IGNOREME: distance functions with golang simd (v1.26)

Makefile

@@ -49,17 +49,23 @@
 #  % MO_CL_CUDA=1 make
 
 # where am I
+ifeq ($(GO),)
+	GO=go
+endif
+
 ROOT_DIR = $(shell dirname $(realpath $(lastword $(MAKEFILE_LIST))))
 BIN_NAME := mo-service
 UNAME_S := $(shell uname -s | tr A-Z a-z)
 UNAME_M := $(shell uname -m)
-GOPATH := $(shell go env GOPATH)
-GO_VERSION=$(shell go version)
+GOPATH := $(shell $(GO) env GOPATH)
+GO_VERSION=$(shell $(GO) version)
 BRANCH_NAME=$(shell git rev-parse --abbrev-ref HEAD)
 LAST_COMMIT_ID=$(shell git rev-parse --short HEAD)
 BUILD_TIME=$(shell date +%s)
 MO_VERSION=$(shell git symbolic-ref -q --short HEAD || git describe --tags --exact-match)
-GO_MODULE=$(shell go list -m)
+GO_MODULE=$(shell $(GO) list -m)
+GO_MAJOR_VERSION = $(shell $(GO) version | cut -c 14- | cut -d' ' -f1 | cut -d'.' -f1)
+GO_MINOR_VERSION = $(shell $(GO) version | cut -c 14- | cut -d' ' -f1 | cut -d'.' -f2)
 
 # check the MUSL_TARGET from https://musl.cc
 # make MUSL_TARGET=aarch64-linux musl to cross make the aarch64 linux executable
@@ -78,6 +84,7 @@ ifneq ($(GOARCH)$(TARGET_ARCH)$(GOOS)$(TARGET_OS),)
 $(error cross compilation has been disabled)
 endif
 
+
 ###############################################################################
 # default target
 ###############################################################################
@@ -151,8 +158,8 @@ help:
 
 .PHONY: vendor-build
 vendor-build:
-	$(info [go mod vendor])
-	@go mod vendor
+	$(info [$(GO) mod vendor])
+	@$(GO) mod vendor
 
 ###############################################################################
 # code generation
@@ -161,7 +168,7 @@ vendor-build:
 .PHONY: config
 config:
 	$(info [Create build config])
-	@go mod tidy
+	@$(GO) mod tidy
 
 .PHONY: generate-pb
 generate-pb:
@@ -178,37 +185,55 @@ pb: vendor-build generate-pb fmt
 
 VERSION_INFO :=-X '$(GO_MODULE)/pkg/version.GoVersion=$(GO_VERSION)' -X '$(GO_MODULE)/pkg/version.BranchName=$(BRANCH_NAME)' -X '$(GO_MODULE)/pkg/version.CommitID=$(LAST_COMMIT_ID)' -X '$(GO_MODULE)/pkg/version.BuildTime=$(BUILD_TIME)' -X '$(GO_MODULE)/pkg/version.Version=$(MO_VERSION)'
 THIRDPARTIES_INSTALL_DIR=$(ROOT_DIR)/thirdparties/install
+CGO_DIR=$(ROOT_DIR)/cgo
 RACE_OPT :=
 DEBUG_OPT :=
 CGO_DEBUG_OPT :=
 TAGS :=
+GOTAGS :=
+GOEXPERIMENT_OPT :=
+
+ifeq ("$(UNAME_M)", "x86_64")
+  ifeq ($(shell expr $(GO_MAJOR_VERSION) \>= 1), 1)
+    ifeq ($(shell expr $(GO_MINOR_VERSION) \>= 26), 1)
+	#GOEXPERIMENT_OPT=GOEXPERIMENT=simd
+    endif
+  endif
+  ifneq ($(GOAMD64),)
+	GOEXPERIMENT_OPT+=GOAMD64=$(GOAMD64)
+  endif
+endif
 
 ifeq ($(MO_CL_CUDA),1)
   ifeq ($(CONDA_PREFIX),)
     $(error CONDA_PREFIX env variable not found.)
   endif
 	CUVS_CFLAGS := -I$(CONDA_PREFIX)/include
-	CUVS_LDFLAGS := -L$(CONDA_PREFIX)/envs/go/lib -lcuvs -lcuvs_c
+	CUVS_LDFLAGS := -L$(CONDA_PREFIX)/lib -lcuvs -lcuvs_c
 	CUDA_CFLAGS := -I/usr/local/cuda/include $(CUVS_CFLAGS)
 	CUDA_LDFLAGS := -L/usr/local/cuda/lib64/stubs -lcuda -L/usr/local/cuda/lib64 -lcudart $(CUVS_LDFLAGS) -lstdc++
-	TAGS += -tags "gpu"
+	TAGS += gpu
 endif
 
