SSE optimization for 565 pixel format

Actually, these optimizations were done based on Android Eclair and Froyo branches. At that time, Android Framework were mainly using rgb565. Currently most are using 32b pixel format by default. However there are still some cases which are using 565 format (e.g. system UI on Android). Also, it's expected that these optimizations will also benefit other embedded platforms.  

The benefit measurement I got on Android platform has several cases:

1. skia_bench -config 565 -forceBlend 1 -forceDither 1 -repeat 30

w/o these opts:
D/skia    ( 1824): running bench [640 480]                  bitmap_8888
D/skia    ( 1824):    565: cmsecs = 157.49

w/ these opts:
D/skia    ( 1728): running bench [640 480]                  bitmap_8888
D/skia    ( 1728):    565: cmsecs =  26.78



2. skia_bench -config 565 -repeat 30
w/o these opts:
D/skia    ( 2954): running bench [640 480]                  bitmap_8888
D/skia    ( 2954):    565: cmsecs =  27.39
D/skia    ( 2954):
D/skia    ( 2954): running bench [640 480]                bitmap_8888_A
D/skia    ( 2954):    565: cmsecs =  79.04

w/ these opts:
D/skia    ( 2573): running bench [640 480]                  bitmap_8888
D/skia    ( 2573):    565: cmsecs =  13.19
D/skia    ( 2573):
D/skia    ( 2573): running bench [640 480]                bitmap_8888_A
D/skia    ( 2573):    565: cmsecs =  27.41

3: skia_bench -config 565 -forceBlend 1 -repeat 30

w/o these opts:
D/skia    ( 2216): running bench [640 480]                  bitmap_8888
D/skia    ( 2216):    565: cmsecs = 115.36
D/skia    ( 2216):
D/skia    ( 2216): running bench [640 480]                bitmap_8888_A
D/skia    ( 2216):    565: cmsecs = 110.25

w/ these opts:
D/skia    ( 2575): running bench [640 480]                  bitmap_8888
D/skia    ( 2575):    565: cmsecs =  23.18
D/skia    ( 2575):
D/skia    ( 2575): running bench [640 480]                bitmap_8888_A
D/skia    ( 2575):    565: cmsecs =  38.36

4: skia_bench -config 565 -forceDither 1 -repeat 30

w/o these opts:
D/skia    ( 1477): running bench [640 480]                  bitmap_8888
D/skia    ( 1477):    565: cmsecs =  92.68
D/skia    ( 1477):
D/skia    ( 1477): running bench [640 480]                bitmap_8888_A
D/skia    ( 1477):    565: cmsecs =  88.33

w/ these opts:
D/skia    ( 3315): running bench [640 480]                  bitmap_8888
D/skia    ( 3315):    565: cmsecs =  16.25
D/skia    ( 3315):
D/skia    ( 3315): running bench [640 480]                bitmap_8888_A
D/skia    ( 3315):    565: cmsecs =  29.54


After running several times and double check between the results, it's stable.

[reed1, 2012-07-09 15:21:54]

lgtm

[reed1, 2012-07-09 15:23:44]

If we don't already have them, we should write some benchmarks to exercise
these, so we can be sure the SSE2 extra complexity is paying off.

We should also amend gm as needed, so we can verify that we get the same (or
very nearly the same) results between a straight-C version and the SSE2 version.

[Lei, 2012-07-16 07:55:04]

Hi Reed, we did these optimizations on Eclair and Froyo platforms. As I said in
another patch review thread, we ran the workload, found the hot functions, and
then optimized with SSE. Of course, we kept the correct functionality when doing
optimization. Actually, the debugging took us more time than just writing SSE
routines. I will summarize the benefits and update the commit description later.
Thanks.

[Lei, 2012-07-24 08:34:15]

Hi Reed, I update the benefit measurement in the description. Please help check.
Thanks very much~

[reed1, 2012-07-24 13:58:17]

nice looking bench results. However, it didn't build for me on top of today's
sources:


/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp: In function 'void
S32A_D565_Opaque_Dither_SSE2(uint16_t*, const SkPMColor*, int, U8CPU, int,
int)':
/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp:1015: error:
'gDitherMatrix_3Bit_16' was not declared in this scope
/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp:1136: error:
'DITHER_565_SCAN' was not declared in this scope
/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp:1142: error:
'DITHER_VALUE' was not declared in this scope
/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp:1146: error:
'SkDITHER_R32_FOR_565' was not declared in this scope
/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp:1147: error:
'SkDITHER_G32_FOR_565' was not declared in this scope
/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp:1148: error:
'SkDITHER_B32_FOR_565' was not declared in this scope
/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp:1157: error:
'DITHER_INC_X' was not declared in this scope
/clean5/trunk/gyp/../src/opts/SkBlitRow_opts_SSE2.cpp: In function 'void
S32_D565_Opaque_Dither_SSE2(uint16_t*, const SkPMColor*, int, U8CPU, int, int)':

[Lei, 2012-07-25 03:04:00]

Thanks for you check. The error is introduced that SkDither.h is not included. I
will update the patch.

[Lei, 2012-07-25 03:04:27]

include SkDither.h

[Lei, 2012-07-27 03:09:00]

Hi Reed, please help check if you can also get the benefit on Mac when you get a
chance. Thanks very much~

[TomH, 2012-07-27 13:05:52]

Header file needs to use SK_RESTRICT, not __restrict__, to work cross-platform.

[reed1, 2012-07-27 13:32:29]

my summary of the issues:

1. cast pointers for bit masking using (intptr_t) instead of (int)
2. use SK_RESTRICT instead of __restrict__
3. the code must respect at compile-time the byte order in SkPMColor. In this
patch, the byte-order for RED and BLUE are reversed between Mac (where this code
runs fine) and windows/linux. See SkColorPrive.h for the SHIFT values for
A,R,G,B

[Lei, 2012-08-23 09:19:32]

replaced __restrict__ with SK_RESTRICT and use intptr_t instead of int for
casting pointers

[Lei, 2012-08-23 09:23:42]

Hi Reed, sorry for the late reply. I fixed the issues according to your
suggestion. For the red and blue order issue, I checked the header file
SkColorPrive.h. The 565 SSE routines are using the standard function SkPackRGB16
in that header file. I haven't seen the issue on Android platform (also based on
Linux). If there are still any issues, just tell me. Thanks a lot~

[Lei, 2012-08-28 03:37:46]

The patch landed in r4799. Thanks, Reed! closed this issue.