switch XvMC to gem

1 files changed, 362 insertions, 272 deletions

diff --git a/src/xvmc/i965_xvmc.c b/src/xvmc/i965_xvmc.c
index 4b1c4812..51a7ae63 100644
--- a/src/xvmc/i965_xvmc.c
+++ b/src/xvmc/i965_xvmc.c
@@ -32,13 +32,15 @@
 #include "i965_hwmc.h"
 #define BATCH_STRUCT(x) intelBatchbufferData(&x, sizeof(x), 0)
 #define URB_SIZE     256        /* XXX */
+
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof(x[0]))
+
 enum interface {
-    INTRA_INTERFACE,            /* non field intra */
+    INTRA_INTERFACE = 0,            /* non field intra */
     NULL_INTERFACE,             /* fill with white, do nothing, for debug */
     FORWARD_INTERFACE,          /* non field forward predict */
     BACKWARD_INTERFACE,         /* non field backward predict */
     F_B_INTERFACE,              /* non field forward and backward predict */
-    FIELD_INTRA_INTERFACE,      /* field intra */
     FIELD_FORWARD_INTERFACE,    /* field forward predict */
     FIELD_BACKWARD_INTERFACE,   /* field backward predict */
     FIELD_F_B_INTERFACE,        /* field forward and backward predict */
@@ -94,65 +96,137 @@ static const uint32_t dual_prime_igd_kernel_static[][4]= {
 	#include "shader/mc/dual_prime_igd.g4b"
 }; 
 
+struct kernel_struct{
+	const uint32_t (*bin)[4];
+	uint32_t size;
+};
+
+struct kernel_struct kernels_igd[] = {
+    {ipicture_kernel_static, sizeof(ipicture_kernel_static)},
+    {null_kernel_static, sizeof(null_kernel_static)},
+    {frame_forward_igd_kernel_static, sizeof(frame_forward_igd_kernel_static)},
+    {frame_backward_igd_kernel_static, sizeof(frame_backward_igd_kernel_static)},
+    {frame_f_b_igd_kernel_static, sizeof(frame_f_b_igd_kernel_static)},
+    {field_forward_igd_kernel_static, sizeof(field_forward_igd_kernel_static)},
+    {field_backward_igd_kernel_static, sizeof(field_backward_igd_kernel_static)},
+    {field_f_b_igd_kernel_static, sizeof(field_f_b_igd_kernel_static)},
+    {dual_prime_igd_kernel_static, sizeof(dual_prime_igd_kernel_static)}
+};
+
+struct kernel_struct kernels_965[] = {
+    {ipicture_kernel_static, sizeof(ipicture_kernel_static)},
+    {null_kernel_static, sizeof(null_kernel_static)},
+    {frame_forward_kernel_static, sizeof(frame_forward_kernel_static)},
+    {frame_backward_kernel_static, sizeof(frame_backward_kernel_static)},
+    {frame_f_b_kernel_static, sizeof(frame_f_b_kernel_static)},
+    {field_forward_kernel_static, sizeof(field_forward_kernel_static)},
+    {field_backward_kernel_static, sizeof(field_backward_kernel_static)},
+    {field_f_b_kernel_static, sizeof(field_f_b_kernel_static)},
+    {dual_prime_kernel_static, sizeof(dual_prime_kernel_static)}
+};
+
 #define ALIGN(i,m)    (((i) + (m) - 1) & ~((m) - 1))
 
 #define MAX_SURFACE_NUM	10
 #define DESCRIPTOR_NUM 12
 
+struct media_kernel_obj {
+    dri_bo *bo;
+};
+
+struct interface_descriptor_obj {
+   dri_bo *bo;
+   struct media_kernel_obj kernels[DESCRIPTOR_NUM];
+};
+
+struct vfe_state_obj {
+   dri_bo *bo;
+   struct interface_descriptor_obj interface;
+};
+
+struct surface_obj {
+     dri_bo *bo; 
+};
+
+struct surface_state_obj {
+      struct surface_obj surface; 
+      dri_bo *bo;
+};
+
+struct binding_table_obj {
+    dri_bo *bo;
+    struct surface_state_obj surface_states[MAX_SURFACE_NUM];
+};
+
+struct indirect_data_obj {
+    dri_bo *bo;
+};
+
 struct media_state {
-    unsigned long state_base;
-    void 	      *state_ptr;
-    unsigned int  binding_table_entry_count;
-    unsigned long vfe_state_offset;
-    unsigned long interface_descriptor_offset[DESCRIPTOR_NUM];
-    unsigned long ipicture_kernel_offset;
-    unsigned long frame_forward_kernel_offset;
-    unsigned long frame_backward_kernel_offset;
-    unsigned long frame_f_b_kernel_offset;
-    unsigned long ipicture_field_kernel_offset;
-    unsigned long field_forward_kernel_offset;
-    unsigned long field_backward_kernel_offset;
-    unsigned long field_f_b_kernel_offset;
-    unsigned long dual_prime_kernel_offset;
-    unsigned long null_kernel_offset;
-    unsigned long surface_offsets[MAX_SURFACE_NUM];
-    unsigned long binding_table_offset;
     unsigned int  is_g4x:1;
     unsigned int  is_965_q:1;
+
+    struct vfe_state_obj vfe_state;
+    struct binding_table_obj binding_table;
+    struct indirect_data_obj indirect_data;
 };
 struct media_state media_state;
 
