mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
30c4c6ff9b
Summary: This diff refactors the code by moving CUDAAllocatorConfig into the header file. This config refactoring is done so that we can use the same config code for CUDA pinned memory as well. Test Plan: sandcastle Differential Revision: D49653265 Pull Request resolved: https://github.com/pytorch/pytorch/pull/110123 Approved by: https://github.com/zdevito
90 lines
2.9 KiB
CMake
90 lines
2.9 KiB
CMake
# Build file for the C10 CUDA.
|
|
#
|
|
# C10 CUDA is a minimal library, but it does depend on CUDA.
|
|
|
|
include(../../cmake/public/utils.cmake)
|
|
include(../../cmake/public/cuda.cmake)
|
|
|
|
# ---[ Configure macro file.
|
|
set(C10_CUDA_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS}) # used in cmake_macros.h.in
|
|
# Probably have to do this :(
|
|
configure_file(
|
|
${CMAKE_CURRENT_LIST_DIR}/impl/cuda_cmake_macros.h.in
|
|
${CMAKE_BINARY_DIR}/c10/cuda/impl/cuda_cmake_macros.h)
|
|
|
|
# Note: if you want to add ANY dependency to the c10 library, make sure you
|
|
# check with the core PyTorch developers as the dependency will be
|
|
# transitively passed on to all libraries dependent on PyTorch.
|
|
|
|
# Note: if you add a new source file/header, you will need to update
|
|
# torch/utils/hipify/cuda_to_hip_mappings.py for new files
|
|
# and headers you add
|
|
set(C10_CUDA_SRCS
|
|
CUDAAllocatorConfig.cpp
|
|
CUDACachingAllocator.cpp
|
|
CUDADeviceAssertionHost.cpp
|
|
CUDAException.cpp
|
|
CUDAFunctions.cpp
|
|
CUDAMallocAsyncAllocator.cpp
|
|
CUDAMiscFunctions.cpp
|
|
CUDAStream.cpp
|
|
impl/CUDAGuardImpl.cpp
|
|
impl/CUDATest.cpp
|
|
driver_api.cpp
|
|
)
|
|
set(C10_CUDA_HEADERS
|
|
CUDAAllocatorConfig.h
|
|
CUDACachingAllocator.h
|
|
CUDADeviceAssertionHost.h
|
|
CUDAException.h
|
|
CUDAFunctions.h
|
|
CUDAGuard.h
|
|
CUDAMacros.h
|
|
CUDAMathCompat.h
|
|
CUDAMiscFunctions.h
|
|
CUDAStream.h
|
|
impl/CUDAGuardImpl.h
|
|
impl/CUDATest.h
|
|
)
|
|
set(CUDA_LINK_LIBRARIES_KEYWORD PRIVATE)
|
|
torch_cuda_based_add_library(c10_cuda ${C10_CUDA_SRCS} ${C10_CUDA_HEADERS})
|
|
set(CUDA_LINK_LIBRARIES_KEYWORD)
|
|
# If building shared library, set dllimport/dllexport proper.
|
|
target_compile_options(c10_cuda PRIVATE "-DC10_CUDA_BUILD_MAIN_LIB")
|
|
# Enable hidden visibility if compiler supports it.
|
|
if(${COMPILER_SUPPORTS_HIDDEN_VISIBILITY})
|
|
target_compile_options(c10_cuda PRIVATE "-fvisibility=hidden")
|
|
endif()
|
|
|
|
# ---[ Dependency of c10_cuda
|
|
target_link_libraries(c10_cuda PUBLIC c10 torch::cudart)
|
|
|
|
if(NOT WIN32)
|
|
target_link_libraries(c10_cuda PRIVATE dl)
|
|
target_compile_options(c10_cuda PRIVATE "-DPYTORCH_C10_DRIVER_API_SUPPORTED")
|
|
endif()
|
|
|
|
target_include_directories(
|
|
c10_cuda PUBLIC
|
|
$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../..>
|
|
$<BUILD_INTERFACE:${CMAKE_BINARY_DIR}>
|
|
$<INSTALL_INTERFACE:include>)
|
|
|
|
add_subdirectory(test)
|
|
|
|
# ---[ Installation
|
|
# Note: for now, we will put all export path into one single Caffe2Targets group
|
|
# to deal with the cmake deployment need. Inside the Caffe2Targets set, the
|
|
# individual libraries like libc10.so and libcaffe2.so are still self-contained.
|
|
install(TARGETS c10_cuda EXPORT Caffe2Targets DESTINATION lib)
|
|
foreach(file ${C10_CUDA_HEADERS})
|
|
get_filename_component( dir ${file} DIRECTORY )
|
|
install( FILES ${file} DESTINATION include/c10/cuda/${dir} )
|
|
endforeach()
|
|
install(FILES ${CMAKE_BINARY_DIR}/c10/cuda/impl/cuda_cmake_macros.h
|
|
DESTINATION include/c10/cuda/impl)
|
|
|
|
if(MSVC AND C10_CUDA_BUILD_SHARED_LIBS)
|
|
install(FILES $<TARGET_PDB_FILE:c10_cuda> DESTINATION lib OPTIONAL)
|
|
endif()
|