i965_drv_video: fix graphics memory usage

i965_drv_video/i965_avc_bsd.h

@@ -29,12 +29,7 @@
 #ifndef __I965_AVC_BSD_H__
 #define __I965_AVC_BSD_H__
 
-struct i965_avc_bsd_surface
-{
-    dri_bo *direct_mv_wr_top_bo;
-    dri_bo *direct_mv_wr_bottom_bo;
-    int direct_mv_flag;
-};
+#define DMV_SIZE        0x88000 /* 557056 bytes for a frame */
 
 struct i965_avc_bsd_context
 {
@@ -46,14 +41,15 @@ struct i965_avc_bsd_context
         dri_bo *bo;
     } mpr_row_store;
 
-    struct {
-        dri_bo *bo;
-    } avc_it_command_mb_info;
+    int init;
+};
 
-    struct {
-        dri_bo *bo;
-        long write_offset;
-    } avc_it_data;
+struct i965_avc_bsd_surface
+{
+    struct i965_avc_bsd_context *ctx;
+    dri_bo *dmv_top;
+    dri_bo *dmv_bottom;
+    int dmv_bottom_flag;
 };
 
 void i965_avc_bsd_pipeline(VADriverContextP, struct decode_state *);

i965_drv_video/i965_avc_hw_scoreboard.c

@@ -176,11 +176,15 @@ i965_avc_hw_scoreboard_upload_constants(struct i965_avc_hw_scoreboard_context *a
 {
     unsigned char *constant_buffer;
 
+    if (avc_hw_scoreboard_context->curbe.upload)
+        return;
+
     dri_bo_map(avc_hw_scoreboard_context->curbe.bo, 1);
     assert(avc_hw_scoreboard_context->curbe.bo->virtual);
     constant_buffer = avc_hw_scoreboard_context->curbe.bo->virtual;
     memcpy(constant_buffer, avc_hw_scoreboard_constants, sizeof(avc_hw_scoreboard_constants));
     dri_bo_unmap(avc_hw_scoreboard_context->curbe.bo);
+    avc_hw_scoreboard_context->curbe.upload = 1;
 }
 
 static void
@@ -358,12 +362,14 @@ i965_avc_hw_scoreboard_decode_init(VADriverContextP ctx)
         struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context = &i965_h264_context->avc_hw_scoreboard_context;
         dri_bo *bo;
 
-        dri_bo_unreference(avc_hw_scoreboard_context->curbe.bo);
-        bo = dri_bo_alloc(i965->intel.bufmgr,
-                          "constant buffer",
-                          4096, 64);
-        assert(bo);
-        avc_hw_scoreboard_context->curbe.bo = bo;
+        if (avc_hw_scoreboard_context->curbe.bo == NULL) {
+            bo = dri_bo_alloc(i965->intel.bufmgr,
+                              "constant buffer",
+                              4096, 64);
+            assert(bo);
+            avc_hw_scoreboard_context->curbe.bo = bo;
+            avc_hw_scoreboard_context->curbe.upload = 0;
+        }
 
         dri_bo_unreference(avc_hw_scoreboard_context->surface.s_bo);
         avc_hw_scoreboard_context->surface.s_bo = i965_h264_context->avc_it_command_mb_info.bo;

i965_drv_video/i965_avc_hw_scoreboard.h

@@ -57,6 +57,7 @@ struct i965_avc_hw_scoreboard_context
 
     struct {
         dri_bo *bo;
+        int upload;
     } curbe;
 
     struct {

i965_drv_video/i965_drv_video.c

@@ -367,11 +367,8 @@ i965_CreateSurfaces(VADriverContextP ctx,
         obj_surface->width = ALIGN(width, 16);
         obj_surface->height = ALIGN(height, 16);
         obj_surface->size = SIZE_YUV420(obj_surface->width, obj_surface->height);
-        obj_surface->bo = dri_bo_alloc(i965->intel.bufmgr,
-                                       "vaapi surface",
-                                       obj_surface->size,
-                                       64);
-        assert(obj_surface->bo);
+        obj_surface->flags = SURFACE_REFERENCED;
+        obj_surface->bo = NULL;
         obj_surface->private_data = NULL;
         obj_surface->free_private_data = NULL;
     }
@@ -422,16 +419,16 @@ i965_QueryImageFormats(VADriverContextP ctx,
 
