i965: Amalgamate surface binding tables

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
(cherry picked from commit 3cc74044ce3546cc7dc2e918cbabbb41a77f4026)

Signed-off-by: Owain G. Ainsworth <oga@openbsd.org>

7 files changed, 265 insertions, 286 deletions

diff --git a/src/i830_render.c b/src/i830_render.c
index 77aab4eb..709cf79d 100644
--- a/src/i830_render.c
+++ b/src/i830_render.c
@@ -860,9 +860,7 @@ i830_composite(PixmapPtr dest, int srcX, int srcY, int maskX, int maskY,
 	intel_batch_end_atomic(scrn);
 }
 
-void i830_batch_flush_notify(ScrnInfoPtr scrn)
+void i830_batch_commit_notify(intel_screen_private *intel)
 {
-	intel_screen_private *intel = intel_get_screen_private(scrn);
-
 	intel->needs_render_state_emit = TRUE;
 }
diff --git a/src/i915_render.c b/src/i915_render.c
index 866d0b8f..0bf6496a 100644
--- a/src/i915_render.c
+++ b/src/i915_render.c
@@ -1189,10 +1189,8 @@ i915_vertex_flush(intel_screen_private *intel)
 }
 
 void
-i915_batch_flush_notify(ScrnInfoPtr scrn)
+i915_batch_commit_notify(intel_screen_private *intel)
 {
-	intel_screen_private *intel = intel_get_screen_private(scrn);
-
 	intel->needs_render_state_emit = TRUE;
 	intel->render_current_dest = NULL;
 	intel->last_floats_per_vertex = 0;
diff --git a/src/i965_render.c b/src/i965_render.c
index acf807f7..61574922 100644
--- a/src/i965_render.c
+++ b/src/i965_render.c
@@ -303,29 +303,6 @@ i965_check_composite_texture(ScreenPtr screen, PicturePtr picture)
 #define URB_SF_ENTRY_SIZE     2
 #define URB_SF_ENTRIES	      1
 
-static const uint32_t sip_kernel_static[][4] = {
-/*    wait (1) a0<1>UW a145<0,1,0>UW { align1 +  } */
-	{0x00000030, 0x20000108, 0x00001220, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-/*    nop (4) g0<1>UD { align1 +  } */
-	{0x0040007e, 0x20000c21, 0x00690000, 0x00000000},
-};
-
 /*
  * this program computes dA/dx and dA/dy for the texture coordinates along
  * with the base texture coordinate. It was extracted from the Mesa driver
@@ -686,8 +663,6 @@ typedef struct brw_surface_state_padded {
 	char pad[32 - sizeof(struct brw_surface_state)];
 } brw_surface_state_padded;
 
-#define PS_BINDING_TABLE_OFFSET	(3 * sizeof(struct brw_surface_state_padded))
-
 struct gen4_cc_unit_state {
 	/* Index by [src_blend][dst_blend] */
 	brw_cc_unit_state_padded cc_state[BRW_BLENDFACTOR_COUNT]
@@ -696,7 +671,6 @@ struct gen4_cc_unit_state {
 
 typedef struct gen4_composite_op {
 	int op;
-	drm_intel_bo *surface_state_binding_table_bo;
 	sampler_state_filter_t src_filter;
 	sampler_state_filter_t mask_filter;
 	sampler_state_extend_t src_extend;
@@ -718,8 +692,6 @@ struct gen4_render_state {
 	    [SAMPLER_STATE_EXTEND_COUNT];
 	drm_intel_bo *wm_kernel_bo[WM_KERNEL_COUNT];
 
-	drm_intel_bo *sip_kernel_bo;
-
 	drm_intel_bo *cc_vp_bo;
 	drm_intel_bo *gen6_blend_bo;
 	drm_intel_bo *gen6_depth_stencil_bo;
@@ -1140,55 +1112,40 @@ static sampler_state_extend_t sampler_state_extend_from_picture(int repeat_type)
  * Sets up the common fields for a surface state buffer for the given
  * picture in the given surface state buffer.
  */
-static void
+static int
 i965_set_picture_surface_state(intel_screen_private *intel,
-			       dri_bo * ss_bo, int ss_index,
 			       PicturePtr picture, PixmapPtr pixmap,
 			       Bool is_dst)
 {
-	struct brw_surface_state_padded *ss;
-	struct brw_surface_state local_ss;
 	struct intel_pixmap *priv = intel_get_pixmap_private(pixmap);
+	struct brw_surface_state *ss;
+	int offset;
 
-	ss = (struct brw_surface_state_padded *)ss_bo->virtual + ss_index;
+	ss = (struct brw_surface_state *)
+		(intel->surface_data + intel->surface_used);
 
-	/* Since ss is a pointer to WC memory, do all of our bit operations
-	 * into a local temporary first.
-	 */
-	memset(&local_ss, 0, sizeof(local_ss));
-	local_ss.ss0.surface_type = BRW_SURFACE_2D;
+	memset(ss, 0, sizeof(*ss));
+	ss->ss0.surface_type = BRW_SURFACE_2D;
 	if (is_dst) {
 		uint32_t dst_format = 0;
-		Bool ret = TRUE;
+		Bool ret;
 
 		ret = i965_get_dest_format(picture, &dst_format);
 		assert(ret == TRUE);
-		local_ss.ss0.surface_format = dst_format;
+		ss->ss0.surface_format = dst_format;
 	} else {
-		local_ss.ss0.surface_format = i965_get_card_format(picture);
+		ss->ss0.surface_format = i965_get_card_format(picture);
 	}
 
