iview->vk_format))
+         srgb_cntl |= (1 << i);
 
-      tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_FLAG_BUFFER_ADDR(i), 3);
-      tu_cs_image_flag_ref(cs, &iview->view, 0);
+      const struct tu_native_format *format =
+         tu6_get_native_format(iview->vk_format);
+      assert(format && format->rb >= 0);
 
-      if (i == 0)
-         mrt0_format = iview->view.SP_FS_MRT_REG & 0xff;
-   }
+      offset = slice->offset + slice->size * iview->base_layer;
+      stride = slice->pitch * vk_format_get_blocksize(iview->vk_format);
+
+      tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_BUF_INFO(i), 6);
+      tu_cs_emit(cs, A6XX_RB_MRT_BUF_INFO_COLOR_FORMAT(format->rb) |
+                        A6XX_RB_MRT_BUF_INFO_COLOR_TILE_MODE(tile_mode) |
+                        A6XX_RB_MRT_BUF_INFO_COLOR_SWAP(format->swap));
+      tu_cs_emit(cs, A6XX_RB_MRT_PITCH(stride));
+      tu_cs_emit(cs, A6XX_RB_MRT_ARRAY_PITCH(slice->size));
+      tu_cs_emit_qw(cs, iview->image->bo->iova + iview->image->bo_offset +
+                           offset); /* BASE_LO/HI */
+      tu_cs_emit(
+         cs, tiling->gmem_offsets[gmem_index++]); /* RB_MRT[i].BASE_GMEM */
+
+      tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_MRT_REG(i), 1);
+      tu_cs_emit(cs, A6XX_SP_FS_MRT_REG_COLOR_FORMAT(format->rb));
+
+#if 0
+      /* when we support UBWC, these would be the system memory
+       * addr/pitch/etc:
+       */
+      tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_FLAG_BUFFER(i), 4);
+      tu_cs_emit(cs, 0x00000000);    /* RB_MRT_FLAG_BUFFER[i].ADDR_LO */
+      tu_cs_emit(cs, 0x00000000);    /* RB_MRT_FLAG_BUFFER[i].ADDR_HI */
+      tu_cs_emit(cs, A6XX_RB_MRT_FLAG_BUFFER_PITCH(0));
+      tu_cs_emit(cs, A6XX_RB_MRT_FLAG_BUFFER_ARRAY_PITCH(0));
+#endif
+   }
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_SRGB_CNTL, 1);
+   tu_cs_emit(cs, srgb_cntl);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_SRGB_CNTL, 1);
+   tu_cs_emit(cs, srgb_cntl);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_RENDER_COMPONENTS, 1);
+   tu_cs_emit(cs, A6XX_RB_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
+                     A6XX_RB_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
+                     A6XX_RB_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
+                     A6XX_RB_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
+                     A6XX_RB_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
+                     A6XX_RB_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
+                     A6XX_RB_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
+                     A6XX_RB_RENDER_COMPONENTS_RT7(mrt_comp[7]));
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_RENDER_COMPONENTS, 1);
+   tu_cs_emit(cs, A6XX_SP_FS_RENDER_COMPONENTS_RT0(mrt_comp[0]) |
+                     A6XX_SP_FS_RENDER_COMPONENTS_RT1(mrt_comp[1]) |
+                     A6XX_SP_FS_RENDER_COMPONENTS_RT2(mrt_comp[2]) |
+                     A6XX_SP_FS_RENDER_COMPONENTS_RT3(mrt_comp[3]) |
+                     A6XX_SP_FS_RENDER_COMPONENTS_RT4(mrt_comp[4]) |
+                     A6XX_SP_FS_RENDER_COMPONENTS_RT5(mrt_comp[5]) |
+                     A6XX_SP_FS_RENDER_COMPONENTS_RT6(mrt_comp[6]) |
+                     A6XX_SP_FS_RENDER_COMPONENTS_RT7(mrt_comp[7]));
+}
 
-   tu_cs_emit_regs(cs, A6XX_GRAS_LRZ_MRT_BUF_INFO_0(.color_format = mrt0_format));
+static void
+tu6_emit_msaa(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+   const struct tu_subpass *subpass = cmd->state.subpass;
+   const enum a3xx_msaa_samples samples =
+      tu6_msaa_samples(subpass->max_sample_count);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_RAS_MSAA_CNTL, 2);
+   tu_cs_emit(cs, A6XX_SP_TP_RAS_MSAA_CNTL_SAMPLES(samples));
+   tu_cs_emit(
+      cs, A6XX_SP_TP_DEST_MSAA_CNTL_SAMPLES(samples) |
+             ((samples == MSAA_ONE) ? A6XX_SP_TP_DEST_MSAA_CNTL_MSAA_DISABLE
+                                    : 0));
 
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_SRGB_CNTL(.dword = subpass->srgb_cntl));
-   tu_cs_emit_regs(cs,
-                   A6XX_SP_SRGB_CNTL(.dword = subpass->srgb_cntl));
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_RAS_MSAA_CNTL, 2);
+   tu_cs_emit(cs, A6XX_GRAS_RAS_MSAA_CNTL_SAMPLES(samples));
+   tu_cs_emit(
+      cs,
+      A6XX_GRAS_DEST_MSAA_CNTL_SAMPLES(samples) |
+         ((samples == MSAA_ONE) ? A6XX_GRAS_DEST_MSAA_CNTL_MSAA_DISABLE : 0));
 
-   unsigned layers = MAX2(fb->layers, util_logbase2(subpass->multiview_mask) + 1);
-   tu_cs_emit_regs(cs, A6XX_GRAS_MAX_LAYER_INDEX(layers - 1));
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_RAS_MSAA_CNTL, 2);
+   tu_cs_emit(cs, A6XX_RB_RAS_MSAA_CNTL_SAMPLES(samples));
+   tu_cs_emit(
+      cs,
+      A6XX_RB_DEST_MSAA_CNTL_SAMPLES(samples) |
+         ((samples == MSAA_ONE) ? A6XX_RB_DEST_MSAA_CNTL_MSAA_DISABLE : 0));
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_MSAA_CNTL, 1);
+   tu_cs_emit(cs, A6XX_RB_MSAA_CNTL_SAMPLES(samples));
 }
 
 static void
-tu6_emit_bin_size(struct tu_cs *cs,
-                  uint32_t bin_w, uint32_t bin_h, uint32_t flags)
+tu6_emit_bin_size(struct tu_cmd_buffer *cmd, struct tu_cs *cs, uint32_t flags)
 {
-   tu_cs_emit_regs(cs,
-                   A6XX_GRAS_BIN_CONTROL(.binw = bin_w,
-                                         .binh = bin_h,
-                                         .dword = flags));
+   const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+   const uint32_t bin_w = tiling->tile0.extent.width;
+   const uint32_t bin_h = tiling->tile0.extent.height;
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_BIN_CONTROL, 1);
+   tu_cs_emit(cs, A6XX_GRAS_BIN_CONTROL_BINW(bin_w) |
+                     A6XX_GRAS_BIN_CONTROL_BINH(bin_h) | flags);
 
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_BIN_CONTROL(.binw = bin_w,
-                                       .binh = bin_h,
-                                       .dword = flags));
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BIN_CONTROL, 1);
+   tu_cs_emit(cs, A6XX_RB_BIN_CONTROL_BINW(bin_w) |
+                     A6XX_RB_BIN_CONTROL_BINH(bin_h) | flags);
 
    /* no flag for RB_BIN_CONTROL2... */
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_BIN_CONTROL2(.binw = bin_w,
-                                        .binh = bin_h));
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BIN_CONTROL2, 1);
+   tu_cs_emit(cs, A6XX_RB_BIN_CONTROL2_BINW(bin_w) |
+                     A6XX_RB_BIN_CONTROL2_BINH(bin_h));
 }
 
 static void
 tu6_emit_render_cntl(struct tu_cmd_buffer *cmd,
-                     const struct tu_subpass *subpass,
                      struct tu_cs *cs,
                      bool binning)
 {
-   /* doesn't RB_RENDER_CNTL set differently for binning pass: */
-   bool no_track = !cmd->device->physical_device->info->a6xx.has_cp_reg_write;
    uint32_t cntl = 0;
-   cntl |= A6XX_RB_RENDER_CNTL_CCUSINGLECACHELINESIZE(2);
-   if (binning) {
-      if (no_track)
-         return;
+   cntl |= A6XX_RB_RENDER_CNTL_UNK4;
+   if (binning)
       cntl |= A6XX_RB_RENDER_CNTL_BINNING;
-   } else {
-      uint32_t mrts_ubwc_enable = 0;
-      for (uint32_t i = 0; i < subpass->color_count; ++i) {
-         uint32_t a = subpass->color_attachments[i].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-
-         const struct tu_image_view *iview = cmd->state.attachments[a];
-         if (iview->view.ubwc_enabled)
-            mrts_ubwc_enable |= 1 << i;
-      }
-
-      cntl |= A6XX_RB_RENDER_CNTL_FLAG_MRTS(mrts_ubwc_enable);
-
-      const uint32_t a = subpass->depth_stencil_attachment.attachment;
-      if (a != VK_ATTACHMENT_UNUSED) {
-         const struct tu_image_view *iview = cmd->state.attachments[a];
-         if (iview->view.ubwc_enabled)
-            cntl |= A6XX_RB_RENDER_CNTL_FLAG_DEPTH;
-      }
-
-      if (no_track) {
-         tu_cs_emit_pkt4(cs, REG_A6XX_RB_RENDER_CNTL, 1);
-         tu_cs_emit(cs, cntl);
-         return;
-      }
-
-      /* In the !binning case, we need to set RB_RENDER_CNTL in the draw_cs
-       * in order to set it correctly for the different subpasses. However,
-       * that means the packets we're emitting also happen during binning. So
-       * we need to guard the write on !BINNING at CP execution time.
-       */
-      tu_cs_reserve(cs, 3 + 4);
-      tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
-      tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
-                     CP_COND_REG_EXEC_0_GMEM | CP_COND_REG_EXEC_0_SYSMEM);
-      tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(4));
-   }
 
    tu_cs_emit_pkt7(cs, CP_REG_WRITE, 3);
-   tu_cs_emit(cs, CP_REG_WRITE_0_TRACKER(TRACK_RENDER_CNTL));
+   tu_cs_emit(cs, 0x2);
    tu_cs_emit(cs, REG_A6XX_RB_RENDER_CNTL);
    tu_cs_emit(cs, cntl);
 }
 
 static void
-tu6_emit_blit_scissor(struct tu_cmd_buffer *cmd, struct tu_cs *cs, bool align)
+tu6_emit_blit_scissor(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
-   struct tu_physical_device *phys_dev = cmd->device->physical_device;
-   const VkRect2D *render_area = &cmd->state.render_area;
-
-   /* Avoid assertion fails with an empty render area at (0, 0) where the
-    * subtraction below wraps around. Empty render areas should be forced to
-    * the sysmem path by use_sysmem_rendering(). It's not even clear whether
-    * an empty scissor here works, and the blob seems to force sysmem too as
-    * it sets something wrong (non-empty) for the scissor.
-    */
-   if (render_area->extent.width == 0 ||
-       render_area->extent.height == 0)
-      return;
-
-   uint32_t x1 = render_area->offset.x;
-   uint32_t y1 = render_area->offset.y;
-   uint32_t x2 = x1 + render_area->extent.width - 1;
-   uint32_t y2 = y1 + render_area->extent.height - 1;
-
-   if (align) {
-      x1 = x1 & ~(phys_dev->info->gmem_align_w - 1);
-      y1 = y1 & ~(phys_dev->info->gmem_align_h - 1);
-      x2 = ALIGN_POT(x2 + 1, phys_dev->info->gmem_align_w) - 1;
-      y2 = ALIGN_POT(y2 + 1, phys_dev->info->gmem_align_h) - 1;
-   }
+   const VkRect2D *render_area = &cmd->state.tiling_config.render_area;
+   const uint32_t x1 = render_area->offset.x;
+   const uint32_t y1 = render_area->offset.y;
+   const uint32_t x2 = x1 + render_area->extent.width - 1;
+   const uint32_t y2 = y1 + render_area->extent.height - 1;
 
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_BLIT_SCISSOR_TL(.x = x1, .y = y1),
-                   A6XX_RB_BLIT_SCISSOR_BR(.x = x2, .y = y2));
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_SCISSOR_TL, 2);
+   tu_cs_emit(cs,
+              A6XX_RB_BLIT_SCISSOR_TL_X(x1) | A6XX_RB_BLIT_SCISSOR_TL_Y(y1));
+   tu_cs_emit(cs,
+              A6XX_RB_BLIT_SCISSOR_BR_X(x2) | A6XX_RB_BLIT_SCISSOR_BR_Y(y2));
 }
 
-void
-tu6_emit_window_scissor(struct tu_cs *cs,
-                        uint32_t x1,
-                        uint32_t y1,
-                        uint32_t x2,
-                        uint32_t y2)
-{
-   tu_cs_emit_regs(cs,
-                   A6XX_GRAS_SC_WINDOW_SCISSOR_TL(.x = x1, .y = y1),
-                   A6XX_GRAS_SC_WINDOW_SCISSOR_BR(.x = x2, .y = y2));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_GRAS_2D_RESOLVE_CNTL_1(.x = x1, .y = y1),
-                   A6XX_GRAS_2D_RESOLVE_CNTL_2(.x = x2, .y = y2));
+static void
+tu6_emit_blit_info(struct tu_cmd_buffer *cmd,
+                   struct tu_cs *cs,
+                   const struct tu_image_view *iview,
+                   uint32_t gmem_offset,
+                   uint32_t blit_info)
+{
+   const struct tu_image_level *slice =
+      &iview->image->levels[iview->base_mip];
+   const uint32_t offset = slice->offset + slice->size * iview->base_layer;
+   const uint32_t stride =
+      slice->pitch * vk_format_get_blocksize(iview->vk_format);
+   const enum a6xx_tile_mode tile_mode = TILE6_LINEAR;
+   const enum a3xx_msaa_samples samples = tu6_msaa_samples(1);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_INFO, 1);
+   tu_cs_emit(cs, blit_info);
+
+   /* tile mode? */
+   const struct tu_native_format *format =
+      tu6_get_native_format(iview->vk_format);
+   assert(format && format->rb >= 0);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_DST_INFO, 5);
+   tu_cs_emit(cs, A6XX_RB_BLIT_DST_INFO_TILE_MODE(tile_mode) |
+                     A6XX_RB_BLIT_DST_INFO_SAMPLES(samples) |
+                     A6XX_RB_BLIT_DST_INFO_COLOR_FORMAT(format->rb) |
+                     A6XX_RB_BLIT_DST_INFO_COLOR_SWAP(format->swap));
+   tu_cs_emit_qw(cs,
+                 iview->image->bo->iova + iview->image->bo_offset + offset);
+   tu_cs_emit(cs, A6XX_RB_BLIT_DST_PITCH(stride));
+   tu_cs_emit(cs, A6XX_RB_BLIT_DST_ARRAY_PITCH(slice->size));
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_BASE_GMEM, 1);
+   tu_cs_emit(cs, gmem_offset);
 }
 
-void
-tu6_emit_window_offset(struct tu_cs *cs, uint32_t x1, uint32_t y1)
-{
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_WINDOW_OFFSET(.x = x1, .y = y1));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_WINDOW_OFFSET2(.x = x1, .y = y1));
+static void
+tu6_emit_blit_clear(struct tu_cmd_buffer *cmd,
+                    struct tu_cs *cs,
+                    const struct tu_image_view *iview,
+                    uint32_t gmem_offset,
+                    const VkClearValue *clear_value)
+{
+   const enum a6xx_tile_mode tile_mode = TILE6_LINEAR;
+   const enum a3xx_msaa_samples samples = tu6_msaa_samples(1);
+
+   const struct tu_native_format *format =
+      tu6_get_native_format(iview->vk_format);
+   assert(format && format->rb >= 0);
+   /* must be WZYX; other values are ignored */
+   const enum a3xx_color_swap swap = WZYX;
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_DST_INFO, 1);
+   tu_cs_emit(cs, A6XX_RB_BLIT_DST_INFO_TILE_MODE(tile_mode) |
+                     A6XX_RB_BLIT_DST_INFO_SAMPLES(samples) |
+                     A6XX_RB_BLIT_DST_INFO_COLOR_FORMAT(format->rb) |
+                     A6XX_RB_BLIT_DST_INFO_COLOR_SWAP(swap));
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_INFO, 1);
+   tu_cs_emit(cs, A6XX_RB_BLIT_INFO_GMEM | A6XX_RB_BLIT_INFO_CLEAR_MASK(0xf));
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_BASE_GMEM, 1);
+   tu_cs_emit(cs, gmem_offset);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_UNKNOWN_88D0, 1);
+   tu_cs_emit(cs, 0);
 
-   tu_cs_emit_regs(cs,
-                   A6XX_SP_WINDOW_OFFSET(.x = x1, .y = y1));
+   /* pack clear_value into WZYX order */
+   uint32_t clear_vals[4] = { 0 };
+   tu_pack_clear_value(clear_value, iview->vk_format, clear_vals);
 
-   tu_cs_emit_regs(cs,
-                   A6XX_SP_TP_WINDOW_OFFSET(.x = x1, .y = y1));
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_CLEAR_COLOR_DW0, 4);
+   tu_cs_emit(cs, clear_vals[0]);
+   tu_cs_emit(cs, clear_vals[1]);
+   tu_cs_emit(cs, clear_vals[2]);
+   tu_cs_emit(cs, clear_vals[3]);
 }
 
-void
-tu6_apply_depth_bounds_workaround(struct tu_device *device,
-                                  uint32_t *rb_depth_cntl)
+static void
+tu6_emit_blit(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
-   if (!device->physical_device->info->a6xx.depth_bounds_require_depth_test_quirk)
-      return;
-
-   /* On some GPUs it is necessary to enable z test for depth bounds test when
-    * UBWC is enabled. Otherwise, the GPU would hang. FUNC_ALWAYS is required to
-    * pass z test. Relevant tests:
-    *  dEQP-VK.pipeline.extended_dynamic_state.two_draws_dynamic.depth_bounds_test_disable
-    *  dEQP-VK.dynamic_state.ds_state.depth_bounds_1
-    */
-   *rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE |
-                     A6XX_RB_DEPTH_CNTL_ZFUNC(FUNC_ALWAYS);
+   tu6_emit_marker(cmd, cs);
+   tu6_emit_event_write(cmd, cs, BLIT, false);
+   tu6_emit_marker(cmd, cs);
 }
 
 static void
-tu_cs_emit_draw_state(struct tu_cs *cs, uint32_t id, struct tu_draw_state state)
+tu6_emit_window_scissor(struct tu_cmd_buffer *cmd,
+                        struct tu_cs *cs,
+                        uint32_t x1,
+                        uint32_t y1,
+                        uint32_t x2,
+                        uint32_t y2)
 {
-   uint32_t enable_mask;
-   switch (id) {
-   case TU_DRAW_STATE_PROGRAM:
-   /* The blob seems to not enable this (DESC_SETS_LOAD) for binning, even
-    * when resources would actually be used in the binning shader.
-    * Presumably the overhead of prefetching the resources isn't
-    * worth it.
-    */
-   case TU_DRAW_STATE_DESC_SETS_LOAD:
-      enable_mask = CP_SET_DRAW_STATE__0_GMEM |
-                    CP_SET_DRAW_STATE__0_SYSMEM;
-      break;
-   case TU_DRAW_STATE_PROGRAM_BINNING:
-      enable_mask = CP_SET_DRAW_STATE__0_BINNING;
-      break;
-   case TU_DRAW_STATE_INPUT_ATTACHMENTS_GMEM:
-   case TU_DRAW_STATE_PRIM_MODE_GMEM:
-      enable_mask = CP_SET_DRAW_STATE__0_GMEM;
-      break;
-   case TU_DRAW_STATE_INPUT_ATTACHMENTS_SYSMEM:
-   case TU_DRAW_STATE_PRIM_MODE_SYSMEM:
-      enable_mask = CP_SET_DRAW_STATE__0_SYSMEM;
-      break;
-   default:
-      enable_mask = CP_SET_DRAW_STATE__0_GMEM |
-                    CP_SET_DRAW_STATE__0_SYSMEM |
-                    CP_SET_DRAW_STATE__0_BINNING;
-      break;
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SC_WINDOW_SCISSOR_TL, 2);
+   tu_cs_emit(cs, A6XX_GRAS_SC_WINDOW_SCISSOR_TL_X(x1) |
+                     A6XX_GRAS_SC_WINDOW_SCISSOR_TL_Y(y1));
+   tu_cs_emit(cs, A6XX_GRAS_SC_WINDOW_SCISSOR_BR_X(x2) |
+                     A6XX_GRAS_SC_WINDOW_SCISSOR_BR_Y(y2));
 
-   STATIC_ASSERT(TU_DRAW_STATE_COUNT <= 32);
-
-   /* We need to reload the descriptors every time the descriptor sets
-    * change. However, the commands we send only depend on the pipeline
-    * because the whole point is to cache descriptors which are used by the
-    * pipeline. There's a problem here, in that the firmware has an
-    * "optimization" which skips executing groups that are set to the same
-    * value as the last draw. This means that if the descriptor sets change
-    * but not the pipeline, we'd try to re-execute the same buffer which
-    * the firmware would ignore and we wouldn't pre-load the new
-    * descriptors. Set the DIRTY bit to avoid this optimization
-    */
-   if (id == TU_DRAW_STATE_DESC_SETS_LOAD)
-      enable_mask |= CP_SET_DRAW_STATE__0_DIRTY;
-
-   tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(state.size) |
-                  enable_mask |
-                  CP_SET_DRAW_STATE__0_GROUP_ID(id) |
-                  COND(!state.size || !state.iova, CP_SET_DRAW_STATE__0_DISABLE));
-   tu_cs_emit_qw(cs, state.iova);
-}
-
-void
-tu6_emit_msaa(struct tu_cs *cs, VkSampleCountFlagBits vk_samples,
-              bool msaa_disable)
-{
-   const enum a3xx_msaa_samples samples = tu_msaa_samples(vk_samples);
-   msaa_disable |= (samples == MSAA_ONE);
-   tu_cs_emit_regs(cs,
-                   A6XX_SP_TP_RAS_MSAA_CNTL(samples),
-                   A6XX_SP_TP_DEST_MSAA_CNTL(.samples = samples,
-                                             .msaa_disable = msaa_disable));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_GRAS_RAS_MSAA_CNTL(samples),
-                   A6XX_GRAS_DEST_MSAA_CNTL(.samples = samples,
-                                            .msaa_disable = msaa_disable));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_RAS_MSAA_CNTL(samples),
-                   A6XX_RB_DEST_MSAA_CNTL(.samples = samples,
-                                          .msaa_disable = msaa_disable));
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_RESOLVE_CNTL_1, 2);
+   tu_cs_emit(
+      cs, A6XX_GRAS_RESOLVE_CNTL_1_X(x1) | A6XX_GRAS_RESOLVE_CNTL_1_Y(y1));
+   tu_cs_emit(
+      cs, A6XX_GRAS_RESOLVE_CNTL_2_X(x2) | A6XX_GRAS_RESOLVE_CNTL_2_Y(y2));
 }
 
 static void
-tu6_update_msaa(struct tu_cmd_buffer *cmd, VkSampleCountFlagBits samples)
-{
-   bool is_line =
-      tu6_primtype_line(cmd->state.primtype) ||
-      (tu6_primtype_patches(cmd->state.primtype) &&
-       cmd->state.pipeline &&
-       cmd->state.pipeline->tess.patch_type == IR3_TESS_ISOLINES);
-   bool msaa_disable = is_line && cmd->state.line_mode == BRESENHAM;
-
-   if (cmd->state.msaa_disable != msaa_disable ||
-       cmd->state.samples != samples) {
-      struct tu_cs cs;
-      cmd->state.msaa = tu_cs_draw_state(&cmd->sub_cs, &cs, 9);
-      tu6_emit_msaa(&cs, samples, msaa_disable);
-      if (!(cmd->state.dirty & TU_CMD_DIRTY_DRAW_STATE)) {
-         tu_cs_emit_pkt7(&cmd->draw_cs, CP_SET_DRAW_STATE, 3);
-         tu_cs_emit_draw_state(&cmd->draw_cs, TU_DRAW_STATE_MSAA, cmd->state.msaa);
-      }
-      cmd->state.msaa_disable = msaa_disable;
-      cmd->state.samples = samples;
-   }
-}
-
-static bool
-use_hw_binning(struct tu_cmd_buffer *cmd)
+tu6_emit_window_offset(struct tu_cmd_buffer *cmd,
+                       struct tu_cs *cs,
+                       uint32_t x1,
+                       uint32_t y1)
 {
-   const struct tu_framebuffer *fb = cmd->state.framebuffer;
-   const struct tu_tiling_config *tiling = &fb->tiling[cmd->state.gmem_layout];
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_WINDOW_OFFSET, 1);
+   tu_cs_emit(cs, A6XX_RB_WINDOW_OFFSET_X(x1) | A6XX_RB_WINDOW_OFFSET_Y(y1));
 
-   /* XFB commands are emitted for BINNING || SYSMEM, which makes it
-    * incompatible with non-hw binning GMEM rendering. this is required because
-    * some of the XFB commands need to only be executed once.
-    * use_sysmem_rendering() should have made sure we only ended up here if no
-    * XFB was used.
-    */
-   if (cmd->state.rp.xfb_used) {
-      assert(tiling->binning_possible);
-      return true;
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_WINDOW_OFFSET2, 1);
+   tu_cs_emit(cs,
+              A6XX_RB_WINDOW_OFFSET2_X(x1) | A6XX_RB_WINDOW_OFFSET2_Y(y1));
 
-   /* VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT emulates GL_PRIMITIVES_GENERATED,
-    * which wasn't designed to care about tilers and expects the result not to
-    * be multiplied by tile count.
-    * See https://gitlab.khronos.org/vulkan/vulkan/-/issues/3131
-    */
-   if (cmd->state.rp.has_prim_generated_query_in_rp ||
-       cmd->state.prim_generated_query_running_before_rp) {
-      assert(tiling->binning_possible);
-      return true;
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_WINDOW_OFFSET, 1);
+   tu_cs_emit(cs, A6XX_SP_WINDOW_OFFSET_X(x1) | A6XX_SP_WINDOW_OFFSET_Y(y1));
 
-   return tiling->binning;
-}
-
-static bool
-use_sysmem_rendering(struct tu_cmd_buffer *cmd,
-                     struct tu_renderpass_result **autotune_result)
-{
-   if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_SYSMEM))
-      return true;
-
-   /* can't fit attachments into gmem */
-   if (!cmd->state.pass->gmem_pixels[cmd->state.gmem_layout])
-      return true;
-
-   if (cmd->state.framebuffer->layers > 1)
-      return true;
-
-   /* Use sysmem for empty render areas */
-   if (cmd->state.render_area.extent.width == 0 ||
-       cmd->state.render_area.extent.height == 0)
-      return true;
-
-   if (cmd->state.rp.has_tess)
-      return true;
-
-   if (cmd->state.rp.disable_gmem)
-      return true;
-
-   /* XFB is incompatible with non-hw binning GMEM rendering, see use_hw_binning */
-   if (cmd->state.rp.xfb_used && !cmd->state.tiling->binning_possible)
-      return true;
-
-   /* QUERY_TYPE_PRIMITIVES_GENERATED is incompatible with non-hw binning
-    * GMEM rendering, see use_hw_binning.
-    */
-   if ((cmd->state.rp.has_prim_generated_query_in_rp ||
-        cmd->state.prim_generated_query_running_before_rp) &&
-       !cmd->state.tiling->binning_possible)
-      return true;
-
-   if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_GMEM))
-      return false;
-
-   bool use_sysmem = tu_autotune_use_bypass(&cmd->device->autotune,
-                                            cmd, autotune_result);
-   if (*autotune_result) {
-      list_addtail(&(*autotune_result)->node, &cmd->renderpass_autotune_results);
-   }
-
-   return use_sysmem;
-}
-
-/* Optimization: there is no reason to load gmem if there is no
- * geometry to process. COND_REG_EXEC predicate is set here,
- * but the actual skip happens in tu6_emit_tile_load() and tile_store_cs,
- * for each blit separately.
- */
-static void
-tu6_emit_cond_for_load_stores(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-                              uint32_t pipe, uint32_t slot, bool wfm)
-{
-   if (cmd->state.tiling->binning_possible) {
-      tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
-      tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(REG_A6XX_VSC_STATE_REG(pipe)) |
-                     A6XX_CP_REG_TEST_0_BIT(slot) |
-                     COND(wfm, A6XX_CP_REG_TEST_0_WAIT_FOR_ME));
-   } else {
-      /* COND_REG_EXECs are not emitted in non-binning case */
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_WINDOW_OFFSET, 1);
+   tu_cs_emit(
+      cs, A6XX_SP_TP_WINDOW_OFFSET_X(x1) | A6XX_SP_TP_WINDOW_OFFSET_Y(y1));
 }
 
 static void
 tu6_emit_tile_select(struct tu_cmd_buffer *cmd,
                      struct tu_cs *cs,
-                     uint32_t tx, uint32_t ty, uint32_t pipe, uint32_t slot)
+                     const struct tu_tile *tile)
 {
-   const struct tu_tiling_config *tiling = cmd->state.tiling;
+   tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
+   tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(0x7));
 
+   tu6_emit_marker(cmd, cs);
    tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
-   tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_GMEM));
+   tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_GMEM) | 0x10);
+   tu6_emit_marker(cmd, cs);
 
-   const uint32_t x1 = tiling->tile0.width * tx;
-   const uint32_t y1 = tiling->tile0.height * ty;
-   const uint32_t x2 = MIN2(x1 + tiling->tile0.width - 1, MAX_VIEWPORT_SIZE - 1);
-   const uint32_t y2 = MIN2(y1 + tiling->tile0.height - 1, MAX_VIEWPORT_SIZE - 1);
-   tu6_emit_window_scissor(cs, x1, y1, x2, y2);
-   tu6_emit_window_offset(cs, x1, y1);
+   const uint32_t x1 = tile->begin.x;
+   const uint32_t y1 = tile->begin.y;
+   const uint32_t x2 = tile->end.x - 1;
+   const uint32_t y2 = tile->end.y - 1;
+   tu6_emit_window_scissor(cmd, cs, x1, y1, x2, y2);
+   tu6_emit_window_offset(cmd, cs, x1, y1);
 
-   bool hw_binning = use_hw_binning(cmd);
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_OVERRIDE, 1);
+   tu_cs_emit(cs, A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
 
-   if (hw_binning) {
-      tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
+   if (false) {
+      /* hw binning? */
+   } else {
+      tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
+      tu_cs_emit(cs, 0x1);
 
       tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
       tu_cs_emit(cs, 0x0);
-
-      tu_cs_emit_pkt7(cs, CP_SET_BIN_DATA5_OFFSET, 4);
-      tu_cs_emit(cs, tiling->pipe_sizes[pipe] |
-                     CP_SET_BIN_DATA5_0_VSC_N(slot));
-      tu_cs_emit(cs, pipe * cmd->vsc_draw_strm_pitch);
-      tu_cs_emit(cs, pipe * 4);
-      tu_cs_emit(cs, pipe * cmd->vsc_prim_strm_pitch);
    }
-
-   tu6_emit_cond_for_load_stores(cmd, cs, pipe, slot, hw_binning);
-
-   tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
-   tu_cs_emit(cs, !hw_binning);
-
-   tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
-   tu_cs_emit(cs, 0x0);
 }
 
 static void
-tu6_emit_sysmem_resolve(struct tu_cmd_buffer *cmd,
-                        struct tu_cs *cs,
-                        uint32_t layer_mask,
-                        uint32_t a,
-                        uint32_t gmem_a)
+tu6_emit_tile_load(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
    const struct tu_framebuffer *fb = cmd->state.framebuffer;
-   const struct tu_image_view *dst = cmd->state.attachments[a];
-   const struct tu_image_view *src = cmd->state.attachments[gmem_a];
-
-   tu_resolve_sysmem(cmd, cs, src, dst, layer_mask, fb->layers, &cmd->state.render_area);
-}
-
-static void
-tu6_emit_sysmem_resolves(struct tu_cmd_buffer *cmd,
-                         struct tu_cs *cs,
-                         const struct tu_subpass *subpass)
-{
-   if (subpass->resolve_attachments) {
-      /* From the documentation for vkCmdNextSubpass, section 7.4 "Render Pass
-       * Commands":
-       *
-       *    End-of-subpass multisample resolves are treated as color
-       *    attachment writes for the purposes of synchronization.
-       *    This applies to resolve operations for both color and
-       *    depth/stencil attachments. That is, they are considered to
-       *    execute in the VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT
-       *    pipeline stage and their writes are synchronized with
-       *    VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT. Synchronization between
-       *    rendering within a subpass and any resolve operations at the end
-       *    of the subpass occurs automatically, without need for explicit
-       *    dependencies or pipeline barriers. However, if the resolve
-       *    attachment is also used in a different subpass, an explicit
-       *    dependency is needed.
-       *
-       * We use the CP_BLIT path for sysmem resolves, which is really a
-       * transfer command, so we have to manually flush similar to the gmem
-       * resolve case. However, a flush afterwards isn't needed because of the
-       * last sentence and the fact that we're in sysmem mode.
-       */
-      tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS);
-      if (subpass->resolve_depth_stencil)
-         tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_DEPTH_TS);
+   const struct tu_subpass *subpass = cmd->state.subpass;
+   const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+   const struct tu_attachment_state *attachments = cmd->state.attachments;
 
-      tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE);
+   tu6_emit_blit_scissor(cmd, cs);
 
-      /* Wait for the flushes to land before using the 2D engine */
-      tu_cs_emit_wfi(cs);
-
-      for (unsigned i = 0; i < subpass->resolve_count; i++) {
-         uint32_t a = subpass->resolve_attachments[i].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-
-         uint32_t gmem_a = tu_subpass_get_attachment_to_resolve(subpass, i);
+   uint32_t gmem_index = 0;
+   for (uint32_t i = 0; i < subpass->color_count; ++i) {
+      const uint32_t a = subpass->color_attachments[i].attachment;
+      if (a == VK_ATTACHMENT_UNUSED)
+         continue;
 
-         tu6_emit_sysmem_resolve(cmd, cs, subpass->multiview_mask, a, gmem_a);
+      const struct tu_image_view *iview = fb->attachments[a].attachment;
+      const struct tu_attachment_state *att = attachments + a;
+      if (att->pending_clear_aspects) {
+         assert(att->pending_clear_aspects == VK_IMAGE_ASPECT_COLOR_BIT);
+         tu6_emit_blit_clear(cmd, cs, iview,
+                             tiling->gmem_offsets[gmem_index++],
+                             &att->clear_value);
+      } else {
+         tu6_emit_blit_info(cmd, cs, iview,
+                            tiling->gmem_offsets[gmem_index++],
+                            A6XX_RB_BLIT_INFO_UNK0 | A6XX_RB_BLIT_INFO_GMEM);
       }
-   }
-}
 
-static void
-tu6_emit_tile_load(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
-   tu6_emit_blit_scissor(cmd, cs, true);
+      tu6_emit_blit(cmd, cs);
+   }
 
-   for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
-      tu_load_gmem_attachment(cmd, cs, i, cmd->state.tiling->binning, false);
+   /* load/clear zs? */
 }
 
 static void
 tu6_emit_tile_store(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
-   const struct tu_render_pass *pass = cmd->state.pass;
-   const struct tu_subpass *subpass = &pass->subpasses[pass->subpass_count-1];
-
-   tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
-   tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_RESOLVE));
-
-   tu6_emit_blit_scissor(cmd, cs, true);
+   const struct tu_framebuffer *fb = cmd->state.framebuffer;
+   const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
 
-   for (uint32_t a = 0; a < pass->attachment_count; ++a) {
-      if (pass->attachments[a].gmem)
-         tu_store_gmem_attachment(cmd, cs, a, a, cmd->state.tiling->binning_possible);
+   if (false) {
+      /* hw binning? */
    }
 
-   if (subpass->resolve_attachments) {
-      for (unsigned i = 0; i < subpass->resolve_count; i++) {
-         uint32_t a = subpass->resolve_attachments[i].attachment;
-         if (a != VK_ATTACHMENT_UNUSED) {
-            uint32_t gmem_a = tu_subpass_get_attachment_to_resolve(subpass, i);
-            tu_store_gmem_attachment(cmd, cs, a, gmem_a, false);
-         }
-      }
-   }
-}
-
-void
-tu_disable_draw_states(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
    tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
    tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(0) |
                      CP_SET_DRAW_STATE__0_DISABLE_ALL_GROUPS |
@@ -881,80 +809,84 @@ tu_disable_draw_states(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
    tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
    tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
 
-   cmd->state.dirty |= TU_CMD_DIRTY_DRAW_STATE;
+   tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
+   tu_cs_emit(cs, 0x0);
+
+   tu6_emit_marker(cmd, cs);
+   tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
+   tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_RESOLVE) | 0x10);
+   tu6_emit_marker(cmd, cs);
+
+   tu6_emit_blit_scissor(cmd, cs);
+
+   uint32_t gmem_index = 0;
+   for (uint32_t i = 0; i < cmd->state.subpass->color_count; ++i) {
+      uint32_t a = cmd->state.subpass->color_attachments[i].attachment;
+      if (a == VK_ATTACHMENT_UNUSED)
+         continue;
+
+      const struct tu_image_view *iview = fb->attachments[a].attachment;
+      tu6_emit_blit_info(cmd, cs, iview, tiling->gmem_offsets[gmem_index++],
+                         0);
+      tu6_emit_blit(cmd, cs);
+   }
+}
+
+static void
+tu6_emit_restart_index(struct tu_cs *cs, uint32_t restart_index)
+{
+   tu_cs_emit_pkt4(cs, REG_A6XX_PC_RESTART_INDEX, 1);
+   tu_cs_emit(cs, restart_index);
 }
 
 static void
 tu6_init_hw(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
-   struct tu_device *dev = cmd->device;
-   const struct tu_physical_device *phys_dev = dev->physical_device;
-
-   tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE);
-
-   tu_cs_emit_regs(cs, A6XX_HLSQ_INVALIDATE_CMD(
-         .vs_state = true,
-         .hs_state = true,
-         .ds_state = true,
-         .gs_state = true,
-         .fs_state = true,
-         .cs_state = true,
-         .gfx_ibo = true,
-         .cs_ibo = true,
-         .gfx_shared_const = true,
-         .cs_shared_const = true,
-         .gfx_bindless = 0x1f,
-         .cs_bindless = 0x1f));
-
-   tu_cs_emit_wfi(cs);
-
-   cmd->state.cache.pending_flush_bits &=
-      ~(TU_CMD_FLAG_WAIT_FOR_IDLE | TU_CMD_FLAG_CACHE_INVALIDATE);
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_CCU_CNTL(.color_offset = phys_dev->ccu_offset_bypass));
-   cmd->state.ccu_state = TU_CMD_CCU_SYSMEM;
-   tu_cs_emit_write_reg(cs, REG_A6XX_RB_DBG_ECO_CNTL, 0x00100000);
-   tu_cs_emit_write_reg(cs, REG_A6XX_SP_FLOAT_CNTL, 0);
-   tu_cs_emit_write_reg(cs, REG_A6XX_SP_DBG_ECO_CNTL,
-                        phys_dev->info->a6xx.magic.SP_DBG_ECO_CNTL);
-   tu_cs_emit_write_reg(cs, REG_A6XX_SP_PERFCTR_ENABLE, 0x3f);
-   tu_cs_emit_write_reg(cs, REG_A6XX_TPL1_UNKNOWN_B605, 0x44);
-   tu_cs_emit_write_reg(cs, REG_A6XX_TPL1_DBG_ECO_CNTL,
-                        phys_dev->info->a6xx.magic.TPL1_DBG_ECO_CNTL);
+   VkResult result = tu_cs_reserve_space(cmd->device, cs, 256);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
+   }
+
+   tu6_emit_cache_flush(cmd, cs);
+
+   tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UPDATE_CNTL, 0xfffff);
+
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_CCU_CNTL, 0x7c400004);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8E04, 0x00100000);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE04, 0x8);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE00, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE0F, 0x3f);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B605, 0x44);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B600, 0x100000);
    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BE00, 0x80);
    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BE01, 0);
 
-   tu_cs_emit_write_reg(cs, REG_A6XX_VPC_DBG_ECO_CNTL,
-                        phys_dev->info->a6xx.magic.VPC_DBG_ECO_CNTL);
-   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_DBG_ECO_CNTL,
-                        phys_dev->info->a6xx.magic.GRAS_DBG_ECO_CNTL);
-   tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_DBG_ECO_CNTL,
-                        phys_dev->info->a6xx.magic.HLSQ_DBG_ECO_CNTL);
-   tu_cs_emit_write_reg(cs, REG_A6XX_SP_CHICKEN_BITS,
-                        phys_dev->info->a6xx.magic.SP_CHICKEN_BITS);
+   tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9600, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8600, 0x880);
+   tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BE04, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE03, 0x00000410);
    tu_cs_emit_write_reg(cs, REG_A6XX_SP_IBO_COUNT, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B182, 0);
-   tu_cs_emit_regs(cs, A6XX_HLSQ_SHARED_CONSTS(.enable = false));
-   tu_cs_emit_write_reg(cs, REG_A6XX_UCHE_UNKNOWN_0E12,
-                        phys_dev->info->a6xx.magic.UCHE_UNKNOWN_0E12);
-   tu_cs_emit_write_reg(cs, REG_A6XX_UCHE_CLIENT_PF,
-                        phys_dev->info->a6xx.magic.UCHE_CLIENT_PF);
-   tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8E01,
-                        phys_dev->info->a6xx.magic.RB_UNKNOWN_8E01);
-   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_A9A8, 0);
-   tu_cs_emit_regs(cs, A6XX_SP_MODE_CONTROL(.constant_demotion_enable = true,
-                                            .isammode = ISAMMODE_GL,
-                                            .shared_consts_enable = false));
-
-   /* TODO: set A6XX_VFD_ADD_OFFSET_INSTANCE and fix ir3 to avoid adding base instance */
-   tu_cs_emit_write_reg(cs, REG_A6XX_VFD_ADD_OFFSET, A6XX_VFD_ADD_OFFSET_VERTEX);
+   tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BB11, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_UCHE_UNKNOWN_0E12, 0x3200000);
+   tu_cs_emit_write_reg(cs, REG_A6XX_UCHE_CLIENT_PF, 4);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8E01, 0x0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AB00, 0x5);
+   tu_cs_emit_write_reg(cs, REG_A6XX_VFD_UNKNOWN_A009, 0x00000001);
    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8811, 0x00000010);
-   tu_cs_emit_write_reg(cs, REG_A6XX_PC_MODE_CNTL,
-                        phys_dev->info->a6xx.magic.PC_MODE_CNTL);
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_MODE_CNTL, 0x1f);
+
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_SRGB_CNTL, 0);
 
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8101, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_SAMPLE_CNTL, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8110, 0);
 
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_RENDER_CONTROL0, 0x401);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_RENDER_CONTROL1, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_FS_OUTPUT_CNTL0, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_SAMPLE_CNTL, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8818, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8819, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881A, 0);
@@ -964,834 +896,705 @@ tu6_init_hw(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881E, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_88F0, 0);
 
-   tu_cs_emit_regs(cs, A6XX_VPC_POINT_COORD_INVERT(false));
+   tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9101, 0xffff00);
+   tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9107, 0);
+
+   tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9236, 1);
    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9300, 0);
 
-   tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(true));
+   tu_cs_emit_write_reg(cs, REG_A6XX_VPC_SO_OVERRIDE,
+                        A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
+
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9801, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9806, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9980, 0);
+
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9B06, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9B06, 0);
+
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_A81B, 0);
 
    tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B183, 0);
 
-   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_SU_CONSERVATIVE_RAS_CNTL, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8099, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_809B, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A0, 2);
    tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80AF, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9210, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9211, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9602, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9981, 0x3);
    tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9E72, 0);
-   tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_MODE_CNTL,
-                        0x000000a0 |
-                        A6XX_SP_TP_MODE_CNTL_ISAMMODE(ISAMMODE_GL));
+   tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9108, 0x3);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_UNKNOWN_B304, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_UNKNOWN_B309, 0x000000a2);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8804, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A4, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A5, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A6, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8805, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8806, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8878, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8879, 0);
    tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_CONTROL_5_REG, 0xfc);
 
+   tu6_emit_marker(cmd, cs);
+
    tu_cs_emit_write_reg(cs, REG_A6XX_VFD_MODE_CNTL, 0x00000000);
 
+   tu_cs_emit_write_reg(cs, REG_A6XX_VFD_UNKNOWN_A008, 0);
+
    tu_cs_emit_write_reg(cs, REG_A6XX_PC_MODE_CNTL, 0x0000001f);
 
-   tu_cs_emit_regs(cs, A6XX_RB_ALPHA_CONTROL()); /* always disable alpha test */
-   tu_cs_emit_regs(cs, A6XX_RB_DITHER_CNTL()); /* always disable dithering */
+   /* we don't use this yet.. probably best to disable.. */
+   tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
+   tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(0) |
+                     CP_SET_DRAW_STATE__0_DISABLE_ALL_GROUPS |
+                     CP_SET_DRAW_STATE__0_GROUP_ID(0));
+   tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
+   tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
 
-   tu_disable_draw_states(cmd, cs);
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_BASE_LO(0), 3);
+   tu_cs_emit(cs, 0x00000000); /* VPC_SO_BUFFER_BASE_LO_0 */
+   tu_cs_emit(cs, 0x00000000); /* VPC_SO_BUFFER_BASE_HI_0 */
+   tu_cs_emit(cs, 0x00000000); /* VPC_SO_BUFFER_SIZE_0 */
 
-   tu_cs_emit_regs(cs,
-                   A6XX_SP_TP_BORDER_COLOR_BASE_ADDR(.bo = dev->global_bo,
-                                                     .bo_offset = gb_offset(bcolor_builtin)));
-   tu_cs_emit_regs(cs,
-                   A6XX_SP_PS_TP_BORDER_COLOR_BASE_ADDR(.bo = dev->global_bo,
-                                                        .bo_offset = gb_offset(bcolor_builtin)));
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_FLUSH_BASE_LO(0), 2);
+   tu_cs_emit(cs, 0x00000000); /* VPC_SO_FLUSH_BASE_LO_0 */
+   tu_cs_emit(cs, 0x00000000); /* VPC_SO_FLUSH_BASE_HI_0 */
 
-   tu_cs_sanity_check(cs);
-}
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUF_CNTL, 1);
+   tu_cs_emit(cs, 0x00000000); /* VPC_SO_BUF_CNTL */
 
-static void
-update_vsc_pipe(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
-   const struct tu_tiling_config *tiling = cmd->state.tiling;
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_OFFSET(0), 1);
+   tu_cs_emit(cs, 0x00000000); /* UNKNOWN_E2AB */
 
-   tu_cs_emit_regs(cs,
-                   A6XX_VSC_BIN_SIZE(.width = tiling->tile0.width,
-                                     .height = tiling->tile0.height));
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_BASE_LO(1), 3);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
 
-   tu_cs_emit_regs(cs,
-                   A6XX_VSC_BIN_COUNT(.nx = tiling->tile_count.width,
-                                      .ny = tiling->tile_count.height));
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_OFFSET(1), 6);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_VSC_PIPE_CONFIG_REG(0), 32);
-   tu_cs_emit_array(cs, tiling->pipe_config, 32);
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_OFFSET(2), 6);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
 
-   tu_cs_emit_regs(cs,
-                   A6XX_VSC_PRIM_STRM_PITCH(cmd->vsc_prim_strm_pitch),
-                   A6XX_VSC_PRIM_STRM_LIMIT(cmd->vsc_prim_strm_pitch - VSC_PAD));
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_OFFSET(3), 3);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
+   tu_cs_emit(cs, 0x00000000);
 
-   tu_cs_emit_regs(cs,
-                   A6XX_VSC_DRAW_STRM_PITCH(cmd->vsc_draw_strm_pitch),
-                   A6XX_VSC_DRAW_STRM_LIMIT(cmd->vsc_draw_strm_pitch - VSC_PAD));
-}
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_HS_CTRL_REG0, 1);
+   tu_cs_emit(cs, 0x00000000);
 
-static void
-emit_vsc_overflow_test(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
-   const struct tu_tiling_config *tiling = cmd->state.tiling;
-   const uint32_t used_pipe_count =
-      tiling->pipe_count.width * tiling->pipe_count.height;
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_CTRL_REG0, 1);
+   tu_cs_emit(cs, 0x00000000);
 
-   for (int i = 0; i < used_pipe_count; i++) {
-      tu_cs_emit_pkt7(cs, CP_COND_WRITE5, 8);
-      tu_cs_emit(cs, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
-            CP_COND_WRITE5_0_WRITE_MEMORY);
-      tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_DRAW_STRM_SIZE_REG(i)));
-      tu_cs_emit(cs, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
-      tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_draw_strm_pitch - VSC_PAD));
-      tu_cs_emit(cs, CP_COND_WRITE5_4_MASK(~0));
-      tu_cs_emit_qw(cs, global_iova(cmd, vsc_draw_overflow));
-      tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(cmd->vsc_draw_strm_pitch));
-
-      tu_cs_emit_pkt7(cs, CP_COND_WRITE5, 8);
-      tu_cs_emit(cs, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
-            CP_COND_WRITE5_0_WRITE_MEMORY);
-      tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_PRIM_STRM_SIZE_REG(i)));
-      tu_cs_emit(cs, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
-      tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_prim_strm_pitch - VSC_PAD));
-      tu_cs_emit(cs, CP_COND_WRITE5_4_MASK(~0));
-      tu_cs_emit_qw(cs, global_iova(cmd, vsc_prim_overflow));
-      tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(cmd->vsc_prim_strm_pitch));
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_LRZ_CNTL, 1);
+   tu_cs_emit(cs, 0x00000000);
 
-   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_LRZ_CNTL, 1);
+   tu_cs_emit(cs, 0x00000000);
+
+   tu_cs_sanity_check(cs);
 }
 
 static void
-tu6_emit_binning_pass(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+tu6_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
 {
-   struct tu_physical_device *phys_dev = cmd->device->physical_device;
-   const struct tu_framebuffer *fb = cmd->state.framebuffer;
-
-   tu6_emit_window_scissor(cs, 0, 0, fb->width - 1, fb->height - 1);
-
-   tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
-   tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_BINNING));
-
-   tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
-   tu_cs_emit(cs, 0x1);
-
-   tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
-   tu_cs_emit(cs, 0x1);
-
-   tu_cs_emit_wfi(cs);
-
-   tu_cs_emit_regs(cs,
-                   A6XX_VFD_MODE_CNTL(.render_mode = BINNING_PASS));
-
-   update_vsc_pipe(cmd, cs);
-
-   tu_cs_emit_regs(cs,
-                   A6XX_PC_POWER_CNTL(phys_dev->info->a6xx.magic.PC_POWER_CNTL));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_VFD_POWER_CNTL(phys_dev->info->a6xx.magic.PC_POWER_CNTL));
-
-   tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
-   tu_cs_emit(cs, UNK_2C);
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_WINDOW_OFFSET(.x = 0, .y = 0));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_SP_TP_WINDOW_OFFSET(.x = 0, .y = 0));
-
-   trace_start_binning_ib(&cmd->trace, cs);
-
-   /* emit IB to binning drawcmds: */
-   tu_cs_emit_call(cs, &cmd->draw_cs);
+   VkResult result = tu_cs_reserve_space(cmd->device, cs, 256);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
+   }
 
-   trace_end_binning_ib(&cmd->trace, cs);
+   tu6_emit_lrz_flush(cmd, cs);
 
-   /* switching from binning pass to GMEM pass will cause a switch from
-    * PROGRAM_BINNING to PROGRAM, which invalidates const state (XS_CONST states)
-    * so make sure these states are re-emitted
-    * (eventually these states shouldn't exist at all with shader prologue)
-    * only VS and GS are invalidated, as FS isn't emitted in binning pass,
-    * and we don't use HW binning when tesselation is used
-    */
-   tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
-   tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(0) |
-                  CP_SET_DRAW_STATE__0_DISABLE |
-                  CP_SET_DRAW_STATE__0_GROUP_ID(TU_DRAW_STATE_CONST));
-   tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
-   tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
+   /* lrz clear? */
 
-   tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
-   tu_cs_emit(cs, UNK_2D);
+   tu6_emit_cache_flush(cmd, cs);
 
-   /* This flush is probably required because the VSC, which produces the
-    * visibility stream, is a client of UCHE, whereas the CP needs to read the
-    * visibility stream (without caching) to do draw skipping. The
-    * WFI+WAIT_FOR_ME combination guarantees that the binning commands
-    * submitted are finished before reading the VSC regs (in
-    * emit_vsc_overflow_test) or the VSC_DATA buffer directly (implicitly as
-    * part of draws).
-    */
-   tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS);
+   tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
+   tu_cs_emit(cs, 0x0);
 
-   tu_cs_emit_wfi(cs);
+   /* 0x10000000 for BYPASS.. 0x7c13c080 for GMEM: */
+   tu6_emit_wfi(cmd, cs);
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_CCU_CNTL, 1);
+   tu_cs_emit(cs, 0x7c400004); /* RB_CCU_CNTL */
 
-   tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
+   tu6_emit_zs(cmd, cs);
+   tu6_emit_mrt(cmd, cs);
+   tu6_emit_msaa(cmd, cs);
 
-   emit_vsc_overflow_test(cmd, cs);
+   if (false) {
+      /* hw binning? */
+   } else {
+      tu6_emit_bin_size(cmd, cs, 0x6000000);
+      /* no draws */
+   }
 
-   tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
-   tu_cs_emit(cs, 0x0);
+   tu6_emit_render_cntl(cmd, cs, false);
 
-   tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
-   tu_cs_emit(cs, 0x0);
+   tu_cs_sanity_check(cs);
 }
 
-static struct tu_draw_state
-tu_emit_input_attachments(struct tu_cmd_buffer *cmd,
-                          const struct tu_subpass *subpass,
-                          bool gmem)
+static void
+tu6_render_tile(struct tu_cmd_buffer *cmd,
+                struct tu_cs *cs,
+                const struct tu_tile *tile)
 {
-   const struct tu_tiling_config *tiling = cmd->state.tiling;
-
-   /* note: we can probably emit input attachments just once for the whole
-    * renderpass, this would avoid emitting both sysmem/gmem versions
-    *
-    * emit two texture descriptors for each input, as a workaround for
-    * d24s8/d32s8, which can be sampled as both float (depth) and integer (stencil)
-    * tu_shader lowers uint input attachment loads to use the 2nd descriptor
-    * in the pair
-    * TODO: a smarter workaround
-    */
-
-   if (!subpass->input_count)
-      return (struct tu_draw_state) {};
-
-   struct tu_cs_memory texture;
-   VkResult result = tu_cs_alloc(&cmd->sub_cs, subpass->input_count * 2,
-                                 A6XX_TEX_CONST_DWORDS, &texture);
+   const uint32_t render_tile_space = 64 + tu_cs_get_call_size(&cmd->draw_cs);
+   VkResult result = tu_cs_reserve_space(cmd->device, cs, render_tile_space);
    if (result != VK_SUCCESS) {
-      vk_command_buffer_set_error(&cmd->vk, result);
-      return (struct tu_draw_state) {};
+      cmd->record_result = result;
+      return;
    }
 
-   for (unsigned i = 0; i < subpass->input_count * 2; i++) {
-      uint32_t a = subpass->input_attachments[i / 2].attachment;
-      if (a == VK_ATTACHMENT_UNUSED)
-         continue;
-
-      const struct tu_image_view *iview = cmd->state.attachments[a];
-      const struct tu_render_pass_attachment *att =
-         &cmd->state.pass->attachments[a];
-      uint32_t *dst = &texture.map[A6XX_TEX_CONST_DWORDS * i];
-      uint32_t gmem_offset = tu_attachment_gmem_offset(cmd, att);
-      uint32_t cpp = att->cpp;
-
-      memcpy(dst, iview->view.descriptor, A6XX_TEX_CONST_DWORDS * 4);
-
-      /* Cube descriptors require a different sampling instruction in shader,
-       * however we don't know whether image is a cube or not until the start
-       * of a renderpass. We have to patch the descriptor to make it compatible
-       * with how it is sampled in shader.
-       */
-      enum a6xx_tex_type tex_type = (dst[2] & A6XX_TEX_CONST_2_TYPE__MASK) >>
-                                    A6XX_TEX_CONST_2_TYPE__SHIFT;
-      if (tex_type == A6XX_TEX_CUBE) {
-         dst[2] &= ~A6XX_TEX_CONST_2_TYPE__MASK;
-         dst[2] |= A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D);
-
-         uint32_t depth = (dst[5] & A6XX_TEX_CONST_5_DEPTH__MASK) >>
-                          A6XX_TEX_CONST_5_DEPTH__SHIFT;
-         dst[5] &= ~A6XX_TEX_CONST_5_DEPTH__MASK;
-         dst[5] |= A6XX_TEX_CONST_5_DEPTH(depth * 6);
-      }
+   tu6_emit_tile_select(cmd, cs, tile);
+   tu_cs_emit_ib(cs, &cmd->state.tile_load_ib);
 
-      if (i % 2 == 1 && att->format == VK_FORMAT_D24_UNORM_S8_UINT) {
-         /* note this works because spec says fb and input attachments
-          * must use identity swizzle
-          *
-          * Also we clear swap to WZYX.  This is because the view might have
-          * picked XYZW to work better with border colors.
-          */
-         dst[0] &= ~(A6XX_TEX_CONST_0_FMT__MASK |
-            A6XX_TEX_CONST_0_SWAP__MASK |
-            A6XX_TEX_CONST_0_SWIZ_X__MASK | A6XX_TEX_CONST_0_SWIZ_Y__MASK |
-            A6XX_TEX_CONST_0_SWIZ_Z__MASK | A6XX_TEX_CONST_0_SWIZ_W__MASK);
-         if (!cmd->device->physical_device->info->a6xx.has_z24uint_s8uint) {
-            dst[0] |= A6XX_TEX_CONST_0_FMT(FMT6_8_8_8_8_UINT) |
-               A6XX_TEX_CONST_0_SWIZ_X(A6XX_TEX_W) |
-               A6XX_TEX_CONST_0_SWIZ_Y(A6XX_TEX_ZERO) |
-               A6XX_TEX_CONST_0_SWIZ_Z(A6XX_TEX_ZERO) |
-               A6XX_TEX_CONST_0_SWIZ_W(A6XX_TEX_ONE);
-         } else {
-            dst[0] |= A6XX_TEX_CONST_0_FMT(FMT6_Z24_UINT_S8_UINT) |
-               A6XX_TEX_CONST_0_SWIZ_X(A6XX_TEX_Y) |
-               A6XX_TEX_CONST_0_SWIZ_Y(A6XX_TEX_ZERO) |
-               A6XX_TEX_CONST_0_SWIZ_Z(A6XX_TEX_ZERO) |
-               A6XX_TEX_CONST_0_SWIZ_W(A6XX_TEX_ONE);
-         }
-      }
+   tu_cs_emit_call(cs, &cmd->draw_cs);
+   cmd->wait_for_idle = true;
 
-      if (i % 2 == 1 && att->format == VK_FORMAT_D32_SFLOAT_S8_UINT) {
-         dst[0] &= ~A6XX_TEX_CONST_0_FMT__MASK;
-         dst[0] |= A6XX_TEX_CONST_0_FMT(FMT6_8_UINT);
-         dst[2] &= ~(A6XX_TEX_CONST_2_PITCHALIGN__MASK | A6XX_TEX_CONST_2_PITCH__MASK);
-         dst[2] |= A6XX_TEX_CONST_2_PITCH(iview->stencil_PITCH << 6);
-         dst[3] = 0;
-         dst[4] = iview->stencil_base_addr;
-         dst[5] = (dst[5] & 0xffff) | iview->stencil_base_addr >> 32;
-
-         cpp = att->samples;
-         gmem_offset = att->gmem_offset_stencil[cmd->state.gmem_layout];
-      }
+   tu_cs_emit_ib(cs, &cmd->state.tile_store_ib);
 
-      if (!gmem || !subpass->input_attachments[i / 2].patch_input_gmem)
-         continue;
+   tu_cs_sanity_check(cs);
+}
 
-      /* patched for gmem */
-      dst[0] &= ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
-      dst[0] |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
-      dst[2] =
-         A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D) |
-         A6XX_TEX_CONST_2_PITCH(tiling->tile0.width * cpp);
-      dst[3] = 0;
-      dst[4] = cmd->device->physical_device->gmem_base + gmem_offset;
-      dst[5] = A6XX_TEX_CONST_5_DEPTH(1);
-      for (unsigned i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
-         dst[i] = 0;
+static void
+tu6_render_end(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+   VkResult result = tu_cs_reserve_space(cmd->device, cs, 16);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
    }
 
-   struct tu_cs cs;
-   struct tu_draw_state ds = tu_cs_draw_state(&cmd->sub_cs, &cs, 9);
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_LRZ_CNTL, 1);
+   tu_cs_emit(cs, A6XX_GRAS_LRZ_CNTL_ENABLE | A6XX_GRAS_LRZ_CNTL_UNK3);
 
-   tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_FRAG, 3);
-   tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
-                  CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                  CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
-                  CP_LOAD_STATE6_0_STATE_BLOCK(SB6_FS_TEX) |
-                  CP_LOAD_STATE6_0_NUM_UNIT(subpass->input_count * 2));
-   tu_cs_emit_qw(&cs, texture.iova);
+   tu6_emit_lrz_flush(cmd, cs);
 
-   tu_cs_emit_regs(&cs, A6XX_SP_FS_TEX_CONST(.qword = texture.iova));
+   tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS, true);
 
-   tu_cs_emit_regs(&cs, A6XX_SP_FS_TEX_COUNT(subpass->input_count * 2));
-
-   assert(cs.cur == cs.end); /* validate draw state size */
-
-   return ds;
+   tu_cs_sanity_check(cs);
 }
 
 static void
-tu_set_input_attachments(struct tu_cmd_buffer *cmd, const struct tu_subpass *subpass)
+tu_cmd_render_tiles(struct tu_cmd_buffer *cmd)
 {
-   struct tu_cs *cs = &cmd->draw_cs;
+   const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
 
-   tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 6);
-   tu_cs_emit_draw_state(cs, TU_DRAW_STATE_INPUT_ATTACHMENTS_GMEM,
-                         tu_emit_input_attachments(cmd, subpass, true));
-   tu_cs_emit_draw_state(cs, TU_DRAW_STATE_INPUT_ATTACHMENTS_SYSMEM,
-                         tu_emit_input_attachments(cmd, subpass, false));
-}
+   tu6_render_begin(cmd, &cmd->cs);
 
+   for (uint32_t y = 0; y < tiling->tile_count.height; y++) {
+      for (uint32_t x = 0; x < tiling->tile_count.width; x++) {
+         struct tu_tile tile;
+         tu_tiling_config_get_tile(tiling, cmd->device, x, y, &tile);
+         tu6_render_tile(cmd, &cmd->cs, &tile);
+      }
+   }
+
+   tu6_render_end(cmd, &cmd->cs);
+}
 
 static void
-tu_emit_renderpass_begin(struct tu_cmd_buffer *cmd,
-                         const VkClearValue *clear_values)
+tu_cmd_prepare_tile_load_ib(struct tu_cmd_buffer *cmd)
 {
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_GMEM);
-
-   tu6_emit_tile_load(cmd, cs);
+   const uint32_t tile_load_space = 16 + 32 * MAX_RTS;
+   const struct tu_subpass *subpass = cmd->state.subpass;
+   struct tu_attachment_state *attachments = cmd->state.attachments;
+   struct tu_cs sub_cs;
 
-   tu6_emit_blit_scissor(cmd, cs, false);
-
-   for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
-      tu_clear_gmem_attachment(cmd, cs, i, &clear_values[i]);
-
-   tu_cond_exec_end(cs);
+   VkResult result = tu_cs_begin_sub_stream(cmd->device, &cmd->tile_cs,
+                                            tile_load_space, &sub_cs);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
+   }
 
-   tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
+   /* emit to tile-load sub_cs */
+   tu6_emit_tile_load(cmd, &sub_cs);
 
-   for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
-      tu_clear_sysmem_attachment(cmd, cs, i, &clear_values[i]);
+   cmd->state.tile_load_ib = tu_cs_end_sub_stream(&cmd->tile_cs, &sub_cs);
 
-   tu_cond_exec_end(cs);
+   for (uint32_t i = 0; i < subpass->color_count; ++i) {
+      const uint32_t a = subpass->color_attachments[i].attachment;
+      if (a != VK_ATTACHMENT_UNUSED)
+         attachments[a].pending_clear_aspects = 0;
+   }
 }
 
 static void
-tu6_sysmem_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-                        struct tu_renderpass_result *autotune_result)
+tu_cmd_prepare_tile_store_ib(struct tu_cmd_buffer *cmd)
 {
-   const struct tu_framebuffer *fb = cmd->state.framebuffer;
-
-   tu_lrz_sysmem_begin(cmd, cs);
-
-   assert(fb->width > 0 && fb->height > 0);
-   tu6_emit_window_scissor(cs, 0, 0, fb->width - 1, fb->height - 1);
-   tu6_emit_window_offset(cs, 0, 0);
-
-   tu6_emit_bin_size(cs, 0, 0,
-                     A6XX_RB_BIN_CONTROL_BUFFERS_LOCATION(BUFFERS_IN_SYSMEM) |
-                     A6XX_RB_BIN_CONTROL_FORCE_LRZ_WRITE_DIS);
-
-   tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
-   tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_BYPASS));
-
-   tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
-   tu_cs_emit(cs, 0x0);
-
-   tu_emit_cache_flush_ccu(cmd, cs, TU_CMD_CCU_SYSMEM);
-
-   tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
-   tu_cs_emit(cs, 0x1);
+   const uint32_t tile_store_space = 32 + 32 * MAX_RTS;
+   struct tu_cs sub_cs;
 
-   tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
-   tu_cs_emit(cs, 0x0);
+   VkResult result = tu_cs_begin_sub_stream(cmd->device, &cmd->tile_cs,
+                                            tile_store_space, &sub_cs);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
+   }
 
-   tu_autotune_begin_renderpass(cmd, cs, autotune_result);
+   /* emit to tile-store sub_cs */
+   tu6_emit_tile_store(cmd, &sub_cs);
 
-   tu_cs_sanity_check(cs);
+   cmd->state.tile_store_ib = tu_cs_end_sub_stream(&cmd->tile_cs, &sub_cs);
 }
 
 static void
-tu6_sysmem_render_end(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-                      struct tu_renderpass_result *autotune_result)
+tu_cmd_update_tiling_config(struct tu_cmd_buffer *cmd,
+                            const VkRect2D *render_area)
 {
-   tu_autotune_end_renderpass(cmd, cs, autotune_result);
-
-   /* Do any resolves of the last subpass. These are handled in the
-    * tile_store_cs in the gmem path.
-    */
-   tu6_emit_sysmem_resolves(cmd, cs, cmd->state.subpass);
+   const struct tu_device *dev = cmd->device;
+   const struct tu_render_pass *pass = cmd->state.pass;
+   const struct tu_subpass *subpass = cmd->state.subpass;
+   struct tu_tiling_config *tiling = &cmd->state.tiling_config;
 
-   tu_cs_emit_call(cs, &cmd->draw_epilogue_cs);
+   uint32_t buffer_cpp[MAX_RTS + 2];
+   uint32_t buffer_count = 0;
 
-   tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
-   tu_cs_emit(cs, 0x0);
+   for (uint32_t i = 0; i < subpass->color_count; ++i) {
+      const uint32_t a = subpass->color_attachments[i].attachment;
+      if (a == VK_ATTACHMENT_UNUSED)
+         continue;
 
-   tu_lrz_sysmem_end(cmd, cs);
+      const struct tu_render_pass_attachment *att = &pass->attachments[a];
+      buffer_cpp[buffer_count++] =
+         vk_format_get_blocksize(att->format) * att->samples;
+   }
 
-   tu_cs_sanity_check(cs);
-}
+   if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
+      const uint32_t a = subpass->depth_stencil_attachment.attachment;
+      const struct tu_render_pass_attachment *att = &pass->attachments[a];
+
+      /* TODO */
+      assert(att->format != VK_FORMAT_D32_SFLOAT_S8_UINT);
+
+      buffer_cpp[buffer_count++] =
+         vk_format_get_blocksize(att->format) * att->samples;
+   }
+
+   tu_tiling_config_update(tiling, dev, buffer_cpp, buffer_count,
+                           render_area);
+}
+
+const struct tu_dynamic_state default_dynamic_state = {
+   .viewport =
+     {
+       .count = 0,
+     },
+   .scissor =
+     {
+       .count = 0,
+     },
+   .line_width = 1.0f,
+   .depth_bias =
+     {
+       .bias = 0.0f,
+       .clamp = 0.0f,
+       .slope = 0.0f,
+     },
+   .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
+   .depth_bounds =
+     {
+       .min = 0.0f,
+       .max = 1.0f,
+     },
+   .stencil_compare_mask =
+     {
+       .front = ~0u,
+       .back = ~0u,
+     },
+   .stencil_write_mask =
+     {
+       .front = ~0u,
+       .back = ~0u,
+     },
+   .stencil_reference =
+     {
+       .front = 0u,
+       .back = 0u,
+     },
+};
 
-static void
-tu6_tile_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-                      struct tu_renderpass_result *autotune_result)
+static void UNUSED /* FINISHME */
+tu_bind_dynamic_state(struct tu_cmd_buffer *cmd_buffer,
+                      const struct tu_dynamic_state *src)
 {
-   struct tu_physical_device *phys_dev = cmd->device->physical_device;
-   const struct tu_tiling_config *tiling = cmd->state.tiling;
-   tu_lrz_tiling_begin(cmd, cs);
+   struct tu_dynamic_state *dest = &cmd_buffer->state.dynamic;
+   uint32_t copy_mask = src->mask;
+   uint32_t dest_mask = 0;
 
-   tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
-   tu_cs_emit(cs, 0x0);
-
-   tu_emit_cache_flush_ccu(cmd, cs, TU_CMD_CCU_GMEM);
+   tu_use_args(cmd_buffer); /* FINISHME */
 
-   if (use_hw_binning(cmd)) {
-      if (!cmd->vsc_initialized) {
-         tu6_lazy_emit_vsc(cmd, cs);
+   /* Make sure to copy the number of viewports/scissors because they can
+    * only be specified at pipeline creation time.
+    */
+   dest->viewport.count = src->viewport.count;
+   dest->scissor.count = src->scissor.count;
+   dest->discard_rectangle.count = src->discard_rectangle.count;
+
+   if (copy_mask & TU_DYNAMIC_VIEWPORT) {
+      if (memcmp(&dest->viewport.viewports, &src->viewport.viewports,
+                 src->viewport.count * sizeof(VkViewport))) {
+         typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
+                      src->viewport.count);
+         dest_mask |= TU_DYNAMIC_VIEWPORT;
       }
+   }
 
-      tu6_emit_bin_size(cs, tiling->tile0.width, tiling->tile0.height,
-                        A6XX_RB_BIN_CONTROL_RENDER_MODE(BINNING_PASS) |
-                        A6XX_RB_BIN_CONTROL_LRZ_FEEDBACK_ZMODE_MASK(0x6));
-
-      tu6_emit_render_cntl(cmd, cmd->state.subpass, cs, true);
+   if (copy_mask & TU_DYNAMIC_SCISSOR) {
+      if (memcmp(&dest->scissor.scissors, &src->scissor.scissors,
+                 src->scissor.count * sizeof(VkRect2D))) {
+         typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
+                      src->scissor.count);
+         dest_mask |= TU_DYNAMIC_SCISSOR;
+      }
+   }
 
-      tu6_emit_binning_pass(cmd, cs);
+   if (copy_mask & TU_DYNAMIC_LINE_WIDTH) {
+      if (dest->line_width != src->line_width) {
+         dest->line_width = src->line_width;
+         dest_mask |= TU_DYNAMIC_LINE_WIDTH;
+      }
+   }
 
-      tu6_emit_bin_size(cs, tiling->tile0.width, tiling->tile0.height,
-                        A6XX_RB_BIN_CONTROL_FORCE_LRZ_WRITE_DIS |
-                        A6XX_RB_BIN_CONTROL_LRZ_FEEDBACK_ZMODE_MASK(0x6));
+   if (copy_mask & TU_DYNAMIC_DEPTH_BIAS) {
+      if (memcmp(&dest->depth_bias, &src->depth_bias,
+                 sizeof(src->depth_bias))) {
+         dest->depth_bias = src->depth_bias;
+         dest_mask |= TU_DYNAMIC_DEPTH_BIAS;
+      }
+   }
 
-      tu_cs_emit_regs(cs,
-                      A6XX_VFD_MODE_CNTL(0));
+   if (copy_mask & TU_DYNAMIC_BLEND_CONSTANTS) {
+      if (memcmp(&dest->blend_constants, &src->blend_constants,
+                 sizeof(src->blend_constants))) {
+         typed_memcpy(dest->blend_constants, src->blend_constants, 4);
+         dest_mask |= TU_DYNAMIC_BLEND_CONSTANTS;
+      }
+   }
 
-      tu_cs_emit_regs(cs,
-                      A6XX_PC_POWER_CNTL(phys_dev->info->a6xx.magic.PC_POWER_CNTL));
+   if (copy_mask & TU_DYNAMIC_DEPTH_BOUNDS) {
+      if (memcmp(&dest->depth_bounds, &src->depth_bounds,
+                 sizeof(src->depth_bounds))) {
+         dest->depth_bounds = src->depth_bounds;
+         dest_mask |= TU_DYNAMIC_DEPTH_BOUNDS;
+      }
+   }
 
-      tu_cs_emit_regs(cs,
-                      A6XX_VFD_POWER_CNTL(phys_dev->info->a6xx.magic.PC_POWER_CNTL));
+   if (copy_mask & TU_DYNAMIC_STENCIL_COMPARE_MASK) {
+      if (memcmp(&dest->stencil_compare_mask, &src->stencil_compare_mask,
+                 sizeof(src->stencil_compare_mask))) {
+         dest->stencil_compare_mask = src->stencil_compare_mask;
+         dest_mask |= TU_DYNAMIC_STENCIL_COMPARE_MASK;
+      }
+   }
 
-      tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
-      tu_cs_emit(cs, 0x1);
-      tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_LOCAL, 1);
-      tu_cs_emit(cs, 0x1);
-   } else {
-      tu6_emit_bin_size(cs, tiling->tile0.width, tiling->tile0.height,
-                        A6XX_RB_BIN_CONTROL_LRZ_FEEDBACK_ZMODE_MASK(0x6));
-
-      if (tiling->binning_possible) {
-         /* Mark all tiles as visible for tu6_emit_cond_for_load_stores(), since
-          * the actual binner didn't run.
-          */
-         int pipe_count = tiling->pipe_count.width * tiling->pipe_count.height;
-         tu_cs_emit_pkt4(cs, REG_A6XX_VSC_STATE_REG(0), pipe_count);
-         for (int i = 0; i < pipe_count; i++)
-            tu_cs_emit(cs, ~0);
+   if (copy_mask & TU_DYNAMIC_STENCIL_WRITE_MASK) {
+      if (memcmp(&dest->stencil_write_mask, &src->stencil_write_mask,
+                 sizeof(src->stencil_write_mask))) {
+         dest->stencil_write_mask = src->stencil_write_mask;
+         dest_mask |= TU_DYNAMIC_STENCIL_WRITE_MASK;
       }
    }
 
-   tu_autotune_begin_renderpass(cmd, cs, autotune_result);
+   if (copy_mask & TU_DYNAMIC_STENCIL_REFERENCE) {
+      if (memcmp(&dest->stencil_reference, &src->stencil_reference,
+                 sizeof(src->stencil_reference))) {
+         dest->stencil_reference = src->stencil_reference;
+         dest_mask |= TU_DYNAMIC_STENCIL_REFERENCE;
+      }
+   }
 
-   tu_cs_sanity_check(cs);
+   if (copy_mask & TU_DYNAMIC_DISCARD_RECTANGLE) {
+      if (memcmp(&dest->discard_rectangle.rectangles,
+                 &src->discard_rectangle.rectangles,
+                 src->discard_rectangle.count * sizeof(VkRect2D))) {
+         typed_memcpy(dest->discard_rectangle.rectangles,
+                      src->discard_rectangle.rectangles,
+                      src->discard_rectangle.count);
+         dest_mask |= TU_DYNAMIC_DISCARD_RECTANGLE;
+      }
+   }
 }
 
-static void
-tu6_render_tile(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-                uint32_t tx, uint32_t ty, uint32_t pipe, uint32_t slot)
+static VkResult
+tu_create_cmd_buffer(struct tu_device *device,
+                     struct tu_cmd_pool *pool,
+                     VkCommandBufferLevel level,
+                     VkCommandBuffer *pCommandBuffer)
 {
-   tu6_emit_tile_select(cmd, &cmd->cs, tx, ty, pipe, slot);
-
-   trace_start_draw_ib_gmem(&cmd->trace, &cmd->cs);
-
-   /* Primitives that passed all tests are still counted in in each
-    * tile even with HW binning beforehand. Do not permit it.
-    */
-   if (cmd->state.prim_generated_query_running_before_rp)
-      tu6_emit_event_write(cmd, cs, STOP_PRIMITIVE_CTRS);
+   struct tu_cmd_buffer *cmd_buffer;
+   cmd_buffer = vk_zalloc(&pool->alloc, sizeof(*cmd_buffer), 8,
+                          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (cmd_buffer == NULL)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   tu_cs_emit_call(cs, &cmd->draw_cs);
+   cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+   cmd_buffer->device = device;
+   cmd_buffer->pool = pool;
+   cmd_buffer->level = level;
 
-   if (cmd->state.prim_generated_query_running_before_rp)
-      tu6_emit_event_write(cmd, cs, START_PRIMITIVE_CTRS);
+   if (pool) {
+      list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
+      cmd_buffer->queue_family_index = pool->queue_family_index;
 
-   if (use_hw_binning(cmd)) {
-      tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
-      tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_ENDVIS));
+   } else {
+      /* Init the pool_link so we can safely call list_del when we destroy
+       * the command buffer
+       */
+      list_inithead(&cmd_buffer->pool_link);
+      cmd_buffer->queue_family_index = TU_QUEUE_GENERAL;
    }
 
-   /* Predicate is changed in draw_cs so we have to re-emit it */
-   if (cmd->state.rp.draw_cs_writes_to_cond_pred)
-      tu6_emit_cond_for_load_stores(cmd, cs, pipe, slot, false);
+   tu_bo_list_init(&cmd_buffer->bo_list);
+   tu_cs_init(&cmd_buffer->cs, TU_CS_MODE_GROW, 4096);
+   tu_cs_init(&cmd_buffer->draw_cs, TU_CS_MODE_GROW, 4096);
+   tu_cs_init(&cmd_buffer->tile_cs, TU_CS_MODE_SUB_STREAM, 1024);
 
-   tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
-   tu_cs_emit(cs, 0x0);
+   *pCommandBuffer = tu_cmd_buffer_to_handle(cmd_buffer);
 
-   tu_cs_emit_call(cs, &cmd->tile_store_cs);
+   list_inithead(&cmd_buffer->upload.list);
 
-   tu_clone_trace_range(cmd, cs, cmd->trace_renderpass_start,
-         cmd->trace_renderpass_end);
+   cmd_buffer->marker_reg = REG_A6XX_CP_SCRATCH_REG(
+      cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY ? 7 : 6);
 
-   tu_cs_sanity_check(cs);
+   VkResult result = tu_bo_init_new(device, &cmd_buffer->scratch_bo, 0x1000);
+   if (result != VK_SUCCESS)
+      return result;
 
-   trace_end_draw_ib_gmem(&cmd->trace, &cmd->cs);
+   return VK_SUCCESS;
 }
 
 static void
-tu6_tile_render_end(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-                    struct tu_renderpass_result *autotune_result)
+tu_cmd_buffer_destroy(struct tu_cmd_buffer *cmd_buffer)
 {
-   tu_autotune_end_renderpass(cmd, cs, autotune_result);
+   tu_bo_finish(cmd_buffer->device, &cmd_buffer->scratch_bo);
 
-   tu_cs_emit_call(cs, &cmd->draw_epilogue_cs);
+   list_del(&cmd_buffer->pool_link);
 
-   tu_lrz_tiling_end(cmd, cs);
+   for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++)
+      free(cmd_buffer->descriptors[i].push_set.set.mapped_ptr);
 
-   tu6_emit_event_write(cmd, cs, PC_CCU_RESOLVE_TS);
+   tu_cs_finish(cmd_buffer->device, &cmd_buffer->cs);
+   tu_cs_finish(cmd_buffer->device, &cmd_buffer->draw_cs);
+   tu_cs_finish(cmd_buffer->device, &cmd_buffer->tile_cs);
 
-   tu_cs_sanity_check(cs);
+   tu_bo_list_destroy(&cmd_buffer->bo_list);
+   vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
 }
 
-static void
-tu_cmd_render_tiles(struct tu_cmd_buffer *cmd,
-                    struct tu_renderpass_result *autotune_result)
+static VkResult
+tu_reset_cmd_buffer(struct tu_cmd_buffer *cmd_buffer)
 {
-   const struct tu_framebuffer *fb = cmd->state.framebuffer;
-   const struct tu_tiling_config *tiling = cmd->state.tiling;
+   cmd_buffer->wait_for_idle = true;
 
-   /* Create gmem stores now (at EndRenderPass time)) because they needed to
-    * know whether to allow their conditional execution, which was tied to a
-    * state that was known only at the end of the renderpass.  They will be
-    * called from tu6_render_tile().
-    */
-   tu_cs_begin(&cmd->tile_store_cs);
-   tu6_emit_tile_store(cmd, &cmd->tile_store_cs);
-   tu_cs_end(&cmd->tile_store_cs);
-
-   cmd->trace_renderpass_end = u_trace_end_iterator(&cmd->trace);
+   cmd_buffer->record_result = VK_SUCCESS;
 
-   tu6_tile_render_begin(cmd, &cmd->cs, autotune_result);
+   tu_bo_list_reset(&cmd_buffer->bo_list);
+   tu_cs_reset(cmd_buffer->device, &cmd_buffer->cs);
+   tu_cs_reset(cmd_buffer->device, &cmd_buffer->draw_cs);
+   tu_cs_reset(cmd_buffer->device, &cmd_buffer->tile_cs);
 
-   /* Note: we reverse the order of walking the pipes and tiles on every
-    * other row, to improve texture cache locality compared to raster order.
-    */
-   for (uint32_t py = 0; py < tiling->pipe_count.height; py++) {
-      uint32_t pipe_row = py * tiling->pipe_count.width;
-      for (uint32_t pipe_row_i = 0; pipe_row_i < tiling->pipe_count.width; pipe_row_i++) {
-         uint32_t px;
-         if (py & 1)
-            px = tiling->pipe_count.width - 1 - pipe_row_i;
-         else
-            px = pipe_row_i;
-         uint32_t pipe = pipe_row + px;
-         uint32_t tx1 = px * tiling->pipe0.width;
-         uint32_t ty1 = py * tiling->pipe0.height;
-         uint32_t tx2 = MIN2(tx1 + tiling->pipe0.width, tiling->tile_count.width);
-         uint32_t ty2 = MIN2(ty1 + tiling->pipe0.height, tiling->tile_count.height);
-         uint32_t tile_row_stride = tx2 - tx1;
-         uint32_t slot_row = 0;
-         for (uint32_t ty = ty1; ty < ty2; ty++) {
-            for (uint32_t tile_row_i = 0; tile_row_i < tile_row_stride; tile_row_i++) {
-               uint32_t tx;
-               if (ty & 1)
-                  tx = tile_row_stride - 1 - tile_row_i;
-               else
-                  tx = tile_row_i;
-               uint32_t slot = slot_row + tx;
-               tu6_render_tile(cmd, &cmd->cs, tx1 + tx, ty, pipe, slot);
-            }
-            slot_row += tile_row_stride;
-         }
-      }
+   for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++) {
+      cmd_buffer->descriptors[i].dirty = 0;
+      cmd_buffer->descriptors[i].valid = 0;
+      cmd_buffer->descriptors[i].push_dirty = false;
    }
 
-   tu6_tile_render_end(cmd, &cmd->cs, autotune_result);
-
-   trace_end_render_pass(&cmd->trace, &cmd->cs, fb, tiling);
-
-   /* tu6_render_tile has cloned these tracepoints for each tile */
-   if (!u_trace_iterator_equal(cmd->trace_renderpass_start, cmd->trace_renderpass_end))
-      u_trace_disable_event_range(cmd->trace_renderpass_start,
-                                  cmd->trace_renderpass_end);
+   cmd_buffer->status = TU_CMD_BUFFER_STATUS_INITIAL;
 
-   /* Reset the gmem store CS entry lists so that the next render pass
-    * does its own stores.
-    */
-   tu_cs_discard_entries(&cmd->tile_store_cs);
+   return cmd_buffer->record_result;
 }
 
-static void
-tu_cmd_render_sysmem(struct tu_cmd_buffer *cmd,
-                     struct tu_renderpass_result *autotune_result)
+static VkResult
+tu_cmd_state_setup_attachments(struct tu_cmd_buffer *cmd_buffer,
+                               const VkRenderPassBeginInfo *info)
 {
-   cmd->trace_renderpass_end = u_trace_end_iterator(&cmd->trace);
-
-   tu6_sysmem_render_begin(cmd, &cmd->cs, autotune_result);
+   struct tu_cmd_state *state = &cmd_buffer->state;
+   const struct tu_framebuffer *fb = state->framebuffer;
+   const struct tu_render_pass *pass = state->pass;
 
-   trace_start_draw_ib_sysmem(&cmd->trace, &cmd->cs);
-
-   tu_cs_emit_call(&cmd->cs, &cmd->draw_cs);
+   for (uint32_t i = 0; i < fb->attachment_count; ++i) {
+      const struct tu_image_view *iview = fb->attachments[i].attachment;
+      tu_bo_list_add(&cmd_buffer->bo_list, iview->image->bo,
+                     MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
+   }
 
-   trace_end_draw_ib_sysmem(&cmd->trace, &cmd->cs);
+   if (pass->attachment_count == 0) {
+      state->attachments = NULL;
+      return VK_SUCCESS;
+   }
 
-   tu6_sysmem_render_end(cmd, &cmd->cs, autotune_result);
+   state->attachments =
+      vk_alloc(&cmd_buffer->pool->alloc,
+               pass->attachment_count * sizeof(state->attachments[0]), 8,
+               VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (state->attachments == NULL) {
+      cmd_buffer->record_result = VK_ERROR_OUT_OF_HOST_MEMORY;
+      return cmd_buffer->record_result;
+   }
 
-   trace_end_render_pass(&cmd->trace, &cmd->cs, cmd->state.framebuffer, cmd->state.tiling);
-}
+   for (uint32_t i = 0; i < pass->attachment_count; ++i) {
+      const struct tu_render_pass_attachment *att = &pass->attachments[i];
+      VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
+      VkImageAspectFlags clear_aspects = 0;
 
-void
-tu_cmd_render(struct tu_cmd_buffer *cmd_buffer)
-{
-   if (cmd_buffer->state.rp.has_tess)
-      tu6_lazy_emit_tessfactor_addr(cmd_buffer);
+      if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
+         /* color attachment */
+         if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+            clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
+         }
+      } else {
+         /* depthstencil attachment */
+         if ((att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
+             att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+            clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
+            if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
+                att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_DONT_CARE)
+               clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
+         }
+         if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
+             att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
+            clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
+         }
+      }
 
-   struct tu_renderpass_result *autotune_result = NULL;
-   if (use_sysmem_rendering(cmd_buffer, &autotune_result))
-      tu_cmd_render_sysmem(cmd_buffer, autotune_result);
-   else
-      tu_cmd_render_tiles(cmd_buffer, autotune_result);
+      state->attachments[i].pending_clear_aspects = clear_aspects;
+      state->attachments[i].cleared_views = 0;
+      if (clear_aspects && info) {
+         assert(info->clearValueCount > i);
+         state->attachments[i].clear_value = info->pClearValues[i];
+      }
 
-   /* Outside of renderpasses we assume all draw states are disabled. We do
-    * this outside the draw CS for the normal case where 3d gmem stores aren't
-    * used.
-    */
-   tu_disable_draw_states(cmd_buffer, &cmd_buffer->cs);
+      state->attachments[i].current_layout = att->initial_layout;
+   }
 
+   return VK_SUCCESS;
 }
 
-static void tu_reset_render_pass(struct tu_cmd_buffer *cmd_buffer)
+VkResult
+tu_AllocateCommandBuffers(VkDevice _device,
+                          const VkCommandBufferAllocateInfo *pAllocateInfo,
+                          VkCommandBuffer *pCommandBuffers)
 {
-   /* discard draw_cs and draw_epilogue_cs entries now that the tiles are
-      rendered */
-   tu_cs_discard_entries(&cmd_buffer->draw_cs);
-   tu_cs_begin(&cmd_buffer->draw_cs);
-   tu_cs_discard_entries(&cmd_buffer->draw_epilogue_cs);
-   tu_cs_begin(&cmd_buffer->draw_epilogue_cs);
+   TU_FROM_HANDLE(tu_device, device, _device);
+   TU_FROM_HANDLE(tu_cmd_pool, pool, pAllocateInfo->commandPool);
 
-   cmd_buffer->state.pass = NULL;
-   cmd_buffer->state.subpass = NULL;
-   cmd_buffer->state.framebuffer = NULL;
-   cmd_buffer->state.attachments = NULL;
-   cmd_buffer->state.gmem_layout = TU_GMEM_LAYOUT_COUNT; /* invalid value to prevent looking up gmem offsets */
-   memset(&cmd_buffer->state.rp, 0, sizeof(cmd_buffer->state.rp));
+   VkResult result = VK_SUCCESS;
+   uint32_t i;
 
-   /* LRZ is not valid next time we use it */
-   cmd_buffer->state.lrz.valid = false;
-   cmd_buffer->state.dirty |= TU_CMD_DIRTY_LRZ;
-}
+   for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
 
-static VkResult
-tu_create_cmd_buffer(struct vk_command_pool *pool,
-                     struct vk_command_buffer **cmd_buffer_out)
-{
-   struct tu_device *device =
-      container_of(pool->base.device, struct tu_device, vk);
-   struct tu_cmd_buffer *cmd_buffer;
+      if (!list_empty(&pool->free_cmd_buffers)) {
+         struct tu_cmd_buffer *cmd_buffer = list_first_entry(
+            &pool->free_cmd_buffers, struct tu_cmd_buffer, pool_link);
 
-   cmd_buffer = vk_zalloc2(&device->vk.alloc, NULL, sizeof(*cmd_buffer), 8,
-                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+         list_del(&cmd_buffer->pool_link);
+         list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
 
-   if (cmd_buffer == NULL)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+         result = tu_reset_cmd_buffer(cmd_buffer);
+         cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+         cmd_buffer->level = pAllocateInfo->level;
 
-   VkResult result = vk_command_buffer_init(pool, &cmd_buffer->vk,
-                                            &tu_cmd_buffer_ops, 0);
-   if (result != VK_SUCCESS) {
-      vk_free2(&device->vk.alloc, NULL, cmd_buffer);
-      return result;
+         pCommandBuffers[i] = tu_cmd_buffer_to_handle(cmd_buffer);
+      } else {
+         result = tu_create_cmd_buffer(device, pool, pAllocateInfo->level,
+                                       &pCommandBuffers[i]);
+      }
+      if (result != VK_SUCCESS)
+         break;
    }
 
-   cmd_buffer->device = device;
-
-   u_trace_init(&cmd_buffer->trace, &device->trace_context);
-   list_inithead(&cmd_buffer->renderpass_autotune_results);
-
-   tu_cs_init(&cmd_buffer->cs, device, TU_CS_MODE_GROW, 4096, "cmd cs");
-   tu_cs_init(&cmd_buffer->draw_cs, device, TU_CS_MODE_GROW, 4096, "draw cs");
-   tu_cs_init(&cmd_buffer->tile_store_cs, device, TU_CS_MODE_GROW, 2048, "tile store cs");
-   tu_cs_init(&cmd_buffer->draw_epilogue_cs, device, TU_CS_MODE_GROW, 4096, "draw epilogue cs");
-   tu_cs_init(&cmd_buffer->sub_cs, device, TU_CS_MODE_SUB_STREAM, 2048, "draw sub cs");
-   tu_cs_init(&cmd_buffer->pre_chain.draw_cs, device, TU_CS_MODE_GROW, 4096, "prechain draw cs");
-   tu_cs_init(&cmd_buffer->pre_chain.draw_epilogue_cs, device, TU_CS_MODE_GROW, 4096, "prechain draw epiligoue cs");
-
-   *cmd_buffer_out = &cmd_buffer->vk;
-
-   return VK_SUCCESS;
-}
+   if (result != VK_SUCCESS) {
+      tu_FreeCommandBuffers(_device, pAllocateInfo->commandPool, i,
+                            pCommandBuffers);
 
-static void
-tu_cmd_buffer_destroy(struct vk_command_buffer *vk_cmd_buffer)
-{
-   struct tu_cmd_buffer *cmd_buffer =
-      container_of(vk_cmd_buffer, struct tu_cmd_buffer, vk);
-
-   tu_cs_finish(&cmd_buffer->cs);
-   tu_cs_finish(&cmd_buffer->draw_cs);
-   tu_cs_finish(&cmd_buffer->tile_store_cs);
-   tu_cs_finish(&cmd_buffer->draw_epilogue_cs);
-   tu_cs_finish(&cmd_buffer->sub_cs);
-   tu_cs_finish(&cmd_buffer->pre_chain.draw_cs);
-   tu_cs_finish(&cmd_buffer->pre_chain.draw_epilogue_cs);
-
-   u_trace_fini(&cmd_buffer->trace);
-
-   tu_autotune_free_results(cmd_buffer->device, &cmd_buffer->renderpass_autotune_results);
-
-   for (unsigned i = 0; i < MAX_BIND_POINTS; i++) {
-      if (cmd_buffer->descriptors[i].push_set.layout)
-         vk_descriptor_set_layout_unref(&cmd_buffer->device->vk,
-                                        &cmd_buffer->descriptors[i].push_set.layout->vk);
-      vk_free(&cmd_buffer->device->vk.alloc,
-              cmd_buffer->descriptors[i].push_set.mapped_ptr);
+      /* From the Vulkan 1.0.66 spec:
+       *
+       * "vkAllocateCommandBuffers can be used to create multiple
+       *  command buffers. If the creation of any of those command
+       *  buffers fails, the implementation must destroy all
+       *  successfully created command buffer objects from this
+       *  command, set all entries of the pCommandBuffers array to
+       *  NULL and return the error."
+       */
+      memset(pCommandBuffers, 0,
+             sizeof(*pCommandBuffers) * pAllocateInfo->commandBufferCount);
    }
 
-   vk_command_buffer_finish(&cmd_buffer->vk);
-   vk_free2(&cmd_buffer->device->vk.alloc, &cmd_buffer->vk.pool->alloc,
-            cmd_buffer);
+   return result;
 }
 
-static void
-tu_reset_cmd_buffer(struct vk_command_buffer *vk_cmd_buffer,
-                    UNUSED VkCommandBufferResetFlags flags)
+void
+tu_FreeCommandBuffers(VkDevice device,
+                      VkCommandPool commandPool,
+                      uint32_t commandBufferCount,
+                      const VkCommandBuffer *pCommandBuffers)
 {
-   struct tu_cmd_buffer *cmd_buffer =
-      container_of(vk_cmd_buffer, struct tu_cmd_buffer, vk);
-
-   vk_command_buffer_reset(&cmd_buffer->vk);
-
-   tu_cs_reset(&cmd_buffer->cs);
-   tu_cs_reset(&cmd_buffer->draw_cs);
-   tu_cs_reset(&cmd_buffer->tile_store_cs);
-   tu_cs_reset(&cmd_buffer->draw_epilogue_cs);
-   tu_cs_reset(&cmd_buffer->sub_cs);
-   tu_cs_reset(&cmd_buffer->pre_chain.draw_cs);
-   tu_cs_reset(&cmd_buffer->pre_chain.draw_epilogue_cs);
-
-   tu_autotune_free_results(cmd_buffer->device, &cmd_buffer->renderpass_autotune_results);
-
-   for (unsigned i = 0; i < MAX_BIND_POINTS; i++) {
-      memset(&cmd_buffer->descriptors[i].sets, 0, sizeof(cmd_buffer->descriptors[i].sets));
-      if (cmd_buffer->descriptors[i].push_set.layout) {
-         vk_descriptor_set_layout_unref(&cmd_buffer->device->vk,
-                                        &cmd_buffer->descriptors[i].push_set.layout->vk);
+   for (uint32_t i = 0; i < commandBufferCount; i++) {
+      TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, pCommandBuffers[i]);
+
+      if (cmd_buffer) {
+         if (cmd_buffer->pool) {
+            list_del(&cmd_buffer->pool_link);
+            list_addtail(&cmd_buffer->pool_link,
+                         &cmd_buffer->pool->free_cmd_buffers);
+         } else
+            tu_cmd_buffer_destroy(cmd_buffer);
       }
-      memset(&cmd_buffer->descriptors[i].push_set, 0, sizeof(cmd_buffer->descriptors[i].push_set));
-      cmd_buffer->descriptors[i].push_set.base.type = VK_OBJECT_TYPE_DESCRIPTOR_SET;
-      cmd_buffer->descriptors[i].max_sets_bound = 0;
-      cmd_buffer->descriptors[i].dynamic_bound = 0;
    }
-
-   u_trace_fini(&cmd_buffer->trace);
-   u_trace_init(&cmd_buffer->trace, &cmd_buffer->device->trace_context);
-
-   cmd_buffer->state.max_vbs_bound = 0;
-   cmd_buffer->state.last_prim_params.valid = false;
-
-   cmd_buffer->vsc_initialized = false;
-
-   cmd_buffer->status = TU_CMD_BUFFER_STATUS_INITIAL;
 }
 
-const struct vk_command_buffer_ops tu_cmd_buffer_ops = {
-   .create = tu_create_cmd_buffer,
-   .reset = tu_reset_cmd_buffer,
-   .destroy = tu_cmd_buffer_destroy,
-};
-
-/* Initialize the cache, assuming all necessary flushes have happened but *not*
- * invalidations.
- */
-static void
-tu_cache_init(struct tu_cache_state *cache)
+VkResult
+tu_ResetCommandBuffer(VkCommandBuffer commandBuffer,
+                      VkCommandBufferResetFlags flags)
 {
-   cache->flush_bits = 0;
-   cache->pending_flush_bits = TU_CMD_FLAG_ALL_INVALIDATE;
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   return tu_reset_cmd_buffer(cmd_buffer);
 }
 
-/* Unlike the public entrypoint, this doesn't handle cache tracking, and
- * tracking the CCU state. It's used for the driver to insert its own command
- * buffer in the middle of a submit.
- */
 VkResult
-tu_cmd_buffer_begin(struct tu_cmd_buffer *cmd_buffer,
-                    VkCommandBufferUsageFlags usage_flags)
+tu_BeginCommandBuffer(VkCommandBuffer commandBuffer,
+                      const VkCommandBufferBeginInfo *pBeginInfo)
 {
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   VkResult result = VK_SUCCESS;
+
    if (cmd_buffer->status != TU_CMD_BUFFER_STATUS_INITIAL) {
       /* If the command buffer has already been resetted with
        * vkResetCommandBuffer, no need to do it again.
        */
-      tu_reset_cmd_buffer(&cmd_buffer->vk, 0);
+      result = tu_reset_cmd_buffer(cmd_buffer);
+      if (result != VK_SUCCESS)
+         return result;
    }
 
    memset(&cmd_buffer->state, 0, sizeof(cmd_buffer->state));
-   cmd_buffer->state.index_size = 0xff; /* dirty restart index */
-   cmd_buffer->state.line_mode = RECTANGULAR;
-   cmd_buffer->state.gmem_layout = TU_GMEM_LAYOUT_COUNT; /* dirty value */
-
-   tu_cache_init(&cmd_buffer->state.cache);
-   tu_cache_init(&cmd_buffer->state.renderpass_cache);
-   cmd_buffer->usage_flags = usage_flags;
+   cmd_buffer->usage_flags = pBeginInfo->flags;
 
    tu_cs_begin(&cmd_buffer->cs);
-   tu_cs_begin(&cmd_buffer->draw_cs);
-   tu_cs_begin(&cmd_buffer->draw_epilogue_cs);
 
-   cmd_buffer->status = TU_CMD_BUFFER_STATUS_RECORDING;
-   return VK_SUCCESS;
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_BeginCommandBuffer(VkCommandBuffer commandBuffer,
-                      const VkCommandBufferBeginInfo *pBeginInfo)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
-   VkResult result = tu_cmd_buffer_begin(cmd_buffer, pBeginInfo->flags);
-   if (result != VK_SUCCESS)
-      return result;
+   cmd_buffer->marker_seqno = 0;
+   cmd_buffer->scratch_seqno = 0;
 
    /* setup initial configuration into command buffer */
-   if (cmd_buffer->vk.level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
-      trace_start_cmd_buffer(&cmd_buffer->trace, &cmd_buffer->cs);
-
+   if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
       switch (cmd_buffer->queue_family_index) {
       case TU_QUEUE_GENERAL:
          tu6_init_hw(cmd_buffer, &cmd_buffer->cs);
@@ -1799,212 +1602,35 @@ tu_BeginCommandBuffer(VkCommandBuffer commandBuffer,
       default:
          break;
       }
-   } else if (cmd_buffer->vk.level == VK_COMMAND_BUFFER_LEVEL_SECONDARY) {
-      const bool pass_continue =
-         pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT;
-
-      trace_start_cmd_buffer(&cmd_buffer->trace,
-            pass_continue ? &cmd_buffer->draw_cs : &cmd_buffer->cs);
-
-      assert(pBeginInfo->pInheritanceInfo);
-
-      cmd_buffer->inherited_pipeline_statistics =
-         pBeginInfo->pInheritanceInfo->pipelineStatistics;
-
-      vk_foreach_struct_const(ext, pBeginInfo->pInheritanceInfo) {
-         switch (ext->sType) {
-         case VK_STRUCTURE_TYPE_COMMAND_BUFFER_INHERITANCE_CONDITIONAL_RENDERING_INFO_EXT: {
-            const VkCommandBufferInheritanceConditionalRenderingInfoEXT *cond_rend = (void *) ext;
-            cmd_buffer->state.predication_active = cond_rend->conditionalRenderingEnable;
-            break;
-         default:
-            break;
-         }
-         }
-      }
-
-      if (pass_continue) {
-         const VkCommandBufferInheritanceRenderingInfo *rendering_info =
-            vk_find_struct_const(pBeginInfo->pInheritanceInfo->pNext,
-                                 COMMAND_BUFFER_INHERITANCE_RENDERING_INFO);
-
-         if (unlikely(cmd_buffer->device->instance->debug_flags & TU_DEBUG_DYNAMIC)) {
-            rendering_info =
-               vk_get_command_buffer_inheritance_rendering_info(cmd_buffer->vk.level,
-                                                                pBeginInfo);
-         }
-
-         if (rendering_info) {
-            tu_setup_dynamic_inheritance(cmd_buffer, rendering_info);
-            cmd_buffer->state.pass = &cmd_buffer->dynamic_pass;
-            cmd_buffer->state.subpass = &cmd_buffer->dynamic_subpass;
-         } else {
-            cmd_buffer->state.pass = tu_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
-            cmd_buffer->state.subpass =
-               &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
-         }
-
-         /* We can't set the gmem layout here, because the state.pass only has
-          * to be compatible (same formats/sample counts) with the primary's
-          * renderpass, rather than exactly equal.
-          */
-
-         tu_lrz_begin_secondary_cmdbuf(cmd_buffer);
-      } else {
-         /* When executing in the middle of another command buffer, the CCU
-          * state is unknown.
-          */
-         cmd_buffer->state.ccu_state = TU_CMD_CCU_UNKNOWN;
-      }
    }
 
-   return VK_SUCCESS;
-}
-
-static void
-tu6_emit_vertex_strides(struct tu_cmd_buffer *cmd, unsigned num_vbs)
-{
-   struct tu_cs cs;
-   cmd->state.dynamic_state[TU_DYNAMIC_STATE_VB_STRIDE].iova =
-      tu_cs_draw_state(&cmd->sub_cs, &cs, 2 * num_vbs).iova;
-
-   for (uint32_t i = 0; i < num_vbs; i++)
-      tu_cs_emit_regs(&cs, A6XX_VFD_FETCH_STRIDE(i, cmd->state.vb[i].stride));
-
-   cmd->state.dirty |= TU_CMD_DIRTY_VB_STRIDE;
-}
-
-static struct tu_cs
-tu_cmd_dynamic_state(struct tu_cmd_buffer *cmd, uint32_t id, uint32_t size)
-{
-   struct tu_cs cs;
-
-   assert(id < ARRAY_SIZE(cmd->state.dynamic_state));
-   cmd->state.dynamic_state[id] = tu_cs_draw_state(&cmd->sub_cs, &cs, size);
-
-   /* note: this also avoids emitting draw states before renderpass clears,
-    * which may use the 3D clear path (for MSAA cases)
-    */
-   if (cmd->state.dirty & TU_CMD_DIRTY_DRAW_STATE)
-      return cs;
-
-   tu_cs_emit_pkt7(&cmd->draw_cs, CP_SET_DRAW_STATE, 3);
-   tu_cs_emit_draw_state(&cmd->draw_cs, TU_DRAW_STATE_DYNAMIC + id, cmd->state.dynamic_state[id]);
-
-   return cs;
-}
-
-static void
-tu_cmd_end_dynamic_state(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
-                         uint32_t id)
-{
-   assert(id < ARRAY_SIZE(cmd->state.dynamic_state));
-   cmd->state.dynamic_state[id] = tu_cs_end_draw_state(&cmd->sub_cs, cs);
-
-   /* note: this also avoids emitting draw states before renderpass clears,
-    * which may use the 3D clear path (for MSAA cases)
-    */
-   if (cmd->state.dirty & TU_CMD_DIRTY_DRAW_STATE)
-      return;
-
-   tu_cs_emit_pkt7(&cmd->draw_cs, CP_SET_DRAW_STATE, 3);
-   tu_cs_emit_draw_state(&cmd->draw_cs, TU_DRAW_STATE_DYNAMIC + id, cmd->state.dynamic_state[id]);
-}
-
-static void
-tu_update_num_vbs(struct tu_cmd_buffer *cmd, unsigned num_vbs)
-{
-   /* the vertex_buffers draw state always contains all the currently
-    * bound vertex buffers. update its size to only emit the vbs which
-    * are actually used by the pipeline
-    * note there is a HW optimization which makes it so the draw state
-    * is not re-executed completely when only the size changes
-    */
-   if (cmd->state.vertex_buffers.size != num_vbs * 4) {
-      cmd->state.vertex_buffers.size = num_vbs * 4;
-      cmd->state.dirty |= TU_CMD_DIRTY_VERTEX_BUFFERS;
-   }
-
-   if (cmd->state.dynamic_state[TU_DYNAMIC_STATE_VB_STRIDE].size != num_vbs * 2) {
-      cmd->state.dynamic_state[TU_DYNAMIC_STATE_VB_STRIDE].size = num_vbs * 2;
-      cmd->state.dirty |= TU_CMD_DIRTY_VB_STRIDE;
-   }
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetVertexInputEXT(VkCommandBuffer commandBuffer,
-                        uint32_t vertexBindingDescriptionCount,
-                        const VkVertexInputBindingDescription2EXT *pVertexBindingDescriptions,
-                        uint32_t vertexAttributeDescriptionCount,
-                        const VkVertexInputAttributeDescription2EXT *pVertexAttributeDescriptions)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs;
-
-   unsigned num_vbs = 0;
-   for (unsigned i = 0; i < vertexBindingDescriptionCount; i++) {
-      const VkVertexInputBindingDescription2EXT *binding =
-         &pVertexBindingDescriptions[i];
-      num_vbs = MAX2(num_vbs, binding->binding + 1);
-      cmd->state.vb[binding->binding].stride = binding->stride;
-   }
-
-   tu6_emit_vertex_strides(cmd, num_vbs);
-   tu_update_num_vbs(cmd, num_vbs);
+   cmd_buffer->status = TU_CMD_BUFFER_STATUS_RECORDING;
 
-   tu_cs_begin_sub_stream(&cmd->sub_cs, TU6_EMIT_VERTEX_INPUT_MAX_DWORDS, &cs);
-   tu6_emit_vertex_input(&cs, vertexBindingDescriptionCount,
-                         pVertexBindingDescriptions,
-                         vertexAttributeDescriptionCount,
-                         pVertexAttributeDescriptions);
-   tu_cmd_end_dynamic_state(cmd, &cs, TU_DYNAMIC_STATE_VERTEX_INPUT);
+   return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdBindVertexBuffers2EXT(VkCommandBuffer commandBuffer,
-                            uint32_t firstBinding,
-                            uint32_t bindingCount,
-                            const VkBuffer* pBuffers,
-                            const VkDeviceSize* pOffsets,
-                            const VkDeviceSize* pSizes,
-                            const VkDeviceSize* pStrides)
+void
+tu_CmdBindVertexBuffers(VkCommandBuffer commandBuffer,
+                        uint32_t firstBinding,
+                        uint32_t bindingCount,
+                        const VkBuffer *pBuffers,
+                        const VkDeviceSize *pOffsets)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs;
 
-   cmd->state.max_vbs_bound = MAX2(
-      cmd->state.max_vbs_bound, firstBinding + bindingCount);
-
-   cmd->state.vertex_buffers.iova =
-      tu_cs_draw_state(&cmd->sub_cs, &cs, 4 * cmd->state.max_vbs_bound).iova;
+   assert(firstBinding + bindingCount <= MAX_VBS);
 
    for (uint32_t i = 0; i < bindingCount; i++) {
-      if (pBuffers[i] == VK_NULL_HANDLE) {
-         cmd->state.vb[firstBinding + i].base = 0;
-         cmd->state.vb[firstBinding + i].size = 0;
-      } else {
-         struct tu_buffer *buf = tu_buffer_from_handle(pBuffers[i]);
-         cmd->state.vb[firstBinding + i].base = buf->iova + pOffsets[i];
-         cmd->state.vb[firstBinding + i].size = pSizes ? pSizes[i] : (buf->vk.size - pOffsets[i]);
-      }
-
-      if (pStrides)
-         cmd->state.vb[firstBinding + i].stride = pStrides[i];
-   }
-
-   for (uint32_t i = 0; i < cmd->state.max_vbs_bound; i++) {
-      tu_cs_emit_regs(&cs,
-                      A6XX_VFD_FETCH_BASE(i, .qword = cmd->state.vb[i].base),
-                      A6XX_VFD_FETCH_SIZE(i, cmd->state.vb[i].size));
+      cmd->state.vb.buffers[firstBinding + i] =
+         tu_buffer_from_handle(pBuffers[i]);
+      cmd->state.vb.offsets[firstBinding + i] = pOffsets[i];
    }
 
+   /* VB states depend on VkPipelineVertexInputStateCreateInfo */
    cmd->state.dirty |= TU_CMD_DIRTY_VERTEX_BUFFERS;
-
-   if (pStrides)
-      tu6_emit_vertex_strides(cmd, cmd->state.max_vbs_bound);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdBindIndexBuffer(VkCommandBuffer commandBuffer,
                       VkBuffer buffer,
                       VkDeviceSize offset,
@@ -2013,42 +1639,31 @@ tu_CmdBindIndexBuffer(VkCommandBuffer commandBuffer,
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
    TU_FROM_HANDLE(tu_buffer, buf, buffer);
 
+   /* initialize/update the restart index */
+   if (!cmd->state.index_buffer || cmd->state.index_type != indexType) {
+      struct tu_cs *draw_cs = &cmd->draw_cs;
+      VkResult result = tu_cs_reserve_space(cmd->device, draw_cs, 2);
+      if (result != VK_SUCCESS) {
+         cmd->record_result = result;
+         return;
+      }
 
+      tu6_emit_restart_index(
+         draw_cs, indexType == VK_INDEX_TYPE_UINT32 ? 0xffffffff : 0xffff);
 
-   uint32_t index_size, index_shift, restart_index;
-
-   switch (indexType) {
-   case VK_INDEX_TYPE_UINT16:
-      index_size = INDEX4_SIZE_16_BIT;
-      index_shift = 1;
-      restart_index = 0xffff;
-      break;
-   case VK_INDEX_TYPE_UINT32:
-      index_size = INDEX4_SIZE_32_BIT;
-      index_shift = 2;
-      restart_index = 0xffffffff;
-      break;
-   case VK_INDEX_TYPE_UINT8_EXT:
-      index_size = INDEX4_SIZE_8_BIT;
-      index_shift = 0;
-      restart_index = 0xff;
-      break;
-   default:
-      unreachable("invalid VkIndexType");
+      tu_cs_sanity_check(draw_cs);
    }
 
-   /* initialize/update the restart index */
-   if (cmd->state.index_size != index_size)
-      tu_cs_emit_regs(&cmd->draw_cs, A6XX_PC_RESTART_INDEX(restart_index));
-
-   assert(buf->vk.size >= offset);
+   /* track the BO */
+   if (cmd->state.index_buffer != buf)
+      tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
 
-   cmd->state.index_va = buf->iova + offset;
-   cmd->state.max_index_count = (buf->vk.size - offset) >> index_shift;
-   cmd->state.index_size = index_size;
+   cmd->state.index_buffer = buf;
+   cmd->state.index_offset = offset;
+   cmd->state.index_type = indexType;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdBindDescriptorSets(VkCommandBuffer commandBuffer,
                          VkPipelineBindPoint pipelineBindPoint,
                          VkPipelineLayout _layout,
@@ -2058,404 +1673,9 @@ tu_CmdBindDescriptorSets(VkCommandBuffer commandBuffer,
                          uint32_t dynamicOffsetCount,
                          const uint32_t *pDynamicOffsets)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_pipeline_layout, layout, _layout);
-   unsigned dyn_idx = 0;
-
-   struct tu_descriptor_state *descriptors_state =
-      tu_get_descriptors_state(cmd, pipelineBindPoint);
-
-   descriptors_state->max_sets_bound =
-      MAX2(descriptors_state->max_sets_bound, firstSet + descriptorSetCount);
-
-   for (unsigned i = 0; i < descriptorSetCount; ++i) {
-      unsigned idx = i + firstSet;
-      TU_FROM_HANDLE(tu_descriptor_set, set, pDescriptorSets[i]);
-
-      descriptors_state->sets[idx] = set;
-
-      if (!set)
-         continue;
-
-      if (!set->layout->dynamic_offset_size)
-         continue;
-
-      uint32_t *src = set->dynamic_descriptors;
-      uint32_t *dst = descriptors_state->dynamic_descriptors +
-         layout->set[idx].dynamic_offset_start / 4;
-      for (unsigned j = 0; j < set->layout->binding_count; j++) {
-         struct tu_descriptor_set_binding_layout *binding =
-            &set->layout->binding[j];
-         if (binding->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC ||
-             binding->type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC) {
-            for (unsigned k = 0; k < binding->array_size; k++, dyn_idx++) {
-               assert(dyn_idx < dynamicOffsetCount);
-               uint32_t offset = pDynamicOffsets[dyn_idx];
-               memcpy(dst, src, binding->size);
-
-               if (binding->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC) {
-                  /* Note: we can assume here that the addition won't roll
-                   * over and change the SIZE field.
-                   */
-                  uint64_t va = src[0] | ((uint64_t)src[1] << 32);
-                  va += offset;
-                  dst[0] = va;
-                  dst[1] = va >> 32;
-               } else {
-                  uint32_t *dst_desc = dst;
-                  for (unsigned i = 0;
-                       i < binding->size / (4 * A6XX_TEX_CONST_DWORDS);
-                       i++, dst_desc += A6XX_TEX_CONST_DWORDS) {
-                     /* Note: A6XX_TEX_CONST_5_DEPTH is always 0 */
-                     uint64_t va = dst_desc[4] | ((uint64_t)dst_desc[5] << 32);
-                     va += offset;
-                     dst_desc[4] = va;
-                     dst_desc[5] = va >> 32;
-                  }
-               }
-
-               dst += binding->size / 4;
-               src += binding->size / 4;
-            }
-         }
-      }
-   }
-   assert(dyn_idx == dynamicOffsetCount);
-
-   uint32_t sp_bindless_base_reg, hlsq_bindless_base_reg, hlsq_invalidate_value;
-   uint64_t addr[MAX_SETS] = {};
-   uint64_t dynamic_addr = 0;
-   struct tu_cs *cs, state_cs;
-
-   for (uint32_t i = 0; i < descriptors_state->max_sets_bound; i++) {
-      struct tu_descriptor_set *set = descriptors_state->sets[i];
-      if (set)
-         addr[i] = set->va | 3;
-   }
-
-   if (layout->dynamic_offset_size) {
-      /* allocate and fill out dynamic descriptor set */
-      struct tu_cs_memory dynamic_desc_set;
-      VkResult result = tu_cs_alloc(&cmd->sub_cs,
-                                    layout->dynamic_offset_size / (4 * A6XX_TEX_CONST_DWORDS),
-                                    A6XX_TEX_CONST_DWORDS, &dynamic_desc_set);
-      if (result != VK_SUCCESS) {
-         vk_command_buffer_set_error(&cmd->vk, result);
-         return;
-      }
-
-      memcpy(dynamic_desc_set.map, descriptors_state->dynamic_descriptors,
-             layout->dynamic_offset_size);
-      dynamic_addr = dynamic_desc_set.iova | 3;
-      descriptors_state->dynamic_bound = true;
-   }
-
-   if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS) {
-      sp_bindless_base_reg = REG_A6XX_SP_BINDLESS_BASE(0);
-      hlsq_bindless_base_reg = REG_A6XX_HLSQ_BINDLESS_BASE(0);
-      hlsq_invalidate_value = A6XX_HLSQ_INVALIDATE_CMD_GFX_BINDLESS(0x1f);
-
-      cmd->state.desc_sets =
-         tu_cs_draw_state(&cmd->sub_cs, &state_cs,
-                          4 + 4 * descriptors_state->max_sets_bound +
-                             (descriptors_state->dynamic_bound ? 6 : 0));
-      cmd->state.dirty |= TU_CMD_DIRTY_DESC_SETS_LOAD;
-      cs = &state_cs;
-   } else {
-      assert(pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE);
-
-      sp_bindless_base_reg = REG_A6XX_SP_CS_BINDLESS_BASE(0);
-      hlsq_bindless_base_reg = REG_A6XX_HLSQ_CS_BINDLESS_BASE(0);
-      hlsq_invalidate_value = A6XX_HLSQ_INVALIDATE_CMD_CS_BINDLESS(0x1f);
-
-      cmd->state.dirty |= TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD;
-      cs = &cmd->cs;
-   }
-
-   tu_cs_emit_pkt4(cs, sp_bindless_base_reg, 2 * descriptors_state->max_sets_bound);
-   tu_cs_emit_array(cs, (const uint32_t*) addr, 2 * descriptors_state->max_sets_bound);
-   tu_cs_emit_pkt4(cs, hlsq_bindless_base_reg, 2 * descriptors_state->max_sets_bound);
-   tu_cs_emit_array(cs, (const uint32_t*) addr, 2 * descriptors_state->max_sets_bound);
-
-   /* Dynamic descriptors get the last descriptor set. */
-   if (descriptors_state->dynamic_bound) {
-      tu_cs_emit_pkt4(cs, sp_bindless_base_reg + 4 * 2, 2);
-      tu_cs_emit_qw(cs, dynamic_addr);
-      tu_cs_emit_pkt4(cs, hlsq_bindless_base_reg + 4 * 2, 2);
-      tu_cs_emit_qw(cs, dynamic_addr);
-   }
-
-   tu_cs_emit_regs(cs, A6XX_HLSQ_INVALIDATE_CMD(.dword = hlsq_invalidate_value));
-
-   if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS) {
-      assert(cs->cur == cs->end); /* validate draw state size */
-      /* note: this also avoids emitting draw states before renderpass clears,
-       * which may use the 3D clear path (for MSAA cases)
-       */
-      if (!(cmd->state.dirty & TU_CMD_DIRTY_DRAW_STATE)) {
-         tu_cs_emit_pkt7(&cmd->draw_cs, CP_SET_DRAW_STATE, 3);
-         tu_cs_emit_draw_state(&cmd->draw_cs, TU_DRAW_STATE_DESC_SETS, cmd->state.desc_sets);
-      }
-   }
 }
 
-static enum VkResult
-tu_push_descriptor_set_update_layout(struct tu_device *device,
-                                     struct tu_descriptor_set *set,
-                                     struct tu_descriptor_set_layout *layout)
-{
-   if (set->layout == layout)
-      return VK_SUCCESS;
-
-   if (set->layout)
-      vk_descriptor_set_layout_unref(&device->vk, &set->layout->vk);
-   vk_descriptor_set_layout_ref(&layout->vk);
-   set->layout = layout;
-
-   if (set->host_size < layout->size) {
-      void *new_buf =
-         vk_realloc(&device->vk.alloc, set->mapped_ptr, layout->size, 8,
-                    VK_QUERY_SCOPE_COMMAND_BUFFER_KHR);
-      if (!new_buf)
-         return VK_ERROR_OUT_OF_HOST_MEMORY;
-      set->mapped_ptr = new_buf;
-      set->host_size = layout->size;
-   }
-   return VK_SUCCESS;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdPushDescriptorSetKHR(VkCommandBuffer commandBuffer,
-                           VkPipelineBindPoint pipelineBindPoint,
-                           VkPipelineLayout _layout,
-                           uint32_t _set,
-                           uint32_t descriptorWriteCount,
-                           const VkWriteDescriptorSet *pDescriptorWrites)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_pipeline_layout, pipe_layout, _layout);
-   struct tu_descriptor_set_layout *layout = pipe_layout->set[_set].layout;
-   struct tu_descriptor_set *set =
-      &tu_get_descriptors_state(cmd, pipelineBindPoint)->push_set;
-
-   struct tu_cs_memory set_mem;
-   VkResult result = tu_cs_alloc(&cmd->sub_cs,
-                                 DIV_ROUND_UP(layout->size, A6XX_TEX_CONST_DWORDS * 4),
-                                 A6XX_TEX_CONST_DWORDS, &set_mem);
-   if (result != VK_SUCCESS) {
-      vk_command_buffer_set_error(&cmd->vk, result);
-      return;
-   }
-
-   result = tu_push_descriptor_set_update_layout(cmd->device, set, layout);
-   if (result != VK_SUCCESS) {
-      vk_command_buffer_set_error(&cmd->vk, result);
-      return;
-   }
-
-   tu_update_descriptor_sets(cmd->device, tu_descriptor_set_to_handle(set),
-                             descriptorWriteCount, pDescriptorWrites, 0, NULL);
-
-   memcpy(set_mem.map, set->mapped_ptr, layout->size);
-   set->va = set_mem.iova;
-
-   tu_CmdBindDescriptorSets(commandBuffer, pipelineBindPoint, _layout, _set,
-                            1, (VkDescriptorSet[]) { tu_descriptor_set_to_handle(set) },
-                            0, NULL);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdPushDescriptorSetWithTemplateKHR(VkCommandBuffer commandBuffer,
-                                       VkDescriptorUpdateTemplate descriptorUpdateTemplate,
-                                       VkPipelineLayout _layout,
-                                       uint32_t _set,
-                                       const void* pData)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_pipeline_layout, pipe_layout, _layout);
-   TU_FROM_HANDLE(tu_descriptor_update_template, templ, descriptorUpdateTemplate);
-   struct tu_descriptor_set_layout *layout = pipe_layout->set[_set].layout;
-   struct tu_descriptor_set *set =
-      &tu_get_descriptors_state(cmd, templ->bind_point)->push_set;
-
-   struct tu_cs_memory set_mem;
-   VkResult result = tu_cs_alloc(&cmd->sub_cs,
-                                 DIV_ROUND_UP(layout->size, A6XX_TEX_CONST_DWORDS * 4),
-                                 A6XX_TEX_CONST_DWORDS, &set_mem);
-   if (result != VK_SUCCESS) {
-      vk_command_buffer_set_error(&cmd->vk, result);
-      return;
-   }
-
-   result = tu_push_descriptor_set_update_layout(cmd->device, set, layout);
-   if (result != VK_SUCCESS) {
-      vk_command_buffer_set_error(&cmd->vk, result);
-      return;
-   }
-
-   tu_update_descriptor_set_with_template(cmd->device, set, descriptorUpdateTemplate, pData);
-
-   memcpy(set_mem.map, set->mapped_ptr, layout->size);
-   set->va = set_mem.iova;
-
-   tu_CmdBindDescriptorSets(commandBuffer, templ->bind_point, _layout, _set,
-                            1, (VkDescriptorSet[]) { tu_descriptor_set_to_handle(set) },
-                            0, NULL);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdBindTransformFeedbackBuffersEXT(VkCommandBuffer commandBuffer,
-                                      uint32_t firstBinding,
-                                      uint32_t bindingCount,
-                                      const VkBuffer *pBuffers,
-                                      const VkDeviceSize *pOffsets,
-                                      const VkDeviceSize *pSizes)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   /* using COND_REG_EXEC for xfb commands matches the blob behavior
-    * presumably there isn't any benefit using a draw state when the
-    * condition is (SYSMEM | BINNING)
-    */
-   tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
-                          CP_COND_REG_EXEC_0_SYSMEM |
-                          CP_COND_REG_EXEC_0_BINNING);
-
-   for (uint32_t i = 0; i < bindingCount; i++) {
-      TU_FROM_HANDLE(tu_buffer, buf, pBuffers[i]);
-      uint64_t iova = buf->iova + pOffsets[i];
-      uint32_t size = buf->bo->size - (iova - buf->bo->iova);
-      uint32_t idx = i + firstBinding;
-
-      if (pSizes && pSizes[i] != VK_WHOLE_SIZE)
-         size = pSizes[i];
-
-      /* BUFFER_BASE is 32-byte aligned, add remaining offset to BUFFER_OFFSET */
-      uint32_t offset = iova & 0x1f;
-      iova &= ~(uint64_t) 0x1f;
-
-      tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_BASE(idx), 3);
-      tu_cs_emit_qw(cs, iova);
-      tu_cs_emit(cs, size + offset);
-
-      cmd->state.streamout_offset[idx] = offset;
-   }
-
-   tu_cond_exec_end(cs);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer,
-                                uint32_t firstCounterBuffer,
-                                uint32_t counterBufferCount,
-                                const VkBuffer *pCounterBuffers,
-                                const VkDeviceSize *pCounterBufferOffsets)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
-                          CP_COND_REG_EXEC_0_SYSMEM |
-                          CP_COND_REG_EXEC_0_BINNING);
-
-   tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(false));
-
-   /* TODO: only update offset for active buffers */
-   for (uint32_t i = 0; i < IR3_MAX_SO_BUFFERS; i++)
-      tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_OFFSET(i, cmd->state.streamout_offset[i]));
-
-   for (uint32_t i = 0; i < (pCounterBuffers ? counterBufferCount : 0); i++) {
-      uint32_t idx = firstCounterBuffer + i;
-      uint32_t offset = cmd->state.streamout_offset[idx];
-      uint64_t counter_buffer_offset = pCounterBufferOffsets ? pCounterBufferOffsets[i] : 0u;
-
-      if (!pCounterBuffers[i])
-         continue;
-
-      TU_FROM_HANDLE(tu_buffer, buf, pCounterBuffers[i]);
-
-      tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
-      tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(idx)) |
-                     CP_MEM_TO_REG_0_UNK31 |
-                     CP_MEM_TO_REG_0_CNT(1));
-      tu_cs_emit_qw(cs, buf->iova + counter_buffer_offset);
-
-      if (offset) {
-         tu_cs_emit_pkt7(cs, CP_REG_RMW, 3);
-         tu_cs_emit(cs, CP_REG_RMW_0_DST_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(idx)) |
-                        CP_REG_RMW_0_SRC1_ADD);
-         tu_cs_emit(cs, 0xffffffff);
-         tu_cs_emit(cs, offset);
-      }
-   }
-
-   tu_cond_exec_end(cs);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdEndTransformFeedbackEXT(VkCommandBuffer commandBuffer,
-                              uint32_t firstCounterBuffer,
-                              uint32_t counterBufferCount,
-                              const VkBuffer *pCounterBuffers,
-                              const VkDeviceSize *pCounterBufferOffsets)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
-                          CP_COND_REG_EXEC_0_SYSMEM |
-                          CP_COND_REG_EXEC_0_BINNING);
-
-   tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(true));
-
-   /* TODO: only flush buffers that need to be flushed */
-   for (uint32_t i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
-      /* note: FLUSH_BASE is always the same, so it could go in init_hw()? */
-      tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_FLUSH_BASE(i), 2);
-      tu_cs_emit_qw(cs, global_iova(cmd, flush_base[i]));
-      tu6_emit_event_write(cmd, cs, FLUSH_SO_0 + i);
-   }
-
-   for (uint32_t i = 0; i < (pCounterBuffers ? counterBufferCount : 0); i++) {
-      uint32_t idx = firstCounterBuffer + i;
-      uint32_t offset = cmd->state.streamout_offset[idx];
-      uint64_t counter_buffer_offset = pCounterBufferOffsets ? pCounterBufferOffsets[i] : 0u;
-
-      if (!pCounterBuffers[i])
-         continue;
-
-      TU_FROM_HANDLE(tu_buffer, buf, pCounterBuffers[i]);
-
-      /* VPC_SO_FLUSH_BASE has dwords counter, but counter should be in bytes */
-      tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
-      tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
-                     CP_MEM_TO_REG_0_SHIFT_BY_2 |
-                     0x40000 | /* ??? */
-                     CP_MEM_TO_REG_0_UNK31 |
-                     CP_MEM_TO_REG_0_CNT(1));
-      tu_cs_emit_qw(cs, global_iova(cmd, flush_base[idx]));
-
-      if (offset) {
-         tu_cs_emit_pkt7(cs, CP_REG_RMW, 3);
-         tu_cs_emit(cs, CP_REG_RMW_0_DST_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
-                        CP_REG_RMW_0_SRC1_ADD);
-         tu_cs_emit(cs, 0xffffffff);
-         tu_cs_emit(cs, -offset);
-      }
-
-      tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
-      tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
-                     CP_REG_TO_MEM_0_CNT(1));
-      tu_cs_emit_qw(cs, buf->iova + counter_buffer_offset);
-   }
-
-   tu_cond_exec_end(cs);
-
-   cmd->state.rp.xfb_used = true;
-}
-
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdPushConstants(VkCommandBuffer commandBuffer,
                     VkPipelineLayout layout,
                     VkShaderStageFlags stageFlags,
@@ -2463,63 +1683,38 @@ tu_CmdPushConstants(VkCommandBuffer commandBuffer,
                     uint32_t size,
                     const void *pValues)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   memcpy((void*) cmd->push_constants + offset, pValues, size);
-   cmd->state.dirty |= TU_CMD_DIRTY_SHADER_CONSTS;
 }
 
-/* Flush everything which has been made available but we haven't actually
- * flushed yet.
- */
-static void
-tu_flush_all_pending(struct tu_cache_state *cache)
-{
-   cache->flush_bits |= cache->pending_flush_bits & TU_CMD_FLAG_ALL_FLUSH;
-   cache->pending_flush_bits &= ~TU_CMD_FLAG_ALL_FLUSH;
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_EndCommandBuffer(VkCommandBuffer commandBuffer)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
 
-   /* We currently flush CCU at the end of the command buffer, like
-    * what the blob does. There's implicit synchronization around every
-    * vkQueueSubmit, but the kernel only flushes the UCHE, and we don't
-    * know yet if this command buffer will be the last in the submit so we
-    * have to defensively flush everything else.
-    *
-    * TODO: We could definitely do better than this, since these flushes
-    * aren't required by Vulkan, but we'd need kernel support to do that.
-    * Ideally, we'd like the kernel to flush everything afterwards, so that we
-    * wouldn't have to do any flushes here, and when submitting multiple
-    * command buffers there wouldn't be any unnecessary flushes in between.
-    */
-   if (cmd_buffer->state.pass) {
-      tu_flush_all_pending(&cmd_buffer->state.renderpass_cache);
-      tu_emit_cache_flush_renderpass(cmd_buffer, &cmd_buffer->draw_cs);
+   if (cmd_buffer->scratch_seqno) {
+      tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->scratch_bo,
+                     MSM_SUBMIT_BO_WRITE);
+   }
 
-      trace_end_cmd_buffer(&cmd_buffer->trace, &cmd_buffer->draw_cs, cmd_buffer);
-   } else {
-      tu_flush_all_pending(&cmd_buffer->state.cache);
-      cmd_buffer->state.cache.flush_bits |=
-         TU_CMD_FLAG_CCU_FLUSH_COLOR |
-         TU_CMD_FLAG_CCU_FLUSH_DEPTH;
-      tu_emit_cache_flush(cmd_buffer, &cmd_buffer->cs);
+   for (uint32_t i = 0; i < cmd_buffer->draw_cs.bo_count; i++) {
+      tu_bo_list_add(&cmd_buffer->bo_list, cmd_buffer->draw_cs.bos[i],
+                     MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+   }
 
-      trace_end_cmd_buffer(&cmd_buffer->trace, &cmd_buffer->cs, cmd_buffer);
+   for (uint32_t i = 0; i < cmd_buffer->tile_cs.bo_count; i++) {
+      tu_bo_list_add(&cmd_buffer->bo_list, cmd_buffer->tile_cs.bos[i],
+                     MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
    }
 
    tu_cs_end(&cmd_buffer->cs);
-   tu_cs_end(&cmd_buffer->draw_cs);
-   tu_cs_end(&cmd_buffer->draw_epilogue_cs);
+
+   assert(!cmd_buffer->state.attachments);
 
    cmd_buffer->status = TU_CMD_BUFFER_STATUS_EXECUTABLE;
 
-   return vk_command_buffer_get_record_result(&cmd_buffer->vk);
+   return cmd_buffer->record_result;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdBindPipeline(VkCommandBuffer commandBuffer,
                    VkPipelineBindPoint pipelineBindPoint,
                    VkPipeline _pipeline)
@@ -2527,2300 +1722,654 @@ tu_CmdBindPipeline(VkCommandBuffer commandBuffer,
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
    TU_FROM_HANDLE(tu_pipeline, pipeline, _pipeline);
 
-   if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE) {
-      cmd->state.compute_pipeline = pipeline;
-      tu_cs_emit_state_ib(&cmd->cs, pipeline->program.state);
-      return;
-   }
-
-   assert(pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS);
-
-   cmd->state.pipeline = pipeline;
-   cmd->state.dirty |= TU_CMD_DIRTY_DESC_SETS_LOAD | TU_CMD_DIRTY_SHADER_CONSTS |
-                       TU_CMD_DIRTY_LRZ | TU_CMD_DIRTY_VS_PARAMS;
-
-   if (pipeline->output.feedback_loop_may_involve_textures &&
-       !cmd->state.rp.disable_gmem) {
-      /* VK_EXT_attachment_feedback_loop_layout allows feedback loop to involve
-       * not only input attachments but also sampled images or image resources.
-       * But we cannot just patch gmem for image in the descriptors.
-       *
-       * At the moment, in context of DXVK, it is expected that only a few
-       * drawcalls in a frame would use feedback loop and they would be wrapped
-       * in their own renderpasses, so it should be ok to force sysmem.
-       *
-       * However, there are two further possible optimizations if need would
-       * arise for other translation layer:
-       * - Tiling could be enabled if we ensure that there is no barrier in
-       *   the renderpass;
-       * - Check that both pipeline and attachments agree that feedback loop
-       *   is needed.
-       */
-      perf_debug(
-         cmd->device,
-         "Disabling gmem due to VK_EXT_attachment_feedback_loop_layout");
-      cmd->state.rp.disable_gmem = true;
-   }
-
-   if (pipeline->prim_order.sysmem_single_prim_mode &&
-       !cmd->state.rp.sysmem_single_prim_mode) {
-      if (pipeline->output.subpass_feedback_loop_color ||
-          pipeline->output.subpass_feedback_loop_ds) {
-         perf_debug(cmd->device, "single_prim_mode due to feedback loop");
-      } else {
-         perf_debug(cmd->device, "single_prim_mode due to rast order access");
-      }
-      cmd->state.rp.sysmem_single_prim_mode = true;
-   }
-
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   /* note: this also avoids emitting draw states before renderpass clears,
-    * which may use the 3D clear path (for MSAA cases)
-    */
-   if (!(cmd->state.dirty & TU_CMD_DIRTY_DRAW_STATE)) {
-      uint32_t mask = ~pipeline->dynamic_state_mask & BITFIELD_MASK(TU_DYNAMIC_STATE_COUNT);
-
-      tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * (6 + util_bitcount(mask)));
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM_CONFIG, pipeline->program.config_state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM, pipeline->program.state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM_BINNING, pipeline->program.binning_state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_RAST, pipeline->rast.state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PRIM_MODE_SYSMEM, pipeline->prim_order.state_sysmem);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PRIM_MODE_GMEM, pipeline->prim_order.state_gmem);
-
-      u_foreach_bit(i, mask)
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + i, pipeline->dynamic_state[i]);
-   }
-
-   if (pipeline->active_stages & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT) {
-      cmd->state.rp.has_tess = true;
-
-      if (!(pipeline->dynamic_state_mask &
-            BIT(TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS))) {
-         cmd->state.patch_control_points = pipeline->tess.patch_control_points;
-         cmd->state.dirty &= ~TU_CMD_DIRTY_PATCH_CONTROL_POINTS;
-      } else {
-         cmd->state.dirty |= TU_CMD_DIRTY_PATCH_CONTROL_POINTS;
-      }
-   }
-
-   cmd->state.line_mode = pipeline->rast.line_mode;
-   if (!(pipeline->dynamic_state_mask &
-         BIT(TU_DYNAMIC_STATE_PRIMITIVE_TOPOLOGY)))
-      cmd->state.primtype = pipeline->ia.primtype;
-
-   tu6_update_msaa(cmd, pipeline->output.samples);
-
-   if ((pipeline->dynamic_state_mask & BIT(VK_DYNAMIC_STATE_VIEWPORT)) &&
-       (pipeline->viewport.z_negative_one_to_one != cmd->state.z_negative_one_to_one)) {
-      cmd->state.z_negative_one_to_one = pipeline->viewport.z_negative_one_to_one;
-      cmd->state.dirty |= TU_CMD_DIRTY_VIEWPORTS;
-   }
-
-   if (!(pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_VERTEX_INPUT)))
-      tu_update_num_vbs(cmd, pipeline->vi.num_vbs);
-
-#define UPDATE_REG(group, X, Y) {                                    \
-   /* note: would be better to have pipeline bits already masked */  \
-   uint32_t pipeline_bits = pipeline->group.X & pipeline->group.X##_mask; \
-   if ((cmd->state.X & pipeline->group.X##_mask) != pipeline_bits) { \
-      cmd->state.X &= ~pipeline->group.X##_mask;                     \
-      cmd->state.X |= pipeline_bits;                                 \
-      cmd->state.dirty |= TU_CMD_DIRTY_##Y;                          \
-   }                                                                 \
-   if (!(pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_##Y)))  \
-      cmd->state.dirty &= ~TU_CMD_DIRTY_##Y;                         \
-}
-
-   /* these registers can have bits set from both pipeline and dynamic state
-    * this updates the bits set by the pipeline
-    * if the pipeline doesn't use a dynamic state for the register, then
-    * the relevant dirty bit is cleared to avoid overriding the non-dynamic
-    * state with a dynamic state the next draw.
-    */
-   UPDATE_REG(rast, gras_su_cntl, GRAS_SU_CNTL);
-   UPDATE_REG(rast_ds, rb_depth_cntl, RB_DEPTH_CNTL);
-   UPDATE_REG(ds, rb_stencil_cntl, RB_STENCIL_CNTL);
-   UPDATE_REG(rast, pc_raster_cntl, RASTERIZER_DISCARD);
-   UPDATE_REG(rast, vpc_unknown_9107, RASTERIZER_DISCARD);
-   UPDATE_REG(blend, sp_blend_cntl, BLEND);
-   UPDATE_REG(blend, rb_blend_cntl, BLEND);
-
-   for (unsigned i = 0; i < pipeline->blend.num_rts; i++) {
-      if ((cmd->state.rb_mrt_control[i] & pipeline->blend.rb_mrt_control_mask) !=
-          pipeline->blend.rb_mrt_control[i]) {
-         cmd->state.rb_mrt_control[i] &= ~pipeline->blend.rb_mrt_control_mask;
-         cmd->state.rb_mrt_control[i] |= pipeline->blend.rb_mrt_control[i];
-         cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-      }
-
-      if (cmd->state.rb_mrt_blend_control[i] != pipeline->blend.rb_mrt_blend_control[i]) {
-         cmd->state.rb_mrt_blend_control[i] = pipeline->blend.rb_mrt_blend_control[i];
-         cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-      }
-   }
-#undef UPDATE_REG
-
-   if (cmd->state.pipeline_color_write_enable != pipeline->blend.color_write_enable) {
-      cmd->state.pipeline_color_write_enable = pipeline->blend.color_write_enable;
-      cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-   }
-   if (cmd->state.pipeline_blend_enable != pipeline->blend.blend_enable) {
-      cmd->state.pipeline_blend_enable = pipeline->blend.blend_enable;
-      cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-   }
-   if (cmd->state.logic_op_enabled != pipeline->blend.logic_op_enabled) {
-      cmd->state.logic_op_enabled = pipeline->blend.logic_op_enabled;
-      cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-   }
-   if (!(pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_LOGIC_OP)) &&
-       cmd->state.rop_reads_dst != pipeline->blend.rop_reads_dst) {
-      cmd->state.rop_reads_dst = pipeline->blend.rop_reads_dst;
-      cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-   }
-   if (cmd->state.dynamic_state[TU_DYNAMIC_STATE_BLEND].size != pipeline->blend.num_rts * 3 + 4) {
-      cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-   }
-   if (!(pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_BLEND))) {
-      cmd->state.dirty &= ~TU_CMD_DIRTY_BLEND;
+   switch (pipelineBindPoint) {
+   case VK_PIPELINE_BIND_POINT_GRAPHICS:
+      cmd->state.pipeline = pipeline;
+      cmd->state.dirty |= TU_CMD_DIRTY_PIPELINE;
+      break;
+   case VK_PIPELINE_BIND_POINT_COMPUTE:
+      tu_finishme("binding compute pipeline");
+      break;
+   default:
+      unreachable("unrecognized pipeline bind point");
+      break;
    }
-
-   if (pipeline->output.rb_depth_cntl_disable)
-      cmd->state.dirty |= TU_CMD_DIRTY_RB_DEPTH_CNTL;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdSetViewport(VkCommandBuffer commandBuffer,
                   uint32_t firstViewport,
                   uint32_t viewportCount,
                   const VkViewport *pViewports)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   struct tu_cs *draw_cs = &cmd->draw_cs;
 
-   memcpy(&cmd->state.viewport[firstViewport], pViewports, viewportCount * sizeof(*pViewports));
-   cmd->state.max_viewport = MAX2(cmd->state.max_viewport, firstViewport + viewportCount);
+   VkResult result = tu_cs_reserve_space(cmd->device, draw_cs, 12);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
+   }
 
-   /* With VK_EXT_depth_clip_control we have to take into account
-    * negativeOneToOne property of the pipeline, so the viewport calculations
-    * are deferred until it is known.
-    */
-   cmd->state.dirty |= TU_CMD_DIRTY_VIEWPORTS;
+   assert(firstViewport == 0 && viewportCount == 1);
+   tu6_emit_viewport(draw_cs, pViewports);
+
+   tu_cs_sanity_check(draw_cs);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdSetScissor(VkCommandBuffer commandBuffer,
                  uint32_t firstScissor,
                  uint32_t scissorCount,
                  const VkRect2D *pScissors)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs;
+   struct tu_cs *draw_cs = &cmd->draw_cs;
+
+   VkResult result = tu_cs_reserve_space(cmd->device, draw_cs, 3);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
+   }
 
-   memcpy(&cmd->state.scissor[firstScissor], pScissors, scissorCount * sizeof(*pScissors));
-   cmd->state.max_scissor = MAX2(cmd->state.max_scissor, firstScissor + scissorCount);
+   assert(firstScissor == 0 && scissorCount == 1);
+   tu6_emit_scissor(draw_cs, pScissors);
 
-   cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_SCISSOR, 1 + 2 * cmd->state.max_scissor);
-   tu6_emit_scissor(&cs, cmd->state.scissor, cmd->state.max_scissor);
+   tu_cs_sanity_check(draw_cs);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdSetLineWidth(VkCommandBuffer commandBuffer, float lineWidth)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
 
-   cmd->state.gras_su_cntl &= ~A6XX_GRAS_SU_CNTL_LINEHALFWIDTH__MASK;
-   cmd->state.gras_su_cntl |= A6XX_GRAS_SU_CNTL_LINEHALFWIDTH(lineWidth / 2.0f);
+   cmd->state.dynamic.line_width = lineWidth;
 
-   cmd->state.dirty |= TU_CMD_DIRTY_GRAS_SU_CNTL;
+   /* line width depends on VkPipelineRasterizationStateCreateInfo */
+   cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdSetDepthBias(VkCommandBuffer commandBuffer,
                    float depthBiasConstantFactor,
                    float depthBiasClamp,
                    float depthBiasSlopeFactor)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_DEPTH_BIAS, 4);
+   struct tu_cs *draw_cs = &cmd->draw_cs;
+
+   VkResult result = tu_cs_reserve_space(cmd->device, draw_cs, 4);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
+   }
+
+   tu6_emit_depth_bias(draw_cs, depthBiasConstantFactor, depthBiasClamp,
+                       depthBiasSlopeFactor);
 
-   tu6_emit_depth_bias(&cs, depthBiasConstantFactor, depthBiasClamp, depthBiasSlopeFactor);
+   tu_cs_sanity_check(draw_cs);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdSetBlendConstants(VkCommandBuffer commandBuffer,
                         const float blendConstants[4])
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_BLEND_CONSTANTS, 5);
+   struct tu_cs *draw_cs = &cmd->draw_cs;
 
-   tu_cs_emit_pkt4(&cs, REG_A6XX_RB_BLEND_RED_F32, 4);
-   tu_cs_emit_array(&cs, (const uint32_t *) blendConstants, 4);
+   VkResult result = tu_cs_reserve_space(cmd->device, draw_cs, 5);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
+      return;
+   }
+
+   tu6_emit_blend_constants(draw_cs, blendConstants);
+
+   tu_cs_sanity_check(draw_cs);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdSetDepthBounds(VkCommandBuffer commandBuffer,
                      float minDepthBounds,
                      float maxDepthBounds)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_DEPTH_BOUNDS, 3);
-
-   tu_cs_emit_regs(&cs,
-                   A6XX_RB_Z_BOUNDS_MIN(minDepthBounds),
-                   A6XX_RB_Z_BOUNDS_MAX(maxDepthBounds));
 }
 
 void
-update_stencil_mask(uint32_t *value, VkStencilFaceFlags face, uint32_t mask)
-{
-   if (face & VK_STENCIL_FACE_FRONT_BIT)
-      *value = (*value & 0xff00) | (mask & 0xff);
-   if (face & VK_STENCIL_FACE_BACK_BIT)
-      *value = (*value & 0xff) | (mask & 0xff) << 8;
-}
-
-VKAPI_ATTR void VKAPI_CALL
 tu_CmdSetStencilCompareMask(VkCommandBuffer commandBuffer,
                             VkStencilFaceFlags faceMask,
                             uint32_t compareMask)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK, 2);
 
-   update_stencil_mask(&cmd->state.dynamic_stencil_mask, faceMask, compareMask);
+   if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
+      cmd->state.dynamic.stencil_compare_mask.front = compareMask;
+   if (faceMask & VK_STENCIL_FACE_BACK_BIT)
+      cmd->state.dynamic.stencil_compare_mask.back = compareMask;
 
-   tu_cs_emit_regs(&cs, A6XX_RB_STENCILMASK(.dword = cmd->state.dynamic_stencil_mask));
+   /* the front/back compare masks must be updated together */
+   cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdSetStencilWriteMask(VkCommandBuffer commandBuffer,
                           VkStencilFaceFlags faceMask,
                           uint32_t writeMask)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_WRITE_MASK, 2);
-
-   update_stencil_mask(&cmd->state.dynamic_stencil_wrmask, faceMask, writeMask);
 
-   tu_cs_emit_regs(&cs, A6XX_RB_STENCILWRMASK(.dword = cmd->state.dynamic_stencil_wrmask));
+   if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
+      cmd->state.dynamic.stencil_write_mask.front = writeMask;
+   if (faceMask & VK_STENCIL_FACE_BACK_BIT)
+      cmd->state.dynamic.stencil_write_mask.back = writeMask;
 
-   cmd->state.dirty |= TU_CMD_DIRTY_LRZ;
+   /* the front/back write masks must be updated together */
+   cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdSetStencilReference(VkCommandBuffer commandBuffer,
                           VkStencilFaceFlags faceMask,
                           uint32_t reference)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_REFERENCE, 2);
-
-   update_stencil_mask(&cmd->state.dynamic_stencil_ref, faceMask, reference);
-
-   tu_cs_emit_regs(&cs, A6XX_RB_STENCILREF(.dword = cmd->state.dynamic_stencil_ref));
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetSampleLocationsEXT(VkCommandBuffer commandBuffer,
-                            const VkSampleLocationsInfoEXT* pSampleLocationsInfo)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs cs = tu_cmd_dynamic_state(cmd, TU_DYNAMIC_STATE_SAMPLE_LOCATIONS, 9);
 
-   assert(pSampleLocationsInfo);
+   if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
+      cmd->state.dynamic.stencil_reference.front = reference;
+   if (faceMask & VK_STENCIL_FACE_BACK_BIT)
+      cmd->state.dynamic.stencil_reference.back = reference;
 
-   tu6_emit_sample_locations(&cs, pSampleLocationsInfo);
+   /* the front/back references must be updated together */
+   cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetCullModeEXT(VkCommandBuffer commandBuffer, VkCullModeFlags cullMode)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.gras_su_cntl &=
-      ~(A6XX_GRAS_SU_CNTL_CULL_FRONT | A6XX_GRAS_SU_CNTL_CULL_BACK);
-
-   if (cullMode & VK_CULL_MODE_FRONT_BIT)
-      cmd->state.gras_su_cntl |= A6XX_GRAS_SU_CNTL_CULL_FRONT;
-   if (cullMode & VK_CULL_MODE_BACK_BIT)
-      cmd->state.gras_su_cntl |= A6XX_GRAS_SU_CNTL_CULL_BACK;
-
-   cmd->state.dirty |= TU_CMD_DIRTY_GRAS_SU_CNTL;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetFrontFaceEXT(VkCommandBuffer commandBuffer, VkFrontFace frontFace)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.gras_su_cntl &= ~A6XX_GRAS_SU_CNTL_FRONT_CW;
-
-   if (frontFace == VK_FRONT_FACE_CLOCKWISE)
-      cmd->state.gras_su_cntl |= A6XX_GRAS_SU_CNTL_FRONT_CW;
-
-   cmd->state.dirty |= TU_CMD_DIRTY_GRAS_SU_CNTL;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetPrimitiveTopologyEXT(VkCommandBuffer commandBuffer,
-                              VkPrimitiveTopology primitiveTopology)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.primtype = tu6_primtype(primitiveTopology);
-   tu6_update_msaa(cmd, cmd->state.samples);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetViewportWithCountEXT(VkCommandBuffer commandBuffer,
-                              uint32_t viewportCount,
-                              const VkViewport* pViewports)
-{
-   tu_CmdSetViewport(commandBuffer, 0, viewportCount, pViewports);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetScissorWithCountEXT(VkCommandBuffer commandBuffer,
-                             uint32_t scissorCount,
-                             const VkRect2D* pScissors)
-{
-   tu_CmdSetScissor(commandBuffer, 0, scissorCount, pScissors);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetDepthTestEnableEXT(VkCommandBuffer commandBuffer,
-                            VkBool32 depthTestEnable)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.rb_depth_cntl &= ~A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE;
-
-   if (depthTestEnable)
-      cmd->state.rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE;
-
-   cmd->state.dirty |= TU_CMD_DIRTY_RB_DEPTH_CNTL;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetDepthWriteEnableEXT(VkCommandBuffer commandBuffer,
-                             VkBool32 depthWriteEnable)
+void
+tu_CmdExecuteCommands(VkCommandBuffer commandBuffer,
+                      uint32_t commandBufferCount,
+                      const VkCommandBuffer *pCmdBuffers)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.rb_depth_cntl &= ~A6XX_RB_DEPTH_CNTL_Z_WRITE_ENABLE;
-
-   if (depthWriteEnable)
-      cmd->state.rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_WRITE_ENABLE;
-
-   cmd->state.dirty |= TU_CMD_DIRTY_RB_DEPTH_CNTL;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetDepthCompareOpEXT(VkCommandBuffer commandBuffer,
-                           VkCompareOp depthCompareOp)
+VkResult
+tu_CreateCommandPool(VkDevice _device,
+                     const VkCommandPoolCreateInfo *pCreateInfo,
+                     const VkAllocationCallbacks *pAllocator,
+                     VkCommandPool *pCmdPool)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.rb_depth_cntl &= ~A6XX_RB_DEPTH_CNTL_ZFUNC__MASK;
+   TU_FROM_HANDLE(tu_device, device, _device);
+   struct tu_cmd_pool *pool;
 
-   cmd->state.rb_depth_cntl |=
-      A6XX_RB_DEPTH_CNTL_ZFUNC(tu6_compare_func(depthCompareOp));
+   pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
+                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (pool == NULL)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   cmd->state.dirty |= TU_CMD_DIRTY_RB_DEPTH_CNTL;
-}
+   if (pAllocator)
+      pool->alloc = *pAllocator;
+   else
+      pool->alloc = device->alloc;
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetDepthBoundsTestEnableEXT(VkCommandBuffer commandBuffer,
-                                  VkBool32 depthBoundsTestEnable)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   list_inithead(&pool->cmd_buffers);
+   list_inithead(&pool->free_cmd_buffers);
 
-   cmd->state.rb_depth_cntl &= ~A6XX_RB_DEPTH_CNTL_Z_BOUNDS_ENABLE;
+   pool->queue_family_index = pCreateInfo->queueFamilyIndex;
 
-   if (depthBoundsTestEnable)
-      cmd->state.rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_BOUNDS_ENABLE;
+   *pCmdPool = tu_cmd_pool_to_handle(pool);
 
-   cmd->state.dirty |= TU_CMD_DIRTY_RB_DEPTH_CNTL;
+   return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetStencilTestEnableEXT(VkCommandBuffer commandBuffer,
-                              VkBool32 stencilTestEnable)
+void
+tu_DestroyCommandPool(VkDevice _device,
+                      VkCommandPool commandPool,
+                      const VkAllocationCallbacks *pAllocator)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.rb_stencil_cntl &= ~(
-      A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE |
-      A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE_BF |
-      A6XX_RB_STENCIL_CONTROL_STENCIL_READ);
-
-   if (stencilTestEnable) {
-      cmd->state.rb_stencil_cntl |=
-         A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE |
-         A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE_BF |
-         A6XX_RB_STENCIL_CONTROL_STENCIL_READ;
-   }
+   TU_FROM_HANDLE(tu_device, device, _device);
+   TU_FROM_HANDLE(tu_cmd_pool, pool, commandPool);
 
