未验证 提交 068d3690 编写于 作者: S Shibo Tao 提交者: GitHub

support CUDA using cmake built-in way (#24395)

* support CUDA using cmake built-in way. test=develop

* test=develop
上级 d0307145
......@@ -20,6 +20,19 @@ set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
include(system)
project(paddle CXX C)
include(init)
# enable language CUDA
# TODO(Shibo Tao): remove find_package(CUDA) completely.
find_package(CUDA QUIET)
option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
if(WITH_GPU AND NOT APPLE)
enable_language(CUDA)
message(STATUS "CUDA compiler: ${CMAKE_CUDA_COMPILER}, version: "
"${CMAKE_CUDA_COMPILER_ID} ${CMAKE_CUDA_COMPILER_VERSION}")
endif()
message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
"${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
......@@ -52,14 +65,12 @@ else(WIN32)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=deprecated-declarations -Wno-deprecated-declarations")
endif(WIN32)
find_package(CUDA QUIET)
find_package(Git REQUIRED)
find_package(Threads REQUIRED)
include(simd)
################################ Exposed Configurations #######################################
option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON)
......
......@@ -76,7 +76,7 @@ if(WITH_GPU)
FIND_PACKAGE(CUDA REQUIRED)
if(${CUDA_VERSION_MAJOR} VERSION_LESS 7)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_LESS 7)
message(FATAL_ERROR "Paddle needs CUDA >= 7.0 to compile")
endif()
......@@ -89,7 +89,7 @@ if(WITH_GPU)
else()
message(STATUS "Cannot find CUPTI, GPU Profiling is incorrect.")
endif()
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS} "-Xcompiler ${SIMD_FLAG}")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Xcompiler=\"${SIMD_FLAG}\"")
# Include cuda and cudnn
include_directories(${CUDNN_INCLUDE_DIR})
......@@ -97,11 +97,11 @@ if(WITH_GPU)
if(TENSORRT_FOUND)
if(WIN32)
if(${CUDA_VERSION_MAJOR} VERSION_LESS 9)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_LESS 9)
message(FATAL_ERROR "TensorRT needs CUDA >= 9.0 to compile on Windows")
endif()
else()
if(${CUDA_VERSION_MAJOR} VERSION_LESS 8)
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_LESS 8)
message(FATAL_ERROR "TensorRT needs CUDA >= 8.0 to compile")
endif()
if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7)
......
......@@ -104,12 +104,12 @@ function(select_nvcc_arch_flags out_variable)
elseif(${CUDA_ARCH_NAME} STREQUAL "Pascal")
set(cuda_arch_bin "60 61")
elseif(${CUDA_ARCH_NAME} STREQUAL "Volta")
if (NOT ${CUDA_VERSION} LESS 10.0)
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} LESS 10.0)
add_definitions("-DSUPPORTS_CUDA_FP16")
endif()
set(cuda_arch_bin "70")
elseif(${CUDA_ARCH_NAME} STREQUAL "Turing")
if (NOT ${CUDA_VERSION} LESS 10.0)
if (NOT ${CMAKE_CUDA_COMPILER_VERSION} LESS 10.0)
add_definitions("-DSUPPORTS_CUDA_FP16")
endif()
set(cuda_arch_bin "75")
......@@ -142,19 +142,19 @@ function(select_nvcc_arch_flags out_variable)
foreach(arch ${cuda_arch_bin})
if(arch MATCHES "([0-9]+)\\(([0-9]+)\\)")
# User explicitly specified PTX for the concrete BIN
list(APPEND nvcc_flags -gencode arch=compute_${CMAKE_MATCH_2},code=sm_${CMAKE_MATCH_1})
list(APPEND nvcc_archs_readable sm_${CMAKE_MATCH_1})
string(APPEND nvcc_flags " -gencode arch=compute_${CMAKE_MATCH_2},code=sm_${CMAKE_MATCH_1}")
string(APPEND nvcc_archs_readable " sm_${CMAKE_MATCH_1}")
else()
# User didn't explicitly specify PTX for the concrete BIN, we assume PTX=BIN
list(APPEND nvcc_flags -gencode arch=compute_${arch},code=sm_${arch})
list(APPEND nvcc_archs_readable sm_${arch})
string(APPEND nvcc_flags " -gencode arch=compute_${arch},code=sm_${arch}")
string(APPEND nvcc_archs_readable " sm_${arch}")
endif()
endforeach()
# Tell NVCC to add PTX intermediate code for the specified architectures
foreach(arch ${cuda_arch_ptx})
list(APPEND nvcc_flags -gencode arch=compute_${arch},code=compute_${arch})
list(APPEND nvcc_archs_readable compute_${arch})
string(APPEND nvcc_flags " -gencode arch=compute_${arch},code=compute_${arch}")
string(APPEND nvcc_archs_readable " compute_${arch}")
endforeach()
string(REPLACE ";" " " nvcc_archs_readable "${nvcc_archs_readable}")
......@@ -162,32 +162,32 @@ function(select_nvcc_arch_flags out_variable)
set(${out_variable}_readable ${nvcc_archs_readable} PARENT_SCOPE)
endfunction()
message(STATUS "CUDA detected: " ${CUDA_VERSION})
if (${CUDA_VERSION} LESS 7.0)
message(STATUS "CUDA detected: " ${CMAKE_CUDA_COMPILER_VERSION})
if (${CMAKE_CUDA_COMPILER_VERSION} LESS 7.0)
set(paddle_known_gpu_archs ${paddle_known_gpu_archs})
elseif (${CUDA_VERSION} LESS 8.0) # CUDA 7.x
elseif (${CMAKE_CUDA_COMPILER_VERSION} LESS 8.0) # CUDA 7.x
set(paddle_known_gpu_archs ${paddle_known_gpu_archs7})
list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__")
elseif (${CUDA_VERSION} LESS 9.0) # CUDA 8.x
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D_MWAITXINTRIN_H_INCLUDED")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D__STRICT_ANSI__")
elseif (${CMAKE_CUDA_COMPILER_VERSION} LESS 9.0) # CUDA 8.x
set(paddle_known_gpu_archs ${paddle_known_gpu_archs8})
list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D_MWAITXINTRIN_H_INCLUDED")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D__STRICT_ANSI__")
