[FFmpeg-devel] [PATCH v1] avcodec/vaapi: set more flags for VASurfaceAttrib

[Fu, Linjie]

Hi,

> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of
> Wang, Fei W
> Sent: Wednesday, November 20, 2019 10:40
> To: mypopy at gmail.com; FFmpeg development discussions and patches
> <ffmpeg-devel at ffmpeg.org>
> Subject: Re: [FFmpeg-devel] [PATCH v1] avcodec/vaapi: set more flags for
> VASurfaceAttrib
> 
> 
> > On Tue, Nov 19, 2019 at 4:24 PM Wangfei <fei.w.wang at intel.com> wrote:
> > >
> > > flags and value.type is needed when pass VASurfaceAttrib to driver.
> > > Otherwise the attribute will be considered invalid in driver.
> > >
> > > Signed-off-by: Wangfei <fei.w.wang at intel.com>
> > > ---
> > >  libavcodec/vaapi_decode.c | 2 ++
> > >  1 file changed, 2 insertions(+)
> > >
> > > diff --git a/libavcodec/vaapi_decode.c b/libavcodec/vaapi_decode.c
> > > index 69512e1d45..a93aba5a5d 100644
> > > --- a/libavcodec/vaapi_decode.c
> > > +++ b/libavcodec/vaapi_decode.c
> > > @@ -350,6 +350,8 @@ static int
> > > vaapi_decode_find_best_format(AVCodecContext *avctx,
> > >
> > >          ctx->pixel_format_attribute = (VASurfaceAttrib) {
> > >              .type          = VASurfaceAttribPixelFormat,
> > > +            .flags         = VA_SURFACE_ATTRIB_SETTABLE,
> > > +            .value.type    = VAGenericValueTypeInteger,
> > I can't found any description in libva about this part, do you have any other
> > docs?
> I checked the logic of vaCreateSurfaces implementation in media-driver, it
> will check flags and value.type in VASurfaceAttrib when using it. If leaks of
> the info, it will consider the attribute is invalid or assert(in this case, flags
> checked first, so no assert occurred.):
> https://github.com/intel/media-
> driver/blob/8a8b8ae263a04cb5dd384d4625215cf296d2cc59/media_driver/lin
> ux/common/ddi/media_libva.cpp#L2149
> 
> On the other hand, there is another place in ffmpeg-vaapi that sets
> VASurfaceAttrib, and it filled flags and value.type:
> https://github.com/FFmpeg/FFmpeg/blob/d73f06270600c37c74beeceac37f5
> 93838ced383/libavutil/hwcontext_vaapi.c#L527
> So I believe here we need do the same thing.
> 
> > >              .value.value.i = best_fourcc,
> > >          };

Setting flags and type allows driver to assign the expected_fourcc/media_format of surface
according to the best FourCC, instead of  determining by the VA_RT_FORMAT_ defined in
vaapi_format_map[].

With this patch, decoder still works well even if a wrong VA_RT_FORMAT_  was assigned.
(For example, VA_RT_FORMAT_YUV444 assigned to VA_FOURCC_NV12, when decoding
NV12 clips)

Patch LGTM since it seems to be more robust.(but also may hide some mistakes on the other hand)

Thanks,

- linjie