-   cmd->state.dirty |= TU_CMD_DIRTY_RB_STENCIL_CNTL;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetStencilOpEXT(VkCommandBuffer commandBuffer,
-                      VkStencilFaceFlags faceMask,
-                      VkStencilOp failOp,
-                      VkStencilOp passOp,
-                      VkStencilOp depthFailOp,
-                      VkCompareOp compareOp)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   if (!pool)
+      return;
 
-   if (faceMask & VK_STENCIL_FACE_FRONT_BIT) {
-      cmd->state.rb_stencil_cntl &= ~(
-         A6XX_RB_STENCIL_CONTROL_FUNC__MASK |
-         A6XX_RB_STENCIL_CONTROL_FAIL__MASK |
-         A6XX_RB_STENCIL_CONTROL_ZPASS__MASK |
-         A6XX_RB_STENCIL_CONTROL_ZFAIL__MASK);
-
-      cmd->state.rb_stencil_cntl |=
-         A6XX_RB_STENCIL_CONTROL_FUNC(tu6_compare_func(compareOp)) |
-         A6XX_RB_STENCIL_CONTROL_FAIL(tu6_stencil_op(failOp)) |
-         A6XX_RB_STENCIL_CONTROL_ZPASS(tu6_stencil_op(passOp)) |
-         A6XX_RB_STENCIL_CONTROL_ZFAIL(tu6_stencil_op(depthFailOp));
+   list_for_each_entry_safe(struct tu_cmd_buffer, cmd_buffer,
+                            &pool->cmd_buffers, pool_link)
+   {
+      tu_cmd_buffer_destroy(cmd_buffer);
    }
 
-   if (faceMask & VK_STENCIL_FACE_BACK_BIT) {
-      cmd->state.rb_stencil_cntl &= ~(
-         A6XX_RB_STENCIL_CONTROL_FUNC_BF__MASK |
-         A6XX_RB_STENCIL_CONTROL_FAIL_BF__MASK |
-         A6XX_RB_STENCIL_CONTROL_ZPASS_BF__MASK |
-         A6XX_RB_STENCIL_CONTROL_ZFAIL_BF__MASK);
-
-      cmd->state.rb_stencil_cntl |=
-         A6XX_RB_STENCIL_CONTROL_FUNC_BF(tu6_compare_func(compareOp)) |
-         A6XX_RB_STENCIL_CONTROL_FAIL_BF(tu6_stencil_op(failOp)) |
-         A6XX_RB_STENCIL_CONTROL_ZPASS_BF(tu6_stencil_op(passOp)) |
-         A6XX_RB_STENCIL_CONTROL_ZFAIL_BF(tu6_stencil_op(depthFailOp));
+   list_for_each_entry_safe(struct tu_cmd_buffer, cmd_buffer,
+                            &pool->free_cmd_buffers, pool_link)
+   {
+      tu_cmd_buffer_destroy(cmd_buffer);
    }
 
-   cmd->state.dirty |= TU_CMD_DIRTY_RB_STENCIL_CNTL;
+   vk_free2(&device->alloc, pAllocator, pool);
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetDepthBiasEnableEXT(VkCommandBuffer commandBuffer,
-                            VkBool32 depthBiasEnable)
+VkResult
+tu_ResetCommandPool(VkDevice device,
+                    VkCommandPool commandPool,
+                    VkCommandPoolResetFlags flags)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   TU_FROM_HANDLE(tu_cmd_pool, pool, commandPool);
+   VkResult result;
 
-   cmd->state.gras_su_cntl &= ~A6XX_GRAS_SU_CNTL_POLY_OFFSET;
-   if (depthBiasEnable)
-      cmd->state.gras_su_cntl |= A6XX_GRAS_SU_CNTL_POLY_OFFSET;
+   list_for_each_entry(struct tu_cmd_buffer, cmd_buffer, &pool->cmd_buffers,
+                       pool_link)
+   {
+      result = tu_reset_cmd_buffer(cmd_buffer);
+      if (result != VK_SUCCESS)
+         return result;
+   }
 
-   cmd->state.dirty |= TU_CMD_DIRTY_GRAS_SU_CNTL;
+   return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetPrimitiveRestartEnableEXT(VkCommandBuffer commandBuffer,
-                                   VkBool32 primitiveRestartEnable)
+void
+tu_TrimCommandPool(VkDevice device,
+                   VkCommandPool commandPool,
+                   VkCommandPoolTrimFlags flags)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   TU_FROM_HANDLE(tu_cmd_pool, pool, commandPool);
 
-   cmd->state.primitive_restart_enable = primitiveRestartEnable;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetRasterizerDiscardEnableEXT(VkCommandBuffer commandBuffer,
-                                    VkBool32 rasterizerDiscardEnable)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   if (!pool)
+      return;
 
-   cmd->state.pc_raster_cntl &= ~A6XX_PC_RASTER_CNTL_DISCARD;
-   cmd->state.vpc_unknown_9107 &= ~A6XX_VPC_UNKNOWN_9107_RASTER_DISCARD;
-   if (rasterizerDiscardEnable) {
-      cmd->state.pc_raster_cntl |= A6XX_PC_RASTER_CNTL_DISCARD;
-      cmd->state.vpc_unknown_9107 |= A6XX_VPC_UNKNOWN_9107_RASTER_DISCARD;
+   list_for_each_entry_safe(struct tu_cmd_buffer, cmd_buffer,
+                            &pool->free_cmd_buffers, pool_link)
+   {
+      tu_cmd_buffer_destroy(cmd_buffer);
    }
-
-   cmd->state.dirty |= TU_CMD_DIRTY_RASTERIZER_DISCARD;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetLogicOpEXT(VkCommandBuffer commandBuffer,
-                    VkLogicOp logicOp)
+void
+tu_CmdBeginRenderPass(VkCommandBuffer commandBuffer,
+                      const VkRenderPassBeginInfo *pRenderPassBegin,
+                      VkSubpassContents contents)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.rb_mrt_control_rop =
-      tu6_rb_mrt_control_rop(logicOp, &cmd->state.rop_reads_dst);
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   TU_FROM_HANDLE(tu_render_pass, pass, pRenderPassBegin->renderPass);
+   TU_FROM_HANDLE(tu_framebuffer, framebuffer, pRenderPassBegin->framebuffer);
+   VkResult result;
 
-   cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-}
+   cmd_buffer->state.pass = pass;
+   cmd_buffer->state.subpass = pass->subpasses;
+   cmd_buffer->state.framebuffer = framebuffer;
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetPatchControlPointsEXT(VkCommandBuffer commandBuffer,
-                               uint32_t patchControlPoints)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   result = tu_cmd_state_setup_attachments(cmd_buffer, pRenderPassBegin);
+   if (result != VK_SUCCESS)
+      return;
 
-   cmd->state.patch_control_points = patchControlPoints;
+   tu_cmd_update_tiling_config(cmd_buffer, &pRenderPassBegin->renderArea);
+   tu_cmd_prepare_tile_load_ib(cmd_buffer);
+   tu_cmd_prepare_tile_store_ib(cmd_buffer);
 
-   cmd->state.dirty |= TU_CMD_DIRTY_PATCH_CONTROL_POINTS;
+   /* draw_cs should contain entries only for this render pass */
+   assert(!cmd_buffer->draw_cs.entry_count);
+   tu_cs_begin(&cmd_buffer->draw_cs);
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetLineStippleEXT(VkCommandBuffer commandBuffer,
-                        uint32_t lineStippleFactor,
-                        uint16_t lineStipplePattern)
+void
+tu_CmdBeginRenderPass2KHR(VkCommandBuffer commandBuffer,
+                          const VkRenderPassBeginInfo *pRenderPassBeginInfo,
+                          const VkSubpassBeginInfoKHR *pSubpassBeginInfo)
 {
-   tu_stub();
+   tu_CmdBeginRenderPass(commandBuffer, pRenderPassBeginInfo,
+                         pSubpassBeginInfo->contents);
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetColorWriteEnableEXT(VkCommandBuffer commandBuffer, uint32_t attachmentCount,
-                             const VkBool32 *pColorWriteEnables)
+void
+tu_CmdNextSubpass(VkCommandBuffer commandBuffer, VkSubpassContents contents)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   uint32_t color_write_enable = 0;
-
-   for (unsigned i = 0; i < attachmentCount; i++) {
-      if (pColorWriteEnables[i])
-         color_write_enable |= BIT(i);
-   }
-
-   cmd->state.color_write_enable = color_write_enable;
-   cmd->state.dirty |= TU_CMD_DIRTY_BLEND;
-}
-
-static void
-tu_flush_for_access(struct tu_cache_state *cache,
-                    enum tu_cmd_access_mask src_mask,
-                    enum tu_cmd_access_mask dst_mask)
-{
-   enum tu_cmd_flush_bits flush_bits = 0;
-
-   if (src_mask & TU_ACCESS_SYSMEM_WRITE) {
-      cache->pending_flush_bits |= TU_CMD_FLAG_ALL_INVALIDATE;
-   }
-
-   if (src_mask & TU_ACCESS_CP_WRITE) {
-      /* Flush the CP write queue.
-       */
-      cache->pending_flush_bits |=
-         TU_CMD_FLAG_WAIT_MEM_WRITES |
-         TU_CMD_FLAG_ALL_INVALIDATE;
-   }
-
-#define SRC_FLUSH(domain, flush, invalidate) \
-   if (src_mask & TU_ACCESS_##domain##_WRITE) {                      \
-      cache->pending_flush_bits |= TU_CMD_FLAG_##flush |             \
-         (TU_CMD_FLAG_ALL_INVALIDATE & ~TU_CMD_FLAG_##invalidate);   \
-   }
 
-   SRC_FLUSH(UCHE, CACHE_FLUSH, CACHE_INVALIDATE)
-   SRC_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
-   SRC_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+   tu_cmd_render_tiles(cmd);
 
-#undef SRC_FLUSH
+   cmd->state.subpass++;
 
-#define SRC_INCOHERENT_FLUSH(domain, flush, invalidate)              \
-   if (src_mask & TU_ACCESS_##domain##_INCOHERENT_WRITE) {           \
-      flush_bits |= TU_CMD_FLAG_##flush;                             \
-      cache->pending_flush_bits |=                                   \
-         (TU_CMD_FLAG_ALL_INVALIDATE & ~TU_CMD_FLAG_##invalidate);   \
-   }
-
-   SRC_INCOHERENT_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
-   SRC_INCOHERENT_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
-
-#undef SRC_INCOHERENT_FLUSH
-
-   /* Treat host & sysmem write accesses the same, since the kernel implicitly
-    * drains the queue before signalling completion to the host.
-    */
-   if (dst_mask & (TU_ACCESS_SYSMEM_READ | TU_ACCESS_SYSMEM_WRITE)) {
-      flush_bits |= cache->pending_flush_bits & TU_CMD_FLAG_ALL_FLUSH;
-   }
-
-#define DST_FLUSH(domain, flush, invalidate) \
-   if (dst_mask & (TU_ACCESS_##domain##_READ |                 \
-                   TU_ACCESS_##domain##_WRITE)) {              \
-      flush_bits |= cache->pending_flush_bits &                \
-         (TU_CMD_FLAG_##invalidate |                           \
-          (TU_CMD_FLAG_ALL_FLUSH & ~TU_CMD_FLAG_##flush));     \
-   }
-
-   DST_FLUSH(UCHE, CACHE_FLUSH, CACHE_INVALIDATE)
-   DST_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
-   DST_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
-
-#undef DST_FLUSH
-
-#define DST_INCOHERENT_FLUSH(domain, flush, invalidate) \
-   if (dst_mask & (TU_ACCESS_##domain##_INCOHERENT_READ |      \
-                   TU_ACCESS_##domain##_INCOHERENT_WRITE)) {   \
-      flush_bits |= TU_CMD_FLAG_##invalidate |                 \
-          (cache->pending_flush_bits &                         \
-           (TU_CMD_FLAG_ALL_FLUSH & ~TU_CMD_FLAG_##flush));    \
-   }
-
-   DST_INCOHERENT_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
-   DST_INCOHERENT_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
-
-#undef DST_INCOHERENT_FLUSH
-
-   cache->flush_bits |= flush_bits;
-   cache->pending_flush_bits &= ~flush_bits;
-}
-
-/* When translating Vulkan access flags to which cache is accessed
- * (CCU/UCHE/sysmem), we should take into account both the access flags and
- * the stage so that accesses with MEMORY_READ_BIT/MEMORY_WRITE_BIT + a
- * specific stage return something sensible. The specification for
- * VK_KHR_synchronization2 says that we should do this:
- *
- *    Additionally, scoping the pipeline stages into the barrier structs
- *    allows the use of the MEMORY_READ and MEMORY_WRITE flags without
- *    sacrificing precision. The per-stage access flags should be used to
- *    disambiguate specific accesses in a given stage or set of stages - for
- *    instance, between uniform reads and sampling operations.
- *
- * Note that while in all known cases the stage is actually enough, we should
- * still narrow things down based on the access flags to handle "old-style"
- * barriers that may specify a wider range of stages but more precise access
- * flags. These helpers allow us to do both.
- */
-
-static bool
-filter_read_access(VkAccessFlags2 flags, VkPipelineStageFlags2 stages,
-                   VkAccessFlags2 tu_flags, VkPipelineStageFlags2 tu_stages)
-{
-   return (flags & (tu_flags | VK_ACCESS_2_MEMORY_READ_BIT)) &&
-      (stages & (tu_stages | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT));
-}
-
-static bool
-filter_write_access(VkAccessFlags2 flags, VkPipelineStageFlags2 stages,
-                    VkAccessFlags2 tu_flags, VkPipelineStageFlags2 tu_stages)
-{
-   return (flags & (tu_flags | VK_ACCESS_2_MEMORY_WRITE_BIT)) &&
-      (stages & (tu_stages | VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT));
+   tu_cmd_update_tiling_config(cmd, NULL);
+   tu_cmd_prepare_tile_load_ib(cmd);
+   tu_cmd_prepare_tile_store_ib(cmd);
 }
 
-static bool
-gfx_read_access(VkAccessFlags2 flags, VkPipelineStageFlags2 stages,
-                VkAccessFlags2 tu_flags, VkPipelineStageFlags2 tu_stages)
+void
+tu_CmdNextSubpass2KHR(VkCommandBuffer commandBuffer,
+                      const VkSubpassBeginInfoKHR *pSubpassBeginInfo,
+                      const VkSubpassEndInfoKHR *pSubpassEndInfo)
 {
-   return filter_read_access(flags, stages, tu_flags,
-                             tu_stages | VK_PIPELINE_STAGE_2_ALL_GRAPHICS_BIT);
+   tu_CmdNextSubpass(commandBuffer, pSubpassBeginInfo->contents);
 }
 
-static bool
-gfx_write_access(VkAccessFlags2 flags, VkPipelineStageFlags2 stages,
-                 VkAccessFlags2 tu_flags, VkPipelineStageFlags2 tu_stages)
-{
-   return filter_write_access(flags, stages, tu_flags,
-                              tu_stages | VK_PIPELINE_STAGE_2_ALL_GRAPHICS_BIT);
-}
-static enum tu_cmd_access_mask
-vk2tu_access(VkAccessFlags2 flags, VkPipelineStageFlags2 stages, bool image_only, bool gmem)
+struct tu_draw_info
 {
-   enum tu_cmd_access_mask mask = 0;
-
-   if (gfx_read_access(flags, stages,
-                       VK_ACCESS_2_INDIRECT_COMMAND_READ_BIT |
-                       VK_ACCESS_2_CONDITIONAL_RENDERING_READ_BIT_EXT |
-                       VK_ACCESS_2_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT |
-                       VK_ACCESS_2_HOST_READ_BIT,
-                       VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT |
-                       VK_PIPELINE_STAGE_2_CONDITIONAL_RENDERING_BIT_EXT |
-                       VK_PIPELINE_STAGE_2_TRANSFORM_FEEDBACK_BIT_EXT |
-                       VK_PIPELINE_STAGE_2_HOST_BIT))
-      mask |= TU_ACCESS_SYSMEM_READ;
-
-   if (gfx_write_access(flags, stages,
-                        VK_ACCESS_2_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT,
-                        VK_PIPELINE_STAGE_2_TRANSFORM_FEEDBACK_BIT_EXT))
-      mask |= TU_ACCESS_CP_WRITE;
-
-   if (gfx_write_access(flags, stages,
-                        VK_ACCESS_2_HOST_WRITE_BIT,
-                        VK_PIPELINE_STAGE_2_HOST_BIT))
-      mask |= TU_ACCESS_SYSMEM_WRITE;
-
-#define SHADER_STAGES \
-   (VK_PIPELINE_STAGE_2_VERTEX_SHADER_BIT | \
-    VK_PIPELINE_STAGE_2_TESSELLATION_CONTROL_SHADER_BIT | \
-    VK_PIPELINE_STAGE_2_TESSELLATION_EVALUATION_SHADER_BIT | \
-    VK_PIPELINE_STAGE_2_GEOMETRY_SHADER_BIT | \
-    VK_PIPELINE_STAGE_2_PRE_RASTERIZATION_SHADERS_BIT | \
-    VK_PIPELINE_STAGE_2_FRAGMENT_SHADER_BIT | \
-    VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT)
-
-
-   if (gfx_read_access(flags, stages,
-                       VK_ACCESS_2_INDEX_READ_BIT |
-                       VK_ACCESS_2_VERTEX_ATTRIBUTE_READ_BIT |
-                       VK_ACCESS_2_UNIFORM_READ_BIT |
-                       VK_ACCESS_2_INPUT_ATTACHMENT_READ_BIT |
-                       VK_ACCESS_2_SHADER_READ_BIT,
-                       VK_PIPELINE_STAGE_2_INDEX_INPUT_BIT |
-                       VK_PIPELINE_STAGE_2_VERTEX_INPUT_BIT |
-                       VK_PIPELINE_STAGE_2_VERTEX_ATTRIBUTE_INPUT_BIT |
-                       SHADER_STAGES))
-       mask |= TU_ACCESS_UCHE_READ;
-
-   if (gfx_write_access(flags, stages,
-                        VK_ACCESS_2_SHADER_WRITE_BIT |
-                        VK_ACCESS_2_TRANSFORM_FEEDBACK_WRITE_BIT_EXT,
-                        VK_PIPELINE_STAGE_2_TRANSFORM_FEEDBACK_BIT_EXT |
-                        SHADER_STAGES))
-       mask |= TU_ACCESS_UCHE_WRITE;
-
-   /* When using GMEM, the CCU is always flushed automatically to GMEM, and
-    * then GMEM is flushed to sysmem. Furthermore, we already had to flush any
-    * previous writes in sysmem mode when transitioning to GMEM. Therefore we
-    * can ignore CCU and pretend that color attachments and transfers use
-    * sysmem directly.
+   /**
+    * Number of vertices.
     */
+   uint32_t count;
 
-   if (gfx_read_access(flags, stages,
-                       VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT |
-                       VK_ACCESS_2_COLOR_ATTACHMENT_READ_NONCOHERENT_BIT_EXT,
-                       VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT)) {
-      if (gmem)
-         mask |= TU_ACCESS_SYSMEM_READ;
-      else
-         mask |= TU_ACCESS_CCU_COLOR_INCOHERENT_READ;
-   }
-
-   if (gfx_read_access(flags, stages,
-                       VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_READ_BIT,
-                       VK_PIPELINE_STAGE_2_EARLY_FRAGMENT_TESTS_BIT |
-                       VK_PIPELINE_STAGE_2_LATE_FRAGMENT_TESTS_BIT)) {
-      if (gmem)
-         mask |= TU_ACCESS_SYSMEM_READ;
-      else
-         mask |= TU_ACCESS_CCU_DEPTH_INCOHERENT_READ;
-   }
-
-   if (gfx_write_access(flags, stages,
-                        VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT,
-                        VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT)) {
-      if (gmem) {
-         mask |= TU_ACCESS_SYSMEM_WRITE;
-      } else {
-         mask |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
-      }
-   }
-
-   if (gfx_write_access(flags, stages,
-                        VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,
-                        VK_PIPELINE_STAGE_2_EARLY_FRAGMENT_TESTS_BIT |
-                        VK_PIPELINE_STAGE_2_LATE_FRAGMENT_TESTS_BIT)) {
-      if (gmem) {
-         mask |= TU_ACCESS_SYSMEM_WRITE;
-      } else {
-         mask |= TU_ACCESS_CCU_DEPTH_INCOHERENT_WRITE;
-      }
-   }
-
-   if (filter_write_access(flags, stages,
-                           VK_ACCESS_2_TRANSFER_WRITE_BIT,
-                           VK_PIPELINE_STAGE_2_COPY_BIT |
-                           VK_PIPELINE_STAGE_2_BLIT_BIT |
-                           VK_PIPELINE_STAGE_2_CLEAR_BIT |
-                           VK_PIPELINE_STAGE_2_RESOLVE_BIT |
-                           VK_PIPELINE_STAGE_2_ALL_TRANSFER_BIT)) {
-      if (gmem) {
-         mask |= TU_ACCESS_SYSMEM_WRITE;
-      } else if (image_only) {
-         /* Because we always split up blits/copies of images involving
-          * multiple layers, we always access each layer in the same way, with
-          * the same base address, same format, etc. This means we can avoid
-          * flushing between multiple writes to the same image. This elides
-          * flushes between e.g. multiple blits to the same image.
-          */
-         mask |= TU_ACCESS_CCU_COLOR_WRITE;
-      } else {
-         mask |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
-      }
-   }
-
-   if (filter_read_access(flags, stages,
-                          VK_ACCESS_2_TRANSFER_READ_BIT,
-                          VK_PIPELINE_STAGE_2_COPY_BIT |
-                          VK_PIPELINE_STAGE_2_BLIT_BIT |
-                          VK_PIPELINE_STAGE_2_RESOLVE_BIT |
-                          VK_PIPELINE_STAGE_2_ALL_TRANSFER_BIT)) {
-      mask |= TU_ACCESS_UCHE_READ;
-   }
-
-   return mask;
-}
-
-/* These helpers deal with legacy BOTTOM_OF_PIPE/TOP_OF_PIPE stages.
- */
-
-static VkPipelineStageFlags2
-sanitize_src_stage(VkPipelineStageFlags2 stage_mask)
-{
-   /* From the Vulkan spec:
-    *
-    *    VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT is ...  equivalent to
-    *    VK_PIPELINE_STAGE_2_NONE in the first scope.
-    *
-    *    VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT is equivalent to
-    *    VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT with VkAccessFlags2 set to 0
-    *    when specified in the first synchronization scope, ...
+   /**
+    * Index of the first vertex.
     */
-   if (stage_mask & VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT)
-      return VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT;
+   int32_t vertex_offset;
 
-   return stage_mask & ~VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT;
-}
-
-static VkPipelineStageFlags2
-sanitize_dst_stage(VkPipelineStageFlags2 stage_mask)
-{
-   /* From the Vulkan spec:
-    *
-    *    VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT is equivalent to
-    *    VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT with VkAccessFlags2 set to 0
-    *    when specified in the second synchronization scope, ...
-    *
-    *    VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT is ... equivalent to
-    *    VK_PIPELINE_STAGE_2_NONE in the second scope.
-    *
+   /**
+    * First instance id.
     */
-   if (stage_mask & VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT)
-      return VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT;
-
-   return stage_mask & ~VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT;
-}
-
-static enum tu_stage
-vk2tu_single_stage(VkPipelineStageFlags2 vk_stage, bool dst)
-{
-   if (vk_stage == VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_CONDITIONAL_RENDERING_BIT_EXT)
-      return TU_STAGE_CP;
-
-   if (vk_stage == VK_PIPELINE_STAGE_2_VERTEX_INPUT_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_INDEX_INPUT_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_VERTEX_ATTRIBUTE_INPUT_BIT)
-      return TU_STAGE_FE;
-
-   if (vk_stage == VK_PIPELINE_STAGE_2_VERTEX_SHADER_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_TESSELLATION_CONTROL_SHADER_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_TESSELLATION_EVALUATION_SHADER_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_GEOMETRY_SHADER_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_PRE_RASTERIZATION_SHADERS_BIT)
-      return TU_STAGE_SP_VS;
-
-   if (vk_stage == VK_PIPELINE_STAGE_2_FRAGMENT_SHADER_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT)
-      return TU_STAGE_SP_PS;
-
-   if (vk_stage == VK_PIPELINE_STAGE_2_TRANSFORM_FEEDBACK_BIT_EXT || /* Yes, really */
-   /* See comment in TU_STAGE_GRAS about early fragment tests */
-       vk_stage == VK_PIPELINE_STAGE_2_EARLY_FRAGMENT_TESTS_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_LATE_FRAGMENT_TESTS_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT)
-
-      return TU_STAGE_PS;
-
-   if (vk_stage == VK_PIPELINE_STAGE_2_COPY_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_BLIT_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_RESOLVE_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_CLEAR_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_ALL_TRANSFER_BIT)
-      /* Blits read in SP_PS and write in PS, in both 2d and 3d cases */
-      return dst ? TU_STAGE_SP_PS : TU_STAGE_PS;
-
-   if (vk_stage == VK_PIPELINE_STAGE_2_ALL_GRAPHICS_BIT ||
-       vk_stage == VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT)
-      /* Be conservative */
-      return dst ? TU_STAGE_CP : TU_STAGE_PS;
-
-   if (vk_stage == VK_PIPELINE_STAGE_2_HOST_BIT)
-      return dst ? TU_STAGE_PS : TU_STAGE_CP;
-
-   unreachable("unknown pipeline stage");
-}
+   uint32_t first_instance;
 
-static enum tu_stage
-vk2tu_src_stage(VkPipelineStageFlags vk_stages)
-{
-   enum tu_stage stage = TU_STAGE_CP;
-   u_foreach_bit (bit, vk_stages) {
-      enum tu_stage new_stage = vk2tu_single_stage(1ull << bit, false);
-      stage = MAX2(stage, new_stage);
-   }
-
-   return stage;
-}
-
-static enum tu_stage
-vk2tu_dst_stage(VkPipelineStageFlags vk_stages)
-{
-   enum tu_stage stage = TU_STAGE_PS;
-   u_foreach_bit (bit, vk_stages) {
-      enum tu_stage new_stage = vk2tu_single_stage(1ull << bit, true);
-      stage = MIN2(stage, new_stage);
-   }
-
-   return stage;
-}
-
-static void
-tu_flush_for_stage(struct tu_cache_state *cache,
-                   enum tu_stage src_stage, enum tu_stage dst_stage)
-{
-   /* As far as we know, flushes take place in the last stage so if there are
-    * any pending flushes then we have to move down the source stage, because
-    * the data only becomes available when the flush finishes. In particular
-    * this can matter when the CP writes something and we need to invalidate
-    * UCHE to read it.
-    */
-   if (cache->flush_bits & (TU_CMD_FLAG_ALL_FLUSH | TU_CMD_FLAG_ALL_INVALIDATE))
-      src_stage = TU_STAGE_PS;
-
-   /* Note: if the destination stage is the CP, then the CP also has to wait
-    * for any WFI's to finish. This is already done for draw calls, including
-    * before indirect param reads, for the most part, so we just need to WFI.
-    *
-    * However, some indirect draw opcodes, depending on firmware, don't have
-    * implicit CP_WAIT_FOR_ME so we have to handle it manually.
-    *
-    * Transform feedback counters are read via CP_MEM_TO_REG, which implicitly
-    * does CP_WAIT_FOR_ME, but we still need a WFI if the GPU writes it.
-    *
-    * Currently we read the draw predicate using CP_MEM_TO_MEM, which
-    * also implicitly does CP_WAIT_FOR_ME. However CP_DRAW_PRED_SET does *not*
-    * implicitly do CP_WAIT_FOR_ME, it seems to only wait for counters to
-    * complete since it's written for DX11 where you can only predicate on the
-    * result of a query object. So if we implement 64-bit comparisons in the
-    * future, or if CP_DRAW_PRED_SET grows the capability to do 32-bit
-    * comparisons, then this will have to be dealt with.
+   /**
+    * Number of instances.
     */
-   if (src_stage > dst_stage) {
-      cache->flush_bits |= TU_CMD_FLAG_WAIT_FOR_IDLE;
-      if (dst_stage == TU_STAGE_CP)
-         cache->pending_flush_bits |= TU_CMD_FLAG_WAIT_FOR_ME;
-   }
-}
+   uint32_t instance_count;
 
-void
-tu_render_pass_state_merge(struct tu_render_pass_state *dst,
-                           const struct tu_render_pass_state *src)
-{
-   dst->xfb_used |= src->xfb_used;
-   dst->has_tess |= src->has_tess;
-   dst->has_prim_generated_query_in_rp |= src->has_prim_generated_query_in_rp;
-   dst->disable_gmem |= src->disable_gmem;
-   dst->sysmem_single_prim_mode |= src->sysmem_single_prim_mode;
-   dst->draw_cs_writes_to_cond_pred |= src->draw_cs_writes_to_cond_pred;
-
-   dst->drawcall_count += src->drawcall_count;
-   dst->drawcall_bandwidth_per_sample_sum +=
-      src->drawcall_bandwidth_per_sample_sum;
-}
-
-void
-tu_restore_suspended_pass(struct tu_cmd_buffer *cmd,
-                          struct tu_cmd_buffer *suspended)
-{
-   cmd->state.pass = suspended->state.suspended_pass.pass;
-   cmd->state.subpass = suspended->state.suspended_pass.subpass;
-   cmd->state.framebuffer = suspended->state.suspended_pass.framebuffer;
-   cmd->state.attachments = suspended->state.suspended_pass.attachments;
-   cmd->state.render_area = suspended->state.suspended_pass.render_area;
-   cmd->state.gmem_layout = suspended->state.suspended_pass.gmem_layout;
-   cmd->state.tiling = &cmd->state.framebuffer->tiling[cmd->state.gmem_layout];
-   cmd->state.lrz = suspended->state.suspended_pass.lrz;
-}
-
-/* Take the saved pre-chain in "secondary" and copy its commands to "cmd",
- * appending it after any saved-up commands in "cmd".
- */
-void
-tu_append_pre_chain(struct tu_cmd_buffer *cmd,
-                    struct tu_cmd_buffer *secondary)
-{
-   tu_cs_add_entries(&cmd->draw_cs, &secondary->pre_chain.draw_cs);
-   tu_cs_add_entries(&cmd->draw_epilogue_cs,
-                     &secondary->pre_chain.draw_epilogue_cs);
-
-   tu_render_pass_state_merge(&cmd->state.rp,
-                              &secondary->pre_chain.state);
-   tu_clone_trace_range(cmd, &cmd->draw_cs, secondary->pre_chain.trace_renderpass_start,
-         secondary->pre_chain.trace_renderpass_end);
-}
-
-/* Take the saved post-chain in "secondary" and copy it to "cmd".
- */
-void
-tu_append_post_chain(struct tu_cmd_buffer *cmd,
-                     struct tu_cmd_buffer *secondary)
-{
-   tu_cs_add_entries(&cmd->draw_cs, &secondary->draw_cs);
-   tu_cs_add_entries(&cmd->draw_epilogue_cs, &secondary->draw_epilogue_cs);
-
-   tu_clone_trace_range(cmd, &cmd->draw_cs, secondary->trace_renderpass_start,
-         secondary->trace_renderpass_end);
-   cmd->state.rp = secondary->state.rp;
-}
-
-/* Assuming "secondary" is just a sequence of suspended and resuming passes,
- * copy its state to "cmd". This also works instead of tu_append_post_chain(),
- * but it's a bit slower because we don't assume that the chain begins in
- * "secondary" and therefore have to care about the command buffer's
- * renderpass state.
- */
-void
-tu_append_pre_post_chain(struct tu_cmd_buffer *cmd,
-                         struct tu_cmd_buffer *secondary)
-{
-   tu_cs_add_entries(&cmd->draw_cs, &secondary->draw_cs);
-   tu_cs_add_entries(&cmd->draw_epilogue_cs, &secondary->draw_epilogue_cs);
-
-   tu_clone_trace_range(cmd, &cmd->draw_cs, secondary->trace_renderpass_start,
-         secondary->trace_renderpass_end);
-   tu_render_pass_state_merge(&cmd->state.rp,
-                              &secondary->state.rp);
-}
-
-/* Take the current render pass state and save it to "pre_chain" to be
- * combined later.
- */
-static void
-tu_save_pre_chain(struct tu_cmd_buffer *cmd)
-{
-   tu_cs_add_entries(&cmd->pre_chain.draw_cs,
-                     &cmd->draw_cs);
-   tu_cs_add_entries(&cmd->pre_chain.draw_epilogue_cs,
-                     &cmd->draw_epilogue_cs);
-   cmd->pre_chain.trace_renderpass_start =
-      cmd->trace_renderpass_start;
-   cmd->pre_chain.trace_renderpass_end =
-      cmd->trace_renderpass_end;
-   cmd->pre_chain.state = cmd->state.rp;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdExecuteCommands(VkCommandBuffer commandBuffer,
-                      uint32_t commandBufferCount,
-                      const VkCommandBuffer *pCmdBuffers)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   VkResult result;
-
-   assert(commandBufferCount > 0);
-
-   /* Emit any pending flushes. */
-   if (cmd->state.pass) {
-      tu_flush_all_pending(&cmd->state.renderpass_cache);
-      tu_emit_cache_flush_renderpass(cmd, &cmd->draw_cs);
-   } else {
-      tu_flush_all_pending(&cmd->state.cache);
-      tu_emit_cache_flush(cmd, &cmd->cs);
-   }
-
-   for (uint32_t i = 0; i < commandBufferCount; i++) {
-      TU_FROM_HANDLE(tu_cmd_buffer, secondary, pCmdBuffers[i]);
-
-      if (secondary->usage_flags &
-          VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
-         assert(tu_cs_is_empty(&secondary->cs));
-
-         result = tu_cs_add_entries(&cmd->draw_cs, &secondary->draw_cs);
-         if (result != VK_SUCCESS) {
-            vk_command_buffer_set_error(&cmd->vk, result);
-            break;
-         }
-
-         result = tu_cs_add_entries(&cmd->draw_epilogue_cs,
-               &secondary->draw_epilogue_cs);
-         if (result != VK_SUCCESS) {
-            vk_command_buffer_set_error(&cmd->vk, result);
-            break;
-         }
-
-         /* If LRZ was made invalid in secondary - we should disable
-          * LRZ retroactively for the whole renderpass.
-          */
-         if (!secondary->state.lrz.valid)
-            cmd->state.lrz.valid = false;
-
-         tu_clone_trace(cmd, &cmd->draw_cs, &secondary->trace);
-         tu_render_pass_state_merge(&cmd->state.rp, &secondary->state.rp);
-      } else {
-         switch (secondary->state.suspend_resume) {
-         case SR_NONE:
-            assert(tu_cs_is_empty(&secondary->draw_cs));
-            assert(tu_cs_is_empty(&secondary->draw_epilogue_cs));
-            tu_cs_add_entries(&cmd->cs, &secondary->cs);
-            tu_clone_trace(cmd, &cmd->cs, &secondary->trace);
-            break;
-
-         case SR_IN_PRE_CHAIN:
-            /* cmd may be empty, which means that the chain begins before cmd
-             * in which case we have to update its state.
-             */
-            if (cmd->state.suspend_resume == SR_NONE) {
-               cmd->state.suspend_resume = SR_IN_PRE_CHAIN;
-               cmd->trace_renderpass_start = u_trace_end_iterator(&cmd->trace);
-            }
-
-            /* The secondary is just a continuous suspend/resume chain so we
-             * just have to append it to the the command buffer.
-             */
-            assert(tu_cs_is_empty(&secondary->cs));
-            tu_append_pre_post_chain(cmd, secondary);
-            break;
-
-         case SR_AFTER_PRE_CHAIN:
-         case SR_IN_CHAIN:
-         case SR_IN_CHAIN_AFTER_PRE_CHAIN:
-            if (secondary->state.suspend_resume == SR_AFTER_PRE_CHAIN ||
-                secondary->state.suspend_resume == SR_IN_CHAIN_AFTER_PRE_CHAIN) {
-               /* In thse cases there is a `pre_chain` in the secondary which
-                * ends that we need to append to the primary.
-                */
-
-               if (cmd->state.suspend_resume == SR_NONE)
-                  cmd->trace_renderpass_start = u_trace_end_iterator(&cmd->trace);
-
-               tu_append_pre_chain(cmd, secondary);
-
-               /* We're about to render, so we need to end the command stream
-                * in case there were any extra commands generated by copying
-                * the trace.
-                */
-               tu_cs_end(&cmd->draw_cs);
-               tu_cs_end(&cmd->draw_epilogue_cs);
-
-               switch (cmd->state.suspend_resume) {
-               case SR_NONE:
-               case SR_IN_PRE_CHAIN:
-                  /* The renderpass chain ends in the secondary but isn't
-                   * started in the primary, so we have to move the state to
-                   * `pre_chain`.
-                   */
-                  cmd->trace_renderpass_end = u_trace_end_iterator(&cmd->trace);
-                  tu_save_pre_chain(cmd);
-                  cmd->state.suspend_resume = SR_AFTER_PRE_CHAIN;
-                  break;
-               case SR_IN_CHAIN:
-               case SR_IN_CHAIN_AFTER_PRE_CHAIN:
-                  /* The renderpass ends in the secondary and starts somewhere
-                   * earlier in this primary. Since the last render pass in
-                   * the chain is in the secondary, we are technically outside
-                   * of a render pass.  Fix that here by reusing the dynamic
-                   * render pass that was setup for the last suspended render
-                   * pass before the secondary.
-                   */
-                  tu_restore_suspended_pass(cmd, cmd);
-
-                  tu_cmd_render(cmd);
-                  if (cmd->state.suspend_resume == SR_IN_CHAIN)
-                     cmd->state.suspend_resume = SR_NONE;
-                  else
-                     cmd->state.suspend_resume = SR_AFTER_PRE_CHAIN;
-                  break;
-               case SR_AFTER_PRE_CHAIN:
-                  unreachable("resuming render pass is not preceded by suspending one");
-               }
-
-               tu_reset_render_pass(cmd);
-            }
-
-            tu_cs_add_entries(&cmd->cs, &secondary->cs);
-
-            if (secondary->state.suspend_resume == SR_IN_CHAIN_AFTER_PRE_CHAIN ||
-                secondary->state.suspend_resume == SR_IN_CHAIN) {
-               /* The secondary ends in a "post-chain" (the opposite of a
-                * pre-chain) that we need to copy into the current command
-                * buffer.
-                */
-               cmd->trace_renderpass_start = u_trace_end_iterator(&cmd->trace);
-               tu_append_post_chain(cmd, secondary);
-               cmd->trace_renderpass_end = u_trace_end_iterator(&cmd->trace);
-               cmd->state.suspended_pass = secondary->state.suspended_pass;
-
-               switch (cmd->state.suspend_resume) {
-               case SR_NONE:
-                  cmd->state.suspend_resume = SR_IN_CHAIN;
-                  break;
-               case SR_AFTER_PRE_CHAIN:
-                  cmd->state.suspend_resume = SR_IN_CHAIN_AFTER_PRE_CHAIN;
-                  break;
-               default:
-                  unreachable("suspending render pass is followed by a not resuming one");
-               }
-            }
-         }
-      }
-
-      cmd->state.index_size = secondary->state.index_size; /* for restart index update */
-   }
-   cmd->state.dirty = ~0u; /* TODO: set dirty only what needs to be */
-
-   if (!cmd->state.lrz.gpu_dir_tracking && cmd->state.pass) {
-      /* After a secondary command buffer is executed, LRZ is not valid
-       * until it is cleared again.
-       */
-      cmd->state.lrz.valid = false;
-   }
-
-   /* After executing secondary command buffers, there may have been arbitrary
-    * flushes executed, so when we encounter a pipeline barrier with a
-    * srcMask, we have to assume that we need to invalidate. Therefore we need
-    * to re-initialize the cache with all pending invalidate bits set.
+   /**
+    * First index (indexed draws only).
     */
-   if (cmd->state.pass) {
-      tu_cache_init(&cmd->state.renderpass_cache);
-   } else {
-      tu_cache_init(&cmd->state.cache);
-   }
-}
+   uint32_t first_index;
 
-static void
-tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
-                   const struct tu_subpass_barrier *barrier,
-                   bool external)
-{
-   /* Note: we don't know until the end of the subpass whether we'll use
-    * sysmem, so assume sysmem here to be safe.
+   /**
+    * Whether it's an indexed draw.
     */
-   struct tu_cache_state *cache =
-      external ? &cmd_buffer->state.cache : &cmd_buffer->state.renderpass_cache;
-   VkPipelineStageFlags2 src_stage_vk =
-      sanitize_src_stage(barrier->src_stage_mask);
-   VkPipelineStageFlags2 dst_stage_vk =
-      sanitize_dst_stage(barrier->dst_stage_mask);
-   enum tu_cmd_access_mask src_flags =
-      vk2tu_access(barrier->src_access_mask, src_stage_vk, false, false);
-   enum tu_cmd_access_mask dst_flags =
-      vk2tu_access(barrier->dst_access_mask, dst_stage_vk, false, false);
-
-   if (barrier->incoherent_ccu_color)
-      src_flags |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
-   if (barrier->incoherent_ccu_depth)
-      src_flags |= TU_ACCESS_CCU_DEPTH_INCOHERENT_WRITE;
-
-   tu_flush_for_access(cache, src_flags, dst_flags);
-
-   enum tu_stage src_stage = vk2tu_src_stage(src_stage_vk);
-   enum tu_stage dst_stage = vk2tu_dst_stage(dst_stage_vk);
-   tu_flush_for_stage(cache, src_stage, dst_stage);
-}
-
-/* emit mrt/zs/msaa/ubwc state for the subpass that is starting (either at
- * vkCmdBeginRenderPass2() or vkCmdNextSubpass2())
- */
-static void
-tu_emit_subpass_begin(struct tu_cmd_buffer *cmd)
-{
-   tu6_emit_zs(cmd, cmd->state.subpass, &cmd->draw_cs);
-   tu6_emit_mrt(cmd, cmd->state.subpass, &cmd->draw_cs);
-   if (cmd->state.subpass->samples != 0)
-      tu6_update_msaa(cmd, cmd->state.subpass->samples);
-   tu6_emit_render_cntl(cmd, cmd->state.subpass, &cmd->draw_cs, false);
+   bool indexed;
 
-   tu_set_input_attachments(cmd, cmd->state.subpass);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdBeginRenderPass2(VkCommandBuffer commandBuffer,
-                       const VkRenderPassBeginInfo *pRenderPassBegin,
-                       const VkSubpassBeginInfo *pSubpassBeginInfo)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   if (unlikely(cmd->device->instance->debug_flags & TU_DEBUG_DYNAMIC)) {
-      vk_common_CmdBeginRenderPass2(commandBuffer, pRenderPassBegin,
-                                    pSubpassBeginInfo);
-      return;
-   }
-
-   TU_FROM_HANDLE(tu_render_pass, pass, pRenderPassBegin->renderPass);
-   TU_FROM_HANDLE(tu_framebuffer, fb, pRenderPassBegin->framebuffer);
-
-   const struct VkRenderPassAttachmentBeginInfo *pAttachmentInfo =
-      vk_find_struct_const(pRenderPassBegin->pNext,
-                           RENDER_PASS_ATTACHMENT_BEGIN_INFO);
-
-   cmd->state.pass = pass;
-   cmd->state.subpass = pass->subpasses;
-   cmd->state.framebuffer = fb;
-   cmd->state.render_area = pRenderPassBegin->renderArea;
-
-   cmd->state.attachments =
-      vk_alloc(&cmd->vk.pool->alloc, pass->attachment_count *
-               sizeof(cmd->state.attachments[0]), 8,
-               VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
-
-   if (!cmd->state.attachments) {
-      vk_command_buffer_set_error(&cmd->vk, VK_ERROR_OUT_OF_HOST_MEMORY);
-      return;
-   }
-
-   for (unsigned i = 0; i < pass->attachment_count; i++) {
-      cmd->state.attachments[i] = pAttachmentInfo ?
-         tu_image_view_from_handle(pAttachmentInfo->pAttachments[i]) :
-         cmd->state.framebuffer->attachments[i].attachment;
-   }
-   tu_choose_gmem_layout(cmd);
-
-   trace_start_render_pass(&cmd->trace, &cmd->cs);
-
-   /* Note: because this is external, any flushes will happen before draw_cs
-    * gets called. However deferred flushes could have to happen later as part
-    * of the subpass.
+   /**
+    * Indirect draw parameters resource.
     */
-   tu_subpass_barrier(cmd, &pass->subpasses[0].start_barrier, true);
-   cmd->state.renderpass_cache.pending_flush_bits =
-      cmd->state.cache.pending_flush_bits;
-   cmd->state.renderpass_cache.flush_bits = 0;
-
-   if (pass->subpasses[0].feedback_invalidate)
-      cmd->state.renderpass_cache.flush_bits |= TU_CMD_FLAG_CACHE_INVALIDATE;
-
-   tu_lrz_begin_renderpass(cmd, pRenderPassBegin->pClearValues);
-
-   cmd->trace_renderpass_start = u_trace_end_iterator(&cmd->trace);
-
-   tu_emit_renderpass_begin(cmd, pRenderPassBegin->pClearValues);
-   tu_emit_subpass_begin(cmd);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdBeginRendering(VkCommandBuffer commandBuffer,
-                     const VkRenderingInfo *pRenderingInfo)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   VkClearValue clear_values[2 * (MAX_RTS + 1)];
-
-   tu_setup_dynamic_render_pass(cmd, pRenderingInfo);
-   tu_setup_dynamic_framebuffer(cmd, pRenderingInfo);
-
-   cmd->state.pass = &cmd->dynamic_pass;
-   cmd->state.subpass = &cmd->dynamic_subpass;
-   cmd->state.framebuffer = &cmd->dynamic_framebuffer;
-   cmd->state.render_area = pRenderingInfo->renderArea;
-
-   cmd->state.attachments = cmd->dynamic_attachments;
-
-   for (unsigned i = 0; i < pRenderingInfo->colorAttachmentCount; i++) {
-      uint32_t a = cmd->dynamic_subpass.color_attachments[i].attachment;
-      if (!pRenderingInfo->pColorAttachments[i].imageView)
-         continue;
-
-      TU_FROM_HANDLE(tu_image_view, view,
-                     pRenderingInfo->pColorAttachments[i].imageView);
-      cmd->state.attachments[a] = view;
-      clear_values[a] = pRenderingInfo->pColorAttachments[i].clearValue;
-
-      a = cmd->dynamic_subpass.resolve_attachments[i].attachment;
-      if (a != VK_ATTACHMENT_UNUSED) {
-         TU_FROM_HANDLE(tu_image_view, resolve_view,
-                        pRenderingInfo->pColorAttachments[i].resolveImageView);
-         cmd->state.attachments[a] = resolve_view;
-      }
-   }
-
-   uint32_t a = cmd->dynamic_subpass.depth_stencil_attachment.attachment;
-   if (pRenderingInfo->pDepthAttachment || pRenderingInfo->pStencilAttachment) {
-      const struct VkRenderingAttachmentInfo *common_info =
-         (pRenderingInfo->pDepthAttachment &&
-          pRenderingInfo->pDepthAttachment->imageView != VK_NULL_HANDLE) ?
-         pRenderingInfo->pDepthAttachment :
-         pRenderingInfo->pStencilAttachment;
-      if (common_info && common_info->imageView != VK_NULL_HANDLE) {
-         TU_FROM_HANDLE(tu_image_view, view, common_info->imageView);
-         cmd->state.attachments[a] = view;
-         if (pRenderingInfo->pDepthAttachment) {
-            clear_values[a].depthStencil.depth =
-               pRenderingInfo->pDepthAttachment->clearValue.depthStencil.depth;
-         }
-
-         if (pRenderingInfo->pStencilAttachment) {
-            clear_values[a].depthStencil.stencil =
-               pRenderingInfo->pStencilAttachment->clearValue.depthStencil.stencil;
-         }
-
-         if (cmd->dynamic_subpass.resolve_count >
-             cmd->dynamic_subpass.color_count) {
-            TU_FROM_HANDLE(tu_image_view, resolve_view,
-                           common_info->resolveImageView);
-            a = cmd->dynamic_subpass.resolve_attachments[cmd->dynamic_subpass.color_count].attachment;
-            cmd->state.attachments[a] = resolve_view;
-         }
-      }
-   }
-
-   if (unlikely(cmd->device->instance->debug_flags & TU_DEBUG_DYNAMIC)) {
-      const VkRenderingSelfDependencyInfoMESA *self_dependency =
-         vk_find_struct_const(pRenderingInfo->pNext, RENDERING_SELF_DEPENDENCY_INFO_MESA);
-      if (self_dependency &&
-          (self_dependency->colorSelfDependencies ||
-           self_dependency->depthSelfDependency ||
-           self_dependency->stencilSelfDependency)) {
-         /* Mesa's renderpass emulation requires us to use normal attachments
-          * for input attachments, and currently doesn't try to keep track of
-          * which color/depth attachment an input attachment corresponds to.
-          * So when there's a self-dependency, we have to use sysmem.
-          */
-         cmd->state.rp.disable_gmem = true;
-      }
-   }
-
-   tu_choose_gmem_layout(cmd);
-
-   cmd->state.renderpass_cache.pending_flush_bits =
-      cmd->state.cache.pending_flush_bits;
-   cmd->state.renderpass_cache.flush_bits = 0;
-
-   bool resuming = pRenderingInfo->flags & VK_RENDERING_RESUMING_BIT;
-   bool suspending = pRenderingInfo->flags & VK_RENDERING_SUSPENDING_BIT;
-   cmd->state.suspending = suspending;
-   cmd->state.resuming = resuming;
+   struct tu_buffer *indirect;
+   uint64_t indirect_offset;
+   uint32_t stride;
 
-   /* We can't track LRZ across command buffer boundaries, so we have to
-    * disable LRZ when resuming/suspending unless we can track on the GPU.
+   /**
+    * Draw count parameters resource.
     */
-   if ((resuming || suspending) &&
-       !cmd->device->physical_device->info->a6xx.has_lrz_dir_tracking) {
-      cmd->state.lrz.valid = false;
-   } else {
-      if (resuming)
-         tu_lrz_begin_resumed_renderpass(cmd, clear_values);
-      else
-         tu_lrz_begin_renderpass(cmd, clear_values);
-   }
-
-
-   if (suspending) {
-      cmd->state.suspended_pass.pass = cmd->state.pass;
-      cmd->state.suspended_pass.subpass = cmd->state.subpass;
-      cmd->state.suspended_pass.framebuffer = cmd->state.framebuffer;
-      cmd->state.suspended_pass.render_area = cmd->state.render_area;
-      cmd->state.suspended_pass.attachments = cmd->state.attachments;
-      cmd->state.suspended_pass.gmem_layout = cmd->state.gmem_layout;
-   }
-
-   if (!resuming) {
-      trace_start_render_pass(&cmd->trace, &cmd->cs);
-   }
-
-   if (!resuming || cmd->state.suspend_resume == SR_NONE) {
-      cmd->trace_renderpass_start = u_trace_end_iterator(&cmd->trace);
-   }
+   struct tu_buffer *count_buffer;
+   uint64_t count_buffer_offset;
+};
 
-   if (!resuming) {
-      tu_emit_renderpass_begin(cmd, clear_values);
-      tu_emit_subpass_begin(cmd);
-   }
+enum tu_draw_state_group_id
+{
+   TU_DRAW_STATE_PROGRAM,
+   TU_DRAW_STATE_PROGRAM_BINNING,
+   TU_DRAW_STATE_VI,
+   TU_DRAW_STATE_VI_BINNING,
+   TU_DRAW_STATE_VP,
+   TU_DRAW_STATE_RAST,
+   TU_DRAW_STATE_DS,
+   TU_DRAW_STATE_BLEND,
 
-   if (suspending && !resuming) {
-      /* entering a chain */
-      switch (cmd->state.suspend_resume) {
-      case SR_NONE:
-         cmd->state.suspend_resume = SR_IN_CHAIN;
-         break;
-      case SR_AFTER_PRE_CHAIN:
-         cmd->state.suspend_resume = SR_IN_CHAIN_AFTER_PRE_CHAIN;
-         break;
-      case SR_IN_PRE_CHAIN:
-      case SR_IN_CHAIN:
-      case SR_IN_CHAIN_AFTER_PRE_CHAIN:
-         unreachable("suspending render pass not followed by resuming pass");
-         break;
-      }
-   }
+   TU_DRAW_STATE_COUNT,
+};
 
-   if (resuming && cmd->state.suspend_resume == SR_NONE)
-      cmd->state.suspend_resume = SR_IN_PRE_CHAIN;
-}
+struct tu_draw_state_group
+{
+   enum tu_draw_state_group_id id;
+   uint32_t enable_mask;
+   const struct tu_cs_entry *ib;
+};
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdNextSubpass2(VkCommandBuffer commandBuffer,
-                   const VkSubpassBeginInfo *pSubpassBeginInfo,
-                   const VkSubpassEndInfo *pSubpassEndInfo)
+static void
+tu6_bind_draw_states(struct tu_cmd_buffer *cmd,
+                     struct tu_cs *cs,
+                     const struct tu_draw_info *draw)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+   const struct tu_pipeline *pipeline = cmd->state.pipeline;
+   const struct tu_dynamic_state *dynamic = &cmd->state.dynamic;
+   struct tu_draw_state_group draw_state_groups[TU_DRAW_STATE_COUNT];
+   uint32_t draw_state_group_count = 0;
 
-   if (unlikely(cmd->device->instance->debug_flags & TU_DEBUG_DYNAMIC)) {
-      vk_common_CmdNextSubpass2(commandBuffer, pSubpassBeginInfo,
-                                pSubpassEndInfo);
+   VkResult result = tu_cs_reserve_space(cmd->device, cs, 256);
+   if (result != VK_SUCCESS) {
+      cmd->record_result = result;
       return;
    }
 
-   const struct tu_render_pass *pass = cmd->state.pass;
-   struct tu_cs *cs = &cmd->draw_cs;
-   const struct tu_subpass *last_subpass = cmd->state.subpass;
-
-   const struct tu_subpass *subpass = cmd->state.subpass++;
-
-   /* Track LRZ valid state
-    *
-    * TODO: Improve this tracking for keeping the state of the past depth/stencil images,
-    * so if they become active again, we reuse its old state.
-    */
-   if (last_subpass->depth_stencil_attachment.attachment != subpass->depth_stencil_attachment.attachment) {
-      cmd->state.lrz.valid = false;
-      cmd->state.dirty |= TU_CMD_DIRTY_LRZ;
-   }
-
-   tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_GMEM);
+   /* TODO lrz */
 
-   if (subpass->resolve_attachments) {
-      tu6_emit_blit_scissor(cmd, cs, true);
+   uint32_t pc_primitive_cntl = 0;
+   if (pipeline->ia.primitive_restart && draw->indexed)
+      pc_primitive_cntl |= A6XX_PC_PRIMITIVE_CNTL_0_PRIMITIVE_RESTART;
 
-      for (unsigned i = 0; i < subpass->resolve_count; i++) {
-         uint32_t a = subpass->resolve_attachments[i].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9806, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9990, 0);
+   tu_cs_emit_write_reg(cs, REG_A6XX_VFD_UNKNOWN_A008, 0);
 
-         uint32_t gmem_a = tu_subpass_get_attachment_to_resolve(subpass, i);
+   tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_0, 1);
+   tu_cs_emit(cs, pc_primitive_cntl);
 
-         tu_store_gmem_attachment(cmd, cs, a, gmem_a, false);
-
-         if (!pass->attachments[a].gmem)
-            continue;
-
-         /* check if the resolved attachment is needed by later subpasses,
-          * if it is, should be doing a GMEM->GMEM resolve instead of GMEM->MEM->GMEM..
-          */
-         perf_debug(cmd->device, "TODO: missing GMEM->GMEM resolve path\n");
-         tu_load_gmem_attachment(cmd, cs, a, false, true);
-      }
+   if (cmd->state.dirty &
+          (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH) &&
+       (pipeline->dynamic_state.mask & TU_DYNAMIC_LINE_WIDTH)) {
+      tu6_emit_gras_su_cntl(cs, pipeline->rast.gras_su_cntl,
+                            dynamic->line_width);
    }
 
-   tu_cond_exec_end(cs);
-
-   tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
-
-   tu6_emit_sysmem_resolves(cmd, cs, subpass);
-
-   tu_cond_exec_end(cs);
-
-   /* Handle dependencies for the next subpass */
-   tu_subpass_barrier(cmd, &cmd->state.subpass->start_barrier, false);
-
-   if (cmd->state.subpass->feedback_invalidate)
-      cmd->state.renderpass_cache.flush_bits |= TU_CMD_FLAG_CACHE_INVALIDATE;
-
-   tu_emit_subpass_begin(cmd);
-}
-
-static uint32_t
-tu6_user_consts_size(const struct tu_pipeline *pipeline,
-                     gl_shader_stage type)
-{
-   const struct tu_program_descriptor_linkage *link =
-      &pipeline->program.link[type];
-   uint32_t dwords = 0;
-
-   if (link->tu_const_state.push_consts.dwords > 0) {
-      unsigned num_units = link->tu_const_state.push_consts.dwords;
-      dwords += 4 + num_units;
+   if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK) &&
+       (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_COMPARE_MASK)) {
+      tu6_emit_stencil_compare_mask(cs, dynamic->stencil_compare_mask.front,
+                                    dynamic->stencil_compare_mask.back);
    }
 
-   return dwords;
-}
-
-static void
-tu6_emit_user_consts(struct tu_cs *cs,
-                     const struct tu_pipeline *pipeline,
-                     gl_shader_stage type,
-                     uint32_t *push_constants)
-{
-   const struct tu_program_descriptor_linkage *link =
-      &pipeline->program.link[type];
-
-   if (link->tu_const_state.push_consts.dwords > 0) {
-      unsigned num_units = link->tu_const_state.push_consts.dwords;
-      unsigned offset = link->tu_const_state.push_consts.lo;
-
-      /* DST_OFF and NUM_UNIT requires vec4 units */
-      tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + num_units);
-      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(offset / 4) |
-            CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-            CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-            CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
-            CP_LOAD_STATE6_0_NUM_UNIT(num_units / 4));
-      tu_cs_emit(cs, 0);
-      tu_cs_emit(cs, 0);
-      for (unsigned i = 0; i < num_units; i++)
-         tu_cs_emit(cs, push_constants[i + offset]);
+   if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK) &&
+       (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_WRITE_MASK)) {
+      tu6_emit_stencil_write_mask(cs, dynamic->stencil_write_mask.front,
+                                  dynamic->stencil_write_mask.back);
    }
-}
 
-static void
-tu6_emit_shared_consts(struct tu_cs *cs,
-                       const struct tu_pipeline *pipeline,
-                       uint32_t *push_constants,
-                       bool compute)
-{
-   if (pipeline->shared_consts.dwords > 0) {
-      /* Offset and num_units for shared consts are in units of dwords. */
-      unsigned num_units = pipeline->shared_consts.dwords;
-      unsigned offset = pipeline->shared_consts.lo;
-
-      enum a6xx_state_type st = compute ? ST6_UBO : ST6_CONSTANTS;
-      uint32_t cp_load_state = compute ? CP_LOAD_STATE6_FRAG : CP_LOAD_STATE6;
-
-      tu_cs_emit_pkt7(cs, cp_load_state, 3 + num_units);
-      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(offset) |
-            CP_LOAD_STATE6_0_STATE_TYPE(st) |
-            CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-            CP_LOAD_STATE6_0_STATE_BLOCK(SB6_IBO) |
-            CP_LOAD_STATE6_0_NUM_UNIT(num_units));
-      tu_cs_emit(cs, 0);
-      tu_cs_emit(cs, 0);
-
-      for (unsigned i = 0; i < num_units; i++)
-         tu_cs_emit(cs, push_constants[i + offset]);
+   if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE) &&
+       (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_REFERENCE)) {
+      tu6_emit_stencil_reference(cs, dynamic->stencil_reference.front,
+                                 dynamic->stencil_reference.back);
    }
-}
 
-static uint32_t
-tu6_const_size(struct tu_cmd_buffer *cmd,
-               const struct tu_pipeline *pipeline,
-               bool compute)
-{
-   uint32_t dwords = 0;
+   if (cmd->state.dirty &
+       (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_VERTEX_BUFFERS)) {
+      for (uint32_t i = 0; i < pipeline->vi.count; i++) {
+         const uint32_t binding = pipeline->vi.bindings[i];
+         const uint32_t stride = pipeline->vi.strides[i];
+         const struct tu_buffer *buf = cmd->state.vb.buffers[binding];
+         const VkDeviceSize offset = buf->bo_offset +
+                                     cmd->state.vb.offsets[binding] +
+                                     pipeline->vi.offsets[i];
+         const VkDeviceSize size =
+            offset < buf->bo->size ? buf->bo->size - offset : 0;
 
-   if (pipeline->shared_consts.dwords > 0) {
-      dwords = pipeline->shared_consts.dwords + 4;
-   } else {
-      if (compute) {
-         dwords = tu6_user_consts_size(pipeline, MESA_SHADER_COMPUTE);
-      } else {
-         for (uint32_t type = MESA_SHADER_VERTEX; type <= MESA_SHADER_FRAGMENT; type++)
-            dwords += tu6_user_consts_size(pipeline, type);
+         tu_cs_emit_pkt4(cs, REG_A6XX_VFD_FETCH(i), 4);
+         tu_cs_emit_qw(cs, buf->bo->iova + offset);
+         tu_cs_emit(cs, size);
+         tu_cs_emit(cs, stride);
       }
    }
 
-   return dwords;
-}
-
-static struct tu_draw_state
-tu6_emit_consts(struct tu_cmd_buffer *cmd,
-                const struct tu_pipeline *pipeline,
-                bool compute)
-{
-   uint32_t dwords = 0;
-
-   dwords = tu6_const_size(cmd, pipeline, compute);
-
-   if (dwords == 0)
-      return (struct tu_draw_state) {};
-
-   struct tu_cs cs;
-   tu_cs_begin_sub_stream(&cmd->sub_cs, dwords, &cs);
-
-   if (pipeline->shared_consts.dwords > 0) {
-      tu6_emit_shared_consts(&cs, pipeline, cmd->push_constants, compute);
-
-      for (uint32_t i = 0; i < ARRAY_SIZE(pipeline->program.link); i++) {
-         const struct tu_program_descriptor_linkage *link =
-            &pipeline->program.link[i];
-         assert(!link->tu_const_state.push_consts.dwords);
-      }
-   } else {
-      if (compute) {
-         tu6_emit_user_consts(&cs, pipeline, MESA_SHADER_COMPUTE, cmd->push_constants);
+   /* TODO shader consts */
+
+   if (cmd->state.dirty & TU_CMD_DIRTY_PIPELINE) {
+      draw_state_groups[draw_state_group_count++] =
+         (struct tu_draw_state_group) {
+            .id = TU_DRAW_STATE_PROGRAM,
+            .enable_mask = 0x6,
+            .ib = &pipeline->program.state_ib,
+         };
+      draw_state_groups[draw_state_group_count++] =
+         (struct tu_draw_state_group) {
+            .id = TU_DRAW_STATE_PROGRAM_BINNING,
+            .enable_mask = 0x1,
+            .ib = &pipeline->program.binning_state_ib,
+         };
+      draw_state_groups[draw_state_group_count++] =
+         (struct tu_draw_state_group) {
+            .id = TU_DRAW_STATE_VI,
+            .enable_mask = 0x6,
+            .ib = &pipeline->vi.state_ib,
+         };
+      draw_state_groups[draw_state_group_count++] =
+         (struct tu_draw_state_group) {
+            .id = TU_DRAW_STATE_VI_BINNING,
+            .enable_mask = 0x1,
+            .ib = &pipeline->vi.binning_state_ib,
+         };
+      draw_state_groups[draw_state_group_count++] =
+         (struct tu_draw_state_group) {
+            .id = TU_DRAW_STATE_VP,
+            .enable_mask = 0x7,
+            .ib = &pipeline->vp.state_ib,
+         };
+      draw_state_groups[draw_state_group_count++] =
+         (struct tu_draw_state_group) {
+            .id = TU_DRAW_STATE_RAST,
+            .enable_mask = 0x7,
+            .ib = &pipeline->rast.state_ib,
+         };
+      draw_state_groups[draw_state_group_count++] =
+         (struct tu_draw_state_group) {
+            .id = TU_DRAW_STATE_DS,
+            .enable_mask = 0x7,
+            .ib = &pipeline->ds.state_ib,
+         };
+      draw_state_groups[draw_state_group_count++] =
+         (struct tu_draw_state_group) {
+            .id = TU_DRAW_STATE_BLEND,
+            .enable_mask = 0x7,
+            .ib = &pipeline->blend.state_ib,
+         };
+   }
+
+   tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * draw_state_group_count);
+   for (uint32_t i = 0; i < draw_state_group_count; i++) {
+      const struct tu_draw_state_group *group = &draw_state_groups[i];
+
+      uint32_t cp_set_draw_state =
+         CP_SET_DRAW_STATE__0_COUNT(group->ib->size / 4) |
+         CP_SET_DRAW_STATE__0_ENABLE_MASK(group->enable_mask) |
+         CP_SET_DRAW_STATE__0_GROUP_ID(group->id);
+      uint64_t iova;
+      if (group->ib->size) {
+         iova = group->ib->bo->iova + group->ib->offset;
       } else {
-         for (uint32_t type = MESA_SHADER_VERTEX; type <= MESA_SHADER_FRAGMENT; type++)
-            tu6_emit_user_consts(&cs, pipeline, type, cmd->push_constants);
+         cp_set_draw_state |= CP_SET_DRAW_STATE__0_DISABLE;
+         iova = 0;
       }
-   }
 
-   return tu_cs_end_draw_state(&cmd->sub_cs, &cs);
-}
-
-/* Various frontends (ANGLE, zink at least) will enable stencil testing with
- * what works out to be no-op writes.  Simplify what they give us into flags
- * that LRZ can use.
- */
-static void
-tu6_update_simplified_stencil_state(struct tu_cmd_buffer *cmd)
-{
-   bool stencil_test_enable =
-      cmd->state.rb_stencil_cntl & A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE;
-
-   if (!stencil_test_enable) {
-      cmd->state.stencil_front_write = false;
-      cmd->state.stencil_back_write = false;
-      return;
-   }
-
-   bool stencil_front_writemask =
-      (cmd->state.pipeline->dynamic_state_mask & BIT(VK_DYNAMIC_STATE_STENCIL_WRITE_MASK)) ?
-      (cmd->state.dynamic_stencil_wrmask & 0xff) :
-      (cmd->state.pipeline->ds.stencil_wrmask & 0xff);
-
-   bool stencil_back_writemask =
-      (cmd->state.pipeline->dynamic_state_mask & BIT(VK_DYNAMIC_STATE_STENCIL_WRITE_MASK)) ?
-      ((cmd->state.dynamic_stencil_wrmask & 0xff00) >> 8) :
-      (cmd->state.pipeline->ds.stencil_wrmask & 0xff00) >> 8;
-
-   VkStencilOp front_fail_op =
-      (cmd->state.rb_stencil_cntl & A6XX_RB_STENCIL_CONTROL_FAIL__MASK) >> A6XX_RB_STENCIL_CONTROL_FAIL__SHIFT;
-   VkStencilOp front_pass_op =
-      (cmd->state.rb_stencil_cntl & A6XX_RB_STENCIL_CONTROL_ZPASS__MASK) >> A6XX_RB_STENCIL_CONTROL_ZPASS__SHIFT;
-   VkStencilOp front_depth_fail_op =
-      (cmd->state.rb_stencil_cntl & A6XX_RB_STENCIL_CONTROL_ZFAIL__MASK) >> A6XX_RB_STENCIL_CONTROL_ZFAIL__SHIFT;
-   VkStencilOp back_fail_op =
-      (cmd->state.rb_stencil_cntl & A6XX_RB_STENCIL_CONTROL_FAIL_BF__MASK) >> A6XX_RB_STENCIL_CONTROL_FAIL_BF__SHIFT;
-   VkStencilOp back_pass_op =
-      (cmd->state.rb_stencil_cntl & A6XX_RB_STENCIL_CONTROL_ZPASS_BF__MASK) >> A6XX_RB_STENCIL_CONTROL_ZPASS_BF__SHIFT;
-   VkStencilOp back_depth_fail_op =
-      (cmd->state.rb_stencil_cntl & A6XX_RB_STENCIL_CONTROL_ZFAIL_BF__MASK) >> A6XX_RB_STENCIL_CONTROL_ZFAIL_BF__SHIFT;
-
-   bool stencil_front_op_writes =
-      front_pass_op != VK_STENCIL_OP_KEEP ||
-      front_fail_op != VK_STENCIL_OP_KEEP ||
-      front_depth_fail_op != VK_STENCIL_OP_KEEP;
-
-   bool stencil_back_op_writes =
-      back_pass_op != VK_STENCIL_OP_KEEP ||
-      back_fail_op != VK_STENCIL_OP_KEEP ||
-      back_depth_fail_op != VK_STENCIL_OP_KEEP;
-
-   cmd->state.stencil_front_write =
-      stencil_front_op_writes && stencil_front_writemask;
-   cmd->state.stencil_back_write =
-      stencil_back_op_writes && stencil_back_writemask;
-}
-
-static bool
-tu6_writes_depth(struct tu_cmd_buffer *cmd, bool depth_test_enable)
-{
-   bool depth_write_enable =
-      cmd->state.rb_depth_cntl & A6XX_RB_DEPTH_CNTL_Z_WRITE_ENABLE;
-
-   VkCompareOp depth_compare_op =
-      (cmd->state.rb_depth_cntl & A6XX_RB_DEPTH_CNTL_ZFUNC__MASK) >> A6XX_RB_DEPTH_CNTL_ZFUNC__SHIFT;
-
-   bool depth_compare_op_writes = depth_compare_op != VK_COMPARE_OP_NEVER;
-
-   return depth_test_enable && depth_write_enable && depth_compare_op_writes;
-}
-
-static bool
-tu6_writes_stencil(struct tu_cmd_buffer *cmd)
-{
-   return cmd->state.stencil_front_write || cmd->state.stencil_back_write;
-}
-
-static void
-tu6_build_depth_plane_z_mode(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
-   enum a6xx_ztest_mode zmode = A6XX_EARLY_Z;
-   bool depth_test_enable = cmd->state.rb_depth_cntl & A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE;
-   bool depth_write = tu6_writes_depth(cmd, depth_test_enable);
-   bool stencil_write = tu6_writes_stencil(cmd);
-
-   if ((cmd->state.pipeline->lrz.fs.has_kill ||
-        cmd->state.pipeline->output.subpass_feedback_loop_ds) &&
-       (depth_write || stencil_write)) {
-      zmode = (cmd->state.lrz.valid && cmd->state.lrz.enabled)
-                 ? A6XX_EARLY_LRZ_LATE_Z
-                 : A6XX_LATE_Z;
+      tu_cs_emit(cs, cp_set_draw_state);
+      tu_cs_emit_qw(cs, iova);
    }
 
-   if ((cmd->state.pipeline->lrz.force_late_z &&
-        !cmd->state.pipeline->lrz.fs.force_early_z) || !depth_test_enable)
-      zmode = A6XX_LATE_Z;
-
-   /* User defined early tests take precedence above all else */
-   if (cmd->state.pipeline->lrz.fs.early_fragment_tests)
-      zmode = A6XX_EARLY_Z;
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_DEPTH_PLANE_CNTL, 1);
-   tu_cs_emit(cs, A6XX_GRAS_SU_DEPTH_PLANE_CNTL_Z_MODE(zmode));
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_PLANE_CNTL, 1);
-   tu_cs_emit(cs, A6XX_RB_DEPTH_PLANE_CNTL_Z_MODE(zmode));
-}
+   tu_cs_sanity_check(cs);
 
-static void
-tu6_emit_blend(struct tu_cs *cs, struct tu_cmd_buffer *cmd)
-{
-   struct tu_pipeline *pipeline = cmd->state.pipeline;
-   uint32_t color_write_enable = cmd->state.pipeline_color_write_enable;
-
-   if (pipeline->dynamic_state_mask &
-       BIT(TU_DYNAMIC_STATE_COLOR_WRITE_ENABLE))
-      color_write_enable &= cmd->state.color_write_enable;
-
-   for (unsigned i = 0; i < pipeline->blend.num_rts; i++) {
-      tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_CONTROL(i), 2);
-      if (color_write_enable & BIT(i)) {
-         tu_cs_emit(cs, cmd->state.rb_mrt_control[i] |
-                        ((cmd->state.logic_op_enabled ?
-                          cmd->state.rb_mrt_control_rop : 0) &
-                         ~pipeline->blend.rb_mrt_control_mask));
-         tu_cs_emit(cs, cmd->state.rb_mrt_blend_control[i]);
-      } else {
-         tu_cs_emit(cs, 0);
-         tu_cs_emit(cs, 0);
+   /* track BOs */
+   if (cmd->state.dirty & TU_CMD_DIRTY_PIPELINE) {
+      tu_bo_list_add(&cmd->bo_list, &pipeline->program.binary_bo,
+                     MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+      for (uint32_t i = 0; i < pipeline->cs.bo_count; i++) {
+         tu_bo_list_add(&cmd->bo_list, pipeline->cs.bos[i],
+                        MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
       }
    }
-
-   uint32_t blend_enable_mask = color_write_enable;
-   if (!(cmd->state.logic_op_enabled && cmd->state.rop_reads_dst))
-      blend_enable_mask &= cmd->state.pipeline_blend_enable;
-
-   tu_cs_emit_regs(cs, A6XX_SP_FS_OUTPUT_CNTL1(.mrt = pipeline->blend.num_rts));
-   tu_cs_emit_regs(cs, A6XX_RB_FS_OUTPUT_CNTL1(.mrt = pipeline->blend.num_rts));
-   tu_cs_emit_pkt4(cs, REG_A6XX_SP_BLEND_CNTL, 1);
-   tu_cs_emit(cs, cmd->state.sp_blend_cntl |
-                  (A6XX_SP_BLEND_CNTL_ENABLE_BLEND(blend_enable_mask) &
-                   ~pipeline->blend.sp_blend_cntl_mask));
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLEND_CNTL, 1);
-   tu_cs_emit(cs, cmd->state.rb_blend_cntl |
-                  (A6XX_RB_BLEND_CNTL_ENABLE_BLEND(blend_enable_mask) &
-                   ~pipeline->blend.rb_blend_cntl_mask));
-}
-
-static VkResult
-tu6_draw_common(struct tu_cmd_buffer *cmd,
-                struct tu_cs *cs,
-                bool indexed,
-                /* note: draw_count is 0 for indirect */
-                uint32_t draw_count)
-{
-   const struct tu_pipeline *pipeline = cmd->state.pipeline;
-   struct tu_render_pass_state *rp = &cmd->state.rp;
-
-   /* Fill draw stats for autotuner */
-   rp->drawcall_count++;
-
-   rp->drawcall_bandwidth_per_sample_sum +=
-      pipeline->output.color_bandwidth_per_sample;
-
-   /* add depth memory bandwidth cost */
-   const uint32_t depth_bandwidth = pipeline->output.depth_cpp_per_sample;
-   if (cmd->state.rb_depth_cntl & A6XX_RB_DEPTH_CNTL_Z_WRITE_ENABLE)
-      rp->drawcall_bandwidth_per_sample_sum += depth_bandwidth;
-   if (cmd->state.rb_depth_cntl & A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE)
-      rp->drawcall_bandwidth_per_sample_sum += depth_bandwidth;
-
-   /* add stencil memory bandwidth cost */
-   const uint32_t stencil_bandwidth = pipeline->output.stencil_cpp_per_sample;
-   if (cmd->state.rb_stencil_cntl & A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE)
-      rp->drawcall_bandwidth_per_sample_sum += stencil_bandwidth * 2;
-
-   tu_emit_cache_flush_renderpass(cmd, cs);
-
-   bool primitive_restart_enabled = pipeline->ia.primitive_restart;
-   if (pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_PRIMITIVE_RESTART_ENABLE))
-      primitive_restart_enabled = cmd->state.primitive_restart_enable;
-
-   bool primitive_restart = primitive_restart_enabled && indexed;
-   bool provoking_vtx_last = pipeline->rast.provoking_vertex_last;
-   bool tess_upper_left_domain_origin =
-      pipeline->tess.upper_left_domain_origin;
-
-   struct tu_primitive_params* prim_params = &cmd->state.last_prim_params;
-
-   if (!prim_params->valid ||
-       prim_params->primitive_restart != primitive_restart ||
-       prim_params->provoking_vtx_last != provoking_vtx_last ||
-       prim_params->tess_upper_left_domain_origin !=
-          tess_upper_left_domain_origin) {
-      tu_cs_emit_regs(
-         cs,
-         A6XX_PC_PRIMITIVE_CNTL_0(.primitive_restart = primitive_restart,
-                                  .provoking_vtx_last = provoking_vtx_last,
-         .tess_upper_left_domain_origin =
-                                     tess_upper_left_domain_origin));
-      prim_params->valid = true;
-      prim_params->primitive_restart = primitive_restart;
-      prim_params->provoking_vtx_last = provoking_vtx_last;
-      prim_params->tess_upper_left_domain_origin = tess_upper_left_domain_origin;
-   }
-
-   /* Early exit if there is nothing to emit, saves CPU cycles */
-   uint32_t dirty = cmd->state.dirty;
-   if (!(dirty & ~TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD))
-      return VK_SUCCESS;
-
-   bool dirty_lrz =
-      dirty & (TU_CMD_DIRTY_LRZ | TU_CMD_DIRTY_RB_DEPTH_CNTL |
-                          TU_CMD_DIRTY_RB_STENCIL_CNTL | TU_CMD_DIRTY_BLEND);
-
-   if (dirty_lrz) {
-      struct tu_cs cs;
-      uint32_t size = cmd->device->physical_device->info->a6xx.lrz_track_quirk ? 10 : 8;
-
-      cmd->state.lrz_and_depth_plane_state =
-         tu_cs_draw_state(&cmd->sub_cs, &cs, size);
-      tu6_update_simplified_stencil_state(cmd);
-      tu6_emit_lrz(cmd, &cs);
-      tu6_build_depth_plane_z_mode(cmd, &cs);
-   }
-
-   if (dirty & TU_CMD_DIRTY_RASTERIZER_DISCARD) {
-      struct tu_cs cs = tu_cmd_dynamic_state(cmd, TU_DYNAMIC_STATE_RASTERIZER_DISCARD, 4);
-      tu_cs_emit_regs(&cs, A6XX_PC_RASTER_CNTL(.dword = cmd->state.pc_raster_cntl));
-      tu_cs_emit_regs(&cs, A6XX_VPC_UNKNOWN_9107(.dword = cmd->state.vpc_unknown_9107));
-   }
-
-   if (dirty & TU_CMD_DIRTY_GRAS_SU_CNTL) {
-      struct tu_cs cs = tu_cmd_dynamic_state(cmd, TU_DYNAMIC_STATE_GRAS_SU_CNTL, 2);
-      tu_cs_emit_regs(&cs, A6XX_GRAS_SU_CNTL(.dword = cmd->state.gras_su_cntl));
-   }
-
-   if (dirty & TU_CMD_DIRTY_RB_DEPTH_CNTL) {
-      struct tu_cs cs = tu_cmd_dynamic_state(cmd, TU_DYNAMIC_STATE_RB_DEPTH_CNTL, 2);
-      uint32_t rb_depth_cntl = cmd->state.rb_depth_cntl;
-
-      if ((rb_depth_cntl & A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE) ||
-          (rb_depth_cntl & A6XX_RB_DEPTH_CNTL_Z_BOUNDS_ENABLE))
-         rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_READ_ENABLE;
-
-      if ((rb_depth_cntl & A6XX_RB_DEPTH_CNTL_Z_BOUNDS_ENABLE) &&
-          !(rb_depth_cntl & A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE))
-         tu6_apply_depth_bounds_workaround(cmd->device, &rb_depth_cntl);
-
-      if (pipeline->output.rb_depth_cntl_disable)
-         rb_depth_cntl = 0;
-
-      tu_cs_emit_regs(&cs, A6XX_RB_DEPTH_CNTL(.dword = rb_depth_cntl));
-   }
-
-   if (dirty & TU_CMD_DIRTY_RB_STENCIL_CNTL) {
-      struct tu_cs cs = tu_cmd_dynamic_state(cmd, TU_DYNAMIC_STATE_RB_STENCIL_CNTL, 2);
-      tu_cs_emit_regs(&cs, A6XX_RB_STENCIL_CONTROL(.dword = cmd->state.rb_stencil_cntl));
-   }
-
-   if (dirty & TU_CMD_DIRTY_SHADER_CONSTS)
-      cmd->state.shader_const = tu6_emit_consts(cmd, pipeline, false);
-
-   if (dirty & TU_CMD_DIRTY_VIEWPORTS) {
-      struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_VIEWPORT, 8 + 10 * cmd->state.max_viewport);
-      tu6_emit_viewport(&cs, cmd->state.viewport, cmd->state.max_viewport,
-                        pipeline->viewport.z_negative_one_to_one);
-   }
-
-   if (dirty & TU_CMD_DIRTY_BLEND) {
-      struct tu_cs cs = tu_cmd_dynamic_state(cmd, TU_DYNAMIC_STATE_BLEND,
-                                             8 + 3 * cmd->state.pipeline->blend.num_rts);
-      tu6_emit_blend(&cs, cmd);
-   }
-
-   if (dirty & TU_CMD_DIRTY_PATCH_CONTROL_POINTS) {
-      bool tess = cmd->state.pipeline->active_stages &
-         VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT;
-      uint32_t state_size = TU6_EMIT_PATCH_CONTROL_POINTS_DWORDS(
-         pipeline->program.hs_param_dwords);
-      struct tu_cs cs = tu_cmd_dynamic_state(
-         cmd, TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS, tess ? state_size : 0);
-      tu6_emit_patch_control_points(&cs, cmd->state.pipeline,
-                                    cmd->state.patch_control_points);
-   }
-
-   /* for the first draw in a renderpass, re-emit all the draw states
-    *
-    * and if a draw-state disabling path (CmdClearAttachments 3D fallback) was
-    * used, then draw states must be re-emitted. note however this only happens
-    * in the sysmem path, so this can be skipped this for the gmem path (TODO)
-    *
-    * the two input attachment states are excluded because secondary command
-    * buffer doesn't have a state ib to restore it, and not re-emitting them
-    * is OK since CmdClearAttachments won't disable/overwrite them
-    */
-   if (dirty & TU_CMD_DIRTY_DRAW_STATE) {
-      tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * (TU_DRAW_STATE_COUNT - 2));
-
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM_CONFIG, pipeline->program.config_state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM, pipeline->program.state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM_BINNING, pipeline->program.binning_state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_RAST, pipeline->rast.state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PRIM_MODE_SYSMEM, pipeline->prim_order.state_sysmem);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PRIM_MODE_GMEM, pipeline->prim_order.state_gmem);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_CONST, cmd->state.shader_const);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DESC_SETS, cmd->state.desc_sets);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DESC_SETS_LOAD, pipeline->load_state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VB, cmd->state.vertex_buffers);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VS_PARAMS, cmd->state.vs_params);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_LRZ_AND_DEPTH_PLANE, cmd->state.lrz_and_depth_plane_state);
-      tu_cs_emit_draw_state(cs, TU_DRAW_STATE_MSAA, cmd->state.msaa);
-
-      for (uint32_t i = 0; i < ARRAY_SIZE(cmd->state.dynamic_state); i++) {
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + i,
-                               ((pipeline->dynamic_state_mask & BIT(i)) ?
-                                cmd->state.dynamic_state[i] :
-                                pipeline->dynamic_state[i]));
-      }
-   } else {
-      /* emit draw states that were just updated
-       * note we eventually don't want to have to emit anything here
-       */
-      bool emit_binding_stride = false, emit_blend = false,
-           emit_patch_control_points = false;
-      uint32_t draw_state_count =
-         ((dirty & TU_CMD_DIRTY_SHADER_CONSTS) ? 1 : 0) +
-         ((dirty & TU_CMD_DIRTY_DESC_SETS_LOAD) ? 1 : 0) +
-         ((dirty & TU_CMD_DIRTY_VERTEX_BUFFERS) ? 1 : 0) +
-         ((dirty & TU_CMD_DIRTY_VS_PARAMS) ? 1 : 0) +
-         (dirty_lrz ? 1 : 0);
-
-      if ((dirty & TU_CMD_DIRTY_VB_STRIDE) &&
-          (pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_VB_STRIDE))) {
-         emit_binding_stride = true;
-         draw_state_count += 1;
-      }
-
-      if ((dirty & TU_CMD_DIRTY_BLEND) &&
-          (pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_BLEND))) {
-         emit_blend = true;
-         draw_state_count += 1;
-      }
-
-      if ((dirty & TU_CMD_DIRTY_PATCH_CONTROL_POINTS) &&
-          (pipeline->dynamic_state_mask &
-           BIT(TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS))) {
-         emit_patch_control_points = true;
-         draw_state_count += 1;
-      }
-
-      if (draw_state_count > 0)
-         tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * draw_state_count);
-
-      if (dirty & TU_CMD_DIRTY_SHADER_CONSTS)
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_CONST, cmd->state.shader_const);
-      if (dirty & TU_CMD_DIRTY_DESC_SETS_LOAD)
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DESC_SETS_LOAD, pipeline->load_state);
-      if (dirty & TU_CMD_DIRTY_VERTEX_BUFFERS)
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VB, cmd->state.vertex_buffers);
-      if (emit_binding_stride) {
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + TU_DYNAMIC_STATE_VB_STRIDE,
-                               cmd->state.dynamic_state[TU_DYNAMIC_STATE_VB_STRIDE]);
-      }
-      if (emit_blend) {
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + TU_DYNAMIC_STATE_BLEND,
-                               cmd->state.dynamic_state[TU_DYNAMIC_STATE_BLEND]);
-      }
-      if (emit_patch_control_points) {
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS,
-                               cmd->state.dynamic_state[TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS]);
-      }
-      if (dirty & TU_CMD_DIRTY_VS_PARAMS)
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VS_PARAMS, cmd->state.vs_params);
-
-      if (dirty_lrz) {
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_LRZ_AND_DEPTH_PLANE, cmd->state.lrz_and_depth_plane_state);
+   if (cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS) {
+      for (uint32_t i = 0; i < MAX_VBS; i++) {
+         const struct tu_buffer *buf = cmd->state.vb.buffers[i];
+         if (buf)
+            tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
       }
    }
 
-   tu_cs_sanity_check(cs);
-
-   /* There are too many graphics dirty bits to list here, so just list the
-    * bits to preserve instead. The only things not emitted here are
-    * compute-related state.
-    */
-   cmd->state.dirty &= TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD;
-   return VK_SUCCESS;
+   cmd->state.dirty = 0;
 }
 
-static uint32_t
-tu_draw_initiator(struct tu_cmd_buffer *cmd, enum pc_di_src_sel src_sel)
+static void
+tu6_emit_draw_direct(struct tu_cmd_buffer *cmd,
+                     struct tu_cs *cs,
+                     const struct tu_draw_info *draw)
 {
-   const struct tu_pipeline *pipeline = cmd->state.pipeline;
-   enum pc_di_primtype primtype = cmd->state.primtype;
 
-   if (primtype == DI_PT_PATCHES0)
-      primtype += cmd->state.patch_control_points;
+   const enum pc_di_primtype primtype = cmd->state.pipeline->ia.primtype;
 
-   uint32_t initiator =
-      CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
-      CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(src_sel) |
-      CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(cmd->state.index_size) |
-      CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY);
+   tu_cs_emit_pkt4(cs, REG_A6XX_VFD_INDEX_OFFSET, 2);
+   tu_cs_emit(cs, draw->vertex_offset);
+   tu_cs_emit(cs, draw->first_instance);
 
-   if (pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT)
-      initiator |= CP_DRAW_INDX_OFFSET_0_GS_ENABLE;
-
-   switch (pipeline->tess.patch_type) {
-   case IR3_TESS_TRIANGLES:
-      initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_TRIANGLES) |
-                   CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
-      break;
-   case IR3_TESS_ISOLINES:
-      initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_ISOLINES) |
-                   CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
-      break;
-   case IR3_TESS_NONE:
-      initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_QUADS);
-      break;
-   case IR3_TESS_QUADS:
-      initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_QUADS) |
-                   CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
-      break;
-   }
-   return initiator;
-}
+   /* TODO hw binning */
+   if (draw->indexed) {
+      const enum a4xx_index_size index_size =
+         tu6_index_size(cmd->state.index_type);
+      const uint32_t index_bytes =
+         (cmd->state.index_type == VK_INDEX_TYPE_UINT32) ? 4 : 2;
+      const struct tu_buffer *buf = cmd->state.index_buffer;
+      const VkDeviceSize offset = buf->bo_offset + cmd->state.index_offset +
+                                  index_bytes * draw->first_index;
+      const uint32_t size = index_bytes * draw->count;
 
+      const uint32_t cp_draw_indx =
+         CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
+         CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_DMA) |
+         CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(index_size) |
+         CP_DRAW_INDX_OFFSET_0_VIS_CULL(IGNORE_VISIBILITY) | 0x2000;
 
-static uint32_t
-vs_params_offset(struct tu_cmd_buffer *cmd)
-{
-   const struct tu_program_descriptor_linkage *link =
-      &cmd->state.pipeline->program.link[MESA_SHADER_VERTEX];
-   const struct ir3_const_state *const_state = &link->const_state;
-
-   if (const_state->offsets.driver_param >= link->constlen)
-      return 0;
-
-   /* this layout is required by CP_DRAW_INDIRECT_MULTI */
-   STATIC_ASSERT(IR3_DP_DRAWID == 0);
-   STATIC_ASSERT(IR3_DP_VTXID_BASE == 1);
-   STATIC_ASSERT(IR3_DP_INSTID_BASE == 2);
-
-   /* 0 means disabled for CP_DRAW_INDIRECT_MULTI */
-   assert(const_state->offsets.driver_param != 0);
-
-   return const_state->offsets.driver_param;
-}
+      tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
+      tu_cs_emit(cs, cp_draw_indx);
+      tu_cs_emit(cs, draw->instance_count);
+      tu_cs_emit(cs, draw->count);
+      tu_cs_emit(cs, 0x0); /* XXX */
+      tu_cs_emit_qw(cs, buf->bo->iova + offset);
+      tu_cs_emit(cs, size);
+   } else {
+      const uint32_t cp_draw_indx =
+         CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
+         CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_AUTO_INDEX) |
+         CP_DRAW_INDX_OFFSET_0_VIS_CULL(IGNORE_VISIBILITY) | 0x2000;
 
-static void
-tu6_emit_empty_vs_params(struct tu_cmd_buffer *cmd)
-{
-   if (cmd->state.vs_params.iova) {
-      cmd->state.vs_params = (struct tu_draw_state) {};
-      cmd->state.dirty |= TU_CMD_DIRTY_VS_PARAMS;
+      tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
+      tu_cs_emit(cs, cp_draw_indx);
+      tu_cs_emit(cs, draw->instance_count);
+      tu_cs_emit(cs, draw->count);
    }
 }
 
 static void
-tu6_emit_vs_params(struct tu_cmd_buffer *cmd,
-                   uint32_t draw_id,
-                   uint32_t vertex_offset,
-                   uint32_t first_instance)
+tu_draw(struct tu_cmd_buffer *cmd, const struct tu_draw_info *draw)
 {
-   uint32_t offset = vs_params_offset(cmd);
+   struct tu_cs *cs = &cmd->draw_cs;
 
-   /* Beside re-emitting params when they are changed, we should re-emit
-    * them after constants are invalidated via HLSQ_INVALIDATE_CMD.
-    */
-   if (!(cmd->state.dirty & (TU_CMD_DIRTY_DRAW_STATE | TU_CMD_DIRTY_VS_PARAMS)) &&
-       (offset == 0 || draw_id == cmd->state.last_vs_params.draw_id) &&
-       vertex_offset == cmd->state.last_vs_params.vertex_offset &&
-       first_instance == cmd->state.last_vs_params.first_instance) {
-      return;
-   }
+   tu6_bind_draw_states(cmd, cs, draw);
 
-   struct tu_cs cs;
-   VkResult result = tu_cs_begin_sub_stream(&cmd->sub_cs, 3 + (offset ? 8 : 0), &cs);
+   VkResult result = tu_cs_reserve_space(cmd->device, cs, 32);
    if (result != VK_SUCCESS) {
-      vk_command_buffer_set_error(&cmd->vk, result);
+      cmd->record_result = result;
       return;
    }
 
-   tu_cs_emit_regs(&cs,
-                   A6XX_VFD_INDEX_OFFSET(vertex_offset),
-                   A6XX_VFD_INSTANCE_START_OFFSET(first_instance));
-
-   if (offset) {
-      tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_GEOM, 3 + 4);
-      tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(offset) |
-            CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-            CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-            CP_LOAD_STATE6_0_STATE_BLOCK(SB6_VS_SHADER) |
-            CP_LOAD_STATE6_0_NUM_UNIT(1));
-      tu_cs_emit(&cs, 0);
-      tu_cs_emit(&cs, 0);
-
-      tu_cs_emit(&cs, draw_id);
-      tu_cs_emit(&cs, vertex_offset);
-      tu_cs_emit(&cs, first_instance);
-      tu_cs_emit(&cs, 0);
+   if (draw->indirect) {
+      tu_finishme("indirect draw");
+      return;
    }
 
-   cmd->state.last_vs_params.vertex_offset = vertex_offset;
-   cmd->state.last_vs_params.first_instance = first_instance;
-   cmd->state.last_vs_params.draw_id = draw_id;
+   /* TODO tu6_emit_marker should pick different regs depending on cs */
+   tu6_emit_marker(cmd, cs);
+   tu6_emit_draw_direct(cmd, cs, draw);
+   tu6_emit_marker(cmd, cs);
 
-   struct tu_cs_entry entry = tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
-   cmd->state.vs_params = (struct tu_draw_state) {entry.bo->iova + entry.offset, entry.size / 4};
+   cmd->wait_for_idle = true;
 
-   cmd->state.dirty |= TU_CMD_DIRTY_VS_PARAMS;
+   tu_cs_sanity_check(cs);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdDraw(VkCommandBuffer commandBuffer,
            uint32_t vertexCount,
            uint32_t instanceCount,
            uint32_t firstVertex,
            uint32_t firstInstance)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu6_emit_vs_params(cmd, 0, firstVertex, firstInstance);
-
-   tu6_draw_common(cmd, cs, false, vertexCount);
-
-   tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
-   tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
-   tu_cs_emit(cs, instanceCount);
-   tu_cs_emit(cs, vertexCount);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdDrawMultiEXT(VkCommandBuffer commandBuffer,
-                   uint32_t drawCount,
-                   const VkMultiDrawInfoEXT *pVertexInfo,
-                   uint32_t instanceCount,
-                   uint32_t firstInstance,
-                   uint32_t stride)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   if (!drawCount)
-      return;
-
-   bool has_tess =
-         cmd->state.pipeline->active_stages & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT;
-
-   uint32_t max_vertex_count = 0;
-   if (has_tess) {
-      uint32_t i = 0;
-      vk_foreach_multi_draw(draw, i, pVertexInfo, drawCount, stride) {
-         max_vertex_count = MAX2(max_vertex_count, draw->vertexCount);
-      }
-   }
-
-   uint32_t i = 0;
-   vk_foreach_multi_draw(draw, i, pVertexInfo, drawCount, stride) {
-      tu6_emit_vs_params(cmd, i, draw->firstVertex, firstInstance);
-
-      if (i == 0)
-         tu6_draw_common(cmd, cs, false, max_vertex_count);
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   struct tu_draw_info info = {};
 
-      if (cmd->state.dirty & TU_CMD_DIRTY_VS_PARAMS) {
-         tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VS_PARAMS, cmd->state.vs_params);
-         cmd->state.dirty &= ~TU_CMD_DIRTY_VS_PARAMS;
-      }
+   info.count = vertexCount;
+   info.instance_count = instanceCount;
+   info.first_instance = firstInstance;
+   info.vertex_offset = firstVertex;
 
-      tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
-      tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
-      tu_cs_emit(cs, instanceCount);
-      tu_cs_emit(cs, draw->vertexCount);
-   }
+   tu_draw(cmd_buffer, &info);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdDrawIndexed(VkCommandBuffer commandBuffer,
                   uint32_t indexCount,
                   uint32_t instanceCount,
@@ -4828,239 +2377,56 @@ tu_CmdDrawIndexed(VkCommandBuffer commandBuffer,
                   int32_t vertexOffset,
                   uint32_t firstInstance)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu6_emit_vs_params(cmd, 0, vertexOffset, firstInstance);
-
-   tu6_draw_common(cmd, cs, true, indexCount);
-
-   tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
-   tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
-   tu_cs_emit(cs, instanceCount);
-   tu_cs_emit(cs, indexCount);
-   tu_cs_emit(cs, firstIndex);
-   tu_cs_emit_qw(cs, cmd->state.index_va);
-   tu_cs_emit(cs, cmd->state.max_index_count);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdDrawMultiIndexedEXT(VkCommandBuffer commandBuffer,
-                          uint32_t drawCount,
-                          const VkMultiDrawIndexedInfoEXT *pIndexInfo,
-                          uint32_t instanceCount,
-                          uint32_t firstInstance,
-                          uint32_t stride,
-                          const int32_t *pVertexOffset)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   if (!drawCount)
-      return;
-
-   bool has_tess =
-         cmd->state.pipeline->active_stages & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT;
-
-   uint32_t max_index_count = 0;
-   if (has_tess) {
-      uint32_t i = 0;
-      vk_foreach_multi_draw_indexed(draw, i, pIndexInfo, drawCount, stride) {
-         max_index_count = MAX2(max_index_count, draw->indexCount);
-      }
-   }
-
-   uint32_t i = 0;
-   vk_foreach_multi_draw_indexed(draw, i, pIndexInfo, drawCount, stride) {
-      int32_t vertexOffset = pVertexOffset ? *pVertexOffset : draw->vertexOffset;
-      tu6_emit_vs_params(cmd, i, vertexOffset, firstInstance);
-
-      if (i == 0)
-         tu6_draw_common(cmd, cs, true, max_index_count);
-
-      if (cmd->state.dirty & TU_CMD_DIRTY_VS_PARAMS) {
-         tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
-         tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VS_PARAMS, cmd->state.vs_params);
-         cmd->state.dirty &= ~TU_CMD_DIRTY_VS_PARAMS;
-      }
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   struct tu_draw_info info = {};
 
-      tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
-      tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
-      tu_cs_emit(cs, instanceCount);
-      tu_cs_emit(cs, draw->indexCount);
-      tu_cs_emit(cs, draw->firstIndex);
-      tu_cs_emit_qw(cs, cmd->state.index_va);
-      tu_cs_emit(cs, cmd->state.max_index_count);
-   }
-}
+   info.indexed = true;
+   info.count = indexCount;
+   info.instance_count = instanceCount;
+   info.first_index = firstIndex;
+   info.vertex_offset = vertexOffset;
+   info.first_instance = firstInstance;
 
-/* Various firmware bugs/inconsistencies mean that some indirect draw opcodes
- * do not wait for WFI's to complete before executing. Add a WAIT_FOR_ME if
- * pending for these opcodes. This may result in a few extra WAIT_FOR_ME's
- * with these opcodes, but the alternative would add unnecessary WAIT_FOR_ME's
- * before draw opcodes that don't need it.
- */
-static void
-draw_wfm(struct tu_cmd_buffer *cmd)
-{
-   cmd->state.renderpass_cache.flush_bits |=
-      cmd->state.renderpass_cache.pending_flush_bits & TU_CMD_FLAG_WAIT_FOR_ME;
-   cmd->state.renderpass_cache.pending_flush_bits &= ~TU_CMD_FLAG_WAIT_FOR_ME;
+   tu_draw(cmd_buffer, &info);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdDrawIndirect(VkCommandBuffer commandBuffer,
                    VkBuffer _buffer,
                    VkDeviceSize offset,
                    uint32_t drawCount,
                    uint32_t stride)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_buffer, buf, _buffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu6_emit_empty_vs_params(cmd);
-
-   if (cmd->device->physical_device->info->a6xx.indirect_draw_wfm_quirk)
-      draw_wfm(cmd);
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
+   struct tu_draw_info info = {};
 
-   tu6_draw_common(cmd, cs, false, 0);
+   info.count = drawCount;
+   info.indirect = buffer;
+   info.indirect_offset = offset;
+   info.stride = stride;
 
-   tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT_MULTI, 6);
-   tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
-   tu_cs_emit(cs, A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_NORMAL) |
-                  A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(vs_params_offset(cmd)));
-   tu_cs_emit(cs, drawCount);
-   tu_cs_emit_qw(cs, buf->iova + offset);
-   tu_cs_emit(cs, stride);
+   tu_draw(cmd_buffer, &info);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdDrawIndexedIndirect(VkCommandBuffer commandBuffer,
                           VkBuffer _buffer,
                           VkDeviceSize offset,
                           uint32_t drawCount,
                           uint32_t stride)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_buffer, buf, _buffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu6_emit_empty_vs_params(cmd);
-
-   if (cmd->device->physical_device->info->a6xx.indirect_draw_wfm_quirk)
-      draw_wfm(cmd);
-
-   tu6_draw_common(cmd, cs, true, 0);
-
-   tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT_MULTI, 9);
-   tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
-   tu_cs_emit(cs, A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_INDEXED) |
-                  A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(vs_params_offset(cmd)));
-   tu_cs_emit(cs, drawCount);
-   tu_cs_emit_qw(cs, cmd->state.index_va);
-   tu_cs_emit(cs, cmd->state.max_index_count);
-   tu_cs_emit_qw(cs, buf->iova + offset);
-   tu_cs_emit(cs, stride);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdDrawIndirectCount(VkCommandBuffer commandBuffer,
-                        VkBuffer _buffer,
-                        VkDeviceSize offset,
-                        VkBuffer countBuffer,
-                        VkDeviceSize countBufferOffset,
-                        uint32_t drawCount,
-                        uint32_t stride)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_buffer, buf, _buffer);
-   TU_FROM_HANDLE(tu_buffer, count_buf, countBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu6_emit_empty_vs_params(cmd);
-
-   /* It turns out that the firmware we have for a650 only partially fixed the
-    * problem with CP_DRAW_INDIRECT_MULTI not waiting for WFI's to complete
-    * before reading indirect parameters. It waits for WFI's before reading
-    * the draw parameters, but after reading the indirect count :(.
-    */
-   draw_wfm(cmd);
-
-   tu6_draw_common(cmd, cs, false, 0);
-
-   tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT_MULTI, 8);
-   tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
-   tu_cs_emit(cs, A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_INDIRECT_COUNT) |
-                  A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(vs_params_offset(cmd)));
-   tu_cs_emit(cs, drawCount);
-   tu_cs_emit_qw(cs, buf->iova + offset);
-   tu_cs_emit_qw(cs, count_buf->iova + countBufferOffset);
-   tu_cs_emit(cs, stride);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdDrawIndexedIndirectCount(VkCommandBuffer commandBuffer,
-                               VkBuffer _buffer,
-                               VkDeviceSize offset,
-                               VkBuffer countBuffer,
-                               VkDeviceSize countBufferOffset,
-                               uint32_t drawCount,
-                               uint32_t stride)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_buffer, buf, _buffer);
-   TU_FROM_HANDLE(tu_buffer, count_buf, countBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   tu6_emit_empty_vs_params(cmd);
-
-   draw_wfm(cmd);
-
-   tu6_draw_common(cmd, cs, true, 0);
-
-   tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT_MULTI, 11);
-   tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
-   tu_cs_emit(cs, A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_INDIRECT_COUNT_INDEXED) |
-                  A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(vs_params_offset(cmd)));
-   tu_cs_emit(cs, drawCount);
-   tu_cs_emit_qw(cs, cmd->state.index_va);
-   tu_cs_emit(cs, cmd->state.max_index_count);
-   tu_cs_emit_qw(cs, buf->iova + offset);
-   tu_cs_emit_qw(cs, count_buf->iova + countBufferOffset);
-   tu_cs_emit(cs, stride);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdDrawIndirectByteCountEXT(VkCommandBuffer commandBuffer,
-                               uint32_t instanceCount,
-                               uint32_t firstInstance,
-                               VkBuffer _counterBuffer,
-                               VkDeviceSize counterBufferOffset,
-                               uint32_t counterOffset,
-                               uint32_t vertexStride)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_buffer, buf, _counterBuffer);
-   struct tu_cs *cs = &cmd->draw_cs;
-
-   /* All known firmware versions do not wait for WFI's with CP_DRAW_AUTO.
-    * Plus, for the common case where the counter buffer is written by
-    * vkCmdEndTransformFeedback, we need to wait for the CP_WAIT_MEM_WRITES to
-    * complete which means we need a WAIT_FOR_ME anyway.
-    */
-   draw_wfm(cmd);
-
-   tu6_emit_vs_params(cmd, 0, 0, firstInstance);
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
+   struct tu_draw_info info = {};
 
-   tu6_draw_common(cmd, cs, false, 0);
+   info.indexed = true;
+   info.count = drawCount;
+   info.indirect = buffer;
+   info.indirect_offset = offset;
+   info.stride = stride;
 
-   tu_cs_emit_pkt7(cs, CP_DRAW_AUTO, 6);
-   tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_XFB));
-   tu_cs_emit(cs, instanceCount);
-   tu_cs_emit_qw(cs, buf->iova + counterBufferOffset);
-   tu_cs_emit(cs, counterOffset);
-   tu_cs_emit(cs, vertexStride);
+   tu_draw(cmd_buffer, &info);
 }
 
 struct tu_dispatch_info
@@ -5088,221 +2454,12 @@ struct tu_dispatch_info
 };
 
 static void
-tu_emit_compute_driver_params(struct tu_cmd_buffer *cmd,
-                              struct tu_cs *cs, struct tu_pipeline *pipeline,
-                              const struct tu_dispatch_info *info)
-{
-   gl_shader_stage type = MESA_SHADER_COMPUTE;
-   const struct tu_program_descriptor_linkage *link =
-      &pipeline->program.link[type];
-   const struct ir3_const_state *const_state = &link->const_state;
-   uint32_t offset = const_state->offsets.driver_param;
-   unsigned subgroup_size = pipeline->compute.subgroup_size;
-   unsigned subgroup_shift = util_logbase2(subgroup_size);
-
-   if (link->constlen <= offset)
-      return;
-
-   uint32_t num_consts = MIN2(const_state->num_driver_params,
-                              (link->constlen - offset) * 4);
-
-   if (!info->indirect) {
-      uint32_t driver_params[12] = {
-         [IR3_DP_NUM_WORK_GROUPS_X] = info->blocks[0],
-         [IR3_DP_NUM_WORK_GROUPS_Y] = info->blocks[1],
-         [IR3_DP_NUM_WORK_GROUPS_Z] = info->blocks[2],
-         [IR3_DP_BASE_GROUP_X] = info->offsets[0],
-         [IR3_DP_BASE_GROUP_Y] = info->offsets[1],
-         [IR3_DP_BASE_GROUP_Z] = info->offsets[2],
-         [IR3_DP_CS_SUBGROUP_SIZE] = subgroup_size,
-         [IR3_DP_SUBGROUP_ID_SHIFT] = subgroup_shift,
-      };
-
-      assert(num_consts <= ARRAY_SIZE(driver_params));
-
-      /* push constants */
-      tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + num_consts);
-      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(offset) |
-                 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-                 CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
-                 CP_LOAD_STATE6_0_NUM_UNIT(num_consts / 4));
-      tu_cs_emit(cs, 0);
-      tu_cs_emit(cs, 0);
-      uint32_t i;
-      for (i = 0; i < num_consts; i++)
-         tu_cs_emit(cs, driver_params[i]);
-   } else if (!(info->indirect_offset & 0xf)) {
-      tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3);
-      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(offset) |
-                  CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                  CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
-                  CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
-                  CP_LOAD_STATE6_0_NUM_UNIT(1));
-      tu_cs_emit_qw(cs, info->indirect->iova + info->indirect_offset);
-   } else {
-      /* Vulkan guarantees only 4 byte alignment for indirect_offset.
-       * However, CP_LOAD_STATE.EXT_SRC_ADDR needs 16 byte alignment.
-       */
-
-      uint64_t indirect_iova = info->indirect->iova + info->indirect_offset;
-
-      for (uint32_t i = 0; i < 3; i++) {
-         tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 5);
-         tu_cs_emit(cs, 0);
-         tu_cs_emit_qw(cs, global_iova(cmd, cs_indirect_xyz[i]));
-         tu_cs_emit_qw(cs, indirect_iova + i * 4);
-      }
-
-      tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-      tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE);
-
-      tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3);
-      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(offset) |
-                  CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                  CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
-                  CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
-                  CP_LOAD_STATE6_0_NUM_UNIT(1));
-      tu_cs_emit_qw(cs, global_iova(cmd, cs_indirect_xyz[0]));
-   }
-
-   /* Fill out IR3_DP_CS_SUBGROUP_SIZE and IR3_DP_SUBGROUP_ID_SHIFT for
-    * indirect dispatch.
-    */
-   if (info->indirect && num_consts > IR3_DP_BASE_GROUP_X) {
-      bool emit_local = num_consts > IR3_DP_LOCAL_GROUP_SIZE_X;
-      tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 7 + (emit_local ? 4 : 0));
-      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(offset + (IR3_DP_BASE_GROUP_X / 4)) |
-                 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-                 CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
-                 CP_LOAD_STATE6_0_NUM_UNIT((num_consts - IR3_DP_BASE_GROUP_X) / 4));
-      tu_cs_emit_qw(cs, 0);
-      tu_cs_emit(cs, 0); /* BASE_GROUP_X */
-      tu_cs_emit(cs, 0); /* BASE_GROUP_Y */
-      tu_cs_emit(cs, 0); /* BASE_GROUP_Z */
-      tu_cs_emit(cs, subgroup_size);
-      if (emit_local) {
-         assert(num_consts == align(IR3_DP_SUBGROUP_ID_SHIFT, 4));
-         tu_cs_emit(cs, 0); /* LOCAL_GROUP_SIZE_X */
-         tu_cs_emit(cs, 0); /* LOCAL_GROUP_SIZE_Y */
-         tu_cs_emit(cs, 0); /* LOCAL_GROUP_SIZE_Z */
-         tu_cs_emit(cs, subgroup_shift);
-      }
-   }
-}
-
-static void
-tu_dispatch(struct tu_cmd_buffer *cmd,
+tu_dispatch(struct tu_cmd_buffer *cmd_buffer,
             const struct tu_dispatch_info *info)
 {
-   if (!info->indirect &&
-       (info->blocks[0] == 0 || info->blocks[1] == 0 || info->blocks[2] == 0))
-      return;
-
-   struct tu_cs *cs = &cmd->cs;
-   struct tu_pipeline *pipeline = cmd->state.compute_pipeline;
-
-   bool emit_instrlen_workaround =
-      pipeline->program.cs_instrlen >
-      cmd->device->physical_device->info->a6xx.instr_cache_size;
-
-   /* There appears to be a HW bug where in some rare circumstances it appears
-    * to accidentally use the FS instrlen instead of the CS instrlen, which
-    * affects all known gens. Based on various experiments it appears that the
-    * issue is that when prefetching a branch destination and there is a cache
-    * miss, when fetching from memory the HW bounds-checks the fetch against
-    * SP_CS_INSTRLEN, except when one of the two register contexts is active
-    * it accidentally fetches SP_FS_INSTRLEN from the other (inactive)
-    * context. To workaround it we set the FS instrlen here and do a dummy
-    * event to roll the context (because it fetches SP_FS_INSTRLEN from the
-    * "wrong" context). Because the bug seems to involve cache misses, we
-    * don't emit this if the entire CS program fits in cache, which will
-    * hopefully be the majority of cases.
-    *
-    * See https://gitlab.freedesktop.org/mesa/mesa/-/issues/5892
-    */
-   if (emit_instrlen_workaround) {
-      tu_cs_emit_regs(cs, A6XX_SP_FS_INSTRLEN(pipeline->program.cs_instrlen));
-      tu6_emit_event_write(cmd, cs, LABEL);
-   }
-
-   /* TODO: We could probably flush less if we add a compute_flush_bits
-    * bitfield.
-    */
-   tu_emit_cache_flush(cmd, cs);
-
-   /* note: no reason to have this in a separate IB */
-   tu_cs_emit_state_ib(cs, tu6_emit_consts(cmd, pipeline, true));
-
-   tu_emit_compute_driver_params(cmd, cs, pipeline, info);
-
-   if (cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD)
-      tu_cs_emit_state_ib(cs, pipeline->load_state);
-
-   cmd->state.dirty &= ~TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD;
-
-   tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
-   tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_COMPUTE));
-
-   const uint32_t *local_size = pipeline->compute.local_size;
-   const uint32_t *num_groups = info->blocks;
-   tu_cs_emit_regs(cs,
-                   A6XX_HLSQ_CS_NDRANGE_0(.kerneldim = 3,
-                                          .localsizex = local_size[0] - 1,
-                                          .localsizey = local_size[1] - 1,
-                                          .localsizez = local_size[2] - 1),
-                   A6XX_HLSQ_CS_NDRANGE_1(.globalsize_x = local_size[0] * num_groups[0]),
-                   A6XX_HLSQ_CS_NDRANGE_2(.globaloff_x = 0),
-                   A6XX_HLSQ_CS_NDRANGE_3(.globalsize_y = local_size[1] * num_groups[1]),
-                   A6XX_HLSQ_CS_NDRANGE_4(.globaloff_y = 0),
-                   A6XX_HLSQ_CS_NDRANGE_5(.globalsize_z = local_size[2] * num_groups[2]),
-                   A6XX_HLSQ_CS_NDRANGE_6(.globaloff_z = 0));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_HLSQ_CS_KERNEL_GROUP_X(1),
-                   A6XX_HLSQ_CS_KERNEL_GROUP_Y(1),
-                   A6XX_HLSQ_CS_KERNEL_GROUP_Z(1));
-
-   trace_start_compute(&cmd->trace, cs);
-
-   if (info->indirect) {
-      uint64_t iova = info->indirect->iova + info->indirect_offset;
-
-      tu_cs_emit_pkt7(cs, CP_EXEC_CS_INDIRECT, 4);
-      tu_cs_emit(cs, 0x00000000);
-      tu_cs_emit_qw(cs, iova);
-      tu_cs_emit(cs,
-                 A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEX(local_size[0] - 1) |
-                 A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEY(local_size[1] - 1) |
-                 A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEZ(local_size[2] - 1));
-   } else {
-      tu_cs_emit_pkt7(cs, CP_EXEC_CS, 4);
-      tu_cs_emit(cs, 0x00000000);
-      tu_cs_emit(cs, CP_EXEC_CS_1_NGROUPS_X(info->blocks[0]));
-      tu_cs_emit(cs, CP_EXEC_CS_2_NGROUPS_Y(info->blocks[1]));
-      tu_cs_emit(cs, CP_EXEC_CS_3_NGROUPS_Z(info->blocks[2]));
-   }
-
-   trace_end_compute(&cmd->trace, cs,
-                     info->indirect != NULL,
-                     local_size[0], local_size[1], local_size[2],
-                     info->blocks[0], info->blocks[1], info->blocks[2]);
-
-   /* For the workaround above, because it's using the "wrong" context for
-    * SP_FS_INSTRLEN we should emit another dummy event write to avoid a
-    * potential race between writing the register and the CP_EXEC_CS we just
-    * did. We don't need to reset the register because it will be re-emitted
-    * anyway when the next renderpass starts.
-    */
-   if (emit_instrlen_workaround) {
-      tu6_emit_event_write(cmd, cs, LABEL);
-   }
-
-   tu_cs_emit_wfi(cs);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdDispatchBase(VkCommandBuffer commandBuffer,
                    uint32_t base_x,
                    uint32_t base_y,
@@ -5324,7 +2481,7 @@ tu_CmdDispatchBase(VkCommandBuffer commandBuffer,
    tu_dispatch(cmd_buffer, &info);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdDispatch(VkCommandBuffer commandBuffer,
                uint32_t x,
                uint32_t y,
@@ -5333,7 +2490,7 @@ tu_CmdDispatch(VkCommandBuffer commandBuffer,
    tu_CmdDispatchBase(commandBuffer, 0, 0, 0, x, y, z);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdDispatchIndirect(VkCommandBuffer commandBuffer,
                        VkBuffer _buffer,
                        VkDeviceSize offset)
@@ -5348,410 +2505,133 @@ tu_CmdDispatchIndirect(VkCommandBuffer commandBuffer,
    tu_dispatch(cmd_buffer, &info);
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdEndRenderPass2(VkCommandBuffer commandBuffer,
-                     const VkSubpassEndInfo *pSubpassEndInfo)
+void
+tu_CmdEndRenderPass(VkCommandBuffer commandBuffer)
 {
    TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
 
-   if (unlikely(cmd_buffer->device->instance->debug_flags & TU_DEBUG_DYNAMIC)) {
-      vk_common_CmdEndRenderPass2(commandBuffer, pSubpassEndInfo);
-      return;
-   }
-
    tu_cs_end(&cmd_buffer->draw_cs);
-   tu_cs_end(&cmd_buffer->draw_epilogue_cs);
-   tu_cmd_render(cmd_buffer);
-
-   cmd_buffer->state.cache.pending_flush_bits |=
-      cmd_buffer->state.renderpass_cache.pending_flush_bits;
-   tu_subpass_barrier(cmd_buffer, &cmd_buffer->state.pass->end_barrier, true);
-
-   vk_free(&cmd_buffer->vk.pool->alloc, cmd_buffer->state.attachments);
-
-   tu_reset_render_pass(cmd_buffer);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdEndRendering(VkCommandBuffer commandBuffer)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
-
-   if (cmd_buffer->state.suspending)
-      cmd_buffer->state.suspended_pass.lrz = cmd_buffer->state.lrz;
 
-   if (!cmd_buffer->state.suspending) {
-      tu_cs_end(&cmd_buffer->draw_cs);
-      tu_cs_end(&cmd_buffer->draw_epilogue_cs);
+   tu_cmd_render_tiles(cmd_buffer);
 
-      if (cmd_buffer->state.suspend_resume == SR_IN_PRE_CHAIN) {
-         cmd_buffer->trace_renderpass_end = u_trace_end_iterator(&cmd_buffer->trace);
-         tu_save_pre_chain(cmd_buffer);
-
-         /* Even we don't call tu_cmd_render here, renderpass is finished
-          * and draw states should be disabled.
-          */
-         tu_disable_draw_states(cmd_buffer, &cmd_buffer->cs);
-      } else {
-         tu_cmd_render(cmd_buffer);
-      }
+   /* discard draw_cs entries now that the tiles are rendered */
+   tu_cs_discard_entries(&cmd_buffer->draw_cs);
 
-      tu_reset_render_pass(cmd_buffer);
-   }
+   vk_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
+   cmd_buffer->state.attachments = NULL;
 
-   if (cmd_buffer->state.resuming && !cmd_buffer->state.suspending) {
-      /* exiting suspend/resume chain */
-      switch (cmd_buffer->state.suspend_resume) {
-      case SR_IN_CHAIN:
-         cmd_buffer->state.suspend_resume = SR_NONE;
-         break;
-      case SR_IN_PRE_CHAIN:
-      case SR_IN_CHAIN_AFTER_PRE_CHAIN:
-         cmd_buffer->state.suspend_resume = SR_AFTER_PRE_CHAIN;
-         break;
-      default:
-         unreachable("suspending render pass not followed by resuming pass");
-      }
-   }
+   cmd_buffer->state.pass = NULL;
+   cmd_buffer->state.subpass = NULL;
+   cmd_buffer->state.framebuffer = NULL;
 }
 
-static void
-tu_barrier(struct tu_cmd_buffer *cmd,
-           const VkDependencyInfo *dep_info)
+void
+tu_CmdEndRenderPass2KHR(VkCommandBuffer commandBuffer,
+                        const VkSubpassEndInfoKHR *pSubpassEndInfo)
 {
-   VkPipelineStageFlags2 srcStage = 0;
-   VkPipelineStageFlags2 dstStage = 0;
-   enum tu_cmd_access_mask src_flags = 0;
-   enum tu_cmd_access_mask dst_flags = 0;
-
-   /* Inside a renderpass, we don't know yet whether we'll be using sysmem
-    * so we have to use the sysmem flushes.
-    */
-   bool gmem = cmd->state.ccu_state == TU_CMD_CCU_GMEM &&
-      !cmd->state.pass;
-
-
-   for (uint32_t i = 0; i < dep_info->memoryBarrierCount; i++) {
-      VkPipelineStageFlags2 sanitized_src_stage =
-         sanitize_src_stage(dep_info->pMemoryBarriers[i].srcStageMask);
-      VkPipelineStageFlags2 sanitized_dst_stage =
-         sanitize_dst_stage(dep_info->pMemoryBarriers[i].dstStageMask);
-      src_flags |= vk2tu_access(dep_info->pMemoryBarriers[i].srcAccessMask,
-                                sanitized_src_stage, false, gmem);
-      dst_flags |= vk2tu_access(dep_info->pMemoryBarriers[i].dstAccessMask,
-                                sanitized_dst_stage, false, gmem);
-      srcStage |= sanitized_src_stage;
-      dstStage |= sanitized_dst_stage;
-   }
-
-   for (uint32_t i = 0; i < dep_info->bufferMemoryBarrierCount; i++) {
-      VkPipelineStageFlags2 sanitized_src_stage =
-         sanitize_src_stage(dep_info->pBufferMemoryBarriers[i].srcStageMask);
-      VkPipelineStageFlags2 sanitized_dst_stage =
-         sanitize_dst_stage(dep_info->pBufferMemoryBarriers[i].dstStageMask);
-      src_flags |= vk2tu_access(dep_info->pBufferMemoryBarriers[i].srcAccessMask,
-                                sanitized_src_stage, false, gmem);
-      dst_flags |= vk2tu_access(dep_info->pBufferMemoryBarriers[i].dstAccessMask,
-                                sanitized_dst_stage, false, gmem);
-      srcStage |= sanitized_src_stage;
-      dstStage |= sanitized_dst_stage;
-   }
-
-   for (uint32_t i = 0; i < dep_info->imageMemoryBarrierCount; i++) {
-      VkImageLayout old_layout = dep_info->pImageMemoryBarriers[i].oldLayout;
-      if (old_layout == VK_IMAGE_LAYOUT_UNDEFINED) {
-         /* The underlying memory for this image may have been used earlier
-          * within the same queue submission for a different image, which
-          * means that there may be old, stale cache entries which are in the
-          * "wrong" location, which could cause problems later after writing
-          * to the image. We don't want these entries being flushed later and
-          * overwriting the actual image, so we need to flush the CCU.
-          */
-         TU_FROM_HANDLE(tu_image, image, dep_info->pImageMemoryBarriers[i].image);
-
-         if (vk_format_is_depth_or_stencil(image->vk.format)) {
-            src_flags |= TU_ACCESS_CCU_DEPTH_INCOHERENT_WRITE;
-         } else {
-            src_flags |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
-         }
-      }
-      VkPipelineStageFlags2 sanitized_src_stage =
-         sanitize_src_stage(dep_info->pImageMemoryBarriers[i].srcStageMask);
-      VkPipelineStageFlags2 sanitized_dst_stage =
-         sanitize_dst_stage(dep_info->pImageMemoryBarriers[i].dstStageMask);
-      src_flags |= vk2tu_access(dep_info->pImageMemoryBarriers[i].srcAccessMask,
-                                sanitized_src_stage, true, gmem);
-      dst_flags |= vk2tu_access(dep_info->pImageMemoryBarriers[i].dstAccessMask,
-                                sanitized_dst_stage, true, gmem);
-      srcStage |= sanitized_src_stage;
-      dstStage |= sanitized_dst_stage;
-   }
-
-   if (cmd->state.pass) {
-      const VkPipelineStageFlags framebuffer_space_stages =
-         VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-         VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
-         VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
-         VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
-
-      /* We cannot have non-by-region "fb-space to fb-space" barriers.
-       *
-       * From the Vulkan 1.2.185 spec, section 7.6.1 "Subpass Self-dependency":
-       *
-       *    If the source and destination stage masks both include
-       *    framebuffer-space stages, then dependencyFlags must include
-       *    VK_DEPENDENCY_BY_REGION_BIT.
-       *    [...]
-       *    Each of the synchronization scopes and access scopes of a
-       *    vkCmdPipelineBarrier2 or vkCmdPipelineBarrier command inside
-       *    a render pass instance must be a subset of the scopes of one of
-       *    the self-dependencies for the current subpass.
-       *
-       *    If the self-dependency has VK_DEPENDENCY_BY_REGION_BIT or
-       *    VK_DEPENDENCY_VIEW_LOCAL_BIT set, then so must the pipeline barrier.
-       *
-       * By-region barriers are ok for gmem. All other barriers would involve
-       * vtx stages which are NOT ok for gmem rendering.
-       * See dep_invalid_for_gmem().
-       */
-      if ((srcStage & ~framebuffer_space_stages) ||
-          (dstStage & ~framebuffer_space_stages)) {
-         cmd->state.rp.disable_gmem = true;
-      }
-   }
-
-   struct tu_cache_state *cache =
-      cmd->state.pass  ? &cmd->state.renderpass_cache : &cmd->state.cache;
-   tu_flush_for_access(cache, src_flags, dst_flags);
-
-   enum tu_stage src_stage = vk2tu_src_stage(srcStage);
-   enum tu_stage dst_stage = vk2tu_dst_stage(dstStage);
-   tu_flush_for_stage(cache, src_stage, dst_stage);
+   tu_CmdEndRenderPass(commandBuffer);
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdPipelineBarrier2(VkCommandBuffer commandBuffer,
-                       const VkDependencyInfo *pDependencyInfo)
+struct tu_barrier_info
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
-
-   tu_barrier(cmd_buffer, pDependencyInfo);
-}
+   uint32_t eventCount;
+   const VkEvent *pEvents;
+   VkPipelineStageFlags srcStageMask;
+};
 
 static void
-write_event(struct tu_cmd_buffer *cmd, struct tu_event *event,
-            VkPipelineStageFlags2 stageMask, unsigned value)
+tu_barrier(struct tu_cmd_buffer *cmd_buffer,
+           uint32_t memoryBarrierCount,
+           const VkMemoryBarrier *pMemoryBarriers,
+           uint32_t bufferMemoryBarrierCount,
+           const VkBufferMemoryBarrier *pBufferMemoryBarriers,
+           uint32_t imageMemoryBarrierCount,
+           const VkImageMemoryBarrier *pImageMemoryBarriers,
+           const struct tu_barrier_info *info)
 {
-   struct tu_cs *cs = &cmd->cs;
-
-   /* vkCmdSetEvent/vkCmdResetEvent cannot be called inside a render pass */
-   assert(!cmd->state.pass);
-
-   tu_emit_cache_flush(cmd, cs);
-
-   /* Flags that only require a top-of-pipe event. DrawIndirect parameters are
-    * read by the CP, so the draw indirect stage counts as top-of-pipe too.
-    */
-   VkPipelineStageFlags2 top_of_pipe_flags =
-      VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT |
-      VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT;
-
-   if (!(stageMask & ~top_of_pipe_flags)) {
-      tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
-      tu_cs_emit_qw(cs, event->bo->iova); /* ADDR_LO/HI */
-      tu_cs_emit(cs, value);
-   } else {
-      /* Use a RB_DONE_TS event to wait for everything to complete. */
-      tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 4);
-      tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(RB_DONE_TS));
-      tu_cs_emit_qw(cs, event->bo->iova);
-      tu_cs_emit(cs, value);
-   }
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetEvent2(VkCommandBuffer commandBuffer,
-                VkEvent _event,
-                const VkDependencyInfo *pDependencyInfo)
+void
+tu_CmdPipelineBarrier(VkCommandBuffer commandBuffer,
+                      VkPipelineStageFlags srcStageMask,
+                      VkPipelineStageFlags destStageMask,
+                      VkBool32 byRegion,
+                      uint32_t memoryBarrierCount,
+                      const VkMemoryBarrier *pMemoryBarriers,
+                      uint32_t bufferMemoryBarrierCount,
+                      const VkBufferMemoryBarrier *pBufferMemoryBarriers,
+                      uint32_t imageMemoryBarrierCount,
+                      const VkImageMemoryBarrier *pImageMemoryBarriers)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_event, event, _event);
-   VkPipelineStageFlags2 src_stage_mask = 0;
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   struct tu_barrier_info info;
 
-   for (uint32_t i = 0; i < pDependencyInfo->memoryBarrierCount; i++)
-      src_stage_mask |= pDependencyInfo->pMemoryBarriers[i].srcStageMask;
-   for (uint32_t i = 0; i < pDependencyInfo->bufferMemoryBarrierCount; i++)
-      src_stage_mask |= pDependencyInfo->pBufferMemoryBarriers[i].srcStageMask;
-   for (uint32_t i = 0; i < pDependencyInfo->imageMemoryBarrierCount; i++)
-      src_stage_mask |= pDependencyInfo->pImageMemoryBarriers[i].srcStageMask;
+   info.eventCount = 0;
+   info.pEvents = NULL;
+   info.srcStageMask = srcStageMask;
 
-   write_event(cmd, event, src_stage_mask, 1);
+   tu_barrier(cmd_buffer, memoryBarrierCount, pMemoryBarriers,
+              bufferMemoryBarrierCount, pBufferMemoryBarriers,
+              imageMemoryBarrierCount, pImageMemoryBarriers, &info);
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdResetEvent2(VkCommandBuffer commandBuffer,
-                  VkEvent _event,
-                  VkPipelineStageFlags2 stageMask)
+static void
+write_event(struct tu_cmd_buffer *cmd_buffer,
+            struct tu_event *event,
+            VkPipelineStageFlags stageMask,
+            unsigned value)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_event, event, _event);
-
-   write_event(cmd, event, stageMask, 0);
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdWaitEvents2(VkCommandBuffer commandBuffer,
-                  uint32_t eventCount,
-                  const VkEvent *pEvents,
-                  const VkDependencyInfo* pDependencyInfos)
+void
+tu_CmdSetEvent(VkCommandBuffer commandBuffer,
+               VkEvent _event,
+               VkPipelineStageFlags stageMask)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   struct tu_cs *cs = cmd->state.pass ? &cmd->draw_cs : &cmd->cs;
-
-   for (uint32_t i = 0; i < eventCount; i++) {
-      TU_FROM_HANDLE(tu_event, event, pEvents[i]);
-
-      tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
-      tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
-                     CP_WAIT_REG_MEM_0_POLL_MEMORY);
-      tu_cs_emit_qw(cs, event->bo->iova); /* POLL_ADDR_LO/HI */
-      tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(1));
-      tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0u));
-      tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(20));
-   }
-
-   tu_barrier(cmd, pDependencyInfos);
-}
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   TU_FROM_HANDLE(tu_event, event, _event);
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdSetDeviceMask(VkCommandBuffer commandBuffer, uint32_t deviceMask)
-{
-   /* No-op */
+   write_event(cmd_buffer, event, stageMask, 1);
 }
 
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdBeginConditionalRenderingEXT(VkCommandBuffer commandBuffer,
-                                   const VkConditionalRenderingBeginInfoEXT *pConditionalRenderingBegin)
+void
+tu_CmdResetEvent(VkCommandBuffer commandBuffer,
+                 VkEvent _event,
+                 VkPipelineStageFlags stageMask)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.predication_active = true;
-
-   struct tu_cs *cs = cmd->state.pass ? &cmd->draw_cs : &cmd->cs;
-
-   tu_cs_emit_pkt7(cs, CP_DRAW_PRED_ENABLE_GLOBAL, 1);
-   tu_cs_emit(cs, 1);
-
-   /* Wait for any writes to the predicate to land */
-   if (cmd->state.pass)
-      tu_emit_cache_flush_renderpass(cmd, cs);
-   else
-      tu_emit_cache_flush(cmd, cs);
-
-   TU_FROM_HANDLE(tu_buffer, buf, pConditionalRenderingBegin->buffer);
-   uint64_t iova = buf->iova + pConditionalRenderingBegin->offset;
-
-   /* qcom doesn't support 32-bit reference values, only 64-bit, but Vulkan
-    * mandates 32-bit comparisons. Our workaround is to copy the the reference
-    * value to the low 32-bits of a location where the high 32 bits are known
-    * to be 0 and then compare that.
-    */
-   tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 5);
-   tu_cs_emit(cs, 0);
-   tu_cs_emit_qw(cs, global_iova(cmd, predicate));
-   tu_cs_emit_qw(cs, iova);
-
-   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-   tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   TU_FROM_HANDLE(tu_event, event, _event);
 
-   bool inv = pConditionalRenderingBegin->flags & VK_CONDITIONAL_RENDERING_INVERTED_BIT_EXT;
-   tu_cs_emit_pkt7(cs, CP_DRAW_PRED_SET, 3);
-   tu_cs_emit(cs, CP_DRAW_PRED_SET_0_SRC(PRED_SRC_MEM) |
-                  CP_DRAW_PRED_SET_0_TEST(inv ? EQ_0_PASS : NE_0_PASS));
-   tu_cs_emit_qw(cs, global_iova(cmd, predicate));
+   write_event(cmd_buffer, event, stageMask, 0);
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdEndConditionalRenderingEXT(VkCommandBuffer commandBuffer)
+void
+tu_CmdWaitEvents(VkCommandBuffer commandBuffer,
+                 uint32_t eventCount,
+                 const VkEvent *pEvents,
+                 VkPipelineStageFlags srcStageMask,
+                 VkPipelineStageFlags dstStageMask,
+                 uint32_t memoryBarrierCount,
+                 const VkMemoryBarrier *pMemoryBarriers,
+                 uint32_t bufferMemoryBarrierCount,
+                 const VkBufferMemoryBarrier *pBufferMemoryBarriers,
+                 uint32_t imageMemoryBarrierCount,
+                 const VkImageMemoryBarrier *pImageMemoryBarriers)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-
-   cmd->state.predication_active = false;
+   TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+   struct tu_barrier_info info;
 
-   struct tu_cs *cs = cmd->state.pass ? &cmd->draw_cs : &cmd->cs;
+   info.eventCount = eventCount;
+   info.pEvents = pEvents;
+   info.srcStageMask = 0;
 
-   tu_cs_emit_pkt7(cs, CP_DRAW_PRED_ENABLE_GLOBAL, 1);
-   tu_cs_emit(cs, 0);
+   tu_barrier(cmd_buffer, memoryBarrierCount, pMemoryBarriers,
+              bufferMemoryBarrierCount, pBufferMemoryBarriers,
+              imageMemoryBarrierCount, pImageMemoryBarriers, &info);
 }
 
 void
-tu_CmdWriteBufferMarker2AMD(VkCommandBuffer commandBuffer,
-                            VkPipelineStageFlagBits2 pipelineStage,
-                            VkBuffer dstBuffer,
-                            VkDeviceSize dstOffset,
-                            uint32_t marker)
+tu_CmdSetDeviceMask(VkCommandBuffer commandBuffer, uint32_t deviceMask)
 {
-   /* Almost the same as write_event, but also allowed in renderpass */
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_buffer, buffer, dstBuffer);
-
-   uint64_t va = buffer->bo->iova + dstOffset;
-
-   struct tu_cs *cs = cmd->state.pass ? &cmd->draw_cs : &cmd->cs;
-   struct tu_cache_state *cache =
-      cmd->state.pass ? &cmd->state.renderpass_cache : &cmd->state.cache;
-
-   /* From the Vulkan 1.2.203 spec:
-    *
-    *    The access scope for buffer marker writes falls under
-    *    the VK_ACCESS_TRANSFER_WRITE_BIT, and the pipeline stages for
-    *    identifying the synchronization scope must include both pipelineStage
-    *    and VK_PIPELINE_STAGE_TRANSFER_BIT.
-    *
-    * Transfer operations use CCU however here we write via CP.
-    * Flush CCU in order to make the results of previous transfer
-    * operation visible to CP.
-    */
-   tu_flush_for_access(cache, 0, TU_ACCESS_SYSMEM_WRITE);
-
-   /* Flags that only require a top-of-pipe event. DrawIndirect parameters are
-    * read by the CP, so the draw indirect stage counts as top-of-pipe too.
-    */
-   VkPipelineStageFlags2 top_of_pipe_flags =
-      VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT |
-      VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT;
-
-   bool is_top_of_pipe = !(pipelineStage & ~top_of_pipe_flags);
-
-   /* We have to WFI only if we flushed CCU here and are using CP_MEM_WRITE.
-    * Otherwise:
-    * - We do CP_EVENT_WRITE(RB_DONE_TS) which should wait for flushes;
-    * - There was a barrier to synchronize other writes with WriteBufferMarkerAMD
-    *   and they had to include our pipelineStage which forces the WFI.
-    */
-   if (cache->flush_bits != 0 && is_top_of_pipe) {
-      cache->flush_bits |= TU_CMD_FLAG_WAIT_FOR_IDLE;
-   }
-
-   if (cmd->state.pass) {
-      tu_emit_cache_flush_renderpass(cmd, cs);
-   } else {
-      tu_emit_cache_flush(cmd, cs);
-   }
-
-   if (is_top_of_pipe) {
-      tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
-      tu_cs_emit_qw(cs, va); /* ADDR_LO/HI */
-      tu_cs_emit(cs, marker);
-   } else {
-      /* Use a RB_DONE_TS event to wait for everything to complete. */
-      tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 4);
-      tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(RB_DONE_TS));
-      tu_cs_emit_qw(cs, va);
-      tu_cs_emit(cs, marker);
-   }
-
-   /* Make sure the result of this write is visible to others. */
-   tu_flush_for_access(cache, TU_ACCESS_CP_WRITE, 0);
+   /* No-op */
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_cs.c b/lib/mesa/src/freedreno/vulkan/tu_cs.c
index 2e6f215f4..48242f813 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_cs.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_cs.c
@@ -1,101 +1,108 @@
 /*
  * Copyright © 2019 Google LLC
- * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
 #include "tu_cs.h"
 
-#include "tu_suballoc.h"
-
 /**
  * Initialize a command stream.
  */
 void
-tu_cs_init(struct tu_cs *cs,
-           struct tu_device *device,
-           enum tu_cs_mode mode,
-           uint32_t initial_size, const char *name)
+tu_cs_init(struct tu_cs *cs, enum tu_cs_mode mode, uint32_t initial_size)
 {
    assert(mode != TU_CS_MODE_EXTERNAL);
 
    memset(cs, 0, sizeof(*cs));
 
-   cs->device = device;
    cs->mode = mode;
    cs->next_bo_size = initial_size;
-   cs->name = name;
 }
 
 /**
  * Initialize a command stream as a wrapper to an external buffer.
  */
 void
-tu_cs_init_external(struct tu_cs *cs, struct tu_device *device,
-                    uint32_t *start, uint32_t *end)
+tu_cs_init_external(struct tu_cs *cs, uint32_t *start, uint32_t *end)
 {
    memset(cs, 0, sizeof(*cs));
 
-   cs->device = device;
    cs->mode = TU_CS_MODE_EXTERNAL;
    cs->start = cs->reserved_end = cs->cur = start;
    cs->end = end;
 }
 
 /**
- * Initialize a sub-command stream as a wrapper to an externally sub-allocated
- * buffer.
+ * Finish and release all resources owned by a command stream.
  */
 void
-tu_cs_init_suballoc(struct tu_cs *cs, struct tu_device *device,
-                    struct tu_suballoc_bo *suballoc_bo)
+tu_cs_finish(struct tu_device *dev, struct tu_cs *cs)
 {
-   uint32_t *start = tu_suballoc_bo_map(suballoc_bo);
-   uint32_t *end = start + (suballoc_bo->size >> 2);
+   for (uint32_t i = 0; i < cs->bo_count; ++i) {
+      tu_bo_finish(dev, cs->bos[i]);
+      free(cs->bos[i]);
+   }
 
-   memset(cs, 0, sizeof(*cs));
-   cs->device = device;
-   cs->mode = TU_CS_MODE_SUB_STREAM;
-   cs->start = cs->reserved_end = cs->cur = start;
-   cs->end = end;
-   cs->refcount_bo = tu_bo_get_ref(suballoc_bo->bo);
+   free(cs->entries);
+   free(cs->bos);
 }
 
 /**
- * Finish and release all resources owned by a command stream.
+ * Get the offset of the command packets emitted since the last call to
+ * tu_cs_add_entry.
  */
-void
-tu_cs_finish(struct tu_cs *cs)
+static uint32_t
+tu_cs_get_offset(const struct tu_cs *cs)
 {
-   for (uint32_t i = 0; i < cs->bo_count; ++i) {
-      tu_bo_finish(cs->device, cs->bos[i]);
-   }
-
-   if (cs->refcount_bo)
-      tu_bo_finish(cs->device, cs->refcount_bo);
+   assert(cs->bo_count);
+   return cs->start - (uint32_t *) cs->bos[cs->bo_count - 1]->map;
+}
 
-   free(cs->entries);
-   free(cs->bos);
+/**
+ * Get the size of the command packets emitted since the last call to
+ * tu_cs_add_entry.
+ */
+static uint32_t
+tu_cs_get_size(const struct tu_cs *cs)
+{
+   return cs->cur - cs->start;
 }
 
-static struct tu_bo *
-tu_cs_current_bo(const struct tu_cs *cs)
+/**
+ * Get the size of the remaining space in the current BO.
+ */
+static uint32_t
+tu_cs_get_space(const struct tu_cs *cs)
 {
-   if (cs->refcount_bo) {
-      return cs->refcount_bo;
-   } else {
-      assert(cs->bo_count);
-      return cs->bos[cs->bo_count - 1];
-   }
+   return cs->end - cs->cur;
 }
 
 /**
- * Get the offset of the command packets emitted since the last call to
+ * Return true if there is no command packet emitted since the last call to
  * tu_cs_add_entry.
  */
 static uint32_t
-tu_cs_get_offset(const struct tu_cs *cs)
+tu_cs_is_empty(const struct tu_cs *cs)
 {
-   return cs->start - (uint32_t *) tu_cs_current_bo(cs)->map;
+   return tu_cs_get_size(cs) == 0;
 }
 
 /*
@@ -103,12 +110,10 @@ tu_cs_get_offset(const struct tu_cs *cs)
  * be emitted to the new BO.
  */
 static VkResult
-tu_cs_add_bo(struct tu_cs *cs, uint32_t size)
+tu_cs_add_bo(struct tu_device *dev, struct tu_cs *cs, uint32_t size)
 {
    /* no BO for TU_CS_MODE_EXTERNAL */
    assert(cs->mode != TU_CS_MODE_EXTERNAL);
-   /* No adding more BOs if suballocating from a suballoc_bo. */
-   assert(!cs->refcount_bo);
 
    /* no dangling command packet */
    assert(tu_cs_is_empty(cs));
@@ -125,18 +130,20 @@ tu_cs_add_bo(struct tu_cs *cs, uint32_t size)
       cs->bos = new_bos;
    }
 
-   struct tu_bo *new_bo;
+   struct tu_bo *new_bo = malloc(sizeof(struct tu_bo));
+   if (!new_bo)
+      return VK_ERROR_OUT_OF_HOST_MEMORY;
 
-   VkResult result =
-      tu_bo_init_new(cs->device, &new_bo, size * sizeof(uint32_t),
-                     TU_BO_ALLOC_GPU_READ_ONLY | TU_BO_ALLOC_ALLOW_DUMP, cs->name);
+   VkResult result = tu_bo_init_new(dev, new_bo, size * sizeof(uint32_t));
    if (result != VK_SUCCESS) {
+      free(new_bo);
       return result;
    }
 
-   result = tu_bo_map(cs->device, new_bo);
+   result = tu_bo_map(dev, new_bo);
    if (result != VK_SUCCESS) {
-      tu_bo_finish(cs->device, new_bo);
+      tu_bo_finish(dev, new_bo);
+      free(new_bo);
       return result;
    }
 
@@ -152,7 +159,7 @@ tu_cs_add_bo(struct tu_cs *cs, uint32_t size)
  * Reserve an IB entry.
  */
 static VkResult
-tu_cs_reserve_entry(struct tu_cs *cs)
+tu_cs_reserve_entry(struct tu_device *dev, struct tu_cs *cs)
 {
    /* entries are only for TU_CS_MODE_GROW */
    assert(cs->mode == TU_CS_MODE_GROW);
@@ -194,7 +201,7 @@ tu_cs_add_entry(struct tu_cs *cs)
 
    /* add an entry for [cs->start, cs->cur] */
    cs->entries[cs->entry_count++] = (struct tu_cs_entry) {
-      .bo = tu_cs_current_bo(cs),
+      .bo = cs->bos[cs->bo_count - 1],
       .size = tu_cs_get_size(cs) * sizeof(uint32_t),
       .offset = tu_cs_get_offset(cs) * sizeof(uint32_t),
    };
@@ -203,30 +210,6 @@ tu_cs_add_entry(struct tu_cs *cs)
 }
 
 /**
- * same behavior as tu_cs_emit_call but without the indirect
- */
-VkResult
-tu_cs_add_entries(struct tu_cs *cs, struct tu_cs *target)
-{
-   VkResult result;
-
-   assert(cs->mode == TU_CS_MODE_GROW);
-   assert(target->mode == TU_CS_MODE_GROW);
-
-   if (!tu_cs_is_empty(cs))
-      tu_cs_add_entry(cs);
-
-   for (unsigned i = 0; i < target->entry_count; i++) {
-      result = tu_cs_reserve_entry(cs);
-      if (result != VK_SUCCESS)
-         return result;
-      cs->entries[cs->entry_count++] = target->entries[i];
-   }
-
-   return VK_SUCCESS;
-}
-
-/**
  * Begin (or continue) command packet emission.  This does nothing but sanity
  * checks currently.  \a cs must not be in TU_CS_MODE_SUB_STREAM mode.
  */
@@ -259,58 +242,27 @@ tu_cs_end(struct tu_cs *cs)
  * emission.
  */
 VkResult
-tu_cs_begin_sub_stream(struct tu_cs *cs, uint32_t size, struct tu_cs *sub_cs)
+tu_cs_begin_sub_stream(struct tu_device *dev,
+                       struct tu_cs *cs,
+                       uint32_t size,
+                       struct tu_cs *sub_cs)
 {
    assert(cs->mode == TU_CS_MODE_SUB_STREAM);
    assert(size);
 
-   VkResult result = tu_cs_reserve_space(cs, size);
+   VkResult result = tu_cs_reserve_space(dev, cs, size);
    if (result != VK_SUCCESS)
       return result;
 
-   tu_cs_init_external(sub_cs, cs->device, cs->cur, cs->reserved_end);
+   tu_cs_init_external(sub_cs, cs->cur, cs->reserved_end);
    tu_cs_begin(sub_cs);
-   result = tu_cs_reserve_space(sub_cs, size);
+   result = tu_cs_reserve_space(dev, sub_cs, size);
    assert(result == VK_SUCCESS);
 
    return VK_SUCCESS;
 }
 
 /**
- * Allocate count*size dwords, aligned to size dwords.
- * \a cs must be in TU_CS_MODE_SUB_STREAM mode.
- *
- */
-VkResult
-tu_cs_alloc(struct tu_cs *cs,
-            uint32_t count,
-            uint32_t size,
-            struct tu_cs_memory *memory)
-{
-   assert(cs->mode == TU_CS_MODE_SUB_STREAM);
-   assert(size && size <= 1024);
-
-   if (!count)
-      return VK_SUCCESS;
-
-   /* TODO: smarter way to deal with alignment? */
-
-   VkResult result = tu_cs_reserve_space(cs, count * size + (size-1));
-   if (result != VK_SUCCESS)
-      return result;
-
-   struct tu_bo *bo = tu_cs_current_bo(cs);
-   size_t offset = align(tu_cs_get_offset(cs), size);
-
-   memory->map = bo->map + offset * sizeof(uint32_t);
-   memory->iova = bo->iova + offset * sizeof(uint32_t);
-
-   cs->start = cs->cur = (uint32_t*) bo->map + offset + count * size;
-
-   return VK_SUCCESS;
-}
-
-/**
  * End command packet emission to a sub-stream.  \a sub_cs becomes invalid
  * after this call.
  *
@@ -321,6 +273,7 @@ struct tu_cs_entry
 tu_cs_end_sub_stream(struct tu_cs *cs, struct tu_cs *sub_cs)
 {
    assert(cs->mode == TU_CS_MODE_SUB_STREAM);
+   assert(cs->bo_count);
    assert(sub_cs->start == cs->cur && sub_cs->end == cs->reserved_end);
    tu_cs_sanity_check(sub_cs);
 
@@ -329,7 +282,7 @@ tu_cs_end_sub_stream(struct tu_cs *cs, struct tu_cs *sub_cs)
    cs->cur = sub_cs->cur;
 
    struct tu_cs_entry entry = {
-      .bo = tu_cs_current_bo(cs),
+      .bo = cs->bos[cs->bo_count - 1],
       .size = tu_cs_get_size(cs) * sizeof(uint32_t),
       .offset = tu_cs_get_offset(cs) * sizeof(uint32_t),
    };
@@ -344,7 +297,9 @@ tu_cs_end_sub_stream(struct tu_cs *cs, struct tu_cs *sub_cs)
  * This never fails when \a cs has mode TU_CS_MODE_EXTERNAL.
  */
 VkResult
-tu_cs_reserve_space(struct tu_cs *cs, uint32_t reserved_size)
+tu_cs_reserve_space(struct tu_device *dev,
+                    struct tu_cs *cs,
+                    uint32_t reserved_size)
 {
    if (tu_cs_get_space(cs) < reserved_size) {
       if (cs->mode == TU_CS_MODE_EXTERNAL) {
@@ -360,39 +315,14 @@ tu_cs_reserve_space(struct tu_cs *cs, uint32_t reserved_size)
          tu_cs_add_entry(cs);
       }
 
-      for (uint32_t i = 0; i < cs->cond_stack_depth; i++) {
-         /* Subtract one here to account for the DWORD field itself. */
-         *cs->cond_dwords[i] = cs->cur - cs->cond_dwords[i] - 1;
-
-         /* space for CP_COND_REG_EXEC in next bo */
-         reserved_size += 3;
-      }
-
       /* switch to a new BO */
       uint32_t new_size = MAX2(cs->next_bo_size, reserved_size);
-      VkResult result = tu_cs_add_bo(cs, new_size);
+      VkResult result = tu_cs_add_bo(dev, cs, new_size);
       if (result != VK_SUCCESS)
          return result;
 
-      if (cs->cond_stack_depth) {
-         cs->reserved_end = cs->cur + reserved_size;
-      }
-
-      /* Re-emit CP_COND_REG_EXECs */
-      for (uint32_t i = 0; i < cs->cond_stack_depth; i++) {
-         tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
-         tu_cs_emit(cs, cs->cond_flags[i]);
-
-         cs->cond_dwords[i] = cs->cur;
-
-         /* Emit dummy DWORD field here */
-         tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(0));
-      }
-
-      /* double the size for the next bo, also there is an upper
-       * bound on IB size, which appears to be 0x0fffff
-       */
-      new_size = MIN2(new_size << 1, 0x0fffff);
+      /* double the size for the next bo */
+      new_size <<= 1;
       if (cs->next_bo_size < new_size)
          cs->next_bo_size = new_size;
    }
@@ -402,7 +332,7 @@ tu_cs_reserve_space(struct tu_cs *cs, uint32_t reserved_size)
 
    if (cs->mode == TU_CS_MODE_GROW) {
       /* reserve an entry for the next call to this function or tu_cs_end */
-      return tu_cs_reserve_entry(cs);
+      return tu_cs_reserve_entry(dev, cs);
    }
 
    return VK_SUCCESS;
@@ -413,16 +343,17 @@ tu_cs_reserve_space(struct tu_cs *cs, uint32_t reserved_size)
  * packets in \a cs, but does not necessarily release all resources.
  */
 void
-tu_cs_reset(struct tu_cs *cs)
+tu_cs_reset(struct tu_device *dev, struct tu_cs *cs)
 {
    if (cs->mode == TU_CS_MODE_EXTERNAL) {
-      assert(!cs->bo_count && !cs->refcount_bo && !cs->entry_count);
+      assert(!cs->bo_count && !cs->entry_count);
       cs->reserved_end = cs->cur = cs->start;
       return;
    }
 
    for (uint32_t i = 0; i + 1 < cs->bo_count; ++i) {
-      tu_bo_finish(cs->device, cs->bos[i]);
+      tu_bo_finish(dev, cs->bos[i]);
+      free(cs->bos[i]);
    }
 
    if (cs->bo_count) {
diff --git a/lib/mesa/src/freedreno/vulkan/tu_descriptor_set.c b/lib/mesa/src/freedreno/vulkan/tu_descriptor_set.c
index 14d8b4b07..0f49d26e2 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_descriptor_set.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_descriptor_set.c
@@ -1,100 +1,67 @@
 /*
  * Copyright © 2016 Red Hat.
  * Copyright © 2016 Bas Nieuwenhuizen
- * SPDX-License-Identifier: MIT
- */
-
-/**
- * @file
  *
- * We use the bindless descriptor model, which maps fairly closely to how
- * Vulkan descriptor sets work. The two exceptions are input attachments and
- * dynamic descriptors, which have to be patched when recording command
- * buffers. We reserve an extra descriptor set for these. This descriptor set
- * contains all the input attachments in the pipeline, in order, and then all
- * the dynamic descriptors. The dynamic descriptors are stored in the CPU-side
- * datastructure for each tu_descriptor_set, and then combined into one big
- * descriptor set at CmdBindDescriptors time/draw time.
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
+#include "tu_private.h"
 
-#include "tu_descriptor_set.h"
-
+#include <assert.h>
 #include <fcntl.h>
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
 
 #include "util/mesa-sha1.h"
-#include "vk_descriptors.h"
 #include "vk_util.h"
 
-#include "tu_device.h"
-#include "tu_image.h"
-
-static inline uint8_t *
-pool_base(struct tu_descriptor_pool *pool)
+static int
+binding_compare(const void *av, const void *bv)
 {
-   return pool->host_bo ?: pool->bo->map;
-}
+   const VkDescriptorSetLayoutBinding *a =
+      (const VkDescriptorSetLayoutBinding *) av;
+   const VkDescriptorSetLayoutBinding *b =
+      (const VkDescriptorSetLayoutBinding *) bv;
 
