mirror of
https://github.com/open-mmlab/mmdeploy.git
synced 2025-01-14 08:09:43 +08:00
6fdf6b8616
* apply cmake-format and refactor cmake scripts of building onnx and tensorrt ops * add static target * suppress 'CMAKE_CUDA_ARCHITECTURES' warning when cmake with version 18 or later is used * fix typo
113 lines
4.0 KiB
CMake
113 lines
4.0 KiB
CMake
set(TARGET_NAME mmlab_tensorrt_ops)
|
|
set(SHARED_TARGET ${TARGET_NAME})
|
|
set(STATIC_TARGET ${TARGET_NAME}_static)
|
|
|
|
# to suppress 'CMAKE_CUDA_ARCHITECTURES' warning when cmake version is >= 18.0
|
|
cmake_policy(SET CMP0104 OLD)
|
|
# cuda
|
|
find_package(CUDA REQUIRED)
|
|
include_directories(${CUDA_INCLUDE_DIRS})
|
|
enable_language(CUDA)
|
|
|
|
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
|
|
|
|
if(MSVC)
|
|
set(CMAKE_CUDA_COMPILER ${CUDA_TOOLKIT_ROOT_DIR}/bin/nvcc.exe)
|
|
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -Xcompiler=/wd4819,/wd4828")
|
|
if(HAVE_CXX_FLAG_UTF_8)
|
|
set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} -Xcompiler=/utf-8")
|
|
endif()
|
|
else()
|
|
set(CMAKE_CUDA_COMPILER ${CUDA_TOOLKIT_ROOT_DIR}/bin/nvcc)
|
|
# Explicitly set the cuda host compiler. Because the default host compiler #
|
|
# selected by cmake maybe wrong.
|
|
set(CMAKE_CUDA_HOST_COMPILER ${CMAKE_CXX_COMPILER})
|
|
set(CUDA_NVCC_FLAGS
|
|
"${CUDA_NVCC_FLAGS} -Xcompiler=-fPIC,-Wall,-fvisibility=hidden")
|
|
endif()
|
|
|
|
# set virtual compute architecture and real ones
|
|
set(_NVCC_FLAGS)
|
|
set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_52,code=sm_52")
|
|
if(CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "8")
|
|
set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_60,code=sm_60")
|
|
set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_61,code=sm_61")
|
|
endif()
|
|
if(CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "9")
|
|
set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_70,code=sm_70")
|
|
endif()
|
|
if(CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "10")
|
|
set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_72,code=sm_72")
|
|
set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_75,code=sm_75")
|
|
endif()
|
|
if(CUDA_VERSION_MAJOR VERSION_GREATER_EQUAL "11")
|
|
set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_80,code=sm_80")
|
|
set(_NVCC_FLAGS "${_NVCC_FLAGS} -gencode arch=compute_86,code=sm_86")
|
|
endif()
|
|
|
|
set(CUDA_NVCC_FLAGS_DEBUG "-g -O0")
|
|
set(CUDA_NVCC_FLAGS_RELEASE "-O3")
|
|
set(CUDA_NVCC_FLAGS "-std=c++11 ${CUDA_NVCC_FLAGS}")
|
|
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${CUDA_NVCC_FLAGS} ${_NVCC_FLAGS}")
|
|
|
|
# tensorrt
|
|
find_path(
|
|
TENSORRT_INCLUDE_DIR NvInfer.h
|
|
HINTS ${TENSORRT_DIR} ${CUDA_TOOLKIT_ROOT_DIR}
|
|
PATH_SUFFIXES include)
|
|
if(TENSORRT_INCLUDE_DIR)
|
|
message(STATUS " Found TensorRT headers at ${TENSORRT_INCLUDE_DIR}")
|
|
else()
|
|
message(ERROR " Cannot found TensorRT headers")
|
|
endif()
|
|
|
|
find_library(
|
|
TENSORRT_LIBRARY_INFER nvinfer
|
|
HINTS ${TENSORRT_DIR} ${TENSORRT_BUILD} ${CUDA_TOOLKIT_ROOT_DIR}
|
|
PATH_SUFFIXES lib lib64 lib/x64)
|
|
find_library(
|
|
TENSORRT_LIBRARY_INFER_PLUGIN nvinfer_plugin
|
|
HINTS ${TENSORRT_DIR} ${TENSORRT_BUILD} ${CUDA_TOOLKIT_ROOT_DIR}
|
|
PATH_SUFFIXES lib lib64 lib/x64)
|
|
set(TENSORRT_LIBRARY ${TENSORRT_LIBRARY_INFER}
|
|
${TENSORRT_LIBRARY_INFER_PLUGIN})
|
|
if(TENSORRT_LIBRARY_INFER
|
|
AND TENSORRT_LIBRARY_INFER_PLUGIN)
|
|
message(STATUS "Find TensorRT libs at ${TENSORRT_LIBRARY}")
|
|
else()
|
|
message(FATAL_ERROR " Cannot found TensorRT libs")
|
|
endif()
|
|
find_package_handle_standard_args(TENSORRT DEFAULT_MSG TENSORRT_INCLUDE_DIR
|
|
TENSORRT_LIBRARY)
|
|
if(NOT TENSORRT_FOUND)
|
|
message(ERROR " Cannot find TensorRT library.")
|
|
endif()
|
|
include_directories(${TENSORRT_INCLUDE_DIR})
|
|
|
|
# cudnn
|
|
include_directories(${CUDNN_DIR}/include)
|
|
link_directories(${CUDNN_DIR}/lib64)
|
|
|
|
# cub
|
|
add_definitions(-DTHRUST_IGNORE_DEPRECATED_CPP_DIALECT)
|
|
if(NOT DEFINED CUB_ROOT_DIR)
|
|
set(CUB_ROOT_DIR "${PROJECT_SOURCE_DIR}/third_party/cub")
|
|
endif()
|
|
include_directories(${CUB_ROOT_DIR})
|
|
|
|
# add plugin source
|
|
# include_directories(${CMAKE_CURRENT_SOURCE_DIR}/common)
|
|
file(GLOB_RECURSE BACKEND_OPS_SRCS *.cpp *.cu)
|
|
|
|
set(INFER_PLUGIN_LIB ${TENSORRT_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} cudnn)
|
|
|
|
cuda_add_library(${SHARED_TARGET} MODULE ${BACKEND_OPS_SRCS})
|
|
target_link_libraries(${SHARED_TARGET} ${INFER_PLUGIN_LIB})
|
|
target_include_directories(${SHARED_TARGET}
|
|
PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/common)
|
|
|
|
cuda_add_library(${STATIC_TARGET} STATIC ${BACKEND_OPS_SRCS})
|
|
target_link_libraries(${STATIC_TARGET} ${INFER_PLUGIN_LIB})
|
|
target_include_directories(${STATIC_TARGET}
|
|
PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/common)
|