[FFmpeg-devel] [PATCH, v2] avcodec/d3d12va_encode: texture array support for HEVC
Araz Iusubov
primeadvice at gmail.com
Fri May 16 12:43:56 EEST 2025
This patch adds support for the texture array feature
used by AMD boards in the D3D12 HEVC encoder.
In texture array mode, a single texture array is shared for all
reference and reconstructed pictures using different subresources.
The implementation ensures compatibility
and has been successfully tested on AMD, Intel, and NVIDIA GPUs.
v2 updates:
1. The reference to MaxL1ReferencesForB for the H.264 codec
was updated to use the corresponding H.264 field instead of the HEVC one.
2. Max_subresource_array_size calculation was adjusted
by removing the D3D12VA_VIDEO_ENC_ASYNC_DEPTH offset.
---
libavcodec/d3d12va_encode.c | 241 +++++++++++++++++++++++++------
libavcodec/d3d12va_encode.h | 29 ++++
libavcodec/d3d12va_encode_hevc.c | 5 +-
3 files changed, 231 insertions(+), 44 deletions(-)
diff --git a/libavcodec/d3d12va_encode.c b/libavcodec/d3d12va_encode.c
index 4d738200fe..d428ad1fd8 100644
--- a/libavcodec/d3d12va_encode.c
+++ b/libavcodec/d3d12va_encode.c
@@ -264,6 +264,11 @@ static int d3d12va_encode_issue(AVCodecContext *avctx,
av_log(avctx, AV_LOG_DEBUG, "Input surface is %p.\n", pic->input_surface->texture);
+ if (ctx->is_texture_array) {
+ base_pic->recon_image->data[0] = ctx->texture_array_frame;
+ pic->subresource_index = (ctx->subresource_used_index++) % ctx->max_subresource_array_size;
+ }
+
pic->recon_surface = (AVD3D12VAFrame *)base_pic->recon_image->data[0];
av_log(avctx, AV_LOG_DEBUG, "Recon surface is %p.\n",
pic->recon_surface->texture);
@@ -325,11 +330,28 @@ static int d3d12va_encode_issue(AVCodecContext *avctx,
goto fail;
}
+ if (ctx->is_texture_array) {
+ d3d12_refs.pSubresources = av_calloc(d3d12_refs.NumTexture2Ds,
+ sizeof(*d3d12_refs.pSubresources));
+ if (!d3d12_refs.pSubresources) {
+ err = AVERROR(ENOMEM);
+ goto fail;
+ }
+ }
+
i = 0;
- for (j = 0; j < base_pic->nb_refs[0]; j++)
- d3d12_refs.ppTexture2Ds[i++] = ((D3D12VAEncodePicture *)base_pic->refs[0][j]->priv)->recon_surface->texture;
- for (j = 0; j < base_pic->nb_refs[1]; j++)
- d3d12_refs.ppTexture2Ds[i++] = ((D3D12VAEncodePicture *)base_pic->refs[1][j]->priv)->recon_surface->texture;
+ for (j = 0; j < base_pic->nb_refs[0]; j++) {
+ d3d12_refs.ppTexture2Ds[i] = ((D3D12VAEncodePicture *)base_pic->refs[0][j]->priv)->recon_surface->texture;
+ if (ctx->is_texture_array)
+ d3d12_refs.pSubresources[i] = ((D3D12VAEncodePicture *)base_pic->refs[0][j]->priv)->subresource_index;
+ i++;
+ }
+ for (j = 0; j < base_pic->nb_refs[1]; j++) {
+ d3d12_refs.ppTexture2Ds[i] = ((D3D12VAEncodePicture *)base_pic->refs[1][j]->priv)->recon_surface->texture;
+ if (ctx->is_texture_array)
+ d3d12_refs.pSubresources[i] = ((D3D12VAEncodePicture *)base_pic->refs[1][j]->priv)->subresource_index;
+ i++;
+ }
}
input_args.PictureControlDesc.IntraRefreshFrameIndex = 0;
@@ -343,7 +365,11 @@ static int d3d12va_encode_issue(AVCodecContext *avctx,