-	local_ss.ss0.data_return_format = BRW_SURFACERETURNFORMAT_FLOAT32;
-	local_ss.ss0.writedisable_alpha = 0;
-	local_ss.ss0.writedisable_red = 0;
-	local_ss.ss0.writedisable_green = 0;
-	local_ss.ss0.writedisable_blue = 0;
-	local_ss.ss0.color_blend = 1;
-	local_ss.ss0.vert_line_stride = 0;
-	local_ss.ss0.vert_line_stride_ofs = 0;
-	local_ss.ss0.mipmap_layout_mode = 0;
-	local_ss.ss0.render_cache_read_mode = 0;
-	local_ss.ss1.base_addr = priv->bo->offset;
-
-	local_ss.ss2.mip_count = 0;
-	local_ss.ss2.render_target_rotation = 0;
-	local_ss.ss2.height = pixmap->drawable.height - 1;
-	local_ss.ss2.width = pixmap->drawable.width - 1;
-	local_ss.ss3.pitch = intel_pixmap_pitch(pixmap) - 1;
-	local_ss.ss3.tile_walk = 0;	/* Tiled X */
-	local_ss.ss3.tiled_surface = intel_pixmap_tiled(pixmap) ? 1 : 0;
-
-	memcpy(ss, &local_ss, sizeof(local_ss));
+	ss->ss0.data_return_format = BRW_SURFACERETURNFORMAT_FLOAT32;
+	ss->ss0.color_blend = 1;
+	ss->ss1.base_addr = priv->bo->offset;
+
+	ss->ss2.height = pixmap->drawable.height - 1;
+	ss->ss2.width = pixmap->drawable.width - 1;
+	ss->ss3.pitch = intel_pixmap_pitch(pixmap) - 1;
+	ss->ss3.tile_walk = 0;	/* Tiled X */
+	ss->ss3.tiled_surface = intel_pixmap_tiled(pixmap) ? 1 : 0;
 
 	if (priv->bo != NULL) {
 		uint32_t write_domain, read_domains;
@@ -1202,12 +1159,18 @@ i965_set_picture_surface_state(intel_screen_private *intel,
 		}
 
 		intel_batch_mark_pixmap_domains(intel, priv, read_domains, write_domain);
-		dri_bo_emit_reloc(ss_bo, read_domains, write_domain,
+		dri_bo_emit_reloc(intel->surface_bo,
+				  read_domains, write_domain,
 				  0,
-				  ss_index * sizeof(*ss) +
+				  intel->surface_used +
 				  offsetof(struct brw_surface_state, ss1),
 				  priv->bo);
 	}
+
+	offset = intel->surface_used;
+	intel->surface_used += sizeof(struct brw_surface_state_padded);
+
+	return offset;
 }
 
 static void i965_emit_composite_state(ScrnInfoPtr scrn)
@@ -1225,71 +1188,45 @@ static void i965_emit_composite_state(ScrnInfoPtr scrn)
 	sampler_state_extend_t src_extend = composite_op->src_extend;
 	sampler_state_extend_t mask_extend = composite_op->mask_extend;
 	Bool is_affine = composite_op->is_affine;
-	int urb_vs_start, urb_vs_size;
-	int urb_gs_start, urb_gs_size;
-	int urb_clip_start, urb_clip_size;
-	int urb_sf_start, urb_sf_size;
-	int urb_cs_start, urb_cs_size;
 	uint32_t src_blend, dst_blend;
-	dri_bo *surface_state_binding_table_bo = composite_op->surface_state_binding_table_bo;
 
 	intel->needs_render_state_emit = FALSE;
 
-	IntelEmitInvarientState(scrn);
-	intel->last_3d = LAST_3D_RENDER;
-
-	/* Mark the destination dirty within this batch */
-	intel_batch_mark_pixmap_domains(intel,
-					intel_get_pixmap_private(dest),
-					I915_GEM_DOMAIN_RENDER,
-					I915_GEM_DOMAIN_RENDER);
-
-	urb_vs_start = 0;
-	urb_vs_size = URB_VS_ENTRIES * URB_VS_ENTRY_SIZE;
-	urb_gs_start = urb_vs_start + urb_vs_size;
-	urb_gs_size = URB_GS_ENTRIES * URB_GS_ENTRY_SIZE;
-	urb_clip_start = urb_gs_start + urb_gs_size;
-	urb_clip_size = URB_CLIP_ENTRIES * URB_CLIP_ENTRY_SIZE;
-	urb_sf_start = urb_clip_start + urb_clip_size;
-	urb_sf_size = URB_SF_ENTRIES * URB_SF_ENTRY_SIZE;
-	urb_cs_start = urb_sf_start + urb_sf_size;
-	urb_cs_size = URB_CS_ENTRIES * URB_CS_ENTRY_SIZE;
-
-	i965_get_blend_cntl(op, mask_picture, dest_picture->format,
-			    &src_blend, &dst_blend);
-
 	/* Begin the long sequence of commands needed to set up the 3D
 	 * rendering pipe
 	 */
 
-	/* URB fence. Erratum (Vol 1a, p32): URB_FENCE must not cross a
-	 * cache-line (64 bytes). Start by aligning this sequence of ops to
-	 * a cache-line...
-	 */
-	ALIGN_BATCH(64);
+	if (intel->needs_3d_invariant) {
+		if (IS_GEN5(intel)) {
+			/* Ironlake errata workaround: Before disabling the clipper,
+			 * you have to MI_FLUSH to get the pipeline idle.
+			 */
+			OUT_BATCH(MI_FLUSH | MI_INHIBIT_RENDER_CACHE_FLUSH);
+		}
 
-	assert(intel->in_batch_atomic);
-	{
 		/* Match Mesa driver setup */
-		OUT_BATCH(MI_FLUSH |
-			  MI_STATE_INSTRUCTION_CACHE_FLUSH |
-			  BRW_MI_GLOBAL_SNAPSHOT_RESET);
 		if (INTEL_INFO(intel)->gen >= 45)
 			OUT_BATCH(NEW_PIPELINE_SELECT | PIPELINE_SELECT_3D);
 		else
 			OUT_BATCH(BRW_PIPELINE_SELECT | PIPELINE_SELECT_3D);
 
-		OUT_BATCH(BRW_CS_URB_STATE | 0);
-		OUT_BATCH((0 << 4) |	/* URB Entry Allocation Size */
-			  (0 << 0));	/* Number of URB Entries */
+		/* Set system instruction pointer */
+		OUT_BATCH(BRW_STATE_SIP | 0);
+		OUT_BATCH(0);
+
+		intel->needs_3d_invariant = FALSE;
+	}
 
