# Copyright (c) 2019-2023 Advanced Micro Devices, Inc. All rights reserved. # Modifications Copyright (c) Microsoft Corporation. Licensed under the MIT License. # CMake version minimum requirements #================================================================================================== cmake_minimum_required(VERSION 3.16) # CMake Toolchain file to define compilers and path to ROCm #================================================================================================== if (NOT CMAKE_TOOLCHAIN_FILE) set(CMAKE_TOOLCHAIN_FILE "${CMAKE_CURRENT_SOURCE_DIR}/toolchain-linux.cmake") message(STATUS "CMAKE_TOOLCHAIN_FILE: ${CMAKE_TOOLCHAIN_FILE}") endif() # RCCL project #================================================================================================== project(rccl CXX) # Build options #================================================================================================== option(BUILD_ADDRESS_SANITIZER "Enable address sanitizer" OFF) option(BUILD_BFD "Enable custom backtrace (if bfd.h exists)" OFF) option(BUILD_FILE_REORG_BACKWARD_COMPATIBILITY "File/folder reorg with backward compatibility" OFF) option(BUILD_LOCAL_GPU_TARGET_ONLY "Build only for GPUs detected on this machine" OFF) option(BUILD_SHARED_LIBS "Build as shared library" ON) option(BUILD_TESTS "Build unit test programs" OFF) option(COLLTRACE "Collective Trace Option" ON) option(ENABLE_MSCCL_KERNEL "Enable MSCCL while compiling" ON) option(ENABLE_MSCCLPP "Enable MSCCL++" ON) option(ENABLE_IFC "Enable indirect function call" OFF) option(INSTALL_DEPENDENCIES "Force install dependencies" OFF) option(ROCTX "Enable ROCTX" ON) option(PROFILE "Enable profiling" OFF) option(TIMETRACE "Enable time-trace during compilation" OFF) option(TRACE "Enable additional tracing" OFF) # Default GPU architectures to build #================================================================================================== set(DEFAULT_GPUS gfx906 gfx908 gfx90a gfx942 gfx1030 gfx1100 gfx1101 gfx1102 gfx1200 gfx1201) # Load CMake modules #================================================================================================== include(CheckIncludeFiles) include(CheckSymbolExists) include(cmake/Dependencies.cmake) # GTest, rocm-cmake, rocm_local_targets include(cmake/CheckSymbolExistsNoWarn.cmake) list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake") # Build only for local GPU architecture if (BUILD_LOCAL_GPU_TARGET_ONLY) message(STATUS "Building only for local GPU target") if (COMMAND rocm_local_targets) rocm_local_targets(DEFAULT_GPUS) else() message(WARNING "Unable to determine local GPU targets. Falling back to default GPUs.") endif() endif() # Determine which GPU architectures to build for set(GPU_TARGETS "${DEFAULT_GPUS}" CACHE STRING "Target default GPUs if GPU_TARGETS is not defined.") # Modify GPU architectures for Address Sanitizer builds by appending "xnack+" if (BUILD_ADDRESS_SANITIZER) SET(amdgpu_targets "") foreach(amdgpu_target IN LISTS GPU_TARGETS) if(NOT amdgpu_target STREQUAL "") string(FIND "${amdgpu_target}" ":xnack+" HAS_XNACK_SUFFIX) if(HAS_XNACK_SUFFIX EQUAL -1) list(APPEND amdgpu_targets "${amdgpu_target}:xnack+") else() list(APPEND amdgpu_targets "${amdgpu_target}") endif() endif() endforeach() SET(GPU_TARGETS "${amdgpu_targets}") endif() # Check if clang compiler can offload to GPU_TARGETS if (COMMAND rocm_check_target_ids) message(STATUS "Checking for ROCm support for GPU targets: " "${GPU_TARGETS}") rocm_check_target_ids(SUPPORTED_GPUS TARGETS ${GPU_TARGETS}) else() message(WARNING "Unable to check for supported GPU targets. Falling back to default GPUs.") set(SUPPORTED_GPUS ${DEFAULT_GPUS}) endif() set(GPU_TARGETS "${SUPPORTED_GPUS}") message(STATUS "Compiling for ${GPU_TARGETS}") ## NOTE: Reload rocm-cmake in order to update GPU_TARGETS include(cmake/Dependencies.cmake) # Reloading to use desired GPU_TARGETS instead of defaults # Try to establish ROCM_PATH (for find_package) #================================================================================================== if(NOT DEFINED ROCM_PATH) # Guess default location set(ROCM_PATH "/opt/rocm") message(WARNING "Unable to find ROCM_PATH: Falling back to ${ROCM_PATH}") else() message(STATUS "ROCM_PATH found: ${ROCM_PATH}") endif() set(ENV{ROCM_PATH} ${ROCM_PATH}) if("${CMAKE_CXX_COMPILER}" MATCHES ".*amdclang\\+\\+") message(STATUS "Compiling with amdclang++") set(COMPILER_EXE_NAME amdclang++) set(COMPILER_GREP_STRING "AMD clang version") set(COMPILER_AWK_CMD "awk -F\" \" '{ printf $4}'") elseif("${CMAKE_CXX_COMPILER}" MATCHES ".