SSE optimization for S32_D16_filter_DX

This optimization was done on Android Gingerbread platform for SoftWeg benchmark.

The benefit can be measured by running following command on Android platform:

skia_bench -config 565 -match bitmap_8888_scale_filter -forceFilter 1 -repeat 30

The result I got on Android platform was below:

w/o this optimization routine: 
D/skia    ( 1868): running bench [640 480]     bitmap_8888_scale_filter
D/skia    ( 1868):    565: cmsecs = 286.50

w/ optimization:
D/skia    ( 1463): running bench [640 480]     bitmap_8888_scale_filter
D/skia    ( 1463):    565: cmsecs = 186.80

The net gain is 34.80%.

[reed1, 2012-07-03 12:15:29]

Did not compile on my machine:


/clean5/trunk/gyp/../src/opts/opts_check_SSE2.cpp: In member function 'void
SkBitmapProcState::platformProcs()':
/clean5/trunk/gyp/../src/opts/opts_check_SSE2.cpp:120: error:
'S32_D16_filter_DX' was not declared in this scope

[Lei, 2012-07-04 07:17:26]

It's strange. I can build on Android build. The definition is located in section
of "D16 functions for SRC == 8888" in SkBitmapProcState.cpp

[Lei, 2012-07-16 07:24:06]

Hi Reed, does the error still exist on your side?

[reed1, 2012-07-16 16:02:06]

/clean5/trunk/gyp/../src/opts/opts_check_SSE2.cpp: In member function 'void
SkBitmapProcState::platformProcs()':
/clean5/trunk/gyp/../src/opts/opts_check_SSE2.cpp:120: error:
'S32_D16_filter_DX' was not declared in this scope

[reed1, 2012-07-16 16:02:48]

I am building on a mac. Can you build cleanly there w/ and w/o your patch?

[Lei, 2012-07-17 06:29:27]

add the declaration of S32_D16_filter_DX in SkBitmapProcState.h

[Lei, 2012-07-17 06:33:17]

On 2012/07/16 16:02:48, reed1 wrote:
> I am building on a mac. Can you build cleanly there w/ and w/o your patch?

Sorry. My fault! The definition is in SkBitmapProcState.cpp. However,
opts_check_SSE2.cpp doesn't see this. It needs to declare it in
SkBitmapProcState.h file header, which is included by opts_check_SSE2.cpp.
Thanks for your patience~

[reed1, 2012-07-17 19:57:40]

Thanks, this builds for me now.

What bench(s) should I run to see the performance difference?

I tried this, but didn't see any improvements:

    out/Release/bench -config 565 -match bitmap_8888 -repeat 100


/clean5/trunk> cat old565.txt 
skia bench: alpha=0xFF antialias=1 filter=0 deferred=0 logperiter=0 rotate=0
scale=0 clip=0 dither=default strokeWidth=none scalar=float system=MAC
running bench [640 480] bitmap_8888_update_scale_rotate_filter   565: cmsecs =
114.09
running bench [640 480] bitmap_8888_update_volatile_scale_rotate_filter   565:
cmsecs = 114.23
running bench [640 480] bitmap_8888_scale_rotate_filter   565: cmsecs = 114.28
running bench [640 480] bitmap_8888_A_scale_rotate_filter   565: cmsecs =  78.11
running bench [640 480] bitmap_8888_update_scale_filter   565: cmsecs =  87.36
running bench [640 480] bitmap_8888_update_volatile_scale_filter   565: cmsecs =
 87.99
running bench [640 480]     bitmap_8888_scale_filter   565: cmsecs =  88.87
running bench [640 480]   bitmap_8888_A_scale_filter   565: cmsecs =  51.13
running bench [640 480]           bitmap_8888_update   565: cmsecs =   8.20
running bench [640 480]  bitmap_8888_update_volatile   565: cmsecs =   8.20
running bench [640 480]                  bitmap_8888   565: cmsecs =   8.19
running bench [640 480]                bitmap_8888_A   565: cmsecs =  24.04



/clean5/trunk> cat new565.txt 
skia bench: alpha=0xFF antialias=1 filter=0 deferred=0 logperiter=0 rotate=0
scale=0 clip=0 dither=default strokeWidth=none scalar=float system=MAC
running bench [640 480] bitmap_8888_update_scale_rotate_filter   565: cmsecs =
114.21
running bench [640 480] bitmap_8888_update_volatile_scale_rotate_filter   565:
cmsecs = 114.04
running bench [640 480] bitmap_8888_scale_rotate_filter   565: cmsecs = 114.04
running bench [640 480] bitmap_8888_A_scale_rotate_filter   565: cmsecs =  78.84
running bench [640 480] bitmap_8888_update_scale_filter   565: cmsecs =  87.25
running bench [640 480] bitmap_8888_update_volatile_scale_filter   565: cmsecs =
 87.18
running bench [640 480]     bitmap_8888_scale_filter   565: cmsecs =  87.25
running bench [640 480]   bitmap_8888_A_scale_filter   565: cmsecs =  51.20
running bench [640 480]           bitmap_8888_update   565: cmsecs =   8.20
running bench [640 480]  bitmap_8888_update_volatile   565: cmsecs =   8.20
running bench [640 480]                  bitmap_8888   565: cmsecs =   8.19
running bench [640 480]                bitmap_8888_A   565: cmsecs =  23.99

[Lei, 2012-07-18 02:06:28]

Hi Reed, thanks for your effort. This has filter option. I will find the config
to show the benefit later. Actually yesterday I tried to run bench on Android
platform. Unfortunately, it reported segmentation fault. I will find the reason
first.

[Lei, 2012-07-24 07:42:09]

update the comments for the bench test case and debugging check

[Lei, 2012-07-24 08:23:45]

Hi Reed, I updated the performance gain in the description of this issue. Please
help check. Actually, when I was trying to find the configuration, I noticed
that SKIA in Android seems not synced to the upstream for a long time.
Previously, it seems to me that it will be synced to the upstream once Android
was upgraded to the next version (e.g. from Gingerbread to Honeycomb, from
Honeycomb to ICS). However, it seems no big difference for SKIA between ICS and
Jelly Bean. The bench tests are not updated, either.

[reed1, 2012-07-24 13:21:52]

Derek, can you comment on android syncing?

[DerekS, 2012-07-24 13:42:33]

On 2012/07/24 13:21:52, reed1 wrote:
> Derek, can you comment on android syncing?

Yes, I can provide some insight there.  Android has not updated Skia in
JellyBean due to some complications with the Android framework.  Going forward
we plan to update Skia in Android before the next major release and will keep it
updated going forward.

[reed1, 2012-07-24 13:53:42]

very good bench summary.

landed in rev. 4729

[Lei, 2012-07-25 02:18:47]

Thanks, Reed and DerekS. I will close this issue.