[FFmpeg-devel] [PATCH][VAAPI][2/6] Add common data structures and helpers (take 3)

[Reimar Döffinger]

On Mon, Mar 09, 2009 at 03:30:26PM +0100, Gwenole Beauchesne wrote:
> Note: for larger blocks (>= 512 KB), fast_memcpy/sse2 was the best on my 
> CPU. I have not fully tested unaligned cases though. Agner's should behave 
> better since he also has SSSE3 optimizations (PALIGNR). BTW, his lib is now 
> GPLv3.

Previously fast_memcpy was used in MPlayer as a general replacement for
memcpy, but that was just nonsense. The most relevant use case is
copying of full images, which usually means aligned and large.
Which is also why the libc version can't beat it in the end if used
properly, only the user knows which places are basically sure to copy
large, aligned blocks of data and can use the most appropriate
implementation.
Still, work on that has mostly stopped since memcpy is rarely the
bottleneck (anymore).