 ifeq ($(TYPECHECK),1)
-	TAGS += -tags "typecheck"
+	TAGS += typecheck
 endif
 
-CGO_OPTS :=CGO_CFLAGS="-I$(THIRDPARTIES_INSTALL_DIR)/include $(CUDA_CFLAGS)"
-GOLDFLAGS=-ldflags="-extldflags '$(CUDA_LDFLAGS) -L$(THIRDPARTIES_INSTALL_DIR)/lib -Wl,-rpath,\$${ORIGIN}/lib -fopenmp' $(VERSION_INFO)"
+CGO_OPTS :=CGO_CFLAGS="-I$(CGO_DIR) -I$(THIRDPARTIES_INSTALL_DIR)/include $(CUDA_CFLAGS)"
+GOLDFLAGS=-ldflags="-extldflags '$(CUDA_LDFLAGS) -L$(CGO_DIR) -lmo -L$(THIRDPARTIES_INSTALL_DIR)/lib -Wl,-rpath,\$${ORIGIN}/lib -fopenmp' $(VERSION_INFO)"
 
 ifeq ("$(UNAME_S)","darwin")
-GOLDFLAGS:=-ldflags="-extldflags '-L$(THIRDPARTIES_INSTALL_DIR)/lib -Wl,-rpath,@executable_path/lib' $(VERSION_INFO)"
+GOLDFLAGS:=-ldflags="-extldflags '-L$(CGO_DIR) -lmo -L$(THIRDPARTIES_INSTALL_DIR)/lib -Wl,-rpath,@executable_path/lib' $(VERSION_INFO)"
 endif
 
 ifeq ($(GOBUILD_OPT),)
 	GOBUILD_OPT :=
 endif
 
+ifneq ($(TAGS),)
+	GOTAGS := -tags "$(TAGS)"
+endif
+
 .PHONY: cgo
 cgo: thirdparties
 	@(cd cgo; ${MAKE} ${CGO_DEBUG_OPT})
@@ -222,7 +247,7 @@ thirdparties:
 .PHONY: build
 build: config cgo thirdparties
 	$(info [Build binary])
-	$(CGO_OPTS) go build $(TAGS) $(RACE_OPT) $(GOLDFLAGS) $(DEBUG_OPT) $(GOBUILD_OPT) -o $(BIN_NAME) ./cmd/mo-service
+	$(GOEXPERIMENT_OPT) $(CGO_OPTS) $(GO) build $(GOTAGS) $(RACE_OPT) $(GOLDFLAGS) $(DEBUG_OPT) $(GOBUILD_OPT) -o $(BIN_NAME) ./cmd/mo-service
 
 # https://wiki.musl-libc.org/getting-started.html
 # https://musl.cc/
@@ -248,17 +273,17 @@ musl-thirdparties: musl-install
 .PHONY: musl
 musl: override CGO_OPTS += CC=$(MUSL_CC)
 musl: override GOLDFLAGS:=-ldflags="--linkmode 'external' --extldflags '-static -L$(THIRDPARTIES_INSTALL_DIR)/lib -lstdc++ -Wl,-rpath,\$${ORIGIN}/lib' $(VERSION_INFO)"
-musl: override TAGS := -tags musl
+musl: override GOTAGS := -tags musl
 musl: musl-install musl-cgo config musl-thirdparties
 musl:
 	$(info [Build binary(musl)])
-	$(CGO_OPTS) go build $(TAGS) $(RACE_OPT) $(GOLDFLAGS) $(DEBUG_OPT) $(GOBUILD_OPT) -o $(BIN_NAME) ./cmd/mo-service
+	$(CGO_OPTS) $(GO) build $(GOTAGS) $(RACE_OPT) $(GOLDFLAGS) $(DEBUG_OPT) $(GOBUILD_OPT) -o $(BIN_NAME) ./cmd/mo-service
 
 # build mo-tool
 .PHONY: mo-tool
 mo-tool: config cgo thirdparties
 	$(info [Build mo-tool tool])
-	$(CGO_OPTS) go build $(GOLDFLAGS) -o mo-tool ./cmd/mo-tool
+	$(CGO_OPTS) $(GO) build $(GOLDFLAGS) -o mo-tool ./cmd/mo-tool
 
