Quoting Mark Thompson (2016-03-10 18:54:30)
> On 10/03/16 16:16, Anton Khirnov wrote:
> > Quoting Mark Thompson (2016-03-10 16:56:21)
> >>>> + ++avfc->nb_surfaces;
> >>>> + if (avfc->nb_surfaces == 1) {
> >>>> + // Test whether vaDeriveImage() works for this type of surface.
> >>>> + VAImageFormat *expected_format;
> >>>> + VAImage test_image;
> >>>> + int err;
> >>>> +
> >>>> + ctx->derive_works = 0;
> >>>
> >>> I think this is unsafe, since this variable is also accessed in
> >>> mapping the frame, which can easily happen concurrently with this
> >>> function. So it would be better to make this test in frames_init.
> >>> Also, this function won't be called at all if the caller supplies his
> >>> own frame pool. The mapping function should work even then.
> >>
> >> So call get_buffer() at the end of frames_init() and try the mapping there?
> >> That might be slightly surprising to the user supplying their own pool, but
> >> otherwise seems entirely reasonable.
> >
> > I think just creating a single dummy surface with a random format and
> > calling derive on it might be simpler.
>
> The dummy surface needs to have the same attributes as the surface you
> actually
> want to test, though - especially VASurfaceAttribMemoryType, which is going to
> be set in many of the obvious use cases for this (e.g. DRM surfaces acquired
> from an external source like DRI2 and made into a buffer pool to use here).
> Since the attributes aren't supplied, it has to call get_buffer() in the
> external case, at which point it might as well call get_buffer() for the
> internal case as well.
Hmm, I thought it would either always work for a given device or always
fails. If it's really the case that it might succeed or fail depending
on the framesctx properties then I guess what you suggest is the way to
go.
--
Anton Khirnov
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
