fix: SimSIMD dynamic dispatch (#5870)

* fix: SimSIMD dynamic dispatch was added

* fix: additional logging added

* fix: additional logging added

* fix: simsimd dynamic dispatch updated

* fix: so library build

* fix: makefile updated

* fix: dynamic dispatch

* fix: cleanup

* fix: dynamic/header only libs build

* fix: Makefile checkout

* fix: cleanup

* fix: review comments

* fix: review comments

* fix: review comments

* fix: review comments

* fix: review comments

* fix: review comments

* fix: review comments

* fix: review comments

* Update src/core/search/vector_utils.cc

Co-authored-by: Roman Gershman <[email protected]>
Signed-off-by: Volodymyr Yavdoshenko <[email protected]>

---------

Signed-off-by: Volodymyr Yavdoshenko <[email protected]>
Co-authored-by: Roman Gershman <[email protected]>

Author: vyavdoshenko

src/CMakeLists.txt

@@ -1,6 +1,6 @@
 option(ENABLE_GIT_VERSION "Build with Git metadata" OFF)
 
-option(USE_SIMSIMD "Enable SimSIMD vector optimizations" OFF)
+option(WITH_SIMSIMD "Enable SimSIMD vector optimizations" OFF)
 option(SIMSIMD_NATIVE_F16 "Enable native float16 support in SimSIMD" OFF)
 option(WITH_SEARCH "Enable compilation of search module" ON)
 

src/core/search/CMakeLists.txt

@@ -12,16 +12,13 @@ add_library(dfly_search_core base.cc ast_expr.cc query_driver.cc search.cc indic
 target_link_libraries(dfly_search_core base redis_lib absl::strings
   TRDP::reflex TRDP::uni-algo TRDP::hnswlib)
 
-if(USE_SIMSIMD)
+if(WITH_SIMSIMD)
   target_link_libraries(dfly_search_core TRDP::simsimd)
-  target_compile_definitions(dfly_search_core PRIVATE USE_SIMSIMD=1)
-
-  # Conditionally disable native float16 support (default behavior)
-  if(NOT SIMSIMD_NATIVE_F16)
-    target_compile_definitions(dfly_search_core PRIVATE
-      SIMSIMD_NATIVE_F16=0
-      SIMSIMD_NATIVE_BF16=0)
-  endif()
+  target_compile_definitions(dfly_search_core PRIVATE
+    WITH_SIMSIMD=1
+    SIMSIMD_DYNAMIC_DISPATCH=1
+    SIMSIMD_NATIVE_F16=$<IF:$<BOOL:${SIMSIMD_NATIVE_F16}>,1,0>
+    SIMSIMD_NATIVE_BF16=$<IF:$<BOOL:${SIMSIMD_NATIVE_F16}>,1,0>)
 endif()
 
 cxx_test(compressed_sorted_set_test dfly_search_core LABELS DFLY)
@@ -31,14 +28,11 @@ cxx_test(rax_tree_test redis_test_lib LABELS DFLY)
 cxx_test(search_parser_test dfly_search_core LABELS DFLY)
 cxx_test(search_test redis_test_lib dfly_search_core LABELS DFLY)
 
-if(USE_SIMSIMD)
+if(WITH_SIMSIMD)
   target_link_libraries(search_test TRDP::simsimd)
-  target_compile_definitions(search_test PRIVATE USE_SIMSIMD=1)
-
-  # Conditionally disable native float16 support (default behavior)
-  if(NOT SIMSIMD_NATIVE_F16)
-    target_compile_definitions(search_test PRIVATE
-      SIMSIMD_NATIVE_F16=0
-      SIMSIMD_NATIVE_BF16=0)
-  endif()
+  target_compile_definitions(search_test PRIVATE
+    WITH_SIMSIMD=1
+    SIMSIMD_DYNAMIC_DISPATCH=1
+    SIMSIMD_NATIVE_F16=$<IF:$<BOOL:${SIMSIMD_NATIVE_F16}>,1,0>
+    SIMSIMD_NATIVE_BF16=$<IF:$<BOOL:${SIMSIMD_NATIVE_F16}>,1,0>)
 endif()

src/core/search/search_test.cc

@@ -149,6 +149,8 @@ class SearchTest : public ::testing::Test {
   static void SetUpTestSuite() {
     auto* tlh = mi_heap_get_backing();
     init_zmalloc_threadlocal(tlh);
+    // Initialize SimSIMD runtime for tests that may exercise vector kernels
+    InitSimSIMD();
   }
 