*clang\\+\\+") message(STATUS "Compiling with clang++") set(COMPILER_EXE_NAME clang++) set(COMPILER_GREP_STRING "AMD clang version") set(COMPILER_AWK_CMD "awk -F\" \" '{ printf $4}'") elseif("${CMAKE_CXX_COMPILER}" MATCHES ".*hipcc$") message(STATUS "Compiling with hipcc") set(COMPILER_EXE_NAME hipcc) set(COMPILER_GREP_STRING "HIP version") set(COMPILER_AWK_CMD "awk -F\" \" '{ printf $3}' | awk -F\"-\" '{ printf $1}'") else() message(FATAL_ERROR "RCCL can be built only with hipcc or amdclang++") endif() # Set CMAKE flags #================================================================================================== set(CMAKE_INSTALL_PREFIX "${ROCM_PATH}" CACHE PATH "") set(CMAKE_CXX_STANDARD 14) # We use C++14 features, this will add compile option: -std=c++14 set(CMAKE_CXX_EXTENSIONS OFF) # Without this line, it will add -std=gnu++14 instead, which has some issues. if(ROCM_PATH) list(APPEND CMAKE_PREFIX_PATH # Add ROCM_PATH to CMake search paths (for finding HIP / HSA ${ROCM_PATH} ${ROCM_PATH}/hip ${ROCM_PATH}/llvm) endif() # Check for required dependencies #================================================================================================== ## Check for Threads set(THREADS_PREFER_PTHREAD_FLAG ON) find_package(Threads REQUIRED) ## Check for HIP find_package(hip REQUIRED) message(STATUS "HIP compiler: ${HIP_COMPILER}") message(STATUS "HIP runtime: ${HIP_RUNTIME}") if(NOT "${HIP_COMPILER}" MATCHES "clang") message(FATAL_ERROR "RCCL requires clang-based compiler (amdclang++ or hipcc)") endif() ## Check for compiler version find_program(compiler_executable ${COMPILER_EXE_NAME}) message(STATUS "${COMPILER_EXE_NAME} executable: ${compiler_executable}") execute_process( COMMAND bash "-c" "${compiler_executable} --version | grep \"${COMPILER_GREP_STRING}\" | ${COMPILER_AWK_CMD}" OUTPUT_VARIABLE compiler_version_string) message(STATUS "${COMPILER_EXE_NAME} version: ${compiler_version_string}") ## Check for HIP version find_program(hipconfig_executable hipconfig) message(STATUS "hipconfig executable: ${hipconfig_executable}") execute_process( COMMAND bash "-c" "${hipconfig_executable} -v | awk -F\"-\" '{ printf $1 }'" OUTPUT_VARIABLE hip_version_string) message(STATUS "${COMPILER_EXE_NAME} HIP version: ${hip_version_string}") ## Check for ROCm version set(EXPLICIT_ROCM_VERSION "" CACHE STRING "Explicit ROCM version to compile to (auto detect if empty)") if(EXPLICIT_ROCM_VERSION) set(rocm_version_string "${EXPLICIT_ROCM_VERSION}") elseif(ROCM_PATH) message(STATUS "Reading ROCM version from ${ROCM_PATH}/.info/version") file(READ "${ROCM_PATH}/.info/version" rocm_version_string) else() message(FATAL_ERROR "Could not determine ROCM version (set EXPLICIT_ROCM_VERSION or set ROCM_PATH to a valid installation)") endif() string(REGEX MATCH "([0-9]+)\\.([0-9]+)\\.([0-9]+)" rocm_version_matches ${rocm_version_string}) if (rocm_version_matches) set(ROCM_MAJOR_VERSION ${CMAKE_MATCH_1}) set(ROCM_MINOR_VERSION ${CMAKE_MATCH_2}) set(ROCM_PATCH_VERSION ${CMAKE_MATCH_3}) message(STATUS "ROCm version: ${ROCM_MAJOR_VERSION}.${ROCM_MINOR_VERSION}.${ROCM_PATCH_VERSION}") # Convert the version components to int for comparison math(EXPR ROCM_VERSION "(10000 * ${ROCM_MAJOR_VERSION}) + (100 * ${ROCM_MINOR_VERSION}) + ${ROCM_PATCH_VERSION}") add_definitions("-DROCM_VERSION=${ROCM_VERSION}") else() message(WARNING "Failed to extract ROCm version.") endif() ### Required for checking HIP device symbols when building with amdclang++ set(CMAKE_REQUIRED_LIBRARIES hip::device) ### Check for hipDeviceMallocUncached support check_symbol_exists("hipDeviceMallocUncached" "hip/hip_runtime_api.h" HIP_UNCACHED_MEMORY) ### Check for hipDeviceMallocContiguous support check_symbol_exists("hipDeviceMallocContiguous" "hip/hip_runtime_api.h" HIP_CONTIGUOUS_MEMORY) unset(CMAKE_REQUIRED_LIBRARIES) ### Check for indirect function call support if(ENABLE_IFC) if("${hip_version_string}" VERSION_GREATER_EQUAL "5.5.30201") set(IFC_ENABLED ON) message(STATUS "Indirect function call enabled") else() set(IFC_ENABLED OFF) message(WARNING "Indirect function call disabled - requires HIP version >= 5.5.30201") endif() else() set(IFC_ENABLED OFF) endif() ## Check for LL128 support if("${hip_version_string}" VERSION_GREATER_EQUAL "6.1.33591") set(LL128_ENABLED ON) message(STATUS "RCCL LL128 protocol enabled") else() message(STATUS "RCCL LL128 protocol disabled - requires