-static int map_buffer(struct  drm_memory_block *mem)
+static int free_object(struct media_state *s)
 {
-    return (drmMap(xvmc_driver->fd, 
-		mem->handle, mem->size, &mem->ptr));
+    int i;
+#define FREE_ONE_BO(bo) drm_intel_bo_unreference(bo)
+    FREE_ONE_BO(s->vfe_state.bo);
+    FREE_ONE_BO(s->vfe_state.interface.bo);
+    for (i = 0; i < DESCRIPTOR_NUM; i++)
+        FREE_ONE_BO(s->vfe_state.interface.kernels[i].bo);
+    FREE_ONE_BO(s->binding_table.bo);
+    for (i = 0; i < MAX_SURFACE_NUM; i++)
+        FREE_ONE_BO(s->binding_table.surface_states[i].bo);
+    FREE_ONE_BO(s->indirect_data.bo);
 }
 
-static void unmap_buffer(struct  drm_memory_block *mem)
+static int alloc_object(struct media_state *s)
 {
-    drmUnmap(mem->ptr, mem->size);
+    int i;
+
+    for (i = 0; i < MAX_SURFACE_NUM; i++) {
+        s->binding_table.surface_states[i].bo =
+            drm_intel_bo_alloc(xvmc_driver->bufmgr, "surface_state", 
+ 		sizeof(struct brw_surface_state), 0x1000);
+        if (!s->binding_table.surface_states[i].bo)
+            goto out;
+    }
+    return 0;
+out:
+    free_object(s);
+    return BadAlloc;
 }
 
+
 static Status destroy_context(Display *display, XvMCContext *context)
 {
     struct i965_xvmc_context *private_context;
     private_context = context->privData;
-    unmap_buffer(&private_context->static_buffer);
-    unmap_buffer(&private_context->blocks);
 
+    free_object(&media_state);
     Xfree(private_context);
     return Success;
 }
 
+#define STRIDE(w)               (w)
+#define SIZE_YUV420(w, h)       (h * (STRIDE(w) + STRIDE(w >> 1)))
+
 static Status create_surface(Display *display,
 	XvMCContext *context, XvMCSurface *surface, int priv_count,
 	CARD32 *priv_data)
 {
     struct i965_xvmc_surface *priv_surface = 
 	(struct i965_xvmc_surface *)priv_data;
-    if (map_buffer(&priv_surface->buffer))
-	return BadAlloc;
+    size_t size = SIZE_YUV420(priv_surface->w, priv_surface->h);
     surface->privData = priv_data;
+    priv_surface->bo = drm_intel_bo_alloc(xvmc_driver->bufmgr, "surface", 
+	    size, 0x1000);
     return Success;
 }
 
@@ -160,7 +234,9 @@ static Status destroy_surface(Display *display, XvMCSurface *surface)
 {
     struct i965_xvmc_surface *priv_surface = 
 	surface->privData;
-    unmap_buffer(&priv_surface->buffer);
+    XSync(display, False);
+
+    drm_intel_bo_unreference(priv_surface->bo);
     return Success;
 }
 
@@ -206,21 +282,6 @@ static void urb_layout()
     OUT_BATCH(BRW_URB_FENCE |
 	    UF0_VFE_REALLOC |
 	    UF0_CS_REALLOC |
-	    1);
-    OUT_BATCH(0);
-    OUT_BATCH(((URB_SIZE)<< UF2_VFE_FENCE_SHIFT) |	/* VFE_SIZE */
-	    ((URB_SIZE)<< UF2_CS_FENCE_SHIFT));		/* CS_SIZE is 0 */
-    ADVANCE_BATCH();
-}
-
-/* clear previous urb layout */
-static void clear_urb_state()
-{
-    BATCH_LOCALS;
-    align_urb_fence();
-    BEGIN_BATCH(3);
-    OUT_BATCH(BRW_URB_FENCE |
-	    UF0_CS_REALLOC |
 	    UF0_SF_REALLOC |
 	    UF0_CLIP_REALLOC |
 	    UF0_GS_REALLOC |
@@ -229,8 +290,9 @@ static void clear_urb_state()
     OUT_BATCH((0 << UF1_CLIP_FENCE_SHIFT) |
 	    (0 << UF1_GS_FENCE_SHIFT) |
 	    (0 << UF1_VS_FENCE_SHIFT));
-    OUT_BATCH((0 << UF2_CS_FENCE_SHIFT) |
-	    (0 << UF2_SF_FENCE_SHIFT));
+
+    OUT_BATCH(((URB_SIZE)<< UF2_VFE_FENCE_SHIFT) |	/* VFE_SIZE */
+	    ((URB_SIZE)<< UF2_CS_FENCE_SHIFT));		/* CS_SIZE is 0 */
     ADVANCE_BATCH();
 }
 
