[MPlayer-dev-eng] DxVA in MPlayer

Georgi Petrov gogothebee at gmail.com
Sat Dec 4 10:07:03 CET 2010

On Fri, Dec 3, 2010 at 6:59 PM, Uoti Urpala <uoti.urpala at pp1.inet.fi> wrote:
> On Fri, 2010-12-03 at 17:30 +0200, Georgi Petrov wrote:
>> After analyzing some code today, I'm willing to listen to Reimar and
>> start from DxVA 1.
> Hmm didn't Reimar primarily argue for a readback implementation, which
> is _incompatible_ with DxVA 1?

Yes, initially. However I opt for implementing non read-back first as
it will be speedier and will fit better the speed requirements I have.
Laurent's test show very measurable difference between the read-back
and non-readback (VLC vs MPC-HC). I'm willing to implement and
read-back once non read-back is ready, so all sides are happy. The
difference between them won't be so big. Right now a good starting
point is more important IMHO. I'm already behind schedule for this
nasty linking bug.

> So if you're starting from DxVA 1 then
> referring to it as Reimar's preferred approach sounds misleading.

On Wed, Nov 24, 2010 at 2:00 AM, Reimar Döffinger
<Reimar.Doeffinger at gmx.de> wrote:
> Well, in principle, since DXVA 2 seems currently unusable with
> composition and you can use DXVA 2 via the DXVA 1 API it might
> be possible to start with the DXVA 1 implementation first.

> Do you
> mean you're starting by adding DxVA 1 acceleration support to vo_directx
> or vo_direct3d?

Yes, exactly. Right now I'm looking at code to decide where to start
from and preferably I'll start from vo_direct3d (if this is possible),
because I've written most of it and I know it much better and the
initial effort will be smaller. vo_directx is 100% candidate, as
overlays are explicit requirement in the specs I have.

> If your only "target" is the acceleration then yes you may get an
> initial version of that done faster; OTOH you could get demonstrably
> working code for the renderer part too. If nobody has tried to use DxVA
> 1 with FFmpeg yet then I think there's a higher risk of hitting
> non-obvious problems with that.

Yes, I thought about it. Let me comment more on these issues once I
have enough information. I'm working on it.

Working renderer basically means implementing Custom EVR without any
form of DxVA and make it stable using the current non-accelerated
ffmpeg codecs. Then start adding DxVA. This is high in the TODO list
as well.

As I have a lot of code to look in, I'm doing the best thing - trying
to understand most of the code in MPC-HC and VLC and afterwards I'll
comment my initial intentions. Give me 2-3 days.

More information about the MPlayer-dev-eng mailing list