Optimize yuv alpha blend AVX2 code to do 32 pixels at time.

Optimize yuv alpha blend AVX2 code to do 32 pixels at time.

out/Release/libyuv_unittest --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=9999 --libyuv_flags=-1 --gtest_filter=*I420Blend_Opt

Was LibYUVPlanarTest.I420Blend_Opt (2335 ms)
Now LibYUVPlanarTest.I420Blend_Opt (1937 ms)

vs SSSE3
LibYUVPlanarTest.I420Blend_Opt (2599 ms)


BUG=libyuv:527
R=dhrosa@google.com

Committed: https://chromium.googlesource.com/libyuv/libyuv/+/dee77a4ebeaebc781cb3acd80aa6627fd1c7c825

[fbarchard, 2015-12-08 06:58:06]

Description was changed from

==========
Optimize yuv alpha blend AVX2 code to do 32 pixels at time.

BUG=libyuv:527
==========

to

==========
Optimize yuv alpha blend AVX2 code to do 32 pixels at time.

out/Release/libyuv_unittest --libyuv_width=1280 --libyuv_height=720
--libyuv_repeat=9999 --libyuv_flags=-1 --gtest_filter=*I420Blend_Opt

Was LibYUVPlanarTest.I420Blend_Opt (2335 ms)
Now LibYUVPlanarTest.I420Blend_Opt (1937 ms)

vs SSSE3
LibYUVPlanarTest.I420Blend_Opt (2599 ms)


BUG=libyuv:527
==========

[fbarchard, 2015-12-08 06:59:03]

fbarchard@chromium.org changed reviewers:
+ dhrosa@google.com, fbarchard@chromium.org

[fbarchard, 2015-12-08 06:59:03]

perf report on avx2 function:

  0.02 │      vpsllw $0x8,%ymm5,%ymm5
  0.02 │      mov    $0x80808080,%eax
  0.06 │      vmovd  %eax,%xmm6
       │      vbroad %xmm6,%ymm6
  0.02 │      mov    $0x807f807f,%eax
  0.03 │      vmovd  %eax,%xmm7
  0.02 │      vbroad %xmm7,%ymm7
  0.06 │      sub    %rdx,%rdi
  0.06 │      sub    %rdx,%rsi
  0.03 │      sub    %rdx,%rcx
  0.98 │2e:   vmovdq (%rdx),%ymm0
  9.06 │      vpunpc %ymm0,%ymm0,%ymm3
  5.20 │      vpunpc %ymm0,%ymm0,%ymm0
  0.97 │      vpxor  %ymm5,%ymm3,%ymm3
  1.02 │      vpxor  %ymm5,%ymm0,%ymm0
  2.39 │      vmovdq (%rdi,%rdx,1),%ymm1
 14.18 │      vmovdq (%rsi,%rdx,1),%ymm2
 14.23 │      vpunpc %ymm2,%ymm1,%ymm4
  5.02 │      vpunpc %ymm2,%ymm1,%ymm1
  2.66 │      vpsubb %ymm6,%ymm1,%ymm4
  4.86 │      vpsubb %ymm6,%ymm1,%ymm1
  1.16 │      vpmadd %ymm4,%ymm3,%ymm3
 10.24 │      vpmadd %ymm1,%ymm0,%ymm0
  3.28 │      vpaddw %ymm7,%ymm3,%ymm3
  3.41 │      vpaddw %ymm7,%ymm0,%ymm0
  2.29 │      vpsrlw $0x8,%ymm3,%ymm3
  2.73 │      vpsrlw $0x8,%ymm0,%ymm0
  4.80 │      vpacku %ymm3,%ymm0,%ymm0
  5.78 │      vmovdq %ymm0,(%rcx,%rdx,1)
  4.55 │      lea    0x20(%rdx),%rdx
  0.69 │      sub    $0x20,%r8d
       │    ↑ jg     2e
       │      vzerou
  0.11 │    ← retq

[harryjin, 2015-12-08 07:48:57]

harryjin@google.com changed reviewers:
+ harryjin@google.com

[harryjin, 2015-12-08 07:48:58]

https://codereview.chromium.org/1505673003/diff/40001/source/cpu_id.cc
File source/cpu_id.cc (right):

https://codereview.chromium.org/1505673003/diff/40001/source/cpu_id.cc#newcod...
source/cpu_id.cc:110: //0 && (defined(_MSC_VER) && !defined(__clang__)) &&
(_MSC_FULL_VER >= 160040219)
? Remove?

https://codereview.chromium.org/1505673003/diff/40001/source/cpu_id.cc#newcod...
source/cpu_id.cc:112: printf("xgetbv xcr0 %d\n", xcr0);
Remove the printf?

https://codereview.chromium.org/1505673003/diff/40001/source/cpu_id.cc#newcod...
source/cpu_id.cc:119: printf("asm xcr0 %d\n", xcr0);
same here?