@@ -240,62 +302,89 @@ static void media_state_pointers(struct media_state *media_state)
     BEGIN_BATCH(3);
     OUT_BATCH(BRW_MEDIA_STATE_POINTERS|1);
     OUT_BATCH(0);
-    OUT_BATCH(media_state->vfe_state_offset);
-    ADVANCE_BATCH();
-}
-
-static void cs_urb_layout()
-{
-    BATCH_LOCALS;
-    BEGIN_BATCH(2);
-    OUT_BATCH(BRW_CS_URB_STATE | 0);
-    OUT_BATCH((0 << 4) |    /* URB Entry Allocation Size */
-	      (0 << 0));    /* Number of URB Entries */
+    OUT_RELOC(media_state->vfe_state.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0);
     ADVANCE_BATCH();
 }
 
 /* setup 2D surface for media_read or media_write 
  */
-static void setup_media_surface(struct media_state *media_state,
-	int surface_num, unsigned long offset, int w, int h)
+static Status setup_media_surface(struct media_state *media_state,
+	int surface_num, dri_bo *bo, unsigned long offset, int w, int h, Bool write)
 {
-    struct brw_surface_state *ss;
-    ss = media_state->state_ptr +
-	(media_state->surface_offsets[surface_num] - media_state->state_base);
+    struct brw_surface_state s, *ss = &s;
+
     memset(ss, 0, sizeof(struct brw_surface_state));
     ss->ss0.surface_type = BRW_SURFACE_2D;
     ss->ss0.surface_format = BRW_SURFACEFORMAT_R8_SINT;
-    ss->ss1.base_addr = offset;
+    ss->ss1.base_addr = offset + bo->offset;
     ss->ss2.width = w - 1;
     ss->ss2.height = h - 1;
     ss->ss3.pitch = w - 1;
+
+    if (media_state->binding_table.surface_states[surface_num].bo)
+        drm_intel_bo_unreference(media_state->binding_table.surface_states[surface_num].bo);
+    media_state->binding_table.surface_states[surface_num].bo =
+            drm_intel_bo_alloc(xvmc_driver->bufmgr, "surface_state", 
+ 		sizeof(struct brw_surface_state), 0x1000);
+    if (!media_state->binding_table.surface_states[surface_num].bo)
+        return BadAlloc;
+
+    drm_intel_bo_subdata(
+	    media_state->binding_table.surface_states[surface_num].bo,
+	    0, sizeof(*ss), ss);
+
+    drm_intel_bo_emit_reloc(media_state->binding_table.surface_states[surface_num].bo, 
+	    offsetof(struct brw_surface_state, ss1),
+	    bo,
+	    offset,
+	    I915_GEM_DOMAIN_RENDER, write?I915_GEM_DOMAIN_RENDER:0);
+
+    return Success;
 }
 
-static void setup_surfaces(struct media_state *media_state, 
-	unsigned long dst_offset, unsigned long past_offset, 
-	unsigned long future_offset, 
+static Status setup_surfaces(struct media_state *media_state, 
+	dri_bo *dst_bo, dri_bo *past_bo, dri_bo *future_bo, 
 	int w, int h)
 {
-    setup_media_surface(media_state, 0, dst_offset, w, h);
-    setup_media_surface(media_state, 1, dst_offset+w*h, w/2, h/2);
-    setup_media_surface(media_state, 2, dst_offset+w*h + w*h/4, w/2, h/2);
-    if (past_offset) {
-	setup_media_surface(media_state, 4, past_offset, w, h);
-	setup_media_surface(media_state, 5, past_offset+w*h, w/2, h/2);
-	setup_media_surface(media_state, 6, past_offset+w*h + w*h/4, w/2, h/2);
+    Status ret;
+    ret = setup_media_surface(media_state, 0, dst_bo, 0, w, h, TRUE);
+    if (ret != Success)
+        return ret;
+    ret = setup_media_surface(media_state, 1, dst_bo, w*h, w/2, h/2, TRUE);
+    if (ret != Success)
+        return ret;
+    ret = setup_media_surface(media_state, 2, dst_bo, w*h + w*h/4, w/2, h/2, TRUE);
+    if (ret != Success)
+        return ret;
+    if (past_bo) {
+	ret = setup_media_surface(media_state, 4, past_bo, 0, w, h, FALSE);
+        if (ret != Success)
+            return ret;
+	ret = setup_media_surface(media_state, 5, past_bo, w*h, w/2, h/2, FALSE);
+        if (ret != Success)
+            return ret;
+	ret = setup_media_surface(media_state, 6, past_bo, w*h + w*h/4, w/2, h/2, FALSE);
+        if (ret != Success)
+            return ret;
     }
-    if (future_offset) {
-	setup_media_surface(media_state, 7, future_offset, w, h);
-	setup_media_surface(media_state, 8, future_offset+w*h, w/2, h/2);
-	setup_media_surface(media_state, 9, future_offset+w*h + w*h/4, w/2, h/2);
+    if (future_bo) {
+	ret = setup_media_surface(media_state, 7, future_bo, 0, w, h, FALSE);
+        if (ret != Success)
+            return ret;
+	ret = setup_media_surface(media_state, 8, future_bo, w*h, w/2, h/2, FALSE);
+        if (ret != Success)
+            return ret;
+	ret = setup_media_surface(media_state, 9, future_bo, w*h + w*h/4, w/2, h/2, FALSE);
+        if (ret != Success)
+            return ret;
     }
+    return Success;
 }
 /* BUFFER SURFACE has a strange format
  * the size of the surface is in part of w h and d component
  */
 