output_args.Bitstream.pBuffer = pic->output_buffer;
output_args.Bitstream.FrameStartOffset = pic->aligned_header_size;
output_args.ReconstructedPicture.pReconstructedPicture = pic->recon_surface->texture;
- output_args.ReconstructedPicture.ReconstructedPictureSubresource = 0;
+ if (ctx->is_texture_array) {
+ output_args.ReconstructedPicture.ReconstructedPictureSubresource = pic->subresource_index;
+ } else {
+ output_args.ReconstructedPicture.ReconstructedPictureSubresource = 0;
+ }
output_args.EncoderOutputMetadata.pBuffer = pic->encoded_metadata;
output_args.EncoderOutputMetadata.Offset = 0;
@@ -381,35 +407,87 @@ static int d3d12va_encode_issue(AVCodecContext *avctx,
}, \
}
+#define TRANSITION_BARRIER_SUBRESOURCE(res, subres,before, after) \
+ (D3D12_RESOURCE_BARRIER) { \
+ .Type = D3D12_RESOURCE_BARRIER_TYPE_TRANSITION, \
+ .Flags = D3D12_RESOURCE_BARRIER_FLAG_NONE, \
+ .Transition = { \
+ .pResource = res, \
+ .Subresource = subres, \
+ .StateBefore = before, \
+ .StateAfter = after, \
+ }, \
+ }
+
barriers[0] = TRANSITION_BARRIER(pic->input_surface->texture,
D3D12_RESOURCE_STATE_COMMON,
D3D12_RESOURCE_STATE_VIDEO_ENCODE_READ);
barriers[1] = TRANSITION_BARRIER(pic->output_buffer,
D3D12_RESOURCE_STATE_COMMON,
D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE);
- barriers[2] = TRANSITION_BARRIER(pic->recon_surface->texture,
+ barriers[2] = TRANSITION_BARRIER(pic->encoded_metadata,
D3D12_RESOURCE_STATE_COMMON,
D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE);
- barriers[3] = TRANSITION_BARRIER(pic->encoded_metadata,
+ barriers[3] = TRANSITION_BARRIER(pic->resolved_metadata,
D3D12_RESOURCE_STATE_COMMON,
D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE);
- barriers[4] = TRANSITION_BARRIER(pic->resolved_metadata,
- D3D12_RESOURCE_STATE_COMMON,
- D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE);
-
- ID3D12VideoEncodeCommandList2_ResourceBarrier(cmd_list, 5, barriers);
-
- if (d3d12_refs.NumTexture2Ds) {
- D3D12_RESOURCE_BARRIER refs_barriers[3];
-
- for (i = 0; i < d3d12_refs.NumTexture2Ds; i++)
- refs_barriers[i] = TRANSITION_BARRIER(d3d12_refs.ppTexture2Ds[i],
- D3D12_RESOURCE_STATE_COMMON,
- D3D12_RESOURCE_STATE_VIDEO_ENCODE_READ);
-
- ID3D12VideoEncodeCommandList2_ResourceBarrier(cmd_list, d3d12_refs.NumTexture2Ds,
- refs_barriers);
+ ID3D12VideoEncodeCommandList2_ResourceBarrier(cmd_list, 4, barriers);
+
+ //set transit barriers for reference pic and recon pic
+ int barriers_ref_index = 0;
+ D3D12_RESOURCE_BARRIER *barriers_ref = NULL;
+ if (ctx->is_texture_array) {
+ barriers_ref = av_calloc(ctx->max_subresource_array_size * ctx->plane_count,
+ sizeof(D3D12_RESOURCE_BARRIER));
+ } else {
+ barriers_ref = av_calloc(MAX_DPB_SIZE,sizeof(D3D12_RESOURCE_BARRIER));
+ }
+
+ if (ctx->is_texture_array) {
+ // In Texture array mode, the D3D12 uses the same texture array for all the input
+ // reference pics in ppTexture2Ds and also for the pReconstructedPicture output allocations,
+ //just different subresources.
