Lack of AVX-512 causing installation errors while installing Executorch

[cad-rlc]

Issue: Lack of AVX-512 on our CPU causing installation errors while installing Executorch on main and viable/strict branch.

Resolution: Redefine _mm_loadu_si64 and _mm_storeu_si64 to their safe AVX2-compatible equivalents using compiler flags. The command used was

CFLAGS="-D_mm_loadu_si64=_mm_loadl_epi64 -D_mm_storeu_si64=_mm_storel_epi64”
CXXFLAGS="-D_mm_loadu_si64=_mm_loadl_epi64 -D_mm_storeu_si64=_mm_storel_epi64"
./install_executorch.sh |& tee executorch_installation.log

Commenting out XNNPACK does not resolve the issue.

Is AVX-512 a requirement for executorch installation? If so, is it documented somewhere?

Thanks

@zonglinpengmeta @mcremon-meta

cc @digantdesai @mcr229 @cbilgin

[digantdesai]

Hey sorry for this experience.

Is AVX-512 a requirement for executorch installation? If so, is it documented somewhere?

No it shouldn't be the case. Can you share where is the error coming from and what is the error?

Redefine _mm_loadu_si64 and _mm_storeu_si64 to their safe AVX2-compatible equivalents using compiler flags.

Sorry can you elaborate a bit more? Is this the error message? or this is how you resolved it?

[cad-rlc]

Hey,

Thanks for the reply.

The error is coming from backends/xnnpack/third-party/XNNPACK/src/qs8-vpreluc/gen/qs8-vpreluc-avx2-u16.c:66:40, with the message:

warning: implicit declaration of function ‘_mm_loadu_si64’; did you mean ‘_mm_loadl_epi64’? [-Wimplicit-function-declaration]
__m256i va0 = _mm256_cvtepi8_epi32(_mm_loadu_si64((const __m128i*) input_a));
^~~~~~~~~~~~~~
_mm_loadl_epi64
error: incompatible type for argument 1 of ‘_mm256_cvtepi8_epi32’
__m256i va0 = _mm256_cvtepi8_epi32(_mm_loadu_si64((const __m128i*) input_a));
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from /usr/lib/gcc/x86_64-redhat-linux/8/include/immintrin.h:43,
from /home/kvariar/ExecuTorch/test/executorch/backends/xnnpack/third-party/XNNPACK/src/qs8-vpreluc/gen/qs8-vpreluc-avx2-u16.c:37:
/usr/lib/gcc/x86_64-redhat-linux/8/include/avx2intrin.h:447:31: note: expected ‘__m128i’ {aka ‘__vector(2) long long int’} but argument is of type ‘int’
_mm256_cvtepi8_epi32 (__m128i __X)
~~~~~~~~^~~

Observing other such error messages too.

The redefinition CFLAGS helped resolve the issue.

[digantdesai]

Using GCC-8 by any chance :)
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78782

We should handle this bug properly in XNNPACK by detecting the version number and switching the intrinsic as you have done.

[cad-rlc]

Hey,

We have been using GCC v12.4.0. We tried with GCC v9.3.0 and faced the same issue.

Could you share the details of where this bug is, or potentially will be, handled?

Thanks

[digantdesai]

If this is an issue with GCC you can take a look at similar fixes - https://github.com/google/XNNPACK/blob/e98747dad01a4b43ad70d43bd87b299dbd7b61e3/CMakeLists.txt#L1071-L1082 (not related this issue just FYI).