[FFmpeg-devel] [PATCH 2/2] lavc/vaapi_av1: Avoid sending the same slice buffer multiple times

David Rosca nowrep at gmail.com
Wed May 8 10:17:03 EEST 2024


On Tue, May 7, 2024 at 10:01 PM Mark Thompson <sw at jkqxz.net> wrote:
>
> On 28/04/2024 08:26, David Rosca wrote:
> > When there are multiple tiles in one slice buffer, use multiple slice
> > params to avoid sending the same slice buffer multiple times and thus
> > increasing the bitstream size the driver will need to upload to hw.
> > ---
> >  libavcodec/vaapi_av1.c | 37 +++++++++++++++++++++++--------------
> >  1 file changed, 23 insertions(+), 14 deletions(-)
>
> Can you confirm that this works on at least iHD (Intel) and Mesa (AMD)?  (I don't expect any issue, but it's good to check in case of something strange going on matching up to what this was previously doing.)

I've tested with Mesa and it works correctly. Gstreamer also does the same.

https://gitlab.freedesktop.org/gstreamer/gstreamer/-/blob/da35ed69164ba8c8599d337b00ba54074215f7e7/subprojects/gst-plugins-bad/sys/va/gstvaav1dec.c#L891-L912

>
> > diff --git a/libavcodec/vaapi_av1.c b/libavcodec/vaapi_av1.c
> > index 4a90db1e09..567f505fbd 100644
> > --- a/libavcodec/vaapi_av1.c
> > +++ b/libavcodec/vaapi_av1.c
> > @@ -19,6 +19,7 @@
> >   */
> >
> >  #include "libavutil/frame.h"
> > +#include "libavutil/mem.h"
> >  #include "hwaccel_internal.h"
> >  #include "vaapi_decode.h"
> >  #include "internal.h"
> > @@ -393,13 +394,17 @@ static int vaapi_av1_decode_slice(AVCodecContext *avctx,
> >  {
> >      const AV1DecContext *s = avctx->priv_data;
> >      VAAPIDecodePicture *pic = s->cur_frame.hwaccel_picture_private;
> > -    VASliceParameterBufferAV1 slice_param;
> > -    int err = 0;
> > +    VASliceParameterBufferAV1 *slice_params;
> > +    int err = 0, nb_params = 0;
>
> Remove the spurious initialisation on err?

Done in v2.

>
> >
> > -    for (int i = s->tg_start; i <= s->tg_end; i++) {
> > -        memset(&slice_param, 0, sizeof(VASliceParameterBufferAV1));
> > +    slice_params = av_calloc(s->tg_end - s->tg_start + 1, sizeof(*slice_params));
>
> I suggest allocating this into VAAPIAV1DecContext to avoid the alloc/free on every call.  (Only reallocate if it needs to be bigger than the previous maximum.)

Done in v2.

Thanks,
David

>
> > +    if (!slice_params) {
> > +        err = AVERROR(ENOMEM);
> > +        goto fail;
> > +    }
> >
> > -        slice_param = (VASliceParameterBufferAV1) {
> > +    for (int i = s->tg_start; i <= s->tg_end; i++) {
> > +        slice_params[nb_params++] = (VASliceParameterBufferAV1) {
> >              .slice_data_size   = s->tile_group_info[i].tile_size,
> >              .slice_data_offset = s->tile_group_info[i].tile_offset,
> >              .slice_data_flag   = VA_SLICE_DATA_FLAG_ALL,
> > @@ -408,18 +413,22 @@ static int vaapi_av1_decode_slice(AVCodecContext *avctx,
> >              .tg_start          = s->tg_start,
> >              .tg_end            = s->tg_end,
> >          };
> > -
> > -        err = ff_vaapi_decode_make_slice_buffer(avctx, pic, &slice_param, 1,
> > -                                                sizeof(VASliceParameterBufferAV1),
> > -                                                buffer,
> > -                                                size);
> > -        if (err) {
> > -            ff_vaapi_decode_cancel(avctx, pic);
> > -            return err;
> > -        }
> >      }
> >
> > +    err = ff_vaapi_decode_make_slice_buffer(avctx, pic, slice_params, nb_params,
> > +                                            sizeof(VASliceParameterBufferAV1),
> > +                                            buffer,
> > +                                            size);
> > +    av_free(slice_params);
> > +
> > +    if (err)
> > +        goto fail;
> > +
> >      return 0;
> > +
> > +fail:
> > +    ff_vaapi_decode_cancel(avctx, pic);
> > +    return err;
> >  }
> >
> >  const FFHWAccel ff_av1_vaapi_hwaccel = {
>
> It's amusing that this quadratic behaviour was around for so long!  (I guess people don't use many tiles.)
>
> Thanks,
>
> - Mark
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".


More information about the ffmpeg-devel mailing list