   SearchTest() {
@@ -967,6 +969,8 @@ TEST_F(SearchTest, VectorDistanceConsistency) {
 }
 
 static void BM_VectorSearch(benchmark::State& state) {
+  // Ensure SimSIMD dynamic dispatch is initialized for the benchmark
+  InitSimSIMD();
   unsigned ndims = state.range(0);
   unsigned nvecs = state.range(1);
 
@@ -1659,6 +1663,8 @@ TEST(CosineDistanceTest, ZeroVectors) {
 
 // Unified vector distance benchmarks using VectorDistance function
 static void BM_VectorDistance(benchmark::State& state) {
+  // Ensure SimSIMD dynamic dispatch is initialized for the benchmark
+  InitSimSIMD();
   size_t dims = state.range(0);
   size_t num_pairs = state.range(1);
   VectorSimilarity sim = static_cast<VectorSimilarity>(state.range(2));
@@ -1691,6 +1697,8 @@ static void BM_VectorDistance(benchmark::State& state) {
 
 // Intensive benchmark with batch processing
 static void BM_VectorDistance_Intensive(benchmark::State& state) {
+  // Ensure SimSIMD dynamic dispatch is initialized for the benchmark
+  InitSimSIMD();
   size_t dims = 512;  // Fixed medium size
   size_t batch_size = 1000;
   VectorSimilarity sim = static_cast<VectorSimilarity>(state.range(0));

src/core/search/vector_utils.cc

@@ -15,9 +15,7 @@ using namespace std;
 
 namespace {
 
-#ifdef USE_SIMSIMD
-// Native float16 support is controlled via CMake option SIMSIMD_NATIVE_F16
-// Default behavior disables native F16/BF16 for maximum compatibility
+#ifdef WITH_SIMSIMD
 #include <simsimd/simsimd.h>
 #endif
 
@@ -29,7 +27,7 @@ namespace {
 
 // Euclidean vector distance: sqrt( sum: (u[i] - v[i])^2  )
 FAST_MATH float L2Distance(const float* u, const float* v, size_t dims) {
-#ifdef USE_SIMSIMD
+#ifdef WITH_SIMSIMD
   simsimd_distance_t distance = 0;
   simsimd_l2_f32(u, v, dims, &distance);
   return static_cast<float>(distance);
@@ -44,7 +42,7 @@ FAST_MATH float L2Distance(const float* u, const float* v, size_t dims) {
 // Inner product distance: 1 - dot_product(u, v)
 // For normalized vectors, this is equivalent to cosine distance
 FAST_MATH float IPDistance(const float* u, const float* v, size_t dims) {
-#ifdef USE_SIMSIMD
+#ifdef WITH_SIMSIMD
   // Use SimSIMD dot product and convert to inner product distance: 1 - dot(u, v).
   simsimd_distance_t dot = 0;
   simsimd_dot_f32(u, v, dims, &dot);
@@ -58,7 +56,7 @@ FAST_MATH float IPDistance(const float* u, const float* v, size_t dims) {
 }
 
 FAST_MATH float CosineDistance(const float* u, const float* v, size_t dims) {
-#ifdef USE_SIMSIMD
+#ifdef WITH_SIMSIMD
   simsimd_distance_t distance = 0;
   simsimd_cos_f32(u, v, dims, &distance);
   return static_cast<float>(distance);
@@ -112,4 +110,10 @@ float VectorDistance(const float* u, const float* v, size_t dims, VectorSimilari
   return 0.0f;
 }
 
+void InitSimSIMD() {
+#if defined(WITH_SIMSIMD)
+  (void)simsimd_capabilities();
+#endif
+}
+
 }  // namespace dfly::search

src/core/search/vector_utils.h

@@ -8,6 +8,9 @@
 
 namespace dfly::search {
 
+// Initializes SimSIMD runtime if dynamic dispatch is enabled.
+void InitSimSIMD();
+
 OwnedFtVector BytesToFtVector(std::string_view value);
 