-static uint32_t
-descriptor_size(struct tu_device *dev,
-                const VkDescriptorSetLayoutBinding *binding,
-                VkDescriptorType type)
-{
-   switch (type) {
-   case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
-      if (unlikely(dev->instance->debug_flags & TU_DEBUG_DYNAMIC))
-         return A6XX_TEX_CONST_DWORDS * 4;
-
-      /* Input attachment doesn't use descriptor sets at all */
-      return 0;
-   case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
-      /* We make offsets and sizes all 16 dwords, to match how the hardware
-       * interprets indices passed to sample/load/store instructions in
-       * multiples of 16 dwords.  This means that "normal" descriptors are all
-       * of size 16, with padding for smaller descriptors like uniform storage
-       * descriptors which are less than 16 dwords. However combined images
-       * and samplers are actually two descriptors, so they have size 2.
-       */
-      return A6XX_TEX_CONST_DWORDS * 4 * 2;
-   case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
-   case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
-      /* When we support 16-bit storage, we need an extra descriptor setup as
-       * a 32-bit array for isam to work.
-       */
-      if (dev->physical_device->info->a6xx.storage_16bit) {
-         return A6XX_TEX_CONST_DWORDS * 4 * 2;
-      } else {
-         return A6XX_TEX_CONST_DWORDS * 4;
-      }
-   case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK:
-      return A6XX_TEX_CONST_DWORDS * 4 +
-         ALIGN(binding->descriptorCount, A6XX_TEX_CONST_DWORDS * 4);
-   default:
-      return A6XX_TEX_CONST_DWORDS * 4;
-   }
+   return (a->binding < b->binding) ? -1 : (a->binding > b->binding) ? 1 : 0;
 }
 
-static bool
-is_dynamic(VkDescriptorType type)
+static VkDescriptorSetLayoutBinding *
+create_sorted_bindings(const VkDescriptorSetLayoutBinding *bindings,
+                       unsigned count)
 {
-   return type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC ||
-          type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC;
-}
+   VkDescriptorSetLayoutBinding *sorted_bindings =
+      malloc(count * sizeof(VkDescriptorSetLayoutBinding));
+   if (!sorted_bindings)
+      return NULL;
 
-static uint32_t
-mutable_descriptor_size(struct tu_device *dev,
-                        const VkMutableDescriptorTypeListEXT *list)
-{
-   uint32_t max_size = 0;
+   memcpy(sorted_bindings, bindings,
+          count * sizeof(VkDescriptorSetLayoutBinding));
 
-   for (uint32_t i = 0; i < list->descriptorTypeCount; i++) {
-      uint32_t size = descriptor_size(dev, NULL, list->pDescriptorTypes[i]);
-      max_size = MAX2(max_size, size);
-   }
+   qsort(sorted_bindings, count, sizeof(VkDescriptorSetLayoutBinding),
+         binding_compare);
 
-   return max_size;
+   return sorted_bindings;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateDescriptorSetLayout(
    VkDevice _device,
    const VkDescriptorSetLayoutCreateInfo *pCreateInfo,
@@ -106,191 +73,178 @@ tu_CreateDescriptorSetLayout(
 
    assert(pCreateInfo->sType ==
           VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO);
-   const VkDescriptorSetLayoutBindingFlagsCreateInfo *variable_flags =
-      vk_find_struct_const(
-         pCreateInfo->pNext,
-         DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO);
-   const VkMutableDescriptorTypeCreateInfoEXT *mutable_info =
+   const VkDescriptorSetLayoutBindingFlagsCreateInfoEXT *variable_flags =
       vk_find_struct_const(
          pCreateInfo->pNext,
-         MUTABLE_DESCRIPTOR_TYPE_CREATE_INFO_EXT);
+         DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT);
 
-   uint32_t num_bindings = 0;
+   uint32_t max_binding = 0;
    uint32_t immutable_sampler_count = 0;
-   uint32_t ycbcr_sampler_count = 0;
    for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
-      num_bindings = MAX2(num_bindings, pCreateInfo->pBindings[j].binding + 1);
-      if ((pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
-           pCreateInfo->pBindings[j].descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) &&
-           pCreateInfo->pBindings[j].pImmutableSamplers) {
+      max_binding = MAX2(max_binding, pCreateInfo->pBindings[j].binding);
+      if (pCreateInfo->pBindings[j].pImmutableSamplers)
          immutable_sampler_count += pCreateInfo->pBindings[j].descriptorCount;
-
-         bool has_ycbcr_sampler = false;
-         for (unsigned i = 0; i < pCreateInfo->pBindings[j].descriptorCount; ++i) {
-            if (tu_sampler_from_handle(pCreateInfo->pBindings[j].pImmutableSamplers[i])->ycbcr_sampler)
-               has_ycbcr_sampler = true;
-         }
-
-         if (has_ycbcr_sampler)
-            ycbcr_sampler_count += pCreateInfo->pBindings[j].descriptorCount;
-      }
    }
 
    uint32_t samplers_offset =
-         offsetof(struct tu_descriptor_set_layout, binding[num_bindings]);
-
-   /* note: only need to store TEX_SAMP_DWORDS for immutable samples,
-    * but using struct tu_sampler makes things simpler */
-   uint32_t size = samplers_offset +
-      immutable_sampler_count * sizeof(struct tu_sampler) +
-      ycbcr_sampler_count * sizeof(struct tu_sampler_ycbcr_conversion);
+      sizeof(struct tu_descriptor_set_layout) +
+      (max_binding + 1) * sizeof(set_layout->binding[0]);
+   size_t size =
+      samplers_offset + immutable_sampler_count * 4 * sizeof(uint32_t);
 
-   set_layout = vk_descriptor_set_layout_zalloc(&device->vk, size);
+   set_layout = vk_alloc2(&device->alloc, pAllocator, size, 8,
+                          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (!set_layout)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    set_layout->flags = pCreateInfo->flags;
 
-   /* We just allocate all the immutable samplers at the end of the struct */
-   struct tu_sampler *samplers = (void*) &set_layout->binding[num_bindings];
-   struct tu_sampler_ycbcr_conversion *ycbcr_samplers =
-      (void*) &samplers[immutable_sampler_count];
-
-   VkDescriptorSetLayoutBinding *bindings = NULL;
-   VkResult result = vk_create_sorted_bindings(
-      pCreateInfo->pBindings, pCreateInfo->bindingCount, &bindings);
-   if (result != VK_SUCCESS) {
-      vk_object_free(&device->vk, pAllocator, set_layout);
-      return vk_error(device, result);
+   /* We just allocate all the samplers at the end of the struct */
+   uint32_t *samplers = (uint32_t *) &set_layout->binding[max_binding + 1];
+   (void) samplers; /* TODO: Use me */
+
+   VkDescriptorSetLayoutBinding *bindings = create_sorted_bindings(
+      pCreateInfo->pBindings, pCreateInfo->bindingCount);
+   if (!bindings) {
+      vk_free2(&device->alloc, pAllocator, set_layout);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
    }
 
-   set_layout->binding_count = num_bindings;
+   set_layout->binding_count = max_binding + 1;
    set_layout->shader_stages = 0;
+   set_layout->dynamic_shader_stages = 0;
    set_layout->has_immutable_samplers = false;
-   set_layout->has_inline_uniforms = false;
    set_layout->size = 0;
 
-   uint32_t dynamic_offset_size = 0;
+   memset(set_layout->binding, 0,
+          size - sizeof(struct tu_descriptor_set_layout));
+
+   uint32_t buffer_count = 0;
+   uint32_t dynamic_offset_count = 0;
 
    for (uint32_t j = 0; j < pCreateInfo->bindingCount; j++) {
       const VkDescriptorSetLayoutBinding *binding = bindings + j;
       uint32_t b = binding->binding;
+      uint32_t alignment;
+      unsigned binding_buffer_count = 0;
 
-      set_layout->binding[b].type = binding->descriptorType;
-      set_layout->binding[b].array_size =
-         binding->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK ?
-         1 : binding->descriptorCount;
-      set_layout->binding[b].offset = set_layout->size;
-      set_layout->binding[b].dynamic_offset_offset = dynamic_offset_size;
-      set_layout->binding[b].shader_stages = binding->stageFlags;
-
-      if (binding->descriptorType == VK_DESCRIPTOR_TYPE_MUTABLE_EXT) {
-         /* For mutable descriptor types we must allocate a size that fits the
-          * largest descriptor type that the binding can mutate to.
-          */
-         set_layout->binding[b].size =
-            mutable_descriptor_size(device, &mutable_info->pMutableDescriptorTypeLists[j]);
-      } else {
-         set_layout->binding[b].size =
-            descriptor_size(device, binding, binding->descriptorType);
+      switch (binding->descriptorType) {
+      case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
+      case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
+         assert(!(pCreateInfo->flags &
+                  VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
+         set_layout->binding[b].dynamic_offset_count = 1;
+         set_layout->dynamic_shader_stages |= binding->stageFlags;
+         set_layout->binding[b].size = 0;
+         binding_buffer_count = 1;
+         alignment = 1;
+         break;
+      case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+      case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+      case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+      case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+         set_layout->binding[b].size = 16;
+         binding_buffer_count = 1;
+         alignment = 16;
+         break;
+      case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
+      case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
+      case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
+         /* main descriptor + fmask descriptor */
+         set_layout->binding[b].size = 64;
+         binding_buffer_count = 1;
+         alignment = 32;
+         break;
+      case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+         /* main descriptor + fmask descriptor + sampler */
+         set_layout->binding[b].size = 96;
+         binding_buffer_count = 1;
+         alignment = 32;
+         break;
+      case VK_DESCRIPTOR_TYPE_SAMPLER:
+         set_layout->binding[b].size = 16;
+         alignment = 16;
+         break;
+      default:
+         unreachable("unknown descriptor type\n");
+         break;
       }
 
-      if (binding->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK)
-         set_layout->has_inline_uniforms = true;
+      set_layout->size = align(set_layout->size, alignment);
+      set_layout->binding[b].type = binding->descriptorType;
+      set_layout->binding[b].array_size = binding->descriptorCount;
+      set_layout->binding[b].offset = set_layout->size;
+      set_layout->binding[b].buffer_offset = buffer_count;
+      set_layout->binding[b].dynamic_offset_offset = dynamic_offset_count;
 
       if (variable_flags && binding->binding < variable_flags->bindingCount &&
           (variable_flags->pBindingFlags[binding->binding] &
-           VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT)) {
+           VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
          assert(!binding->pImmutableSamplers); /* Terribly ill defined  how
                                                   many samplers are valid */
-         assert(binding->binding == num_bindings - 1);
+         assert(binding->binding == max_binding);
 
          set_layout->has_variable_descriptors = true;
       }
 
-      if ((binding->descriptorType == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
-           binding->descriptorType == VK_DESCRIPTOR_TYPE_SAMPLER) &&
-          binding->pImmutableSamplers) {
+      if (binding->pImmutableSamplers) {
          set_layout->binding[b].immutable_samplers_offset = samplers_offset;
          set_layout->has_immutable_samplers = true;
-
-         for (uint32_t i = 0; i < binding->descriptorCount; i++)
-            samplers[i] = *tu_sampler_from_handle(binding->pImmutableSamplers[i]);
-
-         samplers += binding->descriptorCount;
-         samplers_offset += sizeof(struct tu_sampler) * binding->descriptorCount;
-
-         bool has_ycbcr_sampler = false;
-         for (unsigned i = 0; i < pCreateInfo->pBindings[j].descriptorCount; ++i) {
-            if (tu_sampler_from_handle(binding->pImmutableSamplers[i])->ycbcr_sampler)
-               has_ycbcr_sampler = true;
-         }
-
-         if (has_ycbcr_sampler) {
-            set_layout->binding[b].ycbcr_samplers_offset =
-               (const char*)ycbcr_samplers - (const char*)set_layout;
-            for (uint32_t i = 0; i < binding->descriptorCount; i++) {
-               struct tu_sampler *sampler = tu_sampler_from_handle(binding->pImmutableSamplers[i]);
-               if (sampler->ycbcr_sampler)
-                  ycbcr_samplers[i] = *sampler->ycbcr_sampler;
-               else
-                  ycbcr_samplers[i].ycbcr_model = VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY;
-            }
-            ycbcr_samplers += binding->descriptorCount;
-         } else {
-            set_layout->binding[b].ycbcr_samplers_offset = 0;
-         }
-      }
-
-      uint32_t size =
-         ALIGN_POT(set_layout->binding[b].array_size * set_layout->binding[b].size, 4 * A6XX_TEX_CONST_DWORDS);
-      if (is_dynamic(binding->descriptorType)) {
-         dynamic_offset_size += size;
-      } else {
-         set_layout->size += size;
       }
 
+      set_layout->size +=
+         binding->descriptorCount * set_layout->binding[b].size;
+      buffer_count += binding->descriptorCount * binding_buffer_count;
+      dynamic_offset_count += binding->descriptorCount *
+                              set_layout->binding[b].dynamic_offset_count;
       set_layout->shader_stages |= binding->stageFlags;
    }
 
    free(bindings);
 
-   set_layout->dynamic_offset_size = dynamic_offset_size;
+   set_layout->buffer_count = buffer_count;
+   set_layout->dynamic_offset_count = dynamic_offset_count;
 
    *pSetLayout = tu_descriptor_set_layout_to_handle(set_layout);
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
+tu_DestroyDescriptorSetLayout(VkDevice _device,
+                              VkDescriptorSetLayout _set_layout,
+                              const VkAllocationCallbacks *pAllocator)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout, _set_layout);
+
+   if (!set_layout)
+      return;
+
+   vk_free2(&device->alloc, pAllocator, set_layout);
+}
+
+void
 tu_GetDescriptorSetLayoutSupport(
-   VkDevice _device,
+   VkDevice device,
    const VkDescriptorSetLayoutCreateInfo *pCreateInfo,
    VkDescriptorSetLayoutSupport *pSupport)
 {
-   TU_FROM_HANDLE(tu_device, device, _device);
-
-   VkDescriptorSetLayoutBinding *bindings = NULL;
-   VkResult result = vk_create_sorted_bindings(
-      pCreateInfo->pBindings, pCreateInfo->bindingCount, &bindings);
-   if (result != VK_SUCCESS) {
+   VkDescriptorSetLayoutBinding *bindings = create_sorted_bindings(
+      pCreateInfo->pBindings, pCreateInfo->bindingCount);
+   if (!bindings) {
       pSupport->supported = false;
       return;
    }
 
-   const VkDescriptorSetLayoutBindingFlagsCreateInfo *variable_flags =
+   const VkDescriptorSetLayoutBindingFlagsCreateInfoEXT *variable_flags =
       vk_find_struct_const(
          pCreateInfo->pNext,
-         DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO);
-   VkDescriptorSetVariableDescriptorCountLayoutSupport *variable_count =
+         DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO_EXT);
+   VkDescriptorSetVariableDescriptorCountLayoutSupportEXT *variable_count =
       vk_find_struct(
          (void *) pCreateInfo->pNext,
-         DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_LAYOUT_SUPPORT);
-   const VkMutableDescriptorTypeCreateInfoEXT *mutable_info =
-      vk_find_struct_const(
-         pCreateInfo->pNext,
-         MUTABLE_DESCRIPTOR_TYPE_CREATE_INFO_EXT);
-
+         DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_LAYOUT_SUPPORT_EXT);
    if (variable_count) {
       variable_count->maxVariableDescriptorCount = 0;
    }
@@ -300,157 +254,71 @@ tu_GetDescriptorSetLayoutSupport(
    for (uint32_t i = 0; i < pCreateInfo->bindingCount; i++) {
       const VkDescriptorSetLayoutBinding *binding = bindings + i;
 
-      uint64_t descriptor_sz;
-
-      if (is_dynamic(binding->descriptorType)) {
-         descriptor_sz = 0;
-      } else if (binding->descriptorType == VK_DESCRIPTOR_TYPE_MUTABLE_EXT) {
-         const VkMutableDescriptorTypeListEXT *list =
-            &mutable_info->pMutableDescriptorTypeLists[i];
-
-         for (uint32_t j = 0; j < list->descriptorTypeCount; j++) {
-            /* Don't support the input attachement and combined image sampler type
-             * for mutable descriptors */
-            if (list->pDescriptorTypes[j] == VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT ||
-                list->pDescriptorTypes[j] == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER ||
-                list->pDescriptorTypes[j] == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
-               supported = false;
-               goto out;
-            }
-         }
-
-         descriptor_sz =
-            mutable_descriptor_size(device, &mutable_info->pMutableDescriptorTypeLists[i]);
-      } else {
-         descriptor_sz = descriptor_size(device, binding, binding->descriptorType);
+      uint64_t descriptor_size = 0;
+      uint64_t descriptor_alignment = 1;
+      switch (binding->descriptorType) {
+      case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
+      case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
+         break;
+      case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+      case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+      case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+      case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+         descriptor_size = 16;
+         descriptor_alignment = 16;
+         break;
+      case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
+      case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
+      case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
+         descriptor_size = 64;
+         descriptor_alignment = 32;
+         break;
+      case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+         descriptor_size = 96;
+         descriptor_alignment = 32;
+         break;
+      case VK_DESCRIPTOR_TYPE_SAMPLER:
+         descriptor_size = 16;
+         descriptor_alignment = 16;
+         break;
+      default:
+         unreachable("unknown descriptor type\n");
+         break;
       }
-      uint64_t descriptor_alignment = 4 * A6XX_TEX_CONST_DWORDS;
 
-      if (size && !ALIGN_POT(size, descriptor_alignment)) {
+      if (size && !align_u64(size, descriptor_alignment)) {
          supported = false;
       }
-      size = ALIGN_POT(size, descriptor_alignment);
-
-      uint64_t max_count = MAX_SET_SIZE;
-      unsigned descriptor_count = binding->descriptorCount;
-      if (binding->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
-         max_count = MAX_SET_SIZE - size;
-         descriptor_count = descriptor_sz;
-         descriptor_sz = 1;
-      } else if (descriptor_sz) {
-         max_count = (MAX_SET_SIZE - size) / descriptor_sz;
-      }
+      size = align_u64(size, descriptor_alignment);
 
-      if (max_count < descriptor_count) {
+      uint64_t max_count = UINT64_MAX;
+      if (descriptor_size)
+         max_count = (UINT64_MAX - size) / descriptor_size;
+
+      if (max_count < binding->descriptorCount) {
          supported = false;
       }
-
       if (variable_flags && binding->binding < variable_flags->bindingCount &&
           variable_count &&
           (variable_flags->pBindingFlags[binding->binding] &
-           VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT)) {
+           VK_DESCRIPTOR_BINDING_VARIABLE_DESCRIPTOR_COUNT_BIT_EXT)) {
          variable_count->maxVariableDescriptorCount =
             MIN2(UINT32_MAX, max_count);
       }
-      size += descriptor_count * descriptor_sz;
+      size += binding->descriptorCount * descriptor_size;
    }
 
-out:
    free(bindings);
 
    pSupport->supported = supported;
 }
 
-/* Note: we must hash any values used in tu_lower_io(). */
-
-#define SHA1_UPDATE_VALUE(ctx, x) _mesa_sha1_update(ctx, &(x), sizeof(x));
-
-static void
-sha1_update_ycbcr_sampler(struct mesa_sha1 *ctx,
-                          const struct tu_sampler_ycbcr_conversion *sampler)
-{
-   SHA1_UPDATE_VALUE(ctx, sampler->ycbcr_model);
-   SHA1_UPDATE_VALUE(ctx, sampler->ycbcr_range);
-   SHA1_UPDATE_VALUE(ctx, sampler->format);
-}
-
-static void
-sha1_update_descriptor_set_binding_layout(struct mesa_sha1 *ctx,
-   const struct tu_descriptor_set_binding_layout *layout,
-   const struct tu_descriptor_set_layout *set_layout)
-{
-   SHA1_UPDATE_VALUE(ctx, layout->type);
-   SHA1_UPDATE_VALUE(ctx, layout->offset);
-   SHA1_UPDATE_VALUE(ctx, layout->size);
-   SHA1_UPDATE_VALUE(ctx, layout->array_size);
-   SHA1_UPDATE_VALUE(ctx, layout->dynamic_offset_offset);
-   SHA1_UPDATE_VALUE(ctx, layout->immutable_samplers_offset);
-
-   const struct tu_sampler_ycbcr_conversion *ycbcr_samplers =
-      tu_immutable_ycbcr_samplers(set_layout, layout);
-
-   if (ycbcr_samplers) {
-      for (unsigned i = 0; i < layout->array_size; i++)
-         sha1_update_ycbcr_sampler(ctx, ycbcr_samplers + i);
-   }
-}
-
-
-static void
-sha1_update_descriptor_set_layout(struct mesa_sha1 *ctx,
-                                  const struct tu_descriptor_set_layout *layout)
-{
-   for (uint16_t i = 0; i < layout->binding_count; i++)
-      sha1_update_descriptor_set_binding_layout(ctx, &layout->binding[i],
-                                                layout);
-}
-
 /*
  * Pipeline layouts.  These have nothing to do with the pipeline.  They are
  * just multiple descriptor set layouts pasted together.
  */
 
-void
-tu_pipeline_layout_init(struct tu_pipeline_layout *layout)
-{
-   unsigned dynamic_offset_size = 0;
-
-   for (uint32_t set = 0; set < layout->num_sets; set++) {
-      assert(set < MAX_SETS);
-      layout->set[set].dynamic_offset_start = dynamic_offset_size;
-
-      if (layout->set[set].layout)
-         dynamic_offset_size += layout->set[set].layout->dynamic_offset_size;
-   }
-
-   layout->dynamic_offset_size = dynamic_offset_size;
-
-   /* We only care about INDEPENDENT_SETS for dynamic-offset descriptors,
-    * where all the descriptors from all the sets are combined into one set
-    * and we have to provide the dynamic_offset_start dynamically with fast
-    * linking.
-    */
-   if (dynamic_offset_size == 0) {
-      layout->independent_sets = false;
-   }
-
-   struct mesa_sha1 ctx;
-   _mesa_sha1_init(&ctx);
-   for (unsigned s = 0; s < layout->num_sets; s++) {
-      if (layout->set[s].layout)
-         sha1_update_descriptor_set_layout(&ctx, layout->set[s].layout);
-      _mesa_sha1_update(&ctx, &layout->set[s].dynamic_offset_start,
-                        sizeof(layout->set[s].dynamic_offset_start));
-   }
-   _mesa_sha1_update(&ctx, &layout->num_sets, sizeof(layout->num_sets));
-   _mesa_sha1_update(&ctx, &layout->push_constant_size,
-                     sizeof(layout->push_constant_size));
-   _mesa_sha1_update(&ctx, &layout->independent_sets,
-                     sizeof(layout->independent_sets));
-   _mesa_sha1_final(&ctx, layout->sha1);
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreatePipelineLayout(VkDevice _device,
                         const VkPipelineLayoutCreateInfo *pCreateInfo,
                         const VkAllocationCallbacks *pAllocator,
@@ -458,26 +326,42 @@ tu_CreatePipelineLayout(VkDevice _device,
 {
    TU_FROM_HANDLE(tu_device, device, _device);
    struct tu_pipeline_layout *layout;
+   struct mesa_sha1 ctx;
 
    assert(pCreateInfo->sType ==
           VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO);
 
-   layout = vk_object_alloc(&device->vk, pAllocator, sizeof(*layout),
-                            VK_OBJECT_TYPE_PIPELINE_LAYOUT);
+   layout = vk_alloc2(&device->alloc, pAllocator, sizeof(*layout), 8,
+                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (layout == NULL)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    layout->num_sets = pCreateInfo->setLayoutCount;
+
+   unsigned dynamic_offset_count = 0;
+
+   _mesa_sha1_init(&ctx);
    for (uint32_t set = 0; set < pCreateInfo->setLayoutCount; set++) {
       TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout,
                      pCreateInfo->pSetLayouts[set]);
-
-      assert(set < MAX_SETS);
       layout->set[set].layout = set_layout;
-      if (set_layout)
-         vk_descriptor_set_layout_ref(&set_layout->vk);
+
+      layout->set[set].dynamic_offset_start = dynamic_offset_count;
+      for (uint32_t b = 0; b < set_layout->binding_count; b++) {
+         dynamic_offset_count += set_layout->binding[b].array_size *
+                                 set_layout->binding[b].dynamic_offset_count;
+         if (set_layout->binding[b].immutable_samplers_offset)
+            _mesa_sha1_update(
+               &ctx,
+               tu_immutable_samplers(set_layout, set_layout->binding + b),
+               set_layout->binding[b].array_size * 4 * sizeof(uint32_t));
+      }
+      _mesa_sha1_update(
+         &ctx, set_layout->binding,
+         sizeof(set_layout->binding[0]) * set_layout->binding_count);
    }
 
+   layout->dynamic_offset_count = dynamic_offset_count;
    layout->push_constant_size = 0;
 
    for (unsigned i = 0; i < pCreateInfo->pushConstantRangeCount; ++i) {
@@ -487,17 +371,15 @@ tu_CreatePipelineLayout(VkDevice _device,
    }
 
    layout->push_constant_size = align(layout->push_constant_size, 16);
-   layout->independent_sets =
-      pCreateInfo->flags & VK_PIPELINE_LAYOUT_CREATE_INDEPENDENT_SETS_BIT_EXT;
-
-   tu_pipeline_layout_init(layout);
-
+   _mesa_sha1_update(&ctx, &layout->push_constant_size,
+                     sizeof(layout->push_constant_size));
+   _mesa_sha1_final(&ctx, layout->sha1);
    *pPipelineLayout = tu_pipeline_layout_to_handle(layout);
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyPipelineLayout(VkDevice _device,
                          VkPipelineLayout _pipelineLayout,
                          const VkAllocationCallbacks *pAllocator)
@@ -507,329 +389,31 @@ tu_DestroyPipelineLayout(VkDevice _device,
 
    if (!pipeline_layout)
       return;
-
-   for (uint32_t i = 0; i < pipeline_layout->num_sets; i++) {
-      if (pipeline_layout->set[i].layout)
-         vk_descriptor_set_layout_unref(&device->vk, &pipeline_layout->set[i].layout->vk);
-   }
-
-   vk_object_free(&device->vk, pAllocator, pipeline_layout);
+   vk_free2(&device->alloc, pAllocator, pipeline_layout);
 }
 
 #define EMPTY 1
 
-static VkResult
-tu_descriptor_set_create(struct tu_device *device,
-            struct tu_descriptor_pool *pool,
-            struct tu_descriptor_set_layout *layout,
-            uint32_t variable_count,
-            struct tu_descriptor_set **out_set)
-{
-   struct tu_descriptor_set *set;
-   unsigned dynamic_offset = sizeof(struct tu_descriptor_set);
-   unsigned mem_size = dynamic_offset + layout->dynamic_offset_size;
-
-   if (pool->host_memory_base) {
-      if (pool->host_memory_end - pool->host_memory_ptr < mem_size)
-         return vk_error(device, VK_ERROR_OUT_OF_POOL_MEMORY);
-
-      set = (struct tu_descriptor_set*)pool->host_memory_ptr;
-      pool->host_memory_ptr += mem_size;
-   } else {
-      set = vk_alloc2(&device->vk.alloc, NULL, mem_size, 8,
-                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
-
-      if (!set)
-         return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
-   }
-
-   memset(set, 0, mem_size);
-   vk_object_base_init(&device->vk, &set->base, VK_OBJECT_TYPE_DESCRIPTOR_SET);
-
-   if (layout->dynamic_offset_size) {
-      set->dynamic_descriptors = (uint32_t *)((uint8_t*)set + dynamic_offset);
-   }
-
-   set->layout = layout;
-   set->pool = pool;
-   uint32_t layout_size = layout->size;
-   if (layout->has_variable_descriptors) {
-      struct tu_descriptor_set_binding_layout *binding =
-         &layout->binding[layout->binding_count - 1];
-      if (binding->type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
-         layout_size = binding->offset + A6XX_TEX_CONST_DWORDS * 4 +
-            ALIGN(variable_count, A6XX_TEX_CONST_DWORDS * 4);
-      } else {
-         uint32_t stride = binding->size;
-         layout_size = binding->offset + variable_count * stride;
-      }
-   }
-
-   if (layout_size) {
-      set->size = layout_size;
-
-      if (!pool->host_memory_base && pool->entry_count == pool->max_entry_count) {
-         vk_object_free(&device->vk, NULL, set);
-         return vk_error(device, VK_ERROR_OUT_OF_POOL_MEMORY);
-      }
-
-      /* try to allocate linearly first, so that we don't spend
-       * time looking for gaps if the app only allocates &
-       * resets via the pool. */
-      if (pool->current_offset + layout_size <= pool->size) {
-         set->mapped_ptr = (uint32_t*)(pool_base(pool) + pool->current_offset);
-         set->va = pool->host_bo ? 0 : pool->bo->iova + pool->current_offset;
-
-         if (!pool->host_memory_base) {
-            pool->entries[pool->entry_count].offset = pool->current_offset;
-            pool->entries[pool->entry_count].size = layout_size;
-            pool->entries[pool->entry_count].set = set;
-            pool->entry_count++;
-         }
-         pool->current_offset += layout_size;
-      } else if (!pool->host_memory_base) {
-         uint64_t offset = 0;
-         int index;
-
-         for (index = 0; index < pool->entry_count; ++index) {
-            if (pool->entries[index].offset - offset >= layout_size)
-               break;
-            offset = pool->entries[index].offset + pool->entries[index].size;
-         }
-
-         if (pool->size - offset < layout_size) {
-            vk_object_free(&device->vk, NULL, set);
-            return vk_error(device, VK_ERROR_OUT_OF_POOL_MEMORY);
-         }
-
-         set->mapped_ptr = (uint32_t*)(pool_base(pool) + offset);
-         set->va = pool->host_bo ? 0 : pool->bo->iova + offset;
-
-         memmove(&pool->entries[index + 1], &pool->entries[index],
-            sizeof(pool->entries[0]) * (pool->entry_count - index));
-         pool->entries[index].offset = offset;
-         pool->entries[index].size = layout_size;
-         pool->entries[index].set = set;
-         pool->entry_count++;
-      } else
-         return vk_error(device, VK_ERROR_OUT_OF_POOL_MEMORY);
-   }
-
-   if (layout->has_immutable_samplers) {
-      for (unsigned i = 0; i < layout->binding_count; ++i) {
-         if (!layout->binding[i].immutable_samplers_offset)
-            continue;
-
-         unsigned offset = layout->binding[i].offset / 4;
-         if (layout->binding[i].type == VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
-            offset += A6XX_TEX_CONST_DWORDS;
-
-         const struct tu_sampler *samplers =
-            (const struct tu_sampler *)((const char *)layout +
-                               layout->binding[i].immutable_samplers_offset);
-         for (unsigned j = 0; j < layout->binding[i].array_size; ++j) {
-            memcpy(set->mapped_ptr + offset, samplers[j].descriptor,
-                   sizeof(samplers[j].descriptor));
-            offset += layout->binding[i].size / 4;
-         }
-      }
-   }
-
-   if (layout->has_inline_uniforms) {
-      for (unsigned i = 0; i < layout->binding_count; i++) {
-         if (layout->binding[i].type != VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK)
-            continue;
-
-         uint32_t *ptr = set->mapped_ptr + layout->binding[i].offset / 4;
-         uint64_t va = set->va + layout->binding[i].offset +
-            A6XX_TEX_CONST_DWORDS * 4;
-         uint32_t size =
-            (layout->has_variable_descriptors && i == layout->binding_count - 1) ?
-            variable_count : layout->binding[i].size - A6XX_TEX_CONST_DWORDS * 4;
-         size = ALIGN_POT(size, 16) / 16;
-
-         ptr[0] = A6XX_UBO_0_BASE_LO(va);
-         ptr[1] = A6XX_UBO_1_BASE_HI(va >> 32) | A6XX_UBO_1_SIZE(size);
-      }
-   }
-
-   vk_descriptor_set_layout_ref(&layout->vk);
-   list_addtail(&set->pool_link, &pool->desc_sets);
-
-   *out_set = set;
-   return VK_SUCCESS;
-}
-
-static void
-tu_descriptor_set_destroy(struct tu_device *device,
-             struct tu_descriptor_pool *pool,
-             struct tu_descriptor_set *set,
-             bool free_bo)
-{
-   assert(!pool->host_memory_base);
-
-   if (free_bo && set->size && !pool->host_memory_base) {
-      uint32_t offset = (uint8_t*)set->mapped_ptr - pool_base(pool);
-
-      for (int i = 0; i < pool->entry_count; ++i) {
-         if (pool->entries[i].offset == offset) {
-            memmove(&pool->entries[i], &pool->entries[i+1],
-               sizeof(pool->entries[i]) * (pool->entry_count - i - 1));
-            --pool->entry_count;
-            break;
-         }
-      }
-   }
-
-   vk_object_free(&device->vk, NULL, set);
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateDescriptorPool(VkDevice _device,
                         const VkDescriptorPoolCreateInfo *pCreateInfo,
                         const VkAllocationCallbacks *pAllocator,
                         VkDescriptorPool *pDescriptorPool)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-   struct tu_descriptor_pool *pool;
-   uint64_t size = sizeof(struct tu_descriptor_pool);
-   uint64_t bo_size = 0, dynamic_size = 0;
-   VkResult ret;
-
-   const VkMutableDescriptorTypeCreateInfoEXT *mutable_info =
-      vk_find_struct_const( pCreateInfo->pNext,
-         MUTABLE_DESCRIPTOR_TYPE_CREATE_INFO_EXT);
-
-   const VkDescriptorPoolInlineUniformBlockCreateInfo *inline_info =
-      vk_find_struct_const(pCreateInfo->pNext,
-                           DESCRIPTOR_POOL_INLINE_UNIFORM_BLOCK_CREATE_INFO);
-
-   if (inline_info) {
-      /* In addition to the size of the descriptors, we have to factor in the
-       * padding for each binding. The sizes are 4 aligned but we have to
-       * align to a descriptor size, and in the worst case each inline
-       * binding has a size of 4 bytes and we have to pad each one out.
-       */
-      bo_size += (2 * 4 * A6XX_TEX_CONST_DWORDS - 4) *
-         inline_info->maxInlineUniformBlockBindings;
-   }
-
-   for (unsigned i = 0; i < pCreateInfo->poolSizeCount; ++i) {
-      const VkDescriptorPoolSize *pool_size = &pCreateInfo->pPoolSizes[i];
-
-      switch (pool_size->type) {
-      case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
-      case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
-         dynamic_size += descriptor_size(device, NULL, pool_size->type) *
-            pool_size->descriptorCount;
-         break;
-      case VK_DESCRIPTOR_TYPE_MUTABLE_EXT:
-         if (mutable_info && i < mutable_info->mutableDescriptorTypeListCount &&
-             mutable_info->pMutableDescriptorTypeLists[i].descriptorTypeCount > 0) {
-            bo_size +=
-               mutable_descriptor_size(device, &mutable_info->pMutableDescriptorTypeLists[i]) *
-                  pool_size->descriptorCount;
-         } else {
-            /* Allocate the maximum size possible. */
-            bo_size += 2 * A6XX_TEX_CONST_DWORDS * 4 *
-                  pool_size->descriptorCount;
-         }
-         break;
-      case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK:
-         bo_size += pool_size->descriptorCount;
-         break;
-      default:
-         bo_size += descriptor_size(device, NULL, pool_size->type) *
-                              pool_size->descriptorCount;
-         break;
-      }
-   }
-
-   if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) {
-      uint64_t host_size = pCreateInfo->maxSets * sizeof(struct tu_descriptor_set);
-      host_size += dynamic_size;
-      size += host_size;
-   } else {
-      size += sizeof(struct tu_descriptor_pool_entry) * pCreateInfo->maxSets;
-   }
-
-   pool = vk_object_zalloc(&device->vk, pAllocator, size,
-                          VK_OBJECT_TYPE_DESCRIPTOR_POOL);
-   if (!pool)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
-
-   if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT)) {
-      pool->host_memory_base = (uint8_t*)pool + sizeof(struct tu_descriptor_pool);
-      pool->host_memory_ptr = pool->host_memory_base;
-      pool->host_memory_end = (uint8_t*)pool + size;
-   }
-
-   if (bo_size) {
-      if (!(pCreateInfo->flags & VK_DESCRIPTOR_POOL_CREATE_HOST_ONLY_BIT_EXT)) {
-         ret = tu_bo_init_new(device, &pool->bo, bo_size, TU_BO_ALLOC_ALLOW_DUMP, "descriptor pool");
-         if (ret)
-            goto fail_alloc;
-
-         ret = tu_bo_map(device, pool->bo);
-         if (ret)
-            goto fail_map;
-      } else {
-         pool->host_bo = vk_alloc2(&device->vk.alloc, pAllocator, bo_size, 8,
-                                   VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
-         if (!pool->host_bo) {
-            ret = VK_ERROR_OUT_OF_HOST_MEMORY;
-            goto fail_alloc;
-         }
-      }
-   }
-   pool->size = bo_size;
-   pool->max_entry_count = pCreateInfo->maxSets;
-
-   list_inithead(&pool->desc_sets);
-
-   *pDescriptorPool = tu_descriptor_pool_to_handle(pool);
+   tu_use_args(device);
+   tu_stub();
    return VK_SUCCESS;
-
-fail_map:
-   tu_bo_finish(device, pool->bo);
-fail_alloc:
-   vk_object_free(&device->vk, pAllocator, pool);
-   return ret;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyDescriptorPool(VkDevice _device,
                          VkDescriptorPool _pool,
                          const VkAllocationCallbacks *pAllocator)
 {
-   TU_FROM_HANDLE(tu_device, device, _device);
-   TU_FROM_HANDLE(tu_descriptor_pool, pool, _pool);
-
-   if (!pool)
-      return;
-
-   list_for_each_entry_safe(struct tu_descriptor_set, set,
-                            &pool->desc_sets, pool_link) {
-      vk_descriptor_set_layout_unref(&device->vk, &set->layout->vk);
-   }
-
-   if (!pool->host_memory_base) {
-      for(int i = 0; i < pool->entry_count; ++i) {
-         tu_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
-      }
-   }
-
-   if (pool->size) {
-      if (pool->host_bo)
-         vk_free2(&device->vk.alloc, pAllocator, pool->host_bo);
-      else
-         tu_bo_finish(device, pool->bo);
-   }
-
-   vk_object_free(&device->vk, pAllocator, pool);
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_ResetDescriptorPool(VkDevice _device,
                        VkDescriptorPool descriptorPool,
                        VkDescriptorPoolResetFlags flags)
@@ -837,26 +421,12 @@ tu_ResetDescriptorPool(VkDevice _device,
    TU_FROM_HANDLE(tu_device, device, _device);
    TU_FROM_HANDLE(tu_descriptor_pool, pool, descriptorPool);
 
-   list_for_each_entry_safe(struct tu_descriptor_set, set,
-                            &pool->desc_sets, pool_link) {
-      vk_descriptor_set_layout_unref(&device->vk, &set->layout->vk);
-   }
-   list_inithead(&pool->desc_sets);
-
-   if (!pool->host_memory_base) {
-      for(int i = 0; i < pool->entry_count; ++i) {
-         tu_descriptor_set_destroy(device, pool, pool->entries[i].set, false);
-      }
-      pool->entry_count = 0;
-   }
-
-   pool->current_offset = 0;
-   pool->host_memory_ptr = pool->host_memory_base;
-
+   tu_use_args(device, pool);
+   tu_stub();
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_AllocateDescriptorSets(VkDevice _device,
                           const VkDescriptorSetAllocateInfo *pAllocateInfo,
                           VkDescriptorSet *pDescriptorSets)
@@ -864,42 +434,12 @@ tu_AllocateDescriptorSets(VkDevice _device,
    TU_FROM_HANDLE(tu_device, device, _device);
    TU_FROM_HANDLE(tu_descriptor_pool, pool, pAllocateInfo->descriptorPool);
 
-   VkResult result = VK_SUCCESS;
-   uint32_t i;
-   struct tu_descriptor_set *set = NULL;
-
-   const VkDescriptorSetVariableDescriptorCountAllocateInfo *variable_counts =
-      vk_find_struct_const(pAllocateInfo->pNext, DESCRIPTOR_SET_VARIABLE_DESCRIPTOR_COUNT_ALLOCATE_INFO);
-   if (variable_counts && !variable_counts->descriptorSetCount)
-      variable_counts = NULL;
-
-   /* allocate a set of buffers for each shader to contain descriptors */
-   for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
-      TU_FROM_HANDLE(tu_descriptor_set_layout, layout,
-             pAllocateInfo->pSetLayouts[i]);
-
-      assert(!(layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
-
-      result = tu_descriptor_set_create(
-         device, pool, layout,
-         variable_counts ? variable_counts->pDescriptorCounts[i] : 0, &set);
-      if (result != VK_SUCCESS)
-         break;
-
-      pDescriptorSets[i] = tu_descriptor_set_to_handle(set);
-   }
-
-   if (result != VK_SUCCESS) {
-      tu_FreeDescriptorSets(_device, pAllocateInfo->descriptorPool,
-               i, pDescriptorSets);
-      for (i = 0; i < pAllocateInfo->descriptorSetCount; i++) {
-         pDescriptorSets[i] = VK_NULL_HANDLE;
-      }
-   }
-   return result;
+   tu_use_args(device, pool);
+   tu_stub();
+   return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_FreeDescriptorSets(VkDevice _device,
                       VkDescriptorPool descriptorPool,
                       uint32_t count,
@@ -908,338 +448,23 @@ tu_FreeDescriptorSets(VkDevice _device,
    TU_FROM_HANDLE(tu_device, device, _device);
    TU_FROM_HANDLE(tu_descriptor_pool, pool, descriptorPool);
 
-   for (uint32_t i = 0; i < count; i++) {
-      TU_FROM_HANDLE(tu_descriptor_set, set, pDescriptorSets[i]);
-
-      if (set) {
-         vk_descriptor_set_layout_unref(&device->vk, &set->layout->vk);
-         list_del(&set->pool_link);
-      }
-
-      if (set && !pool->host_memory_base)
-         tu_descriptor_set_destroy(device, pool, set, true);
-   }
+   tu_use_args(device, pool);
+   tu_stub();
    return VK_SUCCESS;
 }
 
-static void
-write_texel_buffer_descriptor(uint32_t *dst, const VkBufferView buffer_view)
-{
-   if (buffer_view == VK_NULL_HANDLE) {
-      memset(dst, 0, A6XX_TEX_CONST_DWORDS * sizeof(uint32_t));
-   } else {
-      TU_FROM_HANDLE(tu_buffer_view, view, buffer_view);
-
-      memcpy(dst, view->descriptor, sizeof(view->descriptor));
-   }
-}
-
-static void
-write_buffer_descriptor(const struct tu_device *device,
-                        uint32_t *dst,
-                        const VkDescriptorBufferInfo *buffer_info)
-{
-   bool storage_16bit = device->physical_device->info->a6xx.storage_16bit;
-   /* newer a6xx allows using 16-bit descriptor for both 16-bit and 32-bit
-    * access, but we need to keep a 32-bit descriptor for readonly access via
-    * isam.
-    */
-   unsigned descriptors = storage_16bit ? 2 : 1;
-   if (buffer_info->buffer == VK_NULL_HANDLE) {
-      memset(dst, 0, descriptors * A6XX_TEX_CONST_DWORDS * sizeof(uint32_t));
-      return;
-   }
-
-   TU_FROM_HANDLE(tu_buffer, buffer, buffer_info->buffer);
-
-   assert((buffer_info->offset & 63) == 0); /* minStorageBufferOffsetAlignment */
-   uint64_t va = buffer->iova + buffer_info->offset;
-   uint32_t range = vk_buffer_range(&buffer->vk, buffer_info->offset, buffer_info->range);
-
-   for (unsigned i = 0; i < descriptors; i++) {
-      if (storage_16bit && i == 0) {
-         dst[0] = A6XX_TEX_CONST_0_TILE_MODE(TILE6_LINEAR) | A6XX_TEX_CONST_0_FMT(FMT6_16_UINT);
-         dst[1] = DIV_ROUND_UP(range, 2);
-      } else {
-         dst[0] = A6XX_TEX_CONST_0_TILE_MODE(TILE6_LINEAR) | A6XX_TEX_CONST_0_FMT(FMT6_32_UINT);
-         dst[1] = DIV_ROUND_UP(range, 4);
-      }
-      dst[2] =
-         A6XX_TEX_CONST_2_BUFFER | A6XX_TEX_CONST_2_TYPE(A6XX_TEX_BUFFER);
-      dst[3] = 0;
-      dst[4] = A6XX_TEX_CONST_4_BASE_LO(va);
-      dst[5] = A6XX_TEX_CONST_5_BASE_HI(va >> 32);
-      for (int j = 6; j < A6XX_TEX_CONST_DWORDS; j++)
-         dst[j] = 0;
-      dst += A6XX_TEX_CONST_DWORDS;
-   }
-}
-
-static void
-write_ubo_descriptor(uint32_t *dst, const VkDescriptorBufferInfo *buffer_info)
-{
-   if (buffer_info->buffer == VK_NULL_HANDLE) {
-      dst[0] = dst[1] = 0;
-      return;
-   }
-
-   TU_FROM_HANDLE(tu_buffer, buffer, buffer_info->buffer);
-
-   uint32_t range = vk_buffer_range(&buffer->vk, buffer_info->offset, buffer_info->range);
-   /* The HW range is in vec4 units */
-   range = ALIGN_POT(range, 16) / 16;
-   uint64_t va = buffer->iova + buffer_info->offset;
-
-   dst[0] = A6XX_UBO_0_BASE_LO(va);
-   dst[1] = A6XX_UBO_1_BASE_HI(va >> 32) | A6XX_UBO_1_SIZE(range);
-}
-
-static void
-write_image_descriptor(uint32_t *dst,
-                       VkDescriptorType descriptor_type,
-                       const VkDescriptorImageInfo *image_info)
-{
-   if (image_info->imageView == VK_NULL_HANDLE) {
-      memset(dst, 0, A6XX_TEX_CONST_DWORDS * sizeof(uint32_t));
-      return;
-   }
-
-   TU_FROM_HANDLE(tu_image_view, iview, image_info->imageView);
-
-   if (descriptor_type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE) {
-      memcpy(dst, iview->view.storage_descriptor, sizeof(iview->view.storage_descriptor));
-   } else {
-      memcpy(dst, iview->view.descriptor, sizeof(iview->view.descriptor));
-   }
-}
-
-static void
-write_combined_image_sampler_descriptor(uint32_t *dst,
-                                        VkDescriptorType descriptor_type,
-                                        const VkDescriptorImageInfo *image_info,
-                                        bool has_sampler)
-{
-   write_image_descriptor(dst, descriptor_type, image_info);
-   /* copy over sampler state */
-   if (has_sampler) {
-      TU_FROM_HANDLE(tu_sampler, sampler, image_info->sampler);
-      memcpy(dst + A6XX_TEX_CONST_DWORDS, sampler->descriptor, sizeof(sampler->descriptor));
-   }
-}
-
-static void
-write_sampler_descriptor(uint32_t *dst, const VkDescriptorImageInfo *image_info)
-{
-   TU_FROM_HANDLE(tu_sampler, sampler, image_info->sampler);
-
-   memcpy(dst, sampler->descriptor, sizeof(sampler->descriptor));
-}
-
-/* note: this is used with immutable samplers in push descriptors */
-static void
-write_sampler_push(uint32_t *dst, const struct tu_sampler *sampler)
-{
-   memcpy(dst, sampler->descriptor, sizeof(sampler->descriptor));
-}
-
 void
-tu_update_descriptor_sets(const struct tu_device *device,
+tu_update_descriptor_sets(struct tu_device *device,
+                          struct tu_cmd_buffer *cmd_buffer,
                           VkDescriptorSet dstSetOverride,
                           uint32_t descriptorWriteCount,
                           const VkWriteDescriptorSet *pDescriptorWrites,
                           uint32_t descriptorCopyCount,
                           const VkCopyDescriptorSet *pDescriptorCopies)
 {
-   uint32_t i, j;
-   for (i = 0; i < descriptorWriteCount; i++) {
-      const VkWriteDescriptorSet *writeset = &pDescriptorWrites[i];
-      TU_FROM_HANDLE(tu_descriptor_set, set, dstSetOverride ?: writeset->dstSet);
-      const struct tu_descriptor_set_binding_layout *binding_layout =
-         set->layout->binding + writeset->dstBinding;
-      uint32_t *ptr = set->mapped_ptr;
-      if (writeset->descriptorType == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC ||
-          writeset->descriptorType == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC) {
-         ptr = set->dynamic_descriptors;
-         ptr += binding_layout->dynamic_offset_offset / 4;
-      } else {
-         ptr = set->mapped_ptr;
-         ptr += binding_layout->offset / 4;
-      }
-
-      /* for immutable samplers with push descriptors: */
-      const bool copy_immutable_samplers =
-         dstSetOverride && binding_layout->immutable_samplers_offset;
-      const struct tu_sampler *samplers =
-         tu_immutable_samplers(set->layout, binding_layout);
-
-      if (writeset->descriptorType == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
-         /* We need to respect this note:
-          *
-          *    The same behavior applies to bindings with a descriptor type of
-          *    VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK where descriptorCount
-          *    specifies the number of bytes to update while dstArrayElement
-          *    specifies the starting byte offset, thus in this case if the
-          *    dstBinding has a smaller byte size than the sum of
-          *    dstArrayElement and descriptorCount, then the remainder will be
-          *    used to update the subsequent binding - dstBinding+1 starting
-          *    at offset zero. This falls out as a special case of the above
-          *    rule.
-          *
-          * This means we can't just do a straight memcpy, because due to
-          * alignment padding and the descriptor itself there are gaps between
-          * sequential bindings. We have to loop over each binding updated.
-          */
-         const VkWriteDescriptorSetInlineUniformBlock *inline_write =
-            vk_find_struct_const(writeset->pNext,
-                                 WRITE_DESCRIPTOR_SET_INLINE_UNIFORM_BLOCK);
-         uint32_t remaining = inline_write->dataSize;
-         const uint8_t *src = inline_write->pData;
-         uint32_t dst_offset = writeset->dstArrayElement;
-         do {
-            uint8_t *dst = (uint8_t *)(ptr + A6XX_TEX_CONST_DWORDS) + dst_offset;
-            uint32_t binding_size =
-               binding_layout->size - A6XX_TEX_CONST_DWORDS * 4 - dst_offset;
-            uint32_t to_write = MIN2(remaining, binding_size);
-            memcpy(dst, src, to_write);
-
-            binding_layout++;
-            ptr = set->mapped_ptr + binding_layout->offset / 4;
-            dst_offset = 0;
-            src += to_write;
-            remaining -= to_write;
-         } while (remaining > 0);
-
-         continue;
-      }
-
-      ptr += binding_layout->size / 4 * writeset->dstArrayElement;
-      for (j = 0; j < writeset->descriptorCount; ++j) {
-         switch(writeset->descriptorType) {
-         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
-         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
-            write_ubo_descriptor(ptr, writeset->pBufferInfo + j);
-            break;
-         case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
-         case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
-            write_buffer_descriptor(device, ptr, writeset->pBufferInfo + j);
-            break;
-         case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
-         case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
-            write_texel_buffer_descriptor(ptr, writeset->pTexelBufferView[j]);
-            break;
-         case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
-         case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
-            write_image_descriptor(ptr, writeset->descriptorType, writeset->pImageInfo + j);
-            break;
-         case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
-            write_combined_image_sampler_descriptor(ptr,
-                                                    writeset->descriptorType,
-                                                    writeset->pImageInfo + j,
-                                                    !binding_layout->immutable_samplers_offset);
-
-            if (copy_immutable_samplers)
-               write_sampler_push(ptr + A6XX_TEX_CONST_DWORDS, &samplers[writeset->dstArrayElement + j]);
-            break;
-         case VK_DESCRIPTOR_TYPE_SAMPLER:
-            if (!binding_layout->immutable_samplers_offset)
-               write_sampler_descriptor(ptr, writeset->pImageInfo + j);
-            else if (copy_immutable_samplers)
-               write_sampler_push(ptr, &samplers[writeset->dstArrayElement + j]);
-            break;
-         case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
-            /* nothing in descriptor set - framebuffer state is used instead */
-            if (unlikely(device->instance->debug_flags & TU_DEBUG_DYNAMIC))
-               write_image_descriptor(ptr, writeset->descriptorType, writeset->pImageInfo + j);
-            break;
-         default:
-            unreachable("unimplemented descriptor type");
-            break;
-         }
-         ptr += binding_layout->size / 4;
-      }
-   }
-
-   for (i = 0; i < descriptorCopyCount; i++) {
-      const VkCopyDescriptorSet *copyset = &pDescriptorCopies[i];
-      TU_FROM_HANDLE(tu_descriptor_set, src_set,
-                       copyset->srcSet);
-      TU_FROM_HANDLE(tu_descriptor_set, dst_set,
-                       copyset->dstSet);
-      const struct tu_descriptor_set_binding_layout *src_binding_layout =
-         src_set->layout->binding + copyset->srcBinding;
-      const struct tu_descriptor_set_binding_layout *dst_binding_layout =
-         dst_set->layout->binding + copyset->dstBinding;
-      uint32_t *src_ptr = src_set->mapped_ptr;
-      uint32_t *dst_ptr = dst_set->mapped_ptr;
-      if (src_binding_layout->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC ||
-          src_binding_layout->type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC) {
-         src_ptr = src_set->dynamic_descriptors;
-         dst_ptr = dst_set->dynamic_descriptors;
-         src_ptr += src_binding_layout->dynamic_offset_offset / 4;
-         dst_ptr += dst_binding_layout->dynamic_offset_offset / 4;
-      } else {
-         src_ptr = src_set->mapped_ptr;
-         dst_ptr = dst_set->mapped_ptr;
-         src_ptr += src_binding_layout->offset / 4;
-         dst_ptr += dst_binding_layout->offset / 4;
-      }
-
-      if (src_binding_layout->type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
-         uint32_t remaining = copyset->descriptorCount;
-         uint32_t src_start = copyset->srcArrayElement;
-         uint32_t dst_start = copyset->dstArrayElement;
-         uint8_t *src = (uint8_t *)(src_ptr + A6XX_TEX_CONST_DWORDS) + src_start;
-         uint8_t *dst = (uint8_t *)(dst_ptr + A6XX_TEX_CONST_DWORDS) + dst_start;
-         uint32_t src_remaining =
-            src_binding_layout->size - src_start - 4 * A6XX_TEX_CONST_DWORDS;
-         uint32_t dst_remaining =
-            dst_binding_layout->size - dst_start - 4 * A6XX_TEX_CONST_DWORDS;
-         do {
-            uint32_t to_write = MIN3(remaining, src_remaining, dst_remaining);
-            memcpy(dst, src, to_write);
-
-            src += to_write;
-            dst += to_write;
-            src_remaining -= to_write;
-            dst_remaining -= to_write;
-            remaining -= to_write;
-            
-            if (src_remaining == 0) {
-               src_binding_layout++;
-               src_ptr = src_set->mapped_ptr + src_binding_layout->offset / 4;
-               src = (uint8_t *)(src_ptr + A6XX_TEX_CONST_DWORDS);
-               src_remaining = src_binding_layout->size - 4 * A6XX_TEX_CONST_DWORDS;
-            }
-
-            if (dst_remaining == 0) {
-               dst_binding_layout++;
-               dst_ptr = dst_set->mapped_ptr + dst_binding_layout->offset / 4;
-               dst = (uint8_t *)(dst_ptr + A6XX_TEX_CONST_DWORDS);
-               dst_remaining = dst_binding_layout->size - 4 * A6XX_TEX_CONST_DWORDS;
-            }
-         } while (remaining > 0);
-
-         continue;
-      }
-
-      src_ptr += src_binding_layout->size * copyset->srcArrayElement / 4;
-      dst_ptr += dst_binding_layout->size * copyset->dstArrayElement / 4;
-
-      /* In case of copies between mutable descriptor types
-       * and non-mutable descriptor types.
-       */
-      uint32_t copy_size = MIN2(src_binding_layout->size, dst_binding_layout->size);
-
-      for (j = 0; j < copyset->descriptorCount; ++j) {
-         memcpy(dst_ptr, src_ptr, copy_size);
-
-         src_ptr += src_binding_layout->size / 4;
-         dst_ptr += dst_binding_layout->size / 4;
-      }
-   }
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_UpdateDescriptorSets(VkDevice _device,
                         uint32_t descriptorWriteCount,
                         const VkWriteDescriptorSet *pDescriptorWrites,
@@ -1247,12 +472,13 @@ tu_UpdateDescriptorSets(VkDevice _device,
                         const VkCopyDescriptorSet *pDescriptorCopies)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-   tu_update_descriptor_sets(device, VK_NULL_HANDLE,
+
+   tu_update_descriptor_sets(device, NULL, VK_NULL_HANDLE,
                              descriptorWriteCount, pDescriptorWrites,
                              descriptorCopyCount, pDescriptorCopies);
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateDescriptorUpdateTemplate(
    VkDevice _device,
    const VkDescriptorUpdateTemplateCreateInfo *pCreateInfo,
@@ -1260,142 +486,28 @@ tu_CreateDescriptorUpdateTemplate(
    VkDescriptorUpdateTemplate *pDescriptorUpdateTemplate)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-   struct tu_descriptor_set_layout *set_layout = NULL;
+   TU_FROM_HANDLE(tu_descriptor_set_layout, set_layout,
+                  pCreateInfo->descriptorSetLayout);
    const uint32_t entry_count = pCreateInfo->descriptorUpdateEntryCount;
-   uint32_t dst_entry_count = 0;
-
-   if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR) {
-      TU_FROM_HANDLE(tu_pipeline_layout, pipeline_layout, pCreateInfo->pipelineLayout);
-
-      /* descriptorSetLayout should be ignored for push descriptors
-       * and instead it refers to pipelineLayout and set.
-       */
-      assert(pCreateInfo->set < MAX_SETS);
-      set_layout = pipeline_layout->set[pCreateInfo->set].layout;
-   } else {
-      TU_FROM_HANDLE(tu_descriptor_set_layout, _set_layout,
-                     pCreateInfo->descriptorSetLayout);
-      set_layout = _set_layout;
-   }
-
-   for (uint32_t i = 0; i < entry_count; i++) {
-      const VkDescriptorUpdateTemplateEntry *entry = &pCreateInfo->pDescriptorUpdateEntries[i];
-      if (entry->descriptorType != VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
-         dst_entry_count++;
-         continue;
-      }
-
-      /* Calculate how many bindings this update steps over, so we can split
-       * up the template entry. This lets the actual update be a simple
-       * memcpy.
-       */
-      uint32_t remaining = entry->descriptorCount;
-      const struct tu_descriptor_set_binding_layout *binding_layout =
-         set_layout->binding + entry->dstBinding;
-      uint32_t dst_start = entry->dstArrayElement;
-      do {
-         uint32_t size = binding_layout->size - A6XX_TEX_CONST_DWORDS * 4;
-         uint32_t count = MIN2(remaining, size - dst_start);
-         remaining -= count;
-         binding_layout++;
-         dst_entry_count++;
-         dst_start = 0;
-      } while (remaining > 0);
-   }
-
    const size_t size =
       sizeof(struct tu_descriptor_update_template) +
-      sizeof(struct tu_descriptor_update_template_entry) * dst_entry_count;
+      sizeof(struct tu_descriptor_update_template_entry) * entry_count;
    struct tu_descriptor_update_template *templ;
 
-   templ = vk_object_alloc(&device->vk, pAllocator, size,
-                           VK_OBJECT_TYPE_DESCRIPTOR_UPDATE_TEMPLATE);
+   templ = vk_alloc2(&device->alloc, pAllocator, size, 8,
+                     VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (!templ)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
-
-   templ->entry_count = dst_entry_count;
-
-   if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR) {
-      templ->bind_point = pCreateInfo->pipelineBindPoint;
-   }
-
-   uint32_t j = 0;
-   for (uint32_t i = 0; i < entry_count; i++) {
-      const VkDescriptorUpdateTemplateEntry *entry = &pCreateInfo->pDescriptorUpdateEntries[i];
-
-      const struct tu_descriptor_set_binding_layout *binding_layout =
-         set_layout->binding + entry->dstBinding;
-      uint32_t dst_offset, dst_stride;
-      const struct tu_sampler *immutable_samplers = NULL;
-
-      /* dst_offset is an offset into dynamic_descriptors when the descriptor 
-       * is dynamic, and an offset into mapped_ptr otherwise.
-       */
-      switch (entry->descriptorType) {
-      case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
-      case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
-         dst_offset = binding_layout->dynamic_offset_offset / 4;
-         break;
-      case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK: {
-         uint32_t remaining = entry->descriptorCount;
-         uint32_t dst_start = entry->dstArrayElement;
-         uint32_t src_offset = entry->offset;
-         /* See comment in update_descriptor_sets() */
-         do {
-            dst_offset =
-               binding_layout->offset + A6XX_TEX_CONST_DWORDS * 4 + dst_start;
-            uint32_t size = binding_layout->size - A6XX_TEX_CONST_DWORDS * 4;
-            uint32_t count = MIN2(remaining, size - dst_start);
-            templ->entry[j++] = (struct tu_descriptor_update_template_entry) {
-               .descriptor_type = entry->descriptorType,
-               .descriptor_count = count,
-               .src_offset = src_offset,
-               .dst_offset = dst_offset,
-            };
-            remaining -= count;
-            src_offset += count;
-            binding_layout++;
-            dst_start = 0;
-         } while (remaining > 0);
-
-         continue;
-      }
-      case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
-      case VK_DESCRIPTOR_TYPE_SAMPLER:
-         if (pCreateInfo->templateType == VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_PUSH_DESCRIPTORS_KHR &&
-             binding_layout->immutable_samplers_offset) {
-            immutable_samplers =
-               tu_immutable_samplers(set_layout, binding_layout) + entry->dstArrayElement;
-         }
-         FALLTHROUGH;
-      default:
-         dst_offset = binding_layout->offset / 4;
-      }
-
-      dst_offset += (binding_layout->size * entry->dstArrayElement) / 4;
-      dst_stride = binding_layout->size / 4;
-
-      templ->entry[j++] = (struct tu_descriptor_update_template_entry) {
-         .descriptor_type = entry->descriptorType,
-         .descriptor_count = entry->descriptorCount,
-         .src_offset = entry->offset,
-         .src_stride = entry->stride,
-         .dst_offset = dst_offset,
-         .dst_stride = dst_stride,
-         .has_sampler = !binding_layout->immutable_samplers_offset,
-         .immutable_samplers = immutable_samplers,
-      };
-   }
-
-   assert(j == dst_entry_count);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    *pDescriptorUpdateTemplate =
       tu_descriptor_update_template_to_handle(templ);
 
+   tu_use_args(set_layout);
+   tu_stub();
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyDescriptorUpdateTemplate(
    VkDevice _device,
    VkDescriptorUpdateTemplate descriptorUpdateTemplate,
@@ -1408,90 +520,23 @@ tu_DestroyDescriptorUpdateTemplate(
    if (!templ)
       return;
 
-   vk_object_free(&device->vk, pAllocator, templ);
+   vk_free2(&device->alloc, pAllocator, templ);
 }
 
 void
 tu_update_descriptor_set_with_template(
-   const struct tu_device *device,
+   struct tu_device *device,
+   struct tu_cmd_buffer *cmd_buffer,
    struct tu_descriptor_set *set,
    VkDescriptorUpdateTemplate descriptorUpdateTemplate,
    const void *pData)
 {
    TU_FROM_HANDLE(tu_descriptor_update_template, templ,
                   descriptorUpdateTemplate);
-
-   for (uint32_t i = 0; i < templ->entry_count; i++) {
-      uint32_t *ptr = set->mapped_ptr;
-      const void *src = ((const char *) pData) + templ->entry[i].src_offset;
-      const struct tu_sampler *samplers = templ->entry[i].immutable_samplers;
-
-      if (templ->entry[i].descriptor_type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
-         memcpy(((uint8_t *) ptr) + templ->entry[i].dst_offset, src,
-                templ->entry[i].descriptor_count);
-         continue;
-      }
-
-      ptr += templ->entry[i].dst_offset;
-      unsigned dst_offset = templ->entry[i].dst_offset;
-      for (unsigned j = 0; j < templ->entry[i].descriptor_count; ++j) {
-         switch(templ->entry[i].descriptor_type) {
-         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC: {
-            assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
-            write_ubo_descriptor(set->dynamic_descriptors + dst_offset, src);
-            break;
-         }
-         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
-            write_ubo_descriptor(ptr, src);
-            break;
-         case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
-            assert(!(set->layout->flags & VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR));
-            write_buffer_descriptor(device, set->dynamic_descriptors + dst_offset, src);
-            break;
-         }
-         case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
-            write_buffer_descriptor(device, ptr, src);
-            break;
-         case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
-         case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
-            write_texel_buffer_descriptor(ptr, *(VkBufferView *) src);
-            break;
-         case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
-         case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: {
-            write_image_descriptor(ptr, templ->entry[i].descriptor_type,  src);
-            break;
-         }
-         case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
-            write_combined_image_sampler_descriptor(ptr,
-                                                    templ->entry[i].descriptor_type,
-                                                    src,
-                                                    templ->entry[i].has_sampler);
-            if (samplers)
-               write_sampler_push(ptr + A6XX_TEX_CONST_DWORDS, &samplers[j]);
-            break;
-         case VK_DESCRIPTOR_TYPE_SAMPLER:
-            if (templ->entry[i].has_sampler)
-               write_sampler_descriptor(ptr, src);
-            else if (samplers)
-               write_sampler_push(ptr, &samplers[j]);
-            break;
-         case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
-            /* nothing in descriptor set - framebuffer state is used instead */
-            if (unlikely(device->instance->debug_flags & TU_DEBUG_DYNAMIC))
-               write_image_descriptor(ptr, templ->entry[i].descriptor_type, src);
-            break;
-         default:
-            unreachable("unimplemented descriptor type");
-            break;
-         }
-         src = (char *) src + templ->entry[i].src_stride;
-         ptr += templ->entry[i].dst_stride;
-         dst_offset += templ->entry[i].dst_stride;
-      }
-   }
+   tu_use_args(templ);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_UpdateDescriptorSetWithTemplate(
    VkDevice _device,
    VkDescriptorSet descriptorSet,
@@ -1501,46 +546,25 @@ tu_UpdateDescriptorSetWithTemplate(
    TU_FROM_HANDLE(tu_device, device, _device);
    TU_FROM_HANDLE(tu_descriptor_set, set, descriptorSet);
 
-   tu_update_descriptor_set_with_template(device, set, descriptorUpdateTemplate, pData);
+   tu_update_descriptor_set_with_template(device, NULL, set,
+                                          descriptorUpdateTemplate, pData);
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateSamplerYcbcrConversion(
-   VkDevice _device,
+   VkDevice device,
    const VkSamplerYcbcrConversionCreateInfo *pCreateInfo,
    const VkAllocationCallbacks *pAllocator,
    VkSamplerYcbcrConversion *pYcbcrConversion)
 {
-   TU_FROM_HANDLE(tu_device, device, _device);
-   struct tu_sampler_ycbcr_conversion *conversion;
-
-   conversion = vk_object_alloc(&device->vk, pAllocator, sizeof(*conversion),
-                                VK_OBJECT_TYPE_SAMPLER_YCBCR_CONVERSION);
-   if (!conversion)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
-
-   conversion->format = pCreateInfo->format;
-   conversion->ycbcr_model = pCreateInfo->ycbcrModel;
-   conversion->ycbcr_range = pCreateInfo->ycbcrRange;
-   conversion->components = pCreateInfo->components;
-   conversion->chroma_offsets[0] = pCreateInfo->xChromaOffset;
-   conversion->chroma_offsets[1] = pCreateInfo->yChromaOffset;
-   conversion->chroma_filter = pCreateInfo->chromaFilter;
-
-   *pYcbcrConversion = tu_sampler_ycbcr_conversion_to_handle(conversion);
+   *pYcbcrConversion = VK_NULL_HANDLE;
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_DestroySamplerYcbcrConversion(VkDevice _device,
+void
+tu_DestroySamplerYcbcrConversion(VkDevice device,
                                  VkSamplerYcbcrConversion ycbcrConversion,
                                  const VkAllocationCallbacks *pAllocator)
 {
-   TU_FROM_HANDLE(tu_device, device, _device);
-   TU_FROM_HANDLE(tu_sampler_ycbcr_conversion, ycbcr_conversion, ycbcrConversion);
-
-   if (!ycbcr_conversion)
-      return;
-
-   vk_object_free(&device->vk, pAllocator, ycbcr_conversion);
+   /* Do nothing. */
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_device.c b/lib/mesa/src/freedreno/vulkan/tu_device.c
index 83f782635..901f02486 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_device.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_device.c
@@ -1,401 +1,358 @@
 /*
  * Copyright © 2016 Red Hat.
  * Copyright © 2016 Bas Nieuwenhuizen
- * SPDX-License-Identifier: MIT
  *
  * based in part on anv driver which is:
  * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_device.h"
+#include "tu_private.h"
 
 #include <fcntl.h>
-#include <poll.h>
+#include <libsync.h>
+#include <stdbool.h>
+#include <string.h>
+#include <sys/mman.h>
 #include <sys/sysinfo.h>
+#include <unistd.h>
+#include <xf86drm.h>
 
-#include "git_sha1.h"
-#include "util/u_debug.h"
+#include "compiler/glsl_types.h"
+#include "util/debug.h"
 #include "util/disk_cache.h"
-#include "util/driconf.h"
-#include "util/os_misc.h"
-#include "vk_shader_module.h"
-#include "vk_sampler.h"
+#include "vk_format.h"
 #include "vk_util.h"
 
-/* for fd_get_driver/device_uuid() */
-#include "freedreno/common/freedreno_uuid.h"
-
-#include "tu_clear_blit.h"
-#include "tu_cmd_buffer.h"
-#include "tu_cs.h"
-#include "tu_descriptor_set.h"
-#include "tu_dynamic_rendering.h"
-#include "tu_image.h"
-#include "tu_pass.h"
-#include "tu_query.h"
-#include "tu_tracepoints.h"
-#include "tu_wsi.h"
-
-#if defined(VK_USE_PLATFORM_WAYLAND_KHR) || \
-     defined(VK_USE_PLATFORM_XCB_KHR) || \
-     defined(VK_USE_PLATFORM_XLIB_KHR) || \
-     defined(VK_USE_PLATFORM_DISPLAY_KHR)
-#define TU_HAS_SURFACE 1
-#else
-#define TU_HAS_SURFACE 0
-#endif
-
+#include "drm-uapi/msm_drm.h"
 
 static int
-tu_device_get_cache_uuid(struct tu_physical_device *device, void *uuid)
-{
-   struct mesa_sha1 ctx;
-   unsigned char sha1[20];
-   /* Note: IR3_SHADER_DEBUG also affects compilation, but it's not
-    * initialized until after compiler creation so we have to add it to the
-    * shader hash instead, since the compiler is only created with the logical
-    * device.
-    */
-   uint64_t driver_flags = device->instance->debug_flags & TU_DEBUG_NOMULTIPOS;
-   uint16_t family = fd_dev_gpu_id(&device->dev_id);
-
+tu_device_get_cache_uuid(uint16_t family, void *uuid)
+{
+   uint32_t mesa_timestamp;
+   uint16_t f = family;
    memset(uuid, 0, VK_UUID_SIZE);
-   _mesa_sha1_init(&ctx);
-
-   if (!disk_cache_get_function_identifier(tu_device_get_cache_uuid, &ctx))
+   if (!disk_cache_get_function_timestamp(tu_device_get_cache_uuid,
+                                          &mesa_timestamp))
       return -1;
 
-   _mesa_sha1_update(&ctx, &family, sizeof(family));
-   _mesa_sha1_update(&ctx, &driver_flags, sizeof(driver_flags));
-   _mesa_sha1_final(&ctx, sha1);
-
-   memcpy(uuid, sha1, VK_UUID_SIZE);
+   memcpy(uuid, &mesa_timestamp, 4);
+   memcpy((char *) uuid + 4, &f, 2);
+   snprintf((char *) uuid + 6, VK_UUID_SIZE - 10, "tu");
    return 0;
 }
 
-#define TU_API_VERSION VK_MAKE_VERSION(1, 3, VK_HEADER_VERSION)
-
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_EnumerateInstanceVersion(uint32_t *pApiVersion)
-{
-    *pApiVersion = TU_API_VERSION;
-    return VK_SUCCESS;
-}
-
-static const struct vk_instance_extension_table tu_instance_extensions_supported = {
-   .KHR_device_group_creation           = true,
-   .KHR_external_fence_capabilities     = true,
-   .KHR_external_memory_capabilities    = true,
-   .KHR_external_semaphore_capabilities = true,
-   .KHR_get_physical_device_properties2 = true,
-   .KHR_surface                         = TU_HAS_SURFACE,
-   .KHR_get_surface_capabilities2       = TU_HAS_SURFACE,
-   .EXT_debug_report                    = true,
-   .EXT_debug_utils                     = true,
-#ifdef VK_USE_PLATFORM_WAYLAND_KHR
-   .KHR_wayland_surface                 = true,
-#endif
-#ifdef VK_USE_PLATFORM_XCB_KHR
-   .KHR_xcb_surface                     = true,
-#endif
-#ifdef VK_USE_PLATFORM_XLIB_KHR
-   .KHR_xlib_surface                    = true,
-#endif
-#ifdef VK_USE_PLATFORM_XLIB_XRANDR_EXT
-   .EXT_acquire_xlib_display            = true,
-#endif
-#ifdef VK_USE_PLATFORM_DISPLAY_KHR
-   .KHR_display                         = true,
-   .KHR_get_display_properties2         = true,
-   .EXT_direct_mode_display             = true,
-   .EXT_display_surface_counter         = true,
-   .EXT_acquire_drm_display             = true,
-#endif
-};
+static void
+tu_get_driver_uuid(void *uuid)
+{
+   memset(uuid, 0, VK_UUID_SIZE);
+   snprintf(uuid, VK_UUID_SIZE, "freedreno");
+}
 
 static void
-get_device_extensions(const struct tu_physical_device *device,
-                      struct vk_device_extension_table *ext)
-{
-   *ext = (struct vk_device_extension_table) {
-      .KHR_16bit_storage = device->info->a6xx.storage_16bit,
-      .KHR_bind_memory2 = true,
-      .KHR_copy_commands2 = true,
-      .KHR_create_renderpass2 = true,
-      .KHR_dedicated_allocation = true,
-      .KHR_depth_stencil_resolve = true,
-      .KHR_descriptor_update_template = true,
-      .KHR_device_group = true,
-      .KHR_draw_indirect_count = true,
-      .KHR_external_fence = true,
-      .KHR_external_fence_fd = true,
-      .KHR_external_memory = true,
-      .KHR_external_memory_fd = true,
-      .KHR_external_semaphore = true,
-      .KHR_external_semaphore_fd = true,
-      .KHR_format_feature_flags2 = true,
-      .KHR_get_memory_requirements2 = true,
-      .KHR_global_priority = true,
-      .KHR_imageless_framebuffer = true,
-      .KHR_incremental_present = TU_HAS_SURFACE,
-      .KHR_image_format_list = true,
-      .KHR_maintenance1 = true,
-      .KHR_maintenance2 = true,
-      .KHR_maintenance3 = true,
-      .KHR_maintenance4 = true,
-      .KHR_multiview = true,
-      .KHR_performance_query = device->instance->debug_flags & TU_DEBUG_PERFC,
-      .KHR_pipeline_executable_properties = true,
-      .KHR_push_descriptor = true,
-      .KHR_relaxed_block_layout = true,
-      .KHR_sampler_mirror_clamp_to_edge = true,
-      .KHR_sampler_ycbcr_conversion = true,
-      .KHR_shader_draw_parameters = true,
-      .KHR_shader_float_controls = true,
-      .KHR_shader_float16_int8 = true,
-      .KHR_shader_subgroup_extended_types = true,
-      .KHR_shader_terminate_invocation = true,
-      .KHR_spirv_1_4 = true,
-      .KHR_storage_buffer_storage_class = true,
-      .KHR_swapchain = TU_HAS_SURFACE,
-      .KHR_swapchain_mutable_format = TU_HAS_SURFACE,
-      .KHR_uniform_buffer_standard_layout = true,
-      .KHR_variable_pointers = true,
-      .KHR_vulkan_memory_model = true,
-      .KHR_driver_properties = true,
-      .KHR_separate_depth_stencil_layouts = true,
-      .KHR_buffer_device_address = true,
-      .KHR_shader_integer_dot_product = true,
-      .KHR_zero_initialize_workgroup_memory = true,
-      .KHR_shader_non_semantic_info = true,
-      .KHR_synchronization2 = true,
-      .KHR_dynamic_rendering = true,
-#ifndef TU_USE_KGSL
-      .KHR_timeline_semaphore = true,
-#endif
-#ifdef VK_USE_PLATFORM_DISPLAY_KHR
-      .EXT_display_control = true,
-#endif
-      .EXT_external_memory_dma_buf = true,
-      .EXT_image_drm_format_modifier = true,
-      .EXT_sample_locations = device->info->a6xx.has_sample_locations,
-      .EXT_sampler_filter_minmax = true,
-      .EXT_transform_feedback = true,
-      .EXT_4444_formats = true,
-      .EXT_border_color_swizzle = true,
-      .EXT_conditional_rendering = true,
-      .EXT_custom_border_color = true,
-      .EXT_depth_clip_control = true,
-      .EXT_depth_clip_enable = true,
-      .EXT_descriptor_indexing = true,
-      .EXT_extended_dynamic_state = true,
-      .EXT_extended_dynamic_state2 = true,
-      .EXT_filter_cubic = device->info->a6xx.has_tex_filter_cubic,
-      .EXT_global_priority = true,
-      .EXT_global_priority_query = true,
-      .EXT_host_query_reset = true,
-      .EXT_index_type_uint8 = true,
-      .EXT_memory_budget = true,
-      .EXT_primitive_topology_list_restart = true,
-      .EXT_private_data = true,
-      .EXT_queue_family_foreign = true,
-      .EXT_robustness2 = true,
-      .EXT_scalar_block_layout = true,
-      .EXT_separate_stencil_usage = true,
-      .EXT_shader_demote_to_helper_invocation = true,
-      .EXT_shader_stencil_export = true,
-      .EXT_shader_viewport_index_layer = true,
-      .EXT_shader_module_identifier = true,
-      .EXT_texel_buffer_alignment = true,
-      .EXT_vertex_attribute_divisor = true,
-      .EXT_provoking_vertex = true,
-      .EXT_line_rasterization = true,
-      .EXT_subgroup_size_control = true,
-      .EXT_image_robustness = true,
-      .EXT_primitives_generated_query = true,
-      .EXT_image_view_min_lod = true,
-      .EXT_pipeline_creation_feedback = true,
-      .EXT_pipeline_creation_cache_control = true,
-      .EXT_vertex_input_dynamic_state = true,
-      .EXT_attachment_feedback_loop_layout = true,
-      .EXT_rasterization_order_attachment_access = true,
-      .EXT_multi_draw = true,
-#ifndef TU_USE_KGSL
-      .EXT_physical_device_drm = true,
-#endif
-      /* For Graphics Flight Recorder (GFR) */
-      .AMD_buffer_marker = true,
-      .ARM_rasterization_order_attachment_access = true,
-#ifdef ANDROID
-      .ANDROID_native_buffer = true,
-#endif
-      .IMG_filter_cubic = device->info->a6xx.has_tex_filter_cubic,
-      .VALVE_mutable_descriptor_type = true,
-      .EXT_image_2d_view_of_3d = true,
-      .EXT_color_write_enable = true,
-      .EXT_load_store_op_none = true,
-      .EXT_non_seamless_cube_map = true,
-      .EXT_tooling_info = true,
-      .EXT_inline_uniform_block = true,
-      .EXT_mutable_descriptor_type = true,
-      .KHR_pipeline_library = true,
-      .EXT_graphics_pipeline_library = true,
+tu_get_device_uuid(void *uuid)
+{
+   memset(uuid, 0, VK_UUID_SIZE);
+}
+
+static VkResult
+tu_bo_init(struct tu_device *dev,
+           struct tu_bo *bo,
+           uint32_t gem_handle,
+           uint64_t size)
+{
+   uint64_t iova = tu_gem_info_iova(dev, gem_handle);
+   if (!iova)
+      return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+
+   *bo = (struct tu_bo) {
+      .gem_handle = gem_handle,
+      .size = size,
+      .iova = iova,
    };
+
+   return VK_SUCCESS;
 }
 
-static const struct vk_pipeline_cache_object_ops *const cache_import_ops[] = {
-   &tu_shaders_ops,
-   &tu_nir_shaders_ops,
-   NULL,
-};
+VkResult
+tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size)
+{
+   /* TODO: Choose better flags. As of 2018-11-12, freedreno/drm/msm_bo.c
+    * always sets `flags = MSM_BO_WC`, and we copy that behavior here.
+    */
+   uint32_t gem_handle = tu_gem_new(dev, size, MSM_BO_WC);
+   if (!gem_handle)
+      return vk_error(dev->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
+
+   VkResult result = tu_bo_init(dev, bo, gem_handle, size);
+   if (result != VK_SUCCESS) {
+      tu_gem_close(dev, gem_handle);
+      return vk_error(dev->instance, result);
+   }
+
+   return VK_SUCCESS;
+}
+
+VkResult
+tu_bo_init_dmabuf(struct tu_device *dev,
+                  struct tu_bo *bo,
+                  uint64_t size,
+                  int fd)
+{
+   uint32_t gem_handle = tu_gem_import_dmabuf(dev, fd, size);
+   if (!gem_handle)
+      return vk_error(dev->instance, VK_ERROR_INVALID_EXTERNAL_HANDLE);
+
+   VkResult result = tu_bo_init(dev, bo, gem_handle, size);
+   if (result != VK_SUCCESS) {
+      tu_gem_close(dev, gem_handle);
+      return vk_error(dev->instance, result);
+   }
+
+   return VK_SUCCESS;
+}
+
+int
+tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo)
+{
+   return tu_gem_export_dmabuf(dev, bo->gem_handle);
+}
 
 VkResult
+tu_bo_map(struct tu_device *dev, struct tu_bo *bo)
+{
+   if (bo->map)
+      return VK_SUCCESS;
+
+   uint64_t offset = tu_gem_info_offset(dev, bo->gem_handle);
+   if (!offset)
+      return vk_error(dev->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
+
+   /* TODO: Should we use the wrapper os_mmap() like Freedreno does? */
+   void *map = mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
+                    dev->physical_device->local_fd, offset);
+   if (map == MAP_FAILED)
+      return vk_error(dev->instance, VK_ERROR_MEMORY_MAP_FAILED);
+
+   bo->map = map;
+   return VK_SUCCESS;
+}
+
+void
+tu_bo_finish(struct tu_device *dev, struct tu_bo *bo)
+{
+   assert(bo->gem_handle);
+
+   if (bo->map)
+      munmap(bo->map, bo->size);
+
+   tu_gem_close(dev, bo->gem_handle);
+}
+
+static VkResult
 tu_physical_device_init(struct tu_physical_device *device,
-                        struct tu_instance *instance)
+                        struct tu_instance *instance,
+                        drmDevicePtr drm_device)
 {
+   const char *path = drm_device->nodes[DRM_NODE_RENDER];
    VkResult result = VK_SUCCESS;
+   drmVersionPtr version;
+   int fd;
+   int master_fd = -1;
+
+   fd = open(path, O_RDWR | O_CLOEXEC);
+   if (fd < 0) {
+      return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
+                       "failed to open device %s", path);
+   }
+
+   /* Version 1.3 added MSM_INFO_IOVA. */
+   const int min_version_major = 1;
+   const int min_version_minor = 3;
+
+   version = drmGetVersion(fd);
+   if (!version) {
+      close(fd);
+      return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
+                       "failed to query kernel driver version for device %s",
+                       path);
+   }
+
+   if (strcmp(version->name, "msm")) {
+      drmFreeVersion(version);
+      close(fd);
+      return vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
+                       "device %s does not use the msm kernel driver", path);
+   }
+
+   if (version->version_major != min_version_major ||
+       version->version_minor < min_version_minor) {
+      result = vk_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
+                         "kernel driver for device %s has version %d.%d, "
+                         "but Vulkan requires version >= %d.%d",
+                         path, version->version_major, version->version_minor,
+                         min_version_major, min_version_minor);
+      drmFreeVersion(version);
+      close(fd);
+      return result;
+   }
 
-   const char *fd_name = fd_dev_name(&device->dev_id);
-   if (strncmp(fd_name, "FD", 2) == 0) {
-      device->name = vk_asprintf(&instance->vk.alloc,
-                                 VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE,
-                                 "Turnip Adreno (TM) %s", &fd_name[2]);
-   } else {
-      device->name = vk_strdup(&instance->vk.alloc, fd_name,
-                               VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
+   drmFreeVersion(version);
+
+   if (instance->debug_flags & TU_DEBUG_STARTUP)
+      tu_logi("Found compatible device '%s'.", path);
 
+   device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
+   device->instance = instance;
+   assert(strlen(path) < ARRAY_SIZE(device->path));
+   strncpy(device->path, path, ARRAY_SIZE(device->path));
+
+   if (instance->enabled_extensions.KHR_display) {
+      master_fd =
+         open(drm_device->nodes[DRM_NODE_PRIMARY], O_RDWR | O_CLOEXEC);
+      if (master_fd >= 0) {
+         /* TODO: free master_fd is accel is not working? */
+      }
    }
-   if (!device->name) {
-      return vk_startup_errorf(instance, VK_ERROR_OUT_OF_HOST_MEMORY,
-                               "device name alloc fail");
+
+   device->master_fd = master_fd;
+   device->local_fd = fd;
+
+   if (tu_drm_get_gpu_id(device, &device->gpu_id)) {
+      if (instance->debug_flags & TU_DEBUG_STARTUP)
+         tu_logi("Could not query the GPU ID");
+      result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
+                         "could not get GPU ID");
+      goto fail;
    }
 
-   const struct fd_dev_info *info = fd_dev_info(&device->dev_id);
-   if (!info) {
-      result = vk_startup_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                                 "device %s is unsupported", device->name);
-      goto fail_free_name;
+   if (tu_drm_get_gmem_size(device, &device->gmem_size)) {
+      if (instance->debug_flags & TU_DEBUG_STARTUP)
+         tu_logi("Could not query the GMEM size");
+      result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
+                         "could not get GMEM size");
+      goto fail;
    }
-   switch (fd_dev_gen(&device->dev_id)) {
-   case 6:
-      device->info = info;
-      device->ccu_offset_bypass = device->info->num_ccu * A6XX_CCU_DEPTH_SIZE;
-      device->ccu_offset_gmem = (device->gmem_size -
-         device->info->num_ccu * A6XX_CCU_GMEM_COLOR_SIZE);
+
+   memset(device->name, 0, sizeof(device->name));
+   sprintf(device->name, "FD%d", device->gpu_id);
+
+   switch (device->gpu_id) {
+   case 630:
+      device->tile_align_w = 32;
+      device->tile_align_h = 32;
       break;
    default:
-      result = vk_startup_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                                 "device %s is unsupported", device->name);
-      goto fail_free_name;
-   }
-   if (tu_device_get_cache_uuid(device, device->cache_uuid)) {
-      result = vk_startup_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                                 "cannot generate UUID");
-      goto fail_free_name;
+      result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
+                         "device %s is unsupported", device->name);
+      goto fail;
    }
-
-   if (device->has_set_iova) {
-      mtx_init(&device->vma_mutex, mtx_plain);
-      util_vma_heap_init(&device->vma, device->va_start,
-                         ROUND_DOWN_TO(device->va_size, 4096));
+   if (tu_device_get_cache_uuid(device->gpu_id, device->cache_uuid)) {
+      result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
+                         "cannot generate UUID");
+      goto fail;
    }
 
-   fd_get_driver_uuid(device->driver_uuid);
-   fd_get_device_uuid(device->device_uuid, &device->dev_id);
+   /* The gpu id is already embedded in the uuid so we just pass "tu"
+    * when creating the cache.
+    */
+   char buf[VK_UUID_SIZE * 2 + 1];
+   disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
+   device->disk_cache = disk_cache_create(device->name, buf, 0);
 
-   struct vk_device_extension_table supported_extensions;
-   get_device_extensions(device, &supported_extensions);
+   fprintf(stderr, "WARNING: tu is not a conformant vulkan implementation, "
+                   "testing use only.\n");
 
-   struct vk_physical_device_dispatch_table dispatch_table;
-   vk_physical_device_dispatch_table_from_entrypoints(
-      &dispatch_table, &tu_physical_device_entrypoints, true);
-   vk_physical_device_dispatch_table_from_entrypoints(
-      &dispatch_table, &wsi_physical_device_entrypoints, false);
+   tu_get_driver_uuid(&device->device_uuid);
+   tu_get_device_uuid(&device->device_uuid);
 
-   result = vk_physical_device_init(&device->vk, &instance->vk,
-                                    &supported_extensions,
-                                    &dispatch_table);
-   if (result != VK_SUCCESS)
-      goto fail_free_vma;
+   tu_fill_device_extension_table(device, &device->supported_extensions);
 
-   device->vk.supported_sync_types = device->sync_types;
+   if (result != VK_SUCCESS) {
+      vk_error(instance, result);
+      goto fail;
+   }
 
-#if TU_HAS_SURFACE
    result = tu_wsi_init(device);
    if (result != VK_SUCCESS) {
-      vk_startup_errorf(instance, result, "WSI init failure");
-      vk_physical_device_finish(&device->vk);
-      goto fail_free_vma;
+      vk_error(instance, result);
+      goto fail;
    }
-#endif
-
-   /* The gpu id is already embedded in the uuid so we just pass "tu"
-    * when creating the cache.
-    */
-   char buf[VK_UUID_SIZE * 2 + 1];
-   disk_cache_format_hex_id(buf, device->cache_uuid, VK_UUID_SIZE * 2);
-   device->vk.disk_cache = disk_cache_create(device->name, buf, 0);
-
-   device->vk.pipeline_cache_import_ops = cache_import_ops;
 
    return VK_SUCCESS;
 
-fail_free_vma:
-   if (device->has_set_iova)
-      util_vma_heap_finish(&device->vma);
-fail_free_name:
-   vk_free(&instance->vk.alloc, (void *)device->name);
+fail:
+   close(fd);
+   if (master_fd != -1)
+      close(master_fd);
    return result;
 }
 
 static void
 tu_physical_device_finish(struct tu_physical_device *device)
 {
-#if TU_HAS_SURFACE
    tu_wsi_finish(device);
-#endif
 
+   disk_cache_destroy(device->disk_cache);
    close(device->local_fd);
    if (device->master_fd != -1)
       close(device->master_fd);
+}
 
-   if (device->has_set_iova)
-      util_vma_heap_finish(&device->vma);
-
-   vk_free(&device->instance->vk.alloc, (void *)device->name);
+static void *
+default_alloc_func(void *pUserData,
+                   size_t size,
+                   size_t align,
+                   VkSystemAllocationScope allocationScope)
+{
+   return malloc(size);
+}
 
-   vk_physical_device_finish(&device->vk);
+static void *
+default_realloc_func(void *pUserData,
+                     void *pOriginal,
+                     size_t size,
+                     size_t align,
+                     VkSystemAllocationScope allocationScope)
+{
+   return realloc(pOriginal, size);
 }
 
 static void
-tu_destroy_physical_device(struct vk_physical_device *device)
+default_free_func(void *pUserData, void *pMemory)
 {
-   tu_physical_device_finish((struct tu_physical_device *) device);
-   vk_free(&device->instance->alloc, device);
+   free(pMemory);
 }
 
+static const VkAllocationCallbacks default_alloc = {
+   .pUserData = NULL,
+   .pfnAllocation = default_alloc_func,
+   .pfnReallocation = default_realloc_func,
+   .pfnFree = default_free_func,
+};
+
 static const struct debug_control tu_debug_options[] = {
    { "startup", TU_DEBUG_STARTUP },
    { "nir", TU_DEBUG_NIR },
-   { "nobin", TU_DEBUG_NOBIN },
-   { "sysmem", TU_DEBUG_SYSMEM },
-   { "gmem", TU_DEBUG_GMEM },
-   { "forcebin", TU_DEBUG_FORCEBIN },
-   { "layout", TU_DEBUG_LAYOUT },
-   { "noubwc", TU_DEBUG_NOUBWC },
-   { "nomultipos", TU_DEBUG_NOMULTIPOS },
-   { "nolrz", TU_DEBUG_NOLRZ },
-   { "nolrzfc", TU_DEBUG_NOLRZFC },
-   { "perf", TU_DEBUG_PERF },
-   { "perfc", TU_DEBUG_PERFC },
-   { "flushall", TU_DEBUG_FLUSHALL },
-   { "syncdraw", TU_DEBUG_SYNCDRAW },
-   { "dontcare_as_load", TU_DEBUG_DONT_CARE_AS_LOAD },
-   { "rast_order", TU_DEBUG_RAST_ORDER },
-   { "unaligned_store", TU_DEBUG_UNALIGNED_STORE },
-   { "log_skip_gmem_ops", TU_DEBUG_LOG_SKIP_GMEM_OPS },
-   { "dynamic", TU_DEBUG_DYNAMIC },
-   { "bos", TU_DEBUG_BOS },
+   { "ir3", TU_DEBUG_IR3 },
    { NULL, 0 }
 };
 
@@ -406,34 +363,17 @@ tu_get_debug_option_name(int id)
    return tu_debug_options[id].string;
 }
 
-static const driOptionDescription tu_dri_options[] = {
-   DRI_CONF_SECTION_PERFORMANCE
-      DRI_CONF_VK_X11_OVERRIDE_MIN_IMAGE_COUNT(0)
-      DRI_CONF_VK_X11_STRICT_IMAGE_COUNT(false)
-      DRI_CONF_VK_X11_ENSURE_MIN_IMAGE_COUNT(false)
-      DRI_CONF_VK_XWAYLAND_WAIT_READY(true)
-   DRI_CONF_SECTION_END
-
-   DRI_CONF_SECTION_DEBUG
-      DRI_CONF_VK_WSI_FORCE_BGRA8_UNORM_FIRST(false)
-      DRI_CONF_VK_DONT_CARE_AS_LOAD(false)
-   DRI_CONF_SECTION_END
-};
-
-static void
-tu_init_dri_options(struct tu_instance *instance)
+static int
+tu_get_instance_extension_index(const char *name)
 {
-   driParseOptionInfo(&instance->available_dri_options, tu_dri_options,
-                      ARRAY_SIZE(tu_dri_options));
-   driParseConfigFiles(&instance->dri_options, &instance->available_dri_options, 0, "turnip", NULL, NULL,
-                       instance->vk.app_info.app_name, instance->vk.app_info.app_version,
-                       instance->vk.app_info.engine_name, instance->vk.app_info.engine_version);
-
-   if (driQueryOptionb(&instance->dri_options, "vk_dont_care_as_load"))
-      instance->debug_flags |= TU_DEBUG_DONT_CARE_AS_LOAD;
+   for (unsigned i = 0; i < TU_INSTANCE_EXTENSION_COUNT; ++i) {
+      if (strcmp(name, tu_instance_extensions[i].extensionName) == 0)
+         return i;
+   }
+   return -1;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateInstance(const VkInstanceCreateInfo *pCreateInfo,
                   const VkAllocationCallbacks *pAllocator,
                   VkInstance *pInstance)
@@ -443,66 +383,63 @@ tu_CreateInstance(const VkInstanceCreateInfo *pCreateInfo,
 
    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO);
 
-   if (pAllocator == NULL)
-      pAllocator = vk_default_allocator();
-
-   instance = vk_zalloc(pAllocator, sizeof(*instance), 8,
-                        VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
+   uint32_t client_version;
+   if (pCreateInfo->pApplicationInfo &&
+       pCreateInfo->pApplicationInfo->apiVersion != 0) {
+      client_version = pCreateInfo->pApplicationInfo->apiVersion;
+   } else {
+      tu_EnumerateInstanceVersion(&client_version);
+   }
 
+   instance = vk_zalloc2(&default_alloc, pAllocator, sizeof(*instance), 8,
+                         VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
    if (!instance)
       return vk_error(NULL, VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   struct vk_instance_dispatch_table dispatch_table;
-   vk_instance_dispatch_table_from_entrypoints(
-      &dispatch_table, &tu_instance_entrypoints, true);
-   vk_instance_dispatch_table_from_entrypoints(
-      &dispatch_table, &wsi_instance_entrypoints, false);
+   instance->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
 
-   result = vk_instance_init(&instance->vk,
-                             &tu_instance_extensions_supported,
-                             &dispatch_table,
-                             pCreateInfo, pAllocator);
-   if (result != VK_SUCCESS) {
-      vk_free(pAllocator, instance);
-      return vk_error(NULL, result);
-   }
+   if (pAllocator)
+      instance->alloc = *pAllocator;
+   else
+      instance->alloc = default_alloc;
 
-#ifndef TU_USE_KGSL
-   instance->vk.physical_devices.try_create_for_drm =
-      tu_physical_device_try_create;
-#else
-   instance->vk.physical_devices.enumerate = tu_enumerate_devices;
-#endif
-   instance->vk.physical_devices.destroy = tu_destroy_physical_device;
+   instance->api_version = client_version;
+   instance->physical_device_count = -1;
 
    instance->debug_flags =
-      parse_debug_string(os_get_option("TU_DEBUG"), tu_debug_options);
-
-#ifdef DEBUG
-   /* Enable startup debugging by default on debug drivers.  You almost always
-    * want to see your startup failures in that case, and it's hard to set
-    * this env var on android.
-    */
-   instance->debug_flags |= TU_DEBUG_STARTUP;
-#endif
+      parse_debug_string(getenv("TU_DEBUG"), tu_debug_options);
 
    if (instance->debug_flags & TU_DEBUG_STARTUP)
-      mesa_logi("Created an instance");
+      tu_logi("Created an instance");
 
-   VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
+   for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
+      const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
+      int index = tu_get_instance_extension_index(ext_name);
+
+      if (index < 0 || !tu_supported_instance_extensions.extensions[index]) {
+         vk_free2(&default_alloc, pAllocator, instance);
+         return vk_error(instance, VK_ERROR_EXTENSION_NOT_PRESENT);
+      }
 
-   tu_init_dri_options(instance);
+      instance->enabled_extensions.extensions[index] = true;
+   }
 
-   *pInstance = tu_instance_to_handle(instance);
+   result = vk_debug_report_instance_init(&instance->debug_report_callbacks);
+   if (result != VK_SUCCESS) {
+      vk_free2(&default_alloc, pAllocator, instance);
+      return vk_error(instance, result);
+   }
+
+   glsl_type_singleton_init_or_ref();
 
-#ifdef HAVE_PERFETTO
-   tu_perfetto_init();
-#endif
+   VG(VALGRIND_CREATE_MEMPOOL(instance, 0, false));
+
+   *pInstance = tu_instance_to_handle(instance);
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyInstance(VkInstance _instance,
                    const VkAllocationCallbacks *pAllocator)
 {
@@ -511,623 +448,272 @@ tu_DestroyInstance(VkInstance _instance,
    if (!instance)
       return;
 
+   for (int i = 0; i < instance->physical_device_count; ++i) {
+      tu_physical_device_finish(instance->physical_devices + i);
+   }
+
    VG(VALGRIND_DESTROY_MEMPOOL(instance));
 
-   driDestroyOptionCache(&instance->dri_options);
-   driDestroyOptionInfo(&instance->available_dri_options);
+   glsl_type_singleton_decref();
 
-   vk_instance_finish(&instance->vk);
-   vk_free(&instance->vk.alloc, instance);
-}
+   vk_debug_report_instance_destroy(&instance->debug_report_callbacks);
 
-static void
-tu_get_physical_device_features_1_1(struct tu_physical_device *pdevice,
-                                    VkPhysicalDeviceVulkan11Features *features)
-{
-   features->storageBuffer16BitAccess            = pdevice->info->a6xx.storage_16bit;
-   features->uniformAndStorageBuffer16BitAccess  = false;
-   features->storagePushConstant16               = false;
-   features->storageInputOutput16                = false;
-   features->multiview                           = true;
-   features->multiviewGeometryShader             = false;
-   features->multiviewTessellationShader         = false;
-   features->variablePointersStorageBuffer       = true;
-   features->variablePointers                    = true;
-   features->protectedMemory                     = false;
-   features->samplerYcbcrConversion              = true;
-   features->shaderDrawParameters                = true;
+   vk_free(&instance->alloc, instance);
 }
 
-static void
-tu_get_physical_device_features_1_2(struct tu_physical_device *pdevice,
-                                    VkPhysicalDeviceVulkan12Features *features)
-{
-   features->samplerMirrorClampToEdge            = true;
-   features->drawIndirectCount                   = true;
-   features->storageBuffer8BitAccess             = false;
-   features->uniformAndStorageBuffer8BitAccess   = false;
-   features->storagePushConstant8                = false;
-   features->shaderBufferInt64Atomics            = false;
-   features->shaderSharedInt64Atomics            = false;
-   features->shaderFloat16                       = true;
-   features->shaderInt8                          = false;
-
-   features->descriptorIndexing                                 = true;
-   features->shaderInputAttachmentArrayDynamicIndexing          = false;
-   features->shaderUniformTexelBufferArrayDynamicIndexing       = true;
-   features->shaderStorageTexelBufferArrayDynamicIndexing       = true;
-   features->shaderUniformBufferArrayNonUniformIndexing         = true;
-   features->shaderSampledImageArrayNonUniformIndexing          = true;
-   features->shaderStorageBufferArrayNonUniformIndexing         = true;
-   features->shaderStorageImageArrayNonUniformIndexing          = true;
-   features->shaderInputAttachmentArrayNonUniformIndexing       = false;
-   features->shaderUniformTexelBufferArrayNonUniformIndexing    = true;
-   features->shaderStorageTexelBufferArrayNonUniformIndexing    = true;
-   features->descriptorBindingUniformBufferUpdateAfterBind      = true;
-   features->descriptorBindingSampledImageUpdateAfterBind       = true;
-   features->descriptorBindingStorageImageUpdateAfterBind       = true;
-   features->descriptorBindingStorageBufferUpdateAfterBind      = true;
-   features->descriptorBindingUniformTexelBufferUpdateAfterBind = true;
-   features->descriptorBindingStorageTexelBufferUpdateAfterBind = true;
-   features->descriptorBindingUpdateUnusedWhilePending          = true;
-   features->descriptorBindingPartiallyBound                    = true;
-   features->descriptorBindingVariableDescriptorCount           = true;
-   features->runtimeDescriptorArray                             = true;
-
-   features->samplerFilterMinmax                 = true;
-   features->scalarBlockLayout                   = true;
-   features->imagelessFramebuffer                = true;
-   features->uniformBufferStandardLayout         = true;
-   features->shaderSubgroupExtendedTypes         = true;
-   features->separateDepthStencilLayouts         = true;
-   features->hostQueryReset                      = true;
-   features->timelineSemaphore                   = true;
-   features->bufferDeviceAddress                 = true;
-   features->bufferDeviceAddressCaptureReplay    = pdevice->has_set_iova;
-   features->bufferDeviceAddressMultiDevice      = false;
-   features->vulkanMemoryModel                   = true;
-   features->vulkanMemoryModelDeviceScope        = true;
-   features->vulkanMemoryModelAvailabilityVisibilityChains = true;
-   features->shaderOutputViewportIndex           = true;
-   features->shaderOutputLayer                   = true;
-   features->subgroupBroadcastDynamicId          = true;
+static VkResult
+tu_enumerate_devices(struct tu_instance *instance)
+{
+   /* TODO: Check for more devices ? */
+   drmDevicePtr devices[8];
+   VkResult result = VK_ERROR_INCOMPATIBLE_DRIVER;
+   int max_devices;
+
+   instance->physical_device_count = 0;
+
+   max_devices = drmGetDevices2(0, devices, ARRAY_SIZE(devices));
+
+   if (instance->debug_flags & TU_DEBUG_STARTUP)
+      tu_logi("Found %d drm nodes", max_devices);
+
+   if (max_devices < 1)
+      return vk_error(instance, VK_ERROR_INCOMPATIBLE_DRIVER);
+
+   for (unsigned i = 0; i < (unsigned) max_devices; i++) {
+      if (devices[i]->available_nodes & 1 << DRM_NODE_RENDER &&
+          devices[i]->bustype == DRM_BUS_PLATFORM) {
+
+         result = tu_physical_device_init(
+            instance->physical_devices + instance->physical_device_count,
+            instance, devices[i]);
+         if (result == VK_SUCCESS)
+            ++instance->physical_device_count;
+         else if (result != VK_ERROR_INCOMPATIBLE_DRIVER)
+            break;
+      }
+   }
+   drmFreeDevices(devices, max_devices);
+
+   return result;
 }
 
-static void
-tu_get_physical_device_features_1_3(struct tu_physical_device *pdevice,
-                                    VkPhysicalDeviceVulkan13Features *features)
-{
-   features->robustImageAccess                   = true;
-   features->inlineUniformBlock                  = true;
-   features->descriptorBindingInlineUniformBlockUpdateAfterBind = true;
-   features->pipelineCreationCacheControl        = true;
-   features->privateData                         = true;
-   features->shaderDemoteToHelperInvocation      = true;
-   features->shaderTerminateInvocation           = true;
-   features->subgroupSizeControl                 = true;
-   features->computeFullSubgroups                = true;
-   features->synchronization2                    = true;
-   features->textureCompressionASTC_HDR          = false;
-   features->shaderZeroInitializeWorkgroupMemory = true;
-   features->dynamicRendering                    = true;
-   features->shaderIntegerDotProduct             = true;
-   features->maintenance4                        = true;
+VkResult
+tu_EnumeratePhysicalDevices(VkInstance _instance,
+                            uint32_t *pPhysicalDeviceCount,
+                            VkPhysicalDevice *pPhysicalDevices)
+{
+   TU_FROM_HANDLE(tu_instance, instance, _instance);
+   VK_OUTARRAY_MAKE(out, pPhysicalDevices, pPhysicalDeviceCount);
+
+   VkResult result;
+
+   if (instance->physical_device_count < 0) {
+      result = tu_enumerate_devices(instance);
+      if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
+         return result;
+   }
+
+   for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
+      vk_outarray_append(&out, p)
+      {
+         *p = tu_physical_device_to_handle(instance->physical_devices + i);
+      }
+   }
+
+   return vk_outarray_status(&out);
 }
 
-void
-tu_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
-                              VkPhysicalDeviceFeatures2 *pFeatures)
+VkResult
+tu_EnumeratePhysicalDeviceGroups(
+   VkInstance _instance,
+   uint32_t *pPhysicalDeviceGroupCount,
+   VkPhysicalDeviceGroupProperties *pPhysicalDeviceGroupProperties)
 {
-   TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
+   TU_FROM_HANDLE(tu_instance, instance, _instance);
+   VK_OUTARRAY_MAKE(out, pPhysicalDeviceGroupProperties,
+                    pPhysicalDeviceGroupCount);
+   VkResult result;
+
+   if (instance->physical_device_count < 0) {
+      result = tu_enumerate_devices(instance);
+      if (result != VK_SUCCESS && result != VK_ERROR_INCOMPATIBLE_DRIVER)
+         return result;
+   }
 
-   pFeatures->features = (VkPhysicalDeviceFeatures) {
-      .robustBufferAccess = true,
-      .fullDrawIndexUint32 = true,
-      .imageCubeArray = true,
-      .independentBlend = true,
-      .geometryShader = true,
-      .tessellationShader = true,
-      .sampleRateShading = true,
-      .dualSrcBlend = true,
-      .logicOp = true,
-      .multiDrawIndirect = true,
-      .drawIndirectFirstInstance = true,
-      .depthClamp = true,
-      .depthBiasClamp = true,
-      .fillModeNonSolid = true,
-      .depthBounds = true,
+   for (uint32_t i = 0; i < instance->physical_device_count; ++i) {
+      vk_outarray_append(&out, p)
+      {
+         p->physicalDeviceCount = 1;
+         p->physicalDevices[0] =
+            tu_physical_device_to_handle(instance->physical_devices + i);
+         p->subsetAllocation = false;
+      }
+   }
+
+   return vk_outarray_status(&out);
+}
+
+void
+tu_GetPhysicalDeviceFeatures(VkPhysicalDevice physicalDevice,
+                             VkPhysicalDeviceFeatures *pFeatures)
+{
+   memset(pFeatures, 0, sizeof(*pFeatures));
+
+   *pFeatures = (VkPhysicalDeviceFeatures) {
+      .robustBufferAccess = false,
+      .fullDrawIndexUint32 = false,
+      .imageCubeArray = false,
+      .independentBlend = false,
+      .geometryShader = false,
+      .tessellationShader = false,
+      .sampleRateShading = false,
+      .dualSrcBlend = false,
+      .logicOp = false,
+      .multiDrawIndirect = false,
+      .drawIndirectFirstInstance = false,
+      .depthClamp = false,
+      .depthBiasClamp = false,
+      .fillModeNonSolid = false,
+      .depthBounds = false,
       .wideLines = false,
-      .largePoints = true,
-      .alphaToOne = true,
-      .multiViewport = true,
-      .samplerAnisotropy = true,
-      .textureCompressionETC2 = true,
-      .textureCompressionASTC_LDR = true,
-      .textureCompressionBC = true,
-      .occlusionQueryPrecise = true,
-      .pipelineStatisticsQuery = true,
-      .vertexPipelineStoresAndAtomics = true,
-      .fragmentStoresAndAtomics = true,
-      .shaderTessellationAndGeometryPointSize = true,
-      .shaderImageGatherExtended = true,
-      .shaderStorageImageExtendedFormats = true,
+      .largePoints = false,
+      .alphaToOne = false,
+      .multiViewport = false,
+      .samplerAnisotropy = false,
+      .textureCompressionETC2 = false,
+      .textureCompressionASTC_LDR = false,
+      .textureCompressionBC = false,
+      .occlusionQueryPrecise = false,
+      .pipelineStatisticsQuery = false,
+      .vertexPipelineStoresAndAtomics = false,
+      .fragmentStoresAndAtomics = false,
+      .shaderTessellationAndGeometryPointSize = false,
+      .shaderImageGatherExtended = false,
+      .shaderStorageImageExtendedFormats = false,
       .shaderStorageImageMultisample = false,
-      .shaderUniformBufferArrayDynamicIndexing = true,
-      .shaderSampledImageArrayDynamicIndexing = true,
-      .shaderStorageBufferArrayDynamicIndexing = true,
-      .shaderStorageImageArrayDynamicIndexing = true,
-      .shaderStorageImageReadWithoutFormat = true,
-      .shaderStorageImageWriteWithoutFormat = true,
-      .shaderClipDistance = true,
-      .shaderCullDistance = true,
+      .shaderUniformBufferArrayDynamicIndexing = false,
+      .shaderSampledImageArrayDynamicIndexing = false,
+      .shaderStorageBufferArrayDynamicIndexing = false,
+      .shaderStorageImageArrayDynamicIndexing = false,
+      .shaderStorageImageReadWithoutFormat = false,
+      .shaderStorageImageWriteWithoutFormat = false,
+      .shaderClipDistance = false,
+      .shaderCullDistance = false,
       .shaderFloat64 = false,
       .shaderInt64 = false,
-      .shaderInt16 = true,
+      .shaderInt16 = false,
       .sparseBinding = false,
-      .variableMultisampleRate = true,
-      .inheritedQueries = true,
+      .variableMultisampleRate = false,
+      .inheritedQueries = false,
    };
+}
 
-   VkPhysicalDeviceVulkan11Features core_1_1 = {
-      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
-   };
-   tu_get_physical_device_features_1_1(pdevice, &core_1_1);
-
-   VkPhysicalDeviceVulkan12Features core_1_2 = {
-      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
-   };
-   tu_get_physical_device_features_1_2(pdevice, &core_1_2);
-
-   VkPhysicalDeviceVulkan13Features core_1_3 = {
-      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_FEATURES,
-   };
-   tu_get_physical_device_features_1_3(pdevice, &core_1_3);
-
+void
+tu_GetPhysicalDeviceFeatures2(VkPhysicalDevice physicalDevice,
+                              VkPhysicalDeviceFeatures2 *pFeatures)
+{
    vk_foreach_struct(ext, pFeatures->pNext)
    {
-      if (vk_get_physical_device_core_1_1_feature_ext(ext, &core_1_1))
-         continue;
-      if (vk_get_physical_device_core_1_2_feature_ext(ext, &core_1_2))
-         continue;
-      if (vk_get_physical_device_core_1_3_feature_ext(ext, &core_1_3))
-         continue;
-
       switch (ext->sType) {
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
-         VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
-            (VkPhysicalDeviceConditionalRenderingFeaturesEXT *) ext;
-         features->conditionalRendering = true;
-         features->inheritedConditionalRendering = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_FEATURES_EXT: {
-         VkPhysicalDeviceTransformFeedbackFeaturesEXT *features =
-            (VkPhysicalDeviceTransformFeedbackFeaturesEXT *) ext;
-         features->transformFeedback = true;
-         features->geometryStreams = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES_EXT: {
-         VkPhysicalDeviceIndexTypeUint8FeaturesEXT *features =
-            (VkPhysicalDeviceIndexTypeUint8FeaturesEXT *)ext;
-         features->indexTypeUint8 = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_FEATURES_EXT: {
-         VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *features =
-            (VkPhysicalDeviceVertexAttributeDivisorFeaturesEXT *)ext;
-         features->vertexAttributeInstanceRateDivisor = true;
-         features->vertexAttributeInstanceRateZeroDivisor = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_ENABLE_FEATURES_EXT: {
-         VkPhysicalDeviceDepthClipEnableFeaturesEXT *features =
-            (VkPhysicalDeviceDepthClipEnableFeaturesEXT *)ext;
-         features->depthClipEnable = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_4444_FORMATS_FEATURES_EXT: {
-         VkPhysicalDevice4444FormatsFeaturesEXT *features = (void *)ext;
-         features->formatA4R4G4B4 = true;
-         features->formatA4B4G4R4 = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_BORDER_COLOR_SWIZZLE_FEATURES_EXT: {
-         VkPhysicalDeviceBorderColorSwizzleFeaturesEXT *features = (void *)ext;
-         features->borderColorSwizzle = true;
-         features->borderColorSwizzleFromImage = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT: {
-         VkPhysicalDeviceCustomBorderColorFeaturesEXT *features = (void *) ext;
-         features->customBorderColors = true;
-         features->customBorderColorWithoutFormat = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT: {
-         VkPhysicalDeviceExtendedDynamicStateFeaturesEXT *features = (void *)ext;
-         features->extendedDynamicState = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_2_FEATURES_EXT: {
-         VkPhysicalDeviceExtendedDynamicState2FeaturesEXT *features =
-            (VkPhysicalDeviceExtendedDynamicState2FeaturesEXT *)ext;
-         features->extendedDynamicState2 = true;
-         features->extendedDynamicState2LogicOp = true;
-         features->extendedDynamicState2PatchControlPoints = true;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VARIABLE_POINTERS_FEATURES: {
+         VkPhysicalDeviceVariablePointersFeatures *features = (void *) ext;
+         features->variablePointersStorageBuffer = false;
+         features->variablePointers = false;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PERFORMANCE_QUERY_FEATURES_KHR: {
-         VkPhysicalDevicePerformanceQueryFeaturesKHR *feature =
-            (VkPhysicalDevicePerformanceQueryFeaturesKHR *)ext;
-         feature->performanceCounterQueryPools = true;
-         feature->performanceCounterMultipleQueryPools = false;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_FEATURES: {
+         VkPhysicalDeviceMultiviewFeatures *features =
+            (VkPhysicalDeviceMultiviewFeatures *) ext;
+         features->multiview = false;
+         features->multiviewGeometryShader = false;
+         features->multiviewTessellationShader = false;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_EXECUTABLE_PROPERTIES_FEATURES_KHR: {
-         VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *features =
-            (VkPhysicalDevicePipelineExecutablePropertiesFeaturesKHR *)ext;
-         features->pipelineExecutableInfo = true;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_DRAW_PARAMETERS_FEATURES: {
+         VkPhysicalDeviceShaderDrawParametersFeatures *features =
+            (VkPhysicalDeviceShaderDrawParametersFeatures *) ext;
+         features->shaderDrawParameters = false;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_FLOAT16_INT8_FEATURES: {
-         VkPhysicalDeviceShaderFloat16Int8Features *features =
-            (VkPhysicalDeviceShaderFloat16Int8Features *) ext;
-         features->shaderFloat16 = true;
-         features->shaderInt8 = false;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROTECTED_MEMORY_FEATURES: {
+         VkPhysicalDeviceProtectedMemoryFeatures *features =
+            (VkPhysicalDeviceProtectedMemoryFeatures *) ext;
+         features->protectedMemory = false;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SCALAR_BLOCK_LAYOUT_FEATURES: {
-         VkPhysicalDeviceScalarBlockLayoutFeatures *features = (void *)ext;
-         features->scalarBlockLayout = true;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_16BIT_STORAGE_FEATURES: {
+         VkPhysicalDevice16BitStorageFeatures *features =
+            (VkPhysicalDevice16BitStorageFeatures *) ext;
+         features->storageBuffer16BitAccess = false;
+         features->uniformAndStorageBuffer16BitAccess = false;
+         features->storagePushConstant16 = false;
+         features->storageInputOutput16 = false;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT: {
-         VkPhysicalDeviceRobustness2FeaturesEXT *features = (void *)ext;
-         features->robustBufferAccess2 = true;
-         features->robustImageAccess2 = true;
-         features->nullDescriptor = true;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLER_YCBCR_CONVERSION_FEATURES: {
+         VkPhysicalDeviceSamplerYcbcrConversionFeatures *features =
+            (VkPhysicalDeviceSamplerYcbcrConversionFeatures *) ext;
+         features->samplerYcbcrConversion = false;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TIMELINE_SEMAPHORE_FEATURES: {
-         VkPhysicalDeviceTimelineSemaphoreFeatures *features =
-            (VkPhysicalDeviceTimelineSemaphoreFeatures *) ext;
-         features->timelineSemaphore = true;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES_EXT: {
+         VkPhysicalDeviceDescriptorIndexingFeaturesEXT *features =
+            (VkPhysicalDeviceDescriptorIndexingFeaturesEXT *) ext;
+         features->shaderInputAttachmentArrayDynamicIndexing = false;
+         features->shaderUniformTexelBufferArrayDynamicIndexing = false;
+         features->shaderStorageTexelBufferArrayDynamicIndexing = false;
+         features->shaderUniformBufferArrayNonUniformIndexing = false;
+         features->shaderSampledImageArrayNonUniformIndexing = false;
+         features->shaderStorageBufferArrayNonUniformIndexing = false;
+         features->shaderStorageImageArrayNonUniformIndexing = false;
+         features->shaderInputAttachmentArrayNonUniformIndexing = false;
+         features->shaderUniformTexelBufferArrayNonUniformIndexing = false;
+         features->shaderStorageTexelBufferArrayNonUniformIndexing = false;
+         features->descriptorBindingUniformBufferUpdateAfterBind = false;
+         features->descriptorBindingSampledImageUpdateAfterBind = false;
+         features->descriptorBindingStorageImageUpdateAfterBind = false;
+         features->descriptorBindingStorageBufferUpdateAfterBind = false;
+         features->descriptorBindingUniformTexelBufferUpdateAfterBind = false;
+         features->descriptorBindingStorageTexelBufferUpdateAfterBind = false;
+         features->descriptorBindingUpdateUnusedWhilePending = false;
+         features->descriptorBindingPartiallyBound = false;
+         features->descriptorBindingVariableDescriptorCount = false;
+         features->runtimeDescriptorArray = false;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_FEATURES_EXT: {
-         VkPhysicalDeviceProvokingVertexFeaturesEXT *features =
-            (VkPhysicalDeviceProvokingVertexFeaturesEXT *)ext;
-         features->provokingVertexLast = true;
-         features->transformFeedbackPreservesProvokingVertex = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MUTABLE_DESCRIPTOR_TYPE_FEATURES_EXT: {
-         VkPhysicalDeviceMutableDescriptorTypeFeaturesEXT *features =
-            (VkPhysicalDeviceMutableDescriptorTypeFeaturesEXT *)ext;
-         features->mutableDescriptorType = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_FEATURES_EXT: {
-         VkPhysicalDeviceLineRasterizationFeaturesEXT *features =
-            (VkPhysicalDeviceLineRasterizationFeaturesEXT *)ext;
-         features->rectangularLines = true;
-         features->bresenhamLines = true;
-         features->smoothLines = false;
-         features->stippledRectangularLines = false;
-         features->stippledBresenhamLines = false;
-         features->stippledSmoothLines = false;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIMITIVE_TOPOLOGY_LIST_RESTART_FEATURES_EXT: {
-         VkPhysicalDevicePrimitiveTopologyListRestartFeaturesEXT *features =
-            (VkPhysicalDevicePrimitiveTopologyListRestartFeaturesEXT *)ext;
-         features->primitiveTopologyListRestart = true;
-         features->primitiveTopologyPatchListRestart = false;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_RASTERIZATION_ORDER_ATTACHMENT_ACCESS_FEATURES_EXT: {
-         VkPhysicalDeviceRasterizationOrderAttachmentAccessFeaturesEXT *features =
-            (VkPhysicalDeviceRasterizationOrderAttachmentAccessFeaturesEXT *)ext;
-         features->rasterizationOrderColorAttachmentAccess = true;
-         features->rasterizationOrderDepthAttachmentAccess = true;
-         features->rasterizationOrderStencilAttachmentAccess = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DEPTH_CLIP_CONTROL_FEATURES_EXT: {
-         VkPhysicalDeviceDepthClipControlFeaturesEXT *features =
-            (VkPhysicalDeviceDepthClipControlFeaturesEXT *)ext;
-         features->depthClipControl = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TEXEL_BUFFER_ALIGNMENT_FEATURES_EXT: {
-         VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *features =
-            (VkPhysicalDeviceTexelBufferAlignmentFeaturesEXT *)ext;
-         features->texelBufferAlignment = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRIMITIVES_GENERATED_QUERY_FEATURES_EXT: {
-         VkPhysicalDevicePrimitivesGeneratedQueryFeaturesEXT *features =
-            (VkPhysicalDevicePrimitivesGeneratedQueryFeaturesEXT *)ext;
-         features->primitivesGeneratedQuery = true;
-         features->primitivesGeneratedQueryWithRasterizerDiscard = false;
-         features->primitivesGeneratedQueryWithNonZeroStreams = false;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_VIEW_MIN_LOD_FEATURES_EXT: {
-         VkPhysicalDeviceImageViewMinLodFeaturesEXT *features =
-            (VkPhysicalDeviceImageViewMinLodFeaturesEXT *)ext;
-         features->minLod = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_2D_VIEW_OF_3D_FEATURES_EXT: {
-         VkPhysicalDeviceImage2DViewOf3DFeaturesEXT *features =
-            (VkPhysicalDeviceImage2DViewOf3DFeaturesEXT *)ext;
-         features->image2DViewOf3D = true;
-         features->sampler2DViewOf3D = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COLOR_WRITE_ENABLE_FEATURES_EXT: {
-         VkPhysicalDeviceColorWriteEnableFeaturesEXT *features =
-            (VkPhysicalDeviceColorWriteEnableFeaturesEXT *)ext;
-         features->colorWriteEnable = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MODULE_IDENTIFIER_FEATURES_EXT: {
-         VkPhysicalDeviceShaderModuleIdentifierFeaturesEXT *features =
-            (VkPhysicalDeviceShaderModuleIdentifierFeaturesEXT *)ext;
-         features->shaderModuleIdentifier = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_INPUT_DYNAMIC_STATE_FEATURES_EXT: {
-         VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT *features =
-            (VkPhysicalDeviceVertexInputDynamicStateFeaturesEXT *)ext;
-         features->vertexInputDynamicState = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_NON_SEAMLESS_CUBE_MAP_FEATURES_EXT: {
-         VkPhysicalDeviceNonSeamlessCubeMapFeaturesEXT *features =
-            (VkPhysicalDeviceNonSeamlessCubeMapFeaturesEXT *)ext;
-         features->nonSeamlessCubeMap = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ATTACHMENT_FEEDBACK_LOOP_LAYOUT_FEATURES_EXT: {
-         VkPhysicalDeviceAttachmentFeedbackLoopLayoutFeaturesEXT *features =
-            (VkPhysicalDeviceAttachmentFeedbackLoopLayoutFeaturesEXT*)ext;
-         features->attachmentFeedbackLoopLayout = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_GLOBAL_PRIORITY_QUERY_FEATURES_KHR: {
-         VkPhysicalDeviceGlobalPriorityQueryFeaturesKHR *features =
-            (VkPhysicalDeviceGlobalPriorityQueryFeaturesKHR*)ext;
-         features->globalPriorityQuery = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTI_DRAW_FEATURES_EXT: {
-         VkPhysicalDeviceMultiDrawFeaturesEXT *features =
-            (VkPhysicalDeviceMultiDrawFeaturesEXT *)ext;
-         features->multiDraw = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_GRAPHICS_PIPELINE_LIBRARY_FEATURES_EXT: {
-         VkPhysicalDeviceGraphicsPipelineLibraryFeaturesEXT *features =
-            (VkPhysicalDeviceGraphicsPipelineLibraryFeaturesEXT *)ext;
-         features->graphicsPipelineLibrary = true;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CONDITIONAL_RENDERING_FEATURES_EXT: {
+         VkPhysicalDeviceConditionalRenderingFeaturesEXT *features =
+            (VkPhysicalDeviceConditionalRenderingFeaturesEXT *) ext;
+         features->conditionalRendering = false;
+         features->inheritedConditionalRendering = false;
          break;
       }
-
       default:
          break;
       }
    }
+   return tu_GetPhysicalDeviceFeatures(physicalDevice, &pFeatures->features);
 }
 
-
-static void
-tu_get_physical_device_properties_1_1(struct tu_physical_device *pdevice,
-                                       VkPhysicalDeviceVulkan11Properties *p)
-{
-   assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES);
-
-   memcpy(p->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
-   memcpy(p->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
-   memset(p->deviceLUID, 0, VK_LUID_SIZE);
-   p->deviceNodeMask = 0;
-   p->deviceLUIDValid = false;
-
-   p->subgroupSize = 128;
-   p->subgroupSupportedStages = VK_SHADER_STAGE_COMPUTE_BIT;
-   p->subgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT |
-                                    VK_SUBGROUP_FEATURE_VOTE_BIT |
-                                    VK_SUBGROUP_FEATURE_BALLOT_BIT |
-                                    VK_SUBGROUP_FEATURE_SHUFFLE_BIT |
-                                    VK_SUBGROUP_FEATURE_SHUFFLE_RELATIVE_BIT |
-                                    VK_SUBGROUP_FEATURE_ARITHMETIC_BIT;
-   if (pdevice->info->a6xx.has_getfiberid) {
-      p->subgroupSupportedStages |= VK_SHADER_STAGE_ALL_GRAPHICS;
-      p->subgroupSupportedOperations |= VK_SUBGROUP_FEATURE_QUAD_BIT;
-   }
-
-   p->subgroupQuadOperationsInAllStages = false;
-
-   p->pointClippingBehavior = VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
-   p->maxMultiviewViewCount = MAX_VIEWS;
-   p->maxMultiviewInstanceIndex = INT_MAX;
-   p->protectedNoFault = false;
-   /* Our largest descriptors are 2 texture descriptors, or a texture and
-    * sampler descriptor.
-    */
-   p->maxPerSetDescriptors = MAX_SET_SIZE / (2 * A6XX_TEX_CONST_DWORDS * 4);
-   /* Our buffer size fields allow only this much */
-   p->maxMemoryAllocationSize = 0xFFFFFFFFull;
-
-}
-
-
-static const size_t max_descriptor_set_size = MAX_SET_SIZE / (4 * A6XX_TEX_CONST_DWORDS);
-static const VkSampleCountFlags sample_counts =
-   VK_SAMPLE_COUNT_1_BIT | VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT;
-
-static void
-tu_get_physical_device_properties_1_2(struct tu_physical_device *pdevice,
-                                       VkPhysicalDeviceVulkan12Properties *p)
-{
-   assert(p->sType == VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES);
-
-   p->driverID = VK_DRIVER_ID_MESA_TURNIP;
-   memset(p->driverName, 0, sizeof(p->driverName));
-   snprintf(p->driverName, VK_MAX_DRIVER_NAME_SIZE,
-            "turnip Mesa driver");
-   memset(p->driverInfo, 0, sizeof(p->driverInfo));
-   snprintf(p->driverInfo, VK_MAX_DRIVER_INFO_SIZE,
-            "Mesa " PACKAGE_VERSION MESA_GIT_SHA1);
-   p->conformanceVersion = (VkConformanceVersion) {
-      .major = 1,
-      .minor = 2,
-      .subminor = 7,
-      .patch = 1,
-   };
-
-   p->denormBehaviorIndependence =
-      VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL;
-   p->roundingModeIndependence =
-      VK_SHADER_FLOAT_CONTROLS_INDEPENDENCE_ALL;
-
-   p->shaderDenormFlushToZeroFloat16         = true;
-   p->shaderDenormPreserveFloat16            = false;
-   p->shaderRoundingModeRTEFloat16           = true;
-   p->shaderRoundingModeRTZFloat16           = false;
-   p->shaderSignedZeroInfNanPreserveFloat16  = true;
-
-   p->shaderDenormFlushToZeroFloat32         = true;
-   p->shaderDenormPreserveFloat32            = false;
-   p->shaderRoundingModeRTEFloat32           = true;
-   p->shaderRoundingModeRTZFloat32           = false;
-   p->shaderSignedZeroInfNanPreserveFloat32  = true;
-
-   p->shaderDenormFlushToZeroFloat64         = false;
-   p->shaderDenormPreserveFloat64            = false;
-   p->shaderRoundingModeRTEFloat64           = false;
-   p->shaderRoundingModeRTZFloat64           = false;
-   p->shaderSignedZeroInfNanPreserveFloat64  = false;
-
-   p->shaderUniformBufferArrayNonUniformIndexingNative   = true;
-   p->shaderSampledImageArrayNonUniformIndexingNative    = true;
-   p->shaderStorageBufferArrayNonUniformIndexingNative   = true;
-   p->shaderStorageImageArrayNonUniformIndexingNative    = true;
-   p->shaderInputAttachmentArrayNonUniformIndexingNative = false;
-   p->robustBufferAccessUpdateAfterBind                  = false;
-   p->quadDivergentImplicitLod                           = false;
-
-   p->maxUpdateAfterBindDescriptorsInAllPools            = max_descriptor_set_size;
-   p->maxPerStageDescriptorUpdateAfterBindSamplers       = max_descriptor_set_size;
-   p->maxPerStageDescriptorUpdateAfterBindUniformBuffers = max_descriptor_set_size;
-   p->maxPerStageDescriptorUpdateAfterBindStorageBuffers = max_descriptor_set_size;
-   p->maxPerStageDescriptorUpdateAfterBindSampledImages  = max_descriptor_set_size;
-   p->maxPerStageDescriptorUpdateAfterBindStorageImages  = max_descriptor_set_size;
-   p->maxPerStageDescriptorUpdateAfterBindInputAttachments = MAX_RTS;
-   p->maxPerStageUpdateAfterBindResources                = max_descriptor_set_size;
-   p->maxDescriptorSetUpdateAfterBindSamplers            = max_descriptor_set_size;
-   p->maxDescriptorSetUpdateAfterBindUniformBuffers      = max_descriptor_set_size;
-   p->maxDescriptorSetUpdateAfterBindUniformBuffersDynamic = MAX_DYNAMIC_UNIFORM_BUFFERS;
-   p->maxDescriptorSetUpdateAfterBindStorageBuffers      = max_descriptor_set_size;
-   p->maxDescriptorSetUpdateAfterBindStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS;
-   p->maxDescriptorSetUpdateAfterBindSampledImages       = max_descriptor_set_size;
-   p->maxDescriptorSetUpdateAfterBindStorageImages       = max_descriptor_set_size;
-   p->maxDescriptorSetUpdateAfterBindInputAttachments    = MAX_RTS;
-
-   p->supportedDepthResolveModes    = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
-   p->supportedStencilResolveModes  = VK_RESOLVE_MODE_SAMPLE_ZERO_BIT;
-   p->independentResolveNone  = false;
-   p->independentResolve      = false;
-
-   p->filterMinmaxSingleComponentFormats  = true;
-   p->filterMinmaxImageComponentMapping   = true;
-
-   p->maxTimelineSemaphoreValueDifference = UINT64_MAX;
-
-   p->framebufferIntegerColorSampleCounts = sample_counts;
-}
-
-static void
-tu_get_physical_device_properties_1_3(struct tu_physical_device *pdevice,
-                                       VkPhysicalDeviceVulkan13Properties *p)
-{
-   /* TODO move threadsize_base and max_waves to fd_dev_info and use them here */
-   p->minSubgroupSize = 64; /* threadsize_base */
-   p->maxSubgroupSize = 128; /* threadsize_base * 2 */
-   p->maxComputeWorkgroupSubgroups = 16; /* max_waves */
-   p->requiredSubgroupSizeStages = VK_SHADER_STAGE_ALL;
-
-   /* Inline uniform buffers are just normal UBOs */
-   p->maxInlineUniformBlockSize = MAX_UNIFORM_BUFFER_RANGE;
-
-   /* Halve the normal limit on the number of descriptors, see below. */
-   p->maxPerStageDescriptorInlineUniformBlocks = max_descriptor_set_size / 2;
-   p->maxPerStageDescriptorUpdateAfterBindInlineUniformBlocks = max_descriptor_set_size / 2;
-   p->maxDescriptorSetInlineUniformBlocks = max_descriptor_set_size / 2;
-   p->maxDescriptorSetUpdateAfterBindInlineUniformBlocks = max_descriptor_set_size / 2;
-   /* Because we halve the normal limit on the number of descriptors, in the
-    * worst case each descriptor takes up half the space, leaving the rest for
-    * the actual data.
-    */
-   p->maxInlineUniformTotalSize = MAX_SET_SIZE / 2;
-
-   p->integerDotProduct8BitUnsignedAccelerated = false;
-   p->integerDotProduct8BitSignedAccelerated = false;
-   p->integerDotProduct8BitMixedSignednessAccelerated = false;
-   p->integerDotProduct4x8BitPackedUnsignedAccelerated =
-      pdevice->info->a6xx.has_dp2acc;
-   /* TODO: we should be able to emulate 4x8BitPackedSigned fast enough */
-   p->integerDotProduct4x8BitPackedSignedAccelerated = false;
-   p->integerDotProduct4x8BitPackedMixedSignednessAccelerated =
-      pdevice->info->a6xx.has_dp2acc;
-   p->integerDotProduct16BitUnsignedAccelerated = false;
-   p->integerDotProduct16BitSignedAccelerated = false;
-   p->integerDotProduct16BitMixedSignednessAccelerated = false;
-   p->integerDotProduct32BitUnsignedAccelerated = false;
-   p->integerDotProduct32BitSignedAccelerated = false;
-   p->integerDotProduct32BitMixedSignednessAccelerated = false;
-   p->integerDotProduct64BitUnsignedAccelerated = false;
-   p->integerDotProduct64BitSignedAccelerated = false;
-   p->integerDotProduct64BitMixedSignednessAccelerated = false;
-   p->integerDotProductAccumulatingSaturating8BitUnsignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating8BitSignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating8BitMixedSignednessAccelerated = false;
-   p->integerDotProductAccumulatingSaturating4x8BitPackedUnsignedAccelerated =
-      pdevice->info->a6xx.has_dp2acc;
-   /* TODO: we should be able to emulate Saturating4x8BitPackedSigned fast enough */
-   p->integerDotProductAccumulatingSaturating4x8BitPackedSignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating4x8BitPackedMixedSignednessAccelerated =
-      pdevice->info->a6xx.has_dp2acc;
-   p->integerDotProductAccumulatingSaturating16BitUnsignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating16BitSignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating16BitMixedSignednessAccelerated = false;
-   p->integerDotProductAccumulatingSaturating32BitUnsignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating32BitSignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating32BitMixedSignednessAccelerated = false;
-   p->integerDotProductAccumulatingSaturating64BitUnsignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating64BitSignedAccelerated = false;
-   p->integerDotProductAccumulatingSaturating64BitMixedSignednessAccelerated = false;
-
-   p->storageTexelBufferOffsetAlignmentBytes = 64;
-   p->storageTexelBufferOffsetSingleTexelAlignment = false;
-   p->uniformTexelBufferOffsetAlignmentBytes = 64;
-   p->uniformTexelBufferOffsetSingleTexelAlignment = false;
-
-   /* The address space is 4GB for current kernels, so there's no point
-    * allowing a larger buffer. Our buffer sizes are 64-bit though, so
-    * GetBufferDeviceRequirements won't fall over if someone actually creates
-    * a 4GB buffer.
-    */
-   p->maxBufferSize = 1ull << 32;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
-                                VkPhysicalDeviceProperties2 *pProperties)
+void
+tu_GetPhysicalDeviceProperties(VkPhysicalDevice physicalDevice,
+                               VkPhysicalDeviceProperties *pProperties)
 {
    TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
+   VkSampleCountFlags sample_counts = 0xf;
+
+   /* make sure that the entire descriptor set is addressable with a signed
+    * 32-bit int. So the sum of all limits scaled by descriptor size has to
+    * be at most 2 GiB. the combined image & samples object count as one of
+    * both. This limit is for the pipeline layout, not for the set layout, but
+    * there is no set limit, so we just set a pipeline limit. I don't think
+    * any app is going to hit this soon. */
+   size_t max_descriptor_set_size =
+      ((1ull << 31) - 16 * MAX_DYNAMIC_BUFFERS) /
+      (32 /* uniform buffer, 32 due to potential space wasted on alignment */ +
+       32 /* storage buffer, 32 due to potential space wasted on alignment */ +
+       32 /* sampler, largest when combined with image */ +
+       64 /* sampled image */ + 64 /* storage image */);
 
    VkPhysicalDeviceLimits limits = {
       .maxImageDimension1D = (1 << 14),
@@ -1136,20 +722,20 @@ tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
       .maxImageDimensionCube = (1 << 14),
       .maxImageArrayLayers = (1 << 11),
       .maxTexelBufferElements = 128 * 1024 * 1024,
-      .maxUniformBufferRange = MAX_UNIFORM_BUFFER_RANGE,
-      .maxStorageBufferRange = MAX_STORAGE_BUFFER_RANGE,
+      .maxUniformBufferRange = UINT32_MAX,
+      .maxStorageBufferRange = UINT32_MAX,
       .maxPushConstantsSize = MAX_PUSH_CONSTANTS_SIZE,
       .maxMemoryAllocationCount = UINT32_MAX,
       .maxSamplerAllocationCount = 64 * 1024,
       .bufferImageGranularity = 64,          /* A cache line */
-      .sparseAddressSpaceSize = 0,
+      .sparseAddressSpaceSize = 0xffffffffu, /* buffer max size */
       .maxBoundDescriptorSets = MAX_SETS,
       .maxPerStageDescriptorSamplers = max_descriptor_set_size,
       .maxPerStageDescriptorUniformBuffers = max_descriptor_set_size,
       .maxPerStageDescriptorStorageBuffers = max_descriptor_set_size,
       .maxPerStageDescriptorSampledImages = max_descriptor_set_size,
       .maxPerStageDescriptorStorageImages = max_descriptor_set_size,
-      .maxPerStageDescriptorInputAttachments = MAX_RTS,
+      .maxPerStageDescriptorInputAttachments = max_descriptor_set_size,
       .maxPerStageResources = max_descriptor_set_size,
       .maxDescriptorSetSamplers = max_descriptor_set_size,
       .maxDescriptorSetUniformBuffers = max_descriptor_set_size,
@@ -1158,10 +744,10 @@ tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
       .maxDescriptorSetStorageBuffersDynamic = MAX_DYNAMIC_STORAGE_BUFFERS,
       .maxDescriptorSetSampledImages = max_descriptor_set_size,
       .maxDescriptorSetStorageImages = max_descriptor_set_size,
-      .maxDescriptorSetInputAttachments = MAX_RTS,
+      .maxDescriptorSetInputAttachments = max_descriptor_set_size,
       .maxVertexInputAttributes = 32,
       .maxVertexInputBindings = 32,
-      .maxVertexInputAttributeOffset = 4095,
+      .maxVertexInputAttributeOffset = 2047,
       .maxVertexInputBindingStride = 2048,
       .maxVertexOutputComponents = 128,
       .maxTessellationGenerationLevel = 64,
@@ -1172,41 +758,41 @@ tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
       .maxTessellationControlTotalOutputComponents = 4096,
       .maxTessellationEvaluationInputComponents = 128,
       .maxTessellationEvaluationOutputComponents = 128,
-      .maxGeometryShaderInvocations = 32,
+      .maxGeometryShaderInvocations = 127,
       .maxGeometryInputComponents = 64,
       .maxGeometryOutputComponents = 128,
       .maxGeometryOutputVertices = 256,
       .maxGeometryTotalOutputComponents = 1024,
-      .maxFragmentInputComponents = 124,
+      .maxFragmentInputComponents = 128,
       .maxFragmentOutputAttachments = 8,
       .maxFragmentDualSrcAttachments = 1,
-      .maxFragmentCombinedOutputResources = MAX_RTS + max_descriptor_set_size * 2,
+      .maxFragmentCombinedOutputResources = 8,
       .maxComputeSharedMemorySize = 32768,
       .maxComputeWorkGroupCount = { 65535, 65535, 65535 },
       .maxComputeWorkGroupInvocations = 2048,
-      .maxComputeWorkGroupSize = { 1024, 1024, 1024 },
-      .subPixelPrecisionBits = 8,
-      .subTexelPrecisionBits = 8,
-      .mipmapPrecisionBits = 8,
+      .maxComputeWorkGroupSize = { 2048, 2048, 2048 },
+      .subPixelPrecisionBits = 4 /* FIXME */,
+      .subTexelPrecisionBits = 4 /* FIXME */,
+      .mipmapPrecisionBits = 4 /* FIXME */,
       .maxDrawIndexedIndexValue = UINT32_MAX,
       .maxDrawIndirectCount = UINT32_MAX,
-      .maxSamplerLodBias = 4095.0 / 256.0, /* [-16, 15.99609375] */
+      .maxSamplerLodBias = 16,
       .maxSamplerAnisotropy = 16,
       .maxViewports = MAX_VIEWPORTS,
-      .maxViewportDimensions = { MAX_VIEWPORT_SIZE, MAX_VIEWPORT_SIZE },
+      .maxViewportDimensions = { (1 << 14), (1 << 14) },
       .viewportBoundsRange = { INT16_MIN, INT16_MAX },
       .viewportSubPixelBits = 8,
       .minMemoryMapAlignment = 4096, /* A page */
-      .minTexelBufferOffsetAlignment = 64,
-      .minUniformBufferOffsetAlignment = 64,
-      .minStorageBufferOffsetAlignment = 64,
-      .minTexelOffset = -16,
-      .maxTexelOffset = 15,
+      .minTexelBufferOffsetAlignment = 1,
+      .minUniformBufferOffsetAlignment = 4,
+      .minStorageBufferOffsetAlignment = 4,
+      .minTexelOffset = -32,
+      .maxTexelOffset = 31,
       .minTexelGatherOffset = -32,
       .maxTexelGatherOffset = 31,
-      .minInterpolationOffset = -0.5,
-      .maxInterpolationOffset = 0.4375,
-      .subPixelInterpolationOffsetBits = 4,
+      .minInterpolationOffset = -2,
+      .maxInterpolationOffset = 2,
+      .subPixelInterpolationOffsetBits = 8,
       .maxFramebufferWidth = (1 << 14),
       .maxFramebufferHeight = (1 << 14),
       .maxFramebufferLayers = (1 << 10),
@@ -1216,65 +802,51 @@ tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
       .framebufferNoAttachmentsSampleCounts = sample_counts,
       .maxColorAttachments = MAX_RTS,
       .sampledImageColorSampleCounts = sample_counts,
-      .sampledImageIntegerSampleCounts = sample_counts,
+      .sampledImageIntegerSampleCounts = VK_SAMPLE_COUNT_1_BIT,
       .sampledImageDepthSampleCounts = sample_counts,
       .sampledImageStencilSampleCounts = sample_counts,
       .storageImageSampleCounts = VK_SAMPLE_COUNT_1_BIT,
       .maxSampleMaskWords = 1,
       .timestampComputeAndGraphics = true,
-      .timestampPeriod = 1000000000.0 / 19200000.0, /* CP_ALWAYS_ON_COUNTER is fixed 19.2MHz */
+      .timestampPeriod = 1,
       .maxClipDistances = 8,
       .maxCullDistances = 8,
       .maxCombinedClipAndCullDistances = 8,
-      .discreteQueuePriorities = 2,
-      .pointSizeRange = { 1, 4092 },
-      .lineWidthRange = { 1.0, 1.0 },
-      .pointSizeGranularity = 	0.0625,
-      .lineWidthGranularity = 0.0,
-      .strictLines = true,
+      .discreteQueuePriorities = 1,
+      .pointSizeRange = { 0.125, 255.875 },
+      .lineWidthRange = { 0.0, 7.9921875 },
+      .pointSizeGranularity = (1.0 / 8.0),
+      .lineWidthGranularity = (1.0 / 128.0),
+      .strictLines = false, /* FINISHME */
       .standardSampleLocations = true,
       .optimalBufferCopyOffsetAlignment = 128,
       .optimalBufferCopyRowPitchAlignment = 128,
       .nonCoherentAtomSize = 64,
    };
 
-   pProperties->properties = (VkPhysicalDeviceProperties) {
-      .apiVersion = TU_API_VERSION,
+   *pProperties = (VkPhysicalDeviceProperties) {
+      .apiVersion = tu_physical_device_api_version(pdevice),
       .driverVersion = vk_get_driver_version(),
-      .vendorID = 0x5143,
-      .deviceID = pdevice->dev_id.chip_id,
+      .vendorID = 0, /* TODO */
+      .deviceID = 0,
       .deviceType = VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU,
       .limits = limits,
       .sparseProperties = { 0 },
    };
 
-   strcpy(pProperties->properties.deviceName, pdevice->name);
-   memcpy(pProperties->properties.pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
-
-   VkPhysicalDeviceVulkan11Properties core_1_1 = {
-      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_PROPERTIES,
-   };
-   tu_get_physical_device_properties_1_1(pdevice, &core_1_1);
-
-   VkPhysicalDeviceVulkan12Properties core_1_2 = {
-      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_PROPERTIES,
-   };
-   tu_get_physical_device_properties_1_2(pdevice, &core_1_2);
+   strcpy(pProperties->deviceName, pdevice->name);
+   memcpy(pProperties->pipelineCacheUUID, pdevice->cache_uuid, VK_UUID_SIZE);
+}
 
-   VkPhysicalDeviceVulkan13Properties core_1_3 = {
-      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_3_PROPERTIES,
-   };
-   tu_get_physical_device_properties_1_3(pdevice, &core_1_3);
+void
+tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
+                                VkPhysicalDeviceProperties2 *pProperties)
+{
+   TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
+   tu_GetPhysicalDeviceProperties(physicalDevice, &pProperties->properties);
 
    vk_foreach_struct(ext, pProperties->pNext)
    {
-      if (vk_get_physical_device_core_1_1_property_ext(ext, &core_1_1))
-         continue;
-      if (vk_get_physical_device_core_1_2_property_ext(ext, &core_1_2))
-         continue;
-      if (vk_get_physical_device_core_1_3_property_ext(ext, &core_1_3))
-         continue;
-
       switch (ext->sType) {
       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PUSH_DESCRIPTOR_PROPERTIES_KHR: {
          VkPhysicalDevicePushDescriptorPropertiesKHR *properties =
@@ -1282,109 +854,36 @@ tu_GetPhysicalDeviceProperties2(VkPhysicalDevice physicalDevice,
          properties->maxPushDescriptors = MAX_PUSH_DESCRIPTORS;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT: {
-         VkPhysicalDeviceTransformFeedbackPropertiesEXT *properties =
-            (VkPhysicalDeviceTransformFeedbackPropertiesEXT *)ext;
-
-         properties->maxTransformFeedbackStreams = IR3_MAX_SO_STREAMS;
-         properties->maxTransformFeedbackBuffers = IR3_MAX_SO_BUFFERS;
-         properties->maxTransformFeedbackBufferSize = UINT32_MAX;
-         properties->maxTransformFeedbackStreamDataSize = 512;
-         properties->maxTransformFeedbackBufferDataSize = 512;
-         properties->maxTransformFeedbackBufferDataStride = 512;
-         properties->transformFeedbackQueries = true;
-         properties->transformFeedbackStreamsLinesTriangles = true;
-         properties->transformFeedbackRasterizationStreamSelect = true;
-         properties->transformFeedbackDraw = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SAMPLE_LOCATIONS_PROPERTIES_EXT: {
-         VkPhysicalDeviceSampleLocationsPropertiesEXT *properties =
-            (VkPhysicalDeviceSampleLocationsPropertiesEXT *)ext;
-         properties->sampleLocationSampleCounts = 0;
-         if (pdevice->vk.supported_extensions.EXT_sample_locations) {
-            properties->sampleLocationSampleCounts =
-               VK_SAMPLE_COUNT_1_BIT | VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT;
-         }
-         properties->maxSampleLocationGridSize = (VkExtent2D) { 1 , 1 };
-         properties->sampleLocationCoordinateRange[0] = 0.0f;
-         properties->sampleLocationCoordinateRange[1] = 0.9375f;
-         properties->sampleLocationSubPixelBits = 4;
-         properties->variableSampleLocations = true;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT: {
-         VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *props =
-            (VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT *)ext;
-         props->maxVertexAttribDivisor = UINT32_MAX;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ID_PROPERTIES: {
+         VkPhysicalDeviceIDProperties *properties =
+            (VkPhysicalDeviceIDProperties *) ext;
+         memcpy(properties->driverUUID, pdevice->driver_uuid, VK_UUID_SIZE);
+         memcpy(properties->deviceUUID, pdevice->device_uuid, VK_UUID_SIZE);
+         properties->deviceLUIDValid = false;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_PROPERTIES_EXT: {
-         VkPhysicalDeviceCustomBorderColorPropertiesEXT *props = (void *)ext;
-         props->maxCustomBorderColorSamplers = TU_BORDER_COLOR_COUNT;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTIVIEW_PROPERTIES: {
+         VkPhysicalDeviceMultiviewProperties *properties =
+            (VkPhysicalDeviceMultiviewProperties *) ext;
+         properties->maxMultiviewViewCount = MAX_VIEWS;
+         properties->maxMultiviewInstanceIndex = INT_MAX;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PERFORMANCE_QUERY_PROPERTIES_KHR: {
-         VkPhysicalDevicePerformanceQueryPropertiesKHR *properties =
-            (VkPhysicalDevicePerformanceQueryPropertiesKHR *)ext;
-         properties->allowCommandBufferQueryCopies = false;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_POINT_CLIPPING_PROPERTIES: {
+         VkPhysicalDevicePointClippingProperties *properties =
+            (VkPhysicalDevicePointClippingProperties *) ext;
+         properties->pointClippingBehavior =
+            VK_POINT_CLIPPING_BEHAVIOR_ALL_CLIP_PLANES;
          break;
       }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_PROPERTIES_EXT: {
-         VkPhysicalDeviceRobustness2PropertiesEXT *props = (void *)ext;
-         /* see write_buffer_descriptor() */
-         props->robustStorageBufferAccessSizeAlignment = 4;
-         /* see write_ubo_descriptor() */
-         props->robustUniformBufferAccessSizeAlignment = 16;
-         break;
-      }
-
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROVOKING_VERTEX_PROPERTIES_EXT: {
-         VkPhysicalDeviceProvokingVertexPropertiesEXT *properties =
-            (VkPhysicalDeviceProvokingVertexPropertiesEXT *)ext;
-         properties->provokingVertexModePerPipeline = true;
-         properties->transformFeedbackPreservesTriangleFanProvokingVertex = false;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_LINE_RASTERIZATION_PROPERTIES_EXT: {
-         VkPhysicalDeviceLineRasterizationPropertiesEXT *props =
-            (VkPhysicalDeviceLineRasterizationPropertiesEXT *)ext;
-         props->lineSubPixelPrecisionBits = 8;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DRM_PROPERTIES_EXT: {
-         VkPhysicalDeviceDrmPropertiesEXT *props =
-            (VkPhysicalDeviceDrmPropertiesEXT *)ext;
-         props->hasPrimary = pdevice->has_master;
-         props->primaryMajor = pdevice->master_major;
-         props->primaryMinor = pdevice->master_minor;
-
-         props->hasRender = pdevice->has_local;
-         props->renderMajor = pdevice->local_major;
-         props->renderMinor = pdevice->local_minor;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_MODULE_IDENTIFIER_PROPERTIES_EXT: {
-         VkPhysicalDeviceShaderModuleIdentifierPropertiesEXT *props =
-            (VkPhysicalDeviceShaderModuleIdentifierPropertiesEXT *)ext;
-         STATIC_ASSERT(sizeof(vk_shaderModuleIdentifierAlgorithmUUID) ==
-                       sizeof(props->shaderModuleIdentifierAlgorithmUUID));
-         memcpy(props->shaderModuleIdentifierAlgorithmUUID,
-                vk_shaderModuleIdentifierAlgorithmUUID,
-                sizeof(props->shaderModuleIdentifierAlgorithmUUID));
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MULTI_DRAW_PROPERTIES_EXT: {
-         VkPhysicalDeviceMultiDrawPropertiesEXT *properties =
-            (VkPhysicalDeviceMultiDrawPropertiesEXT *)ext;
-         properties->maxMultiDrawCount = 2048;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_GRAPHICS_PIPELINE_LIBRARY_PROPERTIES_EXT: {
-         VkPhysicalDeviceGraphicsPipelineLibraryPropertiesEXT *props =
-            (VkPhysicalDeviceGraphicsPipelineLibraryPropertiesEXT *)ext;
-         props->graphicsPipelineLibraryFastLinking = true;
-         props->graphicsPipelineLibraryIndependentInterpolationDecoration = true;
+      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_3_PROPERTIES: {
+         VkPhysicalDeviceMaintenance3Properties *properties =
+            (VkPhysicalDeviceMaintenance3Properties *) ext;
+         /* Make sure everything is addressable by a signed 32-bit int, and
+          * our largest descriptors are 96 bytes. */
+         properties->maxPerSetDescriptors = (1ull << 31) / 96;
+         /* Our buffer size fields allow only this much */
+         properties->maxMemoryAllocationSize = 0xFFFFFFFFull;
          break;
       }
       default:
@@ -1397,99 +896,36 @@ static const VkQueueFamilyProperties tu_queue_family_properties = {
    .queueFlags =
       VK_QUEUE_GRAPHICS_BIT | VK_QUEUE_COMPUTE_BIT | VK_QUEUE_TRANSFER_BIT,
    .queueCount = 1,
-   .timestampValidBits = 48,
+   .timestampValidBits = 64,
    .minImageTransferGranularity = { 1, 1, 1 },
 };
 