+	if (intel->surface_reloc == 0) {
 		/* Zero out the two base address registers so all offsets are
 		 * absolute.
 		 */
 		if (IS_GEN5(intel)) {
 			OUT_BATCH(BRW_STATE_BASE_ADDRESS | 6);
 			OUT_BATCH(0 | BASE_ADDRESS_MODIFY);	/* Generate state base address */
-			OUT_RELOC(surface_state_binding_table_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, BASE_ADDRESS_MODIFY); /* Surface state base address */
+			intel->surface_reloc = intel->batch_used;
+			intel_batch_emit_dword(intel,
+					       intel->surface_bo->offset | BASE_ADDRESS_MODIFY);
 			OUT_BATCH(0 | BASE_ADDRESS_MODIFY);	/* media base addr, don't care */
 			OUT_BATCH(0 | BASE_ADDRESS_MODIFY);	/* Instruction base address */
 			/* general state max addr, disabled */
@@ -1301,42 +1238,21 @@ static void i965_emit_composite_state(ScrnInfoPtr scrn)
 		} else {
 			OUT_BATCH(BRW_STATE_BASE_ADDRESS | 4);
 			OUT_BATCH(0 | BASE_ADDRESS_MODIFY);	/* Generate state base address */
-			OUT_RELOC(surface_state_binding_table_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, BASE_ADDRESS_MODIFY); /* Surface state base address */
+			intel->surface_reloc = intel->batch_used;
+			intel_batch_emit_dword(intel,
+					       intel->surface_bo->offset | BASE_ADDRESS_MODIFY);
 			OUT_BATCH(0 | BASE_ADDRESS_MODIFY);	/* media base addr, don't care */
 			/* general state max addr, disabled */
 			OUT_BATCH(0 | BASE_ADDRESS_MODIFY);
 			/* media object state max addr, disabled */
 			OUT_BATCH(0 | BASE_ADDRESS_MODIFY);
 		}
-		/* Set system instruction pointer */
-		OUT_BATCH(BRW_STATE_SIP | 0);
-		OUT_RELOC(render_state->sip_kernel_bo,
-			  I915_GEM_DOMAIN_INSTRUCTION, 0, 0);
 	}
 
-	if (IS_GEN5(intel)) {
-		/* Ironlake errata workaround: Before disabling the clipper,
-		 * you have to MI_FLUSH to get the pipeline idle.
-		 */
-		OUT_BATCH(MI_FLUSH);
-	}
+	i965_get_blend_cntl(op, mask_picture, dest_picture->format,
+			    &src_blend, &dst_blend);
 
 	{
-		int pipe_ctrl;
-		/* Pipe control */
-
-		if (IS_GEN5(intel))
-			pipe_ctrl = BRW_PIPE_CONTROL_NOWRITE;
-		else
-			pipe_ctrl =
-			    BRW_PIPE_CONTROL_NOWRITE |
-			    BRW_PIPE_CONTROL_IS_FLUSH;
-
-		OUT_BATCH(BRW_PIPE_CONTROL | pipe_ctrl | 2);
-		OUT_BATCH(0);	/* Destination address */
-		OUT_BATCH(0);	/* Immediate data low DW */
-		OUT_BATCH(0);	/* Immediate data high DW */
-
 		/* Binding table pointers */
 		OUT_BATCH(BRW_3DSTATE_BINDING_TABLE_POINTERS | 4);
 		OUT_BATCH(0);	/* vs */
@@ -1344,14 +1260,15 @@ static void i965_emit_composite_state(ScrnInfoPtr scrn)
 		OUT_BATCH(0);	/* clip */
 		OUT_BATCH(0);	/* sf */
 		/* Only the PS uses the binding table */
-		OUT_BATCH(PS_BINDING_TABLE_OFFSET);
+		OUT_BATCH(intel->surface_table);
 
 		/* The drawing rectangle clipping is always on.  Set it to values that
 		 * shouldn't do any clipping.
 		 */
 		OUT_BATCH(BRW_3DSTATE_DRAWING_RECTANGLE | 2);
 		OUT_BATCH(0x00000000);	/* ymin, xmin */
-		OUT_BATCH(DRAW_YMAX(dest->drawable.height - 1) | DRAW_XMAX(dest->drawable.width - 1));	/* ymax, xmax */
+		OUT_BATCH(DRAW_YMAX(dest->drawable.height - 1) |
+			  DRAW_XMAX(dest->drawable.width - 1));	/* ymax, xmax */
 		OUT_BATCH(0x00000000);	/* yorigin, xorigin */
 
 		/* skip the depth buffer */
@@ -1382,42 +1299,54 @@ static void i965_emit_composite_state(ScrnInfoPtr scrn)
 			  I915_GEM_DOMAIN_INSTRUCTION, 0,
 			  offsetof(struct gen4_cc_unit_state,
 				   cc_state[src_blend][dst_blend]));
+	}
 
-		/* URB fence. Erratum (Vol 1a, p32): URB_FENCE must not cross a
-		 * cache-line (64 bytes).
-		 *
-		 * 21 preceding dwords since start of section: 84 bytes.
-		 * 12 bytes for URB_FENCE, implies that the end-of-instruction
-		 * does not cross the cache-line boundary...
-		 *
-		 * A total of 33 or 35 dwords since alignment: 132, 140 bytes.
-		 * Again, the URB_FENCE will not cross a cache-line.
+	{
+		int urb_vs_start, urb_vs_size;
+		int urb_gs_start, urb_gs_size;
+		int urb_clip_start, urb_clip_size;
+		int urb_sf_start, urb_sf_size;
+		int urb_cs_start, urb_cs_size;
+
+		urb_vs_start = 0;
+		urb_vs_size = URB_VS_ENTRIES * URB_VS_ENTRY_SIZE;
+		urb_gs_start = urb_vs_start + urb_vs_size;
+		urb_gs_size = URB_GS_ENTRIES * URB_GS_ENTRY_SIZE;
+		urb_clip_start = urb_gs_start + urb_gs_size;
+		urb_clip_size = URB_CLIP_ENTRIES * URB_CLIP_ENTRY_SIZE;
+		urb_sf_start = urb_clip_start + urb_clip_size;
+		urb_sf_size = URB_SF_ENTRIES * URB_SF_ENTRY_SIZE;
+		urb_cs_start = urb_sf_start + urb_sf_size;
+		urb_cs_size = URB_CS_ENTRIES * URB_CS_ENTRY_SIZE;
+
+		/* Erratum (Vol 1a, p32):
+		 *   URB_FENCE must not cross a cache-line (64 bytes).
 		 */
+		if ((intel->batch_used & 15) > (16 - 3)) {
+			int cnt = 16 - (intel->batch_used & 15);
+			while (cnt--)
+				OUT_BATCH(MI_NOOP);
+		}
+
 		OUT_BATCH(BRW_URB_FENCE |
 			  UF0_CS_REALLOC |
 			  UF0_SF_REALLOC |
 			  UF0_CLIP_REALLOC |
-			  UF0_GS_REALLOC | UF0_VS_REALLOC | 1);
-		OUT_BATCH(((urb_clip_start +
-			    urb_clip_size) << UF1_CLIP_FENCE_SHIFT) |
-			  ((urb_gs_start +
-			    urb_gs_size) << UF1_GS_FENCE_SHIFT) | ((urb_vs_start
-								    +
-								    urb_vs_size)
-								   <<
-								   UF1_VS_FENCE_SHIFT));
-		OUT_BATCH(((urb_cs_start +
-			    urb_cs_size) << UF2_CS_FENCE_SHIFT) | ((urb_sf_start
-								    +
-								    urb_sf_size)
-								   <<
-								   UF2_SF_FENCE_SHIFT));
+			  UF0_GS_REALLOC |
+			  UF0_VS_REALLOC |
+			  1);
+		OUT_BATCH(((urb_clip_start + urb_clip_size) << UF1_CLIP_FENCE_SHIFT) |
+			  ((urb_gs_start + urb_gs_size) << UF1_GS_FENCE_SHIFT) |
+			  ((urb_vs_start + urb_vs_size) << UF1_VS_FENCE_SHIFT));
+		OUT_BATCH(((urb_cs_start + urb_cs_size) << UF2_CS_FENCE_SHIFT) |
+			  ((urb_sf_start + urb_sf_size) << UF2_SF_FENCE_SHIFT));
 
