[FFmpeg-devel] [PATCH 31/92] Vulkan patchset part 1 - common code changes

[Lynne]

Apr 22, 2023, 14:38 by anton at khirnov.net:

> Quoting Lynne (2023-04-19 16:47:11)
>
>> Apr 17, 2023, 11:07 by anton at khirnov.net:
>>
>> > Quoting Anton Khirnov (2023-04-16 22:38:16)
>> >
>> >> Quoting Lynne (2023-03-14 07:33:43)
>> >> > From 61f412eea3fbcb1e2a8625796760c0e24fa3fb83 Mon Sep 17 00:00:00 2001
>> >> > From: Lynne <dev at lynne.ee>
>> >> > Date: Tue, 17 Jan 2023 05:01:45 +0100
>> >> > Subject: [PATCH 27/92] h264dec: add hwaccel_params_buf
>> >> > 
>> >> > diff --git a/libavcodec/h264dec.c b/libavcodec/h264dec.c
>> >> > index 0aee724c0d..6559593195 100644
>> >> > --- a/libavcodec/h264dec.c
>> >> > +++ b/libavcodec/h264dec.c
>> >> > @@ -341,6 +341,7 @@ static av_cold int h264_decode_end(AVCodecContext *avctx)
>> >> >      H264Context *h = avctx->priv_data;
>> >> >      int i;
>> >> > 
>> >> > +    av_buffer_unref(&h->hwaccel_params_buf);
>> >> >      ff_h264_remove_all_refs(h);
>> >> >      ff_h264_free_tables(h);
>> >> > 
>> >> > @@ -470,6 +471,7 @@ static void h264_decode_flush(AVCodecContext *avctx)
>> >> > 
>> >> >      ff_h264_flush_change(h);
>> >> >      ff_h264_sei_uninit(&h->sei);
>> >> > +    av_buffer_unref(&h->hwaccel_params_buf);
>> >> > 
>> >> >      for (i = 0; i < H264_MAX_PICTURE_COUNT; i++)
>> >> >          ff_h264_unref_picture(h, &h->DPB[i]);
>> >> > @@ -669,6 +671,7 @@ static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size)
>> >> >              avpriv_request_sample(avctx, "data partitioning");
>> >> >              break;
>> >> >          case H264_NAL_SEI:
>> >> > +            av_buffer_unref(&h->hwaccel_params_buf);
>> >> >              if (h->setup_finished) {
>> >> >                  avpriv_request_sample(avctx, "Late SEI");
>> >> >                  break;
>> >> > @@ -682,6 +685,7 @@ static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size)
>> >> >              break;
>> >> >          case H264_NAL_SPS: {
>> >> >              GetBitContext tmp_gb = nal->gb;
>> >> > +            av_buffer_unref(&h->hwaccel_params_buf);
>> >> >              if (avctx->hwaccel && avctx->hwaccel->decode_params) {
>> >> >                  ret = avctx->hwaccel->decode_params(avctx,
>> >> >                                                      nal->type,
>> >>
>> >> Removed on SEI and SPS? That seems bizarre and highly arbitrary. Why
>> >> specifically those two?
>> >>
>> >
>> > Found the explanation in the cover letter - I assume SEI is a mistake
>> > and should really be PPS. The points where it is reset still seem highly
>> > Vulkan-specific, so I think it'd be better for the decoder to set a flag
>> > on seeing new parameter sets. Vulkan could would then check and clear
>> > the flag as needed.
>> >
>>
>> It was a typo, changed to remove it on SPS and PPS.
>> We still have to store the compiled bufferref somewhere, and synchronize
>> it between frames, and having it in the frame context itself solves
>> those issues.
>>
>
> Except the way it is handled makes no sense for anything other than
> vulkan.
>
> Having per-thread hwaccel contexts with a callback for propagating the
> data would solve this more cleanly.
>

How would that look?