-static void setup_blocks(struct media_state *media_state, 
-	unsigned long offset, unsigned int block_size)
+static Status setup_blocks(struct media_state *media_state, unsigned int block_size)
 {
     union element{
 		struct {
@@ -306,22 +395,39 @@ static void setup_blocks(struct media_state *media_state,
 		}whd;
 		unsigned int size;	
     }e;
-    struct brw_surface_state *ss;
-    ss = media_state->state_ptr +
-	(media_state->surface_offsets[3] - media_state->state_base);
-    memset(ss, 0, sizeof(struct brw_surface_state));
-    ss->ss0.surface_type = BRW_SURFACE_BUFFER;
-    ss->ss0.surface_format = BRW_SURFACEFORMAT_R8_UINT;
-    ss->ss1.base_addr = offset;
+    struct brw_surface_state ss;
+    memset(&ss, 0, sizeof(struct brw_surface_state));
+    ss.ss0.surface_type = BRW_SURFACE_BUFFER;
+    ss.ss0.surface_format = BRW_SURFACEFORMAT_R8_UINT;
+    ss.ss1.base_addr = media_state->indirect_data.bo->offset;
+
     e.size = block_size - 1;
-    ss->ss2.width = e.whd.w;
-    ss->ss2.height = e.whd.h;
-    ss->ss3.depth = e.whd.d;
-    ss->ss3.pitch = block_size - 1;
+    ss.ss2.width = e.whd.w;
+    ss.ss2.height = e.whd.h;
+    ss.ss3.depth = e.whd.d;
+    ss.ss3.pitch = block_size - 1;
+  
+    if (media_state->binding_table.surface_states[3].bo)
+        drm_intel_bo_unreference(media_state->binding_table.surface_states[3].bo);
+
+    media_state->binding_table.surface_states[3].bo =
+            drm_intel_bo_alloc(xvmc_driver->bufmgr, "surface_state", 
+ 		sizeof(struct brw_surface_state), 0x1000);
+    if (!media_state->binding_table.surface_states[3].bo)
+        return BadAlloc;
+  
+    drm_intel_bo_subdata(media_state->binding_table.surface_states[3].bo, 0, 
+	    sizeof(ss), &ss);
+
+    drm_intel_bo_emit_reloc(media_state->binding_table.surface_states[3].bo,
+        offsetof(struct brw_surface_state, ss1),
+        media_state->indirect_data.bo, 0,
+        I915_GEM_DOMAIN_SAMPLER, 0);
+    return Success;
 }
 
 /* setup state base address */
-static void state_base_address(int offset)
+static void state_base_address()
 {
     BATCH_LOCALS;
     BEGIN_BATCH(6);
@@ -330,7 +436,7 @@ static void state_base_address(int offset)
     OUT_BATCH(0 | BASE_ADDRESS_MODIFY); 
     OUT_BATCH(0 | BASE_ADDRESS_MODIFY);
     OUT_BATCH(0 | BASE_ADDRESS_MODIFY);
-    OUT_BATCH((0xFFFFF<<12) | BASE_ADDRESS_MODIFY);
+    OUT_BATCH(BASE_ADDRESS_MODIFY);
     ADVANCE_BATCH();
 }
 
@@ -358,12 +464,13 @@ static void send_media_object(XvMCMacroBlock *mb, int offset, enum interface int
 	OUT_BATCH(0);
     }else {
 	OUT_BATCH(6*128);
-	OUT_BATCH(offset);
+        OUT_RELOC(media_state.indirect_data.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, offset);
     }
     
     OUT_BATCH(mb->x<<4);                 //g1.0
     OUT_BATCH(mb->y<<4);
-    OUT_BATCH(offset);               //g1.8
+    OUT_RELOC(media_state.indirect_data.bo, //g1.8
+            I915_GEM_DOMAIN_INSTRUCTION, 0, offset);
     OUT_BATCH_SHORT(mb->coded_block_pattern);  //g1.12
     OUT_BATCH_SHORT(mb->PMV[0][0][0]);         //g1.14
     OUT_BATCH_SHORT(mb->PMV[0][0][1]);         //g1.16
@@ -384,178 +491,152 @@ static void send_media_object(XvMCMacroBlock *mb, int offset, enum interface int
     ADVANCE_BATCH();
 }
 