# CUDA 8 may complain that sm_20 is no longer supported. Suppress the
# warning for now.
list(APPEND CUDA_NVCC_FLAGS "-Wno-deprecated-gpu-targets")
elseif (${CUDA_VERSION} LESS 10.0) # CUDA 9.x
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -Wno-deprecated-gpu-targets")
elseif (${CMAKE_CUDA_COMPILER_VERSION} LESS 10.0) # CUDA 9.x
set(paddle_known_gpu_archs ${paddle_known_gpu_archs9})
list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__")
elseif (${CUDA_VERSION} LESS 11.0) # CUDA 10.x
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D_MWAITXINTRIN_H_INCLUDED")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D__STRICT_ANSI__")
elseif (${CMAKE_CUDA_COMPILER_VERSION} LESS 11.0) # CUDA 10.x
set(paddle_known_gpu_archs ${paddle_known_gpu_archs10})
list(APPEND CUDA_NVCC_FLAGS "-D_MWAITXINTRIN_H_INCLUDED")
list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D_MWAITXINTRIN_H_INCLUDED")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D__STRICT_ANSI__")
endif()
add_definitions("-DPADDLE_CUDA_BINVER=\"${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}\"")
include_directories(${CUDA_INCLUDE_DIRS})
message(STATUS "PADDLE_CUDA_BINVER=${CUDA_VERSION_MAJOR}${CUDA_VERSION_MINOR}")
if(NOT WITH_DSO)
if(WIN32)
set_property(GLOBAL PROPERTY CUDA_MODULES ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${CUDA_cusolver_LIBRARY})
......@@ -196,37 +196,24 @@ endif(NOT WITH_DSO)
# setting nvcc arch flags
select_nvcc_arch_flags(NVCC_FLAGS_EXTRA)
list(APPEND CUDA_NVCC_FLAGS ${NVCC_FLAGS_EXTRA})
message(STATUS "Added CUDA NVCC flags for: ${NVCC_FLAGS_EXTRA_readable}")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${NVCC_FLAGS_EXTRA}")
message(STATUS "NVCC_FLAGS_EXTRA: ${NVCC_FLAGS_EXTRA}")
# Set C++11 support
set(CUDA_PROPAGATE_HOST_FLAGS OFF)
# Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
# So, don't set these flags here.
if (NOT WIN32) # windows msvc2015 support c++11 natively.
# -std=c++11 -fPIC not recoginize by msvc, -Xcompiler will be added by cmake.
list(APPEND CUDA_NVCC_FLAGS "-std=c++11")
list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC")
# -std=c++11 -fPIC not recoginize by msvc, -Xcompiler will be added by cmake.
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -std=c++11")
endif(NOT WIN32)
# in cuda9, suppress cuda warning on eigen
list(APPEND CUDA_NVCC_FLAGS "-w")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -w")
# Set :expt-relaxed-constexpr to suppress Eigen warnings
list(APPEND CUDA_NVCC_FLAGS "--expt-relaxed-constexpr")
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --expt-relaxed-constexpr")
if (NOT WIN32)
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG})
elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE})
elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel")
# nvcc 9 does not support -Os. Use Release flags instead
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE})
endif()
else(NOT WIN32)
if (WIN32)
list(APPEND CUDA_NVCC_FLAGS "-Xcompiler \"/wd 4244 /wd 4267 /wd 4819\"")
list(APPEND CUDA_NVCC_FLAGS "--compiler-options;/bigobj")
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
......@@ -237,8 +224,8 @@ else(NOT WIN32)
list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
else()
message(FATAL "Windows only support Release or Debug build now. Please set visual studio build type to Release/Debug, x64 build.")