 		/* Constant buffer state */
 		OUT_BATCH(BRW_CS_URB_STATE | 0);
 		OUT_BATCH(((URB_CS_ENTRY_SIZE - 1) << 4) |
 			  (URB_CS_ENTRIES << 0));
 	}
+
 	{
 		/*
 		 * number of extra parameters per vertex
@@ -1546,7 +1475,7 @@ static Bool i965_composite_check_aperture(ScrnInfoPtr scrn)
 	drm_intel_bo *bo_table[] = {
 		intel->batch_bo,
 		intel->vertex_bo,
-		composite_op->surface_state_binding_table_bo,
+		intel->surface_bo,
 		render_state->vs_state_bo,
 		render_state->sf_state_bo,
 		render_state->sf_mask_state_bo,
@@ -1556,12 +1485,11 @@ static Bool i965_composite_check_aperture(ScrnInfoPtr scrn)
 		    [composite_op->mask_filter]
 		    [composite_op->mask_extend],
 		render_state->cc_state_bo,
-		render_state->sip_kernel_bo,
 	};
 	drm_intel_bo *gen6_bo_table[] = {
 		intel->batch_bo,
 		intel->vertex_bo,
-		composite_op->surface_state_binding_table_bo,
+		intel->surface_bo,
 		render_state->wm_kernel_bo[composite_op->wm_kernel],
 		render_state->ps_sampler_state_bo[composite_op->src_filter]
 		    [composite_op->src_extend]
@@ -1581,6 +1509,26 @@ static Bool i965_composite_check_aperture(ScrnInfoPtr scrn)
 							ARRAY_SIZE(bo_table)) == 0;
 }
 
+static void i965_surface_flush(struct intel_screen_private *intel)
+{
+	drm_intel_bo_subdata(intel->surface_bo,
+			     0, intel->surface_used,
+			     intel->surface_data);
+	intel->surface_used = 0;
+
+	assert (intel->surface_reloc != 0);
+	drm_intel_bo_emit_reloc(intel->batch_bo,
+				intel->surface_reloc * 4,
+				intel->surface_bo, BASE_ADDRESS_MODIFY,
+				I915_GEM_DOMAIN_INSTRUCTION, 0);
+	intel->surface_reloc = 0;
+
+	drm_intel_bo_unreference(intel->surface_bo);
+	intel->surface_bo =
+		drm_intel_bo_alloc(intel->bufmgr, "surface data",
+				   sizeof(intel->surface_data), 4096);
+}
+
 Bool
 i965_prepare_composite(int op, PicturePtr source_picture,
 		       PicturePtr mask_picture, PicturePtr dest_picture,
@@ -1590,8 +1538,6 @@ i965_prepare_composite(int op, PicturePtr source_picture,
 	intel_screen_private *intel = intel_get_screen_private(scrn);
 	struct gen4_render_state *render_state = intel->gen4_render_state;
 	gen4_composite_op *composite_op = &render_state->composite_op;
-	uint32_t *binding_table;
-	drm_intel_bo *surface_state_binding_table_bo;
 
 	composite_op->src_filter =
 	    sampler_state_filter_from_picture(source_picture->filter);
@@ -1649,39 +1595,6 @@ i965_prepare_composite(int op, PicturePtr source_picture,
 	    (mask && intel_pixmap_is_dirty(mask)))
 		intel_batch_emit_flush(scrn);
 
-
-	/* Set up the surface states. */
-	surface_state_binding_table_bo = dri_bo_alloc(intel->bufmgr, "surface_state",
-					3 * (sizeof(struct brw_surface_state_padded) + sizeof(uint32_t)),
-					4096);
-	if (dri_bo_map(surface_state_binding_table_bo, 1) != 0) {
-		dri_bo_unreference(surface_state_binding_table_bo);
-		return FALSE;
-	}
-	/* Set up the state buffer for the destination surface */
-	i965_set_picture_surface_state(intel, surface_state_binding_table_bo, 0,
-				       dest_picture, dest, TRUE);
-	/* Set up the source surface state buffer */
-	i965_set_picture_surface_state(intel, surface_state_binding_table_bo, 1,
-				       source_picture, source, FALSE);
-	if (mask) {
-		/* Set up the mask surface state buffer */
-		i965_set_picture_surface_state(intel, surface_state_binding_table_bo, 2,
-					       mask_picture, mask, FALSE);
-	}
-
-	/* Set up the binding table of surface indices to surface state. */
-	binding_table = (uint32_t *)((char *)surface_state_binding_table_bo->virtual + PS_BINDING_TABLE_OFFSET);
-	binding_table[0] = 0;
-	binding_table[1] = sizeof(struct brw_surface_state_padded);
-
-	if (mask) {
-		binding_table[2] = 2 * sizeof(struct brw_surface_state_padded);
-	} else {
-		binding_table[2] = 0;
-	}
-	dri_bo_unmap(surface_state_binding_table_bo);
-
 	composite_op->op = op;
 	intel->render_source_picture = source_picture;
 	intel->render_mask_picture = mask_picture;
@@ -1689,8 +1602,6 @@ i965_prepare_composite(int op, PicturePtr source_picture,
 	intel->render_source = source;
 	intel->render_mask = mask;
 	intel->render_dest = dest;
-	drm_intel_bo_unreference(composite_op->surface_state_binding_table_bo);
-	composite_op->surface_state_binding_table_bo = surface_state_binding_table_bo;
 