+ D3D12_RESOURCE_DESC references_tex_array_desc = { 0 };
+ pic->recon_surface->texture->lpVtbl->GetDesc(pic->recon_surface->texture, &references_tex_array_desc);
+
+ for (uint32_t reference_subresource = 0; reference_subresource < references_tex_array_desc.DepthOrArraySize;
+ reference_subresource++) {
+
+ //D3D12 DecomposeSubresource
+ uint32_t mip_slice, plane_slice, array_slice, array_size;
+ array_size = references_tex_array_desc.DepthOrArraySize;
+ mip_slice = reference_subresource % references_tex_array_desc.MipLevels;
+ array_slice = (reference_subresource / references_tex_array_desc.MipLevels) % array_size;
+
+ for (plane_slice = 0; plane_slice < ctx->plane_count; plane_slice++) {
+ //Calculate the subresource index
+ uint32_t planeOutputSubresource = mip_slice + array_slice * references_tex_array_desc.MipLevels +
+ plane_slice * references_tex_array_desc.MipLevels * array_size;
+
+ if (reference_subresource == pic->subresource_index) {
+ barriers_ref[barriers_ref_index++] = TRANSITION_BARRIER_SUBRESOURCE(pic->recon_surface->texture, planeOutputSubresource,
+ D3D12_RESOURCE_STATE_COMMON,
+ D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE);
+ } else {
+ barriers_ref[barriers_ref_index++] = TRANSITION_BARRIER_SUBRESOURCE(pic->recon_surface->texture, planeOutputSubresource,
+ D3D12_RESOURCE_STATE_COMMON,
+ D3D12_RESOURCE_STATE_VIDEO_ENCODE_READ);
+ }
+ }
+ }
+ } else {
+ barriers_ref[barriers_ref_index++] = TRANSITION_BARRIER(pic->recon_surface->texture,
+ D3D12_RESOURCE_STATE_COMMON,
+ D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE);
+
+ if (d3d12_refs.NumTexture2Ds) {
+ for (i = 0; i < d3d12_refs.NumTexture2Ds; i++)
+ barriers_ref[barriers_ref_index++] = TRANSITION_BARRIER(d3d12_refs.ppTexture2Ds[i],
+ D3D12_RESOURCE_STATE_COMMON,
+ D3D12_RESOURCE_STATE_VIDEO_ENCODE_READ);
+ }
}
+ ID3D12VideoEncodeCommandList2_ResourceBarrier(cmd_list, barriers_ref_index, barriers_ref);
ID3D12VideoEncodeCommandList2_EncodeFrame(cmd_list, ctx->encoder, ctx->encoder_heap,
&input_args, &output_args);
@@ -422,16 +500,15 @@ static int d3d12va_encode_issue(AVCodecContext *avctx,
ID3D12VideoEncodeCommandList2_ResolveEncoderOutputMetadata(cmd_list, &input_metadata, &output_metadata);
- if (d3d12_refs.NumTexture2Ds) {
- D3D12_RESOURCE_BARRIER refs_barriers[3];
-
- for (i = 0; i < d3d12_refs.NumTexture2Ds; i++)
- refs_barriers[i] = TRANSITION_BARRIER(d3d12_refs.ppTexture2Ds[i],
- D3D12_RESOURCE_STATE_VIDEO_ENCODE_READ,
- D3D12_RESOURCE_STATE_COMMON);
-
- ID3D12VideoEncodeCommandList2_ResourceBarrier(cmd_list, d3d12_refs.NumTexture2Ds,
- refs_barriers);
+ //swap the barriers_ref transition state
+ if (barriers_ref_index > 0) {
+ for (i = 0; i < barriers_ref_index; i++) {