[fbarchard, 2015-12-08 19:24:04]

Add comment about xgetbv bug.  Only affects AVX2 on vs2013 and earlier.
Add version number.

https://codereview.chromium.org/1505673003/diff/40001/source/cpu_id.cc
File source/cpu_id.cc (right):

https://codereview.chromium.org/1505673003/diff/40001/source/cpu_id.cc#newcod...
source/cpu_id.cc:110: //0 && (defined(_MSC_VER) && !defined(__clang__)) &&
(_MSC_FULL_VER >= 160040219)
On 2015/12/08 07:48:58, harryjin wrote:
> ? Remove?

Done.

https://codereview.chromium.org/1505673003/diff/40001/source/cpu_id.cc#newcod...
source/cpu_id.cc:112: printf("xgetbv xcr0 %d\n", xcr0);
On 2015/12/08 07:48:58, harryjin wrote:
> Remove the printf?

Done.

https://codereview.chromium.org/1505673003/diff/40001/source/cpu_id.cc#newcod...
source/cpu_id.cc:119: printf("asm xcr0 %d\n", xcr0);
On 2015/12/08 07:48:58, harryjin wrote:
> same here?

Done.

[fbarchard, 2015-12-08 19:36:39]

BlendPlaneRow_AVX2 takes less time, meaning ScaleRowDown2Box_SSE2 takes more:
Samples: 7K of event 'cycles', Event count (approx.): 7177582504
 79.73%  libyuv_unittest  libyuv_unittest      [.] BlendPlaneRow_AVX2           
                                                                                
       
 18.77%  libyuv_unittest  libyuv_unittest      [.] ScaleRowDown2Box_SSE2        
                                                                                
       
  0.58%  libyuv_unittest  libyuv_unittest      [.] I420Blend                    
                                                                                
       
  0.31%  libyuv_unittest  libyuv_unittest      [.] libyuv::TestI420Blend(int,
int, int, int, int, int, int) [clone .isra.28]                                  
          
  0.18%  libyuv_unittest  [kernel.kallsyms]    [k] 0xffffffff8104f45a           
                                                                                
       
  0.10%  libyuv_unittest  libyuv_unittest      [.] testing::AssertionResult
testing::internal::CmpHelperEQ<unsigned char, unsigned char>(char const*, char
const*, unsign
  0.08%  libyuv_unittest  libyuv_unittest      [.]
testing::internal::DoubleNearPredFormat(char const*, char const*, char const*,
double, double, double)            
  0.06%  libyuv_unittest  libyuv_unittest      [.] testing::AssertionSuccess()  
                                                                                
       
  0.06%  libyuv_unittest  libc-2.19.so         [.] _int_malloc                  
                                                                                
       
  0.05%  libyuv_unittest  libc-2.19.so         [.] free

[Diony Rosa, 2015-12-09 01:28:51]

lgtm

[fbarchard1, 2015-12-09 02:20:34]

Description was changed from

==========
Optimize yuv alpha blend AVX2 code to do 32 pixels at time.

out/Release/libyuv_unittest --libyuv_width=1280 --libyuv_height=720
--libyuv_repeat=9999 --libyuv_flags=-1 --gtest_filter=*I420Blend_Opt

Was LibYUVPlanarTest.I420Blend_Opt (2335 ms)
Now LibYUVPlanarTest.I420Blend_Opt (1937 ms)

vs SSSE3
LibYUVPlanarTest.I420Blend_Opt (2599 ms)


BUG=libyuv:527
==========

to

==========
Optimize yuv alpha blend AVX2 code to do 32 pixels at time.

out/Release/libyuv_unittest --libyuv_width=1280 --libyuv_height=720
--libyuv_repeat=9999 --libyuv_flags=-1 --gtest_filter=*I420Blend_Opt

Was LibYUVPlanarTest.I420Blend_Opt (2335 ms)
Now LibYUVPlanarTest.I420Blend_Opt (1937 ms)

vs SSSE3
LibYUVPlanarTest.I420Blend_Opt (2599 ms)


BUG=libyuv:527
R=dhrosa@google.com

Committed:
https://chromium.googlesource.com/libyuv/libyuv/+/dee77a4ebeaebc781cb3acd80aa...
==========

[fbarchard1, 2015-12-09 02:20:35]

Committed patchset #6 (id:100001) manually as
dee77a4ebeaebc781cb3acd80aa6627fd1c7c825 (presubmit successful).