[FFmpeg-devel] [RFC] ffmpeg: default to single thread when hwaccel is active

[Dennis Mungai]

On Wed, 29 Sept 2021 at 20:02, Timo Rothenpieler <timo at rothenpieler.org>
wrote:

> ---
>  fftools/ffmpeg.c | 16 ++++++++++------
>  1 file changed, 10 insertions(+), 6 deletions(-)
>
> Not sure if there is any downside to this.
> Threading for hwaccel does not make a whole lot of sense, and at least
> in case of nvdec wastes a lot of VRAM for no performance gain, and
> specially on high core count systems by default can even exhaust the
> maximum frame pool size.
>
> diff --git a/fftools/ffmpeg.c b/fftools/ffmpeg.c
> index 98c2421938..d007d55173 100644
> --- a/fftools/ffmpeg.c
> +++ b/fftools/ffmpeg.c
> @@ -2995,12 +2995,6 @@ FF_ENABLE_DEPRECATION_WARNINGS
>           * audio, and video decoders such as cuvid or mediacodec */
>          ist->dec_ctx->pkt_timebase = ist->st->time_base;
>
> -        if (!av_dict_get(ist->decoder_opts, "threads", NULL, 0))
> -            av_dict_set(&ist->decoder_opts, "threads", "auto", 0);
> -        /* Attached pics are sparse, therefore we would not want to delay
> their decoding till EOF. */
> -        if (ist->st->disposition & AV_DISPOSITION_ATTACHED_PIC)
> -            av_dict_set(&ist->decoder_opts, "threads", "1", 0);
> -
>          ret = hw_device_setup_for_decode(ist);
>          if (ret < 0) {
>              snprintf(error, error_len, "Device setup failed for "
> @@ -3009,6 +3003,16 @@ FF_ENABLE_DEPRECATION_WARNINGS
>              return ret;
>          }
>
> +        if (!av_dict_get(ist->decoder_opts, "threads", NULL, 0)) {
> +            if (ist->dec_ctx->hw_device_ctx)
> +                av_dict_set(&ist->decoder_opts, "threads", "1", 0);
> +            else
> +                av_dict_set(&ist->decoder_opts, "threads", "auto", 0);
> +        }
> +        /* Attached pics are sparse, therefore we would not want to delay
> their decoding till EOF. */
> +        if (ist->st->disposition & AV_DISPOSITION_ATTACHED_PIC)
> +            av_dict_set(&ist->decoder_opts, "threads", "1", 0);
> +
>          if ((ret = avcodec_open2(ist->dec_ctx, codec,
> &ist->decoder_opts)) < 0) {
>              if (ret == AVERROR_EXPERIMENTAL)
>                  abort_codec_experimental(codec, 0);
> --
> 2.25.1
>
>
>
A potential downside to this would be on QSV's side, which needs at least
2  threads to prevent dead-locking, see
https://github.com/Intel-Media-SDK/MediaSDK/blob/master/_studio/mfx_lib/scheduler/linux/src/mfx_scheduler_core_ischeduler.cpp#L90-L93

Regards,

Dennis.