+ D3D12_RESOURCE_STATES temp_statue = barriers_ref[i].Transition.StateBefore;
+ barriers_ref[i].Transition.StateBefore = barriers_ref[i].Transition.StateAfter;
+ barriers_ref[i].Transition.StateAfter = temp_statue;
+ }
+ ID3D12VideoEncodeCommandList2_ResourceBarrier(cmd_list, barriers_ref_index,
+ barriers_ref);
}
barriers[0] = TRANSITION_BARRIER(pic->input_surface->texture,
@@ -440,17 +517,14 @@ static int d3d12va_encode_issue(AVCodecContext *avctx,
barriers[1] = TRANSITION_BARRIER(pic->output_buffer,
D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE,
D3D12_RESOURCE_STATE_COMMON);
- barriers[2] = TRANSITION_BARRIER(pic->recon_surface->texture,
- D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE,
- D3D12_RESOURCE_STATE_COMMON);
- barriers[3] = TRANSITION_BARRIER(pic->encoded_metadata,
+ barriers[2] = TRANSITION_BARRIER(pic->encoded_metadata,
D3D12_RESOURCE_STATE_VIDEO_ENCODE_READ,
D3D12_RESOURCE_STATE_COMMON);
- barriers[4] = TRANSITION_BARRIER(pic->resolved_metadata,
+ barriers[3] = TRANSITION_BARRIER(pic->resolved_metadata,
D3D12_RESOURCE_STATE_VIDEO_ENCODE_WRITE,
D3D12_RESOURCE_STATE_COMMON);
- ID3D12VideoEncodeCommandList2_ResourceBarrier(cmd_list, 5, barriers);
+ ID3D12VideoEncodeCommandList2_ResourceBarrier(cmd_list, 4, barriers);
hr = ID3D12VideoEncodeCommandList2_Close(cmd_list);
if (FAILED(hr)) {
@@ -489,6 +563,14 @@ static int d3d12va_encode_issue(AVCodecContext *avctx,
if (d3d12_refs.ppTexture2Ds)
av_freep(&d3d12_refs.ppTexture2Ds);
+ if (ctx->is_texture_array) {
+ if (d3d12_refs.pSubresources)
+ av_freep(&d3d12_refs.pSubresources);
+ }
+
+ if (barriers_ref)
+ av_freep(&barriers_ref);
+
return 0;
fail:
@@ -498,6 +580,14 @@ fail:
if (d3d12_refs.ppTexture2Ds)
av_freep(&d3d12_refs.ppTexture2Ds);
+ if (ctx->is_texture_array) {
+ if (d3d12_refs.pSubresources)
+ av_freep(&d3d12_refs.pSubresources);
+ }
+
+ if (barriers_ref)
+ av_freep(&barriers_ref);
+
if (ctx->codec->free_picture_params)
ctx->codec->free_picture_params(pic);
@@ -1088,13 +1178,15 @@ static int d3d12va_encode_init_gop_structure(AVCodecContext *avctx)
switch (ctx->codec->d3d12_codec) {
case D3D12_VIDEO_ENCODER_CODEC_H264:
ref_l0 = FFMIN(support.PictureSupport.pH264Support->MaxL0ReferencesForP,
- support.PictureSupport.pH264Support->MaxL1ReferencesForB);
+ support.PictureSupport.pH264Support->MaxL1ReferencesForB ?
+ support.PictureSupport.pH264Support->MaxL1ReferencesForB : UINT_MAX);
ref_l1 = support.PictureSupport.pH264Support->MaxL1ReferencesForB;
break;
case D3D12_VIDEO_ENCODER_CODEC_HEVC:
ref_l0 = FFMIN(support.PictureSupport.pHEVCSupport->MaxL0ReferencesForP,
- support.PictureSupport.pHEVCSupport->MaxL1ReferencesForB);
+ support.PictureSupport.pHEVCSupport->MaxL1ReferencesForB ?
