Implemented SSE2 version of loadAlphaData

With this patch,my HTML5 2D canvas benchmark with Chromium on Windows7 with GPU acceleration can boost about 4% though the most execution time is in GPU.

[dgkoch, 2012-04-18 13:21:59]

Thanks for contributing!

I noticed there seems to be some formatting issues -- likely caused by tab
characters. Please use 4-spaces for indenting.

+nicolas to review.

[Jin Yang, 2012-04-18 13:24:15]

correct the format

[Jin Yang, 2012-04-20 05:56:08]

Hi nicolas,

Please help review it when you are available.

Thanks,
Jin

[nicolas, 2012-04-20 13:06:31]

Spotted a little bug and an easy optimization.

http://codereview.appspot.com/6050054/diff/5001/src/libGLESv2/Texture.cpp
File src/libGLESv2/Texture.cpp (right):

http://codereview.appspot.com/6050054/diff/5001/src/libGLESv2/Texture.cpp#new...
src/libGLESv2/Texture.cpp:455: dest[x] = source[x] << 24;
Shifting a byte by 24 bit results in zero. You probably want "dest[x] =
(unsigned int)source[x] << 24;".

http://codereview.appspot.com/6050054/diff/5001/src/libGLESv2/Texture.cpp#new...
src/libGLESv2/Texture.cpp:458: __m128i zeroWide = _mm_setzero_si128();
It's best to initialize this outside of the y loop.

http://codereview.appspot.com/6050054/diff/5001/src/libGLESv2/Texture.cpp#new...
src/libGLESv2/Texture.cpp:474: dest[x] = source[x] << 24;
Same as above.

[Jin Yang, 2012-04-24 02:29:04]

patch 3

[Jin Yang, 2012-04-24 02:30:50]

Thanks for your nice comments! I have fixed them in the new patch.

Jin

http://codereview.appspot.com/6050054/diff/5001/src/libGLESv2/Texture.cpp
File src/libGLESv2/Texture.cpp (right):

http://codereview.appspot.com/6050054/diff/5001/src/libGLESv2/Texture.cpp#new...
src/libGLESv2/Texture.cpp:455: dest[x] = source[x] << 24;
On 2012/04/20 13:06:31, nicolas wrote:
> Shifting a byte by 24 bit results in zero. You probably want "dest[x] =
> (unsigned int)source[x] << 24;".
Yes, thanks for reminder.

Done.

http://codereview.appspot.com/6050054/diff/5001/src/libGLESv2/Texture.cpp#new...
src/libGLESv2/Texture.cpp:458: __m128i zeroWide = _mm_setzero_si128();
On 2012/04/20 13:06:31, nicolas wrote:
> It's best to initialize this outside of the y loop.

Done.

http://codereview.appspot.com/6050054/diff/5001/src/libGLESv2/Texture.cpp#new...
src/libGLESv2/Texture.cpp:474: dest[x] = source[x] << 24;
On 2012/04/20 13:06:31, nicolas wrote:
> Same as above.

Done.

[nicolas, 2012-04-24 19:43:39]

Looks good to me.

By the way, have you tried optimizing the non-SSE2 version by writing 32-bit
values (instead of four individual bytes) and compared the performance against
the SSE2 version? It's possible the SSE2 version isn't significantly faster if
it's bandwidth limited. Just wondering.

[Jin Yang, 2012-04-25 02:23:54]

On 2012/04/24 19:43:39, nicolas wrote:
> Looks good to me.
> 
> By the way, have you tried optimizing the non-SSE2 version by writing 32-bit
> values (instead of four individual bytes) and compared the performance against
> the SSE2 version? It's possible the SSE2 version isn't significantly faster if
> it's bandwidth limited. Just wondering.

I can't observer obvious performance difference with  the optimized version with
none SSE vs original version. The SSE version do have some improvement but
should be limited though.

[Jin Yang, 2012-04-28 03:37:58]

Is it possible to commit the patch into the repo? Thanks

[dgkoch, 2012-05-01 15:27:30]

On 2012/04/28 03:37:58, Jin Yang wrote:
> Is it possible to commit the patch into the repo? Thanks

Committed to ANGLE as r1067:
http://code.google.com/p/angleproject/source/detail?r=1067
(you'll have to manually close this review)