Quoting wm4 (2017-07-21 22:56:43)
> On Fri, 21 Jul 2017 22:32:43 +0200
> Anton Khirnov <an...@khirnov.net> wrote:
>
> > Quoting wm4 (2017-07-12 17:02:49)
> > > +int av_frame_apply_cropping(AVFrame *frame, int flags)
> > > +{
> > > + const AVPixFmtDescriptor *desc;
> > > + size_t offsets[4];
> > > + int i;
> > > +
> > > + if (!(frame->width > 0 && frame->height > 0))
> > > + return AVERROR(EINVAL);
> > > +
> > > + /* make sure we are noisy about decoders returning invalid cropping
> > > data */
> >
> > The comment does not apply anymore.
> >
> > > diff --git a/libavutil/frame.h b/libavutil/frame.h
> > > index f9ffb5bbbf..eb1fa0cba5 100644
> > > --- a/libavutil/frame.h
> > > +++ b/libavutil/frame.h
> > > @@ -580,6 +580,38 @@ AVFrameSideData *av_frame_get_side_data(const
> > > AVFrame *frame,
> > > */
> > > void av_frame_remove_side_data(AVFrame *frame, enum AVFrameSideDataType
> > > type);
> > >
> > > +
> > > +/**
> > > + * Flags for frame cropping.
> > > + */
> > > +enum {
> > > + /**
> > > + * Apply the maximum possible cropping, even if it requires setting
> > > the
> > > + * AVFrame.data[] entries to unaligned pointers. Using this is
> > > strongly
> > > + * discouraged, especially if the data is to be fed back to Libav
> > > API. It
> > > + * can cause performance issues, and in some cases crashes.
> >
> > 'discouraged' is too weak IMO, to me it means that it's something which
> > is allowed but suboptimal. Whereas passing invalid data to libav* is
> > outright forbidden as UB, and is very likely to crash. Suggested
> > wording:
> > 'Note that all Libav APIs (except those that explicitly specify
> > otherwise) expect the frame data to be aligned, so when this flag is
> > used, the resulting frame must not be fed back to Libav.'
> >
> > Otherwise looks ok, modulo apichanges/bump
> >
>
> Isn't that wording to strict? I'm fairly sure you'd still be allowed to
> e.g. copy the frame data, or pass it to swscale. (Yes, libswscale is
> officially fine with unaligned data, just that it might switch to C
> code paths.)
Then we should document those functions as allowing unaligned data and
then what I wrote will still be true. And for sws i would almost expect
a copy+simd to be faster than the C paths.
--
Anton Khirnov
_______________________________________________
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
