Skip to content

N
NiuTrans.Tensor
Commit b4785f07 by liyinqiao
Options

Update the CMake file. 

1. Support to generate the dynamic link library on GPU.
2. Support to compile the project on the MKL and OpenBLAS.
parent 2cc0a82d
正在显示 包含 200 行增加 和 67 行删除
CMakeLists.txt
......@@ -6,39 +6,64 @@
# cmake minimum version
cmake_minimum_required(VERSION 2.8)
# project's name
# Project's name
project(NiuTensor)
# the prefix of the generated executable file
set(NIUTRANS_EXE "NiuTensor")
# The name of the generated executable file
# The name of the dynamic link library
set(NIUTENSOR_EXE "NiuTensor")
set(NIUTENSOR_DLL "${NIUTENSOR_EXE}")
# generated file path
# Generated file path
set(EXECUTABLE_OUTPUT_PATH ../bin)
set(LIBRARY_OUTPUT_PATH ../lib)
# 0 - Linux or macOS
# 1 - Windows
set(ON_WINDOWS 0)
# Use CMAKE_MACOSX_RPATH for macOS
set(CMAKE_MACOSX_RPATH 1)
# Open floder manage
set_property(GLOBAL PROPERTY USE_FOLDERS ON)
set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
option(USE_CUDA "Use CUDA" OFF)
option(USE_MKL "Use MKL" OFF)
option(USE_OPENBLAS "Use OpenBLAS" OFF)
option(GEN_DLL "Generate Dynamic Link Library" OFF)
# 0 - Not use
# 1 - Use, if set 1, please modify the CUDA_VERSION and CUDA_TOOLKIT_ROOT_DIR below
set(USE_CUDA 0)
# If set USE_CUDA ON, please modify CUDA_TOOLKIT_ROOT_DIR below.
# If set USE_MKL ON, please modify the INTEL_ROOT below.
# If set USE_OPENBLAS ON, please modify the OPENBLAS_ROOT below.
if (USE_CUDA)
set(CUDA_VERSION 9.2)
set(CUDA_TOOLKIT_ROOT_DIR "/usr/local/cuda")
if(NOT EXISTS ${CUDA_TOOLKIT_ROOT_DIR})
if(WIN32)
set(CUDA_TOOLKIT_ROOT_DIR "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v9.2")
else()
set(CUDA_TOOLKIT_ROOT_DIR "/usr/local/cuda-9.2")
endif()
endif()
message(STATUS "CUDA_TOOLKIT_ROOT_DIR: ${CUDA_TOOLKIT_ROOT_DIR}")
endif()
if(USE_MKL)
if(NOT DEFINED INTEL_ROOT)
if(WIN32)
set(INTEL_ROOT "C:/Program Files (x86)/IntelSWTools/compilers_and_libraries_2020.2.254/windows")
else()
set(INTEL_ROOT "/opt/intel/compilers_and_libraries_2020.2.254/linux")
endif()
endif()
message(STATUS "INTEL_ROOT: ${INTEL_ROOT}")
endif()
if(USE_OPENBLAS)
if(NOT DEFINED OPENBLAS_ROOT)
if(WIN32)
set(OPENBLAS_ROOT "C:/Program Files/OpenBLAS")
else()
set(OPENBLAS_ROOT "/opt/OpenBLAS")
endif()
endif()
message(STATUS "OPENBLAS_ROOT: ${OPENBLAS_ROOT}")
endif()
# 0 - Not use half precision in CUDA codes
# 1 - Use, note the architecture of GPU must be newer than Pascal(including Pascal)
set(USE_HALF_PRECISION 0)
# use CMAKE_MACOSX_RPATH for macOS
set(CMAKE_MACOSX_RPATH 1)
# open floder manage
set_property(GLOBAL PROPERTY USE_FOLDERS On)
add_definitions(-std=c++11)
# find all the .cpp .h .cu .chu files in source folder
# Find all the .cpp .h .cu .chu files in source folder
file(GLOB_RECURSE CPP_FILES source/*.cpp)
file(GLOB_RECURSE H_FILES source/*.h)
file(GLOB_RECURSE CU_FILES source/*.cu)
......@@ -68,15 +93,14 @@ function(my_add_executable)
endif()
endfunction(my_add_executable)
# Set libs and compiler options for CUDA
if(USE_CUDA)
set(NIUTRANS_EXE "${NIUTRANS_EXE}.GPU")
add_definitions(-DUSE_CUDA)
if(USE_HALF_PRECISION)
add_definitions(-DHALF_PRECISION)
endif()
if(ON_WINDOWS)
find_package(CUDA ${CUDA_VERSION} REQUIRED)
