Add ARGBExtractAlpha_AVX2 function

Add ARGBExtractAlpha_AVX2 function

Port SSE2 version to AVX2.
BUG=libyuv:572
TEST=/usr/local/google/home/fbarchard/intelsde/sde -skx -- out/Release/libyuv_unittest --gtest_filter=*Extract*
R=wangcheng@google.com, magjed@chromium.org

Committed: https://chromium.googlesource.com/libyuv/libyuv/+/a5e93766a20ac5fff9e0ef2f5bc7c4bb1a0fdb8e

[fbarchard1, 2016-10-13 02:07:55]

Description was changed from

==========
Add ARGBExtractAlpha_AVX2 function

Port SSE2 version to AVX2.
BUG=libyuv:572
R=magjed@chromium.org
==========

to

==========
Add ARGBExtractAlpha_AVX2 function

Port SSE2 version to AVX2.
BUG=libyuv:572
TEST=/usr/local/google/home/fbarchard/intelsde/sde -skx --
out/Release/libyuv_unittest --gtest_filter=*Extract*
R=magjed@chromium.org
==========

[fbarchard1, 2016-10-13 02:08:22]

/usr/local/google/home/fbarchard/intelsde/sde -skx --
out/Release/libyuv_unittest --gtest_filter=*Extract*
Note: Google Test filter = *Extract*
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from LibYUVPlanarTest
[ RUN      ] LibYUVPlanarTest.TestARGBExtractAlpha
[       OK ] LibYUVPlanarTest.TestARGBExtractAlpha (52 ms)
[----------] 1 test from LibYUVPlanarTest (149 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test case ran. (261 ms total)
[  PASSED  ] 1 test.

[fbarchard1, 2016-10-13 17:31:31]

try bots pass.
This function is used in media canvas capture to convert ARGB to YUVA
https://cs.chromium.org/chromium/src/content/renderer/media/canvas_capture_ha...

[fbarchard1, 2016-10-13 18:01:08]

Benchmark on Macbook Air 11" early 2015  broadwelll 5650u 2.2GHz  i7

LIBYUV_DISABLE_AVX2=1 out/Release/libyuv_unittest --gtest_filter=*ExtractAlpha*
--libyuv_width=640 --libyuv_height=360 --libyuv_repeat=9999

Was SSE2
TestARGBExtractAlpha (444 ms)

Npw AVX2
TestARGBExtractAlpha (404 ms)

9.9% faster.

[fbarchard1, 2016-10-13 18:48:29]

tried unrolling:
void ARGBExtractAlphaRow_AVX2(const uint8* src_argb, uint8* dst_a, int width) {
 asm volatile (
    LABELALIGN
  "1:                                          \n"
    "vmovdqu   " MEMACCESS(0) ", %%ymm0        \n"
    "vmovdqu   " MEMACCESS2(0x20, 0) ", %%ymm1 \n"
	"vmovdqu   " MEMACCESS2(0x40, 0) ", %%ymm2 \n"
    "vmovdqu   " MEMACCESS2(0x60, 0) ", %%ymm3 \n"
    "lea       " MEMLEA(0x80, 0) ", %0         \n"
    "vpsrld     $0x18, %%ymm0, %%ymm0          \n"
    "vpsrld     $0x18, %%ymm1, %%ymm1          \n"
    "vpsrld     $0x18, %%ymm2, %%ymm2          \n"
    "vpsrld     $0x18, %%ymm3, %%ymm3          \n"
    "vpackssdw  %%ymm1, %%ymm0, %%ymm0         \n"
    "vpackssdw  %%ymm3, %%ymm2, %%ymm2         \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
    "vpacksswb  %%ymm2, %%ymm0, %%ymm0         \n"
    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
    "lea        " MEMLEA(0x20, 1) ", %1        \n"
    "sub        $0x20, %2                      \n"
    "jg         1b                             \n"
    "vzeroupper                                \n"
  : "+r"(src_argb),  // %0
    "+r"(dst_a),     // %1
    "+rm"(width)     // %2
  :
  : "memory", "cc"
    , "xmm0", "xmm1", "xmm2", "xmm3"
  );
}
#endif  // HAS_ARGBEXTRACTALPHAROW_AVX2

[fbarchard1, 2016-10-13 19:28:20]

replaced 2 vpermq with 1 vpermd as done in the ARGBToY function using the same
constant.

Was SSE2
TestARGBExtractAlpha (432 ms)

Now AVX2
TestARGBExtractAlpha (378 ms)

Best of 3 runs.

[wangcheng, 2016-10-13 19:52:18]

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

[wangcheng, 2016-10-13 19:52:18]

https://codereview.chromium.org/2420553002/diff/60001/source/row_gcc.cc
File source/row_gcc.cc (right):

https://codereview.chromium.org/2420553002/diff/60001/source/row_gcc.cc#newco...
source/row_gcc.cc:2884: "+rm"(width)     // %2
Do you need to check "width" is multiply of 16?

