[FFmpeg-devel] [PATCH v4 1/5] libavutil: VAAPI infrastructure

[wm4]

On Sun, 24 Jan 2016 13:58:59 +0100
Timo Rothenpieler <timo at rothenpieler.org> wrote:

> >> +    void *address;
> >> +    // On current Intel drivers, derive gives you memory which is very slow
> >> +    // to read (uncached?).  It can be better for write-only cases, but for
> >> +    // now play it safe and never use derive.  
> > 
> > Still subject to debate when we introduce a "GPU memcpy".
> > 
> > I think someone on this list once even posted a patch to add such a
> > memcpy, but it didn't get in yet.  
> 
> Related article from Intel:
> 
> https://software.intel.com/en-us/articles/copying-accelerated-video-decode-frame-buffers
> 
> It contains a ready-to-use copy-function, which just has to be converted
> from intrinsics to yasm.
> And propably also needs a reversed counterpart, as it's designed to copy
> from uswc memory to normal system memory, while ffmpeg also needs to
> copy data into such uswc buffers.
> 

Yes, though apparently we can get by without this additional "cache"
suggested on this page.