[vlc-devel] [PATCH] [RFC] always output an opaque D3D9 surface from DXVA

[Steve Lhomme]

On Mon, Apr 27, 2015 at 12:30 PM, Rémi Denis-Courmont <remi at remlab.net> wrote:
> Le 2015-04-27 12:31, Steve Lhomme a écrit :
>>>
>>> I think it is preferable for the hardware context to be made available
>>> already
>>> at Open(). (Also I would pass the picture_sys_t cast to void, rather than
>>> the
>>> whole picture_t. It would be slightly less ugly.)
>>>
>>> Besides Setup needs to go away separately. It should be folded into
>>> Open()
>>> (for all plugins). libavcodec has to call get_format() again if the
>>> resolution
>>> changes.
>>
>>
>> I'll give it a try. Thanks for helping moving this forward.
>
>
> There are still some issues left:
>
> - We need to ensure that fmt_out does not vary between get_format() calls -
> so that the underlying video output does not get destroyed and recreated.
> This is a bit intricate because we also need to support unaccelerated
> decoders that do not call get_format() at all.

I think I can't help much there.

> - Passing opaque handles from the vout to the decoder only works if there is
> no format conversion (i.e. vout format and decoder output are identical).
> Otherwise, the lack of custom pool for filters hits.

Yes, from what I saw, the filter uses the decoder pool although it
should use the output of the conversion. Not sure how to fix that
properly.

> - And obviously I have not adjusted vlc_va_New() to get the picture_sys_t or
> whatever yet.

Patch sent to add this feature. It works fine with D3D9.

>> Also the big remaining question is how to handle the D3D9 picture_t
>> pool in the core. I did it the quick & dirty way with some if on the
>> opaque pixel format and surrounded by _WIN32 tests.
>
>
> I'm not sure what you mean by that.

That it's kinda ugly and not very generic code. There might be a
similar hack for D3D11 soon and maybe other platforms could make use
of more opaque formats.

> If you don't do conversion and the video
> output supports subpictures, there should be no major issues. Except maybe
> for the corner case of snapshots.

If you mean it's working, yes it is. For snapshots, at least for my
case, the DXA9 to planes filter should be able to help here.

> If you do conversion, then you either have to use work-arounds like VDPAU,
> or implement filter pools or an usable subset of them.


I'll check what's currently done with VDPAU.