AVX512 compilation error on MSVC 19.16.27035.0 (VS 2017)

[pitrou]

While trying to compile code using xsimd::batch<uint32_t, 16> with AVX512 enabled, I got these errors on AppVeyor:
https://ci.appveyor.com/project/pitrou/arrow/builds/37935012/job/5k8hyx5vm69f5bav#L1169

c:\projects\arrow\cpp\build\xsimd_ep\src\xsimd_ep-install\include\xsimd\types\xsimd_sse_conversion.hpp(90): error C3861: '_mm_cvtepu32_ps': identifier not found
c:\projects\arrow\cpp\build\xsimd_ep\src\xsimd_ep-install\include\xsimd\types\xsimd_avx_conversion.hpp(126): error C3861: '_mm256_cvtepu32_ps': identifier not found
c:\projects\arrow\cpp\build\xsimd_ep\src\xsimd_ep-install\include\xsimd\types\xsimd_avx512_double.hpp(452): error C2440: 'type cast': cannot convert from '__m512d' to '__m512i'
c:\projects\arrow\cpp\build\xsimd_ep\src\xsimd_ep-install\include\xsimd\types\xsimd_avx512_double.hpp(452): note: No constructor could take the source type, or constructor overload resolution was ambiguous
c:\projects\arrow\cpp\build\xsimd_ep\src\xsimd_ep-install\include\xsimd\types\xsimd_avx512_double.hpp(452): error C2660: '_mm512_maskz_and_epi64': function does not take 2 arguments
C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\include\zmmintrin.h(1121): note: see declaration of '_mm512_maskz_and_epi64'
[etc.]

While generated, the code which is being compiled is really simple:
https://github.com/pitrou/arrow/blob/xsimd-bpacking/cpp/src/arrow/util/bpacking_simd512_generated.h

[pitrou]

Here is a similar issue with MSVC 19.28.29337.0 (VS 2019):
https://github.com/pitrou/arrow/runs/1972721711?check_suite_focus=true#step:7:777

[pitrou]

Some thought after looking at the first few errors and checking in the online Intel intrinsics guide:

• _mm_cvtepu32_ps and _mm256_cvtepu32_ps indeed don't exist
• _mm512_maskz_and_epi64 takes 3 arguments not 2
• _mm512_and_epi64 takes 2 arguments not 1
• ...

It seems that somehow xsimd is assuming to find weird intrinsics that are not officially documented by Intel...?

[serge-sans-paille]

@pitrou the github link to bpacking_simd512_generated.h is no longer valid, could you share the faulty source code again, or provide an equivalent reproducer?

[pitrou]

@serge-sans-paille Ah, right, the name of the git branch changed, sorry: https://github.com/pitrou/arrow/blob/ARROW-11839-xsimd-bpacking/cpp/src/arrow/util/bpacking_simd512_generated.h

[serge-sans-paille]

Some thought after looking at the first few errors and checking in the online Intel intrinsics guide:

* `_mm_cvtepu32_ps` and `_mm256_cvtepu32_ps` indeed don't exist

They actually do ( see https://officedaytime.com/simd512e/simdimg/convert.php?f=vcvtudq2ps)
Under Linux, they are defined in avx512vlintrin.h, see https://github.com/gcc-mirror/gcc/blob/master/gcc/config/i386/avx512vlintrin.h#L1487 for the coresponding intrinsic.

Using a more recent version of MSVC should fix the issue, as showcased by https://godbolt.org/z/fGq6hKE18

Can you share the compiler flags you're using?

[JohanMabille]

xsimd is supposed to work starting with VS2015, so I think we should provide a workaround here.

[pitrou]

They actually do ( see https://officedaytime.com/simd512e/simdimg/convert.php?f=vcvtudq2ps)

That's not an official Intel site.
If you look for them in the official Intrinsics guide, you won't find them: https://software.intel.com/sites/landingpage/IntrinsicsGuide/#

So for some reason, non-official docs and Linux headers have an intrinsic that's neither in the official Intel docs nor in the MSVC headers. I don't know why.

As for compiler flags, I tried /arch:AVX512 (see the AppVeyor logs I linked to above). Perhaps something more specific must be used?

[pitrou]

/arch:AVX512 is supposed to enable all desired extensions according to https://docs.microsoft.com/fr-fr/cpp/build/reference/arch-x64?view=msvc-160

[pitrou]

I took at the header (zmmintrin.h) provided with Visual Studio 2017:

• _mm_cvtepu32_ps is not defined
• _mm_cvtepu32_pd is defined as:

#define _mm_cvtepu32_pd(v1) \
        _mm_maskz_cvtepu32_pd((__mmask8)_MM_K0_REG, (v1))

(but an equivalent _mm_maskz_cvtepu32_ps doesn't exist anyway)

• _mm512_cvtepu32_ps is defined as:

#define _mm512_cvtepu32_ps(v) \
        _mm512_cvt_roundepu32_ps((v), _MM_FROUND_CUR_DIRECTION)

[pitrou]

Also, looks like this is not the only case of missing intrinsics:
https://developercommunity.visualstudio.com/t/missing-intrinsic-definitions-in-zmmintrinh/828399

[serge-sans-paille]

Thanksfor the detailed information. Note that the latest version of MSVC, as hinted by the godbolt link above, seems to include support for these non-official intrinsics. @JohanMabille should we implement a portability layer with older MSVC version?

[pitrou]

More spelunking:

• _mm512_ceil_ps(v) should probably be a shortcut for _mm512_roundscale_ps(v, _MM_FROUND_TO_POS_INF)
• ditto for _mm512_ceil_pd to _mm512_roundscale_pd
• similarly, _mm512_floor_ps(v) -> _mm512_roundscale_ps(v, _MM_FROUND_TO_NEG_INF)
• ditto for _mm512_floor_pd

[pitrou]

As for float <-> uint32 conversions, I think it would be ok to leave them undefined on MSVC. It's probably not a common use case.