HIP version >= 6.1.33591") endif() ## Check for hsa-runtime64 find_package(hsa-runtime64 REQUIRED) get_target_property(HSA_INCLUDE_PATH hsa-runtime64::hsa-runtime64 INTERFACE_INCLUDE_DIRECTORIES) message(STATUS "HSA runtime: ${HSA_INCLUDE_PATH}") ## Check for ROCM-smi find_package(rocm_smi PATHS ${ROCM_PATH}/lib/cmake/rocm_smi) if (rocm_smi_FOUND) message(STATUS "Found rocm_smi at ${ROCM_SMI_INCLUDE_DIR}") else() message(STATUS "Checking old include directory structure for rocm_smi") set(ROCM_SMI_INCLUDE_DIR "${ROCM_PATH}/rocm_smi/include") set(ROCM_SMI_LIB_DIR "${ROCM_PATH}/rocm_smi/lib") set(ROCM_SMI_LIBRARIES rocm_smi64) endif() check_include_file_cxx("${ROCM_SMI_INCLUDE_DIR}/rocm_smi/rocm_smi64Config.h" HAVE_ROCM_SMI64CONFIG) ### Check for RSMI_INIT_FLAG_THRAD_ONLY_MUTEX support file(READ "${ROCM_SMI_INCLUDE_DIR}/rocm_smi/rocm_smi.h" rocm_smi_incl) string(FIND "${rocm_smi_incl}" "RSMI_INIT_FLAG_THRAD_ONLY_MUTEX" matchres) if(${matchres} EQUAL -1) message(STATUS "RSMI_INIT_FLAG_THRAD_ONLY_MUTEX not supported") else() message(STATUS "RSMI_INIT_FLAG_THRAD_ONLY_MUTEX supported") set(HAVE_ROCM_SMI_THREAD_ONLY_MUTEX True) endif () ## Check for BFD library if custom backtrace is requested if(BUILD_BFD) enable_language(C) check_include_files(bfd.h HAVE_BFD) if (HAVE_BFD) message(STATUS "-- Found BFD support") ### Required for checking HIP device symbols when building with amdclang++ set(CMAKE_REQUIRED_LIBRARIES hip::device) # Check for specific BFD feature support CHECK_SYMBOL_EXISTS(bfd_get_section_flags "bfd.h" HAVE_DECL_BFD_GET_SECTION_FLAGS) CHECK_SYMBOL_EXISTS(bfd_get_section_vma "bfd.h" HAVE_DECL_BFD_GET_SECTION_VMA) CHECK_CXX_SOURCE_COMPILES( "#include int main (int argc, char **argv){ bfd_size_type size; bfd abfd; asection sec; size = bfd_section_size(&abfd, &sec); return (int)(size); }" HAVE_TWO_ARG_BFD_SECTION_SIZE) unset(CMAKE_REQUIRED_LIBRARIES) # Check for iberty support find_library(HAVE_IBERTY iberty PATHS /usr/lib64 /usr/lib/ PATH_SUFFIXES x86_64-linux-gnu) if(HAVE_IBERTY) message(STATUS "iberty found @ ${HAVE_IBERTY}") endif() # Check for demangle support find_path(DEMANGLE_DIR demangle.h PATHS /usr/include PATH_SUFFIXES libiberty) if(NOT DEMANGLE_DIR) message(WARNING "Could not find demangle.h ${DEMANGLE_DIR}") else() message(STATUS "Found demangle.h in ${DEMANGLE_DIR}") endif() else() message(WARNING "bfd.h header not found - Disabling custom backtrace") endif() endif() # Check for --amdgpu-kernarg-preload-count check_cxx_compiler_flag("-mllvm --amdgpu-kernarg-preload-count=16" HAVE_KERNARG_PRELOAD) if (HAVE_KERNARG_PRELOAD) message(STATUS "Kernarg preloading to SGPR enabled") endif() check_cxx_compiler_flag("-parallel-jobs=12" HAVE_PARALLEL_JOBS) if (HAVE_PARALLEL_JOBS) message(STATUS "Parallel jobs enabled") endif() ## Disable building MSCCL++ if the build environment is invalid ## Currently MSCCL++ is supported only on gfx942, and only on Ubuntu and CentOS if (ENABLE_MSCCLPP AND NOT ("gfx942" IN_LIST GPU_TARGETS OR "gfx942:xnack-" IN_LIST GPU_TARGETS OR "gfx942:xnack+" IN_LIST GPU_TARGETS)) set(ENABLE_MSCCLPP OFF) message(WARNING "Can only build MSCCL++ for gfx942; disabling MSCCL++ build") endif() if (ENABLE_MSCCLPP AND ROCM_VERSION VERSION_LESS "60200") set(ENABLE_MSCCLPP OFF) message(WARNING "MSCCL++ integration only supported on ROCm 6.2 or greater; disabling MSCCL++ build") endif() # cmake_host_system_information(RESULT HOST_OS_ID QUERY DISTRIB_ID) ## Requires cmake 3.22 execute_process( COMMAND bash -c "grep '^ID=' /etc/os-release | cut -d'=' -f2" OUTPUT_VARIABLE HOST_OS_ID OUTPUT_STRIP_TRAILING_WHITESPACE ) if (ENABLE_MSCCLPP AND NOT(${HOST_OS_ID} STREQUAL "ubuntu" OR ${HOST_OS_ID} STREQUAL "centos")) set(ENABLE_MSCCLPP OFF) message(WARNING "MSCCL++ integration not supported on this OS (${HOST_OS_ID}); disabling MSCCL++ build") endif() # Check for ROCTX if(ROCTX) find_library(ROCTX_LIB NAMES roctx64) if(ROCTX_LIB) set(ROCTX_ENABLE ON) message(STATUS "ROCTX library found: ${ROCTX_LIB}") else() message(WARNING "ROCTX library not found. Skipping ROCTX linking.") endif() endif() # Determine version from makefiles/version.mk and fill in templates #================================================================================================== ## parse version from Makefile NCCL_MAJOR, NCCL_MINOR, NCCL_PATCH must exist ## NCCL_SUFFIX is optional ## NCCL_VERSION formatting is ((X) * 1000 + (Y) * 100 + (Z)) so we must first detect one or two digits first file(READ makefiles/version.mk