[FFmpeg-devel] [PATCH 2/2] avcodec/mjpegdec: add support for frame threading

[Andreas Rheinhardt]

Timo Rothenpieler:
> In my tests, this lead to a notable speed increase with the amount
> of threads used. Decoding a 720p sample gave the following results:
> 
> 1 Thread: 1428 FPS
> 2 Threads: 2501 FPS
> 8 Threads: 7575 FPS
> Automatic: 11326 FPS (On a 16 Core/32 Threads system)
> ---
>  libavcodec/jpeglsdec.c |  2 +-
>  libavcodec/mjpegdec.c  | 13 +++++++------
>  libavcodec/sp5xdec.c   |  4 ++--
>  3 files changed, 10 insertions(+), 9 deletions(-)
> 
> diff --git a/libavcodec/jpeglsdec.c b/libavcodec/jpeglsdec.c
> index 2e6d018ea6..c0642e8e30 100644
> --- a/libavcodec/jpeglsdec.c
> +++ b/libavcodec/jpeglsdec.c
> @@ -559,7 +559,7 @@ const FFCodec ff_jpegls_decoder = {
>      .init           = ff_mjpeg_decode_init,
>      .close          = ff_mjpeg_decode_end,
>      FF_CODEC_RECEIVE_FRAME_CB(ff_mjpeg_receive_frame),
> -    .p.capabilities = AV_CODEC_CAP_DR1,
> +    .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
>      .caps_internal  = FF_CODEC_CAP_INIT_CLEANUP |
>                        FF_CODEC_CAP_SETS_PKT_DTS,
>  };
> diff --git a/libavcodec/mjpegdec.c b/libavcodec/mjpegdec.c
> index 9b7465abe7..d30d722398 100644
> --- a/libavcodec/mjpegdec.c
> +++ b/libavcodec/mjpegdec.c
> @@ -54,6 +54,7 @@
>  #include "exif.h"
>  #include "bytestream.h"
>  #include "tiff_common.h"
> +#include "thread.h"
>  
>  
>  static int init_default_huffman_tables(MJpegDecodeContext *s)
> @@ -713,7 +714,7 @@ int ff_mjpeg_decode_sof(MJpegDecodeContext *s)
>                  s->avctx->pix_fmt,
>                  AV_PIX_FMT_NONE,
>              };
> -            s->hwaccel_pix_fmt = ff_get_format(s->avctx, pix_fmts);
> +            s->hwaccel_pix_fmt = ff_thread_get_format(s->avctx, pix_fmts);
>              if (s->hwaccel_pix_fmt < 0)
>                  return AVERROR(EINVAL);
>  
> @@ -729,7 +730,7 @@ int ff_mjpeg_decode_sof(MJpegDecodeContext *s)
>          }
>  
>          av_frame_unref(s->picture_ptr);
> -        if (ff_get_buffer(s->avctx, s->picture_ptr, AV_GET_BUFFER_FLAG_REF) < 0)
> +        if (ff_thread_get_buffer(s->avctx, s->picture_ptr, AV_GET_BUFFER_FLAG_REF) < 0)
>              return -1;
>          s->picture_ptr->pict_type = AV_PICTURE_TYPE_I;
>          s->picture_ptr->key_frame = 1;
> @@ -2388,7 +2389,7 @@ static int mjpeg_get_packet(AVCodecContext *avctx)
>      int ret;
>  
>      av_packet_unref(s->pkt);
> -    ret = ff_decode_get_packet(avctx, s->pkt);
> +    ret = ff_thread_decode_get_packet(avctx, s->pkt);
>      if (ret < 0)
>          return ret;
>  
> @@ -3020,7 +3021,7 @@ const FFCodec ff_mjpeg_decoder = {
>      .close          = ff_mjpeg_decode_end,
>      FF_CODEC_RECEIVE_FRAME_CB(ff_mjpeg_receive_frame),
>      .flush          = decode_flush,
> -    .p.capabilities = AV_CODEC_CAP_DR1,
> +    .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
>      .p.max_lowres   = 3,
>      .p.priv_class   = &mjpegdec_class,
>      .p.profiles     = NULL_IF_CONFIG_SMALL(ff_mjpeg_profiles),
> @@ -3050,7 +3051,7 @@ const FFCodec ff_thp_decoder = {
>      .close          = ff_mjpeg_decode_end,
>      FF_CODEC_RECEIVE_FRAME_CB(ff_mjpeg_receive_frame),
>      .flush          = decode_flush,
> -    .p.capabilities = AV_CODEC_CAP_DR1,
> +    .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
>      .p.max_lowres   = 3,
>      .caps_internal  = FF_CODEC_CAP_INIT_CLEANUP |
>                        FF_CODEC_CAP_SETS_PKT_DTS,
> @@ -3068,7 +3069,7 @@ const FFCodec ff_smvjpeg_decoder = {
>      .close          = ff_mjpeg_decode_end,
>      FF_CODEC_RECEIVE_FRAME_CB(ff_mjpeg_receive_frame),
>      .flush          = decode_flush,
> -    .p.capabilities = AV_CODEC_CAP_DR1,
> +    .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
>      .caps_internal  = FF_CODEC_CAP_EXPORTS_CROPPING |
>                        FF_CODEC_CAP_SETS_PKT_DTS | FF_CODEC_CAP_INIT_CLEANUP,
>  };
> diff --git a/libavcodec/sp5xdec.c b/libavcodec/sp5xdec.c
> index 394448c5a9..8b08dc672a 100644
> --- a/libavcodec/sp5xdec.c
> +++ b/libavcodec/sp5xdec.c
> @@ -101,7 +101,7 @@ const FFCodec ff_sp5x_decoder = {
>      .init           = ff_mjpeg_decode_init,
>      .close          = ff_mjpeg_decode_end,
>      FF_CODEC_RECEIVE_FRAME_CB(ff_mjpeg_receive_frame),
> -    .p.capabilities = AV_CODEC_CAP_DR1,
> +    .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
>      .p.max_lowres   = 3,
>      .caps_internal  = FF_CODEC_CAP_INIT_CLEANUP |
>                        FF_CODEC_CAP_SETS_PKT_DTS,
> @@ -118,7 +118,7 @@ const FFCodec ff_amv_decoder = {
>      .close          = ff_mjpeg_decode_end,
>      FF_CODEC_RECEIVE_FRAME_CB(ff_mjpeg_receive_frame),
>      .p.max_lowres   = 3,
> -    .p.capabilities = AV_CODEC_CAP_DR1,
> +    .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
>      .caps_internal  = FF_CODEC_CAP_INIT_CLEANUP |
>                        FF_CODEC_CAP_SETS_PKT_DTS,
>  };

Can you test the sample that Michael mentioned here:
https://patchwork.ffmpeg.org/project/ffmpeg/patch/AS8PR01MB7944E105BE990A5D01EF89208FEF9@AS8PR01MB7944.eurprd01.prod.exchangelabs.com/?
(I never got around to analyzing this, but if I am not mistaken, it
shows that this decoder is not a simple one-in, one-out decoder, so that
making it multithreaded will be more complicated than just adding the
flag and using ff_thread_get_buffer/format.)

- Andreas