[MPlayer-users] Disable colospace conversion
Reimar Döffinger
Reimar.Doeffinger at gmx.de
Sun Aug 25 14:32:22 CEST 2013
On Sat, Jun 29, 2013 at 09:26:43AM +0200, Reimar Döffinger wrote:
> On 28.06.2013, at 20:53, Andy Furniss <adf.lists at gmail.com> wrote:
> > Carl Eugen Hoyos wrote:
> >> Andy Furniss <adf.lists <at> gmail.com> writes:
> >>> I assume -vo null will also avoid a mem copy as well.
> >>
> >> This is not generally correct, the memcopy isn't always
> >> needed for actual playback either (direct rendering).
> >
> > Yea, I was just thinking from the PC + GPU card point of view here.
> >
> > What I was trying to say was that the difference seen when measuring perf with -vo null (on PC + card) is likely down to avoiding copying the yuv across whatever bus to the GPU card rather than CSC - but it totally doesn't make sense now I know it's ARM not PC.
>
> Well, which one is worse depends on the specifics of the system.
> The copying to the GPU (if it is a discrete one) is handled by a DMA engine and requires only read bandwidth.
> Things are actually worse on ARM systems: they usually support only GLES, which at best provides vendor-specific ways of copying, and the memory is shared for both GPU and CPU, and often memory bandwidth is too little for normal use-cases.
> So while in theory in a well-designed system with shared CPU and GPU memory copies should be less of an issue, the current implementations are so incomplete or just crap (in some cases only due to software/APIs lacking or at least being vendor-specific) that they actually suffer a lot more from performance issues due to pointless copies.
By the way, GLES 3.0 finally supports PBOs!
Unfortunately I don't have any device with GLES 3.0 support, and
considering how much of a pain it is to get Linux with hardware OpenGL
to work on any of those devices I'm not sure I'll even bother to try.
More information about the MPlayer-users
mailing list