version_mk_text) if("${version_mk_text}" MATCHES "NCCL_MAJOR *:= *([0-9]*)") set(NCCL_MAJOR ${CMAKE_MATCH_1}) else() message(FATAL_ERROR "Failed to parse NCCL_MAJOR") endif() if("${version_mk_text}" MATCHES "NCCL_MINOR *:= *([0-9]*)") set(NCCL_MINOR ${CMAKE_MATCH_1}) else() message(FATAL_ERROR "Failed to parse NCCL_MINOR") endif() if("${version_mk_text}" MATCHES "NCCL_PATCH *:= *([0-9]*)") set(NCCL_PATCH ${CMAKE_MATCH_1}) else() message(FATAL_ERROR "Failed to parse NCCL_PATCH") endif() if("${version_mk_text}" MATCHES "NCCL_SUFFIX *:= *([0-9]*)") set(NCCL_SUFFIX ${CMAKE_MATCH_1}) else() set(NCCL_SUFFIX) endif() if("${version_mk_text}" MATCHES "PKG_REVISION *:= *([0-9]*)") set(PKG_REVISION ${CMAKE_MATCH_1}) else() message(FATAL_ERROR "Failed to parse PKG_REVISION") endif() if("${NCCL_PATCH}" MATCHES "[0-9][0-9]") set(NCCL_VERSION "${NCCL_MAJOR}${NCCL_MINOR}${NCCL_PATCH}") else() set(NCCL_VERSION "${NCCL_MAJOR}${NCCL_MINOR}0${NCCL_PATCH}") endif() ## Setup VERSION set(VERSION_STRING "${NCCL_MAJOR}.${NCCL_MINOR}.${NCCL_PATCH}") rocm_setup_version(VERSION ${VERSION_STRING}) ## Fill in version information for main header file configure_file(src/nccl.h.in ${PROJECT_BINARY_DIR}/include/rccl/rccl.h) # For external linking configure_file(src/nccl.h.in ${PROJECT_BINARY_DIR}/include/nccl.h) # Used by some internal files # Collect list of all source files #================================================================================================== # E.g: find src -type f \( -name "*.cc" -o -name "*.h" -o -name "*.hpp" \) | sort set(SRC_FILES src/bootstrap.cc src/channel.cc src/collectives.cc src/debug.cc src/enqueue.cc src/group.cc src/init.cc src/init_nvtx.cc src/net.cc src/msccl.cc src/proxy.cc src/register.cc src/transport.cc # src/clique/AllReduceCliqueKernel.h # src/clique/CliqueCommon.h # src/clique/CliqueManager.cc # src/clique/CliqueManager.h # src/clique/CliqueShmNames.h # src/clique/HandleCache.cc # src/clique/HandleCache.h # src/clique/HandleShm.cc # src/clique/HandleShm.h # src/clique/Hash.cc # src/clique/Hash.h # src/clique/MsgQueue.cc # src/clique/MsgQueue.h # src/clique/SharedMemHelper.h # src/clique/ShmObject.cc # src/clique/ShmObject.h src/device/all_gather.h src/device/all_reduce.h src/device/alltoall_pivot.h src/device/broadcast.h src/device/common.h src/device/common_kernel.h src/device/op128.h src/device/primitives.h src/device/prims_ll128.h src/device/prims_ll.h src/device/prims_simple.h src/device/reduce.h src/device/reduce_kernel.h src/device/reduce_scatter.h src/device/sendrecv.h src/device/common.cu src/device/onerank.cu src/device/network/unpack/unpack_defs.h src/device/network/unpack/unpack.h src/graph/connect.cc src/graph/paths.cc src/graph/rings.cc src/graph/rings.h src/graph/rome_models.cc src/graph/rome_models.h src/graph/search.cc src/graph/topo.cc src/graph/topo.h src/graph/trees.cc src/graph/tuning.cc src/graph/xml.cc src/graph/xml.h src/include/alloc.h src/include/alt_rsmi.h src/include/archinfo.h src/include/api_trace.h src/include/argcheck.h src/include/BfdBacktrace.hpp src/include/bitops.h src/include/bootstrap.h src/include/channel.h src/include/checks.h src/include/collectives.h src/include/coll_net.h src/include/comm.h src/include/core.h src/include/cpuset.h # src/include/cudawrap.h src/include/debug.h src/include/device.h src/include/enqueue.h src/include/gdrwrap.h src/include/git_version.h src/include/graph.h src/include/group.h src/include/hip_rocm_version_info.h src/include/ibvcore.h src/include/ibvsymbols.h src/include/ibvwrap.h src/include/info.h src/include/ipcsocket.h src/include/nccl_common.h src/include/nccl_net.h src/include/nccl_tuner.h src/include/net_device.h src/include/net.h src/include/nvmlwrap.h src/include/nvtx.h src/include/nvtx_stub.h src/include/p2p.h src/include/param.h src/include/profiler.h src/include/proxy.h src/include/rccl_vars.h src/include/register.h src/include/rccl_float8.h src/include/rocm_smi_wrap.h src/include/rocmwrap.h src/include/roctx.h src/include/shm.h src/include/signals.h src/include/socket.h src/include/strongstream.h src/include/timer.h src/include/transport.h src/include/trees.h src/include/tuner.h src/include/utils.h src/include/msccl/msccl_lifecycle.h src/include/msccl/msccl_parser.h src/include/msccl/msccl_scheduler.h src/include/msccl/msccl_setup.h src/include/msccl/msccl_status.h src/include/msccl/msccl_struct.h src/include/npkit/npkit.h src/include/npkit/npkit_event.h src/include/npkit/npkit_struct.h src/include/nvtx3/nvToolsExt.h