 # build mo-service binary for debugging with go's race detector enabled
 # produced executable is 10x slower and consumes much more memory
@@ -1007,7 +1032,7 @@ launch-minio-debug: debug dev-up-minio-local
 clean:
 	$(info [Clean up])
 	$(info Clean go test cache)
-	@go clean -testcache
+	@$(GO) clean -testcache
 	rm -f $(BIN_NAME)
 	rm -rf $(ROOT_DIR)/vendor
 	rm -rf $(MUSL_DIR)
@@ -1027,12 +1052,12 @@ fmt:
 .PHONY: install-static-check-tools
 install-static-check-tools:
 	@curl -sSfL https://raw.githubusercontent.com/golangci/golangci-lint/master/install.sh | bash -s -- -b $(GOPATH)/bin v2.6.2
-	@go install github.com/matrixorigin/linter/cmd/molint@latest
-	@go install github.com/apache/skywalking-eyes/cmd/license-eye@v0.4.0
+	@$(GO) install github.com/matrixorigin/linter/cmd/molint@latest
+	@$(GO) install github.com/apache/skywalking-eyes/cmd/license-eye@v0.4.0
 
 .PHONY: static-check
 static-check: config err-check
-	$(CGO_OPTS) go vet -vettool=`which molint` ./...
+	$(CGO_OPTS) $(GO) vet -vettool=`which molint` ./...
 	$(CGO_OPTS) license-eye -c .licenserc.yml header check
 	$(CGO_OPTS) license-eye -c .licenserc.yml dep check
 	$(CGO_OPTS) golangci-lint run -v -c .golangci.yml ./...

cgo/Makefile

@@ -1,48 +1,77 @@
 DEBUG_OPT :=
 UNAME_M := $(shell uname -m)
+UNAME_S := $(shell uname -s)
+CC ?= gcc
 
 # Yeah, fast math.  We want it to be fast, for all xcall, 
 # IEEE compliance should not be an issue.
 OPT_LV := -O3 -ffast-math -ftree-vectorize -funroll-loops
-CFLAGS=-std=c99 -g ${OPT_LV} -Wall -Werror -I../thirdparties/install/include
-OBJS=mo.o arith.o compare.o logic.o xcall.o usearchex.o bloom.o
-CUDA_OBJS=
+COMMON_CFLAGS := -g $(OPT_LV) -Wall -Werror -fPIC -I../thirdparties/install/include
+CFLAGS := -std=c99 $(COMMON_CFLAGS)
+OBJS := mo.o arith.o compare.o logic.o xcall.o usearchex.o bloom.o
+CUDA_OBJS :=
+LDFLAGS := -L../thirdparties/install/lib -lusearch_c
+TARGET_LIB := libmo.so
+
+ifeq ($(UNAME_S),Darwin)
+	TARGET_LIB := libmo.dylib
+	LDFLAGS += -dynamiclib -undefined dynamic_lookup -install_name @rpath/$(TARGET_LIB)
+else
+	LDFLAGS += -shared
+endif
 
 ifeq ($(UNAME_M), x86_64)
-	CFLAGS+= -march=haswell
+	CFLAGS += -march=haswell
 endif
 
 ifeq ($(MO_CL_CUDA),1)
+	ifeq ($(CONDA_PREFIX),)
+		$(error CONDA_PREFIX env variable not found. Please activate your conda environment.)
+	endif
 	CC = /usr/local/cuda/bin/nvcc 
-	CFLAGS = -ccbin g++ -m64 --shared -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_89,code=sm_89 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_90,code=compute_90
+	CFLAGS = -ccbin g++ -m64 -Xcompiler -fPIC -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_89,code=sm_89 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_90,code=compute_90
 	CFLAGS += -I../thirdparties/install/include -DMO_CL_CUDA
 	CUDA_OBJS += cuda/cuda.o
-	CUDA_LDFLAGS := -L/usr/local/cuda/lib64/stubs -lcuda -L/usr/local/cuda/lib64 -lcudart -lstdc++
+	# Explicitly include all needed libraries for shared library linking
+	CUDA_LDFLAGS := -L/usr/local/cuda/lib64/stubs -lcuda -L/usr/local/cuda/lib64 -lcudart -L$(CONDA_PREFIX)/lib -lcuvs -lcuvs_c -ldl -lrmm -lstdc++
+	LDFLAGS += $(CUDA_LDFLAGS)
 endif
 