 	intel->scale_units[0][0] = 1. / source->drawable.width;
 	intel->scale_units[0][1] = 1. / source->drawable.height;
@@ -1756,6 +1667,10 @@ i965_prepare_composite(int op, PicturePtr source_picture,
 		}
 	}
 
+	if (sizeof(intel->surface_data) - intel->surface_used <
+	    4 * sizeof(struct brw_surface_state_padded))
+		i965_surface_flush(intel);
+
 	intel->needs_render_state_emit = TRUE;
 
 	return TRUE;
@@ -1792,6 +1707,35 @@ static void i965_select_vertex_buffer(struct intel_screen_private *intel)
 	intel->last_floats_per_vertex = vertex_size;
 }
 
+static void i965_bind_surfaces(struct intel_screen_private *intel)
+{
+	uint32_t *binding_table;
+
+	assert(intel->surface_used + 4 * sizeof(struct brw_surface_state_padded) <
+	       sizeof(intel->surface_data));
+
+	binding_table = (uint32_t*) (intel->surface_data + intel->surface_used);
+	intel->surface_table = intel->surface_used;
+	intel->surface_used += sizeof(struct brw_surface_state_padded);
+
+	binding_table[0] =
+		i965_set_picture_surface_state(intel,
+					       intel->render_dest_picture,
+					       intel->render_dest,
+					       TRUE);
+	binding_table[1] =
+		i965_set_picture_surface_state(intel,
+					       intel->render_source_picture,
+					       intel->render_source,
+					       FALSE);
+	if (intel->render_mask) {
+		binding_table[2] =
+			i965_set_picture_surface_state(intel,
+						       intel->render_mask_picture,
+						       intel->render_mask,
+						       FALSE);
+	}
+}
 
 void
 i965_composite(PixmapPtr dest, int srcX, int srcY, int maskX, int maskY,
@@ -1880,12 +1824,12 @@ i965_composite(PixmapPtr dest, int srcX, int srcY, int maskX, int maskY,
 
 	intel_batch_start_atomic(scrn, 200);
 	if (intel->needs_render_state_emit) {
+		i965_bind_surfaces(intel);
+
 		if (INTEL_INFO(intel)->gen >= 60)
 			gen6_emit_composite_state(scrn);
 		else
 			i965_emit_composite_state(scrn);
-	} else {
-		OUT_BATCH(MI_FLUSH);
 	}
 
 	if (intel->vertex_used &&
@@ -1967,11 +1911,10 @@ i965_composite(PixmapPtr dest, int srcX, int srcY, int maskX, int maskY,
 	intel_batch_end_atomic(scrn);
 }
 
-void i965_batch_flush_notify(ScrnInfoPtr scrn)
+void i965_batch_commit_notify(intel_screen_private *intel)
 {
-	intel_screen_private *intel = intel_get_screen_private(scrn);
-
 	intel->needs_render_state_emit = TRUE;
+	intel->needs_3d_invariant = TRUE;
 	intel->last_floats_per_vertex = 0;
 	intel->vertex_index = 0;
 }
@@ -1987,6 +1930,13 @@ void gen4_render_state_init(ScrnInfoPtr scrn)
 	drm_intel_bo *sf_kernel_bo, *sf_kernel_mask_bo;
 	drm_intel_bo *border_color_bo;
 
+	intel->needs_3d_invariant = TRUE;
+
+	intel->surface_bo =
+		drm_intel_bo_alloc(intel->bufmgr, "surface data",
+				   sizeof(intel->surface_data), 4096);
+	intel->surface_used = 0;
+
 	if (INTEL_INFO(intel)->gen >= 60)
 		return gen6_render_state_init(scrn);
 
@@ -2089,11 +2039,6 @@ void gen4_render_state_init(ScrnInfoPtr scrn)
 	drm_intel_bo_unreference(border_color_bo);
 
 	render_state->cc_state_bo = gen4_create_cc_unit_state(scrn);
-	render_state->sip_kernel_bo = intel_bo_alloc_for_data(scrn,
-							      sip_kernel_static,
-							      sizeof
-							      (sip_kernel_static),
-							      "sip kernel");
 }
 
 /**
@@ -2104,9 +2049,8 @@ void gen4_render_state_cleanup(ScrnInfoPtr scrn)
 	intel_screen_private *intel = intel_get_screen_private(scrn);
 	struct gen4_render_state *render_state = intel->gen4_render_state;
 	int i, j, k, l, m;
-	gen4_composite_op *composite_op = &render_state->composite_op;
 
-	drm_intel_bo_unreference(composite_op->surface_state_binding_table_bo);
+	drm_intel_bo_unreference(intel->surface_bo);
 	drm_intel_bo_unreference(render_state->vs_state_bo);
 	drm_intel_bo_unreference(render_state->sf_state_bo);
 	drm_intel_bo_unreference(render_state->sf_mask_state_bo);
@@ -2131,7 +2075,6 @@ void gen4_render_state_cleanup(ScrnInfoPtr scrn)
 					drm_intel_bo_unreference(render_state->ps_sampler_state_bo[i][j][k][l]);
 
 	drm_intel_bo_unreference(render_state->cc_state_bo);
-	drm_intel_bo_unreference(render_state->sip_kernel_bo);
 
 	drm_intel_bo_unreference(render_state->cc_vp_bo);
 	drm_intel_bo_unreference(render_state->gen6_blend_bo);
@@ -2232,9 +2175,6 @@ gen6_composite_invarient_states(ScrnInfoPtr scrn)
 {
 	intel_screen_private *intel = intel_get_screen_private(scrn);
 