add_compile_options(-Wno-dev)
find_package(CUDA REQUIRED)
if(WIN32)
SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /wd4819")
set(CMAKE_CUDA_FLAGS ${CMAKE_CUDA_FLAGS} "-maxrregcount=0 -m64 --disable-warnings -use_fast_math -DUSE_CUDA")
if(USE_HALF_PRECISION)
......@@ -86,7 +110,8 @@ if(USE_CUDA)
-gencode=arch=compute_61,code=sm_61
-gencode=arch=compute_62,code=sm_62
-gencode=arch=compute_70,code=sm_70
-gencode=arch=compute_70,code=compute_70)
-gencode=arch=compute_70,code=compute_70
)
else()
set(CMAKE_CUDA_FLAGS ${CMAKE_CUDA_FLAGS} -arch=sm_30
-gencode=arch=compute_30,code=sm_30
......@@ -96,20 +121,19 @@ if(USE_CUDA)
-gencode=arch=compute_61,code=sm_61
-gencode=arch=compute_62,code=sm_62
-gencode=arch=compute_70,code=sm_70
-gencode=arch=compute_70,code=compute_70)
-gencode=arch=compute_70,code=compute_70
)
endif()
set(CMAKE_POLICY_DEFAULT_CMP0028 NEW)
link_directories("${CUDA_TOOLKIT_ROOT_DIR}/lib/x64")
include_directories("${CUDA_TOOLKIT_ROOT_DIR}/include")
set(CUDA_LIB_PATH_ME "${CUDA_TOOLKIT_ROOT_DIR}/lib/x64/")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}cublas.lib")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}npps.lib")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}nppc.lib")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}cudadevrt.lib")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}curand.lib")
my_add_executable(${NIUTRANS_EXE} ${CPP_FILES} ${H_FILES} ${CU_FILES} ${CUH_FILES})
else()
find_package(CUDA ${CUDA_VERSION} REQUIRED)
set(CUDA_LIB_DIR "${CUDA_TOOLKIT_ROOT_DIR}/lib/x64/")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}cublas.lib")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}npps.lib")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}nppc.lib")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}cudadevrt.lib")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}curand.lib")
else()
set(CMAKE_CXX_FLAGS "-fPIC -msse4.2 -w -march=native -Wno-enum-compare -Wno-sign-compare -Wno-format -Wno-dev -O3 -DNDEBUG -rdynamic")
set(CUDA_NVCC_FLAGS "-Xcompiler -fPIC -maxrregcount=0 --disable-warnings -use_fast_math -DUSE_CUDA -Wno-deprecated-gpu-targets -std=c++11")
if(USE_HALF_PRECISION)
......@@ -119,7 +143,8 @@ if(USE_CUDA)
-gencode=arch=compute_61,code=sm_61
-gencode=arch=compute_62,code=sm_62
-gencode=arch=compute_70,code=sm_70
-gencode=arch=compute_70,code=compute_70)
-gencode=arch=compute_70,code=compute_70
)
else()
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} -arch=sm_30
-gencode=arch=compute_30,code=sm_30
......@@ -129,44 +154,152 @@ if(USE_CUDA)
-gencode=arch=compute_61,code=sm_61
-gencode=arch=compute_62,code=sm_62
-gencode=arch=compute_70,code=sm_70
-gencode=arch=compute_70,code=compute_70)
-gencode=arch=compute_70,code=compute_70
)
endif()
link_directories(${CUDA_TOOLKIT_ROOT_DIR}/lib64)
include_directories(${CUDA_TOOLKIT_ROOT_DIR}/include)
set(CUDA_LIB_PATH_ME "${CUDA_TOOLKIT_ROOT_DIR}/lib64/")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}libcublas_static.a")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}libculibos.a")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}libnpps_static.a")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}libnppc_static.a")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}libcudadevrt.a")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "${CUDA_LIB_PATH_ME}libcurand_static.a")
