feat(vectordb): integrate KRL for ARM Kunpeng vector search optimization

src/CMakeLists.txt

@@ -39,6 +39,12 @@ set(Python3_ARCH_INCLUDE_DIR "/usr/include/${CMAKE_SYSTEM_PROCESSOR}-linux-gnu/"
 
 find_package(Python3 COMPONENTS Interpreter Development REQUIRED)
 
+# On Linux, pybind11 modules don't need to link against libpython
+# This prevents issues with static libpython that wasn't built with -fPIC
+if(UNIX AND NOT APPLE)
+    set(Python3_LIBRARIES "")
+endif()
+
 find_package(pybind11 REQUIRED)
 find_package(Threads REQUIRED)
 
@@ -59,11 +65,24 @@ endif()
 
 add_subdirectory(../third_party/spdlog-1.14.1 ${CMAKE_BINARY_DIR}/spdlog_build)
 
+# ARM platform detection and KRL integration
+set(OV_PLATFORM_ARM OFF)
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|ARM64|arm64")
+    set(OV_PLATFORM_ARM ON)
+    message(STATUS "Building for ARM platform with KRL support")
+    add_subdirectory(../third_party/krl ${CMAKE_BINARY_DIR}/krl_build)
+endif()
+
 include_directories(.)
 include_directories(../third_party/)
 include_directories(../third_party/leveldb-1.23/include/)
 include_directories(../third_party/spdlog-1.14.1/include/)
 
+# Add KRL include directory for ARM platform
+if(OV_PLATFORM_ARM)
+    include_directories(../third_party/krl/include/)
+endif()
+
 if(NOT DEFINED Python3_INCLUDE_DIRS)
     set(Python3_INCLUDE_DIRS 
         ${Python3_ARCH_INCLUDE_DIR}
@@ -88,6 +107,11 @@ target_link_libraries(engine_impl PRIVATE
     leveldb
 )
 
+# Link KRL library for ARM platform
+if(OV_PLATFORM_ARM)
+    target_link_libraries(engine_impl PRIVATE krl)
+endif()
+
 if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
     if(CMAKE_CXX_COMPILER_VERSION VERSION_LESS "9.0")
         target_link_libraries(engine_impl PRIVATE stdc++fs)

src/index/detail/vector/common/space_ip.h

@@ -108,10 +108,27 @@ static float inner_product_sse(const void* v1, const void* v2,
 }
 #endif
 
+#if defined(OV_SIMD_NEON)
+#include "krl.h"
+
+// ARM NEON optimized inner product using KRL library
+static float inner_product_neon(const void* v1, const void* v2,
+                                const void* params) {
+  const float* pv1 = static_cast<const float*>(v1);
+  const float* pv2 = static_cast<const float*>(v2);
+  size_t dim = *static_cast<const size_t*>(params);
+  float dis = 0;
+  krl_ipdis(pv1, pv2, dim, &dis, 1);
+  return dis;
+}
+#endif
+
 class InnerProductSpace : public VectorSpace<float> {
  public:
   explicit InnerProductSpace(size_t dim) : dim_(dim) {
-#if defined(OV_SIMD_AVX512)
+#if defined(OV_SIMD_NEON)
+    metric_func_ = inner_product_neon;
+#elif defined(OV_SIMD_AVX512)
     metric_func_ = inner_product_avx512;
 #elif defined(OV_SIMD_AVX)
     metric_func_ = inner_product_avx;

src/index/detail/vector/common/space_l2.h

@@ -121,12 +121,28 @@ static float l2_sqr_sse(const void* v1, const void* v2, const void* params) {
 }
 #endif
 
+#if defined(OV_SIMD_NEON)
+#include "krl.h"
+
+// ARM NEON optimized L2 squared distance using KRL library
+static float l2_sqr_neon(const void* v1, const void* v2, const void* params) {
+  const float* pv1 = static_cast<const float*>(v1);
+  const float* pv2 = static_cast<const float*>(v2);
+  size_t dim = *static_cast<const size_t*>(params);
+  float dis = 0;
+  krl_L2sqr(pv1, pv2, dim, &dis, 1);
+  return 1.0f - dis;
+}
+#endif
+
 class L2Space : public VectorSpace<float> {
  public:
   explicit L2Space(size_t dim) : dim_(dim) {
     // Select best implementation at runtime based on compile-time flags
     // In a real scenario, we might want dynamic dispatch based on CPUID
-#if defined(OV_SIMD_AVX512)
+#if defined(OV_SIMD_NEON)
+    metric_func_ = l2_sqr_neon;
+#elif defined(OV_SIMD_AVX512)
     metric_func_ = l2_sqr_avx512;
 #elif defined(OV_SIMD_AVX)
     metric_func_ = l2_sqr_avx;

src/index/detail/vector/common/vector_base.h

@@ -21,6 +21,15 @@
 #endif
 #endif
 
+// ARM Platform Detection
+#if defined(__aarch64__) || defined(_M_ARM64)
+#define OV_PLATFORM_ARM
+#define OV_SIMD_NEON
+#if defined(__ARM_FEATURE_SVE)
+#define OV_SIMD_SVE
+#endif
+#endif
+
 // Memory Alignment Macros
 #if defined(_MSC_VER)
 #define OV_ALIGN_32 __declspec(align(32))

third_party/krl/CMakeLists.txt

@@ -0,0 +1,33 @@
+cmake_minimum_required(VERSION 3.12)
+
+project(krl CXX)
+
+# Only build on ARM platform
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|ARM64|arm64")
+    # Set C++ standard
+    set(CMAKE_CXX_STANDARD 17)
+    set(CMAKE_CXX_STANDARD_REQUIRED ON)
+
+    # ARM compile options - use more conservative flags
+    add_compile_options(-O2 -fPIC -fvisibility=hidden)
+
+    # Minimal set for OpenViking: only krl_L2sqr and krl_ipdis (float, single-vector)
+    # C++ sources following OpenViking code style
+    set(KRL_SOURCES
+        ${CMAKE_CURRENT_SOURCE_DIR}/src/L2distance_simd.cpp
+        ${CMAKE_CURRENT_SOURCE_DIR}/src/IPdistance_simd.cpp
+    )
+
+    # Create static library
+    add_library(krl STATIC ${KRL_SOURCES})
+
+    # Include directories
+    target_include_directories(krl PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+
+    # PIC for static library
+    set_target_properties(krl PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
+    message(STATUS "KRL library configured for ARM platform (core distance functions only)")
+else()
+    message(STATUS "KRL library skipped - not ARM platform")
+endif()