AVX2 support doesn't compile with MSVC

[borrrden]

Summary

The AVX2 implementation of FAISS appears to use intrinsics that are GNU specific (or at least don't port to Windows via MSVC).

Platform

Windows 10.0.19045
MSVC 2022

Faiss version: git tag v1.7.4

Installed from: Attempting to compile

Faiss compilation options:

Probably doesn't much matter but using OpenBLAS compiled myself
CMake set(FAISS_OPT_LEVEL "avx2" CACHE INTERNAL "") and link with faiss_avx2 instead of faiss

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

1. Get any BLAS implementation in order to satisfy the dependency requirements for building this library
2. Standard CMake dance (build directory, etc)
3. cmake -DFAISS_OPT_LEVEL=avx2 -DFAISS_ENABLE_GPU=OFF -DFAISS_ENABLE_PYTHON=OFF .. (assuming the source directory is one level up from the build directory)
4. cmake --build . --config Release --target faiss_avx2 ..

Fix instructions

This boils down to 3 distinct issues:

1. The definition of _mm_prefetch in MSVC (and actually in the Intel Intrinsics Guide suggesting GCC / clang are mistaken) takes a const char * as the first argument, not const void *. This base is peppered with calls directly passing float pointers.
2. __m128i_u and __m256i_u are not portable. However they are the exact same datatype as the non _u variants, just with some implied side effects so I'm guessing it is safe to define them away to their non _u variants #if defined(_MSC_VER) && !defined(__clang__).
3. By default Visual Studio implemented OpenMP 2.0, which doesn't allow signed indexes in for loops. The following code violates this

   faiss/faiss/utils/distances_fused/simdlib_based.cpp

   Lines 262 to 266 in d87888b

   	#pragma omp parallel for schedule(dynamic)
   	for (size_t i = 0; i < nx_p; i += NX_POINTS_PER_LOOP) {
   	kernel<DIM, NX_POINTS_PER_LOOP, NY_POINTS_PER_LOOP>(
   	x, y, y_transposed.data(), ny, res, y_norms, i);
   	}

However, issue 3 can be worked around with some trickery, involving override CMake's FindMP module by setting OpenMP_CXX_FLAGS="-openmp:llvm". This turns on the "better" support (which is OpenMP 2.5 with a few 3.1 features including unsigned index types) for OpenMP in Visual Studio as noted here

[algoriddle]

I will add AVX2 to our contbuild to verify.

[algoriddle]

Confirmed that MSVC is not supported by the current implementation of AVX2. We welcome any PRs.

[borrrden]

Any opinions on the way to approach this? I had simply cast all the prefetch calls blindly from float * to char *. I'm assuming that this function, taking a void * in the case of Linux, is not going to be doing anything special due to the fact that it is a float. It compiles and appears to function but if someone who knows better than me has a different approach let me know.

Also do you care if the iterator above is signed or unsigned? Does it have any implications? Because as I mentioned it can be either fixed in code or in build option (though the build option will create a dependency on LLVM libomp which I think at this point might be included in the MSVC runtime but I haven't looked carefully into that yet)