set(CUDA_LIBS_ME ${CUDA_LIBS_ME} "/usr/lib64/libdl.so.2")
my_add_executable(${NIUTRANS_EXE} ${CPP_FILES} ${H_FILES} ${CU_FILES} ${CUH_FILES})
set(CUDA_LIB_DIR "${CUDA_TOOLKIT_ROOT_DIR}/lib64/")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libcublas_static.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libculibos.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libnpps_static.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libnppc_static.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libcudadevrt.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "${CUDA_LIB_DIR}libcurand_static.a")
set(CUDA_LIB_PATH ${CUDA_LIB_PATH} "/usr/lib64/libdl.so.2")
endif()
else()
set(NIUTRANS_EXE "${NIUTRANS_EXE}.CPU")
my_add_executable(${NIUTRANS_EXE} ${CPP_FILES} ${H_FILES})
endif()
set(CUDA_LIB ${CUDA_LIBS_ME})
# Set libs and compiler options for MKL
if(USE_MKL)
add_definitions(-DUSE_BLAS -DMKL)
set(COMPILER_DIR "${INTEL_ROOT}/compiler")
set(MKL_DIR "${INTEL_ROOT}/mkl")
set(CPU_ARCH intel64)
if(WIN32)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DNDEBUG")
link_directories(${MKL_DIR}/lib/intel64/)
link_directories(${COMPILER_DIR}/lib/intel64)
include_directories(${MKL_DIR}/include)
set(COMPILER_LIB_DIR "${COMPILER_DIR}/lib/intel64/")
set(MKL_LIB_DIR "${MKL_DIR}/lib/intel64/")
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${MKL_LIB_DIR}mkl_intel_lp64.lib")
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${MKL_LIB_DIR}mkl_core.lib")
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${MKL_LIB_DIR}mkl_intel_thread.lib")
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${COMPILER_LIB_DIR}libiomp5md.lib")
else()
if(USE_CUDA)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-reorder")
else()
set(CMAKE_CXX_FLAGS "-std=c++11 -msse4.2 -w -march=native -Wno-enum-compare -Wno-sign-compare -Wno-reorder -Wno-format -O3 -flto -DNDEBUG -rdynamic -fkeep-inline-functions -fPIC")
endif(USE_CUDA)
link_directories(${MKL_DIR}/lib/intel64/)
link_directories(${COMPILER_DIR}/lib/intel64)
include_directories(${MKL_DIR}/include)
set(COMPILER_LIB_DIR "${COMPILER_DIR}/lib/intel64/")
set(MKL_LIB_DIR "${MKL_DIR}/lib/intel64/")
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${MKL_LIB_DIR}libmkl_intel_lp64.a")
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${MKL_LIB_DIR}libmkl_core.a")
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${MKL_LIB_DIR}libmkl_intel_thread.a")
set(MKL_LIB_PATH ${MKL_LIB_PATH} "${COMPILER_LIB_DIR}libiomp5.a")
endif()
endif()
if(ON_WINDOWS)
add_definitions(-DWIN32) # for windows specially
# Set libs and compiler options for OpenBLAS
if(USE_OPENBLAS)
add_definitions(-DUSE_BLAS -DOPENBLAS)
set(OPENBLAS_INCLUDE_DIR "${OPENBLAS_ROOT}/include")
set(OPENBLAS_LIB_DIR "${OPENBLAS_ROOT}/lib")
if(WIN32)
link_directories(${OPENBLAS_LIB_DIR})
include_directories(${OPENBLAS_INCLUDE_DIR})
set(OPENBLAS_LIB_PATH ${OPENBLAS_LIB_PATH} "${OPENBLAS_LIB_DIR}/libopenblas.lib")
else()
link_directories(${OPENBLAS_LIB_DIR})