-static void binding_tables(struct media_state *media_state)
+static Status binding_tables(struct media_state *media_state)
 {
-    unsigned int *binding_table;
+    unsigned int binding_table[MAX_SURFACE_NUM];
     int i;
-    binding_table = media_state->state_ptr +
-	(media_state->binding_table_offset - media_state->state_base);
+
+    if (media_state->binding_table.bo)
+        drm_intel_bo_unreference(media_state->binding_table.bo);
+    media_state->binding_table.bo = 
+ 	drm_intel_bo_alloc(xvmc_driver->bufmgr, "binding_table", 
+ 		MAX_SURFACE_NUM*4, 0x1000);
+    if (!media_state->binding_table.bo)
+        return BadAlloc;
+
+    for (i = 0; i < MAX_SURFACE_NUM; i++)
+	binding_table[i] = media_state->binding_table.surface_states[i].bo->offset;
+    drm_intel_bo_subdata(media_state->binding_table.bo, 0, sizeof(binding_table),
+        binding_table);
+
     for (i = 0; i < MAX_SURFACE_NUM; i++)
-	binding_table[i] = media_state->surface_offsets[i];
+        drm_intel_bo_emit_reloc(media_state->binding_table.bo, 
+	    i * sizeof(unsigned int),
+	    media_state->binding_table.surface_states[i].bo, 0,
+	    I915_GEM_DOMAIN_INSTRUCTION, 0);
+     return Success;
 }
 
-static void media_kernels(struct media_state *media_state)
+static int media_kernels(struct media_state *media_state)
 {
-	void *kernel; 
-#define LOAD_KERNEL(name) kernel = media_state->state_ptr +\
-	(media_state->name##_kernel_offset - media_state->state_base);\
-	memcpy(kernel, name##_kernel_static, sizeof(name##_kernel_static));
-#define LOAD_KERNEL_IGD(name) kernel = media_state->state_ptr +\
-	(media_state->name##_kernel_offset - media_state->state_base);\
-	memcpy(kernel, name##_igd_kernel_static, sizeof(name##_igd_kernel_static));
-
-	LOAD_KERNEL(ipicture);
-	LOAD_KERNEL(null);
+	struct kernel_struct *kernels;
+	int kernel_array_size, i;
+
 	if (media_state->is_g4x) {
-		LOAD_KERNEL_IGD(frame_forward);
-		LOAD_KERNEL_IGD(field_forward);
-		LOAD_KERNEL_IGD(frame_backward);
-		LOAD_KERNEL_IGD(field_backward);
-		LOAD_KERNEL_IGD(frame_f_b);
-		LOAD_KERNEL_IGD(field_f_b);
-		LOAD_KERNEL_IGD(dual_prime);
-
-	}else {
-		LOAD_KERNEL(frame_forward);
-		LOAD_KERNEL(field_forward);
-		LOAD_KERNEL(frame_backward);
-		LOAD_KERNEL(field_backward);
-		LOAD_KERNEL(frame_f_b);
-		LOAD_KERNEL(field_f_b);
-		LOAD_KERNEL(dual_prime);
+		kernels = kernels_igd;
+		kernel_array_size = ARRAY_SIZE(kernels_igd);
+	} else {
+		kernels = kernels_965;
+		kernel_array_size = ARRAY_SIZE(kernels_965);
+	}
+
+	for (i = 0; i < kernel_array_size; i++) {
+		media_state->vfe_state.interface.kernels[i].bo =
+			drm_intel_bo_alloc(xvmc_driver->bufmgr, "kernel",
+					kernels[i].size, 0x1000);
+		if (!media_state->vfe_state.interface.kernels[i].bo)
+			goto out;
 	}
+
+	for (i = 0; i < kernel_array_size; i++) {
+		dri_bo *bo = media_state->vfe_state.interface.kernels[i].bo;
+		drm_intel_bo_subdata(bo, 0, kernels[i].size, kernels[i].bin);
+	}
+	return 0;
+out:
+	free_object(media_state);
+	return BadAlloc;
 }
 
 static void setup_interface(struct media_state *media_state, 
-	enum interface interface, unsigned int kernel_offset)
+	enum interface i)
 {
-    struct brw_interface_descriptor *desc;
-    desc = media_state->state_ptr +
-	(media_state->interface_descriptor_offset[interface] 
-	 - media_state->state_base);
-    memset(desc, 0, sizeof(*desc));
-    desc->desc0.grf_reg_blocks = 15;
-    desc->desc0.kernel_start_pointer = kernel_offset >> 6;
-    desc->desc1.floating_point_mode = BRW_FLOATING_POINT_NON_IEEE_754;
+    struct brw_interface_descriptor desc;
+    memset(&desc, 0, sizeof(desc));
+
+    desc.desc0.grf_reg_blocks = 15;
+    desc.desc0.kernel_start_pointer =
+        media_state->vfe_state.interface.kernels[i].bo->offset >> 6;
+
+    desc.desc1.floating_point_mode = BRW_FLOATING_POINT_NON_IEEE_754;
 
