On Mon, Mar 20, 2023 at 07:07:59PM -0300, James Almer wrote:
> On 3/19/2023 4:42 PM, Michael Niedermayer wrote:
> > Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
> > ---
> >   libavcodec/snowenc.c | 10 +++++++++-
> >   1 file changed, 9 insertions(+), 1 deletion(-)
> > 
> > diff --git a/libavcodec/snowenc.c b/libavcodec/snowenc.c
> > index 658684c575..5fb5906ed8 100644
> > --- a/libavcodec/snowenc.c
> > +++ b/libavcodec/snowenc.c
> > @@ -26,6 +26,7 @@
> >   #include "avcodec.h"
> >   #include "codec_internal.h"
> >   #include "encode.h"
> > +#include "internal.h" //For AVCodecInternal.recon_frame
> >   #include "me_cmp.h"
> >   #include "packet_internal.h"
> >   #include "snow_dwt.h"
> > @@ -1576,6 +1577,7 @@ static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
> >   {
> >       SnowContext *s = avctx->priv_data;
> >       RangeCoder * const c= &s->c;
> > +    AVCodecInternal *avci = avctx->internal;
> >       AVFrame *pic;
> >       const int width= s->avctx->width;
> >       const int height= s->avctx->height;
> > @@ -1877,6 +1879,10 @@ redo_frame:
> >                                      s->encoding_error,
> >                                      (s->avctx->flags&AV_CODEC_FLAG_PSNR) ? SNOW_MAX_PLANES : 0,
> >                                      s->current_picture->pict_type);
> > +    if (s->avctx->flags & AV_CODEC_FLAG_RECON_FRAME) {
> > +        av_frame_unref(avci->recon_frame);
> > +        av_frame_ref(avci->recon_frame, s->current_picture);
> > +    }
> >       pkt->size = ff_rac_terminate(c, 0);
> >       if (s->current_picture->key_frame)
> > @@ -1934,7 +1940,9 @@ const FFCodec ff_snow_encoder = {
> >       CODEC_LONG_NAME("Snow"),
> >       .p.type         = AVMEDIA_TYPE_VIDEO,
> >       .p.id           = AV_CODEC_ID_SNOW,
> > -    .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE,
> > +    .p.capabilities = AV_CODEC_CAP_DR1 |
> > +                      AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE |
> > +                      AV_CODEC_CAP_ENCODER_RECON_FRAME,
> >       .priv_data_size = sizeof(SnowContext),
> >       .init           = encode_init,
> >       FF_CODEC_ENCODE_CB(encode_frame),
> 
> Testing with the tool from https://git.khirnov.net/libav.git/commit/?h=recon&id=d78e17840347abfe10b2bc7e789c60665de56e1f
> (Which i hope will be upstreamed eventually) i get
> 
> $ tools/enc_recon_frame_test input.y4m snow flags=+bitexact 10
> Checksum mismatch in frame ts=0, plane 0

I cannot replicate this, using matrixbench and above command line i get no 
mismatches, where can i find the input.y4m you used ?

also would be nice if this tool would write a image showing the difference

thx

> 
> I don't know if this really matters for snow (is it mean to be bitexact?),
> but it seems the very first frame returned by the reconstructed frame API is
> not the same as the one output by the snow decoder. The rest seem to match,
> though.

This sounds rather strange, the first frame would be used as reference for
motion compensation so the subsequent frames matching after a mismatch is
"surprising"

thx

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Concerning the gods, I have no means of knowing whether they exist or not
or of what sort they may be, because of the obscurity of the subject, and
the brevity of human life -- Protagoras