|
| 1 | +cmake_minimum_required(VERSION 3.25) # ipp6 is using 3.28 |
| 2 | + |
| 3 | +# Version information |
| 4 | +# Read makefiles/version.mk file |
| 5 | +file(READ ${CMAKE_SOURCE_DIR}/makefiles/version.mk VERSION_CONTENT) |
| 6 | +string(REGEX REPLACE ".*NCCL_MAJOR[ ]*:=[ ]*([0-9]+).*" "\\1" NCCL_MAJOR "${VERSION_CONTENT}") |
| 7 | +string(REGEX REPLACE ".*NCCL_MINOR[ ]*:=[ ]*([0-9]+).*" "\\1" NCCL_MINOR "${VERSION_CONTENT}") |
| 8 | +string(REGEX REPLACE ".*NCCL_PATCH[ ]*:=[ ]*([0-9]+).*" "\\1" NCCL_PATCH "${VERSION_CONTENT}") |
| 9 | +string(REGEX REPLACE ".*NCCL_SUFFIX[ ]*:=[ ]*([a-zA-Z0-9]*).*" "\\1" NCCL_SUFFIX "${VERSION_CONTENT}") |
| 10 | +string(REGEX REPLACE ".*PKG_REVISION[ ]*:=[ ]*([0-9]+).*" "\\1" PKG_REVISION "${VERSION_CONTENT}") |
| 11 | +math(EXPR NCCL_VERSION_CODE "(${NCCL_MAJOR} * 10000) + (${NCCL_MINOR} * 100) + ${NCCL_PATCH}") |
| 12 | + |
| 13 | +# Make version information available to C++ source files |
| 14 | +add_compile_definitions( |
| 15 | + NCCL_USE_CMAKE |
| 16 | + NCCL_MAJOR=${NCCL_MAJOR} |
| 17 | + NCCL_MINOR=${NCCL_MINOR} |
| 18 | + NCCL_PATCH=${NCCL_PATCH} |
| 19 | + NCCL_VERSION_CODE=${NCCL_VERSION_CODE} |
| 20 | +) |
| 21 | + |
| 22 | +set(ENV{NCCL_USE_CMAKE} "1") |
| 23 | + |
| 24 | +project(NCCL VERSION ${NCCL_MAJOR}.${NCCL_MINOR}.${NCCL_PATCH} |
| 25 | + LANGUAGES CUDA CXX C) |
| 26 | + |
| 27 | +# Make CMAKE_BUILD_TYPE to release by default if not set |
| 28 | +if(NOT CMAKE_BUILD_TYPE) |
| 29 | + set(CMAKE_BUILD_TYPE "Release") |
| 30 | +endif() |
| 31 | + |
| 32 | +option(VERBOSE "Enable verbose output" OFF) |
| 33 | +option(KEEP "Keep intermediate files" OFF) |
| 34 | +option(DEBUG "Enable debug build" OFF) |
| 35 | +option(ASAN "Enable Address Sanitizer" OFF) |
| 36 | +option(UBSAN "Enable Undefined Behavior Sanitizer" OFF) |
| 37 | +option(TRACE "Enable tracing" OFF) |
| 38 | +option(WERROR "Treat warnings as errors" OFF) |
| 39 | +option(PROFAPI "Enable profiling API" ON) |
| 40 | +option(NVTX "Enable NVTX" ON) |
| 41 | +option(RDMA_CORE "Enable RDMA core" OFF) |
| 42 | +option(NET_PROFILER "Enable network profiler" OFF) |
| 43 | +option(MLX5DV "Enable MLX5DV" OFF) |
| 44 | +option(MAX_EXT_NET_PLUGINS "Maximum external network plugins" 0) |
| 45 | + |
| 46 | +find_package(CUDAToolkit REQUIRED) |
| 47 | + |
| 48 | +# CUDA version detection |
| 49 | +string(REGEX MATCH "([0-9]+\\.[0-9]+)" CUDA_VERSION "${CUDAToolkit_VERSION}") |
| 50 | + |
| 51 | +# Extract major and minor version numbers |
| 52 | +string(REGEX MATCH "([0-9]+)" CUDA_MAJOR "${CUDA_VERSION}") |
| 53 | +string(REGEX MATCH "([0-9]+)$" CUDA_MINOR "${CUDA_VERSION}") |
| 54 | +string(REGEX REPLACE ".*\\.([0-9]+)$" "\\1" CUDA_MINOR "${CUDA_VERSION}") |
| 55 | + |
| 56 | +# Add CUDA version definitions after find_package |
| 57 | +add_compile_definitions( |
| 58 | + CUDA_MAJOR=${CUDA_MAJOR} |
| 59 | + CUDA_MINOR=${CUDA_MINOR} |
| 60 | +) |
| 61 | + |
| 62 | +# CUDA architecture flags |
| 63 | +if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES OR CMAKE_CUDA_ARCHITECTURES STREQUAL "") |
| 64 | + message(STATUS "CMAKE_CUDA_ARCHITECTURES not defined or empty, setting default values based on CUDA version") |
| 65 | + |
| 66 | + if(${CUDA_MAJOR} LESS 9) |
| 67 | + set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61") |
| 68 | + elseif(${CUDA_MAJOR} EQUAL 9) |
| 69 | + set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61;70") |
| 70 | + elseif(${CUDA_MAJOR} EQUAL 10) |
| 71 | + set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61;70") |
| 72 | + elseif(${CUDA_MAJOR} EQUAL 11) |
| 73 | + if(${CUDA_MINOR} LESS 8) |
| 74 | + set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61;70;80") |