-static void
-tu_physical_device_get_global_priority_properties(const struct tu_physical_device *pdevice,
-                                                  VkQueueFamilyGlobalPriorityPropertiesKHR *props)
+void
+tu_GetPhysicalDeviceQueueFamilyProperties(
+   VkPhysicalDevice physicalDevice,
+   uint32_t *pQueueFamilyPropertyCount,
+   VkQueueFamilyProperties *pQueueFamilyProperties)
 {
-   props->priorityCount = MIN2(pdevice->submitqueue_priority_count, 3);
-   switch (props->priorityCount) {
-   case 1:
-      props->priorities[0] = VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_KHR;
-      break;
-   case 2:
-      props->priorities[0] = VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_KHR;
-      props->priorities[1] = VK_QUEUE_GLOBAL_PRIORITY_HIGH_KHR;
-      break;
-   case 3:
-      props->priorities[0] = VK_QUEUE_GLOBAL_PRIORITY_LOW_KHR;
-      props->priorities[1] = VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_KHR;
-      props->priorities[2] = VK_QUEUE_GLOBAL_PRIORITY_HIGH_KHR;
-      break;
-   default:
-      unreachable("unexpected priority count");
-      break;
-   }
-}
+   VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 
-static int
-tu_physical_device_get_submitqueue_priority(const struct tu_physical_device *pdevice,
-                                            VkQueueGlobalPriorityKHR global_priority,
-                                            bool global_priority_query)
-{
-   if (global_priority_query) {
-      VkQueueFamilyGlobalPriorityPropertiesKHR props;
-      tu_physical_device_get_global_priority_properties(pdevice, &props);
-
-      bool valid = false;
-      for (uint32_t i = 0; i < props.priorityCount; i++) {
-         if (props.priorities[i] == global_priority) {
-            valid = true;
-            break;
-         }
-      }
-
-      if (!valid)
-         return -1;
-   }
-
-   /* Valid values are from 0 to (pdevice->submitqueue_priority_count - 1),
-    * with 0 being the highest priority.  This matches what freedreno does.
-    */
-   int priority;
-   if (global_priority == VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_KHR)
-      priority = pdevice->submitqueue_priority_count / 2;
-   else if (global_priority < VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_KHR)
-      priority = pdevice->submitqueue_priority_count - 1;
-   else
-      priority = 0;
-
-   return priority;
+   vk_outarray_append(&out, p) { *p = tu_queue_family_properties; }
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_GetPhysicalDeviceQueueFamilyProperties2(
    VkPhysicalDevice physicalDevice,
    uint32_t *pQueueFamilyPropertyCount,
    VkQueueFamilyProperties2 *pQueueFamilyProperties)
 {
-   TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
+   VK_OUTARRAY_MAKE(out, pQueueFamilyProperties, pQueueFamilyPropertyCount);
 
-   VK_OUTARRAY_MAKE_TYPED(VkQueueFamilyProperties2, out,
-                          pQueueFamilyProperties, pQueueFamilyPropertyCount);
-
-   vk_outarray_append_typed(VkQueueFamilyProperties2, &out, p)
+   vk_outarray_append(&out, p)
    {
       p->queueFamilyProperties = tu_queue_family_properties;
-
-      vk_foreach_struct(ext, p->pNext) {
-         switch (ext->sType) {
-         case VK_STRUCTURE_TYPE_QUEUE_FAMILY_GLOBAL_PRIORITY_PROPERTIES_KHR: {
-            VkQueueFamilyGlobalPriorityPropertiesKHR *props = (void *)ext;
-            tu_physical_device_get_global_priority_properties(pdevice, props);
-            break;
-         }
-         default:
-            break;
-         }
-      }
    }
 }
 
-uint64_t
+static uint64_t
 tu_get_system_heap_size()
 {
    struct sysinfo info;
@@ -1509,101 +945,50 @@ tu_get_system_heap_size()
    return available_ram;
 }
 
-static VkDeviceSize
-tu_get_budget_memory(struct tu_physical_device *physical_device)
-{
-   uint64_t heap_size = physical_device->heap.size;
-   uint64_t heap_used = physical_device->heap.used;
-   uint64_t sys_available;
-   ASSERTED bool has_available_memory =
-      os_get_available_system_memory(&sys_available);
-   assert(has_available_memory);
-
-   /*
-    * Let's not incite the app to starve the system: report at most 90% of
-    * available system memory.
-    */
-   uint64_t heap_available = sys_available * 9 / 10;
-   return MIN2(heap_size, heap_used + heap_available);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_GetPhysicalDeviceMemoryProperties2(VkPhysicalDevice pdev,
-                                      VkPhysicalDeviceMemoryProperties2 *props2)
+void
+tu_GetPhysicalDeviceMemoryProperties(
+   VkPhysicalDevice physicalDevice,
+   VkPhysicalDeviceMemoryProperties *pMemoryProperties)
 {
-   TU_FROM_HANDLE(tu_physical_device, physical_device, pdev);
-
-   VkPhysicalDeviceMemoryProperties *props = &props2->memoryProperties;
-   props->memoryHeapCount = 1;
-   props->memoryHeaps[0].size = physical_device->heap.size;
-   props->memoryHeaps[0].flags = physical_device->heap.flags;
+   pMemoryProperties->memoryHeapCount = 1;
+   pMemoryProperties->memoryHeaps[0].size = tu_get_system_heap_size();
+   pMemoryProperties->memoryHeaps[0].flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
 
-   props->memoryTypeCount = 1;
-   props->memoryTypes[0].propertyFlags =
+   pMemoryProperties->memoryTypeCount = 1;
+   pMemoryProperties->memoryTypes[0].propertyFlags =
       VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
       VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
       VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-   props->memoryTypes[0].heapIndex = 0;
+   pMemoryProperties->memoryTypes[0].heapIndex = 0;
+}
 
-   vk_foreach_struct(ext, props2->pNext)
-   {
-      switch (ext->sType) {
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT: {
-         VkPhysicalDeviceMemoryBudgetPropertiesEXT *memory_budget_props =
-            (VkPhysicalDeviceMemoryBudgetPropertiesEXT *) ext;
-         memory_budget_props->heapUsage[0] = physical_device->heap.used;
-         memory_budget_props->heapBudget[0] = tu_get_budget_memory(physical_device);
-
-         /* The heapBudget and heapUsage values must be zero for array elements
-          * greater than or equal to VkPhysicalDeviceMemoryProperties::memoryHeapCount
-          */
-         for (unsigned i = 1; i < VK_MAX_MEMORY_HEAPS; i++) {
-            memory_budget_props->heapBudget[i] = 0u;
-            memory_budget_props->heapUsage[i] = 0u;
-         }
-         break;
-      }
-      default:
-         break;
-      }
-   }
+void
+tu_GetPhysicalDeviceMemoryProperties2(
+   VkPhysicalDevice physicalDevice,
+   VkPhysicalDeviceMemoryProperties2 *pMemoryProperties)
+{
+   return tu_GetPhysicalDeviceMemoryProperties(
+      physicalDevice, &pMemoryProperties->memoryProperties);
 }
 
 static VkResult
 tu_queue_init(struct tu_device *device,
               struct tu_queue *queue,
+              uint32_t queue_family_index,
               int idx,
-              const VkDeviceQueueCreateInfo *create_info,
-              bool global_priority_query)
-{
-   const VkDeviceQueueGlobalPriorityCreateInfoKHR *priority_info =
-      vk_find_struct_const(create_info->pNext,
-            DEVICE_QUEUE_GLOBAL_PRIORITY_CREATE_INFO_KHR);
-   const enum VkQueueGlobalPriorityKHR global_priority = priority_info ?
-      priority_info->globalPriority : VK_QUEUE_GLOBAL_PRIORITY_MEDIUM_KHR;
-
-   const int priority = tu_physical_device_get_submitqueue_priority(
-         device->physical_device, global_priority, global_priority_query);
-   if (priority < 0) {
-      return vk_startup_errorf(device->instance, VK_ERROR_INITIALIZATION_FAILED,
-                               "invalid global priority");
-   }
-
-   VkResult result = vk_queue_init(&queue->vk, &device->vk, create_info, idx);
-   if (result != VK_SUCCESS)
-      return result;
-
+              VkDeviceQueueCreateFlags flags)
+{
+   queue->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
    queue->device = device;
-#ifndef TU_USE_KGSL
-   queue->vk.driver_submit = tu_queue_submit;
-#endif
+   queue->queue_family_index = queue_family_index;
+   queue->queue_idx = idx;
+   queue->flags = flags;
 
-   int ret = tu_drm_submitqueue_new(device, priority, &queue->msm_queue_id);
+   int ret = tu_drm_submitqueue_new(device, 0, &queue->msm_queue_id);
    if (ret)
-      return vk_startup_errorf(device->instance, VK_ERROR_INITIALIZATION_FAILED,
-                               "submitqueue create failed");
+      return VK_ERROR_INITIALIZATION_FAILED;
 
-   queue->fence = -1;
+   tu_fence_init(&queue->submit_fence, false);
 
    return VK_SUCCESS;
 }
@@ -1611,271 +996,21 @@ tu_queue_init(struct tu_device *device,
 static void
 tu_queue_finish(struct tu_queue *queue)
 {
-   vk_queue_finish(&queue->vk);
-   if (queue->fence >= 0)
-      close(queue->fence);
+   tu_fence_finish(&queue->submit_fence);
    tu_drm_submitqueue_close(queue->device, queue->msm_queue_id);
 }
 
-uint64_t
-tu_device_ticks_to_ns(struct tu_device *dev, uint64_t ts)
-{
-   /* This is based on the 19.2MHz always-on rbbm timer.
-    *
-    * TODO we should probably query this value from kernel..
-    */
-   return ts * (1000000000 / 19200000);
-}
-
-static void*
-tu_trace_create_ts_buffer(struct u_trace_context *utctx, uint32_t size)
-{
-   struct tu_device *device =
-      container_of(utctx, struct tu_device, trace_context);
-
-   struct tu_bo *bo;
-   tu_bo_init_new(device, &bo, size, false, "trace");
-
-   return bo;
-}
-
-static void
-tu_trace_destroy_ts_buffer(struct u_trace_context *utctx, void *timestamps)
-{
-   struct tu_device *device =
-      container_of(utctx, struct tu_device, trace_context);
-   struct tu_bo *bo = timestamps;
-
-   tu_bo_finish(device, bo);
-}
-
-static void
-tu_trace_record_ts(struct u_trace *ut, void *cs, void *timestamps,
-                   unsigned idx, bool end_of_pipe)
-{
-   struct tu_bo *bo = timestamps;
-   struct tu_cs *ts_cs = cs;
-
-   unsigned ts_offset = idx * sizeof(uint64_t);
-   tu_cs_emit_pkt7(ts_cs, CP_EVENT_WRITE, 4);
-   tu_cs_emit(ts_cs, CP_EVENT_WRITE_0_EVENT(RB_DONE_TS) | CP_EVENT_WRITE_0_TIMESTAMP);
-   tu_cs_emit_qw(ts_cs, bo->iova + ts_offset);
-   tu_cs_emit(ts_cs, 0x00000000);
-}
-
-static uint64_t
-tu_trace_read_ts(struct u_trace_context *utctx,
-                 void *timestamps, unsigned idx, void *flush_data)
-{
-   struct tu_device *device =
-      container_of(utctx, struct tu_device, trace_context);
-   struct tu_bo *bo = timestamps;
-   struct tu_u_trace_submission_data *submission_data = flush_data;
-
-   /* Only need to stall on results for the first entry: */
-   if (idx == 0) {
-      tu_device_wait_u_trace(device, submission_data->syncobj);
-   }
-
-   if (tu_bo_map(device, bo) != VK_SUCCESS) {
-      return U_TRACE_NO_TIMESTAMP;
-   }
-
-   uint64_t *ts = bo->map;
-
-   /* Don't translate the no-timestamp marker: */
-   if (ts[idx] == U_TRACE_NO_TIMESTAMP)
-      return U_TRACE_NO_TIMESTAMP;
-
-   return tu_device_ticks_to_ns(device, ts[idx]);
-}
-
-static void
-tu_trace_delete_flush_data(struct u_trace_context *utctx, void *flush_data)
-{
-   struct tu_device *device =
-      container_of(utctx, struct tu_device, trace_context);
-   struct tu_u_trace_submission_data *submission_data = flush_data;
-
-   tu_u_trace_submission_data_finish(device, submission_data);
-}
-
-void
-tu_copy_timestamp_buffer(struct u_trace_context *utctx, void *cmdstream,
-                         void *ts_from, uint32_t from_offset,
-                         void *ts_to, uint32_t to_offset,
-                         uint32_t count)
-{
-   struct tu_cs *cs = cmdstream;
-   struct tu_bo *bo_from = ts_from;
-   struct tu_bo *bo_to = ts_to;
-
-   tu_cs_emit_pkt7(cs, CP_MEMCPY, 5);
-   tu_cs_emit(cs, count * sizeof(uint64_t) / sizeof(uint32_t));
-   tu_cs_emit_qw(cs, bo_from->iova + from_offset * sizeof(uint64_t));
-   tu_cs_emit_qw(cs, bo_to->iova + to_offset * sizeof(uint64_t));
-}
-
-/* Special helpers instead of u_trace_begin_iterator()/u_trace_end_iterator()
- * that ignore tracepoints at the beginning/end that are part of a
- * suspend/resume chain.
- */
-static struct u_trace_iterator
-tu_cmd_begin_iterator(struct tu_cmd_buffer *cmdbuf)
-{
-   switch (cmdbuf->state.suspend_resume) {
-   case SR_IN_PRE_CHAIN:
-      return cmdbuf->trace_renderpass_end;
-   case SR_AFTER_PRE_CHAIN:
-   case SR_IN_CHAIN_AFTER_PRE_CHAIN:
-      return cmdbuf->pre_chain.trace_renderpass_end;
-   default:
-      return u_trace_begin_iterator(&cmdbuf->trace);
-   }
-}
-
-static struct u_trace_iterator
-tu_cmd_end_iterator(struct tu_cmd_buffer *cmdbuf)
-{
-   switch (cmdbuf->state.suspend_resume) {
-   case SR_IN_PRE_CHAIN:
-      return cmdbuf->trace_renderpass_end;
-   case SR_IN_CHAIN:
-   case SR_IN_CHAIN_AFTER_PRE_CHAIN:
-      return cmdbuf->trace_renderpass_start;
-   default:
-      return u_trace_end_iterator(&cmdbuf->trace);
-   }
-}
-VkResult
-tu_create_copy_timestamp_cs(struct tu_cmd_buffer *cmdbuf, struct tu_cs** cs,
-                            struct u_trace **trace_copy)
+static int
+tu_get_device_extension_index(const char *name)
 {
-   *cs = vk_zalloc(&cmdbuf->device->vk.alloc, sizeof(struct tu_cs), 8,
-                   VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
-   if (*cs == NULL) {
-      return VK_ERROR_OUT_OF_HOST_MEMORY;
-   }
-
-   tu_cs_init(*cs, cmdbuf->device, TU_CS_MODE_GROW,
-              list_length(&cmdbuf->trace.trace_chunks) * 6 + 3, "trace copy timestamp cs");
-
-   tu_cs_begin(*cs);
-
-   tu_cs_emit_wfi(*cs);
-   tu_cs_emit_pkt7(*cs, CP_WAIT_FOR_ME, 0);
-
-   *trace_copy = vk_zalloc(&cmdbuf->device->vk.alloc, sizeof(struct u_trace), 8,
-                           VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
-   if (*trace_copy == NULL) {
-      return VK_ERROR_OUT_OF_HOST_MEMORY;
+   for (unsigned i = 0; i < TU_DEVICE_EXTENSION_COUNT; ++i) {
+      if (strcmp(name, tu_device_extensions[i].extensionName) == 0)
+         return i;
    }
-
-   u_trace_init(*trace_copy, cmdbuf->trace.utctx);
-   u_trace_clone_append(tu_cmd_begin_iterator(cmdbuf),
-                        tu_cmd_end_iterator(cmdbuf),
-                        *trace_copy, *cs,
-                        tu_copy_timestamp_buffer);
-
-   tu_cs_emit_wfi(*cs);
-
-   tu_cs_end(*cs);
-
-   return VK_SUCCESS;
+   return -1;
 }
 
 VkResult
-tu_u_trace_submission_data_create(
-   struct tu_device *device,
-   struct tu_cmd_buffer **cmd_buffers,
-   uint32_t cmd_buffer_count,
-   struct tu_u_trace_submission_data **submission_data)
-{
-   *submission_data =
-      vk_zalloc(&device->vk.alloc,
-                sizeof(struct tu_u_trace_submission_data), 8,
-                VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
-   if (!(*submission_data)) {
-      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
-   }
-
-   struct tu_u_trace_submission_data *data = *submission_data;
-
-   data->cmd_trace_data =
-      vk_zalloc(&device->vk.alloc,
-                cmd_buffer_count * sizeof(struct tu_u_trace_cmd_data), 8,
-                VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
-   if (!data->cmd_trace_data) {
-      goto fail;
-   }
-
-   data->cmd_buffer_count = cmd_buffer_count;
-   data->last_buffer_with_tracepoints = -1;
-
-   for (uint32_t i = 0; i < cmd_buffer_count; ++i) {
-      struct tu_cmd_buffer *cmdbuf = cmd_buffers[i];
-
-      if (!u_trace_has_points(&cmdbuf->trace))
-         continue;
-
-      data->last_buffer_with_tracepoints = i;
-
-      if (!(cmdbuf->usage_flags & VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT)) {
-         /* A single command buffer could be submitted several times, but we
-          * already baked timestamp iova addresses and trace points are
-          * single-use. Therefor we have to copy trace points and create
-          * a new timestamp buffer on every submit of reusable command buffer.
-          */
-         if (tu_create_copy_timestamp_cs(cmdbuf,
-               &data->cmd_trace_data[i].timestamp_copy_cs,
-               &data->cmd_trace_data[i].trace) != VK_SUCCESS) {
-            goto fail;
-         }
-
-         assert(data->cmd_trace_data[i].timestamp_copy_cs->entry_count == 1);
-      } else {
-         data->cmd_trace_data[i].trace = &cmdbuf->trace;
-      }
-   }
-
-   assert(data->last_buffer_with_tracepoints != -1);
-
-   return VK_SUCCESS;
-
-fail:
-   tu_u_trace_submission_data_finish(device, data);
-   *submission_data = NULL;
-
-   return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
-}
-
-void
-tu_u_trace_submission_data_finish(
-   struct tu_device *device,
-   struct tu_u_trace_submission_data *submission_data)
-{
-   for (uint32_t i = 0; i < submission_data->cmd_buffer_count; ++i) {
-      /* Only if we had to create a copy of trace we should free it */
-      struct tu_u_trace_cmd_data *cmd_data = &submission_data->cmd_trace_data[i];
-      if (cmd_data->timestamp_copy_cs) {
-         tu_cs_finish(cmd_data->timestamp_copy_cs);
-         vk_free(&device->vk.alloc, cmd_data->timestamp_copy_cs);
-
-         u_trace_fini(cmd_data->trace);
-         vk_free(&device->vk.alloc, cmd_data->trace);
-      }
-   }
-
-   vk_free(&device->vk.alloc, submission_data->cmd_trace_data);
-   vk_free(&device->vk.alloc, submission_data->syncobj);
-   vk_free(&device->vk.alloc, submission_data);
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
 tu_CreateDevice(VkPhysicalDevice physicalDevice,
                 const VkDeviceCreateInfo *pCreateInfo,
                 const VkAllocationCallbacks *pAllocator,
@@ -1884,92 +1019,59 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
    VkResult result;
    struct tu_device *device;
-   bool custom_border_colors = false;
-   bool perf_query_pools = false;
-   bool robust_buffer_access2 = false;
-   bool border_color_without_format = false;
-   bool global_priority_query = false;
 
-   vk_foreach_struct_const(ext, pCreateInfo->pNext) {
-      switch (ext->sType) {
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_CUSTOM_BORDER_COLOR_FEATURES_EXT: {
-         const VkPhysicalDeviceCustomBorderColorFeaturesEXT *border_color_features = (const void *)ext;
-         custom_border_colors = border_color_features->customBorderColors;
-         border_color_without_format =
-            border_color_features->customBorderColorWithoutFormat;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PERFORMANCE_QUERY_FEATURES_KHR: {
-         const VkPhysicalDevicePerformanceQueryFeaturesKHR *feature =
-            (VkPhysicalDevicePerformanceQueryFeaturesKHR *)ext;
-         perf_query_pools = feature->performanceCounterQueryPools;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_ROBUSTNESS_2_FEATURES_EXT: {
-         VkPhysicalDeviceRobustness2FeaturesEXT *features = (void *)ext;
-         robust_buffer_access2 = features->robustBufferAccess2;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_GLOBAL_PRIORITY_QUERY_FEATURES_KHR: {
-         VkPhysicalDeviceGlobalPriorityQueryFeaturesKHR *features = (void *)ext;
-         global_priority_query = features->globalPriorityQuery;
-         break;
-      }
-      default:
-         break;
+   /* Check enabled features */
+   if (pCreateInfo->pEnabledFeatures) {
+      VkPhysicalDeviceFeatures supported_features;
+      tu_GetPhysicalDeviceFeatures(physicalDevice, &supported_features);
+      VkBool32 *supported_feature = (VkBool32 *) &supported_features;
+      VkBool32 *enabled_feature = (VkBool32 *) pCreateInfo->pEnabledFeatures;
+      unsigned num_features =
+         sizeof(VkPhysicalDeviceFeatures) / sizeof(VkBool32);
+      for (uint32_t i = 0; i < num_features; i++) {
+         if (enabled_feature[i] && !supported_feature[i])
+            return vk_error(physical_device->instance,
+                            VK_ERROR_FEATURE_NOT_PRESENT);
       }
    }
 
-   device = vk_zalloc2(&physical_device->instance->vk.alloc, pAllocator,
+   device = vk_zalloc2(&physical_device->instance->alloc, pAllocator,
                        sizeof(*device), 8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
    if (!device)
-      return vk_startup_errorf(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY, "OOM");
-
-   struct vk_device_dispatch_table dispatch_table;
-   vk_device_dispatch_table_from_entrypoints(
-      &dispatch_table, &tu_device_entrypoints, true);
-   vk_device_dispatch_table_from_entrypoints(
-      &dispatch_table, &wsi_device_entrypoints, false);
-
-   result = vk_device_init(&device->vk, &physical_device->vk,
-                           &dispatch_table, pCreateInfo, pAllocator);
-   if (result != VK_SUCCESS) {
-      vk_free(&device->vk.alloc, device);
-      return vk_startup_errorf(physical_device->instance, result,
-                               "vk_device_init failed");
-   }
+      return vk_error(physical_device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
+   device->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
    device->instance = physical_device->instance;
    device->physical_device = physical_device;
-   device->fd = physical_device->local_fd;
-   device->vk.command_buffer_ops = &tu_cmd_buffer_ops;
-   device->vk.check_status = tu_device_check_status;
 
-   mtx_init(&device->bo_mutex, mtx_plain);
-   mtx_init(&device->pipeline_mutex, mtx_plain);
-   mtx_init(&device->autotune_mutex, mtx_plain);
-   u_rwlock_init(&device->dma_bo_lock);
-   pthread_mutex_init(&device->submit_mutex, NULL);
+   if (pAllocator)
+      device->alloc = *pAllocator;
+   else
+      device->alloc = physical_device->instance->alloc;
 
-   if (device->instance->debug_flags & TU_DEBUG_BOS)
-      device->bo_sizes = _mesa_hash_table_create(NULL, _mesa_hash_string, _mesa_key_string_equal);
+   for (uint32_t i = 0; i < pCreateInfo->enabledExtensionCount; i++) {
+      const char *ext_name = pCreateInfo->ppEnabledExtensionNames[i];
+      int index = tu_get_device_extension_index(ext_name);
+      if (index < 0 ||
+          !physical_device->supported_extensions.extensions[index]) {
+         vk_free(&device->alloc, device);
+         return vk_error(physical_device->instance,
+                         VK_ERROR_EXTENSION_NOT_PRESENT);
+      }
 
-#ifndef TU_USE_KGSL
-   vk_device_set_drm_fd(&device->vk, device->fd);
-#endif
+      device->enabled_extensions.extensions[index] = true;
+   }
 
    for (unsigned i = 0; i < pCreateInfo->queueCreateInfoCount; i++) {
       const VkDeviceQueueCreateInfo *queue_create =
          &pCreateInfo->pQueueCreateInfos[i];
       uint32_t qfi = queue_create->queueFamilyIndex;
       device->queues[qfi] = vk_alloc(
-         &device->vk.alloc, queue_create->queueCount * sizeof(struct tu_queue),
+         &device->alloc, queue_create->queueCount * sizeof(struct tu_queue),
          8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
       if (!device->queues[qfi]) {
-         result = vk_startup_errorf(physical_device->instance,
-                                    VK_ERROR_OUT_OF_HOST_MEMORY,
-                                    "OOM");
-         goto fail_queues;
+         result = VK_ERROR_OUT_OF_HOST_MEMORY;
+         goto fail;
       }
 
       memset(device->queues[qfi], 0,
@@ -1978,221 +1080,50 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
       device->queue_count[qfi] = queue_create->queueCount;
 
       for (unsigned q = 0; q < queue_create->queueCount; q++) {
-         result = tu_queue_init(device, &device->queues[qfi][q], q,
-                                queue_create, global_priority_query);
-         if (result != VK_SUCCESS) {
-            device->queue_count[qfi] = q;
-            goto fail_queues;
-         }
-      }
-   }
-
-   device->compiler =
-      ir3_compiler_create(NULL, &physical_device->dev_id,
-                          &(struct ir3_compiler_options) {
-                              .robust_buffer_access2 = robust_buffer_access2,
-                              .push_ubo_with_preamble = true,
-                              .disable_cache = true,
-                           });
-   if (!device->compiler) {
-      result = vk_startup_errorf(physical_device->instance,
-                                 VK_ERROR_INITIALIZATION_FAILED,
-                                 "failed to initialize ir3 compiler");
-      goto fail_queues;
-   }
-
-   /* Initialize sparse array for refcounting imported BOs */
-   util_sparse_array_init(&device->bo_map, sizeof(struct tu_bo), 512);
-
-   /* initial sizes, these will increase if there is overflow */
-   device->vsc_draw_strm_pitch = 0x1000 + VSC_PAD;
-   device->vsc_prim_strm_pitch = 0x4000 + VSC_PAD;
-
-   uint32_t global_size = sizeof(struct tu6_global);
-   if (custom_border_colors)
-      global_size += TU_BORDER_COLOR_COUNT * sizeof(struct bcolor_entry);
-
-   tu_bo_suballocator_init(&device->pipeline_suballoc, device,
-                           128 * 1024, TU_BO_ALLOC_GPU_READ_ONLY | TU_BO_ALLOC_ALLOW_DUMP);
-   tu_bo_suballocator_init(&device->autotune_suballoc, device,
-                           128 * 1024, 0);
-
-   result = tu_bo_init_new(device, &device->global_bo, global_size,
-                           TU_BO_ALLOC_ALLOW_DUMP, "global");
-   if (result != VK_SUCCESS) {
-      vk_startup_errorf(device->instance, result, "BO init");
-      goto fail_global_bo;
-   }
-
-   result = tu_bo_map(device, device->global_bo);
-   if (result != VK_SUCCESS) {
-      vk_startup_errorf(device->instance, result, "BO map");
-      goto fail_global_bo_map;
-   }
-
-   struct tu6_global *global = device->global_bo->map;
-   tu_init_clear_blit_shaders(device);
-   global->predicate = 0;
-   global->vtx_stats_query_not_running = 1;
-   global->dbg_one = (uint32_t)-1;
-   global->dbg_gmem_total_loads = 0;
-   global->dbg_gmem_taken_loads = 0;
-   global->dbg_gmem_total_stores = 0;
-   global->dbg_gmem_taken_stores = 0;
-   for (int i = 0; i < TU_BORDER_COLOR_BUILTIN; i++) {
-      VkClearColorValue border_color = vk_border_color_value(i);
-      tu6_pack_border_color(&global->bcolor_builtin[i], &border_color,
-                            vk_border_color_is_int(i));
-   }
-
-   /* initialize to ones so ffs can be used to find unused slots */
-   BITSET_ONES(device->custom_border_color);
-
-   result = tu_init_dynamic_rendering(device);
-   if (result != VK_SUCCESS) {
-      vk_startup_errorf(device->instance, result, "dynamic rendering");
-      goto fail_dynamic_rendering;
-   }
-
-   struct vk_pipeline_cache_create_info pcc_info = { };
-   device->mem_cache = vk_pipeline_cache_create(&device->vk, &pcc_info,
-                                                false);
-   if (!device->mem_cache) {
-      result = VK_ERROR_OUT_OF_HOST_MEMORY;
-      vk_startup_errorf(device->instance, result, "create pipeline cache failed");
-      goto fail_pipeline_cache;
-   }
-
-   if (perf_query_pools) {
-      /* Prepare command streams setting pass index to the PERF_CNTRS_REG
-       * from 0 to 31. One of these will be picked up at cmd submit time
-       * when the perf query is executed.
-       */
-      struct tu_cs *cs;
-
-      if (!(device->perfcntrs_pass_cs = calloc(1, sizeof(struct tu_cs)))) {
-         result = vk_startup_errorf(device->instance,
-               VK_ERROR_OUT_OF_HOST_MEMORY, "OOM");
-         goto fail_perfcntrs_pass_alloc;
-      }
-
-      device->perfcntrs_pass_cs_entries = calloc(32, sizeof(struct tu_cs_entry));
-      if (!device->perfcntrs_pass_cs_entries) {
-         result = vk_startup_errorf(device->instance,
-               VK_ERROR_OUT_OF_HOST_MEMORY, "OOM");
-         goto fail_perfcntrs_pass_entries_alloc;
-      }
-
-      cs = device->perfcntrs_pass_cs;
-      tu_cs_init(cs, device, TU_CS_MODE_SUB_STREAM, 96, "perfcntrs cs");
-
-      for (unsigned i = 0; i < 32; i++) {
-         struct tu_cs sub_cs;
-
-         result = tu_cs_begin_sub_stream(cs, 3, &sub_cs);
-         if (result != VK_SUCCESS) {
-            vk_startup_errorf(device->instance, result,
-                  "failed to allocate commands streams");
-            goto fail_prepare_perfcntrs_pass_cs;
-         }
-
-         tu_cs_emit_regs(&sub_cs, A6XX_CP_SCRATCH_REG(PERF_CNTRS_REG, 1 << i));
-         tu_cs_emit_pkt7(&sub_cs, CP_WAIT_FOR_ME, 0);
-
-         device->perfcntrs_pass_cs_entries[i] = tu_cs_end_sub_stream(cs, &sub_cs);
+         result = tu_queue_init(device, &device->queues[qfi][q], qfi, q,
+                                queue_create->flags);
+         if (result != VK_SUCCESS)
+            goto fail;
       }
    }
 
-   /* Initialize a condition variable for timeline semaphore */
-   pthread_condattr_t condattr;
-   if (pthread_condattr_init(&condattr) != 0) {
-      result = vk_startup_errorf(physical_device->instance,
-                                 VK_ERROR_INITIALIZATION_FAILED,
-                                 "pthread condattr init");
-      goto fail_timeline_cond;
-   }
-   if (pthread_condattr_setclock(&condattr, CLOCK_MONOTONIC) != 0) {
-      pthread_condattr_destroy(&condattr);
-      result = vk_startup_errorf(physical_device->instance,
-                                 VK_ERROR_INITIALIZATION_FAILED,
-                                 "pthread condattr clock setup");
-      goto fail_timeline_cond;
-   }
-   if (pthread_cond_init(&device->timeline_cond, &condattr) != 0) {
-      pthread_condattr_destroy(&condattr);
-      result = vk_startup_errorf(physical_device->instance,
-                                 VK_ERROR_INITIALIZATION_FAILED,
-                                 "pthread cond init");
-      goto fail_timeline_cond;
-   }
-   pthread_condattr_destroy(&condattr);
-
-   result = tu_autotune_init(&device->autotune, device);
-   if (result != VK_SUCCESS) {
-      goto fail_timeline_cond;
-   }
-
-   for (unsigned i = 0; i < ARRAY_SIZE(device->scratch_bos); i++)
-      mtx_init(&device->scratch_bos[i].construct_mtx, mtx_plain);
-
-   mtx_init(&device->fiber_pvtmem_bo.mtx, mtx_plain);
-   mtx_init(&device->wave_pvtmem_bo.mtx, mtx_plain);
-
-   mtx_init(&device->mutex, mtx_plain);
-
-   device->use_z24uint_s8uint =
-      physical_device->info->a6xx.has_z24uint_s8uint &&
-      !border_color_without_format;
-
-   tu_gpu_tracepoint_config_variable();
+   device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
+   if (!device->compiler)
+      goto fail;
 
-   device->submit_count = 0;
-   u_trace_context_init(&device->trace_context, device,
-                     tu_trace_create_ts_buffer,
-                     tu_trace_destroy_ts_buffer,
-                     tu_trace_record_ts,
-                     tu_trace_read_ts,
-                     tu_trace_delete_flush_data);
+   VkPipelineCacheCreateInfo ci;
+   ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
+   ci.pNext = NULL;
+   ci.flags = 0;
+   ci.pInitialData = NULL;
+   ci.initialDataSize = 0;
+   VkPipelineCache pc;
+   result =
+      tu_CreatePipelineCache(tu_device_to_handle(device), &ci, NULL, &pc);
+   if (result != VK_SUCCESS)
+      goto fail;
 
-   tu_breadcrumbs_init(device);
+   device->mem_cache = tu_pipeline_cache_from_handle(pc);
 
    *pDevice = tu_device_to_handle(device);
    return VK_SUCCESS;
 
-fail_timeline_cond:
-fail_prepare_perfcntrs_pass_cs:
-   free(device->perfcntrs_pass_cs_entries);
-   tu_cs_finish(device->perfcntrs_pass_cs);
-fail_perfcntrs_pass_entries_alloc:
-   free(device->perfcntrs_pass_cs);
-fail_perfcntrs_pass_alloc:
-   vk_pipeline_cache_destroy(device->mem_cache, &device->vk.alloc);
-fail_pipeline_cache:
-   tu_destroy_dynamic_rendering(device);
-fail_dynamic_rendering:
-   tu_destroy_clear_blit_shaders(device);
-fail_global_bo_map:
-   tu_bo_finish(device, device->global_bo);
-   vk_free(&device->vk.alloc, device->bo_list);
-fail_global_bo:
-   ir3_compiler_destroy(device->compiler);
-   util_sparse_array_finish(&device->bo_map);
-
-fail_queues:
+fail:
    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
       for (unsigned q = 0; q < device->queue_count[i]; q++)
          tu_queue_finish(&device->queues[i][q]);
-      if (device->queues[i])
-         vk_free(&device->vk.alloc, device->queues[i]);
+      if (device->queue_count[i])
+         vk_free(&device->alloc, device->queues[i]);
    }
 
-   u_rwlock_destroy(&device->dma_bo_lock);
-   vk_device_finish(&device->vk);
-   vk_free(&device->vk.alloc, device);
+   if (device->compiler)
+      ralloc_free(device->compiler);
+
+   vk_free(&device->alloc, device);
    return result;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
@@ -2200,158 +1131,231 @@ tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
    if (!device)
       return;
 
-   tu_breadcrumbs_finish(device);
-
-   u_trace_context_fini(&device->trace_context);
-
    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
       for (unsigned q = 0; q < device->queue_count[i]; q++)
          tu_queue_finish(&device->queues[i][q]);
       if (device->queue_count[i])
-         vk_free(&device->vk.alloc, device->queues[i]);
+         vk_free(&device->alloc, device->queues[i]);
    }
 
-   for (unsigned i = 0; i < ARRAY_SIZE(device->scratch_bos); i++) {
-      if (device->scratch_bos[i].initialized)
-         tu_bo_finish(device, device->scratch_bos[i].bo);
-   }
+   /* the compiler does not use pAllocator */
+   ralloc_free(device->compiler);
 
-   if (device->fiber_pvtmem_bo.bo)
-      tu_bo_finish(device, device->fiber_pvtmem_bo.bo);
-   
-   if (device->wave_pvtmem_bo.bo)
-      tu_bo_finish(device, device->wave_pvtmem_bo.bo);
+   VkPipelineCache pc = tu_pipeline_cache_to_handle(device->mem_cache);
+   tu_DestroyPipelineCache(tu_device_to_handle(device), pc, NULL);
 
-   tu_destroy_clear_blit_shaders(device);
-
-   tu_destroy_dynamic_rendering(device);
+   vk_free(&device->alloc, device);
+}
 
-   ir3_compiler_destroy(device->compiler);
+VkResult
+tu_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount,
+                                    VkLayerProperties *pProperties)
+{
+   *pPropertyCount = 0;
+   return VK_SUCCESS;
+}
 
-   vk_pipeline_cache_destroy(device->mem_cache, &device->vk.alloc);
+VkResult
+tu_EnumerateDeviceLayerProperties(VkPhysicalDevice physicalDevice,
+                                  uint32_t *pPropertyCount,
+                                  VkLayerProperties *pProperties)
+{
+   *pPropertyCount = 0;
+   return VK_SUCCESS;
+}
 
-   if (device->perfcntrs_pass_cs) {
-      free(device->perfcntrs_pass_cs_entries);
-      tu_cs_finish(device->perfcntrs_pass_cs);
-      free(device->perfcntrs_pass_cs);
+void
+tu_GetDeviceQueue2(VkDevice _device,
+                   const VkDeviceQueueInfo2 *pQueueInfo,
+                   VkQueue *pQueue)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   struct tu_queue *queue;
+
+   queue =
+      &device->queues[pQueueInfo->queueFamilyIndex][pQueueInfo->queueIndex];
+   if (pQueueInfo->flags != queue->flags) {
+      /* From the Vulkan 1.1.70 spec:
+       *
+       * "The queue returned by vkGetDeviceQueue2 must have the same
+       * flags value from this structure as that used at device
+       * creation time in a VkDeviceQueueCreateInfo instance. If no
+       * matching flags were specified at device creation time then
+       * pQueue will return VK_NULL_HANDLE."
+       */
+      *pQueue = VK_NULL_HANDLE;
+      return;
    }
 
-   tu_autotune_fini(&device->autotune, device);
-
-   tu_bo_suballocator_finish(&device->pipeline_suballoc);
-   tu_bo_suballocator_finish(&device->autotune_suballoc);
+   *pQueue = tu_queue_to_handle(queue);
+}
 
-   util_sparse_array_finish(&device->bo_map);
-   u_rwlock_destroy(&device->dma_bo_lock);
+void
+tu_GetDeviceQueue(VkDevice _device,
+                  uint32_t queueFamilyIndex,
+                  uint32_t queueIndex,
+                  VkQueue *pQueue)
+{
+   const VkDeviceQueueInfo2 info =
+      (VkDeviceQueueInfo2) { .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_INFO_2,
+                             .queueFamilyIndex = queueFamilyIndex,
+                             .queueIndex = queueIndex };
 
-   pthread_cond_destroy(&device->timeline_cond);
-   _mesa_hash_table_destroy(device->bo_sizes, NULL);
-   vk_free(&device->vk.alloc, device->bo_list);
-   vk_device_finish(&device->vk);
-   vk_free(&device->vk.alloc, device);
+   tu_GetDeviceQueue2(_device, &info, pQueue);
 }
 
 VkResult
-tu_get_scratch_bo(struct tu_device *dev, uint64_t size, struct tu_bo **bo)
+tu_QueueSubmit(VkQueue _queue,
+               uint32_t submitCount,
+               const VkSubmitInfo *pSubmits,
+               VkFence _fence)
 {
-   unsigned size_log2 = MAX2(util_logbase2_ceil64(size), MIN_SCRATCH_BO_SIZE_LOG2);
-   unsigned index = size_log2 - MIN_SCRATCH_BO_SIZE_LOG2;
-   assert(index < ARRAY_SIZE(dev->scratch_bos));
+   TU_FROM_HANDLE(tu_queue, queue, _queue);
+
+   for (uint32_t i = 0; i < submitCount; ++i) {
+      const VkSubmitInfo *submit = pSubmits + i;
+      const bool last_submit = (i == submitCount - 1);
+      struct tu_bo_list bo_list;
+      tu_bo_list_init(&bo_list);
+
+      uint32_t entry_count = 0;
+      for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
+         TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
+         entry_count += cmdbuf->cs.entry_count;
+      }
+
+      struct drm_msm_gem_submit_cmd cmds[entry_count];
+      uint32_t entry_idx = 0;
+      for (uint32_t j = 0; j < submit->commandBufferCount; ++j) {
+         TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, submit->pCommandBuffers[j]);
+         struct tu_cs *cs = &cmdbuf->cs;
+         for (unsigned i = 0; i < cs->entry_count; ++i, ++entry_idx) {
+            cmds[entry_idx].type = MSM_SUBMIT_CMD_BUF;
+            cmds[entry_idx].submit_idx =
+               tu_bo_list_add(&bo_list, cs->entries[i].bo,
+                              MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+            cmds[entry_idx].submit_offset = cs->entries[i].offset;
+            cmds[entry_idx].size = cs->entries[i].size;
+            cmds[entry_idx].pad = 0;
+            cmds[entry_idx].nr_relocs = 0;
+            cmds[entry_idx].relocs = 0;
+         }
 
-   for (unsigned i = index; i < ARRAY_SIZE(dev->scratch_bos); i++) {
-      if (p_atomic_read(&dev->scratch_bos[i].initialized)) {
-         /* Fast path: just return the already-allocated BO. */
-         *bo = dev->scratch_bos[i].bo;
-         return VK_SUCCESS;
+         tu_bo_list_merge(&bo_list, &cmdbuf->bo_list);
       }
-   }
 
-   /* Slow path: actually allocate the BO. We take a lock because the process
-    * of allocating it is slow, and we don't want to block the CPU while it
-    * finishes.
-   */
-   mtx_lock(&dev->scratch_bos[index].construct_mtx);
+      uint32_t flags = MSM_PIPE_3D0;
+      if (last_submit) {
+         flags |= MSM_SUBMIT_FENCE_FD_OUT;
+      }
 
-   /* Another thread may have allocated it already while we were waiting on
-    * the lock. We need to check this in order to avoid double-allocating.
-    */
-   if (dev->scratch_bos[index].initialized) {
-      mtx_unlock(&dev->scratch_bos[index].construct_mtx);
-      *bo = dev->scratch_bos[index].bo;
-      return VK_SUCCESS;
-   }
+      struct drm_msm_gem_submit req = {
+         .flags = flags,
+         .queueid = queue->msm_queue_id,
+         .bos = (uint64_t)(uintptr_t) bo_list.bo_infos,
+         .nr_bos = bo_list.count,
+         .cmds = (uint64_t)(uintptr_t)cmds,
+         .nr_cmds = entry_count,
+      };
 
-   unsigned bo_size = 1ull << size_log2;
-   VkResult result = tu_bo_init_new(dev, &dev->scratch_bos[index].bo, bo_size,
-                                    TU_BO_ALLOC_NO_FLAGS, "scratch");
-   if (result != VK_SUCCESS) {
-      mtx_unlock(&dev->scratch_bos[index].construct_mtx);
-      return result;
-   }
+      int ret = drmCommandWriteRead(queue->device->physical_device->local_fd,
+                                    DRM_MSM_GEM_SUBMIT,
+                                    &req, sizeof(req));
+      if (ret) {
+         fprintf(stderr, "submit failed: %s\n", strerror(errno));
+         abort();
+      }
 
-   p_atomic_set(&dev->scratch_bos[index].initialized, true);
+      tu_bo_list_destroy(&bo_list);
 
-   mtx_unlock(&dev->scratch_bos[index].construct_mtx);
+      if (last_submit) {
+         /* no need to merge fences as queue execution is serialized */
+         tu_fence_update_fd(&queue->submit_fence, req.fence_fd);
+      }
+   }
 
-   *bo = dev->scratch_bos[index].bo;
-   return VK_SUCCESS;
-}
+   if (_fence != VK_NULL_HANDLE) {
+      TU_FROM_HANDLE(tu_fence, fence, _fence);
+      tu_fence_copy(fence, &queue->submit_fence);
+   }
 
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_EnumerateInstanceLayerProperties(uint32_t *pPropertyCount,
-                                    VkLayerProperties *pProperties)
-{
-   *pPropertyCount = 0;
    return VK_SUCCESS;
 }
 
-/* Only used for kgsl since drm started using common implementation */
-#ifdef TU_USE_KGSL
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_QueueWaitIdle(VkQueue _queue)
 {
    TU_FROM_HANDLE(tu_queue, queue, _queue);
 
-   if (vk_device_is_lost(&queue->device->vk))
-      return VK_ERROR_DEVICE_LOST;
+   tu_fence_wait_idle(&queue->submit_fence);
 
-   if (queue->fence < 0)
-      return VK_SUCCESS;
-
-   struct pollfd fds = { .fd = queue->fence, .events = POLLIN };
-   int ret;
-   do {
-      ret = poll(&fds, 1, -1);
-   } while (ret == -1 && (errno == EINTR || errno == EAGAIN));
+   return VK_SUCCESS;
+}
 
-   /* TODO: otherwise set device lost ? */
-   assert(ret == 1 && !(fds.revents & (POLLERR | POLLNVAL)));
+VkResult
+tu_DeviceWaitIdle(VkDevice _device)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
 
-   close(queue->fence);
-   queue->fence = -1;
+   for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
+      for (unsigned q = 0; q < device->queue_count[i]; q++) {
+         tu_QueueWaitIdle(tu_queue_to_handle(&device->queues[i][q]));
+      }
+   }
    return VK_SUCCESS;
 }
-#endif
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_EnumerateInstanceExtensionProperties(const char *pLayerName,
                                         uint32_t *pPropertyCount,
                                         VkExtensionProperties *pProperties)
 {
+   VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
+
+   /* We spport no lyaers */
    if (pLayerName)
       return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
 
-   return vk_enumerate_instance_extension_properties(
-      &tu_instance_extensions_supported, pPropertyCount, pProperties);
+   for (int i = 0; i < TU_INSTANCE_EXTENSION_COUNT; i++) {
+      if (tu_supported_instance_extensions.extensions[i]) {
+         vk_outarray_append(&out, prop) { *prop = tu_instance_extensions[i]; }
+      }
+   }
+
+   return vk_outarray_status(&out);
 }
 
-VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
+VkResult
+tu_EnumerateDeviceExtensionProperties(VkPhysicalDevice physicalDevice,
+                                      const char *pLayerName,
+                                      uint32_t *pPropertyCount,
+                                      VkExtensionProperties *pProperties)
+{
+   /* We spport no lyaers */
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
+   VK_OUTARRAY_MAKE(out, pProperties, pPropertyCount);
+
+   /* We spport no lyaers */
+   if (pLayerName)
+      return vk_error(NULL, VK_ERROR_LAYER_NOT_PRESENT);
+
+   for (int i = 0; i < TU_DEVICE_EXTENSION_COUNT; i++) {
+      if (device->supported_extensions.extensions[i]) {
+         vk_outarray_append(&out, prop) { *prop = tu_device_extensions[i]; }
+      }
+   }
+
+   return vk_outarray_status(&out);
+}
+
+PFN_vkVoidFunction
 tu_GetInstanceProcAddr(VkInstance _instance, const char *pName)
 {
    TU_FROM_HANDLE(tu_instance, instance, _instance);
-   return vk_instance_get_proc_addr(&instance->vk,
-                                    &tu_instance_entrypoints,
-                                    pName);
+
+   return tu_lookup_entrypoint_checked(
+      pName, instance ? instance->api_version : 0,
+      instance ? &instance->enabled_extensions : NULL, NULL);
 }
 
 /* The loader wants us to expose a second GetInstanceProcAddr function
@@ -2368,30 +1372,22 @@ vk_icdGetInstanceProcAddr(VkInstance instance, const char *pName)
    return tu_GetInstanceProcAddr(instance, pName);
 }
 
-/* With version 4+ of the loader interface the ICD should expose
- * vk_icdGetPhysicalDeviceProcAddr()
- */
-PUBLIC
-VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
-vk_icdGetPhysicalDeviceProcAddr(VkInstance  _instance,
-                                const char* pName);
-
 PFN_vkVoidFunction
-vk_icdGetPhysicalDeviceProcAddr(VkInstance  _instance,
-                                const char* pName)
+tu_GetDeviceProcAddr(VkDevice _device, const char *pName)
 {
-   TU_FROM_HANDLE(tu_instance, instance, _instance);
+   TU_FROM_HANDLE(tu_device, device, _device);
 
-   return vk_instance_get_physical_device_proc_addr(&instance->vk, pName);
+   return tu_lookup_entrypoint_checked(pName, device->instance->api_version,
+                                       &device->instance->enabled_extensions,
+                                       &device->enabled_extensions);
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_AllocateMemory(VkDevice _device,
-                  const VkMemoryAllocateInfo *pAllocateInfo,
-                  const VkAllocationCallbacks *pAllocator,
-                  VkDeviceMemory *pMem)
+static VkResult
+tu_alloc_memory(struct tu_device *device,
+                const VkMemoryAllocateInfo *pAllocateInfo,
+                const VkAllocationCallbacks *pAllocator,
+                VkDeviceMemory *pMem)
 {
-   TU_FROM_HANDLE(tu_device, device, _device);
    struct tu_device_memory *mem;
    VkResult result;
 
@@ -2403,15 +1399,10 @@ tu_AllocateMemory(VkDevice _device,
       return VK_SUCCESS;
    }
 
-   struct tu_memory_heap *mem_heap = &device->physical_device->heap;
-   uint64_t mem_heap_used = p_atomic_read(&mem_heap->used);
-   if (mem_heap_used > mem_heap->size)
-      return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
-
-   mem = vk_object_alloc(&device->vk, pAllocator, sizeof(*mem),
-                         VK_OBJECT_TYPE_DEVICE_MEMORY);
+   mem = vk_alloc2(&device->alloc, pAllocator, sizeof(*mem), 8,
+                   VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (mem == NULL)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    const VkImportMemoryFdInfoKHR *fd_info =
       vk_find_struct_const(pAllocateInfo->pNext, IMPORT_MEMORY_FD_INFO_KHR);
@@ -2436,69 +1427,37 @@ tu_AllocateMemory(VkDevice _device,
          close(fd_info->fd);
       }
    } else {
-      uint64_t client_address = 0;
-      enum tu_bo_alloc_flags alloc_flags = TU_BO_ALLOC_NO_FLAGS;
-
-      const VkMemoryOpaqueCaptureAddressAllocateInfo *replay_info =
-         vk_find_struct_const(pAllocateInfo->pNext,
-                              MEMORY_OPAQUE_CAPTURE_ADDRESS_ALLOCATE_INFO);
-      if (replay_info && replay_info->opaqueCaptureAddress) {
-         client_address = replay_info->opaqueCaptureAddress;
-         alloc_flags |= TU_BO_ALLOC_REPLAYABLE;
-      }
-
-      const VkMemoryAllocateFlagsInfo *flags_info = vk_find_struct_const(
-         pAllocateInfo->pNext, MEMORY_ALLOCATE_FLAGS_INFO);
-      if (flags_info &&
-          (flags_info->flags &
-           VK_MEMORY_ALLOCATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT)) {
-         alloc_flags |= TU_BO_ALLOC_REPLAYABLE;
-      }
-
-      char name[64] = "vkAllocateMemory()";
-      if (device->bo_sizes)
-         snprintf(name, ARRAY_SIZE(name), "vkAllocateMemory(%ldkb)",
-                  (long)DIV_ROUND_UP(pAllocateInfo->allocationSize, 1024));
-      result = tu_bo_init_new_explicit_iova(
-         device, &mem->bo, pAllocateInfo->allocationSize, client_address,
-         alloc_flags, name);
-   }
-
-   if (result == VK_SUCCESS) {
-      mem_heap_used = p_atomic_add_return(&mem_heap->used, mem->bo->size);
-      if (mem_heap_used > mem_heap->size) {
-         p_atomic_add(&mem_heap->used, -mem->bo->size);
-         tu_bo_finish(device, mem->bo);
-         result = vk_errorf(device, VK_ERROR_OUT_OF_DEVICE_MEMORY,
-                            "Out of heap memory");
-      }
+      result =
+         tu_bo_init_new(device, &mem->bo, pAllocateInfo->allocationSize);
    }
 
    if (result != VK_SUCCESS) {
-      vk_object_free(&device->vk, pAllocator, mem);
+      vk_free2(&device->alloc, pAllocator, mem);
       return result;
    }
 
-   /* Track in the device whether our BO list contains any implicit-sync BOs, so
-    * we can suppress implicit sync on non-WSI usage.
-    */
-   const struct wsi_memory_allocate_info *wsi_info =
-      vk_find_struct_const(pAllocateInfo->pNext, WSI_MEMORY_ALLOCATE_INFO_MESA);
-   if (wsi_info && wsi_info->implicit_sync) {
-      mtx_lock(&device->bo_mutex);
-      if (!mem->bo->implicit_sync) {
-         mem->bo->implicit_sync = true;
-         device->implicit_sync_bo_count++;
-      }
-      mtx_unlock(&device->bo_mutex);
-   }
+   mem->size = pAllocateInfo->allocationSize;
+   mem->type_index = pAllocateInfo->memoryTypeIndex;
+
+   mem->map = NULL;
+   mem->user_ptr = NULL;
 
    *pMem = tu_device_memory_to_handle(mem);
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+VkResult
+tu_AllocateMemory(VkDevice _device,
+                  const VkMemoryAllocateInfo *pAllocateInfo,
+                  const VkAllocationCallbacks *pAllocator,
+                  VkDeviceMemory *pMem)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   return tu_alloc_memory(device, pAllocateInfo, pAllocator, pMem);
+}
+
+void
 tu_FreeMemory(VkDevice _device,
               VkDeviceMemory _mem,
               const VkAllocationCallbacks *pAllocator)
@@ -2509,12 +1468,11 @@ tu_FreeMemory(VkDevice _device,
    if (mem == NULL)
       return;
 
-   p_atomic_add(&device->physical_device->heap.used, -mem->bo->size);
-   tu_bo_finish(device, mem->bo);
-   vk_object_free(&device->vk, pAllocator, mem);
+   tu_bo_finish(device, &mem->bo);
+   vk_free2(&device->alloc, pAllocator, mem);
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_MapMemory(VkDevice _device,
              VkDeviceMemory _memory,
              VkDeviceSize offset,
@@ -2531,23 +1489,31 @@ tu_MapMemory(VkDevice _device,
       return VK_SUCCESS;
    }
 
-   if (!mem->bo->map) {
-      result = tu_bo_map(device, mem->bo);
+   if (mem->user_ptr) {
+      *ppData = mem->user_ptr;
+   } else if (!mem->map) {
+      result = tu_bo_map(device, &mem->bo);
       if (result != VK_SUCCESS)
          return result;
+      *ppData = mem->map = mem->bo.map;
+   } else
+      *ppData = mem->map;
+
+   if (*ppData) {
+      *ppData += offset;
+      return VK_SUCCESS;
    }
 
-   *ppData = mem->bo->map + offset;
-   return VK_SUCCESS;
+   return vk_error(device->instance, VK_ERROR_MEMORY_MAP_FAILED);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_UnmapMemory(VkDevice _device, VkDeviceMemory _memory)
 {
-   /* TODO: unmap here instead of waiting for FreeMemory */
+   /* I do not see any unmapping done by the freedreno Gallium driver. */
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_FlushMappedMemoryRanges(VkDevice _device,
                            uint32_t memoryRangeCount,
                            const VkMappedMemoryRange *pMemoryRanges)
@@ -2555,7 +1521,7 @@ tu_FlushMappedMemoryRanges(VkDevice _device,
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_InvalidateMappedMemoryRanges(VkDevice _device,
                                 uint32_t memoryRangeCount,
                                 const VkMappedMemoryRange *pMemoryRanges)
@@ -2563,52 +1529,71 @@ tu_InvalidateMappedMemoryRanges(VkDevice _device,
    return VK_SUCCESS;
 }
 
-static void
-tu_get_buffer_memory_requirements(uint64_t size,
-                                  VkMemoryRequirements2 *pMemoryRequirements)
+void
+tu_GetBufferMemoryRequirements(VkDevice _device,
+                               VkBuffer _buffer,
+                               VkMemoryRequirements *pMemoryRequirements)
 {
-   pMemoryRequirements->memoryRequirements = (VkMemoryRequirements) {
-      .memoryTypeBits = 1,
-      .alignment = 64,
-      .size = MAX2(align64(size, 64), size),
-   };
+   TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
 
-   vk_foreach_struct(ext, pMemoryRequirements->pNext) {
-      switch (ext->sType) {
-      case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
-         VkMemoryDedicatedRequirements *req =
-            (VkMemoryDedicatedRequirements *) ext;
-         req->requiresDedicatedAllocation = false;
-         req->prefersDedicatedAllocation = req->requiresDedicatedAllocation;
-         break;
-      }
-      default:
-         break;
-      }
-   }
+   pMemoryRequirements->memoryTypeBits = 1;
+   pMemoryRequirements->alignment = 16;
+   pMemoryRequirements->size =
+      align64(buffer->size, pMemoryRequirements->alignment);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_GetBufferMemoryRequirements2(
    VkDevice device,
    const VkBufferMemoryRequirementsInfo2 *pInfo,
    VkMemoryRequirements2 *pMemoryRequirements)
 {
-   TU_FROM_HANDLE(tu_buffer, buffer, pInfo->buffer);
+   tu_GetBufferMemoryRequirements(device, pInfo->buffer,
+                                  &pMemoryRequirements->memoryRequirements);
+}
+
+void
+tu_GetImageMemoryRequirements(VkDevice _device,
+                              VkImage _image,
+                              VkMemoryRequirements *pMemoryRequirements)
+{
+   TU_FROM_HANDLE(tu_image, image, _image);
+
+   pMemoryRequirements->memoryTypeBits = 1;
+   pMemoryRequirements->size = image->size;
+   pMemoryRequirements->alignment = image->alignment;
+}
+
+void
+tu_GetImageMemoryRequirements2(VkDevice device,
+                               const VkImageMemoryRequirementsInfo2 *pInfo,
+                               VkMemoryRequirements2 *pMemoryRequirements)
+{
+   tu_GetImageMemoryRequirements(device, pInfo->image,
+                                 &pMemoryRequirements->memoryRequirements);
+}
 
-   tu_get_buffer_memory_requirements(buffer->vk.size, pMemoryRequirements);
+void
+tu_GetImageSparseMemoryRequirements(
+   VkDevice device,
+   VkImage image,
+   uint32_t *pSparseMemoryRequirementCount,
+   VkSparseImageMemoryRequirements *pSparseMemoryRequirements)
+{
+   tu_stub();
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_GetDeviceBufferMemoryRequirements(
+void
+tu_GetImageSparseMemoryRequirements2(
    VkDevice device,
-   const VkDeviceBufferMemoryRequirements *pInfo,
-   VkMemoryRequirements2 *pMemoryRequirements)
+   const VkImageSparseMemoryRequirementsInfo2 *pInfo,
+   uint32_t *pSparseMemoryRequirementCount,
+   VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
 {
-   tu_get_buffer_memory_requirements(pInfo->pCreateInfo->size, pMemoryRequirements);
+   tu_stub();
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_GetDeviceMemoryCommitment(VkDevice device,
                              VkDeviceMemory memory,
                              VkDeviceSize *pCommittedMemoryInBytes)
@@ -2616,7 +1601,7 @@ tu_GetDeviceMemoryCommitment(VkDevice device,
    *pCommittedMemoryInBytes = 0;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_BindBufferMemory2(VkDevice device,
                      uint32_t bindInfoCount,
                      const VkBindBufferMemoryInfo *pBindInfos)
@@ -2626,8 +1611,8 @@ tu_BindBufferMemory2(VkDevice device,
       TU_FROM_HANDLE(tu_buffer, buffer, pBindInfos[i].buffer);
 
       if (mem) {
-         buffer->bo = mem->bo;
-         buffer->iova = mem->bo->iova + pBindInfos[i].memoryOffset;
+         buffer->bo = &mem->bo;
+         buffer->bo_offset = pBindInfos[i].memoryOffset;
       } else {
          buffer->bo = NULL;
       }
@@ -2635,7 +1620,23 @@ tu_BindBufferMemory2(VkDevice device,
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
+tu_BindBufferMemory(VkDevice device,
+                    VkBuffer buffer,
+                    VkDeviceMemory memory,
+                    VkDeviceSize memoryOffset)
+{
+   const VkBindBufferMemoryInfo info = {
+      .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
+      .buffer = buffer,
+      .memory = memory,
+      .memoryOffset = memoryOffset
+   };
+
+   return tu_BindBufferMemory2(device, 1, &info);
+}
+
+VkResult
 tu_BindImageMemory2(VkDevice device,
                     uint32_t bindInfoCount,
                     const VkBindImageMemoryInfo *pBindInfos)
@@ -2645,18 +1646,34 @@ tu_BindImageMemory2(VkDevice device,
       TU_FROM_HANDLE(tu_device_memory, mem, pBindInfos[i].memory);
 
       if (mem) {
-         image->bo = mem->bo;
-         image->iova = mem->bo->iova + pBindInfos[i].memoryOffset;
+         image->bo = &mem->bo;
+         image->bo_offset = pBindInfos[i].memoryOffset;
       } else {
          image->bo = NULL;
-         image->iova = 0;
+         image->bo_offset = 0;
       }
    }
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
+tu_BindImageMemory(VkDevice device,
+                   VkImage image,
+                   VkDeviceMemory memory,
+                   VkDeviceSize memoryOffset)
+{
+   const VkBindImageMemoryInfo info = {
+      .sType = VK_STRUCTURE_TYPE_BIND_BUFFER_MEMORY_INFO,
+      .image = image,
+      .memory = memory,
+      .memoryOffset = memoryOffset
+   };
+
+   return tu_BindImageMemory2(device, 1, &info);
+}
+
+VkResult
 tu_QueueBindSparse(VkQueue _queue,
                    uint32_t bindInfoCount,
                    const VkBindSparseInfo *pBindInfo,
@@ -2665,41 +1682,59 @@ tu_QueueBindSparse(VkQueue _queue,
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+// Queue semaphore functions
+
+VkResult
+tu_CreateSemaphore(VkDevice _device,
+                   const VkSemaphoreCreateInfo *pCreateInfo,
+                   const VkAllocationCallbacks *pAllocator,
+                   VkSemaphore *pSemaphore)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+
+   struct tu_semaphore *sem =
+      vk_alloc2(&device->alloc, pAllocator, sizeof(*sem), 8,
+                VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (!sem)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   *pSemaphore = tu_semaphore_to_handle(sem);
+   return VK_SUCCESS;
+}
+
+void
+tu_DestroySemaphore(VkDevice _device,
+                    VkSemaphore _semaphore,
+                    const VkAllocationCallbacks *pAllocator)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   TU_FROM_HANDLE(tu_semaphore, sem, _semaphore);
+   if (!_semaphore)
+      return;
+
+   vk_free2(&device->alloc, pAllocator, sem);
+}
+
+VkResult
 tu_CreateEvent(VkDevice _device,
                const VkEventCreateInfo *pCreateInfo,
                const VkAllocationCallbacks *pAllocator,
                VkEvent *pEvent)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-
    struct tu_event *event =
-         vk_object_alloc(&device->vk, pAllocator, sizeof(*event),
-                         VK_OBJECT_TYPE_EVENT);
-   if (!event)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
-
-   VkResult result = tu_bo_init_new(device, &event->bo, 0x1000,
-                                    TU_BO_ALLOC_NO_FLAGS, "event");
-   if (result != VK_SUCCESS)
-      goto fail_alloc;
+      vk_alloc2(&device->alloc, pAllocator, sizeof(*event), 8,
+                VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 
-   result = tu_bo_map(device, event->bo);
-   if (result != VK_SUCCESS)
-      goto fail_map;
+   if (!event)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    *pEvent = tu_event_to_handle(event);
 
    return VK_SUCCESS;
-
-fail_map:
-   tu_bo_finish(device, event->bo);
-fail_alloc:
-   vk_object_free(&device->vk, pAllocator, event);
-   return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyEvent(VkDevice _device,
                 VkEvent _event,
                 const VkAllocationCallbacks *pAllocator)
@@ -2709,40 +1744,38 @@ tu_DestroyEvent(VkDevice _device,
 
    if (!event)
       return;
-
-   tu_bo_finish(device, event->bo);
-   vk_object_free(&device->vk, pAllocator, event);
+   vk_free2(&device->alloc, pAllocator, event);
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_GetEventStatus(VkDevice _device, VkEvent _event)
 {
    TU_FROM_HANDLE(tu_event, event, _event);
 
-   if (*(uint64_t*) event->bo->map == 1)
+   if (*event->map == 1)
       return VK_EVENT_SET;
    return VK_EVENT_RESET;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_SetEvent(VkDevice _device, VkEvent _event)
 {
    TU_FROM_HANDLE(tu_event, event, _event);
-   *(uint64_t*) event->bo->map = 1;
+   *event->map = 1;
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_ResetEvent(VkDevice _device, VkEvent _event)
 {
    TU_FROM_HANDLE(tu_event, event, _event);
-   *(uint64_t*) event->bo->map = 0;
+   *event->map = 0;
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateBuffer(VkDevice _device,
                 const VkBufferCreateInfo *pCreateInfo,
                 const VkAllocationCallbacks *pAllocator,
@@ -2751,17 +1784,23 @@ tu_CreateBuffer(VkDevice _device,
    TU_FROM_HANDLE(tu_device, device, _device);
    struct tu_buffer *buffer;
 
-   buffer = vk_buffer_create(&device->vk, pCreateInfo, pAllocator,
-         sizeof(*buffer));
+   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO);
+
+   buffer = vk_alloc2(&device->alloc, pAllocator, sizeof(*buffer), 8,
+                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (buffer == NULL)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   buffer->size = pCreateInfo->size;
+   buffer->usage = pCreateInfo->usage;
+   buffer->flags = pCreateInfo->flags;
 
    *pBuffer = tu_buffer_to_handle(buffer);
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyBuffer(VkDevice _device,
                  VkBuffer _buffer,
                  const VkAllocationCallbacks *pAllocator)
@@ -2772,90 +1811,65 @@ tu_DestroyBuffer(VkDevice _device,
    if (!buffer)
       return;
 
-   vk_buffer_destroy(&device->vk, pAllocator, &buffer->vk);
+   vk_free2(&device->alloc, pAllocator, buffer);
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+static uint32_t
+tu_surface_max_layer_count(struct tu_image_view *iview)
+{
+   return iview->type == VK_IMAGE_VIEW_TYPE_3D
+             ? iview->extent.depth
+             : (iview->base_layer + iview->layer_count);
+}
+
+VkResult
 tu_CreateFramebuffer(VkDevice _device,
                      const VkFramebufferCreateInfo *pCreateInfo,
                      const VkAllocationCallbacks *pAllocator,
                      VkFramebuffer *pFramebuffer)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-
-   if (unlikely(device->instance->debug_flags & TU_DEBUG_DYNAMIC))
-      return vk_common_CreateFramebuffer(_device, pCreateInfo, pAllocator,
-                                         pFramebuffer);
-
-   TU_FROM_HANDLE(tu_render_pass, pass, pCreateInfo->renderPass);
    struct tu_framebuffer *framebuffer;
 
    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO);
 
-   bool imageless = pCreateInfo->flags & VK_FRAMEBUFFER_CREATE_IMAGELESS_BIT;
-
-   size_t size = sizeof(*framebuffer);
-   if (!imageless)
-      size += sizeof(struct tu_attachment_info) * pCreateInfo->attachmentCount;
-   framebuffer = vk_object_alloc(&device->vk, pAllocator, size,
-                                 VK_OBJECT_TYPE_FRAMEBUFFER);
+   size_t size = sizeof(*framebuffer) + sizeof(struct tu_attachment_info) *
+                                           pCreateInfo->attachmentCount;
+   framebuffer = vk_alloc2(&device->alloc, pAllocator, size, 8,
+                           VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (framebuffer == NULL)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    framebuffer->attachment_count = pCreateInfo->attachmentCount;
    framebuffer->width = pCreateInfo->width;
    framebuffer->height = pCreateInfo->height;
    framebuffer->layers = pCreateInfo->layers;
+   for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
+      VkImageView _iview = pCreateInfo->pAttachments[i];
+      struct tu_image_view *iview = tu_image_view_from_handle(_iview);
+      framebuffer->attachments[i].attachment = iview;
 
-   if (!imageless) {
-      for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
-         VkImageView _iview = pCreateInfo->pAttachments[i];
-         struct tu_image_view *iview = tu_image_view_from_handle(_iview);
-         framebuffer->attachments[i].attachment = iview;
-      }
+      framebuffer->width = MIN2(framebuffer->width, iview->extent.width);
+      framebuffer->height = MIN2(framebuffer->height, iview->extent.height);
+      framebuffer->layers =
+         MIN2(framebuffer->layers, tu_surface_max_layer_count(iview));
    }
 
-   tu_framebuffer_tiling_config(framebuffer, device, pass);
-
    *pFramebuffer = tu_framebuffer_to_handle(framebuffer);
    return VK_SUCCESS;
 }
 
 void
-tu_setup_dynamic_framebuffer(struct tu_cmd_buffer *cmd_buffer,
-                             const VkRenderingInfo *pRenderingInfo)
-{
-   struct tu_render_pass *pass = &cmd_buffer->dynamic_pass;
-   struct tu_framebuffer *framebuffer = &cmd_buffer->dynamic_framebuffer;
-
-   framebuffer->attachment_count = pass->attachment_count;
-   framebuffer->width = pRenderingInfo->renderArea.offset.x +
-      pRenderingInfo->renderArea.extent.width;
-   framebuffer->height = pRenderingInfo->renderArea.offset.y +
-      pRenderingInfo->renderArea.extent.height;
-   framebuffer->layers = pRenderingInfo->layerCount;
-
-   tu_framebuffer_tiling_config(framebuffer, cmd_buffer->device, pass);
-}
-
-VKAPI_ATTR void VKAPI_CALL
 tu_DestroyFramebuffer(VkDevice _device,
                       VkFramebuffer _fb,
                       const VkAllocationCallbacks *pAllocator)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-
-   if (unlikely(device->instance->debug_flags & TU_DEBUG_DYNAMIC)) {
-      vk_common_DestroyFramebuffer(_device, _fb, pAllocator);
-      return;
-   }
-
    TU_FROM_HANDLE(tu_framebuffer, fb, _fb);
 
    if (!fb)
       return;
-
-   vk_object_free(&device->vk, pAllocator, fb);
+   vk_free2(&device->alloc, pAllocator, fb);
 }
 
 static void
@@ -2863,89 +1877,9 @@ tu_init_sampler(struct tu_device *device,
                 struct tu_sampler *sampler,
                 const VkSamplerCreateInfo *pCreateInfo)
 {
-   const struct VkSamplerReductionModeCreateInfo *reduction =
-      vk_find_struct_const(pCreateInfo->pNext, SAMPLER_REDUCTION_MODE_CREATE_INFO);
-   const struct VkSamplerYcbcrConversionInfo *ycbcr_conversion =
-      vk_find_struct_const(pCreateInfo->pNext,  SAMPLER_YCBCR_CONVERSION_INFO);
-   const VkSamplerCustomBorderColorCreateInfoEXT *custom_border_color =
-      vk_find_struct_const(pCreateInfo->pNext, SAMPLER_CUSTOM_BORDER_COLOR_CREATE_INFO_EXT);
-   /* for non-custom border colors, the VK enum is translated directly to an offset in
-    * the border color buffer. custom border colors are located immediately after the
-    * builtin colors, and thus an offset of TU_BORDER_COLOR_BUILTIN is added.
-    */
-   uint32_t border_color = (unsigned) pCreateInfo->borderColor;
-   if (pCreateInfo->borderColor == VK_BORDER_COLOR_FLOAT_CUSTOM_EXT ||
-       pCreateInfo->borderColor == VK_BORDER_COLOR_INT_CUSTOM_EXT) {
-      mtx_lock(&device->mutex);
-      border_color = BITSET_FFS(device->custom_border_color) - 1;
-      assert(border_color < TU_BORDER_COLOR_COUNT);
-      BITSET_CLEAR(device->custom_border_color, border_color);
-      mtx_unlock(&device->mutex);
-
-      VkClearColorValue color = custom_border_color->customBorderColor;
-      if (custom_border_color->format == VK_FORMAT_D24_UNORM_S8_UINT &&
-          pCreateInfo->borderColor == VK_BORDER_COLOR_INT_CUSTOM_EXT &&
-          device->use_z24uint_s8uint) {
-         /* When sampling stencil using the special Z24UINT_S8UINT format, the
-          * border color is in the second component. Note: if
-          * customBorderColorWithoutFormat is enabled, we may miss doing this
-          * here if the format isn't specified, which is why we don't use that
-          * format.
-          */
-         color.uint32[1] = color.uint32[0];
-      }
-
-      tu6_pack_border_color(device->global_bo->map + gb_offset(bcolor[border_color]),
-                            &color,
-                            pCreateInfo->borderColor == VK_BORDER_COLOR_INT_CUSTOM_EXT);
-      border_color += TU_BORDER_COLOR_BUILTIN;
-   }
-
-   unsigned aniso = pCreateInfo->anisotropyEnable ?
-      util_last_bit(MIN2((uint32_t)pCreateInfo->maxAnisotropy >> 1, 8)) : 0;
-   bool miplinear = (pCreateInfo->mipmapMode == VK_SAMPLER_MIPMAP_MODE_LINEAR);
-   float min_lod = CLAMP(pCreateInfo->minLod, 0.0f, 4095.0f / 256.0f);
-   float max_lod = CLAMP(pCreateInfo->maxLod, 0.0f, 4095.0f / 256.0f);
-
-   sampler->descriptor[0] =
-      COND(miplinear, A6XX_TEX_SAMP_0_MIPFILTER_LINEAR_NEAR) |
-      A6XX_TEX_SAMP_0_XY_MAG(tu6_tex_filter(pCreateInfo->magFilter, aniso)) |
-      A6XX_TEX_SAMP_0_XY_MIN(tu6_tex_filter(pCreateInfo->minFilter, aniso)) |
-      A6XX_TEX_SAMP_0_ANISO(aniso) |
-      A6XX_TEX_SAMP_0_WRAP_S(tu6_tex_wrap(pCreateInfo->addressModeU)) |
-      A6XX_TEX_SAMP_0_WRAP_T(tu6_tex_wrap(pCreateInfo->addressModeV)) |
-      A6XX_TEX_SAMP_0_WRAP_R(tu6_tex_wrap(pCreateInfo->addressModeW)) |
-      A6XX_TEX_SAMP_0_LOD_BIAS(pCreateInfo->mipLodBias);
-   sampler->descriptor[1] =
-      COND(pCreateInfo->flags & VK_SAMPLER_CREATE_NON_SEAMLESS_CUBE_MAP_BIT_EXT,
-           A6XX_TEX_SAMP_1_CUBEMAPSEAMLESSFILTOFF) |
-      COND(pCreateInfo->unnormalizedCoordinates, A6XX_TEX_SAMP_1_UNNORM_COORDS) |
-      A6XX_TEX_SAMP_1_MIN_LOD(min_lod) |
-      A6XX_TEX_SAMP_1_MAX_LOD(max_lod) |
-      COND(pCreateInfo->compareEnable,
-           A6XX_TEX_SAMP_1_COMPARE_FUNC(tu6_compare_func(pCreateInfo->compareOp)));
-   sampler->descriptor[2] = A6XX_TEX_SAMP_2_BCOLOR(border_color);
-   sampler->descriptor[3] = 0;
-
-   if (reduction) {
-      sampler->descriptor[2] |= A6XX_TEX_SAMP_2_REDUCTION_MODE(
-         tu6_reduction_mode(reduction->reductionMode));
-   }
-
-   sampler->ycbcr_sampler = ycbcr_conversion ?
-      tu_sampler_ycbcr_conversion_from_handle(ycbcr_conversion->conversion) : NULL;
-
-   if (sampler->ycbcr_sampler &&
-       sampler->ycbcr_sampler->chroma_filter == VK_FILTER_LINEAR) {
-      sampler->descriptor[2] |= A6XX_TEX_SAMP_2_CHROMA_LINEAR;
-   }
-
-   /* TODO:
-    * A6XX_TEX_SAMP_1_MIPFILTER_LINEAR_FAR disables mipmapping, but vk has no NONE mipfilter?
-    */
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateSampler(VkDevice _device,
                  const VkSamplerCreateInfo *pCreateInfo,
                  const VkAllocationCallbacks *pAllocator,
@@ -2956,10 +1890,10 @@ tu_CreateSampler(VkDevice _device,
 
    assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO);
 
-   sampler = vk_object_alloc(&device->vk, pAllocator, sizeof(*sampler),
-                             VK_OBJECT_TYPE_SAMPLER);
+   sampler = vk_alloc2(&device->alloc, pAllocator, sizeof(*sampler), 8,
+                       VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (!sampler)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    tu_init_sampler(device, sampler, pCreateInfo);
    *pSampler = tu_sampler_to_handle(sampler);
@@ -2967,29 +1901,17 @@ tu_CreateSampler(VkDevice _device,
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroySampler(VkDevice _device,
                   VkSampler _sampler,
                   const VkAllocationCallbacks *pAllocator)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
    TU_FROM_HANDLE(tu_sampler, sampler, _sampler);
-   uint32_t border_color;
 
    if (!sampler)
       return;
-
-   border_color = (sampler->descriptor[2] & A6XX_TEX_SAMP_2_BCOLOR__MASK) >> A6XX_TEX_SAMP_2_BCOLOR__SHIFT;
-   if (border_color >= TU_BORDER_COLOR_BUILTIN) {
-      border_color -= TU_BORDER_COLOR_BUILTIN;
-      /* if the sampler had a custom border color, free it. TODO: no lock */
-      mtx_lock(&device->mutex);
-      assert(!BITSET_TEST(device->custom_border_color, border_color));
-      BITSET_SET(device->custom_border_color, border_color);
-      mtx_unlock(&device->mutex);
-   }
-
-   vk_object_free(&device->vk, pAllocator, sampler);
+   vk_free2(&device->alloc, pAllocator, sampler);
 }
 
 /* vk_icd.h does not declare this function, so we declare it here to
@@ -3032,21 +1954,12 @@ vk_icdNegotiateLoaderICDInterfaceVersion(uint32_t *pSupportedVersion)
     *        - The ICD must implement vkCreate{PLATFORM}SurfaceKHR(),
     *          vkDestroySurfaceKHR(), and other API which uses VKSurfaceKHR,
     *          because the loader no longer does so.
-    *
-    *    - Loader interface v4 differs from v3 in:
-    *        - The ICD must implement vk_icdGetPhysicalDeviceProcAddr().
-    *
-    *    - Loader interface v5 differs from v4 in:
-    *        - The ICD must support Vulkan API version 1.1 and must not return
-    *          VK_ERROR_INCOMPATIBLE_DRIVER from vkCreateInstance() unless a
-    *          Vulkan Loader with interface v4 or smaller is being used and the
-    *          application provides an API version that is greater than 1.0.
     */
-   *pSupportedVersion = MIN2(*pSupportedVersion, 5u);
+   *pSupportedVersion = MIN2(*pSupportedVersion, 3u);
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_GetMemoryFdKHR(VkDevice _device,
                   const VkMemoryGetFdInfoKHR *pGetFdInfo,
                   int *pFd)
@@ -3062,15 +1975,15 @@ tu_GetMemoryFdKHR(VkDevice _device,
           pGetFdInfo->handleType ==
              VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT);
 
-   int prime_fd = tu_bo_export_dmabuf(device, memory->bo);
+   int prime_fd = tu_bo_export_dmabuf(device, &memory->bo);
    if (prime_fd < 0)
-      return vk_error(device, VK_ERROR_OUT_OF_DEVICE_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_DEVICE_MEMORY);
 
    *pFd = prime_fd;
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_GetMemoryFdPropertiesKHR(VkDevice _device,
                             VkExternalMemoryHandleTypeFlagBits handleType,
                             int fd,
@@ -3081,160 +1994,78 @@ tu_GetMemoryFdPropertiesKHR(VkDevice _device,
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_GetDeviceGroupPeerMemoryFeatures(
-   VkDevice device,
-   uint32_t heapIndex,
-   uint32_t localDeviceIndex,
-   uint32_t remoteDeviceIndex,
-   VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
-{
-   assert(localDeviceIndex == remoteDeviceIndex);
-
-   *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
-                          VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
-                          VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
-                          VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_GetPhysicalDeviceMultisamplePropertiesEXT(
-   VkPhysicalDevice                            physicalDevice,
-   VkSampleCountFlagBits                       samples,
-   VkMultisamplePropertiesEXT*                 pMultisampleProperties)
-{
-   TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
-
-   if (samples <= VK_SAMPLE_COUNT_4_BIT && pdevice->vk.supported_extensions.EXT_sample_locations)
-      pMultisampleProperties->maxSampleLocationGridSize = (VkExtent2D){ 1, 1 };
-   else
-      pMultisampleProperties->maxSampleLocationGridSize = (VkExtent2D){ 0, 0 };
-}
-
-VkDeviceAddress
-tu_GetBufferDeviceAddress(VkDevice _device,
-                          const VkBufferDeviceAddressInfo* pInfo)
-{
-   TU_FROM_HANDLE(tu_buffer, buffer, pInfo->buffer);
-
-   return buffer->iova;
-}
-
-uint64_t tu_GetBufferOpaqueCaptureAddress(
-    VkDevice                                    device,
-    const VkBufferDeviceAddressInfo*            pInfo)
+void
+tu_GetPhysicalDeviceExternalSemaphoreProperties(
+   VkPhysicalDevice physicalDevice,
+   const VkPhysicalDeviceExternalSemaphoreInfo *pExternalSemaphoreInfo,
+   VkExternalSemaphoreProperties *pExternalSemaphoreProperties)
 {
-   /* We care only about memory allocation opaque addresses */
-   return 0;
+   pExternalSemaphoreProperties->exportFromImportedHandleTypes = 0;
+   pExternalSemaphoreProperties->compatibleHandleTypes = 0;
+   pExternalSemaphoreProperties->externalSemaphoreFeatures = 0;
 }
 
-uint64_t tu_GetDeviceMemoryOpaqueCaptureAddress(
-    VkDevice                                    device,
-    const VkDeviceMemoryOpaqueCaptureAddressInfo* pInfo)
+void
+tu_GetPhysicalDeviceExternalFenceProperties(
+   VkPhysicalDevice physicalDevice,
+   const VkPhysicalDeviceExternalFenceInfo *pExternalFenceInfo,
+   VkExternalFenceProperties *pExternalFenceProperties)
 {
-   TU_FROM_HANDLE(tu_device_memory, mem, pInfo->memory);
-   return mem->bo->iova;
+   pExternalFenceProperties->exportFromImportedHandleTypes = 0;
+   pExternalFenceProperties->compatibleHandleTypes = 0;
+   pExternalFenceProperties->externalFenceFeatures = 0;
 }
 
-struct tu_debug_bos_entry {
-   uint32_t count;
-   uint64_t size;
-   const char *name;
-};
-
-const char *
-tu_debug_bos_add(struct tu_device *dev, uint64_t size, const char *name)
+VkResult
+tu_CreateDebugReportCallbackEXT(
+   VkInstance _instance,
+   const VkDebugReportCallbackCreateInfoEXT *pCreateInfo,
+   const VkAllocationCallbacks *pAllocator,
+   VkDebugReportCallbackEXT *pCallback)
 {
-   assert(name);
-
-   if (likely(!dev->bo_sizes))
-      return NULL;
-
-   mtx_lock(&dev->bo_mutex);
-   struct hash_entry *entry = _mesa_hash_table_search(dev->bo_sizes, name);
-   struct tu_debug_bos_entry *debug_bos;
-
-   if (!entry) {
-      debug_bos = calloc(1, sizeof(struct tu_debug_bos_entry));
-      debug_bos->name = strdup(name);
-      _mesa_hash_table_insert(dev->bo_sizes, debug_bos->name, debug_bos);
-   } else {
-      debug_bos = entry->data;
-   }
-
-   debug_bos->count++;
-   debug_bos->size += align(size, 4096);
-   mtx_unlock(&dev->bo_mutex);
-
-   return debug_bos->name;
+   TU_FROM_HANDLE(tu_instance, instance, _instance);
+   return vk_create_debug_report_callback(&instance->debug_report_callbacks,
+                                          pCreateInfo, pAllocator,
+                                          &instance->alloc, pCallback);
 }
 
 void
-tu_debug_bos_del(struct tu_device *dev, struct tu_bo *bo)
+tu_DestroyDebugReportCallbackEXT(VkInstance _instance,
+                                 VkDebugReportCallbackEXT _callback,
+                                 const VkAllocationCallbacks *pAllocator)
 {
-   if (likely(!dev->bo_sizes) || !bo->name)
-      return;
-
-   mtx_lock(&dev->bo_mutex);
-   struct hash_entry *entry =
-      _mesa_hash_table_search(dev->bo_sizes, bo->name);
-   /* If we're finishing the BO, it should have been added already */
-   assert(entry);
-
-   struct tu_debug_bos_entry *debug_bos = entry->data;
-   debug_bos->count--;
-   debug_bos->size -= align(bo->size, 4096);
-   if (!debug_bos->count) {
-      _mesa_hash_table_remove(dev->bo_sizes, entry);
-      free((void *) debug_bos->name);
-      free(debug_bos);
-   }
-   mtx_unlock(&dev->bo_mutex);
+   TU_FROM_HANDLE(tu_instance, instance, _instance);
+   vk_destroy_debug_report_callback(&instance->debug_report_callbacks,
+                                    _callback, pAllocator, &instance->alloc);
 }
 
-static int debug_bos_count_compare(const void *in_a, const void *in_b)
+void
+tu_DebugReportMessageEXT(VkInstance _instance,
+                         VkDebugReportFlagsEXT flags,
+                         VkDebugReportObjectTypeEXT objectType,
+                         uint64_t object,
+                         size_t location,
+                         int32_t messageCode,
+                         const char *pLayerPrefix,
+                         const char *pMessage)
 {
-   struct tu_debug_bos_entry *a = *(struct tu_debug_bos_entry **)in_a;
-   struct tu_debug_bos_entry *b = *(struct tu_debug_bos_entry **)in_b;
-   return a->count - b->count;
+   TU_FROM_HANDLE(tu_instance, instance, _instance);
+   vk_debug_report(&instance->debug_report_callbacks, flags, objectType,
+                   object, location, messageCode, pLayerPrefix, pMessage);
 }
 
 void
-tu_debug_bos_print_stats(struct tu_device *dev)
+tu_GetDeviceGroupPeerMemoryFeatures(
+   VkDevice device,
+   uint32_t heapIndex,
+   uint32_t localDeviceIndex,
+   uint32_t remoteDeviceIndex,
+   VkPeerMemoryFeatureFlags *pPeerMemoryFeatures)
 {
-   if (likely(!dev->bo_sizes))
-      return;
-
-   mtx_lock(&dev->bo_mutex);
-
-   /* Put the HT's sizes data in an array so we can sort by number of allocations. */
-   struct util_dynarray dyn;
-   util_dynarray_init(&dyn, NULL);
-
-   uint32_t size = 0;
-   uint32_t count = 0;
-   hash_table_foreach(dev->bo_sizes, entry)
-   {
-      struct tu_debug_bos_entry *debug_bos = (void *) entry->data;
-      util_dynarray_append(&dyn, struct tu_debug_bos_entry *, debug_bos);
-      size += debug_bos->size / 1024;
-      count += debug_bos->count;
-   }
-
-   qsort(dyn.data,
-         util_dynarray_num_elements(&dyn, struct tu_debug_bos_entry *),
-         sizeof(struct tu_debug_bos_entryos_entry *), debug_bos_count_compare);
-
-   util_dynarray_foreach(&dyn, struct tu_debug_bos_entry *, entryp)
-   {
-      struct tu_debug_bos_entry *debug_bos = *entryp;
-      mesa_logi("%30s: %4d bos, %lld kb\n", debug_bos->name, debug_bos->count,
-                (long long) (debug_bos->size / 1024));
-   }
-
-   mesa_logi("submitted %d bos (%d MB)\n", count, DIV_ROUND_UP(size, 1024));
-
-   util_dynarray_fini(&dyn);
+   assert(localDeviceIndex == remoteDeviceIndex);
 
-   mtx_unlock(&dev->bo_mutex);
+   *pPeerMemoryFeatures = VK_PEER_MEMORY_FEATURE_COPY_SRC_BIT |
+                          VK_PEER_MEMORY_FEATURE_COPY_DST_BIT |
+                          VK_PEER_MEMORY_FEATURE_GENERIC_SRC_BIT |
+                          VK_PEER_MEMORY_FEATURE_GENERIC_DST_BIT;
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_drm.c b/lib/mesa/src/freedreno/vulkan/tu_drm.c
index 9a57c6644..9b2e6f788 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_drm.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_drm.c
@@ -1,61 +1,36 @@
 /*
  * Copyright © 2018 Google, Inc.
  * Copyright © 2015 Intel Corporation
- * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_drm.h"
+#include "tu_private.h"
 
 #include <errno.h>
 #include <fcntl.h>
+#include <stdint.h>
 #include <sys/ioctl.h>
-#include <sys/mman.h>
 #include <xf86drm.h>
 
-#ifdef MAJOR_IN_MKDEV
-#include <sys/mkdev.h>
-#endif
-#ifdef MAJOR_IN_SYSMACROS
-#include <sys/sysmacros.h>
-#endif
-
-#include "vk_util.h"
-
 #include "drm-uapi/msm_drm.h"
-#include "util/u_debug.h"
-#include "util/hash_table.h"
-#include "util/timespec.h"
-#include "util/os_time.h"
-
-#include "tu_cmd_buffer.h"
-#include "tu_cs.h"
-#include "tu_device.h"
-#include "tu_dynamic_rendering.h"
-
-struct tu_queue_submit
-{
-   struct vk_queue_submit *vk_submit;
-   struct tu_u_trace_submission_data *u_trace_submission_data;
-
-   struct tu_cmd_buffer **cmd_buffers;
-   struct drm_msm_gem_submit_cmd *cmds;
-   struct drm_msm_gem_submit_syncobj *in_syncobjs;
-   struct drm_msm_gem_submit_syncobj *out_syncobjs;
-
-   uint32_t nr_cmd_buffers;
-   uint32_t nr_in_syncobjs;
-   uint32_t nr_out_syncobjs;
-   uint32_t entry_count;
-   uint32_t perf_pass_index;
-
-   bool     autotune_fence;
-};
-
-struct tu_u_trace_syncobj
-{
-   uint32_t msm_queue_id;
-   uint32_t fence;
-};
 
 static int
 tu_drm_get_param(const struct tu_physical_device *dev,
@@ -80,7 +55,7 @@ tu_drm_get_param(const struct tu_physical_device *dev,
    return 0;
 }
 
-static int
+int
 tu_drm_get_gpu_id(const struct tu_physical_device *dev, uint32_t *id)
 {
    uint64_t value;
@@ -92,7 +67,7 @@ tu_drm_get_gpu_id(const struct tu_physical_device *dev, uint32_t *id)
    return 0;
 }
 
-static int
+int
 tu_drm_get_gmem_size(const struct tu_physical_device *dev, uint32_t *size)
 {
    uint64_t value;
@@ -104,85 +79,17 @@ tu_drm_get_gmem_size(const struct tu_physical_device *dev, uint32_t *size)
    return 0;
 }
 
-static int
-tu_drm_get_gmem_base(const struct tu_physical_device *dev, uint64_t *base)
-{
-   return tu_drm_get_param(dev, MSM_PARAM_GMEM_BASE, base);
-}
-
-static int
-tu_drm_get_va_prop(const struct tu_physical_device *dev,
-                   uint64_t *va_start, uint64_t *va_size)
-{
-   uint64_t value;
-   int ret = tu_drm_get_param(dev, MSM_PARAM_VA_START, &value);
-   if (ret)
-      return ret;
-
-   *va_start = value;
-
-   ret = tu_drm_get_param(dev, MSM_PARAM_VA_SIZE, &value);
-   if (ret)
-      return ret;
-
-   *va_size = value;
-
-   return 0;
-}
-
-static uint32_t
-tu_drm_get_priorities(const struct tu_physical_device *dev)
-{
-   uint64_t val = 1;
-   tu_drm_get_param(dev, MSM_PARAM_PRIORITIES, &val);
-   assert(val >= 1);
-
-   return val;
-}
-
-int
-tu_device_get_gpu_timestamp(struct tu_device *dev, uint64_t *ts)
-{
-   return tu_drm_get_param(dev->physical_device, MSM_PARAM_TIMESTAMP, ts);
-}
-
-int
-tu_device_get_suspend_count(struct tu_device *dev, uint64_t *suspend_count)
-{
-   int ret = tu_drm_get_param(dev->physical_device, MSM_PARAM_SUSPENDS, suspend_count);
-   return ret;
-}
-
-VkResult
-tu_device_check_status(struct vk_device *vk_device)
-{
-   struct tu_device *device = container_of(vk_device, struct tu_device, vk);
-   struct tu_physical_device *physical_device = device->physical_device;
-
-   uint64_t last_fault_count = physical_device->fault_count;
-   int ret = tu_drm_get_param(physical_device, MSM_PARAM_FAULTS, &physical_device->fault_count);
-   if (ret != 0)
-      return vk_device_set_lost(&device->vk, "error getting GPU fault count: %d", ret);
-
-   if (last_fault_count != physical_device->fault_count)
-      return vk_device_set_lost(&device->vk, "GPU faulted or hung");
-
-   return VK_SUCCESS;
-}
-
 int
 tu_drm_submitqueue_new(const struct tu_device *dev,
                        int priority,
                        uint32_t *queue_id)
 {
-   assert(priority >= 0 &&
-          priority < dev->physical_device->submitqueue_priority_count);
    struct drm_msm_submitqueue req = {
       .flags = 0,
       .prio = priority,
    };
 
-   int ret = drmCommandWriteRead(dev->fd,
+   int ret = drmCommandWriteRead(dev->physical_device->local_fd,
                                  DRM_MSM_SUBMITQUEUE_NEW, &req, sizeof(req));
    if (ret)
       return ret;
@@ -194,1156 +101,94 @@ tu_drm_submitqueue_new(const struct tu_device *dev,
 void
 tu_drm_submitqueue_close(const struct tu_device *dev, uint32_t queue_id)
 {
-   drmCommandWrite(dev->fd, DRM_MSM_SUBMITQUEUE_CLOSE,
+   drmCommandWrite(dev->physical_device->local_fd, DRM_MSM_SUBMITQUEUE_CLOSE,
                    &queue_id, sizeof(uint32_t));
 }
 
-static void
-tu_gem_close(const struct tu_device *dev, uint32_t gem_handle)
-{
-   struct drm_gem_close req = {
-      .handle = gem_handle,
-   };
-
-   drmIoctl(dev->fd, DRM_IOCTL_GEM_CLOSE, &req);
-}
-
-/** Helper for DRM_MSM_GEM_INFO, returns 0 on error. */
-static uint64_t
-tu_gem_info(const struct tu_device *dev, uint32_t gem_handle, uint32_t info)
-{
-   struct drm_msm_gem_info req = {
-      .handle = gem_handle,
-      .info = info,
-   };
-
-   int ret = drmCommandWriteRead(dev->fd,
-                                 DRM_MSM_GEM_INFO, &req, sizeof(req));
-   if (ret < 0)
-      return 0;
-
-   return req.value;
-}
-
-
-static VkResult
-tu_allocate_userspace_iova(struct tu_device *dev,
-                           uint32_t gem_handle,
-                           uint64_t size,
-                           uint64_t client_iova,
-                           enum tu_bo_alloc_flags flags,
-                           uint64_t *iova)
-{
-   mtx_lock(&dev->physical_device->vma_mutex);
-
-   *iova = 0;
-
-   if (flags & TU_BO_ALLOC_REPLAYABLE) {
-      if (client_iova) {
-         if (util_vma_heap_alloc_addr(&dev->physical_device->vma, client_iova,
-                                      size)) {
-            *iova = client_iova;
-         } else {
-            return VK_ERROR_INVALID_OPAQUE_CAPTURE_ADDRESS;
-         }
-      } else {
-         /* We have to separate replayable IOVAs from ordinary one in order to
-          * for them not to clash. The easiest way to do this is to allocate
-          * them from the other end of the address space.
-          */
-         dev->physical_device->vma.alloc_high = true;
-         *iova =
-            util_vma_heap_alloc(&dev->physical_device->vma, size, 0x1000);
-      }
-   } else {
-      dev->physical_device->vma.alloc_high = false;
-      *iova = util_vma_heap_alloc(&dev->physical_device->vma, size, 0x1000);
-   }
-
-   mtx_unlock(&dev->physical_device->vma_mutex);
-
-   if (!*iova)
-      return VK_ERROR_OUT_OF_DEVICE_MEMORY;
-
-   struct drm_msm_gem_info req = {
-      .handle = gem_handle,
-      .info = MSM_INFO_SET_IOVA,
-      .value = *iova,
-   };
-
-   int ret =
-      drmCommandWriteRead(dev->fd, DRM_MSM_GEM_INFO, &req, sizeof(req));
-   if (ret < 0)
-      return VK_ERROR_OUT_OF_HOST_MEMORY;
-
-   return VK_SUCCESS;
-}
-
-static VkResult
-tu_allocate_kernel_iova(struct tu_device *dev,
-                        uint32_t gem_handle,
-                        uint64_t *iova)
-{
-   *iova = tu_gem_info(dev, gem_handle, MSM_INFO_GET_IOVA);
-   if (!*iova)
-      return VK_ERROR_OUT_OF_DEVICE_MEMORY;
-
-   return VK_SUCCESS;
-}
-
-static VkResult
-tu_bo_init(struct tu_device *dev,
-           struct tu_bo *bo,
-           uint32_t gem_handle,
-           uint64_t size,
-           uint64_t client_iova,
-           enum tu_bo_alloc_flags flags,
-           const char *name)
-{
-   VkResult result = VK_SUCCESS;
-   uint64_t iova = 0;
-
-   assert(!client_iova || dev->physical_device->has_set_iova);
-
-   if (dev->physical_device->has_set_iova) {
-      result = tu_allocate_userspace_iova(dev, gem_handle, size, client_iova,
-                                          flags, &iova);
-   } else {
-      result = tu_allocate_kernel_iova(dev, gem_handle, &iova);
-   }
-
-   if (result != VK_SUCCESS)
-      goto fail_bo_list;
-
-   name = tu_debug_bos_add(dev, size, name);
-
-   mtx_lock(&dev->bo_mutex);
-   uint32_t idx = dev->bo_count++;
-
-   /* grow the bo list if needed */
-   if (idx >= dev->bo_list_size) {
-      uint32_t new_len = idx + 64;
-      struct drm_msm_gem_submit_bo *new_ptr =
-         vk_realloc(&dev->vk.alloc, dev->bo_list, new_len * sizeof(*dev->bo_list),
-                    8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-      if (!new_ptr) {
-         result = VK_ERROR_OUT_OF_HOST_MEMORY;
-         goto fail_bo_list;
-      }
-
-      dev->bo_list = new_ptr;
-      dev->bo_list_size = new_len;
-   }
-
-   bool dump = flags & TU_BO_ALLOC_ALLOW_DUMP;
-   dev->bo_list[idx] = (struct drm_msm_gem_submit_bo) {
-      .flags = MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE |
-               COND(dump, MSM_SUBMIT_BO_DUMP),
-      .handle = gem_handle,
-      .presumed = iova,
-   };
-
-   *bo = (struct tu_bo) {
-      .gem_handle = gem_handle,
-      .size = size,
-      .iova = iova,
-      .refcnt = 1,
-      .bo_list_idx = idx,
-      .name = name,
-   };
-
-   mtx_unlock(&dev->bo_mutex);
-
-   return VK_SUCCESS;
-
-fail_bo_list:
-   tu_gem_close(dev, gem_handle);
-   return result;
-}
-
 /**
- * Sets the name in the kernel so that the contents of /debug/dri/0/gem are more
- * useful.
- *
- * We skip this on release builds (when we're also not doing BO debugging) to
- * reduce overhead.
+ * Return gem handle on success. Return 0 on failure.
  */
-static void
-tu_bo_set_kernel_name(struct tu_device *dev, struct tu_bo *bo, const char *name)
+uint32_t
+tu_gem_new(const struct tu_device *dev, uint64_t size, uint32_t flags)
 {
-   bool kernel_bo_names = dev->bo_sizes != NULL;
-#ifdef DEBUG
-   kernel_bo_names = true;
-#endif
-   if (!kernel_bo_names)
-      return;
-
-   struct drm_msm_gem_info req = {
-      .handle = bo->gem_handle,
-      .info = MSM_INFO_SET_NAME,
-      .value = (uintptr_t)(void *)name,
-      .len = strlen(name),
-   };
-
-   int ret = drmCommandWrite(dev->fd, DRM_MSM_GEM_INFO, &req, sizeof(req));
-   if (ret) {
-      mesa_logw_once("Failed to set BO name with DRM_MSM_GEM_INFO: %d",
-                     ret);
-   }
-}
-
-VkResult
-tu_bo_init_new_explicit_iova(struct tu_device *dev,
-                             struct tu_bo **out_bo,
-                             uint64_t size,
-                             uint64_t client_iova,
-                             enum tu_bo_alloc_flags flags,
-                             const char *name)
-{
-   /* TODO: Choose better flags. As of 2018-11-12, freedreno/drm/msm_bo.c
-    * always sets `flags = MSM_BO_WC`, and we copy that behavior here.
-    */
    struct drm_msm_gem_new req = {
       .size = size,
-      .flags = MSM_BO_WC
+      .flags = flags,
    };
 
-   if (flags & TU_BO_ALLOC_GPU_READ_ONLY)
-      req.flags |= MSM_BO_GPU_READONLY;
-
-   int ret = drmCommandWriteRead(dev->fd,
+   int ret = drmCommandWriteRead(dev->physical_device->local_fd,
                                  DRM_MSM_GEM_NEW, &req, sizeof(req));
    if (ret)
-      return vk_error(dev, VK_ERROR_OUT_OF_DEVICE_MEMORY);
-
-   struct tu_bo* bo = tu_device_lookup_bo(dev, req.handle);
-   assert(bo && bo->gem_handle == 0);
-
-   VkResult result =
-      tu_bo_init(dev, bo, req.handle, size, client_iova, flags, name);
-
-   if (result != VK_SUCCESS)
-      memset(bo, 0, sizeof(*bo));
-   else
-      *out_bo = bo;
-
-   /* We don't use bo->name here because for the !TU_DEBUG=bo case bo->name is NULL. */
-   tu_bo_set_kernel_name(dev, bo, name);
+      return 0;
 
-   return result;
+   return req.handle;
 }
 
-VkResult
-tu_bo_init_dmabuf(struct tu_device *dev,
-                  struct tu_bo **out_bo,
-                  uint64_t size,
-                  int prime_fd)
+uint32_t
+tu_gem_import_dmabuf(const struct tu_device *dev, int prime_fd, uint64_t size)
 {
    /* lseek() to get the real size */
    off_t real_size = lseek(prime_fd, 0, SEEK_END);
    lseek(prime_fd, 0, SEEK_SET);
    if (real_size < 0 || (uint64_t) real_size < size)
-      return vk_error(dev, VK_ERROR_INVALID_EXTERNAL_HANDLE);
-
-   /* Importing the same dmabuf several times would yield the same
-    * gem_handle. Thus there could be a race when destroying
-    * BO and importing the same dmabuf from different threads.
-    * We must not permit the creation of dmabuf BO and its release
-    * to happen in parallel.
-    */
-   u_rwlock_wrlock(&dev->dma_bo_lock);
+      return 0;
 
    uint32_t gem_handle;
-   int ret = drmPrimeFDToHandle(dev->fd, prime_fd,
+   int ret = drmPrimeFDToHandle(dev->physical_device->local_fd, prime_fd,
                                 &gem_handle);
-   if (ret) {
-      u_rwlock_wrunlock(&dev->dma_bo_lock);
-      return vk_error(dev, VK_ERROR_INVALID_EXTERNAL_HANDLE);
-   }
-
-   struct tu_bo* bo = tu_device_lookup_bo(dev, gem_handle);
-
-   if (bo->refcnt != 0) {
-      p_atomic_inc(&bo->refcnt);
-      u_rwlock_wrunlock(&dev->dma_bo_lock);
-
-      *out_bo = bo;
-      return VK_SUCCESS;
-   }
-
-   VkResult result =
-      tu_bo_init(dev, bo, gem_handle, size, 0, TU_BO_ALLOC_NO_FLAGS, "dmabuf");
-
-   if (result != VK_SUCCESS)
-      memset(bo, 0, sizeof(*bo));
-   else
-      *out_bo = bo;
-
-   u_rwlock_wrunlock(&dev->dma_bo_lock);
+   if (ret)
+      return 0;
 
-   return result;
+   return gem_handle;
 }
 
 int
-tu_bo_export_dmabuf(struct tu_device *dev, struct tu_bo *bo)
+tu_gem_export_dmabuf(const struct tu_device *dev, uint32_t gem_handle)
 {
    int prime_fd;
-   int ret = drmPrimeHandleToFD(dev->fd, bo->gem_handle,
-                                DRM_CLOEXEC | DRM_RDWR, &prime_fd);
+   int ret = drmPrimeHandleToFD(dev->physical_device->local_fd, gem_handle,
+                                DRM_CLOEXEC, &prime_fd);
 
    return ret == 0 ? prime_fd : -1;
 }
 
-VkResult
-tu_bo_map(struct tu_device *dev, struct tu_bo *bo)
-{
-   if (bo->map)
-      return VK_SUCCESS;
-
-   uint64_t offset = tu_gem_info(dev, bo->gem_handle, MSM_INFO_GET_OFFSET);
-   if (!offset)
-      return vk_error(dev, VK_ERROR_OUT_OF_DEVICE_MEMORY);
-
-   /* TODO: Should we use the wrapper os_mmap() like Freedreno does? */
-   void *map = mmap(0, bo->size, PROT_READ | PROT_WRITE, MAP_SHARED,
-                    dev->fd, offset);
-   if (map == MAP_FAILED)
-      return vk_error(dev, VK_ERROR_MEMORY_MAP_FAILED);
-
-   bo->map = map;
-   return VK_SUCCESS;
-}
-
 void
-tu_bo_finish(struct tu_device *dev, struct tu_bo *bo)
-{
-   assert(bo->gem_handle);
-
-   u_rwlock_rdlock(&dev->dma_bo_lock);
-
-   if (!p_atomic_dec_zero(&bo->refcnt)) {
-      u_rwlock_rdunlock(&dev->dma_bo_lock);
-      return;
-   }
-
-   if (bo->map)
-      munmap(bo->map, bo->size);
-
-   tu_debug_bos_del(dev, bo);
-
-   mtx_lock(&dev->bo_mutex);
-   dev->bo_count--;
-   dev->bo_list[bo->bo_list_idx] = dev->bo_list[dev->bo_count];
-
-   struct tu_bo* exchanging_bo = tu_device_lookup_bo(dev, dev->bo_list[bo->bo_list_idx].handle);
-   exchanging_bo->bo_list_idx = bo->bo_list_idx;
-
-   if (bo->implicit_sync)
-      dev->implicit_sync_bo_count--;
-
-   mtx_unlock(&dev->bo_mutex);
-
-   if (dev->physical_device->has_set_iova) {
-      mtx_lock(&dev->physical_device->vma_mutex);
-      util_vma_heap_free(&dev->physical_device->vma, bo->iova, bo->size);
-      mtx_unlock(&dev->physical_device->vma_mutex);
-   }
-
-   /* Our BO structs are stored in a sparse array in the physical device,
-    * so we don't want to free the BO pointer, instead we want to reset it
-    * to 0, to signal that array entry as being free.
-    */
-   uint32_t gem_handle = bo->gem_handle;
-   memset(bo, 0, sizeof(*bo));
-
-   tu_gem_close(dev, gem_handle);
-
-   u_rwlock_rdunlock(&dev->dma_bo_lock);
-}
-
-extern const struct vk_sync_type tu_timeline_sync_type;
-
-static inline bool
-vk_sync_is_tu_timeline_sync(const struct vk_sync *sync)
-{
-   return sync->type == &tu_timeline_sync_type;
-}
-
-static struct tu_timeline_sync *
-to_tu_timeline_sync(struct vk_sync *sync)
-{
-   assert(sync->type == &tu_timeline_sync_type);
-   return container_of(sync, struct tu_timeline_sync, base);
-}
-
-static uint32_t
-tu_syncobj_from_vk_sync(struct vk_sync *sync)
-{
-   uint32_t syncobj = -1;
-   if (vk_sync_is_tu_timeline_sync(sync)) {
-      syncobj = to_tu_timeline_sync(sync)->syncobj;
-   } else if (vk_sync_type_is_drm_syncobj(sync->type)) {
-      syncobj = vk_sync_as_drm_syncobj(sync)->syncobj;
-   }
-
-   assert(syncobj != -1);
-
-   return syncobj;
-}
-
-static VkResult
-tu_timeline_sync_init(struct vk_device *vk_device,
-                      struct vk_sync *vk_sync,
-                      uint64_t initial_value)
-{
-   struct tu_device *device = container_of(vk_device, struct tu_device, vk);
-   struct tu_timeline_sync *sync = to_tu_timeline_sync(vk_sync);
-   uint32_t flags = 0;
-
-   assert(device->fd >= 0);
-
-   int err = drmSyncobjCreate(device->fd, flags, &sync->syncobj);
-
-   if (err < 0) {
-        return vk_error(device, VK_ERROR_DEVICE_LOST);
-   }
-
-   sync->state = initial_value ? TU_TIMELINE_SYNC_STATE_SIGNALED :
-                                    TU_TIMELINE_SYNC_STATE_RESET;
-
-   return VK_SUCCESS;
-}
-
-static void
-tu_timeline_sync_finish(struct vk_device *vk_device,
-                   struct vk_sync *vk_sync)
-{
-   struct tu_device *dev = container_of(vk_device, struct tu_device, vk);
-   struct tu_timeline_sync *sync = to_tu_timeline_sync(vk_sync);
-
-   assert(dev->fd >= 0);
-   ASSERTED int err = drmSyncobjDestroy(dev->fd, sync->syncobj);
-   assert(err == 0);
-}
-
-static VkResult
-tu_timeline_sync_reset(struct vk_device *vk_device,
-                  struct vk_sync *vk_sync)
-{
-   struct tu_device *dev = container_of(vk_device, struct tu_device, vk);
-   struct tu_timeline_sync *sync = to_tu_timeline_sync(vk_sync);
-
-   int err = drmSyncobjReset(dev->fd, &sync->syncobj, 1);
-   if (err) {
-      return vk_errorf(dev, VK_ERROR_UNKNOWN,
-                       "DRM_IOCTL_SYNCOBJ_RESET failed: %m");
-   } else {
-       sync->state = TU_TIMELINE_SYNC_STATE_RESET;
-   }
-
-   return VK_SUCCESS;
-}
-
-static VkResult
-drm_syncobj_wait(struct tu_device *device,
-                 uint32_t *handles, uint32_t count_handles,
-                 uint64_t timeout_nsec, bool wait_all)
-{
-   uint32_t syncobj_wait_flags = DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT;
-   if (wait_all) syncobj_wait_flags |= DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL;
-
-   /* syncobj absolute timeouts are signed.  clamp OS_TIMEOUT_INFINITE down. */
-   timeout_nsec = MIN2(timeout_nsec, (uint64_t)INT64_MAX);
-
-   int err = drmSyncobjWait(device->fd, handles,
-                            count_handles, timeout_nsec,
-                            syncobj_wait_flags,
-                            NULL /* first_signaled */);
-   if (err && errno == ETIME) {
-      return VK_TIMEOUT;
-   } else if (err) {
-      return vk_errorf(device, VK_ERROR_UNKNOWN,
-                       "DRM_IOCTL_SYNCOBJ_WAIT failed: %m");
-   }
-
-   return VK_SUCCESS;
-}
-
-/* Based on anv_bo_sync_wait */
-static VkResult
-tu_timeline_sync_wait(struct vk_device *vk_device,
-                 uint32_t wait_count,
-                 const struct vk_sync_wait *waits,
-                 enum vk_sync_wait_flags wait_flags,
-                 uint64_t abs_timeout_ns)
-{
-   struct tu_device *dev = container_of(vk_device, struct tu_device, vk);
-   bool wait_all = !(wait_flags & VK_SYNC_WAIT_ANY);
-
-   uint32_t handles[wait_count];
-   uint32_t submit_count;
-   VkResult ret = VK_SUCCESS;
-   uint32_t pending = wait_count;
-   struct tu_timeline_sync *submitted_syncs[wait_count];
-
-   while (pending) {
-      pending = 0;
-      submit_count = 0;
-
-      for (unsigned i = 0; i < wait_count; ++i) {
-         struct tu_timeline_sync *sync = to_tu_timeline_sync(waits[i].sync);
-
-         if (sync->state == TU_TIMELINE_SYNC_STATE_RESET) {
-            assert(!(wait_flags & VK_SYNC_WAIT_PENDING));
-            pending++;
-         } else if (sync->state == TU_TIMELINE_SYNC_STATE_SIGNALED) {
-            if (wait_flags & VK_SYNC_WAIT_ANY)
-               return VK_SUCCESS;
-         } else if (sync->state == TU_TIMELINE_SYNC_STATE_SUBMITTED) {
-            if (!(wait_flags & VK_SYNC_WAIT_PENDING)) {
-               handles[submit_count] = sync->syncobj;
-               submitted_syncs[submit_count++] = sync;
-            }
-         }
-      }
-
-      if (submit_count > 0) {
-         do {
-            ret = drm_syncobj_wait(dev, handles, submit_count, abs_timeout_ns, wait_all);
-         } while (ret == VK_TIMEOUT && os_time_get_nano() < abs_timeout_ns);
-
-         if (ret == VK_SUCCESS) {
-            for (unsigned i = 0; i < submit_count; ++i) {
-               struct tu_timeline_sync *sync = submitted_syncs[i];
-               sync->state = TU_TIMELINE_SYNC_STATE_SIGNALED;
-            }
-         } else {
-            /* return error covering timeout */
-            return ret;
-         }
-      } else if (pending > 0) {
-         /* If we've hit this then someone decided to vkWaitForFences before
-          * they've actually submitted any of them to a queue.  This is a
-          * fairly pessimal case, so it's ok to lock here and use a standard
-          * pthreads condition variable.
-          */
-         pthread_mutex_lock(&dev->submit_mutex);
-
-         /* It's possible that some of the fences have changed state since the
-          * last time we checked.  Now that we have the lock, check for
-          * pending fences again and don't wait if it's changed.
-          */
-         uint32_t now_pending = 0;
-         for (uint32_t i = 0; i < wait_count; i++) {
-            struct tu_timeline_sync *sync = to_tu_timeline_sync(waits[i].sync);
-            if (sync->state == TU_TIMELINE_SYNC_STATE_RESET)
-               now_pending++;
-         }
-         assert(now_pending <= pending);
-
-         if (now_pending == pending) {
-            struct timespec abstime = {
-               .tv_sec = abs_timeout_ns / NSEC_PER_SEC,
-               .tv_nsec = abs_timeout_ns % NSEC_PER_SEC,
-            };
-
-            ASSERTED int ret;
-            ret = pthread_cond_timedwait(&dev->timeline_cond,
-                                         &dev->submit_mutex, &abstime);
-            assert(ret != EINVAL);
-            if (os_time_get_nano() >= abs_timeout_ns) {
-               pthread_mutex_unlock(&dev->submit_mutex);
-               return VK_TIMEOUT;
-            }
-         }
-
-         pthread_mutex_unlock(&dev->submit_mutex);
-      }
-   }
-
-   return ret;
-}
-
-const struct vk_sync_type tu_timeline_sync_type = {
-   .size = sizeof(struct tu_timeline_sync),
-   .features = VK_SYNC_FEATURE_BINARY |
-               VK_SYNC_FEATURE_GPU_WAIT |
-               VK_SYNC_FEATURE_GPU_MULTI_WAIT |
-               VK_SYNC_FEATURE_CPU_WAIT |
-               VK_SYNC_FEATURE_CPU_RESET |
-               VK_SYNC_FEATURE_WAIT_ANY |
-               VK_SYNC_FEATURE_WAIT_PENDING,
-   .init = tu_timeline_sync_init,
-   .finish = tu_timeline_sync_finish,
-   .reset = tu_timeline_sync_reset,
-   .wait_many = tu_timeline_sync_wait,
-};
-
-VkResult
-tu_physical_device_try_create(struct vk_instance *vk_instance,
-                              struct _drmDevice *drm_device,
-                              struct vk_physical_device **out)
-{
-   struct tu_instance *instance =
-      container_of(vk_instance, struct tu_instance, vk);
-
-   if (!(drm_device->available_nodes & (1 << DRM_NODE_RENDER)) ||
-       drm_device->bustype != DRM_BUS_PLATFORM)
-      return VK_ERROR_INCOMPATIBLE_DRIVER;
-
-   const char *primary_path = drm_device->nodes[DRM_NODE_PRIMARY];
-   const char *path = drm_device->nodes[DRM_NODE_RENDER];
-   VkResult result = VK_SUCCESS;
-   drmVersionPtr version;
-   int fd;
-   int master_fd = -1;
-
-   fd = open(path, O_RDWR | O_CLOEXEC);
-   if (fd < 0) {
-      return vk_startup_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
-                               "failed to open device %s", path);
-   }
-
-   /* Version 1.6 added SYNCOBJ support. */
-   const int min_version_major = 1;
-   const int min_version_minor = 6;
-
-   version = drmGetVersion(fd);
-   if (!version) {
-      close(fd);
-      return vk_startup_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
-                               "failed to query kernel driver version for device %s",
-                               path);
-   }
-
-   if (strcmp(version->name, "msm")) {
-      drmFreeVersion(version);
-      close(fd);
-      return vk_startup_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
-                               "device %s does not use the msm kernel driver",
-                               path);
-   }
-
-   if (version->version_major != min_version_major ||
-       version->version_minor < min_version_minor) {
-      result = vk_startup_errorf(instance, VK_ERROR_INCOMPATIBLE_DRIVER,
-                                 "kernel driver for device %s has version %d.%d, "
-                                 "but Vulkan requires version >= %d.%d",
-                                 path,
-                                 version->version_major, version->version_minor,
-                                 min_version_major, min_version_minor);
-      drmFreeVersion(version);
-      close(fd);
-      return result;
-   }
-
-   struct tu_physical_device *device =
-      vk_zalloc(&instance->vk.alloc, sizeof(*device), 8,
-                VK_SYSTEM_ALLOCATION_SCOPE_INSTANCE);
-   if (!device) {
-      result = vk_error(instance, VK_ERROR_OUT_OF_HOST_MEMORY);
-      drmFreeVersion(version);
-      goto fail;
-   }
-
-   device->msm_major_version = version->version_major;
-   device->msm_minor_version = version->version_minor;
-
-   drmFreeVersion(version);
-
-   if (instance->debug_flags & TU_DEBUG_STARTUP)
-      mesa_logi("Found compatible device '%s'.", path);
-
-   device->instance = instance;
-
-   if (instance->vk.enabled_extensions.KHR_display) {
-      master_fd = open(primary_path, O_RDWR | O_CLOEXEC);
-      if (master_fd >= 0) {
-         /* TODO: free master_fd is accel is not working? */
-      }
-   }
-
-   device->master_fd = master_fd;
-   device->local_fd = fd;
-
-   if (tu_drm_get_gpu_id(device, &device->dev_id.gpu_id)) {
-      result = vk_startup_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                                 "could not get GPU ID");
-      goto fail;
-   }
-
-   if (tu_drm_get_param(device, MSM_PARAM_CHIP_ID, &device->dev_id.chip_id)) {
-      result = vk_startup_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                                 "could not get CHIP ID");
-      goto fail;
-   }
-
-   if (tu_drm_get_gmem_size(device, &device->gmem_size)) {
-      result = vk_startup_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                                "could not get GMEM size");
-      goto fail;
-   }
-   device->gmem_size = debug_get_num_option("TU_GMEM", device->gmem_size);
-
-   if (tu_drm_get_gmem_base(device, &device->gmem_base)) {
-      result = vk_startup_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                                 "could not get GMEM size");
-      goto fail;
-   }
-
-   /*
-    * device->has_set_iova = !tu_drm_get_va_prop(device, &device->va_start,
-    *                                            &device->va_size);
-    *
-    * If BO is freed while kernel considers it busy, our VMA state gets
-    * desynchronized from kernel's VMA state, because kernel waits
-    * until BO stops being busy. And whether BO is busy decided at
-    * submission granularity.
-    *
-    * Disable this capability until solution is found.
-    */
-   device->has_set_iova = false;
-
-   struct stat st;
-
-   if (stat(primary_path, &st) == 0) {
-      device->has_master = true;
-      device->master_major = major(st.st_rdev);
-      device->master_minor = minor(st.st_rdev);
-   } else {
-      device->has_master = false;
-      device->master_major = 0;
-      device->master_minor = 0;
-   }
-
-   if (stat(path, &st) == 0) {
-      device->has_local = true;
-      device->local_major = major(st.st_rdev);
-      device->local_minor = minor(st.st_rdev);
-   } else {
-      result = vk_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                         "failed to stat DRM render node %s", path);
-      goto fail;
-   }
-
-   int ret = tu_drm_get_param(device, MSM_PARAM_FAULTS, &device->fault_count);
-   if (ret != 0) {
-      result = vk_startup_errorf(instance, VK_ERROR_INITIALIZATION_FAILED,
-                                 "Failed to get initial fault count: %d", ret);
-      goto fail;
-   }
-
-   device->submitqueue_priority_count = tu_drm_get_priorities(device);
-
-   device->syncobj_type = vk_drm_syncobj_get_type(fd);
-   /* we don't support DRM_CAP_SYNCOBJ_TIMELINE, but drm-shim does */
-   if (!(device->syncobj_type.features & VK_SYNC_FEATURE_TIMELINE))
-      device->timeline_type = vk_sync_timeline_get_type(&tu_timeline_sync_type);
-
-   device->sync_types[0] = &device->syncobj_type;
-   device->sync_types[1] = &device->timeline_type.sync;
-   device->sync_types[2] = NULL;
-
-   device->heap.size = tu_get_system_heap_size();
-   device->heap.used = 0u;
-   device->heap.flags = VK_MEMORY_HEAP_DEVICE_LOCAL_BIT;
-
-   result = tu_physical_device_init(device, instance);
-
-   if (result == VK_SUCCESS) {
-      *out = &device->vk;
-      return result;
-   }
-
-fail:
-   if (device)
-      vk_free(&instance->vk.alloc, device);
-   close(fd);
-   if (master_fd != -1)
-      close(master_fd);
-   return result;
-}
-
-static VkResult
-tu_queue_submit_create_locked(struct tu_queue *queue,
-                              struct vk_queue_submit *vk_submit,
-                              const uint32_t nr_in_syncobjs,
-                              const uint32_t nr_out_syncobjs,
-                              uint32_t perf_pass_index,
-                              struct tu_queue_submit *new_submit)
-{
-   VkResult result;
-
-   bool u_trace_enabled = u_trace_context_actively_tracing(&queue->device->trace_context);
-   bool has_trace_points = false;
-
-   struct vk_command_buffer **vk_cmd_buffers = vk_submit->command_buffers;
-
-   memset(new_submit, 0, sizeof(struct tu_queue_submit));
-
-   new_submit->cmd_buffers = (void *)vk_cmd_buffers;
-   new_submit->nr_cmd_buffers = vk_submit->command_buffer_count;
-   tu_insert_dynamic_cmdbufs(queue->device, &new_submit->cmd_buffers,
-                             &new_submit->nr_cmd_buffers);
-
-   uint32_t entry_count = 0;
-   for (uint32_t j = 0; j < new_submit->nr_cmd_buffers; ++j) {
-      struct tu_cmd_buffer *cmdbuf = new_submit->cmd_buffers[j];
-
-      if (perf_pass_index != ~0)
-         entry_count++;
-
-      entry_count += cmdbuf->cs.entry_count;
-
-      if (u_trace_enabled && u_trace_has_points(&cmdbuf->trace)) {
-         if (!(cmdbuf->usage_flags & VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT))
-            entry_count++;
-
-         has_trace_points = true;
-      }
-   }
-
-   new_submit->autotune_fence =
-      tu_autotune_submit_requires_fence(new_submit->cmd_buffers, new_submit->nr_cmd_buffers);
-   if (new_submit->autotune_fence)
-      entry_count++;
-
-   new_submit->cmds = vk_zalloc(&queue->device->vk.alloc,
-         entry_count * sizeof(*new_submit->cmds), 8,
-         VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
-   if (new_submit->cmds == NULL) {
-      result = vk_error(queue, VK_ERROR_OUT_OF_HOST_MEMORY);
-      goto fail_cmds;
-   }
-
-   if (has_trace_points) {
-      result =
-         tu_u_trace_submission_data_create(
-            queue->device, new_submit->cmd_buffers,
-            new_submit->nr_cmd_buffers,
-            &new_submit->u_trace_submission_data);
-
-      if (result != VK_SUCCESS) {
-         goto fail_u_trace_submission_data;
-      }
-   }
-
-   /* Allocate without wait timeline semaphores */
-   new_submit->in_syncobjs = vk_zalloc(&queue->device->vk.alloc,
-         nr_in_syncobjs * sizeof(*new_submit->in_syncobjs), 8,
-         VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
-   if (new_submit->in_syncobjs == NULL) {
-      result = vk_error(queue, VK_ERROR_OUT_OF_HOST_MEMORY);
-      goto fail_in_syncobjs;
-   }
-
-   /* Allocate with signal timeline semaphores considered */
-   new_submit->out_syncobjs = vk_zalloc(&queue->device->vk.alloc,
-         nr_out_syncobjs * sizeof(*new_submit->out_syncobjs), 8,
-         VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-
-   if (new_submit->out_syncobjs == NULL) {
-      result = vk_error(queue, VK_ERROR_OUT_OF_HOST_MEMORY);
-      goto fail_out_syncobjs;
-   }
-
-   new_submit->entry_count = entry_count;
-   new_submit->nr_in_syncobjs = nr_in_syncobjs;
-   new_submit->nr_out_syncobjs = nr_out_syncobjs;
-   new_submit->perf_pass_index = perf_pass_index;
-   new_submit->vk_submit = vk_submit;
-
-   return VK_SUCCESS;
-
-fail_out_syncobjs:
-   vk_free(&queue->device->vk.alloc, new_submit->in_syncobjs);
-fail_in_syncobjs:
-   if (new_submit->u_trace_submission_data)
-      tu_u_trace_submission_data_finish(queue->device,
-                                        new_submit->u_trace_submission_data);
-fail_u_trace_submission_data:
-   vk_free(&queue->device->vk.alloc, new_submit->cmds);
-fail_cmds:
-   return result;
-}
-
-static void
-tu_queue_submit_finish(struct tu_queue *queue, struct tu_queue_submit *submit)
-{
-   vk_free(&queue->device->vk.alloc, submit->cmds);
-   vk_free(&queue->device->vk.alloc, submit->in_syncobjs);
-   vk_free(&queue->device->vk.alloc, submit->out_syncobjs);
-   if (submit->cmd_buffers != (void *) submit->vk_submit->command_buffers)
-      vk_free(&queue->device->vk.alloc, submit->cmd_buffers);
-}
-
-static void
-tu_fill_msm_gem_submit(struct tu_device *dev,
-                       struct drm_msm_gem_submit_cmd *cmd,
-                       struct tu_cs_entry *cs_entry)
-{
-   cmd->type = MSM_SUBMIT_CMD_BUF;
-   cmd->submit_idx = cs_entry->bo->bo_list_idx;
-   cmd->submit_offset = cs_entry->offset;
-   cmd->size = cs_entry->size;
-   cmd->pad = 0;
-   cmd->nr_relocs = 0;
-   cmd->relocs = 0;
-}
-
-static void
-tu_queue_build_msm_gem_submit_cmds(struct tu_queue *queue,
-                                   struct tu_queue_submit *submit,
-                                   struct tu_cs *autotune_cs)
+tu_gem_close(const struct tu_device *dev, uint32_t gem_handle)
 {
-   struct tu_device *dev = queue->device;
-   struct drm_msm_gem_submit_cmd *cmds = submit->cmds;
-
-   uint32_t entry_idx = 0;
-   for (uint32_t j = 0; j < submit->nr_cmd_buffers; ++j) {
-      struct tu_device *dev = queue->device;
-      struct tu_cmd_buffer *cmdbuf = submit->cmd_buffers[j];
-      struct tu_cs *cs = &cmdbuf->cs;
-
-      if (submit->perf_pass_index != ~0) {
-         struct tu_cs_entry *perf_cs_entry =
-            &dev->perfcntrs_pass_cs_entries[submit->perf_pass_index];
-
-         tu_fill_msm_gem_submit(dev, &cmds[entry_idx], perf_cs_entry);
-         entry_idx++;
-      }
-
-      for (unsigned i = 0; i < cs->entry_count; ++i, ++entry_idx) {
-         tu_fill_msm_gem_submit(dev, &cmds[entry_idx], &cs->entries[i]);
-      }
-
-      if (submit->u_trace_submission_data) {
-         struct tu_cs *ts_cs =
-            submit->u_trace_submission_data->cmd_trace_data[j].timestamp_copy_cs;
-         if (ts_cs) {
-            tu_fill_msm_gem_submit(dev, &cmds[entry_idx], &ts_cs->entries[0]);
-            entry_idx++;
-         }
-      }
-   }
+   struct drm_gem_close req = {
+      .handle = gem_handle,
+   };
 
-   if (autotune_cs) {
-      assert(autotune_cs->entry_count == 1);
-      tu_fill_msm_gem_submit(dev, &cmds[entry_idx], &autotune_cs->entries[0]);
-      entry_idx++;
-   }
+   drmIoctl(dev->physical_device->local_fd, DRM_IOCTL_GEM_CLOSE, &req);
 }
 
-static VkResult
-tu_queue_submit_locked(struct tu_queue *queue, struct tu_queue_submit *submit)
+/** Return UINT64_MAX on error. */
+static uint64_t
+tu_gem_info(const struct tu_device *dev, uint32_t gem_handle, uint32_t info)
 {
-   queue->device->submit_count++;
-
-   struct tu_cs *autotune_cs = NULL;
-   if (submit->autotune_fence) {
-      autotune_cs = tu_autotune_on_submit(queue->device,
-                                          &queue->device->autotune,
-                                          submit->cmd_buffers,
-                                          submit->nr_cmd_buffers);
-   }
-
-   uint32_t flags = MSM_PIPE_3D0;
-
-   if (submit->vk_submit->wait_count)
-      flags |= MSM_SUBMIT_SYNCOBJ_IN;
-
-   if (submit->vk_submit->signal_count)
-      flags |= MSM_SUBMIT_SYNCOBJ_OUT;
-
-   mtx_lock(&queue->device->bo_mutex);
-
-   if (queue->device->implicit_sync_bo_count == 0)
-      flags |= MSM_SUBMIT_NO_IMPLICIT;
-
-   /* drm_msm_gem_submit_cmd requires index of bo which could change at any
-    * time when bo_mutex is not locked. So we build submit cmds here the real
-    * place to submit.
-    */
-   tu_queue_build_msm_gem_submit_cmds(queue, submit, autotune_cs);
-
-   struct drm_msm_gem_submit req = {
-      .flags = flags,
-      .queueid = queue->msm_queue_id,
-      .bos = (uint64_t)(uintptr_t) queue->device->bo_list,
-      .nr_bos = submit->entry_count ? queue->device->bo_count : 0,
-      .cmds = (uint64_t)(uintptr_t)submit->cmds,
-      .nr_cmds = submit->entry_count,
-      .in_syncobjs = (uint64_t)(uintptr_t)submit->in_syncobjs,
-      .out_syncobjs = (uint64_t)(uintptr_t)submit->out_syncobjs,
-      .nr_in_syncobjs = submit->nr_in_syncobjs,
-      .nr_out_syncobjs = submit->nr_out_syncobjs,
-      .syncobj_stride = sizeof(struct drm_msm_gem_submit_syncobj),
+   struct drm_msm_gem_info req = {
+      .handle = gem_handle,
+      .info = info,
    };
 
-   int ret = drmCommandWriteRead(queue->device->fd,
-                                 DRM_MSM_GEM_SUBMIT,
-                                 &req, sizeof(req));
-
-   mtx_unlock(&queue->device->bo_mutex);
-
-   tu_debug_bos_print_stats(queue->device);
-
-   if (ret)
-      return vk_device_set_lost(&queue->device->vk, "submit failed: %m");
-
-#if HAVE_PERFETTO
-   tu_perfetto_submit(queue->device, queue->device->submit_count);
-#endif
-
-   if (submit->u_trace_submission_data) {
-      struct tu_u_trace_submission_data *submission_data =
-         submit->u_trace_submission_data;
-      submission_data->submission_id = queue->device->submit_count;
-      /* We have to allocate it here since it is different between drm/kgsl */
-      submission_data->syncobj =
-         vk_alloc(&queue->device->vk.alloc, sizeof(struct tu_u_trace_syncobj),
-               8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
-      submission_data->syncobj->fence = req.fence;
-      submission_data->syncobj->msm_queue_id = queue->msm_queue_id;
-
-      submit->u_trace_submission_data = NULL;
-
-      for (uint32_t i = 0; i < submission_data->cmd_buffer_count; i++) {
-         bool free_data = i == submission_data->last_buffer_with_tracepoints;
-         if (submission_data->cmd_trace_data[i].trace)
-            u_trace_flush(submission_data->cmd_trace_data[i].trace,
-                          submission_data, free_data);
-
-         if (!submission_data->cmd_trace_data[i].timestamp_copy_cs) {
-            /* u_trace is owned by cmd_buffer */
-            submission_data->cmd_trace_data[i].trace = NULL;
-         }
-      }
-   }
-
-   for (uint32_t i = 0; i < submit->vk_submit->wait_count; i++) {
-      if (!vk_sync_is_tu_timeline_sync(submit->vk_submit->waits[i].sync))
-         continue;
-
-      struct tu_timeline_sync *sync =
-         container_of(submit->vk_submit->waits[i].sync, struct tu_timeline_sync, base);
-
-      assert(sync->state != TU_TIMELINE_SYNC_STATE_RESET);
-
-      /* Set SIGNALED to the state of the wait timeline sync since this means the syncobj
-       * is done and ready again so this can be garbage-collectioned later.
-       */
-      sync->state = TU_TIMELINE_SYNC_STATE_SIGNALED;
-   }
-
-   for (uint32_t i = 0; i < submit->vk_submit->signal_count; i++) {
-      if (!vk_sync_is_tu_timeline_sync(submit->vk_submit->signals[i].sync))
-         continue;
-
-      struct tu_timeline_sync *sync =
-         container_of(submit->vk_submit->signals[i].sync, struct tu_timeline_sync, base);
-
-      assert(sync->state == TU_TIMELINE_SYNC_STATE_RESET);
-      /* Set SUBMITTED to the state of the signal timeline sync so we could wait for
-       * this timeline sync until completed if necessary.
-       */
-      sync->state = TU_TIMELINE_SYNC_STATE_SUBMITTED;
-   }
-
-   pthread_cond_broadcast(&queue->device->timeline_cond);
-
-   return VK_SUCCESS;
-}
+   int ret = drmCommandWriteRead(dev->physical_device->local_fd,
+                                 DRM_MSM_GEM_INFO, &req, sizeof(req));
+   if (ret == -1)
+      return UINT64_MAX;
 
-static inline void
-get_abs_timeout(struct drm_msm_timespec *tv, uint64_t ns)
-{
-   struct timespec t;
-   clock_gettime(CLOCK_MONOTONIC, &t);
-   tv->tv_sec = t.tv_sec + ns / 1000000000;
-   tv->tv_nsec = t.tv_nsec + ns % 1000000000;
+   return req.value;
 }
 
-VkResult
-tu_device_wait_u_trace(struct tu_device *dev, struct tu_u_trace_syncobj *syncobj)
+/** Return UINT64_MAX on error. */
+uint64_t
+tu_gem_info_offset(const struct tu_device *dev, uint32_t gem_handle)
 {
-   struct drm_msm_wait_fence req = {
-      .fence = syncobj->fence,
-      .queueid = syncobj->msm_queue_id,
-   };
-   int ret;
-
-   get_abs_timeout(&req.timeout, 1000000000);
-
-   ret = drmCommandWrite(dev->fd, DRM_MSM_WAIT_FENCE, &req, sizeof(req));
-   if (ret && (ret != -ETIMEDOUT)) {
-      fprintf(stderr, "wait-fence failed! %d (%s)", ret, strerror(errno));
-      return VK_TIMEOUT;
-   }
-
-   return VK_SUCCESS;
+   return tu_gem_info(dev, gem_handle, MSM_INFO_GET_OFFSET);
 }
 
-VkResult
-tu_queue_submit(struct vk_queue *vk_queue, struct vk_queue_submit *submit)
+/** Return UINT64_MAX on error. */
+uint64_t
+tu_gem_info_iova(const struct tu_device *dev, uint32_t gem_handle)
 {
-   MESA_TRACE_FUNC();
-   struct tu_queue *queue = container_of(vk_queue, struct tu_queue, vk);
-   uint32_t perf_pass_index = queue->device->perfcntrs_pass_cs ?
-                              submit->perf_pass_index : ~0;
-   struct tu_queue_submit submit_req;
-
-   if (unlikely(queue->device->physical_device->instance->debug_flags &
-                 TU_DEBUG_LOG_SKIP_GMEM_OPS)) {
-      tu_dbg_log_gmem_load_store_skips(queue->device);
-   }
-
-   pthread_mutex_lock(&queue->device->submit_mutex);
-
-   VkResult ret = tu_queue_submit_create_locked(queue, submit,
-         submit->wait_count, submit->signal_count,
-         perf_pass_index, &submit_req);
-
-   if (ret != VK_SUCCESS) {
-      pthread_mutex_unlock(&queue->device->submit_mutex);
-      return ret;
-   }
-
-   /* note: assuming there won't be any very large semaphore counts */
-   struct drm_msm_gem_submit_syncobj *in_syncobjs = submit_req.in_syncobjs;
-   struct drm_msm_gem_submit_syncobj *out_syncobjs = submit_req.out_syncobjs;
-
-   uint32_t nr_in_syncobjs = 0, nr_out_syncobjs = 0;
-
-   for (uint32_t i = 0; i < submit->wait_count; i++) {
-      struct vk_sync *sync = submit->waits[i].sync;
-
-      in_syncobjs[nr_in_syncobjs++] = (struct drm_msm_gem_submit_syncobj) {
-         .handle = tu_syncobj_from_vk_sync(sync),
-         .flags = 0,
-      };
-   }
-
-   for (uint32_t i = 0; i < submit->signal_count; i++) {
-      struct vk_sync *sync = submit->signals[i].sync;
-
-      out_syncobjs[nr_out_syncobjs++] = (struct drm_msm_gem_submit_syncobj) {
-         .handle = tu_syncobj_from_vk_sync(sync),
-         .flags = 0,
-      };
-   }
-
-   ret = tu_queue_submit_locked(queue, &submit_req);
-
-   pthread_mutex_unlock(&queue->device->submit_mutex);
-   tu_queue_submit_finish(queue, &submit_req);
-
-   if (ret != VK_SUCCESS)
-       return ret;
-
-   u_trace_context_process(&queue->device->trace_context, true);
-
-   return VK_SUCCESS;
+   return tu_gem_info(dev, gem_handle, MSM_INFO_GET_IOVA);
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_formats.c b/lib/mesa/src/freedreno/vulkan/tu_formats.c
index 75a3ce74e..537b59d25 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_formats.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_formats.c
@@ -1,395 +1,661 @@
+
 /*
  * Copyright © 2016 Red Hat.
  * Copyright © 2016 Bas Nieuwenhuizen
- * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_formats.h"
+#include "tu_private.h"
 
-#include "fdl/fd6_format_table.h"
+#include "registers/adreno_common.xml.h"
+#include "registers/a6xx.xml.h"
 
+#include "util/format_r11g11b10f.h"
+#include "util/format_srgb.h"
+#include "util/u_half.h"
+#include "vk_format.h"
 #include "vk_util.h"
-#include "drm-uapi/drm_fourcc.h"
 
-#include "tu_device.h"
-#include "tu_image.h"
+/**
+ * Declare a format table.  A format table is an array of tu_native_format.
+ * It can map a consecutive range of VkFormat to the corresponding
+ * tu_native_format.
+ *
+ * TU_FORMAT_TABLE_FIRST and TU_FORMAT_TABLE_LAST must already be defined and
+ * have the values of the first and last VkFormat of the array respectively.
+ */
+#define TU_FORMAT_TABLE(var)                                                 \
+   static const VkFormat var##_first = TU_FORMAT_TABLE_FIRST;                \
+   static const VkFormat var##_last = TU_FORMAT_TABLE_LAST;                  \
+   static const struct tu_native_format var[TU_FORMAT_TABLE_LAST - TU_FORMAT_TABLE_FIRST + 1]
+#undef TU_FORMAT_TABLE_FIRST
+#undef TU_FORMAT_TABLE_LAST
+
+#define VFMT6_x -1
+#define TFMT6_x -1
+#define RB6_x -1
+
+#define TU6_FMT(vkfmt, vtxfmt, texfmt, rbfmt, swapfmt, valid)                \
+   [VK_FORMAT_##vkfmt - TU_FORMAT_TABLE_FIRST] = {                           \
+      .vtx = VFMT6_##vtxfmt,                                                 \
+      .tex = TFMT6_##texfmt,                                                 \
+      .rb = RB6_##rbfmt,                                                     \
+      .swap = swapfmt,                                                       \
+      .present = valid,                                                      \
+   }
 