 VAStatus 
 i965_PutImage(VADriverContextP ctx,
-               VASurfaceID surface,
-               VAImageID image,
-               int src_x,
-               int src_y,
-               unsigned int src_width,
-               unsigned int src_height,
-               int dest_x,
-               int dest_y,
-               unsigned int dest_width,
-               unsigned int dest_height)
+              VASurfaceID surface,
+              VAImageID image,
+              int src_x,
+              int src_y,
+              unsigned int src_width,
+              unsigned int src_height,
+              int dest_x,
+              int dest_y,
+              unsigned int dest_width,
+              unsigned int dest_height)
 {
     return VA_STATUS_SUCCESS;
 }
@@ -474,7 +471,7 @@ i965_CreateSubpicture(VADriverContextP ctx,
     struct i965_driver_data *i965 = i965_driver_data(ctx);
     VASubpictureID subpicID = NEW_SUBPIC_ID()
 	
-    struct object_subpic *obj_subpic = SUBPIC(subpicID);
+        struct object_subpic *obj_subpic = SUBPIC(subpicID);
     if (!obj_subpic)
         return VA_STATUS_ERROR_ALLOCATION_FAILED;
 
@@ -794,8 +791,8 @@ i965_CreateBuffer(VADriverContextP ctx,
 
     if (type == VASliceDataBufferType || type == VAImageBufferType) {
         buffer_store->bo = dri_bo_alloc(i965->intel.bufmgr, 
-                                      "Buffer", 
-                                      size * num_elements, 64);
+                                        "Buffer", 
+                                        size * num_elements, 64);
         assert(buffer_store->bo);
 
         if (data)
@@ -977,7 +974,7 @@ i965_render_bit_plane_buffer(VADriverContextP ctx,
     assert(obj_buffer->buffer_store->buffer);
     i965_release_buffer_store(&obj_context->decode_state.bit_plane);
     i965_reference_buffer_store(&obj_context->decode_state.bit_plane,
-                                 obj_buffer->buffer_store);
+                                obj_buffer->buffer_store);
     
     return VA_STATUS_SUCCESS;
 }
@@ -992,7 +989,7 @@ i965_render_slice_parameter_buffer(VADriverContextP ctx,
     
     if (obj_context->decode_state.num_slice_params == obj_context->decode_state.max_slice_params) {
         obj_context->decode_state.slice_params = realloc(obj_context->decode_state.slice_params,
-                                                        (obj_context->decode_state.max_slice_params + NUM_SLICES) * sizeof(*obj_context->decode_state.slice_params));
+                                                         (obj_context->decode_state.max_slice_params + NUM_SLICES) * sizeof(*obj_context->decode_state.slice_params));
         memset(obj_context->decode_state.slice_params + obj_context->decode_state.max_slice_params, 0, NUM_SLICES * sizeof(*obj_context->decode_state.slice_params));
         obj_context->decode_state.max_slice_params += NUM_SLICES;
     }
@@ -1327,7 +1324,7 @@ i965_CreateImage(VADriverContextP ctx,
     *out_image                  = *image;
     return VA_STATUS_SUCCESS;
 
- error:
+error:
     i965_DestroyImage(ctx, image_id);
     return va_status;
 }
@@ -1393,7 +1390,7 @@ i965_SetImagePalette(VADriverContextP ctx,
     for (i = 0; i < obj_image->image.num_palette_entries; i++)
         obj_image->palette[i] = (((unsigned int)palette[3*i + 0] << 16) |
                                  ((unsigned int)palette[3*i + 1] <<  8) |
-                                  (unsigned int)palette[3*i + 2]);
+                                 (unsigned int)palette[3*i + 2]);
     return VA_STATUS_SUCCESS;
 }
 
@@ -1432,13 +1429,20 @@ i965_PutSurface(VADriverContextP ctx,
     union dri_buffer *buffer;
     struct intel_region *dest_region;
     struct object_surface *obj_surface; 
-	int ret;
+    int ret;
     uint32_t name;
     Bool new_region = False;
     /* Currently don't support DRI1 */
     if (dri_state->driConnectedFlag != VA_DRI2)
         return VA_STATUS_ERROR_UNKNOWN;
 
+    /* Some broken sources such as H.264 conformance case FM2_SVA_C
+     * will get here
+     */
+    obj_surface = SURFACE(surface);
+    if (obj_surface->bo == NULL)
+        return VA_STATUS_SUCCESS;
+
     dri_drawable = dri_get_drawable(ctx, draw);
     assert(dri_drawable);
 