-all: libmo.a
+.PHONY: all clean test debug
+
+all: $(TARGET_LIB) libmo.a
 
-libmo.a: $(OBJS) 
+$(TARGET_LIB): $(OBJS) 
 ifeq ($(MO_CL_CUDA),1)
-	make -C cuda
+	$(MAKE) -C cuda
+	$(MAKE) -C cuvs
+	$(CC) $(LDFLAGS) -o $@ $(OBJS) $(CUDA_OBJS) cuvs/*.o
+else
+	$(CC) $(LDFLAGS) -o $@ $(OBJS)
 endif
-	ar -rcs libmo.a $(OBJS) $(CUDA_OBJS)
 
-# 
-#	$(CC) -o libmo.a $(OBJS) $(CUDA_OBJS) $(CUDA_LDFLAGS)
+libmo.a: $(OBJS)
+ifeq ($(MO_CL_CUDA),1)
+	$(MAKE) -C cuda
+	$(MAKE) -C cuvs
+	ar -rcs $@ $(OBJS) $(CUDA_OBJS) cuvs/*.o
+else
+	ar -rcs $@ $(OBJS)
+endif
 
+%.o: %.c
+	$(CC) $(CFLAGS) -c $< -o $@
 
-test: libmo.a
-	make -C test
+test: $(TARGET_LIB)
+	$(MAKE) -C test
 
-.PHONY: debug
 debug: override OPT_LV := -O0
 debug: override DEBUG_OPT := debug
 debug: all
 
-.PHONY: clean
 clean:
-	rm -f *.o *.a *.so
+	rm -f *.o *.a *.so *.dylib
 ifeq ($(MO_CL_CUDA),1)
-	make -C cuda clean
+	$(MAKE) -C cuda clean
+	$(MAKE) -C cuvs clean
 endif

cgo/README.md

@@ -1,25 +1,28 @@
 MatrixOne CGO Kernel
 ===============================
 
-This directory contains cgo source code for MO.   Running
-make should produce two files to be used by go code.
-On go side, go will `include "mo.h"` and `-lmo`.   
+This directory contains CGO source code for MatrixOne. Running `make` produces the core library files used by Go code.
+
+On the Go side, the integration typically uses `mo.h` and links against the generated libraries:
 ```
 mo.h
-libmo.a
+libmo.a / libmo.so
 ```
 
-`mo.h` should be pristine, meaning it only contains C function
-prototype used by go.  The only datatypes that can be passed 
-between go and c code are int and float/double and pointer.   
-Always explicitly specify int size such as `int32_t`, `uint64_t`.
-Do not use `int`, `long`, etc.
+`mo.h` should remain pristine, containing only C function prototypes for Go to consume. Data passed between Go and C should be limited to standard types (int, float, double, pointers). Always specify explicit integer sizes (e.g., `int32_t`, `uint64_t`) and avoid platform-dependent types like `int` or `long`.
+
+GPU Support (CUDA & cuVS)
+-------------------------
+The kernel supports GPU acceleration for certain operations (e.g., vector search) via NVIDIA CUDA and the cuVS library.
+
+- **Build Flag:** GPU support is enabled by setting `MO_CL_CUDA=1` during the build.
+- **Environment:** Requires a working CUDA installation and a Conda environment with `cuvs` and `rmm` installed.
+- **Source Code:** GPU-specific code resides in the `cuda/` and `cuvs/` subdirectories.
 
 Implementation Notes
---------------------------------
+--------------------
 
-1. Pure C.
-2. Use memory passed from go.  Try not allocate memory in C code.
-3. Only depends on libc and libm.
-4. If 3rd party lib is absolutely necessary, import source code 
-   and build from source. If 3rd party lib is C++, wrap it completely in C.
+1. **Language:** Core kernel is Pure C. GPU extensions use C++ and CUDA, wrapped in a C-compatible interface.
+2. **Memory Management:** Prefer using memory allocated and passed from Go. Minimize internal allocations in C/C++ code.
+3. **Dependencies:** The base kernel depends only on `libc`, `libm`, and `libusearch`. GPU builds introduce dependencies on CUDA, `cuvs`, and `rmm`.
+4. **Third-party Libraries:** If a third-party library is necessary, it should be built from source (see `thirdparties/` directory). C++ libraries must be fully wrapped in C before being exposed to Go.