+ support.PictureSupport.pHEVCSupport->MaxL1ReferencesForB : UINT_MAX);
ref_l1 = support.PictureSupport.pHEVCSupport->MaxL1ReferencesForB;
break;
@@ -1336,6 +1428,47 @@ fail:
return err;
}
+static int d3d12va_create_texture_array(AVHWFramesContext *ctx, D3D12VAEncodeContext *encode_context)
+{
+ AVD3D12VAFramesContext *hwctx = ctx->hwctx;
+ AVD3D12VADeviceContext *device_hwctx = ctx->device_ctx->hwctx;
+
+ AVD3D12VAFrame *frame;
+ D3D12_HEAP_PROPERTIES props = { .Type = D3D12_HEAP_TYPE_DEFAULT };
+
+ encode_context->max_subresource_array_size = MAX_DPB_SIZE + 1;
+
+ D3D12_RESOURCE_DESC desc = {
+ .Dimension = D3D12_RESOURCE_DIMENSION_TEXTURE2D,
+ .Alignment = 0,
+ .Width = ctx->width,
+ .Height = ctx->height,
+ .DepthOrArraySize = encode_context->max_subresource_array_size,
+ .MipLevels = 1,
+ .Format = hwctx->format,
+ .SampleDesc = {.Count = 1, .Quality = 0 },
+ .Layout = D3D12_TEXTURE_LAYOUT_UNKNOWN,
+ .Flags = hwctx->flags,
+ };
+
+ frame = av_mallocz(sizeof(AVD3D12VAFrame));
+ if (!frame)
+ return AVERROR(ENOMEM);
+
+ if (FAILED(ID3D12Device_CreateCommittedResource(device_hwctx->device, &props, D3D12_HEAP_FLAG_NONE, &desc,
+ D3D12_RESOURCE_STATE_COMMON, NULL, &IID_ID3D12Resource, (void **)&frame->texture))) {
+ av_log(ctx, AV_LOG_ERROR, "Could not create the texture\n");
+ return AVERROR(EINVAL);
+ }
+
+ ID3D12Device_CreateFence(device_hwctx->device, 0, D3D12_FENCE_FLAG_NONE,
+ &IID_ID3D12Fence, (void **)&frame->sync_ctx.fence);
+
+ frame->sync_ctx.event = CreateEvent(NULL, FALSE, FALSE, NULL);
+ encode_context->texture_array_frame = frame;
+ return 0;
+}
+
static int d3d12va_encode_create_recon_frames(AVCodecContext *avctx)
{
FFHWBaseEncodeContext *base_ctx = avctx->priv_data;
@@ -1394,6 +1527,7 @@ int ff_d3d12va_encode_init(AVCodecContext *avctx)
FFHWBaseEncodeContext *base_ctx = avctx->priv_data;
D3D12VAEncodeContext *ctx = avctx->priv_data;
D3D12_FEATURE_DATA_VIDEO_FEATURE_AREA_SUPPORT support = { 0 };
+ D3D12_FEATURE_DATA_FORMAT_INFO format_info = {0};
int err;
HRESULT hr;
@@ -1429,6 +1563,15 @@ int ff_d3d12va_encode_init(AVCodecContext *avctx)
goto fail;
}
+ format_info.Format = ((AVD3D12VAFramesContext*)base_ctx->input_frames->hwctx)->format;
+ if (FAILED(ID3D12VideoDevice_CheckFeatureSupport(ctx->hwctx->device, D3D12_FEATURE_FORMAT_INFO,
+ &format_info, sizeof(format_info)))) {
+ av_log(avctx, AV_LOG_ERROR, "Failed to query format plane count: 0x%x\n", hr);
+ err = AVERROR_EXTERNAL;
+ goto fail;
+ }
+ ctx->plane_count = format_info.PlaneCount;
+
err = d3d12va_encode_set_profile(avctx);
if (err < 0)
goto fail;
@@ -1485,6 +1628,10 @@ int ff_d3d12va_encode_init(AVCodecContext *avctx)
goto fail;
}
+ if (ctx->is_texture_array) {
+ d3d12va_create_texture_array(base_ctx->recon_frames, avctx->priv_data);