src/include/nvtx3/nvToolsExtCounters.h src/include/nvtx3/nvToolsExtCuda.h src/include/nvtx3/nvToolsExtCudaRt.h src/include/nvtx3/nvToolsExtMem.h src/include/nvtx3/nvToolsExtMemCudaRt.h src/include/nvtx3/nvToolsExtOpenCL.h src/include/nvtx3/nvToolsExtPayload.h src/include/nvtx3/nvToolsExtPayloadHelper.h src/include/nvtx3/nvToolsExtSemanticsCounters.h src/include/nvtx3/nvToolsExtSemanticsScope.h src/include/nvtx3/nvToolsExtSync.h src/include/nvtx3/nvtx3.hpp src/include/nvtx3/nvtxDetail/nvtxExtHelperMacros.h src/include/nvtx3/nvtxDetail/nvtxExtImpl.h src/include/nvtx3/nvtxDetail/nvtxExtImplCounters_v1.h src/include/nvtx3/nvtxDetail/nvtxExtImplMem_v1.h src/include/nvtx3/nvtxDetail/nvtxExtImplMemCudaRt_v1.h src/include/nvtx3/nvtxDetail/nvtxExtImplPayload_v1.h src/include/nvtx3/nvtxDetail/nvtxExtInit.h src/include/nvtx3/nvtxDetail/nvtxExtPayloadHelperInternal.h src/include/nvtx3/nvtxDetail/nvtxExtPayloadTypeInfo.h src/include/nvtx3/nvtxDetail/nvtxExtTypes.h src/include/nvtx3/nvtxDetail/nvtxImpl.h src/include/nvtx3/nvtxDetail/nvtxImplCore.h src/include/nvtx3/nvtxDetail/nvtxImplCuda_v3.h src/include/nvtx3/nvtxDetail/nvtxImplCudaRt_v3.h src/include/nvtx3/nvtxDetail/nvtxImplOpenCL_v3.h src/include/nvtx3/nvtxDetail/nvtxImplSync_v3.h src/include/nvtx3/nvtxDetail/nvtxInit.h src/include/nvtx3/nvtxDetail/nvtxInitDecls.h src/include/nvtx3/nvtxDetail/nvtxInitDefs.h src/include/nvtx3/nvtxDetail/nvtxLinkOnce.h src/include/nvtx3/nvtxDetail/nvtxTypes.h src/misc/alt_rsmi.cc src/misc/archinfo.cc src/misc/argcheck.cc src/misc/api_trace.c src/misc/api_trace.cc # src/misc/cudawrap.cc # src/misc/gdrwrap.cc src/misc/ibvsymbols.cc src/misc/ibvwrap.cc src/misc/ipcsocket.cc src/misc/npkit.cc # src/misc/nvmlwrap.cc src/misc/nvmlwrap_stub.cc src/misc/param.cc src/misc/profiler.cc src/misc/rocm_smi_wrap.cc src/misc/rocmwrap.cc src/misc/roctx.cc src/misc/shmutils.cc src/misc/signals.cc src/misc/socket.cc src/misc/strongstream.cc src/misc/tuner.cc src/misc/utils.cc src/misc/msccl/msccl_lifecycle.cc src/misc/msccl/msccl_parser.cc src/misc/msccl/msccl_setup.cc src/misc/msccl/msccl_status.cc src/transport/coll_net.cc src/transport/generic.cc src/transport/net.cc src/transport/net_ib.cc src/transport/net_socket.cc src/transport/nvls.cc src/transport/p2p.cc src/transport/shm.cc ) if (ENABLE_MSCCL_KERNEL) set(MSCCL_KERNEL_SOURCES src/device/msccl_kernel_impl.h src/include/msccl/msccl_kernel.h ) list(APPEND SRC_FILES ${MSCCL_KERNEL_SOURCES}) endif() if (ENABLE_MSCCLPP) set(MSCCLPP_SOURCES src/include/mscclpp/mscclpp_nccl.h src/misc/mscclpp/mscclpp_nccl.cc ) list(APPEND SRC_FILES ${MSCCLPP_SOURCES}) endif() # Hipify source files (copy of source generated into hipify directory) #================================================================================================== find_program(hipify-perl_executable hipify-perl) if(NOT hipify-perl_executable) message(FATAL_ERROR "hipify-perl not found") endif() set(HIPIFY_DIR "${CMAKE_CURRENT_BINARY_DIR}/hipify") ## Loop over each source file to hipify foreach(SRC_FILE ${SRC_FILES}) # Check that file exists if (NOT EXISTS ${CMAKE_SOURCE_DIR}/${SRC_FILE}) message(FATAL_ERROR "Unable to find file listed in CMakeLists.txt: ${CMAKE_SOURCE_DIR}/${SRC_FILE}") endif() # Establish hipified copy of the source file set(HIP_FILE "${HIPIFY_DIR}/${SRC_FILE}") get_filename_component(HIP_FILE_DIR ${HIP_FILE} DIRECTORY) # Make sure the file name is unique and there is no duplicate add_file_unique(HIP_SOURCES ${HIP_FILE}) # Convert .cu files to .cpp so that they get processed properly string(REPLACE "\.cu" "\.cu.cpp" HIP_FILE ${HIP_FILE}) list(APPEND HIP_SOURCES ${HIP_FILE}) # Create a custom command to create hipified source code add_custom_command( OUTPUT ${HIP_FILE} COMMAND mkdir -p ${HIP_FILE_DIR} && ${hipify-perl_executable} -quiet-warnings ${CMAKE_SOURCE_DIR}/${SRC_FILE} -o ${HIP_FILE} && ${CMAKE_COMMAND} -E env bash ${CMAKE_CURRENT_SOURCE_DIR}/cmake/scripts/add_unroll.sh ${HIP_FILE} MAIN_DEPENDENCY ${SRC_FILE} COMMENT "Hipifying ${SRC_FILE} -> ${HIP_FILE}" ) endforeach() # Generate device/host tables and all the collective functions that are going to be in librccl.so #================================================================================================== find_package(Python3 COMPONENTS Interpreter REQUIRED) if (NOT Python3_FOUND) message(FATAL_ERROR "RCCL requires Python3 for generating host/device tables") endif() set(GEN_DIR "${HIPIFY_DIR}/gensrc") # Execute the python script to generate required