     /* use same binding table for all interface
      * may change this if it affect performance
      */
-    desc->desc3.binding_table_entry_count = MAX_SURFACE_NUM;
-    desc->desc3.binding_table_pointer = media_state->binding_table_offset >> 5;
+    desc.desc3.binding_table_entry_count = MAX_SURFACE_NUM;
+    desc.desc3.binding_table_pointer = media_state->binding_table.bo->offset >> 5;
+
+    drm_intel_bo_subdata(media_state->vfe_state.interface.bo, i*sizeof(desc), 
+	    sizeof(desc), &desc);
+
+    drm_intel_bo_emit_reloc(
+	    media_state->vfe_state.interface.bo,
+	    i * sizeof(desc) + 
+	    offsetof(struct brw_interface_descriptor, desc0),
+	    media_state->vfe_state.interface.kernels[i].bo,
+	    desc.desc0.grf_reg_blocks,
+	    I915_GEM_DOMAIN_INSTRUCTION, 0);
+
+    drm_intel_bo_emit_reloc(
+	    media_state->vfe_state.interface.bo,
+	    i * sizeof(desc) + 
+	    offsetof(struct brw_interface_descriptor, desc3),
+	    media_state->binding_table.bo,
+	    desc.desc3.binding_table_entry_count,
+	    I915_GEM_DOMAIN_INSTRUCTION, 0);
 }
 
-static void interface_descriptor(struct media_state *media_state)
+static Status interface_descriptor(struct media_state *media_state)
 {
-	setup_interface(media_state, INTRA_INTERFACE, 
-		media_state->ipicture_kernel_offset);
-	setup_interface(media_state, NULL_INTERFACE, 
-		media_state->null_kernel_offset);
-	setup_interface(media_state, FORWARD_INTERFACE, 
-		media_state->frame_forward_kernel_offset);
-	setup_interface(media_state, FIELD_FORWARD_INTERFACE, 
-		media_state->field_forward_kernel_offset);
-	setup_interface(media_state, BACKWARD_INTERFACE, 
-		media_state->frame_backward_kernel_offset);
-	setup_interface(media_state, FIELD_BACKWARD_INTERFACE, 
-		media_state->field_backward_kernel_offset);
-	setup_interface(media_state, F_B_INTERFACE, 
-		media_state->frame_f_b_kernel_offset);
-	setup_interface(media_state, FIELD_F_B_INTERFACE, 
-		media_state->field_f_b_kernel_offset);
-	setup_interface(media_state, DUAL_PRIME_INTERFACE,
-		media_state->dual_prime_kernel_offset);
+	if (media_state->vfe_state.interface.bo)
+		drm_intel_bo_unreference(media_state->vfe_state.interface.bo);
+	media_state->vfe_state.interface.bo = drm_intel_bo_alloc(xvmc_driver->bufmgr,
+			"interfaces", DESCRIPTOR_NUM*sizeof(struct brw_interface_descriptor),
+			0x1000);
+	if (!media_state->vfe_state.interface.bo)
+		return BadAlloc;
+
+	setup_interface(media_state, INTRA_INTERFACE);
+	setup_interface(media_state, NULL_INTERFACE);
+	setup_interface(media_state, FORWARD_INTERFACE);
+	setup_interface(media_state, FIELD_FORWARD_INTERFACE);
+	setup_interface(media_state, BACKWARD_INTERFACE);
+	setup_interface(media_state, FIELD_BACKWARD_INTERFACE);
+	setup_interface(media_state, F_B_INTERFACE);
+	setup_interface(media_state, FIELD_F_B_INTERFACE);
+	setup_interface(media_state, DUAL_PRIME_INTERFACE);
+	return Success;
 }
 
