[FFmpeg-devel] [PATCH v2] nvdec/vc1: add marker insertion logic

[Timo Rothenpieler]

On 08/04/2025 22:50, averne wrote:
> This mirrors existing code in d3dxx and dxva hwaccels
> 
> ---
> Changes since v1: use the codec-specific handler
> for wmv3.
> 
> Insert the relevant marker into the bitstream on
> slice submission.
> This is analogous to the logic found in DXVA and
> D3D hwaccels.
> 
> Fixes decoding of various VC-1 streams, eg.:
> https://drive.google.com/file/d/1WJyiRhcdU4FHTW3sVMitS7UdrZM1NBy-/view?usp=sharing
> 
> This was investigated using my nvdec tracing tool:
> https://github.com/averne/NvdecTrace
> 
> Signed-off-by: averne <averne381 at gmail.com>
> ---
>   libavcodec/nvdec_vc1.c | 45 ++++++++++++++++++++++++++++++++++++++++++++-
>   1 file changed, 44 insertions(+), 1 deletion(-)
> 
> diff --git a/libavcodec/nvdec_vc1.c b/libavcodec/nvdec_vc1.c
> index fbfba1ecb43421573ef8fea1e37a2425c272edc9..a88327833df55054a42e7ced8cdd4c99fa6b6584 100644
> --- a/libavcodec/nvdec_vc1.c
> +++ b/libavcodec/nvdec_vc1.c
> @@ -22,6 +22,7 @@
>   
>   #include "config_components.h"
>   
> +#include "libavutil/mem.h"
>   #include "avcodec.h"
>   #include "hwaccel_internal.h"
>   #include "internal.h"
> @@ -107,6 +108,48 @@ static int nvdec_vc1_start_frame(AVCodecContext *avctx, const uint8_t *buffer, u
>       return 0;
>   }
>   
> +static int nvdec_vc1_decode_slice(AVCodecContext *avctx, const uint8_t *buffer,
> +                                  uint32_t size)
> +{
> +    NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
> +    const VC1Context *v = avctx->priv_data;
> +    uint32_t marker;
> +    int marker_size;
> +    void *tmp;
> +
> +    if (ctx->bitstream_len)
> +        marker = VC1_CODE_SLICE;
> +    else if (v->profile == PROFILE_ADVANCED && v->fcm == ILACE_FIELD && v->second_field)
> +        marker = VC1_CODE_FIELD;
> +    else
> +        marker = VC1_CODE_FRAME;
> +
> +    /* Only insert the marker if not already present in the bitstream */
> +    marker_size = (size >= sizeof(marker) && AV_RB32(buffer) != marker) ? sizeof(marker) : 0;
> +
> +    tmp = av_fast_realloc(ctx->bitstream_internal, &ctx->bitstream_allocated,
> +                          ctx->bitstream_len + size + marker_size);
> +    if (!tmp)
> +        return AVERROR(ENOMEM);
> +    ctx->bitstream = ctx->bitstream_internal = tmp;
> +
> +    tmp = av_fast_realloc(ctx->slice_offsets, &ctx->slice_offsets_allocated,
> +                          (ctx->nb_slices + 1) * sizeof(*ctx->slice_offsets));
> +    if (!tmp)
> +        return AVERROR(ENOMEM);
> +    ctx->slice_offsets = tmp;
> +
> +    if (marker_size)
> +        AV_WB32(ctx->bitstream_internal + ctx->bitstream_len, marker);
> +
> +    memcpy(ctx->bitstream_internal + ctx->bitstream_len + marker_size, buffer, size);
> +    ctx->slice_offsets[ctx->nb_slices] = ctx->bitstream_len ;
> +    ctx->bitstream_len += size + marker_size;
> +    ctx->nb_slices++;
> +
> +    return 0;
> +}
> +
>   static int nvdec_vc1_frame_params(AVCodecContext *avctx,
>                                     AVBufferRef *hw_frames_ctx)
>   {
> @@ -121,7 +164,7 @@ const FFHWAccel ff_vc1_nvdec_hwaccel = {
>       .p.pix_fmt            = AV_PIX_FMT_CUDA,
>       .start_frame          = nvdec_vc1_start_frame,
>       .end_frame            = ff_nvdec_simple_end_frame,
> -    .decode_slice         = ff_nvdec_simple_decode_slice,
> +    .decode_slice         = nvdec_vc1_decode_slice,
>       .frame_params         = nvdec_vc1_frame_params,
>       .init                 = ff_nvdec_decode_init,
>       .uninit               = ff_nvdec_decode_uninit,
> 
> ---
> base-commit: f76195ff656d6bea68feee783160652e2b3e3d60
> change-id: 20250408-nvdec-vc1-fix-7fb117513851
> 
> Best regards,

I'll give this a test later and apply if everything works fine.