-struct tu_native_format
-tu6_format_vtx(VkFormat vk_format)
-{
-   enum pipe_format format = vk_format_to_pipe_format(vk_format);
-   struct tu_native_format fmt = {
-      .fmt = fd6_vertex_format(format),
-      .swap = fd6_vertex_swap(format),
-   };
-   assert(fmt.fmt != FMT6_NONE);
-   return fmt;
-}
+/**
+ * fmt/alias/swap are derived from VkFormat mechanically (and might not even
+ * exist).  It is the macro of choice that decides whether a VkFormat is
+ * supported and how.
+ */
+#define TU6_VTC(vk, fmt, alias, swap) TU6_FMT(vk, fmt, fmt, alias, swap, true)
+#define TU6_xTC(vk, fmt, alias, swap) TU6_FMT(vk, x, fmt, alias, swap, true)
+#define TU6_VTx(vk, fmt, alias, swap) TU6_FMT(vk, fmt, fmt, x, swap, true)
+#define TU6_Vxx(vk, fmt, alias, swap) TU6_FMT(vk, fmt, x, x, swap, true)
+#define TU6_xTx(vk, fmt, alias, swap) TU6_FMT(vk, x, fmt, x, swap, true)
+#define TU6_xxx(vk, fmt, alias, swap) TU6_FMT(vk, x, x, x, WZYX, false)
+
+#define TU_FORMAT_TABLE_FIRST VK_FORMAT_UNDEFINED
+#define TU_FORMAT_TABLE_LAST VK_FORMAT_ASTC_12x12_SRGB_BLOCK
+TU_FORMAT_TABLE(tu6_format_table0) = {
+   TU6_xxx(UNDEFINED,                  x,                 x,                  x),    /* 0 */
+
+   /* 8-bit packed */
+   TU6_xxx(R4G4_UNORM_PACK8,           4_4_UNORM,         R4G4_UNORM,         WZXY), /* 1 */
+
+   /* 16-bit packed */
+   TU6_xTC(R4G4B4A4_UNORM_PACK16,      4_4_4_4_UNORM,     R4G4B4A4_UNORM,     XYZW), /* 2 */
+   TU6_xTC(B4G4R4A4_UNORM_PACK16,      4_4_4_4_UNORM,     R4G4B4A4_UNORM,     ZYXW), /* 3 */
+   TU6_xTC(R5G6B5_UNORM_PACK16,        5_6_5_UNORM,       R5G6B5_UNORM,       WXYZ), /* 4 */
+   TU6_xTC(B5G6R5_UNORM_PACK16,        5_6_5_UNORM,       R5G6B5_UNORM,       WXYZ), /* 5 */
+   TU6_xxx(R5G5B5A1_UNORM_PACK16,      1_5_5_5_UNORM,     A1R5G5B5_UNORM,     XYZW), /* 6 */
+   TU6_xxx(B5G5R5A1_UNORM_PACK16,      1_5_5_5_UNORM,     A1R5G5B5_UNORM,     XYZW), /* 7 */
+   TU6_xTC(A1R5G5B5_UNORM_PACK16,      5_5_5_1_UNORM,     R5G5B5A1_UNORM,     WXYZ), /* 8 */
+
+   /* 8-bit R */
+   TU6_VTC(R8_UNORM,                   8_UNORM,           R8_UNORM,           WZYX), /* 9 */
+   TU6_VTC(R8_SNORM,                   8_SNORM,           R8_SNORM,           WZYX), /* 10 */
+   TU6_Vxx(R8_USCALED,                 8_UINT,            R8_UINT,            WZYX), /* 11 */
+   TU6_Vxx(R8_SSCALED,                 8_SINT,            R8_SINT,            WZYX), /* 12 */
+   TU6_VTC(R8_UINT,                    8_UINT,            R8_UINT,            WZYX), /* 13 */
+   TU6_VTC(R8_SINT,                    8_SINT,            R8_SINT,            WZYX), /* 14 */
+   TU6_xTC(R8_SRGB,                    8_UNORM,           R8_UNORM,           WZYX), /* 15 */
+
+   /* 16-bit RG */
+   TU6_VTC(R8G8_UNORM,                 8_8_UNORM,         R8G8_UNORM,         WZYX), /* 16 */
+   TU6_VTC(R8G8_SNORM,                 8_8_SNORM,         R8G8_SNORM,         WZYX), /* 17 */
+   TU6_Vxx(R8G8_USCALED,               8_8_UINT,          R8G8_UINT,          WZYX), /* 18 */
+   TU6_Vxx(R8G8_SSCALED,               8_8_SINT,          R8G8_SINT,          WZYX), /* 19 */
+   TU6_VTC(R8G8_UINT,                  8_8_UINT,          R8G8_UINT,          WZYX), /* 20 */
+   TU6_VTC(R8G8_SINT,                  8_8_SINT,          R8G8_SINT,          WZYX), /* 21 */
+   TU6_xTC(R8G8_SRGB,                  8_8_UNORM,         R8G8_UNORM,         WZYX), /* 22 */
+
+   /* 24-bit RGB */
+   TU6_Vxx(R8G8B8_UNORM,               8_8_8_UNORM,       R8G8B8_UNORM,       WZYX), /* 23 */
+   TU6_Vxx(R8G8B8_SNORM,               8_8_8_SNORM,       R8G8B8_SNORM,       WZYX), /* 24 */
+   TU6_Vxx(R8G8B8_USCALED,             8_8_8_UINT,        R8G8B8_UINT,        WZYX), /* 25 */
+   TU6_Vxx(R8G8B8_SSCALED,             8_8_8_SINT,        R8G8B8_SINT,        WZYX), /* 26 */
+   TU6_Vxx(R8G8B8_UINT,                8_8_8_UINT,        R8G8B8_UINT,        WZYX), /* 27 */
+   TU6_Vxx(R8G8B8_SINT,                8_8_8_SINT,        R8G8B8_SINT,        WZYX), /* 28 */
+   TU6_xxx(R8G8B8_SRGB,                8_8_8_UNORM,       R8G8B8_UNORM,       WZYX), /* 29 */
+
+   /* 24-bit BGR */
+   TU6_Vxx(B8G8R8_UNORM,               8_8_8_UNORM,       R8G8B8_UNORM,       WXYZ), /* 30 */
+   TU6_Vxx(B8G8R8_SNORM,               8_8_8_SNORM,       R8G8B8_SNORM,       WXYZ), /* 31 */
+   TU6_Vxx(B8G8R8_USCALED,             8_8_8_UINT,        R8G8B8_UINT,        WXYZ), /* 32 */
+   TU6_Vxx(B8G8R8_SSCALED,             8_8_8_SINT,        R8G8B8_SINT,        WXYZ), /* 33 */
+   TU6_Vxx(B8G8R8_UINT,                8_8_8_UINT,        R8G8B8_UINT,        WXYZ), /* 34 */
+   TU6_Vxx(B8G8R8_SINT,                8_8_8_SINT,        R8G8B8_SINT,        WXYZ), /* 35 */
+   TU6_xxx(B8G8R8_SRGB,                8_8_8_UNORM,       R8G8B8_UNORM,       WXYZ), /* 36 */
+
+   /* 32-bit RGBA */
+   TU6_VTC(R8G8B8A8_UNORM,             8_8_8_8_UNORM,     R8G8B8A8_UNORM,     WZYX), /* 37 */
+   TU6_VTC(R8G8B8A8_SNORM,             8_8_8_8_SNORM,     R8G8B8A8_SNORM,     WZYX), /* 38 */
+   TU6_Vxx(R8G8B8A8_USCALED,           8_8_8_8_UINT,      R8G8B8A8_UINT,      WZYX), /* 39 */
+   TU6_Vxx(R8G8B8A8_SSCALED,           8_8_8_8_SINT,      R8G8B8A8_SINT,      WZYX), /* 40 */
+   TU6_VTC(R8G8B8A8_UINT,              8_8_8_8_UINT,      R8G8B8A8_UINT,      WZYX), /* 41 */
+   TU6_VTC(R8G8B8A8_SINT,              8_8_8_8_SINT,      R8G8B8A8_SINT,      WZYX), /* 42 */
+   TU6_xTC(R8G8B8A8_SRGB,              8_8_8_8_UNORM,     R8G8B8A8_UNORM,     WZYX), /* 43 */
+
+   /* 32-bit BGRA */
+   TU6_VTC(B8G8R8A8_UNORM,             8_8_8_8_UNORM,     R8G8B8A8_UNORM,     WXYZ), /* 44 */
+   TU6_VTC(B8G8R8A8_SNORM,             8_8_8_8_SNORM,     R8G8B8A8_SNORM,     WXYZ), /* 45 */
+   TU6_Vxx(B8G8R8A8_USCALED,           8_8_8_8_UINT,      R8G8B8A8_UINT,      WXYZ), /* 46 */
+   TU6_Vxx(B8G8R8A8_SSCALED,           8_8_8_8_SINT,      R8G8B8A8_SINT,      WXYZ), /* 47 */
+   TU6_VTC(B8G8R8A8_UINT,              8_8_8_8_UINT,      R8G8B8A8_UINT,      WXYZ), /* 48 */
+   TU6_VTC(B8G8R8A8_SINT,              8_8_8_8_SINT,      R8G8B8A8_SINT,      WXYZ), /* 49 */
+   TU6_xTC(B8G8R8A8_SRGB,              8_8_8_8_UNORM,     R8G8B8A8_UNORM,     WXYZ), /* 50 */
+
+   /* 32-bit packed */
+   TU6_VTC(A8B8G8R8_UNORM_PACK32,      8_8_8_8_UNORM,     R8G8B8A8_UNORM,     WZYX), /* 51 */
+   TU6_VTC(A8B8G8R8_SNORM_PACK32,      8_8_8_8_SNORM,     R8G8B8A8_SNORM,     WZYX), /* 52 */
+   TU6_Vxx(A8B8G8R8_USCALED_PACK32,    8_8_8_8_UINT,      R8G8B8A8_UINT,      WZYX), /* 53 */
+   TU6_Vxx(A8B8G8R8_SSCALED_PACK32,    8_8_8_8_SINT,      R8G8B8A8_SINT,      WZYX), /* 54 */
+   TU6_VTC(A8B8G8R8_UINT_PACK32,       8_8_8_8_UINT,      R8G8B8A8_UINT,      WZYX), /* 55 */
+   TU6_VTC(A8B8G8R8_SINT_PACK32,       8_8_8_8_SINT,      R8G8B8A8_SINT,      WZYX), /* 56 */
+   TU6_xTC(A8B8G8R8_SRGB_PACK32,       8_8_8_8_UNORM,     R8G8B8A8_UNORM,     WZYX), /* 57 */
+   TU6_VTC(A2R10G10B10_UNORM_PACK32,   10_10_10_2_UNORM,  R10G10B10A2_UNORM,  WXYZ), /* 58 */
+   TU6_Vxx(A2R10G10B10_SNORM_PACK32,   10_10_10_2_SNORM,  R10G10B10A2_SNORM,  WXYZ), /* 59 */
+   TU6_Vxx(A2R10G10B10_USCALED_PACK32, 10_10_10_2_UINT,   R10G10B10A2_UINT,   WXYZ), /* 60 */
+   TU6_Vxx(A2R10G10B10_SSCALED_PACK32, 10_10_10_2_SINT,   R10G10B10A2_SINT,   WXYZ), /* 61 */
+   TU6_VTC(A2R10G10B10_UINT_PACK32,    10_10_10_2_UINT,   R10G10B10A2_UINT,   WXYZ), /* 62 */
+   TU6_Vxx(A2R10G10B10_SINT_PACK32,    10_10_10_2_SINT,   R10G10B10A2_SINT,   WXYZ), /* 63 */
+   TU6_VTC(A2B10G10R10_UNORM_PACK32,   10_10_10_2_UNORM,  R10G10B10A2_UNORM,  WZYX), /* 64 */
+   TU6_Vxx(A2B10G10R10_SNORM_PACK32,   10_10_10_2_SNORM,  R10G10B10A2_SNORM,  WZYX), /* 65 */
+   TU6_Vxx(A2B10G10R10_USCALED_PACK32, 10_10_10_2_UINT,   R10G10B10A2_UINT,   WZYX), /* 66 */
+   TU6_Vxx(A2B10G10R10_SSCALED_PACK32, 10_10_10_2_SINT,   R10G10B10A2_SINT,   WZYX), /* 67 */
+   TU6_VTC(A2B10G10R10_UINT_PACK32,    10_10_10_2_UINT,   R10G10B10A2_UINT,   WZYX), /* 68 */
+   TU6_Vxx(A2B10G10R10_SINT_PACK32,    10_10_10_2_SINT,   R10G10B10A2_SINT,   WZYX), /* 69 */
+
+   /* 16-bit R */
+   TU6_VTC(R16_UNORM,                  16_UNORM,          R16_UNORM,          WZYX), /* 70 */
+   TU6_VTC(R16_SNORM,                  16_SNORM,          R16_SNORM,          WZYX), /* 71 */
+   TU6_Vxx(R16_USCALED,                16_UINT,           R16_UINT,           WZYX), /* 72 */
+   TU6_Vxx(R16_SSCALED,                16_SINT,           R16_SINT,           WZYX), /* 73 */
+   TU6_VTC(R16_UINT,                   16_UINT,           R16_UINT,           WZYX), /* 74 */
+   TU6_VTC(R16_SINT,                   16_SINT,           R16_SINT,           WZYX), /* 75 */
+   TU6_VTC(R16_SFLOAT,                 16_FLOAT,          R16_FLOAT,          WZYX), /* 76 */
+
+   /* 32-bit RG */
+   TU6_VTC(R16G16_UNORM,               16_16_UNORM,       R16G16_UNORM,       WZYX), /* 77 */
+   TU6_VTC(R16G16_SNORM,               16_16_SNORM,       R16G16_SNORM,       WZYX), /* 78 */
+   TU6_VTx(R16G16_USCALED,             16_16_UINT,        R16G16_UINT,        WZYX), /* 79 */
+   TU6_VTx(R16G16_SSCALED,             16_16_SINT,        R16G16_SINT,        WZYX), /* 80 */
+   TU6_VTC(R16G16_UINT,                16_16_UINT,        R16G16_UINT,        WZYX), /* 81 */
+   TU6_VTC(R16G16_SINT,                16_16_SINT,        R16G16_SINT,        WZYX), /* 82 */
+   TU6_VTC(R16G16_SFLOAT,              16_16_FLOAT,       R16G16_FLOAT,       WZYX), /* 83 */
+
+   /* 48-bit RGB */
+   TU6_Vxx(R16G16B16_UNORM,            16_16_16_UNORM,    R16G16B16_UNORM,    WZYX), /* 84 */
+   TU6_Vxx(R16G16B16_SNORM,            16_16_16_SNORM,    R16G16B16_SNORM,    WZYX), /* 85 */
+   TU6_Vxx(R16G16B16_USCALED,          16_16_16_UINT,     R16G16B16_UINT,     WZYX), /* 86 */
+   TU6_Vxx(R16G16B16_SSCALED,          16_16_16_SINT,     R16G16B16_SINT,     WZYX), /* 87 */
+   TU6_Vxx(R16G16B16_UINT,             16_16_16_UINT,     R16G16B16_UINT,     WZYX), /* 88 */
+   TU6_Vxx(R16G16B16_SINT,             16_16_16_SINT,     R16G16B16_SINT,     WZYX), /* 89 */
+   TU6_Vxx(R16G16B16_SFLOAT,           16_16_16_FLOAT,    R16G16B16_FLOAT,    WZYX), /* 90 */
+
+   /* 64-bit RGBA */
+   TU6_VTC(R16G16B16A16_UNORM,         16_16_16_16_UNORM, R16G16B16A16_UNORM, WZYX), /* 91 */
+   TU6_VTC(R16G16B16A16_SNORM,         16_16_16_16_SNORM, R16G16B16A16_SNORM, WZYX), /* 92 */
+   TU6_VTx(R16G16B16A16_USCALED,       16_16_16_16_UINT,  R16G16B16A16_UINT,  WZYX), /* 93 */
+   TU6_VTx(R16G16B16A16_SSCALED,       16_16_16_16_SINT,  R16G16B16A16_SINT,  WZYX), /* 94 */
+   TU6_VTC(R16G16B16A16_UINT,          16_16_16_16_UINT,  R16G16B16A16_UINT,  WZYX), /* 95 */
+   TU6_VTC(R16G16B16A16_SINT,          16_16_16_16_SINT,  R16G16B16A16_SINT,  WZYX), /* 96 */
+   TU6_VTC(R16G16B16A16_SFLOAT,        16_16_16_16_FLOAT, R16G16B16A16_FLOAT, WZYX), /* 97 */
+
+   /* 32-bit R */
+   TU6_VTC(R32_UINT,                   32_UINT,           R32_UINT,           WZYX), /* 98 */
+   TU6_VTC(R32_SINT,                   32_SINT,           R32_SINT,           WZYX), /* 99 */
+   TU6_VTC(R32_SFLOAT,                 32_FLOAT,          R32_FLOAT,          WZYX), /* 100 */
+
+   /* 64-bit RG */
+   TU6_VTC(R32G32_UINT,                32_32_UINT,        R32G32_UINT,        WZYX), /* 101 */
+   TU6_VTC(R32G32_SINT,                32_32_SINT,        R32G32_SINT,        WZYX), /* 102 */
+   TU6_VTC(R32G32_SFLOAT,              32_32_FLOAT,       R32G32_FLOAT,       WZYX), /* 103 */
+
+   /* 96-bit RGB */
+   TU6_VTx(R32G32B32_UINT,             32_32_32_UINT,     R32G32B32_UINT,     WZYX), /* 104 */
+   TU6_VTx(R32G32B32_SINT,             32_32_32_SINT,     R32G32B32_SINT,     WZYX), /* 105 */
+   TU6_VTx(R32G32B32_SFLOAT,           32_32_32_FLOAT,    R32G32B32_FLOAT,    WZYX), /* 106 */
+
+   /* 128-bit RGBA */
+   TU6_VTC(R32G32B32A32_UINT,          32_32_32_32_UINT,  R32G32B32A32_UINT,  WZYX), /* 107 */
+   TU6_VTC(R32G32B32A32_SINT,          32_32_32_32_SINT,  R32G32B32A32_SINT,  WZYX), /* 108 */
+   TU6_VTC(R32G32B32A32_SFLOAT,        32_32_32_32_FLOAT, R32G32B32A32_FLOAT, WZYX), /* 109 */
+
+   /* 64-bit R */
+   TU6_xxx(R64_UINT,                   64_UINT,           R64_UINT,           WZYX), /* 110 */
+   TU6_xxx(R64_SINT,                   64_SINT,           R64_SINT,           WZYX), /* 111 */
+   TU6_xxx(R64_SFLOAT,                 64_FLOAT,          R64_FLOAT,          WZYX), /* 112 */
+
+   /* 128-bit RG */
+   TU6_xxx(R64G64_UINT,                64_64_UINT,        R64G64_UINT,        WZYX), /* 113 */
+   TU6_xxx(R64G64_SINT,                64_64_SINT,        R64G64_SINT,        WZYX), /* 114 */
+   TU6_xxx(R64G64_SFLOAT,              64_64_FLOAT,       R64G64_FLOAT,       WZYX), /* 115 */
+
+   /* 192-bit RGB */
+   TU6_xxx(R64G64B64_UINT,             64_64_64_UINT,     R64G64B64_UINT,     WZYX), /* 116 */
+   TU6_xxx(R64G64B64_SINT,             64_64_64_SINT,     R64G64B64_SINT,     WZYX), /* 117 */
+   TU6_xxx(R64G64B64_SFLOAT,           64_64_64_FLOAT,    R64G64B64_FLOAT,    WZYX), /* 118 */
+
+   /* 256-bit RGBA */
+   TU6_xxx(R64G64B64A64_UINT,          64_64_64_64_UINT,  R64G64B64A64_UINT,  WZYX), /* 119 */
+   TU6_xxx(R64G64B64A64_SINT,          64_64_64_64_SINT,  R64G64B64A64_SINT,  WZYX), /* 120 */
+   TU6_xxx(R64G64B64A64_SFLOAT,        64_64_64_64_FLOAT, R64G64B64A64_FLOAT, WZYX), /* 121 */
+
+   /* 32-bit packed float */
+   TU6_VTC(B10G11R11_UFLOAT_PACK32,    11_11_10_FLOAT,    R11G11B10_FLOAT,    WZYX), /* 122 */
+   TU6_xTx(E5B9G9R9_UFLOAT_PACK32,     9_9_9_E5_FLOAT,    R9G9B9E5_FLOAT,     WZYX), /* 123 */
+
+   /* depth/stencil */
+   TU6_xTC(D16_UNORM,                  16_UNORM,          R16_UNORM,          WZYX), /* 124 */
+   TU6_xTC(X8_D24_UNORM_PACK32,        X8Z24_UNORM,       X8Z24_UNORM,        WZYX), /* 125 */
+   TU6_xTC(D32_SFLOAT,                 32_FLOAT,          R32_FLOAT,          WZYX), /* 126 */
+   TU6_xTC(S8_UINT,                    8_UINT,            R8_UNORM,           WZYX), /* 127 */
+   TU6_xxx(D16_UNORM_S8_UINT,          X8Z16_UNORM,       X8Z16_UNORM,        WZYX), /* 128 */
+   TU6_xTC(D24_UNORM_S8_UINT,          X8Z24_UNORM,       X8Z24_UNORM,        WZYX), /* 129 */
+   TU6_xTC(D32_SFLOAT_S8_UINT,         32_FLOAT,          R32_FLOAT,          WZYX), /* 130 */
+
+   /* compressed */
+   TU6_xTx(BC1_RGB_UNORM_BLOCK,        DXT1,              DXT1,               WZYX), /* 131 */
+   TU6_xTx(BC1_RGB_SRGB_BLOCK,         DXT1,              DXT1,               WZYX), /* 132 */
+   TU6_xTx(BC1_RGBA_UNORM_BLOCK,       DXT1,              DXT1,               WZYX), /* 133 */
+   TU6_xTx(BC1_RGBA_SRGB_BLOCK,        DXT1,              DXT1,               WZYX), /* 134 */
+   TU6_xTx(BC2_UNORM_BLOCK,            DXT3,              DXT3,               WZYX), /* 135 */
+   TU6_xTx(BC2_SRGB_BLOCK,             DXT3,              DXT3,               WZYX), /* 136 */
+   TU6_xTx(BC3_UNORM_BLOCK,            DXT5,              DXT5,               WZYX), /* 137 */
+   TU6_xTx(BC3_SRGB_BLOCK,             DXT5,              DXT5,               WZYX), /* 138 */
+   TU6_xTx(BC4_UNORM_BLOCK,            RGTC1_UNORM,       RGTC1_UNORM,        WZYX), /* 139 */
+   TU6_xTx(BC4_SNORM_BLOCK,            RGTC1_SNORM,       RGTC1_SNORM,        WZYX), /* 140 */
+   TU6_xTx(BC5_UNORM_BLOCK,            RGTC2_UNORM,       RGTC2_UNORM,        WZYX), /* 141 */
+   TU6_xTx(BC5_SNORM_BLOCK,            RGTC2_SNORM,       RGTC2_SNORM,        WZYX), /* 142 */
+   TU6_xTx(BC6H_UFLOAT_BLOCK,          BPTC_UFLOAT,       BPTC_UFLOAT,        WZYX), /* 143 */
+   TU6_xTx(BC6H_SFLOAT_BLOCK,          BPTC_FLOAT,        BPTC_FLOAT,         WZYX), /* 144 */
+   TU6_xTx(BC7_UNORM_BLOCK,            BPTC,              BPTC,               WZYX), /* 145 */
+   TU6_xTx(BC7_SRGB_BLOCK,             BPTC,              BPTC,               WZYX), /* 146 */
+   TU6_xTx(ETC2_R8G8B8_UNORM_BLOCK,    ETC2_RGB8,         ETC2_RGB8,          WZYX), /* 147 */
+   TU6_xTx(ETC2_R8G8B8_SRGB_BLOCK,     ETC2_RGB8,         ETC2_RGB8,          WZYX), /* 148 */
+   TU6_xTx(ETC2_R8G8B8A1_UNORM_BLOCK,  ETC2_RGB8A1,       ETC2_RGB8A1,        WZYX), /* 149 */
+   TU6_xTx(ETC2_R8G8B8A1_SRGB_BLOCK,   ETC2_RGB8A1,       ETC2_RGB8A1,        WZYX), /* 150 */
+   TU6_xTx(ETC2_R8G8B8A8_UNORM_BLOCK,  ETC2_RGBA8,        ETC2_RGBA8,         WZYX), /* 151 */
+   TU6_xTx(ETC2_R8G8B8A8_SRGB_BLOCK,   ETC2_RGBA8,        ETC2_RGBA8,         WZYX), /* 152 */
+   TU6_xTx(EAC_R11_UNORM_BLOCK,        ETC2_R11_UNORM,    ETC2_R11_UNORM,     WZYX), /* 153 */
+   TU6_xTx(EAC_R11_SNORM_BLOCK,        ETC2_R11_SNORM,    ETC2_R11_SNORM,     WZYX), /* 154 */
+   TU6_xTx(EAC_R11G11_UNORM_BLOCK,     ETC2_RG11_UNORM,   ETC2_RG11_UNORM,    WZYX), /* 155 */
+   TU6_xTx(EAC_R11G11_SNORM_BLOCK,     ETC2_RG11_SNORM,   ETC2_RG11_SNORM,    WZYX), /* 156 */
+   TU6_xTx(ASTC_4x4_UNORM_BLOCK,       ASTC_4x4,          ASTC_4x4,           WZYX), /* 157 */
+   TU6_xTx(ASTC_4x4_SRGB_BLOCK,        ASTC_4x4,          ASTC_4x4,           WZYX), /* 158 */
+   TU6_xTx(ASTC_5x4_UNORM_BLOCK,       ASTC_5x4,          ASTC_5x4,           WZYX), /* 159 */
+   TU6_xTx(ASTC_5x4_SRGB_BLOCK,        ASTC_5x4,          ASTC_5x4,           WZYX), /* 160 */
+   TU6_xTx(ASTC_5x5_UNORM_BLOCK,       ASTC_5x5,          ASTC_5x5,           WZYX), /* 161 */
+   TU6_xTx(ASTC_5x5_SRGB_BLOCK,        ASTC_5x5,          ASTC_5x5,           WZYX), /* 162 */
+   TU6_xTx(ASTC_6x5_UNORM_BLOCK,       ASTC_6x5,          ASTC_6x5,           WZYX), /* 163 */
+   TU6_xTx(ASTC_6x5_SRGB_BLOCK,        ASTC_6x5,          ASTC_6x5,           WZYX), /* 164 */
+   TU6_xTx(ASTC_6x6_UNORM_BLOCK,       ASTC_6x6,          ASTC_6x6,           WZYX), /* 165 */
+   TU6_xTx(ASTC_6x6_SRGB_BLOCK,        ASTC_6x6,          ASTC_6x6,           WZYX), /* 166 */
+   TU6_xTx(ASTC_8x5_UNORM_BLOCK,       ASTC_8x5,          ASTC_8x5,           WZYX), /* 167 */
+   TU6_xTx(ASTC_8x5_SRGB_BLOCK,        ASTC_8x5,          ASTC_8x5,           WZYX), /* 168 */
+   TU6_xTx(ASTC_8x6_UNORM_BLOCK,       ASTC_8x6,          ASTC_8x6,           WZYX), /* 169 */
+   TU6_xTx(ASTC_8x6_SRGB_BLOCK,        ASTC_8x6,          ASTC_8x6,           WZYX), /* 170 */
+   TU6_xTx(ASTC_8x8_UNORM_BLOCK,       ASTC_8x8,          ASTC_8x8,           WZYX), /* 171 */
+   TU6_xTx(ASTC_8x8_SRGB_BLOCK,        ASTC_8x8,          ASTC_8x8,           WZYX), /* 172 */
+   TU6_xTx(ASTC_10x5_UNORM_BLOCK,      ASTC_10x5,         ASTC_10x5,          WZYX), /* 173 */
+   TU6_xTx(ASTC_10x5_SRGB_BLOCK,       ASTC_10x5,         ASTC_10x5,          WZYX), /* 174 */
+   TU6_xTx(ASTC_10x6_UNORM_BLOCK,      ASTC_10x6,         ASTC_10x6,          WZYX), /* 175 */
+   TU6_xTx(ASTC_10x6_SRGB_BLOCK,       ASTC_10x6,         ASTC_10x6,          WZYX), /* 176 */
+   TU6_xTx(ASTC_10x8_UNORM_BLOCK,      ASTC_10x8,         ASTC_10x8,          WZYX), /* 177 */
+   TU6_xTx(ASTC_10x8_SRGB_BLOCK,       ASTC_10x8,         ASTC_10x8,          WZYX), /* 178 */
+   TU6_xTx(ASTC_10x10_UNORM_BLOCK,     ASTC_10x10,        ASTC_10x10,         WZYX), /* 179 */
+   TU6_xTx(ASTC_10x10_SRGB_BLOCK,      ASTC_10x10,        ASTC_10x10,         WZYX), /* 180 */
+   TU6_xTx(ASTC_12x10_UNORM_BLOCK,     ASTC_12x10,        ASTC_12x10,         WZYX), /* 181 */
+   TU6_xTx(ASTC_12x10_SRGB_BLOCK,      ASTC_12x10,        ASTC_12x10,         WZYX), /* 182 */
+   TU6_xTx(ASTC_12x12_UNORM_BLOCK,     ASTC_12x12,        ASTC_12x12,         WZYX), /* 183 */
+   TU6_xTx(ASTC_12x12_SRGB_BLOCK,      ASTC_12x12,        ASTC_12x12,         WZYX), /* 184 */
+};
+#undef TU_FORMAT_TABLE_FIRST
+#undef TU_FORMAT_TABLE_LAST
 
-bool
-tu6_format_vtx_supported(VkFormat vk_format)
+const struct tu_native_format *
+tu6_get_native_format(VkFormat format)
 {
-   enum pipe_format format = vk_format_to_pipe_format(vk_format);
-   return fd6_vertex_format(format) != FMT6_NONE;
+   const struct tu_native_format *fmt = NULL;
+
+   if (format >= tu6_format_table0_first && format <= tu6_format_table0_last)
+      fmt = &tu6_format_table0[format - tu6_format_table0_first];
+
+   return (fmt && fmt->present) ? fmt : NULL;
 }
 
-/* Map non-colorspace-converted YUV formats to RGB pipe formats where we can,
- * since our hardware doesn't support colorspace conversion.
- *
- * Really, we should probably be returning the RGB formats in
- * vk_format_to_pipe_format, but we don't have all the equivalent pipe formats
- * for VK RGB formats yet, and we'd have to switch all consumers of that
- * function at once.
- */
-enum pipe_format
-tu_vk_format_to_pipe_format(VkFormat vk_format)
+enum a6xx_2d_ifmt
+tu6_rb_fmt_to_ifmt(enum a6xx_color_fmt fmt)
 {
-   switch (vk_format) {
-   case VK_FORMAT_G8B8G8R8_422_UNORM: /* YUYV */
-      return PIPE_FORMAT_R8G8_R8B8_UNORM;
-   case VK_FORMAT_B8G8R8G8_422_UNORM: /* UYVY */
-      return PIPE_FORMAT_G8R8_B8R8_UNORM;
-   case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
-      return PIPE_FORMAT_G8_B8R8_420_UNORM;
-   case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
-      return PIPE_FORMAT_G8_B8_R8_420_UNORM;
+   switch (fmt) {
+   case RB6_A8_UNORM:
+   case RB6_R8_UNORM:
+   case RB6_R8_SNORM:
+   case RB6_R8G8_UNORM:
+   case RB6_R8G8_SNORM:
+   case RB6_R8G8B8A8_UNORM:
+   case RB6_R8G8B8_UNORM:
+   case RB6_R8G8B8A8_SNORM:
+      return R2D_UNORM8;
+
+   case RB6_R32_UINT:
+   case RB6_R32_SINT:
+   case RB6_R32G32_UINT:
+   case RB6_R32G32_SINT:
+   case RB6_R32G32B32A32_UINT:
+   case RB6_R32G32B32A32_SINT:
+      return R2D_INT32;
+
+   case RB6_R16_UINT:
+   case RB6_R16_SINT:
+   case RB6_R16G16_UINT:
+   case RB6_R16G16_SINT:
+   case RB6_R16G16B16A16_UINT:
+   case RB6_R16G16B16A16_SINT:
+      return R2D_INT16;
+
+   case RB6_R8_UINT:
+   case RB6_R8_SINT:
+   case RB6_R8G8_UINT:
+   case RB6_R8G8_SINT:
+   case RB6_R8G8B8A8_UINT:
+   case RB6_R8G8B8A8_SINT:
+      return R2D_INT8;
+
+   case RB6_R16_UNORM:
+   case RB6_R16_SNORM:
+   case RB6_R16G16_UNORM:
+   case RB6_R16G16_SNORM:
+   case RB6_R16G16B16A16_UNORM:
+   case RB6_R16G16B16A16_SNORM:
+   case RB6_R32_FLOAT:
+   case RB6_R32G32_FLOAT:
+   case RB6_R32G32B32A32_FLOAT:
+      return R2D_FLOAT32;
+
+   case RB6_R16_FLOAT:
+   case RB6_R16G16_FLOAT:
+   case RB6_R16G16B16A16_FLOAT:
+      return R2D_FLOAT16;
+
+   case RB6_R4G4B4A4_UNORM:
+   case RB6_R5G5B5A1_UNORM:
+   case RB6_R5G6B5_UNORM:
+   case RB6_R10G10B10A2_UNORM:
+   case RB6_R10G10B10A2_UINT:
+   case RB6_R11G11B10_FLOAT:
+   case RB6_X8Z24_UNORM:
+      // ???
+      return 0;
    default:
-      return vk_format_to_pipe_format(vk_format);
+      unreachable("bad format");
+      return 0;
    }
 }
 
-static struct tu_native_format
-tu6_format_color_unchecked(enum pipe_format format, enum a6xx_tile_mode tile_mode)
+static uint32_t
+tu_pack_mask(int bits)
 {
-   struct tu_native_format fmt = {
-      .fmt = fd6_color_format(format, tile_mode),
-      .swap = fd6_color_swap(format, tile_mode),
-   };
-
-   switch (format) {
-   case PIPE_FORMAT_Z24X8_UNORM:
-   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
-      fmt.fmt = FMT6_8_8_8_8_UNORM;
-      break;
-
-   default:
-      break;
-   }
-
-   return fmt;
+   assert(bits <= 32);
+   return (1ull << bits) - 1;
 }
 
-bool
-tu6_format_color_supported(enum pipe_format format)
+static uint32_t
+tu_pack_float32_for_unorm(float val, int bits)
 {
-   return tu6_format_color_unchecked(format, TILE6_LINEAR).fmt != FMT6_NONE;
+   const uint32_t max = tu_pack_mask(bits);
+   if (val < 0.0f)
+      return 0;
+   else if (val > 1.0f)
+      return max;
+   else
+      return _mesa_lroundevenf(val * (float) max);
 }
 
-struct tu_native_format
-tu6_format_color(enum pipe_format format, enum a6xx_tile_mode tile_mode)
+static uint32_t
+tu_pack_float32_for_snorm(float val, int bits)
 {
-   struct tu_native_format fmt = tu6_format_color_unchecked(format, tile_mode);
-   assert(fmt.fmt != FMT6_NONE);
-   return fmt;
+   const int32_t max = tu_pack_mask(bits - 1);
+   int32_t tmp;
+   if (val < -1.0f)
+      tmp = -max;
+   else if (val > 1.0f)
+      tmp = max;
+   else
+      tmp = _mesa_lroundevenf(val * (float) max);
+
+   return tmp & tu_pack_mask(bits);
 }
 
-static struct tu_native_format
-tu6_format_texture_unchecked(enum pipe_format format, enum a6xx_tile_mode tile_mode)
+static uint32_t
+tu_pack_float32_for_uscaled(float val, int bits)
 {
-   struct tu_native_format fmt = {
-      .fmt = fd6_texture_format(format, tile_mode),
-      .swap = fd6_texture_swap(format, tile_mode),
-   };
-
-   switch (format) {
-   case PIPE_FORMAT_Z24X8_UNORM:
-   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
-      /* freedreno uses Z24_UNORM_S8_UINT (sampling) or
-       * FMT6_Z24_UNORM_S8_UINT_AS_R8G8B8A8 (blits) for this format, while we use
-       * FMT6_8_8_8_8_UNORM or FMT6_Z24_UNORM_S8_UINT_AS_R8G8B8A8
-       */
-      fmt.fmt = FMT6_8_8_8_8_UNORM;
-      break;
+   const uint32_t max = tu_pack_mask(bits);
+   if (val < 0.0f)
+      return 0;
+   else if (val > (float) max)
+      return max;
+   else
+      return (uint32_t) val;
+}
 
-   default:
-      break;
-   }
+static uint32_t
+tu_pack_float32_for_sscaled(float val, int bits)
+{
+   const int32_t max = tu_pack_mask(bits - 1);
+   const int32_t min = -max - 1;
+   int32_t tmp;
+   if (val < (float) min)
+      tmp = min;
+   else if (val > (float) max)
+      tmp = max;
+   else
+      tmp = (int32_t) val;
+
+   return tmp & tu_pack_mask(bits);
+}
 
-   return fmt;
+static uint32_t
+tu_pack_uint32_for_uint(uint32_t val, int bits)
+{
+   return val & tu_pack_mask(bits);
 }
 
-struct tu_native_format
-tu6_format_texture(enum pipe_format format, enum a6xx_tile_mode tile_mode)
+static uint32_t
+tu_pack_int32_for_sint(int32_t val, int bits)
 {
-   struct tu_native_format fmt = tu6_format_texture_unchecked(format, tile_mode);
-   assert(fmt.fmt != FMT6_NONE);
-   return fmt;
+   return val & tu_pack_mask(bits);
 }
 
-bool
-tu6_format_texture_supported(enum pipe_format format)
+static uint32_t
+tu_pack_float32_for_sfloat(float val, int bits)
 {
-   return tu6_format_texture_unchecked(format, TILE6_LINEAR).fmt != FMT6_NONE;
+   assert(bits == 16 || bits == 32);
+   return bits == 16 ? util_float_to_half(val) : fui(val);
 }
 
-enum tu6_ubwc_compat_type {
-   TU6_UBWC_UNKNOWN_COMPAT,
-   TU6_UBWC_R8G8_UNORM,
-   TU6_UBWC_R8G8_INT,
-   TU6_UBWC_R8G8B8A8_UNORM,
-   TU6_UBWC_R8G8B8A8_INT,
-   TU6_UBWC_B8G8R8A8_UNORM,
-   TU6_UBWC_R16G16_INT,
-   TU6_UBWC_R16G16B16A16_INT,
-   TU6_UBWC_R32_INT,
-   TU6_UBWC_R32G32_INT,
-   TU6_UBWC_R32G32B32A32_INT,
-   TU6_UBWC_R32_FLOAT,
+union tu_clear_component_value {
+   float float32;
+   int32_t int32;
+   uint32_t uint32;
 };
 
-static enum tu6_ubwc_compat_type
-tu6_ubwc_compat_mode(VkFormat format)
+static uint32_t
+tu_pack_clear_component_value(union tu_clear_component_value val,
+                              const struct vk_format_channel_description *ch)
 {
-   switch (format) {
-   case VK_FORMAT_R8G8_UNORM:
-   case VK_FORMAT_R8G8_SRGB:
-      return TU6_UBWC_R8G8_UNORM;
-
-   case VK_FORMAT_R8G8_UINT:
-   case VK_FORMAT_R8G8_SINT:
-      return TU6_UBWC_R8G8_INT;
-
-   case VK_FORMAT_R8G8B8A8_UNORM:
-   case VK_FORMAT_R8G8B8A8_SRGB:
-   case VK_FORMAT_A8B8G8R8_UNORM_PACK32:
-   case VK_FORMAT_A8B8G8R8_SRGB_PACK32:
-      return TU6_UBWC_R8G8B8A8_UNORM;
-
-   case VK_FORMAT_R8G8B8A8_UINT:
-   case VK_FORMAT_R8G8B8A8_SINT:
-   case VK_FORMAT_A8B8G8R8_UINT_PACK32:
-   case VK_FORMAT_A8B8G8R8_SINT_PACK32:
-      return TU6_UBWC_R8G8B8A8_INT;
-
-   case VK_FORMAT_R16G16_UINT:
-   case VK_FORMAT_R16G16_SINT:
-      return TU6_UBWC_R16G16_INT;
-
-   case VK_FORMAT_R16G16B16A16_UINT:
-   case VK_FORMAT_R16G16B16A16_SINT:
-      return TU6_UBWC_R16G16B16A16_INT;
-
-   case VK_FORMAT_R32_UINT:
-   case VK_FORMAT_R32_SINT:
-      return TU6_UBWC_R32_INT;
-
-   case VK_FORMAT_R32G32_UINT:
-   case VK_FORMAT_R32G32_SINT:
-      return TU6_UBWC_R32G32_INT;
-
-   case VK_FORMAT_R32G32B32A32_UINT:
-   case VK_FORMAT_R32G32B32A32_SINT:
-      return TU6_UBWC_R32G32B32A32_INT;
-
-   case VK_FORMAT_D32_SFLOAT:
-   case VK_FORMAT_R32_SFLOAT:
-      /* TODO: a630 blob allows these, but not a660.  When is it legal? */
-      return TU6_UBWC_UNKNOWN_COMPAT;
-
-   case VK_FORMAT_B8G8R8A8_UNORM:
-   case VK_FORMAT_B8G8R8A8_SRGB:
-      /* The blob doesn't list these as compatible, but they surely are.
-       * freedreno's happy to cast between them, and zink would really like
-       * to.
-       */
-      return TU6_UBWC_B8G8R8A8_UNORM;
-
+   uint32_t packed;
+
+   switch (ch->type) {
+   case VK_FORMAT_TYPE_UNSIGNED:
+      /* normalized, scaled, or pure integer */
+      assert(ch->normalized + ch->scaled + ch->pure_integer == 1);
+      if (ch->normalized)
+         packed = tu_pack_float32_for_unorm(val.float32, ch->size);
+      else if (ch->scaled)
+         packed = tu_pack_float32_for_uscaled(val.float32, ch->size);
+      else
+         packed = tu_pack_uint32_for_uint(val.uint32, ch->size);
+      break;
+   case VK_FORMAT_TYPE_SIGNED:
+      /* normalized, scaled, or pure integer */
+      assert(ch->normalized + ch->scaled + ch->pure_integer == 1);
+      if (ch->normalized)
+         packed = tu_pack_float32_for_snorm(val.float32, ch->size);
+      else if (ch->scaled)
+         packed = tu_pack_float32_for_sscaled(val.float32, ch->size);
+      else
+         packed = tu_pack_int32_for_sint(val.int32, ch->size);
+      break;
+   case VK_FORMAT_TYPE_FLOAT:
+      packed = tu_pack_float32_for_sfloat(val.float32, ch->size);
+      break;
    default:
-      return TU6_UBWC_UNKNOWN_COMPAT;
+      unreachable("unexpected channel type");
+      packed = 0;
+      break;
    }
+
+   assert((packed & tu_pack_mask(ch->size)) == packed);
+   return packed;
 }
 
-bool
-tu6_mutable_format_list_ubwc_compatible(const VkImageFormatListCreateInfo *fmt_list)
+static const struct vk_format_channel_description *
+tu_get_format_channel_description(const struct vk_format_description *desc,
+                                  int comp)
 {
-   if (!fmt_list || !fmt_list->viewFormatCount)
-      return false;
-
-   /* We're only looking at format list cross compatibility here, check
-    * ubwc_possible() for the base "is the format UBWC-able at all?"
-    */
-   if (fmt_list->viewFormatCount == 1)
-      return true;
-
-   enum tu6_ubwc_compat_type type =
-      tu6_ubwc_compat_mode(fmt_list->pViewFormats[0]);
-   if (type == TU6_UBWC_UNKNOWN_COMPAT)
-      return false;
-
-   for (uint32_t i = 1; i < fmt_list->viewFormatCount; i++) {
-      if (tu6_ubwc_compat_mode(fmt_list->pViewFormats[i]) != type)
-         return false;
+   switch (desc->swizzle[comp]) {
+   case VK_SWIZZLE_X:
+      return &desc->channel[0];
+   case VK_SWIZZLE_Y:
+      return &desc->channel[1];
+   case VK_SWIZZLE_Z:
+      return &desc->channel[2];
+   case VK_SWIZZLE_W:
+      return &desc->channel[3];
+   default:
+      return NULL;
    }
-
-   return true;
 }
 
-static void
-tu_physical_device_get_format_properties(
-   struct tu_physical_device *physical_device,
-   VkFormat vk_format,
-   VkFormatProperties3 *out_properties)
+static union tu_clear_component_value
+tu_get_clear_component_value(const VkClearValue *val, int comp, bool color)
 {
-   VkFormatFeatureFlags2 linear = 0, optimal = 0, buffer = 0;
-   enum pipe_format format = tu_vk_format_to_pipe_format(vk_format);
-   const struct util_format_description *desc = util_format_description(format);
-
-   bool supported_vtx = tu6_format_vtx_supported(vk_format);
-   bool supported_color = tu6_format_color_supported(format);
-   bool supported_tex = tu6_format_texture_supported(format);
-   bool is_npot = !util_is_power_of_two_or_zero(desc->block.bits);
-
-   if (format == PIPE_FORMAT_NONE ||
-       !(supported_vtx || supported_color || supported_tex)) {
-      goto end;
+   union tu_clear_component_value tmp;
+   if (color) {
+      assert(comp < 4);
+      tmp.uint32 = val->color.uint32[comp];
+   } else {
+      assert(comp < 2);
+      if (comp == 0)
+         tmp.float32 = val->depthStencil.depth;
+      else
+         tmp.uint32 = val->depthStencil.stencil;
    }
 
-   /* We don't support BufferToImage/ImageToBuffer for npot formats */
-   if (!is_npot)
-      buffer |= VK_FORMAT_FEATURE_TRANSFER_SRC_BIT | VK_FORMAT_FEATURE_TRANSFER_DST_BIT;
+   return tmp;
+}
 
-   if (supported_vtx)
-      buffer |= VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT;
+/**
+ * Pack a VkClearValue into a 128-bit buffer.  \a format is respected except
+ * for the component order.  The components are always packed in WZYX order
+ * (i.e., msb is white and lsb is red).
+ *
+ * Return the number of uint32_t's used.
+ */
+int
+tu_pack_clear_value(const VkClearValue *val, VkFormat format, uint32_t buf[4])
+{
+   const struct vk_format_description *desc = vk_format_description(format);
+   assert(desc && desc->layout == VK_FORMAT_LAYOUT_PLAIN);
+
+   /* S8_UINT is special and has no depth */
+   const int max_components =
+      format == VK_FORMAT_S8_UINT ? 2 : desc->nr_channels;
+
+   int buf_offset = 0;
+   int bit_shift = 0;
+   for (int comp = 0; comp < max_components; comp++) {
+      const struct vk_format_channel_description *ch =
+         tu_get_format_channel_description(desc, comp);
+      if (!ch) {
+         assert(format == VK_FORMAT_S8_UINT && comp == 0);
+         continue;
+      }
 
-   if (supported_tex)
-      buffer |= VK_FORMAT_FEATURE_UNIFORM_TEXEL_BUFFER_BIT;
+      union tu_clear_component_value v = tu_get_clear_component_value(
+         val, comp, desc->colorspace != VK_FORMAT_COLORSPACE_ZS);
 
-   /* Don't support anything but texel buffers for non-power-of-two formats
-    * with 3 components. We'd need several workarounds for copying and
-    * clearing them because they're not renderable.
-    */
-   if (supported_tex && !is_npot) {
-      optimal |= VK_FORMAT_FEATURE_TRANSFER_SRC_BIT |
-                 VK_FORMAT_FEATURE_TRANSFER_DST_BIT |
-                 VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
-                 VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_MINMAX_BIT |
-                 VK_FORMAT_FEATURE_COSITED_CHROMA_SAMPLES_BIT |
-                 VK_FORMAT_FEATURE_MIDPOINT_CHROMA_SAMPLES_BIT;
-
-      /* no blit src bit for YUYV/NV12/I420 formats */
-      if (desc->layout != UTIL_FORMAT_LAYOUT_SUBSAMPLED &&
-          desc->layout != UTIL_FORMAT_LAYOUT_PLANAR2 &&
-          desc->layout != UTIL_FORMAT_LAYOUT_PLANAR3)
-         optimal |= VK_FORMAT_FEATURE_BLIT_SRC_BIT;
-
-      if (desc->layout != UTIL_FORMAT_LAYOUT_SUBSAMPLED)
-         optimal |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_YCBCR_CONVERSION_LINEAR_FILTER_BIT;
-
-      if (!vk_format_is_int(vk_format)) {
-         optimal |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT;
-
-         if (physical_device->vk.supported_extensions.EXT_filter_cubic)
-            optimal |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_CUBIC_BIT_EXT;
+      /* move to the next uint32_t when there is not enough space */
+      assert(ch->size <= 32);
+      if (bit_shift + ch->size > 32) {
+         buf_offset++;
+         bit_shift = 0;
       }
-   }
 
-   if (supported_color) {
-      assert(supported_tex);
-      optimal |= VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT |
-                 VK_FORMAT_FEATURE_BLIT_DST_BIT |
-                 VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT |
-                 VK_FORMAT_FEATURE_2_STORAGE_READ_WITHOUT_FORMAT_BIT |
-                 VK_FORMAT_FEATURE_2_STORAGE_WRITE_WITHOUT_FORMAT_BIT;
+      if (bit_shift == 0)
+         buf[buf_offset] = 0;
 
-      buffer |= VK_FORMAT_FEATURE_STORAGE_TEXEL_BUFFER_BIT |
-                VK_FORMAT_FEATURE_2_STORAGE_READ_WITHOUT_FORMAT_BIT |
-                VK_FORMAT_FEATURE_2_STORAGE_WRITE_WITHOUT_FORMAT_BIT;
+      buf[buf_offset] |= tu_pack_clear_component_value(v, ch) << bit_shift;
+      bit_shift += ch->size;
+   }
 
-      /* TODO: The blob also exposes these for R16G16_UINT/R16G16_SINT, but we
-       * don't have any tests for those.
-       */
-      if (vk_format == VK_FORMAT_R32_UINT || vk_format == VK_FORMAT_R32_SINT) {
-         optimal |= VK_FORMAT_FEATURE_STORAGE_IMAGE_ATOMIC_BIT;
-         buffer |= VK_FORMAT_FEATURE_STORAGE_TEXEL_BUFFER_ATOMIC_BIT;
-      }
+   return buf_offset + 1;
+}
 
-      if (!util_format_is_pure_integer(format))
-         optimal |= VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BLEND_BIT;
+static void
+tu_physical_device_get_format_properties(
+   struct tu_physical_device *physical_device,
+   VkFormat format,
+   VkFormatProperties *out_properties)
+{
+   VkFormatFeatureFlags linear = 0, tiled = 0, buffer = 0;
+   const struct vk_format_description *desc = vk_format_description(format);
+   const struct tu_native_format *native_fmt = tu6_get_native_format(format);
+   if (!desc || !native_fmt) {
+      out_properties->linearTilingFeatures = linear;
+      out_properties->optimalTilingFeatures = tiled;
+      out_properties->bufferFeatures = buffer;
+      return;
    }
 
-   /* For the most part, we can do anything with a linear image that we could
-    * do with a tiled image. However, we can't support sysmem rendering with a
-    * linear depth texture, because we don't know if there's a bit to control
-    * the tiling of the depth buffer in BYPASS mode, and the blob also
-    * disables linear depth rendering, so there's no way to discover it. We
-    * also can't force GMEM mode, because there are other situations where we
-    * have to use sysmem rendering. So follow the blob here, and only enable
-    * DEPTH_STENCIL_ATTACHMENT_BIT for the optimal features.
-    */
-   linear = optimal;
-   if (tu6_pipe2depth(vk_format) != (enum a6xx_depth_format)~0)
-      optimal |= VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT;
-
-   if (!tiling_possible(vk_format) &&
-       /* We don't actually support tiling for this format, but we need to
-        * fake it as it's required by VK_KHR_sampler_ycbcr_conversion.
-        */
-       vk_format != VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM) {
-      optimal = 0;
-   }
+   linear |= VK_FORMAT_FEATURE_TRANSFER_SRC_BIT | VK_FORMAT_FEATURE_TRANSFER_DST_BIT;
+   tiled |= VK_FORMAT_FEATURE_TRANSFER_SRC_BIT | VK_FORMAT_FEATURE_TRANSFER_DST_BIT;
+   buffer |= VK_FORMAT_FEATURE_TRANSFER_SRC_BIT | VK_FORMAT_FEATURE_TRANSFER_DST_BIT;
 
-   if (vk_format == VK_FORMAT_G8B8G8R8_422_UNORM ||
-       vk_format == VK_FORMAT_B8G8R8G8_422_UNORM ||
-       vk_format == VK_FORMAT_G8_B8R8_2PLANE_420_UNORM ||
-       vk_format == VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM) {
-      /* Disable buffer texturing of subsampled (422) and planar YUV textures.
-       * The subsampling requirement comes from "If format is a block-compressed
-       * format, then bufferFeatures must not support any features for the
-       * format" plus the specification of subsampled as 2x1 compressed block
-       * format.  I couldn't find the citation for planar, but 1D access of
-       * planar YUV would be really silly.
-       */
-      buffer = 0;
+   if (native_fmt->tex >= 0) {
+      linear |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT;
+      tiled |= VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT;
+      buffer |= VK_FORMAT_FEATURE_UNIFORM_TEXEL_BUFFER_BIT;
    }
 
-   /* We don't support writing into VK__FORMAT_*_PACK16 images/buffers  */
-   if (desc->nr_channels > 2 && desc->block.bits == 16) {
-      buffer &= VK_FORMAT_FEATURE_UNIFORM_TEXEL_BUFFER_BIT;
-      linear &= ~(VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT |
-                  VK_FORMAT_FEATURE_STORAGE_IMAGE_ATOMIC_BIT);
-      optimal &= ~(VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT |
-                   VK_FORMAT_FEATURE_STORAGE_IMAGE_ATOMIC_BIT);
+   if (native_fmt->rb >= 0) {
+      linear |= VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT;
+      tiled |= VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT;
    }
 
-   /* All our depth formats support shadow comparisons. */
-   if (vk_format_has_depth(vk_format) && (optimal & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT)) {
-      optimal |= VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_DEPTH_COMPARISON_BIT;
-      linear |= VK_FORMAT_FEATURE_2_SAMPLED_IMAGE_DEPTH_COMPARISON_BIT;
+   if (native_fmt->vtx >= 0) {
+      buffer |= VK_FORMAT_FEATURE_VERTEX_BUFFER_BIT;
    }
 
-   /* From the Vulkan 1.3.205 spec, section 19.3 "43.3. Required Format Support":
-    *
-    *    Mandatory format support: depth/stencil with VkImageType
-    *    VK_IMAGE_TYPE_2D
-    *    [...]
-    *    bufferFeatures must not support any features for these formats
-    */
-   if (vk_format_is_depth_or_stencil(vk_format))
-      buffer = 0;
-
-   /* D32_SFLOAT_S8_UINT is tiled as two images, so no linear format
-    * blob enables some linear features, but its not useful, so don't bother.
-    */
-   if (vk_format == VK_FORMAT_D32_SFLOAT_S8_UINT)
-      linear = 0;
-
-end:
    out_properties->linearTilingFeatures = linear;
-   out_properties->optimalTilingFeatures = optimal;
+   out_properties->optimalTilingFeatures = tiled;
    out_properties->bufferFeatures = buffer;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
+tu_GetPhysicalDeviceFormatProperties(VkPhysicalDevice physicalDevice,
+                                     VkFormat format,
+                                     VkFormatProperties *pFormatProperties)
+{
+   TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
+
+   tu_physical_device_get_format_properties(physical_device, format,
+                                            pFormatProperties);
+}
+
+void
 tu_GetPhysicalDeviceFormatProperties2(
    VkPhysicalDevice physicalDevice,
    VkFormat format,
@@ -397,61 +663,18 @@ tu_GetPhysicalDeviceFormatProperties2(
 {
    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
 
-   VkFormatProperties3 local_props3;
-   VkFormatProperties3 *props3 =
-      vk_find_struct(pFormatProperties->pNext, FORMAT_PROPERTIES_3);
-   if (!props3)
-      props3 = &local_props3;
-
    tu_physical_device_get_format_properties(
-      physical_device, format, props3);
-
-   pFormatProperties->formatProperties = (VkFormatProperties) {
-      .linearTilingFeatures = props3->linearTilingFeatures,
-      .optimalTilingFeatures = props3->optimalTilingFeatures,
-      .bufferFeatures = props3->bufferFeatures,
-   };
-
-   VkDrmFormatModifierPropertiesListEXT *list =
-      vk_find_struct(pFormatProperties->pNext, DRM_FORMAT_MODIFIER_PROPERTIES_LIST_EXT);
-   if (list) {
-      VK_OUTARRAY_MAKE_TYPED(VkDrmFormatModifierPropertiesEXT, out,
-                             list->pDrmFormatModifierProperties,
-                             &list->drmFormatModifierCount);
-
-      if (pFormatProperties->formatProperties.linearTilingFeatures) {
-         vk_outarray_append_typed(VkDrmFormatModifierPropertiesEXT, &out, mod_props) {
-            mod_props->drmFormatModifier = DRM_FORMAT_MOD_LINEAR;
-            mod_props->drmFormatModifierPlaneCount = tu6_plane_count(format);
-            mod_props->drmFormatModifierTilingFeatures =
-               pFormatProperties->formatProperties.linearTilingFeatures;
-         }
-      }
-
-      /* note: ubwc_possible() argument values to be ignored except for format */
-      if (pFormatProperties->formatProperties.optimalTilingFeatures &&
-          tiling_possible(format) &&
-          ubwc_possible(NULL, format, VK_IMAGE_TYPE_2D, 0, 0,
-                        physical_device->info, VK_SAMPLE_COUNT_1_BIT,
-                        false)) {
-         vk_outarray_append_typed(VkDrmFormatModifierPropertiesEXT, &out, mod_props) {
-            mod_props->drmFormatModifier = DRM_FORMAT_MOD_QCOM_COMPRESSED;
-            mod_props->drmFormatModifierPlaneCount = tu6_plane_count(format);
-            mod_props->drmFormatModifierTilingFeatures =
-               pFormatProperties->formatProperties.optimalTilingFeatures;
-         }
-      }
-   }
+      physical_device, format, &pFormatProperties->formatProperties);
 }
 
 static VkResult
 tu_get_image_format_properties(
    struct tu_physical_device *physical_device,
    const VkPhysicalDeviceImageFormatInfo2 *info,
-   VkImageFormatProperties *pImageFormatProperties,
-   VkFormatFeatureFlags *p_feature_flags)
+   VkImageFormatProperties *pImageFormatProperties)
+
 {
-   VkFormatProperties3 format_props;
+   VkFormatProperties format_props;
    VkFormatFeatureFlags format_feature_flags;
    VkExtent3D maxExtent;
    uint32_t maxMipLevels;
@@ -460,53 +683,12 @@ tu_get_image_format_properties(
 
    tu_physical_device_get_format_properties(physical_device, info->format,
                                             &format_props);
-
-   switch (info->tiling) {
-   case VK_IMAGE_TILING_LINEAR:
+   if (info->tiling == VK_IMAGE_TILING_LINEAR) {
       format_feature_flags = format_props.linearTilingFeatures;
-      break;
-
-   case VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT: {
-      const VkPhysicalDeviceImageDrmFormatModifierInfoEXT *drm_info =
-         vk_find_struct_const(info->pNext, PHYSICAL_DEVICE_IMAGE_DRM_FORMAT_MODIFIER_INFO_EXT);
-
-      switch (drm_info->drmFormatModifier) {
-      case DRM_FORMAT_MOD_QCOM_COMPRESSED:
-         /* falling back to linear/non-UBWC isn't possible with explicit modifier */
-
-         /* formats which don't support tiling */
-         if (!format_props.optimalTilingFeatures ||
-             !tiling_possible(info->format))
-            return VK_ERROR_FORMAT_NOT_SUPPORTED;
-
-         if (info->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) {
-            const VkImageFormatListCreateInfo *format_list =
-               vk_find_struct_const(info->pNext,
-                                    IMAGE_FORMAT_LIST_CREATE_INFO);
-            if (!tu6_mutable_format_list_ubwc_compatible(format_list))
-               return VK_ERROR_FORMAT_NOT_SUPPORTED;
-         }
-
-         if (!ubwc_possible(NULL, info->format, info->type, info->usage,
-                            info->usage, physical_device->info, sampleCounts,
-                            false)) {
-            return VK_ERROR_FORMAT_NOT_SUPPORTED;
-         }
-
-         format_feature_flags = format_props.optimalTilingFeatures;
-         break;
-      case DRM_FORMAT_MOD_LINEAR:
-         format_feature_flags = format_props.linearTilingFeatures;
-         break;
-      default:
-         return VK_ERROR_FORMAT_NOT_SUPPORTED;
-      }
-   } break;
-   case VK_IMAGE_TILING_OPTIMAL:
+   } else if (info->tiling == VK_IMAGE_TILING_OPTIMAL) {
       format_feature_flags = format_props.optimalTilingFeatures;
-      break;
-   default:
-      unreachable("bad VkPhysicalDeviceImageFormatInfo2");
+   } else {
+      unreachable("bad VkImageTiling");
    }
 
    if (format_feature_flags == 0)
@@ -549,50 +731,29 @@ tu_get_image_format_properties(
          VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT)) &&
        !(info->flags & VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT) &&
        !(info->usage & VK_IMAGE_USAGE_STORAGE_BIT)) {
-      sampleCounts |= VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT;
-      /* note: most operations support 8 samples (GMEM render/resolve do at least)
-       * but some do not (which ones?), just disable 8 samples completely,
-       * (no 8x msaa matches the blob driver behavior)
-       */
+      sampleCounts |= VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT |
+                      VK_SAMPLE_COUNT_8_BIT;
    }
 
-   /* From the Vulkan 1.3.206 spec:
-    *
-    * "VK_IMAGE_CREATE_EXTENDED_USAGE_BIT specifies that the image can be
-    * created with usage flags that are not supported for the format the image
-    * is created with but are supported for at least one format a VkImageView
-    * created from the image can have."
-    *
-    * This means we should relax checks that only depend on the
-    * format_feature_flags, to allow the user to create images that may be
-    * e.g. reinterpreted as storage when the original format doesn't allow it.
-    * The user will have to check against the format features anyway.
-    * Otherwise we'd unnecessarily disallow it.
-    */
-
-   VkImageUsageFlags image_usage = info->usage;
-   if (info->flags & VK_IMAGE_CREATE_EXTENDED_USAGE_BIT)
-      image_usage = 0;
-
-   if (image_usage & VK_IMAGE_USAGE_SAMPLED_BIT) {
+   if (info->usage & VK_IMAGE_USAGE_SAMPLED_BIT) {
       if (!(format_feature_flags & VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT)) {
          goto unsupported;
       }
    }
 
-   if (image_usage & VK_IMAGE_USAGE_STORAGE_BIT) {
+   if (info->usage & VK_IMAGE_USAGE_STORAGE_BIT) {
       if (!(format_feature_flags & VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT)) {
          goto unsupported;
       }
    }
 
-   if (image_usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
+   if (info->usage & VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
       if (!(format_feature_flags & VK_FORMAT_FEATURE_COLOR_ATTACHMENT_BIT)) {
          goto unsupported;
       }
    }
 
-   if (image_usage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
+   if (info->usage & VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
       if (!(format_feature_flags &
             VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT)) {
          goto unsupported;
@@ -611,9 +772,6 @@ tu_get_image_format_properties(
       .maxResourceSize = UINT32_MAX,
    };
 
-   if (p_feature_flags)
-      *p_feature_flags = format_feature_flags;
-
    return VK_SUCCESS;
 unsupported:
    *pImageFormatProperties = (VkImageFormatProperties) {
@@ -627,12 +785,38 @@ unsupported:
    return VK_ERROR_FORMAT_NOT_SUPPORTED;
 }
 
+VkResult
+tu_GetPhysicalDeviceImageFormatProperties(
+   VkPhysicalDevice physicalDevice,
+   VkFormat format,
+   VkImageType type,
+   VkImageTiling tiling,
+   VkImageUsageFlags usage,
+   VkImageCreateFlags createFlags,
+   VkImageFormatProperties *pImageFormatProperties)
+{
+   TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
+
+   const VkPhysicalDeviceImageFormatInfo2 info = {
+      .sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_FORMAT_INFO_2,
+      .pNext = NULL,
+      .format = format,
+      .type = type,
+      .tiling = tiling,
+      .usage = usage,
+      .flags = createFlags,
+   };
+
+   return tu_get_image_format_properties(physical_device, &info,
+                                         pImageFormatProperties);
+}
+
 static VkResult
 tu_get_external_image_format_properties(
    const struct tu_physical_device *physical_device,
    const VkPhysicalDeviceImageFormatInfo2 *pImageFormatInfo,
    VkExternalMemoryHandleTypeFlagBits handleType,
-   VkExternalImageFormatProperties *external_properties)
+   VkExternalMemoryProperties *external_properties)
 {
    VkExternalMemoryFeatureFlagBits flags = 0;
    VkExternalMemoryHandleTypeFlags export_flags = 0;
@@ -659,7 +843,7 @@ tu_get_external_image_format_properties(
             VK_EXTERNAL_MEMORY_HANDLE_TYPE_DMA_BUF_BIT_EXT;
          break;
       default:
-         return vk_errorf(physical_device, VK_ERROR_FORMAT_NOT_SUPPORTED,
+         return vk_errorf(physical_device->instance, VK_ERROR_FORMAT_NOT_SUPPORTED,
                           "VkExternalMemoryTypeFlagBits(0x%x) unsupported for VkImageType(%d)",
                           handleType, pImageFormatInfo->type);
       }
@@ -669,24 +853,21 @@ tu_get_external_image_format_properties(
       compat_flags = VK_EXTERNAL_MEMORY_HANDLE_TYPE_HOST_ALLOCATION_BIT_EXT;
       break;
    default:
-      return vk_errorf(physical_device, VK_ERROR_FORMAT_NOT_SUPPORTED,
+      return vk_errorf(physical_device->instance, VK_ERROR_FORMAT_NOT_SUPPORTED,
                        "VkExternalMemoryTypeFlagBits(0x%x) unsupported",
                        handleType);
    }
 
-   if (external_properties) {
-      external_properties->externalMemoryProperties =
-         (VkExternalMemoryProperties) {
-            .externalMemoryFeatures = flags,
-            .exportFromImportedHandleTypes = export_flags,
-            .compatibleHandleTypes = compat_flags,
-         };
-   }
+   *external_properties = (VkExternalMemoryProperties) {
+      .externalMemoryFeatures = flags,
+      .exportFromImportedHandleTypes = export_flags,
+      .compatibleHandleTypes = compat_flags,
+   };
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_GetPhysicalDeviceImageFormatProperties2(
    VkPhysicalDevice physicalDevice,
    const VkPhysicalDeviceImageFormatInfo2 *base_info,
@@ -694,15 +875,11 @@ tu_GetPhysicalDeviceImageFormatProperties2(
 {
    TU_FROM_HANDLE(tu_physical_device, physical_device, physicalDevice);
    const VkPhysicalDeviceExternalImageFormatInfo *external_info = NULL;
-   const VkPhysicalDeviceImageViewImageFormatInfoEXT *image_view_info = NULL;
    VkExternalImageFormatProperties *external_props = NULL;
-   VkFilterCubicImageViewImageFormatPropertiesEXT *cubic_props = NULL;
-   VkFormatFeatureFlags format_feature_flags;
-   VkSamplerYcbcrConversionImageFormatProperties *ycbcr_props = NULL;
    VkResult result;
 
-   result = tu_get_image_format_properties(physical_device,
-      base_info, &base_props->imageFormatProperties, &format_feature_flags);
+   result = tu_get_image_format_properties(
+      physical_device, base_info, &base_props->imageFormatProperties);
    if (result != VK_SUCCESS)
       return result;
 
@@ -713,9 +890,6 @@ tu_GetPhysicalDeviceImageFormatProperties2(
       case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTERNAL_IMAGE_FORMAT_INFO:
          external_info = (const void *) s;
          break;
-      case VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_IMAGE_VIEW_IMAGE_FORMAT_INFO_EXT:
-         image_view_info = (const void *) s;
-         break;
       default:
          break;
       }
@@ -728,12 +902,6 @@ tu_GetPhysicalDeviceImageFormatProperties2(
       case VK_STRUCTURE_TYPE_EXTERNAL_IMAGE_FORMAT_PROPERTIES:
          external_props = (void *) s;
          break;
-      case VK_STRUCTURE_TYPE_FILTER_CUBIC_IMAGE_VIEW_IMAGE_FORMAT_PROPERTIES_EXT:
-         cubic_props = (void *) s;
-         break;
-      case VK_STRUCTURE_TYPE_SAMPLER_YCBCR_CONVERSION_IMAGE_FORMAT_PROPERTIES:
-         ycbcr_props = (void *) s;
-         break;
       default:
          break;
       }
@@ -748,29 +916,11 @@ tu_GetPhysicalDeviceImageFormatProperties2(
    if (external_info && external_info->handleType != 0) {
       result = tu_get_external_image_format_properties(
          physical_device, base_info, external_info->handleType,
-         external_props);
+         &external_props->externalMemoryProperties);
       if (result != VK_SUCCESS)
          goto fail;
    }
 
-   if (cubic_props) {
-      /* note: blob only allows cubic filtering for 2D and 2D array views
-       * its likely we can enable it for 1D and CUBE, needs testing however
-       */
-      if ((image_view_info->imageViewType == VK_IMAGE_VIEW_TYPE_2D ||
-           image_view_info->imageViewType == VK_IMAGE_VIEW_TYPE_2D_ARRAY) &&
-          (format_feature_flags & VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_CUBIC_BIT_EXT)) {
-         cubic_props->filterCubic = true;
-         cubic_props->filterCubicMinmax = true;
-      } else {
-         cubic_props->filterCubic = false;
-         cubic_props->filterCubicMinmax = false;
-      }
-   }
-
-   if (ycbcr_props)
-      ycbcr_props->combinedImageSamplerDescriptorCount = 1;
-
    return VK_SUCCESS;
 
 fail:
@@ -782,13 +932,28 @@ fail:
        *    the implementation for use in vkCreateImage, then all members of
        *    imageFormatProperties will be filled with zero.
        */
-      base_props->imageFormatProperties = (VkImageFormatProperties) {};
+      base_props->imageFormatProperties = (VkImageFormatProperties) { 0 };
    }
 
    return result;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
+tu_GetPhysicalDeviceSparseImageFormatProperties(
+   VkPhysicalDevice physicalDevice,
+   VkFormat format,
+   VkImageType type,
+   uint32_t samples,
+   VkImageUsageFlags usage,
+   VkImageTiling tiling,
+   uint32_t *pNumProperties,
+   VkSparseImageFormatProperties *pProperties)
+{
+   /* Sparse images are not yet supported. */
+   *pNumProperties = 0;
+}
+
+void
 tu_GetPhysicalDeviceSparseImageFormatProperties2(
    VkPhysicalDevice physicalDevice,
    const VkPhysicalDeviceSparseImageFormatInfo2 *pFormatInfo,
@@ -799,7 +964,7 @@ tu_GetPhysicalDeviceSparseImageFormatProperties2(
    *pPropertyCount = 0;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_GetPhysicalDeviceExternalBufferProperties(
    VkPhysicalDevice physicalDevice,
    const VkPhysicalDeviceExternalBufferInfo *pExternalBufferInfo,
diff --git a/lib/mesa/src/freedreno/vulkan/tu_image.c b/lib/mesa/src/freedreno/vulkan/tu_image.c
index 15a0649a2..657612d42 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_image.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_image.c
@@ -1,733 +1,266 @@
 /*
  * Copyright © 2016 Red Hat.
  * Copyright © 2016 Bas Nieuwenhuizen
- * SPDX-License-Identifier: MIT
  *
  * based in part on anv driver which is:
  * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_image.h"
-
-#include "fdl/fd6_format_table.h"
+#include "tu_private.h"
 
-#include "util/u_debug.h"
-#include "util/format/u_format.h"
+#include "util/debug.h"
+#include "util/u_atomic.h"
+#include "vk_format.h"
 #include "vk_util.h"
-#include "drm-uapi/drm_fourcc.h"
 
-#include "tu_android.h"
-#include "tu_cs.h"
-#include "tu_descriptor_set.h"
-#include "tu_device.h"
-#include "tu_formats.h"
-
-uint32_t
-tu6_plane_count(VkFormat format)
+static inline bool
+image_level_linear(struct tu_image *image, int level)
 {
-   switch (format) {
-   default:
-      return 1;
-   case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
-   case VK_FORMAT_D32_SFLOAT_S8_UINT:
-      return 2;
-   case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
-      return 3;
-   }
-}
-
-enum pipe_format
-tu6_plane_format(VkFormat format, uint32_t plane)
-{
-   switch (format) {
-   case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
-      return plane ? PIPE_FORMAT_R8G8_UNORM : PIPE_FORMAT_Y8_UNORM;
-   case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
-      return PIPE_FORMAT_R8_UNORM;
-   case VK_FORMAT_D32_SFLOAT_S8_UINT:
-      return plane ? PIPE_FORMAT_S8_UINT : PIPE_FORMAT_Z32_FLOAT;
-   default:
-      return tu_vk_format_to_pipe_format(format);
-   }
-}
-
-uint32_t
-tu6_plane_index(VkFormat format, VkImageAspectFlags aspect_mask)
-{
-   switch (aspect_mask) {
-   default:
-      assert(aspect_mask != VK_IMAGE_ASPECT_MEMORY_PLANE_3_BIT_EXT);
-      return 0;
-   case VK_IMAGE_ASPECT_PLANE_1_BIT:
-   case VK_IMAGE_ASPECT_MEMORY_PLANE_1_BIT_EXT:
-      return 1;
-   case VK_IMAGE_ASPECT_PLANE_2_BIT:
-   case VK_IMAGE_ASPECT_MEMORY_PLANE_2_BIT_EXT:
-      return 2;
-   case VK_IMAGE_ASPECT_STENCIL_BIT:
-      return format == VK_FORMAT_D32_SFLOAT_S8_UINT;
-   }
-}
-
-enum pipe_format
-tu_format_for_aspect(enum pipe_format format, VkImageAspectFlags aspect_mask)
-{
-   switch (format) {
-   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
-      if (aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT)
-         return PIPE_FORMAT_Z24_UNORM_S8_UINT_AS_R8G8B8A8;
-      if (aspect_mask & VK_IMAGE_ASPECT_STENCIL_BIT) {
-         if (aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT)
-            return PIPE_FORMAT_Z24_UNORM_S8_UINT;
-         else
-            return PIPE_FORMAT_X24S8_UINT;
-      } else {
-         return PIPE_FORMAT_Z24X8_UNORM;
-      }
-   case PIPE_FORMAT_Z24X8_UNORM:
-      if (aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT)
-         return PIPE_FORMAT_Z24_UNORM_S8_UINT_AS_R8G8B8A8;
-      return PIPE_FORMAT_Z24X8_UNORM;
-   default:
-      return format;
-   }
+   unsigned w = u_minify(image->extent.width, level);
+   return w < 16;
 }
 
-static bool
-tu_is_r8g8(enum pipe_format format)
+/* indexed by cpp: */
+static const struct
 {
-   return (util_format_get_blocksize(format) == 2) &&
-          (util_format_get_nr_components(format) == 2);
-}
-
-static bool
-tu_is_r8g8_compatible(enum pipe_format format)
-{
-   return (util_format_get_blocksize(format) == 2) &&
-          !util_format_is_depth_or_stencil(format);
-}
-
-void
-tu_cs_image_ref(struct tu_cs *cs, const struct fdl6_view *iview, uint32_t layer)
-{
-   tu_cs_emit(cs, iview->PITCH);
-   tu_cs_emit(cs, iview->layer_size >> 6);
-   tu_cs_emit_qw(cs, iview->base_addr + iview->layer_size * layer);
-}
-
-void
-tu_cs_image_stencil_ref(struct tu_cs *cs, const struct tu_image_view *iview, uint32_t layer)
-{
-   tu_cs_emit(cs, iview->stencil_PITCH);
-   tu_cs_emit(cs, iview->stencil_layer_size >> 6);
-   tu_cs_emit_qw(cs, iview->stencil_base_addr + iview->stencil_layer_size * layer);
-}
-
-void
-tu_cs_image_depth_ref(struct tu_cs *cs, const struct tu_image_view *iview, uint32_t layer)
-{
-   tu_cs_emit(cs, iview->depth_PITCH);
-   tu_cs_emit(cs, iview->depth_layer_size >> 6);
-   tu_cs_emit_qw(cs, iview->depth_base_addr + iview->depth_layer_size * layer);
-}
-
-void
-tu_cs_image_ref_2d(struct tu_cs *cs, const struct fdl6_view *iview, uint32_t layer, bool src)
-{
-   tu_cs_emit_qw(cs, iview->base_addr + iview->layer_size * layer);
-   /* SP_PS_2D_SRC_PITCH has shifted pitch field */
-   tu_cs_emit(cs, iview->PITCH << (src ? 9 : 0));
-}
-
-void
-tu_cs_image_flag_ref(struct tu_cs *cs, const struct fdl6_view *iview, uint32_t layer)
-{
-   tu_cs_emit_qw(cs, iview->ubwc_addr + iview->ubwc_layer_size * layer);
-   tu_cs_emit(cs, iview->FLAG_BUFFER_PITCH);
-}
+   unsigned pitchalign;
+   unsigned heightalign;
+} tile_alignment[] = {
+   [1] = { 128, 32 }, [2] = { 128, 16 }, [3] = { 128, 16 }, [4] = { 64, 16 },
+   [8] = { 64, 16 },  [12] = { 64, 16 }, [16] = { 64, 16 },
+};
 
 static void
-tu_image_view_init(struct tu_device *device,
-                   struct tu_image_view *iview,
-                   const VkImageViewCreateInfo *pCreateInfo,
-                   bool has_z24uint_s8uint)
+setup_slices(struct tu_image *image, const VkImageCreateInfo *pCreateInfo)
 {
-   TU_FROM_HANDLE(tu_image, image, pCreateInfo->image);
-   const VkImageSubresourceRange *range = &pCreateInfo->subresourceRange;
-   VkFormat vk_format = pCreateInfo->format;
-   VkImageAspectFlagBits aspect_mask = pCreateInfo->subresourceRange.aspectMask;
-
-   const struct VkSamplerYcbcrConversionInfo *ycbcr_conversion =
-      vk_find_struct_const(pCreateInfo->pNext, SAMPLER_YCBCR_CONVERSION_INFO);
-   const struct tu_sampler_ycbcr_conversion *conversion = ycbcr_conversion ?
-      tu_sampler_ycbcr_conversion_from_handle(ycbcr_conversion->conversion) : NULL;
-
-   vk_image_view_init(&device->vk, &iview->vk, false, pCreateInfo);
-
-   iview->image = image;
-
-   const struct fdl_layout *layouts[3];
-
-   layouts[0] = &image->layout[tu6_plane_index(image->vk.format, aspect_mask)];
-
-   enum pipe_format format;
-   if (aspect_mask != VK_IMAGE_ASPECT_COLOR_BIT)
-      format = tu6_plane_format(vk_format, tu6_plane_index(vk_format, aspect_mask));
-   else
-      format = tu_vk_format_to_pipe_format(vk_format);
-
-   if (image->vk.format == VK_FORMAT_G8_B8R8_2PLANE_420_UNORM &&
-       aspect_mask == VK_IMAGE_ASPECT_PLANE_0_BIT) {
-      if (vk_format == VK_FORMAT_R8_UNORM) {
-         /* The 0'th plane of this format has a different UBWC compression. */
-         format = PIPE_FORMAT_Y8_UNORM;
+   enum vk_format_layout layout =
+      vk_format_description(pCreateInfo->format)->layout;
+   uint32_t layer_size = 0;
+   uint32_t width = pCreateInfo->extent.width;
+   uint32_t height = pCreateInfo->extent.height;
+   uint32_t depth = pCreateInfo->extent.depth;
+   bool layer_first = pCreateInfo->imageType != VK_IMAGE_TYPE_3D;
+   uint32_t alignment = pCreateInfo->imageType == VK_IMAGE_TYPE_3D ? 4096 : 1;
+   uint32_t cpp = vk_format_get_blocksize(pCreateInfo->format);
+
+   uint32_t heightalign = tile_alignment[cpp].heightalign;
+
+   for (unsigned level = 0; level < pCreateInfo->mipLevels; level++) {
+      struct tu_image_level *slice = &image->levels[level];
+      bool linear_level = image_level_linear(image, level);
+      uint32_t aligned_height = height;
+      uint32_t blocks;
+      uint32_t pitchalign;
+
+      if (image->tile_mode && !linear_level) {
+         pitchalign = tile_alignment[cpp].pitchalign;
+         aligned_height = align(aligned_height, heightalign);
       } else {
-         /* If the user wants to reinterpret this plane, then they should've
-          * set MUTABLE_FORMAT_BIT which should disable UBWC and tiling.
+         pitchalign = 64;
+
+         /* The blits used for mem<->gmem work at a granularity of
+          * 32x32, which can cause faults due to over-fetch on the
+          * last level.  The simple solution is to over-allocate a
+          * bit the last level to ensure any over-fetch is harmless.
+          * The pitch is already sufficiently aligned, but height
+          * may not be:
           */
-         assert(!layouts[0]->ubwc);
+         if ((level + 1 == pCreateInfo->mipLevels))
+            aligned_height = align(aligned_height, 32);
       }
-   }
 
-   if (aspect_mask == VK_IMAGE_ASPECT_COLOR_BIT &&
-       (vk_format == VK_FORMAT_G8_B8R8_2PLANE_420_UNORM ||
-        vk_format == VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM)) {
-      layouts[1] = &image->layout[1];
-      layouts[2] = &image->layout[2];
-   }
-
-   struct fdl_view_args args = {};
-   args.iova = image->iova;
-   args.base_array_layer = range->baseArrayLayer;
-   args.base_miplevel = range->baseMipLevel;
-   args.layer_count = vk_image_subresource_layer_count(&image->vk, range);
-   args.level_count = vk_image_subresource_level_count(&image->vk, range);
-   args.min_lod_clamp = iview->vk.min_lod;
-   args.format = tu_format_for_aspect(format, aspect_mask);
-   vk_component_mapping_to_pipe_swizzle(pCreateInfo->components, args.swiz);
-   if (conversion) {
-      unsigned char conversion_swiz[4], create_swiz[4];
-      memcpy(create_swiz, args.swiz, sizeof(create_swiz));
-      vk_component_mapping_to_pipe_swizzle(conversion->components,
-                                           conversion_swiz);
-      util_format_compose_swizzles(create_swiz, conversion_swiz, args.swiz);
-   }
-
-   switch (pCreateInfo->viewType) {
-   case VK_IMAGE_VIEW_TYPE_1D:
-   case VK_IMAGE_VIEW_TYPE_1D_ARRAY:
-      args.type = FDL_VIEW_TYPE_1D;
-      break;
-   case VK_IMAGE_VIEW_TYPE_2D:
-   case VK_IMAGE_VIEW_TYPE_2D_ARRAY:
-      args.type = FDL_VIEW_TYPE_2D;
-      break;
-   case VK_IMAGE_VIEW_TYPE_CUBE:
-   case VK_IMAGE_VIEW_TYPE_CUBE_ARRAY:
-      args.type = FDL_VIEW_TYPE_CUBE;
-      break;
-   case VK_IMAGE_VIEW_TYPE_3D:
-      args.type = FDL_VIEW_TYPE_3D;
-      break;
-   default:
-      unreachable("unknown view type");
-   }
+      if (layout == VK_FORMAT_LAYOUT_ASTC)
+         slice->pitch = util_align_npot(
+            width,
+            pitchalign * vk_format_get_blockwidth(pCreateInfo->format));
+      else
+         slice->pitch = align(width, pitchalign);
 
-   STATIC_ASSERT((unsigned)VK_CHROMA_LOCATION_COSITED_EVEN == (unsigned)FDL_CHROMA_LOCATION_COSITED_EVEN);
-   STATIC_ASSERT((unsigned)VK_CHROMA_LOCATION_MIDPOINT == (unsigned)FDL_CHROMA_LOCATION_MIDPOINT);
-   if (conversion) {
-      args.chroma_offsets[0] = (enum fdl_chroma_location) conversion->chroma_offsets[0];
-      args.chroma_offsets[1] = (enum fdl_chroma_location) conversion->chroma_offsets[1];
-   }
+      slice->offset = layer_size;
+      blocks = vk_format_get_block_count(pCreateInfo->format, slice->pitch,
+                                         aligned_height);
 
-   fdl6_view_init(&iview->view, layouts, &args, has_z24uint_s8uint);
+      /* 1d array and 2d array textures must all have the same layer size
+       * for each miplevel on a3xx. 3d textures can have different layer
+       * sizes for high levels, but the hw auto-sizer is buggy (or at least
+       * different than what this code does), so as soon as the layer size
+       * range gets into range, we stop reducing it.
+       */
+      if (pCreateInfo->imageType == VK_IMAGE_TYPE_3D &&
+          (level == 1 ||
+           (level > 1 && image->levels[level - 1].size > 0xf000)))
+         slice->size = align(blocks * cpp, alignment);
+      else if (level == 0 || layer_first || alignment == 1)
+         slice->size = align(blocks * cpp, alignment);
+      else
+         slice->size = image->levels[level - 1].size;
 
-   if (image->vk.format == VK_FORMAT_D32_SFLOAT_S8_UINT) {
-      struct fdl_layout *layout = &image->layout[0];
-      iview->depth_base_addr = image->iova +
-         fdl_surface_offset(layout, range->baseMipLevel, range->baseArrayLayer);
-      iview->depth_layer_size = fdl_layer_stride(layout, range->baseMipLevel);
-      iview->depth_PITCH = A6XX_RB_DEPTH_BUFFER_PITCH(fdl_pitch(layout, range->baseMipLevel)).value;
+      layer_size += slice->size * depth;
 
-      layout = &image->layout[1];
-      iview->stencil_base_addr = image->iova +
-         fdl_surface_offset(layout, range->baseMipLevel, range->baseArrayLayer);
-      iview->stencil_layer_size = fdl_layer_stride(layout, range->baseMipLevel);
-      iview->stencil_PITCH = A6XX_RB_STENCIL_BUFFER_PITCH(fdl_pitch(layout, range->baseMipLevel)).value;
+      width = u_minify(width, 1);
+      height = u_minify(height, 1);
+      depth = u_minify(depth, 1);
    }
-}
 
-bool
-tiling_possible(VkFormat format)
-{
-   if (format == VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM ||
-       format == VK_FORMAT_G8B8G8R8_422_UNORM ||
-       format == VK_FORMAT_B8G8R8G8_422_UNORM)
-      return false;
-
-   return true;
+   image->layer_size = layer_size;
 }
 
-/* Checks if we should advertise UBWC support for the given usage.
- *
- * Used by both vkCreateImage and vkGetPhysicalDeviceFormatProperties2, so the
- * logical tu_device may be NULL.
- */
-bool
-ubwc_possible(struct tu_device *device,
-              VkFormat format,
-              VkImageType type,
-              VkImageUsageFlags usage,
-              VkImageUsageFlags stencil_usage,
-              const struct fd_dev_info *info,
-              VkSampleCountFlagBits samples,
-              bool use_z24uint_s8uint)
+VkResult
+tu_image_create(VkDevice _device,
+                const struct tu_image_create_info *create_info,
+                const VkAllocationCallbacks *alloc,
+                VkImage *pImage)
 {
-   /* no UBWC with compressed formats, E5B9G9R9, S8_UINT
-    * (S8_UINT because separate stencil doesn't have UBWC-enable bit)
-    */
-   if (vk_format_is_compressed(format) ||
-       format == VK_FORMAT_E5B9G9R9_UFLOAT_PACK32 ||
-       format == VK_FORMAT_S8_UINT)
-      return false;
-
-   /* In copy_format, we treat snorm as unorm to avoid clamping.  But snorm
-    * and unorm are UBWC incompatible for special values such as all 0's or
-    * all 1's.  Disable UBWC for snorm.
-    */
-   if (vk_format_is_snorm(format))
-      return false;
-
-   if (!info->a6xx.has_8bpp_ubwc &&
-       (format == VK_FORMAT_R8_UNORM ||
-        format == VK_FORMAT_R8_SNORM ||
-        format == VK_FORMAT_R8_UINT ||
-        format == VK_FORMAT_R8_SINT ||
-        format == VK_FORMAT_R8_SRGB))
-      return false;
-
-   if (type == VK_IMAGE_TYPE_3D) {
-      if (device) {
-         perf_debug(device,
-                    "Disabling UBWC for %s 3D image, but it should be "
-                    "possible to support.",
-                    util_format_name(vk_format_to_pipe_format(format)));
-      }
-      return false;
-   }
-
-   /* Disable UBWC for storage images.
-    *
-    * The closed GL driver skips UBWC for storage images (and additionally
-    * uses linear for writeonly images).  We seem to have image tiling working
-    * in freedreno in general, so turnip matches that.  freedreno also enables
-    * UBWC on images, but it's not really tested due to the lack of
-    * UBWC-enabled mipmaps in freedreno currently.  Just match the closed GL
-    * behavior of no UBWC.
-   */
-   if ((usage | stencil_usage) & VK_IMAGE_USAGE_STORAGE_BIT) {
-      if (device) {
-         perf_debug(device,
-                    "Disabling UBWC for %s storage image, but should be "
-                    "possible to support",
-                    util_format_name(vk_format_to_pipe_format(format)));
-      }
-      return false;
+   TU_FROM_HANDLE(tu_device, device, _device);
+   const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
+   struct tu_image *image = NULL;
+   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO);
+
+   tu_assert(pCreateInfo->mipLevels > 0);
+   tu_assert(pCreateInfo->arrayLayers > 0);
+   tu_assert(pCreateInfo->samples > 0);
+   tu_assert(pCreateInfo->extent.width > 0);
+   tu_assert(pCreateInfo->extent.height > 0);
+   tu_assert(pCreateInfo->extent.depth > 0);
+
+   image = vk_zalloc2(&device->alloc, alloc, sizeof(*image), 8,
+                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (!image)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   image->type = pCreateInfo->imageType;
+
+   image->vk_format = pCreateInfo->format;
+   image->tiling = pCreateInfo->tiling;
+   image->usage = pCreateInfo->usage;
+   image->flags = pCreateInfo->flags;
+   image->extent = pCreateInfo->extent;
+   image->level_count = pCreateInfo->mipLevels;
+   image->layer_count = pCreateInfo->arrayLayers;
+
+   image->exclusive = pCreateInfo->sharingMode == VK_SHARING_MODE_EXCLUSIVE;
+   if (pCreateInfo->sharingMode == VK_SHARING_MODE_CONCURRENT) {
+      for (uint32_t i = 0; i < pCreateInfo->queueFamilyIndexCount; ++i)
+         if (pCreateInfo->pQueueFamilyIndices[i] ==
+             VK_QUEUE_FAMILY_EXTERNAL)
+            image->queue_family_mask |= (1u << TU_MAX_QUEUE_FAMILIES) - 1u;
+         else
+            image->queue_family_mask |=
+               1u << pCreateInfo->pQueueFamilyIndices[i];
    }
 
-   /* Disable UBWC for D24S8 on A630 in some cases
-    *
-    * VK_IMAGE_ASPECT_STENCIL_BIT image view requires to be able to sample
-    * from the stencil component as UINT, however no format allows this
-    * on a630 (the special FMT6_Z24_UINT_S8_UINT format is missing)
-    *
-    * It must be sampled as FMT6_8_8_8_8_UINT, which is not UBWC-compatible
-    *
-    * If we wish to get the border colors correct without knowing the format
-    * when creating the sampler, we also have to use the A630 workaround.
-    *
-    * Additionally, the special AS_R8G8B8A8 format is broken without UBWC,
-    * so we have to fallback to 8_8_8_8_UNORM when UBWC is disabled
-    */
-   if (!use_z24uint_s8uint &&
-       format == VK_FORMAT_D24_UNORM_S8_UINT &&
-       (stencil_usage & (VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT)))
-      return false;
-
-   /* This meant to disable UBWC for MSAA z24s8, but accidentally disables it
-    * for all MSAA.  https://gitlab.freedesktop.org/mesa/mesa/-/issues/7438
-    */
-   if (!info->a6xx.has_z24uint_s8uint && samples > VK_SAMPLE_COUNT_1_BIT) {
-      if (device) {
-         perf_debug(device,
-                    "Disabling UBWC for %d-sample %s image, but it should be "
-                    "possible to support",
-                    samples,
-                    util_format_name(vk_format_to_pipe_format(format)));
-      }
-      return false;
-   }
+   image->shareable =
+      vk_find_struct_const(pCreateInfo->pNext,
+                           EXTERNAL_MEMORY_IMAGE_CREATE_INFO) != NULL;
 
-   return true;
-}
+   image->tile_mode = pCreateInfo->tiling == VK_IMAGE_TILING_OPTIMAL ? 3 : 0;
+   setup_slices(image, pCreateInfo);
 
-/* R8G8 have a different block width/height and height alignment from other
- * formats that would normally be compatible (like R16), and so if we are
- * trying to, for example, sample R16 as R8G8 we need to demote to linear.
- */
-static bool
-format_list_reinterprets_r8g8_r16(enum pipe_format format, const VkImageFormatListCreateInfo *fmt_list)
-{
-   /* Check if it's actually a 2-cpp color format. */
-   if (!tu_is_r8g8_compatible(format))
-      return false;
-
-   /* If there's no format list, then the app may reinterpret to any compatible
-    * format.
-    */
-   if (!fmt_list || !fmt_list->viewFormatCount)
-      return true;
-
-   bool has_r8g8 = false;
-   bool has_non_r8g8 = false;
-   for (uint32_t i = 0; i < fmt_list->viewFormatCount; i++) {
-      enum pipe_format format =
-         tu_vk_format_to_pipe_format(fmt_list->pViewFormats[i]);
-      if (tu_is_r8g8(format))
-         has_r8g8 = true;
-      else
-         has_non_r8g8 = true;
-   }
-   return has_r8g8 && has_non_r8g8;
-}
+   image->size = image->layer_size * pCreateInfo->arrayLayers;
+   *pImage = tu_image_to_handle(image);
 
-static bool
-format_list_has_swaps(const VkImageFormatListCreateInfo *fmt_list)
-{
-   /* If there's no format list, then the app may reinterpret to any compatible
-    * format, and presumably one would have the swap set.
-    */
-   if (!fmt_list || !fmt_list->viewFormatCount)
-      return true;
-
-   for (uint32_t i = 0; i < fmt_list->viewFormatCount; i++) {
-      enum pipe_format format =
-         tu_vk_format_to_pipe_format(fmt_list->pViewFormats[i]);
-
-      if (tu6_format_texture(format, TILE6_LINEAR).swap)
-         return true;
-   }
-   return false;
+   return VK_SUCCESS;
 }
 
-static VkResult
-tu_image_init(struct tu_device *device, struct tu_image *image,
-              const VkImageCreateInfo *pCreateInfo, uint64_t modifier,
-              const VkSubresourceLayout *plane_layouts)
+void
+tu_image_view_init(struct tu_image_view *iview,
+                   struct tu_device *device,
+                   const VkImageViewCreateInfo *pCreateInfo)
 {
-   vk_image_init(&device->vk, &image->vk, pCreateInfo);
-   image->vk.drm_format_mod = modifier;
-
-   enum a6xx_tile_mode tile_mode = TILE6_3;
-   bool ubwc_enabled = true;
-
-   /* use linear tiling if requested */
-   if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR || modifier == DRM_FORMAT_MOD_LINEAR) {
-      tile_mode = TILE6_LINEAR;
-      ubwc_enabled = false;
-   }
-
-   /* Force linear tiling for formats with "fake" optimalTilingFeatures */
-   if (!tiling_possible(image->vk.format)) {
-      tile_mode = TILE6_LINEAR;
-      ubwc_enabled = false;
-   }
-
-   /* No sense in tiling a 1D image, you'd just waste space and cache locality. */
-   if (pCreateInfo->imageType == VK_IMAGE_TYPE_1D) {
-      tile_mode = TILE6_LINEAR;
-      ubwc_enabled = false;
-   }
-
-   enum pipe_format format =
-      tu_vk_format_to_pipe_format(image->vk.format);
-   /* Whether a view of the image with an R8G8 format could be made. */
-   bool has_r8g8 = tu_is_r8g8(format);
-
-   if (ubwc_enabled &&
-       !ubwc_possible(device, image->vk.format, pCreateInfo->imageType,
-                      pCreateInfo->usage, image->vk.stencil_usage,
-                      device->physical_device->info, pCreateInfo->samples,
-                      device->use_z24uint_s8uint))
-      ubwc_enabled = false;
-
-   /* Mutable images can be reinterpreted as any other compatible format.
-    * This is a problem with UBWC (compression for different formats is different),
-    * but also tiling ("swap" affects how tiled formats are stored in memory)
-    * Depth and stencil formats cannot be reintepreted as another format, and
-    * cannot be linear with sysmem rendering, so don't fall back for those.
-    *
-    * TODO:
-    * - if the fmt_list contains only formats which are swapped, but compatible
-    *   with each other (B8G8R8A8_UNORM and B8G8R8A8_UINT for example), then
-    *   tiling is still possible
-    * - figure out which UBWC compressions are compatible to keep it enabled
-    */
-   if ((pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) &&
-       !vk_format_is_depth_or_stencil(image->vk.format)) {
-      const VkImageFormatListCreateInfo *fmt_list =
-         vk_find_struct_const(pCreateInfo->pNext, IMAGE_FORMAT_LIST_CREATE_INFO);
-      if (!tu6_mutable_format_list_ubwc_compatible(fmt_list)) {
-         if (ubwc_enabled) {
-            if (fmt_list && fmt_list->viewFormatCount == 2) {
-               perf_debug(
-                  device,
-                  "Disabling UBWC on %dx%d %s resource due to mutable formats "
-                  "(fmt list %s, %s)",
-                  image->vk.extent.width, image->vk.extent.height,
-                  util_format_name(vk_format_to_pipe_format(image->vk.format)),
-                  util_format_name(vk_format_to_pipe_format(fmt_list->pViewFormats[0])),
-                  util_format_name(vk_format_to_pipe_format(fmt_list->pViewFormats[1])));
-            } else {
-               perf_debug(
-                  device,
-                  "Disabling UBWC on %dx%d %s resource due to mutable formats "
-                  "(fmt list %s)",
-                  image->vk.extent.width, image->vk.extent.height,
-                  util_format_name(vk_format_to_pipe_format(image->vk.format)),
-                  fmt_list ? "present" : "missing");
-            }
-            ubwc_enabled = false;
-         }
-
-         if (format_list_reinterprets_r8g8_r16(format, fmt_list) ||
-            format_list_has_swaps(fmt_list)) {
-            tile_mode = TILE6_LINEAR;
-         }
-      }
-   }
+   TU_FROM_HANDLE(tu_image, image, pCreateInfo->image);
+   const VkImageSubresourceRange *range = &pCreateInfo->subresourceRange;
 
-   /* expect UBWC enabled if we asked for it */
-   if (modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED)
-      assert(ubwc_enabled);
-   else if (device->physical_device->instance->debug_flags & TU_DEBUG_NOUBWC)
-      ubwc_enabled = false;
-
-   /* Non-UBWC tiled R8G8 is probably buggy since media formats are always
-    * either linear or UBWC. There is no simple test to reproduce the bug.
-    * However it was observed in the wild leading to an unrecoverable hang
-    * on a650/a660.
-    */
-   if (has_r8g8 && tile_mode == TILE6_3 && !ubwc_enabled) {
-      tile_mode = TILE6_LINEAR;
+   switch (image->type) {
+   case VK_IMAGE_TYPE_1D:
+   case VK_IMAGE_TYPE_2D:
+      assert(range->baseArrayLayer + tu_get_layerCount(image, range) <=
+             image->layer_count);
+      break;
+   case VK_IMAGE_TYPE_3D:
+      assert(range->baseArrayLayer + tu_get_layerCount(image, range) <=
+             tu_minify(image->extent.depth, range->baseMipLevel));
+      break;
+   default:
+      unreachable("bad VkImageType");
    }
 
-   for (uint32_t i = 0; i < tu6_plane_count(image->vk.format); i++) {
-      struct fdl_layout *layout = &image->layout[i];
-      enum pipe_format format = tu6_plane_format(image->vk.format, i);
-      uint32_t width0 = pCreateInfo->extent.width;
-      uint32_t height0 = pCreateInfo->extent.height;
-
-      if (i > 0) {
-         switch (image->vk.format) {
-         case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
-         case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
-            /* half width/height on chroma planes */
-            width0 = (width0 + 1) >> 1;
-            height0 = (height0 + 1) >> 1;
-            break;
-         case VK_FORMAT_D32_SFLOAT_S8_UINT:
-            /* no UBWC for separate stencil */
-            ubwc_enabled = false;
-            break;
-         default:
-            break;
-         }
-      }
-
-      struct fdl_explicit_layout plane_layout;
-
-      if (plane_layouts) {
-         /* only expect simple 2D images for now */
-         if (pCreateInfo->mipLevels != 1 ||
-            pCreateInfo->arrayLayers != 1 ||
-            pCreateInfo->extent.depth != 1)
-            return vk_error(device, VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT);
-
-         plane_layout.offset = plane_layouts[i].offset;
-         plane_layout.pitch = plane_layouts[i].rowPitch;
-         /* note: use plane_layouts[0].arrayPitch to support array formats */
-      }
-
-      layout->tile_mode = tile_mode;
-      layout->ubwc = ubwc_enabled;
-
-      if (!fdl6_layout(layout, format,
-                       pCreateInfo->samples,
-                       width0, height0,
-                       pCreateInfo->extent.depth,
-                       pCreateInfo->mipLevels,
-                       pCreateInfo->arrayLayers,
-                       pCreateInfo->imageType == VK_IMAGE_TYPE_3D,
-                       plane_layouts ? &plane_layout : NULL)) {
-         assert(plane_layouts); /* can only fail with explicit layout */
-         return vk_error(device, VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT);
-      }
-
-      if (device->instance->debug_flags & TU_DEBUG_LAYOUT)
-         fdl_dump_layout(layout);
-
-      /* fdl6_layout can't take explicit offset without explicit pitch
-       * add offset manually for extra layouts for planes
-       */
-      if (!plane_layouts && i > 0) {
-         uint32_t offset = ALIGN_POT(image->total_size, 4096);
-         for (int i = 0; i < pCreateInfo->mipLevels; i++) {
-            layout->slices[i].offset += offset;
-            layout->ubwc_slices[i].offset += offset;
-         }
-         layout->size += offset;
-      }
-
-      image->total_size = MAX2(image->total_size, layout->size);
+   iview->image = image;
+   iview->type = pCreateInfo->viewType;
+   iview->vk_format = pCreateInfo->format;
+   iview->aspect_mask = pCreateInfo->subresourceRange.aspectMask;
+
+   if (iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) {
+      iview->vk_format = vk_format_stencil_only(iview->vk_format);
+   } else if (iview->aspect_mask == VK_IMAGE_ASPECT_DEPTH_BIT) {
+      iview->vk_format = vk_format_depth_only(iview->vk_format);
    }
 
-   const struct util_format_description *desc = util_format_description(image->layout[0].format);
-   if (util_format_has_depth(desc) && !(device->instance->debug_flags & TU_DEBUG_NOLRZ))
-   {
-      /* Depth plane is the first one */
-      struct fdl_layout *layout = &image->layout[0];
-      unsigned width = layout->width0;
-      unsigned height = layout->height0;
-
-      /* LRZ buffer is super-sampled */
-      switch (layout->nr_samples) {
-      case 4:
-         width *= 2;
-         FALLTHROUGH;
-      case 2:
-         height *= 2;
-         break;
-      default:
-         break;
-      }
-
-      unsigned lrz_pitch  = align(DIV_ROUND_UP(width, 8), 32);
-      unsigned lrz_height = align(DIV_ROUND_UP(height, 8), 16);
-
-      image->lrz_height = lrz_height;
-      image->lrz_pitch = lrz_pitch;
-      image->lrz_offset = image->total_size;
-      unsigned lrz_size = lrz_pitch * lrz_height * 2;
-      image->total_size += lrz_size;
-
-      unsigned nblocksx = DIV_ROUND_UP(DIV_ROUND_UP(width, 8), 16);
-      unsigned nblocksy = DIV_ROUND_UP(DIV_ROUND_UP(height, 8), 4);
-
-      /* Fast-clear buffer is 1bit/block */
-      image->lrz_fc_size = DIV_ROUND_UP(nblocksx * nblocksy, 8);
-
-      /* Fast-clear buffer cannot be larger than 512 bytes (HW limitation) */
-      bool has_lrz_fc = image->lrz_fc_size <= 512 &&
-         device->physical_device->info->a6xx.enable_lrz_fast_clear &&
-         !unlikely(device->physical_device->instance->debug_flags & TU_DEBUG_NOLRZFC);
+   // should we minify?
+   iview->extent = image->extent;
 
-      if (has_lrz_fc || device->physical_device->info->a6xx.has_lrz_dir_tracking) {
-         image->lrz_fc_offset = image->total_size;
-         image->total_size += 512;
-
-         if (device->physical_device->info->a6xx.has_lrz_dir_tracking) {
-            /* Direction tracking uses 1 byte */
-            image->total_size += 1;
-            /* GRAS_LRZ_DEPTH_VIEW needs 5 bytes: 4 for view data and 1 for padding */
-            image->total_size += 5;
-         }
-      }
-
-      if (!has_lrz_fc) {
-         image->lrz_fc_size = 0;
-      }
-   } else {
-      image->lrz_height = 0;
-   }
+   iview->base_layer = range->baseArrayLayer;
+   iview->layer_count = tu_get_layerCount(image, range);
+   iview->base_mip = range->baseMipLevel;
+   iview->level_count = tu_get_levelCount(image, range);
+}
 
-   return VK_SUCCESS;
+unsigned
+tu_image_queue_family_mask(const struct tu_image *image,
+                           uint32_t family,
+                           uint32_t queue_family)
+{
+   if (!image->exclusive)
+      return image->queue_family_mask;
+   if (family == VK_QUEUE_FAMILY_EXTERNAL)
+      return (1u << TU_MAX_QUEUE_FAMILIES) - 1u;
+   if (family == VK_QUEUE_FAMILY_IGNORED)
+      return 1u << queue_family;
+   return 1u << family;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_CreateImage(VkDevice _device,
+VkResult
+tu_CreateImage(VkDevice device,
                const VkImageCreateInfo *pCreateInfo,
-               const VkAllocationCallbacks *alloc,
+               const VkAllocationCallbacks *pAllocator,
                VkImage *pImage)
 {
-   uint64_t modifier = DRM_FORMAT_MOD_INVALID;
-   const VkSubresourceLayout *plane_layouts = NULL;
-
-   TU_FROM_HANDLE(tu_device, device, _device);
-   struct tu_image *image =
-      vk_object_zalloc(&device->vk, alloc, sizeof(*image), VK_OBJECT_TYPE_IMAGE);
-
-   if (!image)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
-
-   if (pCreateInfo->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT) {
-      const VkImageDrmFormatModifierListCreateInfoEXT *mod_info =
-         vk_find_struct_const(pCreateInfo->pNext,
-                              IMAGE_DRM_FORMAT_MODIFIER_LIST_CREATE_INFO_EXT);
-      const VkImageDrmFormatModifierExplicitCreateInfoEXT *drm_explicit_info =
-         vk_find_struct_const(pCreateInfo->pNext,
-                              IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT);
-
-      assert(mod_info || drm_explicit_info);
-
-      if (mod_info) {
-         modifier = DRM_FORMAT_MOD_LINEAR;
-         for (unsigned i = 0; i < mod_info->drmFormatModifierCount; i++) {
-            if (mod_info->pDrmFormatModifiers[i] == DRM_FORMAT_MOD_QCOM_COMPRESSED)
-               modifier = DRM_FORMAT_MOD_QCOM_COMPRESSED;
-         }
-      } else {
-         modifier = drm_explicit_info->drmFormatModifier;
-         assert(modifier == DRM_FORMAT_MOD_LINEAR ||
-                modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED);
-         plane_layouts = drm_explicit_info->pPlaneLayouts;
-      }
-   } else {
-      const struct wsi_image_create_info *wsi_info =
-         vk_find_struct_const(pCreateInfo->pNext, WSI_IMAGE_CREATE_INFO_MESA);
-      if (wsi_info && wsi_info->scanout)
-         modifier = DRM_FORMAT_MOD_LINEAR;
-   }
-
 #ifdef ANDROID
    const VkNativeBufferANDROID *gralloc_info =
       vk_find_struct_const(pCreateInfo->pNext, NATIVE_BUFFER_ANDROID);
-   int dma_buf;
-   if (gralloc_info) {
-      VkResult result = tu_gralloc_info(device, gralloc_info, &dma_buf, &modifier);
-      if (result != VK_SUCCESS)
-         return result;
-   }
-#endif
 
-   VkResult result = tu_image_init(device, image, pCreateInfo, modifier,
-                                   plane_layouts);
-   if (result != VK_SUCCESS) {
-      vk_object_free(&device->vk, alloc, image);
-      return result;
-   }
-
-   *pImage = tu_image_to_handle(image);
-
-#ifdef ANDROID
    if (gralloc_info)
-      return tu_import_memory_from_gralloc_handle(_device, dma_buf, alloc,
-                                                  *pImage);
+      return tu_image_from_gralloc(device, pCreateInfo, gralloc_info,
+                                   pAllocator, pImage);
 #endif
-   return VK_SUCCESS;
+
+   return tu_image_create(device,
+                          &(struct tu_image_create_info) {
+                             .vk_info = pCreateInfo,
+                             .scanout = false,
+                          },
+                          pAllocator, pImage);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyImage(VkDevice _device,
                 VkImage _image,
                 const VkAllocationCallbacks *pAllocator)
@@ -738,87 +271,13 @@ tu_DestroyImage(VkDevice _device,
    if (!image)
       return;
 
-#ifdef ANDROID
    if (image->owned_memory != VK_NULL_HANDLE)
       tu_FreeMemory(_device, image->owned_memory, pAllocator);
-#endif
-
-   vk_object_free(&device->vk, pAllocator, image);
-}
-
-static void
-tu_get_image_memory_requirements(struct tu_image *image,
-                                 VkMemoryRequirements2 *pMemoryRequirements)
-{
-   pMemoryRequirements->memoryRequirements = (VkMemoryRequirements) {
-      .memoryTypeBits = 1,
-      .alignment = image->layout[0].base_align,
-      .size = image->total_size
-   };
-
-   vk_foreach_struct(ext, pMemoryRequirements->pNext) {
-      switch (ext->sType) {
-      case VK_STRUCTURE_TYPE_MEMORY_DEDICATED_REQUIREMENTS: {
-         VkMemoryDedicatedRequirements *req =
-            (VkMemoryDedicatedRequirements *) ext;
-         req->requiresDedicatedAllocation =
-            image->vk.external_handle_types != 0;
-         req->prefersDedicatedAllocation = req->requiresDedicatedAllocation;
-         break;
-      }
-      default:
-         break;
-      }
-   }
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_GetImageMemoryRequirements2(VkDevice device,
-                               const VkImageMemoryRequirementsInfo2 *pInfo,
-                               VkMemoryRequirements2 *pMemoryRequirements)
-{
-   TU_FROM_HANDLE(tu_image, image, pInfo->image);
-
-   tu_get_image_memory_requirements(image, pMemoryRequirements);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_GetImageSparseMemoryRequirements2(
-   VkDevice device,
-   const VkImageSparseMemoryRequirementsInfo2 *pInfo,
-   uint32_t *pSparseMemoryRequirementCount,
-   VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
-{
-   tu_stub();
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_GetDeviceImageMemoryRequirements(
-   VkDevice _device,
-   const VkDeviceImageMemoryRequirements *pInfo,
-   VkMemoryRequirements2 *pMemoryRequirements)
-{
-   TU_FROM_HANDLE(tu_device, device, _device);
-
-   struct tu_image image = {0};
 
-   tu_image_init(device, &image, pInfo->pCreateInfo, DRM_FORMAT_MOD_INVALID,
-                 NULL);
-
-   tu_get_image_memory_requirements(&image, pMemoryRequirements);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_GetDeviceImageSparseMemoryRequirements(
-    VkDevice device,
-    const VkDeviceImageMemoryRequirements *pInfo,
-    uint32_t *pSparseMemoryRequirementCount,
-    VkSparseImageMemoryRequirements2 *pSparseMemoryRequirements)
-{
-   tu_stub();
+   vk_free2(&device->alloc, pAllocator, image);
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_GetImageSubresourceLayout(VkDevice _device,
                              VkImage _image,
                              const VkImageSubresource *pSubresource,
@@ -826,26 +285,19 @@ tu_GetImageSubresourceLayout(VkDevice _device,
 {
    TU_FROM_HANDLE(tu_image, image, _image);
 
-   struct fdl_layout *layout =
-      &image->layout[tu6_plane_index(image->vk.format, pSubresource->aspectMask)];
-   const struct fdl_slice *slice = layout->slices + pSubresource->mipLevel;
-
-   pLayout->offset =
-      fdl_surface_offset(layout, pSubresource->mipLevel, pSubresource->arrayLayer);
-   pLayout->rowPitch = fdl_pitch(layout, pSubresource->mipLevel);
-   pLayout->arrayPitch = fdl_layer_stride(layout, pSubresource->mipLevel);
-   pLayout->depthPitch = slice->size0;
-   pLayout->size = pLayout->depthPitch * layout->depth0;
-
-   if (fdl_ubwc_enabled(layout, pSubresource->mipLevel)) {
-      /* UBWC starts at offset 0 */
-      pLayout->offset = 0;
-      /* UBWC scanout won't match what the kernel wants if we have levels/layers */
-      assert(image->vk.mip_levels == 1 && image->vk.array_layers == 1);
-   }
+   const uint32_t layer_offset = image->layer_size * pSubresource->arrayLayer;
+   const struct tu_image_level *level =
+      image->levels + pSubresource->mipLevel;
+
+   pLayout->offset = layer_offset + level->offset;
+   pLayout->size = level->size;
+   pLayout->rowPitch =
+      level->pitch * vk_format_get_blocksize(image->vk_format);
+   pLayout->arrayPitch = image->layer_size;
+   pLayout->depthPitch = level->size;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateImageView(VkDevice _device,
                    const VkImageViewCreateInfo *pCreateInfo,
                    const VkAllocationCallbacks *pAllocator,
@@ -854,19 +306,19 @@ tu_CreateImageView(VkDevice _device,
    TU_FROM_HANDLE(tu_device, device, _device);
    struct tu_image_view *view;
 
-   view = vk_object_alloc(&device->vk, pAllocator, sizeof(*view),
-                          VK_OBJECT_TYPE_IMAGE_VIEW);
+   view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8,
+                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (view == NULL)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   tu_image_view_init(device, view, pCreateInfo, device->use_z24uint_s8uint);
+   tu_image_view_init(view, device, pCreateInfo);
 
    *pView = tu_image_view_to_handle(view);
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyImageView(VkDevice _device,
                     VkImageView _iview,
                     const VkAllocationCallbacks *pAllocator)
@@ -876,8 +328,7 @@ tu_DestroyImageView(VkDevice _device,
 
    if (!iview)
       return;
-
-   vk_object_free(&device->vk, pAllocator, iview);
+   vk_free2(&device->alloc, pAllocator, iview);
 }
 
 void
@@ -887,19 +338,13 @@ tu_buffer_view_init(struct tu_buffer_view *view,
 {
    TU_FROM_HANDLE(tu_buffer, buffer, pCreateInfo->buffer);
 
-   view->buffer = buffer;
-
-   uint32_t range = vk_buffer_range(&buffer->vk, pCreateInfo->offset,
-         pCreateInfo->range);
-   uint8_t swiz[4] = { PIPE_SWIZZLE_X, PIPE_SWIZZLE_Y, PIPE_SWIZZLE_Z,
-                       PIPE_SWIZZLE_W };
-
-   fdl6_buffer_view_init(
-      view->descriptor, tu_vk_format_to_pipe_format(pCreateInfo->format),
-      swiz, buffer->iova + pCreateInfo->offset, range);
+   view->range = pCreateInfo->range == VK_WHOLE_SIZE
+                    ? buffer->size - pCreateInfo->offset
+                    : pCreateInfo->range;
+   view->vk_format = pCreateInfo->format;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateBufferView(VkDevice _device,
                     const VkBufferViewCreateInfo *pCreateInfo,
                     const VkAllocationCallbacks *pAllocator,
@@ -908,10 +353,10 @@ tu_CreateBufferView(VkDevice _device,
    TU_FROM_HANDLE(tu_device, device, _device);
    struct tu_buffer_view *view;
 
-   view = vk_object_alloc(&device->vk, pAllocator, sizeof(*view),
-                          VK_OBJECT_TYPE_BUFFER_VIEW);
+   view = vk_alloc2(&device->alloc, pAllocator, sizeof(*view), 8,
+                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (!view)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
    tu_buffer_view_init(view, device, pCreateInfo);
 
@@ -920,7 +365,7 @@ tu_CreateBufferView(VkDevice _device,
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyBufferView(VkDevice _device,
                      VkBufferView bufferView,
                      const VkAllocationCallbacks *pAllocator)
@@ -931,5 +376,5 @@ tu_DestroyBufferView(VkDevice _device,
    if (!view)
       return;
 
-   vk_object_free(&device->vk, pAllocator, view);
+   vk_free2(&device->alloc, pAllocator, view);
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_pass.c b/lib/mesa/src/freedreno/vulkan/tu_pass.c
index 84c1c3061..e3d9f23df 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_pass.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_pass.c
@@ -1,796 +1,245 @@
 /*
  * Copyright © 2016 Red Hat.
  * Copyright © 2016 Bas Nieuwenhuizen
- * SPDX-License-Identifier: MIT
  *
  * based in part on anv driver which is:
  * Copyright © 2015 Intel Corporation
- */
-
-#include "tu_pass.h"
-
-#include "vk_util.h"
-
-#include "tu_cmd_buffer.h"
-#include "tu_device.h"
-#include "tu_image.h"
-
-/* Return true if we have to fallback to sysmem rendering because the
- * dependency can't be satisfied with tiled rendering.
- */
-
-static bool
-dep_invalid_for_gmem(const VkSubpassDependency2 *dep,
-                     VkPipelineStageFlags2 src_stage_mask,
-                     VkPipelineStageFlags2 dst_stage_mask)
-{
-   /* External dependencies don't matter here. */
-   if (dep->srcSubpass == VK_SUBPASS_EXTERNAL ||
-       dep->dstSubpass == VK_SUBPASS_EXTERNAL)
-      return false;
-
-   /* We can conceptually break down the process of rewriting a sysmem
-    * renderpass into a gmem one into two parts:
-    *
-    * 1. Split each draw and multisample resolve into N copies, one for each
-    * bin. (If hardware binning, add one more copy where the FS is disabled
-    * for the binning pass). This is always allowed because the vertex stage
-    * is allowed to run an arbitrary number of times and there are no extra
-    * ordering constraints within a draw.
-    * 2. Take the last copy of the second-to-last draw and slide it down to
-    * before the last copy of the last draw. Repeat for each earlier draw
-    * until the draw pass for the last bin is complete, then repeat for each
-    * earlier bin until we finish with the first bin.
-    *
-    * During this rearranging process, we can't slide draws past each other in
-    * a way that breaks the subpass dependencies. For each draw, we must slide
-    * it past (copies of) the rest of the draws in the renderpass. We can
-    * slide a draw past another if there isn't a dependency between them, or
-    * if the dependenc(ies) are dependencies between framebuffer-space stages
-    * only with the BY_REGION bit set. Note that this includes
-    * self-dependencies, since these may result in pipeline barriers that also
-    * break the rearranging process.
-    */
-
-   /* This is straight from the Vulkan 1.2 spec, section 6.1.4 "Framebuffer
-    * Region Dependencies":
-    */
-   const VkPipelineStageFlags2 framebuffer_space_stages =
-      VK_PIPELINE_STAGE_2_FRAGMENT_SHADER_BIT |
-      VK_PIPELINE_STAGE_2_EARLY_FRAGMENT_TESTS_BIT |
-      VK_PIPELINE_STAGE_2_LATE_FRAGMENT_TESTS_BIT |
-      VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT;
-
-   return
-      (src_stage_mask & ~(framebuffer_space_stages | VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT)) ||
-      (dst_stage_mask & ~(framebuffer_space_stages | VK_PIPELINE_STAGE_2_BOTTOM_OF_PIPE_BIT)) ||
-      !(dep->dependencyFlags & VK_DEPENDENCY_BY_REGION_BIT);
-}
-
-static void
-tu_render_pass_add_subpass_dep(struct tu_render_pass *pass,
-                               const VkSubpassDependency2 *dep)
-{
-   uint32_t src = dep->srcSubpass;
-   uint32_t dst = dep->dstSubpass;
-
-   /* Ignore subpass self-dependencies as they allow the app to call
-    * vkCmdPipelineBarrier() inside the render pass and the driver should only
-    * do the barrier when called, not when starting the render pass.
-    *
-    * We cannot decide whether to allow gmem rendering before a barrier
-    * is actually emitted, so we delay the decision until then.
-    */
-   if (src == dst)
-      return;
-
-   /* From the Vulkan 1.2.195 spec:
-    *
-    * "If an instance of VkMemoryBarrier2 is included in the pNext chain, srcStageMask,
-    *  dstStageMask, srcAccessMask, and dstAccessMask parameters are ignored. The synchronization
-    *  and access scopes instead are defined by the parameters of VkMemoryBarrier2."
-    */
-   const VkMemoryBarrier2 *barrier =
-      vk_find_struct_const(dep->pNext, MEMORY_BARRIER_2);
-   VkPipelineStageFlags2 src_stage_mask = barrier ? barrier->srcStageMask : dep->srcStageMask;
-   VkAccessFlags2 src_access_mask = barrier ? barrier->srcAccessMask : dep->srcAccessMask;
-   VkPipelineStageFlags2 dst_stage_mask = barrier ? barrier->dstStageMask : dep->dstStageMask;
-   VkAccessFlags2 dst_access_mask = barrier ? barrier->dstAccessMask : dep->dstAccessMask;
-
-   if (dep_invalid_for_gmem(dep, src_stage_mask, dst_stage_mask)) {
-      perf_debug((struct tu_device *)pass->base.device, "Disabling gmem rendering due to invalid subpass dependency");
-      for (int i = 0; i < ARRAY_SIZE(pass->gmem_pixels); i++)
-         pass->gmem_pixels[i] = 0;
-   }
-
-   struct tu_subpass_barrier *dst_barrier;
-   if (dst == VK_SUBPASS_EXTERNAL) {
-      dst_barrier = &pass->end_barrier;
-   } else {
-      dst_barrier = &pass->subpasses[dst].start_barrier;
-   }
-
-   dst_barrier->src_stage_mask |= src_stage_mask;
-   dst_barrier->dst_stage_mask |= dst_stage_mask;
-   dst_barrier->src_access_mask |= src_access_mask;
-   dst_barrier->dst_access_mask |= dst_access_mask;
-}
-
-/* We currently only care about undefined layouts, because we have to
- * flush/invalidate CCU for those. PREINITIALIZED is the same thing as
- * UNDEFINED for anything not linear tiled, but we don't know yet whether the
- * images used are tiled, so just assume they are.
- */
-
-static bool
-layout_undefined(VkImageLayout layout)
-{
-   return layout == VK_IMAGE_LAYOUT_UNDEFINED ||
-          layout == VK_IMAGE_LAYOUT_PREINITIALIZED;
-}
-
-/* This implements the following bit of spec text:
- *
- *    If there is no subpass dependency from VK_SUBPASS_EXTERNAL to the
- *    first subpass that uses an attachment, then an implicit subpass
- *    dependency exists from VK_SUBPASS_EXTERNAL to the first subpass it is
- *    used in. The implicit subpass dependency only exists if there
- *    exists an automatic layout transition away from initialLayout.
- *    The subpass dependency operates as if defined with the
- *    following parameters:
- *
- *    VkSubpassDependency implicitDependency = {
- *        .srcSubpass = VK_SUBPASS_EXTERNAL;
- *        .dstSubpass = firstSubpass; // First subpass attachment is used in
- *        .srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
- *        .dstStageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
- *        .srcAccessMask = 0;
- *        .dstAccessMask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT |
- *                         VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
- *                         VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
- *                         VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
- *                         VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
- *        .dependencyFlags = 0;
- *    };
  *
- *    Similarly, if there is no subpass dependency from the last subpass
- *    that uses an attachment to VK_SUBPASS_EXTERNAL, then an implicit
- *    subpass dependency exists from the last subpass it is used in to
- *    VK_SUBPASS_EXTERNAL. The implicit subpass dependency only exists
- *    if there exists an automatic layout transition into finalLayout.
- *    The subpass dependency operates as if defined with the following
- *    parameters:
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
  *
- *    VkSubpassDependency implicitDependency = {
- *        .srcSubpass = lastSubpass; // Last subpass attachment is used in
- *        .dstSubpass = VK_SUBPASS_EXTERNAL;
- *        .srcStageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
- *        .dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT;
- *        .srcAccessMask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT |
- *                         VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
- *                         VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
- *                         VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
- *                         VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
- *        .dstAccessMask = 0;
- *        .dependencyFlags = 0;
- *    };
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
  *
- * Note: currently this is the only use we have for layout transitions,
- * besides needing to invalidate CCU at the beginning, so we also flag
- * transitions from UNDEFINED here.
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
-static void
-tu_render_pass_add_implicit_deps(struct tu_render_pass *pass,
-                                 const VkRenderPassCreateInfo2 *info)
-{
-   const VkAttachmentDescription2* att = info->pAttachments;
-   bool has_external_src[info->subpassCount];
-   bool has_external_dst[info->subpassCount];
-   bool att_used[pass->attachment_count];
-
-   memset(has_external_src, 0, sizeof(has_external_src));
-   memset(has_external_dst, 0, sizeof(has_external_dst));
-
-   for (uint32_t i = 0; i < info->dependencyCount; i++) {
-      uint32_t src = info->pDependencies[i].srcSubpass;
-      uint32_t dst = info->pDependencies[i].dstSubpass;
-
-      if (src == dst)
-         continue;
+#include "tu_private.h"
 
-      if (src == VK_SUBPASS_EXTERNAL)
-         has_external_src[dst] = true;
-      if (dst == VK_SUBPASS_EXTERNAL)
-         has_external_dst[src] = true;
-   }
-
-   memset(att_used, 0, sizeof(att_used));
-
-   for (unsigned i = 0; i < info->subpassCount; i++) {
-      const VkSubpassDescription2 *subpass = &info->pSubpasses[i];
-      bool src_implicit_dep = false;
-
-      for (unsigned j = 0; j < subpass->inputAttachmentCount; j++) {
-         uint32_t a = subpass->pInputAttachments[j].attachment;
-
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-
-         uint32_t stencil_layout = vk_format_has_stencil(att[a].format) ?
-               vk_att_ref_stencil_layout(&subpass->pInputAttachments[j], att) :
-               VK_IMAGE_LAYOUT_UNDEFINED;
-         uint32_t stencil_initial_layout = vk_att_desc_stencil_layout(&att[a], false);
-
-         if ((att[a].initialLayout != subpass->pInputAttachments[j].layout ||
-             stencil_initial_layout != stencil_layout) &&
-             !att_used[a] && !has_external_src[i])
-            src_implicit_dep = true;
-         att_used[a] = true;
-      }
-
-      for (unsigned j = 0; j < subpass->colorAttachmentCount; j++) {
-         uint32_t a = subpass->pColorAttachments[j].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-         if (att[a].initialLayout != subpass->pColorAttachments[j].layout &&
-             !att_used[a] && !has_external_src[i])
-            src_implicit_dep = true;
-         att_used[a] = true;
-      }
-
-      if (subpass->pDepthStencilAttachment &&
-          subpass->pDepthStencilAttachment->attachment != VK_ATTACHMENT_UNUSED) {
-         uint32_t a = subpass->pDepthStencilAttachment->attachment;
-         uint32_t stencil_layout = vk_att_ref_stencil_layout(subpass->pDepthStencilAttachment, att);
-         uint32_t stencil_initial_layout = vk_att_desc_stencil_layout(&att[a], false);
+#include "vk_util.h"
 
-         if ((att[a].initialLayout != subpass->pDepthStencilAttachment->layout ||
-             stencil_initial_layout != stencil_layout) &&
-             !att_used[a] && !has_external_src[i]) {
-            src_implicit_dep = true;
-         }
-         att_used[a] = true;
-      }
+VkResult
+tu_CreateRenderPass(VkDevice _device,
+                    const VkRenderPassCreateInfo *pCreateInfo,
+                    const VkAllocationCallbacks *pAllocator,
+                    VkRenderPass *pRenderPass)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   struct tu_render_pass *pass;
+   size_t size;
+   size_t attachments_offset;
+   VkRenderPassMultiviewCreateInfo *multiview_info = NULL;
 
-      if (subpass->pResolveAttachments) {
-         for (unsigned j = 0; j < subpass->colorAttachmentCount; j++) {
-            uint32_t a = subpass->pResolveAttachments[j].attachment;
-            if (a == VK_ATTACHMENT_UNUSED)
-               continue;
-            if (att[a].initialLayout != subpass->pResolveAttachments[j].layout &&
-               !att_used[a] && !has_external_src[i])
-               src_implicit_dep = true;
-            att_used[a] = true;
-         }
-      }
+   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO);
 
-      const VkSubpassDescriptionDepthStencilResolve *ds_resolve =
-         vk_find_struct_const(subpass->pNext, SUBPASS_DESCRIPTION_DEPTH_STENCIL_RESOLVE);
+   size = sizeof(*pass);
+   size += pCreateInfo->subpassCount * sizeof(pass->subpasses[0]);
+   attachments_offset = size;
+   size += pCreateInfo->attachmentCount * sizeof(pass->attachments[0]);
 
-      if (ds_resolve && ds_resolve->pDepthStencilResolveAttachment &&
-          ds_resolve->pDepthStencilResolveAttachment->attachment != VK_ATTACHMENT_UNUSED) {
-            uint32_t a = ds_resolve->pDepthStencilResolveAttachment->attachment;
-            uint32_t stencil_layout = vk_att_ref_stencil_layout(ds_resolve->pDepthStencilResolveAttachment, att);
-            uint32_t stencil_initial_layout = vk_att_desc_stencil_layout(&att[a], false);
+   pass = vk_alloc2(&device->alloc, pAllocator, size, 8,
+                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (pass == NULL)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
-            if ((att[a].initialLayout != subpass->pDepthStencilAttachment->layout ||
-                stencil_initial_layout != stencil_layout) &&
-                !att_used[a] && !has_external_src[i])
-               src_implicit_dep = true;
-            att_used[a] = true;
-      }
+   memset(pass, 0, size);
+   pass->attachment_count = pCreateInfo->attachmentCount;
+   pass->subpass_count = pCreateInfo->subpassCount;
+   pass->attachments = (void *) pass + attachments_offset;
 
-      if (src_implicit_dep) {
-         tu_render_pass_add_subpass_dep(pass, &(VkSubpassDependency2) {
-            .srcSubpass = VK_SUBPASS_EXTERNAL,
-            .dstSubpass = i,
-            .srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
-            .dstStageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
-            .srcAccessMask = 0,
-            .dstAccessMask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT |
-                             VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
-                             VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
-                             VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
-                             VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,
-            .dependencyFlags = 0,
-         });
+   vk_foreach_struct(ext, pCreateInfo->pNext)
+   {
+      switch (ext->sType) {
+      case VK_STRUCTURE_TYPE_RENDER_PASS_MULTIVIEW_CREATE_INFO:
+         multiview_info = (VkRenderPassMultiviewCreateInfo *) ext;
+         break;
+      default:
+         break;
       }
    }
 
-   memset(att_used, 0, sizeof(att_used));
-
-   for (int i = info->subpassCount - 1; i >= 0; i--) {
-      const VkSubpassDescription2 *subpass = &info->pSubpasses[i];
-      bool dst_implicit_dep = false;
-
-      for (unsigned j = 0; j < subpass->inputAttachmentCount; j++) {
-         uint32_t a = subpass->pInputAttachments[j].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-
-         uint32_t stencil_layout = vk_format_has_stencil(att[a].format) ?
-               vk_att_ref_stencil_layout(&subpass->pInputAttachments[j], att) :
-               VK_IMAGE_LAYOUT_UNDEFINED;
-         uint32_t stencil_final_layout = vk_att_desc_stencil_layout(&att[a], true);
-
-         if ((att[a].finalLayout != subpass->pInputAttachments[j].layout ||
-             stencil_final_layout != stencil_layout) &&
-             !att_used[a] && !has_external_dst[i])
-            dst_implicit_dep = true;
-         att_used[a] = true;
-      }
-
-      for (unsigned j = 0; j < subpass->colorAttachmentCount; j++) {
-         uint32_t a = subpass->pColorAttachments[j].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-         if (att[a].finalLayout != subpass->pColorAttachments[j].layout &&
-             !att_used[a] && !has_external_dst[i])
-            dst_implicit_dep = true;
-         att_used[a] = true;
-      }
-
-      if (subpass->pDepthStencilAttachment &&
-          subpass->pDepthStencilAttachment->attachment != VK_ATTACHMENT_UNUSED) {
-         uint32_t a = subpass->pDepthStencilAttachment->attachment;
-         uint32_t stencil_layout = vk_att_ref_stencil_layout(subpass->pDepthStencilAttachment, att);
-         uint32_t stencil_final_layout = vk_att_desc_stencil_layout(&att[a], true);
-
-         if ((att[a].finalLayout != subpass->pDepthStencilAttachment->layout ||
-             stencil_final_layout != stencil_layout) &&
-             !att_used[a] && !has_external_dst[i]) {
-            dst_implicit_dep = true;
-         }
-         att_used[a] = true;
-      }
-
-      if (subpass->pResolveAttachments) {
-         for (unsigned j = 0; j < subpass->colorAttachmentCount; j++) {
-            uint32_t a = subpass->pResolveAttachments[j].attachment;
-            if (a == VK_ATTACHMENT_UNUSED)
-               continue;
-            if (att[a].finalLayout != subpass->pResolveAttachments[j].layout &&
-                !att_used[a] && !has_external_dst[i])
-               dst_implicit_dep = true;
-            att_used[a] = true;
-         }
-      }
-
-      const VkSubpassDescriptionDepthStencilResolve *ds_resolve =
-         vk_find_struct_const(subpass->pNext, SUBPASS_DESCRIPTION_DEPTH_STENCIL_RESOLVE);
-
-      if (ds_resolve && ds_resolve->pDepthStencilResolveAttachment &&
-          ds_resolve->pDepthStencilResolveAttachment->attachment != VK_ATTACHMENT_UNUSED) {
-            uint32_t a = ds_resolve->pDepthStencilResolveAttachment->attachment;
-            uint32_t stencil_layout = vk_att_ref_stencil_layout(ds_resolve->pDepthStencilResolveAttachment, att);
-            uint32_t stencil_final_layout = vk_att_desc_stencil_layout(&att[a], true);
-
-            if ((att[a].finalLayout != subpass->pDepthStencilAttachment->layout ||
-                stencil_final_layout != stencil_layout) &&
-                !att_used[a] && !has_external_src[i])
-               dst_implicit_dep = true;
-            att_used[a] = true;
-      }
+   for (uint32_t i = 0; i < pCreateInfo->attachmentCount; i++) {
+      struct tu_render_pass_attachment *att = &pass->attachments[i];
 
-      if (dst_implicit_dep) {
-         tu_render_pass_add_subpass_dep(pass, &(VkSubpassDependency2) {
-            .srcSubpass = i,
-            .dstSubpass = VK_SUBPASS_EXTERNAL,
-            .srcStageMask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
-            .dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT,
-            .srcAccessMask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT |
-                             VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
-                             VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
-                             VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
-                             VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,
-            .dstAccessMask = 0,
-            .dependencyFlags = 0,
-         });
-      }
+      att->format = pCreateInfo->pAttachments[i].format;
+      att->samples = pCreateInfo->pAttachments[i].samples;
+      att->load_op = pCreateInfo->pAttachments[i].loadOp;
+      att->stencil_load_op = pCreateInfo->pAttachments[i].stencilLoadOp;
+      att->initial_layout = pCreateInfo->pAttachments[i].initialLayout;
+      att->final_layout = pCreateInfo->pAttachments[i].finalLayout;
+      // att->store_op = pCreateInfo->pAttachments[i].storeOp;
+      // att->stencil_store_op = pCreateInfo->pAttachments[i].stencilStoreOp;
    }
+   uint32_t subpass_attachment_count = 0;
+   struct tu_subpass_attachment *p;
+   for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
+      const VkSubpassDescription *desc = &pCreateInfo->pSubpasses[i];
 
-   /* Handle UNDEFINED transitions, similar to the handling in tu_barrier().
-    * Assume that if an attachment has an initial layout of UNDEFINED, it gets
-    * transitioned eventually.
-    */
-   for (unsigned i = 0; i < info->attachmentCount; i++) {
-      if (layout_undefined(att[i].initialLayout)) {
-         if (vk_format_is_depth_or_stencil(att[i].format)) {
-            pass->subpasses[0].start_barrier.incoherent_ccu_depth = true;
-         } else {
-            pass->subpasses[0].start_barrier.incoherent_ccu_color = true;
-         }
-      }
+      subpass_attachment_count +=
+         desc->inputAttachmentCount + desc->colorAttachmentCount +
+         (desc->pResolveAttachments ? desc->colorAttachmentCount : 0) +
+         (desc->pDepthStencilAttachment != NULL);
    }
-}
 
-/* If an input attachment is used without an intervening write to the same
- * attachment, then we can just use the original image, even in GMEM mode.
- * This is an optimization, but it's also important because it allows us to
- * avoid having to invalidate UCHE at the beginning of each tile due to it
- * becoming invalid. The only reads of GMEM via UCHE should be after an
- * earlier subpass modified it, which only works if there's already an
- * appropriate dependency that will add the CACHE_INVALIDATE anyway. We
- * don't consider this in the dependency code, so this is also required for
- * correctness.
- */
-static void
-tu_render_pass_patch_input_gmem(struct tu_render_pass *pass)
-{
-   bool written[pass->attachment_count];
-
-   memset(written, 0, sizeof(written));
-
-   for (unsigned i = 0; i < pass->subpass_count; i++) {
-      struct tu_subpass *subpass = &pass->subpasses[i];
-
-      for (unsigned j = 0; j < subpass->input_count; j++) {
-         uint32_t a = subpass->input_attachments[j].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-         subpass->input_attachments[j].patch_input_gmem = written[a];
+   if (subpass_attachment_count) {
+      pass->subpass_attachments = vk_alloc2(
+         &device->alloc, pAllocator,
+         subpass_attachment_count * sizeof(struct tu_subpass_attachment), 8,
+         VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+      if (pass->subpass_attachments == NULL) {
+         vk_free2(&device->alloc, pAllocator, pass);
+         return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
       }
+   } else
+      pass->subpass_attachments = NULL;
 
-      for (unsigned j = 0; j < subpass->color_count; j++) {
-         uint32_t a = subpass->color_attachments[j].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-         written[a] = true;
+   p = pass->subpass_attachments;
+   for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
+      const VkSubpassDescription *desc = &pCreateInfo->pSubpasses[i];
+      uint32_t color_sample_count = 1, depth_sample_count = 1;
+      struct tu_subpass *subpass = &pass->subpasses[i];
 
-         for (unsigned k = 0; k < subpass->input_count; k++) {
-            if (subpass->input_attachments[k].attachment == a &&
-                !subpass->input_attachments[k].patch_input_gmem) {
-               /* For render feedback loops, we have no idea whether the use
-                * as a color attachment or input attachment will come first,
-                * so we have to always use GMEM in case the color attachment
-                * comes first and defensively invalidate UCHE in case the
-                * input attachment comes first.
-                */
-               subpass->feedback_invalidate = true;
-               subpass->input_attachments[k].patch_input_gmem = true;
-            }
-         }
-      }
+      subpass->input_count = desc->inputAttachmentCount;
+      subpass->color_count = desc->colorAttachmentCount;
+      if (multiview_info)
+         subpass->view_mask = multiview_info->pViewMasks[i];
 
-      for (unsigned j = 0; j < subpass->resolve_count; j++) {
-         uint32_t a = subpass->resolve_attachments[j].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-         written[a] = true;
-      }
+      if (desc->inputAttachmentCount > 0) {
+         subpass->input_attachments = p;
+         p += desc->inputAttachmentCount;
 
-      if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
-         written[subpass->depth_stencil_attachment.attachment] = true;
-         for (unsigned k = 0; k < subpass->input_count; k++) {
-            if (subpass->input_attachments[k].attachment ==
-                subpass->depth_stencil_attachment.attachment &&
-                !subpass->input_attachments[k].patch_input_gmem) {
-               subpass->feedback_invalidate = true;
-               subpass->input_attachments[k].patch_input_gmem = true;
-            }
+         for (uint32_t j = 0; j < desc->inputAttachmentCount; j++) {
+            subpass->input_attachments[j] = (struct tu_subpass_attachment) {
+               .attachment = desc->pInputAttachments[j].attachment,
+               .layout = desc->pInputAttachments[j].layout,
+            };
+            if (desc->pInputAttachments[j].attachment != VK_ATTACHMENT_UNUSED)
+               pass->attachments[desc->pInputAttachments[j].attachment]
+                  .view_mask |= subpass->view_mask;
          }
       }
-   }
-}
-
-static void
-tu_render_pass_check_feedback_loop(struct tu_render_pass *pass)
-{
-   for (unsigned i = 0; i < pass->subpass_count; i++) {
-      struct tu_subpass *subpass = &pass->subpasses[i];
 
-      for (unsigned j = 0; j < subpass->color_count; j++) {
-         uint32_t a = subpass->color_attachments[j].attachment;
-         if (a == VK_ATTACHMENT_UNUSED)
-            continue;
-         for (unsigned k = 0; k < subpass->input_count; k++) {
-            if (subpass->input_attachments[k].attachment == a) {
-               subpass->feedback_loop_color = true;
-               break;
-            }
-         }
-      }
+      if (desc->colorAttachmentCount > 0) {
+         subpass->color_attachments = p;
+         p += desc->colorAttachmentCount;
 
-      if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
-         for (unsigned k = 0; k < subpass->input_count; k++) {
-            if (subpass->input_attachments[k].attachment ==
-                subpass->depth_stencil_attachment.attachment) {
-               subpass->feedback_loop_ds = true;
-               break;
+         for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
+            subpass->color_attachments[j] = (struct tu_subpass_attachment) {
+               .attachment = desc->pColorAttachments[j].attachment,
+               .layout = desc->pColorAttachments[j].layout,
+            };
+            if (desc->pColorAttachments[j].attachment !=
+                VK_ATTACHMENT_UNUSED) {
+               pass->attachments[desc->pColorAttachments[j].attachment]
+                  .view_mask |= subpass->view_mask;
+               color_sample_count =
+                  pCreateInfo
+                     ->pAttachments[desc->pColorAttachments[j].attachment]
+                     .samples;
             }
          }
       }
-   }
-}
-
-static void update_samples(struct tu_subpass *subpass,
-                           VkSampleCountFlagBits samples)
-{
-   assert(subpass->samples == 0 || subpass->samples == samples);
-   subpass->samples = samples;
-}
 
-static void
-tu_render_pass_calc_hash(struct tu_render_pass *pass)
-{
-   #define HASH(hash, data) XXH64(&(data), sizeof(data), hash)
-
-   uint64_t hash = HASH(0, pass->attachment_count);
-   hash = XXH64(pass->attachments,
-         pass->attachment_count * sizeof(pass->attachments[0]), hash);
-   hash = HASH(hash, pass->subpass_count);
-   for (unsigned i = 0; i < pass->subpass_count; i++) {
-      hash = HASH(hash, pass->subpasses[i].samples);
-      hash = HASH(hash, pass->subpasses[i].input_count);
-      hash = HASH(hash, pass->subpasses[i].color_count);
-      hash = HASH(hash, pass->subpasses[i].resolve_count);
-   }
-
-   pass->autotune_hash = hash;
-
-   #undef HASH
-}
-
-static void
-tu_render_pass_cond_config(struct tu_render_pass *pass)
-{
-   for (uint32_t i = 0; i < pass->attachment_count; i++) {
-      struct tu_render_pass_attachment *att = &pass->attachments[i];
-
-      att->cond_load_allowed =
-         (att->load || att->load_stencil) && !att->clear_mask && !att->will_be_resolved;
-      att->cond_store_allowed =
-         (att->store || att->store_stencil) && !att->clear_mask;
-   }
-}
-
-static void
-tu_render_pass_gmem_config(struct tu_render_pass *pass,
-                           const struct tu_physical_device *phys_dev)
-{
-   for (enum tu_gmem_layout layout = 0; layout < TU_GMEM_LAYOUT_COUNT;
-        layout++) {
-      /* From the VK_KHR_multiview spec:
-       *
-       *    Multiview is all-or-nothing for a render pass - that is, either all
-       *    subpasses must have a non-zero view mask (though some subpasses may
-       *    have only one view) or all must be zero.
-       *
-       * This means we only have to check one of the view masks.
-       */
-      if (pass->subpasses[0].multiview_mask) {
-         /* It seems multiview must use sysmem rendering. */
-         pass->gmem_pixels[layout] = 0;
-         continue;
-      }
-
-      /* log2(gmem_align/(tile_align_w*tile_align_h)) */
-      uint32_t block_align_shift = 3;
-      uint32_t tile_align_w = phys_dev->info->tile_align_w;
-      uint32_t gmem_align = (1 << block_align_shift) * tile_align_w *
-                            phys_dev->info->tile_align_h;
-
-      /* calculate total bytes per pixel */
-      uint32_t cpp_total = 0;
-      for (uint32_t i = 0; i < pass->attachment_count; i++) {
-         struct tu_render_pass_attachment *att = &pass->attachments[i];
-         bool cpp1 = (att->cpp == 1);
-         if (att->gmem) {
-            cpp_total += att->cpp;
-
-            /* take into account the separate stencil: */
-            if (att->format == VK_FORMAT_D32_SFLOAT_S8_UINT) {
-               cpp1 = (att->samples == 1);
-               cpp_total += att->samples;
-            }
+      subpass->has_resolve = false;
+      if (desc->pResolveAttachments) {
+         subpass->resolve_attachments = p;
+         p += desc->colorAttachmentCount;
 
-            /* texture pitch must be aligned to 64, use a tile_align_w that is
-             * a multiple of 64 for cpp==1 attachment to work as input
-             * attachment
-             */
-            if (cpp1 && tile_align_w % 64 != 0) {
-               tile_align_w *= 2;
-               block_align_shift -= 1;
+         for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
+            uint32_t a = desc->pResolveAttachments[j].attachment;
+            subpass->resolve_attachments[j] = (struct tu_subpass_attachment) {
+               .attachment = desc->pResolveAttachments[j].attachment,
+               .layout = desc->pResolveAttachments[j].layout,
+            };
+            if (a != VK_ATTACHMENT_UNUSED) {
+               subpass->has_resolve = true;
+               pass->attachments[desc->pResolveAttachments[j].attachment]
+                  .view_mask |= subpass->view_mask;
             }
          }
       }
 
-      pass->tile_align_w = tile_align_w;
-
-      /* no gmem attachments */
-      if (cpp_total == 0) {
-         /* any value non-zero value so tiling config works with no
-          * attachments
-          */
-         pass->gmem_pixels[layout] = 1024 * 1024;
-         continue;
-      }
-
-      /* TODO: this algorithm isn't optimal
-       * for example, two attachments with cpp = {1, 4}
-       * result:  nblocks = {12, 52}, pixels = 196608
-       * optimal: nblocks = {13, 51}, pixels = 208896
-       */
-      uint32_t gmem_size = layout == TU_GMEM_LAYOUT_FULL
-                              ? phys_dev->gmem_size
-                              : phys_dev->ccu_offset_gmem;
-      uint32_t gmem_blocks = gmem_size / gmem_align;
-      uint32_t offset = 0, pixels = ~0u, i;
-      for (i = 0; i < pass->attachment_count; i++) {
-         struct tu_render_pass_attachment *att = &pass->attachments[i];
-         if (!att->gmem)
-            continue;
-
-         att->gmem_offset[layout] = offset;
-
-         uint32_t align = MAX2(1, att->cpp >> block_align_shift);
-         uint32_t nblocks =
-            MAX2((gmem_blocks * att->cpp / cpp_total) & ~(align - 1), align);
-
-         if (nblocks > gmem_blocks)
-            break;
-
-         gmem_blocks -= nblocks;
-         cpp_total -= att->cpp;
-         offset += nblocks * gmem_align;
-         pixels = MIN2(pixels, nblocks * gmem_align / att->cpp);
-
-         /* repeat the same for separate stencil */
-         if (att->format == VK_FORMAT_D32_SFLOAT_S8_UINT) {
-            att->gmem_offset_stencil[layout] = offset;
-
-            /* note: for s8_uint, block align is always 1 */
-            uint32_t nblocks = gmem_blocks * att->samples / cpp_total;
-            if (nblocks > gmem_blocks)
-               break;
-
-            gmem_blocks -= nblocks;
-            cpp_total -= att->samples;
-            offset += nblocks * gmem_align;
-            pixels = MIN2(pixels, nblocks * gmem_align / att->samples);
+      if (desc->pDepthStencilAttachment) {
+         subpass->depth_stencil_attachment = (struct tu_subpass_attachment) {
+            .attachment = desc->pDepthStencilAttachment->attachment,
+            .layout = desc->pDepthStencilAttachment->layout,
+         };
+         if (desc->pDepthStencilAttachment->attachment !=
+             VK_ATTACHMENT_UNUSED) {
+            pass->attachments[desc->pDepthStencilAttachment->attachment]
+               .view_mask |= subpass->view_mask;
+            depth_sample_count =
+               pCreateInfo
+                  ->pAttachments[desc->pDepthStencilAttachment->attachment]
+                  .samples;
          }
+      } else {
+         subpass->depth_stencil_attachment.attachment = VK_ATTACHMENT_UNUSED;
       }
 
-      /* if the loop didn't complete then the gmem config is impossible */
-      if (i == pass->attachment_count)
-         pass->gmem_pixels[layout] = pixels;
+      subpass->max_sample_count =
+         MAX2(color_sample_count, depth_sample_count);
    }
-}
 
-static void
-tu_render_pass_bandwidth_config(struct tu_render_pass *pass)
-{
-   pass->gmem_bandwidth_per_pixel = 0;
-   pass->sysmem_bandwidth_per_pixel = 0;
-
-   for (uint32_t i = 0; i < pass->attachment_count; i++) {
-      const struct tu_render_pass_attachment *att = &pass->attachments[i];
-
-      /* approximate tu_load_gmem_attachment */
-      if (att->load)
-         pass->gmem_bandwidth_per_pixel += att->cpp;
-
-      /* approximate tu_store_gmem_attachment */
-      if (att->store)
-         pass->gmem_bandwidth_per_pixel += att->cpp;
-
-      /* approximate tu_clear_sysmem_attachment */
-      if (att->clear_mask)
-         pass->sysmem_bandwidth_per_pixel += att->cpp;
-
-      /* approximate tu6_emit_sysmem_resolves */
-      if (att->will_be_resolved) {
-         pass->sysmem_bandwidth_per_pixel +=
-            att->cpp + att->cpp / att->samples;
+   for (unsigned i = 0; i < pCreateInfo->dependencyCount; ++i) {
+      uint32_t dst = pCreateInfo->pDependencies[i].dstSubpass;
+      if (dst == VK_SUBPASS_EXTERNAL) {
+         pass->end_barrier.src_stage_mask =
+            pCreateInfo->pDependencies[i].srcStageMask;
+         pass->end_barrier.src_access_mask =
+            pCreateInfo->pDependencies[i].srcAccessMask;
+         pass->end_barrier.dst_access_mask =
+            pCreateInfo->pDependencies[i].dstAccessMask;
+      } else {
+         pass->subpasses[dst].start_barrier.src_stage_mask =
+            pCreateInfo->pDependencies[i].srcStageMask;
+         pass->subpasses[dst].start_barrier.src_access_mask =
+            pCreateInfo->pDependencies[i].srcAccessMask;
+         pass->subpasses[dst].start_barrier.dst_access_mask =
+            pCreateInfo->pDependencies[i].dstAccessMask;
       }
    }
-}
-
-static void
-attachment_set_ops(struct tu_device *device,
-                   struct tu_render_pass_attachment *att,
-                   VkAttachmentLoadOp load_op,
-                   VkAttachmentLoadOp stencil_load_op,
-                   VkAttachmentStoreOp store_op,
-                   VkAttachmentStoreOp stencil_store_op)
-{
-   if (device->instance->debug_flags & TU_DEBUG_DONT_CARE_AS_LOAD) {
-      if (load_op == VK_ATTACHMENT_LOAD_OP_DONT_CARE)
-         load_op = VK_ATTACHMENT_LOAD_OP_LOAD;
-      if (stencil_load_op == VK_ATTACHMENT_LOAD_OP_DONT_CARE)
-         stencil_load_op = VK_ATTACHMENT_LOAD_OP_LOAD;
-   }
-
-   /* load/store ops */
-   att->clear_mask =
-      (load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) ? VK_IMAGE_ASPECT_COLOR_BIT : 0;
-   att->load = (load_op == VK_ATTACHMENT_LOAD_OP_LOAD);
-   att->store = (store_op == VK_ATTACHMENT_STORE_OP_STORE);
-
-   bool stencil_clear = (stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR);
-   bool stencil_load = (stencil_load_op == VK_ATTACHMENT_LOAD_OP_LOAD);
-   bool stencil_store = (stencil_store_op == VK_ATTACHMENT_STORE_OP_STORE);
 
-   switch (att->format) {
-   case VK_FORMAT_D24_UNORM_S8_UINT: /* || stencil load/store */
-      if (att->clear_mask)
-         att->clear_mask = VK_IMAGE_ASPECT_DEPTH_BIT;
-      if (stencil_clear)
-         att->clear_mask |= VK_IMAGE_ASPECT_STENCIL_BIT;
-      if (stencil_load)
-         att->load = true;
-      if (stencil_store)
-         att->store = true;
-      break;
-   case VK_FORMAT_S8_UINT: /* replace load/store with stencil load/store */
-      att->clear_mask = stencil_clear ? VK_IMAGE_ASPECT_COLOR_BIT : 0;
-      att->load = stencil_load;
-      att->store = stencil_store;
-      break;
-   case VK_FORMAT_D32_SFLOAT_S8_UINT: /* separate stencil */
-      if (att->clear_mask)
-         att->clear_mask = VK_IMAGE_ASPECT_DEPTH_BIT;
-      if (stencil_clear)
-         att->clear_mask |= VK_IMAGE_ASPECT_STENCIL_BIT;
-      if (stencil_load)
-         att->load_stencil = true;
-      if (stencil_store)
-         att->store_stencil = true;
-      break;
-   default:
-      break;
-   }
-}
-
-static bool
-is_depth_stencil_resolve_enabled(const VkSubpassDescriptionDepthStencilResolve *depth_stencil_resolve)
-{
-   if (depth_stencil_resolve &&
-       depth_stencil_resolve->pDepthStencilResolveAttachment &&
-       depth_stencil_resolve->pDepthStencilResolveAttachment->attachment != VK_ATTACHMENT_UNUSED) {
-      return true;
-   }
-   return false;
-}
-
-static void
-tu_subpass_use_attachment(struct tu_render_pass *pass, int i, uint32_t a, const VkRenderPassCreateInfo2 *pCreateInfo)
-{
-   struct tu_subpass *subpass = &pass->subpasses[i];
+   *pRenderPass = tu_render_pass_to_handle(pass);
 
-   pass->attachments[a].gmem = true;
-   update_samples(subpass, pCreateInfo->pAttachments[a].samples);
-   pass->attachments[a].clear_views |= subpass->multiview_mask;
+   return VK_SUCCESS;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_CreateRenderPass2(VkDevice _device,
-                     const VkRenderPassCreateInfo2 *pCreateInfo,
-                     const VkAllocationCallbacks *pAllocator,
-                     VkRenderPass *pRenderPass)
+VkResult
+tu_CreateRenderPass2KHR(VkDevice _device,
+                        const VkRenderPassCreateInfo2KHR *pCreateInfo,
+                        const VkAllocationCallbacks *pAllocator,
+                        VkRenderPass *pRenderPass)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-
-   if (unlikely(device->instance->debug_flags & TU_DEBUG_DYNAMIC))
-      return vk_common_CreateRenderPass2(_device, pCreateInfo, pAllocator,
-                                         pRenderPass);
-
    struct tu_render_pass *pass;
    size_t size;
    size_t attachments_offset;
 
-   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO_2);
+   assert(pCreateInfo->sType ==
+          VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO_2_KHR);
 
    size = sizeof(*pass);
    size += pCreateInfo->subpassCount * sizeof(pass->subpasses[0]);
    attachments_offset = size;
    size += pCreateInfo->attachmentCount * sizeof(pass->attachments[0]);
 
-   pass = vk_object_zalloc(&device->vk, pAllocator, size,
-                           VK_OBJECT_TYPE_RENDER_PASS);
+   pass = vk_alloc2(&device->alloc, pAllocator, size, 8,
+                    VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
    if (pass == NULL)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
+   memset(pass, 0, size);
    pass->attachment_count = pCreateInfo->attachmentCount;
    pass->subpass_count = pCreateInfo->subpassCount;
    pass->attachments = (void *) pass + attachments_offset;
@@ -800,82 +249,58 @@ tu_CreateRenderPass2(VkDevice _device,
 
       att->format = pCreateInfo->pAttachments[i].format;
       att->samples = pCreateInfo->pAttachments[i].samples;
-      /* for d32s8, cpp is for the depth image, and
-       * att->samples will be used as the cpp for the stencil image
-       */
-      if (att->format == VK_FORMAT_D32_SFLOAT_S8_UINT)
-         att->cpp = 4 * att->samples;
-      else
-         att->cpp = vk_format_get_blocksize(att->format) * att->samples;
-      /* Initially not allocated into gmem, tu_subpass_use_attachment() will move it there. */
-      att->gmem = false;
-
-      VkAttachmentLoadOp loadOp = pCreateInfo->pAttachments[i].loadOp;
-      VkAttachmentLoadOp stencilLoadOp = pCreateInfo->pAttachments[i].stencilLoadOp;
-
-      attachment_set_ops(device, att, loadOp, stencilLoadOp,
-                         pCreateInfo->pAttachments[i].storeOp,
-                         pCreateInfo->pAttachments[i].stencilStoreOp);
+      att->load_op = pCreateInfo->pAttachments[i].loadOp;
+      att->stencil_load_op = pCreateInfo->pAttachments[i].stencilLoadOp;
+      att->initial_layout = pCreateInfo->pAttachments[i].initialLayout;
+      att->final_layout = pCreateInfo->pAttachments[i].finalLayout;
+      // att->store_op = pCreateInfo->pAttachments[i].storeOp;
+      // att->stencil_store_op = pCreateInfo->pAttachments[i].stencilStoreOp;
    }
    uint32_t subpass_attachment_count = 0;
    struct tu_subpass_attachment *p;
    for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
-      const VkSubpassDescription2 *desc = &pCreateInfo->pSubpasses[i];
-      const VkSubpassDescriptionDepthStencilResolve *ds_resolve =
-         vk_find_struct_const(desc->pNext, SUBPASS_DESCRIPTION_DEPTH_STENCIL_RESOLVE);
+      const VkSubpassDescription2KHR *desc = &pCreateInfo->pSubpasses[i];
 
       subpass_attachment_count +=
          desc->inputAttachmentCount + desc->colorAttachmentCount +
          (desc->pResolveAttachments ? desc->colorAttachmentCount : 0) +
-         (is_depth_stencil_resolve_enabled(ds_resolve) ? 1 : 0);
+         (desc->pDepthStencilAttachment != NULL);
    }
 
    if (subpass_attachment_count) {
       pass->subpass_attachments = vk_alloc2(
-         &device->vk.alloc, pAllocator,
+         &device->alloc, pAllocator,
          subpass_attachment_count * sizeof(struct tu_subpass_attachment), 8,
          VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
       if (pass->subpass_attachments == NULL) {
-         vk_object_free(&device->vk, pAllocator, pass);
-         return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
+         vk_free2(&device->alloc, pAllocator, pass);
+         return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
       }
    } else
       pass->subpass_attachments = NULL;
 
    p = pass->subpass_attachments;
    for (uint32_t i = 0; i < pCreateInfo->subpassCount; i++) {
-      const VkSubpassDescription2 *desc = &pCreateInfo->pSubpasses[i];
-      const VkSubpassDescriptionDepthStencilResolve *ds_resolve =
-         vk_find_struct_const(desc->pNext, SUBPASS_DESCRIPTION_DEPTH_STENCIL_RESOLVE);
+      const VkSubpassDescription2KHR *desc = &pCreateInfo->pSubpasses[i];
+      uint32_t color_sample_count = 1, depth_sample_count = 1;
       struct tu_subpass *subpass = &pass->subpasses[i];
 
       subpass->input_count = desc->inputAttachmentCount;
       subpass->color_count = desc->colorAttachmentCount;
-      subpass->resolve_count = 0;
-      subpass->resolve_depth_stencil = is_depth_stencil_resolve_enabled(ds_resolve);
-      subpass->samples = 0;
-      subpass->srgb_cntl = 0;
-
-      const VkSubpassDescriptionFlagBits raster_order_access_bits =
-         VK_SUBPASS_DESCRIPTION_RASTERIZATION_ORDER_ATTACHMENT_COLOR_ACCESS_BIT_EXT |
-         VK_SUBPASS_DESCRIPTION_RASTERIZATION_ORDER_ATTACHMENT_DEPTH_ACCESS_BIT_EXT |
-         VK_SUBPASS_DESCRIPTION_RASTERIZATION_ORDER_ATTACHMENT_STENCIL_ACCESS_BIT_EXT;
-
-      subpass->raster_order_attachment_access = desc->flags & raster_order_access_bits;
-
-      subpass->multiview_mask = desc->viewMask;
+      subpass->view_mask = desc->viewMask;
 
       if (desc->inputAttachmentCount > 0) {
          subpass->input_attachments = p;
          p += desc->inputAttachmentCount;
 
          for (uint32_t j = 0; j < desc->inputAttachmentCount; j++) {
-            uint32_t a = desc->pInputAttachments[j].attachment;
-            subpass->input_attachments[j].attachment = a;
-            /* Note: attachments only used as input attachments will be read
-             * directly instead of through gmem, so we don't mark input
-             * attachments as needing gmem.
-             */
+            subpass->input_attachments[j] = (struct tu_subpass_attachment) {
+               .attachment = desc->pInputAttachments[j].attachment,
+               .layout = desc->pInputAttachments[j].layout,
+            };
+            if (desc->pInputAttachments[j].attachment != VK_ATTACHMENT_UNUSED)
+               pass->attachments[desc->pInputAttachments[j].attachment]
+                  .view_mask |= subpass->view_mask;
          }
       }
 
@@ -884,313 +309,108 @@ tu_CreateRenderPass2(VkDevice _device,
          p += desc->colorAttachmentCount;
 
          for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
-            uint32_t a = desc->pColorAttachments[j].attachment;
-            subpass->color_attachments[j].attachment = a;
-
-            if (a != VK_ATTACHMENT_UNUSED) {
-               tu_subpass_use_attachment(pass, i, a, pCreateInfo);
-
-               if (vk_format_is_srgb(pass->attachments[a].format))
-                  subpass->srgb_cntl |= 1 << j;
+            subpass->color_attachments[j] = (struct tu_subpass_attachment) {
+               .attachment = desc->pColorAttachments[j].attachment,
+               .layout = desc->pColorAttachments[j].layout,
+            };
+            if (desc->pColorAttachments[j].attachment !=
+                VK_ATTACHMENT_UNUSED) {
+               pass->attachments[desc->pColorAttachments[j].attachment]
+                  .view_mask |= subpass->view_mask;
+               color_sample_count =
+                  pCreateInfo
+                     ->pAttachments[desc->pColorAttachments[j].attachment]
+                     .samples;
             }
          }
       }
 
-      subpass->resolve_attachments = (desc->pResolveAttachments || subpass->resolve_depth_stencil) ? p : NULL;
+      subpass->has_resolve = false;
       if (desc->pResolveAttachments) {
+         subpass->resolve_attachments = p;
          p += desc->colorAttachmentCount;
-         subpass->resolve_count += desc->colorAttachmentCount;
-         for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
-            subpass->resolve_attachments[j].attachment =
-                  desc->pResolveAttachments[j].attachment;
 
-            uint32_t src_a = desc->pColorAttachments[j].attachment;
-            if (src_a != VK_ATTACHMENT_UNUSED) {
-               pass->attachments[src_a].will_be_resolved =
-                  desc->pResolveAttachments[j].attachment != VK_ATTACHMENT_UNUSED;
+         for (uint32_t j = 0; j < desc->colorAttachmentCount; j++) {
+            uint32_t a = desc->pResolveAttachments[j].attachment;
+            subpass->resolve_attachments[j] = (struct tu_subpass_attachment) {
+               .attachment = desc->pResolveAttachments[j].attachment,
+               .layout = desc->pResolveAttachments[j].layout,
+            };
+            if (a != VK_ATTACHMENT_UNUSED) {
+               subpass->has_resolve = true;
+               pass->attachments[desc->pResolveAttachments[j].attachment]
+                  .view_mask |= subpass->view_mask;
             }
          }
       }
 
-      if (subpass->resolve_depth_stencil) {
-         p++;
-         subpass->resolve_count++;
-         uint32_t a = ds_resolve->pDepthStencilResolveAttachment->attachment;
-         subpass->resolve_attachments[subpass->resolve_count - 1].attachment = a;
-
-         uint32_t src_a = desc->pDepthStencilAttachment->attachment;
-         if (src_a != VK_ATTACHMENT_UNUSED) {
-            pass->attachments[src_a].will_be_resolved = a != VK_ATTACHMENT_UNUSED;
+      if (desc->pDepthStencilAttachment) {
+         subpass->depth_stencil_attachment = (struct tu_subpass_attachment) {
+            .attachment = desc->pDepthStencilAttachment->attachment,
+            .layout = desc->pDepthStencilAttachment->layout,
+         };
+         if (desc->pDepthStencilAttachment->attachment !=
+             VK_ATTACHMENT_UNUSED) {
+            pass->attachments[desc->pDepthStencilAttachment->attachment]
+               .view_mask |= subpass->view_mask;
+            depth_sample_count =
+               pCreateInfo
+                  ->pAttachments[desc->pDepthStencilAttachment->attachment]
+                  .samples;
          }
+      } else {
+         subpass->depth_stencil_attachment.attachment = VK_ATTACHMENT_UNUSED;
       }
 
-      uint32_t a = desc->pDepthStencilAttachment ?
-         desc->pDepthStencilAttachment->attachment : VK_ATTACHMENT_UNUSED;
-      subpass->depth_stencil_attachment.attachment = a;
-      if (a != VK_ATTACHMENT_UNUSED)
-         tu_subpass_use_attachment(pass, i, a, pCreateInfo);
-   }
-
-   tu_render_pass_patch_input_gmem(pass);
-
-   tu_render_pass_check_feedback_loop(pass);
-
-   /* disable unused attachments */
-   for (uint32_t i = 0; i < pass->attachment_count; i++) {
-      struct tu_render_pass_attachment *att = &pass->attachments[i];
-      if (!att->gmem) {
-         att->clear_mask = 0;
-         att->load = false;
-      }
+      subpass->max_sample_count =
+         MAX2(color_sample_count, depth_sample_count);
    }
 
-   tu_render_pass_cond_config(pass);
-   tu_render_pass_gmem_config(pass, device->physical_device);
-   tu_render_pass_bandwidth_config(pass);
-   tu_render_pass_calc_hash(pass);
-
    for (unsigned i = 0; i < pCreateInfo->dependencyCount; ++i) {
-      tu_render_pass_add_subpass_dep(pass, &pCreateInfo->pDependencies[i]);
+      uint32_t dst = pCreateInfo->pDependencies[i].dstSubpass;
+      if (dst == VK_SUBPASS_EXTERNAL) {
+         pass->end_barrier.src_stage_mask =
+            pCreateInfo->pDependencies[i].srcStageMask;
+         pass->end_barrier.src_access_mask =
+            pCreateInfo->pDependencies[i].srcAccessMask;
+         pass->end_barrier.dst_access_mask =
+            pCreateInfo->pDependencies[i].dstAccessMask;
+      } else {
+         pass->subpasses[dst].start_barrier.src_stage_mask =
+            pCreateInfo->pDependencies[i].srcStageMask;
+         pass->subpasses[dst].start_barrier.src_access_mask =
+            pCreateInfo->pDependencies[i].srcAccessMask;
+         pass->subpasses[dst].start_barrier.dst_access_mask =
+            pCreateInfo->pDependencies[i].dstAccessMask;
+      }
    }
 