-static void vfe_state(struct media_state *media_state)
+static Status vfe_state(struct media_state *media_state)
 {
-	struct brw_vfe_state *state;
-	state = media_state->state_ptr +
-	    (media_state->vfe_state_offset - media_state->state_base);
-	memset(state, 0, sizeof(*state));
+	struct brw_vfe_state state;
+	memset(&state, 0, sizeof(state));
+
 	/* no scratch space */
-	state->vfe1.vfe_mode = VFE_GENERIC_MODE;
-	state->vfe1.num_urb_entries = 1; 
+	state.vfe1.vfe_mode = VFE_GENERIC_MODE;
+	state.vfe1.num_urb_entries = 1;
 	/* XXX TODO */
 	/* should carefully caculate those values for performance */
-	state->vfe1.urb_entry_alloc_size = 2; 
-	state->vfe1.max_threads = 31; 
-	state->vfe2.interface_descriptor_base = 
-		media_state->interface_descriptor_offset[0] >> 4;
-}
-
-static void calc_state_layouts(struct media_state *media_state)
-{
-    int i;
-    media_state->vfe_state_offset = ALIGN(media_state->state_base, 64);
-    media_state->interface_descriptor_offset[0] = 
-	ALIGN(media_state->vfe_state_offset + sizeof(struct brw_vfe_state), 64);
-    for (i = 1; i < DESCRIPTOR_NUM; i++)
-	media_state->interface_descriptor_offset[i] = 
-	    media_state->interface_descriptor_offset[i-1]
-	    + sizeof(struct brw_interface_descriptor);
-    media_state->binding_table_offset = 
-	ALIGN(media_state->interface_descriptor_offset[DESCRIPTOR_NUM - 1]
-		+ sizeof(struct brw_interface_descriptor), 64);
-    media_state->surface_offsets[0] = 
-	ALIGN(media_state->binding_table_offset
-		+ 4*media_state->binding_table_entry_count , 32);
-    for (i = 1; i < MAX_SURFACE_NUM; i++)
-	media_state->surface_offsets[i] = 
-	    ALIGN(media_state->surface_offsets[i - 1] 
-		    + sizeof(struct brw_surface_state) , 32);
-    media_state->ipicture_kernel_offset = 
-	ALIGN(media_state->surface_offsets[MAX_SURFACE_NUM - 1] 
-		+ sizeof(struct brw_surface_state) , 64);
-
-    media_state->frame_forward_kernel_offset = 
-	    ALIGN(media_state->ipicture_kernel_offset + 
-			    sizeof(ipicture_kernel_static), 64);
-    if(!media_state->is_g4x) {
-	    media_state->field_forward_kernel_offset = 
-		    ALIGN(media_state->frame_forward_kernel_offset + 
-				    sizeof(frame_forward_kernel_static), 64);
-	    media_state->frame_backward_kernel_offset = 
-		    ALIGN(media_state->field_forward_kernel_offset + 
-				    sizeof(field_forward_kernel_static), 64);
-	    media_state->field_backward_kernel_offset = 
-		    ALIGN(media_state->frame_backward_kernel_offset + 
-				    sizeof(frame_backward_kernel_static), 64);
-	    media_state->frame_f_b_kernel_offset = 
-		    ALIGN(media_state->field_backward_kernel_offset + 
-				    sizeof(field_backward_kernel_static), 64);
-	    media_state->field_f_b_kernel_offset = 
-		    ALIGN(media_state->frame_f_b_kernel_offset + 
-				    sizeof(frame_f_b_kernel_static), 64);
-	    media_state->null_kernel_offset =
-		    ALIGN(media_state->field_f_b_kernel_offset +
-				    sizeof(field_f_b_kernel_static), 64);
-	    media_state->dual_prime_kernel_offset =
-		    ALIGN(media_state->null_kernel_offset +
-				    sizeof(null_kernel_static), 64);
-    } else {
-	    media_state->field_forward_kernel_offset = 
-		    ALIGN(media_state->frame_forward_kernel_offset + 
-				    sizeof(frame_forward_igd_kernel_static), 64);
-	    media_state->frame_backward_kernel_offset = 
-		    ALIGN(media_state->field_forward_kernel_offset + 
-				    sizeof(field_forward_igd_kernel_static), 64);
-	    media_state->field_backward_kernel_offset = 
-		    ALIGN(media_state->frame_backward_kernel_offset + 
-				    sizeof(frame_backward_igd_kernel_static), 64);
-	    media_state->frame_f_b_kernel_offset = 
-		    ALIGN(media_state->field_backward_kernel_offset + 
-				    sizeof(field_backward_igd_kernel_static), 64);
-	    media_state->field_f_b_kernel_offset = 
-		    ALIGN(media_state->frame_f_b_kernel_offset + 
-				    sizeof(frame_f_b_igd_kernel_static), 64);
-	    media_state->null_kernel_offset =
-		    ALIGN(media_state->field_f_b_kernel_offset +
-				    sizeof(field_f_b_igd_kernel_static), 64);
-	    media_state->dual_prime_kernel_offset =
-		    ALIGN(media_state->null_kernel_offset +
-				    sizeof(null_kernel_static), 64);
-    }
+	state.vfe1.urb_entry_alloc_size = 2;
+	state.vfe1.max_threads = 31;
+        state.vfe2.interface_descriptor_base =
+		media_state->vfe_state.interface.bo->offset >> 4;
+
+	if (media_state->vfe_state.bo)
+		drm_intel_bo_unreference(media_state->vfe_state.bo);
+	media_state->vfe_state.bo = drm_intel_bo_alloc(xvmc_driver->bufmgr,
+		"vfe state", sizeof(struct brw_vfe_state), 0x1000);
+	if (!media_state->vfe_state.bo)
+		return BadAlloc;
+
+	drm_intel_bo_subdata(media_state->vfe_state.bo, 0, sizeof(state), &state);
+
+	drm_intel_bo_emit_reloc(media_state->vfe_state.bo,
+		offsetof(struct brw_vfe_state, vfe2),
+		media_state->vfe_state.interface.bo, 0,
+		I915_GEM_DOMAIN_INSTRUCTION, 0);
+	return Success;
 }
 