@@ -1480,13 +1484,24 @@ i965_PutSurface(VADriverContextP ctx,
     i965_render_put_surface(ctx, surface,
                             srcx, srcy, srcw, srch,
                             destx, desty, destw, desth);
-    obj_surface = SURFACE(surface);
+
     if(obj_surface->subpic != VA_INVALID_ID) {	
 	i965_render_put_subpic(ctx, surface,
-                           srcx, srcy, srcw, srch,
-                           destx, desty, destw, desth);
+                               srcx, srcy, srcw, srch,
+                               destx, desty, destw, desth);
     } 
+
     dri_swap_buffer(ctx, dri_drawable);
+    obj_surface->flags |= SURFACE_DISPLAYED;
+
+    if (!(obj_surface->flags & SURFACE_REFERENCED)) {
+        dri_bo_unreference(obj_surface->bo);
+        obj_surface->bo = NULL;
+        obj_surface->flags = 0;
+
+        if (obj_surface->free_private_data)
+            obj_surface->free_private_data(&obj_surface->private_data);
+    }
 
     return VA_STATUS_SUCCESS;
 }
@@ -1610,6 +1625,6 @@ __vaDriverInit_0_31(  VADriverContextP ctx )
                               sizeof(struct object_subpic), 
                               SUBPIC_ID_OFFSET);
     assert(result == 0);
-
+    
     return i965_Init(ctx);
 }

i965_drv_video/i965_drv_video.h

@@ -94,6 +94,9 @@ struct object_context
     struct decode_state decode_state;
 };
 
+#define SURFACE_REFERENCED      (1 << 0)
+#define SURFACE_DISPLAYED       (1 << 1)
+
 struct object_surface 
 {
     struct object_base base;
@@ -102,6 +105,7 @@ struct object_surface
     int width;
     int height;
     int size;
+    int flags;
     dri_bo *bo;
     void (*free_private_data)(void **data);
     void *private_data;
@@ -137,8 +141,6 @@ struct object_subpic
     dri_bo *bo;
 };
 
-
-
 struct i965_driver_data 
 {
     struct intel_driver_data intel;

i965_drv_video/i965_media.c

@@ -199,7 +199,7 @@ i965_media_pipeline_setup(VADriverContextP ctx, struct decode_state *decode_stat
 }
 
 static void 
-i965_media_decode_init(VADriverContextP ctx, VAProfile profile)
+i965_media_decode_init(VADriverContextP ctx, VAProfile profile, struct decode_state *decode_state)
 {
     int i;
     struct i965_driver_data *i965 = i965_driver_data(ctx);
@@ -250,13 +250,13 @@ i965_media_decode_init(VADriverContextP ctx, VAProfile profile)
     switch (profile) {
     case VAProfileMPEG2Simple:
     case VAProfileMPEG2Main:
-        i965_media_mpeg2_decode_init(ctx);
+        i965_media_mpeg2_decode_init(ctx, decode_state);
         break;
         
     case VAProfileH264Baseline:
     case VAProfileH264Main:
     case VAProfileH264High:
-        i965_media_h264_decode_init(ctx);
+        i965_media_h264_decode_init(ctx, decode_state);
         break;
 
     default:
@@ -273,7 +273,7 @@ i965_media_decode_picture(VADriverContextP ctx,
     struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct i965_media_state *media_state = &i965->media_state;
 
-    i965_media_decode_init(ctx, profile);
+    i965_media_decode_init(ctx, profile, decode_state);
     assert(media_state->media_states_setup);
     media_state->media_states_setup(ctx, decode_state);
     i965_media_pipeline_setup(ctx, decode_state);

i965_drv_video/i965_media_h264.c

@@ -702,21 +702,9 @@ i965_media_h264_states_setup(VADriverContextP ctx, struct decode_state *decode_s
     struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct i965_media_state *media_state = &i965->media_state;
     struct i965_h264_context *i965_h264_context;
-    VAPictureParameterBufferH264 *pic_param;
 