include_directories(${OPENBLAS_INCLUDE_DIR})
set(OPENBLAS_LIB_PATH ${OPENBLAS_LIB_PATH} "${OPENBLAS_LIB_DIR}/libopenblas.a")
set(OPENBLAS_LIB_PATH ${OPENBLAS_LIB_PATH} "${OPENBLAS_LIB_DIR}/libopenblas_haswellp-r0.3.10.a")
endif()
endif()
# Integrate all libs
set(CUDA_LIB ${CUDA_LIB_PATH})
set(MKL_LIB ${MKL_LIB_PATH})
set(OPENBLAS_LIB ${OPENBLAS_LIB_PATH})
# Add executable files to project
# Generate dynamic link library about project
if(USE_CUDA)
if(GEN_DLL)
cuda_add_library(${NIUTENSOR_DLL} SHARED ${CPP_FILES} ${H_FILES} ${CU_FILES} ${CUH_FILES})
else()
my_add_executable(${NIUTENSOR_EXE} ${CPP_FILES} ${H_FILES} ${CU_FILES} ${CUH_FILES})
endif()
else()
if(GEN_DLL)
add_library(${NIUTENSOR_DLL} SHARED ${CPP_FILES} ${H_FILES})
else()
my_add_executable(${NIUTENSOR_EXE} ${CPP_FILES} ${H_FILES})
endif()
endif()
# Link external libs to executable files
# Link external libs to dynamic link library
if(WIN32)
add_definitions(-DWIN32)
set(MESS ${MESS} "On Windows")
if(USE_CUDA)
message(STATUS "On Windows and use cuda")
target_link_libraries(${NIUTRANS_EXE} ${CUDA_LIB})
set(MESS ${MESS} " Use CUDA")
set(ALL_LIB ${ALL_LIB} ${CUDA_LIB})
endif()
if(USE_MKL)
set(MESS ${MESS} " Use MKL")
set(ALL_LIB ${ALL_LIB} ${MKL_LIB})
elseif(USE_OPENBLAS)
set(MESS ${MESS} " Use OpenBLAS")
set(ALL_LIB ${ALL_LIB} ${OPENBLAS_LIB})
else()
endif()
if(GEN_DLL)
message(STATUS "Generate Dynamic Link Library")
message(STATUS "Name of Dynamic Link Library: " ${NIUTENSOR_DLL})
target_link_libraries(${NIUTENSOR_DLL} ${ALL_LIB})
else()
message(STATUS "On Windows and not use cuda")
target_link_libraries(${NIUTRANS_EXE} ${CUDA_LIB})
message(STATUS "Generate Makefile For Executable File")
message(STATUS "Name of Executable File :" ${NIUTENSOR_EXE})
target_link_libraries(${NIUTENSOR_EXE} ${ALL_LIB})
endif()
message(STATUS "${MESS}")
else()
add_definitions(-std=c++11)
set(MESS ${MESS} "On Linux or macOS")
if(USE_CUDA)
message(STATUS "On Linux and use cuda")
target_link_libraries(${NIUTRANS_EXE} ${CUDA_LIB} -lpthread -lcudart -lnvidia-ml)
set(MESS ${MESS} " Use CUDA")
set(ALL_LIB ${ALL_LIB} ${CUDA_LIB})
set(FLAG ${FLAG} "-lpthread -lcudart -lnvidia-ml")
else()
set(FLAG ${FLAG} "-lpthread")
endif()
if(USE_MKL)
set(MESS ${MESS} " Use MKL")
set(ALL_LIB ${ALL_LIB} ${MKL_LIB})
set(FLAG ${FLAG} "-liomp5 -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -ldl")
elseif(USE_OPENBLAS)
set(MESS ${MESS} " Use OpenBLAS")
set(ALL_LIB ${ALL_LIB} ${OPENBLAS_LIB})
set(FLAG ${FLAG} "-lopenblas")
else()
message(STATUS "On Linux and not use cuda")
target_link_libraries(${NIUTRANS_EXE} ${CUDA_LIB} -lpthread)
endif()
endif()
message(STATUS "name of executable file:" ${NIUTRANS_EXE})
\ No newline at end of file
if(GEN_DLL)
message(STATUS "Generate Dynamic Link Library")
message(STATUS "Name of Dynamic Link Library: " ${NIUTENSOR_DLL})
target_link_libraries(${NIUTENSOR_DLL} ${ALL_LIB} ${FLAG})
else()
message(STATUS "Generate Makefile For Executable File")
message(STATUS "Name of Executable File: " ${NIUTENSOR_EXE})
target_link_libraries(${NIUTENSOR_EXE} ${ALL_LIB} ${FLAG})
endif()
message(STATUS "${MESS}")
endif()
\ No newline at end of file
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论