 static Status render_surface(Display *display, 
@@ -588,13 +669,35 @@ static Status render_surface(Display *display,
 	XVMC_ERR("Can't find intel xvmc context\n");
 	return BadValue;
     }
+
+    if (media_state.indirect_data.bo) {
+        if (xvmc_driver->kernel_exec_fencing)
+            drm_intel_gem_bo_unmap_gtt(media_state.indirect_data.bo);
+        else
+            drm_intel_bo_unmap(media_state.indirect_data.bo);
+
+        drm_intel_bo_unreference(media_state.indirect_data.bo);
+    }
+    media_state.indirect_data.bo = drm_intel_bo_alloc(xvmc_driver->bufmgr,
+	    "indirect data", 128*6*num_macroblocks, 64);
+    if (!media_state.indirect_data.bo)
+        return BadAlloc;
     setup_surfaces(&media_state, 
-	    priv_target_surface->buffer.offset, 
-	    past_surface? priv_past_surface->buffer.offset:0, 
-	    future_surface?priv_future_surface->buffer.offset:0, 
+	    priv_target_surface->bo, 
+	    past_surface? priv_past_surface->bo:NULL, 
+	    future_surface?priv_future_surface->bo:NULL, 
 	    context->width, context->height);
+    setup_blocks(&media_state, 128*6*num_macroblocks);
+    binding_tables(&media_state);
+    interface_descriptor(&media_state);
+    vfe_state(&media_state);
+
+    if (xvmc_driver->kernel_exec_fencing)
+        drm_intel_gem_bo_map_gtt(media_state.indirect_data.bo);
+    else
+        drm_intel_bo_map(media_state.indirect_data.bo, 1);
 
-    block_ptr = i965_ctx->blocks.ptr;
+    block_ptr = media_state.indirect_data.bo->virtual;
     for (i = first_macroblock; 
 	    i < num_macroblocks + first_macroblock; i++) {
 	unsigned short *mb_block_ptr;
@@ -635,20 +738,15 @@ static Status render_surface(Display *display,
 	    memcpy(block_ptr, mb_block_ptr, 128);
 	block_ptr += 64;
     }
-
     {
-	int block_offset;
-	block_offset = media_state.is_965_q?0:i965_ctx->blocks.offset;
+	int block_offset = 0;
 	LOCK_HARDWARE(intel_ctx->hw_context);
-	state_base_address(block_offset);
+	state_base_address();
 	flush();	
 	clear_sf_state();
-	clear_urb_state();
 	pipeline_select(&media_state);
 	urb_layout();	
 	media_state_pointers(&media_state);
-	cs_urb_layout();
-
 	for (i = first_macroblock; 
 		i < num_macroblocks + first_macroblock; 
 		i++, block_offset += 128*6) {
@@ -700,8 +798,11 @@ static Status put_surface(Display *display,XvMCSurface *surface,
 {
 	struct i965_xvmc_surface *private_surface =
 		surface->privData;
+	uint32_t handle = 0;
+
+	drm_intel_bo_flink(private_surface->bo, &handle);
+	data->handle = handle;
 
-	data->surf_offset = private_surface->buffer.offset;
 	return Success;
 }
 
@@ -718,25 +819,14 @@ static Status create_context(Display *display, XvMCContext *context,
     struct i965_xvmc_context *i965_ctx;
     i965_ctx = (struct i965_xvmc_context *)priv_data;
     context->privData = i965_ctx;
-    if (map_buffer(&i965_ctx->static_buffer))
-	return BadAlloc;
-    if(map_buffer(&i965_ctx->blocks))
-	return BadAlloc;
-    {
-	media_state.state_base = i965_ctx->static_buffer.offset;
-	media_state.state_ptr = i965_ctx->static_buffer.ptr;
-	media_state.is_g4x = i965_ctx->is_g4x;
-	media_state.is_965_q = i965_ctx->is_965_q;
-	media_state.binding_table_entry_count = MAX_SURFACE_NUM;
-	calc_state_layouts(&media_state);
-	vfe_state(&media_state);
-	interface_descriptor(&media_state); 
-	media_kernels(&media_state);
-	setup_blocks(&media_state, 
-		i965_ctx->blocks.offset, 
-		6*context->width*context->height*sizeof(short));
-	binding_tables(&media_state);
-    }
+
+    media_state.is_g4x = i965_ctx->is_g4x;
+    media_state.is_965_q = i965_ctx->is_965_q;
+
+    if (alloc_object(&media_state))
+        return BadAlloc;
+    if (media_kernels(&media_state))
+        return BadAlloc;
     return Success;
 }