-	OUT_BATCH(MI_FLUSH |
-		MI_STATE_INSTRUCTION_CACHE_FLUSH |
-		BRW_MI_GLOBAL_SNAPSHOT_RESET);
 	OUT_BATCH(NEW_PIPELINE_SELECT | PIPELINE_SELECT_3D);
 
 	OUT_BATCH(GEN6_3DSTATE_MULTISAMPLE | (3 - 2));
@@ -2251,13 +2191,15 @@ gen6_composite_invarient_states(ScrnInfoPtr scrn)
 }
 
 static void
-gen6_composite_state_base_address(ScrnInfoPtr scrn, drm_intel_bo *surface_state_binding_table_bo)
+gen6_composite_state_base_address(ScrnInfoPtr scrn)
 {
 	intel_screen_private *intel = intel_get_screen_private(scrn);
 
 	OUT_BATCH(BRW_STATE_BASE_ADDRESS | (10 - 2));
 	OUT_BATCH(BASE_ADDRESS_MODIFY); /* General state base address */
-	OUT_RELOC(surface_state_binding_table_bo, I915_GEM_DOMAIN_INSTRUCTION, 0, BASE_ADDRESS_MODIFY); /* Surface state base address */
+	intel->surface_reloc = intel->batch_used;
+	intel_batch_emit_dword(intel,
+			       intel->surface_bo->offset | BASE_ADDRESS_MODIFY);
 	OUT_BATCH(BASE_ADDRESS_MODIFY); /* Dynamic state base address */
 	OUT_BATCH(BASE_ADDRESS_MODIFY); /* Indirect object base address */
 	OUT_BATCH(BASE_ADDRESS_MODIFY); /* Instruction base address */
@@ -2447,7 +2389,7 @@ gen6_composite_binding_table_pointers(ScrnInfoPtr scrn)
 	OUT_BATCH(0);		/* vs */
 	OUT_BATCH(0);		/* gs */
 	/* Only the PS uses the binding table */
-	OUT_BATCH(PS_BINDING_TABLE_OFFSET);
+	OUT_BATCH(intel->surface_table);
 }
 
 static void
@@ -2572,26 +2514,21 @@ gen6_emit_composite_state(ScrnInfoPtr scrn)
 	sampler_state_extend_t mask_extend = composite_op->mask_extend;
 	Bool is_affine = composite_op->is_affine;
 	uint32_t src_blend, dst_blend;
-	drm_intel_bo *surface_state_binding_table_bo = composite_op->surface_state_binding_table_bo;
 	drm_intel_bo *ps_sampler_state_bo = render_state->ps_sampler_state_bo[src_filter][src_extend][mask_filter][mask_extend];
 
 	intel->needs_render_state_emit = FALSE;
 	IntelEmitInvarientState(scrn);
 	intel->last_3d = LAST_3D_RENDER;
 
-	/* Mark the destination dirty within this batch */
-	intel_batch_mark_pixmap_domains(intel,
-					intel_get_pixmap_private(dest),
-					I915_GEM_DOMAIN_RENDER,
-					I915_GEM_DOMAIN_RENDER);
-	i965_get_blend_cntl(op, 
-			mask_picture, 
-			dest_picture->format,
-			&src_blend, 
-			&dst_blend);
+	i965_get_blend_cntl(op,
+			    mask_picture,
+			    dest_picture->format,
+			    &src_blend,
+			    &dst_blend);
 	assert(intel->in_batch_atomic);
 	gen6_composite_invarient_states(scrn);
-	gen6_composite_state_base_address(scrn, surface_state_binding_table_bo);
+	if (intel->surface_reloc == 0)
+		gen6_composite_state_base_address(scrn);
 	gen6_composite_viewport_state_pointers(scrn, render_state->cc_vp_bo);
 	gen6_composite_urb(scrn);
 	gen6_composite_cc_state_pointers(scrn,
@@ -2665,3 +2602,9 @@ void i965_vertex_flush(struct intel_screen_private *intel)
 		intel->vertex_offset = 0;
 	}
 }
+
+void i965_batch_flush(struct intel_screen_private *intel)
+{
+	if (intel->surface_used)
+		i965_surface_flush(intel);
+}
diff --git a/src/intel.h b/src/intel.h
index 6f76045b..91b63845 100644
--- a/src/intel.h
+++ b/src/intel.h
@@ -560,8 +560,11 @@ typedef struct intel_screen_private {
 
 	CloseScreenProcPtr CloseScreen;
 
+	void (*context_switch) (struct intel_screen_private *intel,
+				int new_mode);
 	void (*vertex_flush) (struct intel_screen_private *intel);
-	void (*batch_flush_notify) (ScrnInfoPtr scrn);
+	void (*batch_flush) (struct intel_screen_private *intel);
+	void (*batch_commit_notify) (struct intel_screen_private *intel);
 
 	uxa_driver_t *uxa_driver;
 	Bool need_sync;
@@ -613,6 +616,7 @@ typedef struct intel_screen_private {
 	PixmapPtr render_current_dest;
 	Bool render_source_is_solid;
 	Bool render_mask_is_solid;
+	Bool needs_3d_invariant;
 	Bool needs_render_state_emit;
 	Bool needs_render_vertex_emit;
 	Bool needs_render_ca_pass;
@@ -646,6 +650,12 @@ typedef struct intel_screen_private {
 	float vertex_ptr[4*1024];
 	dri_bo *vertex_bo;
 
+	uint8_t surface_data[16*1024];
+	uint16_t surface_used;
+	uint16_t surface_table;
+	uint32_t surface_reloc;
+	dri_bo *surface_bo;
+
 	/* 965 render acceleration state */
 	struct gen4_render_state *gen4_render_state;
 
@@ -969,8 +979,8 @@ Bool i915_prepare_composite(int op, PicturePtr sourcec, PicturePtr mask,
 void i915_composite(PixmapPtr dest, int srcX, int srcY,
 		    int maskX, int maskY, int dstX, int dstY, int w, int h);
 void i915_vertex_flush(intel_screen_private *intel);
-void i915_batch_flush_notify(ScrnInfoPtr scrn);
-void i830_batch_flush_notify(ScrnInfoPtr scrn);
+void i915_batch_commit_notify(intel_screen_private *intel);
+void i830_batch_commit_notify(intel_screen_private *intel);
 /* i965_render.c */
 unsigned int gen4_render_state_size(ScrnInfoPtr scrn);
 void gen4_render_state_init(ScrnInfoPtr scrn);
@@ -985,8 +995,9 @@ Bool i965_prepare_composite(int op, PicturePtr sourcec, PicturePtr mask,
 void i965_composite(PixmapPtr dest, int srcX, int srcY,
 		    int maskX, int maskY, int dstX, int dstY, int w, int h);
 
