From 5a95d47858e98e1325e01c63a94688cb9bea8b67 Mon Sep 17 00:00:00 2001 From: Alexandr Guzhva Date: Tue, 1 Aug 2023 06:08:44 -0700 Subject: [PATCH] Upgrade AVX2 code for SQ8 (#2942) Summary: More efficient code for SQ8 for AVX2. For clang-15, improves a number of Instructions per cycle (IPC) from 2.49 to 3.20 Pull Request resolved: https://github.com/facebookresearch/faiss/pull/2942 Reviewed By: algoriddle Differential Revision: D47946167 Pulled By: mdouze fbshipit-source-id: da864bac8d452f2eb111ca356e54a8a69cd03dbf --- faiss/impl/ScalarQuantizer.cpp | 21 +++++++++------------ 1 file changed, 9 insertions(+), 12 deletions(-) diff --git a/faiss/impl/ScalarQuantizer.cpp b/faiss/impl/ScalarQuantizer.cpp index 4ffec5e6ca..cff2004603 100644 --- a/faiss/impl/ScalarQuantizer.cpp +++ b/faiss/impl/ScalarQuantizer.cpp @@ -74,18 +74,15 @@ struct Codec8bit { } #ifdef __AVX2__ - static __m256 decode_8_components(const uint8_t* code, int i) { - uint64_t c8 = *(uint64_t*)(code + i); - __m128i c4lo = _mm_cvtepu8_epi32(_mm_set1_epi32(c8)); - __m128i c4hi = _mm_cvtepu8_epi32(_mm_set1_epi32(c8 >> 32)); - // __m256i i8 = _mm256_set_m128i(c4lo, c4hi); - __m256i i8 = _mm256_castsi128_si256(c4lo); - i8 = _mm256_insertf128_si256(i8, c4hi, 1); - __m256 f8 = _mm256_cvtepi32_ps(i8); - __m256 half = _mm256_set1_ps(0.5f); - f8 = _mm256_add_ps(f8, half); - __m256 one_255 = _mm256_set1_ps(1.f / 255.f); - return _mm256_mul_ps(f8, one_255); + static inline __m256 decode_8_components(const uint8_t* code, int i) { + const uint64_t c8 = *(uint64_t*)(code + i); + + const __m128i i8 = _mm_set1_epi64x(c8); + const __m256i i32 = _mm256_cvtepu8_epi32(i8); + const __m256 f8 = _mm256_cvtepi32_ps(i32); + const __m256 half_one_255 = _mm256_set1_ps(0.5f / 255.f); + const __m256 one_255 = _mm256_set1_ps(1.f / 255.f); + return _mm256_fmadd_ps(f8, one_255, half_one_255); } #endif };