[FFmpeg-devel] [PATCH] avcodec/cuviddec: update amount of decoder surfaces from within sequence decode callback

[Anton Khirnov]

Quoting Roman Arzumanyan (2023-06-05 09:30:07)
> Hello,
> 
> This patch reduces vRAM usage by cuvid decoder implementation.
> The number of surfaces used for decoding is updated within the parser
> sequence decode callback.
> Also the "surfaces" AVDictionary option specific to cuvid was removed in
> favor of "extra_hw_surfaces".

This can break existing workflows, you should deprecated the option
instead and only remove it after some time has passed.

> 
> vRAM consumption was tested on various videos and savings are between 1%
> for 360p resolution up to 21% for some 1080p H.264 videos.
> Decoding performance was tested on various H.264 and H.265 videos in
> different resolutions from 360p and higher, no performance penalty was
> found.
> 
> From 32a1b016e88fa40b983318d4583750ef250a78d9 Mon Sep 17 00:00:00 2001
> From: Roman Arzumanyan <r.arzumanyan at visionlabs.ai>
> Date: Thu, 1 Jun 2023 11:17:39 +0300
> Subject: [PATCH] libavcodec/cuviddec: determine DPB size from within cuvid
>  parser
> 
> ---
>  libavcodec/cuviddec.c | 29 +++++++++++++++++++++++++++--
>  1 file changed, 27 insertions(+), 2 deletions(-)
> 
> diff --git a/libavcodec/cuviddec.c b/libavcodec/cuviddec.c
> index 3d43bbd466..759ed49870 100644
> --- a/libavcodec/cuviddec.c
> +++ b/libavcodec/cuviddec.c
> @@ -115,6 +115,12 @@ typedef struct CuvidParsedFrame
>  
>  #define CHECK_CU(x) FF_CUDA_CHECK_DL(avctx, ctx->cudl, x)
>  
> +// NV recommends [2;4] range
> +#define CUVID_MAX_DISPLAY_DELAY (4)
> +
> +// Actual DPB size will be determined by parser.
> +#define CUVID_DEFAULT_NUM_SURFACES (CUVID_MAX_DISPLAY_DELAY + 1)
> +
>  static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* format)
>  {
>      AVCodecContext *avctx = opaque;
> @@ -309,6 +315,25 @@ static int CUDAAPI cuvid_handle_video_sequence(void *opaque, CUVIDEOFORMAT* form
>          return 0;
>      }
>  
> +    if (ctx->nb_surfaces < format->min_num_decode_surfaces + 3)
> +        ctx->nb_surfaces = format->min_num_decode_surfaces + 3;

FFMAX()

> +
> +    if (avctx->extra_hw_frames > 0)
> +        ctx->nb_surfaces += avctx->extra_hw_frames;
> +
> +    if (0 > av_fifo_realloc2(ctx->frame_queue, ctx->nb_surfaces * sizeof(CuvidParsedFrame))) {

this is the old deprecated AVFifoBuffer API, you cannot use it with
AVFifo objects

you should also forward the actual error code

> +        av_log(avctx, AV_LOG_ERROR, "Failed to recreate frame queue on video sequence callback\n");
> +        ctx->internal_error = AVERROR(EINVAL);
> +        return 0;
> +    }
> +
> +    ctx->key_frame = av_realloc_array(ctx->key_frame, ctx->nb_surfaces, sizeof(int));
> +    if (!ctx->key_frame) {
> +        av_log(avctx, AV_LOG_ERROR, "Failed to recreate key frame queue on video sequence callback\n");
> +        ctx->internal_error = AVERROR(EINVAL);

Leaks key_frame on failure and should be ENOMEM.

-- 
Anton Khirnov