Nccl sync v2 28 9 clean

CMakeLists.txt

@@ -0,0 +1,170 @@
+cmake_minimum_required(VERSION 3.25) # ipp6 is using 3.28
+
+# Version information
+# Read makefiles/version.mk file
+file(READ ${CMAKE_SOURCE_DIR}/makefiles/version.mk VERSION_CONTENT)
+string(REGEX REPLACE ".*NCCL_MAJOR[ ]*:=[ ]*([0-9]+).*" "\\1" NCCL_MAJOR "${VERSION_CONTENT}")
+string(REGEX REPLACE ".*NCCL_MINOR[ ]*:=[ ]*([0-9]+).*" "\\1" NCCL_MINOR "${VERSION_CONTENT}")
+string(REGEX REPLACE ".*NCCL_PATCH[ ]*:=[ ]*([0-9]+).*" "\\1" NCCL_PATCH "${VERSION_CONTENT}")
+string(REGEX REPLACE ".*NCCL_SUFFIX[ ]*:=[ ]*([a-zA-Z0-9]*).*" "\\1" NCCL_SUFFIX "${VERSION_CONTENT}")
+string(REGEX REPLACE ".*PKG_REVISION[ ]*:=[ ]*([0-9]+).*" "\\1" PKG_REVISION "${VERSION_CONTENT}")
+math(EXPR NCCL_VERSION_CODE "(${NCCL_MAJOR} * 10000) + (${NCCL_MINOR} * 100) + ${NCCL_PATCH}")
+
+# Make version information available to C++ source files
+add_compile_definitions(
+    NCCL_USE_CMAKE
+    NCCL_MAJOR=${NCCL_MAJOR}
+    NCCL_MINOR=${NCCL_MINOR}
+    NCCL_PATCH=${NCCL_PATCH}
+    NCCL_VERSION_CODE=${NCCL_VERSION_CODE}
+)
+
+set(ENV{NCCL_USE_CMAKE} "1")
+
+project(NCCL VERSION ${NCCL_MAJOR}.${NCCL_MINOR}.${NCCL_PATCH}
+        LANGUAGES CUDA CXX C)
+
+# Make CMAKE_BUILD_TYPE to release by default if not set
+if(NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE "Release")
+endif()
+
+option(VERBOSE "Enable verbose output" OFF)
+option(KEEP "Keep intermediate files" OFF)
+option(DEBUG "Enable debug build" OFF)
+option(ASAN "Enable Address Sanitizer" OFF)
+option(UBSAN "Enable Undefined Behavior Sanitizer" OFF)
+option(TRACE "Enable tracing" OFF)
+option(WERROR "Treat warnings as errors" OFF)
+option(PROFAPI "Enable profiling API" ON)
+option(NVTX "Enable NVTX" ON)
+option(RDMA_CORE "Enable RDMA core" OFF)
+option(NET_PROFILER "Enable network profiler" OFF)
+option(MLX5DV "Enable MLX5DV" OFF)
+option(MAX_EXT_NET_PLUGINS "Maximum external network plugins" 0)
+
+find_package(CUDAToolkit REQUIRED)
+
+# CUDA version detection
+string(REGEX MATCH "([0-9]+\\.[0-9]+)" CUDA_VERSION "${CUDAToolkit_VERSION}")
+
+# Extract major and minor version numbers
+string(REGEX MATCH "([0-9]+)" CUDA_MAJOR "${CUDA_VERSION}")
+string(REGEX MATCH "([0-9]+)$" CUDA_MINOR "${CUDA_VERSION}")
+string(REGEX REPLACE ".*\\.([0-9]+)$" "\\1" CUDA_MINOR "${CUDA_VERSION}")
+
+# Add CUDA version definitions after find_package
+add_compile_definitions(
+    CUDA_MAJOR=${CUDA_MAJOR}
+    CUDA_MINOR=${CUDA_MINOR}
+)
+
+# CUDA architecture flags
+if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES OR CMAKE_CUDA_ARCHITECTURES STREQUAL "")
+    message(STATUS "CMAKE_CUDA_ARCHITECTURES not defined or empty, setting default values based on CUDA version")
+
+    if(${CUDA_MAJOR} LESS 9)
+        set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61")
+    elseif(${CUDA_MAJOR} EQUAL 9)
+        set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61;70")
+    elseif(${CUDA_MAJOR} EQUAL 10)
+        set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61;70")
+    elseif(${CUDA_MAJOR} EQUAL 11)
+        if(${CUDA_MINOR} LESS 8)
+            set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61;70;80")
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61;70;80;90")
+        endif()
+    elseif(${CUDA_MAJOR} EQUAL 12)
+        if(${CUDA_MINOR} LESS 8)
+            set(CMAKE_CUDA_ARCHITECTURES "50;60;61;70;80;90")
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "50;60;61;70;80;90;100;120")
+        endif()
+    elseif(${CUDA_MAJOR} EQUAL 13)
+        set(CMAKE_CUDA_ARCHITECTURES "50;60;61;70;80;90;100;110;120")
+    else()
+        # For future CUDA versions, include all architectures up to the latest known
+        set(CMAKE_CUDA_ARCHITECTURES "50;60;61;70;80;90;100;110;120")
+    endif()
+endif()
+message(STATUS "Using CUDA_ARCHITECTURES: ${CMAKE_CUDA_ARCHITECTURES}")
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fPIC -fvisibility=hidden -Wall -Wno-unused-function -Wno-sign-compare -Wvla -g")
+set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --expt-extended-lambda -Xptxas -maxrregcount=96 -Xfatbin -compress-all -fPIC")
+
+# Sanitizer options
+if(ASAN)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address")
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fsanitize=address -static-libasan")
+endif()
+
+if(UBSAN)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=undefined")
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fsanitize=undefined -static-libubsan")
+endif()
+
+# Additional options
+if(TRACE)
+    add_definitions(-DENABLE_TRACE)
+endif()
+
+if(NOT NVTX)
+    add_definitions(-DNVTX_DISABLE)
+endif()
+
+if(WERROR)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror")
+endif()
+
+if(PROFAPI)
+    add_definitions(-DPROFAPI)
+endif()
+
+set(EXTRA_LIBS)
+
+# RDMA and MLX5DV are Linux-specific features
+if(RDMA_CORE)
+    add_definitions(-DNCCL_BUILD_RDMA_CORE=1)
+    find_library(VERBS_LIBRARY NAMES verbs)
+    if(VERBS_LIBRARY)
+        list(APPEND EXTRA_LIBS ${VERBS_LIBRARY})
+    endif()
+endif()
+
+if(MLX5DV)
+    add_definitions(-DNCCL_BUILD_MLX5DV=1)
+    find_library(MLX5_LIBRARY NAMES mlx5)
+    if(MLX5_LIBRARY)
+        list(APPEND EXTRA_LIBS ${MLX5_LIBRARY})
+    endif()
+endif()
+
+if(NET_PROFILER)
+    add_definitions(-DNCCL_ENABLE_NET_PROFILING=1)
+endif()
+
+if(MAX_EXT_NET_PLUGINS GREATER 0)
+    add_definitions(-DNCCL_NET_MAX_PLUGINS=${MAX_EXT_NET_PLUGINS})
+endif()
+
+add_definitions(-DDOCA_VERBS_USE_CUDA_WRAPPER)
+add_definitions(-DDOCA_VERBS_USE_NET_WRAPPER)
+add_definitions(-DNCCL_GIN_PROXY_ENABLE=1)
+
+# Library dependencies
+find_library(RT_LIBRARY NAMES rt)
+if(RT_LIBRARY)
+    list(APPEND EXTRA_LIBS ${RT_LIBRARY})
+endif()
+
+# Debug/Release specific flags
+set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS} -O0")
+set(CMAKE_CUDA_FLAGS_DEBUG "${CMAKE_CUDA_FLAGS} -O0 -G -g")
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS} -O3")
+set(CMAKE_CUDA_FLAGS_RELEASE "${CMAKE_CUDA_FLAGS} -O3")
+
+add_subdirectory(ext-net)
+add_subdirectory(ext-profiler/example)
+add_subdirectory(ext-tuner/example)
+add_subdirectory(src)