cgo/cuda/Makefile

@@ -395,7 +395,7 @@ $(FATBIN_FILE): mocl.cu
 	$(EXEC) $(NVCC) $(INCLUDES) $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -fatbin $<
 
 cuda.o: cuda.cpp
-	$(EXEC) $(NVCC) $(INCLUDES) -O3 --shared $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
+	$(EXEC) $(NVCC) $(INCLUDES) -O3 --shared -Xcompiler -fPIC $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<
 
 mytest.o: cuda.cpp $(FATBIN_FILE)
 	$(EXEC) $(NVCC) $(INCLUDES) -DTEST_RUN -g -O0 $(ALL_CCFLAGS) $(GENCODE_FLAGS) -o $@ -c $<

cgo/cuvs/Makefile

@@ -0,0 +1,71 @@
+# Makefile for MatrixOne cuVS C Wrapper
+
+UNAME_M := $(shell uname -m)
+CUDA_PATH ?= /usr/local/cuda
+NVCC := $(CUDA_PATH)/bin/nvcc
+
+ifeq ($(CONDA_PREFIX),)
+  $(error CONDA_PREFIX env variable not found. Please activate your conda environment.)
+endif
+
+# Compilation flags
+# Added --extended-lambda because raft/core/copy.cuh requires it for some internal headers
+NVCC_FLAGS := -std=c++17 -x cu -Xcompiler "-Wall -Wextra -fPIC -O2" --extended-lambda --expt-relaxed-constexpr
+NVCC_FLAGS += -I. -I$(CUDA_PATH)/include -I$(CONDA_PREFIX)/include -I$(CONDA_PREFIX)/include/rapids -I$(CONDA_PREFIX)/include/raft -I$(CONDA_PREFIX)/include/cuvs
+NVCC_FLAGS += -DLIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE -DRAFT_SYSTEM_LITTLE_ENDIAN=1
+
+# Linking flags
+LDFLAGS := -shared
+LDFLAGS += -L$(CUDA_PATH)/lib64/stubs -lcuda -L$(CUDA_PATH)/lib64 -lcudart
+LDFLAGS += -L$(CONDA_PREFIX)/lib -lcuvs -lcuvs_c -ldl -lrmm -lrapids_logger
+LDFLAGS += -Xlinker -lpthread -Xlinker -lm
+
+# Target library
+TARGET := libmocuvs.so
+
+# Source files
+SRCS := brute_force_c.cpp ivf_flat_c.cpp cagra_c.cpp kmeans_c.cpp helper.cpp
+OBJS := $(SRCS:.cpp=.o)
+
+# Test configuration
+TESTDIR := test
+OBJDIR := obj
+TEST_EXE := test_cuvs_worker
+TEST_SRCS := $(TESTDIR)/main_test.cu \
+             $(TESTDIR)/brute_force_test.cu \
+             $(TESTDIR)/ivf_flat_test.cu \
+             $(TESTDIR)/cagra_test.cu \
+             $(TESTDIR)/kmeans_test.cu
+
+TEST_OBJS := $(patsubst $(TESTDIR)/%.cu, $(OBJDIR)/test/%.o, $(TEST_SRCS))
+
+.PHONY: all clean test
+
+all: $(OBJS)
+
+$(TARGET): $(OBJS)
+	@echo "Linking shared library $@"
+	$(NVCC) $(LDFLAGS) $^ -o $@
+
+%.o: %.cpp
+	@echo "Compiling $< with NVCC"
+	$(NVCC) $(NVCC_FLAGS) -c $< -o $@
+
+# Test targets
+test: $(TEST_EXE)
+	@echo "Running tests..."
+	./$(TEST_EXE)
+
+$(TEST_EXE): $(TEST_OBJS)
+	@echo "NVCCLD $@"
+	$(NVCC) $(subst -x cu,,$(NVCC_FLAGS)) $^ $(subst -shared,,$(LDFLAGS)) -o $@
+
+$(OBJDIR)/test/%.o: $(TESTDIR)/%.cu
+	@mkdir -p $(@D)
+	@echo "NVCC $<"
+	$(NVCC) -std=c++17 -Xcompiler "-Wall -Wextra -fPIC -O2" --extended-lambda --expt-relaxed-constexpr -I. -I$(CUDA_PATH)/include -I$(CONDA_PREFIX)/include -I$(CONDA_PREFIX)/include/rapids -I$(CONDA_PREFIX)/include/raft -I$(CONDA_PREFIX)/include/cuvs -DLIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE -DRAFT_SYSTEM_LITTLE_ENDIAN=1 -c $< -o $@
+
+clean:
+	@echo "Cleaning up..."
+	rm -f $(TARGET) *.o $(TEST_EXE)
+	rm -rf $(OBJDIR)