Xfermode: SSE2 implementation of colordodge&colorburn modes

Xfermode: SSE2 implementation of colordodge&colorburn modes

With SSE2 optimization, performance of the related benchmarks will improve
about 45% for Xfermode_ColorDodge and little for Xfermode_ColorBurn on
desktop i7-3770. The little performance improvement for
Xfermode_ColorBurn is due to the portable version may mostly go the fast
if branch while the SSE2 version do the calculation for all the three
if-else branches. Here are the data:
before:
Xfermode_ColorDodge   8888:  cmsecs =  73.71   565:  cmsecs =  82.88
 Xfermode_ColorBurn   8888:  cmsecs =  46.46   565:  cmsecs =  52.23
after:
Xfermode_ColorDodge   8888:  cmsecs =  39.70   565:  cmsecs =  47.45
 Xfermode_ColorBurn   8888:  cmsecs =  45.02   565:  cmsecs =  51.15

BUG=skia:

Committed: http://code.google.com/p/skia/source/detail?r=14377

[qiankun, 2014-04-11 07:47:01]

PTAL.

[mtklein, 2014-04-24 18:13:30]

lgtm with one suggestion for readability, provided it doesn't hurt performance.

I think I've hit every one of the CLs.  Please let me know if I missed any?

https://codereview.chromium.org/224823004/diff/1/src/opts/SkXfermode_opts_SSE...
File src/opts/SkXfermode_opts_SSE2.cpp (right):

https://codereview.chromium.org/224823004/diff/1/src/opts/SkXfermode_opts_SSE...
src/opts/SkXfermode_opts_SSE2.cpp:126: diff = _mm_cvttps_epi32(_mm_div_ps(x,
y));
Think it's worth pulling these three lines out as a little static inline
function shim_mm_div_epi32 or something like that, and use it here and in
colorburn?  Took me a minute to remember SSE doesn't have integer division.

This may be a dumb question, but is there a way for us to get support for SVML
in GCC and Clang?  _mm_div_epi32 is of course what I really want.

[qiankun, 2014-04-25 09:24:37]

https://codereview.chromium.org/224823004/diff/1/src/opts/SkXfermode_opts_SSE...
File src/opts/SkXfermode_opts_SSE2.cpp (right):

https://codereview.chromium.org/224823004/diff/1/src/opts/SkXfermode_opts_SSE...
src/opts/SkXfermode_opts_SSE2.cpp:126: diff = _mm_cvttps_epi32(_mm_div_ps(x,
y));
On 2014/04/24 18:13:31, mtklein wrote:
> Think it's worth pulling these three lines out as a little static inline
> function shim_mm_div_epi32 or something like that, and use it here and in
> colorburn?  Took me a minute to remember SSE doesn't have integer division.

Done.
> 
> This may be a dumb question, but is there a way for us to get support for SVML
Sorry, I have no idea about this either.
> in GCC and Clang?  _mm_div_epi32 is of course what I really want.

[qiankun, 2014-04-25 09:26:19]

The CQ bit was checked by qiankun.miao@intel.com

[commit-bot: I haz the power, 2014-04-25 09:26:40]

CQ is trying da patch. Follow status at
https://skia-tree-status.appspot.com/cq/qiankun.miao@intel.com/224823004/20001

[commit-bot: I haz the power, 2014-04-25 09:45:03]

Change committed as 14377