-   tu_render_pass_add_implicit_deps(pass, pCreateInfo);
-
    *pRenderPass = tu_render_pass_to_handle(pass);
 
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyRenderPass(VkDevice _device,
                      VkRenderPass _pass,
                      const VkAllocationCallbacks *pAllocator)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-
-   if (unlikely(device->instance->debug_flags & TU_DEBUG_DYNAMIC)) {
-      vk_common_DestroyRenderPass(_device, _pass, pAllocator);
-      return;
-   }
-
    TU_FROM_HANDLE(tu_render_pass, pass, _pass);
 
    if (!_pass)
       return;
-
-   vk_free2(&device->vk.alloc, pAllocator, pass->subpass_attachments);
-   vk_object_free(&device->vk, pAllocator, pass);
-}
-
-static void
-tu_setup_dynamic_attachment(struct tu_render_pass_attachment *att,
-                            struct tu_image_view *view)
-{
-   att->format = view->vk.format;
-   att->samples = view->image->layout->nr_samples;
-
-   /* for d32s8, cpp is for the depth image, and
-    * att->samples will be used as the cpp for the stencil image
-    */
-   if (att->format == VK_FORMAT_D32_SFLOAT_S8_UINT)
-      att->cpp = 4 * att->samples;
-   else
-      att->cpp = vk_format_get_blocksize(att->format) * att->samples;
-}
-
-void
-tu_setup_dynamic_render_pass(struct tu_cmd_buffer *cmd_buffer,
-                             const VkRenderingInfo *info)
-{
-   struct tu_device *device = cmd_buffer->device;
-   struct tu_render_pass *pass = &cmd_buffer->dynamic_pass;
-   struct tu_subpass *subpass = &cmd_buffer->dynamic_subpass;
-
-   pass->subpass_count = 1;
-   pass->attachments = cmd_buffer->dynamic_rp_attachments;
-
-   subpass->color_count = subpass->resolve_count = info->colorAttachmentCount;
-   subpass->resolve_depth_stencil = false;
-   subpass->color_attachments = cmd_buffer->dynamic_color_attachments;
-   subpass->resolve_attachments = cmd_buffer->dynamic_resolve_attachments;
-   subpass->feedback_invalidate = false;
-   subpass->feedback_loop_ds = subpass->feedback_loop_color = false;
-   subpass->input_count = 0;
-   subpass->samples = 0;
-   subpass->srgb_cntl = 0;
-   subpass->raster_order_attachment_access = false;
-   subpass->multiview_mask = info->viewMask;
-
-   uint32_t a = 0;
-   for (uint32_t i = 0; i < info->colorAttachmentCount; i++) {
-      struct tu_render_pass_attachment *att = &pass->attachments[a];
-      const VkRenderingAttachmentInfo *att_info = &info->pColorAttachments[i];
-
-      if (att_info->imageView == VK_NULL_HANDLE) {
-         subpass->color_attachments[i].attachment = VK_ATTACHMENT_UNUSED;
-         subpass->resolve_attachments[i].attachment = VK_ATTACHMENT_UNUSED;
-         continue;
-      }
-
-      TU_FROM_HANDLE(tu_image_view, view, att_info->imageView);
-      tu_setup_dynamic_attachment(att, view);
-      att->gmem = true;
-      att->clear_views = info->viewMask;
-      attachment_set_ops(device, att, att_info->loadOp, 0,
-                         att_info->storeOp, 0);
-      subpass->color_attachments[i].attachment = a++;
-
-      subpass->samples = view->image->layout->nr_samples;
-
-      if (vk_format_is_srgb(view->vk.format))
-         subpass->srgb_cntl |= 1 << i;
-
-      if (att_info->resolveMode != VK_RESOLVE_MODE_NONE) {
-         struct tu_render_pass_attachment *resolve_att = &pass->attachments[a];
-         TU_FROM_HANDLE(tu_image_view, resolve_view, att_info->resolveImageView);
-         tu_setup_dynamic_attachment(resolve_att, resolve_view);
-         resolve_att->gmem = false;
-         attachment_set_ops(device, resolve_att,
-                            VK_ATTACHMENT_LOAD_OP_DONT_CARE, 0,
-                            VK_ATTACHMENT_STORE_OP_STORE, 0);
-         subpass->resolve_attachments[i].attachment = a++;
-         att->will_be_resolved = true;
-      } else {
-         subpass->resolve_attachments[i].attachment = VK_ATTACHMENT_UNUSED;
-         att->will_be_resolved = false;
-      }
-   }
-
-   if (info->pDepthAttachment || info->pStencilAttachment) {
-      const struct VkRenderingAttachmentInfo *common_info =
-         (info->pDepthAttachment &&
-          info->pDepthAttachment->imageView != VK_NULL_HANDLE) ?
-         info->pDepthAttachment :
-         info->pStencilAttachment;
-
-      if (common_info && common_info->imageView != VK_NULL_HANDLE) {
-         TU_FROM_HANDLE(tu_image_view, view, common_info->imageView);
-
-         struct tu_render_pass_attachment *att = &pass->attachments[a];
-         tu_setup_dynamic_attachment(att, view);
-         att->gmem = true;
-         att->clear_views = info->viewMask;
-         subpass->depth_stencil_attachment.attachment = a++;
-
-         attachment_set_ops(device, att,
-                            info->pDepthAttachment ? info->pDepthAttachment->loadOp : 0,
-                            info->pStencilAttachment ? info->pStencilAttachment->loadOp : 0,
-                            info->pDepthAttachment ? info->pDepthAttachment->storeOp : 0,
-                            info->pStencilAttachment ? info->pStencilAttachment->storeOp : 0);
-
-         subpass->samples = view->image->layout->nr_samples;
-
-         if (common_info->resolveMode != VK_RESOLVE_MODE_NONE) {
-            unsigned i = subpass->resolve_count++;
-            struct tu_render_pass_attachment *resolve_att = &pass->attachments[a];
-            TU_FROM_HANDLE(tu_image_view, resolve_view,
-                           common_info->resolveImageView);
-            tu_setup_dynamic_attachment(resolve_att, resolve_view);
-            resolve_att->gmem = false;
-            attachment_set_ops(device, resolve_att,
-                               VK_ATTACHMENT_LOAD_OP_DONT_CARE,
-                               VK_ATTACHMENT_LOAD_OP_DONT_CARE,
-                               VK_ATTACHMENT_STORE_OP_STORE,
-                               VK_ATTACHMENT_STORE_OP_STORE);
-            subpass->resolve_attachments[i].attachment = a++;
-            att->will_be_resolved = true;
-            subpass->resolve_depth_stencil = true;
-         } else {
-            att->will_be_resolved = false;
-         }
-      } else {
-         subpass->depth_stencil_attachment.attachment = VK_ATTACHMENT_UNUSED;
-      }
-   } else {
-      subpass->depth_stencil_attachment.attachment = VK_ATTACHMENT_UNUSED;
-   }
-
-   pass->attachment_count = a;
-
-   tu_render_pass_cond_config(pass);
-   tu_render_pass_gmem_config(pass, device->physical_device);
-   tu_render_pass_bandwidth_config(pass);
-   tu_render_pass_calc_hash(pass);
+   vk_free2(&device->alloc, pAllocator, pass->subpass_attachments);
+   vk_free2(&device->alloc, pAllocator, pass);
 }
 
 void
-tu_setup_dynamic_inheritance(struct tu_cmd_buffer *cmd_buffer,
-                             const VkCommandBufferInheritanceRenderingInfo *info)
-{
-   struct tu_render_pass *pass = &cmd_buffer->dynamic_pass;
-   struct tu_subpass *subpass = &cmd_buffer->dynamic_subpass;
-
-   pass->subpass_count = 1;
-   pass->attachments = cmd_buffer->dynamic_rp_attachments;
-
-   subpass->color_count = info->colorAttachmentCount;
-   subpass->resolve_count = 0;
-   subpass->resolve_depth_stencil = false;
-   subpass->color_attachments = cmd_buffer->dynamic_color_attachments;
-   subpass->resolve_attachments = NULL;
-   subpass->feedback_invalidate = false;
-   subpass->feedback_loop_ds = subpass->feedback_loop_color = false;
-   subpass->input_count = 0;
-   subpass->samples = 0;
-   subpass->srgb_cntl = 0;
-   subpass->raster_order_attachment_access = false;
-   subpass->multiview_mask = info->viewMask;
-   subpass->samples = info->rasterizationSamples;
-
-   unsigned a = 0;
-   for (unsigned i = 0; i < info->colorAttachmentCount; i++) {
-      struct tu_render_pass_attachment *att = &pass->attachments[a];
-      VkFormat format = info->pColorAttachmentFormats[i];
-
-      if (format == VK_FORMAT_UNDEFINED) {
-         subpass->color_attachments[i].attachment = VK_ATTACHMENT_UNUSED;
-         continue;
-      }
-
-      att->format = format;
-      att->samples = info->rasterizationSamples;
-      subpass->samples = info->rasterizationSamples;
-      subpass->color_attachments[i].attachment = a++;
-
-      /* conservatively assume that the attachment may be conditionally
-       * loaded/stored.
-       */
-      att->cond_load_allowed = att->cond_store_allowed = true;
-   }
-
-   if (info->depthAttachmentFormat != VK_FORMAT_UNDEFINED ||
-       info->stencilAttachmentFormat != VK_FORMAT_UNDEFINED) {
-      struct tu_render_pass_attachment *att = &pass->attachments[a];
-      att->format = info->depthAttachmentFormat != VK_FORMAT_UNDEFINED ?
-         info->depthAttachmentFormat : info->stencilAttachmentFormat;
-      att->samples = info->rasterizationSamples;
-      subpass->depth_stencil_attachment.attachment = a++;
-      att->cond_load_allowed = att->cond_store_allowed = true;
-   } else {
-      subpass->depth_stencil_attachment.attachment = VK_ATTACHMENT_UNUSED;
-   }
-}
-
-VKAPI_ATTR void VKAPI_CALL
 tu_GetRenderAreaGranularity(VkDevice _device,
                             VkRenderPass renderPass,
                             VkExtent2D *pGranularity)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-   pGranularity->width = device->physical_device->info->gmem_align_w;
-   pGranularity->height = device->physical_device->info->gmem_align_h;
-}
-
-uint32_t
-tu_subpass_get_attachment_to_resolve(const struct tu_subpass *subpass, uint32_t index)
-{
-   if (subpass->resolve_depth_stencil &&
-       index == (subpass->resolve_count - 1))
-      return subpass->depth_stencil_attachment.attachment;
 
-   return subpass->color_attachments[index].attachment;
+   pGranularity->width = device->physical_device->tile_align_w;
+   pGranularity->height = device->physical_device->tile_align_h;
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_pipeline.c b/lib/mesa/src/freedreno/vulkan/tu_pipeline.c
index d4d3c9735..9964020a8 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_pipeline.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_pipeline.c
@@ -1,297 +1,140 @@
 /*
  * Copyright © 2016 Red Hat.
  * Copyright © 2016 Bas Nieuwenhuizen
- * SPDX-License-Identifier: MIT
  *
  * based in part on anv driver which is:
  * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_pipeline.h"
+#include "tu_private.h"
 
-#include "common/freedreno_guardband.h"
-
-#include "ir3/ir3_nir.h"
 #include "main/menums.h"
 #include "nir/nir.h"
 #include "nir/nir_builder.h"
-#include "nir/nir_serialize.h"
 #include "spirv/nir_spirv.h"
-#include "util/u_debug.h"
+#include "util/debug.h"
 #include "util/mesa-sha1.h"
-#include "vk_pipeline.h"
-#include "vk_render_pass.h"
+#include "util/u_atomic.h"
+#include "vk_format.h"
 #include "vk_util.h"
 
-#include "tu_cmd_buffer.h"
 #include "tu_cs.h"
-#include "tu_device.h"
-#include "tu_drm.h"
-#include "tu_formats.h"
-#include "tu_lrz.h"
-#include "tu_pass.h"
-
-/* Emit IB that preloads the descriptors that the shader uses */
-
-static void
-emit_load_state(struct tu_cs *cs, unsigned opcode, enum a6xx_state_type st,
-                enum a6xx_state_block sb, unsigned base, unsigned offset,
-                unsigned count)
-{
-   /* Note: just emit one packet, even if count overflows NUM_UNIT. It's not
-    * clear if emitting more packets will even help anything. Presumably the
-    * descriptor cache is relatively small, and these packets stop doing
-    * anything when there are too many descriptors.
-    */
-   tu_cs_emit_pkt7(cs, opcode, 3);
-   tu_cs_emit(cs,
-              CP_LOAD_STATE6_0_STATE_TYPE(st) |
-              CP_LOAD_STATE6_0_STATE_SRC(SS6_BINDLESS) |
-              CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
-              CP_LOAD_STATE6_0_NUM_UNIT(MIN2(count, 1024-1)));
-   tu_cs_emit_qw(cs, offset | (base << 28));
-}
-
-static unsigned
-tu6_load_state_size(struct tu_pipeline *pipeline,
-                    struct tu_pipeline_layout *layout)
-{
-   const unsigned load_state_size = 4;
-   unsigned size = 0;
-   for (unsigned i = 0; i < layout->num_sets; i++) {
-      if (!(pipeline->active_desc_sets & (1u << i)))
-         continue;
-
-      struct tu_descriptor_set_layout *set_layout = layout->set[i].layout;
-      for (unsigned j = 0; j < set_layout->binding_count; j++) {
-         struct tu_descriptor_set_binding_layout *binding = &set_layout->binding[j];
-         unsigned count = 0;
-         /* See comment in tu6_emit_load_state(). */
-         VkShaderStageFlags stages = pipeline->active_stages & binding->shader_stages;
-         unsigned stage_count = util_bitcount(stages);
-
-         if (!binding->array_size)
-            continue;
-
-         switch (binding->type) {
-         case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
-         case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
-         case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
-         case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
-            /* IBO-backed resources only need one packet for all graphics stages */
-            if (stage_count)
-               count += 1;
-            break;
-         case VK_DESCRIPTOR_TYPE_SAMPLER:
-         case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
-         case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
-         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
-         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
-         case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT:
-            /* Textures and UBO's needs a packet for each stage */
-            count = stage_count;
-            break;
-         case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
-            /* Because of how we pack combined images and samplers, we
-             * currently can't use one packet for the whole array.
-             */
-            count = stage_count * binding->array_size * 2;
-            break;
-         case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
-         case VK_DESCRIPTOR_TYPE_MUTABLE_EXT:
-            break;
-         default:
-            unreachable("bad descriptor type");
-         }
-         size += count * load_state_size;
-      }
-   }
-   return size;
-}
-
-static void
-tu6_emit_load_state(struct tu_pipeline *pipeline,
-                    struct tu_pipeline_layout *layout)
-{
-   unsigned size = tu6_load_state_size(pipeline, layout);
-   if (size == 0)
-      return;
-
-   struct tu_cs cs;
-   tu_cs_begin_sub_stream(&pipeline->cs, size, &cs);
-
-   for (unsigned i = 0; i < layout->num_sets; i++) {
-      /* From 13.2.7. Descriptor Set Binding:
-       *
-       *    A compatible descriptor set must be bound for all set numbers that
-       *    any shaders in a pipeline access, at the time that a draw or
-       *    dispatch command is recorded to execute using that pipeline.
-       *    However, if none of the shaders in a pipeline statically use any
-       *    bindings with a particular set number, then no descriptor set need
-       *    be bound for that set number, even if the pipeline layout includes
-       *    a non-trivial descriptor set layout for that set number.
-       *
-       * This means that descriptor sets unused by the pipeline may have a
-       * garbage or 0 BINDLESS_BASE register, which will cause context faults
-       * when prefetching descriptors from these sets. Skip prefetching for
-       * descriptors from them to avoid this. This is also an optimization,
-       * since these prefetches would be useless.
-       */
-      if (!(pipeline->active_desc_sets & (1u << i)))
-         continue;
-
-      struct tu_descriptor_set_layout *set_layout = layout->set[i].layout;
-      for (unsigned j = 0; j < set_layout->binding_count; j++) {
-         struct tu_descriptor_set_binding_layout *binding = &set_layout->binding[j];
-         unsigned base = i;
-         unsigned offset = binding->offset / 4;
-         /* Note: amber sets VK_SHADER_STAGE_ALL for its descriptor layout, and
-          * zink has descriptors for each stage in the push layout even if some
-          * stages aren't present in a used pipeline.  We don't want to emit
-          * loads for unused descriptors.
-          */
-         VkShaderStageFlags stages = pipeline->active_stages & binding->shader_stages;
-         unsigned count = binding->array_size;
-
-         /* If this is a variable-count descriptor, then the array_size is an
-          * upper bound on the size, but we don't know how many descriptors
-          * will actually be used. Therefore we can't pre-load them here.
-          */
-         if (j == set_layout->binding_count - 1 &&
-             set_layout->has_variable_descriptors)
-            continue;
-
-         if (count == 0 || stages == 0)
-            continue;
-         switch (binding->type) {
-         case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
-            base = MAX_SETS;
-            offset = (layout->set[i].dynamic_offset_start +
-                      binding->dynamic_offset_offset) / 4;
-            FALLTHROUGH;
-         case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
-         case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
-         case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER: {
-            unsigned mul = binding->size / (A6XX_TEX_CONST_DWORDS * 4);
-            /* IBO-backed resources only need one packet for all graphics stages */
-            if (stages & ~VK_SHADER_STAGE_COMPUTE_BIT) {
-               emit_load_state(&cs, CP_LOAD_STATE6, ST6_SHADER, SB6_IBO,
-                               base, offset, count * mul);
-            }
-            if (stages & VK_SHADER_STAGE_COMPUTE_BIT) {
-               emit_load_state(&cs, CP_LOAD_STATE6_FRAG, ST6_IBO, SB6_CS_SHADER,
-                               base, offset, count * mul);
-            }
-            break;
-         }
-         case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
-         case VK_DESCRIPTOR_TYPE_MUTABLE_EXT:
-            /* nothing - input attachment doesn't use bindless */
-            break;
-         case VK_DESCRIPTOR_TYPE_SAMPLER:
-         case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
-         case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER: {
-            tu_foreach_stage(stage, stages) {
-               emit_load_state(&cs, tu6_stage2opcode(stage),
-                               binding->type == VK_DESCRIPTOR_TYPE_SAMPLER ?
-                               ST6_SHADER : ST6_CONSTANTS,
-                               tu6_stage2texsb(stage), base, offset, count);
-            }
-            break;
-         }
-         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
-            base = MAX_SETS;
-            offset = (layout->set[i].dynamic_offset_start +
-                      binding->dynamic_offset_offset) / 4;
-            FALLTHROUGH;
-         case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
-         case VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK_EXT: {
-            tu_foreach_stage(stage, stages) {
-               emit_load_state(&cs, tu6_stage2opcode(stage), ST6_UBO,
-                               tu6_stage2shadersb(stage), base, offset, count);
-            }
-            break;
-         }
-         case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER: {
-            tu_foreach_stage(stage, stages) {
-               /* TODO: We could emit less CP_LOAD_STATE6 if we used
-                * struct-of-arrays instead of array-of-structs.
-                */
-               for (unsigned i = 0; i < count; i++) {
-                  unsigned tex_offset = offset + 2 * i * A6XX_TEX_CONST_DWORDS;
-                  unsigned sam_offset = offset + (2 * i + 1) * A6XX_TEX_CONST_DWORDS;
-                  emit_load_state(&cs, tu6_stage2opcode(stage),
-                                  ST6_CONSTANTS, tu6_stage2texsb(stage),
-                                  base, tex_offset, 1);
-                  emit_load_state(&cs, tu6_stage2opcode(stage),
-                                  ST6_SHADER, tu6_stage2texsb(stage),
-                                  base, sam_offset, 1);
-               }
-            }
-            break;
-         }
-         default:
-            unreachable("bad descriptor type");
-         }
-      }
-   }
-
-   pipeline->load_state = tu_cs_end_draw_state(&pipeline->cs, &cs);
-}
 
 struct tu_pipeline_builder
 {
    struct tu_device *device;
-   void *mem_ctx;
-   struct vk_pipeline_cache *cache;
+   struct tu_pipeline_cache *cache;
    const VkAllocationCallbacks *alloc;
    const VkGraphicsPipelineCreateInfo *create_info;
 
-   struct tu_pipeline_layout layout;
-
-   struct tu_compiled_shaders *compiled_shaders;
-
-   struct tu_const_state const_state[MESA_SHADER_FRAGMENT + 1];
-   struct ir3_shader_variant *variants[MESA_SHADER_FRAGMENT + 1];
-   struct ir3_shader_variant *binning_variant;
-   uint64_t shader_iova[MESA_SHADER_FRAGMENT + 1];
-   uint64_t binning_vs_iova;
-
-   uint32_t additional_cs_reserve_size;
-
-   struct tu_pvtmem_config pvtmem;
+   struct tu_shader *shaders[MESA_SHADER_STAGES];
+   uint32_t shader_offsets[MESA_SHADER_STAGES];
+   uint32_t binning_vs_offset;
+   uint32_t shader_total_size;
 
    bool rasterizer_discard;
    /* these states are affectd by rasterizer_discard */
+   VkSampleCountFlagBits samples;
+   bool use_depth_stencil_attachment;
    bool use_color_attachments;
-   bool attachment_state_valid;
+   uint32_t color_attachment_count;
    VkFormat color_attachment_formats[MAX_RTS];
-   VkFormat depth_attachment_format;
-   uint32_t multiview_mask;
-
-   bool subpass_raster_order_attachment_access;
-   bool subpass_feedback_loop_color;
-   bool subpass_feedback_loop_ds;
-   bool feedback_loop_may_involve_textures;
-
-   /* Each library defines at least one piece of state in
-    * VkGraphicsPipelineLibraryFlagsEXT, and libraries cannot overlap, so
-    * there can be at most as many libraries as pieces of state, of which
-    * there are currently 4.
-    */
-#define MAX_LIBRARIES 4
+};
 
-   unsigned num_libraries;
-   struct tu_pipeline *libraries[MAX_LIBRARIES];
+static enum tu_dynamic_state_bits
+tu_dynamic_state_bit(VkDynamicState state)
+{
+   switch (state) {
+   case VK_DYNAMIC_STATE_VIEWPORT:
+      return TU_DYNAMIC_VIEWPORT;
+   case VK_DYNAMIC_STATE_SCISSOR:
+      return TU_DYNAMIC_SCISSOR;
+   case VK_DYNAMIC_STATE_LINE_WIDTH:
+      return TU_DYNAMIC_LINE_WIDTH;
+   case VK_DYNAMIC_STATE_DEPTH_BIAS:
+      return TU_DYNAMIC_DEPTH_BIAS;
+   case VK_DYNAMIC_STATE_BLEND_CONSTANTS:
+      return TU_DYNAMIC_BLEND_CONSTANTS;
+   case VK_DYNAMIC_STATE_DEPTH_BOUNDS:
+      return TU_DYNAMIC_DEPTH_BOUNDS;
+   case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK:
+      return TU_DYNAMIC_STENCIL_COMPARE_MASK;
+   case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK:
+      return TU_DYNAMIC_STENCIL_WRITE_MASK;
+   case VK_DYNAMIC_STATE_STENCIL_REFERENCE:
+      return TU_DYNAMIC_STENCIL_REFERENCE;
+   default:
+      unreachable("invalid dynamic state");
+      return 0;
+   }
+}
 
-   /* This is just the state that we are compiling now, whereas the final
-    * pipeline will include the state from the libraries.
-    */
-   VkGraphicsPipelineLibraryFlagsEXT state;
+static gl_shader_stage
+tu_shader_stage(VkShaderStageFlagBits stage)
+{
+   switch (stage) {
+   case VK_SHADER_STAGE_VERTEX_BIT:
+      return MESA_SHADER_VERTEX;
+   case VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT:
+      return MESA_SHADER_TESS_CTRL;
+   case VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT:
+      return MESA_SHADER_TESS_EVAL;
+   case VK_SHADER_STAGE_GEOMETRY_BIT:
+      return MESA_SHADER_GEOMETRY;
+   case VK_SHADER_STAGE_FRAGMENT_BIT:
+      return MESA_SHADER_FRAGMENT;
+   case VK_SHADER_STAGE_COMPUTE_BIT:
+      return MESA_SHADER_COMPUTE;
+   default:
+      unreachable("invalid VkShaderStageFlagBits");
+      return MESA_SHADER_NONE;
+   }
+}
 
-   /* The stages we are compiling now. */
-   VkShaderStageFlags active_stages;
-};
+static const VkVertexInputAttributeDescription *
+tu_find_vertex_input_attribute(
+   const VkPipelineVertexInputStateCreateInfo *vi_info, uint32_t slot)
+{
+   assert(slot >= VERT_ATTRIB_GENERIC0);
+   slot -= VERT_ATTRIB_GENERIC0;
+   for (uint32_t i = 0; i < vi_info->vertexAttributeDescriptionCount; i++) {
+      if (vi_info->pVertexAttributeDescriptions[i].location == slot)
+         return &vi_info->pVertexAttributeDescriptions[i];
+   }
+   return NULL;
+}
+
+static const VkVertexInputBindingDescription *
+tu_find_vertex_input_binding(
+   const VkPipelineVertexInputStateCreateInfo *vi_info,
+   const VkVertexInputAttributeDescription *vi_attr)
+{
+   assert(vi_attr);
+   for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
+      if (vi_info->pVertexBindingDescriptions[i].binding == vi_attr->binding)
+         return &vi_info->pVertexBindingDescriptions[i];
+   }
+   return NULL;
+}
 
 static bool
 tu_logic_op_reads_dst(VkLogicOp op)
@@ -321,732 +164,418 @@ tu_blend_factor_no_dst_alpha(VkBlendFactor factor)
    }
 }
 
-static bool tu_blend_factor_is_dual_src(VkBlendFactor factor)
-{
-   switch (factor) {
-   case VK_BLEND_FACTOR_SRC1_COLOR:
-   case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
-   case VK_BLEND_FACTOR_SRC1_ALPHA:
-   case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
-      return true;
+static enum pc_di_primtype
+tu6_primtype(VkPrimitiveTopology topology)
+{
+   switch (topology) {
+   case VK_PRIMITIVE_TOPOLOGY_POINT_LIST:
+      return DI_PT_POINTLIST;
+   case VK_PRIMITIVE_TOPOLOGY_LINE_LIST:
+      return DI_PT_LINELIST;
+   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP:
+      return DI_PT_LINESTRIP;
+   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST:
+      return DI_PT_TRILIST;
+   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP:
+      return DI_PT_TRILIST;
+   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN:
+      return DI_PT_TRIFAN;
+   case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY:
+      return DI_PT_LINE_ADJ;
+   case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY:
+      return DI_PT_LINESTRIP_ADJ;
+   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY:
+      return DI_PT_TRI_ADJ;
+   case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY:
+      return DI_PT_TRISTRIP_ADJ;
+   case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST:
    default:
-      return false;
+      unreachable("invalid primitive topology");
+      return DI_PT_NONE;
    }
 }
 
-static bool
-tu_blend_state_is_dual_src(const VkPipelineColorBlendStateCreateInfo *info)
+static enum adreno_compare_func
+tu6_compare_func(VkCompareOp op)
 {
-   if (!info)
-      return false;
-
-   for (unsigned i = 0; i < info->attachmentCount; i++) {
-      const VkPipelineColorBlendAttachmentState *blend = &info->pAttachments[i];
-      if (tu_blend_factor_is_dual_src(blend->srcColorBlendFactor) ||
-          tu_blend_factor_is_dual_src(blend->dstColorBlendFactor) ||
-          tu_blend_factor_is_dual_src(blend->srcAlphaBlendFactor) ||
-          tu_blend_factor_is_dual_src(blend->dstAlphaBlendFactor))
-         return true;
+   switch (op) {
+   case VK_COMPARE_OP_NEVER:
+      return FUNC_NEVER;
+   case VK_COMPARE_OP_LESS:
+      return FUNC_LESS;
+   case VK_COMPARE_OP_EQUAL:
+      return FUNC_EQUAL;
+   case VK_COMPARE_OP_LESS_OR_EQUAL:
+      return FUNC_LEQUAL;
+   case VK_COMPARE_OP_GREATER:
+      return FUNC_GREATER;
+   case VK_COMPARE_OP_NOT_EQUAL:
+      return FUNC_NOTEQUAL;
+   case VK_COMPARE_OP_GREATER_OR_EQUAL:
+      return FUNC_GEQUAL;
+   case VK_COMPARE_OP_ALWAYS:
+      return FUNC_ALWAYS;
+   default:
+      unreachable("invalid VkCompareOp");
+      return FUNC_NEVER;
    }
-
-   return false;
 }
 
-static const struct xs_config {
-   uint16_t reg_sp_xs_ctrl;
-   uint16_t reg_sp_xs_config;
-   uint16_t reg_sp_xs_instrlen;
-   uint16_t reg_hlsq_xs_ctrl;
-   uint16_t reg_sp_xs_first_exec_offset;
-   uint16_t reg_sp_xs_pvt_mem_hw_stack_offset;
-} xs_config[] = {
-   [MESA_SHADER_VERTEX] = {
-      REG_A6XX_SP_VS_CTRL_REG0,
-      REG_A6XX_SP_VS_CONFIG,
-      REG_A6XX_SP_VS_INSTRLEN,
-      REG_A6XX_HLSQ_VS_CNTL,
-      REG_A6XX_SP_VS_OBJ_FIRST_EXEC_OFFSET,
-      REG_A6XX_SP_VS_PVT_MEM_HW_STACK_OFFSET,
-   },
-   [MESA_SHADER_TESS_CTRL] = {
-      REG_A6XX_SP_HS_CTRL_REG0,
-      REG_A6XX_SP_HS_CONFIG,
-      REG_A6XX_SP_HS_INSTRLEN,
-      REG_A6XX_HLSQ_HS_CNTL,
-      REG_A6XX_SP_HS_OBJ_FIRST_EXEC_OFFSET,
-      REG_A6XX_SP_HS_PVT_MEM_HW_STACK_OFFSET,
-   },
-   [MESA_SHADER_TESS_EVAL] = {
-      REG_A6XX_SP_DS_CTRL_REG0,
-      REG_A6XX_SP_DS_CONFIG,
-      REG_A6XX_SP_DS_INSTRLEN,
-      REG_A6XX_HLSQ_DS_CNTL,
-      REG_A6XX_SP_DS_OBJ_FIRST_EXEC_OFFSET,
-      REG_A6XX_SP_DS_PVT_MEM_HW_STACK_OFFSET,
-   },
-   [MESA_SHADER_GEOMETRY] = {
-      REG_A6XX_SP_GS_CTRL_REG0,
-      REG_A6XX_SP_GS_CONFIG,
-      REG_A6XX_SP_GS_INSTRLEN,
-      REG_A6XX_HLSQ_GS_CNTL,
-      REG_A6XX_SP_GS_OBJ_FIRST_EXEC_OFFSET,
-      REG_A6XX_SP_GS_PVT_MEM_HW_STACK_OFFSET,
-   },
-   [MESA_SHADER_FRAGMENT] = {
-      REG_A6XX_SP_FS_CTRL_REG0,
-      REG_A6XX_SP_FS_CONFIG,
-      REG_A6XX_SP_FS_INSTRLEN,
-      REG_A6XX_HLSQ_FS_CNTL,
-      REG_A6XX_SP_FS_OBJ_FIRST_EXEC_OFFSET,
-      REG_A6XX_SP_FS_PVT_MEM_HW_STACK_OFFSET,
-   },
-   [MESA_SHADER_COMPUTE] = {
-      REG_A6XX_SP_CS_CTRL_REG0,
-      REG_A6XX_SP_CS_CONFIG,
-      REG_A6XX_SP_CS_INSTRLEN,
-      REG_A6XX_HLSQ_CS_CNTL,
-      REG_A6XX_SP_CS_OBJ_FIRST_EXEC_OFFSET,
-      REG_A6XX_SP_CS_PVT_MEM_HW_STACK_OFFSET,
-   },
-};
-
-static uint32_t
-tu_xs_get_immediates_packet_size_dwords(const struct ir3_shader_variant *xs)
+static enum adreno_stencil_op
+tu6_stencil_op(VkStencilOp op)
 {
-   const struct ir3_const_state *const_state = ir3_const_state(xs);
-   uint32_t base = const_state->offsets.immediate;
-   int32_t size = DIV_ROUND_UP(const_state->immediates_count, 4);
-
-   /* truncate size to avoid writing constants that shader
-    * does not use:
-    */
-   size = MIN2(size + base, xs->constlen) - base;
-
-   return MAX2(size, 0) * 4;
+   switch (op) {
+   case VK_STENCIL_OP_KEEP:
+      return STENCIL_KEEP;
+   case VK_STENCIL_OP_ZERO:
+      return STENCIL_ZERO;
+   case VK_STENCIL_OP_REPLACE:
+      return STENCIL_REPLACE;
+   case VK_STENCIL_OP_INCREMENT_AND_CLAMP:
+      return STENCIL_INCR_CLAMP;
+   case VK_STENCIL_OP_DECREMENT_AND_CLAMP:
+      return STENCIL_DECR_CLAMP;
+   case VK_STENCIL_OP_INVERT:
+      return STENCIL_INVERT;
+   case VK_STENCIL_OP_INCREMENT_AND_WRAP:
+      return STENCIL_INCR_WRAP;
+   case VK_STENCIL_OP_DECREMENT_AND_WRAP:
+      return STENCIL_DECR_WRAP;
+   default:
+      unreachable("invalid VkStencilOp");
+      return STENCIL_KEEP;
+   }
 }
 
-/* We allocate fixed-length substreams for shader state, however some
- * parts of the state may have unbound length. Their additional space
- * requirements should be calculated here.
- */
-static uint32_t
-tu_xs_get_additional_cs_size_dwords(const struct ir3_shader_variant *xs)
+static enum a3xx_rop_code
+tu6_rop(VkLogicOp op)
 {
-   const struct ir3_const_state *const_state = ir3_const_state(xs);
-
-   uint32_t size = tu_xs_get_immediates_packet_size_dwords(xs);
-
-   /* Variable number of UBO upload ranges. */
-   size += 4 * const_state->ubo_state.num_enabled;
-
-   /* Variable number of dwords for the primitive map */
-   size += xs->input_size;
-
-   size += xs->constant_data_size / 4;
-
-   return size;
+   switch (op) {
+   case VK_LOGIC_OP_CLEAR:
+      return ROP_CLEAR;
+   case VK_LOGIC_OP_AND:
+      return ROP_AND;
+   case VK_LOGIC_OP_AND_REVERSE:
+      return ROP_AND_REVERSE;
+   case VK_LOGIC_OP_COPY:
+      return ROP_COPY;
+   case VK_LOGIC_OP_AND_INVERTED:
+      return ROP_AND_INVERTED;
+   case VK_LOGIC_OP_NO_OP:
+      return ROP_NOOP;
+   case VK_LOGIC_OP_XOR:
+      return ROP_XOR;
+   case VK_LOGIC_OP_OR:
+      return ROP_OR;
+   case VK_LOGIC_OP_NOR:
+      return ROP_NOR;
+   case VK_LOGIC_OP_EQUIVALENT:
+      return ROP_EQUIV;
+   case VK_LOGIC_OP_INVERT:
+      return ROP_INVERT;
+   case VK_LOGIC_OP_OR_REVERSE:
+      return ROP_OR_REVERSE;
+   case VK_LOGIC_OP_COPY_INVERTED:
+      return ROP_COPY_INVERTED;
+   case VK_LOGIC_OP_OR_INVERTED:
+      return ROP_OR_INVERTED;
+   case VK_LOGIC_OP_NAND:
+      return ROP_NAND;
+   case VK_LOGIC_OP_SET:
+      return ROP_SET;
+   default:
+      unreachable("invalid VkLogicOp");
+      return ROP_NOOP;
+   }
 }
 
-void
-tu6_emit_xs_config(struct tu_cs *cs,
-                   gl_shader_stage stage, /* xs->type, but xs may be NULL */
-                   const struct ir3_shader_variant *xs)
+static enum adreno_rb_blend_factor
+tu6_blend_factor(VkBlendFactor factor)
 {
-   const struct xs_config *cfg = &xs_config[stage];
-
-   if (!xs) {
-      /* shader stage disabled */
-      tu_cs_emit_pkt4(cs, cfg->reg_sp_xs_config, 1);
-      tu_cs_emit(cs, 0);
-
-      tu_cs_emit_pkt4(cs, cfg->reg_hlsq_xs_ctrl, 1);
-      tu_cs_emit(cs, 0);
-      return;
+   switch (factor) {
+   case VK_BLEND_FACTOR_ZERO:
+      return FACTOR_ZERO;
+   case VK_BLEND_FACTOR_ONE:
+      return FACTOR_ONE;
+   case VK_BLEND_FACTOR_SRC_COLOR:
+      return FACTOR_SRC_COLOR;
+   case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR:
+      return FACTOR_ONE_MINUS_SRC_COLOR;
+   case VK_BLEND_FACTOR_DST_COLOR:
+      return FACTOR_DST_COLOR;
+   case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR:
+      return FACTOR_ONE_MINUS_DST_COLOR;
+   case VK_BLEND_FACTOR_SRC_ALPHA:
+      return FACTOR_SRC_ALPHA;
+   case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA:
+      return FACTOR_ONE_MINUS_SRC_ALPHA;
+   case VK_BLEND_FACTOR_DST_ALPHA:
+      return FACTOR_DST_ALPHA;
+   case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA:
+      return FACTOR_ONE_MINUS_DST_ALPHA;
+   case VK_BLEND_FACTOR_CONSTANT_COLOR:
+      return FACTOR_CONSTANT_COLOR;
+   case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR:
+      return FACTOR_ONE_MINUS_CONSTANT_COLOR;
+   case VK_BLEND_FACTOR_CONSTANT_ALPHA:
+      return FACTOR_CONSTANT_ALPHA;
+   case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA:
+      return FACTOR_ONE_MINUS_CONSTANT_ALPHA;
+   case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE:
+      return FACTOR_SRC_ALPHA_SATURATE;
+   case VK_BLEND_FACTOR_SRC1_COLOR:
+      return FACTOR_SRC1_COLOR;
+   case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR:
+      return FACTOR_ONE_MINUS_SRC1_COLOR;
+   case VK_BLEND_FACTOR_SRC1_ALPHA:
+      return FACTOR_SRC1_ALPHA;
+   case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA:
+      return FACTOR_ONE_MINUS_SRC1_ALPHA;
+   default:
+      unreachable("invalid VkBlendFactor");
+      return FACTOR_ZERO;
    }
-
-   tu_cs_emit_pkt4(cs, cfg->reg_sp_xs_config, 1);
-   tu_cs_emit(cs, A6XX_SP_VS_CONFIG_ENABLED |
-                  COND(xs->bindless_tex, A6XX_SP_VS_CONFIG_BINDLESS_TEX) |
-                  COND(xs->bindless_samp, A6XX_SP_VS_CONFIG_BINDLESS_SAMP) |
-                  COND(xs->bindless_ibo, A6XX_SP_VS_CONFIG_BINDLESS_IBO) |
-                  COND(xs->bindless_ubo, A6XX_SP_VS_CONFIG_BINDLESS_UBO) |
-                  A6XX_SP_VS_CONFIG_NTEX(xs->num_samp) |
-                  A6XX_SP_VS_CONFIG_NSAMP(xs->num_samp));
-
-   tu_cs_emit_pkt4(cs, cfg->reg_hlsq_xs_ctrl, 1);
-   tu_cs_emit(cs, A6XX_HLSQ_VS_CNTL_CONSTLEN(xs->constlen) |
-                  A6XX_HLSQ_VS_CNTL_ENABLED);
 }
 
-void
-tu6_emit_xs(struct tu_cs *cs,
-            gl_shader_stage stage, /* xs->type, but xs may be NULL */
-            const struct ir3_shader_variant *xs,
-            const struct tu_pvtmem_config *pvtmem,
-            uint64_t binary_iova)
+static enum a3xx_rb_blend_opcode
+tu6_blend_op(VkBlendOp op)
 {
-   const struct xs_config *cfg = &xs_config[stage];
-
-   if (!xs) {
-      /* shader stage disabled */
-      return;
-   }
-
-   enum a6xx_threadsize thrsz =
-      xs->info.double_threadsize ? THREAD128 : THREAD64;
-   switch (stage) {
-   case MESA_SHADER_VERTEX:
-      tu_cs_emit_regs(cs, A6XX_SP_VS_CTRL_REG0(
-               .fullregfootprint = xs->info.max_reg + 1,
-               .halfregfootprint = xs->info.max_half_reg + 1,
-               .branchstack = ir3_shader_branchstack_hw(xs),
-               .mergedregs = xs->mergedregs,
-      ));
-      break;
-   case MESA_SHADER_TESS_CTRL:
-      tu_cs_emit_regs(cs, A6XX_SP_HS_CTRL_REG0(
-               .fullregfootprint = xs->info.max_reg + 1,
-               .halfregfootprint = xs->info.max_half_reg + 1,
-               .branchstack = ir3_shader_branchstack_hw(xs),
-      ));
-      break;
-   case MESA_SHADER_TESS_EVAL:
-      tu_cs_emit_regs(cs, A6XX_SP_DS_CTRL_REG0(
-               .fullregfootprint = xs->info.max_reg + 1,
-               .halfregfootprint = xs->info.max_half_reg + 1,
-               .branchstack = ir3_shader_branchstack_hw(xs),
-      ));
-      break;
-   case MESA_SHADER_GEOMETRY:
-      tu_cs_emit_regs(cs, A6XX_SP_GS_CTRL_REG0(
-               .fullregfootprint = xs->info.max_reg + 1,
-               .halfregfootprint = xs->info.max_half_reg + 1,
-               .branchstack = ir3_shader_branchstack_hw(xs),
-      ));
-      break;
-   case MESA_SHADER_FRAGMENT:
-      tu_cs_emit_regs(cs, A6XX_SP_FS_CTRL_REG0(
-               .fullregfootprint = xs->info.max_reg + 1,
-               .halfregfootprint = xs->info.max_half_reg + 1,
-               .branchstack = ir3_shader_branchstack_hw(xs),
-               .mergedregs = xs->mergedregs,
-               .threadsize = thrsz,
-               .pixlodenable = xs->need_pixlod,
-               .diff_fine = xs->need_fine_derivatives,
-               .varying = xs->total_in != 0,
-               /* unknown bit, seems unnecessary */
-               .unk24 = true,
-      ));
-      break;
-   case MESA_SHADER_COMPUTE:
-      tu_cs_emit_regs(cs, A6XX_SP_CS_CTRL_REG0(
-               .fullregfootprint = xs->info.max_reg + 1,
-               .halfregfootprint = xs->info.max_half_reg + 1,
-               .branchstack = ir3_shader_branchstack_hw(xs),
-               .mergedregs = xs->mergedregs,
-               .threadsize = thrsz,
-      ));
-      break;
+   switch (op) {
+   case VK_BLEND_OP_ADD:
+      return BLEND_DST_PLUS_SRC;
+   case VK_BLEND_OP_SUBTRACT:
+      return BLEND_SRC_MINUS_DST;
+   case VK_BLEND_OP_REVERSE_SUBTRACT:
+      return BLEND_DST_MINUS_SRC;
+   case VK_BLEND_OP_MIN:
+      return BLEND_MIN_DST_SRC;
+   case VK_BLEND_OP_MAX:
+      return BLEND_MAX_DST_SRC;
    default:
-      unreachable("bad shader stage");
+      unreachable("invalid VkBlendOp");
+      return BLEND_DST_PLUS_SRC;
    }
+}
 
-   tu_cs_emit_pkt4(cs, cfg->reg_sp_xs_instrlen, 1);
-   tu_cs_emit(cs, xs->instrlen);
-
-   /* emit program binary & private memory layout
-    * binary_iova should be aligned to 1 instrlen unit (128 bytes)
-    */
+static void
+tu6_emit_vs_config(struct tu_cs *cs, const struct ir3_shader_variant *vs)
+{
+   uint32_t sp_vs_ctrl =
+      A6XX_SP_VS_CTRL_REG0_THREADSIZE(FOUR_QUADS) |
+      A6XX_SP_VS_CTRL_REG0_FULLREGFOOTPRINT(vs->info.max_reg + 1) |
+      A6XX_SP_VS_CTRL_REG0_MERGEDREGS |
+      A6XX_SP_VS_CTRL_REG0_BRANCHSTACK(vs->branchstack);
+   if (vs->num_samp)
+      sp_vs_ctrl |= A6XX_SP_VS_CTRL_REG0_PIXLODENABLE;
 
-   assert((binary_iova & 0x7f) == 0);
-   assert((pvtmem->iova & 0x1f) == 0);
+   uint32_t sp_vs_config = A6XX_SP_VS_CONFIG_NTEX(vs->num_samp) |
+                           A6XX_SP_VS_CONFIG_NSAMP(vs->num_samp);
+   if (vs->instrlen)
+      sp_vs_config |= A6XX_SP_VS_CONFIG_ENABLED;
 
-   tu_cs_emit_pkt4(cs, cfg->reg_sp_xs_first_exec_offset, 7);
-   tu_cs_emit(cs, 0);
-   tu_cs_emit_qw(cs, binary_iova);
-   tu_cs_emit(cs,
-              A6XX_SP_VS_PVT_MEM_PARAM_MEMSIZEPERITEM(pvtmem->per_fiber_size));
-   tu_cs_emit_qw(cs, pvtmem->iova);
-   tu_cs_emit(cs, A6XX_SP_VS_PVT_MEM_SIZE_TOTALPVTMEMSIZE(pvtmem->per_sp_size) |
-                  COND(pvtmem->per_wave, A6XX_SP_VS_PVT_MEM_SIZE_PERWAVEMEMLAYOUT));
-
-   tu_cs_emit_pkt4(cs, cfg->reg_sp_xs_pvt_mem_hw_stack_offset, 1);
-   tu_cs_emit(cs, A6XX_SP_VS_PVT_MEM_HW_STACK_OFFSET_OFFSET(pvtmem->per_sp_size));
-
-   uint32_t shader_preload_size =
-      MIN2(xs->instrlen, cs->device->physical_device->info->a6xx.instr_cache_size);
-
-   tu_cs_emit_pkt7(cs, tu6_stage2opcode(stage), 3);
-   tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(0) |
-                  CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
-                  CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
-                  CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(stage)) |
-                  CP_LOAD_STATE6_0_NUM_UNIT(shader_preload_size));
-   tu_cs_emit_qw(cs, binary_iova);
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_VS_CTRL_REG0, 1);
+   tu_cs_emit(cs, sp_vs_ctrl);
 
-   /* emit immediates */
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_VS_CONFIG, 2);
+   tu_cs_emit(cs, sp_vs_config);
+   tu_cs_emit(cs, vs->instrlen);
 
-   const struct ir3_const_state *const_state = ir3_const_state(xs);
-   uint32_t base = const_state->offsets.immediate;
-   unsigned immediate_size = tu_xs_get_immediates_packet_size_dwords(xs);
-
-   if (immediate_size > 0) {
-      tu_cs_emit_pkt7(cs, tu6_stage2opcode(stage), 3 + immediate_size);
-      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(base) |
-                 CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-                 CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(stage)) |
-                 CP_LOAD_STATE6_0_NUM_UNIT(immediate_size / 4));
-      tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
-      tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
+   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_VS_CNTL, 1);
+   tu_cs_emit(cs, A6XX_HLSQ_VS_CNTL_CONSTLEN(align(vs->constlen, 4)) | 0x100);
+}
 
-      tu_cs_emit_array(cs, const_state->immediates, immediate_size);
-   }
+static void
+tu6_emit_hs_config(struct tu_cs *cs, const struct ir3_shader_variant *hs)
+{
+   uint32_t sp_hs_config = 0;
+   if (hs->instrlen)
+      sp_hs_config |= A6XX_SP_HS_CONFIG_ENABLED;
 
-   if (const_state->constant_data_ubo != -1) {
-      uint64_t iova = binary_iova + xs->info.constant_data_offset;
-
-      /* Upload UBO state for the constant data. */
-      tu_cs_emit_pkt7(cs, tu6_stage2opcode(stage), 5);
-      tu_cs_emit(cs,
-                 CP_LOAD_STATE6_0_DST_OFF(const_state->constant_data_ubo) |
-                 CP_LOAD_STATE6_0_STATE_TYPE(ST6_UBO)|
-                 CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-                 CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(stage)) |
-                 CP_LOAD_STATE6_0_NUM_UNIT(1));
-      tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
-      tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
-      int size_vec4s = DIV_ROUND_UP(xs->constant_data_size, 16);
-      tu_cs_emit_qw(cs,
-                    iova |
-                    (uint64_t)A6XX_UBO_1_SIZE(size_vec4s) << 32);
-
-      /* Upload the constant data to the const file if needed. */
-      const struct ir3_ubo_analysis_state *ubo_state = &const_state->ubo_state;
-
-      for (int i = 0; i < ubo_state->num_enabled; i++) {
-         if (ubo_state->range[i].ubo.block != const_state->constant_data_ubo ||
-             ubo_state->range[i].ubo.bindless) {
-            continue;
-         }
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_HS_UNKNOWN_A831, 1);
+   tu_cs_emit(cs, 0);
 
-         uint32_t start = ubo_state->range[i].start;
-         uint32_t end = ubo_state->range[i].end;
-         uint32_t size = MIN2(end - start,
-                              (16 * xs->constlen) - ubo_state->range[i].offset);
-
-         tu_cs_emit_pkt7(cs, tu6_stage2opcode(stage), 3);
-         tu_cs_emit(cs,
-                    CP_LOAD_STATE6_0_DST_OFF(ubo_state->range[i].offset / 16) |
-                    CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                    CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
-                    CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(stage)) |
-                    CP_LOAD_STATE6_0_NUM_UNIT(size / 16));
-         tu_cs_emit_qw(cs, iova + start);
-      }
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_HS_CONFIG, 2);
+   tu_cs_emit(cs, sp_hs_config);
+   tu_cs_emit(cs, hs->instrlen);
 
-   /* emit FS driver param */
-   if (stage == MESA_SHADER_FRAGMENT && const_state->num_driver_params > 0) {
-      uint32_t base = const_state->offsets.driver_param;
-      int32_t size = DIV_ROUND_UP(const_state->num_driver_params, 4);
-      size = MAX2(MIN2(size + base, xs->constlen) - base, 0);
-
-      if (size > 0) {
-         tu_cs_emit_pkt7(cs, tu6_stage2opcode(stage), 3 + size * 4);
-         tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(base) |
-                    CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-                    CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-                    CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(stage)) |
-                    CP_LOAD_STATE6_0_NUM_UNIT(size));
-         tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
-         tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
-
-         assert(size == 1);
-         tu_cs_emit(cs, xs->info.double_threadsize ? 128 : 64);
-         tu_cs_emit(cs, 0);
-         tu_cs_emit(cs, 0);
-         tu_cs_emit(cs, 0);
-      }
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_HS_CNTL, 1);
+   tu_cs_emit(cs, A6XX_HLSQ_HS_CNTL_CONSTLEN(align(hs->constlen, 4)));
 }
 
 static void
-tu6_emit_dynamic_offset(struct tu_cs *cs,
-                        const struct ir3_shader_variant *xs,
-                        struct tu_pipeline_builder *builder)
+tu6_emit_ds_config(struct tu_cs *cs, const struct ir3_shader_variant *ds)
 {
-   if (!xs || builder->const_state[xs->type].dynamic_offset_loc == UINT32_MAX)
-      return;
+   uint32_t sp_ds_config = 0;
+   if (ds->instrlen)
+      sp_ds_config |= A6XX_SP_DS_CONFIG_ENABLED;
 
-   tu_cs_emit_pkt7(cs, tu6_stage2opcode(xs->type), 3 + MAX_SETS);
-   tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(builder->const_state[xs->type].dynamic_offset_loc / 4) |
-              CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-              CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-              CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(xs->type)) |
-              CP_LOAD_STATE6_0_NUM_UNIT(DIV_ROUND_UP(MAX_SETS, 4)));
-   tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
-   tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
-
-   for (unsigned i = 0; i < MAX_SETS; i++) {
-      unsigned dynamic_offset_start =
-         builder->layout.set[i].dynamic_offset_start / (A6XX_TEX_CONST_DWORDS * 4);
-      tu_cs_emit(cs, i < builder->layout.num_sets ? dynamic_offset_start : 0);
-   }
-}
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_DS_CONFIG, 2);
+   tu_cs_emit(cs, sp_ds_config);
+   tu_cs_emit(cs, ds->instrlen);
 
-static void
-tu6_emit_shared_consts_enable(struct tu_cs *cs, bool enable)
-{
-   /* Enable/disable shared constants */
-   tu_cs_emit_regs(cs, A6XX_HLSQ_SHARED_CONSTS(.enable = enable));
-   tu_cs_emit_regs(cs, A6XX_SP_MODE_CONTROL(.constant_demotion_enable = true,
-                                            .isammode = ISAMMODE_GL,
-                                            .shared_consts_enable = enable));
+   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_DS_CNTL, 1);
+   tu_cs_emit(cs, A6XX_HLSQ_DS_CNTL_CONSTLEN(align(ds->constlen, 4)));
 }
 
 static void
-tu6_emit_cs_config(struct tu_cs *cs,
-                   const struct ir3_shader_variant *v,
-                   const struct tu_pvtmem_config *pvtmem,
-                   uint64_t binary_iova)
+tu6_emit_gs_config(struct tu_cs *cs, const struct ir3_shader_variant *gs)
 {
-   bool shared_consts_enable = ir3_const_state(v)->shared_consts_enable;
-   tu6_emit_shared_consts_enable(cs, shared_consts_enable);
-
-   tu_cs_emit_regs(cs, A6XX_HLSQ_INVALIDATE_CMD(
-         .cs_state = true,
-         .cs_ibo = true,
-         .cs_shared_const = shared_consts_enable));
-
-   tu6_emit_xs_config(cs, MESA_SHADER_COMPUTE, v);
-   tu6_emit_xs(cs, MESA_SHADER_COMPUTE, v, pvtmem, binary_iova);
-
-   uint32_t shared_size = MAX2(((int)v->shared_size - 1) / 1024, 1);
-   tu_cs_emit_pkt4(cs, REG_A6XX_SP_CS_UNKNOWN_A9B1, 1);
-   tu_cs_emit(cs, A6XX_SP_CS_UNKNOWN_A9B1_SHARED_SIZE(shared_size) |
-                  A6XX_SP_CS_UNKNOWN_A9B1_UNK6);
-
-   if (cs->device->physical_device->info->a6xx.has_lpac) {
-      tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CS_UNKNOWN_B9D0, 1);
-      tu_cs_emit(cs, A6XX_HLSQ_CS_UNKNOWN_B9D0_SHARED_SIZE(shared_size) |
-                     A6XX_HLSQ_CS_UNKNOWN_B9D0_UNK6);
-   }
+   uint32_t sp_gs_config = 0;
+   if (gs->instrlen)
+      sp_gs_config |= A6XX_SP_GS_CONFIG_ENABLED;
 
-   uint32_t local_invocation_id =
-      ir3_find_sysval_regid(v, SYSTEM_VALUE_LOCAL_INVOCATION_ID);
-   uint32_t work_group_id =
-      ir3_find_sysval_regid(v, SYSTEM_VALUE_WORKGROUP_ID);
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_UNKNOWN_A871, 1);
+   tu_cs_emit(cs, 0);
 
-   enum a6xx_threadsize thrsz = v->info.double_threadsize ? THREAD128 : THREAD64;
-   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CS_CNTL_0, 2);
-   tu_cs_emit(cs,
-              A6XX_HLSQ_CS_CNTL_0_WGIDCONSTID(work_group_id) |
-              A6XX_HLSQ_CS_CNTL_0_WGSIZECONSTID(regid(63, 0)) |
-              A6XX_HLSQ_CS_CNTL_0_WGOFFSETCONSTID(regid(63, 0)) |
-              A6XX_HLSQ_CS_CNTL_0_LOCALIDREGID(local_invocation_id));
-   tu_cs_emit(cs, A6XX_HLSQ_CS_CNTL_1_LINEARLOCALIDREGID(regid(63, 0)) |
-                  A6XX_HLSQ_CS_CNTL_1_THREADSIZE(thrsz));
-
-   if (cs->device->physical_device->info->a6xx.has_lpac) {
-      tu_cs_emit_pkt4(cs, REG_A6XX_SP_CS_CNTL_0, 2);
-      tu_cs_emit(cs,
-                 A6XX_SP_CS_CNTL_0_WGIDCONSTID(work_group_id) |
-                 A6XX_SP_CS_CNTL_0_WGSIZECONSTID(regid(63, 0)) |
-                 A6XX_SP_CS_CNTL_0_WGOFFSETCONSTID(regid(63, 0)) |
-                 A6XX_SP_CS_CNTL_0_LOCALIDREGID(local_invocation_id));
-      tu_cs_emit(cs, A6XX_SP_CS_CNTL_1_LINEARLOCALIDREGID(regid(63, 0)) |
-                     A6XX_SP_CS_CNTL_1_THREADSIZE(thrsz));
-   }
-}
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_CONFIG, 2);
+   tu_cs_emit(cs, sp_gs_config);
+   tu_cs_emit(cs, gs->instrlen);
 
-#define TU6_EMIT_VFD_DEST_MAX_DWORDS (MAX_VERTEX_ATTRIBS + 2)
+   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_GS_CNTL, 1);
+   tu_cs_emit(cs, A6XX_HLSQ_GS_CNTL_CONSTLEN(align(gs->constlen, 4)));
+}
 
 static void
-tu6_emit_vfd_dest(struct tu_cs *cs,
-                  const struct ir3_shader_variant *vs)
-{
-   int32_t input_for_attr[MAX_VERTEX_ATTRIBS];
-   uint32_t attr_count = 0;
+tu6_emit_fs_config(struct tu_cs *cs, const struct ir3_shader_variant *fs)
+{
+   uint32_t sp_fs_ctrl =
+      A6XX_SP_FS_CTRL_REG0_THREADSIZE(FOUR_QUADS) | 0x1000000 |
+      A6XX_SP_FS_CTRL_REG0_FULLREGFOOTPRINT(fs->info.max_reg + 1) |
+      A6XX_SP_FS_CTRL_REG0_MERGEDREGS |
+      A6XX_SP_FS_CTRL_REG0_BRANCHSTACK(fs->branchstack);
+   if (fs->total_in > 0 || fs->frag_coord)
+      sp_fs_ctrl |= A6XX_SP_FS_CTRL_REG0_VARYING;
+   if (fs->num_samp > 0)
+      sp_fs_ctrl |= A6XX_SP_FS_CTRL_REG0_PIXLODENABLE;
+
+   uint32_t sp_fs_config = A6XX_SP_FS_CONFIG_NTEX(fs->num_samp) |
+                           A6XX_SP_FS_CONFIG_NSAMP(fs->num_samp);
+   if (fs->instrlen)
+      sp_fs_config |= A6XX_SP_FS_CONFIG_ENABLED;
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_UNKNOWN_A99E, 1);
+   tu_cs_emit(cs, 0x7fc0);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_UNKNOWN_A9A8, 1);
+   tu_cs_emit(cs, 0);
 
-   for (unsigned i = 0; i < MAX_VERTEX_ATTRIBS; i++)
-      input_for_attr[i] = -1;
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_UNKNOWN_AB00, 1);
+   tu_cs_emit(cs, 0x5);
 
-   for (unsigned i = 0; i < vs->inputs_count; i++) {
-      if (vs->inputs[i].sysval || vs->inputs[i].regid == regid(63, 0))
-         continue;
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_CTRL_REG0, 1);
+   tu_cs_emit(cs, sp_fs_ctrl);
 
-      assert(vs->inputs[i].slot >= VERT_ATTRIB_GENERIC0);
-      unsigned loc = vs->inputs[i].slot - VERT_ATTRIB_GENERIC0;
-      input_for_attr[loc] = i;
-      attr_count = MAX2(attr_count, loc + 1);
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_CONFIG, 2);
+   tu_cs_emit(cs, sp_fs_config);
+   tu_cs_emit(cs, fs->instrlen);
 
-   tu_cs_emit_regs(cs,
-                   A6XX_VFD_CONTROL_0(
-                     .fetch_cnt = attr_count, /* decode_cnt for binning pass ? */
-                     .decode_cnt = attr_count));
-
-   if (attr_count)
-      tu_cs_emit_pkt4(cs, REG_A6XX_VFD_DEST_CNTL_INSTR(0), attr_count);
-
-   for (unsigned i = 0; i < attr_count; i++) {
-      if (input_for_attr[i] >= 0) {
-            unsigned input_idx = input_for_attr[i];
-            tu_cs_emit(cs, A6XX_VFD_DEST_CNTL_INSTR(0,
-                             .writemask = vs->inputs[input_idx].compmask,
-                             .regid = vs->inputs[input_idx].regid).value);
-      } else {
-            tu_cs_emit(cs, A6XX_VFD_DEST_CNTL_INSTR(0,
-                             .writemask = 0,
-                             .regid = regid(63, 0)).value);
-      }
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_FS_CNTL, 1);
+   tu_cs_emit(cs, A6XX_HLSQ_FS_CNTL_CONSTLEN(align(fs->constlen, 4)) | 0x100);
 }
 
 static void
 tu6_emit_vs_system_values(struct tu_cs *cs,
-                          const struct ir3_shader_variant *vs,
-                          const struct ir3_shader_variant *hs,
-                          const struct ir3_shader_variant *ds,
-                          const struct ir3_shader_variant *gs,
-                          bool primid_passthru)
+                          const struct ir3_shader_variant *vs)
 {
    const uint32_t vertexid_regid =
-         ir3_find_sysval_regid(vs, SYSTEM_VALUE_VERTEX_ID);
+      ir3_find_sysval_regid(vs, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE);
    const uint32_t instanceid_regid =
-         ir3_find_sysval_regid(vs, SYSTEM_VALUE_INSTANCE_ID);
-   const uint32_t tess_coord_x_regid = hs ?
-         ir3_find_sysval_regid(ds, SYSTEM_VALUE_TESS_COORD) :
-         regid(63, 0);
-   const uint32_t tess_coord_y_regid = VALIDREG(tess_coord_x_regid) ?
-         tess_coord_x_regid + 1 :
-         regid(63, 0);
-   const uint32_t hs_rel_patch_regid = hs ?
-         ir3_find_sysval_regid(hs, SYSTEM_VALUE_REL_PATCH_ID_IR3) :
-         regid(63, 0);
-   const uint32_t ds_rel_patch_regid = hs ?
-         ir3_find_sysval_regid(ds, SYSTEM_VALUE_REL_PATCH_ID_IR3) :
-         regid(63, 0);
-   const uint32_t hs_invocation_regid = hs ?
-         ir3_find_sysval_regid(hs, SYSTEM_VALUE_TCS_HEADER_IR3) :
-         regid(63, 0);
-   const uint32_t gs_primitiveid_regid = gs ?
-         ir3_find_sysval_regid(gs, SYSTEM_VALUE_PRIMITIVE_ID) :
-         regid(63, 0);
-   const uint32_t vs_primitiveid_regid = hs ?
-         ir3_find_sysval_regid(hs, SYSTEM_VALUE_PRIMITIVE_ID) :
-         gs_primitiveid_regid;
-   const uint32_t ds_primitiveid_regid = ds ?
-         ir3_find_sysval_regid(ds, SYSTEM_VALUE_PRIMITIVE_ID) :
-         regid(63, 0);
-   const uint32_t gsheader_regid = gs ?
-         ir3_find_sysval_regid(gs, SYSTEM_VALUE_GS_HEADER_IR3) :
-         regid(63, 0);
-
-   /* Note: we currently don't support multiview with tess or GS. If we did,
-    * and the HW actually works, then we'd have to somehow share this across
-    * stages. Note that the blob doesn't support this either.
-    */
-   const uint32_t viewid_regid =
-      ir3_find_sysval_regid(vs, SYSTEM_VALUE_VIEW_INDEX);
+      ir3_find_sysval_regid(vs, SYSTEM_VALUE_INSTANCE_ID);
 
    tu_cs_emit_pkt4(cs, REG_A6XX_VFD_CONTROL_1, 6);
    tu_cs_emit(cs, A6XX_VFD_CONTROL_1_REGID4VTX(vertexid_regid) |
-                  A6XX_VFD_CONTROL_1_REGID4INST(instanceid_regid) |
-                  A6XX_VFD_CONTROL_1_REGID4PRIMID(vs_primitiveid_regid) |
-                  A6XX_VFD_CONTROL_1_REGID4VIEWID(viewid_regid));
-   tu_cs_emit(cs, A6XX_VFD_CONTROL_2_REGID_HSRELPATCHID(hs_rel_patch_regid) |
-                  A6XX_VFD_CONTROL_2_REGID_INVOCATIONID(hs_invocation_regid));
-   tu_cs_emit(cs, A6XX_VFD_CONTROL_3_REGID_DSRELPATCHID(ds_rel_patch_regid) |
-                  A6XX_VFD_CONTROL_3_REGID_TESSX(tess_coord_x_regid) |
-                  A6XX_VFD_CONTROL_3_REGID_TESSY(tess_coord_y_regid) |
-                  A6XX_VFD_CONTROL_3_REGID_DSPRIMID(ds_primitiveid_regid));
+                     A6XX_VFD_CONTROL_1_REGID4INST(instanceid_regid) |
+                     0xfcfc0000);
+   tu_cs_emit(cs, 0x0000fcfc); /* VFD_CONTROL_2 */
+   tu_cs_emit(cs, 0xfcfcfcfc); /* VFD_CONTROL_3 */
    tu_cs_emit(cs, 0x000000fc); /* VFD_CONTROL_4 */
-   tu_cs_emit(cs, A6XX_VFD_CONTROL_5_REGID_GSHEADER(gsheader_regid) |
-                  0xfc00); /* VFD_CONTROL_5 */
-   tu_cs_emit(cs, COND(primid_passthru, A6XX_VFD_CONTROL_6_PRIMID_PASSTHRU)); /* VFD_CONTROL_6 */
+   tu_cs_emit(cs, 0x0000fcfc); /* VFD_CONTROL_5 */
+   tu_cs_emit(cs, 0x00000000); /* VFD_CONTROL_6 */
 }
 
 static void
-tu6_setup_streamout(struct tu_cs *cs,
-                    const struct ir3_shader_variant *v,
-                    struct ir3_shader_linkage *l)
+tu6_emit_vpc(struct tu_cs *cs,
+             const struct ir3_shader_variant *vs,
+             const struct ir3_shader_variant *fs,
+             bool binning_pass)
 {
-   const struct ir3_stream_output_info *info = &v->stream_output;
-   /* Note: 64 here comes from the HW layout of the program RAM. The program
-    * for stream N is at DWORD 64 * N.
-    */
-#define A6XX_SO_PROG_DWORDS 64
-   uint32_t prog[A6XX_SO_PROG_DWORDS * IR3_MAX_SO_STREAMS] = {};
-   BITSET_DECLARE(valid_dwords, A6XX_SO_PROG_DWORDS * IR3_MAX_SO_STREAMS) = {0};
-
-   /* TODO: streamout state should be in a non-GMEM draw state */
-
-   /* no streamout: */
-   if (info->num_outputs == 0) {
-      unsigned sizedw = 4;
-      if (cs->device->physical_device->info->a6xx.tess_use_shared)
-         sizedw += 2;
-
-      tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, sizedw);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
-      tu_cs_emit(cs, 0);
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_STREAM_CNTL);
-      tu_cs_emit(cs, 0);
+   struct ir3_shader_linkage linkage = { 0 };
+   ir3_link_shaders(&linkage, vs, fs);
 
-      if (cs->device->physical_device->info->a6xx.tess_use_shared) {
-         tu_cs_emit(cs, REG_A6XX_PC_SO_STREAM_CNTL);
-         tu_cs_emit(cs, 0);
-      }
+   if (vs->shader->stream_output.num_outputs && !binning_pass)
+      tu_finishme("stream output");
 
-      return;
+   BITSET_DECLARE(vpc_var_enables, 128) = { 0 };
+   for (uint32_t i = 0; i < linkage.cnt; i++) {
+      const uint32_t comp_count = util_last_bit(linkage.var[i].compmask);
+      for (uint32_t j = 0; j < comp_count; j++)
+         BITSET_SET(vpc_var_enables, linkage.var[i].loc + j);
    }
 
-   for (unsigned i = 0; i < info->num_outputs; i++) {
-      const struct ir3_stream_output *out = &info->output[i];
-      unsigned k = out->register_index;
-      unsigned idx;
-
-      /* Skip it, if it's an output that was never assigned a register. */
-      if (k >= v->outputs_count || v->outputs[k].regid == INVALID_REG)
-         continue;
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VAR_DISABLE(0), 4);
+   tu_cs_emit(cs, ~vpc_var_enables[0]);
+   tu_cs_emit(cs, ~vpc_var_enables[1]);
+   tu_cs_emit(cs, ~vpc_var_enables[2]);
+   tu_cs_emit(cs, ~vpc_var_enables[3]);
 
-      /* linkage map sorted by order frag shader wants things, so
-       * a bit less ideal here..
-       */
-      for (idx = 0; idx < l->cnt; idx++)
-         if (l->var[idx].slot == v->outputs[k].slot)
-            break;
-
-      assert(idx < l->cnt);
-
-      for (unsigned j = 0; j < out->num_components; j++) {
-         unsigned c   = j + out->start_component;
-         unsigned loc = l->var[idx].loc + c;
-         unsigned off = j + out->dst_offset;  /* in dwords */
-
-         assert(loc < A6XX_SO_PROG_DWORDS * 2);
-         unsigned dword = out->stream * A6XX_SO_PROG_DWORDS + loc/2;
-         if (loc & 1) {
-            prog[dword] |= A6XX_VPC_SO_PROG_B_EN |
-                           A6XX_VPC_SO_PROG_B_BUF(out->output_buffer) |
-                           A6XX_VPC_SO_PROG_B_OFF(off * 4);
-         } else {
-            prog[dword] |= A6XX_VPC_SO_PROG_A_EN |
-                           A6XX_VPC_SO_PROG_A_BUF(out->output_buffer) |
-                           A6XX_VPC_SO_PROG_A_OFF(off * 4);
-         }
-         BITSET_SET(valid_dwords, dword);
-      }
+   /* a6xx finds position/pointsize at the end */
+   const uint32_t position_regid =
+      ir3_find_output_regid(vs, VARYING_SLOT_POS);
+   const uint32_t pointsize_regid =
+      ir3_find_output_regid(vs, VARYING_SLOT_PSIZ);
+   uint32_t pointsize_loc = 0xff;
+   if (position_regid != regid(63, 0))
+      ir3_link_add(&linkage, position_regid, 0xf, linkage.max_loc);
+   if (pointsize_regid != regid(63, 0)) {
+      pointsize_loc = linkage.max_loc;
+      ir3_link_add(&linkage, pointsize_regid, 0x1, linkage.max_loc);
    }
 
-   unsigned prog_count = 0;
-   unsigned start, end;
-   BITSET_FOREACH_RANGE(start, end, valid_dwords,
-                        A6XX_SO_PROG_DWORDS * IR3_MAX_SO_STREAMS) {
-      prog_count += end - start + 1;
+   /* map vs outputs to VPC */
+   assert(linkage.cnt <= 32);
+   const uint32_t sp_vs_out_count = (linkage.cnt + 1) / 2;
+   const uint32_t sp_vs_vpc_dst_count = (linkage.cnt + 3) / 4;
+   uint32_t sp_vs_out[16];
+   uint32_t sp_vs_vpc_dst[8];
+   sp_vs_out[sp_vs_out_count - 1] = 0;
+   sp_vs_vpc_dst[sp_vs_vpc_dst_count - 1] = 0;
+   for (uint32_t i = 0; i < linkage.cnt; i++) {
+      ((uint16_t *) sp_vs_out)[i] =
+         A6XX_SP_VS_OUT_REG_A_REGID(linkage.var[i].regid) |
+         A6XX_SP_VS_OUT_REG_A_COMPMASK(linkage.var[i].compmask);
+      ((uint8_t *) sp_vs_vpc_dst)[i] =
+         A6XX_SP_VS_VPC_DST_REG_OUTLOC0(linkage.var[i].loc);
    }
 
-   const bool emit_pc_so_stream_cntl =
-      cs->device->physical_device->info->a6xx.tess_use_shared &&
-      v->type == MESA_SHADER_TESS_EVAL;
-
-   if (emit_pc_so_stream_cntl)
-      prog_count += 1;
-
-   tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 10 + 2 * prog_count);
-   tu_cs_emit(cs, REG_A6XX_VPC_SO_STREAM_CNTL);
-   tu_cs_emit(cs, A6XX_VPC_SO_STREAM_CNTL_STREAM_ENABLE(info->streams_written) |
-                  COND(info->stride[0] > 0,
-                       A6XX_VPC_SO_STREAM_CNTL_BUF0_STREAM(1 + info->buffer_to_stream[0])) |
-                  COND(info->stride[1] > 0,
-                       A6XX_VPC_SO_STREAM_CNTL_BUF1_STREAM(1 + info->buffer_to_stream[1])) |
-                  COND(info->stride[2] > 0,
-                       A6XX_VPC_SO_STREAM_CNTL_BUF2_STREAM(1 + info->buffer_to_stream[2])) |
-                  COND(info->stride[3] > 0,
-                       A6XX_VPC_SO_STREAM_CNTL_BUF3_STREAM(1 + info->buffer_to_stream[3])));
-   for (uint32_t i = 0; i < 4; i++) {
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_BUFFER_STRIDE(i));
-      tu_cs_emit(cs, info->stride[i]);
-   }
-   bool first = true;
-   BITSET_FOREACH_RANGE(start, end, valid_dwords,
-                        A6XX_SO_PROG_DWORDS * IR3_MAX_SO_STREAMS) {
-      tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
-      tu_cs_emit(cs, COND(first, A6XX_VPC_SO_CNTL_RESET) |
-                     A6XX_VPC_SO_CNTL_ADDR(start));
-      for (unsigned i = start; i < end; i++) {
-         tu_cs_emit(cs, REG_A6XX_VPC_SO_PROG);
-         tu_cs_emit(cs, prog[i]);
-      }
-      first = false;
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_VS_OUT_REG(0), sp_vs_out_count);
+   tu_cs_emit_array(cs, sp_vs_out, sp_vs_out_count);
 
-   if (emit_pc_so_stream_cntl) {
-      /* Possibly not tess_use_shared related, but the combination of
-       * tess + xfb fails some tests if we don't emit this.
-       */
-      tu_cs_emit(cs, REG_A6XX_PC_SO_STREAM_CNTL);
-      tu_cs_emit(cs, A6XX_PC_SO_STREAM_CNTL_STREAM_ENABLE(info->streams_written));
-   }
-}
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_VS_VPC_DST_REG(0), sp_vs_vpc_dst_count);
+   tu_cs_emit_array(cs, sp_vs_vpc_dst, sp_vs_vpc_dst_count);
 
-static void
-tu6_emit_const(struct tu_cs *cs, uint32_t opcode, uint32_t base,
-               enum a6xx_state_block block, uint32_t offset,
-               uint32_t size, const uint32_t *dwords) {
-   assert(size % 4 == 0);
-
-   tu_cs_emit_pkt7(cs, opcode, 3 + size);
-   tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(base) |
-         CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
-         CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
-         CP_LOAD_STATE6_0_STATE_BLOCK(block) |
-         CP_LOAD_STATE6_0_NUM_UNIT(size / 4));
-
-   tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
-   tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
-   dwords = (uint32_t *)&((uint8_t *)dwords)[offset];
-
-   tu_cs_emit_array(cs, dwords, size);
-}
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_CNTL_0, 1);
+   tu_cs_emit(cs, A6XX_VPC_CNTL_0_NUMNONPOSVAR(fs->total_in) |
+                     (fs->total_in > 0 ? A6XX_VPC_CNTL_0_VARYING : 0) |
+                     0xff00ff00);
 
-static void
-tu6_emit_link_map(struct tu_cs *cs,
-                  const struct ir3_shader_variant *producer,
-                  const struct ir3_shader_variant *consumer,
-                  enum a6xx_state_block sb)
-{
-   const struct ir3_const_state *const_state = ir3_const_state(consumer);
-   uint32_t base = const_state->offsets.primitive_map;
-   int size = DIV_ROUND_UP(consumer->input_size, 4);
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_PACK, 1);
+   tu_cs_emit(cs, A6XX_VPC_PACK_NUMNONPOSVAR(fs->total_in) |
+                     A6XX_VPC_PACK_PSIZELOC(pointsize_loc) |
+                     A6XX_VPC_PACK_STRIDE_IN_VPC(linkage.max_loc));
 
-   size = (MIN2(size + base, consumer->constlen) - base) * 4;
-   if (size <= 0)
-      return;
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_GS_SIV_CNTL, 1);
+   tu_cs_emit(cs, 0x0000ffff); /* XXX */
 
-   tu6_emit_const(cs, CP_LOAD_STATE6_GEOM, base, sb, 0, size,
-                         producer->output_loc);
-}
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_PRIMITIVE_CNTL, 1);
+   tu_cs_emit(cs, A6XX_SP_PRIMITIVE_CNTL_VSOUT(linkage.cnt));
 
-static uint16_t
-primitive_to_tess(enum shader_prim primitive) {
-   switch (primitive) {
-   case SHADER_PRIM_POINTS:
-      return TESS_POINTS;
-   case SHADER_PRIM_LINE_STRIP:
-      return TESS_LINES;
-   case SHADER_PRIM_TRIANGLE_STRIP:
-      return TESS_CW_TRIS;
-   default:
-      unreachable("");
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_1, 1);
+   tu_cs_emit(cs, A6XX_PC_PRIMITIVE_CNTL_1_STRIDE_IN_VPC(linkage.max_loc) |
+                     (vs->writes_psize ? A6XX_PC_PRIMITIVE_CNTL_1_PSIZE : 0));
 }
 
 static int
 tu6_vpc_varying_mode(const struct ir3_shader_variant *fs,
-                     const struct ir3_shader_variant *last_shader,
                      uint32_t index,
                      uint8_t *interp_mode,
                      uint8_t *ps_repl_mode)
@@ -1091,18 +620,8 @@ tu6_vpc_varying_mode(const struct ir3_shader_variant *fs,
          *interp_mode |= INTERP_ONE << 6;
          shift += 2;
       }
-   } else if (fs->inputs[index].slot == VARYING_SLOT_LAYER ||
-              fs->inputs[index].slot == VARYING_SLOT_VIEWPORT) {
-      /* If the last geometry shader doesn't statically write these, they're
-       * implicitly zero and the FS is supposed to read zero.
-       */
-      if (ir3_find_output(last_shader, fs->inputs[index].slot) < 0 &&
-          (compmask & 0x1)) {
-         *interp_mode |= INTERP_ZERO;
-      } else {
-         *interp_mode |= INTERP_FLAT;
-      }
-   } else if (fs->inputs[index].flat) {
+   } else if ((fs->inputs[index].interpolate == INTERP_MODE_FLAT) ||
+              fs->inputs[index].rasterflat) {
       for (int i = 0; i < 4; i++) {
          if (compmask & (1 << i)) {
             *interp_mode |= INTERP_FLAT << shift;
@@ -1111,19 +630,18 @@ tu6_vpc_varying_mode(const struct ir3_shader_variant *fs,
       }
    }
 
-   return util_bitcount(compmask) * 2;
+   return shift;
 }
 
 static void
 tu6_emit_vpc_varying_modes(struct tu_cs *cs,
                            const struct ir3_shader_variant *fs,
-                           const struct ir3_shader_variant *last_shader)
+                           bool binning_pass)
 {
    uint32_t interp_modes[8] = { 0 };
    uint32_t ps_repl_modes[8] = { 0 };
-   uint32_t interp_regs = 0;
 
-   if (fs) {
+   if (!binning_pass) {
       for (int i = -1;
            (i = ir3_next_varying(fs, i)) < (int) fs->inputs_count;) {
 
@@ -1131,7 +649,7 @@ tu6_emit_vpc_varying_modes(struct tu_cs *cs,
          uint8_t interp_mode;
          uint8_t ps_repl_mode;
          const int bits =
-            tu6_vpc_varying_mode(fs, last_shader, i, &interp_mode, &ps_repl_mode);
+            tu6_vpc_varying_mode(fs, i, &interp_mode, &ps_repl_mode);
 
          /* OR the mode into the array */
          const uint32_t inloc = fs->inputs[i].inloc * 2;
@@ -1146,1043 +664,445 @@ tu6_emit_vpc_varying_modes(struct tu_cs *cs,
             interp_modes[n] |= interp_mode >> shift;
             ps_repl_modes[n] |= ps_repl_mode >> shift;
          }
-         interp_regs = MAX2(interp_regs, n + 1);
       }
    }
 
-   if (interp_regs) {
-      tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VARYING_INTERP_MODE(0), interp_regs);
-      tu_cs_emit_array(cs, interp_modes, interp_regs);
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VARYING_INTERP_MODE(0), 8);
+   tu_cs_emit_array(cs, interp_modes, 8);
 
-      tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VARYING_PS_REPL_MODE(0), interp_regs);
-      tu_cs_emit_array(cs, ps_repl_modes, interp_regs);
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VARYING_PS_REPL_MODE(0), 8);
+   tu_cs_emit_array(cs, ps_repl_modes, 8);
 }
 
-void
-tu6_emit_vpc(struct tu_cs *cs,
-             const struct ir3_shader_variant *vs,
-             const struct ir3_shader_variant *hs,
-             const struct ir3_shader_variant *ds,
-             const struct ir3_shader_variant *gs,
-             const struct ir3_shader_variant *fs)
-{
-   /* note: doesn't compile as static because of the array regs.. */
-   const struct reg_config {
-      uint16_t reg_sp_xs_out_reg;
-      uint16_t reg_sp_xs_vpc_dst_reg;
-      uint16_t reg_vpc_xs_pack;
-      uint16_t reg_vpc_xs_clip_cntl;
-      uint16_t reg_gras_xs_cl_cntl;
-      uint16_t reg_pc_xs_out_cntl;
-      uint16_t reg_sp_xs_primitive_cntl;
-      uint16_t reg_vpc_xs_layer_cntl;
-      uint16_t reg_gras_xs_layer_cntl;
-   } reg_config[] = {
-      [MESA_SHADER_VERTEX] = {
-         REG_A6XX_SP_VS_OUT_REG(0),
-         REG_A6XX_SP_VS_VPC_DST_REG(0),
-         REG_A6XX_VPC_VS_PACK,
-         REG_A6XX_VPC_VS_CLIP_CNTL,
-         REG_A6XX_GRAS_VS_CL_CNTL,
-         REG_A6XX_PC_VS_OUT_CNTL,
-         REG_A6XX_SP_VS_PRIMITIVE_CNTL,
-         REG_A6XX_VPC_VS_LAYER_CNTL,
-         REG_A6XX_GRAS_VS_LAYER_CNTL
-      },
-      [MESA_SHADER_TESS_CTRL] = {
-         0,
-         0,
-         0,
-         0,
-         0,
-         REG_A6XX_PC_HS_OUT_CNTL,
-         0,
-         0,
-         0
-      },
-      [MESA_SHADER_TESS_EVAL] = {
-         REG_A6XX_SP_DS_OUT_REG(0),
-         REG_A6XX_SP_DS_VPC_DST_REG(0),
-         REG_A6XX_VPC_DS_PACK,
-         REG_A6XX_VPC_DS_CLIP_CNTL,
-         REG_A6XX_GRAS_DS_CL_CNTL,
-         REG_A6XX_PC_DS_OUT_CNTL,
-         REG_A6XX_SP_DS_PRIMITIVE_CNTL,
-         REG_A6XX_VPC_DS_LAYER_CNTL,
-         REG_A6XX_GRAS_DS_LAYER_CNTL
-      },
-      [MESA_SHADER_GEOMETRY] = {
-         REG_A6XX_SP_GS_OUT_REG(0),
-         REG_A6XX_SP_GS_VPC_DST_REG(0),
-         REG_A6XX_VPC_GS_PACK,
-         REG_A6XX_VPC_GS_CLIP_CNTL,
-         REG_A6XX_GRAS_GS_CL_CNTL,
-         REG_A6XX_PC_GS_OUT_CNTL,
-         REG_A6XX_SP_GS_PRIMITIVE_CNTL,
-         REG_A6XX_VPC_GS_LAYER_CNTL,
-         REG_A6XX_GRAS_GS_LAYER_CNTL
-      },
-   };
-
-   const struct ir3_shader_variant *last_shader;
-   if (gs) {
-      last_shader = gs;
-   } else if (hs) {
-      last_shader = ds;
-   } else {
-      last_shader = vs;
-   }
-
-   const struct reg_config *cfg = &reg_config[last_shader->type];
-
-   struct ir3_shader_linkage linkage = {
-      .primid_loc = 0xff,
-      .clip0_loc = 0xff,
-      .clip1_loc = 0xff,
-   };
-   if (fs)
-      ir3_link_shaders(&linkage, last_shader, fs, true);
-
-   if (last_shader->stream_output.num_outputs)
-      ir3_link_stream_out(&linkage, last_shader);
-
-   /* We do this after linking shaders in order to know whether PrimID
-    * passthrough needs to be enabled.
-    */
-   bool primid_passthru = linkage.primid_loc != 0xff;
-   tu6_emit_vs_system_values(cs, vs, hs, ds, gs, primid_passthru);
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_VAR_DISABLE(0), 4);
-   tu_cs_emit(cs, ~linkage.varmask[0]);
-   tu_cs_emit(cs, ~linkage.varmask[1]);
-   tu_cs_emit(cs, ~linkage.varmask[2]);
-   tu_cs_emit(cs, ~linkage.varmask[3]);
-
-   /* a6xx finds position/pointsize at the end */
-   const uint32_t pointsize_regid =
-      ir3_find_output_regid(last_shader, VARYING_SLOT_PSIZ);
-   const uint32_t layer_regid =
-      ir3_find_output_regid(last_shader, VARYING_SLOT_LAYER);
-   const uint32_t view_regid =
-      ir3_find_output_regid(last_shader, VARYING_SLOT_VIEWPORT);
-   const uint32_t clip0_regid =
-      ir3_find_output_regid(last_shader, VARYING_SLOT_CLIP_DIST0);
-   const uint32_t clip1_regid =
-      ir3_find_output_regid(last_shader, VARYING_SLOT_CLIP_DIST1);
-   uint32_t flags_regid = gs ?
-      ir3_find_output_regid(gs, VARYING_SLOT_GS_VERTEX_FLAGS_IR3) : 0;
-
-   uint32_t pointsize_loc = 0xff, position_loc = 0xff, layer_loc = 0xff, view_loc = 0xff;
-
-   if (layer_regid != regid(63, 0)) {
-      layer_loc = linkage.max_loc;
-      ir3_link_add(&linkage, VARYING_SLOT_LAYER, layer_regid, 0x1, linkage.max_loc);
-   }
-
-   if (view_regid != regid(63, 0)) {
-      view_loc = linkage.max_loc;
-      ir3_link_add(&linkage, VARYING_SLOT_VIEWPORT, view_regid, 0x1, linkage.max_loc);
-   }
-
-   unsigned extra_pos = 0;
-
-   for (unsigned i = 0; i < last_shader->outputs_count; i++) {
-      if (last_shader->outputs[i].slot != VARYING_SLOT_POS)
-         continue;
-
-      if (position_loc == 0xff)
-         position_loc = linkage.max_loc;
-
-      ir3_link_add(&linkage, last_shader->outputs[i].slot,
-                   last_shader->outputs[i].regid,
-                   0xf, position_loc + 4 * last_shader->outputs[i].view);
-      extra_pos = MAX2(extra_pos, last_shader->outputs[i].view);
-   }
-
-   if (pointsize_regid != regid(63, 0)) {
-      pointsize_loc = linkage.max_loc;
-      ir3_link_add(&linkage, VARYING_SLOT_PSIZ, pointsize_regid, 0x1, linkage.max_loc);
-   }
-
-   uint8_t clip_cull_mask = last_shader->clip_mask | last_shader->cull_mask;
-
-   /* Handle the case where clip/cull distances aren't read by the FS */
-   uint32_t clip0_loc = linkage.clip0_loc, clip1_loc = linkage.clip1_loc;
-   if (clip0_loc == 0xff && clip0_regid != regid(63, 0)) {
-      clip0_loc = linkage.max_loc;
-      ir3_link_add(&linkage, VARYING_SLOT_CLIP_DIST0, clip0_regid,
-                   clip_cull_mask & 0xf, linkage.max_loc);
-   }
-   if (clip1_loc == 0xff && clip1_regid != regid(63, 0)) {
-      clip1_loc = linkage.max_loc;
-      ir3_link_add(&linkage, VARYING_SLOT_CLIP_DIST1, clip1_regid,
-                   clip_cull_mask >> 4, linkage.max_loc);
-   }
-
-   tu6_setup_streamout(cs, last_shader, &linkage);
-
-   /* The GPU hangs on some models when there are no outputs (xs_pack::CNT),
-    * at least when a DS is the last stage, so add a dummy output to keep it
-    * happy if there aren't any. We do this late in order to avoid emitting
-    * any unused code and make sure that optimizations don't remove it.
-    */
-   if (linkage.cnt == 0)
-      ir3_link_add(&linkage, 0, 0, 0x1, linkage.max_loc);
-
-   /* map outputs of the last shader to VPC */
-   assert(linkage.cnt <= 32);
-   const uint32_t sp_out_count = DIV_ROUND_UP(linkage.cnt, 2);
-   const uint32_t sp_vpc_dst_count = DIV_ROUND_UP(linkage.cnt, 4);
-   uint32_t sp_out[16] = {0};
-   uint32_t sp_vpc_dst[8] = {0};
-   for (uint32_t i = 0; i < linkage.cnt; i++) {
-      ((uint16_t *) sp_out)[i] =
-         A6XX_SP_VS_OUT_REG_A_REGID(linkage.var[i].regid) |
-         A6XX_SP_VS_OUT_REG_A_COMPMASK(linkage.var[i].compmask);
-      ((uint8_t *) sp_vpc_dst)[i] =
-         A6XX_SP_VS_VPC_DST_REG_OUTLOC0(linkage.var[i].loc);
-   }
-
-   tu_cs_emit_pkt4(cs, cfg->reg_sp_xs_out_reg, sp_out_count);
-   tu_cs_emit_array(cs, sp_out, sp_out_count);
-
-   tu_cs_emit_pkt4(cs, cfg->reg_sp_xs_vpc_dst_reg, sp_vpc_dst_count);
-   tu_cs_emit_array(cs, sp_vpc_dst, sp_vpc_dst_count);
-
-   tu_cs_emit_pkt4(cs, cfg->reg_vpc_xs_pack, 1);
-   tu_cs_emit(cs, A6XX_VPC_VS_PACK_POSITIONLOC(position_loc) |
-                  A6XX_VPC_VS_PACK_PSIZELOC(pointsize_loc) |
-                  A6XX_VPC_VS_PACK_STRIDE_IN_VPC(linkage.max_loc) |
-                  A6XX_VPC_VS_PACK_EXTRAPOS(extra_pos));
-
-   tu_cs_emit_pkt4(cs, cfg->reg_vpc_xs_clip_cntl, 1);
-   tu_cs_emit(cs, A6XX_VPC_VS_CLIP_CNTL_CLIP_MASK(clip_cull_mask) |
-                  A6XX_VPC_VS_CLIP_CNTL_CLIP_DIST_03_LOC(clip0_loc) |
-                  A6XX_VPC_VS_CLIP_CNTL_CLIP_DIST_47_LOC(clip1_loc));
-
-   tu_cs_emit_pkt4(cs, cfg->reg_gras_xs_cl_cntl, 1);
-   tu_cs_emit(cs, A6XX_GRAS_VS_CL_CNTL_CLIP_MASK(last_shader->clip_mask) |
-                  A6XX_GRAS_VS_CL_CNTL_CULL_MASK(last_shader->cull_mask));
-
-   const struct ir3_shader_variant *geom_shaders[] = { vs, hs, ds, gs };
-
-   for (unsigned i = 0; i < ARRAY_SIZE(geom_shaders); i++) {
-      const struct ir3_shader_variant *shader = geom_shaders[i];
-      if (!shader)
-         continue;
-
-      bool primid = shader->type != MESA_SHADER_VERTEX &&
-         VALIDREG(ir3_find_sysval_regid(shader, SYSTEM_VALUE_PRIMITIVE_ID));
-
-      tu_cs_emit_pkt4(cs, reg_config[shader->type].reg_pc_xs_out_cntl, 1);
-      if (shader == last_shader) {
-         tu_cs_emit(cs, A6XX_PC_VS_OUT_CNTL_STRIDE_IN_VPC(linkage.max_loc) |
-                        CONDREG(pointsize_regid, A6XX_PC_VS_OUT_CNTL_PSIZE) |
-                        CONDREG(layer_regid, A6XX_PC_VS_OUT_CNTL_LAYER) |
-                        CONDREG(view_regid, A6XX_PC_VS_OUT_CNTL_VIEW) |
-                        COND(primid, A6XX_PC_VS_OUT_CNTL_PRIMITIVE_ID) |
-                        A6XX_PC_VS_OUT_CNTL_CLIP_MASK(clip_cull_mask));
-      } else {
-         tu_cs_emit(cs, COND(primid, A6XX_PC_VS_OUT_CNTL_PRIMITIVE_ID));
-      }
-   }
-
-   /* if vertex_flags somehow gets optimized out, your gonna have a bad time: */
-   if (gs)
-      assert(flags_regid != INVALID_REG);
-
-   tu_cs_emit_pkt4(cs, cfg->reg_sp_xs_primitive_cntl, 1);
-   tu_cs_emit(cs, A6XX_SP_VS_PRIMITIVE_CNTL_OUT(linkage.cnt) |
-                  A6XX_SP_GS_PRIMITIVE_CNTL_FLAGS_REGID(flags_regid));
-
-   tu_cs_emit_pkt4(cs, cfg->reg_vpc_xs_layer_cntl, 1);
-   tu_cs_emit(cs, A6XX_VPC_VS_LAYER_CNTL_LAYERLOC(layer_loc) |
-                  A6XX_VPC_VS_LAYER_CNTL_VIEWLOC(view_loc));
-
-   tu_cs_emit_pkt4(cs, cfg->reg_gras_xs_layer_cntl, 1);
-   tu_cs_emit(cs, CONDREG(layer_regid, A6XX_GRAS_GS_LAYER_CNTL_WRITES_LAYER) |
-                  CONDREG(view_regid, A6XX_GRAS_GS_LAYER_CNTL_WRITES_VIEW));
-
-   tu_cs_emit_regs(cs, A6XX_PC_PRIMID_PASSTHRU(primid_passthru));
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_VPC_CNTL_0, 1);
-   tu_cs_emit(cs, A6XX_VPC_CNTL_0_NUMNONPOSVAR(fs ? fs->total_in : 0) |
-                  COND(fs && fs->total_in, A6XX_VPC_CNTL_0_VARYING) |
-                  A6XX_VPC_CNTL_0_PRIMIDLOC(linkage.primid_loc) |
-                  A6XX_VPC_CNTL_0_VIEWIDLOC(linkage.viewid_loc));
-
-   if (hs) {
-      tu_cs_emit_pkt4(cs, REG_A6XX_PC_TESS_NUM_VERTEX, 1);
-      tu_cs_emit(cs, hs->tess.tcs_vertices_out);
-
-      /* In SPIR-V generated from GLSL, the tessellation primitive params are
-       * are specified in the tess eval shader, but in SPIR-V generated from
-       * HLSL, they are specified in the tess control shader. */
-      const struct ir3_shader_variant *tess =
-         ds->tess.spacing == TESS_SPACING_UNSPECIFIED ? hs : ds;
-      tu_cs_emit_pkt4(cs, REG_A6XX_PC_TESS_CNTL, 1);
-      uint32_t output;
-      if (tess->tess.point_mode)
-         output = TESS_POINTS;
-      else if (tess->tess.primitive_mode == TESS_PRIMITIVE_ISOLINES)
-         output = TESS_LINES;
-      else if (tess->tess.ccw)
-         output = TESS_CCW_TRIS;
-      else
-         output = TESS_CW_TRIS;
-
-      enum a6xx_tess_spacing spacing;
-      switch (tess->tess.spacing) {
-      case TESS_SPACING_EQUAL:
-         spacing = TESS_EQUAL;
-         break;
-      case TESS_SPACING_FRACTIONAL_ODD:
-         spacing = TESS_FRACTIONAL_ODD;
-         break;
-      case TESS_SPACING_FRACTIONAL_EVEN:
-         spacing = TESS_FRACTIONAL_EVEN;
-         break;
-      case TESS_SPACING_UNSPECIFIED:
-      default:
-         unreachable("invalid tess spacing");
-      }
-      tu_cs_emit(cs, A6XX_PC_TESS_CNTL_SPACING(spacing) |
-            A6XX_PC_TESS_CNTL_OUTPUT(output));
-
-      tu6_emit_link_map(cs, vs, hs, SB6_HS_SHADER);
-      tu6_emit_link_map(cs, hs, ds, SB6_DS_SHADER);
-   }
-
-
-   if (gs) {
-      uint32_t vertices_out, invocations, output, vec4_size;
-      uint32_t prev_stage_output_size = ds ? ds->output_size : vs->output_size;
-
-      if (hs) {
-         tu6_emit_link_map(cs, ds, gs, SB6_GS_SHADER);
-      } else {
-         tu6_emit_link_map(cs, vs, gs, SB6_GS_SHADER);
-      }
-      vertices_out = gs->gs.vertices_out - 1;
-      output = primitive_to_tess(gs->gs.output_primitive);
-      invocations = gs->gs.invocations - 1;
-      /* Size of per-primitive alloction in ldlw memory in vec4s. */
-      vec4_size = gs->gs.vertices_in *
-                  DIV_ROUND_UP(prev_stage_output_size, 4);
-
-      tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_5, 1);
-      tu_cs_emit(cs,
-            A6XX_PC_PRIMITIVE_CNTL_5_GS_VERTICES_OUT(vertices_out) |
-            A6XX_PC_PRIMITIVE_CNTL_5_GS_OUTPUT(output) |
-            A6XX_PC_PRIMITIVE_CNTL_5_GS_INVOCATIONS(invocations));
-
-      tu_cs_emit_pkt4(cs, REG_A6XX_VPC_GS_PARAM, 1);
-      tu_cs_emit(cs, 0xff);
-
-      tu_cs_emit_pkt4(cs, REG_A6XX_PC_PRIMITIVE_CNTL_6, 1);
-      tu_cs_emit(cs, A6XX_PC_PRIMITIVE_CNTL_6_STRIDE_IN_VPC(vec4_size));
-
-      uint32_t prim_size = prev_stage_output_size;
-      if (prim_size > 64)
-         prim_size = 64;
-      else if (prim_size == 64)
-         prim_size = 63;
-      tu_cs_emit_pkt4(cs, REG_A6XX_SP_GS_PRIM_SIZE, 1);
-      tu_cs_emit(cs, prim_size);
-   }
+static void
+tu6_emit_fs_system_values(struct tu_cs *cs,
+                          const struct ir3_shader_variant *fs)
+{
+   const uint32_t frontfacing_regid =
+      ir3_find_sysval_regid(fs, SYSTEM_VALUE_FRONT_FACE);
+   const uint32_t sampleid_regid =
+      ir3_find_sysval_regid(fs, SYSTEM_VALUE_SAMPLE_ID);
+   const uint32_t samplemaskin_regid =
+      ir3_find_sysval_regid(fs, SYSTEM_VALUE_SAMPLE_MASK_IN);
+   const uint32_t fragcoord_xy_regid =
+      ir3_find_sysval_regid(fs, SYSTEM_VALUE_FRAG_COORD);
+   const uint32_t fragcoord_zw_regid = (fragcoord_xy_regid != regid(63, 0))
+                                          ? (fragcoord_xy_regid + 2)
+                                          : fragcoord_xy_regid;
+   const uint32_t varyingcoord_regid =
+      ir3_find_sysval_regid(fs, SYSTEM_VALUE_BARYCENTRIC_PIXEL);
 
-   tu6_emit_vpc_varying_modes(cs, fs, last_shader);
+   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CONTROL_1_REG, 5);
+   tu_cs_emit(cs, 0x7);
+   tu_cs_emit(cs, A6XX_HLSQ_CONTROL_2_REG_FACEREGID(frontfacing_regid) |
+                     A6XX_HLSQ_CONTROL_2_REG_SAMPLEID(sampleid_regid) |
+                     A6XX_HLSQ_CONTROL_2_REG_SAMPLEMASK(samplemaskin_regid) |
+                     A6XX_HLSQ_CONTROL_2_REG_SIZE(regid(63, 0)));
+   tu_cs_emit(cs,
+                 A6XX_HLSQ_CONTROL_3_REG_BARY_IJ_PIXEL(varyingcoord_regid) |
+                 A6XX_HLSQ_CONTROL_3_REG_BARY_IJ_CENTROID(regid(63, 0)) |
+                 0xfc00fc00);
+   tu_cs_emit(cs,
+              A6XX_HLSQ_CONTROL_4_REG_XYCOORDREGID(fragcoord_xy_regid) |
+                 A6XX_HLSQ_CONTROL_4_REG_ZWCOORDREGID(fragcoord_zw_regid) |
+                 A6XX_HLSQ_CONTROL_4_REG_BARY_IJ_PIXEL_PERSAMP(regid(63, 0)) |
+                 0x0000fc00);
+   tu_cs_emit(cs, 0xfc);
 }
 
-void
+static void
 tu6_emit_fs_inputs(struct tu_cs *cs, const struct ir3_shader_variant *fs)
 {
-   uint32_t face_regid, coord_regid, zwcoord_regid, samp_id_regid;
-   uint32_t ij_regid[IJ_COUNT];
-   uint32_t smask_in_regid;
-
-   bool sample_shading = fs->per_samp | fs->key.sample_shading;
-   bool enable_varyings = fs->total_in > 0;
-
-   samp_id_regid   = ir3_find_sysval_regid(fs, SYSTEM_VALUE_SAMPLE_ID);
-   smask_in_regid  = ir3_find_sysval_regid(fs, SYSTEM_VALUE_SAMPLE_MASK_IN);
-   face_regid      = ir3_find_sysval_regid(fs, SYSTEM_VALUE_FRONT_FACE);
-   coord_regid     = ir3_find_sysval_regid(fs, SYSTEM_VALUE_FRAG_COORD);
-   zwcoord_regid   = VALIDREG(coord_regid) ? coord_regid + 2 : regid(63, 0);
-   for (unsigned i = 0; i < ARRAY_SIZE(ij_regid); i++)
-      ij_regid[i] = ir3_find_sysval_regid(fs, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL + i);
-
-   if (fs->num_sampler_prefetch > 0) {
-      assert(VALIDREG(ij_regid[IJ_PERSP_PIXEL]));
-      /* also, it seems like ij_pix is *required* to be r0.x */
-      assert(ij_regid[IJ_PERSP_PIXEL] == regid(0, 0));
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_UNKNOWN_B980, 1);
+   tu_cs_emit(cs, fs->total_in > 0 ? 3 : 1);
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_PREFETCH_CNTL, 1 + fs->num_sampler_prefetch);
-   tu_cs_emit(cs, A6XX_SP_FS_PREFETCH_CNTL_COUNT(fs->num_sampler_prefetch) |
-         A6XX_SP_FS_PREFETCH_CNTL_UNK4(regid(63, 0)) |
-         0x7000);    // XXX);
-   for (int i = 0; i < fs->num_sampler_prefetch; i++) {
-      const struct ir3_sampler_prefetch *prefetch = &fs->sampler_prefetch[i];
-      tu_cs_emit(cs, A6XX_SP_FS_PREFETCH_CMD_SRC(prefetch->src) |
-                     A6XX_SP_FS_PREFETCH_CMD_SAMP_ID(prefetch->samp_id) |
-                     A6XX_SP_FS_PREFETCH_CMD_TEX_ID(prefetch->tex_id) |
-                     A6XX_SP_FS_PREFETCH_CMD_DST(prefetch->dst) |
-                     A6XX_SP_FS_PREFETCH_CMD_WRMASK(prefetch->wrmask) |
-                     COND(prefetch->half_precision, A6XX_SP_FS_PREFETCH_CMD_HALF) |
-                     A6XX_SP_FS_PREFETCH_CMD_CMD(prefetch->cmd));
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_UNKNOWN_A982, 1);
+   tu_cs_emit(cs, 0); /* XXX */
 
-   if (fs->num_sampler_prefetch > 0) {
-      tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_BINDLESS_PREFETCH_CMD(0), fs->num_sampler_prefetch);
-      for (int i = 0; i < fs->num_sampler_prefetch; i++) {
-         const struct ir3_sampler_prefetch *prefetch = &fs->sampler_prefetch[i];
-         tu_cs_emit(cs,
-                    A6XX_SP_FS_BINDLESS_PREFETCH_CMD_SAMP_ID(prefetch->samp_bindless_id) |
-                    A6XX_SP_FS_BINDLESS_PREFETCH_CMD_TEX_ID(prefetch->tex_bindless_id));
-      }
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_UPDATE_CNTL, 1);
+   tu_cs_emit(cs, 0xff); /* XXX */
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CONTROL_1_REG, 5);
-   tu_cs_emit(cs, 0x7);
-   tu_cs_emit(cs, A6XX_HLSQ_CONTROL_2_REG_FACEREGID(face_regid) |
-                  A6XX_HLSQ_CONTROL_2_REG_SAMPLEID(samp_id_regid) |
-                  A6XX_HLSQ_CONTROL_2_REG_SAMPLEMASK(smask_in_regid) |
-                  A6XX_HLSQ_CONTROL_2_REG_CENTERRHW(ij_regid[IJ_PERSP_CENTER_RHW]));
-   tu_cs_emit(cs, A6XX_HLSQ_CONTROL_3_REG_IJ_PERSP_PIXEL(ij_regid[IJ_PERSP_PIXEL]) |
-                  A6XX_HLSQ_CONTROL_3_REG_IJ_LINEAR_PIXEL(ij_regid[IJ_LINEAR_PIXEL]) |
-                  A6XX_HLSQ_CONTROL_3_REG_IJ_PERSP_CENTROID(ij_regid[IJ_PERSP_CENTROID]) |
-                  A6XX_HLSQ_CONTROL_3_REG_IJ_LINEAR_CENTROID(ij_regid[IJ_LINEAR_CENTROID]));
-   tu_cs_emit(cs, A6XX_HLSQ_CONTROL_4_REG_XYCOORDREGID(coord_regid) |
-                  A6XX_HLSQ_CONTROL_4_REG_ZWCOORDREGID(zwcoord_regid) |
-                  A6XX_HLSQ_CONTROL_4_REG_IJ_PERSP_SAMPLE(ij_regid[IJ_PERSP_SAMPLE]) |
-                  A6XX_HLSQ_CONTROL_4_REG_IJ_LINEAR_SAMPLE(ij_regid[IJ_LINEAR_SAMPLE]));
-   tu_cs_emit(cs, 0xfcfc);
-
-   enum a6xx_threadsize thrsz = fs->info.double_threadsize ? THREAD128 : THREAD64;
-   tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_FS_CNTL_0, 1);
-   tu_cs_emit(cs, A6XX_HLSQ_FS_CNTL_0_THREADSIZE(thrsz) |
-                  COND(enable_varyings, A6XX_HLSQ_FS_CNTL_0_VARYINGS));
-
-   bool need_size = fs->frag_face || fs->fragcoord_compmask != 0;
-   bool need_size_persamp = false;
-   if (VALIDREG(ij_regid[IJ_PERSP_CENTER_RHW])) {
-      if (sample_shading)
-         need_size_persamp = true;
-      else
-         need_size = true;
+   uint32_t gras_cntl = 0;
+   if (fs->total_in > 0)
+      gras_cntl |= A6XX_GRAS_CNTL_VARYING;
+   if (fs->frag_coord) {
+      gras_cntl |= A6XX_GRAS_CNTL_SIZE | A6XX_GRAS_CNTL_XCOORD |
+                   A6XX_GRAS_CNTL_YCOORD | A6XX_GRAS_CNTL_ZCOORD |
+                   A6XX_GRAS_CNTL_WCOORD;
    }
 
    tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_CNTL, 1);
-   tu_cs_emit(cs,
-         CONDREG(ij_regid[IJ_PERSP_PIXEL], A6XX_GRAS_CNTL_IJ_PERSP_PIXEL) |
-         CONDREG(ij_regid[IJ_PERSP_CENTROID], A6XX_GRAS_CNTL_IJ_PERSP_CENTROID) |
-         CONDREG(ij_regid[IJ_PERSP_SAMPLE], A6XX_GRAS_CNTL_IJ_PERSP_SAMPLE) |
-         CONDREG(ij_regid[IJ_LINEAR_PIXEL], A6XX_GRAS_CNTL_IJ_LINEAR_PIXEL) |
-         CONDREG(ij_regid[IJ_LINEAR_CENTROID], A6XX_GRAS_CNTL_IJ_LINEAR_CENTROID) |
-         CONDREG(ij_regid[IJ_LINEAR_SAMPLE], A6XX_GRAS_CNTL_IJ_LINEAR_SAMPLE) |
-         COND(need_size, A6XX_GRAS_CNTL_IJ_LINEAR_PIXEL) |
-         COND(need_size_persamp, A6XX_GRAS_CNTL_IJ_LINEAR_SAMPLE) |
-         COND(fs->fragcoord_compmask != 0, A6XX_GRAS_CNTL_COORD_MASK(fs->fragcoord_compmask)));
+   tu_cs_emit(cs, gras_cntl);
+
+   uint32_t rb_render_control = 0;
+   if (fs->total_in > 0) {
+      rb_render_control =
+         A6XX_RB_RENDER_CONTROL0_VARYING | A6XX_RB_RENDER_CONTROL0_UNK10;
+   }
+   if (fs->frag_coord) {
+      rb_render_control |=
+         A6XX_RB_RENDER_CONTROL0_SIZE | A6XX_RB_RENDER_CONTROL0_XCOORD |
+         A6XX_RB_RENDER_CONTROL0_YCOORD | A6XX_RB_RENDER_CONTROL0_ZCOORD |
+         A6XX_RB_RENDER_CONTROL0_WCOORD;
+   }
 
    tu_cs_emit_pkt4(cs, REG_A6XX_RB_RENDER_CONTROL0, 2);
-   tu_cs_emit(cs,
-         CONDREG(ij_regid[IJ_PERSP_PIXEL], A6XX_RB_RENDER_CONTROL0_IJ_PERSP_PIXEL) |
-         CONDREG(ij_regid[IJ_PERSP_CENTROID], A6XX_RB_RENDER_CONTROL0_IJ_PERSP_CENTROID) |
-         CONDREG(ij_regid[IJ_PERSP_SAMPLE], A6XX_RB_RENDER_CONTROL0_IJ_PERSP_SAMPLE) |
-         CONDREG(ij_regid[IJ_LINEAR_PIXEL], A6XX_RB_RENDER_CONTROL0_IJ_LINEAR_PIXEL) |
-         CONDREG(ij_regid[IJ_LINEAR_CENTROID], A6XX_RB_RENDER_CONTROL0_IJ_LINEAR_CENTROID) |
-         CONDREG(ij_regid[IJ_LINEAR_SAMPLE], A6XX_RB_RENDER_CONTROL0_IJ_LINEAR_SAMPLE) |
-         COND(need_size, A6XX_RB_RENDER_CONTROL0_IJ_LINEAR_PIXEL) |
-         COND(enable_varyings, A6XX_RB_RENDER_CONTROL0_UNK10) |
-         COND(need_size_persamp, A6XX_RB_RENDER_CONTROL0_IJ_LINEAR_SAMPLE) |
-         COND(fs->fragcoord_compmask != 0,
-                           A6XX_RB_RENDER_CONTROL0_COORD_MASK(fs->fragcoord_compmask)));
-   tu_cs_emit(cs,
-         A6XX_RB_RENDER_CONTROL1_FRAGCOORDSAMPLEMODE(
-            sample_shading ? FRAGCOORD_SAMPLE : FRAGCOORD_CENTER) |
-         CONDREG(smask_in_regid, A6XX_RB_RENDER_CONTROL1_SAMPLEMASK) |
-         CONDREG(samp_id_regid, A6XX_RB_RENDER_CONTROL1_SAMPLEID) |
-         CONDREG(ij_regid[IJ_PERSP_CENTER_RHW], A6XX_RB_RENDER_CONTROL1_CENTERRHW) |
-         COND(fs->frag_face, A6XX_RB_RENDER_CONTROL1_FACENESS));
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_CNTL, 1);
-   tu_cs_emit(cs, COND(sample_shading, A6XX_RB_SAMPLE_CNTL_PER_SAMP_MODE));
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_LRZ_PS_INPUT_CNTL, 1);
-   tu_cs_emit(cs, CONDREG(samp_id_regid, A6XX_GRAS_LRZ_PS_INPUT_CNTL_SAMPLEID) |
-              A6XX_GRAS_LRZ_PS_INPUT_CNTL_FRAGCOORDSAMPLEMODE(
-                 sample_shading ? FRAGCOORD_SAMPLE : FRAGCOORD_CENTER));
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SAMPLE_CNTL, 1);
-   tu_cs_emit(cs, COND(sample_shading, A6XX_GRAS_SAMPLE_CNTL_PER_SAMP_MODE));
+   tu_cs_emit(cs, rb_render_control);
+   tu_cs_emit(cs, (fs->frag_face ? A6XX_RB_RENDER_CONTROL1_FACENESS : 0));
 }
 
 static void
 tu6_emit_fs_outputs(struct tu_cs *cs,
                     const struct ir3_shader_variant *fs,
-                    struct tu_pipeline *pipeline)
+                    uint32_t mrt_count)
 {
-   uint32_t smask_regid, posz_regid, stencilref_regid;
-
-   posz_regid      = ir3_find_output_regid(fs, FRAG_RESULT_DEPTH);
-   smask_regid     = ir3_find_output_regid(fs, FRAG_RESULT_SAMPLE_MASK);
-   stencilref_regid = ir3_find_output_regid(fs, FRAG_RESULT_STENCIL);
-
-   int output_reg_count = 0;
+   const uint32_t fragdepth_regid =
+      ir3_find_output_regid(fs, FRAG_RESULT_DEPTH);
    uint32_t fragdata_regid[8];
-
-   assert(!fs->color0_mrt);
-   for (uint32_t i = 0; i < ARRAY_SIZE(fragdata_regid); i++) {
-      fragdata_regid[i] = ir3_find_output_regid(fs, FRAG_RESULT_DATA0 + i);
-      if (VALIDREG(fragdata_regid[i]))
-         output_reg_count = i + 1;
+   if (fs->color0_mrt) {
+      fragdata_regid[0] = ir3_find_output_regid(fs, FRAG_RESULT_COLOR);
+      for (uint32_t i = 1; i < ARRAY_SIZE(fragdata_regid); i++)
+         fragdata_regid[i] = fragdata_regid[0];
+   } else {
+      for (uint32_t i = 0; i < ARRAY_SIZE(fragdata_regid); i++)
+         fragdata_regid[i] = ir3_find_output_regid(fs, FRAG_RESULT_DATA0 + i);
    }
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_OUTPUT_CNTL0, 1);
-   tu_cs_emit(cs, A6XX_SP_FS_OUTPUT_CNTL0_DEPTH_REGID(posz_regid) |
-                  A6XX_SP_FS_OUTPUT_CNTL0_SAMPMASK_REGID(smask_regid) |
-                  A6XX_SP_FS_OUTPUT_CNTL0_STENCILREF_REGID(stencilref_regid) |
-                  COND(fs->dual_src_blend, A6XX_SP_FS_OUTPUT_CNTL0_DUAL_COLOR_IN_ENABLE));
-
-   /* There is no point in having component enabled which is not written
-    * by the shader. Per VK spec it is an UB, however a few apps depend on
-    * attachment not being changed if FS doesn't have corresponding output.
-    */
-   uint32_t fs_render_components = 0;
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_OUTPUT_CNTL0, 2);
+   tu_cs_emit(
+      cs, A6XX_SP_FS_OUTPUT_CNTL0_DEPTH_REGID(fragdepth_regid) | 0xfcfc0000);
+   tu_cs_emit(cs, A6XX_SP_FS_OUTPUT_CNTL1_MRT(mrt_count));
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_OUTPUT_REG(0), output_reg_count);
-   for (uint32_t i = 0; i < output_reg_count; i++) {
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_OUTPUT_REG(0), 8);
+   for (uint32_t i = 0; i < ARRAY_SIZE(fragdata_regid); i++) {
+      // TODO we could have a mix of half and full precision outputs,
+      // we really need to figure out half-precision from IR3_REG_HALF
       tu_cs_emit(cs, A6XX_SP_FS_OUTPUT_REG_REGID(fragdata_regid[i]) |
-                     (COND(fragdata_regid[i] & HALF_REG_ID,
-                           A6XX_SP_FS_OUTPUT_REG_HALF_PRECISION)));
-
-      if (VALIDREG(fragdata_regid[i])) {
-         fs_render_components |= 0xf << (i * 4);
-      }
+                        (false ? A6XX_SP_FS_OUTPUT_REG_HALF_PRECISION : 0));
    }
 
-   tu_cs_emit_regs(cs,
-                   A6XX_SP_FS_RENDER_COMPONENTS(.dword = fs_render_components));
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_RB_FS_OUTPUT_CNTL0, 1);
-   tu_cs_emit(cs, COND(fs->writes_pos, A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_Z) |
-                  COND(fs->writes_smask, A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_SAMPMASK) |
-                  COND(fs->writes_stencilref, A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_STENCILREF) |
-                  COND(fs->dual_src_blend, A6XX_RB_FS_OUTPUT_CNTL0_DUAL_COLOR_IN_ENABLE));
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_FS_OUTPUT_CNTL0, 2);
+   tu_cs_emit(cs, fs->writes_pos ? A6XX_RB_FS_OUTPUT_CNTL0_FRAG_WRITES_Z : 0);
+   tu_cs_emit(cs, A6XX_RB_FS_OUTPUT_CNTL1_MRT(mrt_count));
 
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_RENDER_COMPONENTS(.dword = fs_render_components));
-
-   if (pipeline) {
-      pipeline->lrz.fs.has_kill = fs->has_kill;
-      pipeline->lrz.fs.early_fragment_tests = fs->fs.early_fragment_tests;
-
-      if (!fs->fs.early_fragment_tests &&
-          (fs->no_earlyz || fs->writes_pos || fs->writes_stencilref || fs->writes_smask)) {
-         pipeline->lrz.force_late_z = true;
-      }
-
-      pipeline->lrz.fs.force_early_z = fs->fs.early_fragment_tests;
+   uint32_t gras_su_depth_plane_cntl = 0;
+   uint32_t rb_depth_plane_cntl = 0;
+   if (fs->no_earlyz | fs->writes_pos) {
+      gras_su_depth_plane_cntl |= A6XX_GRAS_SU_DEPTH_PLANE_CNTL_FRAG_WRITES_Z;
+      rb_depth_plane_cntl |= A6XX_RB_DEPTH_PLANE_CNTL_FRAG_WRITES_Z;
    }
-}
 
-static void
-tu6_emit_vs_params(struct tu_cs *cs,
-                   const struct ir3_const_state *const_state,
-                   unsigned constlen,
-                   unsigned param_stride,
-                   unsigned num_vertices)
-{
-   uint32_t vs_params[4] = {
-      param_stride * num_vertices * 4,  /* vs primitive stride */
-      param_stride * 4,                 /* vs vertex stride */
-      0,
-      0,
-   };
-   uint32_t vs_base = const_state->offsets.primitive_param;
-   tu6_emit_const(cs, CP_LOAD_STATE6_GEOM, vs_base, SB6_VS_SHADER, 0,
-                  ARRAY_SIZE(vs_params), vs_params);
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_DEPTH_PLANE_CNTL, 1);
+   tu_cs_emit(cs, gras_su_depth_plane_cntl);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_PLANE_CNTL, 1);
+   tu_cs_emit(cs, rb_depth_plane_cntl);
 }
 
 static void
-tu_get_tess_iova(struct tu_device *dev,
-                 uint64_t *tess_factor_iova,
-                 uint64_t *tess_param_iova)
-{
-   /* Create the shared tess factor BO the first time tess is used on the device. */
-   if (!dev->tess_bo) {
-      mtx_lock(&dev->mutex);
-      if (!dev->tess_bo)
-         tu_bo_init_new(dev, &dev->tess_bo, TU_TESS_BO_SIZE, TU_BO_ALLOC_NO_FLAGS, "tess");
-      mtx_unlock(&dev->mutex);
+tu6_emit_shader_object(struct tu_cs *cs,
+                       gl_shader_stage stage,
+                       const struct ir3_shader_variant *variant,
+                       const struct tu_bo *binary_bo,
+                       uint32_t binary_offset)
+{
+   uint16_t reg;
+   uint8_t opcode;
+   enum a6xx_state_block sb;
+   switch (stage) {
+   case MESA_SHADER_VERTEX:
+      reg = REG_A6XX_SP_VS_OBJ_START_LO;
+      opcode = CP_LOAD_STATE6_GEOM;
+      sb = SB6_VS_SHADER;
+      break;
+   case MESA_SHADER_TESS_CTRL:
+      reg = REG_A6XX_SP_HS_OBJ_START_LO;
+      opcode = CP_LOAD_STATE6_GEOM;
+      sb = SB6_HS_SHADER;
+      break;
+   case MESA_SHADER_TESS_EVAL:
+      reg = REG_A6XX_SP_DS_OBJ_START_LO;
+      opcode = CP_LOAD_STATE6_GEOM;
+      sb = SB6_DS_SHADER;
+      break;
+   case MESA_SHADER_GEOMETRY:
+      reg = REG_A6XX_SP_GS_OBJ_START_LO;
+      opcode = CP_LOAD_STATE6_GEOM;
+      sb = SB6_GS_SHADER;
+      break;
+   case MESA_SHADER_FRAGMENT:
+      reg = REG_A6XX_SP_FS_OBJ_START_LO;
+      opcode = CP_LOAD_STATE6_FRAG;
+      sb = SB6_FS_SHADER;
+      break;
+   case MESA_SHADER_COMPUTE:
+      reg = REG_A6XX_SP_CS_OBJ_START_LO;
+      opcode = CP_LOAD_STATE6_FRAG;
+      sb = SB6_CS_SHADER;
+      break;
+   default:
+      unreachable("invalid gl_shader_stage");
+      opcode = CP_LOAD_STATE6_GEOM;
+      sb = SB6_VS_SHADER;
+      break;
    }
 
-   *tess_factor_iova = dev->tess_bo->iova;
-   *tess_param_iova = dev->tess_bo->iova + TU_TESS_FACTOR_SIZE;
-}
-
-void
-tu6_emit_patch_control_points(struct tu_cs *cs,
-                              const struct tu_pipeline *pipeline,
-                              unsigned patch_control_points)
-{
-   if (!(pipeline->active_stages & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT))
+   if (!variant->instrlen) {
+      tu_cs_emit_pkt4(cs, reg, 2);
+      tu_cs_emit_qw(cs, 0);
       return;
-
-   struct tu_device *dev = cs->device;
-
-   tu6_emit_vs_params(cs,
-                      &pipeline->program.link[MESA_SHADER_VERTEX].const_state,
-                      pipeline->program.link[MESA_SHADER_VERTEX].constlen,
-                      pipeline->program.vs_param_stride,
-                      patch_control_points);
-
-   uint64_t tess_factor_iova, tess_param_iova;
-   tu_get_tess_iova(dev, &tess_factor_iova, &tess_param_iova);
-
-   uint32_t hs_params[8] = {
-      pipeline->program.vs_param_stride * patch_control_points * 4,  /* hs primitive stride */
-      pipeline->program.vs_param_stride * 4,                         /* hs vertex stride */
-      pipeline->program.hs_param_stride,
-      patch_control_points,
-      tess_param_iova,
-      tess_param_iova >> 32,
-      tess_factor_iova,
-      tess_factor_iova >> 32,
-   };
-
-   const struct ir3_const_state *hs_const =
-      &pipeline->program.link[MESA_SHADER_TESS_CTRL].const_state;
-   uint32_t hs_base = hs_const->offsets.primitive_param;
-   tu6_emit_const(cs, CP_LOAD_STATE6_GEOM, hs_base, SB6_HS_SHADER, 0,
-                  pipeline->program.hs_param_dwords, hs_params);
-
-   uint32_t patch_local_mem_size_16b =
-      patch_control_points * pipeline->program.vs_param_stride / 4;
-
-   /* Total attribute slots in HS incoming patch. */
-   tu_cs_emit_pkt4(cs, REG_A6XX_PC_HS_INPUT_SIZE, 1);
-   tu_cs_emit(cs, patch_local_mem_size_16b);
-
-   const uint32_t wavesize = 64;
-   const uint32_t vs_hs_local_mem_size = 16384;
-
-   uint32_t max_patches_per_wave;
-   if (dev->physical_device->info->a6xx.tess_use_shared) {
-      /* HS invocations for a patch are always within the same wave,
-       * making barriers less expensive. VS can't have barriers so we
-       * don't care about VS invocations being in the same wave.
-       */
-      max_patches_per_wave = wavesize / pipeline->program.hs_vertices_out;
-   } else {
-      /* VS is also in the same wave */
-      max_patches_per_wave =
-         wavesize / MAX2(patch_control_points,
-                         pipeline->program.hs_vertices_out);
    }
 
-   uint32_t patches_per_wave =
-      MIN2(vs_hs_local_mem_size / (patch_local_mem_size_16b * 16),
-           max_patches_per_wave);
-
-   uint32_t wave_input_size = DIV_ROUND_UP(
-      patches_per_wave * patch_local_mem_size_16b * 16, 256);
+   assert(variant->type == stage);
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_SP_HS_WAVE_INPUT_SIZE, 1);
-   tu_cs_emit(cs, wave_input_size);
-
-   /* maximum number of patches that can fit in tess factor/param buffers */
-   uint32_t subdraw_size = MIN2(TU_TESS_FACTOR_SIZE / ir3_tess_factor_stride(pipeline->tess.patch_type),
-                        TU_TESS_PARAM_SIZE / (pipeline->program.hs_param_stride * 4));
-   /* convert from # of patches to draw count */
-   subdraw_size *= patch_control_points;
-
-   tu_cs_emit_pkt7(cs, CP_SET_SUBDRAW_SIZE, 1);
-   tu_cs_emit(cs, subdraw_size);
-}
+   const uint64_t binary_iova = binary_bo->iova + binary_offset;
+   assert((binary_iova & 0x3) == 0);
 
-static void
-tu6_emit_geom_tess_consts(struct tu_cs *cs,
-                          const struct ir3_shader_variant *vs,
-                          const struct ir3_shader_variant *hs,
-                          const struct ir3_shader_variant *ds,
-                          const struct ir3_shader_variant *gs)
-{
-   struct tu_device *dev = cs->device;
-
-   if (gs && !hs) {
-      tu6_emit_vs_params(cs, ir3_const_state(vs), vs->constlen,
-                         vs->output_size, gs->gs.vertices_in);
-   }
+   tu_cs_emit_pkt4(cs, reg, 2);
+   tu_cs_emit_qw(cs, binary_iova);
 
-   if (hs) {
-      uint64_t tess_factor_iova, tess_param_iova;
-      tu_get_tess_iova(dev, &tess_factor_iova, &tess_param_iova);
-
-      uint32_t ds_params[8] = {
-         gs ? ds->output_size * gs->gs.vertices_in * 4 : 0,  /* ds primitive stride */
-         ds->output_size * 4,                                /* ds vertex stride */
-         hs->output_size,                                    /* hs vertex stride (dwords) */
-         hs->tess.tcs_vertices_out,
-         tess_param_iova,
-         tess_param_iova >> 32,
-         tess_factor_iova,
-         tess_factor_iova >> 32,
-      };
-
-      uint32_t ds_base = ds->const_state->offsets.primitive_param;
-      uint32_t ds_param_dwords = MIN2((ds->constlen - ds_base) * 4, ARRAY_SIZE(ds_params));
-      tu6_emit_const(cs, CP_LOAD_STATE6_GEOM, ds_base, SB6_DS_SHADER, 0,
-                     ds_param_dwords, ds_params);
-   }
+   /* always indirect */
+   const bool indirect = true;
+   if (indirect) {
+      tu_cs_emit_pkt7(cs, opcode, 3);
+      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(0) |
+                        CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
+                        CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
+                        CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
+                        CP_LOAD_STATE6_0_NUM_UNIT(variant->instrlen));
+      tu_cs_emit_qw(cs, binary_iova);
+   } else {
+      const void *binary = binary_bo->map + binary_offset;
 
-   if (gs) {
-      const struct ir3_shader_variant *prev = ds ? ds : vs;
-      uint32_t gs_params[4] = {
-         prev->output_size * gs->gs.vertices_in * 4,  /* gs primitive stride */
-         prev->output_size * 4,                 /* gs vertex stride */
-         0,
-         0,
-      };
-      uint32_t gs_base = gs->const_state->offsets.primitive_param;
-      tu6_emit_const(cs, CP_LOAD_STATE6_GEOM, gs_base, SB6_GS_SHADER, 0,
-                     ARRAY_SIZE(gs_params), gs_params);
+      tu_cs_emit_pkt7(cs, opcode, 3 + variant->info.sizedwords);
+      tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(0) |
+                        CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
+                        CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
+                        CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
+                        CP_LOAD_STATE6_0_NUM_UNIT(variant->instrlen));
+      tu_cs_emit_qw(cs, 0);
+      tu_cs_emit_array(cs, binary, variant->info.sizedwords);
    }
 }
 
 static void
-tu6_emit_program_config(struct tu_cs *cs,
-                        struct tu_pipeline_builder *builder)
+tu6_emit_program(struct tu_cs *cs,
+                 const struct tu_pipeline_builder *builder,
+                 const struct tu_bo *binary_bo,
+                 bool binning_pass)
 {
-   gl_shader_stage stage = MESA_SHADER_VERTEX;
-
-   STATIC_ASSERT(MESA_SHADER_VERTEX == 0);
-
-   bool shared_consts_enable = tu6_shared_constants_enable(&builder->layout,
-         builder->device->compiler);
-   tu6_emit_shared_consts_enable(cs, shared_consts_enable);
-
-   tu_cs_emit_regs(cs, A6XX_HLSQ_INVALIDATE_CMD(
-         .vs_state = true,
-         .hs_state = true,
-         .ds_state = true,
-         .gs_state = true,
-         .fs_state = true,
-         .gfx_ibo = true,
-         .gfx_shared_const = shared_consts_enable));
-   for (; stage < ARRAY_SIZE(builder->shader_iova); stage++) {
-      tu6_emit_xs_config(cs, stage, builder->variants[stage]);
-   }
+   static const struct ir3_shader_variant dummy_variant = {
+      .type = MESA_SHADER_NONE
+   };
+   assert(builder->shaders[MESA_SHADER_VERTEX]);
+   const struct ir3_shader_variant *vs =
+      &builder->shaders[MESA_SHADER_VERTEX]->variants[0];
+   const struct ir3_shader_variant *hs =
+      builder->shaders[MESA_SHADER_TESS_CTRL]
+         ? &builder->shaders[MESA_SHADER_TESS_CTRL]->variants[0]
+         : &dummy_variant;
+   const struct ir3_shader_variant *ds =
+      builder->shaders[MESA_SHADER_TESS_EVAL]
+         ? &builder->shaders[MESA_SHADER_TESS_EVAL]->variants[0]
+         : &dummy_variant;
+   const struct ir3_shader_variant *gs =
+      builder->shaders[MESA_SHADER_GEOMETRY]
+         ? &builder->shaders[MESA_SHADER_GEOMETRY]->variants[0]
+         : &dummy_variant;
+   const struct ir3_shader_variant *fs =
+      builder->shaders[MESA_SHADER_FRAGMENT]
+         ? &builder->shaders[MESA_SHADER_FRAGMENT]->variants[0]
+         : &dummy_variant;
+
+   if (binning_pass) {
+      vs = &builder->shaders[MESA_SHADER_VERTEX]->variants[1];
+      fs = &dummy_variant;
+   }
+
+   tu6_emit_vs_config(cs, vs);
+   tu6_emit_hs_config(cs, hs);
+   tu6_emit_ds_config(cs, ds);
+   tu6_emit_gs_config(cs, gs);
+   tu6_emit_fs_config(cs, fs);
+
+   tu6_emit_vs_system_values(cs, vs);
+   tu6_emit_vpc(cs, vs, fs, binning_pass);
+   tu6_emit_vpc_varying_modes(cs, fs, binning_pass);
+   tu6_emit_fs_system_values(cs, fs);
+   tu6_emit_fs_inputs(cs, fs);
+   tu6_emit_fs_outputs(cs, fs, builder->color_attachment_count);
+
+   tu6_emit_shader_object(cs, MESA_SHADER_VERTEX, vs, binary_bo,
+                          builder->shader_offsets[MESA_SHADER_VERTEX]);
+
+   tu6_emit_shader_object(cs, MESA_SHADER_FRAGMENT, fs, binary_bo,
+                          builder->shader_offsets[MESA_SHADER_FRAGMENT]);
 }
 
 static void
-tu6_emit_program(struct tu_cs *cs,
-                 struct tu_pipeline_builder *builder,
-                 bool binning_pass,
-                 struct tu_pipeline *pipeline)
-{
-   const struct ir3_shader_variant *vs = builder->variants[MESA_SHADER_VERTEX];
-   const struct ir3_shader_variant *bs = builder->binning_variant;
-   const struct ir3_shader_variant *hs = builder->variants[MESA_SHADER_TESS_CTRL];
-   const struct ir3_shader_variant *ds = builder->variants[MESA_SHADER_TESS_EVAL];
-   const struct ir3_shader_variant *gs = builder->variants[MESA_SHADER_GEOMETRY];
-   const struct ir3_shader_variant *fs = builder->variants[MESA_SHADER_FRAGMENT];
-   gl_shader_stage stage = MESA_SHADER_VERTEX;
-   bool multi_pos_output = vs->multi_pos_output;
-
-  /* Don't use the binning pass variant when GS is present because we don't
-   * support compiling correct binning pass variants with GS.
-   */
-   if (binning_pass && !gs) {
-      vs = bs;
-      tu6_emit_xs(cs, stage, bs, &builder->pvtmem, builder->binning_vs_iova);
-      tu6_emit_dynamic_offset(cs, bs, builder);
-      stage++;
-   }
-
-   for (; stage < ARRAY_SIZE(builder->shader_iova); stage++) {
-      const struct ir3_shader_variant *xs = builder->variants[stage];
-
-      if (stage == MESA_SHADER_FRAGMENT && binning_pass)
-         fs = xs = NULL;
+tu6_emit_vertex_input(struct tu_cs *cs,
+                      const struct ir3_shader_variant *vs,
+                      const VkPipelineVertexInputStateCreateInfo *vi_info,
+                      uint8_t bindings[MAX_VERTEX_ATTRIBS],
+                      uint16_t strides[MAX_VERTEX_ATTRIBS],
+                      uint16_t offsets[MAX_VERTEX_ATTRIBS],
+                      uint32_t *count)
+{
+   uint32_t vfd_decode_idx = 0;
+
+   /* why do we go beyond inputs_count? */
+   assert(vs->inputs_count + 1 <= MAX_VERTEX_ATTRIBS);
+   for (uint32_t i = 0; i <= vs->inputs_count; i++) {
+      if (vs->inputs[i].sysval || !vs->inputs[i].compmask)
+         continue;
 
-      tu6_emit_xs(cs, stage, xs, &builder->pvtmem, builder->shader_iova[stage]);
-      tu6_emit_dynamic_offset(cs, xs, builder);
-   }
+      const VkVertexInputAttributeDescription *vi_attr =
+         tu_find_vertex_input_attribute(vi_info, vs->inputs[i].slot);
+      const VkVertexInputBindingDescription *vi_binding =
+         tu_find_vertex_input_binding(vi_info, vi_attr);
+      assert(vi_attr && vi_binding);
 
-   uint32_t multiview_views = util_logbase2(pipeline->rast.multiview_mask) + 1;
-   uint32_t multiview_cntl = pipeline->rast.multiview_mask ?
-      A6XX_PC_MULTIVIEW_CNTL_ENABLE |
-      A6XX_PC_MULTIVIEW_CNTL_VIEWS(multiview_views) |
-      COND(!multi_pos_output, A6XX_PC_MULTIVIEW_CNTL_DISABLEMULTIPOS)
-      : 0;
+      const struct tu_native_format *format =
+         tu6_get_native_format(vi_attr->format);
+      assert(format && format->vtx >= 0);
 
-   /* Copy what the blob does here. This will emit an extra 0x3f
-    * CP_EVENT_WRITE when multiview is disabled. I'm not exactly sure what
-    * this is working around yet.
-    */
-   if (builder->device->physical_device->info->a6xx.has_cp_reg_write) {
-      tu_cs_emit_pkt7(cs, CP_REG_WRITE, 3);
-      tu_cs_emit(cs, CP_REG_WRITE_0_TRACKER(UNK_EVENT_WRITE));
-      tu_cs_emit(cs, REG_A6XX_PC_MULTIVIEW_CNTL);
-   } else {
-      tu_cs_emit_pkt4(cs, REG_A6XX_PC_MULTIVIEW_CNTL, 1);
-   }
-   tu_cs_emit(cs, multiview_cntl);
+      uint32_t vfd_decode = A6XX_VFD_DECODE_INSTR_IDX(vfd_decode_idx) |
+                            A6XX_VFD_DECODE_INSTR_FORMAT(format->vtx) |
+                            A6XX_VFD_DECODE_INSTR_SWAP(format->swap) |
+                            A6XX_VFD_DECODE_INSTR_UNK30;
+      if (vi_binding->inputRate == VK_VERTEX_INPUT_RATE_INSTANCE)
+         vfd_decode |= A6XX_VFD_DECODE_INSTR_INSTANCED;
+      if (!vk_format_is_int(vi_attr->format))
+         vfd_decode |= A6XX_VFD_DECODE_INSTR_FLOAT;
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_VFD_MULTIVIEW_CNTL, 1);
-   tu_cs_emit(cs, multiview_cntl);
+      const uint32_t vfd_decode_step_rate = 1;
 
-   if (multiview_cntl &&
-       builder->device->physical_device->info->a6xx.supports_multiview_mask) {
-      tu_cs_emit_pkt4(cs, REG_A6XX_PC_MULTIVIEW_MASK, 1);
-      tu_cs_emit(cs, pipeline->rast.multiview_mask);
-   }
+      const uint32_t vfd_dest_cntl =
+         A6XX_VFD_DEST_CNTL_INSTR_WRITEMASK(vs->inputs[i].compmask) |
+         A6XX_VFD_DEST_CNTL_INSTR_REGID(vs->inputs[i].regid);
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_SP_HS_WAVE_INPUT_SIZE, 1);
-   tu_cs_emit(cs, 0);
+      tu_cs_emit_pkt4(cs, REG_A6XX_VFD_DECODE(vfd_decode_idx), 2);
+      tu_cs_emit(cs, vfd_decode);
+      tu_cs_emit(cs, vfd_decode_step_rate);
 
-   tu6_emit_vfd_dest(cs, vs);
+      tu_cs_emit_pkt4(cs, REG_A6XX_VFD_DEST_CNTL(vfd_decode_idx), 1);
+      tu_cs_emit(cs, vfd_dest_cntl);
 
-   tu6_emit_vpc(cs, vs, hs, ds, gs, fs);
+      bindings[vfd_decode_idx] = vi_binding->binding;
+      strides[vfd_decode_idx] = vi_binding->stride;
+      offsets[vfd_decode_idx] = vi_attr->offset;
 
-   if (fs) {
-      tu6_emit_fs_inputs(cs, fs);
-      tu6_emit_fs_outputs(cs, fs, pipeline);
-   } else {
-      /* TODO: check if these can be skipped if fs is disabled */
-      struct ir3_shader_variant dummy_variant = {};
-      tu6_emit_fs_inputs(cs, &dummy_variant);
-      tu6_emit_fs_outputs(cs, &dummy_variant, NULL);
+      vfd_decode_idx++;
    }
 
-   if (gs || hs) {
-      tu6_emit_geom_tess_consts(cs, vs, hs, ds, gs);
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_VFD_CONTROL_0, 1);
+   tu_cs_emit(
+      cs, A6XX_VFD_CONTROL_0_VTXCNT(vfd_decode_idx) | (vfd_decode_idx << 8));
+
+   *count = vfd_decode_idx;
 }
 
-void
-tu6_emit_vertex_input(struct tu_cs *cs,
-                      uint32_t binding_count,
-                      const VkVertexInputBindingDescription2EXT *bindings,
-                      uint32_t unsorted_attr_count,
-                      const VkVertexInputAttributeDescription2EXT *unsorted_attrs)
+static uint32_t
+tu6_guardband_adj(uint32_t v)
 {
-   uint32_t binding_instanced = 0; /* bitmask of instanced bindings */
-   uint32_t step_rate[MAX_VBS];
-
-   for (uint32_t i = 0; i < binding_count; i++) {
-      const VkVertexInputBindingDescription2EXT *binding = &bindings[i];
-
-      if (binding->inputRate == VK_VERTEX_INPUT_RATE_INSTANCE)
-         binding_instanced |= 1u << binding->binding;
-
-      step_rate[binding->binding] = binding->divisor;
-   }
-
-   const VkVertexInputAttributeDescription2EXT *attrs[MAX_VERTEX_ATTRIBS] = { };
-   unsigned attr_count = 0;
-   for (uint32_t i = 0; i < unsorted_attr_count; i++) {
-      const VkVertexInputAttributeDescription2EXT *attr = &unsorted_attrs[i];
-      attrs[attr->location] = attr;
-      attr_count = MAX2(attr_count, attr->location + 1);
-   }
-
-   if (attr_count != 0)
-      tu_cs_emit_pkt4(cs, REG_A6XX_VFD_DECODE_INSTR(0), attr_count * 2);
-
-   for (uint32_t loc = 0; loc < attr_count; loc++) {
-      const VkVertexInputAttributeDescription2EXT *attr = attrs[loc];
-
-      if (attr) {
-         const struct tu_native_format format = tu6_format_vtx(attr->format);
-         tu_cs_emit(cs, A6XX_VFD_DECODE_INSTR(0,
-                          .idx = attr->binding,
-                          .offset = attr->offset,
-                          .instanced = binding_instanced & (1 << attr->binding),
-                          .format = format.fmt,
-                          .swap = format.swap,
-                          .unk30 = 1,
-                          ._float = !vk_format_is_int(attr->format)).value);
-         tu_cs_emit(cs, A6XX_VFD_DECODE_STEP_RATE(0, step_rate[attr->binding]).value);
-      } else {
-         tu_cs_emit(cs, 0);
-         tu_cs_emit(cs, 0);
-      }
-   }
+   if (v > 256)
+      return (uint32_t)(511.0 - 65.0 * (log2(v) - 8.0));
+   else
+      return 511;
 }
 
 void
-tu6_emit_viewport(struct tu_cs *cs, const VkViewport *viewports, uint32_t num_viewport,
-                  bool z_negative_one_to_one)
-{
-   VkExtent2D guardband = {511, 511};
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_CL_VPORT_XOFFSET(0), num_viewport * 6);
-   for (uint32_t i = 0; i < num_viewport; i++) {
-      const VkViewport *viewport = &viewports[i];
-      float offsets[3];
-      float scales[3];
-      scales[0] = viewport->width / 2.0f;
-      scales[1] = viewport->height / 2.0f;
-      if (z_negative_one_to_one) {
-         scales[2] = 0.5 * (viewport->maxDepth - viewport->minDepth);
-      } else {
-         scales[2] = viewport->maxDepth - viewport->minDepth;
-      }
-
-      offsets[0] = viewport->x + scales[0];
-      offsets[1] = viewport->y + scales[1];
-      if (z_negative_one_to_one) {
-         offsets[2] = 0.5 * (viewport->minDepth + viewport->maxDepth);
-      } else {
-         offsets[2] = viewport->minDepth;
-      }
-
-      for (uint32_t j = 0; j < 3; j++) {
-         tu_cs_emit(cs, fui(offsets[j]));
-         tu_cs_emit(cs, fui(scales[j]));
-      }
-
-      guardband.width =
-         MIN2(guardband.width, fd_calc_guardband(offsets[0], scales[0], false));
-      guardband.height =
-         MIN2(guardband.height, fd_calc_guardband(offsets[1], scales[1], false));
-   }
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL(0), num_viewport * 2);
-   for (uint32_t i = 0; i < num_viewport; i++) {
-      const VkViewport *viewport = &viewports[i];
-      VkOffset2D min;
-      VkOffset2D max;
-      min.x = (int32_t) viewport->x;
-      max.x = (int32_t) ceilf(viewport->x + viewport->width);
-      if (viewport->height >= 0.0f) {
-         min.y = (int32_t) viewport->y;
-         max.y = (int32_t) ceilf(viewport->y + viewport->height);
-      } else {
-         min.y = (int32_t)(viewport->y + viewport->height);
-         max.y = (int32_t) ceilf(viewport->y);
-      }
-      /* the spec allows viewport->height to be 0.0f */
-      if (min.y == max.y)
-         max.y++;
-      /* allow viewport->width = 0.0f for un-initialized viewports: */
-      if (min.x == max.x)
-         max.x++;
-
-      min.x = MAX2(min.x, 0);
-      min.y = MAX2(min.y, 0);
-      max.x = MAX2(max.x, 1);
-      max.y = MAX2(max.y, 1);
-
-      assert(min.x < max.x);
-      assert(min.y < max.y);
-
-      tu_cs_emit(cs, A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_X(min.x) |
-                     A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_Y(min.y));
-      tu_cs_emit(cs, A6XX_GRAS_SC_VIEWPORT_SCISSOR_BR_X(max.x - 1) |
-                     A6XX_GRAS_SC_VIEWPORT_SCISSOR_BR_Y(max.y - 1));
-   }
+tu6_emit_viewport(struct tu_cs *cs, const VkViewport *viewport)
+{
+   float offsets[3];
+   float scales[3];
+   scales[0] = viewport->width / 2.0f;
+   scales[1] = viewport->height / 2.0f;
+   scales[2] = viewport->maxDepth - viewport->minDepth;
+   offsets[0] = viewport->x + scales[0];
+   offsets[1] = viewport->y + scales[1];
+   offsets[2] = viewport->minDepth;
+
+   VkOffset2D min;
+   VkOffset2D max;
+   min.x = (int32_t) viewport->x;
+   max.x = (int32_t) ceilf(viewport->x + viewport->width);
+   if (viewport->height >= 0.0f) {
+      min.y = (int32_t) viewport->y;
+      max.y = (int32_t) ceilf(viewport->y + viewport->height);
+   } else {
+      min.y = (int32_t)(viewport->y + viewport->height);
+      max.y = (int32_t) ceilf(viewport->y);
+   }
+   /* the spec allows viewport->height to be 0.0f */
+   if (min.y == max.y)
+      max.y++;
+   assert(min.x >= 0 && min.x < max.x);
+   assert(min.y >= 0 && min.y < max.y);
+
+   VkExtent2D guardband_adj;
+   guardband_adj.width = tu6_guardband_adj(max.x - min.x);
+   guardband_adj.height = tu6_guardband_adj(max.y - min.y);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_CL_VPORT_XOFFSET_0, 6);
+   tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_XOFFSET_0(offsets[0]));
+   tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_XSCALE_0(scales[0]));
+   tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_YOFFSET_0(offsets[1]));
+   tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_YSCALE_0(scales[1]));
+   tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_ZOFFSET_0(offsets[2]));
+   tu_cs_emit(cs, A6XX_GRAS_CL_VPORT_ZSCALE_0(scales[2]));
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0, 2);
+   tu_cs_emit(cs, A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_X(min.x) |
+                     A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_Y(min.y));
+   tu_cs_emit(cs, A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_X(max.x - 1) |
+                     A6XX_GRAS_SC_VIEWPORT_SCISSOR_TL_0_Y(max.y - 1));
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_CL_Z_CLAMP(0), num_viewport * 2);
-   for (uint32_t i = 0; i < num_viewport; i++) {
-      const VkViewport *viewport = &viewports[i];
-      tu_cs_emit(cs, fui(MIN2(viewport->minDepth, viewport->maxDepth)));
-      tu_cs_emit(cs, fui(MAX2(viewport->minDepth, viewport->maxDepth)));
-   }
    tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ, 1);
-   tu_cs_emit(cs, A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ_HORZ(guardband.width) |
-                  A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ_VERT(guardband.height));
-
-   /* TODO: what to do about this and multi viewport ? */
-   float z_clamp_min = num_viewport ? MIN2(viewports[0].minDepth, viewports[0].maxDepth) : 0;
-   float z_clamp_max = num_viewport ? MAX2(viewports[0].minDepth, viewports[0].maxDepth) : 0;
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_Z_CLAMP_MIN(z_clamp_min),
-                   A6XX_RB_Z_CLAMP_MAX(z_clamp_max));
+   tu_cs_emit(cs,
+              A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ_HORZ(guardband_adj.width) |
+                 A6XX_GRAS_CL_GUARDBAND_CLIP_ADJ_VERT(guardband_adj.height));
 }
 
 void
-tu6_emit_scissor(struct tu_cs *cs, const VkRect2D *scissors, uint32_t scissor_count)
+tu6_emit_scissor(struct tu_cs *cs, const VkRect2D *scissor)
 {
-   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SC_SCREEN_SCISSOR_TL(0), scissor_count * 2);
-
-   for (uint32_t i = 0; i < scissor_count; i++) {
-      const VkRect2D *scissor = &scissors[i];
-
-      uint32_t min_x = scissor->offset.x;
-      uint32_t min_y = scissor->offset.y;
-      uint32_t max_x = min_x + scissor->extent.width - 1;
-      uint32_t max_y = min_y + scissor->extent.height - 1;
-
-      if (!scissor->extent.width || !scissor->extent.height) {
-         min_x = min_y = 1;
-         max_x = max_y = 0;
-      } else {
-         /* avoid overflow */
-         uint32_t scissor_max = BITFIELD_MASK(15);
-         min_x = MIN2(scissor_max, min_x);
-         min_y = MIN2(scissor_max, min_y);
-         max_x = MIN2(scissor_max, max_x);
-         max_y = MIN2(scissor_max, max_y);
-      }
+   const VkOffset2D min = scissor->offset;
+   const VkOffset2D max = {
+      scissor->offset.x + scissor->extent.width,
+      scissor->offset.y + scissor->extent.height,
+   };
 
-      tu_cs_emit(cs, A6XX_GRAS_SC_SCREEN_SCISSOR_TL_X(min_x) |
-                     A6XX_GRAS_SC_SCREEN_SCISSOR_TL_Y(min_y));
-      tu_cs_emit(cs, A6XX_GRAS_SC_SCREEN_SCISSOR_BR_X(max_x) |
-                     A6XX_GRAS_SC_SCREEN_SCISSOR_BR_Y(max_y));
-   }
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0, 2);
+   tu_cs_emit(cs, A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_X(min.x) |
+                     A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_Y(min.y));
+   tu_cs_emit(cs, A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_X(max.x - 1) |
+                     A6XX_GRAS_SC_SCREEN_SCISSOR_TL_0_Y(max.y - 1));
 }
 
-void
-tu6_emit_sample_locations(struct tu_cs *cs, const VkSampleLocationsInfoEXT *samp_loc)
+static void
+tu6_emit_gras_unknowns(struct tu_cs *cs)
 {
-   if (!samp_loc) {
-      tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SAMPLE_CONFIG, 1);
-      tu_cs_emit(cs, 0);
-
-      tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_CONFIG, 1);
-      tu_cs_emit(cs, 0);
-
-      tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_SAMPLE_CONFIG, 1);
-      tu_cs_emit(cs, 0);
-      return;
-   }
-
-   assert(samp_loc->sampleLocationsPerPixel == samp_loc->sampleLocationsCount);
-   assert(samp_loc->sampleLocationGridSize.width == 1);
-   assert(samp_loc->sampleLocationGridSize.height == 1);
-
-   uint32_t sample_config =
-      A6XX_RB_SAMPLE_CONFIG_LOCATION_ENABLE;
-   uint32_t sample_locations = 0;
-   for (uint32_t i = 0; i < samp_loc->sampleLocationsCount; i++) {
-      sample_locations |=
-         (A6XX_RB_SAMPLE_LOCATION_0_SAMPLE_0_X(samp_loc->pSampleLocations[i].x) |
-          A6XX_RB_SAMPLE_LOCATION_0_SAMPLE_0_Y(samp_loc->pSampleLocations[i].y)) << i*8;
-   }
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SAMPLE_CONFIG, 2);
-   tu_cs_emit(cs, sample_config);
-   tu_cs_emit(cs, sample_locations);
-
-   tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_CONFIG, 2);
-   tu_cs_emit(cs, sample_config);
-   tu_cs_emit(cs, sample_locations);
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_UNKNOWN_8000, 1);
+   tu_cs_emit(cs, 0x80);
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_UNKNOWN_8001, 1);
+   tu_cs_emit(cs, 0x0);
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_UNKNOWN_8004, 1);
+   tu_cs_emit(cs, 0x0);
+}
 
-   tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_SAMPLE_CONFIG, 2);
-   tu_cs_emit(cs, sample_config);
-   tu_cs_emit(cs, sample_locations);
+static void
+tu6_emit_point_size(struct tu_cs *cs)
+{
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_POINT_MINMAX, 2);
+   tu_cs_emit(cs, A6XX_GRAS_SU_POINT_MINMAX_MIN(1.0f / 16.0f) |
+                     A6XX_GRAS_SU_POINT_MINMAX_MAX(4092.0f));
+   tu_cs_emit(cs, A6XX_GRAS_SU_POINT_SIZE(1.0f));
 }
 
 static uint32_t
 tu6_gras_su_cntl(const VkPipelineRasterizationStateCreateInfo *rast_info,
-                 enum a5xx_line_mode line_mode,
-                 bool multiview)
+                 VkSampleCountFlagBits samples)
 {
    uint32_t gras_su_cntl = 0;
 
@@ -2194,33 +1114,117 @@ tu6_gras_su_cntl(const VkPipelineRasterizationStateCreateInfo *rast_info,
    if (rast_info->frontFace == VK_FRONT_FACE_CLOCKWISE)
       gras_su_cntl |= A6XX_GRAS_SU_CNTL_FRONT_CW;
 
-   gras_su_cntl |=
-      A6XX_GRAS_SU_CNTL_LINEHALFWIDTH(rast_info->lineWidth / 2.0f);
+   /* don't set A6XX_GRAS_SU_CNTL_LINEHALFWIDTH */
 
    if (rast_info->depthBiasEnable)
       gras_su_cntl |= A6XX_GRAS_SU_CNTL_POLY_OFFSET;
 
