[FFmpeg-devel] [PATCH] x86: kill fpel_mmx.c
michaelni at gmx.at
Tue May 20 23:33:22 CEST 2014
On Tue, May 20, 2014 at 05:23:00PM -0300, James Almer wrote:
> On 20/05/14 3:58 PM, Michael Niedermayer wrote:
> > also did you benchmark the change this patch makes ?
> > the original code was quite finetuned IIRC someone back then
> > tested all kinds of instruction orders and complex addressing vs
> > increasing the pointers, and what we had was the best found i that
> > testing ...
> > that was of course with the cpus of that time, but still a mmx only
> > or mmx2 only box for which some of this gets used could be similar
> > So we should make sure we dont worsen this ...
> mmx only boxes are Pentium 2 amd AMD K6 computers. mmxext only boxes
> are AMD K7 (pre Athlon XP) computers. The amount of such computers in
> use out there that would end up using ffmpeg 2.3 or newer in any way
> is minimal.
> So even if this patch means a small performance hit for those or even
> newer computers, as long as it's not a big one it's a very welcome
> change. Especially for compilers like msvc.
we maybe misunderstand each other but I see 2 things
1. change instruction order / implementation
2. change inline to yasm
iam concerned about 1. without any benchmarks, this is speed
relevant code, some of it is used on modern computers as sse* isnt
good for 8x8 blocks so this would affect practically all desktop
machines if its slower or maybe its faster on modern boxes, but to
know that it needs to be benchmarked ...
2. is a good thing, supporting more compilers is good, i fully agree
if 1. causes a slowdown, then the code still could be converted to
yasm to achive 2. without changing the implementation, it would
possible end up with more text in the .asm file though
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
What does censorship reveal? It reveals fear. -- Julian Assange
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 198 bytes
Desc: Digital signature
More information about the ffmpeg-devel