(In reply to Christian König from comment #8)
> (In reply to k.philipp from comment #7)
> > I did think about that, but it has the problem of breaking deinterlacing on
> > all clients that do *not* set VA_SURFACE_ATTRIB_USAGE_HINT_VPP_READ (which I
> > assume to be most) since there is - correct me if I'm wrong - at present no
> > code to un-weave a decoded frame to fields and go back to interlaced/field
> > mode, causing the initial issue in the first place.
> 
> Yeah, we can't do this without at least adding this frame to field
> conversion.
> 
> How about VA_SURFACE_ATTRIB_USAGE_HINT_EXPORT then? The VA-API hints seem to
> describe the use case instead of the effect, so that would match.
I filed a PR with libva (https://github.com/intel/libva/pull/196), we'll see
how it goes.