-   gras_su_cntl |= A6XX_GRAS_SU_CNTL_LINE_MODE(line_mode);
-
-   if (multiview) {
-      gras_su_cntl |=
-         A6XX_GRAS_SU_CNTL_UNK17 |
-         A6XX_GRAS_SU_CNTL_MULTIVIEW_ENABLE;
-   }
+   if (samples > VK_SAMPLE_COUNT_1_BIT)
+      gras_su_cntl |= A6XX_GRAS_SU_CNTL_MSAA_ENABLE;
 
    return gras_su_cntl;
 }
 
 void
+tu6_emit_gras_su_cntl(struct tu_cs *cs,
+                      uint32_t gras_su_cntl,
+                      float line_width)
+{
+   assert((gras_su_cntl & A6XX_GRAS_SU_CNTL_LINEHALFWIDTH__MASK) == 0);
+   gras_su_cntl |= A6XX_GRAS_SU_CNTL_LINEHALFWIDTH(line_width / 2.0f);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_CNTL, 1);
+   tu_cs_emit(cs, gras_su_cntl);
+}
+
+void
 tu6_emit_depth_bias(struct tu_cs *cs,
                     float constant_factor,
                     float clamp,
                     float slope_factor)
 {
    tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_POLY_OFFSET_SCALE, 3);
-   tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_SCALE(slope_factor).value);
-   tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_OFFSET(constant_factor).value);
-   tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_OFFSET_CLAMP(clamp).value);
+   tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_SCALE(slope_factor));
+   tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_OFFSET(constant_factor));
+   tu_cs_emit(cs, A6XX_GRAS_SU_POLY_OFFSET_OFFSET_CLAMP(clamp));
+}
+
+static void
+tu6_emit_alpha_control_disable(struct tu_cs *cs)
+{
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_ALPHA_CONTROL, 1);
+   tu_cs_emit(cs, 0);
+}
+
+static void
+tu6_emit_depth_control(struct tu_cs *cs,
+                       const VkPipelineDepthStencilStateCreateInfo *ds_info)
+{
+   assert(!ds_info->depthBoundsTestEnable);
+
+   uint32_t rb_depth_cntl = 0;
+   if (ds_info->depthTestEnable) {
+      rb_depth_cntl |=
+         A6XX_RB_DEPTH_CNTL_Z_ENABLE |
+         A6XX_RB_DEPTH_CNTL_ZFUNC(tu6_compare_func(ds_info->depthCompareOp)) |
+         A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE;
+
+      if (ds_info->depthWriteEnable)
+         rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_WRITE_ENABLE;
+   }
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_CNTL, 1);
+   tu_cs_emit(cs, rb_depth_cntl);
+}
+
+static void
+tu6_emit_stencil_control(struct tu_cs *cs,
+                         const VkPipelineDepthStencilStateCreateInfo *ds_info)
+{
+   uint32_t rb_stencil_control = 0;
+   if (ds_info->stencilTestEnable) {
+      const VkStencilOpState *front = &ds_info->front;
+      const VkStencilOpState *back = &ds_info->back;
+      rb_stencil_control |=
+         A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE |
+         A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE_BF |
+         A6XX_RB_STENCIL_CONTROL_STENCIL_READ |
+         A6XX_RB_STENCIL_CONTROL_FUNC(tu6_compare_func(front->compareOp)) |
+         A6XX_RB_STENCIL_CONTROL_FAIL(tu6_stencil_op(front->failOp)) |
+         A6XX_RB_STENCIL_CONTROL_ZPASS(tu6_stencil_op(front->passOp)) |
+         A6XX_RB_STENCIL_CONTROL_ZFAIL(tu6_stencil_op(front->depthFailOp)) |
+         A6XX_RB_STENCIL_CONTROL_FUNC_BF(tu6_compare_func(back->compareOp)) |
+         A6XX_RB_STENCIL_CONTROL_FAIL_BF(tu6_stencil_op(back->failOp)) |
+         A6XX_RB_STENCIL_CONTROL_ZPASS_BF(tu6_stencil_op(back->passOp)) |
+         A6XX_RB_STENCIL_CONTROL_ZFAIL_BF(tu6_stencil_op(back->depthFailOp));
+   }
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCIL_CONTROL, 1);
+   tu_cs_emit(cs, rb_stencil_control);
+}
+
+void
+tu6_emit_stencil_compare_mask(struct tu_cs *cs, uint32_t front, uint32_t back)
+{
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCILMASK, 1);
+   tu_cs_emit(
+      cs, A6XX_RB_STENCILMASK_MASK(front) | A6XX_RB_STENCILMASK_BFMASK(back));
+}
+
+void
+tu6_emit_stencil_write_mask(struct tu_cs *cs, uint32_t front, uint32_t back)
+{
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCILWRMASK, 1);
+   tu_cs_emit(cs, A6XX_RB_STENCILWRMASK_WRMASK(front) |
+                     A6XX_RB_STENCILWRMASK_BFWRMASK(back));
+}
+
+void
+tu6_emit_stencil_reference(struct tu_cs *cs, uint32_t front, uint32_t back)
+{
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCILREF, 1);
+   tu_cs_emit(cs,
+              A6XX_RB_STENCILREF_REF(front) | A6XX_RB_STENCILREF_BFREF(back));
 }
 
 static uint32_t
@@ -2251,11 +1255,18 @@ tu6_rb_mrt_blend_control(const VkPipelineColorBlendAttachmentState *att,
 static uint32_t
 tu6_rb_mrt_control(const VkPipelineColorBlendAttachmentState *att,
                    uint32_t rb_mrt_control_rop,
+                   bool is_int,
                    bool has_alpha)
 {
    uint32_t rb_mrt_control =
       A6XX_RB_MRT_CONTROL_COMPONENT_ENABLE(att->colorWriteMask);
 
+   /* ignore blending and logic op for integer attachments */
+   if (is_int) {
+      rb_mrt_control |= A6XX_RB_MRT_CONTROL_ROP_CODE(ROP_COPY);
+      return rb_mrt_control;
+   }
+
    rb_mrt_control |= rb_mrt_control_rop;
 
    if (att->blendEnable) {
@@ -2268,44 +1279,23 @@ tu6_rb_mrt_control(const VkPipelineColorBlendAttachmentState *att,
    return rb_mrt_control;
 }
 
-uint32_t
-tu6_rb_mrt_control_rop(VkLogicOp op, bool *rop_reads_dst)
-{
-   *rop_reads_dst = tu_logic_op_reads_dst(op);
-   return A6XX_RB_MRT_CONTROL_ROP_ENABLE |
-          A6XX_RB_MRT_CONTROL_ROP_CODE(tu6_rop(op));
-}
-
 static void
-tu6_emit_rb_mrt_controls(struct tu_pipeline *pipeline,
+tu6_emit_rb_mrt_controls(struct tu_cs *cs,
                          const VkPipelineColorBlendStateCreateInfo *blend_info,
                          const VkFormat attachment_formats[MAX_RTS],
-                         bool *rop_reads_dst,
-                         uint32_t *color_bandwidth_per_sample)
+                         uint32_t *blend_enable_mask)
 {
-   const VkPipelineColorWriteCreateInfoEXT *color_info =
-      vk_find_struct_const(blend_info->pNext,
-                           PIPELINE_COLOR_WRITE_CREATE_INFO_EXT);
-
-   /* The static state is ignored if it's dynamic. In that case assume
-    * everything is enabled and then the appropriate registers will be zero'd
-    * dynamically.
-    */
-   if (pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_COLOR_WRITE_ENABLE))
-      color_info = NULL;
-
-   *rop_reads_dst = false;
-   *color_bandwidth_per_sample = 0;
+   *blend_enable_mask = 0;
 
+   bool rop_reads_dst = false;
    uint32_t rb_mrt_control_rop = 0;
    if (blend_info->logicOpEnable) {
-      pipeline->blend.logic_op_enabled = true;
-      rb_mrt_control_rop = tu6_rb_mrt_control_rop(blend_info->logicOp,
-                                                  rop_reads_dst);
+      rop_reads_dst = tu_logic_op_reads_dst(blend_info->logicOp);
+      rb_mrt_control_rop =
+         A6XX_RB_MRT_CONTROL_ROP_ENABLE |
+         A6XX_RB_MRT_CONTROL_ROP_CODE(tu6_rop(blend_info->logicOp));
    }
 
-   uint32_t total_bpp = 0;
-   pipeline->blend.num_rts = blend_info->attachmentCount;
    for (uint32_t i = 0; i < blend_info->attachmentCount; i++) {
       const VkPipelineColorBlendAttachmentState *att =
          &blend_info->pAttachments[i];
@@ -2313,1273 +1303,179 @@ tu6_emit_rb_mrt_controls(struct tu_pipeline *pipeline,
 
       uint32_t rb_mrt_control = 0;
       uint32_t rb_mrt_blend_control = 0;
-      if (format != VK_FORMAT_UNDEFINED &&
-          (!color_info || color_info->pColorWriteEnables[i])) {
+      if (format != VK_FORMAT_UNDEFINED) {
+         const bool is_int = vk_format_is_int(format);
          const bool has_alpha = vk_format_has_alpha(format);
 
          rb_mrt_control =
-            tu6_rb_mrt_control(att, rb_mrt_control_rop, has_alpha);
+            tu6_rb_mrt_control(att, rb_mrt_control_rop, is_int, has_alpha);
          rb_mrt_blend_control = tu6_rb_mrt_blend_control(att, has_alpha);
 
-         /* calculate bpp based on format and write mask */
-         uint32_t write_bpp = 0;
-         if (att->colorWriteMask == 0xf) {
-            write_bpp = vk_format_get_blocksizebits(format);
-         } else {
-            const enum pipe_format pipe_format = vk_format_to_pipe_format(format);
-            for (uint32_t i = 0; i < 4; i++) {
-               if (att->colorWriteMask & (1 << i)) {
-                  write_bpp += util_format_get_component_bits(pipe_format,
-                        UTIL_FORMAT_COLORSPACE_RGB, i);
-               }
-            }
-         }
-         total_bpp += write_bpp;
-
-         pipeline->blend.color_write_enable |= BIT(i);
-         if (att->blendEnable)
-            pipeline->blend.blend_enable |= BIT(i);
-
-         if (att->blendEnable || *rop_reads_dst) {
-            total_bpp += write_bpp;
-         }
+         if (att->blendEnable || rop_reads_dst)
+            *blend_enable_mask |= 1 << i;
       }
 
-      pipeline->blend.rb_mrt_control[i] = rb_mrt_control & pipeline->blend.rb_mrt_control_mask;
-      pipeline->blend.rb_mrt_blend_control[i] = rb_mrt_blend_control;
+      tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_CONTROL(i), 2);
+      tu_cs_emit(cs, rb_mrt_control);
+      tu_cs_emit(cs, rb_mrt_blend_control);
    }
 
-   *color_bandwidth_per_sample = total_bpp / 8;
+   for (uint32_t i = blend_info->attachmentCount; i < MAX_RTS; i++) {
+      tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_CONTROL(i), 2);
+      tu_cs_emit(cs, 0);
+      tu_cs_emit(cs, 0);
+   }
 }
 
 static void
-tu6_emit_blend_control(struct tu_pipeline *pipeline,
+tu6_emit_blend_control(struct tu_cs *cs,
                        uint32_t blend_enable_mask,
-                       bool dual_src_blend,
                        const VkPipelineMultisampleStateCreateInfo *msaa_info)
 {
-   const uint32_t sample_mask =
-      msaa_info->pSampleMask ? (*msaa_info->pSampleMask & 0xffff)
-                             : ((1 << msaa_info->rasterizationSamples) - 1);
+   assert(!msaa_info->sampleShadingEnable);
+   assert(!msaa_info->alphaToOneEnable);
 
+   uint32_t sp_blend_cntl = A6XX_SP_BLEND_CNTL_UNK8;
+   if (blend_enable_mask)
+      sp_blend_cntl |= A6XX_SP_BLEND_CNTL_ENABLED;
+   if (msaa_info->alphaToCoverageEnable)
+      sp_blend_cntl |= A6XX_SP_BLEND_CNTL_ALPHA_TO_COVERAGE;
 
-   pipeline->blend.sp_blend_cntl =
-       A6XX_SP_BLEND_CNTL(.enable_blend = blend_enable_mask,
-                          .dual_color_in_enable = dual_src_blend,
-                          .alpha_to_coverage = msaa_info->alphaToCoverageEnable,
-                          .unk8 = true).value & pipeline->blend.sp_blend_cntl_mask;
+   const uint32_t sample_mask =
+      msaa_info->pSampleMask ? *msaa_info->pSampleMask
+                             : ((1 << msaa_info->rasterizationSamples) - 1);
 
    /* set A6XX_RB_BLEND_CNTL_INDEPENDENT_BLEND only when enabled? */
-   pipeline->blend.rb_blend_cntl =
-       A6XX_RB_BLEND_CNTL(.enable_blend = blend_enable_mask,
-                          .independent_blend = true,
-                          .sample_mask = sample_mask,
-                          .dual_color_in_enable = dual_src_blend,
-                          .alpha_to_coverage = msaa_info->alphaToCoverageEnable,
-                          .alpha_to_one = msaa_info->alphaToOneEnable).value &
-      pipeline->blend.rb_blend_cntl_mask;
-}
+   uint32_t rb_blend_cntl =
+      A6XX_RB_BLEND_CNTL_ENABLE_BLEND(blend_enable_mask) |
+      A6XX_RB_BLEND_CNTL_INDEPENDENT_BLEND |
+      A6XX_RB_BLEND_CNTL_SAMPLE_MASK(sample_mask);
+   if (msaa_info->alphaToCoverageEnable)
+      rb_blend_cntl |= A6XX_RB_BLEND_CNTL_ALPHA_TO_COVERAGE;
 
-static void
-tu6_emit_blend(struct tu_cs *cs,
-               struct tu_pipeline *pipeline)
-{
-   tu_cs_emit_regs(cs, A6XX_SP_FS_OUTPUT_CNTL1(.mrt = pipeline->blend.num_rts));
-   tu_cs_emit_regs(cs, A6XX_RB_FS_OUTPUT_CNTL1(.mrt = pipeline->blend.num_rts));
-   tu_cs_emit_regs(cs, A6XX_SP_BLEND_CNTL(.dword = pipeline->blend.sp_blend_cntl));
-   tu_cs_emit_regs(cs, A6XX_RB_BLEND_CNTL(.dword = pipeline->blend.rb_blend_cntl));
-
-   for (unsigned i = 0; i < pipeline->blend.num_rts; i++) {
-      tu_cs_emit_regs(cs,
-                      A6XX_RB_MRT_CONTROL(i, .dword = pipeline->blend.rb_mrt_control[i]),
-                      A6XX_RB_MRT_BLEND_CONTROL(i, .dword = pipeline->blend.rb_mrt_blend_control[i]));
-   }
-}
+   tu_cs_emit_pkt4(cs, REG_A6XX_SP_BLEND_CNTL, 1);
+   tu_cs_emit(cs, sp_blend_cntl);
 
-static VkResult
-tu_setup_pvtmem(struct tu_device *dev,
-                struct tu_pipeline *pipeline,
-                struct tu_pvtmem_config *config,
-                uint32_t pvtmem_bytes,
-                bool per_wave)
-{
-   if (!pvtmem_bytes) {
-      memset(config, 0, sizeof(*config));
-      return VK_SUCCESS;
-   }
-
-   /* There is a substantial memory footprint from private memory BOs being
-    * allocated on a per-pipeline basis and it isn't required as the same
-    * BO can be utilized by multiple pipelines as long as they have the
-    * private memory layout (sizes and per-wave/per-fiber) to avoid being
-    * overwritten by other active pipelines using the same BO with differing
-    * private memory layouts resulting memory corruption.
-    *
-    * To avoid this, we create private memory BOs on a per-device level with
-    * an associated private memory layout then dynamically grow them when
-    * needed and reuse them across pipelines. Growth is done in terms of
-    * powers of two so that we can avoid frequent reallocation of the
-    * private memory BOs.
-    */
-
-   struct tu_pvtmem_bo *pvtmem_bo =
-      per_wave ? &dev->wave_pvtmem_bo : &dev->fiber_pvtmem_bo;
-   mtx_lock(&pvtmem_bo->mtx);
-
-   if (pvtmem_bo->per_fiber_size < pvtmem_bytes) {
-      if (pvtmem_bo->bo)
-         tu_bo_finish(dev, pvtmem_bo->bo);
-
-      pvtmem_bo->per_fiber_size =
-         util_next_power_of_two(ALIGN(pvtmem_bytes, 512));
-      pvtmem_bo->per_sp_size =
-         ALIGN(pvtmem_bo->per_fiber_size *
-                  dev->physical_device->info->a6xx.fibers_per_sp,
-               1 << 12);
-      uint32_t total_size =
-         dev->physical_device->info->num_sp_cores * pvtmem_bo->per_sp_size;
-
-      VkResult result = tu_bo_init_new(dev, &pvtmem_bo->bo, total_size,
-                                       TU_BO_ALLOC_NO_FLAGS, "pvtmem");
-      if (result != VK_SUCCESS) {
-         mtx_unlock(&pvtmem_bo->mtx);
-         return result;
-      }
-   }
-
-   config->per_wave = per_wave;
-   config->per_fiber_size = pvtmem_bo->per_fiber_size;
-   config->per_sp_size = pvtmem_bo->per_sp_size;
-
-   pipeline->pvtmem_bo = tu_bo_get_ref(pvtmem_bo->bo);
-   config->iova = pipeline->pvtmem_bo->iova;
-
-   mtx_unlock(&pvtmem_bo->mtx);
-
-   return VK_SUCCESS;
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLEND_CNTL, 1);
+   tu_cs_emit(cs, rb_blend_cntl);
 }
 
-static bool
-contains_all_shader_state(VkGraphicsPipelineLibraryFlagsEXT state)
-{
-   return (state &
-      (VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-       VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT)) ==
-      (VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-       VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT);
-}
-
-/* Return true if this pipeline contains all of the GPL stages listed but none
- * of the libraries it uses do, so this is "the first time" that all of them
- * are defined together. This is useful for state that needs to be combined
- * from multiple GPL stages.
- */
-
-static bool
-set_combined_state(struct tu_pipeline_builder *builder,
-                   struct tu_pipeline *pipeline,
-                   VkGraphicsPipelineLibraryFlagsEXT state)
+void
+tu6_emit_blend_constants(struct tu_cs *cs, const float constants[4])
 {
-   if ((pipeline->state & state) != state)
-      return false;
-
-   for (unsigned i = 0; i < builder->num_libraries; i++) {
-      if ((builder->libraries[i]->state & state) == state)
-         return false;
-   }
-
-   return true;
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLEND_RED_F32, 4);
+   tu_cs_emit_array(cs, (const uint32_t *) constants, 4);
 }
 
 static VkResult
-tu_pipeline_allocate_cs(struct tu_device *dev,
-                        struct tu_pipeline *pipeline,
-                        struct tu_pipeline_layout *layout,
-                        struct tu_pipeline_builder *builder,
-                        struct ir3_shader_variant *compute)
+tu_pipeline_builder_create_pipeline(struct tu_pipeline_builder *builder,
+                                    struct tu_pipeline **out_pipeline)
 {
-   uint32_t size = 1024;
-
-   /* graphics case: */
-   if (builder) {
-      if (builder->state &
-          VK_GRAPHICS_PIPELINE_LIBRARY_VERTEX_INPUT_INTERFACE_BIT_EXT) {
-         size += TU6_EMIT_VERTEX_INPUT_MAX_DWORDS;
-      }
-
-      if (set_combined_state(builder, pipeline,
-                             VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-                             VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT)) {
-         size += 2 * TU6_EMIT_VFD_DEST_MAX_DWORDS;
-         size += tu6_load_state_size(pipeline, layout);
+   struct tu_device *dev = builder->device;
 
-         for (uint32_t i = 0; i < ARRAY_SIZE(builder->variants); i++) {
-            if (builder->variants[i]) {
-               size += builder->variants[i]->info.size / 4;
-            }
-         }
-
-         size += builder->binning_variant->info.size / 4;
-
-         builder->additional_cs_reserve_size = 0;
-         for (unsigned i = 0; i < ARRAY_SIZE(builder->variants); i++) {
-            struct ir3_shader_variant *variant = builder->variants[i];
-            if (variant) {
-               builder->additional_cs_reserve_size +=
-                  tu_xs_get_additional_cs_size_dwords(variant);
-
-               if (variant->binning) {
-                  builder->additional_cs_reserve_size +=
-                     tu_xs_get_additional_cs_size_dwords(variant->binning);
-               }
-            }
-         }
-
-         /* The additional size is used twice, once per tu6_emit_program() call. */
-         size += builder->additional_cs_reserve_size * 2;
-      }
-   } else {
-      size += tu6_load_state_size(pipeline, layout);
+   struct tu_pipeline *pipeline =
+      vk_zalloc2(&dev->alloc, builder->alloc, sizeof(*pipeline), 8,
+                 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (!pipeline)
+      return VK_ERROR_OUT_OF_HOST_MEMORY;
 
-      size += compute->info.size / 4;
+   tu_cs_init(&pipeline->cs, TU_CS_MODE_SUB_STREAM, 2048);
 
-      size += tu_xs_get_additional_cs_size_dwords(compute);
-   }
-
-   /* Allocate the space for the pipeline out of the device's RO suballocator.
-    *
-    * Sub-allocating BOs saves memory and also kernel overhead in refcounting of
-    * BOs at exec time.
-    *
-    * The pipeline cache would seem like a natural place to stick the
-    * suballocator, except that it is not guaranteed to outlive the pipelines
-    * created from it, so you can't store any long-lived state there, and you
-    * can't use its EXTERNALLY_SYNCHRONIZED flag to avoid atomics because
-    * pipeline destroy isn't synchronized by the cache.
-    */
-   pthread_mutex_lock(&dev->pipeline_mutex);
-   VkResult result = tu_suballoc_bo_alloc(&pipeline->bo, &dev->pipeline_suballoc,
-                                          size * 4, 128);
-   pthread_mutex_unlock(&dev->pipeline_mutex);
-   if (result != VK_SUCCESS)
+   /* reserve the space now such that tu_cs_begin_sub_stream never fails */
+   VkResult result = tu_cs_reserve_space(dev, &pipeline->cs, 2048);
+   if (result != VK_SUCCESS) {
+      vk_free2(&dev->alloc, builder->alloc, pipeline);
       return result;
-
-   tu_cs_init_suballoc(&pipeline->cs, dev, &pipeline->bo);
-
-   return VK_SUCCESS;
-}
-
-static void
-tu_pipeline_shader_key_init(struct ir3_shader_key *key,
-                            const struct tu_pipeline *pipeline,
-                            struct tu_pipeline_builder *builder,
-                            nir_shader **nir)
-{
-   /* We set this after we compile to NIR because we need the prim mode */
-   key->tessellation = IR3_TESS_NONE;
-
-   for (unsigned i = 0; i < builder->num_libraries; i++) {
-      if (!(builder->libraries[i]->state &
-            (VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-             VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT)))
-         continue;
-
-      const struct ir3_shader_key *library_key =
-         &builder->libraries[i]->ir3_key;
-
-      if (library_key->tessellation != IR3_TESS_NONE)
-         key->tessellation = library_key->tessellation;
-      key->has_gs |= library_key->has_gs;
-      key->sample_shading |= library_key->sample_shading;
-   }
-
-   for (uint32_t i = 0; i < builder->create_info->stageCount; i++) {
-      if (builder->create_info->pStages[i].stage == VK_SHADER_STAGE_GEOMETRY_BIT) {
-         key->has_gs = true;
-         break;
-      }
-   }
-
-   if (!(builder->state & VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT))
-      return;
-
-   if (builder->rasterizer_discard)
-      return;
-
-   const VkPipelineMultisampleStateCreateInfo *msaa_info =
-      builder->create_info->pMultisampleState;
-
-   /* The 1.3.215 spec says:
-    *
-    *    Sample shading can be used to specify a minimum number of unique
-    *    samples to process for each fragment. If sample shading is enabled,
-    *    an implementation must provide a minimum of
-    *
-    *       max(ceil(minSampleShadingFactor * totalSamples), 1)
-    *
-    *    unique associated data for each fragment, where
-    *    minSampleShadingFactor is the minimum fraction of sample shading.
-    *
-    * The definition is pretty much the same as OpenGL's GL_SAMPLE_SHADING.
-    * They both require unique associated data.
-    *
-    * There are discussions to change the definition, such that
-    * sampleShadingEnable does not imply unique associated data.  Before the
-    * discussions are settled and before apps (i.e., ANGLE) are fixed to
-    * follow the new and incompatible definition, we should stick to the
-    * current definition.
-    *
-    * Note that ir3_shader_key::sample_shading is not actually used by ir3,
-    * just checked in tu6_emit_fs_inputs.  We will also copy the value to
-    * tu_shader_key::force_sample_interp in a bit.
-    */
-   if (msaa_info && msaa_info->sampleShadingEnable &&
-       (msaa_info->minSampleShading * msaa_info->rasterizationSamples) > 1.0f)
-      key->sample_shading = true;
-}
-
-static uint32_t
-tu6_get_tessmode(struct tu_shader* shader)
-{
-   enum tess_primitive_mode primitive_mode = shader->ir3_shader->nir->info.tess._primitive_mode;
-   switch (primitive_mode) {
-   case TESS_PRIMITIVE_ISOLINES:
-      return IR3_TESS_ISOLINES;
-   case TESS_PRIMITIVE_TRIANGLES:
-      return IR3_TESS_TRIANGLES;
-   case TESS_PRIMITIVE_QUADS:
-      return IR3_TESS_QUADS;
-   case TESS_PRIMITIVE_UNSPECIFIED:
-      return IR3_TESS_NONE;
-   default:
-      unreachable("bad tessmode");
-   }
-}
-
-static uint64_t
-tu_upload_variant(struct tu_pipeline *pipeline,
-                  const struct ir3_shader_variant *variant)
-{
-   struct tu_cs_memory memory;
-
-   if (!variant)
-      return 0;
-
-   /* this expects to get enough alignment because shaders are allocated first
-    * and total size is always aligned correctly
-    * note: an assert in tu6_emit_xs_config validates the alignment
-    */
-   tu_cs_alloc(&pipeline->cs, variant->info.size / 4, 1, &memory);
-
-   memcpy(memory.map, variant->bin, variant->info.size);
-   return memory.iova;
-}
-
-static void
-tu_append_executable(struct tu_pipeline *pipeline, struct ir3_shader_variant *variant,
-                     char *nir_from_spirv)
-{
-   struct tu_pipeline_executable exe = {
-      .stage = variant->type,
-      .nir_from_spirv = nir_from_spirv,
-      .nir_final = ralloc_strdup(pipeline->executables_mem_ctx, variant->disasm_info.nir),
-      .disasm = ralloc_strdup(pipeline->executables_mem_ctx, variant->disasm_info.disasm),
-      .stats = variant->info,
-      .is_binning = variant->binning_pass,
-   };
-
-   util_dynarray_append(&pipeline->executables, struct tu_pipeline_executable, exe);
-}
-
-static bool
-can_remove_out_var(nir_variable *var, void *data)
-{
-   return !var->data.explicit_xfb_buffer && !var->data.explicit_xfb_stride;
-}
-
-static void
-tu_link_shaders(struct tu_pipeline_builder *builder,
-                nir_shader **shaders, unsigned shaders_count)
-{
-   nir_shader *consumer = NULL;
-   for (gl_shader_stage stage = shaders_count - 1;
-        stage >= MESA_SHADER_VERTEX; stage--) {
-      if (!shaders[stage])
-         continue;
-
-      nir_shader *producer = shaders[stage];
-      if (!consumer) {
-         consumer = producer;
-         continue;
-      }
-
-      if (nir_link_opt_varyings(producer, consumer)) {
-         NIR_PASS_V(consumer, nir_opt_constant_folding);
-         NIR_PASS_V(consumer, nir_opt_algebraic);
-         NIR_PASS_V(consumer, nir_opt_dce);
-      }
-
-      const nir_remove_dead_variables_options out_var_opts = {
-         .can_remove_var = can_remove_out_var,
-      };
-      NIR_PASS_V(producer, nir_remove_dead_variables, nir_var_shader_out, &out_var_opts);
-
-      NIR_PASS_V(consumer, nir_remove_dead_variables, nir_var_shader_in, NULL);
-
-      bool progress = nir_remove_unused_varyings(producer, consumer);
-
-      nir_compact_varyings(producer, consumer, true);
-      if (progress) {
-         if (nir_lower_global_vars_to_local(producer)) {
-            /* Remove dead writes, which can remove input loads */
-            NIR_PASS_V(producer, nir_remove_dead_variables, nir_var_shader_temp, NULL);
-            NIR_PASS_V(producer, nir_opt_dce);
-         }
-         nir_lower_global_vars_to_local(consumer);
-      }
-
-      consumer = producer;
-   }
-}
-
-static void
-tu_shader_key_init(struct tu_shader_key *key,
-                   const VkPipelineShaderStageCreateInfo *stage_info,
-                   struct tu_device *dev)
-{
-   enum ir3_wavesize_option api_wavesize, real_wavesize;
-
-   if (stage_info) {
-      if (stage_info->flags &
-          VK_PIPELINE_SHADER_STAGE_CREATE_ALLOW_VARYING_SUBGROUP_SIZE_BIT) {
-         api_wavesize = real_wavesize = IR3_SINGLE_OR_DOUBLE;
-      } else {
-         const VkPipelineShaderStageRequiredSubgroupSizeCreateInfo *size_info =
-            vk_find_struct_const(stage_info->pNext,
-                                 PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO);
-
-         if (size_info) {
-            if (size_info->requiredSubgroupSize == dev->compiler->threadsize_base) {
-               api_wavesize = IR3_SINGLE_ONLY;
-            } else {
-               assert(size_info->requiredSubgroupSize == dev->compiler->threadsize_base * 2);
-               api_wavesize = IR3_DOUBLE_ONLY;
-            }
-         } else {
-            /* Match the exposed subgroupSize. */
-            api_wavesize = IR3_DOUBLE_ONLY;
-         }
-
-         if (stage_info->flags &
-             VK_PIPELINE_SHADER_STAGE_CREATE_REQUIRE_FULL_SUBGROUPS_BIT)
-            real_wavesize = api_wavesize;
-         else if (api_wavesize == IR3_SINGLE_ONLY)
-            real_wavesize = IR3_SINGLE_ONLY;
-         else
-            real_wavesize = IR3_SINGLE_OR_DOUBLE;
-      }
-   } else {
-      api_wavesize = real_wavesize = IR3_SINGLE_OR_DOUBLE;
-   }
-
-   key->api_wavesize = api_wavesize;
-   key->real_wavesize = real_wavesize;
-}
-
-static void
-tu_hash_stage(struct mesa_sha1 *ctx,
-              const VkPipelineShaderStageCreateInfo *stage,
-              const nir_shader *nir,
-              const struct tu_shader_key *key)
-{
-
-   if (nir) {
-      struct blob blob;
-      blob_init(&blob);
-      nir_serialize(&blob, nir, true);
-      _mesa_sha1_update(ctx, blob.data, blob.size);
-      blob_finish(&blob);
-   } else {
-      unsigned char stage_hash[SHA1_DIGEST_LENGTH];
-      vk_pipeline_hash_shader_stage(stage, NULL, stage_hash);
-      _mesa_sha1_update(ctx, stage_hash, sizeof(stage_hash));
-   }
-   _mesa_sha1_update(ctx, key, sizeof(*key));
-}
-
-/* Hash flags which can affect ir3 shader compilation which aren't known until
- * logical device creation.
- */
-static void
-tu_hash_compiler(struct mesa_sha1 *ctx, const struct ir3_compiler *compiler)
-{
-   _mesa_sha1_update(ctx, &compiler->robust_buffer_access2,
-                     sizeof(compiler->robust_buffer_access2));
-   _mesa_sha1_update(ctx, &ir3_shader_debug, sizeof(ir3_shader_debug));
-}
-
-static void
-tu_hash_shaders(unsigned char *hash,
-                const VkPipelineShaderStageCreateInfo **stages,
-                nir_shader *const *nir,
-                const struct tu_pipeline_layout *layout,
-                const struct tu_shader_key *keys,
-                const struct ir3_shader_key *ir3_key,
-                VkGraphicsPipelineLibraryFlagsEXT state,
-                const struct ir3_compiler *compiler)
-{
-   struct mesa_sha1 ctx;
-
-   _mesa_sha1_init(&ctx);
-
-   if (layout)
-      _mesa_sha1_update(&ctx, layout->sha1, sizeof(layout->sha1));
-
-   _mesa_sha1_update(&ctx, ir3_key, sizeof(ir3_key));
-
-   for (int i = 0; i < MESA_SHADER_STAGES; ++i) {
-      if (stages[i] || nir[i]) {
-         tu_hash_stage(&ctx, stages[i], nir[i], &keys[i]);
-      }
-   }
-   _mesa_sha1_update(&ctx, &state, sizeof(state));
-   tu_hash_compiler(&ctx, compiler);
-   _mesa_sha1_final(&ctx, hash);
-}
-
-static void
-tu_hash_compute(unsigned char *hash,
-                const VkPipelineShaderStageCreateInfo *stage,
-                const struct tu_pipeline_layout *layout,
-                const struct tu_shader_key *key,
-                const struct ir3_compiler *compiler)
-{
-   struct mesa_sha1 ctx;
-
-   _mesa_sha1_init(&ctx);
-
-   if (layout)
-      _mesa_sha1_update(&ctx, layout->sha1, sizeof(layout->sha1));
-
-   tu_hash_stage(&ctx, stage, NULL, key);
-
-   tu_hash_compiler(&ctx, compiler);
-   _mesa_sha1_final(&ctx, hash);
-}
-
-static bool
-tu_shaders_serialize(struct vk_pipeline_cache_object *object,
-                     struct blob *blob);
-
-static struct vk_pipeline_cache_object *
-tu_shaders_deserialize(struct vk_device *device,
-                       const void *key_data, size_t key_size,
-                       struct blob_reader *blob);
-
-static void
-tu_shaders_destroy(struct vk_pipeline_cache_object *object)
-{
-   struct tu_compiled_shaders *shaders =
-      container_of(object, struct tu_compiled_shaders, base);
-
-   for (unsigned i = 0; i < ARRAY_SIZE(shaders->variants); i++)
-      ralloc_free(shaders->variants[i]);
-
-   for (unsigned i = 0; i < ARRAY_SIZE(shaders->safe_const_variants); i++)
-      ralloc_free(shaders->safe_const_variants[i]);
-
-   vk_pipeline_cache_object_finish(&shaders->base);
-   vk_free(&object->device->alloc, shaders);
-}
-
-const struct vk_pipeline_cache_object_ops tu_shaders_ops = {
-   .serialize = tu_shaders_serialize,
-   .deserialize = tu_shaders_deserialize,
-   .destroy = tu_shaders_destroy,
-};
-
-static struct tu_compiled_shaders *
-tu_shaders_init(struct tu_device *dev, const void *key_data, size_t key_size)
-{
-   VK_MULTIALLOC(ma);
-   VK_MULTIALLOC_DECL(&ma, struct tu_compiled_shaders, shaders, 1);
-   VK_MULTIALLOC_DECL_SIZE(&ma, void, obj_key_data, key_size);
-
-   if (!vk_multialloc_zalloc(&ma, &dev->vk.alloc,
-                             VK_SYSTEM_ALLOCATION_SCOPE_DEVICE))
-      return NULL;
-
-   memcpy(obj_key_data, key_data, key_size);
-   vk_pipeline_cache_object_init(&dev->vk, &shaders->base,
-                                 &tu_shaders_ops, obj_key_data, key_size);
-
-   return shaders;
-}
-
-static bool
-tu_shaders_serialize(struct vk_pipeline_cache_object *object,
-                     struct blob *blob)
-{
-   struct tu_compiled_shaders *shaders =
-      container_of(object, struct tu_compiled_shaders, base);
-
-   blob_write_bytes(blob, shaders->const_state, sizeof(shaders->const_state));
-   blob_write_uint8(blob, shaders->active_desc_sets);
-
-   for (unsigned i = 0; i < ARRAY_SIZE(shaders->variants); i++) {
-      if (shaders->variants[i]) {
-         blob_write_uint8(blob, 1);
-         ir3_store_variant(blob, shaders->variants[i]);
-      } else {
-         blob_write_uint8(blob, 0);
-      }
-
-      if (shaders->safe_const_variants[i]) {
-         blob_write_uint8(blob, 1);
-         ir3_store_variant(blob, shaders->safe_const_variants[i]);
-      } else {
-         blob_write_uint8(blob, 0);
-      }
    }
 
-   return true;
-}
-
-static struct vk_pipeline_cache_object *
-tu_shaders_deserialize(struct vk_device *_device,
-                       const void *key_data, size_t key_size,
-                       struct blob_reader *blob)
-{
-   struct tu_device *dev = container_of(_device, struct tu_device, vk);
-   struct tu_compiled_shaders *shaders =
-      tu_shaders_init(dev, key_data, key_size);
-
-   if (!shaders)
-      return NULL;
-
-   blob_copy_bytes(blob, shaders->const_state, sizeof(shaders->const_state));
-   shaders->active_desc_sets = blob_read_uint8(blob);
-
-   for (unsigned i = 0; i < ARRAY_SIZE(shaders->variants); i++) {
-      if (blob_read_uint8(blob)) {
-         shaders->variants[i] = ir3_retrieve_variant(blob, dev->compiler, NULL);
-      }
-
-      if (blob_read_uint8(blob)) {
-         shaders->safe_const_variants[i] = ir3_retrieve_variant(blob, dev->compiler, NULL);
-      }
-   }
-
-   return &shaders->base;
-}
-
-static struct tu_compiled_shaders *
-tu_pipeline_cache_lookup(struct vk_pipeline_cache *cache,
-                         const void *key_data, size_t key_size,
-                         bool *application_cache_hit)
-{
-   struct vk_pipeline_cache_object *object =
-      vk_pipeline_cache_lookup_object(cache, key_data, key_size,
-                                      &tu_shaders_ops, application_cache_hit);
-   if (object)
-      return container_of(object, struct tu_compiled_shaders, base);
-   else
-      return NULL;
-}
-
-static struct tu_compiled_shaders *
-tu_pipeline_cache_insert(struct vk_pipeline_cache *cache,
-                         struct tu_compiled_shaders *shaders)
-{
-   struct vk_pipeline_cache_object *object =
-      vk_pipeline_cache_add_object(cache, &shaders->base);
-   return container_of(object, struct tu_compiled_shaders, base);
-}
-
-static bool
-tu_nir_shaders_serialize(struct vk_pipeline_cache_object *object,
-                         struct blob *blob);
-
-static struct vk_pipeline_cache_object *
-tu_nir_shaders_deserialize(struct vk_device *device,
-                           const void *key_data, size_t key_size,
-                           struct blob_reader *blob);
-
-static void
-tu_nir_shaders_destroy(struct vk_pipeline_cache_object *object)
-{
-   struct tu_nir_shaders *shaders =
-      container_of(object, struct tu_nir_shaders, base);
-
-   for (unsigned i = 0; i < ARRAY_SIZE(shaders->nir); i++)
-      ralloc_free(shaders->nir[i]);
-
-   vk_pipeline_cache_object_finish(&shaders->base);
-   vk_free(&object->device->alloc, shaders);
-}
-
-const struct vk_pipeline_cache_object_ops tu_nir_shaders_ops = {
-   .serialize = tu_nir_shaders_serialize,
-   .deserialize = tu_nir_shaders_deserialize,
-   .destroy = tu_nir_shaders_destroy,
-};
-
-static struct tu_nir_shaders *
-tu_nir_shaders_init(struct tu_device *dev, const void *key_data, size_t key_size)
-{
-   VK_MULTIALLOC(ma);
-   VK_MULTIALLOC_DECL(&ma, struct tu_nir_shaders, shaders, 1);
-   VK_MULTIALLOC_DECL_SIZE(&ma, void, obj_key_data, key_size);
-
-   if (!vk_multialloc_zalloc(&ma, &dev->vk.alloc,
-                             VK_SYSTEM_ALLOCATION_SCOPE_DEVICE))
-      return NULL;
-
-   memcpy(obj_key_data, key_data, key_size);
-   vk_pipeline_cache_object_init(&dev->vk, &shaders->base,
-                                 &tu_nir_shaders_ops, obj_key_data, key_size);
-
-   return shaders;
-}
-
-static bool
-tu_nir_shaders_serialize(struct vk_pipeline_cache_object *object,
-                         struct blob *blob)
-{
-   struct tu_nir_shaders *shaders =
-      container_of(object, struct tu_nir_shaders, base);
-
-   for (unsigned i = 0; i < ARRAY_SIZE(shaders->nir); i++) {
-      if (shaders->nir[i]) {
-         blob_write_uint8(blob, 1);
-         nir_serialize(blob, shaders->nir[i], true);
-      } else {
-         blob_write_uint8(blob, 0);
-      }
-   }
-
-   return true;
-}
-
-static struct vk_pipeline_cache_object *
-tu_nir_shaders_deserialize(struct vk_device *_device,
-                       const void *key_data, size_t key_size,
-                       struct blob_reader *blob)
-{
-   struct tu_device *dev = container_of(_device, struct tu_device, vk);
-   struct tu_nir_shaders *shaders =
-      tu_nir_shaders_init(dev, key_data, key_size);
-
-   if (!shaders)
-      return NULL;
-
-   for (unsigned i = 0; i < ARRAY_SIZE(shaders->nir); i++) {
-      if (blob_read_uint8(blob)) {
-         shaders->nir[i] =
-            nir_deserialize(NULL, ir3_get_compiler_options(dev->compiler), blob);
-      }
-   }
-
-   return &shaders->base;
-}
+   *out_pipeline = pipeline;
 
-static struct tu_nir_shaders *
-tu_nir_cache_lookup(struct vk_pipeline_cache *cache,
-                    const void *key_data, size_t key_size,
-                    bool *application_cache_hit)
-{
-   struct vk_pipeline_cache_object *object =
-      vk_pipeline_cache_lookup_object(cache, key_data, key_size,
-                                      &tu_nir_shaders_ops, application_cache_hit);
-   if (object)
-      return container_of(object, struct tu_nir_shaders, base);
-   else
-      return NULL;
-}
-
-static struct tu_nir_shaders *
-tu_nir_cache_insert(struct vk_pipeline_cache *cache,
-                    struct tu_nir_shaders *shaders)
-{
-   struct vk_pipeline_cache_object *object =
-      vk_pipeline_cache_add_object(cache, &shaders->base);
-   return container_of(object, struct tu_nir_shaders, base);
+   return VK_SUCCESS;
 }
 
-
 static VkResult
-tu_pipeline_builder_compile_shaders(struct tu_pipeline_builder *builder,
-                                    struct tu_pipeline *pipeline)
+tu_pipeline_builder_compile_shaders(struct tu_pipeline_builder *builder)
 {
-   VkResult result = VK_SUCCESS;
-   const struct ir3_compiler *compiler = builder->device->compiler;
    const VkPipelineShaderStageCreateInfo *stage_infos[MESA_SHADER_STAGES] = {
       NULL
    };
-   VkPipelineCreationFeedback pipeline_feedback = {
-      .flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT,
-   };
-   VkPipelineCreationFeedback stage_feedbacks[MESA_SHADER_STAGES] = { 0 };
-
-   int64_t pipeline_start = os_time_get_nano();
-
-   const VkPipelineCreationFeedbackCreateInfo *creation_feedback =
-      vk_find_struct_const(builder->create_info->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO);
-
-   bool must_compile =
-      builder->state & VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT;
    for (uint32_t i = 0; i < builder->create_info->stageCount; i++) {
       gl_shader_stage stage =
-         vk_to_mesa_shader_stage(builder->create_info->pStages[i].stage);
+         tu_shader_stage(builder->create_info->pStages[i].stage);
       stage_infos[stage] = &builder->create_info->pStages[i];
-      must_compile = true;
-   }
-
-   if (tu6_shared_constants_enable(&builder->layout, builder->device->compiler)) {
-      pipeline->shared_consts = (struct tu_push_constant_range) {
-         .lo = 0,
-         .dwords = builder->layout.push_constant_size / 4,
-      };
-   }
-
-   nir_shader *nir[ARRAY_SIZE(stage_infos)] = { NULL };
-
-   struct tu_shader_key keys[ARRAY_SIZE(stage_infos)] = { };
-   for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-        stage < ARRAY_SIZE(keys); stage++) {
-      tu_shader_key_init(&keys[stage], stage_infos[stage], builder->device);
-   }
-
-   if (builder->create_info->flags &
-       VK_PIPELINE_CREATE_LINK_TIME_OPTIMIZATION_BIT_EXT) {
-      for (unsigned i = 0; i < builder->num_libraries; i++) {
-         struct tu_pipeline *library = builder->libraries[i];
-
-         for (unsigned j = 0; j < ARRAY_SIZE(library->shaders); j++) {
-            if (library->shaders[j].nir) {
-               assert(!nir[j]);
-               nir[j] = nir_shader_clone(builder->mem_ctx,
-                     library->shaders[j].nir);
-               keys[j] = library->shaders[j].key;
-               must_compile = true;
-            }
-         }
-      }
    }
 
-   struct ir3_shader_key ir3_key = {};
-   tu_pipeline_shader_key_init(&ir3_key, pipeline, builder, nir);
+   struct tu_shader_compile_options options;
+   tu_shader_compile_options_init(&options, builder->create_info);
 
-   struct tu_compiled_shaders *compiled_shaders = NULL;
-   struct tu_nir_shaders *nir_shaders = NULL;
-   if (!must_compile)
-      goto done;
-
-   if (builder->state &
-       VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT) {
-      keys[MESA_SHADER_VERTEX].multiview_mask = builder->multiview_mask;
-   }
-
-   if (builder->state & VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT) {
-      keys[MESA_SHADER_FRAGMENT].multiview_mask = builder->multiview_mask;
-      keys[MESA_SHADER_FRAGMENT].force_sample_interp = ir3_key.sample_shading;
-   }
-
-   unsigned char pipeline_sha1[20];
-   tu_hash_shaders(pipeline_sha1, stage_infos, nir, &builder->layout, keys,
-                   &ir3_key, builder->state, compiler);
-
-   unsigned char nir_sha1[21];
-   memcpy(nir_sha1, pipeline_sha1, sizeof(pipeline_sha1));
-   nir_sha1[20] = 'N';
-
-   const bool executable_info = builder->create_info->flags &
-      VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR;
-
-   char *nir_initial_disasm[ARRAY_SIZE(stage_infos)] = { NULL };
-
-   if (!executable_info) {
-      bool cache_hit = false;
-      bool application_cache_hit = false;
-
-      compiled_shaders =
-         tu_pipeline_cache_lookup(builder->cache, &pipeline_sha1,
-                                  sizeof(pipeline_sha1),
-                                  &application_cache_hit);
-
-      cache_hit = !!compiled_shaders;
-
-      /* If the user asks us to keep the NIR around, we need to have it for a
-       * successful cache hit. If we only have a "partial" cache hit, then we
-       * still need to recompile in order to get the NIR.
-       */
-      if (compiled_shaders &&
-          (builder->create_info->flags &
-           VK_PIPELINE_CREATE_RETAIN_LINK_TIME_OPTIMIZATION_INFO_BIT_EXT)) {
-         bool nir_application_cache_hit = false;
-         nir_shaders =
-            tu_nir_cache_lookup(builder->cache, &nir_sha1,
-                                sizeof(nir_sha1),
-                                &nir_application_cache_hit);
-
-         application_cache_hit &= nir_application_cache_hit;
-         cache_hit &= !!nir_shaders;
-      }
-
-      if (application_cache_hit && builder->cache != builder->device->mem_cache) {
-         pipeline_feedback.flags |=
-            VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT;
-      }
-
-      if (cache_hit)
-         goto done;
-   }
-
-   if (builder->create_info->flags &
-       VK_PIPELINE_CREATE_FAIL_ON_PIPELINE_COMPILE_REQUIRED_BIT) {
-      return VK_PIPELINE_COMPILE_REQUIRED;
-   }
-
-   struct tu_shader *shaders[ARRAY_SIZE(nir)] = { NULL };
-
-   for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-        stage < ARRAY_SIZE(nir); stage++) {
+   /* compile shaders in reverse order */
+   struct tu_shader *next_stage_shader = NULL;
+   for (gl_shader_stage stage = MESA_SHADER_STAGES - 1;
+        stage > MESA_SHADER_NONE; stage--) {
       const VkPipelineShaderStageCreateInfo *stage_info = stage_infos[stage];
       if (!stage_info)
          continue;
 
-      int64_t stage_start = os_time_get_nano();
-
-      nir[stage] = tu_spirv_to_nir(builder->device, builder->mem_ctx, stage_info, stage);
-      if (!nir[stage]) {
-         result = VK_ERROR_OUT_OF_HOST_MEMORY;
-         goto fail;
-      }
-
-      stage_feedbacks[stage].flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT;
-      stage_feedbacks[stage].duration += os_time_get_nano() - stage_start;
-   }
-
-   if (!nir[MESA_SHADER_FRAGMENT] &&
-       (builder->state & VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT)) {
-         const nir_shader_compiler_options *nir_options =
-            ir3_get_compiler_options(builder->device->compiler);
-         nir_builder fs_b = nir_builder_init_simple_shader(MESA_SHADER_FRAGMENT,
-                                                           nir_options,
-                                                           "noop_fs");
-         nir[MESA_SHADER_FRAGMENT] = fs_b.shader;
-   }
-
-   if (executable_info) {
-      for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-            stage < ARRAY_SIZE(nir); stage++) {
-         if (!nir[stage])
-            continue;
-
-         nir_initial_disasm[stage] =
-            nir_shader_as_str(nir[stage], pipeline->executables_mem_ctx);
-      }
-   }
-
-   tu_link_shaders(builder, nir, ARRAY_SIZE(nir));
-
-   if (builder->create_info->flags &
-       VK_PIPELINE_CREATE_RETAIN_LINK_TIME_OPTIMIZATION_INFO_BIT_EXT) {
-      nir_shaders =
-         tu_nir_shaders_init(builder->device, &nir_sha1, sizeof(nir_sha1));
-      for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-           stage < ARRAY_SIZE(nir); stage++) {
-         if (!nir[stage])
-            continue;
-
-         nir_shaders->nir[stage] = nir_shader_clone(NULL, nir[stage]);
-      }
-
-      nir_shaders = tu_nir_cache_insert(builder->cache, nir_shaders);
-
-      if (compiled_shaders)
-         goto done;
-   }
-
-   compiled_shaders =
-      tu_shaders_init(builder->device, &pipeline_sha1, sizeof(pipeline_sha1));
-
-   if (!compiled_shaders) {
-      result = VK_ERROR_OUT_OF_HOST_MEMORY;
-      goto fail;
-   }
-
-   uint32_t desc_sets = 0;
-   for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-        stage < ARRAY_SIZE(nir); stage++) {
-      if (!nir[stage])
-         continue;
-
-      int64_t stage_start = os_time_get_nano();
-
       struct tu_shader *shader =
-         tu_shader_create(builder->device, nir[stage], &keys[stage],
-                          &builder->layout, builder->alloc);
-      if (!shader) {
-         result = VK_ERROR_OUT_OF_HOST_MEMORY;
-         goto fail;
-      }
-
-      /* In SPIR-V generated from GLSL, the primitive mode is specified in the
-       * tessellation evaluation shader, but in SPIR-V generated from HLSL,
-       * the mode is specified in the tessellation control shader. */
-      if ((stage == MESA_SHADER_TESS_EVAL || stage == MESA_SHADER_TESS_CTRL) &&
-          ir3_key.tessellation == IR3_TESS_NONE) {
-         ir3_key.tessellation = tu6_get_tessmode(shader);
-      }
-
-      if (stage > MESA_SHADER_TESS_CTRL) {
-         if (stage == MESA_SHADER_FRAGMENT) {
-            ir3_key.tcs_store_primid = ir3_key.tcs_store_primid ||
-               (nir[stage]->info.inputs_read & (1ull << VARYING_SLOT_PRIMITIVE_ID));
-         } else {
-            ir3_key.tcs_store_primid = ir3_key.tcs_store_primid ||
-               BITSET_TEST(nir[stage]->info.system_values_read, SYSTEM_VALUE_PRIMITIVE_ID);
-         }
-      }
-
-      /* Keep track of the status of each shader's active descriptor sets,
-       * which is set in tu_lower_io. */
-      desc_sets |= shader->active_desc_sets;
-
-      shaders[stage] = shader;
-
-      stage_feedbacks[stage].duration += os_time_get_nano() - stage_start;
-   }
-
-   /* In the the tess-but-not-FS case we don't know whether the FS will read
-    * PrimID so we need to unconditionally store it.
-    */
-   if (nir[MESA_SHADER_TESS_CTRL] && !nir[MESA_SHADER_FRAGMENT])
-      ir3_key.tcs_store_primid = true;
-
-   struct tu_shader *last_shader = shaders[MESA_SHADER_GEOMETRY];
-   if (!last_shader)
-      last_shader = shaders[MESA_SHADER_TESS_EVAL];
-   if (!last_shader)
-      last_shader = shaders[MESA_SHADER_VERTEX];
-
-   compiled_shaders->active_desc_sets = desc_sets;
-
-   for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-        stage < ARRAY_SIZE(shaders); stage++) {
-      if (!shaders[stage])
-         continue;
-
-      int64_t stage_start = os_time_get_nano();
-
-      compiled_shaders->variants[stage] =
-         ir3_shader_create_variant(shaders[stage]->ir3_shader, &ir3_key,
-                                   executable_info);
-      if (!compiled_shaders->variants[stage])
+         tu_shader_create(builder->device, stage, stage_info, builder->alloc);
+      if (!shader)
          return VK_ERROR_OUT_OF_HOST_MEMORY;
 
-      compiled_shaders->const_state[stage] = shaders[stage]->const_state;
-
-      stage_feedbacks[stage].duration += os_time_get_nano() - stage_start;
-   }
-
-   uint32_t safe_constlens = ir3_trim_constlen(compiled_shaders->variants, compiler);
-
-   ir3_key.safe_constlen = true;
-
-   for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-        stage < ARRAY_SIZE(shaders); stage++) {
-      if (!shaders[stage])
-         continue;
-
-      if (safe_constlens & (1 << stage)) {
-         int64_t stage_start = os_time_get_nano();
-
-         ralloc_free(compiled_shaders->variants[stage]);
-         compiled_shaders->variants[stage] =
-            ir3_shader_create_variant(shaders[stage]->ir3_shader, &ir3_key,
-                                      executable_info);
-         if (!compiled_shaders->variants[stage]) {
-            result = VK_ERROR_OUT_OF_HOST_MEMORY;
-            goto fail;
-         }
-
-         stage_feedbacks[stage].duration += os_time_get_nano() - stage_start;
-      } else if (contains_all_shader_state(builder->state)) {
-         compiled_shaders->safe_const_variants[stage] =
-            ir3_shader_create_variant(shaders[stage]->ir3_shader, &ir3_key,
-                                      executable_info);
-         if (!compiled_shaders->variants[stage]) {
-            result = VK_ERROR_OUT_OF_HOST_MEMORY;
-            goto fail;
-         }
-      }
-   }
-
-   ir3_key.safe_constlen = false;
-
-   for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-         stage < ARRAY_SIZE(nir); stage++) {
-      if (shaders[stage]) {
-         tu_shader_destroy(builder->device, shaders[stage], builder->alloc);
-      }
-   }
-
-   compiled_shaders =
-      tu_pipeline_cache_insert(builder->cache, compiled_shaders);
-
-done:;
-
-   struct ir3_shader_variant *safe_const_variants[ARRAY_SIZE(nir)] = { NULL };
-   nir_shader *post_link_nir[ARRAY_SIZE(nir)] = { NULL };
-
-   if (compiled_shaders) {
-      for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-            stage < ARRAY_SIZE(nir); stage++) {
-         if (compiled_shaders->variants[stage]) {
-            tu_append_executable(pipeline, compiled_shaders->variants[stage],
-               nir_initial_disasm[stage]);
-            builder->variants[stage] = compiled_shaders->variants[stage];
-            safe_const_variants[stage] =
-               compiled_shaders->safe_const_variants[stage];
-            builder->const_state[stage] =
-               compiled_shaders->const_state[stage];
-         }
-      }
-   }
-
-   if (nir_shaders) {
-      for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-            stage < ARRAY_SIZE(nir); stage++) {
-         if (nir_shaders->nir[stage]) {
-            post_link_nir[stage] = nir_shaders->nir[stage];
-         }
-      }
-   }
-
-   /* In the case where we're building a library without link-time
-    * optimization but with sub-libraries that retain LTO info, we should
-    * retain it ourselves in case another pipeline includes us with LTO.
-    */
-   for (unsigned i = 0; i < builder->num_libraries; i++) {
-      struct tu_pipeline *library = builder->libraries[i];
-      for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-            stage < ARRAY_SIZE(library->shaders); stage++) {
-         if (!post_link_nir[stage] && library->shaders[stage].nir) {
-            post_link_nir[stage] = library->shaders[stage].nir;
-            keys[stage] = library->shaders[stage].key;
-         }
-      }
-   }
+      VkResult result =
+         tu_shader_compile(builder->device, shader, next_stage_shader,
+                           &options, builder->alloc);
+      if (result != VK_SUCCESS)
+         return result;
 
-   if (!(builder->create_info->flags &
-         VK_PIPELINE_CREATE_LINK_TIME_OPTIMIZATION_BIT_EXT)) {
-      for (unsigned i = 0; i < builder->num_libraries; i++) {
-         struct tu_pipeline *library = builder->libraries[i];
-         for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-               stage < ARRAY_SIZE(library->shaders); stage++) {
-            if (library->shaders[stage].variant) {
-               assert(!builder->variants[stage]);
-               builder->variants[stage] = library->shaders[stage].variant;
-               safe_const_variants[stage] =
-                  library->shaders[stage].safe_const_variant;
-               builder->const_state[stage] =
-                  library->shaders[stage].const_state;
-               post_link_nir[stage] = library->shaders[stage].nir;
-            }
-         }
-      }
+      builder->shaders[stage] = shader;
+      builder->shader_offsets[stage] = builder->shader_total_size;
+      builder->shader_total_size +=
+         sizeof(uint32_t) * shader->variants[0].info.sizedwords;
 
-      /* Because we added more variants, we need to trim constlen again.
-       */
-      if (builder->num_libraries > 0) {
-         uint32_t safe_constlens = ir3_trim_constlen(builder->variants, compiler);
-         for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-              stage < ARRAY_SIZE(builder->variants); stage++) {
-            if (safe_constlens & (1u << stage))
-               builder->variants[stage] = safe_const_variants[stage];
-         }
-      }
+      next_stage_shader = shader;
    }
 
-   if (compiled_shaders)
-      pipeline->active_desc_sets = compiled_shaders->active_desc_sets;
-
-   for (unsigned i = 0; i < builder->num_libraries; i++) {
-      struct tu_pipeline *library = builder->libraries[i];
-      pipeline->active_desc_sets |= library->active_desc_sets;
+   if (builder->shaders[MESA_SHADER_VERTEX]->has_binning_pass) {
+      const struct tu_shader *vs = builder->shaders[MESA_SHADER_VERTEX];
+      builder->binning_vs_offset = builder->shader_total_size;
+      builder->shader_total_size +=
+         sizeof(uint32_t) * vs->variants[1].info.sizedwords;
    }
 
-   if (compiled_shaders && compiled_shaders->variants[MESA_SHADER_TESS_CTRL]) {
-      pipeline->tess.patch_type =
-         compiled_shaders->variants[MESA_SHADER_TESS_CTRL]->key.tessellation;
-   }
+   return VK_SUCCESS;
+}
 
-   if (contains_all_shader_state(pipeline->state)) {
-      struct ir3_shader_variant *vs =
-         builder->variants[MESA_SHADER_VERTEX];
+static VkResult
+tu_pipeline_builder_upload_shaders(struct tu_pipeline_builder *builder,
+                                   struct tu_pipeline *pipeline)
+{
+   struct tu_bo *bo = &pipeline->program.binary_bo;
 
-      struct ir3_shader_variant *variant;
-      if (!vs->stream_output.num_outputs && ir3_has_binning_vs(&vs->key)) {
-         tu_append_executable(pipeline, vs->binning, NULL);
-         variant = vs->binning;
-      } else {
-         variant = vs;
-      }
+   VkResult result =
+      tu_bo_init_new(builder->device, bo, builder->shader_total_size);
+   if (result != VK_SUCCESS)
+      return result;
 
-      builder->binning_variant = variant;
+   result = tu_bo_map(builder->device, bo);
+   if (result != VK_SUCCESS)
+      return result;
 
-      builder->compiled_shaders = compiled_shaders;
+   for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
+      const struct tu_shader *shader = builder->shaders[i];
+      if (!shader)
+         continue;
 
-      /* It doesn't make much sense to use RETAIN_LINK_TIME_OPTIMIZATION_INFO
-       * when compiling all stages, but make sure we don't leak.
-       */
-      if (nir_shaders)
-         vk_pipeline_cache_object_unref(&nir_shaders->base);
-   } else {
-      pipeline->compiled_shaders = compiled_shaders;
-      pipeline->nir_shaders = nir_shaders;
-      pipeline->ir3_key = ir3_key;
-      for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-           stage < ARRAY_SIZE(pipeline->shaders); stage++) {
-         pipeline->shaders[stage].nir = post_link_nir[stage];
-         pipeline->shaders[stage].key = keys[stage];
-         pipeline->shaders[stage].const_state = builder->const_state[stage];
-         pipeline->shaders[stage].variant = builder->variants[stage];
-         pipeline->shaders[stage].safe_const_variant =
-            safe_const_variants[stage];
-      }
+      memcpy(bo->map + builder->shader_offsets[i], shader->binary,
+             sizeof(uint32_t) * shader->variants[0].info.sizedwords);
    }
 
-   pipeline_feedback.duration = os_time_get_nano() - pipeline_start;
-   if (creation_feedback) {
-      *creation_feedback->pPipelineCreationFeedback = pipeline_feedback;
-
-      for (uint32_t i = 0; i < builder->create_info->stageCount; i++) {
-         gl_shader_stage s =
-            vk_to_mesa_shader_stage(builder->create_info->pStages[i].stage);
-         creation_feedback->pPipelineStageCreationFeedbacks[i] = stage_feedbacks[s];
-      }
+   if (builder->shaders[MESA_SHADER_VERTEX]->has_binning_pass) {
+      const struct tu_shader *vs = builder->shaders[MESA_SHADER_VERTEX];
+      memcpy(bo->map + builder->binning_vs_offset, vs->binning_binary,
+             sizeof(uint32_t) * vs->variants[1].info.sizedwords);
    }
 
    return VK_SUCCESS;
-
-fail:
-   for (gl_shader_stage stage = MESA_SHADER_VERTEX;
-         stage < ARRAY_SIZE(nir); stage++) {
-      if (shaders[stage]) {
-         tu_shader_destroy(builder->device, shaders[stage], builder->alloc);
-      }
-   }
-
-   if (compiled_shaders)
-      vk_pipeline_cache_object_unref(&compiled_shaders->base);
-
-   if (nir_shaders)
-      vk_pipeline_cache_object_unref(&nir_shaders->base);
-
-   return result;
 }
 
 static void
@@ -3589,449 +1485,56 @@ tu_pipeline_builder_parse_dynamic(struct tu_pipeline_builder *builder,
    const VkPipelineDynamicStateCreateInfo *dynamic_info =
       builder->create_info->pDynamicState;
 
-   pipeline->rast.gras_su_cntl_mask = ~0u;
-   pipeline->rast.pc_raster_cntl_mask = ~0u;
-   pipeline->rast.vpc_unknown_9107_mask = ~0u;
-   pipeline->ds.rb_depth_cntl_mask = ~0u;
-   pipeline->ds.rb_stencil_cntl_mask = ~0u;
-   pipeline->blend.sp_blend_cntl_mask = ~0u;
-   pipeline->blend.rb_blend_cntl_mask = ~0u;
-   pipeline->blend.rb_mrt_control_mask = ~0u;
-
    if (!dynamic_info)
       return;
 
    for (uint32_t i = 0; i < dynamic_info->dynamicStateCount; i++) {
-      VkDynamicState state = dynamic_info->pDynamicStates[i];
-      switch (state) {
-      case VK_DYNAMIC_STATE_VIEWPORT ... VK_DYNAMIC_STATE_STENCIL_REFERENCE:
-         if (state == VK_DYNAMIC_STATE_LINE_WIDTH)
-            pipeline->rast.gras_su_cntl_mask &= ~A6XX_GRAS_SU_CNTL_LINEHALFWIDTH__MASK;
-         pipeline->dynamic_state_mask |= BIT(state);
-         break;
-      case VK_DYNAMIC_STATE_SAMPLE_LOCATIONS_EXT:
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_SAMPLE_LOCATIONS);
-         break;
-      case VK_DYNAMIC_STATE_CULL_MODE:
-         pipeline->rast.gras_su_cntl_mask &=
-            ~(A6XX_GRAS_SU_CNTL_CULL_BACK | A6XX_GRAS_SU_CNTL_CULL_FRONT);
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_GRAS_SU_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_FRONT_FACE:
-         pipeline->rast.gras_su_cntl_mask &= ~A6XX_GRAS_SU_CNTL_FRONT_CW;
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_GRAS_SU_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_PRIMITIVE_TOPOLOGY:
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_PRIMITIVE_TOPOLOGY);
-         break;
-      case VK_DYNAMIC_STATE_VERTEX_INPUT_BINDING_STRIDE:
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_VB_STRIDE);
-         break;
-      case VK_DYNAMIC_STATE_VIEWPORT_WITH_COUNT:
-         pipeline->dynamic_state_mask |= BIT(VK_DYNAMIC_STATE_VIEWPORT);
-         break;
-      case VK_DYNAMIC_STATE_SCISSOR_WITH_COUNT:
-         pipeline->dynamic_state_mask |= BIT(VK_DYNAMIC_STATE_SCISSOR);
-         break;
-      case VK_DYNAMIC_STATE_DEPTH_TEST_ENABLE:
-         pipeline->ds.rb_depth_cntl_mask &=
-            ~(A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE | A6XX_RB_DEPTH_CNTL_Z_READ_ENABLE);
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_RB_DEPTH_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_DEPTH_WRITE_ENABLE:
-         pipeline->ds.rb_depth_cntl_mask &= ~A6XX_RB_DEPTH_CNTL_Z_WRITE_ENABLE;
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_RB_DEPTH_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_DEPTH_COMPARE_OP:
-         pipeline->ds.rb_depth_cntl_mask &= ~A6XX_RB_DEPTH_CNTL_ZFUNC__MASK;
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_RB_DEPTH_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_DEPTH_BOUNDS_TEST_ENABLE:
-         pipeline->ds.rb_depth_cntl_mask &=
-            ~(A6XX_RB_DEPTH_CNTL_Z_BOUNDS_ENABLE | A6XX_RB_DEPTH_CNTL_Z_READ_ENABLE);
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_RB_DEPTH_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_STENCIL_TEST_ENABLE:
-         pipeline->ds.rb_stencil_cntl_mask &= ~(A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE |
-                                                A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE_BF |
-                                                A6XX_RB_STENCIL_CONTROL_STENCIL_READ);
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_RB_STENCIL_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_STENCIL_OP:
-         pipeline->ds.rb_stencil_cntl_mask &= ~(A6XX_RB_STENCIL_CONTROL_FUNC__MASK |
-                                                A6XX_RB_STENCIL_CONTROL_FAIL__MASK |
-                                                A6XX_RB_STENCIL_CONTROL_ZPASS__MASK |
-                                                A6XX_RB_STENCIL_CONTROL_ZFAIL__MASK |
-                                                A6XX_RB_STENCIL_CONTROL_FUNC_BF__MASK |
-                                                A6XX_RB_STENCIL_CONTROL_FAIL_BF__MASK |
-                                                A6XX_RB_STENCIL_CONTROL_ZPASS_BF__MASK |
-                                                A6XX_RB_STENCIL_CONTROL_ZFAIL_BF__MASK);
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_RB_STENCIL_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_DEPTH_BIAS_ENABLE:
-         pipeline->rast.gras_su_cntl_mask &= ~A6XX_GRAS_SU_CNTL_POLY_OFFSET;
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_GRAS_SU_CNTL);
-         break;
-      case VK_DYNAMIC_STATE_PRIMITIVE_RESTART_ENABLE:
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_PRIMITIVE_RESTART_ENABLE);
-         break;
-      case VK_DYNAMIC_STATE_RASTERIZER_DISCARD_ENABLE:
-         pipeline->rast.pc_raster_cntl_mask &= ~A6XX_PC_RASTER_CNTL_DISCARD;
-         pipeline->rast.vpc_unknown_9107_mask &= ~A6XX_VPC_UNKNOWN_9107_RASTER_DISCARD;
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_RASTERIZER_DISCARD);
-         break;
-      case VK_DYNAMIC_STATE_LOGIC_OP_EXT:
-         pipeline->blend.sp_blend_cntl_mask &= ~A6XX_SP_BLEND_CNTL_ENABLE_BLEND__MASK;
-         pipeline->blend.rb_blend_cntl_mask &= ~A6XX_RB_BLEND_CNTL_ENABLE_BLEND__MASK;
-         pipeline->blend.rb_mrt_control_mask &= ~A6XX_RB_MRT_CONTROL_ROP_CODE__MASK;
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_BLEND);
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_LOGIC_OP);
-         break;
-      case VK_DYNAMIC_STATE_COLOR_WRITE_ENABLE_EXT:
-         pipeline->blend.sp_blend_cntl_mask &= ~A6XX_SP_BLEND_CNTL_ENABLE_BLEND__MASK;
-         pipeline->blend.rb_blend_cntl_mask &= ~A6XX_RB_BLEND_CNTL_ENABLE_BLEND__MASK;
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_BLEND);
-
-         /* Dynamic color write enable doesn't directly change any of the
-          * registers, but it causes us to make some of the registers 0, so we
-          * set this dynamic state instead of making the register dynamic.
-          */
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_COLOR_WRITE_ENABLE);
-         break;
-      case VK_DYNAMIC_STATE_VERTEX_INPUT_EXT:
-         pipeline->dynamic_state_mask |= BIT(TU_DYNAMIC_STATE_VERTEX_INPUT) |
-            BIT(TU_DYNAMIC_STATE_VB_STRIDE);
-         break;
-      case VK_DYNAMIC_STATE_PATCH_CONTROL_POINTS_EXT:
-         pipeline->dynamic_state_mask |=
-            BIT(TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS);
-         break;
-      default:
-         assert(!"unsupported dynamic state");
-         break;
-      }
-   }
-}
-
-static void
-tu_pipeline_builder_parse_libraries(struct tu_pipeline_builder *builder,
-                                    struct tu_pipeline *pipeline)
-{
-   const VkPipelineLibraryCreateInfoKHR *library_info =
-      vk_find_struct_const(builder->create_info->pNext,
-                           PIPELINE_LIBRARY_CREATE_INFO_KHR);
-
-   if (library_info) {
-      assert(library_info->libraryCount <= MAX_LIBRARIES);
-      builder->num_libraries = library_info->libraryCount;
-      for (unsigned i = 0; i < library_info->libraryCount; i++) {
-         TU_FROM_HANDLE(tu_pipeline, library, library_info->pLibraries[i]);
-         builder->libraries[i] = library;
-      }
-   }
-
-   /* Merge in the state from libraries. The program state is a bit special
-    * and is handled separately.
-    */
-   pipeline->state = builder->state;
-   for (unsigned i = 0; i < builder->num_libraries; i++) {
-      struct tu_pipeline *library = builder->libraries[i];
-      pipeline->state |= library->state;
-
-      uint32_t library_dynamic_state = 0;
-      if (library->state &
-          VK_GRAPHICS_PIPELINE_LIBRARY_VERTEX_INPUT_INTERFACE_BIT_EXT) {
-         pipeline->vi = library->vi;
-         pipeline->ia = library->ia;
-         library_dynamic_state |=
-            BIT(TU_DYNAMIC_STATE_VERTEX_INPUT) |
-            BIT(TU_DYNAMIC_STATE_VB_STRIDE) |
-            BIT(TU_DYNAMIC_STATE_PRIMITIVE_TOPOLOGY) |
-            BIT(TU_DYNAMIC_STATE_PRIMITIVE_RESTART_ENABLE);
-         pipeline->shared_consts = library->shared_consts;
-      }
-
-      if (library->state &
-          VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT) {
-         pipeline->tess = library->tess;
-         pipeline->rast = library->rast;
-         pipeline->viewport = library->viewport;
-         library_dynamic_state |=
-            BIT(VK_DYNAMIC_STATE_VIEWPORT) |
-            BIT(VK_DYNAMIC_STATE_SCISSOR) |
-            BIT(VK_DYNAMIC_STATE_LINE_WIDTH) |
-            BIT(VK_DYNAMIC_STATE_DEPTH_BIAS) |
-            BIT(TU_DYNAMIC_STATE_RASTERIZER_DISCARD) |
-            BIT(TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS);
-      }
-
-      if (library->state &
-          VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT) {
-         pipeline->ds = library->ds;
-         pipeline->lrz.fs = library->lrz.fs;
-         pipeline->lrz.force_disable_mask |= library->lrz.force_disable_mask;
-         pipeline->lrz.force_late_z |= library->lrz.force_late_z;
-         library_dynamic_state |=
-            BIT(VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK) |
-            BIT(VK_DYNAMIC_STATE_STENCIL_WRITE_MASK) |
-            BIT(VK_DYNAMIC_STATE_STENCIL_REFERENCE) |
-            BIT(TU_DYNAMIC_STATE_RB_DEPTH_CNTL) |
-            BIT(TU_DYNAMIC_STATE_RB_STENCIL_CNTL) |
-            BIT(VK_DYNAMIC_STATE_DEPTH_BOUNDS);
-         pipeline->shared_consts = library->shared_consts;
-      }
-
-      if (library->state &
-          VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT) {
-         pipeline->blend = library->blend;
-         pipeline->output = library->output;
-         pipeline->lrz.force_disable_mask |= library->lrz.force_disable_mask;
-         pipeline->lrz.force_late_z |= library->lrz.force_late_z;
-         pipeline->prim_order = library->prim_order;
-         library_dynamic_state |=
-            BIT(VK_DYNAMIC_STATE_BLEND_CONSTANTS) |
-            BIT(TU_DYNAMIC_STATE_SAMPLE_LOCATIONS) |
-            BIT(TU_DYNAMIC_STATE_BLEND) |
-            BIT(TU_DYNAMIC_STATE_LOGIC_OP) |
-            BIT(TU_DYNAMIC_STATE_COLOR_WRITE_ENABLE);
-      }
-
-      if ((library->state &
-           VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT) &&
-          (library->state &
-           VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT)) {
-         pipeline->prim_order = library->prim_order;
-      }
-
-      if ((library->state &
-           VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT) &&
-          (library->state &
-           VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT) &&
-          (library->state &
-           VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT)) {
-         pipeline->rast_ds = library->rast_ds;
-      }
-
-      pipeline->dynamic_state_mask =
-         (pipeline->dynamic_state_mask & ~library_dynamic_state) |
-         (library->dynamic_state_mask & library_dynamic_state);
-
-      u_foreach_bit (i, library_dynamic_state & ~library->dynamic_state_mask) {
-         if (i >= TU_DYNAMIC_STATE_COUNT)
-            break;
-
-         pipeline->dynamic_state[i] = library->dynamic_state[i];
-      }
-
-      if (contains_all_shader_state(library->state)) {
-         pipeline->program = library->program;
-         pipeline->load_state = library->load_state;
-      }
+      pipeline->dynamic_state.mask |=
+         tu_dynamic_state_bit(dynamic_info->pDynamicStates[i]);
    }
 }
 
 static void
-tu_pipeline_builder_parse_layout(struct tu_pipeline_builder *builder,
-                                 struct tu_pipeline *pipeline)
-{
-   TU_FROM_HANDLE(tu_pipeline_layout, layout, builder->create_info->layout);
-
-   if (layout) {
-      /* Note: it's still valid to have a layout even if there are libraries.
-       * This allows the app to e.g. overwrite an INDEPENDENT_SET layout with
-       * a non-INDEPENDENT_SET layout which may make us use a faster path,
-       * currently this just affects dynamic offset descriptors.
-       */
-      builder->layout = *layout;
-   } else {
-      for (unsigned i = 0; i < builder->num_libraries; i++) {
-         struct tu_pipeline *library = builder->libraries[i];
-         builder->layout.num_sets = MAX2(builder->layout.num_sets,
-                                         library->num_sets);
-         for (unsigned j = 0; j < library->num_sets; j++) {
-            if (library->layouts[i])
-               builder->layout.set[i].layout = library->layouts[i];
-         }
-
-         builder->layout.push_constant_size = pipeline->push_constant_size;
-         builder->layout.independent_sets |= pipeline->independent_sets;
-      }
-
-      tu_pipeline_layout_init(&builder->layout);
-   }
-
-   if (builder->create_info->flags & VK_PIPELINE_CREATE_LIBRARY_BIT_KHR) {
-      pipeline->num_sets = builder->layout.num_sets;
-      for (unsigned i = 0; i < pipeline->num_sets; i++) {
-         pipeline->layouts[i] = builder->layout.set[i].layout;
-         if (pipeline->layouts[i])
-            vk_descriptor_set_layout_ref(&pipeline->layouts[i]->vk);
-      }
-      pipeline->push_constant_size = builder->layout.push_constant_size;
-      pipeline->independent_sets = builder->layout.independent_sets;
-   }
-}
-
-static void
-tu_pipeline_set_linkage(struct tu_program_descriptor_linkage *link,
-                        struct tu_const_state *const_state,
-                        struct ir3_shader_variant *v)
-{
-   link->const_state = *ir3_const_state(v);
-   link->tu_const_state = *const_state;
-   link->constlen = v->constlen;
-}
-
-static bool
-tu_pipeline_static_state(struct tu_pipeline *pipeline, struct tu_cs *cs,
-                         uint32_t id, uint32_t size)
-{
-   assert(id < ARRAY_SIZE(pipeline->dynamic_state));
-
-   if (pipeline->dynamic_state_mask & BIT(id))
-      return false;
-
-   pipeline->dynamic_state[id] = tu_cs_draw_state(&pipeline->cs, cs, size);
-   return true;
-}
-
-static void
 tu_pipeline_builder_parse_shader_stages(struct tu_pipeline_builder *builder,
                                         struct tu_pipeline *pipeline)
 {
    struct tu_cs prog_cs;
+   tu_cs_begin_sub_stream(builder->device, &pipeline->cs, 512, &prog_cs);
+   tu6_emit_program(&prog_cs, builder, &pipeline->program.binary_bo, false);
+   pipeline->program.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &prog_cs);
 
-   /* Emit HLSQ_xS_CNTL/HLSQ_SP_xS_CONFIG *first*, before emitting anything
-    * else that could depend on that state (like push constants)
-    *
-    * Note also that this always uses the full VS even in binning pass.  The
-    * binning pass variant has the same const layout as the full VS, and
-    * the constlen for the VS will be the same or greater than the constlen
-    * for the binning pass variant.  It is required that the constlen state
-    * matches between binning and draw passes, as some parts of the push
-    * consts are emitted in state groups that are shared between the binning
-    * and draw passes.
-    */
-   tu_cs_begin_sub_stream(&pipeline->cs, 512, &prog_cs);
-   tu6_emit_program_config(&prog_cs, builder);
-   pipeline->program.config_state = tu_cs_end_draw_state(&pipeline->cs, &prog_cs);
-
-   tu_cs_begin_sub_stream(&pipeline->cs, 512 + builder->additional_cs_reserve_size, &prog_cs);
-   tu6_emit_program(&prog_cs, builder, false, pipeline);
-   pipeline->program.state = tu_cs_end_draw_state(&pipeline->cs, &prog_cs);
-
-   tu_cs_begin_sub_stream(&pipeline->cs, 512 + builder->additional_cs_reserve_size, &prog_cs);
-   tu6_emit_program(&prog_cs, builder, true, pipeline);
-   pipeline->program.binning_state = tu_cs_end_draw_state(&pipeline->cs, &prog_cs);
-
-   for (unsigned i = 0; i < ARRAY_SIZE(builder->variants); i++) {
-      if (!builder->variants[i])
-         continue;
-
-      tu_pipeline_set_linkage(&pipeline->program.link[i],
-                              &builder->const_state[i],
-                              builder->variants[i]);
-   }
-
-   struct ir3_shader_variant *vs = builder->variants[MESA_SHADER_VERTEX];
-   struct ir3_shader_variant *hs = builder->variants[MESA_SHADER_TESS_CTRL];
-   if (hs) {
-      pipeline->program.vs_param_stride = vs->output_size;
-      pipeline->program.hs_param_stride = hs->output_size;
-      pipeline->program.hs_vertices_out = hs->tess.tcs_vertices_out;
-
-      const struct ir3_const_state *hs_const =
-         &pipeline->program.link[MESA_SHADER_TESS_CTRL].const_state;
-      unsigned hs_constlen =
-         pipeline->program.link[MESA_SHADER_TESS_CTRL].constlen;
-      uint32_t hs_base = hs_const->offsets.primitive_param;
-      pipeline->program.hs_param_dwords =
-         MIN2((hs_constlen - hs_base) * 4, 8);
-
-      uint32_t state_size = TU6_EMIT_PATCH_CONTROL_POINTS_DWORDS(
-         pipeline->program.hs_param_dwords);
-
-      struct tu_cs cs;
-      if (tu_pipeline_static_state(pipeline, &cs,
-                                   TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS,
-                                   state_size)) {
-         tu6_emit_patch_control_points(&cs, pipeline,
-                                       pipeline->tess.patch_control_points);
-      }
-   }
+   tu_cs_begin_sub_stream(builder->device, &pipeline->cs, 512, &prog_cs);
+   tu6_emit_program(&prog_cs, builder, &pipeline->program.binary_bo, true);
+   pipeline->program.binning_state_ib =
+      tu_cs_end_sub_stream(&pipeline->cs, &prog_cs);
 }
 
 static void
 tu_pipeline_builder_parse_vertex_input(struct tu_pipeline_builder *builder,
                                        struct tu_pipeline *pipeline)
 {
-   if (pipeline->dynamic_state_mask & BIT(TU_DYNAMIC_STATE_VERTEX_INPUT))
-      return;
-
    const VkPipelineVertexInputStateCreateInfo *vi_info =
       builder->create_info->pVertexInputState;
+   const struct tu_shader *vs = builder->shaders[MESA_SHADER_VERTEX];
 
-   struct tu_cs cs;
-   if (tu_pipeline_static_state(pipeline, &cs, TU_DYNAMIC_STATE_VB_STRIDE,
-                                2 * vi_info->vertexBindingDescriptionCount)) {
-      for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
-         const VkVertexInputBindingDescription *binding =
-            &vi_info->pVertexBindingDescriptions[i];
+   struct tu_cs vi_cs;
+   tu_cs_begin_sub_stream(builder->device, &pipeline->cs,
+                          MAX_VERTEX_ATTRIBS * 5 + 2, &vi_cs);
+   tu6_emit_vertex_input(&vi_cs, &vs->variants[0], vi_info,
+                         pipeline->vi.bindings, pipeline->vi.strides,
+                         pipeline->vi.offsets, &pipeline->vi.count);
+   pipeline->vi.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &vi_cs);
 
-         tu_cs_emit_regs(&cs,
-                         A6XX_VFD_FETCH_STRIDE(binding->binding, binding->stride));
-      }
-   }
-
-   VkVertexInputBindingDescription2EXT bindings[MAX_VBS];
-   VkVertexInputAttributeDescription2EXT attrs[MAX_VERTEX_ATTRIBS];
-
-   for (unsigned i = 0; i < vi_info->vertexBindingDescriptionCount; i++) {
-      const VkVertexInputBindingDescription *binding =
-         &vi_info->pVertexBindingDescriptions[i];
-      bindings[i] = (VkVertexInputBindingDescription2EXT) {
-         .sType = VK_STRUCTURE_TYPE_VERTEX_INPUT_BINDING_DESCRIPTION_2_EXT,
-         .pNext = NULL,
-         .binding = binding->binding,
-         .inputRate = binding->inputRate,
-         .stride = binding->stride,
-         .divisor = 1,
-      };
-
-      /* Bindings may contain holes */
-      pipeline->vi.num_vbs = MAX2(pipeline->vi.num_vbs, binding->binding + 1);
-   }
-
-   const VkPipelineVertexInputDivisorStateCreateInfoEXT *div_state =
-      vk_find_struct_const(vi_info->pNext, PIPELINE_VERTEX_INPUT_DIVISOR_STATE_CREATE_INFO_EXT);
-   if (div_state) {
-      for (uint32_t i = 0; i < div_state->vertexBindingDivisorCount; i++) {
-         const VkVertexInputBindingDivisorDescriptionEXT *desc =
-            &div_state->pVertexBindingDivisors[i];
-         bindings[desc->binding].divisor = desc->divisor;
-      }
+   if (vs->has_binning_pass) {
+      tu_cs_begin_sub_stream(builder->device, &pipeline->cs,
+                             MAX_VERTEX_ATTRIBS * 5 + 2, &vi_cs);
+      tu6_emit_vertex_input(
+         &vi_cs, &vs->variants[1], vi_info, pipeline->vi.binning_bindings,
+         pipeline->vi.binning_strides, pipeline->vi.binning_offsets,
+         &pipeline->vi.binning_count);
+      pipeline->vi.binning_state_ib =
+         tu_cs_end_sub_stream(&pipeline->cs, &vi_cs);
    }
-
-   for (unsigned i = 0; i < vi_info->vertexAttributeDescriptionCount; i++) {
-      const VkVertexInputAttributeDescription *attr =
-         &vi_info->pVertexAttributeDescriptions[i];
-      attrs[i] = (VkVertexInputAttributeDescription2EXT) {
-         .sType = VK_STRUCTURE_TYPE_VERTEX_INPUT_ATTRIBUTE_DESCRIPTION_2_EXT,
-         .pNext = NULL,
-         .binding = attr->binding,
-         .location = attr->location,
-         .offset = attr->offset,
-         .format = attr->format,
-      };
-   }
-
-   tu_cs_begin_sub_stream(&pipeline->cs,
-                          TU6_EMIT_VERTEX_INPUT_MAX_DWORDS, &cs);
-   tu6_emit_vertex_input(&cs,
-                         vi_info->vertexBindingDescriptionCount, bindings,
-                         vi_info->vertexAttributeDescriptionCount, attrs);
-   pipeline->dynamic_state[TU_DYNAMIC_STATE_VERTEX_INPUT] =
-      tu_cs_end_draw_state(&pipeline->cs, &cs);
 }
 
 static void
@@ -4046,29 +1549,6 @@ tu_pipeline_builder_parse_input_assembly(struct tu_pipeline_builder *builder,
 }
 
 static void
-tu_pipeline_builder_parse_tessellation(struct tu_pipeline_builder *builder,
-                                       struct tu_pipeline *pipeline)
-{
-   if (!(builder->active_stages & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT) ||
-       !(builder->active_stages & VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT))
-      return;
-
-   const VkPipelineTessellationStateCreateInfo *tess_info =
-      builder->create_info->pTessellationState;
-
-   if (!(pipeline->dynamic_state_mask &
-         BIT(TU_DYNAMIC_STATE_PATCH_CONTROL_POINTS))) {
-      assert(tess_info->patchControlPoints <= 32);
-      pipeline->tess.patch_control_points = tess_info->patchControlPoints;
-   }
-
-   const VkPipelineTessellationDomainOriginStateCreateInfo *domain_info =
-         vk_find_struct_const(tess_info->pNext, PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO);
-   pipeline->tess.upper_left_domain_origin = !domain_info ||
-         domain_info->domainOrigin == VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT;
-}
-
-static void
 tu_pipeline_builder_parse_viewport(struct tu_pipeline_builder *builder,
                                    struct tu_pipeline *pipeline)
 {
@@ -4085,17 +1565,21 @@ tu_pipeline_builder_parse_viewport(struct tu_pipeline_builder *builder,
 
    const VkPipelineViewportStateCreateInfo *vp_info =
       builder->create_info->pViewportState;
-   const VkPipelineViewportDepthClipControlCreateInfoEXT *depth_clip_info =
-         vk_find_struct_const(vp_info->pNext, PIPELINE_VIEWPORT_DEPTH_CLIP_CONTROL_CREATE_INFO_EXT);
-   pipeline->viewport.z_negative_one_to_one = depth_clip_info ? depth_clip_info->negativeOneToOne : false;
 
-   struct tu_cs cs;
+   struct tu_cs vp_cs;
+   tu_cs_begin_sub_stream(builder->device, &pipeline->cs, 15, &vp_cs);
 
-   if (tu_pipeline_static_state(pipeline, &cs, VK_DYNAMIC_STATE_VIEWPORT, 8 + 10 * vp_info->viewportCount))
-      tu6_emit_viewport(&cs, vp_info->pViewports, vp_info->viewportCount, pipeline->viewport.z_negative_one_to_one);
+   if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_VIEWPORT)) {
+      assert(vp_info->viewportCount == 1);
+      tu6_emit_viewport(&vp_cs, vp_info->pViewports);
+   }
 
-   if (tu_pipeline_static_state(pipeline, &cs, VK_DYNAMIC_STATE_SCISSOR, 1 + 2 * vp_info->scissorCount))
-      tu6_emit_scissor(&cs, vp_info->pScissors, vp_info->scissorCount);
+   if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_SCISSOR)) {
+      assert(vp_info->scissorCount == 1);
+      tu6_emit_scissor(&vp_cs, vp_info->pScissors);
+   }
+
+   pipeline->vp.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &vp_cs);
 }
 
 static void
@@ -4105,95 +1589,31 @@ tu_pipeline_builder_parse_rasterization(struct tu_pipeline_builder *builder,
    const VkPipelineRasterizationStateCreateInfo *rast_info =
       builder->create_info->pRasterizationState;
 
-   enum a6xx_polygon_mode mode = tu6_polygon_mode(rast_info->polygonMode);
-
-   bool depth_clip_disable = rast_info->depthClampEnable;
-
-   const VkPipelineRasterizationDepthClipStateCreateInfoEXT *depth_clip_state =
-      vk_find_struct_const(rast_info, PIPELINE_RASTERIZATION_DEPTH_CLIP_STATE_CREATE_INFO_EXT);
-   if (depth_clip_state)
-      depth_clip_disable = !depth_clip_state->depthClipEnable;
-
-   pipeline->rast.rb_depth_cntl =
-      COND(rast_info->depthClampEnable, A6XX_RB_DEPTH_CNTL_Z_CLAMP_ENABLE);
-
-   pipeline->rast.line_mode = RECTANGULAR;
-
-   const VkPipelineRasterizationLineStateCreateInfoEXT *rast_line_state =
-      vk_find_struct_const(rast_info->pNext,
-                           PIPELINE_RASTERIZATION_LINE_STATE_CREATE_INFO_EXT);
-
-   if (rast_line_state &&
-       rast_line_state->lineRasterizationMode == VK_LINE_RASTERIZATION_MODE_BRESENHAM_EXT) {
-      pipeline->rast.line_mode = BRESENHAM;
-   }
-
-   struct tu_cs cs;
-   uint32_t cs_size = 9 +
-      (builder->device->physical_device->info->a6xx.has_shading_rate ? 8 : 0);
-   pipeline->rast.state = tu_cs_draw_state(&pipeline->cs, &cs, cs_size);
-
-   tu_cs_emit_regs(&cs,
-                   A6XX_GRAS_CL_CNTL(
-                     .znear_clip_disable = depth_clip_disable,
-                     .zfar_clip_disable = depth_clip_disable,
-                     .z_clamp_enable = rast_info->depthClampEnable,
-                     .zero_gb_scale_z = pipeline->viewport.z_negative_one_to_one ? 0 : 1,
-                     .vp_clip_code_ignore = 1));
+   assert(!rast_info->depthClampEnable);
+   assert(rast_info->polygonMode == VK_POLYGON_MODE_FILL);
 
-   tu_cs_emit_regs(&cs,
-                   A6XX_VPC_POLYGON_MODE(mode));
-
-   tu_cs_emit_regs(&cs,
-                   A6XX_PC_POLYGON_MODE(mode));
+   struct tu_cs rast_cs;
+   tu_cs_begin_sub_stream(builder->device, &pipeline->cs, 20, &rast_cs);
 
    /* move to hw ctx init? */
-   tu_cs_emit_regs(&cs,
-                   A6XX_GRAS_SU_POINT_MINMAX(.min = 1.0f / 16.0f, .max = 4092.0f),
-                   A6XX_GRAS_SU_POINT_SIZE(1.0f));
-
-   if (builder->device->physical_device->info->a6xx.has_shading_rate) {
-      tu_cs_emit_regs(&cs, A6XX_RB_UNKNOWN_8A00());
-      tu_cs_emit_regs(&cs, A6XX_RB_UNKNOWN_8A10());
-      tu_cs_emit_regs(&cs, A6XX_RB_UNKNOWN_8A20());
-      tu_cs_emit_regs(&cs, A6XX_RB_UNKNOWN_8A30());
-   }
-
-   const VkPipelineRasterizationStateStreamCreateInfoEXT *stream_info =
-      vk_find_struct_const(rast_info->pNext,
-                           PIPELINE_RASTERIZATION_STATE_STREAM_CREATE_INFO_EXT);
-   unsigned stream = stream_info ? stream_info->rasterizationStream : 0;
-
-   pipeline->rast.pc_raster_cntl = A6XX_PC_RASTER_CNTL_STREAM(stream);
-   pipeline->rast.vpc_unknown_9107 = 0;
-   if (rast_info->rasterizerDiscardEnable) {
-      pipeline->rast.pc_raster_cntl |= A6XX_PC_RASTER_CNTL_DISCARD;
-      pipeline->rast.vpc_unknown_9107 |= A6XX_VPC_UNKNOWN_9107_RASTER_DISCARD;
-   }
+   tu6_emit_gras_unknowns(&rast_cs);
+   tu6_emit_point_size(&rast_cs);
 
-   if (tu_pipeline_static_state(pipeline, &cs, TU_DYNAMIC_STATE_RASTERIZER_DISCARD, 4)) {
-      tu_cs_emit_regs(&cs, A6XX_PC_RASTER_CNTL(.dword = pipeline->rast.pc_raster_cntl));
-      tu_cs_emit_regs(&cs, A6XX_VPC_UNKNOWN_9107(.dword = pipeline->rast.vpc_unknown_9107));
-   }
-
-   pipeline->rast.gras_su_cntl =
-      tu6_gras_su_cntl(rast_info, pipeline->rast.line_mode, builder->multiview_mask != 0);
+   const uint32_t gras_su_cntl =
+      tu6_gras_su_cntl(rast_info, builder->samples);
 
-   if (tu_pipeline_static_state(pipeline, &cs, TU_DYNAMIC_STATE_GRAS_SU_CNTL, 2))
-      tu_cs_emit_regs(&cs, A6XX_GRAS_SU_CNTL(.dword = pipeline->rast.gras_su_cntl));
+   if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_LINE_WIDTH))
+      tu6_emit_gras_su_cntl(&rast_cs, gras_su_cntl, rast_info->lineWidth);
 
-   if (tu_pipeline_static_state(pipeline, &cs, VK_DYNAMIC_STATE_DEPTH_BIAS, 4)) {
-      tu6_emit_depth_bias(&cs, rast_info->depthBiasConstantFactor,
+   if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_DEPTH_BIAS)) {
+      tu6_emit_depth_bias(&rast_cs, rast_info->depthBiasConstantFactor,
                           rast_info->depthBiasClamp,
                           rast_info->depthBiasSlopeFactor);
    }
 
-   const struct VkPipelineRasterizationProvokingVertexStateCreateInfoEXT *provoking_vtx_state =
-      vk_find_struct_const(rast_info->pNext, PIPELINE_RASTERIZATION_PROVOKING_VERTEX_STATE_CREATE_INFO_EXT);
-   pipeline->rast.provoking_vertex_last = provoking_vtx_state &&
-      provoking_vtx_state->provokingVertexMode == VK_PROVOKING_VERTEX_MODE_LAST_VERTEX_EXT;
+   pipeline->rast.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &rast_cs);
 
-   pipeline->rast.multiview_mask = builder->multiview_mask;
+   pipeline->rast.gras_su_cntl = gras_su_cntl;
 }
 
 static void
@@ -4207,128 +1627,38 @@ tu_pipeline_builder_parse_depth_stencil(struct tu_pipeline_builder *builder,
     *    the pipeline has rasterization disabled or if the subpass of the
     *    render pass the pipeline is created against does not use a
     *    depth/stencil attachment.
+    *
+    * We disable both depth and stenil tests in those cases.
     */
+   static const VkPipelineDepthStencilStateCreateInfo dummy_ds_info;
    const VkPipelineDepthStencilStateCreateInfo *ds_info =
-      builder->create_info->pDepthStencilState;
-   uint32_t rb_depth_cntl = 0, rb_stencil_cntl = 0;
-   struct tu_cs cs;
-
-   if (!builder->attachment_state_valid ||
-       (builder->depth_attachment_format != VK_FORMAT_UNDEFINED &&
-        builder->depth_attachment_format != VK_FORMAT_S8_UINT)) {
-      if (ds_info->depthTestEnable) {
-         rb_depth_cntl |=
-            A6XX_RB_DEPTH_CNTL_Z_TEST_ENABLE |
-            A6XX_RB_DEPTH_CNTL_ZFUNC(tu6_compare_func(ds_info->depthCompareOp)) |
-            A6XX_RB_DEPTH_CNTL_Z_READ_ENABLE; /* TODO: don't set for ALWAYS/NEVER */
-
-         if (ds_info->depthWriteEnable)
-            rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_WRITE_ENABLE;
-      }
-
-      if (ds_info->depthBoundsTestEnable)
-         rb_depth_cntl |= A6XX_RB_DEPTH_CNTL_Z_BOUNDS_ENABLE | A6XX_RB_DEPTH_CNTL_Z_READ_ENABLE;
+      builder->use_depth_stencil_attachment
+         ? builder->create_info->pDepthStencilState
+         : &dummy_ds_info;
 
-      if (ds_info->depthBoundsTestEnable && !ds_info->depthTestEnable)
-         tu6_apply_depth_bounds_workaround(builder->device, &rb_depth_cntl);
-   }
-
-   if (!builder->attachment_state_valid ||
-       builder->depth_attachment_format != VK_FORMAT_UNDEFINED) {
-      const VkStencilOpState *front = &ds_info->front;
-      const VkStencilOpState *back = &ds_info->back;
-
-      rb_stencil_cntl |=
-         A6XX_RB_STENCIL_CONTROL_FUNC(tu6_compare_func(front->compareOp)) |
-         A6XX_RB_STENCIL_CONTROL_FAIL(tu6_stencil_op(front->failOp)) |
-         A6XX_RB_STENCIL_CONTROL_ZPASS(tu6_stencil_op(front->passOp)) |
-         A6XX_RB_STENCIL_CONTROL_ZFAIL(tu6_stencil_op(front->depthFailOp)) |
-         A6XX_RB_STENCIL_CONTROL_FUNC_BF(tu6_compare_func(back->compareOp)) |
-         A6XX_RB_STENCIL_CONTROL_FAIL_BF(tu6_stencil_op(back->failOp)) |
-         A6XX_RB_STENCIL_CONTROL_ZPASS_BF(tu6_stencil_op(back->passOp)) |
-         A6XX_RB_STENCIL_CONTROL_ZFAIL_BF(tu6_stencil_op(back->depthFailOp));
-
-      if (ds_info->stencilTestEnable) {
-         rb_stencil_cntl |=
-            A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE |
-            A6XX_RB_STENCIL_CONTROL_STENCIL_ENABLE_BF |
-            A6XX_RB_STENCIL_CONTROL_STENCIL_READ;
-      }
-
-      pipeline->ds.raster_order_attachment_access =
-         ds_info->flags &
-         (VK_PIPELINE_DEPTH_STENCIL_STATE_CREATE_RASTERIZATION_ORDER_ATTACHMENT_DEPTH_ACCESS_BIT_ARM |
-          VK_PIPELINE_DEPTH_STENCIL_STATE_CREATE_RASTERIZATION_ORDER_ATTACHMENT_STENCIL_ACCESS_BIT_ARM);
-
-      pipeline->ds.write_enable = 
-         ds_info->depthWriteEnable || ds_info->stencilTestEnable;
-   }
-
-   pipeline->ds.rb_depth_cntl = rb_depth_cntl;
-
-   if (tu_pipeline_static_state(pipeline, &cs, TU_DYNAMIC_STATE_RB_STENCIL_CNTL, 2)) {
-      tu_cs_emit_pkt4(&cs, REG_A6XX_RB_STENCIL_CONTROL, 1);
-      tu_cs_emit(&cs, rb_stencil_cntl);
-   }
-   pipeline->ds.rb_stencil_cntl = rb_stencil_cntl;
+   struct tu_cs ds_cs;
+   tu_cs_begin_sub_stream(builder->device, &pipeline->cs, 12, &ds_cs);
 
-   /* the remaining draw states arent used if there is no d/s, leave them empty */
-   if (builder->depth_attachment_format == VK_FORMAT_UNDEFINED &&
-       builder->attachment_state_valid)
-      return;
+   /* move to hw ctx init? */
+   tu6_emit_alpha_control_disable(&ds_cs);
 
-   if (tu_pipeline_static_state(pipeline, &cs, VK_DYNAMIC_STATE_DEPTH_BOUNDS, 3)) {
-      tu_cs_emit_regs(&cs,
-                      A6XX_RB_Z_BOUNDS_MIN(ds_info->minDepthBounds),
-                      A6XX_RB_Z_BOUNDS_MAX(ds_info->maxDepthBounds));
-   }
+   tu6_emit_depth_control(&ds_cs, ds_info);
+   tu6_emit_stencil_control(&ds_cs, ds_info);
 
-   if (tu_pipeline_static_state(pipeline, &cs, VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK, 2)) {
-      tu_cs_emit_regs(&cs, A6XX_RB_STENCILMASK(.mask = ds_info->front.compareMask & 0xff,
-                                               .bfmask = ds_info->back.compareMask & 0xff));
+   if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_COMPARE_MASK)) {
+      tu6_emit_stencil_compare_mask(&ds_cs, ds_info->front.compareMask,
+                                    ds_info->back.compareMask);
    }
-
-   if (tu_pipeline_static_state(pipeline, &cs, VK_DYNAMIC_STATE_STENCIL_WRITE_MASK, 2)) {
-      update_stencil_mask(&pipeline->ds.stencil_wrmask,  VK_STENCIL_FACE_FRONT_BIT, ds_info->front.writeMask);
-      update_stencil_mask(&pipeline->ds.stencil_wrmask,  VK_STENCIL_FACE_BACK_BIT, ds_info->back.writeMask);
-      tu_cs_emit_regs(&cs, A6XX_RB_STENCILWRMASK(.dword = pipeline->ds.stencil_wrmask));
+   if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_WRITE_MASK)) {
+      tu6_emit_stencil_write_mask(&ds_cs, ds_info->front.writeMask,
+                                  ds_info->back.writeMask);
    }
-
-   if (tu_pipeline_static_state(pipeline, &cs, VK_DYNAMIC_STATE_STENCIL_REFERENCE, 2)) {
-      tu_cs_emit_regs(&cs, A6XX_RB_STENCILREF(.ref = ds_info->front.reference & 0xff,
-                                              .bfref = ds_info->back.reference & 0xff));
+   if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_REFERENCE)) {
+      tu6_emit_stencil_reference(&ds_cs, ds_info->front.reference,
+                                 ds_info->back.reference);
    }
 
-   if (builder->variants[MESA_SHADER_FRAGMENT]) {
-      const struct ir3_shader_variant *fs = builder->variants[MESA_SHADER_FRAGMENT];
-      if (fs->has_kill) {
-         pipeline->lrz.force_disable_mask |= TU_LRZ_FORCE_DISABLE_WRITE;
-      }
-      if (fs->no_earlyz || fs->writes_pos) {
-         pipeline->lrz.force_disable_mask = TU_LRZ_FORCE_DISABLE_LRZ;
-      }
-   }
-}
-
-static void
-tu_pipeline_builder_parse_rast_ds(struct tu_pipeline_builder *builder,
-                                  struct tu_pipeline *pipeline)
-{
-   if (builder->rasterizer_discard)
-      return;
-
-   pipeline->rast_ds.rb_depth_cntl =
-      pipeline->rast.rb_depth_cntl | pipeline->ds.rb_depth_cntl;
-   pipeline->rast_ds.rb_depth_cntl_mask = pipeline->ds.rb_depth_cntl_mask;
-
-   struct tu_cs cs;
-   if (tu_pipeline_static_state(pipeline, &cs, TU_DYNAMIC_STATE_RB_DEPTH_CNTL, 2)) {
-      tu_cs_emit_pkt4(&cs, REG_A6XX_RB_DEPTH_CNTL, 1);
-      if (pipeline->output.rb_depth_cntl_disable)
-         tu_cs_emit(&cs, 0);
-      else
-         tu_cs_emit(&cs, pipeline->rast_ds.rb_depth_cntl);
-   }
+   pipeline->ds.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &ds_cs);
 }
 
 static void
@@ -4351,189 +1681,31 @@ tu_pipeline_builder_parse_multisample_and_color_blend(
     *
     * We leave the relevant registers stale when rasterization is disabled.
     */
-   if (builder->rasterizer_discard) {
-      pipeline->output.samples = VK_SAMPLE_COUNT_1_BIT;
+   if (builder->rasterizer_discard)
       return;
-   }
-
-   pipeline->output.feedback_loop_may_involve_textures =
-      builder->feedback_loop_may_involve_textures;
 
    static const VkPipelineColorBlendStateCreateInfo dummy_blend_info;
    const VkPipelineMultisampleStateCreateInfo *msaa_info =
       builder->create_info->pMultisampleState;
-   pipeline->output.samples = msaa_info->rasterizationSamples;
-
    const VkPipelineColorBlendStateCreateInfo *blend_info =
       builder->use_color_attachments ? builder->create_info->pColorBlendState
                                      : &dummy_blend_info;
 
-   bool no_earlyz = builder->depth_attachment_format == VK_FORMAT_S8_UINT ||
-      /* alpha to coverage can behave like a discard */
-      msaa_info->alphaToCoverageEnable;
-   pipeline->lrz.force_late_z |= no_earlyz;
-
-   pipeline->output.subpass_feedback_loop_color =
-      builder->subpass_feedback_loop_color;
-   pipeline->output.subpass_feedback_loop_ds =
-      builder->subpass_feedback_loop_ds;
-
-   if (builder->use_color_attachments) {
-      pipeline->blend.raster_order_attachment_access =
-         blend_info->flags &
-         VK_PIPELINE_COLOR_BLEND_STATE_CREATE_RASTERIZATION_ORDER_ATTACHMENT_ACCESS_BIT_ARM;
-   }
-
-   const enum pipe_format ds_pipe_format =
-      vk_format_to_pipe_format(builder->depth_attachment_format);
+   struct tu_cs blend_cs;
+   tu_cs_begin_sub_stream(builder->device, &pipeline->cs, MAX_RTS * 3 + 9,
+                          &blend_cs);
 
-   if (builder->depth_attachment_format != VK_FORMAT_UNDEFINED &&
-       builder->depth_attachment_format != VK_FORMAT_S8_UINT) {
-      pipeline->output.depth_cpp_per_sample = util_format_get_component_bits(
-            ds_pipe_format, UTIL_FORMAT_COLORSPACE_ZS, 0) / 8;
-   } else {
-      /* We need to make sure RB_DEPTH_CNTL is set to 0 when this pipeline is
-       * used, regardless of whether it's linked with a fragment shader
-       * pipeline that has an enabled depth test or if RB_DEPTH_CNTL is set
-       * dynamically.
-       */
-      pipeline->output.rb_depth_cntl_disable = true;
-   }
-
-   if (builder->depth_attachment_format != VK_FORMAT_UNDEFINED) {
-      pipeline->output.stencil_cpp_per_sample = util_format_get_component_bits(
-            ds_pipe_format, UTIL_FORMAT_COLORSPACE_ZS, 1) / 8;
-   }
-
-   struct tu_cs cs;
-   tu6_emit_rb_mrt_controls(pipeline, blend_info,
+   uint32_t blend_enable_mask;
+   tu6_emit_rb_mrt_controls(&blend_cs, blend_info,
                             builder->color_attachment_formats,
-                            &pipeline->blend.rop_reads_dst,
-                            &pipeline->output.color_bandwidth_per_sample);
-
-   if (msaa_info->alphaToCoverageEnable && pipeline->blend.num_rts == 0) {
-      /* In addition to changing the *_OUTPUT_CNTL1 registers, this will also
-       * make sure we disable memory writes for MRT0 rather than using
-       * whatever setting was leftover.
-       */
-      pipeline->blend.num_rts = 1;
-   }
-
-   uint32_t blend_enable_mask =
-      pipeline->blend.rop_reads_dst ? 
-      pipeline->blend.color_write_enable :
-      pipeline->blend.blend_enable;
-   tu6_emit_blend_control(pipeline, blend_enable_mask,
-                          tu_blend_state_is_dual_src(blend_info), msaa_info);
-
-   if (tu_pipeline_static_state(pipeline, &cs, TU_DYNAMIC_STATE_BLEND,
-                                pipeline->blend.num_rts * 3 + 8)) {
-      tu6_emit_blend(&cs, pipeline);
-      assert(cs.cur == cs.end); /* validate draw state size */
-   }
-
-   /* Disable LRZ writes when blend or logic op that reads the destination is
-    * enabled, since the resulting pixel value from the blend-draw depends on
-    * an earlier draw, which LRZ in the draw pass could early-reject if the
-    * previous blend-enabled draw wrote LRZ.
-    *
-    * TODO: We need to disable LRZ writes only for the binning pass.
-    * Therefore, we need to emit it in a separate draw state. We keep
-    * it disabled for sysmem path as well for the moment.
-    */
-   if (blend_enable_mask)
-      pipeline->lrz.force_disable_mask |= TU_LRZ_FORCE_DISABLE_WRITE;
-
-   for (int i = 0; i < blend_info->attachmentCount; i++) {
-      VkPipelineColorBlendAttachmentState blendAttachment = blend_info->pAttachments[i];
-      /* From the PoV of LRZ, having masked color channels is
-       * the same as having blend enabled, in that the draw will
-       * care about the fragments from an earlier draw.
-       */
-      VkFormat format = builder->color_attachment_formats[i];
-      unsigned mask = MASK(vk_format_get_nr_components(format));
-      if (format != VK_FORMAT_UNDEFINED &&
-          ((blendAttachment.colorWriteMask & mask) != mask ||
-           !(pipeline->blend.color_write_enable & BIT(i)))) {
-         pipeline->lrz.force_disable_mask |= TU_LRZ_FORCE_DISABLE_WRITE;
-      }
-   }
-
-   if (tu_pipeline_static_state(pipeline, &cs, VK_DYNAMIC_STATE_BLEND_CONSTANTS, 5)) {
-      tu_cs_emit_pkt4(&cs, REG_A6XX_RB_BLEND_RED_F32, 4);
-      tu_cs_emit_array(&cs, (const uint32_t *) blend_info->blendConstants, 4);
-   }
+                            &blend_enable_mask);
 
-   const struct VkPipelineSampleLocationsStateCreateInfoEXT *sample_locations =
-      vk_find_struct_const(msaa_info->pNext, PIPELINE_SAMPLE_LOCATIONS_STATE_CREATE_INFO_EXT);
-   const VkSampleLocationsInfoEXT *samp_loc = NULL;
+   if (!(pipeline->dynamic_state.mask & TU_DYNAMIC_BLEND_CONSTANTS))
+      tu6_emit_blend_constants(&blend_cs, blend_info->blendConstants);
 
-   if (sample_locations && sample_locations->sampleLocationsEnable)
-      samp_loc = &sample_locations->sampleLocationsInfo;
+   tu6_emit_blend_control(&blend_cs, blend_enable_mask, msaa_info);
 
-    if (tu_pipeline_static_state(pipeline, &cs, TU_DYNAMIC_STATE_SAMPLE_LOCATIONS,
-                                 samp_loc ? 9 : 6)) {
-      tu6_emit_sample_locations(&cs, samp_loc);
-    }
-}
-
-static void
-tu_pipeline_builder_parse_rasterization_order(
-   struct tu_pipeline_builder *builder, struct tu_pipeline *pipeline)
-{
-   if (builder->rasterizer_discard)
-      return;
-
-   bool raster_order_attachment_access =
-      pipeline->blend.raster_order_attachment_access ||
-      pipeline->ds.raster_order_attachment_access ||
-      unlikely(builder->device->physical_device->instance->debug_flags & TU_DEBUG_RAST_ORDER);
-
-   /* VK_EXT_blend_operation_advanced would also require ordered access
-    * when implemented in the future.
-    */
-
-   uint32_t sysmem_prim_mode = NO_FLUSH;
-   uint32_t gmem_prim_mode = NO_FLUSH;
-
-   if (raster_order_attachment_access) {
-      /* VK_EXT_rasterization_order_attachment_access:
-       *
-       * This extension allow access to framebuffer attachments when used as
-       * both input and color attachments from one fragment to the next,
-       * in rasterization order, without explicit synchronization.
-       */
-      sysmem_prim_mode = FLUSH_PER_OVERLAP_AND_OVERWRITE;
-      gmem_prim_mode = FLUSH_PER_OVERLAP;
-      pipeline->prim_order.sysmem_single_prim_mode = true;
-   } else {
-      /* If there is a feedback loop, then the shader can read the previous value
-       * of a pixel being written out. It can also write some components and then
-       * read different components without a barrier in between. This is a
-       * problem in sysmem mode with UBWC, because the main buffer and flags
-       * buffer can get out-of-sync if only one is flushed. We fix this by
-       * setting the SINGLE_PRIM_MODE field to the same value that the blob does
-       * for advanced_blend in sysmem mode if a feedback loop is detected.
-       */
-      if (pipeline->output.subpass_feedback_loop_color ||
-          (pipeline->output.subpass_feedback_loop_ds &&
-           pipeline->ds.write_enable)) {
-         sysmem_prim_mode = FLUSH_PER_OVERLAP_AND_OVERWRITE;
-         pipeline->prim_order.sysmem_single_prim_mode = true;
-      }
-   }
-
-   struct tu_cs cs;
-
-   pipeline->prim_order.state_gmem = tu_cs_draw_state(&pipeline->cs, &cs, 2);
-   tu_cs_emit_write_reg(&cs, REG_A6XX_GRAS_SC_CNTL,
-                        A6XX_GRAS_SC_CNTL_CCUSINGLECACHELINESIZE(2) |
-                        A6XX_GRAS_SC_CNTL_SINGLE_PRIM_MODE(gmem_prim_mode));
-
-   pipeline->prim_order.state_sysmem = tu_cs_draw_state(&pipeline->cs, &cs, 2);
-   tu_cs_emit_write_reg(&cs, REG_A6XX_GRAS_SC_CNTL,
-                        A6XX_GRAS_SC_CNTL_CCUSINGLECACHELINESIZE(2) |
-                        A6XX_GRAS_SC_CNTL_SINGLE_PRIM_MODE(sysmem_prim_mode));
+   pipeline->blend.state_ib = tu_cs_end_sub_stream(&pipeline->cs, &blend_cs);
 }
 
 static void
@@ -4541,161 +1713,45 @@ tu_pipeline_finish(struct tu_pipeline *pipeline,
                    struct tu_device *dev,
                    const VkAllocationCallbacks *alloc)
 {
-   tu_cs_finish(&pipeline->cs);
-   pthread_mutex_lock(&dev->pipeline_mutex);
-   tu_suballoc_bo_free(&dev->pipeline_suballoc, &pipeline->bo);
-   pthread_mutex_unlock(&dev->pipeline_mutex);
-
-   if (pipeline->pvtmem_bo)
-      tu_bo_finish(dev, pipeline->pvtmem_bo);
-
-   if (pipeline->compiled_shaders)
-      vk_pipeline_cache_object_unref(&pipeline->compiled_shaders->base);
-
-   if (pipeline->nir_shaders)
-      vk_pipeline_cache_object_unref(&pipeline->nir_shaders->base);
+   tu_cs_finish(dev, &pipeline->cs);
 
-   for (unsigned i = 0; i < pipeline->num_sets; i++) {
-      if (pipeline->layouts[i])
-         vk_descriptor_set_layout_unref(&dev->vk, &pipeline->layouts[i]->vk);
-   }
-
-   ralloc_free(pipeline->executables_mem_ctx);
+   if (pipeline->program.binary_bo.gem_handle)
+      tu_bo_finish(dev, &pipeline->program.binary_bo);
 }
 
-
 static VkResult
 tu_pipeline_builder_build(struct tu_pipeline_builder *builder,
                           struct tu_pipeline **pipeline)
 {
-   VkResult result;
-
-   *pipeline = vk_object_zalloc(&builder->device->vk, builder->alloc,
-                                sizeof(**pipeline), VK_OBJECT_TYPE_PIPELINE);
-   if (!*pipeline)
-      return VK_ERROR_OUT_OF_HOST_MEMORY;
-
-   (*pipeline)->executables_mem_ctx = ralloc_context(NULL);
-   util_dynarray_init(&(*pipeline)->executables, (*pipeline)->executables_mem_ctx);
-
-   tu_pipeline_builder_parse_dynamic(builder, *pipeline);
-   tu_pipeline_builder_parse_libraries(builder, *pipeline);
-
-   VkShaderStageFlags stages = 0;
-   for (unsigned i = 0; i < builder->create_info->stageCount; i++) {
-      stages |= builder->create_info->pStages[i].stage;
-   }
-   builder->active_stages = stages;
-
-   (*pipeline)->active_stages = stages;
-   for (unsigned i = 0; i < builder->num_libraries; i++)
-      (*pipeline)->active_stages |= builder->libraries[i]->active_stages;
+   VkResult result = tu_pipeline_builder_create_pipeline(builder, pipeline);
+   if (result != VK_SUCCESS)
+      return result;
 
-   /* Compile and upload shaders unless a library has already done that. */
-   if ((*pipeline)->program.state.size == 0) {
-      tu_pipeline_builder_parse_layout(builder, *pipeline);
+   /* compile and upload shaders */
+   result = tu_pipeline_builder_compile_shaders(builder);
+   if (result == VK_SUCCESS)
+      result = tu_pipeline_builder_upload_shaders(builder, *pipeline);
+   if (result != VK_SUCCESS) {
+      tu_pipeline_finish(*pipeline, builder->device, builder->alloc);
+      vk_free2(&builder->device->alloc, builder->alloc, *pipeline);
+      *pipeline = VK_NULL_HANDLE;
 
-      result = tu_pipeline_builder_compile_shaders(builder, *pipeline);
-      if (result != VK_SUCCESS) {
-         vk_object_free(&builder->device->vk, builder->alloc, *pipeline);
-         return result;
-      }
+      return result;
    }
 
-   result = tu_pipeline_allocate_cs(builder->device, *pipeline,
-                                    &builder->layout, builder, NULL);
-
-
-   /* This has to come before emitting the program so that
-    * pipeline->tess.patch_control_points and pipeline->rast.multiview_mask
-    * are always set.
+   tu_pipeline_builder_parse_dynamic(builder, *pipeline);
+   tu_pipeline_builder_parse_shader_stages(builder, *pipeline);
+   tu_pipeline_builder_parse_vertex_input(builder, *pipeline);
+   tu_pipeline_builder_parse_input_assembly(builder, *pipeline);
+   tu_pipeline_builder_parse_viewport(builder, *pipeline);
+   tu_pipeline_builder_parse_rasterization(builder, *pipeline);
+   tu_pipeline_builder_parse_depth_stencil(builder, *pipeline);
+   tu_pipeline_builder_parse_multisample_and_color_blend(builder, *pipeline);
+
+   /* we should have reserved enough space upfront such that the CS never
+    * grows
     */
-   if (builder->state &
-       VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT) {
-      tu_pipeline_builder_parse_tessellation(builder, *pipeline);
-      (*pipeline)->rast.multiview_mask = builder->multiview_mask;
-   }
-
-   if (set_combined_state(builder, *pipeline,
-                          VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-                          VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT)) {
-      if (result != VK_SUCCESS) {
-         vk_object_free(&builder->device->vk, builder->alloc, *pipeline);
-         return result;
-      }
-
-      for (uint32_t i = 0; i < ARRAY_SIZE(builder->shader_iova); i++)
-         builder->shader_iova[i] =
-            tu_upload_variant(*pipeline, builder->variants[i]);
-
-      builder->binning_vs_iova =
-         tu_upload_variant(*pipeline, builder->binning_variant);
-
-      /* Setup private memory. Note that because we're sharing the same private
-       * memory for all stages, all stages must use the same config, or else
-       * fibers from one stage might overwrite fibers in another.
-       */
-
-      uint32_t pvtmem_size = 0;
-      bool per_wave = true;
-      for (uint32_t i = 0; i < ARRAY_SIZE(builder->variants); i++) {
-         if (builder->variants[i]) {
-            pvtmem_size = MAX2(pvtmem_size, builder->variants[i]->pvtmem_size);
-            if (!builder->variants[i]->pvtmem_per_wave)
-               per_wave = false;
-         }
-      }
-
-      if (builder->binning_variant) {
-         pvtmem_size = MAX2(pvtmem_size, builder->binning_variant->pvtmem_size);
-         if (!builder->binning_variant->pvtmem_per_wave)
-            per_wave = false;
-      }
-
-      result = tu_setup_pvtmem(builder->device, *pipeline, &builder->pvtmem,
-                               pvtmem_size, per_wave);
-      if (result != VK_SUCCESS) {
-         vk_object_free(&builder->device->vk, builder->alloc, *pipeline);
-         return result;
-      }
-
-      tu_pipeline_builder_parse_shader_stages(builder, *pipeline);
-      tu6_emit_load_state(*pipeline, &builder->layout);
-   }
-
-   if (builder->state &
-       VK_GRAPHICS_PIPELINE_LIBRARY_VERTEX_INPUT_INTERFACE_BIT_EXT) {
-      tu_pipeline_builder_parse_vertex_input(builder, *pipeline);
-      tu_pipeline_builder_parse_input_assembly(builder, *pipeline);
-   }
-
-   if (builder->state &
-       VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT) {
-      tu_pipeline_builder_parse_viewport(builder, *pipeline);
-      tu_pipeline_builder_parse_rasterization(builder, *pipeline);
-   }
-
-   if (builder->state & VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT) {
-      tu_pipeline_builder_parse_depth_stencil(builder, *pipeline);
-   }
-
-   if (builder->state &
-       VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT) {
-      tu_pipeline_builder_parse_multisample_and_color_blend(builder, *pipeline);
-   }
-
-   if (set_combined_state(builder, *pipeline,
-                          VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT |
-                          VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT)) {
-      tu_pipeline_builder_parse_rasterization_order(builder, *pipeline);
-   }
-
-   if (set_combined_state(builder, *pipeline,
-                          VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-                          VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT |
-                          VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT)) {
-      tu_pipeline_builder_parse_rast_ds(builder, *pipeline);
-   }
+   assert((*pipeline)->cs.bo_count == 1);
 
    return VK_SUCCESS;
 }
@@ -4703,231 +1759,59 @@ tu_pipeline_builder_build(struct tu_pipeline_builder *builder,
 static void
 tu_pipeline_builder_finish(struct tu_pipeline_builder *builder)
 {
-   if (builder->compiled_shaders)
-      vk_pipeline_cache_object_unref(&builder->compiled_shaders->base);
-   ralloc_free(builder->mem_ctx);
+   for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) {
+      if (!builder->shaders[i])
+         continue;
+      tu_shader_destroy(builder->device, builder->shaders[i], builder->alloc);
+   }
 }
 
 static void
 tu_pipeline_builder_init_graphics(
    struct tu_pipeline_builder *builder,
    struct tu_device *dev,
-   struct vk_pipeline_cache *cache,
+   struct tu_pipeline_cache *cache,
    const VkGraphicsPipelineCreateInfo *create_info,
    const VkAllocationCallbacks *alloc)
 {
    *builder = (struct tu_pipeline_builder) {
       .device = dev,
-      .mem_ctx = ralloc_context(NULL),
       .cache = cache,
       .create_info = create_info,
       .alloc = alloc,
    };
 
-   const VkGraphicsPipelineLibraryCreateInfoEXT *gpl_info =
-      vk_find_struct_const(builder->create_info->pNext, 
-                           GRAPHICS_PIPELINE_LIBRARY_CREATE_INFO_EXT);
-
-   const VkPipelineLibraryCreateInfoKHR *library_info =
-      vk_find_struct_const(builder->create_info->pNext, 
-                           PIPELINE_LIBRARY_CREATE_INFO_KHR);
-
-   if (gpl_info) {
-      builder->state = gpl_info->flags;
-   } else {
-      /* Implement this bit of spec text:
-       *
-       *    If this structure is omitted, and either
-       *    VkGraphicsPipelineCreateInfo::flags includes
-       *    VK_PIPELINE_CREATE_LIBRARY_BIT_KHR or the
-       *    VkGraphicsPipelineCreateInfo::pNext chain includes a
-       *    VkPipelineLibraryCreateInfoKHR structure with a libraryCount
-       *    greater than 0, it is as if flags is 0. Otherwise if this
-       *    structure is omitted, it is as if flags includes all possible
-       *    subsets of the graphics pipeline (i.e. a complete graphics
-       *    pipeline).
-       */
-      if ((library_info && library_info->libraryCount > 0) ||
-          (builder->create_info->flags & VK_PIPELINE_CREATE_LIBRARY_BIT_KHR)) {
-         builder->state = 0;
-      } else {
-         builder->state =
-            VK_GRAPHICS_PIPELINE_LIBRARY_VERTEX_INPUT_INTERFACE_BIT_EXT |
-            VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-            VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT |
-            VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT;
-      }
-   }
-
-   bool rasterizer_discard_dynamic = false;
-   if (create_info->pDynamicState) {
-      for (uint32_t i = 0; i < create_info->pDynamicState->dynamicStateCount; i++) {
-         if (create_info->pDynamicState->pDynamicStates[i] ==
-               VK_DYNAMIC_STATE_RASTERIZER_DISCARD_ENABLE) {
-            rasterizer_discard_dynamic = true;
-            break;
-         }
-      }
-   }
-
    builder->rasterizer_discard =
-      (builder->state & VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT) &&
-      builder->create_info->pRasterizationState->rasterizerDiscardEnable &&
-      !rasterizer_discard_dynamic;
-
-   if (builder->state &
-       (VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-        VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT |
-        VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT)) {
-      const VkPipelineRenderingCreateInfo *rendering_info =
-         vk_find_struct_const(create_info->pNext, PIPELINE_RENDERING_CREATE_INFO);
-
-      if (unlikely(dev->instance->debug_flags & TU_DEBUG_DYNAMIC) && !rendering_info)
-         rendering_info = vk_get_pipeline_rendering_create_info(create_info);
-
-      /* Get multiview_mask, which is only used for shaders */
-      if (builder->state &
-          (VK_GRAPHICS_PIPELINE_LIBRARY_PRE_RASTERIZATION_SHADERS_BIT_EXT |
-           VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT)) {
-         if (rendering_info) {
-            builder->multiview_mask = rendering_info->viewMask;
-         } else {
-            const struct tu_render_pass *pass =
-               tu_render_pass_from_handle(create_info->renderPass);
-            const struct tu_subpass *subpass =
-               &pass->subpasses[create_info->subpass];
-            builder->multiview_mask = subpass->multiview_mask;
-         }
-      }
-
-      /* Get the attachment state. This is valid:
-       *
-       * - With classic renderpasses, when either fragment shader or fragment
-       *   output interface state is being compiled. This includes when we
-       *   emulate classic renderpasses with dynamic rendering with the debug
-       *   flag.
-       * - With dynamic rendering (renderPass is NULL) only when compiling the
-       *   output interface state.
-       *
-       * We only actually need this for the fragment output interface state,
-       * but the spec also requires us to skip parsing depth/stencil state
-       * when the attachment state is defined *and* no depth/stencil
-       * attachment is not used, so we have to parse it for fragment shader
-       * state when possible. Life is pain.
-       */
-      if (((builder->state &
-            VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT) ||
-           ((builder->state &
-            VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT) &&
-            builder->create_info->renderPass)) &&
-          rendering_info) {
-         builder->subpass_raster_order_attachment_access = false;
-         builder->subpass_feedback_loop_ds = false;
-         builder->subpass_feedback_loop_color = false;
-
-         const VkRenderingSelfDependencyInfoMESA *self_dependency =
-            vk_find_struct_const(rendering_info->pNext, RENDERING_SELF_DEPENDENCY_INFO_MESA);
-
-         if (self_dependency) {
-            builder->subpass_feedback_loop_ds =
-               self_dependency->depthSelfDependency ||
-               self_dependency->stencilSelfDependency;
-            builder->subpass_feedback_loop_color =
-               self_dependency->colorSelfDependencies;
-         }
-
-         if (!builder->rasterizer_discard) {
-            builder->depth_attachment_format =
-               rendering_info->depthAttachmentFormat == VK_FORMAT_UNDEFINED ?
-               rendering_info->stencilAttachmentFormat :
-               rendering_info->depthAttachmentFormat;
-
-            for (unsigned i = 0; i < rendering_info->colorAttachmentCount; i++) {
-               builder->color_attachment_formats[i] =
-                  rendering_info->pColorAttachmentFormats[i];
-               if (builder->color_attachment_formats[i] != VK_FORMAT_UNDEFINED) {
-                  builder->use_color_attachments = true;
-               }
-            }
-         }
+      create_info->pRasterizationState->rasterizerDiscardEnable;
 
-         builder->attachment_state_valid = true;
-      } else if ((builder->state &
-                  (VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_SHADER_BIT_EXT |
-                   VK_GRAPHICS_PIPELINE_LIBRARY_FRAGMENT_OUTPUT_INTERFACE_BIT_EXT)) &&
-                 create_info->renderPass != VK_NULL_HANDLE) {
-         const struct tu_render_pass *pass =
-            tu_render_pass_from_handle(create_info->renderPass);
-         const struct tu_subpass *subpass =
-            &pass->subpasses[create_info->subpass];
-
-         builder->subpass_raster_order_attachment_access =
-            subpass->raster_order_attachment_access;
-         builder->subpass_feedback_loop_color = subpass->feedback_loop_color;
-         builder->subpass_feedback_loop_ds = subpass->feedback_loop_ds;
-
-         if (!builder->rasterizer_discard) {
-            const uint32_t a = subpass->depth_stencil_attachment.attachment;
-            builder->depth_attachment_format = (a != VK_ATTACHMENT_UNUSED) ?
-               pass->attachments[a].format : VK_FORMAT_UNDEFINED;
-
-            assert(subpass->color_count == 0 ||
-                   !create_info->pColorBlendState ||
-                   subpass->color_count == create_info->pColorBlendState->attachmentCount);
-            for (uint32_t i = 0; i < subpass->color_count; i++) {
-               const uint32_t a = subpass->color_attachments[i].attachment;
-               if (a == VK_ATTACHMENT_UNUSED)
-                  continue;
-
-               builder->color_attachment_formats[i] = pass->attachments[a].format;
-               builder->use_color_attachments = true;
-            }
-         }
+   if (builder->rasterizer_discard) {
+      builder->samples = VK_SAMPLE_COUNT_1_BIT;
+   } else {
+      builder->samples = create_info->pMultisampleState->rasterizationSamples;
+
+      const struct tu_render_pass *pass =
+         tu_render_pass_from_handle(create_info->renderPass);
+      const struct tu_subpass *subpass =
+         &pass->subpasses[create_info->subpass];
+
+      builder->use_depth_stencil_attachment =
+         subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED;
+
+      assert(subpass->color_count ==
+             create_info->pColorBlendState->attachmentCount);
+      builder->color_attachment_count = subpass->color_count;
+      for (uint32_t i = 0; i < subpass->color_count; i++) {
+         const uint32_t a = subpass->color_attachments[i].attachment;
+         if (a == VK_ATTACHMENT_UNUSED)
+            continue;
 
-         builder->attachment_state_valid = true;
+         builder->color_attachment_formats[i] = pass->attachments[a].format;
+         builder->use_color_attachments = true;
       }
    }
-
-   if (builder->create_info->flags & VK_PIPELINE_CREATE_COLOR_ATTACHMENT_FEEDBACK_LOOP_BIT_EXT) {
-      builder->subpass_feedback_loop_color = true;
-      builder->feedback_loop_may_involve_textures = true;
-   }
-
-   if (builder->create_info->flags & VK_PIPELINE_CREATE_DEPTH_STENCIL_ATTACHMENT_FEEDBACK_LOOP_BIT_EXT) {
-      builder->subpass_feedback_loop_ds = true;
-      builder->feedback_loop_may_involve_textures = true;
-   }
 }
 
-static VkResult
-tu_graphics_pipeline_create(VkDevice device,
-                            VkPipelineCache pipelineCache,
-                            const VkGraphicsPipelineCreateInfo *pCreateInfo,
-                            const VkAllocationCallbacks *pAllocator,
-                            VkPipeline *pPipeline)
-{
-   TU_FROM_HANDLE(tu_device, dev, device);
-   TU_FROM_HANDLE(vk_pipeline_cache, cache, pipelineCache);
-
-   cache = cache ? cache : dev->mem_cache;
-
-   struct tu_pipeline_builder builder;
-   tu_pipeline_builder_init_graphics(&builder, dev, cache,
-                                     pCreateInfo, pAllocator);
-
-   struct tu_pipeline *pipeline = NULL;
-   VkResult result = tu_pipeline_builder_build(&builder, &pipeline);
-   tu_pipeline_builder_finish(&builder);
-
-   if (result == VK_SUCCESS)
-      *pPipeline = tu_pipeline_to_handle(pipeline);
-   else
-      *pPipeline = VK_NULL_HANDLE;
-
-   return result;
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateGraphicsPipelines(VkDevice device,
                            VkPipelineCache pipelineCache,
                            uint32_t count,
@@ -4935,242 +1819,68 @@ tu_CreateGraphicsPipelines(VkDevice device,
                            const VkAllocationCallbacks *pAllocator,
                            VkPipeline *pPipelines)
 {
-   MESA_TRACE_FUNC();
-   VkResult final_result = VK_SUCCESS;
-   uint32_t i = 0;
+   TU_FROM_HANDLE(tu_device, dev, device);
+   TU_FROM_HANDLE(tu_pipeline_cache, cache, pipelineCache);
 
-   for (; i < count; i++) {
-      VkResult result = tu_graphics_pipeline_create(device, pipelineCache,
-                                                    &pCreateInfos[i], pAllocator,
-                                                    &pPipelines[i]);
+   for (uint32_t i = 0; i < count; i++) {
+      struct tu_pipeline_builder builder;
+      tu_pipeline_builder_init_graphics(&builder, dev, cache,
+                                        &pCreateInfos[i], pAllocator);
+
+      struct tu_pipeline *pipeline;
+      VkResult result = tu_pipeline_builder_build(&builder, &pipeline);
+      tu_pipeline_builder_finish(&builder);
 
       if (result != VK_SUCCESS) {
-         final_result = result;
-         pPipelines[i] = VK_NULL_HANDLE;
+         for (uint32_t j = 0; j < i; j++) {
+            tu_DestroyPipeline(device, pPipelines[j], pAllocator);
+            pPipelines[j] = VK_NULL_HANDLE;
+         }
 
-         if (pCreateInfos[i].flags &
-             VK_PIPELINE_CREATE_EARLY_RETURN_ON_FAILURE_BIT)
-            break;
+         return result;
       }
-   }
 
-   for (; i < count; i++)
-      pPipelines[i] = VK_NULL_HANDLE;
+      pPipelines[i] = tu_pipeline_to_handle(pipeline);
+   }
 
-   return final_result;
+   return VK_SUCCESS;
 }
 
 static VkResult
-tu_compute_pipeline_create(VkDevice device,
-                           VkPipelineCache pipelineCache,
+tu_compute_pipeline_create(VkDevice _device,
+                           VkPipelineCache _cache,
                            const VkComputePipelineCreateInfo *pCreateInfo,
                            const VkAllocationCallbacks *pAllocator,
                            VkPipeline *pPipeline)
 {
-   TU_FROM_HANDLE(tu_device, dev, device);
-   TU_FROM_HANDLE(vk_pipeline_cache, cache, pipelineCache);
-   TU_FROM_HANDLE(tu_pipeline_layout, layout, pCreateInfo->layout);
-   const VkPipelineShaderStageCreateInfo *stage_info = &pCreateInfo->stage;
-   VkResult result;
-
-   cache = cache ? cache : dev->mem_cache;
-
-   struct tu_pipeline *pipeline;
-
-   *pPipeline = VK_NULL_HANDLE;
-
-   VkPipelineCreationFeedback pipeline_feedback = {
-      .flags = VK_PIPELINE_CREATION_FEEDBACK_VALID_BIT,
-   };
-
-   const VkPipelineCreationFeedbackCreateInfo *creation_feedback =
-      vk_find_struct_const(pCreateInfo->pNext, PIPELINE_CREATION_FEEDBACK_CREATE_INFO);
-
-   int64_t pipeline_start = os_time_get_nano();
-
-   pipeline = vk_object_zalloc(&dev->vk, pAllocator, sizeof(*pipeline),
-                               VK_OBJECT_TYPE_PIPELINE);
-   if (!pipeline)
-      return VK_ERROR_OUT_OF_HOST_MEMORY;
-
-   pipeline->executables_mem_ctx = ralloc_context(NULL);
-   util_dynarray_init(&pipeline->executables, pipeline->executables_mem_ctx);
-   pipeline->active_stages = VK_SHADER_STAGE_COMPUTE_BIT;
-
-   struct tu_shader_key key = { };
-   tu_shader_key_init(&key, stage_info, dev);
-
-   void *pipeline_mem_ctx = ralloc_context(NULL);
-
-   unsigned char pipeline_sha1[20];
-   tu_hash_compute(pipeline_sha1, stage_info, layout, &key, dev->compiler);
-
-   struct tu_compiled_shaders *compiled = NULL;
-
-   const bool executable_info = pCreateInfo->flags &
-      VK_PIPELINE_CREATE_CAPTURE_INTERNAL_REPRESENTATIONS_BIT_KHR;
-
-   bool application_cache_hit = false;
-
-   if (!executable_info) {
-      compiled =
-         tu_pipeline_cache_lookup(cache, pipeline_sha1, sizeof(pipeline_sha1),
-                                  &application_cache_hit);
-   }
-
-   if (application_cache_hit && cache != dev->mem_cache) {
-      pipeline_feedback.flags |=
-         VK_PIPELINE_CREATION_FEEDBACK_APPLICATION_PIPELINE_CACHE_HIT_BIT;
-   }
-
-   if (tu6_shared_constants_enable(layout, dev->compiler)) {
-      pipeline->shared_consts = (struct tu_push_constant_range) {
-         .lo = 0,
-         .dwords = layout->push_constant_size / 4,
-      };
-   }
-
-   char *nir_initial_disasm = NULL;
-
-   if (!compiled) {
-      if (pCreateInfo->flags &
-          VK_PIPELINE_CREATE_FAIL_ON_PIPELINE_COMPILE_REQUIRED_BIT) {
-         result = VK_PIPELINE_COMPILE_REQUIRED;
-         goto fail;
-      }
-
-      struct ir3_shader_key ir3_key = {};
-
-      nir_shader *nir = tu_spirv_to_nir(dev, pipeline_mem_ctx, stage_info,
-                                        MESA_SHADER_COMPUTE);
-
-      nir_initial_disasm = executable_info ?
-         nir_shader_as_str(nir, pipeline->executables_mem_ctx) : NULL;
-
-      struct tu_shader *shader =
-         tu_shader_create(dev, nir, &key, layout, pAllocator);
-      if (!shader) {
-         result = VK_ERROR_OUT_OF_HOST_MEMORY;
-         goto fail;
-      }
-
-      compiled = tu_shaders_init(dev, &pipeline_sha1, sizeof(pipeline_sha1));
-      if (!compiled) {
-         tu_shader_destroy(dev, shader, pAllocator);
-         result = VK_ERROR_OUT_OF_HOST_MEMORY;
-         goto fail;
-      }
-
-      compiled->active_desc_sets = shader->active_desc_sets;
-      compiled->const_state[MESA_SHADER_COMPUTE] = shader->const_state;
-
-      struct ir3_shader_variant *v =
-         ir3_shader_create_variant(shader->ir3_shader, &ir3_key, executable_info);
-
-      tu_shader_destroy(dev, shader, pAllocator);
-
-      if (!v) {
-         result = VK_ERROR_OUT_OF_HOST_MEMORY;
-         goto fail;
-      }
-
-      compiled->variants[MESA_SHADER_COMPUTE] = v;
-
-      compiled = tu_pipeline_cache_insert(cache, compiled);
-   }
-
-   pipeline_feedback.duration = os_time_get_nano() - pipeline_start;
-
-   if (creation_feedback) {
-      *creation_feedback->pPipelineCreationFeedback = pipeline_feedback;
-      assert(creation_feedback->pipelineStageCreationFeedbackCount == 1);
-      creation_feedback->pPipelineStageCreationFeedbacks[0] = pipeline_feedback;
-   }
-
-   pipeline->active_desc_sets = compiled->active_desc_sets;
-
-   struct ir3_shader_variant *v = compiled->variants[MESA_SHADER_COMPUTE];
-
-   tu_pipeline_set_linkage(&pipeline->program.link[MESA_SHADER_COMPUTE],
-                           &compiled->const_state[MESA_SHADER_COMPUTE], v);
-
-   result = tu_pipeline_allocate_cs(dev, pipeline, layout, NULL, v);
-   if (result != VK_SUCCESS)
-      goto fail;
-
-   uint64_t shader_iova = tu_upload_variant(pipeline, v);
-
-   struct tu_pvtmem_config pvtmem;
-   tu_setup_pvtmem(dev, pipeline, &pvtmem, v->pvtmem_size, v->pvtmem_per_wave);
-
-   for (int i = 0; i < 3; i++)
-      pipeline->compute.local_size[i] = v->local_size[i];
-
-   pipeline->compute.subgroup_size = v->info.double_threadsize ? 128 : 64;
-
-   struct tu_cs prog_cs;
-   uint32_t additional_reserve_size = tu_xs_get_additional_cs_size_dwords(v);
-   tu_cs_begin_sub_stream(&pipeline->cs, 64 + additional_reserve_size, &prog_cs);
-   tu6_emit_cs_config(&prog_cs, v, &pvtmem, shader_iova);
-   pipeline->program.state = tu_cs_end_draw_state(&pipeline->cs, &prog_cs);
-
-   tu6_emit_load_state(pipeline, layout);
-
-   tu_append_executable(pipeline, v, nir_initial_disasm);
-
-   pipeline->program.cs_instrlen = v->instrlen;
-
-   vk_pipeline_cache_object_unref(&compiled->base);
-   ralloc_free(pipeline_mem_ctx);
-
-   *pPipeline = tu_pipeline_to_handle(pipeline);
-
    return VK_SUCCESS;
-
-fail:
-   if (compiled)
-      vk_pipeline_cache_object_unref(&compiled->base);
-
-   ralloc_free(pipeline_mem_ctx);
-
-   vk_object_free(&dev->vk, pAllocator, pipeline);
-
-   return result;
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_CreateComputePipelines(VkDevice device,
+VkResult
+tu_CreateComputePipelines(VkDevice _device,
                           VkPipelineCache pipelineCache,
                           uint32_t count,
                           const VkComputePipelineCreateInfo *pCreateInfos,
                           const VkAllocationCallbacks *pAllocator,
                           VkPipeline *pPipelines)
 {
-   MESA_TRACE_FUNC();
-   VkResult final_result = VK_SUCCESS;
-   uint32_t i = 0;
+   VkResult result = VK_SUCCESS;
 
+   unsigned i = 0;
    for (; i < count; i++) {
-      VkResult result = tu_compute_pipeline_create(device, pipelineCache,
-                                                   &pCreateInfos[i],
-                                                   pAllocator, &pPipelines[i]);
-      if (result != VK_SUCCESS) {
-         final_result = result;
+      VkResult r;
+      r = tu_compute_pipeline_create(_device, pipelineCache, &pCreateInfos[i],
+                                     pAllocator, &pPipelines[i]);
+      if (r != VK_SUCCESS) {
+         result = r;
          pPipelines[i] = VK_NULL_HANDLE;
-
-         if (pCreateInfos[i].flags &
-             VK_PIPELINE_CREATE_EARLY_RETURN_ON_FAILURE_BIT)
-            break;
       }
    }
 
-   for (; i < count; i++)
-      pPipelines[i] = VK_NULL_HANDLE;
-
-   return final_result;
+   return result;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyPipeline(VkDevice _device,
                    VkPipeline _pipeline,
                    const VkAllocationCallbacks *pAllocator)
@@ -5182,274 +1892,5 @@ tu_DestroyPipeline(VkDevice _device,
       return;
 
    tu_pipeline_finish(pipeline, dev, pAllocator);
-   vk_object_free(&dev->vk, pAllocator, pipeline);
-}
-
-#define WRITE_STR(field, ...) ({                                \
-   memset(field, 0, sizeof(field));                             \
-   UNUSED int _i = snprintf(field, sizeof(field), __VA_ARGS__); \
-   assert(_i > 0 && _i < sizeof(field));                        \
-})
-
-static const struct tu_pipeline_executable *
-tu_pipeline_get_executable(struct tu_pipeline *pipeline, uint32_t index)
-{
-   assert(index < util_dynarray_num_elements(&pipeline->executables,
-                                             struct tu_pipeline_executable));
-   return util_dynarray_element(
-      &pipeline->executables, struct tu_pipeline_executable, index);
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_GetPipelineExecutablePropertiesKHR(
-      VkDevice _device,
-      const VkPipelineInfoKHR* pPipelineInfo,
-      uint32_t* pExecutableCount,
-      VkPipelineExecutablePropertiesKHR* pProperties)
-{
-   TU_FROM_HANDLE(tu_device, dev, _device);
-   TU_FROM_HANDLE(tu_pipeline, pipeline, pPipelineInfo->pipeline);
-   VK_OUTARRAY_MAKE_TYPED(VkPipelineExecutablePropertiesKHR, out,
-                          pProperties, pExecutableCount);
-
-   util_dynarray_foreach (&pipeline->executables, struct tu_pipeline_executable, exe) {
-      vk_outarray_append_typed(VkPipelineExecutablePropertiesKHR, &out, props) {
-         gl_shader_stage stage = exe->stage;
-         props->stages = mesa_to_vk_shader_stage(stage);
-
-         if (!exe->is_binning)
-            WRITE_STR(props->name, "%s", _mesa_shader_stage_to_abbrev(stage));
-         else
-            WRITE_STR(props->name, "Binning VS");
-
-         WRITE_STR(props->description, "%s", _mesa_shader_stage_to_string(stage));
-
-         props->subgroupSize =
-            dev->compiler->threadsize_base * (exe->stats.double_threadsize ? 2 : 1);
-      }
-   }
-
-   return vk_outarray_status(&out);
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_GetPipelineExecutableStatisticsKHR(
-      VkDevice _device,
-      const VkPipelineExecutableInfoKHR* pExecutableInfo,
-      uint32_t* pStatisticCount,
-      VkPipelineExecutableStatisticKHR* pStatistics)
-{
-   TU_FROM_HANDLE(tu_pipeline, pipeline, pExecutableInfo->pipeline);
-   VK_OUTARRAY_MAKE_TYPED(VkPipelineExecutableStatisticKHR, out,
-                          pStatistics, pStatisticCount);
-
-   const struct tu_pipeline_executable *exe =
-      tu_pipeline_get_executable(pipeline, pExecutableInfo->executableIndex);
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Max Waves Per Core");
-      WRITE_STR(stat->description,
-                "Maximum number of simultaneous waves per core.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.max_waves;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Instruction Count");
-      WRITE_STR(stat->description,
-                "Total number of IR3 instructions in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.instrs_count;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Code size");
-      WRITE_STR(stat->description,
-                "Total number of dwords in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.sizedwords;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "NOPs Count");
-      WRITE_STR(stat->description,
-                "Number of NOP instructions in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.nops_count;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "MOV Count");
-      WRITE_STR(stat->description,
-                "Number of MOV instructions in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.mov_count;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "COV Count");
-      WRITE_STR(stat->description,
-                "Number of COV instructions in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.cov_count;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Registers used");
-      WRITE_STR(stat->description,
-                "Number of registers used in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.max_reg + 1;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Half-registers used");
-      WRITE_STR(stat->description,
-                "Number of half-registers used in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.max_half_reg + 1;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Instructions with SS sync bit");
-      WRITE_STR(stat->description,
-                "SS bit is set for instructions which depend on a result "
-                "of \"long\" instructions to prevent RAW hazard.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.ss;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Instructions with SY sync bit");
-      WRITE_STR(stat->description,
-                "SY bit is set for instructions which depend on a result "
-                "of loads from global memory to prevent RAW hazard.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.sy;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Estimated cycles stalled on SS");
-      WRITE_STR(stat->description,
-                "A better metric to estimate the impact of SS syncs.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.sstall;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "Estimated cycles stalled on SY");
-      WRITE_STR(stat->description,
-                "A better metric to estimate the impact of SY syncs.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.systall;
-   }
-
-   for (int i = 0; i < ARRAY_SIZE(exe->stats.instrs_per_cat); i++) {
-      vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-         WRITE_STR(stat->name, "cat%d instructions", i);
-         WRITE_STR(stat->description,
-                  "Number of cat%d instructions.", i);
-         stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-         stat->value.u64 = exe->stats.instrs_per_cat[i];
-      }
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "STP Count");
-      WRITE_STR(stat->description,
-                "Number of STore Private instructions in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.stp_count;
-   }
-
-   vk_outarray_append_typed(VkPipelineExecutableStatisticKHR, &out, stat) {
-      WRITE_STR(stat->name, "LDP Count");
-      WRITE_STR(stat->description,
-                "Number of LoaD Private instructions in the final generated "
-                "shader executable.");
-      stat->format = VK_PIPELINE_EXECUTABLE_STATISTIC_FORMAT_UINT64_KHR;
-      stat->value.u64 = exe->stats.ldp_count;
-   }
-
-   return vk_outarray_status(&out);
-}
-
-static bool
-write_ir_text(VkPipelineExecutableInternalRepresentationKHR* ir,
-              const char *data)
-{
-   ir->isText = VK_TRUE;
-
-   size_t data_len = strlen(data) + 1;
-
-   if (ir->pData == NULL) {
-      ir->dataSize = data_len;
-      return true;
-   }
-
-   strncpy(ir->pData, data, ir->dataSize);
-   if (ir->dataSize < data_len)
-      return false;
-
-   ir->dataSize = data_len;
-   return true;
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_GetPipelineExecutableInternalRepresentationsKHR(
-    VkDevice _device,
-    const VkPipelineExecutableInfoKHR* pExecutableInfo,
-    uint32_t* pInternalRepresentationCount,
-    VkPipelineExecutableInternalRepresentationKHR* pInternalRepresentations)
-{
-   TU_FROM_HANDLE(tu_pipeline, pipeline, pExecutableInfo->pipeline);
-   VK_OUTARRAY_MAKE_TYPED(VkPipelineExecutableInternalRepresentationKHR, out,
-                          pInternalRepresentations, pInternalRepresentationCount);
-   bool incomplete_text = false;
-
-   const struct tu_pipeline_executable *exe =
-      tu_pipeline_get_executable(pipeline, pExecutableInfo->executableIndex);
-
-   if (exe->nir_from_spirv) {
-      vk_outarray_append_typed(VkPipelineExecutableInternalRepresentationKHR, &out, ir) {
-         WRITE_STR(ir->name, "NIR from SPIRV");
-         WRITE_STR(ir->description,
-                   "Initial NIR before any optimizations");
-
-         if (!write_ir_text(ir, exe->nir_from_spirv))
-            incomplete_text = true;
-      }
-   }
-
-   if (exe->nir_final) {
-      vk_outarray_append_typed(VkPipelineExecutableInternalRepresentationKHR, &out, ir) {
-         WRITE_STR(ir->name, "Final NIR");
-         WRITE_STR(ir->description,
-                   "Final NIR before going into the back-end compiler");
-
-         if (!write_ir_text(ir, exe->nir_final))
-            incomplete_text = true;
-      }
-   }
-
-   if (exe->disasm) {
-      vk_outarray_append_typed(VkPipelineExecutableInternalRepresentationKHR, &out, ir) {
-         WRITE_STR(ir->name, "IR3 Assembly");
-         WRITE_STR(ir->description,
-                   "Final IR3 assembly for the generated shader binary");
-
-         if (!write_ir_text(ir, exe->disasm))
-            incomplete_text = true;
-      }
-   }
-
-   return incomplete_text ? VK_INCOMPLETE : vk_outarray_status(&out);
+   vk_free2(&dev->alloc, pAllocator, pipeline);
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_query.c b/lib/mesa/src/freedreno/vulkan/tu_query.c
index 6da5102cc..2cb710fb1 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_query.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_query.c
@@ -1,339 +1,57 @@
 /*
  * Copyrigh 2016 Red Hat Inc.
- * SPDX-License-Identifier: MIT
- *
  * Based on anv:
  * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_query.h"
+#include "tu_private.h"
 
+#include <assert.h>
 #include <fcntl.h>
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
 