     assert(media_state->private_context);
     i965_h264_context = (struct i965_h264_context *)media_state->private_context;
-    assert(decode_state->pic_param && decode_state->pic_param->buffer);
-    pic_param = (VAPictureParameterBufferH264 *)decode_state->pic_param->buffer;
-
-    i965_h264_context->picture.width_in_mbs = ((pic_param->picture_width_in_mbs_minus1 + 1) & 0xff);
-    i965_h264_context->picture.height_in_mbs = ((pic_param->picture_height_in_mbs_minus1 + 1) & 0xff) / 
-        (1 + !!pic_param->pic_fields.bits.field_pic_flag); /* picture height */
-    i965_h264_context->picture.mbaff_frame_flag = (pic_param->seq_fields.bits.mb_adaptive_frame_field_flag &&
-                                                   !pic_param->pic_fields.bits.field_pic_flag);
-
-    i965_h264_context->avc_it_command_mb_info.mbs = (i965_h264_context->picture.width_in_mbs * 
-                                                     i965_h264_context->picture.height_in_mbs);
 
     i965_avc_bsd_pipeline(ctx, decode_state);
 
@@ -792,13 +780,14 @@ i965_media_h264_free_private_context(void **data)
 }
 
 void
-i965_media_h264_decode_init(VADriverContextP ctx)
+i965_media_h264_decode_init(VADriverContextP ctx, struct decode_state *decode_state)
 {
     struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct i965_media_state *media_state = &i965->media_state;
     struct i965_h264_context *i965_h264_context;
     dri_bo *bo;
     int i;
+    VAPictureParameterBufferH264 *pic_param;
 
     i965_h264_context = media_state->private_context;
 
@@ -829,7 +818,7 @@ i965_media_h264_decode_init(VADriverContextP ctx)
             struct media_kernel *kernel = &h264_avc_kernels[i];
             kernel->bo = dri_bo_alloc(i965->intel.bufmgr, 
                                       kernel->name, 
-                                      kernel->size, 64);
+                                      kernel->size, 0x1000);
             assert(kernel->bo);
             dri_bo_subdata(kernel->bo, 0, kernel->size, kernel->bin);
         }
@@ -865,12 +854,20 @@ i965_media_h264_decode_init(VADriverContextP ctx)
         media_state->media_objects = i965_media_h264_objects;
     }
 
-    i965_h264_context->enable_avc_ildb = 0;
+    assert(decode_state->pic_param && decode_state->pic_param->buffer);
+    pic_param = (VAPictureParameterBufferH264 *)decode_state->pic_param->buffer;
+    i965_h264_context->picture.width_in_mbs = ((pic_param->picture_width_in_mbs_minus1 + 1) & 0xff);
+    i965_h264_context->picture.height_in_mbs = ((pic_param->picture_height_in_mbs_minus1 + 1) & 0xff) / 
+        (1 + !!pic_param->pic_fields.bits.field_pic_flag); /* picture height */
+    i965_h264_context->picture.mbaff_frame_flag = (pic_param->seq_fields.bits.mb_adaptive_frame_field_flag &&
+                                                   !pic_param->pic_fields.bits.field_pic_flag);
+    i965_h264_context->avc_it_command_mb_info.mbs = (i965_h264_context->picture.width_in_mbs * 
+                                                     i965_h264_context->picture.height_in_mbs);
 
     dri_bo_unreference(i965_h264_context->avc_it_command_mb_info.bo);
     bo = dri_bo_alloc(i965->intel.bufmgr,
                       "avc it command mb info",
-                      0x80000 * (1 + i965_h264_context->use_avc_hw_scoreboard),  /* at least 522240 bytes */
+                      i965_h264_context->avc_it_command_mb_info.mbs * MB_CMD_IN_BYTES * (1 + i965_h264_context->use_avc_hw_scoreboard) + 4,
                       0x1000);
     assert(bo);
     i965_h264_context->avc_it_command_mb_info.bo = bo;
@@ -878,8 +875,10 @@ i965_media_h264_decode_init(VADriverContextP ctx)
     dri_bo_unreference(i965_h264_context->avc_it_data.bo);
     bo = dri_bo_alloc(i965->intel.bufmgr,
                       "avc it data",
-                      0x1000000, /* at least 16711680 bytes */
-                      4096);
+                      i965_h264_context->avc_it_command_mb_info.mbs * 
+                      0x800 * 
+                      (1 + !!pic_param->pic_fields.bits.field_pic_flag),
+                      0x1000);
     assert(bo);
     i965_h264_context->avc_it_data.bo = bo;
     i965_h264_context->avc_it_data.write_offset = 0;
@@ -891,8 +890,8 @@ i965_media_h264_decode_init(VADriverContextP ctx)
     dri_bo_unreference(i965_h264_context->avc_ildb_data.bo);
     bo = dri_bo_alloc(i965->intel.bufmgr,
                       "AVC-ILDB Data Buffer",
-                      0x100000, /* at least 1044480 bytes */
-                      64);
+                      i965_h264_context->avc_it_command_mb_info.mbs * 64 * 2,
+                      0x1000);
     assert(bo);
     i965_h264_context->avc_ildb_data.bo = bo;
 

i965_drv_video/i965_media_h264.h

@@ -60,6 +60,6 @@ struct i965_h264_context
     } fsid_list[16];
 };
 
