Unit tests fail when building UB2 binaries on Mac.

[stephen-yee-adsk]

Hi there,

I'm having trouble compiling OpenEXR for UB2 binaries, from an Intel based Mac. (To reproduce the issue build with cmake and set CMAKE_OSX_ARCHTECTURES=x86_64;arm64)

The problem appears to be that this check in openexr/cmake/CMakeLists.txt fails:

check_cxx_source_compiles(
"
int main()
{
#if defined(__SSE2__)
    int n   = 0;
    int eax = 0;
    int edx = 0;
    __asm__(
        \"xgetbv     ;\"
        \"vzeroupper  \"
        : \"=a\"(eax), \"=d\"(edx) : \"c\"(n) : );
#else
#    error No SSE support enabled to query AVX support
#endif
}
" OPENEXR_IMF_HAVE_GCC_INLINE_ASM_AVX
)

Arm doesn't support SSE so I think that's why it fails in UB2 compilation.

However OPENEXR_IMF_HAVE_GCC_INLINE_ASM_AVX gets "burned in" to OpenEXRConfigInternal.h, even when compiling the x86_64 "side" of the universal binaries.

This causes some tests to fail. In particular testCPUIdent:

  7/127 Test   #7: OpenEXRCore.testCPUIdent ..................Subprocess aborted***Exception:   0.01 sec
tempDir = '/var/tmp/OpenEXR_D8YKiB/': 24

=======
Running testCPUIdent
CPU Id test f16c mismatch: 1 vs 0
Core Test failed: false
           file:/openexr/src/test/OpenEXRCoreTest/base_units.cpp
           line:375
       function:void testCPUIdent(const std::string &)

CpuId is affected by OPENEXR_IMF_HAVE_GCC_INLINE_ASM_AVX and reports that f16c is not available. That is incorrect, my cpu should support it. Whereas check_for_x86_simd does not depend on that macro, and correctly reports that f16c is available.

This also seems to affect two other tests related to DWA compression

53/127 Test  #53: OpenEXRCore.testDWAACompression ...........Subprocess aborted***Exception:   1.08 sec
 tempDir = '/var/tmp/OpenEXR_9HJaaa/': 24
 
 =======
 Running testDWAACompression
   zeroes tiled: no sampling 1, 1 comp 8
   zeroes tiled: yes sampling 1, 1 comp 8
   zeroes tiled: no sampling 1, 2 comp 8
   zeroes tiled: no sampling 2, 1 comp 8
   zeroes tiled: no sampling 2, 2 comp 8
   pattern1 tiled: no sampling 1, 1 comp 8
   pattern1 tiled: yes sampling 1, 1 comp 8
   pattern1 tiled: no sampling 1, 2 comp 8
   pattern1 tiled: no sampling 2, 1 comp 8
   pattern1 tiled: no sampling 2, 2 comp 8
   pattern2 tiled: no sampling 1, 1 comp 8
 B half at 1276, 1 not equal: C++ loaded C 0x3c4b (1.07324) vs C loaded C 0x3c4e (1.07617)
 Core Test failed: a == b
            file:/openexr/src/test/OpenEXRCoreTest/compression.cpp
            line:475
        function:static void pixels::compareExact(uint16_t, uint16_t, int, int, const char *, const char *, const char *)

The following tests FAILED:
	 53 - OpenEXRCore.testDWAACompression (Subprocess aborted)
	 54 - OpenEXRCore.testDWABCompression (Subprocess aborted)

If I hack the check_cxx_source_compiles code snippet to always pass, then all of the tests pass in x86_64 mode. I have not tried running in arm64 mode though, as I do not have access to an Apple silicon Mac.

System info:

>cmake --version
cmake version 3.20.3

>clang++ -v 
clang version 14.0.3 
Target: x86_64-apple-darwin21.6.0
Thread model: posix

>sw_vers
ProductName:	macOS
ProductVersion:	12.7
BuildVersion:	21G816

>sysctl -a|grep brand
machdep.cpu.brand_string: Intel(R) Core(TM) i9-9880H CPU @ 2.30GHz
machdep.cpu.brand: 0```

[cary-ilm]

Thanks for the report, we'll investigate shortly.