 #include "nir/nir_builder.h"
-#include "util/os_time.h"
-
-#include "vk_util.h"
-
-#include "tu_cmd_buffer.h"
-#include "tu_cs.h"
-#include "tu_device.h"
-
-#define NSEC_PER_SEC 1000000000ull
-#define WAIT_TIMEOUT 5
-#define STAT_COUNT ((REG_A6XX_RBBM_PRIMCTR_10_LO - REG_A6XX_RBBM_PRIMCTR_0_LO) / 2 + 1)
-
-struct PACKED query_slot {
-   uint64_t available;
-};
-
-struct PACKED occlusion_slot_value {
-   /* Seems sample counters are placed to be 16-byte aligned
-    * even though this query needs an 8-byte slot. */
-   uint64_t value;
-   uint64_t _padding;
-};
-
-struct PACKED occlusion_query_slot {
-   struct query_slot common;
-   uint64_t result;
-
-   struct occlusion_slot_value begin;
-   struct occlusion_slot_value end;
-};
-
-struct PACKED timestamp_query_slot {
-   struct query_slot common;
-   uint64_t result;
-};
-
-struct PACKED primitive_slot_value {
-   uint64_t values[2];
-};
-
-struct PACKED pipeline_stat_query_slot {
-   struct query_slot common;
-   uint64_t results[STAT_COUNT];
-
-   uint64_t begin[STAT_COUNT];
-   uint64_t end[STAT_COUNT];
-};
-
-struct PACKED primitive_query_slot {
-   struct query_slot common;
-   /* The result of transform feedback queries is two integer values:
-    *   results[0] is the count of primitives written,
-    *   results[1] is the count of primitives generated.
-    * Also a result for each stream is stored at 4 slots respectively.
-    */
-   uint64_t results[2];
-
-   /* Primitive counters also need to be 16-byte aligned. */
-   uint64_t _padding;
-
-   struct primitive_slot_value begin[4];
-   struct primitive_slot_value end[4];
-};
-
-struct PACKED perfcntr_query_slot {
-   uint64_t result;
-   uint64_t begin;
-   uint64_t end;
-};
-
-struct PACKED perf_query_slot {
-   struct query_slot common;
-   struct perfcntr_query_slot perfcntr;
-};
-
-struct PACKED primitives_generated_query_slot {
-   struct query_slot common;
-   uint64_t result;
-   uint64_t begin;
-   uint64_t end;
-};
-
-/* Returns the IOVA of a given uint64_t field in a given slot of a query
- * pool. */
-#define query_iova(type, pool, query, field)                         \
-   pool->bo->iova + pool->stride * (query) + offsetof(type, field)
-
-#define occlusion_query_iova(pool, query, field)                     \
-   query_iova(struct occlusion_query_slot, pool, query, field)
-
-#define pipeline_stat_query_iova(pool, query, field)                 \
-   pool->bo->iova + pool->stride * (query) +                            \
-   offsetof(struct pipeline_stat_query_slot, field)
-
-#define primitive_query_iova(pool, query, field, i)                  \
-   query_iova(struct primitive_query_slot, pool, query, field) +     \
-   offsetof(struct primitive_slot_value, values[i])
-
-#define perf_query_iova(pool, query, field, i)                          \
-   pool->bo->iova + pool->stride * (query) +                             \
-   sizeof(struct query_slot) +                                   \
-   sizeof(struct perfcntr_query_slot) * (i) +                          \
-   offsetof(struct perfcntr_query_slot, field)
-
-#define primitives_generated_query_iova(pool, query, field)               \
-   query_iova(struct primitives_generated_query_slot, pool, query, field)
-
-#define query_available_iova(pool, query)                            \
-   query_iova(struct query_slot, pool, query, available)
-
-#define query_result_iova(pool, query, type, i)                            \
-   pool->bo->iova + pool->stride * (query) +                          \
-   sizeof(struct query_slot) + sizeof(type) * (i)
-
-#define query_result_addr(pool, query, type, i)                            \
-   pool->bo->map + pool->stride * (query) +                             \
-   sizeof(struct query_slot) + sizeof(type) * (i)
-
-#define query_is_available(slot) slot->available
-
-static const VkPerformanceCounterUnitKHR
-fd_perfcntr_type_to_vk_unit[] = {
-   [FD_PERFCNTR_TYPE_UINT]         = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-   [FD_PERFCNTR_TYPE_UINT64]       = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-   [FD_PERFCNTR_TYPE_FLOAT]        = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-   [FD_PERFCNTR_TYPE_PERCENTAGE]   = VK_PERFORMANCE_COUNTER_UNIT_PERCENTAGE_KHR,
-   [FD_PERFCNTR_TYPE_BYTES]        = VK_PERFORMANCE_COUNTER_UNIT_BYTES_KHR,
-   /* TODO. can be UNIT_NANOSECONDS_KHR with a logic to compute */
-   [FD_PERFCNTR_TYPE_MICROSECONDS] = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-   [FD_PERFCNTR_TYPE_HZ]           = VK_PERFORMANCE_COUNTER_UNIT_HERTZ_KHR,
-   [FD_PERFCNTR_TYPE_DBM]          = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-   [FD_PERFCNTR_TYPE_TEMPERATURE]  = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-   [FD_PERFCNTR_TYPE_VOLTS]        = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-   [FD_PERFCNTR_TYPE_AMPS]         = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-   [FD_PERFCNTR_TYPE_WATTS]        = VK_PERFORMANCE_COUNTER_UNIT_GENERIC_KHR,
-};
-
-/* TODO. Basically this comes from the freedreno implementation where
- * only UINT64 is used. We'd better confirm this by the blob vulkan driver
- * when it starts supporting perf query.
- */
-static const VkPerformanceCounterStorageKHR
-fd_perfcntr_type_to_vk_storage[] = {
-   [FD_PERFCNTR_TYPE_UINT]         = VK_PERFORMANCE_COUNTER_STORAGE_UINT32_KHR,
-   [FD_PERFCNTR_TYPE_UINT64]       = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
-   [FD_PERFCNTR_TYPE_FLOAT]        = VK_PERFORMANCE_COUNTER_STORAGE_FLOAT32_KHR,
-   [FD_PERFCNTR_TYPE_PERCENTAGE]   = VK_PERFORMANCE_COUNTER_STORAGE_FLOAT32_KHR,
-   [FD_PERFCNTR_TYPE_BYTES]        = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
-   [FD_PERFCNTR_TYPE_MICROSECONDS] = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
-   [FD_PERFCNTR_TYPE_HZ]           = VK_PERFORMANCE_COUNTER_STORAGE_UINT64_KHR,
-   [FD_PERFCNTR_TYPE_DBM]          = VK_PERFORMANCE_COUNTER_STORAGE_FLOAT32_KHR,
-   [FD_PERFCNTR_TYPE_TEMPERATURE]  = VK_PERFORMANCE_COUNTER_STORAGE_FLOAT32_KHR,
-   [FD_PERFCNTR_TYPE_VOLTS]        = VK_PERFORMANCE_COUNTER_STORAGE_FLOAT32_KHR,
-   [FD_PERFCNTR_TYPE_AMPS]         = VK_PERFORMANCE_COUNTER_STORAGE_FLOAT32_KHR,
-   [FD_PERFCNTR_TYPE_WATTS]        = VK_PERFORMANCE_COUNTER_STORAGE_FLOAT32_KHR,
-};
-
-/*
- * Returns a pointer to a given slot in a query pool.
- */
-static void* slot_address(struct tu_query_pool *pool, uint32_t query)
-{
-   return (char*)pool->bo->map + query * pool->stride;
-}
-
-static void
-perfcntr_index(const struct fd_perfcntr_group *group, uint32_t group_count,
-               uint32_t index, uint32_t *gid, uint32_t *cid)
-
-{
-   uint32_t i;
-
-   for (i = 0; i < group_count; i++) {
-      if (group[i].num_countables > index) {
-         *gid = i;
-         *cid = index;
-         break;
-      }
-      index -= group[i].num_countables;
-   }
-
-   assert(i < group_count);
-}
 
-static int
-compare_perfcntr_pass(const void *a, const void *b)
-{
-   return ((struct tu_perf_query_data *)a)->pass -
-          ((struct tu_perf_query_data *)b)->pass;
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_CreateQueryPool(VkDevice _device,
                    const VkQueryPoolCreateInfo *pCreateInfo,
                    const VkAllocationCallbacks *pAllocator,
                    VkQueryPool *pQueryPool)
 {
    TU_FROM_HANDLE(tu_device, device, _device);
-   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO);
-   assert(pCreateInfo->queryCount > 0);
-
-   uint32_t pool_size, slot_size;
-   const VkQueryPoolPerformanceCreateInfoKHR *perf_query_info = NULL;
-
-   pool_size = sizeof(struct tu_query_pool);
-
-   switch (pCreateInfo->queryType) {
-   case VK_QUERY_TYPE_OCCLUSION:
-      slot_size = sizeof(struct occlusion_query_slot);
-      break;
-   case VK_QUERY_TYPE_TIMESTAMP:
-      slot_size = sizeof(struct timestamp_query_slot);
-      break;
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-      slot_size = sizeof(struct primitive_query_slot);
-      break;
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-      slot_size = sizeof(struct primitives_generated_query_slot);
-      break;
-   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR: {
-      perf_query_info =
-            vk_find_struct_const(pCreateInfo->pNext,
-                                 QUERY_POOL_PERFORMANCE_CREATE_INFO_KHR);
-      assert(perf_query_info);
-
-      slot_size = sizeof(struct perf_query_slot) +
-                  sizeof(struct perfcntr_query_slot) *
-                  (perf_query_info->counterIndexCount - 1);
-
-      /* Size of the array pool->tu_perf_query_data */
-      pool_size += sizeof(struct tu_perf_query_data) *
-                   perf_query_info->counterIndexCount;
-      break;
-   }
-   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
-      slot_size = sizeof(struct pipeline_stat_query_slot);
-      break;
-   default:
-      unreachable("Invalid query type");
-   }
-
    struct tu_query_pool *pool =
-         vk_object_alloc(&device->vk, pAllocator, pool_size,
-                         VK_OBJECT_TYPE_QUERY_POOL);
-   if (!pool)
-      return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY);
-
-   if (pCreateInfo->queryType == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
-      pool->perf_group = fd_perfcntrs(&device->physical_device->dev_id,
-                                      &pool->perf_group_count);
-
-      pool->counter_index_count = perf_query_info->counterIndexCount;
-
-      /* Build all perf counters data that is requested, so we could get
-       * correct group id, countable id, counter register and pass index with
-       * only a counter index provided by applications at each command submit.
-       *
-       * Also, since this built data will be sorted by pass index later, we
-       * should keep the original indices and store perfcntrs results according
-       * to them so apps can get correct results with their own indices.
-       */
-      uint32_t regs[pool->perf_group_count], pass[pool->perf_group_count];
-      memset(regs, 0x00, pool->perf_group_count * sizeof(regs[0]));
-      memset(pass, 0x00, pool->perf_group_count * sizeof(pass[0]));
-
-      for (uint32_t i = 0; i < pool->counter_index_count; i++) {
-         uint32_t gid = 0, cid = 0;
-
-         perfcntr_index(pool->perf_group, pool->perf_group_count,
-                        perf_query_info->pCounterIndices[i], &gid, &cid);
-
-         pool->perf_query_data[i].gid = gid;
-         pool->perf_query_data[i].cid = cid;
-         pool->perf_query_data[i].app_idx = i;
-
-         /* When a counter register is over the capacity(num_counters),
-          * reset it for next pass.
-          */
-         if (regs[gid] < pool->perf_group[gid].num_counters) {
-            pool->perf_query_data[i].cntr_reg = regs[gid]++;
-            pool->perf_query_data[i].pass = pass[gid];
-         } else {
-            pool->perf_query_data[i].pass = ++pass[gid];
-            pool->perf_query_data[i].cntr_reg = regs[gid] = 0;
-            regs[gid]++;
-         }
-      }
+      vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
+                VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
 
-      /* Sort by pass index so we could easily prepare a command stream
-       * with the ascending order of pass index.
-       */
-      qsort(pool->perf_query_data, pool->counter_index_count,
-            sizeof(pool->perf_query_data[0]),
-            compare_perfcntr_pass);
-   }
-
-   VkResult result = tu_bo_init_new(device, &pool->bo,
-         pCreateInfo->queryCount * slot_size, TU_BO_ALLOC_NO_FLAGS, "query pool");
-   if (result != VK_SUCCESS) {
-      vk_object_free(&device->vk, pAllocator, pool);
-      return result;
-   }
-
-   result = tu_bo_map(device, pool->bo);
-   if (result != VK_SUCCESS) {
-      tu_bo_finish(device, pool->bo);
-      vk_object_free(&device->vk, pAllocator, pool);
-      return result;
-   }
-
-   /* Initialize all query statuses to unavailable */
-   memset(pool->bo->map, 0, pool->bo->size);
+   if (!pool)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   pool->type = pCreateInfo->queryType;
-   pool->stride = slot_size;
-   pool->size = pCreateInfo->queryCount;
-   pool->pipeline_statistics = pCreateInfo->pipelineStatistics;
    *pQueryPool = tu_query_pool_to_handle(pool);
-
    return VK_SUCCESS;
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_DestroyQueryPool(VkDevice _device,
                     VkQueryPool _pool,
                     const VkAllocationCallbacks *pAllocator)
@@ -344,211 +62,10 @@ tu_DestroyQueryPool(VkDevice _device,
    if (!pool)
       return;
 
-   tu_bo_finish(device, pool->bo);
-   vk_object_free(&device->vk, pAllocator, pool);
-}
-
-static uint32_t
-get_result_count(struct tu_query_pool *pool)
-{
-   switch (pool->type) {
-   /* Occulusion and timestamp queries write one integer value */
-   case VK_QUERY_TYPE_OCCLUSION:
-   case VK_QUERY_TYPE_TIMESTAMP:
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-      return 1;
-   /* Transform feedback queries write two integer values */
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-      return 2;
-   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
-      return util_bitcount(pool->pipeline_statistics);
-   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR:
-      return pool->counter_index_count;
-   default:
-      assert(!"Invalid query type");
-      return 0;
-   }
-}
-
-static uint32_t
-statistics_index(uint32_t *statistics)
-{
-   uint32_t stat;
-   stat = u_bit_scan(statistics);
-
-   switch (1 << stat) {
-   case VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_VERTICES_BIT:
-   case VK_QUERY_PIPELINE_STATISTIC_VERTEX_SHADER_INVOCATIONS_BIT:
-      return 0;
-   case VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_PRIMITIVES_BIT:
-      return 1;
-   case VK_QUERY_PIPELINE_STATISTIC_TESSELLATION_CONTROL_SHADER_PATCHES_BIT:
-      return 2;
-   case VK_QUERY_PIPELINE_STATISTIC_TESSELLATION_EVALUATION_SHADER_INVOCATIONS_BIT:
-      return 4;
-   case VK_QUERY_PIPELINE_STATISTIC_GEOMETRY_SHADER_INVOCATIONS_BIT:
-      return 5;
-   case VK_QUERY_PIPELINE_STATISTIC_GEOMETRY_SHADER_PRIMITIVES_BIT:
-      return 6;
-   case VK_QUERY_PIPELINE_STATISTIC_CLIPPING_INVOCATIONS_BIT:
-      return 7;
-   case VK_QUERY_PIPELINE_STATISTIC_CLIPPING_PRIMITIVES_BIT:
-      return 8;
-   case VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT:
-      return 9;
-   case VK_QUERY_PIPELINE_STATISTIC_COMPUTE_SHADER_INVOCATIONS_BIT:
-      return 10;
-   default:
-      return 0;
-   }
-}
-
-static bool
-is_pipeline_query_with_vertex_stage(uint32_t pipeline_statistics)
-{
-   return pipeline_statistics &
-          (VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_VERTICES_BIT |
-           VK_QUERY_PIPELINE_STATISTIC_INPUT_ASSEMBLY_PRIMITIVES_BIT |
-           VK_QUERY_PIPELINE_STATISTIC_VERTEX_SHADER_INVOCATIONS_BIT |
-           VK_QUERY_PIPELINE_STATISTIC_GEOMETRY_SHADER_INVOCATIONS_BIT |
-           VK_QUERY_PIPELINE_STATISTIC_GEOMETRY_SHADER_PRIMITIVES_BIT |
-           VK_QUERY_PIPELINE_STATISTIC_CLIPPING_INVOCATIONS_BIT |
-           VK_QUERY_PIPELINE_STATISTIC_CLIPPING_PRIMITIVES_BIT |
-           VK_QUERY_PIPELINE_STATISTIC_TESSELLATION_CONTROL_SHADER_PATCHES_BIT |
-           VK_QUERY_PIPELINE_STATISTIC_TESSELLATION_EVALUATION_SHADER_INVOCATIONS_BIT);
-}
-
-static bool
-is_pipeline_query_with_fragment_stage(uint32_t pipeline_statistics)
-{
-   return pipeline_statistics &
-          VK_QUERY_PIPELINE_STATISTIC_FRAGMENT_SHADER_INVOCATIONS_BIT;
-}
-
-static bool
-is_pipeline_query_with_compute_stage(uint32_t pipeline_statistics)
-{
-   return pipeline_statistics &
-          VK_QUERY_PIPELINE_STATISTIC_COMPUTE_SHADER_INVOCATIONS_BIT;
-}
-
-/* Wait on the the availability status of a query up until a timeout. */
-static VkResult
-wait_for_available(struct tu_device *device, struct tu_query_pool *pool,
-                   uint32_t query)
-{
-   /* TODO: Use the MSM_IOVA_WAIT ioctl to wait on the available bit in a
-    * scheduler friendly way instead of busy polling once the patch has landed
-    * upstream. */
-   struct query_slot *slot = slot_address(pool, query);
-   uint64_t abs_timeout = os_time_get_absolute_timeout(
-         WAIT_TIMEOUT * NSEC_PER_SEC);
-   while(os_time_get_nano() < abs_timeout) {
-      if (query_is_available(slot))
-         return VK_SUCCESS;
-   }
-   return vk_error(device, VK_TIMEOUT);
+   vk_free2(&device->alloc, pAllocator, pool);
 }
 
-/* Writes a query value to a buffer from the CPU. */
-static void
-write_query_value_cpu(char* base,
-                      uint32_t offset,
-                      uint64_t value,
-                      VkQueryResultFlags flags)
-{
-   if (flags & VK_QUERY_RESULT_64_BIT) {
-      *(uint64_t*)(base + (offset * sizeof(uint64_t))) = value;
-   } else {
-      *(uint32_t*)(base + (offset * sizeof(uint32_t))) = value;
-   }
-}
-
-static VkResult
-get_query_pool_results(struct tu_device *device,
-                       struct tu_query_pool *pool,
-                       uint32_t firstQuery,
-                       uint32_t queryCount,
-                       size_t dataSize,
-                       void *pData,
-                       VkDeviceSize stride,
-                       VkQueryResultFlags flags)
-{
-   assert(dataSize >= stride * queryCount);
-
-   char *result_base = pData;
-   VkResult result = VK_SUCCESS;
-   for (uint32_t i = 0; i < queryCount; i++) {
-      uint32_t query = firstQuery + i;
-      struct query_slot *slot = slot_address(pool, query);
-      bool available = query_is_available(slot);
-      uint32_t result_count = get_result_count(pool);
-      uint32_t statistics = pool->pipeline_statistics;
-
-      if ((flags & VK_QUERY_RESULT_WAIT_BIT) && !available) {
-         VkResult wait_result = wait_for_available(device, pool, query);
-         if (wait_result != VK_SUCCESS)
-            return wait_result;
-         available = true;
-      } else if (!(flags & VK_QUERY_RESULT_PARTIAL_BIT) && !available) {
-         /* From the Vulkan 1.1.130 spec:
-          *
-          *    If VK_QUERY_RESULT_WAIT_BIT and VK_QUERY_RESULT_PARTIAL_BIT are
-          *    both not set then no result values are written to pData for
-          *    queries that are in the unavailable state at the time of the
-          *    call, and vkGetQueryPoolResults returns VK_NOT_READY. However,
-          *    availability state is still written to pData for those queries
-          *    if VK_QUERY_RESULT_WITH_AVAILABILITY_BIT is set.
-          */
-         result = VK_NOT_READY;
-         if (!(flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)) {
-            result_base += stride;
-            continue;
-         }
-      }
-
-      for (uint32_t k = 0; k < result_count; k++) {
-         if (available) {
-            uint64_t *result;
-
-            if (pool->type == VK_QUERY_TYPE_PIPELINE_STATISTICS) {
-               uint32_t stat_idx = statistics_index(&statistics);
-               result = query_result_addr(pool, query, uint64_t, stat_idx);
-            } else if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
-               result = query_result_addr(pool, query, struct perfcntr_query_slot, k);
-            } else {
-               result = query_result_addr(pool, query, uint64_t, k);
-            }
-
-            write_query_value_cpu(result_base, k, *result, flags);
-         } else if (flags & VK_QUERY_RESULT_PARTIAL_BIT)
-             /* From the Vulkan 1.1.130 spec:
-              *
-              *   If VK_QUERY_RESULT_PARTIAL_BIT is set, VK_QUERY_RESULT_WAIT_BIT
-              *   is not set, and the query’s status is unavailable, an
-              *   intermediate result value between zero and the final result
-              *   value is written to pData for that query.
-              *
-              * Just return 0 here for simplicity since it's a valid result.
-              */
-            write_query_value_cpu(result_base, k, 0, flags);
-      }
-
-      if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT)
-         /* From the Vulkan 1.1.130 spec:
-          *
-          *    If VK_QUERY_RESULT_WITH_AVAILABILITY_BIT is set, the final
-          *    integer value written for each query is non-zero if the query’s
-          *    status was available or zero if the status was unavailable.
-          */
-         write_query_value_cpu(result_base, result_count, available, flags);
-
-      result_base += stride;
-   }
-   return result;
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
+VkResult
 tu_GetQueryPoolResults(VkDevice _device,
                        VkQueryPool queryPool,
                        uint32_t firstQuery,
@@ -558,140 +75,10 @@ tu_GetQueryPoolResults(VkDevice _device,
                        VkDeviceSize stride,
                        VkQueryResultFlags flags)
 {
-   TU_FROM_HANDLE(tu_device, device, _device);
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-   assert(firstQuery + queryCount <= pool->size);
-
-   if (vk_device_is_lost(&device->vk))
-      return VK_ERROR_DEVICE_LOST;
-
-   switch (pool->type) {
-   case VK_QUERY_TYPE_OCCLUSION:
-   case VK_QUERY_TYPE_TIMESTAMP:
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
-   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR:
-      return get_query_pool_results(device, pool, firstQuery, queryCount,
-                                    dataSize, pData, stride, flags);
-   default:
-      assert(!"Invalid query type");
-   }
    return VK_SUCCESS;
 }
 
-/* Copies a query value from one buffer to another from the GPU. */
-static void
-copy_query_value_gpu(struct tu_cmd_buffer *cmdbuf,
-                     struct tu_cs *cs,
-                     uint64_t src_iova,
-                     uint64_t base_write_iova,
-                     uint32_t offset,
-                     VkQueryResultFlags flags) {
-   uint32_t element_size = flags & VK_QUERY_RESULT_64_BIT ?
-         sizeof(uint64_t) : sizeof(uint32_t);
-   uint64_t write_iova = base_write_iova + (offset * element_size);
-
-   tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 5);
-   uint32_t mem_to_mem_flags = flags & VK_QUERY_RESULT_64_BIT ?
-         CP_MEM_TO_MEM_0_DOUBLE : 0;
-   tu_cs_emit(cs, mem_to_mem_flags);
-   tu_cs_emit_qw(cs, write_iova);
-   tu_cs_emit_qw(cs, src_iova);
-}
-
-static void
-emit_copy_query_pool_results(struct tu_cmd_buffer *cmdbuf,
-                             struct tu_cs *cs,
-                             struct tu_query_pool *pool,
-                             uint32_t firstQuery,
-                             uint32_t queryCount,
-                             struct tu_buffer *buffer,
-                             VkDeviceSize dstOffset,
-                             VkDeviceSize stride,
-                             VkQueryResultFlags flags)
-{
-   /* From the Vulkan 1.1.130 spec:
-    *
-    *    vkCmdCopyQueryPoolResults is guaranteed to see the effect of previous
-    *    uses of vkCmdResetQueryPool in the same queue, without any additional
-    *    synchronization.
-    *
-    * To ensure that previous writes to the available bit are coherent, first
-    * wait for all writes to complete.
-    */
-   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-
-   for (uint32_t i = 0; i < queryCount; i++) {
-      uint32_t query = firstQuery + i;
-      uint64_t available_iova = query_available_iova(pool, query);
-      uint64_t buffer_iova = buffer->iova + dstOffset + i * stride;
-      uint32_t result_count = get_result_count(pool);
-      uint32_t statistics = pool->pipeline_statistics;
-
-      /* Wait for the available bit to be set if executed with the
-       * VK_QUERY_RESULT_WAIT_BIT flag. */
-      if (flags & VK_QUERY_RESULT_WAIT_BIT) {
-         tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
-         tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
-                        CP_WAIT_REG_MEM_0_POLL_MEMORY);
-         tu_cs_emit_qw(cs, available_iova);
-         tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(0x1));
-         tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0));
-         tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(16));
-      }
-
-      for (uint32_t k = 0; k < result_count; k++) {
-         uint64_t result_iova;
-
-         if (pool->type == VK_QUERY_TYPE_PIPELINE_STATISTICS) {
-            uint32_t stat_idx = statistics_index(&statistics);
-            result_iova = query_result_iova(pool, query, uint64_t, stat_idx);
-         } else if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
-            result_iova = query_result_iova(pool, query,
-                                            struct perfcntr_query_slot, k);
-         } else {
-            result_iova = query_result_iova(pool, query, uint64_t, k);
-         }
-
-         if (flags & VK_QUERY_RESULT_PARTIAL_BIT) {
-            /* Unconditionally copying the bo->result into the buffer here is
-             * valid because we only set bo->result on vkCmdEndQuery. Thus, even
-             * if the query is unavailable, this will copy the correct partial
-             * value of 0.
-             */
-            copy_query_value_gpu(cmdbuf, cs, result_iova, buffer_iova,
-                                 k /* offset */, flags);
-         } else {
-            /* Conditionally copy bo->result into the buffer based on whether the
-             * query is available.
-             *
-             * NOTE: For the conditional packets to be executed, CP_COND_EXEC
-             * tests that ADDR0 != 0 and ADDR1 < REF. The packet here simply tests
-             * that 0 < available < 2, aka available == 1.
-             */
-            tu_cs_reserve(cs, 7 + 6);
-            tu_cs_emit_pkt7(cs, CP_COND_EXEC, 6);
-            tu_cs_emit_qw(cs, available_iova);
-            tu_cs_emit_qw(cs, available_iova);
-            tu_cs_emit(cs, CP_COND_EXEC_4_REF(0x2));
-            tu_cs_emit(cs, 6); /* Cond execute the next 6 DWORDS */
-
-            /* Start of conditional execution */
-            copy_query_value_gpu(cmdbuf, cs, result_iova, buffer_iova,
-                              k /* offset */, flags);
-            /* End of conditional execution */
-         }
-      }
-
-      if (flags & VK_QUERY_RESULT_WITH_AVAILABILITY_BIT) {
-         copy_query_value_gpu(cmdbuf, cs, available_iova, buffer_iova,
-                              result_count /* offset */, flags);
-      }
-   }
-}
-
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdCopyQueryPoolResults(VkCommandBuffer commandBuffer,
                            VkQueryPool queryPool,
                            uint32_t firstQuery,
@@ -701,1032 +88,35 @@ tu_CmdCopyQueryPoolResults(VkCommandBuffer commandBuffer,
                            VkDeviceSize stride,
                            VkQueryResultFlags flags)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, commandBuffer);
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-   TU_FROM_HANDLE(tu_buffer, buffer, dstBuffer);
-   struct tu_cs *cs = &cmdbuf->cs;
-   assert(firstQuery + queryCount <= pool->size);
-
-   switch (pool->type) {
-   case VK_QUERY_TYPE_OCCLUSION:
-   case VK_QUERY_TYPE_TIMESTAMP:
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
-      return emit_copy_query_pool_results(cmdbuf, cs, pool, firstQuery,
-               queryCount, buffer, dstOffset, stride, flags);
-   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR:
-      unreachable("allowCommandBufferQueryCopies is false");
-   default:
-      assert(!"Invalid query type");
-   }
-}
-
-static void
-emit_reset_query_pool(struct tu_cmd_buffer *cmdbuf,
-                      struct tu_query_pool *pool,
-                      uint32_t firstQuery,
-                      uint32_t queryCount)
-{
-   struct tu_cs *cs = &cmdbuf->cs;
-
-   for (uint32_t i = 0; i < queryCount; i++) {
-      uint32_t query = firstQuery + i;
-      uint32_t statistics = pool->pipeline_statistics;
-
-      tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-      tu_cs_emit_qw(cs, query_available_iova(pool, query));
-      tu_cs_emit_qw(cs, 0x0);
-
-      for (uint32_t k = 0; k < get_result_count(pool); k++) {
-         uint64_t result_iova;
-
-         if (pool->type == VK_QUERY_TYPE_PIPELINE_STATISTICS) {
-            uint32_t stat_idx = statistics_index(&statistics);
-            result_iova = query_result_iova(pool, query, uint64_t, stat_idx);
-         } else if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
-            result_iova = query_result_iova(pool, query,
-                                            struct perfcntr_query_slot, k);
-         } else {
-            result_iova = query_result_iova(pool, query, uint64_t, k);
-         }
-
-         tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-         tu_cs_emit_qw(cs, result_iova);
-         tu_cs_emit_qw(cs, 0x0);
-      }
-   }
-
 }
 
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdResetQueryPool(VkCommandBuffer commandBuffer,
                      VkQueryPool queryPool,
                      uint32_t firstQuery,
                      uint32_t queryCount)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, commandBuffer);
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-
-   switch (pool->type) {
-   case VK_QUERY_TYPE_TIMESTAMP:
-   case VK_QUERY_TYPE_OCCLUSION:
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
-   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR:
-      emit_reset_query_pool(cmdbuf, pool, firstQuery, queryCount);
-      break;
-   default:
-      assert(!"Invalid query type");
-   }
 }
 
-VKAPI_ATTR void VKAPI_CALL
-tu_ResetQueryPool(VkDevice device,
-                  VkQueryPool queryPool,
-                  uint32_t firstQuery,
-                  uint32_t queryCount)
-{
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-
-   for (uint32_t i = 0; i < queryCount; i++) {
-      struct query_slot *slot = slot_address(pool, i + firstQuery);
-      slot->available = 0;
-
-      for (uint32_t k = 0; k < get_result_count(pool); k++) {
-         uint64_t *res;
-
-         if (pool->type == VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR) {
-            res = query_result_addr(pool, i + firstQuery,
-                                    struct perfcntr_query_slot, k);
-         } else {
-            res = query_result_addr(pool, i + firstQuery, uint64_t, k);
-         }
-
-         *res = 0;
-      }
-   }
-}
-
-static void
-emit_begin_occlusion_query(struct tu_cmd_buffer *cmdbuf,
-                           struct tu_query_pool *pool,
-                           uint32_t query)
-{
-   /* From the Vulkan 1.1.130 spec:
-    *
-    *    A query must begin and end inside the same subpass of a render pass
-    *    instance, or must both begin and end outside of a render pass
-    *    instance.
-    *
-    * Unlike on an immediate-mode renderer, Turnip renders all tiles on
-    * vkCmdEndRenderPass, not individually on each vkCmdDraw*. As such, if a
-    * query begins/ends inside the same subpass of a render pass, we need to
-    * record the packets on the secondary draw command stream. cmdbuf->draw_cs
-    * is then run on every tile during render, so we just need to accumulate
-    * sample counts in slot->result to compute the query result.
-    */
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-
-   uint64_t begin_iova = occlusion_query_iova(pool, query, begin);
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_SAMPLE_COUNT_CONTROL(.copy = true));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_SAMPLE_COUNT_ADDR(.qword = begin_iova));
-
-   tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
-   tu_cs_emit(cs, ZPASS_DONE);
-}
-
-static void
-emit_begin_stat_query(struct tu_cmd_buffer *cmdbuf,
-                      struct tu_query_pool *pool,
-                      uint32_t query)
-{
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-   uint64_t begin_iova = pipeline_stat_query_iova(pool, query, begin);
-
-   if (is_pipeline_query_with_vertex_stage(pool->pipeline_statistics)) {
-      bool need_cond_exec = cmdbuf->state.pass && cmdbuf->state.prim_counters_running;
-      cmdbuf->state.prim_counters_running++;
-
-      /* Prevent starting primitive counters when it is supposed to be stopped
-       * for outer VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT query.
-       */
-      if (need_cond_exec) {
-         tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
-                        CP_COND_REG_EXEC_0_SYSMEM |
-                        CP_COND_REG_EXEC_0_BINNING);
-      }
-
-      tu6_emit_event_write(cmdbuf, cs, START_PRIMITIVE_CTRS);
-
-      tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
-      tu_cs_emit_qw(cs, global_iova(cmdbuf, vtx_stats_query_not_running));
-      tu_cs_emit(cs, 0);
-
-      if (need_cond_exec) {
-         tu_cond_exec_end(cs);
-      }
-   }
-
-   if (is_pipeline_query_with_fragment_stage(pool->pipeline_statistics)) {
-      tu6_emit_event_write(cmdbuf, cs, START_FRAGMENT_CTRS);
-   }
-
-   if (is_pipeline_query_with_compute_stage(pool->pipeline_statistics)) {
-      tu6_emit_event_write(cmdbuf, cs, START_COMPUTE_CTRS);
-   }
-
-   tu_cs_emit_wfi(cs);
-
-   tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
-   tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_RBBM_PRIMCTR_0_LO) |
-                  CP_REG_TO_MEM_0_CNT(STAT_COUNT * 2) |
-                  CP_REG_TO_MEM_0_64B);
-   tu_cs_emit_qw(cs, begin_iova);
-}
-
-static void
-emit_perfcntrs_pass_start(struct tu_cs *cs, uint32_t pass)
-{
-   tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
-   tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(
-                        REG_A6XX_CP_SCRATCH_REG(PERF_CNTRS_REG)) |
-                  A6XX_CP_REG_TEST_0_BIT(pass) |
-                  A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
-   tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
-}
-
-static void
-emit_begin_perf_query(struct tu_cmd_buffer *cmdbuf,
-                           struct tu_query_pool *pool,
-                           uint32_t query)
-{
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-   uint32_t last_pass = ~0;
-
-   if (cmdbuf->state.pass) {
-      cmdbuf->state.rp.draw_cs_writes_to_cond_pred = true;
-   }
-
-   /* Querying perf counters happens in these steps:
-    *
-    *  0) There's a scratch reg to set a pass index for perf counters query.
-    *     Prepare cmd streams to set each pass index to the reg at device
-    *     creation time. See tu_CreateDevice in tu_device.c
-    *  1) Emit command streams to read all requested perf counters at all
-    *     passes in begin/end query with CP_REG_TEST/CP_COND_REG_EXEC, which
-    *     reads the scratch reg where pass index is set.
-    *     See emit_perfcntrs_pass_start.
-    *  2) Pick the right cs setting proper pass index to the reg and prepend
-    *     it to the command buffer at each submit time.
-    *     See tu_QueueSubmit in tu_drm.c
-    *  3) If the pass index in the reg is true, then executes the command
-    *     stream below CP_COND_REG_EXEC.
-    */
-
-   tu_cs_emit_wfi(cs);
-
-   for (uint32_t i = 0; i < pool->counter_index_count; i++) {
-      struct tu_perf_query_data *data = &pool->perf_query_data[i];
-
-      if (last_pass != data->pass) {
-         last_pass = data->pass;
-
-         if (data->pass != 0)
-            tu_cond_exec_end(cs);
-         emit_perfcntrs_pass_start(cs, data->pass);
-      }
-
-      const struct fd_perfcntr_counter *counter =
-            &pool->perf_group[data->gid].counters[data->cntr_reg];
-      const struct fd_perfcntr_countable *countable =
-            &pool->perf_group[data->gid].countables[data->cid];
-
-      tu_cs_emit_pkt4(cs, counter->select_reg, 1);
-      tu_cs_emit(cs, countable->selector);
-   }
-   tu_cond_exec_end(cs);
-
-   last_pass = ~0;
-   tu_cs_emit_wfi(cs);
-
-   for (uint32_t i = 0; i < pool->counter_index_count; i++) {
-      struct tu_perf_query_data *data = &pool->perf_query_data[i];
-
-      if (last_pass != data->pass) {
-         last_pass = data->pass;
-
-         if (data->pass != 0)
-            tu_cond_exec_end(cs);
-         emit_perfcntrs_pass_start(cs, data->pass);
-      }
-
-      const struct fd_perfcntr_counter *counter =
-            &pool->perf_group[data->gid].counters[data->cntr_reg];
-
-      uint64_t begin_iova = perf_query_iova(pool, 0, begin, data->app_idx);
-
-      tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
-      tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(counter->counter_reg_lo) |
-                     CP_REG_TO_MEM_0_64B);
-      tu_cs_emit_qw(cs, begin_iova);
-   }
-   tu_cond_exec_end(cs);
-}
-
-static void
-emit_begin_xfb_query(struct tu_cmd_buffer *cmdbuf,
-                     struct tu_query_pool *pool,
-                     uint32_t query,
-                     uint32_t stream_id)
-{
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-   uint64_t begin_iova = primitive_query_iova(pool, query, begin[0], 0);
-
-   tu_cs_emit_regs(cs, A6XX_VPC_SO_STREAM_COUNTS(.qword = begin_iova));
-   tu6_emit_event_write(cmdbuf, cs, WRITE_PRIMITIVE_COUNTS);
-}
-
-static void
-emit_begin_prim_generated_query(struct tu_cmd_buffer *cmdbuf,
-                                struct tu_query_pool *pool,
-                                uint32_t query)
-{
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-   uint64_t begin_iova = primitives_generated_query_iova(pool, query, begin);
-
-   if (cmdbuf->state.pass) {
-      cmdbuf->state.rp.has_prim_generated_query_in_rp = true;
-   } else {
-      cmdbuf->state.prim_generated_query_running_before_rp = true;
-   }
-
-   cmdbuf->state.prim_counters_running++;
-
-   if (cmdbuf->state.pass) {
-      /* Primitives that passed all tests are still counted in in each
-       * tile even with HW binning beforehand. Do not permit it.
-       */
-      tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
-                           CP_COND_REG_EXEC_0_SYSMEM |
-                           CP_COND_REG_EXEC_0_BINNING);
-   }
-
-   tu6_emit_event_write(cmdbuf, cs, START_PRIMITIVE_CTRS);
-
-   tu_cs_emit_wfi(cs);
-
-   tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
-   tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_RBBM_PRIMCTR_7_LO) |
-                  CP_REG_TO_MEM_0_CNT(2) |
-                  CP_REG_TO_MEM_0_64B);
-   tu_cs_emit_qw(cs, begin_iova);
-
-   if (cmdbuf->state.pass) {
-      tu_cond_exec_end(cs);
-   }
-}
-
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdBeginQuery(VkCommandBuffer commandBuffer,
                  VkQueryPool queryPool,
                  uint32_t query,
                  VkQueryControlFlags flags)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, commandBuffer);
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-   assert(query < pool->size);
-
-   switch (pool->type) {
-   case VK_QUERY_TYPE_OCCLUSION:
-      /* In freedreno, there is no implementation difference between
-       * GL_SAMPLES_PASSED and GL_ANY_SAMPLES_PASSED, so we can similarly
-       * ignore the VK_QUERY_CONTROL_PRECISE_BIT flag here.
-       */
-      emit_begin_occlusion_query(cmdbuf, pool, query);
-      break;
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-      emit_begin_xfb_query(cmdbuf, pool, query, 0);
-      break;
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-      emit_begin_prim_generated_query(cmdbuf, pool, query);
-      break;
-   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR:
-      emit_begin_perf_query(cmdbuf, pool, query);
-      break;
-   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
-      emit_begin_stat_query(cmdbuf, pool, query);
-      break;
-   case VK_QUERY_TYPE_TIMESTAMP:
-      unreachable("Unimplemented query type");
-   default:
-      assert(!"Invalid query type");
-   }
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdBeginQueryIndexedEXT(VkCommandBuffer commandBuffer,
-                           VkQueryPool queryPool,
-                           uint32_t query,
-                           VkQueryControlFlags flags,
-                           uint32_t index)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, commandBuffer);
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-   assert(query < pool->size);
-
-   switch (pool->type) {
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-      emit_begin_xfb_query(cmdbuf, pool, query, index);
-      break;
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-      emit_begin_prim_generated_query(cmdbuf, pool, query);
-      break;
-   default:
-      assert(!"Invalid query type");
-   }
-}
-
-static void
-emit_end_occlusion_query(struct tu_cmd_buffer *cmdbuf,
-                         struct tu_query_pool *pool,
-                         uint32_t query)
-{
-   /* Ending an occlusion query happens in a few steps:
-    *    1) Set the slot->end to UINT64_MAX.
-    *    2) Set up the SAMPLE_COUNT registers and trigger a CP_EVENT_WRITE to
-    *       write the current sample count value into slot->end.
-    *    3) Since (2) is asynchronous, wait until slot->end is not equal to
-    *       UINT64_MAX before continuing via CP_WAIT_REG_MEM.
-    *    4) Accumulate the results of the query (slot->end - slot->begin) into
-    *       slot->result.
-    *    5) If vkCmdEndQuery is *not* called from within the scope of a render
-    *       pass, set the slot's available bit since the query is now done.
-    *    6) If vkCmdEndQuery *is* called from within the scope of a render
-    *       pass, we cannot mark as available yet since the commands in
-    *       draw_cs are not run until vkCmdEndRenderPass.
-    */
-   const struct tu_render_pass *pass = cmdbuf->state.pass;
-   struct tu_cs *cs = pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-
-   uint64_t available_iova = query_available_iova(pool, query);
-   uint64_t begin_iova = occlusion_query_iova(pool, query, begin);
-   uint64_t end_iova = occlusion_query_iova(pool, query, end);
-   uint64_t result_iova = query_result_iova(pool, query, uint64_t, 0);
-   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-   tu_cs_emit_qw(cs, end_iova);
-   tu_cs_emit_qw(cs, 0xffffffffffffffffull);
-
-   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_SAMPLE_COUNT_CONTROL(.copy = true));
-
-   tu_cs_emit_regs(cs,
-                   A6XX_RB_SAMPLE_COUNT_ADDR(.qword = end_iova));
-
-   tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
-   tu_cs_emit(cs, ZPASS_DONE);
-
-   tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
-   tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_NE) |
-                  CP_WAIT_REG_MEM_0_POLL_MEMORY);
-   tu_cs_emit_qw(cs, end_iova);
-   tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(0xffffffff));
-   tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0));
-   tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(16));
-
-   /* result (dst) = result (srcA) + end (srcB) - begin (srcC) */
-   tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 9);
-   tu_cs_emit(cs, CP_MEM_TO_MEM_0_DOUBLE | CP_MEM_TO_MEM_0_NEG_C);
-   tu_cs_emit_qw(cs, result_iova);
-   tu_cs_emit_qw(cs, result_iova);
-   tu_cs_emit_qw(cs, end_iova);
-   tu_cs_emit_qw(cs, begin_iova);
-
-   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-
-   if (pass)
-      /* Technically, queries should be tracked per-subpass, but here we track
-       * at the render pass level to simply the code a bit. This is safe
-       * because the only commands that use the available bit are
-       * vkCmdCopyQueryPoolResults and vkCmdResetQueryPool, both of which
-       * cannot be invoked from inside a render pass scope.
-       */
-      cs = &cmdbuf->draw_epilogue_cs;
-
-   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-   tu_cs_emit_qw(cs, available_iova);
-   tu_cs_emit_qw(cs, 0x1);
-}
-
-/* PRIMITIVE_CTRS is used for two distinct queries:
- * - VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT
- * - VK_QUERY_TYPE_PIPELINE_STATISTICS
- * If one is nested inside other - STOP_PRIMITIVE_CTRS should be emitted
- * only for outer query.
- *
- * Also, pipeline stat query could run outside of renderpass and prim gen
- * query inside of secondary cmd buffer - for such case we ought to track
- * the status of pipeline stats query.
- */
-static void
-emit_stop_primitive_ctrs(struct tu_cmd_buffer *cmdbuf,
-                         struct tu_cs *cs,
-                         enum VkQueryType query_type)
-{
-   bool is_secondary = cmdbuf->vk.level == VK_COMMAND_BUFFER_LEVEL_SECONDARY;
-   cmdbuf->state.prim_counters_running--;
-   if (cmdbuf->state.prim_counters_running == 0) {
-      bool need_cond_exec =
-         is_secondary &&
-         query_type == VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT &&
-         is_pipeline_query_with_vertex_stage(cmdbuf->inherited_pipeline_statistics);
-
-      if (!need_cond_exec) {
-         tu6_emit_event_write(cmdbuf, cs, STOP_PRIMITIVE_CTRS);
-      } else {
-         tu_cs_reserve(cs, 7 + 2);
-         /* Check that pipeline stats query is not running, only then
-          * we count stop the counter.
-          */
-         tu_cs_emit_pkt7(cs, CP_COND_EXEC, 6);
-         tu_cs_emit_qw(cs, global_iova(cmdbuf, vtx_stats_query_not_running));
-         tu_cs_emit_qw(cs, global_iova(cmdbuf, vtx_stats_query_not_running));
-         tu_cs_emit(cs, CP_COND_EXEC_4_REF(0x2));
-         tu_cs_emit(cs, 2); /* Cond execute the next 2 DWORDS */
-
-         tu6_emit_event_write(cmdbuf, cs, STOP_PRIMITIVE_CTRS);
-      }
-   }
-
-   if (query_type == VK_QUERY_TYPE_PIPELINE_STATISTICS) {
-      tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
-      tu_cs_emit_qw(cs, global_iova(cmdbuf, vtx_stats_query_not_running));
-      tu_cs_emit(cs, 1);
-   }
 }
 
-static void
-emit_end_stat_query(struct tu_cmd_buffer *cmdbuf,
-                    struct tu_query_pool *pool,
-                    uint32_t query)
-{
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-   uint64_t end_iova = pipeline_stat_query_iova(pool, query, end);
-   uint64_t available_iova = query_available_iova(pool, query);
-   uint64_t result_iova;
-   uint64_t stat_start_iova;
-   uint64_t stat_stop_iova;
-
-   if (is_pipeline_query_with_vertex_stage(pool->pipeline_statistics)) {
-      /* No need to conditionally execute STOP_PRIMITIVE_CTRS when
-       * we are inside VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT inside of a
-       * renderpass, because it is already stopped.
-       */
-      emit_stop_primitive_ctrs(cmdbuf, cs, VK_QUERY_TYPE_PIPELINE_STATISTICS);
-   }
-
-   if (is_pipeline_query_with_fragment_stage(pool->pipeline_statistics)) {
-      tu6_emit_event_write(cmdbuf, cs, STOP_FRAGMENT_CTRS);
-   }
-
-   if (is_pipeline_query_with_compute_stage(pool->pipeline_statistics)) {
-      tu6_emit_event_write(cmdbuf, cs, STOP_COMPUTE_CTRS);
-   }
-
-   tu_cs_emit_wfi(cs);
-
-   tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
-   tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_RBBM_PRIMCTR_0_LO) |
-                  CP_REG_TO_MEM_0_CNT(STAT_COUNT * 2) |
-                  CP_REG_TO_MEM_0_64B);
-   tu_cs_emit_qw(cs, end_iova);
-
-   for (int i = 0; i < STAT_COUNT; i++) {
-      result_iova = query_result_iova(pool, query, uint64_t, i);
-      stat_start_iova = pipeline_stat_query_iova(pool, query, begin[i]);
-      stat_stop_iova = pipeline_stat_query_iova(pool, query, end[i]);
-
-      tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 9);
-      tu_cs_emit(cs, CP_MEM_TO_MEM_0_WAIT_FOR_MEM_WRITES |
-                     CP_MEM_TO_MEM_0_DOUBLE |
-                     CP_MEM_TO_MEM_0_NEG_C);
-
-      tu_cs_emit_qw(cs, result_iova);
-      tu_cs_emit_qw(cs, result_iova);
-      tu_cs_emit_qw(cs, stat_stop_iova);
-      tu_cs_emit_qw(cs, stat_start_iova);
-   }
-
-   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-
-   if (cmdbuf->state.pass)
-      cs = &cmdbuf->draw_epilogue_cs;
-
-   /* Set the availability to 1 */
-   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-   tu_cs_emit_qw(cs, available_iova);
-   tu_cs_emit_qw(cs, 0x1);
-}
-
-static void
-emit_end_perf_query(struct tu_cmd_buffer *cmdbuf,
-                         struct tu_query_pool *pool,
-                         uint32_t query)
-{
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-   uint64_t available_iova = query_available_iova(pool, query);
-   uint64_t end_iova;
-   uint64_t begin_iova;
-   uint64_t result_iova;
-   uint32_t last_pass = ~0;
-
-   for (uint32_t i = 0; i < pool->counter_index_count; i++) {
-      struct tu_perf_query_data *data = &pool->perf_query_data[i];
-
-      if (last_pass != data->pass) {
-         last_pass = data->pass;
-
-         if (data->pass != 0)
-            tu_cond_exec_end(cs);
-         emit_perfcntrs_pass_start(cs, data->pass);
-      }
-
-      const struct fd_perfcntr_counter *counter =
-            &pool->perf_group[data->gid].counters[data->cntr_reg];
-
-      end_iova = perf_query_iova(pool, 0, end, data->app_idx);
-
-      tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
-      tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(counter->counter_reg_lo) |
-                     CP_REG_TO_MEM_0_64B);
-      tu_cs_emit_qw(cs, end_iova);
-   }
-   tu_cond_exec_end(cs);
-
-   last_pass = ~0;
-   tu_cs_emit_wfi(cs);
-
-   for (uint32_t i = 0; i < pool->counter_index_count; i++) {
-      struct tu_perf_query_data *data = &pool->perf_query_data[i];
-
-      if (last_pass != data->pass) {
-         last_pass = data->pass;
-
-
-         if (data->pass != 0)
-            tu_cond_exec_end(cs);
-         emit_perfcntrs_pass_start(cs, data->pass);
-      }
-
-      result_iova = query_result_iova(pool, 0, struct perfcntr_query_slot,
-             data->app_idx);
-      begin_iova = perf_query_iova(pool, 0, begin, data->app_idx);
-      end_iova = perf_query_iova(pool, 0, end, data->app_idx);
-
-      /* result += end - begin */
-      tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 9);
-      tu_cs_emit(cs, CP_MEM_TO_MEM_0_WAIT_FOR_MEM_WRITES |
-                     CP_MEM_TO_MEM_0_DOUBLE |
-                     CP_MEM_TO_MEM_0_NEG_C);
-
-      tu_cs_emit_qw(cs, result_iova);
-      tu_cs_emit_qw(cs, result_iova);
-      tu_cs_emit_qw(cs, end_iova);
-      tu_cs_emit_qw(cs, begin_iova);
-   }
-   tu_cond_exec_end(cs);
-
-   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-
-   if (cmdbuf->state.pass)
-      cs = &cmdbuf->draw_epilogue_cs;
-
-   /* Set the availability to 1 */
-   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-   tu_cs_emit_qw(cs, available_iova);
-   tu_cs_emit_qw(cs, 0x1);
-}
-
-static void
-emit_end_xfb_query(struct tu_cmd_buffer *cmdbuf,
-                   struct tu_query_pool *pool,
-                   uint32_t query,
-                   uint32_t stream_id)
-{
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-
-   uint64_t end_iova = primitive_query_iova(pool, query, end[0], 0);
-   uint64_t result_written_iova = query_result_iova(pool, query, uint64_t, 0);
-   uint64_t result_generated_iova = query_result_iova(pool, query, uint64_t, 1);
-   uint64_t begin_written_iova = primitive_query_iova(pool, query, begin[stream_id], 0);
-   uint64_t begin_generated_iova = primitive_query_iova(pool, query, begin[stream_id], 1);
-   uint64_t end_written_iova = primitive_query_iova(pool, query, end[stream_id], 0);
-   uint64_t end_generated_iova = primitive_query_iova(pool, query, end[stream_id], 1);
-   uint64_t available_iova = query_available_iova(pool, query);
-
-   tu_cs_emit_regs(cs, A6XX_VPC_SO_STREAM_COUNTS(.qword = end_iova));
-   tu6_emit_event_write(cmdbuf, cs, WRITE_PRIMITIVE_COUNTS);
-
-   tu_cs_emit_wfi(cs);
-   tu6_emit_event_write(cmdbuf, cs, CACHE_FLUSH_TS);
-
-   /* Set the count of written primitives */
-   tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 9);
-   tu_cs_emit(cs, CP_MEM_TO_MEM_0_DOUBLE | CP_MEM_TO_MEM_0_NEG_C |
-                  CP_MEM_TO_MEM_0_WAIT_FOR_MEM_WRITES | 0x80000000);
-   tu_cs_emit_qw(cs, result_written_iova);
-   tu_cs_emit_qw(cs, result_written_iova);
-   tu_cs_emit_qw(cs, end_written_iova);
-   tu_cs_emit_qw(cs, begin_written_iova);
-
-   tu6_emit_event_write(cmdbuf, cs, CACHE_FLUSH_TS);
-
-   /* Set the count of generated primitives */
-   tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 9);
-   tu_cs_emit(cs, CP_MEM_TO_MEM_0_DOUBLE | CP_MEM_TO_MEM_0_NEG_C |
-                  CP_MEM_TO_MEM_0_WAIT_FOR_MEM_WRITES | 0x80000000);
-   tu_cs_emit_qw(cs, result_generated_iova);
-   tu_cs_emit_qw(cs, result_generated_iova);
-   tu_cs_emit_qw(cs, end_generated_iova);
-   tu_cs_emit_qw(cs, begin_generated_iova);
-
-   /* Set the availability to 1 */
-   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-   tu_cs_emit_qw(cs, available_iova);
-   tu_cs_emit_qw(cs, 0x1);
-}
-
-static void
-emit_end_prim_generated_query(struct tu_cmd_buffer *cmdbuf,
-                              struct tu_query_pool *pool,
-                              uint32_t query)
-{
-   struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
-
-   if (!cmdbuf->state.pass) {
-      cmdbuf->state.prim_generated_query_running_before_rp = false;
-   }
-
-   uint64_t begin_iova = primitives_generated_query_iova(pool, query, begin);
-   uint64_t end_iova = primitives_generated_query_iova(pool, query, end);
-   uint64_t result_iova = primitives_generated_query_iova(pool, query, result);
-   uint64_t available_iova = query_available_iova(pool, query);
-
-   if (cmdbuf->state.pass) {
-      tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
-                             CP_COND_REG_EXEC_0_SYSMEM |
-                             CP_COND_REG_EXEC_0_BINNING);
-   }
-
-   tu_cs_emit_wfi(cs);
-
-   tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
-   tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_RBBM_PRIMCTR_7_LO) |
-                  CP_REG_TO_MEM_0_CNT(2) |
-                  CP_REG_TO_MEM_0_64B);
-   tu_cs_emit_qw(cs, end_iova);
-
-   tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 9);
-   tu_cs_emit(cs, CP_MEM_TO_MEM_0_DOUBLE | CP_MEM_TO_MEM_0_NEG_C |
-                  CP_MEM_TO_MEM_0_WAIT_FOR_MEM_WRITES);
-   tu_cs_emit_qw(cs, result_iova);
-   tu_cs_emit_qw(cs, result_iova);
-   tu_cs_emit_qw(cs, end_iova);
-   tu_cs_emit_qw(cs, begin_iova);
-
-   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-
-   /* Should be after waiting for mem writes to have up to date info
-    * about which query is running.
-    */
-   emit_stop_primitive_ctrs(cmdbuf, cs, VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT);
-
-   if (cmdbuf->state.pass) {
-      tu_cond_exec_end(cs);
-   }
-
-   if (cmdbuf->state.pass)
-      cs = &cmdbuf->draw_epilogue_cs;
-
-   /* Set the availability to 1 */
-   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-   tu_cs_emit_qw(cs, available_iova);
-   tu_cs_emit_qw(cs, 0x1);
-}
-
-/* Implement this bit of spec text from section 17.2 "Query Operation":
- *
- *     If queries are used while executing a render pass instance that has
- *     multiview enabled, the query uses N consecutive query indices in the
- *     query pool (starting at query) where N is the number of bits set in the
- *     view mask in the subpass the query is used in. How the numerical
- *     results of the query are distributed among the queries is
- *     implementation-dependent. For example, some implementations may write
- *     each view’s results to a distinct query, while other implementations
- *     may write the total result to the first query and write zero to the
- *     other queries. However, the sum of the results in all the queries must
- *     accurately reflect the total result of the query summed over all views.
- *     Applications can sum the results from all the queries to compute the
- *     total result.
- *
- * Since we execute all views at once, we write zero to the other queries.
- * Furthermore, because queries must be reset before use, and we set the
- * result to 0 in vkCmdResetQueryPool(), we just need to mark it as available.
- */
-
-static void
-handle_multiview_queries(struct tu_cmd_buffer *cmd,
-                         struct tu_query_pool *pool,
-                         uint32_t query)
-{
-   if (!cmd->state.pass || !cmd->state.subpass->multiview_mask)
-      return;
-
-   unsigned views = util_bitcount(cmd->state.subpass->multiview_mask);
-   struct tu_cs *cs = &cmd->draw_epilogue_cs;
-
-   for (uint32_t i = 1; i < views; i++) {
-      tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-      tu_cs_emit_qw(cs, query_available_iova(pool, query + i));
-      tu_cs_emit_qw(cs, 0x1);
-   }
-}
-
-VKAPI_ATTR void VKAPI_CALL
+void
 tu_CmdEndQuery(VkCommandBuffer commandBuffer,
                VkQueryPool queryPool,
                uint32_t query)
 {
-   TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, commandBuffer);
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-   assert(query < pool->size);
-
-   switch (pool->type) {
-   case VK_QUERY_TYPE_OCCLUSION:
-      emit_end_occlusion_query(cmdbuf, pool, query);
-      break;
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-      emit_end_xfb_query(cmdbuf, pool, query, 0);
-      break;
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-      emit_end_prim_generated_query(cmdbuf, pool, query);
-      break;
-   case VK_QUERY_TYPE_PERFORMANCE_QUERY_KHR:
-      emit_end_perf_query(cmdbuf, pool, query);
-      break;
-   case VK_QUERY_TYPE_PIPELINE_STATISTICS:
-      emit_end_stat_query(cmdbuf, pool, query);
-      break;
-   case VK_QUERY_TYPE_TIMESTAMP:
-      unreachable("Unimplemented query type");
-   default:
-      assert(!"Invalid query type");
-   }
-
-   handle_multiview_queries(cmdbuf, pool, query);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdEndQueryIndexedEXT(VkCommandBuffer commandBuffer,
-                         VkQueryPool queryPool,
-                         uint32_t query,
-                         uint32_t index)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, commandBuffer);
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-   assert(query < pool->size);
-
-   switch (pool->type) {
-   case VK_QUERY_TYPE_TRANSFORM_FEEDBACK_STREAM_EXT:
-      assert(index <= 4);
-      emit_end_xfb_query(cmdbuf, pool, query, index);
-      break;
-   case VK_QUERY_TYPE_PRIMITIVES_GENERATED_EXT:
-      emit_end_prim_generated_query(cmdbuf, pool, query);
-      break;
-   default:
-      assert(!"Invalid query type");
-   }
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_CmdWriteTimestamp2(VkCommandBuffer commandBuffer,
-                      VkPipelineStageFlagBits2 pipelineStage,
-                      VkQueryPool queryPool,
-                      uint32_t query)
-{
-   TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
-   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
-
-   /* Inside a render pass, just write the timestamp multiple times so that
-    * the user gets the last one if we use GMEM. There isn't really much
-    * better we can do, and this seems to be what the blob does too.
-    */
-   struct tu_cs *cs = cmd->state.pass ? &cmd->draw_cs : &cmd->cs;
-
-   /* Stages that will already have been executed by the time the CP executes
-    * the REG_TO_MEM. DrawIndirect parameters are read by the CP, so the draw
-    * indirect stage counts as top-of-pipe too.
-    */
-   VkPipelineStageFlags2 top_of_pipe_flags =
-      VK_PIPELINE_STAGE_2_TOP_OF_PIPE_BIT |
-      VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT;
-
-   if (pipelineStage & ~top_of_pipe_flags) {
-      /* Execute a WFI so that all commands complete. Note that CP_REG_TO_MEM
-       * does CP_WAIT_FOR_ME internally, which will wait for the WFI to
-       * complete.
-       *
-       * Stalling the CP like this is really unfortunate, but I don't think
-       * there's a better solution that allows all 48 bits of precision
-       * because CP_EVENT_WRITE doesn't support 64-bit timestamps.
-       */
-      tu_cs_emit_wfi(cs);
-   }
-
-   tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
-   tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_CP_ALWAYS_ON_COUNTER) |
-                  CP_REG_TO_MEM_0_CNT(2) |
-                  CP_REG_TO_MEM_0_64B);
-   tu_cs_emit_qw(cs, query_result_iova(pool, query, uint64_t, 0));
-
-   /* Only flag availability once the entire renderpass is done, similar to
-    * the begin/end path.
-    */
-   cs = cmd->state.pass ? &cmd->draw_epilogue_cs : &cmd->cs;
-
-   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
-   tu_cs_emit_qw(cs, query_available_iova(pool, query));
-   tu_cs_emit_qw(cs, 0x1);
-
-   /* From the spec for vkCmdWriteTimestamp:
-    *
-    *    If vkCmdWriteTimestamp is called while executing a render pass
-    *    instance that has multiview enabled, the timestamp uses N consecutive
-    *    query indices in the query pool (starting at query) where N is the
-    *    number of bits set in the view mask of the subpass the command is
-    *    executed in. The resulting query values are determined by an
-    *    implementation-dependent choice of one of the following behaviors:
-    *
-    *    -   The first query is a timestamp value and (if more than one bit is
-    *        set in the view mask) zero is written to the remaining queries.
-    *        If two timestamps are written in the same subpass, the sum of the
-    *        execution time of all views between those commands is the
-    *        difference between the first query written by each command.
-    *
-    *    -   All N queries are timestamp values. If two timestamps are written
-    *        in the same subpass, the sum of the execution time of all views
-    *        between those commands is the sum of the difference between
-    *        corresponding queries written by each command. The difference
-    *        between corresponding queries may be the execution time of a
-    *        single view.
-    *
-    * We execute all views in the same draw call, so we implement the first
-    * option, the same as regular queries.
-    */
-   handle_multiview_queries(cmd, pool, query);
-}
-
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_EnumeratePhysicalDeviceQueueFamilyPerformanceQueryCountersKHR(
-    VkPhysicalDevice                            physicalDevice,
-    uint32_t                                    queueFamilyIndex,
-    uint32_t*                                   pCounterCount,
-    VkPerformanceCounterKHR*                    pCounters,
-    VkPerformanceCounterDescriptionKHR*         pCounterDescriptions)
-{
-   TU_FROM_HANDLE(tu_physical_device, phydev, physicalDevice);
-
-   uint32_t desc_count = *pCounterCount;
-   uint32_t group_count;
-   const struct fd_perfcntr_group *group =
-         fd_perfcntrs(&phydev->dev_id, &group_count);
-
-   VK_OUTARRAY_MAKE_TYPED(VkPerformanceCounterKHR, out, pCounters, pCounterCount);
-   VK_OUTARRAY_MAKE_TYPED(VkPerformanceCounterDescriptionKHR, out_desc,
-                          pCounterDescriptions, &desc_count);
-
-   for (int i = 0; i < group_count; i++) {
-      for (int j = 0; j < group[i].num_countables; j++) {
-
-         vk_outarray_append_typed(VkPerformanceCounterKHR, &out, counter) {
-            counter->scope = VK_PERFORMANCE_COUNTER_SCOPE_COMMAND_BUFFER_KHR;
-            counter->unit =
-                  fd_perfcntr_type_to_vk_unit[group[i].countables[j].query_type];
-            counter->storage =
-                  fd_perfcntr_type_to_vk_storage[group[i].countables[j].query_type];
-
-            unsigned char sha1_result[20];
-            _mesa_sha1_compute(group[i].countables[j].name,
-                               strlen(group[i].countables[j].name),
-                               sha1_result);
-            memcpy(counter->uuid, sha1_result, sizeof(counter->uuid));
-         }
-
-         vk_outarray_append_typed(VkPerformanceCounterDescriptionKHR, &out_desc, desc) {
-            desc->flags = 0;
-
-            snprintf(desc->name, sizeof(desc->name),
-                     "%s", group[i].countables[j].name);
-            snprintf(desc->category, sizeof(desc->category), "%s", group[i].name);
-            snprintf(desc->description, sizeof(desc->description),
-                     "%s: %s performance counter",
-                     group[i].name, group[i].countables[j].name);
-         }
-      }
-   }
-
-   return vk_outarray_status(&out);
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_GetPhysicalDeviceQueueFamilyPerformanceQueryPassesKHR(
-      VkPhysicalDevice                            physicalDevice,
-      const VkQueryPoolPerformanceCreateInfoKHR*  pPerformanceQueryCreateInfo,
-      uint32_t*                                   pNumPasses)
-{
-   TU_FROM_HANDLE(tu_physical_device, phydev, physicalDevice);
-   uint32_t group_count = 0;
-   uint32_t gid = 0, cid = 0, n_passes;
-   const struct fd_perfcntr_group *group =
-         fd_perfcntrs(&phydev->dev_id, &group_count);
-
-   uint32_t counters_requested[group_count];
-   memset(counters_requested, 0x0, sizeof(counters_requested));
-   *pNumPasses = 1;
-
-   for (unsigned i = 0; i < pPerformanceQueryCreateInfo->counterIndexCount; i++) {
-      perfcntr_index(group, group_count,
-                     pPerformanceQueryCreateInfo->pCounterIndices[i],
-                     &gid, &cid);
-
-      counters_requested[gid]++;
-   }
-
-   for (uint32_t i = 0; i < group_count; i++) {
-      n_passes = DIV_ROUND_UP(counters_requested[i], group[i].num_counters);
-      *pNumPasses = MAX2(*pNumPasses, n_passes);
-   }
 }
 
-VKAPI_ATTR VkResult VKAPI_CALL
-tu_AcquireProfilingLockKHR(VkDevice device,
-                           const VkAcquireProfilingLockInfoKHR* pInfo)
-{
-   /* TODO. Probably there's something to do for kgsl. */
-   return VK_SUCCESS;
-}
-
-VKAPI_ATTR void VKAPI_CALL
-tu_ReleaseProfilingLockKHR(VkDevice device)
+void
+tu_CmdWriteTimestamp(VkCommandBuffer commandBuffer,
+                     VkPipelineStageFlagBits pipelineStage,
+                     VkQueryPool queryPool,
+                     uint32_t query)
 {
-   /* TODO. Probably there's something to do for kgsl. */
-   return;
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_shader.c b/lib/mesa/src/freedreno/vulkan/tu_shader.c
index e485f8f5c..f6e13d7c4 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_shader.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_shader.c
@@ -1,894 +1,336 @@
 /*
  * Copyright © 2019 Google LLC
- * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_shader.h"
+#include "tu_private.h"
 
 #include "spirv/nir_spirv.h"
 #include "util/mesa-sha1.h"
-#include "nir/nir_xfb_info.h"
-#include "nir/nir_vulkan.h"
-#include "vk_pipeline.h"
-#include "vk_util.h"
 
 #include "ir3/ir3_nir.h"
 
-#include "tu_device.h"
-#include "tu_descriptor_set.h"
-#include "tu_pipeline.h"
-
-nir_shader *
-tu_spirv_to_nir(struct tu_device *dev,
-                void *mem_ctx,
-                const VkPipelineShaderStageCreateInfo *stage_info,
-                gl_shader_stage stage)
+static nir_shader *
+tu_spirv_to_nir(struct ir3_compiler *compiler,
+                const uint32_t *words,
+                size_t word_count,
+                gl_shader_stage stage,
+                const char *entry_point_name,
+                const VkSpecializationInfo *spec_info)
 {
    /* TODO these are made-up */
    const struct spirv_to_nir_options spirv_options = {
-      .ubo_addr_format = nir_address_format_vec2_index_32bit_offset,
-      .ssbo_addr_format = nir_address_format_vec2_index_32bit_offset,
-
-      /* Accessed via stg/ldg */
-      .phys_ssbo_addr_format = nir_address_format_64bit_global,
-
-      /* Accessed via the const register file */
-      .push_const_addr_format = nir_address_format_logical,
-
-      /* Accessed via ldl/stl */
-      .shared_addr_format = nir_address_format_32bit_offset,
-
-      /* Accessed via stg/ldg (not used with Vulkan?) */
-      .global_addr_format = nir_address_format_64bit_global,
-
-      /* Use 16-bit math for RelaxedPrecision ALU ops */
-      .mediump_16bit_alu = true,
-
-      /* ViewID is a sysval in geometry stages and an input in the FS */
-      .view_index_is_input = stage == MESA_SHADER_FRAGMENT,
-      .caps = {
-         .transform_feedback = true,
-         .tessellation = true,
-         .draw_parameters = true,
-         .image_read_without_format = true,
-         .image_write_without_format = true,
-         .variable_pointers = true,
-         .stencil_export = true,
-         .multiview = true,
-         .shader_viewport_index_layer = true,
-         .geometry_streams = true,
-         .device_group = true,
-         .descriptor_indexing = true,
-         .descriptor_array_dynamic_indexing = true,
-         .descriptor_array_non_uniform_indexing = true,
-         .runtime_descriptor_array = true,
-         .float_controls = true,
-         .float16 = true,
-         .int16 = true,
-         .storage_16bit = dev->physical_device->info->a6xx.storage_16bit,
-         .demote_to_helper_invocation = true,
-         .vk_memory_model = true,
-         .vk_memory_model_device_scope = true,
-         .subgroup_basic = true,
-         .subgroup_ballot = true,
-         .subgroup_vote = true,
-         .subgroup_quad = true,
-         .subgroup_shuffle = true,
-         .subgroup_arithmetic = true,
-         .physical_storage_buffer_address = true,
-      },
+      .lower_ubo_ssbo_access_to_offsets = true,
+      .caps = { false },
    };
-
    const nir_shader_compiler_options *nir_options =
-      ir3_get_compiler_options(dev->compiler);
-
-   nir_shader *nir;
-   VkResult result =
-      vk_pipeline_shader_stage_to_nir(&dev->vk, stage_info, &spirv_options,
-                                      nir_options, mem_ctx, &nir);
-   if (result != VK_SUCCESS)
-      return NULL;
-
-   if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_NIR)) {
-      fprintf(stderr, "translated nir:\n");
-      nir_print_shader(nir, stderr);
-   }
-
-   const struct nir_lower_sysvals_to_varyings_options sysvals_to_varyings = {
-      .point_coord = true,
-   };
-   NIR_PASS_V(nir, nir_lower_sysvals_to_varyings, &sysvals_to_varyings);
-
-   NIR_PASS_V(nir, nir_lower_global_vars_to_local);
-
-   /* Older glslang missing bf6efd0316d8 ("SPV: Fix #2293: keep relaxed
-    * precision on arg passed to relaxed param") will pass function args through
-    * a highp temporary, so we need the nir_opt_find_array_copies() and a copy
-    * prop before we lower mediump vars, or you'll be unable to optimize out
-    * array copies after lowering.  We do this before splitting copies, since
-    * that works against nir_opt_find_array_copies().
-    * */
-   NIR_PASS_V(nir, nir_opt_find_array_copies);
-   NIR_PASS_V(nir, nir_opt_copy_prop_vars);
-   NIR_PASS_V(nir, nir_opt_dce);
-
-   NIR_PASS_V(nir, nir_split_var_copies);
-   NIR_PASS_V(nir, nir_lower_var_copies);
-
-   NIR_PASS_V(nir, nir_lower_mediump_vars, nir_var_function_temp | nir_var_shader_temp | nir_var_mem_shared);
-   NIR_PASS_V(nir, nir_opt_copy_prop_vars);
-   NIR_PASS_V(nir, nir_opt_combine_stores, nir_var_all);
-
-   NIR_PASS_V(nir, nir_lower_is_helper_invocation);
-
-   NIR_PASS_V(nir, nir_lower_system_values);
-
-   NIR_PASS_V(nir, nir_lower_frexp);
-
-   ir3_optimize_loop(dev->compiler, nir);
-
-   NIR_PASS_V(nir, nir_opt_conditional_discard);
-
-   return nir;
-}
-
-static void
-lower_load_push_constant(struct tu_device *dev,
-                         nir_builder *b,
-                         nir_intrinsic_instr *instr,
-                         struct tu_shader *shader,
-                         const struct tu_pipeline_layout *layout)
-{
-   uint32_t base = nir_intrinsic_base(instr);
-   assert(base % 4 == 0);
-
-   if (tu6_shared_constants_enable(layout, dev->compiler)) {
-      /* All stages share the same range.  We could potentially add
-       * push_constant_offset to layout and apply it, but this is good for
-       * now.
-       */
-      base += dev->compiler->shared_consts_base_offset * 4;
-   } else {
-      assert(base >= shader->const_state.push_consts.lo * 4);
-      base -= shader->const_state.push_consts.lo * 4;
-   }
-
-   nir_ssa_def *load =
-      nir_load_uniform(b, instr->num_components,
-            instr->dest.ssa.bit_size,
-            nir_ushr(b, instr->src[0].ssa, nir_imm_int(b, 2)),
-            .base = base);
-
-   nir_ssa_def_rewrite_uses(&instr->dest.ssa, load);
-
-   nir_instr_remove(&instr->instr);
-}
-
-static void
-lower_vulkan_resource_index(nir_builder *b, nir_intrinsic_instr *instr,
-                            struct tu_shader *shader,
-                            const struct tu_pipeline_layout *layout)
-{
-   nir_ssa_def *vulkan_idx = instr->src[0].ssa;
-
-   unsigned set = nir_intrinsic_desc_set(instr);
-   unsigned binding = nir_intrinsic_binding(instr);
-   struct tu_descriptor_set_layout *set_layout = layout->set[set].layout;
-   struct tu_descriptor_set_binding_layout *binding_layout =
-      &set_layout->binding[binding];
-   nir_ssa_def *base;
-
-   shader->active_desc_sets |= 1u << set;
-
-   switch (binding_layout->type) {
-   case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
-   case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
-      if (layout->independent_sets) {
-         /* With independent sets, we don't know
-          * layout->set[set].dynamic_offset_start until after link time which
-          * with fast linking means after the shader is compiled. We have to
-          * get it from the const file instead.
-          */
-         base = nir_imm_int(b, binding_layout->dynamic_offset_offset / (4 * A6XX_TEX_CONST_DWORDS));
-         nir_ssa_def *dynamic_offset_start =
-            nir_load_uniform(b, 1, 32, nir_imm_int(b, 0),
-                             .base = shader->const_state.dynamic_offset_loc + set);
-         base = nir_iadd(b, base, dynamic_offset_start);
-      } else {
-         base = nir_imm_int(b, (layout->set[set].dynamic_offset_start +
-            binding_layout->dynamic_offset_offset) / (4 * A6XX_TEX_CONST_DWORDS));
+      ir3_get_compiler_options(compiler);
+
+   /* convert VkSpecializationInfo */
+   struct nir_spirv_specialization *spec = NULL;
+   uint32_t num_spec = 0;
+   if (spec_info && spec_info->mapEntryCount) {
+      spec = malloc(sizeof(*spec) * spec_info->mapEntryCount);
+      if (!spec)
+         return NULL;
+
+      for (uint32_t i = 0; i < spec_info->mapEntryCount; i++) {
+         const VkSpecializationMapEntry *entry = &spec_info->pMapEntries[i];
+         const void *data = spec_info->pData + entry->offset;
+         assert(data + entry->size <= spec_info->pData + spec_info->dataSize);
+         spec[i].id = entry->constantID;
+         if (entry->size == 8)
+            spec[i].data64 = *(const uint64_t *) data;
+         else
+            spec[i].data32 = *(const uint32_t *) data;
+         spec[i].defined_on_module = false;
       }
-      set = MAX_SETS;
-      break;
-   default:
-      base = nir_imm_int(b, binding_layout->offset / (4 * A6XX_TEX_CONST_DWORDS));
-      break;
-   }
-
-   nir_ssa_def *shift;
 
-   if (binding_layout->type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
-      /* Inline uniform blocks cannot have arrays so the stride is unused */
-      shift = nir_imm_int(b, 0);
-   } else {
-      unsigned stride = binding_layout->size / (4 * A6XX_TEX_CONST_DWORDS);
-      assert(util_is_power_of_two_nonzero(stride));
-      shift = nir_imm_int(b, util_logbase2(stride));
+      num_spec = spec_info->mapEntryCount;
    }
 
-   nir_ssa_def *def = nir_vec3(b, nir_imm_int(b, set),
-                               nir_iadd(b, base,
-                                        nir_ishl(b, vulkan_idx, shift)),
-                               shift);
+   nir_shader *nir =
+      spirv_to_nir(words, word_count, spec, num_spec, stage, entry_point_name,
+                   &spirv_options, nir_options);
 
-   nir_ssa_def_rewrite_uses(&instr->dest.ssa, def);
-   nir_instr_remove(&instr->instr);
-}
+   free(spec);
 
-static void
-lower_vulkan_resource_reindex(nir_builder *b, nir_intrinsic_instr *instr)
-{
-   nir_ssa_def *old_index = instr->src[0].ssa;
-   nir_ssa_def *delta = instr->src[1].ssa;
-   nir_ssa_def *shift = nir_channel(b, old_index, 2);
-
-   nir_ssa_def *new_index =
-      nir_vec3(b, nir_channel(b, old_index, 0),
-               nir_iadd(b, nir_channel(b, old_index, 1),
-                        nir_ishl(b, delta, shift)),
-               shift);
-
-   nir_ssa_def_rewrite_uses(&instr->dest.ssa, new_index);
-   nir_instr_remove(&instr->instr);
-}
+   assert(nir->info.stage == stage);
+   nir_validate_shader(nir, "after spirv_to_nir");
 
-static void
-lower_load_vulkan_descriptor(nir_builder *b, nir_intrinsic_instr *intrin)
-{
-   nir_ssa_def *old_index = intrin->src[0].ssa;
-   /* Loading the descriptor happens as part of the load/store instruction so
-    * this is a no-op. We just need to turn the shift into an offset of 0.
-    */
-   nir_ssa_def *new_index =
-      nir_vec3(b, nir_channel(b, old_index, 0),
-               nir_channel(b, old_index, 1),
-               nir_imm_int(b, 0));
-   nir_ssa_def_rewrite_uses(&intrin->dest.ssa, new_index);
-   nir_instr_remove(&intrin->instr);
+   return nir;
 }
 
 static void
-lower_ssbo_ubo_intrinsic(struct tu_device *dev,
-                         nir_builder *b, nir_intrinsic_instr *intrin)
+tu_sort_variables_by_location(struct exec_list *variables)
 {
-   const nir_intrinsic_info *info = &nir_intrinsic_infos[intrin->intrinsic];
-
-   /* The bindless base is part of the instruction, which means that part of
-    * the "pointer" has to be constant. We solve this in the same way the blob
-    * does, by generating a bunch of if-statements. In the usual case where
-    * the descriptor set is constant we can skip that, though).
-    */
-
-   unsigned buffer_src;
-   if (intrin->intrinsic == nir_intrinsic_store_ssbo) {
-      /* This has the value first */
-      buffer_src = 1;
-   } else {
-      buffer_src = 0;
-   }
-
-   nir_ssa_scalar scalar_idx = nir_ssa_scalar_resolved(intrin->src[buffer_src].ssa, 0);
-   nir_ssa_def *descriptor_idx = nir_channel(b, intrin->src[buffer_src].ssa, 1);
-
-   /* For isam, we need to use the appropriate descriptor if 16-bit storage is
-    * enabled. Descriptor 0 is the 16-bit one, descriptor 1 is the 32-bit one.
-    */
-   if (dev->physical_device->info->a6xx.storage_16bit &&
-       intrin->intrinsic == nir_intrinsic_load_ssbo &&
-       (nir_intrinsic_access(intrin) & ACCESS_CAN_REORDER) &&
-       intrin->dest.ssa.bit_size > 16) {
-      descriptor_idx = nir_iadd(b, descriptor_idx, nir_imm_int(b, 1));
-   }
-
-   nir_ssa_def *results[MAX_SETS + 1] = { NULL };
-
-   if (nir_ssa_scalar_is_const(scalar_idx)) {
-      nir_ssa_def *bindless =
-         nir_bindless_resource_ir3(b, 32, descriptor_idx, .desc_set = nir_ssa_scalar_as_uint(scalar_idx));
-      nir_instr_rewrite_src_ssa(&intrin->instr, &intrin->src[buffer_src], bindless);
-      return;
-   }
-
-   nir_ssa_def *base_idx = nir_channel(b, scalar_idx.def, scalar_idx.comp);
-   for (unsigned i = 0; i < MAX_SETS + 1; i++) {
-      /* if (base_idx == i) { ... */
-      nir_if *nif = nir_push_if(b, nir_ieq_imm(b, base_idx, i));
-
-      nir_ssa_def *bindless =
-         nir_bindless_resource_ir3(b, 32, descriptor_idx, .desc_set = i);
-
-      nir_intrinsic_instr *copy =
-         nir_intrinsic_instr_create(b->shader, intrin->intrinsic);
-
-      copy->num_components = intrin->num_components;
-
-      for (unsigned src = 0; src < info->num_srcs; src++) {
-         if (src == buffer_src)
-            copy->src[src] = nir_src_for_ssa(bindless);
-         else
-            copy->src[src] = nir_src_for_ssa(intrin->src[src].ssa);
-      }
-
-      for (unsigned idx = 0; idx < info->num_indices; idx++) {
-         copy->const_index[idx] = intrin->const_index[idx];
-      }
-
-      if (info->has_dest) {
-         nir_ssa_dest_init(&copy->instr, &copy->dest,
-                           intrin->dest.ssa.num_components,
-                           intrin->dest.ssa.bit_size,
-                           NULL);
-         results[i] = &copy->dest.ssa;
+   struct exec_list sorted;
+   exec_list_make_empty(&sorted);
+
+   nir_foreach_variable_safe(var, variables)
+   {
+      exec_node_remove(&var->node);
+
+      /* insert the variable into the sorted list */
+      nir_variable *next = NULL;
+      nir_foreach_variable(tmp, &sorted)
+      {
+         if (var->data.location < tmp->data.location) {
+            next = tmp;
+            break;
+         }
       }
-
-      nir_builder_instr_insert(b, &copy->instr);
-
-      /* } else { ... */
-      nir_push_else(b, nif);
+      if (next)
+         exec_node_insert_node_before(&next->node, &var->node);
+      else
+         exec_list_push_tail(&sorted, &var->node);
    }
 
-   nir_ssa_def *result =
-      nir_ssa_undef(b, intrin->dest.ssa.num_components, intrin->dest.ssa.bit_size);
-   for (int i = MAX_SETS; i >= 0; i--) {
-      nir_pop_if(b, NULL);
-      if (info->has_dest)
-         result = nir_if_phi(b, results[i], result);
-   }
-
-   if (info->has_dest)
-      nir_ssa_def_rewrite_uses(&intrin->dest.ssa, result);
-   nir_instr_remove(&intrin->instr);
+   exec_list_move_nodes_to(&sorted, variables);
 }
 
-static nir_ssa_def *
-build_bindless(struct tu_device *dev, nir_builder *b,
-               nir_deref_instr *deref, bool is_sampler,
-               struct tu_shader *shader,
-               const struct tu_pipeline_layout *layout)
-{
-   nir_variable *var = nir_deref_instr_get_variable(deref);
-
-   unsigned set = var->data.descriptor_set;
-   unsigned binding = var->data.binding;
-   const struct tu_descriptor_set_binding_layout *bind_layout =
-      &layout->set[set].layout->binding[binding];
-
-   /* input attachments use non bindless workaround */
-   if (bind_layout->type == VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT &&
-       likely(!(dev->instance->debug_flags & TU_DEBUG_DYNAMIC))) {
-      const struct glsl_type *glsl_type = glsl_without_array(var->type);
-      uint32_t idx = var->data.index * 2;
-
-      BITSET_SET_RANGE_INSIDE_WORD(b->shader->info.textures_used, idx, (idx + bind_layout->array_size * 2) - 1);
-
-      /* D24S8 workaround: stencil of D24S8 will be sampled as uint */
-      if (glsl_get_sampler_result_type(glsl_type) == GLSL_TYPE_UINT)
-         idx += 1;
-
-      if (deref->deref_type == nir_deref_type_var)
-         return nir_imm_int(b, idx);
-
-      nir_ssa_def *arr_index = nir_ssa_for_src(b, deref->arr.index, 1);
-      return nir_iadd(b, nir_imm_int(b, idx),
-                      nir_imul_imm(b, arr_index, 2));
-   }
-
-   shader->active_desc_sets |= 1u << set;
-
-   nir_ssa_def *desc_offset;
-   unsigned descriptor_stride;
-   unsigned offset = 0;
-   /* Samplers come second in combined image/sampler descriptors, see
-      * write_combined_image_sampler_descriptor().
-      */
-   if (is_sampler && bind_layout->type ==
-         VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER) {
-      offset = 1;
-   }
-   desc_offset =
-      nir_imm_int(b, (bind_layout->offset / (4 * A6XX_TEX_CONST_DWORDS)) +
-                  offset);
-   descriptor_stride = bind_layout->size / (4 * A6XX_TEX_CONST_DWORDS);
-
-   if (deref->deref_type != nir_deref_type_var) {
-      assert(deref->deref_type == nir_deref_type_array);
-
-      nir_ssa_def *arr_index = nir_ssa_for_src(b, deref->arr.index, 1);
-      desc_offset = nir_iadd(b, desc_offset,
-                             nir_imul_imm(b, arr_index, descriptor_stride));
-   }
-
-   return nir_bindless_resource_ir3(b, 32, desc_offset, .desc_set = set);
-}
-
-static void
-lower_image_deref(struct tu_device *dev, nir_builder *b,
-                  nir_intrinsic_instr *instr, struct tu_shader *shader,
-                  const struct tu_pipeline_layout *layout)
+struct tu_shader *
+tu_shader_create(struct tu_device *dev,
+                 gl_shader_stage stage,
+                 const VkPipelineShaderStageCreateInfo *stage_info,
+                 const VkAllocationCallbacks *alloc)
 {
-   nir_deref_instr *deref = nir_src_as_deref(instr->src[0]);
-   nir_ssa_def *bindless = build_bindless(dev, b, deref, false, shader, layout);
-   nir_rewrite_image_intrinsic(instr, bindless, true);
-}
+   const struct tu_shader_module *module =
+      tu_shader_module_from_handle(stage_info->module);
+   struct tu_shader *shader;
 
-static bool
-lower_intrinsic(nir_builder *b, nir_intrinsic_instr *instr,
-                struct tu_device *dev,
-                struct tu_shader *shader,
-                const struct tu_pipeline_layout *layout)
-{
-   switch (instr->intrinsic) {
-   case nir_intrinsic_load_push_constant:
-      lower_load_push_constant(dev, b, instr, shader, layout);
-      return true;
-
-   case nir_intrinsic_load_vulkan_descriptor:
-      lower_load_vulkan_descriptor(b, instr);
-      return true;
-
-   case nir_intrinsic_vulkan_resource_index:
-      lower_vulkan_resource_index(b, instr, shader, layout);
-      return true;
-   case nir_intrinsic_vulkan_resource_reindex:
-      lower_vulkan_resource_reindex(b, instr);
-      return true;
-
-   case nir_intrinsic_load_ubo:
-   case nir_intrinsic_load_ssbo:
-   case nir_intrinsic_store_ssbo:
-   case nir_intrinsic_ssbo_atomic_add:
-   case nir_intrinsic_ssbo_atomic_imin:
-   case nir_intrinsic_ssbo_atomic_umin:
-   case nir_intrinsic_ssbo_atomic_imax:
-   case nir_intrinsic_ssbo_atomic_umax:
-   case nir_intrinsic_ssbo_atomic_and:
-   case nir_intrinsic_ssbo_atomic_or:
-   case nir_intrinsic_ssbo_atomic_xor:
-   case nir_intrinsic_ssbo_atomic_exchange:
-   case nir_intrinsic_ssbo_atomic_comp_swap:
-   case nir_intrinsic_ssbo_atomic_fadd:
-   case nir_intrinsic_ssbo_atomic_fmin:
-   case nir_intrinsic_ssbo_atomic_fmax:
-   case nir_intrinsic_ssbo_atomic_fcomp_swap:
-   case nir_intrinsic_get_ssbo_size:
-      lower_ssbo_ubo_intrinsic(dev, b, instr);
-      return true;
-
-   case nir_intrinsic_image_deref_load:
-   case nir_intrinsic_image_deref_store:
-   case nir_intrinsic_image_deref_atomic_add:
-   case nir_intrinsic_image_deref_atomic_imin:
-   case nir_intrinsic_image_deref_atomic_umin:
-   case nir_intrinsic_image_deref_atomic_imax:
-   case nir_intrinsic_image_deref_atomic_umax:
-   case nir_intrinsic_image_deref_atomic_and:
-   case nir_intrinsic_image_deref_atomic_or:
-   case nir_intrinsic_image_deref_atomic_xor:
-   case nir_intrinsic_image_deref_atomic_exchange:
-   case nir_intrinsic_image_deref_atomic_comp_swap:
-   case nir_intrinsic_image_deref_size:
-   case nir_intrinsic_image_deref_samples:
-      lower_image_deref(dev, b, instr, shader, layout);
-      return true;
+   const uint32_t max_variant_count = (stage == MESA_SHADER_VERTEX) ? 2 : 1;
+   shader = vk_zalloc2(
+      &dev->alloc, alloc,
+      sizeof(*shader) + sizeof(struct ir3_shader_variant) * max_variant_count,
+      8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+   if (!shader)
+      return NULL;
 
-   default:
-      return false;
+   /* translate SPIR-V to NIR */
+   assert(module->code_size % 4 == 0);
+   nir_shader *nir = tu_spirv_to_nir(
+      dev->compiler, (const uint32_t *) module->code, module->code_size / 4,
+      stage, stage_info->pName, stage_info->pSpecializationInfo);
+   if (!nir) {
+      vk_free2(&dev->alloc, alloc, shader);
+      return NULL;
    }
-}
 
-static void
-lower_tex_ycbcr(const struct tu_pipeline_layout *layout,
-                nir_builder *builder,
-                nir_tex_instr *tex)
-{
-   int deref_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
-   assert(deref_src_idx >= 0);
-   nir_deref_instr *deref = nir_src_as_deref(tex->src[deref_src_idx].src);
-
-   nir_variable *var = nir_deref_instr_get_variable(deref);
-   const struct tu_descriptor_set_layout *set_layout =
-      layout->set[var->data.descriptor_set].layout;
-   const struct tu_descriptor_set_binding_layout *binding =
-      &set_layout->binding[var->data.binding];
-   const struct tu_sampler_ycbcr_conversion *ycbcr_samplers =
-      tu_immutable_ycbcr_samplers(set_layout, binding);
-
-   if (!ycbcr_samplers)
-      return;
-
-   /* For the following instructions, we don't apply any change */
-   if (tex->op == nir_texop_txs ||
-       tex->op == nir_texop_query_levels ||
-       tex->op == nir_texop_lod)
-      return;
-
-   assert(tex->texture_index == 0);
-   unsigned array_index = 0;
-   if (deref->deref_type != nir_deref_type_var) {
-      assert(deref->deref_type == nir_deref_type_array);
-      if (!nir_src_is_const(deref->arr.index))
-         return;
-      array_index = nir_src_as_uint(deref->arr.index);
-      array_index = MIN2(array_index, binding->array_size - 1);
+   if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_NIR)) {
+      fprintf(stderr, "translated nir:\n");
+      nir_print_shader(nir, stderr);
    }
-   const struct tu_sampler_ycbcr_conversion *ycbcr_sampler = ycbcr_samplers + array_index;
-
-   if (ycbcr_sampler->ycbcr_model == VK_SAMPLER_YCBCR_MODEL_CONVERSION_RGB_IDENTITY)
-      return;
 
-   builder->cursor = nir_after_instr(&tex->instr);
+   /* TODO what needs to happen? */
 
-   uint8_t bits = vk_format_get_component_bits(ycbcr_sampler->format,
-                                               UTIL_FORMAT_COLORSPACE_RGB,
-                                               PIPE_SWIZZLE_X);
-
-   switch (ycbcr_sampler->format) {
-   case VK_FORMAT_G8B8G8R8_422_UNORM:
-   case VK_FORMAT_B8G8R8G8_422_UNORM:
-   case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM:
-   case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM:
-      /* util_format_get_component_bits doesn't return what we want */
-      bits = 8;
+   switch (stage) {
+   case MESA_SHADER_VERTEX:
+      tu_sort_variables_by_location(&nir->outputs);
+      break;
+   case MESA_SHADER_TESS_CTRL:
+   case MESA_SHADER_TESS_EVAL:
+   case MESA_SHADER_GEOMETRY:
+      tu_sort_variables_by_location(&nir->inputs);
+      tu_sort_variables_by_location(&nir->outputs);
+      break;
+   case MESA_SHADER_FRAGMENT:
+      tu_sort_variables_by_location(&nir->inputs);
+      break;
+   case MESA_SHADER_COMPUTE:
       break;
    default:
+      unreachable("invalid gl_shader_stage");
       break;
    }
 
-   uint32_t bpcs[3] = {bits, bits, bits}; /* TODO: use right bpc for each channel ? */
-   nir_ssa_def *result = nir_convert_ycbcr_to_rgb(builder,
-                                                  ycbcr_sampler->ycbcr_model,
-                                                  ycbcr_sampler->ycbcr_range,
-                                                  &tex->dest.ssa,
-                                                  bpcs);
-   nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, result,
-                                  result->parent_instr);
+   nir_assign_var_locations(&nir->inputs, &nir->num_inputs,
+                            ir3_glsl_type_size);
+   nir_assign_var_locations(&nir->outputs, &nir->num_outputs,
+                            ir3_glsl_type_size);
+   nir_assign_var_locations(&nir->uniforms, &nir->num_uniforms,
+                            ir3_glsl_type_size);
 
-   builder->cursor = nir_before_instr(&tex->instr);
-}
+   NIR_PASS_V(nir, nir_lower_system_values);
+   NIR_PASS_V(nir, nir_lower_frexp);
+   NIR_PASS_V(nir, nir_lower_io, nir_var_all, ir3_glsl_type_size, 0);
 
-static bool
-lower_tex(nir_builder *b, nir_tex_instr *tex, struct tu_device *dev,
-          struct tu_shader *shader, const struct tu_pipeline_layout *layout)
-{
-   lower_tex_ycbcr(layout, b, tex);
-
-   int sampler_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_deref);
-   if (sampler_src_idx >= 0) {
-      nir_deref_instr *deref = nir_src_as_deref(tex->src[sampler_src_idx].src);
-      nir_ssa_def *bindless = build_bindless(dev, b, deref, true, shader, layout);
-      nir_instr_rewrite_src(&tex->instr, &tex->src[sampler_src_idx].src,
-                            nir_src_for_ssa(bindless));
-      tex->src[sampler_src_idx].src_type = nir_tex_src_sampler_handle;
-   }
+   nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
 
-   int tex_src_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_deref);
-   if (tex_src_idx >= 0) {
-      nir_deref_instr *deref = nir_src_as_deref(tex->src[tex_src_idx].src);
-      nir_ssa_def *bindless = build_bindless(dev, b, deref, false, shader, layout);
-      nir_instr_rewrite_src(&tex->instr, &tex->src[tex_src_idx].src,
-                            nir_src_for_ssa(bindless));
-      tex->src[tex_src_idx].src_type = nir_tex_src_texture_handle;
-
-      /* for the input attachment case: */
-      if (bindless->parent_instr->type != nir_instr_type_intrinsic)
-         tex->src[tex_src_idx].src_type = nir_tex_src_texture_offset;
-   }
+   shader->ir3_shader.compiler = dev->compiler;
+   shader->ir3_shader.type = stage;
+   shader->ir3_shader.nir = nir;
 
-   return true;
+   return shader;
 }
 
-struct lower_instr_params {
-   struct tu_device *dev;
-   struct tu_shader *shader;
-   const struct tu_pipeline_layout *layout;
-};
-
-static bool
-lower_instr(nir_builder *b, nir_instr *instr, void *cb_data)
+void
+tu_shader_destroy(struct tu_device *dev,
+                  struct tu_shader *shader,
+                  const VkAllocationCallbacks *alloc)
 {
-   struct lower_instr_params *params = cb_data;
-   b->cursor = nir_before_instr(instr);
-   switch (instr->type) {
-   case nir_instr_type_tex:
-      return lower_tex(b, nir_instr_as_tex(instr), params->dev, params->shader, params->layout);
-   case nir_instr_type_intrinsic:
-      return lower_intrinsic(b, nir_instr_as_intrinsic(instr), params->dev, params->shader, params->layout);
-   default:
-      return false;
-   }
-}
+   if (shader->ir3_shader.nir)
+      ralloc_free(shader->ir3_shader.nir);
 
-/* Figure out the range of push constants that we're actually going to push to
- * the shader, and tell the backend to reserve this range when pushing UBO
- * constants.
- */
-
-static void
-gather_push_constants(nir_shader *shader, struct tu_shader *tu_shader)
-{
-   uint32_t min = UINT32_MAX, max = 0;
-   nir_foreach_function(function, shader) {
-      if (!function->impl)
-         continue;
-
-      nir_foreach_block(block, function->impl) {
-         nir_foreach_instr_safe(instr, block) {
-            if (instr->type != nir_instr_type_intrinsic)
-               continue;
-
-            nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
-            if (intrin->intrinsic != nir_intrinsic_load_push_constant)
-               continue;
-
-            uint32_t base = nir_intrinsic_base(intrin);
-            uint32_t range = nir_intrinsic_range(intrin);
-            min = MIN2(min, base);
-            max = MAX2(max, base + range);
-            break;
-         }
-      }
+   for (uint32_t i = 0; i < 1 + shader->has_binning_pass; i++) {
+      if (shader->variants[i].ir)
+         ir3_destroy(shader->variants[i].ir);
    }
 
-   if (min >= max) {
-      tu_shader->const_state.push_consts.lo = 0;
-      tu_shader->const_state.push_consts.dwords = 0;
-      return;
-   }
+   if (shader->ir3_shader.const_state.immediates)
+	   free(shader->ir3_shader.const_state.immediates);
+   if (shader->binary)
+      free(shader->binary);
+   if (shader->binning_binary)
+      free(shader->binning_binary);
 
-   /* CP_LOAD_STATE OFFSET and NUM_UNIT for SHARED_CONSTS are in units of
-    * dwords while loading regular consts is in units of vec4's.
-    * So we unify the unit here as dwords for tu_push_constant_range, then
-    * we should consider correct unit when emitting.
-    *
-    * Note there's an alignment requirement of 16 dwords on OFFSET. Expand
-    * the range and change units accordingly.
-    */
-   tu_shader->const_state.push_consts.lo = (min / 4) / 4 * 4;
-   tu_shader->const_state.push_consts.dwords =
-      align(max, 16) / 4 - tu_shader->const_state.push_consts.lo;
+   vk_free2(&dev->alloc, alloc, shader);
 }
 
-static bool
-tu_lower_io(nir_shader *shader, struct tu_device *dev,
-            struct tu_shader *tu_shader,
-            const struct tu_pipeline_layout *layout)
+void
+tu_shader_compile_options_init(
+   struct tu_shader_compile_options *options,
+   const VkGraphicsPipelineCreateInfo *pipeline_info)
 {
-   if (!tu6_shared_constants_enable(layout, dev->compiler))
-      gather_push_constants(shader, tu_shader);
-
-   struct tu_const_state *const_state = &tu_shader->const_state;
-   unsigned reserved_consts_vec4 =
-      align(DIV_ROUND_UP(const_state->push_consts.dwords, 4),
-            dev->compiler->const_upload_unit);
-
-   if (layout->independent_sets) {
-      const_state->dynamic_offset_loc = reserved_consts_vec4 * 4;
-      reserved_consts_vec4 += DIV_ROUND_UP(MAX_SETS, 4);
-   } else {
-      const_state->dynamic_offset_loc = UINT32_MAX;
-   }
-
-   tu_shader->reserved_user_consts_vec4 = reserved_consts_vec4;
+   *options = (struct tu_shader_compile_options) {
+      /* TODO ir3_key */
 
-   struct lower_instr_params params = {
-      .dev = dev,
-      .shader = tu_shader,
-      .layout = layout,
+      .optimize = !(pipeline_info->flags &
+                    VK_PIPELINE_CREATE_DISABLE_OPTIMIZATION_BIT),
+      .include_binning_pass = true,
    };
-
-   bool progress = nir_shader_instructions_pass(shader,
-                                                lower_instr,
-                                                nir_metadata_none,
-                                                &params);
-
-   /* Remove now-unused variables so that when we gather the shader info later
-    * they won't be counted.
-    */
-
-   if (progress)
-      nir_opt_dce(shader);
-
-   progress |=
-      nir_remove_dead_variables(shader,
-                                nir_var_uniform | nir_var_mem_ubo | nir_var_mem_ssbo,
-                                NULL);
-
-   return progress;
 }
 
-static void
-shared_type_info(const struct glsl_type *type, unsigned *size, unsigned *align)
+static uint32_t *
+tu_compile_shader_variant(struct ir3_shader *shader,
+                          const struct ir3_shader_key *key,
+                          bool binning_pass,
+                          struct ir3_shader_variant *variant)
 {
-   assert(glsl_type_is_vector_or_scalar(type));
-
-   unsigned comp_size =
-      glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
-   unsigned length = glsl_get_vector_elements(type);
-   *size = comp_size * length;
-   *align = comp_size;
-}
-
-static void
-tu_gather_xfb_info(nir_shader *nir, struct ir3_stream_output_info *info)
-{
-   nir_shader_gather_xfb_info(nir);
-
-   if (!nir->xfb_info)
-      return;
-
-   nir_xfb_info *xfb = nir->xfb_info;
+   variant->shader = shader;
+   variant->type = shader->type;
+   variant->key = *key;
+   variant->binning_pass = binning_pass;
 
-   uint8_t output_map[VARYING_SLOT_TESS_MAX];
-   memset(output_map, 0, sizeof(output_map));
-
-   nir_foreach_shader_out_variable(var, nir) {
-      unsigned slots =
-         var->data.compact ? DIV_ROUND_UP(glsl_get_length(var->type), 4)
-                           : glsl_count_attribute_slots(var->type, false);
-      for (unsigned i = 0; i < slots; i++)
-         output_map[var->data.location + i] = var->data.driver_location + i;
-   }
-
-   assert(xfb->output_count <= IR3_MAX_SO_OUTPUTS);
-   info->num_outputs = xfb->output_count;
-
-   for (int i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
-      info->stride[i] = xfb->buffers[i].stride / 4;
-      info->buffer_to_stream[i] = xfb->buffer_to_stream[i];
-   }
-
-   info->streams_written = xfb->streams_written;
+   int ret = ir3_compile_shader_nir(shader->compiler, variant);
+   if (ret)
+      return NULL;
 
-   for (int i = 0; i < xfb->output_count; i++) {
-      info->output[i].register_index = output_map[xfb->outputs[i].location];
-      info->output[i].start_component = xfb->outputs[i].component_offset;
-      info->output[i].num_components =
-                           util_bitcount(xfb->outputs[i].component_mask);
-      info->output[i].output_buffer  = xfb->outputs[i].buffer;
-      info->output[i].dst_offset = xfb->outputs[i].offset / 4;
-      info->output[i].stream = xfb->buffer_to_stream[xfb->outputs[i].buffer];
-   }
+   /* when assemble fails, we rely on tu_shader_destroy to clean up the
+    * variant
+    */
+   return ir3_shader_assemble(variant, shader->compiler->gpu_id);
 }
 
-struct tu_shader *
-tu_shader_create(struct tu_device *dev,
-                 nir_shader *nir,
-                 const struct tu_shader_key *key,
-                 struct tu_pipeline_layout *layout,
-                 const VkAllocationCallbacks *alloc)
+VkResult
+tu_shader_compile(struct tu_device *dev,
+                  struct tu_shader *shader,
+                  const struct tu_shader *next_stage,
+                  const struct tu_shader_compile_options *options,
+                  const VkAllocationCallbacks *alloc)
 {
-   struct tu_shader *shader;
-
-   shader = vk_zalloc2(
-      &dev->vk.alloc, alloc,
-      sizeof(*shader),
-      8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
-   if (!shader)
-      return NULL;
-
-   NIR_PASS_V(nir, nir_opt_access, &(nir_opt_access_options) {
-               .is_vulkan = true,
-             });
-
-   if (nir->info.stage == MESA_SHADER_FRAGMENT) {
-      NIR_PASS_V(nir, nir_lower_input_attachments,
-                 &(nir_input_attachment_options) {
-                     .use_fragcoord_sysval = true,
-                     .use_layer_id_sysval = false,
-                     /* When using multiview rendering, we must use
-                      * gl_ViewIndex as the layer id to pass to the texture
-                      * sampling function. gl_Layer doesn't work when
-                      * multiview is enabled.
-                      */
-                     .use_view_id_for_layer = key->multiview_mask != 0,
-                 });
+   if (options->optimize) {
+      /* ignore the key for the first pass of optimization */
+      ir3_optimize_nir(&shader->ir3_shader, shader->ir3_shader.nir, NULL);
+
+      if (unlikely(dev->physical_device->instance->debug_flags &
+                   TU_DEBUG_NIR)) {
+         fprintf(stderr, "optimized nir:\n");
+         nir_print_shader(shader->ir3_shader.nir, stderr);
+      }
    }
 
-   /* This needs to happen before multiview lowering which rewrites store
-    * instructions of the position variable, so that we can just rewrite one
-    * store at the end instead of having to rewrite every store specified by
-    * the user.
-    */
-   ir3_nir_lower_io_to_temporaries(nir);
+   shader->binary = tu_compile_shader_variant(
+      &shader->ir3_shader, &options->key, false, &shader->variants[0]);
+   if (!shader->binary)
+      return VK_ERROR_OUT_OF_HOST_MEMORY;
 
-   if (nir->info.stage == MESA_SHADER_VERTEX && key->multiview_mask) {
-      tu_nir_lower_multiview(nir, key->multiview_mask, dev);
-   }
+   /* compile another variant for the binning pass */
+   if (options->include_binning_pass &&
+       shader->ir3_shader.type == MESA_SHADER_VERTEX) {
+      shader->binning_binary = tu_compile_shader_variant(
+         &shader->ir3_shader, &options->key, true, &shader->variants[1]);
+      if (!shader->binning_binary)
+         return VK_ERROR_OUT_OF_HOST_MEMORY;
 
-   if (nir->info.stage == MESA_SHADER_FRAGMENT && key->force_sample_interp) {
-      nir_foreach_shader_in_variable(var, nir) {
-         if (!var->data.centroid)
-            var->data.sample = true;
-      }
+      shader->has_binning_pass = true;
    }
 
-   NIR_PASS_V(nir, nir_lower_explicit_io, nir_var_mem_push_const,
-              nir_address_format_32bit_offset);
-
-   NIR_PASS_V(nir, nir_lower_explicit_io,
-              nir_var_mem_ubo | nir_var_mem_ssbo,
-              nir_address_format_vec2_index_32bit_offset);
-
-   NIR_PASS_V(nir, nir_lower_explicit_io,
-              nir_var_mem_global,
-              nir_address_format_64bit_global);
-
-   if (nir->info.stage == MESA_SHADER_COMPUTE) {
-      NIR_PASS_V(nir, nir_lower_vars_to_explicit_types,
-                 nir_var_mem_shared, shared_type_info);
-      NIR_PASS_V(nir, nir_lower_explicit_io,
-                 nir_var_mem_shared,
-                 nir_address_format_32bit_offset);
-
-      if (nir->info.zero_initialize_shared_memory && nir->info.shared_size > 0) {
-         const unsigned chunk_size = 16; /* max single store size */
-         /* Shared memory is allocated in 1024b chunks in HW, but the zero-init
-          * extension only requires us to initialize the memory that the shader
-          * is allocated at the API level, and it's up to the user to ensure
-          * that accesses are limited to those bounds.
-          */
-         const unsigned shared_size = ALIGN(nir->info.shared_size, chunk_size);
-         NIR_PASS_V(nir, nir_zero_initialize_shared_memory, shared_size, chunk_size);
+   if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_IR3)) {
+      fprintf(stderr, "disassembled ir3:\n");
+      fprintf(stderr, "shader: %s\n",
+              gl_shader_stage_name(shader->ir3_shader.type));
+      ir3_shader_disasm(&shader->variants[0], shader->binary, stderr);
+
+      if (shader->has_binning_pass) {
+         fprintf(stderr, "disassembled ir3:\n");
+         fprintf(stderr, "shader: %s (binning)\n",
+                 gl_shader_stage_name(shader->ir3_shader.type));
+         ir3_shader_disasm(&shader->variants[1], shader->binning_binary,
+                           stderr);
       }
-
-      const struct nir_lower_compute_system_values_options compute_sysval_options = {
-         .has_base_workgroup_id = true,
-      };
-      NIR_PASS_V(nir, nir_lower_compute_system_values, &compute_sysval_options);
    }
 
-   nir_assign_io_var_locations(nir, nir_var_shader_in, &nir->num_inputs, nir->info.stage);
-   nir_assign_io_var_locations(nir, nir_var_shader_out, &nir->num_outputs, nir->info.stage);
+   return VK_SUCCESS;
+}
 
-  /* Gather information for transform feedback. This should be called after:
-    * - nir_split_per_member_structs.
-    * - nir_remove_dead_variables with varyings, so that we could align
-    *   stream outputs correctly.
-    * - nir_assign_io_var_locations - to have valid driver_location
-    */
-   struct ir3_stream_output_info so_info = {};
-   if (nir->info.stage == MESA_SHADER_VERTEX ||
-         nir->info.stage == MESA_SHADER_TESS_EVAL ||
-         nir->info.stage == MESA_SHADER_GEOMETRY)
-      tu_gather_xfb_info(nir, &so_info);
+VkResult
+tu_CreateShaderModule(VkDevice _device,
+                      const VkShaderModuleCreateInfo *pCreateInfo,
+                      const VkAllocationCallbacks *pAllocator,
+                      VkShaderModule *pShaderModule)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   struct tu_shader_module *module;
 
-   NIR_PASS_V(nir, tu_lower_io, dev, shader, layout);
+   assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO);
+   assert(pCreateInfo->flags == 0);
+   assert(pCreateInfo->codeSize % 4 == 0);
 
-   nir_shader_gather_info(nir, nir_shader_get_entrypoint(nir));
+   module = vk_alloc2(&device->alloc, pAllocator,
+                      sizeof(*module) + pCreateInfo->codeSize, 8,
+                      VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+   if (module == NULL)
+      return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   ir3_finalize_nir(dev->compiler, nir);
+   module->code_size = pCreateInfo->codeSize;
+   memcpy(module->code, pCreateInfo->pCode, pCreateInfo->codeSize);
 
-   bool shared_consts_enable = tu6_shared_constants_enable(layout, dev->compiler);
-   if (shared_consts_enable)
-      assert(!shader->const_state.push_consts.dwords);
+   _mesa_sha1_compute(module->code, module->code_size, module->sha1);
 
-   shader->ir3_shader =
-      ir3_shader_from_nir(dev->compiler, nir, &(struct ir3_shader_options) {
-                           .reserved_user_consts = shader->reserved_user_consts_vec4,
-                           .shared_consts_enable = shared_consts_enable,
-                           .api_wavesize = key->api_wavesize,
-                           .real_wavesize = key->real_wavesize,
-                          }, &so_info);
+   *pShaderModule = tu_shader_module_to_handle(module);
 
-   return shader;
+   return VK_SUCCESS;
 }
 
 void
-tu_shader_destroy(struct tu_device *dev,
-                  struct tu_shader *shader,
-                  const VkAllocationCallbacks *alloc)
+tu_DestroyShaderModule(VkDevice _device,
+                       VkShaderModule _module,
+                       const VkAllocationCallbacks *pAllocator)
 {
-   ir3_shader_destroy(shader->ir3_shader);
+   TU_FROM_HANDLE(tu_device, device, _device);
+   TU_FROM_HANDLE(tu_shader_module, module, _module);
+
+   if (!module)
+      return;
 
-   vk_free2(&dev->vk.alloc, alloc, shader);
+   vk_free2(&device->alloc, pAllocator, module);
 }
diff --git a/lib/mesa/src/freedreno/vulkan/tu_util.c b/lib/mesa/src/freedreno/vulkan/tu_util.c
index 9b0b9a420..e630460fb 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_util.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_util.c
@@ -1,21 +1,79 @@
 /*
  * Copyright © 2015 Intel Corporation
- * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_util.h"
+#include "tu_private.h"
 
+#include <assert.h>
 #include <errno.h>
 #include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
 
 #include "util/u_math.h"
-#include "util/timespec.h"
 #include "vk_enum_to_str.h"
 
-#include "tu_device.h"
-#include "tu_pass.h"
+/* TODO: Add Android support to tu_log funcs */
 
-void PRINTFLIKE(3, 4)
+/** Log an error message.  */
+void tu_printflike(1, 2) tu_loge(const char *format, ...)
+{
+   va_list va;
+
+   va_start(va, format);
+   tu_loge_v(format, va);
+   va_end(va);
+}
+
+/** \see tu_loge() */
+void
+tu_loge_v(const char *format, va_list va)
+{
+   fprintf(stderr, "vk: error: ");
+   vfprintf(stderr, format, va);
+   fprintf(stderr, "\n");
+}
+
+/** Log an error message.  */
+void tu_printflike(1, 2) tu_logi(const char *format, ...)
+{
+   va_list va;
+
+   va_start(va, format);
+   tu_logi_v(format, va);
+   va_end(va);
+}
+
+/** \see tu_logi() */
+void
+tu_logi_v(const char *format, va_list va)
+{
+   fprintf(stderr, "tu: info: ");
+   vfprintf(stderr, format, va);
+   fprintf(stderr, "\n");
+}
+
+void tu_printflike(3, 4)
    __tu_finishme(const char *file, int line, const char *format, ...)
 {
    va_list ap;
@@ -25,17 +83,16 @@ void PRINTFLIKE(3, 4)
    vsnprintf(buffer, sizeof(buffer), format, ap);
    va_end(ap);
 
-   mesa_loge("%s:%d: FINISHME: %s\n", file, line, buffer);
+   fprintf(stderr, "%s:%d: FINISHME: %s\n", file, line, buffer);
 }
 
 VkResult
-__vk_startup_errorf(struct tu_instance *instance,
-                    VkResult error,
-                    bool always_print,
-                    const char *file,
-                    int line,
-                    const char *format,
-                    ...)
+__vk_errorf(struct tu_instance *instance,
+            VkResult error,
+            const char *file,
+            int line,
+            const char *format,
+            ...)
 {
    va_list ap;
    char buffer[256];
@@ -43,8 +100,7 @@ __vk_startup_errorf(struct tu_instance *instance,
    const char *error_str = vk_Result_to_str(error);
 
 #ifndef DEBUG
-   if (!always_print)
-      return error;
+   return error;
 #endif
 
    if (format) {
@@ -52,236 +108,10 @@ __vk_startup_errorf(struct tu_instance *instance,
       vsnprintf(buffer, sizeof(buffer), format, ap);
       va_end(ap);
 
-      mesa_loge("%s:%d: %s (%s)\n", file, line, buffer, error_str);
+      fprintf(stderr, "%s:%d: %s (%s)\n", file, line, buffer, error_str);
    } else {
-      mesa_loge("%s:%d: %s\n", file, line, error_str);
+      fprintf(stderr, "%s:%d: %s\n", file, line, error_str);
    }
 
    return error;
 }
-
-static void
-tu_tiling_config_update_tile_layout(struct tu_framebuffer *fb,
-                                    const struct tu_device *dev,
-                                    const struct tu_render_pass *pass,
-                                    enum tu_gmem_layout gmem_layout)
-{
-   const uint32_t tile_align_w = pass->tile_align_w;
-   const uint32_t tile_align_h = dev->physical_device->info->tile_align_h;
-   const uint32_t max_tile_width = dev->physical_device->info->tile_max_w;
-   const uint32_t max_tile_height = dev->physical_device->info->tile_max_h;
-   struct tu_tiling_config *tiling = &fb->tiling[gmem_layout];
-
-   /* start from 1 tile */
-   tiling->tile_count = (VkExtent2D) {
-      .width = 1,
-      .height = 1,
-   };
-   tiling->tile0 = (VkExtent2D) {
-      .width = util_align_npot(fb->width, tile_align_w),
-      .height = align(fb->height, tile_align_h),
-   };
-
-   /* will force to sysmem, don't bother trying to have a valid tile config
-    * TODO: just skip all GMEM stuff when sysmem is forced?
-    */
-   if (!pass->gmem_pixels[gmem_layout])
-      return;
-
-   if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN)) {
-      /* start with 2x2 tiles */
-      tiling->tile_count.width = 2;
-      tiling->tile_count.height = 2;
-      tiling->tile0.width = util_align_npot(DIV_ROUND_UP(fb->width, 2), tile_align_w);
-      tiling->tile0.height = align(DIV_ROUND_UP(fb->height, 2), tile_align_h);
-   }
-
-   /* do not exceed max tile width */
-   while (tiling->tile0.width > max_tile_width) {
-      tiling->tile_count.width++;
-      tiling->tile0.width =
-         util_align_npot(DIV_ROUND_UP(fb->width, tiling->tile_count.width), tile_align_w);
-   }
-
-   /* do not exceed max tile height */
-   while (tiling->tile0.height > max_tile_height) {
-      tiling->tile_count.height++;
-      tiling->tile0.height =
-         util_align_npot(DIV_ROUND_UP(fb->height, tiling->tile_count.height), tile_align_h);
-   }
-
-   /* do not exceed gmem size */
-   while (tiling->tile0.width * tiling->tile0.height > pass->gmem_pixels[gmem_layout]) {
-      if (tiling->tile0.width > MAX2(tile_align_w, tiling->tile0.height)) {
-         tiling->tile_count.width++;
-         tiling->tile0.width =
-            util_align_npot(DIV_ROUND_UP(fb->width, tiling->tile_count.width), tile_align_w);
-      } else {
-         /* if this assert fails then layout is impossible.. */
-         assert(tiling->tile0.height > tile_align_h);
-         tiling->tile_count.height++;
-         tiling->tile0.height =
-            align(DIV_ROUND_UP(fb->height, tiling->tile_count.height), tile_align_h);
-      }
-   }
-}
-
-static void
-tu_tiling_config_update_pipe_layout(struct tu_tiling_config *tiling,
-                                    const struct tu_device *dev)
-{
-   const uint32_t max_pipe_count = 32; /* A6xx */
-
-   /* start from 1 tile per pipe */
-   tiling->pipe0 = (VkExtent2D) {
-      .width = 1,
-      .height = 1,
-   };
-   tiling->pipe_count = tiling->tile_count;
-
-   while (tiling->pipe_count.width * tiling->pipe_count.height > max_pipe_count) {
-      if (tiling->pipe0.width < tiling->pipe0.height) {
-         tiling->pipe0.width += 1;
-         tiling->pipe_count.width =
-            DIV_ROUND_UP(tiling->tile_count.width, tiling->pipe0.width);
-      } else {
-         tiling->pipe0.height += 1;
-         tiling->pipe_count.height =
-            DIV_ROUND_UP(tiling->tile_count.height, tiling->pipe0.height);
-      }
-   }
-}
-
-static void
-tu_tiling_config_update_pipes(struct tu_tiling_config *tiling,
-                              const struct tu_device *dev)
-{
-   const uint32_t max_pipe_count = 32; /* A6xx */
-   const uint32_t used_pipe_count =
-      tiling->pipe_count.width * tiling->pipe_count.height;
-   const VkExtent2D last_pipe = {
-      .width = (tiling->tile_count.width - 1) % tiling->pipe0.width + 1,
-      .height = (tiling->tile_count.height - 1) % tiling->pipe0.height + 1,
-   };
-
-   assert(used_pipe_count <= max_pipe_count);
-   assert(max_pipe_count <= ARRAY_SIZE(tiling->pipe_config));
-
-   for (uint32_t y = 0; y < tiling->pipe_count.height; y++) {
-      for (uint32_t x = 0; x < tiling->pipe_count.width; x++) {
-         const uint32_t pipe_x = tiling->pipe0.width * x;
-         const uint32_t pipe_y = tiling->pipe0.height * y;
-         const uint32_t pipe_w = (x == tiling->pipe_count.width - 1)
-                                    ? last_pipe.width
-                                    : tiling->pipe0.width;
-         const uint32_t pipe_h = (y == tiling->pipe_count.height - 1)
-                                    ? last_pipe.height
-                                    : tiling->pipe0.height;
-         const uint32_t n = tiling->pipe_count.width * y + x;
-
-         tiling->pipe_config[n] = A6XX_VSC_PIPE_CONFIG_REG_X(pipe_x) |
-                                  A6XX_VSC_PIPE_CONFIG_REG_Y(pipe_y) |
-                                  A6XX_VSC_PIPE_CONFIG_REG_W(pipe_w) |
-                                  A6XX_VSC_PIPE_CONFIG_REG_H(pipe_h);
-         tiling->pipe_sizes[n] = CP_SET_BIN_DATA5_0_VSC_SIZE(pipe_w * pipe_h);
-      }
-   }
-
-   memset(tiling->pipe_config + used_pipe_count, 0,
-          sizeof(uint32_t) * (max_pipe_count - used_pipe_count));
-}
-
-static bool
-is_hw_binning_possible(const struct tu_tiling_config *tiling)
-{
-   /* Similar to older gens, # of tiles per pipe cannot be more than 32.
-    * But there are no hangs with 16 or more tiles per pipe in either
-    * X or Y direction, so that limit does not seem to apply.
-    */
-   uint32_t tiles_per_pipe = tiling->pipe0.width * tiling->pipe0.height;
-   return tiles_per_pipe <= 32;
-}
-
-static void
-tu_tiling_config_update_binning(struct tu_tiling_config *tiling, const struct tu_device *device)
-{
-   tiling->binning_possible = is_hw_binning_possible(tiling);
-
-   if (tiling->binning_possible) {
-      tiling->binning = (tiling->tile_count.width * tiling->tile_count.height) > 2;
-
-      if (unlikely(device->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN))
-         tiling->binning = true;
-      if (unlikely(device->physical_device->instance->debug_flags &
-                   TU_DEBUG_NOBIN))
-         tiling->binning = false;
-   } else {
-      tiling->binning = false;
-   }
-}
-
-void
-tu_framebuffer_tiling_config(struct tu_framebuffer *fb,
-                             const struct tu_device *device,
-                             const struct tu_render_pass *pass)
-{
-   for (int gmem_layout = 0; gmem_layout < TU_GMEM_LAYOUT_COUNT; gmem_layout++) {
-      struct tu_tiling_config *tiling = &fb->tiling[gmem_layout];
-      tu_tiling_config_update_tile_layout(fb, device, pass, gmem_layout);
-      tu_tiling_config_update_pipe_layout(tiling, device);
-      tu_tiling_config_update_pipes(tiling, device);
-      tu_tiling_config_update_binning(tiling, device);
-   }
-}
-
-void
-tu_dbg_log_gmem_load_store_skips(struct tu_device *device)
-{
-   static uint32_t last_skipped_loads = 0;
-   static uint32_t last_skipped_stores = 0;
-   static uint32_t last_total_loads = 0;
-   static uint32_t last_total_stores = 0;
-   static struct timespec last_time = {};
-
-   pthread_mutex_lock(&device->submit_mutex);
-
-   struct timespec current_time;
-   clock_gettime(CLOCK_MONOTONIC, &current_time);
-
-   if (timespec_sub_to_nsec(&current_time, &last_time) > 1000 * 1000 * 1000) {
-      last_time = current_time;
-   } else {
-      pthread_mutex_unlock(&device->submit_mutex);
-      return;
-   }
-
-   struct tu6_global *global = device->global_bo->map;
-
-   uint32_t current_taken_loads = global->dbg_gmem_taken_loads;
-   uint32_t current_taken_stores = global->dbg_gmem_taken_stores;
-   uint32_t current_total_loads = global->dbg_gmem_total_loads;
-   uint32_t current_total_stores = global->dbg_gmem_total_stores;
-
-   uint32_t skipped_loads = current_total_loads - current_taken_loads;
-   uint32_t skipped_stores = current_total_stores - current_taken_stores;
-
-   uint32_t current_time_frame_skipped_loads = skipped_loads - last_skipped_loads;
-   uint32_t current_time_frame_skipped_stores = skipped_stores - last_skipped_stores;
-
-   uint32_t current_time_frame_total_loads = current_total_loads - last_total_loads;
-   uint32_t current_time_frame_total_stores = current_total_stores - last_total_stores;
-
-   mesa_logi("[GMEM] loads total: %u skipped: %.1f%%\n",
-         current_time_frame_total_loads,
-         current_time_frame_skipped_loads / (float) current_time_frame_total_loads * 100.f);
-   mesa_logi("[GMEM] stores total: %u skipped: %.1f%%\n",
-         current_time_frame_total_stores,
-         current_time_frame_skipped_stores / (float) current_time_frame_total_stores * 100.f);
-
-   last_skipped_loads = skipped_loads;
-   last_skipped_stores = skipped_stores;
-   last_total_loads = current_total_loads;
-   last_total_stores = current_total_stores;
-
-   pthread_mutex_unlock(&device->submit_mutex);
-}
diff --git a/lib/mesa/src/freedreno/vulkan/tu_wsi.c b/lib/mesa/src/freedreno/vulkan/tu_wsi.c
index cf09cf9b6..21466108b 100644
--- a/lib/mesa/src/freedreno/vulkan/tu_wsi.c
+++ b/lib/mesa/src/freedreno/vulkan/tu_wsi.c
@@ -1,62 +1,272 @@
 /*
  * Copyright © 2016 Red Hat
- * SPDX-License-Identifier: MIT
- *
  * based on intel anv code:
  * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
  */
 
-#include "tu_wsi.h"
+#include "tu_private.h"
 
 #include "vk_util.h"
-#include "wsi_common_drm.h"
-#include "drm-uapi/drm_fourcc.h"
-
-#include "tu_device.h"
+#include "wsi_common.h"
 
-static VKAPI_ATTR PFN_vkVoidFunction VKAPI_CALL
+static PFN_vkVoidFunction
 tu_wsi_proc_addr(VkPhysicalDevice physicalDevice, const char *pName)
 {
-   TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
-   return vk_instance_get_proc_addr_unchecked(&pdevice->instance->vk, pName);
+   return tu_lookup_entrypoint_unchecked(pName);
 }
 
-static bool
-tu_wsi_can_present_on_device(VkPhysicalDevice physicalDevice, int fd)
+VkResult
+tu_wsi_init(struct tu_physical_device *physical_device)
 {
-   TU_FROM_HANDLE(tu_physical_device, pdevice, physicalDevice);
+   return wsi_device_init(&physical_device->wsi_device,
+                          tu_physical_device_to_handle(physical_device),
+                          tu_wsi_proc_addr, &physical_device->instance->alloc,
+                          physical_device->master_fd, NULL);
+}
 
-   return wsi_common_drm_devices_equal(fd, pdevice->local_fd);
+void
+tu_wsi_finish(struct tu_physical_device *physical_device)
+{
+   wsi_device_finish(&physical_device->wsi_device,
+                     &physical_device->instance->alloc);
+}
+
+void
+tu_DestroySurfaceKHR(VkInstance _instance,
+                     VkSurfaceKHR _surface,
+                     const VkAllocationCallbacks *pAllocator)
+{
+   TU_FROM_HANDLE(tu_instance, instance, _instance);
+   ICD_FROM_HANDLE(VkIcdSurfaceBase, surface, _surface);
+
+   vk_free2(&instance->alloc, pAllocator, surface);
 }
 
 VkResult
-tu_wsi_init(struct tu_physical_device *physical_device)
+tu_GetPhysicalDeviceSurfaceSupportKHR(VkPhysicalDevice physicalDevice,
+                                      uint32_t queueFamilyIndex,
+                                      VkSurfaceKHR surface,
+                                      VkBool32 *pSupported)
 {
-   VkResult result;
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
 
-   result = wsi_device_init(&physical_device->wsi_device,
-                            tu_physical_device_to_handle(physical_device),
-                            tu_wsi_proc_addr,
-                            &physical_device->instance->vk.alloc,
-                            physical_device->master_fd,
-                            &physical_device->instance->dri_options,
-                            false);
-   if (result != VK_SUCCESS)
-      return result;
+   return wsi_common_get_surface_support(
+      &device->wsi_device, queueFamilyIndex, surface, pSupported);
+}
 
-   physical_device->wsi_device.supports_modifiers = true;
-   physical_device->wsi_device.can_present_on_device =
-      tu_wsi_can_present_on_device;
+VkResult
+tu_GetPhysicalDeviceSurfaceCapabilitiesKHR(
+   VkPhysicalDevice physicalDevice,
+   VkSurfaceKHR surface,
+   VkSurfaceCapabilitiesKHR *pSurfaceCapabilities)
+{
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
 
-   physical_device->vk.wsi_device = &physical_device->wsi_device;
+   return wsi_common_get_surface_capabilities(&device->wsi_device, surface,
+                                              pSurfaceCapabilities);
+}
 
-   return VK_SUCCESS;
+VkResult
+tu_GetPhysicalDeviceSurfaceCapabilities2KHR(
+   VkPhysicalDevice physicalDevice,
+   const VkPhysicalDeviceSurfaceInfo2KHR *pSurfaceInfo,
+   VkSurfaceCapabilities2KHR *pSurfaceCapabilities)
+{
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
+
+   return wsi_common_get_surface_capabilities2(
+      &device->wsi_device, pSurfaceInfo, pSurfaceCapabilities);
+}
+
+VkResult
+tu_GetPhysicalDeviceSurfaceCapabilities2EXT(
+   VkPhysicalDevice physicalDevice,
+   VkSurfaceKHR surface,
+   VkSurfaceCapabilities2EXT *pSurfaceCapabilities)
+{
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
+
+   return wsi_common_get_surface_capabilities2ext(
+      &device->wsi_device, surface, pSurfaceCapabilities);
+}
+
+VkResult
+tu_GetPhysicalDeviceSurfaceFormatsKHR(VkPhysicalDevice physicalDevice,
+                                      VkSurfaceKHR surface,
+                                      uint32_t *pSurfaceFormatCount,
+                                      VkSurfaceFormatKHR *pSurfaceFormats)
+{
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
+
+   return wsi_common_get_surface_formats(
+      &device->wsi_device, surface, pSurfaceFormatCount, pSurfaceFormats);
+}
+
+VkResult
+tu_GetPhysicalDeviceSurfaceFormats2KHR(
+   VkPhysicalDevice physicalDevice,
+   const VkPhysicalDeviceSurfaceInfo2KHR *pSurfaceInfo,
+   uint32_t *pSurfaceFormatCount,
+   VkSurfaceFormat2KHR *pSurfaceFormats)
+{
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
+
+   return wsi_common_get_surface_formats2(&device->wsi_device, pSurfaceInfo,
+                                          pSurfaceFormatCount,
+                                          pSurfaceFormats);
+}
+
+VkResult
+tu_GetPhysicalDeviceSurfacePresentModesKHR(VkPhysicalDevice physicalDevice,
+                                           VkSurfaceKHR surface,
+                                           uint32_t *pPresentModeCount,
+                                           VkPresentModeKHR *pPresentModes)
+{
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
+
+   return wsi_common_get_surface_present_modes(
+      &device->wsi_device, surface, pPresentModeCount, pPresentModes);
+}
+
+VkResult
+tu_CreateSwapchainKHR(VkDevice _device,
+                      const VkSwapchainCreateInfoKHR *pCreateInfo,
+                      const VkAllocationCallbacks *pAllocator,
+                      VkSwapchainKHR *pSwapchain)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   const VkAllocationCallbacks *alloc;
+   if (pAllocator)
+      alloc = pAllocator;
+   else
+      alloc = &device->alloc;
+
+   return wsi_common_create_swapchain(&device->physical_device->wsi_device,
+                                      tu_device_to_handle(device),
+                                      pCreateInfo, alloc, pSwapchain);
 }
 
 void
-tu_wsi_finish(struct tu_physical_device *physical_device)
+tu_DestroySwapchainKHR(VkDevice _device,
+                       VkSwapchainKHR swapchain,
+                       const VkAllocationCallbacks *pAllocator)
 {
-   physical_device->vk.wsi_device = NULL;
-   wsi_device_finish(&physical_device->wsi_device,
-                     &physical_device->instance->vk.alloc);
+   TU_FROM_HANDLE(tu_device, device, _device);
+   const VkAllocationCallbacks *alloc;
+
+   if (pAllocator)
+      alloc = pAllocator;
+   else
+      alloc = &device->alloc;
+
+   wsi_common_destroy_swapchain(_device, swapchain, alloc);
+}
+
+VkResult
+tu_GetSwapchainImagesKHR(VkDevice device,
+                         VkSwapchainKHR swapchain,
+                         uint32_t *pSwapchainImageCount,
+                         VkImage *pSwapchainImages)
+{
+   return wsi_common_get_images(swapchain, pSwapchainImageCount,
+                                pSwapchainImages);
+}
+
+VkResult
+tu_AcquireNextImageKHR(VkDevice device,
+                       VkSwapchainKHR swapchain,
+                       uint64_t timeout,
+                       VkSemaphore semaphore,
+                       VkFence fence,
+                       uint32_t *pImageIndex)
+{
+   VkAcquireNextImageInfoKHR acquire_info = {
+      .sType = VK_STRUCTURE_TYPE_ACQUIRE_NEXT_IMAGE_INFO_KHR,
+      .swapchain = swapchain,
+      .timeout = timeout,
+      .semaphore = semaphore,
+      .fence = fence,
+      .deviceMask = 0,
+   };
+
+   return tu_AcquireNextImage2KHR(device, &acquire_info, pImageIndex);
+}
+
+VkResult
+tu_AcquireNextImage2KHR(VkDevice _device,
+                        const VkAcquireNextImageInfoKHR *pAcquireInfo,
+                        uint32_t *pImageIndex)
+{
+   TU_FROM_HANDLE(tu_device, device, _device);
+   struct tu_physical_device *pdevice = device->physical_device;
+
+   VkResult result = wsi_common_acquire_next_image2(
+      &pdevice->wsi_device, _device, pAcquireInfo, pImageIndex);
+
+   /* TODO signal fence and semaphore */
+
+   return result;
+}
+
+VkResult
+tu_QueuePresentKHR(VkQueue _queue, const VkPresentInfoKHR *pPresentInfo)
+{
+   TU_FROM_HANDLE(tu_queue, queue, _queue);
+   return wsi_common_queue_present(
+      &queue->device->physical_device->wsi_device,
+      tu_device_to_handle(queue->device), _queue, queue->queue_family_index,
+      pPresentInfo);
+}
+
+VkResult
+tu_GetDeviceGroupPresentCapabilitiesKHR(
+   VkDevice device, VkDeviceGroupPresentCapabilitiesKHR *pCapabilities)
+{
+   memset(pCapabilities->presentMask, 0, sizeof(pCapabilities->presentMask));
+   pCapabilities->presentMask[0] = 0x1;
+   pCapabilities->modes = VK_DEVICE_GROUP_PRESENT_MODE_LOCAL_BIT_KHR;
+
+   return VK_SUCCESS;
+}
+
+VkResult
+tu_GetDeviceGroupSurfacePresentModesKHR(
+   VkDevice device,
+   VkSurfaceKHR surface,
+   VkDeviceGroupPresentModeFlagsKHR *pModes)
+{
+   *pModes = VK_DEVICE_GROUP_PRESENT_MODE_LOCAL_BIT_KHR;
+
+   return VK_SUCCESS;
+}
+
+VkResult
+tu_GetPhysicalDevicePresentRectanglesKHR(VkPhysicalDevice physicalDevice,
+                                         VkSurfaceKHR surface,
+                                         uint32_t *pRectCount,
+                                         VkRect2D *pRects)
+{
+   TU_FROM_HANDLE(tu_physical_device, device, physicalDevice);
+
+   return wsi_common_get_present_rectangles(&device->wsi_device, surface,
+                                            pRectCount, pRects);
 }
diff --git a/lib/mesa/src/gallium/drivers/freedreno/a6xx/fd6_compute.c b/lib/mesa/src/gallium/drivers/freedreno/a6xx/fd6_compute.c
index 774e1603c..2e6a7fd21 100644
--- a/lib/mesa/src/gallium/drivers/freedreno/a6xx/fd6_compute.c
+++ b/lib/mesa/src/gallium/drivers/freedreno/a6xx/fd6_compute.c
@@ -24,190 +24,190 @@
  *    Rob Clark <robclark@freedesktop.org>
  */
 
-#define FD_BO_NO_HARDPIN 1
-
 #include "pipe/p_state.h"
-#include "util/u_dump.h"
-#include "u_tracepoints.h"
 
 #include "freedreno_resource.h"
-#include "freedreno_tracepoints.h"
 
 #include "fd6_compute.h"
-#include "fd6_const.h"
 #include "fd6_context.h"
 #include "fd6_emit.h"
-#include "fd6_pack.h"
+
+struct fd6_compute_stateobj {
+	struct ir3_shader *shader;
+};
+
+
+static void *
+fd6_create_compute_state(struct pipe_context *pctx,
+		const struct pipe_compute_state *cso)
+{
+	struct fd_context *ctx = fd_context(pctx);
+
+	/* req_input_mem will only be non-zero for cl kernels (ie. clover).
+	 * This isn't a perfect test because I guess it is possible (but
+	 * uncommon) for none for the kernel parameters to be a global,
+	 * but ctx->set_global_bindings() can't fail, so this is the next
+	 * best place to fail if we need a newer version of kernel driver:
+	 */
+	if ((cso->req_input_mem > 0) &&
+			fd_device_version(ctx->dev) < FD_VERSION_BO_IOVA) {
+		return NULL;
+	}
+
+	struct ir3_compiler *compiler = ctx->screen->compiler;
+	struct fd6_compute_stateobj *so = CALLOC_STRUCT(fd6_compute_stateobj);
+	so->shader = ir3_shader_create_compute(compiler, cso, &ctx->debug, pctx->screen);
+	return so;
+}
+
+static void
+fd6_delete_compute_state(struct pipe_context *pctx, void *hwcso)
+{
+	struct fd6_compute_stateobj *so = hwcso;
+	ir3_shader_destroy(so->shader);
+	free(so);
+}
 
 /* maybe move to fd6_program? */
 static void
-cs_program_emit(struct fd_context *ctx, struct fd_ringbuffer *ring,
-                struct ir3_shader_variant *v) assert_dt
+cs_program_emit(struct fd_ringbuffer *ring, struct ir3_shader_variant *v,
+		const struct pipe_grid_info *info)
 {
-   const struct ir3_info *i = &v->info;
-   enum a6xx_threadsize thrsz = i->double_threadsize ? THREAD128 : THREAD64;
-
-   OUT_REG(ring, A6XX_HLSQ_INVALIDATE_CMD(.vs_state = true, .hs_state = true,
-                                          .ds_state = true, .gs_state = true,
-                                          .fs_state = true, .cs_state = true,
-                                          .gfx_ibo = true, .cs_ibo = true, ));
-
-   OUT_PKT4(ring, REG_A6XX_HLSQ_CS_CNTL, 1);
-   OUT_RING(ring, A6XX_HLSQ_CS_CNTL_CONSTLEN(v->constlen) |
-                     A6XX_HLSQ_CS_CNTL_ENABLED);
-
-   OUT_PKT4(ring, REG_A6XX_SP_CS_CONFIG, 2);
-   OUT_RING(ring, A6XX_SP_CS_CONFIG_ENABLED |
-                     A6XX_SP_CS_CONFIG_NIBO(ir3_shader_nibo(v)) |
-                     A6XX_SP_CS_CONFIG_NTEX(v->num_samp) |
-                     A6XX_SP_CS_CONFIG_NSAMP(v->num_samp)); /* SP_VS_CONFIG */
-   OUT_RING(ring, v->instrlen);                             /* SP_VS_INSTRLEN */
-
-   OUT_PKT4(ring, REG_A6XX_SP_CS_CTRL_REG0, 1);
-   OUT_RING(ring,
-            A6XX_SP_CS_CTRL_REG0_THREADSIZE(thrsz) |
-               A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT(i->max_reg + 1) |
-               A6XX_SP_CS_CTRL_REG0_HALFREGFOOTPRINT(i->max_half_reg + 1) |
-               COND(v->mergedregs, A6XX_SP_CS_CTRL_REG0_MERGEDREGS) |
-               A6XX_SP_CS_CTRL_REG0_BRANCHSTACK(ir3_shader_branchstack_hw(v)));
-
-   uint32_t shared_size = MAX2(((int)v->cs.req_local_mem - 1) / 1024, 1);
-   OUT_PKT4(ring, REG_A6XX_SP_CS_UNKNOWN_A9B1, 1);
-   OUT_RING(ring, A6XX_SP_CS_UNKNOWN_A9B1_SHARED_SIZE(shared_size) |
-                     A6XX_SP_CS_UNKNOWN_A9B1_UNK6);
-
-   if (ctx->screen->info->a6xx.has_lpac) {
-      OUT_PKT4(ring, REG_A6XX_HLSQ_CS_UNKNOWN_B9D0, 1);
-      OUT_RING(ring, A6XX_HLSQ_CS_UNKNOWN_B9D0_SHARED_SIZE(shared_size) |
-                        A6XX_HLSQ_CS_UNKNOWN_B9D0_UNK6);
-   }
-
-   uint32_t local_invocation_id, work_group_id;
-   local_invocation_id =
-      ir3_find_sysval_regid(v, SYSTEM_VALUE_LOCAL_INVOCATION_ID);
-   work_group_id = ir3_find_sysval_regid(v, SYSTEM_VALUE_WORKGROUP_ID);
-
-   OUT_PKT4(ring, REG_A6XX_HLSQ_CS_CNTL_0, 2);
-   OUT_RING(ring, A6XX_HLSQ_CS_CNTL_0_WGIDCONSTID(work_group_id) |
-                     A6XX_HLSQ_CS_CNTL_0_WGSIZECONSTID(regid(63, 0)) |
-                     A6XX_HLSQ_CS_CNTL_0_WGOFFSETCONSTID(regid(63, 0)) |
-                     A6XX_HLSQ_CS_CNTL_0_LOCALIDREGID(local_invocation_id));
-   OUT_RING(ring, A6XX_HLSQ_CS_CNTL_1_LINEARLOCALIDREGID(regid(63, 0)) |
-                     A6XX_HLSQ_CS_CNTL_1_THREADSIZE(thrsz));
-
-   if (ctx->screen->info->a6xx.has_lpac) {
-      OUT_PKT4(ring, REG_A6XX_SP_CS_CNTL_0, 2);
-      OUT_RING(ring, A6XX_SP_CS_CNTL_0_WGIDCONSTID(work_group_id) |
-                        A6XX_SP_CS_CNTL_0_WGSIZECONSTID(regid(63, 0)) |
-                        A6XX_SP_CS_CNTL_0_WGOFFSETCONSTID(regid(63, 0)) |
-                        A6XX_SP_CS_CNTL_0_LOCALIDREGID(local_invocation_id));
-      OUT_RING(ring, A6XX_SP_CS_CNTL_1_LINEARLOCALIDREGID(regid(63, 0)) |
-                        A6XX_SP_CS_CNTL_1_THREADSIZE(thrsz));
-   }
-
-   OUT_PKT4(ring, REG_A6XX_SP_CS_OBJ_START, 2);
-   OUT_RELOC(ring, v->bo, 0, 0, 0); /* SP_CS_OBJ_START_LO/HI */
-
-   if (v->instrlen > 0)
-      fd6_emit_shader(ctx, ring, v);
+	const struct ir3_info *i = &v->info;
+	enum a3xx_threadsize thrsz = FOUR_QUADS;
+
+	OUT_PKT4(ring, REG_A6XX_HLSQ_UPDATE_CNTL, 1);
+	OUT_RING(ring, 0xff);
+
+	unsigned constlen = align(v->constlen, 4);
+	OUT_PKT4(ring, REG_A6XX_HLSQ_CS_CNTL, 1);
+	OUT_RING(ring, A6XX_HLSQ_CS_CNTL_CONSTLEN(constlen) |
+			A6XX_HLSQ_CS_CNTL_ENABLED);
+
+	OUT_PKT4(ring, REG_A6XX_SP_CS_CONFIG, 2);
+	OUT_RING(ring, A6XX_SP_CS_CONFIG_ENABLED |
+			A6XX_SP_CS_CONFIG_NIBO(v->image_mapping.num_ibo) |
+			A6XX_SP_CS_CONFIG_NTEX(v->num_samp) |
+			A6XX_SP_CS_CONFIG_NSAMP(v->num_samp));    /* SP_VS_CONFIG */
+	OUT_RING(ring, v->instrlen);                      /* SP_VS_INSTRLEN */
+
+	OUT_PKT4(ring, REG_A6XX_SP_CS_CTRL_REG0, 1);
+	OUT_RING(ring, A6XX_SP_CS_CTRL_REG0_THREADSIZE(thrsz) |
+			A6XX_SP_CS_CTRL_REG0_FULLREGFOOTPRINT(i->max_reg + 1) |
+			A6XX_SP_CS_CTRL_REG0_MERGEDREGS |
+			A6XX_SP_CS_CTRL_REG0_BRANCHSTACK(v->branchstack) |
+			COND(v->need_pixlod, A6XX_SP_CS_CTRL_REG0_PIXLODENABLE));
+
+	OUT_PKT4(ring, REG_A6XX_SP_CS_UNKNOWN_A9B1, 1);
+	OUT_RING(ring, 0x41);
+
+	uint32_t local_invocation_id, work_group_id;
+	local_invocation_id = ir3_find_sysval_regid(v, SYSTEM_VALUE_LOCAL_INVOCATION_ID);
+	work_group_id = ir3_find_sysval_regid(v, SYSTEM_VALUE_WORK_GROUP_ID);
+
+	OUT_PKT4(ring, REG_A6XX_HLSQ_CS_CNTL_0, 2);
+	OUT_RING(ring, A6XX_HLSQ_CS_CNTL_0_WGIDCONSTID(work_group_id) |
+		A6XX_HLSQ_CS_CNTL_0_UNK0(regid(63, 0)) |
+		A6XX_HLSQ_CS_CNTL_0_UNK1(regid(63, 0)) |
+		A6XX_HLSQ_CS_CNTL_0_LOCALIDREGID(local_invocation_id));
+	OUT_RING(ring, 0x2fc);             /* HLSQ_CS_UNKNOWN_B998 */
+
+	OUT_PKT4(ring, REG_A6XX_SP_CS_OBJ_START_LO, 2);
+	OUT_RELOC(ring, v->bo, 0, 0, 0);   /* SP_CS_OBJ_START_LO/HI */
+
+	if (v->instrlen > 0)
+		fd6_emit_shader(ring, v);
 }
 
 static void
-fd6_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info) in_dt
+fd6_launch_grid(struct fd_context *ctx, const struct pipe_grid_info *info)
 {
-   struct ir3_shader_key key = {};
-   struct ir3_shader_variant *v;
-   struct fd_ringbuffer *ring = ctx->batch->draw;
-   unsigned nglobal = 0;
-
-   v = ir3_shader_variant(ir3_get_shader(ctx->compute), key, false, &ctx->debug);
-   if (!v)
-      return;
-
-   if (ctx->dirty_shader[PIPE_SHADER_COMPUTE] & FD_DIRTY_SHADER_PROG)
-      cs_program_emit(ctx, ring, v);
-
-   fd6_emit_cs_state(ctx, ring, v);
-   fd6_emit_cs_consts(v, ring, ctx, info);
-
-   u_foreach_bit (i, ctx->global_bindings.enabled_mask)
-      nglobal++;
-
-   if (nglobal > 0) {
-      /* global resources don't otherwise get an OUT_RELOC(), since
-       * the raw ptr address is emitted in ir3_emit_cs_consts().
-       * So to make the kernel aware that these buffers are referenced
-       * by the batch, emit dummy reloc's as part of a no-op packet
-       * payload:
-       */
-      OUT_PKT7(ring, CP_NOP, 2 * nglobal);
-      u_foreach_bit (i, ctx->global_bindings.enabled_mask) {
-         struct pipe_resource *prsc = ctx->global_bindings.buf[i];
-         OUT_RELOC(ring, fd_resource(prsc)->bo, 0, 0, 0);
-      }
-   }
-
-   OUT_PKT7(ring, CP_SET_MARKER, 1);
-   OUT_RING(ring, A6XX_CP_SET_MARKER_0_MODE(RM6_COMPUTE));
-
-   const unsigned *local_size =
-      info->block; // v->shader->nir->info->workgroup_size;
-   const unsigned *num_groups = info->grid;
-   /* for some reason, mesa/st doesn't set info->work_dim, so just assume 3: */
-   const unsigned work_dim = info->work_dim ? info->work_dim : 3;
-   OUT_PKT4(ring, REG_A6XX_HLSQ_CS_NDRANGE_0, 7);
-   OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_0_KERNELDIM(work_dim) |
-                  A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEX(local_size[0] - 1) |
-                  A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEY(local_size[1] - 1) |
-                  A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEZ(local_size[2] - 1));
-   OUT_RING(ring,
-            A6XX_HLSQ_CS_NDRANGE_1_GLOBALSIZE_X(local_size[0] * num_groups[0]));
-   OUT_RING(ring, 0); /* HLSQ_CS_NDRANGE_2_GLOBALOFF_X */
-   OUT_RING(ring,
-            A6XX_HLSQ_CS_NDRANGE_3_GLOBALSIZE_Y(local_size[1] * num_groups[1]));
-   OUT_RING(ring, 0); /* HLSQ_CS_NDRANGE_4_GLOBALOFF_Y */
-   OUT_RING(ring,
-            A6XX_HLSQ_CS_NDRANGE_5_GLOBALSIZE_Z(local_size[2] * num_groups[2]));
-   OUT_RING(ring, 0); /* HLSQ_CS_NDRANGE_6_GLOBALOFF_Z */
-
-   OUT_PKT4(ring, REG_A6XX_HLSQ_CS_KERNEL_GROUP_X, 3);
-   OUT_RING(ring, 1); /* HLSQ_CS_KERNEL_GROUP_X */
-   OUT_RING(ring, 1); /* HLSQ_CS_KERNEL_GROUP_Y */
-   OUT_RING(ring, 1); /* HLSQ_CS_KERNEL_GROUP_Z */
-
-   trace_grid_info(&ctx->batch->trace, ring, info);
-   trace_start_compute(&ctx->batch->trace, ring);
-
-   if (info->indirect) {
-      struct fd_resource *rsc = fd_resource(info->indirect);
-
-      OUT_PKT7(ring, CP_EXEC_CS_INDIRECT, 4);
-      OUT_RING(ring, 0x00000000);
-      OUT_RELOC(ring, rsc->bo, info->indirect_offset, 0, 0); /* ADDR_LO/HI */
-      OUT_RING(ring,
-               A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEX(local_size[0] - 1) |
-                  A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEY(local_size[1] - 1) |
-                  A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEZ(local_size[2] - 1));
-   } else {
-      OUT_PKT7(ring, CP_EXEC_CS, 4);
-      OUT_RING(ring, 0x00000000);
-      OUT_RING(ring, CP_EXEC_CS_1_NGROUPS_X(info->grid[0]));
-      OUT_RING(ring, CP_EXEC_CS_2_NGROUPS_Y(info->grid[1]));
-      OUT_RING(ring, CP_EXEC_CS_3_NGROUPS_Z(info->grid[2]));
-   }
-
-   trace_end_compute(&ctx->batch->trace, ring);
-
-   OUT_WFI5(ring);
-
-   fd6_cache_flush(ctx->batch, ring);
+	struct fd6_compute_stateobj *so = ctx->compute;
+	struct ir3_shader_key key = {};
+	struct ir3_shader_variant *v;
+	struct fd_ringbuffer *ring = ctx->batch->draw;
+	unsigned i, nglobal = 0;
+
+	fd6_emit_restore(ctx->batch, ring);
+
+	v = ir3_shader_variant(so->shader, key, false, &ctx->debug);
+	if (!v)
+		return;
+
+	if (ctx->dirty_shader[PIPE_SHADER_COMPUTE] & FD_DIRTY_SHADER_PROG)
+		cs_program_emit(ring, v, info);
+
+	fd6_emit_cs_state(ctx, ring, v);
+	ir3_emit_cs_consts(v, ring, ctx, info);
+
+	foreach_bit(i, ctx->global_bindings.enabled_mask)
+		nglobal++;
+
+	if (nglobal > 0) {
+		/* global resources don't otherwise get an OUT_RELOC(), since
+		 * the raw ptr address is emitted in ir3_emit_cs_consts().
+		 * So to make the kernel aware that these buffers are referenced
+		 * by the batch, emit dummy reloc's as part of a no-op packet
+		 * payload:
+		 */
+		OUT_PKT7(ring, CP_NOP, 2 * nglobal);
+		foreach_bit(i, ctx->global_bindings.enabled_mask) {
+			struct pipe_resource *prsc = ctx->global_bindings.buf[i];
+			OUT_RELOCW(ring, fd_resource(prsc)->bo, 0, 0, 0);
+		}
+	}
+
+	OUT_PKT7(ring, CP_SET_MARKER, 1);
+	OUT_RING(ring, A6XX_CP_SET_MARKER_0_MODE(0x8));
+
+	const unsigned *local_size = info->block; // v->shader->nir->info->cs.local_size;
+	const unsigned *num_groups = info->grid;
+	/* for some reason, mesa/st doesn't set info->work_dim, so just assume 3: */
+	const unsigned work_dim = info->work_dim ? info->work_dim : 3;
+	OUT_PKT4(ring, REG_A6XX_HLSQ_CS_NDRANGE_0, 7);
+	OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_0_KERNELDIM(work_dim) |
+		A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEX(local_size[0] - 1) |
+		A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEY(local_size[1] - 1) |
+		A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEZ(local_size[2] - 1));
+	OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_1_GLOBALSIZE_X(local_size[0] * num_groups[0]));
+	OUT_RING(ring, 0);            /* HLSQ_CS_NDRANGE_2_GLOBALOFF_X */
+	OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_3_GLOBALSIZE_Y(local_size[1] * num_groups[1]));
+	OUT_RING(ring, 0);            /* HLSQ_CS_NDRANGE_4_GLOBALOFF_Y */
+	OUT_RING(ring, A6XX_HLSQ_CS_NDRANGE_5_GLOBALSIZE_Z(local_size[2] * num_groups[2]));
+	OUT_RING(ring, 0);            /* HLSQ_CS_NDRANGE_6_GLOBALOFF_Z */
+
+	OUT_PKT4(ring, REG_A6XX_HLSQ_CS_KERNEL_GROUP_X, 3);
+	OUT_RING(ring, 1);            /* HLSQ_CS_KERNEL_GROUP_X */
+	OUT_RING(ring, 1);            /* HLSQ_CS_KERNEL_GROUP_Y */
+	OUT_RING(ring, 1);            /* HLSQ_CS_KERNEL_GROUP_Z */
+
+	if (info->indirect) {
+		struct fd_resource *rsc = fd_resource(info->indirect);
+
+		OUT_PKT7(ring, CP_EXEC_CS_INDIRECT, 4);
+		OUT_RING(ring, 0x00000000);
+		OUT_RELOC(ring, rsc->bo, info->indirect_offset, 0, 0);  /* ADDR_LO/HI */
+		OUT_RING(ring, A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEX(local_size[0] - 1) |
+				A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEY(local_size[1] - 1) |
+				A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEZ(local_size[2] - 1));
+	} else {
+		OUT_PKT7(ring, CP_EXEC_CS, 4);
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, CP_EXEC_CS_1_NGROUPS_X(info->grid[0]));
+		OUT_RING(ring, CP_EXEC_CS_2_NGROUPS_Y(info->grid[1]));
+		OUT_RING(ring, CP_EXEC_CS_3_NGROUPS_Z(info->grid[2]));
+	}
+
+	OUT_WFI5(ring);
+
+	fd6_cache_flush(ctx->batch, ring);
 }
 
 void
-fd6_compute_init(struct pipe_context *pctx) disable_thread_safety_analysis
+fd6_compute_init(struct pipe_context *pctx)
 {
-   struct fd_context *ctx = fd_context(pctx);
-   ctx->launch_grid = fd6_launch_grid;
-   pctx->create_compute_state = ir3_shader_compute_state_create;
-   pctx->delete_compute_state = ir3_shader_state_delete;
+	struct fd_context *ctx = fd_context(pctx);
+	ctx->launch_grid = fd6_launch_grid;
+	pctx->create_compute_state = fd6_create_compute_state;
+	pctx->delete_compute_state = fd6_delete_compute_state;
 }