[FFmpeg-devel] [PATCH 2/7] hwcontext_vaapi: Add support for legacy DRM mapping

[Rostislav Pehlivanov]

On 12 March 2018 at 22:46, Mark Thompson <sw at jkqxz.net> wrote:

> On 12/03/18 17:40, Rostislav Pehlivanov wrote:
> > On 12 March 2018 at 00:23, Mark Thompson <sw at jkqxz.net> wrote:
> >
> >> On 12/03/18 00:01, Rostislav Pehlivanov wrote:
> >>> On 11 March 2018 at 22:41, Mark Thompson <sw at jkqxz.net> wrote:
> >>>
> >>>> The old vaAcquireBufferHandle() API works in fewer cases and provides
> >>>> less information than the current vaExportSurfaceHandle(), but it
> exists
> >>>> on older versions and is already used by the OpenCL code.  This
> probably
> >>>> doesn't have much use directly, but it will be used to replace the
> >> ad-hoc
> >>>> OpenCL code in a following patch.
> >>>> ---
> >>>>  libavutil/hwcontext_vaapi.c | 191 ++++++++++++++++++++++++++++++
> >>>> +++++++++++---
> >>>>  1 file changed, 179 insertions(+), 12 deletions(-)
> >>>>
> >>>> ...
> >>>
> >>> Why? libva 1.1.0 is 6 years old now. It doesn't make sense to have 2
> >> pieces
> >>> of code to do the same thing.
> >>
> >> VAAPI 1.1.0 != libva 1.1.0.  VAAPI 1.1.0 is in libva 2.1.0, released
> last
> >> month.
> >>
> >> (This is not confusing at all.)
> >>
> >
> > ABH seems pretty much worse than ESH, and considering you can't get the
> > format modifiers makes it useless for Vulkan. I don't see why we need to
> > support it in addition to ESH.
>
> It's what it currently used by the Beignet/OpenCL interop (since it was
> the only thing which existed when that was written, and the format modifier
> isn't needed), and I don't want to break that.
>
> Since it was only released a few weeks ago it will probably be a while
> before it lands in all distributions, and it isn't implemented at all by
> the Intel ex-blob driver yet.
>
> - Mark
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
>
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>

In this case wouldn't it be better if the ABH mapping function is moved to
hwcontext_opencl? Then hwcontext_vaapi can be trusted to always produce
completely mapped to DRM frames.