missing auto-vectorization for AVX512

[jinjunzh]

PR #4025 introduced a new architecture mode, avx512_spr, which enables the use of features available since Intel® Sapphire Rapids. During benchmarking OpenSearch on Sapphire Rapids server, it's observed that some key hot functions are not built as expected, such as faiss::fvec_inner_product and faiss::fvec_inner_product_batch_4.
Since this is definitely my FAISS library built with FAISS_OPT_LEVEL=avx512_spr, the two functions in the library should have been auto-vectorized using AVX-512 instructions, rather than AVX2.​

After conducting an in-depth analysis, I‘ve identified that the root cause is the absence of the compiler flag ​​-mprefer-vector-width=512​​ in the ​​CMakeLists.txt​​ file. A patch to fix this issue is provided below. Please help confirm whether this is indeed a bug.

diff --git a/faiss/CMakeLists.txt b/faiss/CMakeLists.txt
index 8ee79bd9..46460501 100644
--- a/faiss/CMakeLists.txt
+++ b/faiss/CMakeLists.txt
@@ -291,7 +291,7 @@ endif()
if(NOT WIN32)
- target_compile_options(faiss_avx512_spr PRIVATE $<$<COMPILE_LANGUAGE:CXX>:-march=sapphirerapids -mtune=sapphirerapids>)
+ target_compile_options(faiss_avx512_spr PRIVATE $<$<COMPILE_LANGUAGE:CXX>:-march=sapphirerapids -mtune=sapphirerapids -mprefer-vector-width=512>)
else()
target_compile_options(faiss_avx512_spr PRIVATE $<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)

[mdouze]

@onestardao, Faiss is a vector search engine and the points above are related to RAG, so they are not very relevant.
Please do not spam Faiss issues with unrelated reports.

[mdouze]

@jinjunzh thanks for the report. I understand that sometimes the compiler prefers to use AVX2 instructions without the additional flags. Could you confirm that issuing AVX512 instructions does indeed improve the runtime for some cases?

[jinjunzh]

@mdouze
I was able to reproduce the performance improvements with AVX-512 as claimed by OpenSearch.

Compared to AVX2, I observed a performance gain of roughly 10% on indexing and 5% on searching. The tests were conducted using the Cohere-1M+HNSW+Innerproduct+FP32 configuration, which aligns with the results detailed in the original sources below:

Blog Post: Boost OpenSearch vector search performance with Intel AVX-512

Pull Request: Opensearch-KNN-2110

[alexanderguzhva]

@jinjunzh yes

[mnorris11]

cc @subhadeepkaran

[mnorris11]

Plan after discussion: Land this after dynamic dispatch stack of changes, or add this into the stack.