anv: Add support for AV1 video decoding on Gfx125 and Xe2

Reviewed-by: Hyunjun Ko <zzoon@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36015>

src/intel/vulkan/anv_private.h

@@ -6503,6 +6503,7 @@ enum anv_vid_mem_av1_types {
    ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_Y,
    ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_U,
    ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_V,
+   ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_ALIGNMENT_RW,
    ANV_VID_MEM_AV1_CDF_DEFAULTS_0,
    ANV_VID_MEM_AV1_CDF_DEFAULTS_1,
    ANV_VID_MEM_AV1_CDF_DEFAULTS_2,

src/intel/vulkan/anv_video.c

@@ -676,6 +676,7 @@ static const uint8_t av1_buffer_size[ANV_VID_MEM_AV1_MAX][4] = {
    { 9 ,   17  ,   11  ,    22 }, /* lrTileColYBuf, */
    { 5 ,   9   ,   6   ,    12 }, /* lrTileColUBuf, */
    { 5 ,   9   ,   6   ,    12 }, /* lrTileColVBuf, */
+   { 4,    8   ,   5   ,    10 }, /* lrTileColAlignBuffer, */
 };
 
 static const uint8_t av1_buffer_size_ext[ANV_VID_MEM_AV1_MAX][4] = {
@@ -710,6 +711,7 @@ static const uint8_t av1_buffer_size_ext[ANV_VID_MEM_AV1_MAX][4] = {
    { 2 ,    2    ,    2    ,    2 },  /* lrTileColYBuf, */
    { 1 ,    1    ,    1    ,    1 },  /* lrTileColUBuf, */
    { 1 ,    1    ,    1    ,    1 },  /* lrTileColVBuf, */
+   { 1,     1    ,    1    ,    1 },  /* lrTileColAlignBuffer, */
 };
 
 const uint32_t av1_mi_size_log2         = 2;
@@ -810,6 +812,7 @@ get_av1_video_session_mem_reqs(struct anv_video_session *vid,
       case ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_Y:
       case ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_U:
       case ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_V:
+      case ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_ALIGNMENT_RW:
       case ANV_VID_MEM_AV1_LOOP_RESTORATION_META_TILE_COLUMN:
          buffer_size = height_in_sb * av1_buffer_size[mem][buf_size_idx] +
             av1_buffer_size_ext[mem][buf_size_idx];

src/intel/vulkan/genX_cmd_video.c

@@ -1857,6 +1857,15 @@ anv_av1_decode_video_tile(struct anv_cmd_buffer *cmd_buffer,
       buf.LoopRestorationFilterTileColumnVBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
          .MOCS = anv_mocs(cmd_buffer->device, vid->vid_mem[ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_V].mem->bo, 0),
       };
+#if GFX_VER >= 20
+      buf.LoopRestorationFilterTileColumnAlignmentReadWriteBufferAddress = (struct anv_address) {
+         vid->vid_mem[ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_ALIGNMENT_RW].mem->bo,
+         vid->vid_mem[ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_ALIGNMENT_RW].offset
+      };
+      buf.LoopRestorationFilterTileColumnAlignmentReadWriteBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, vid->vid_mem[ANV_VID_MEM_AV1_LOOP_RESTORATION_FILTER_TILE_COLUMN_ALIGNMENT_RW].mem->bo, 0),
+      };
+#endif
 
       struct anv_bo *ref_bo = NULL;
       struct anv_bo *collocated_bo = NULL;
@@ -1944,6 +1953,43 @@ anv_av1_decode_video_tile(struct anv_cmd_buffer *cmd_buffer,
       buf.DecodedBlockDataStreamoutBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
          .MOCS = anv_mocs(cmd_buffer->device, vid->vid_mem[ANV_VID_MEM_AV1_DBD_BUFFER].mem->bo, 0),
       };
+#if GFX_VERx10 >= 125
+      buf.OriginalUncompressedPictureSourceBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.DownscaledUncompressedPictureSourceBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.TileSizeStreamoutBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.TileStatisticsStreamoutBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.CUStreamoutBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.SSELineReadWriteBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.SSETileLineReadWriteBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.PostCDEFPixelsBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+#endif
+#if GFX_VER >= 20
+      buf.FilmGrainInjectedOutputFrameBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.FilmGrainSampleTemplateAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+      buf.FilmGrainTileColumnDataReadWriteBufferAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
+      };
+#endif
    };
 
    anv_batch_emit(&cmd_buffer->batch, GENX(AVP_IND_OBJ_BASE_ADDR_STATE), ind) {
@@ -1953,11 +1999,9 @@ anv_av1_decode_video_tile(struct anv_cmd_buffer *cmd_buffer,
          .MOCS = anv_mocs(cmd_buffer->device, src_buffer->address.bo, 0),
       };
 #if GFX_VERx10 >= 125
-      /* FIXME.
+      ind.AVPIndirectCUObjectMemoryAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
-      ind.AVPIndirectCUObjectAddressAttributes = (struct GENX(MEMORYADDRESSATTRIBUTES)) {
+         .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
-      .MOCS = anv_mocs(cmd_buffer->device, NULL, 0),
       };
-      */
 #endif
    }