Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

fix(vectordb): default x86 build to AVX2 and disable AVX512 (#291)

kkkwjx5cf378dc00c3
diff --git a/setup.py b/setup.py
index 92e5b45cf..364d78ddd 100644
--- a/setup.py
+++ b/setup.py
@@ -115,6 +115,7 @@ class CMakeBuildExtension(build_ext):
f"-Dpybind11_DIR={pybind11.get_cmake_dir()}",
f"-DCMAKE_C_COMPILER={C_COMPILER_PATH}",
f"-DCMAKE_CXX_COMPILER={CXX_COMPILER_PATH}",
+ f"-DOV_X86_SIMD_LEVEL={os.environ.get('OV_X86_SIMD_LEVEL', 'AVX2')}",
]
if sys.platform == "darwin":
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index add7ac929..d35f64a0b 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -3,16 +3,73 @@ cmake_minimum_required(VERSION 3.12)
project(openviking_cpp)
include(CheckCXXCompilerFlag)
-check_cxx_compiler_flag("-march=native" HAVE_MARCH_NATIVE)
-if(HAVE_MARCH_NATIVE)
- add_compile_options(-march=native)
-else()
- if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64|AMD64")
+
+set(OV_X86_SIMD_LEVEL "AVX2" CACHE STRING "x86 SIMD level: SSE3|AVX2|AVX512|NATIVE")
+set_property(CACHE OV_X86_SIMD_LEVEL PROPERTY STRINGS SSE3 AVX2 AVX512 NATIVE)
+
+set(OV_PLATFORM_X86 OFF)
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64|AMD64|i[3-6]86")
+ set(OV_PLATFORM_X86 ON)
+endif()
+
+if(OV_PLATFORM_X86)
+ string(TOUPPER "${OV_X86_SIMD_LEVEL}" OV_X86_SIMD_LEVEL_UPPER)
+ set(OV_X86_COMPILE_FLAGS)
+
+ if(OV_X86_SIMD_LEVEL_UPPER STREQUAL "NATIVE")
+ check_cxx_compiler_flag("-march=native" HAVE_MARCH_NATIVE)
+ if(HAVE_MARCH_NATIVE)
+ list(APPEND OV_X86_COMPILE_FLAGS -march=native)
+ else()
+ message(WARNING "-march=native is not supported by this compiler; falling back to AVX2")
+ set(OV_X86_SIMD_LEVEL_UPPER "AVX2")
+ endif()
+ endif()
+
+ if(OV_X86_SIMD_LEVEL_UPPER STREQUAL "AVX512")
+ foreach(FLAG -mavx512f -mavx512bw -mavx512dq -mavx512vl)
+ string(REPLACE "-" "_" FLAG_VAR_SUFFIX "${FLAG}")
+ set(FLAG_VAR "HAVE_${FLAG_VAR_SUFFIX}")
+ check_cxx_compiler_flag("${FLAG}" ${FLAG_VAR})
+ if(NOT ${FLAG_VAR})
+ message(FATAL_ERROR "Compiler does not support ${FLAG}; cannot build with OV_X86_SIMD_LEVEL=AVX512")
+ endif()
+ list(APPEND OV_X86_COMPILE_FLAGS ${FLAG})
+ endforeach()
+ elseif(OV_X86_SIMD_LEVEL_UPPER STREQUAL "AVX2")
+ check_cxx_compiler_flag("-mavx2" HAVE_MAVX2)
+ if(HAVE_MAVX2)
+ list(APPEND OV_X86_COMPILE_FLAGS -mavx2)
+ add_compile_definitions(OV_DISABLE_AVX512=1)
+ foreach(FLAG -mno-avx512f -mno-avx512bw -mno-avx512dq -mno-avx512vl)
+ string(REPLACE "-" "_" FLAG_VAR_SUFFIX "${FLAG}")
+ set(FLAG_VAR "HAVE_${FLAG_VAR_SUFFIX}")
+ check_cxx_compiler_flag("${FLAG}" ${FLAG_VAR})
+ if(${FLAG_VAR})
+ list(APPEND OV_X86_COMPILE_FLAGS ${FLAG})
+ endif()
+ endforeach()
+ else()
+ message(WARNING "Compiler does not support -mavx2; falling back to SSE3")
+ set(OV_X86_SIMD_LEVEL_UPPER "SSE3")
+ endif()
+ endif()
+
+ if(OV_X86_SIMD_LEVEL_UPPER STREQUAL "SSE3")
check_cxx_compiler_flag("-msse3" HAVE_SSE3)
if(HAVE_SSE3)