[fbarchard1, 2016-10-13 21:46:27]

https://codereview.chromium.org/2420553002/diff/60001/source/row_gcc.cc
File source/row_gcc.cc (right):

https://codereview.chromium.org/2420553002/diff/60001/source/row_gcc.cc#newco...
source/row_gcc.cc:2884: "+rm"(width)     // %2
On 2016/10/13 19:52:18, wangcheng wrote:
> Do you need to check "width" is multiply of 16?

planar_functions.cc does that check:

#if defined(HAS_ARGBEXTRACTALPHAROW_AVX2)
  if (TestCpuFlag(kCpuHasAVX2)) {
    ARGBExtractAlphaRow = IS_ALIGNED(width, 16) ? ARGBExtractAlphaRow_AVX2
                                                : ARGBExtractAlphaRow_Any_AVX2;
  }
#endif

[fbarchard1, 2016-10-13 22:04:47]

PTAL

Old SSE2 4618
Old AVX2 4109
New AVX2 3792

doing 640x360 100k iterations.
21.7% faster than SSE2.

[wangcheng, 2016-10-13 22:38:27]

lgtm

[fbarchard1, 2016-10-13 22:59:35]

Note that vpshufb can replace vpsrld but appears to be slower
3872 vs 3818 with shift

#ifdef HAS_ARGBEXTRACTALPHAROW_AVX2
static const uvec8 kShuffleAlphaShort_AVX2 = {
  3u, 128u, 128u, 128u, 7u, 128u, 128u, 128u,
  11u, 128u, 128u, 128u, 15u, 128u, 128u, 128u
};

void ARGBExtractAlphaRow_AVX2(const uint8* src_argb, uint8* dst_a, int width) {
 asm volatile (
    "vmovdqa    %3,%%ymm4                      \n"
    "vbroadcastf128 %4,%%ymm5                  \n"
    LABELALIGN
  "1:                                          \n"
    "vmovdqu   " MEMACCESS(0) ", %%ymm0        \n"
    "vmovdqu   " MEMACCESS2(0x20, 0) ", %%ymm1 \n"
    "vpshufb    %%ymm5,%%ymm0,%%ymm0           \n" // vpsrld $0x18, %%ymm0
    "vpshufb    %%ymm5,%%ymm1,%%ymm1           \n"
    "vmovdqu   " MEMACCESS2(0x40, 0) ", %%ymm2 \n"
    "vmovdqu   " MEMACCESS2(0x60, 0) ", %%ymm3 \n"
    "lea       " MEMLEA(0x80, 0) ", %0         \n"
    "vpackssdw  %%ymm1, %%ymm0, %%ymm0         \n"  // mutates
    "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
    "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
    "vpackssdw  %%ymm3, %%ymm2, %%ymm2         \n"  // mutates
    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
    "vpermd     %%ymm0,%%ymm4,%%ymm0           \n"  // unmutate.
    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
    "lea       " MEMLEA(0x20,1) ",%1           \n"
    "sub        $0x20, %2                      \n"
    "jg         1b                             \n"
    "vzeroupper                                \n"
  : "+r"(src_argb),  // %0
    "+r"(dst_a),     // %1
    "+rm"(width)     // %2
  : "m"(kPermdARGBToY_AVX),  // %3
    "m"(kShuffleAlphaShort_AVX2)  // %4
  : "memory", "cc"
    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
  );
}
#endif  // HAS_ARGBEXTRACTALPHAROW_AVX2

[fbarchard1, 2016-10-13 23:03:46]

Description was changed from

==========
Add ARGBExtractAlpha_AVX2 function

Port SSE2 version to AVX2.
BUG=libyuv:572
TEST=/usr/local/google/home/fbarchard/intelsde/sde -skx --
out/Release/libyuv_unittest --gtest_filter=*Extract*
R=magjed@chromium.org
==========

to

==========
Add ARGBExtractAlpha_AVX2 function

Port SSE2 version to AVX2.
BUG=libyuv:572
TEST=/usr/local/google/home/fbarchard/intelsde/sde -skx --
out/Release/libyuv_unittest --gtest_filter=*Extract*
R=wangcheng@google.com, magjed@chromium.org

Committed:
https://chromium.googlesource.com/libyuv/libyuv/+/a5e93766a20ac5fff9e0ef2f5bc...
==========

[fbarchard1, 2016-10-13 23:03:47]

Committed patchset #7 (id:120001) manually as
a5e93766a20ac5fff9e0ef2f5bc7c4bb1a0fdb8e (presubmit successful).