Assess performance bottlenecks of the gather instruction in JVector

[r-devulap]

Due to security vulnerabilities in Intel processors up to the Ice Lake generation, the gather instruction was microcode patched and is now extremely slow. Intel advisory: https://www.intel.com/content/www/us/en/security-center/advisory/intel-sa-00828.html. JVector uses gather instructions in multiple places that are worth looking into:

jvector/jvector-native/src/main/c/jvector_simd.c

Line 292 in f0c52d4

float assemble_and_sum_f32_512(const float* data, int dataBase, const unsigned char* baseOffsets, int baseOffsetsOffset, int baseOffsetsLength) {

jvector/jvector-native/src/main/c/jvector_simd.c

Line 323 in f0c52d4

float pq_decoded_cosine_similarity_f32_512(const unsigned char* baseOffsets, int baseOffsetsOffset, int baseOffsetsLength, int clusterCount, const float* partialSums, const float* aMagnitude, float bMagnitude) {

‣ Ref: other libraries (e.g., NumPy’s x86 simd sort) improved performance by replacing gather with scalar loads: numpy/x86-simd-sort#65