ARGBToI411 test fails on ARM platform

[GoogleCodeExporter]

What steps will reproduce the problem?
1. build libyuvtest from master branch
2. run libyuvtest with following command on Android device:
LIBYUV_WIDTH=130 LIBYUV_HEIGHT=80 LIBYUV_REPEAT=1 LIBYUV_DISABLE_NEON=0 
./libyuv_unittest
3.

What is the expected output? What do you see instead?
Expected: Pass all tests.
Instead:
[  FAILED  ] 4 tests, listed below:
[  FAILED  ] libyuvTest.ARGBToI411_Any
[  FAILED  ] libyuvTest.ARGBToI411_Unaligned
[  FAILED  ] libyuvTest.ARGBToI411_Invert
[  FAILED  ] libyuvTest.ARGBToI411_Opt

What version of the product are you using? On what operating system?

Please provide any additional information below.

Original issue reported on code.google.com by zhongwei...@linaro.org on 15 Sep 2015 at 7:55

• Blocked on: #504

[GoogleCodeExporter]

Passes on Intel

LIBYUV_WIDTH=130 LIBYUV_HEIGHT=80 out/Release/libyuv_unittest 
--gtest_filter=*ARGBToI411*
Note: Google Test filter = *ARGBToI411*
[==========] Running 4 tests from 1 test case.
[----------] Global test environment set-up.
[----------] 4 tests from libyuvTest
[ RUN      ] libyuvTest.ARGBToI411_Any
[       OK ] libyuvTest.ARGBToI411_Any (2 ms)
[ RUN      ] libyuvTest.ARGBToI411_Unaligned
[       OK ] libyuvTest.ARGBToI411_Unaligned (1 ms)
[ RUN      ] libyuvTest.ARGBToI411_Invert
[       OK ] libyuvTest.ARGBToI411_Invert (1 ms)
[ RUN      ] libyuvTest.ARGBToI411_Opt
[       OK ] libyuvTest.ARGBToI411_Opt (1 ms)
[----------] 4 tests from libyuvTest (5 ms total)

[----------] Global test environment tear-down
[==========] 4 tests from 1 test case ran. (5 ms total)
[  PASSED  ] 4 tests.

Original comment by fbarch...@google.com on 17 Sep 2015 at 8:04

[GoogleCodeExporter]

Original comment by fbarch...@google.com on 17 Sep 2015 at 8:09

• Changed title: ARGBToI411 test fails on ARM platform
• Changed state: Accepted

[GoogleCodeExporter]

Original comment by fbarch...@chromium.org on 12 Oct 2015 at 11:48

• Now blocked on: #504

[GoogleCodeExporter]

With default parameters, this function passes
[ RUN      ] LibYUVConvertTest.ARGBToI411_Any
[       OK ] LibYUVConvertTest.ARGBToI411_Any (5 ms)
[ RUN      ] LibYUVConvertTest.ARGBToI411_Unaligned
[       OK ] LibYUVConvertTest.ARGBToI411_Unaligned (4 ms)
[ RUN      ] LibYUVConvertTest.ARGBToI411_Invert
[       OK ] LibYUVConvertTest.ARGBToI411_Invert (4 ms)
[ RUN      ] LibYUVConvertTest.ARGBToI411_Opt
[       OK ] LibYUVConvertTest.ARGBToI411_Opt (4 ms)
[----------] 4 tests from LibYUVConvertTest (18 ms total)

Original comment by fbarch...@chromium.org on 14 Oct 2015 at 10:33

[GoogleCodeExporter]

LIBYUV_WIDTH=130 LIBYUV_HEIGHT=80 LIBYUV_REPEAT=1 LIBYUV_DISABLE_ASM=1 
out/Release/libyuv_unittest