- add_compile_options(-msse3)
+ list(APPEND OV_X86_COMPILE_FLAGS -msse3)
+ add_compile_definitions(OV_DISABLE_AVX512=1)
+ else()
+ message(FATAL_ERROR "Compiler does not support -msse3 on x86 target")
endif()
endif()
+
+ if(OV_X86_COMPILE_FLAGS)
+ add_compile_options(${OV_X86_COMPILE_FLAGS})
+ endif()
+
+ message(STATUS "OpenViking x86 SIMD level: ${OV_X86_SIMD_LEVEL_UPPER}")
endif()
set(CMAKE_CXX_STANDARD 17)
diff --git a/src/index/detail/vector/common/space_ip.h b/src/index/detail/vector/common/space_ip.h
index 51959017d..56c86910d 100644
--- a/src/index/detail/vector/common/space_ip.h
+++ b/src/index/detail/vector/common/space_ip.h
@@ -33,7 +33,7 @@ static float inner_product_avx512(const void* v1, const void* v2,
for (; i + 16 <= dim; i += 16) {
__m512 a = _mm512_loadu_ps(pv1 + i);
__m512 b = _mm512_loadu_ps(pv2 + i);
- sum = _mm512_fmadd_ps(a, b, sum);
+ sum = _mm512_add_ps(sum, _mm512_mul_ps(a, b));
}
float res = _mm512_reduce_add_ps(sum);
@@ -59,7 +59,7 @@ static float inner_product_avx(const void* v1, const void* v2,
for (; i + 8 <= dim; i += 8) {
__m256 a = _mm256_loadu_ps(pv1 + i);
__m256 b = _mm256_loadu_ps(pv2 + i);
- sum = _mm256_fmadd_ps(a, b, sum);
+ sum = _mm256_add_ps(sum, _mm256_mul_ps(a, b));
}
__m128 sum_low = _mm256_extractf128_ps(sum, 0);
diff --git a/src/index/detail/vector/common/space_l2.h b/src/index/detail/vector/common/space_l2.h
index 69c402ad0..773b642ec 100644
--- a/src/index/detail/vector/common/space_l2.h
+++ b/src/index/detail/vector/common/space_l2.h
@@ -35,7 +35,7 @@ static float l2_sqr_avx512(const void* v1, const void* v2, const void* params) {
__m512 a = _mm512_loadu_ps(pv1 + i);
__m512 b = _mm512_loadu_ps(pv2 + i);
__m512 diff = _mm512_sub_ps(a, b);
- sum = _mm512_fmadd_ps(diff, diff, sum);
+ sum = _mm512_add_ps(sum, _mm512_mul_ps(diff, diff));
}
float res = _mm512_reduce_add_ps(sum);
@@ -64,7 +64,7 @@ static float l2_sqr_avx(const void* v1, const void* v2, const void* params) {
__m256 a = _mm256_loadu_ps(pv1 + i);
__m256 b = _mm256_loadu_ps(pv2 + i);
__m256 diff = _mm256_sub_ps(a, b);
- sum = _mm256_fmadd_ps(diff, diff, sum);
+ sum = _mm256_add_ps(sum, _mm256_mul_ps(diff, diff));
}
// Reduce AVX register
@@ -131,7 +131,7 @@ static float l2_sqr_neon(const void* v1, const void* v2, const void* params) {
size_t dim = *static_cast<const size_t*>(params);
float dis = 0;
krl_L2sqr(pv1, pv2, dim, &dis, 1);
- return 1.0f - dis;
+ return dis;
}
#endif
diff --git a/src/index/detail/vector/common/vector_base.h b/src/index/detail/vector/common/vector_base.h
index 2f02e3629..27526d429 100644
--- a/src/index/detail/vector/common/vector_base.h
+++ b/src/index/detail/vector/common/vector_base.h
@@ -10,7 +10,8 @@
#if defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || \
defined(_M_IX86)
#define OV_PLATFORM_X86
-#if defined(__AVX512F__)
+#include <immintrin.h>
+#if defined(__AVX512F__) && !defined(OV_DISABLE_AVX512)
#define OV_SIMD_AVX512
#endif
#if defined(__AVX__)