-void i965_media_h264_decode_init(VADriverContextP ctx);
+void i965_media_h264_decode_init(VADriverContextP ctx, struct decode_state *decode_state);
 
 #endif /* _I965_MEDIA_H264_H_ */

i965_drv_video/i965_media_mpeg2.c

@@ -518,6 +518,15 @@ i965_media_mpeg2_surface_setup(VADriverContextP ctx,
     int w = obj_surface->width;
     int h = obj_surface->height;
 
+    if (obj_surface->bo == NULL) {
+        struct i965_driver_data *i965 = i965_driver_data(ctx);  
+        
+        obj_surface->bo = dri_bo_alloc(i965->intel.bufmgr,
+                                       "vaapi surface",
+                                       obj_surface->size,
+                                       0x1000);
+    }
+
     if (picture_structure == MPEG_FRAME) {
 	i965_media_mpeg2_surface_state(ctx, base_index + 0, obj_surface,
 					0, w, h, 
@@ -897,7 +906,7 @@ i965_media_mpeg2_free_private_context(void **data)
 }
 
 void 
-i965_media_mpeg2_decode_init(VADriverContextP ctx)
+i965_media_mpeg2_decode_init(VADriverContextP ctx, struct decode_state *decode_state)
 {
     struct i965_driver_data *i965 = i965_driver_data(ctx);
     struct i965_media_state *media_state = &i965->media_state;

i965_drv_video/i965_media_mpeg2.h

@@ -39,6 +39,6 @@
 
 struct decode_state;
 
-void i965_media_mpeg2_decode_init(VADriverContextP ctx);
+void i965_media_mpeg2_decode_init(VADriverContextP ctx, struct decode_state * decode_state);
 
 #endif /* _I965_MEDIA_MPEG2_H_ */

i965_drv_video/i965_render.c

@@ -921,6 +921,9 @@ i965_render_upload_constants(VADriverContextP ctx)
     struct i965_render_state *render_state = &i965->render_state;
     unsigned short *constant_buffer;
 
+    if (render_state->curbe.upload)
+        return;
+
     dri_bo_map(render_state->curbe.bo, 1);
     assert(render_state->curbe.bo->virtual);
     constant_buffer = render_state->curbe.bo->virtual;
@@ -931,6 +934,7 @@ i965_render_upload_constants(VADriverContextP ctx)
         *constant_buffer = 0;
 
     dri_bo_unmap(render_state->curbe.bo);
+    render_state->curbe.upload = 1;
 }
 
 static void
@@ -1508,7 +1512,7 @@ i965_render_init(VADriverContextP ctx)
         struct render_kernel *kernel = &render_kernels[i];
         kernel->bo = dri_bo_alloc(i965->intel.bufmgr, 
                                   kernel->name, 
-                                  kernel->size, 64);
+                                  kernel->size, 0x1000);
         assert(kernel->bo);
         dri_bo_subdata(kernel->bo, 0, kernel->size, kernel->bin);
     }
@@ -1518,6 +1522,7 @@ i965_render_init(VADriverContextP ctx)
                       "constant buffer",
                       4096, 64);
     assert(render_state->curbe.bo);
+    render_state->curbe.upload = 0;
 
     return True;
 }

i965_drv_video/i965_render.h

@@ -43,7 +43,6 @@ struct i965_render_state
     
     struct {
         dri_bo *state;
-        dri_bo *prog;
     } sf;
 
     struct {
@@ -52,7 +51,6 @@ struct i965_render_state
         dri_bo *surface[MAX_RENDER_SURFACES];
         dri_bo *binding_table;
         dri_bo *state;
-        dri_bo *prog;
     } wm;
 
     struct {
@@ -62,6 +60,7 @@ struct i965_render_state
 
     struct {
         dri_bo *bo;
+        int upload;
     } curbe;
 
     int interleaved_uv;