This is the Intel code

     0000000000556f00 <ARGBToUV411Row_C>:
       push   %r13
       cmp    $0x3,%ecx
       push   %r12
       push   %rbp
       push   %rbx
     ↓ jle    2c3
       lea    -0x4(%rcx),%r12d
       mov    %rsi,%rbx
       mov    %rdi,%rax
       shr    $0x2,%r12d
       add    $0x1,%r12
       lea    (%rdx,%r12,1),%rbp
       nop

 28:   movzbl (%rax),%r9d
       movzbl 0x4(%rax),%r8d
       add    $0x1,%rdx
       movzbl 0x1(%rax),%r10d
       movzbl 0x2(%rax),%r11d
       add    $0x1,%rbx
       add    $0x10,%rax
       add    %r9d,%r8d
       movzbl -0x8(%rax),%r9d
       add    %r9d,%r8d
       movzbl -0x4(%rax),%r9d
       add    %r9d,%r8d
       movzbl -0xb(%rax),%r9d
       sar    $0x2,%r8d
       mov    %r8d,%r13d
       lea    (%r8,%r8,8),%r8d
       add    %r10d,%r9d
       movzbl -0x7(%rax),%r10d
       add    %r8d,%r8d
       add    %r10d,%r9d
       movzbl -0x3(%rax),%r10d
       add    %r10d,%r9d
       movzbl -0xa(%rax),%r10d
       sar    $0x2,%r9d
       add    %r11d,%r10d
       movzbl -0x6(%rax),%r11d
       add    %r11d,%r10d
       movzbl -0x2(%rax),%r11d
       add    %r11d,%r10d
       mov    $0x70,%r11d
       imul   %r11d,%r13d
       sar    $0x2,%r10d
       imul   $0xffffffb6,%r9d,%r11d
       add    %r11d,%r13d
       imul   $0xffffffda,%r10d,%r11d
       lea    0x8080(%r13,%r11,1),%r11d
       mov    $0xffffffa2,%r13d
       imul   %r13d,%r9d
       sar    $0x8,%r11d
       mov    %r11b,-0x1(%rbx)
       mov    $0x70,%r11d
       imul   %r11d,%r10d
       add    %r10d,%r9d
       sub    %r8d,%r9d
       add    $0x8080,%r9d
       sar    $0x8,%r9d
       mov    %r9b,-0x1(%rdx)
       cmp    %rbp,%rdx
     ↑ jne    28

       mov    %r12,%rax
       add    %r12,%rsi
       shl    $0x4,%rax
       add    %rax,%rdi
104:   and    $0x3,%ecx
       cmp    $0x3,%ecx
     ↓ je     198
       cmp    $0x2,%ecx
     ↓ je     238
       cmp    $0x1,%ecx
     ↓ je     128
       pop    %rbx
       pop    %rbp
       pop    %r12
       pop    %r13
     ← retq
       nop

Original comment by fbarch...@chromium.org on 16 Oct 2015 at 6:16

[GoogleCodeExporter]

I've reproduced the error... had to hard code the unittest/unittest.cc 
dimensions to 130x80.  130x72 also fails... its the width not being a multiple 
of 4.

411 subsampling is really not used in practice, so it may be better to remove 
the functionality than spend time fixing it.

Original comment by fbarch...@google.com on 20 Oct 2015 at 1:37

• Changed state: Started

[GoogleCodeExporter]

#if defined(HAS_ARGBTOYROW_NEON)
  if (TestCpuFlag(kCpuHasNEON)) {
    ARGBToYRow = ARGBToYRow_Any_NEON;
    if (IS_ALIGNED(width, 8)) {
      ARGBToYRow = ARGBToYRow_NEON;
    }
  }
#endif
#if defined(HAS_ARGBTOUV411ROW_NEON)
  if (TestCpuFlag(kCpuHasNEON)) {
    ARGBToUV411Row = ARGBToUV411Row_Any_NEON;
    if (IS_ALIGNED(width, 32)) {
      ARGBToUV411Row = ARGBToUV411Row_NEON;
    }
  }
#endif

Original comment by fbarch...@google.com on 20 Oct 2015 at 1:45