examples/06_device_api/02_gin_alltoall_pure/Makefile

@@ -0,0 +1,84 @@
+#
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+#
+# See LICENSE.txt for license information
+#
+
+# Include common build rules
+include ../../../makefiles/common.mk
+include ../../../makefiles/examples.mk
+
+# Target executable
+TARGET = gin_alltoall_pure_device_api
+
+# Common utilities
+COMMON_INC = ../../common/include
+COMMON_SRC = ../../common/src
+
+# Build configuration
+INCLUDES += -I$(COMMON_INC)
+
+# Source files
+SOURCES = main.cu $(COMMON_SRC)/utils.cc
+OBJECTS = $(SOURCES:.cu=.o)
+OBJECTS := $(OBJECTS:.cc=.o)
+
+# Default target
+all: $(TARGET)
+
+# Build executable
+$(TARGET): $(OBJECTS)
+ifeq ($(MPI),1)
+	$(MPICXX) $(CXXFLAGS) $(OBJECTS) $(LIBRARIES) $(LDFLAGS) -o $@
+else
+	$(CXX) $(CXXFLAGS) $(OBJECTS) $(LIBRARIES) $(LDFLAGS) -lpthread -o $@
+endif
+	@echo "Built target $@"
+
+# Compile source files
+%.o: %.cu
+	$(NVCC) $(NVCUFLAGS) $(INCLUDES) -c $< -o $@
+
+%.o: %.cc
+ifeq ($(MPI),1)
+	$(MPICXX) $(CXXFLAGS) $(INCLUDES) -c $< -o $@
+else
+	$(CXX) $(CXXFLAGS) $(INCLUDES) -c $< -o $@
+endif
+
+# Test target
+test: $(TARGET)
+	@echo "Testing $(TARGET)..."
+ifeq ($(MPI),1)
+	@echo "Running with 2 processes"
+	$(MPIRUN) -np 2 ./$(TARGET)
+else
+	@echo "Running with all available GPUs"
+	./$(TARGET)
+endif
+
+# Clean build artifacts
+clean:
+	rm -f $(OBJECTS) $(TARGET)
+
+# Install target
+install: $(TARGET)
+	@mkdir -p $(PREFIX)/bin
+	cp $(TARGET) $(PREFIX)/bin/
+
+# Help
+help:
+	@echo "NCCL Example: Pure GIN AlltoAll Device API"
+	@echo "=============================================="
+	@echo ""
+	@echo "This example demonstrates pure GPU-Initiated Networking (GIN)"
+	@echo "for AlltoAll operations without LSA optimizations."
+	@echo ""
+	@echo "Targets:"
+	@echo "  all       - Build the example (default)"
+	@echo "  test      - Build and run test with all GPUs"
+	@echo "  clean     - Remove build artifacts"
+	@echo "  install   - Install to PREFIX/bin (default: /usr/local/bin)"
+	@echo "  help      - Show this help"
+
+.PHONY: all test clean install help