endif()
endif(NOT WIN32)
endif()
endif(WIN32)
mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD)
mark_as_advanced(CUDA_SDK_ROOT_DIR CUDA_SEPARABLE_COMPILATION)
......@@ -76,7 +76,7 @@ macro(safe_set_nvflag flag_name)
CHECK_C_COMPILER_FLAG(${flag_name} C_COMPILER_SUPPORT_FLAG_${safe_name})
set(safe_name C_COMPILER_SUPPORT_FLAG_${safe_name})
if(${safe_name})
LIST(APPEND CUDA_NVCC_FLAGS -Xcompiler ${flag_name})
set(SAFE_GPU_COMMON_FLAGS "${SAFE_GPU_COMMON_FLAGS} -Xcompiler=\"${flag_name}\"")
endif()
endmacro()
......@@ -212,10 +212,14 @@ foreach(flag ${COMMON_FLAGS})
safe_set_cxxflag(CMAKE_CXX_FLAGS ${flag})
endforeach()
set(SAFE_GPU_COMMON_FLAGS "")
foreach(flag ${GPU_COMMON_FLAGS})
safe_set_nvflag(${flag})
endforeach()
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${SAFE_GPU_COMMON_FLAGS}")
if(WIN32 AND MSVC_STATIC_CRT)
# windows build turn off warnings.
safe_set_static_flag()
......
......@@ -412,10 +412,14 @@ function(nv_library TARGET_NAME)
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(nv_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
if(nv_library_SRCS)
# Attention:
# 1. cuda_add_library is deprecated after cmake v3.10, use add_library for CUDA please.
# 2. cuda_add_library does not support ccache.
# Reference: https://cmake.org/cmake/help/v3.10/module/FindCUDA.html
if (nv_library_SHARED OR nv_library_shared) # build *.so
cuda_add_library(${TARGET_NAME} SHARED ${nv_library_SRCS})
add_library(${TARGET_NAME} SHARED ${nv_library_SRCS})
else()
cuda_add_library(${TARGET_NAME} STATIC ${nv_library_SRCS})
add_library(${TARGET_NAME} STATIC ${nv_library_SRCS})
find_fluid_modules(${TARGET_NAME})
endif()
if (nv_library_DEPS)
......@@ -450,7 +454,7 @@ function(nv_binary TARGET_NAME)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(nv_binary "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cuda_add_executable(${TARGET_NAME} ${nv_binary_SRCS})
add_executable(${TARGET_NAME} ${nv_binary_SRCS})
if(nv_binary_DEPS)
target_link_libraries(${TARGET_NAME} ${nv_binary_DEPS})
add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
......@@ -464,7 +468,11 @@ function(nv_test TARGET_NAME)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(nv_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
# Attention:
# 1. cuda_add_executable is deprecated after cmake v3.10, use cuda_add_executable for CUDA please.
# 2. cuda_add_executable does not support ccache.
# Reference: https://cmake.org/cmake/help/v3.10/module/FindCUDA.html
add_executable(${TARGET_NAME} ${nv_test_SRCS})
get_property(os_dependency_modules GLOBAL PROPERTY OS_DEPENDENCY_MODULES)
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog ${os_dependency_modules})
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
......
# Attention: cmake will append these flags to compile command automatically.
# So if you want to add global option, change this file rather than flags.cmake
# default: "-g"
set(CMAKE_C_FLAGS_DEBUG "-g")
# default: "-O3 -DNDEBUG"
set(CMAKE_C_FLAGS_RELEASE "-O3 -DNDEBUG")
# default: "-O2 -g -DNDEBUG"
set(CMAKE_C_FLAGS_RELWITHDEBINFO "-O2 -g -DNDEBUG")
# default: "-Os -DNDEBUG"
set(CMAKE_C_FLAGS_MINSIZEREL "-Os -DNDEBUG")
# default: "-g"
set(CMAKE_CXX_FLAGS_DEBUG "-g")
# default: "-O3 -DNDEBUG"
set(CMAKE_CXX_FLAGS_RELEASE "-O3 -DNDEBUG")
# default: "-O2 -g -DNDEBUG"
set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O2 -g -DNDEBUG")
# default: "-Os -DNDEBUG"
set(CMAKE_CXX_FLAGS_MINSIZEREL "-Os -DNDEBUG")
# default: "-g"
set(CMAKE_CUDA_FLAGS_DEBUG "-g")
# default: "-O3 -DNDEBUG"
set(CMAKE_CUDA_FLAGS_RELEASE "-O3 -DNDEBUG")
# default: "-O2 -g -DNDEBUG"
set(CMAKE_CUDA_FLAGS_RELWITHDEBINFO "-O2 -g -DNDEBUG")
# default: "-O1 -DNDEBUG"
set(CMAKE_CUDA_FLAGS_MINSIZEREL "-O1 -DNDEBUG")
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册