| 75 | + else() |
| 76 | + set(CMAKE_CUDA_ARCHITECTURES "35;50;60;61;70;80;90") |
| 77 | + endif() |
| 78 | + elseif(${CUDA_MAJOR} EQUAL 12) |
| 79 | + if(${CUDA_MINOR} LESS 8) |
| 80 | + set(CMAKE_CUDA_ARCHITECTURES "50;60;61;70;80;90") |
| 81 | + else() |
| 82 | + set(CMAKE_CUDA_ARCHITECTURES "50;60;61;70;80;90;100;120") |
| 83 | + endif() |
| 84 | + elseif(${CUDA_MAJOR} EQUAL 13) |
| 85 | + set(CMAKE_CUDA_ARCHITECTURES "50;60;61;70;80;90;100;110;120") |
| 86 | + else() |
| 87 | + # For future CUDA versions, include all architectures up to the latest known |
| 88 | + set(CMAKE_CUDA_ARCHITECTURES "50;60;61;70;80;90;100;110;120") |
| 89 | + endif() |
| 90 | +endif() |
| 91 | +message(STATUS "Using CUDA_ARCHITECTURES: ${CMAKE_CUDA_ARCHITECTURES}") |
| 92 | + |
| 93 | +set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fPIC -fvisibility=hidden -Wall -Wno-unused-function -Wno-sign-compare -Wvla -g") |
| 94 | +set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --expt-extended-lambda -Xptxas -maxrregcount=96 -Xfatbin -compress-all -fPIC") |
| 95 | + |
| 96 | +# Sanitizer options |
| 97 | +if(ASAN) |
| 98 | + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address") |
| 99 | + set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fsanitize=address -static-libasan") |
| 100 | +endif() |
| 101 | + |
| 102 | +if(UBSAN) |
| 103 | + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=undefined") |
| 104 | + set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -fsanitize=undefined -static-libubsan") |
| 105 | +endif() |
| 106 | + |
| 107 | +# Additional options |
| 108 | +if(TRACE) |
| 109 | + add_definitions(-DENABLE_TRACE) |
| 110 | +endif() |
| 111 | + |
| 112 | +if(NOT NVTX) |
| 113 | + add_definitions(-DNVTX_DISABLE) |
| 114 | +endif() |
| 115 | + |
| 116 | +if(WERROR) |
| 117 | + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror") |
| 118 | +endif() |
| 119 | + |
| 120 | +if(PROFAPI) |
| 121 | + add_definitions(-DPROFAPI) |
| 122 | +endif() |
| 123 | + |
| 124 | +set(EXTRA_LIBS) |
| 125 | + |
| 126 | +# RDMA and MLX5DV are Linux-specific features |
| 127 | +if(RDMA_CORE) |
| 128 | + add_definitions(-DNCCL_BUILD_RDMA_CORE=1) |
| 129 | + find_library(VERBS_LIBRARY NAMES verbs) |
| 130 | + if(VERBS_LIBRARY) |
| 131 | + list(APPEND EXTRA_LIBS ${VERBS_LIBRARY}) |
| 132 | + endif() |
| 133 | +endif() |
| 134 | + |
| 135 | +if(MLX5DV) |
| 136 | + add_definitions(-DNCCL_BUILD_MLX5DV=1) |
| 137 | + find_library(MLX5_LIBRARY NAMES mlx5) |
| 138 | + if(MLX5_LIBRARY) |
| 139 | + list(APPEND EXTRA_LIBS ${MLX5_LIBRARY}) |
| 140 | + endif() |
| 141 | +endif() |
| 142 | + |
| 143 | +if(NET_PROFILER) |
| 144 | + add_definitions(-DNCCL_ENABLE_NET_PROFILING=1) |
| 145 | +endif() |
| 146 | + |
| 147 | +if(MAX_EXT_NET_PLUGINS GREATER 0) |
| 148 | + add_definitions(-DNCCL_NET_MAX_PLUGINS=${MAX_EXT_NET_PLUGINS}) |
| 149 | +endif() |
| 150 | + |
| 151 | +# Library dependencies |
| 152 | +find_library(RT_LIBRARY NAMES rt) |
| 153 | +if(RT_LIBRARY) |
| 154 | + list(APPEND EXTRA_LIBS ${RT_LIBRARY}) |
| 155 | +endif() |
| 156 | + |
| 157 | +# Debug/Release specific flags |
| 158 | +set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS} -O0") |
| 159 | +set(CMAKE_CUDA_FLAGS_DEBUG "${CMAKE_CUDA_FLAGS} -O0 -G -g") |
| 160 | +set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS} -O3") |
| 161 | +set(CMAKE_CUDA_FLAGS_RELEASE "${CMAKE_CUDA_FLAGS} -O3") |
| 162 | + |
| 163 | +add_subdirectory(ext-net) |
| 164 | +add_subdirectory(ext-profiler/example) |
| 165 | +add_subdirectory(ext-tuner/example) |
| 166 | +add_subdirectory(src) |
0 commit comments