-void i965_vertex_flush(struct intel_screen_private *intel);
-void i965_batch_flush_notify(ScrnInfoPtr scrn);
+void i965_vertex_flush(intel_screen_private *intel);
+void i965_batch_flush(intel_screen_private *intel);
+void i965_batch_commit_notify(intel_screen_private *intel);
 
 Bool intel_transform_is_affine(PictTransformPtr t);
 Bool
diff --git a/src/intel_batchbuffer.c b/src/intel_batchbuffer.c
index 420c900d..3f8053b0 100644
--- a/src/intel_batchbuffer.c
+++ b/src/intel_batchbuffer.c
@@ -40,7 +40,7 @@
 #include "i915_drm.h"
 #include "i965_reg.h"
 
-#define DUMP_BATCHBUFFERS NULL /* "/tmp/i915-batchbuffers.dump" */
+#define DUMP_BATCHBUFFERS NULL // "/tmp/i915-batchbuffers.dump"
 
 static void intel_end_vertex(intel_screen_private *intel)
 {
@@ -149,23 +149,23 @@ void intel_batch_emit_flush(ScrnInfoPtr scrn)
 	assert (!intel->in_batch_atomic);
 
 	/* Big hammer, look to the pipelined flushes in future. */
-	if (intel->current_batch == BLT_BATCH) {
-		BEGIN_BATCH_BLT(4);
-		OUT_BATCH(MI_FLUSH_DW | 2);
-		OUT_BATCH(0);
-		OUT_BATCH(0);
-		OUT_BATCH(0);
-		ADVANCE_BATCH();
-	} else if ((INTEL_INFO(intel)->gen >= 60)) {
-		BEGIN_BATCH(4);
-		OUT_BATCH(BRW_PIPE_CONTROL | (4 - 2)); /* Mesa does so */
-		OUT_BATCH(BRW_PIPE_CONTROL_IS_FLUSH |
-			  BRW_PIPE_CONTROL_WC_FLUSH |
-			  BRW_PIPE_CONTROL_DEPTH_CACHE_FLUSH |
-			  BRW_PIPE_CONTROL_NOWRITE);
-		OUT_BATCH(0); /* write address */
-		OUT_BATCH(0); /* write data */
-		ADVANCE_BATCH();
+	if ((INTEL_INFO(intel)->gen >= 60)) {
+		if (intel->current_batch == BLT_BATCH) {
+			BEGIN_BATCH_BLT(4);
+			OUT_BATCH(MI_FLUSH_DW | 2);
+			OUT_BATCH(0);
+			OUT_BATCH(0);
+			OUT_BATCH(0);
+			ADVANCE_BATCH();
+		} else  {
+			BEGIN_BATCH(4);
+			OUT_BATCH(BRW_PIPE_CONTROL | (4 - 2));
+			OUT_BATCH(BRW_PIPE_CONTROL_WC_FLUSH |
+				  BRW_PIPE_CONTROL_NOWRITE);
+			OUT_BATCH(0); /* write address */
+			OUT_BATCH(0); /* write data */
+			ADVANCE_BATCH();
+		}
 	} else {
 		flags = MI_WRITE_DIRTY_STATE | MI_INVALIDATE_MAP_CACHE;
 		if (INTEL_INFO(intel)->gen >= 40)
@@ -189,6 +189,9 @@ void intel_batch_submit(ScrnInfoPtr scrn, int flush)
 		intel->vertex_flush(intel);
 	intel_end_vertex(intel);
 
+	if (intel->batch_flush)
+		intel->batch_flush(intel);
+
 	if (flush)
 		intel_batch_emit_flush(scrn);
 
@@ -210,11 +213,12 @@ void intel_batch_submit(ScrnInfoPtr scrn, int flush)
 	}
 
 	ret = dri_bo_subdata(intel->batch_bo, 0, intel->batch_used*4, intel->batch_ptr);
-	if (ret == 0)
+	if (ret == 0) {
 		ret = drm_intel_bo_mrb_exec(intel->batch_bo,
 				intel->batch_used*4,
 				NULL, 0, 0xffffffff,
-				intel->current_batch);
+				IS_GEN6(intel) ? intel->current_batch: I915_EXEC_DEFAULT);
+	}
 
 	if (ret != 0) {
 		if (ret == -EIO) {
@@ -275,8 +279,10 @@ void intel_batch_submit(ScrnInfoPtr scrn, int flush)
 	if (intel->debug_flush & DEBUG_FLUSH_WAIT)
 		intel_batch_wait_last(scrn);
 
-	if (intel->batch_flush_notify)
-		intel->batch_flush_notify(scrn);
+	if (intel->batch_commit_notify)
+		intel->batch_commit_notify(intel);
+
+	intel->current_batch = 0;
 }
 
 /** Waits on the last emitted batchbuffer to be completed. */
diff --git a/src/intel_batchbuffer.h b/src/intel_batchbuffer.h
index 2bafbd33..cd6d8e8d 100644
--- a/src/intel_batchbuffer.h
+++ b/src/intel_batchbuffer.h
@@ -183,16 +183,18 @@ do {									\
 		FatalError("%s: BEGIN_BATCH called without closing "	\
 			   "ADVANCE_BATCH\n", __FUNCTION__);		\
 	assert(!intel->in_batch_atomic);				\
-	if (intel->current_batch != batch_idx)				\
-		intel_batch_submit(scrn, FALSE);			\
+	if (intel->current_batch != batch_idx) {			\
+		if (intel->current_batch && intel->context_switch)	\
+			intel->context_switch(intel, batch_idx);	\
+	}								\
 	intel_batch_require_space(scrn, intel, (n) * 4);		\
 	intel->current_batch = batch_idx;				\
 	intel->batch_emitting = (n);					\
 	intel->batch_emit_start = intel->batch_used;			\
 } while (0)
 