 // Returns std::nullopt if value can not be converted to the vector

src/external_libs.cmake

@@ -164,13 +164,33 @@ add_third_party(
   LIB libhdr_histogram_static.a
 )
 
-if(USE_SIMSIMD)
+if(WITH_SIMSIMD)
+  # Compute integer macros for native half-precision support.
+  set(SIMSIMD_NATIVE_F16_VAL 0)
+  set(SIMSIMD_NATIVE_BF16_VAL 0)
+  if(SIMSIMD_NATIVE_F16)
+    set(SIMSIMD_NATIVE_F16_VAL 1)
+    set(SIMSIMD_NATIVE_BF16_VAL 1)
+  endif()
+
+  # Build statically via add_third_party using the C shim with dynamic dispatch.
   add_third_party(
     simsimd
     URL https://github.com/ashvardanian/SimSIMD/archive/refs/tags/v6.5.3.tar.gz
-    BUILD_COMMAND echo SKIP
-    INSTALL_COMMAND cp -R <SOURCE_DIR>/include ${THIRD_PARTY_LIB_DIR}/simsimd/
-    LIB "none"
+    BUILD_IN_SOURCE 1
+    BUILD_COMMAND bash -c "\
+      mkdir -p ${THIRD_PARTY_LIB_DIR}/simsimd/lib && \
+      ${CMAKE_C_COMPILER} -O3 -fPIC -DNDEBUG \
+        -DSIMSIMD_DYNAMIC_DISPATCH=1 \
+        -DSIMSIMD_NATIVE_F16=${SIMSIMD_NATIVE_F16_VAL} \
+        -DSIMSIMD_NATIVE_BF16=${SIMSIMD_NATIVE_BF16_VAL} \
+        -I<SOURCE_DIR>/include -c <SOURCE_DIR>/c/lib.c -o <SOURCE_DIR>/lib.o && \
+      ar rcs <SOURCE_DIR>/libsimsimd.a <SOURCE_DIR>/lib.o"
+    INSTALL_COMMAND bash -c "\
+      mkdir -p ${THIRD_PARTY_LIB_DIR}/simsimd/include ${THIRD_PARTY_LIB_DIR}/simsimd/lib && \
+      cp -R <SOURCE_DIR>/include/* ${THIRD_PARTY_LIB_DIR}/simsimd/include/ && \
+      cp <SOURCE_DIR>/libsimsimd.a ${THIRD_PARTY_LIB_DIR}/simsimd/lib/"
+    LIB libsimsimd.a
   )
 endif()
 
@@ -196,10 +216,3 @@ add_library(TRDP::fast_float INTERFACE IMPORTED)
 add_dependencies(TRDP::fast_float fast_float_project)
 set_target_properties(TRDP::fast_float PROPERTIES
                       INTERFACE_INCLUDE_DIRECTORIES "${FAST_FLOAT_INCLUDE_DIR}")
-
-if(USE_SIMSIMD)
-  add_library(TRDP::simsimd INTERFACE IMPORTED)
-  add_dependencies(TRDP::simsimd simsimd_project)
-  set_target_properties(TRDP::simsimd PROPERTIES
-                        INTERFACE_INCLUDE_DIRECTORIES "${SIMSIMD_INCLUDE_DIR}")
-endif()

src/server/main_service.cc

@@ -34,6 +34,7 @@ extern "C" {
 #include "base/flag_utils.h"
 #include "base/flags.h"
 #include "base/logging.h"
+#include "core/search/vector_utils.h"
 #include "facade/dragonfly_connection.h"
 #include "facade/error.h"
 #include "facade/reply_builder.h"
@@ -914,6 +915,11 @@ void Service::Init(util::AcceptServer* acceptor, std::vector<facade::Listener*>
   InitRedisTables();
   facade::Connection::Init(pp_.size());
 
+#if defined(WITH_SEARCH)
+  // Initialize SimSIMD runtime if needed (explicit, avoids implicit static initializers)
+  dfly::search::InitSimSIMD();
+#endif
+
   config_registry.RegisterMutable("dbfilename");
   config_registry.Register("dbnum");  // equivalent to databases in redis.
   config_registry.Register("dir");