+ }
+
base_ctx->output_delay = base_ctx->b_per_p;
base_ctx->decode_delay = base_ctx->max_b_depth;
@@ -1528,6 +1675,18 @@ int ff_d3d12va_encode_close(AVCodecContext *avctx)
av_buffer_pool_uninit(&ctx->output_buffer_pool);
+ if (ctx->is_texture_array) {
+ ID3D12Resource *pResource = ctx->texture_array_frame->texture;
+ if (pResource) {
+ D3D12_OBJECT_RELEASE(pResource);
+ ctx->texture_array_frame->texture = NULL;
+ }
+ D3D12_OBJECT_RELEASE(ctx->texture_array_frame->sync_ctx.fence);
+ if (ctx->texture_array_frame->sync_ctx.event)
+ CloseHandle(ctx->texture_array_frame->sync_ctx.event);
+ av_free(ctx->texture_array_frame);
+ }
+
D3D12_OBJECT_RELEASE(ctx->command_list);
D3D12_OBJECT_RELEASE(ctx->command_queue);
diff --git a/libavcodec/d3d12va_encode.h b/libavcodec/d3d12va_encode.h
index 3b0b8153d5..fc31857f1a 100644
--- a/libavcodec/d3d12va_encode.h
+++ b/libavcodec/d3d12va_encode.h
@@ -52,6 +52,8 @@ typedef struct D3D12VAEncodePicture {
ID3D12Resource *encoded_metadata;
ID3D12Resource *resolved_metadata;
+ int subresource_index;
+
D3D12_VIDEO_ENCODER_PICTURE_CONTROL_CODEC_DATA pic_ctl;
int fence_value;
@@ -189,6 +191,33 @@ typedef struct D3D12VAEncodeContext {
*/
AVBufferPool *output_buffer_pool;
+ /**
+ * Flag indicates if the HW is texture array mode.
+ */
+ int is_texture_array;
+
+ /**
+ * In texture array mode, the D3D12 uses the same texture array for all the input
+ * reference pics in ppTexture2Ds and also for the pReconstructedPicture output
+ * allocations, just different subresources.
+ */
+ AVD3D12VAFrame *texture_array_frame;
+
+ /**
+ * The max number of subresources in the texture array.
+ */
+ int max_subresource_array_size;
+
+ /**
+ * The used subresource index for pic in the texture array.
+ */
+ int subresource_used_index;
+
+ /**
+ * The number of planes in the input DXGI FORMAT .
+ */
+ int plane_count;
+
/**
* D3D12 video encoder.
*/
diff --git a/libavcodec/d3d12va_encode_hevc.c b/libavcodec/d3d12va_encode_hevc.c
index 938ba01f54..7e1d973f7e 100644
--- a/libavcodec/d3d12va_encode_hevc.c
+++ b/libavcodec/d3d12va_encode_hevc.c
@@ -280,9 +280,8 @@ static int d3d12va_encode_hevc_init_sequence_params(AVCodecContext *avctx)
}
if (support.SupportFlags & D3D12_VIDEO_ENCODER_SUPPORT_FLAG_RECONSTRUCTED_FRAMES_REQUIRE_TEXTURE_ARRAYS) {
- av_log(avctx, AV_LOG_ERROR, "D3D12 video encode on this device requires texture array support, "
- "but it's not implemented.\n");
- return AVERROR_PATCHWELCOME;
+ ctx->is_texture_array = 1;
+ av_log(avctx, AV_LOG_DEBUG, "D3D12 video encode on this device uses texture array mode.\n");
}
desc = av_pix_fmt_desc_get(base_ctx->input_frames->sw_format);
--
2.45.2.windows.1
More information about the ffmpeg-devel
mailing list