-#define BEGIN_BATCH(n) 	__BEGIN_BATCH(n,RENDER_BATCH)
-#define BEGIN_BATCH_BLT(n) 	__BEGIN_BATCH(n,BLT_BATCH)
+#define BEGIN_BATCH(n)	__BEGIN_BATCH(n,RENDER_BATCH)
+#define BEGIN_BATCH_BLT(n)	__BEGIN_BATCH(n,BLT_BATCH)
 
 #define ADVANCE_BATCH() do {						\
 	if (intel->batch_emitting == 0)					\
diff --git a/src/intel_uxa.c b/src/intel_uxa.c
index 4cf42daa..a9cb5b65 100644
--- a/src/intel_uxa.c
+++ b/src/intel_uxa.c
@@ -86,10 +86,16 @@ int uxa_pixmap_index;
 #endif
 
 static void
-ironlake_blt_workaround(ScrnInfoPtr scrn)
+gen6_context_switch(intel_screen_private *intel,
+		    int new_mode)
 {
-	intel_screen_private *intel = intel_get_screen_private(scrn);
+	intel_batch_submit(intel->scrn, FALSE);
+}
 
+static void
+gen5_context_switch(intel_screen_private *intel,
+		    int new_mode)
+{
 	/* Ironlake has a limitation that a 3D or Media command can't
 	 * be the first command after a BLT, unless it's
 	 * non-pipelined.  Instead of trying to track it and emit a
@@ -97,11 +103,24 @@ ironlake_blt_workaround(ScrnInfoPtr scrn)
 	 * non-pipelined 3D instruction after each blit.
 	 */
 
-	if (IS_GEN5(intel)) {
-		BEGIN_BATCH(2);
+	if (new_mode == I915_EXEC_BLT) {
+		OUT_BATCH(MI_FLUSH |
+			  MI_STATE_INSTRUCTION_CACHE_FLUSH |
+			  MI_INHIBIT_RENDER_CACHE_FLUSH);
+	} else {
 		OUT_BATCH(CMD_POLY_STIPPLE_OFFSET << 16);
 		OUT_BATCH(0);
-		ADVANCE_BATCH();
+	}
+}
+
+static void
+gen4_context_switch(intel_screen_private *intel,
+		    int new_mode)
+{
+	if (new_mode == I915_EXEC_BLT) {
+		OUT_BATCH(MI_FLUSH |
+			  MI_STATE_INSTRUCTION_CACHE_FLUSH |
+			  MI_INHIBIT_RENDER_CACHE_FLUSH);
 	}
 }
 
@@ -293,10 +312,7 @@ static void intel_uxa_solid(PixmapPtr pixmap, int x1, int y1, int x2, int y2)
 	pitch = intel_pixmap_pitch(pixmap);
 
 	{
-		if (IS_GEN6(intel))
-			BEGIN_BATCH_BLT(6);
-		else
-			BEGIN_BATCH(6);
+		BEGIN_BATCH_BLT(6);
 
 		cmd = XY_COLOR_BLT_CMD;
 
@@ -320,8 +336,6 @@ static void intel_uxa_solid(PixmapPtr pixmap, int x1, int y1, int x2, int y2)
 		OUT_BATCH(intel->BR[16]);
 		ADVANCE_BATCH();
 	}
-
-	ironlake_blt_workaround(scrn);
 }
 
 static void intel_uxa_done_solid(PixmapPtr pixmap)
@@ -443,10 +457,7 @@ intel_uxa_copy(PixmapPtr dest, int src_x1, int src_y1, int dst_x1,
 	src_pitch = intel_pixmap_pitch(intel->render_source);
 
 	{
-		if (IS_GEN6(intel))
-			BEGIN_BATCH_BLT(8);
-		else
-			BEGIN_BATCH(8);
+		BEGIN_BATCH_BLT(8);
 
 		cmd = XY_SRC_COPY_BLT_CMD;
 
@@ -486,7 +497,6 @@ intel_uxa_copy(PixmapPtr dest, int src_x1, int src_y1, int dst_x1,
 
 		ADVANCE_BATCH();
 	}
-	ironlake_blt_workaround(scrn);
 }
 
 static void intel_uxa_done_copy(PixmapPtr dest)
@@ -1193,6 +1203,8 @@ Bool intel_uxa_init(ScreenPtr screen)
 	intel->floats_per_vertex = 0;
 	intel->last_floats_per_vertex = 0;
 	intel->vertex_bo = NULL;
+	intel->surface_used = 0;
+	intel->surface_reloc = 0;
 
 	/* Solid fill */
 	intel->uxa_driver->check_solid = intel_uxa_check_solid;
@@ -1215,7 +1227,7 @@ Bool intel_uxa_init(ScreenPtr screen)
 		intel->uxa_driver->composite = i830_composite;
 		intel->uxa_driver->done_composite = i830_done_composite;
 
-		intel->batch_flush_notify = i830_batch_flush_notify;
+		intel->batch_commit_notify = i830_batch_commit_notify;
 	} else if (IS_GEN3(intel)) {
 		intel->uxa_driver->check_composite = i915_check_composite;
 		intel->uxa_driver->check_composite_target = i915_check_composite_target;
@@ -1225,7 +1237,7 @@ Bool intel_uxa_init(ScreenPtr screen)
 		intel->uxa_driver->done_composite = i830_done_composite;
 
 		intel->vertex_flush = i915_vertex_flush;
-		intel->batch_flush_notify = i915_batch_flush_notify;
+		intel->batch_commit_notify = i915_batch_commit_notify;
 	} else {
 		intel->uxa_driver->check_composite = i965_check_composite;
 		intel->uxa_driver->check_composite_texture = i965_check_composite_texture;
@@ -1234,7 +1246,16 @@ Bool intel_uxa_init(ScreenPtr screen)
 		intel->uxa_driver->done_composite = i830_done_composite;
 
 		intel->vertex_flush = i965_vertex_flush;
-		intel->batch_flush_notify = i965_batch_flush_notify;
+		intel->batch_flush = i965_batch_flush;
+		intel->batch_commit_notify = i965_batch_commit_notify;
+
+		if (IS_GEN4(intel)) {
+			intel->context_switch = gen4_context_switch;
+		} else if (IS_GEN5(intel)) {
+			intel->context_switch = gen5_context_switch;
+		} else {
+			intel->context_switch = gen6_context_switch;
+		}
 	}
 
 	/* PutImage */