‘_mm256_cmpge_epi32_mask’ was not declared for fbgemm when compiling CPU only pytorch

[pytorch]

‘_mm256_cmpge_epi32_mask’ was not declared for fbgemm when compiling CPU only pytorch

[dskhudia]

Actually just searched about it a bit more and found that it's a gcc bug. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79932
Essentially the code works if compiled under -Og, -O2, or -O3 but fails with -O0.

The fix is back ported to gcc 5.5 but not to 5.4.

[dskhudia]

Seems like gcc 5.4 doesn't have the intrinsic _mm256_cmpge_epi32_mask implemented.

Code containing _mm256_cmpge_epi32_mask works with 5.5 but doesn't with gcc 5.4 : https://godbolt.org/z/f5bsrrjW1

Could you use a more recent version of compiler?

[imaginary-person]

Which compiler & what version of it are you using?

[zengzengsun]

Which compiler & what version of it are you using?

here is my compiler info:
g++ (Ubuntu 5.4.0-6ubuntu1~16.04.12) 5.4.0 20160609

[malfet]

Hmm, but to be fair, FBGEMM cmake scripts should detect whether AVX512 support is available and skip respective codeparts

[dskhudia]

General AVX512 support is available but it seems gcc 5.4 has missing intrinsic APIs.

[dskhudia]

@malfet, do we build Pytorch with gcc 5.4 in CI?

[malfet]

@dskhudia, no we do not, but it doesn't mean we should not check wether or not compiler extensions we want to use is available.

[dskhudia]

@malfet Agreed. Just want to make sure that the minimum version of gcc supported by PyTorch is also supported by FBGEMM. Do you happen to know the minimum version of gcc supported by PyTorch?

[zengzengsun]

Actually just searched about it a bit more and found that it's a gcc bug. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79932
Essentially the code works if compiled under -Og, -O2, or -O3 but fails with -O0.

The fix is back ported to gcc 5.5 but not to 5.4.

@dskhudia yes~ you're right. I was just going to tell you that the issue is only happened in debug building.
Thanks all guys for this issue