files execute_process( COMMAND ${Python3_EXECUTABLE} ${CMAKE_SOURCE_DIR}/src/device/generate.py ${GEN_DIR} ${IFC_ENABLED} ${COLLTRACE} ${ENABLE_MSCCL_KERNEL} ${BUILD_LOCAL_GPU_TARGET_ONLY} ${ONLY_FUNCS} WORKING_DIRECTORY ${CMAKE_SOURCE_DIR} RESULT_VARIABLE gen_py_result ERROR_VARIABLE gen_py_error ) if (gen_py_result) message(SEND_ERROR "Error: ${gen_py_error}") message(FATAL_ERROR "${CMAKE_SOURCE_DIR}/src/device/generate.py failed") endif() # Find the generated files in the output directory file(GLOB GENERATED_FILES "${GEN_DIR}/*") # Append all found generated files to the list foreach(file ${GENERATED_FILES}) list(APPEND HIP_SOURCES ${file}) endforeach() # Create an initial git_version.cpp file (that will be updated with latest git version) #================================================================================================== file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/git_version.cpp "") list(APPEND HIP_SOURCES ${CMAKE_CURRENT_BINARY_DIR}/git_version.cpp) # Create a custom target that updates git_version.cpp and executes whenever rccl is built add_custom_target(git_version_check COMMENT "Updating git_version.cpp if necessary" COMMAND ${CMAKE_COMMAND} -P ${CMAKE_CURRENT_SOURCE_DIR}/cmake/scripts/git_version.cmake VERBATIM ) # Set up RCCL library #================================================================================================== ## Set RCCL source files add_library(rccl ${HIP_SOURCES}) ## Set RCCL dependencies add_dependencies(rccl git_version_check) # Execute git_version_check during build ## Set RCCL include directories target_include_directories(rccl PRIVATE ${PROJECT_BINARY_DIR}/include) # for generated rccl.h header target_include_directories(rccl PRIVATE ${HIPIFY_DIR}/src) # for hipfied headers target_include_directories(rccl PRIVATE ${HIPIFY_DIR}/src/include) target_include_directories(rccl PRIVATE ${HIPIFY_DIR}/src/device) target_include_directories(rccl PRIVATE ${HIPIFY_DIR}/src/device/network/unpack) target_include_directories(rccl PRIVATE ${HIPIFY_DIR}/gensrc) target_include_directories(rccl PRIVATE ${HSA_INCLUDE_PATH}) target_include_directories(rccl PRIVATE ${ROCM_SMI_INCLUDE_DIR}) if(DEMANGLE_DIR) target_include_directories(rccl PRIVATE ${DEMANGLE_DIR}) endif() ## Set RCCL compile definitions if(COLLTRACE) target_compile_definitions(rccl PRIVATE ENABLE_COLLTRACE) endif() if(ENABLE_MSCCL_KERNEL) target_compile_definitions(rccl PRIVATE COMPILE_MSCCL_KERNEL) endif() if(ENABLE_MSCCLPP) target_compile_definitions(rccl PRIVATE ENABLE_MSCCLPP) endif() if(HAVE_ROCM_SMI64CONFIG) target_compile_definitions(rccl PRIVATE USE_ROCM_SMI64CONFIG) endif() if(HAVE_ROCM_SMI_THREAD_ONLY_MUTEX) target_compile_definitions(rccl PRIVATE USE_ROCM_SMI_THREAD_ONLY_MUTEX) endif() if(NPKIT_FLAGS) target_compile_definitions(rccl PRIVATE ${NPKIT_FLAGS}) endif() if(PROFILE) target_compile_definitions(rccl PRIVATE ENABLE_PROFILING) endif() if(ROCTX_ENABLE) target_compile_definitions(rccl PRIVATE ROCTX_ENABLE) else() target_compile_definitions(rccl PRIVATE NVTX_NO_IMPL) target_compile_definitions(rccl PRIVATE NVTX_DISABLE) endif() if(TRACE) target_compile_definitions(rccl PRIVATE ENABLE_TRACE) endif() if(${HIP_CONTIGUOUS_MEMORY}) target_compile_definitions(rccl PRIVATE HIP_CONTIGUOUS_MEMORY) message(STATUS "HIP_CONTIGUOUS_MEMORY enabled") else() message(STATUS "HIP_CONTIGUOUS_MEMORY disabled") endif() if("${hip_version_string}" VERSION_GREATER_EQUAL "5.7.31920") target_compile_definitions(rccl PRIVATE HIP_UNCACHED_MEMORY) message(STATUS "HIP_UNCACHED_MEMORY enabled") else() message(STATUS "HIP_UNCACHED_MEMORY disabled - requires HIP version >= 5.7.31920") # keep --hipcc-func-supp on older HIP and compiler if(NOT IFC_ENABLED) target_compile_options(rccl PRIVATE --hipcc-func-supp) message(STATUS "--hipcc-func-supp enabled") else() message(STATUS "--hipcc-func-supp disabled") endif() endif() if (BUILD_BFD) if (HAVE_BFD) target_compile_definitions(rccl PRIVATE HAVE_BFD) endif() if (HAVE_DECL_BFD_GET_SECTION_FLAGS) target_compile_definitions(rccl PRIVATE HAVE_DECL_BFD_GET_SECTION_FLAGS) endif() if (HAVE_DECL_BFD_GET_SECTION_VMA) target_compile_definitions(rccl PRIVATE HAVE_DECL_BFD_GET_SECTION_VMA) endif() if (HAVE_TWO_ARG_BFD_SECTION_SIZE) target_compile_definitions(rccl PRIVATE HAVE_TWO_ARG_BFD_SECTION_SIZE) endif() endif() if (IFC_ENABLED) target_compile_definitions(rccl PRIVATE USE_INDIRECT_FUNCTION_CALL) endif() if(DEMANGLE_DIR) target_compile_definitions(rccl PRIVATE "HAVE_CPLUS_DEMANGLE=1") target_compile_definitions(rccl PRIVATE "HAVE_DECL_BASENAME=1") endif() if(LL128_ENABLED) target_compile_definitions(rccl PRIVATE ENABLE_LL128) endif() ## Set RCCL compile options if (HAVE_PARALLEL_JOBS) target_compile_options(rccl PRIVATE -parallel-jobs=12) endif() target_compile_options(rccl PRIVATE -Werror=uninitialized -Werror=sometimes-uninitialized) target_compile_options(rccl PRIVATE -Wno-format-nonliteral) target_compile_options(rccl PRIVATE -fgpu-rdc) # Generate relocatable device code (required for extern __shared__) target_compile_options(rccl PRIVATE -fvisibility=hidden) # Set symbol visibility to hidden by default if (HAVE_KERNARG_PRELOAD) target_compile_options(rccl PRIVATE -mllvm --amdgpu-kernarg-preload-count=16) endif() ## NOTE: This is currently being handled by rocm-cmake, however may need to be re-enabled in the future #foreach(target ${GPU_TARGETS}) # target_compile_options(rccl PRIVATE --offload-arch=${target}) #endforeach() if(BUILD_ADDRESS_SANITIZER) target_compile_options(rccl PRIVATE -fsanitize=address -shared-libasan) endif() if(TIMETRACE) target_compile_options(rccl PRIVATE -ftime-trace) endif() ## Set RCCL linked library directories target_link_directories(rccl PRIVATE ${ROCM_SMI_LIB_DIR}) if (ROCM_VERSION VERSION_GREATER_EQUAL "60100") option(RCCL_ROCPROFILER_REGISTER "Enable rocprofiler-register support" ON) else() if(RCCL_ROCPROFILER_REGISTER) message(AUTHOR_WARNING "RCCL_ROCPROFILER_REGISTER is not valid option for ROCm < 6.2. Current ROCm version: ${ROCM_VERSION}") endif() set(RCCL_ROCPROFILER_REGISTER OFF CACHE BOOL "" FORCE) endif() if(RCCL_ROCPROFILER_REGISTER) find_package(rocprofiler-register REQUIRED) target_compile_definitions(rccl PRIVATE RCCL_ROCPROFILER_REGISTER=1) target_link_libraries( rccl PRIVATE rocprofiler-register::rocprofiler-register) endif() ## Set RCCL linked libraries if (HAVE_BFD) target_link_libraries(rccl PRIVATE bfd) if(HAVE_IBERTY) target_link_libraries(rccl PRIVATE iberty z) endif() endif() if (ROCTX_ENABLE) target_link_libraries(rccl PRIVATE -lroctx64) endif() target_link_libraries(rccl PRIVATE -fgpu-rdc) # Required when linking relocatable device code target_link_libraries(rccl PRIVATE Threads::Threads) target_link_libraries(rccl INTERFACE hip::host) target_link_libraries(rccl PRIVATE hip::device) target_link_libraries(rccl PRIVATE dl) target_link_libraries(rccl PRIVATE ${ROCM_SMI_LIBRARIES}) if(ENABLE_MSCCLPP) target_link_libraries(rccl PRIVATE mscclpp_nccl) endif() ## Set RCCL link options ## Find out available memory execute_process( COMMAND bash "-c" "cat /sys/fs/cgroup/memory.max" OUTPUT_VARIABLE memory_max_string) if (${memory_max_string} MATCHES "^[0-9]+") math(EXPR memory_in_gb "${memory_max_string} / (1024 * 1024 * 1024)") else() execute_process( COMMAND bash "-c" "free | grep -o '[[:digit:]]*' | head -1" OUTPUT_VARIABLE memory_max_string) ## memory_max_string holds the free memory in KB if (${memory_max_string} MATCHES "^[0-9]+") math(EXPR memory_in_gb "${memory_max_string} / (1024 * 1024)") ## KB to GB conversion else() cmake_host_system_information(RESULT memory_max_string QUERY AVAILABLE_PHYSICAL_MEMORY ) math(EXPR memory_in_gb "${memory_max_string} / 1024") endif() endif() ## Reserve 16GB for each linker job. Limit max number of linker jobs to 16 if (HAVE_PARALLEL_JOBS) math(EXPR num_linker_jobs "(${memory_in_gb} + 15) / 16") if (${num_linker_jobs} GREATER_EQUAL "16") set(num_linker_jobs "16") endif() message(STATUS "Use ${num_linker_jobs} jobs for linking") target_link_options(rccl PRIVATE -parallel-jobs=${num_linker_jobs}) # Use multiple threads to link endif() if(BUILD_ADDRESS_SANITIZER) target_link_options(rccl PRIVATE -fuse-ld=lld) endif() if(TIMETRACE) target_link_options(rccl PRIVATE -ftime-trace) endif() if(NOT BUILD_SHARED_LIBS) message(STATUS "Building static RCCL library") else() message(STATUS "Building shared RCCL library") endif() if (HAVE_KERNARG_PRELOAD) target_link_options(rccl PRIVATE "SHELL:-Xoffload-linker -mllvm=-amdgpu-kernarg-preload-count=16") endif() if(ENABLE_MSCCLPP) include(cmake/MSCCLPP.cmake) endif() ## Track linking time set_property(TARGET rccl PROPERTY RULE_LAUNCH_LINK "${CMAKE_COMMAND} -E time") ## Setup librccl.so version rocm_set_soversion(rccl "1.0") if(NOT BUILD_SHARED_LIBS) # To create a static lib with `-fgpu-rdc`, you need `--emit-static-lib` and `--hip-link`. # You also need to invoke amdclang++ again to trigger GPU code generation. set(static_link_flags ${CXXFLAGS} --hip-link -fgpu-rdc --emit-static-lib ) # Find all the libraries we need to link at link time to include them in the clang link # command line. get_target_property(rccl_libs rccl LINK_LIBRARIES) foreach(target ${rccl_libs}) if(TARGET ${target}) get_target_property(location ${target} LOCATION) if(location) LIST(APPEND static_link_flags -l${location}) endif() endif() endforeach() foreach(target ${GPU_TARGETS}) list(APPEND static_link_flags --offload-arch=${target}) endforeach() list(JOIN static_link_flags " " flags_str) # Invoking amdclang++ this way will produce a static archive, so just override ARCHIVE_CREATE. set(CMAKE_CXX_ARCHIVE_CREATE " ${flags_str} -o ") endif() # Install settings #================================================================================================== ## Specify install targets rocm_install_targets(TARGETS rccl) rocm_install(FILES ${PROJECT_BINARY_DIR}/include/rccl/rccl.h src/include/nccl_net.h DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/rccl) rocm_install(FILES src/include/api_trace.h DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/rccl/amd_detail) file(COPY tools/msccl-algorithms DESTINATION ${PROJECT_BINARY_DIR}) file(COPY tools/msccl-unit-test-algorithms DESTINATION ${PROJECT_BINARY_DIR}) ## Install Algorithm files under share folder rocm_install(DIRECTORY ${PROJECT_BINARY_DIR}/msccl-algorithms DESTINATION ${CMAKE_INSTALL_DATADIR}/rccl) rocm_install(DIRECTORY ${PROJECT_BINARY_DIR}/msccl-unit-test-algorithms DESTINATION ${CMAKE_INSTALL_DATADIR}/rccl) rocm_export_targets( NAMESPACE roc:: TARGETS rccl DEPENDS hip) ## Build with backwards compatibility if requested if(BUILD_FILE_REORG_BACKWARD_COMPATIBILITY) ### Create wrapper files rocm_wrap_header_dir( "${PROJECT_BINARY_DIR}/include/rccl" PATTERNS "rccl.h" GUARDS SYMLINK WRAPPER WRAPPER_LOCATIONS ${CMAKE_INSTALL_INCLUDEDIR} rccl/${CMAKE_INSTALL_INCLUDEDIR}) ### install the wrapper header file to package rocm_install( FILES ${PROJECT_BINARY_DIR}/rccl/include/rccl.h src/include/nccl_net.h DESTINATION "./rccl/${CMAKE_INSTALL_INCLUDEDIR}/" ) rocm_install( FILES ${PROJECT_BINARY_DIR}/include/rccl.h src/include/nccl_net.h DESTINATION "${CMAKE_INSTALL_INCLUDEDIR}/" ) endif() ## Set package dependencies if(BUILD_ADDRESS_SANITIZER) set(DEPENDS_HIP_RUNTIME "hip-runtime-amd-asan" ) else() set(DEPENDS_HIP_RUNTIME "hip-runtime-amd" ) endif() rocm_package_add_dependencies(DEPENDS "${DEPENDS_HIP_RUNTIME} >= 4.5.0" "rocm-smi-lib >= 4.0.0") set(CPACK_DEBIAN_PACKAGE_SHLIBDEPS ON) set(CPACK_RPM_EXCLUDE_FROM_AUTO_FILELIST_ADDITION "/opt" "${ROCM_PATH}") find_file (DEBIAN debian_version debconf.conf PATHS /etc) if(DEBIAN) # Write copyright file file(WRITE "${CMAKE_BINARY_DIR}/copyright" "Format: https://www.debian.org/doc/packaging-manuals/copyright-format/1.0/ Upstream-Name: rccl Source: https://github.com/ROCm/rccl Files: * Copyright: (c) 2016-2020, NVIDIA CORPORATION. All rights reserved. Modifications Copyright (c) 2020-2023 Advanced Micro Devices, Inc. All rights reserved. Modifications Copyright (c) Microsoft Corporation. Licensed under the MIT License. License: See LICENSE.txt for license information\n") rocm_install(FILES "${CMAKE_BINARY_DIR}/copyright" DESTINATION ${CMAKE_INSTALL_DATADIR}/rccl) # Write changelog file find_program( date_executable date ) execute_process(COMMAND ${date_executable} -R OUTPUT_VARIABLE TIMESTAMP) file(WRITE "${CMAKE_BINARY_DIR}/changelog" "rccl (${VERSION_STRING}-1) unstable; urgency=medium * Initial release. -- RCCL Maintainer ${TIMESTAMP}\n") find_program( gzip_executable gzip ) execute_process(COMMAND bash "-c" "${gzip_executable} -9 -c ${CMAKE_BINARY_DIR}/changelog" WORKING_DIRECTORY ${CMAKE_BINARY_DIR} OUTPUT_FILE "${CMAKE_BINARY_DIR}/changelog.Debian.gz") rocm_install(FILES "${CMAKE_BINARY_DIR}/changelog.Debian.gz" DESTINATION ${CMAKE_INSTALL_DATADIR}/rccl) set(CPACK_DEBIAN_PACKAGE_DESCRIPTION "ROCm Communication Collectives Library Optimized primitives for collective multi-GPU communication") endif() if(BUILD_TESTS) rocm_package_setup_component(clients) rocm_package_setup_client_component(tests PACKAGE_NAME unittests) add_subdirectory(test) if(BUILD_SHARED_LIBS) add_custom_command(TARGET rccl POST_BUILD COMMENT "Extracting metadata from librccl.so" COMMAND COMMAND ${CMAKE_COMMAND} -P ${CMAKE_CURRENT_SOURCE_DIR}/cmake/scripts/extract_metadata.cmake VERBATIM ) endif() endif() rocm_create_package( NAME rccl DESCRIPTION "ROCm Communication Collectives Library" MAINTAINER "RCCL Maintainer " LDCONFIG)