mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-06 12:20:52 +01:00
Split libtorch.so back into libtorch_{cpu,cuda,hip} (#30315)
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/30315 The new structure is that libtorch_cpu contains the bulk of our code, and libtorch depends on libtorch_cpu and libtorch_cuda. This is a reland of https://github.com/pytorch/pytorch/pull/29731 but I've extracted all of the prep work into separate PRs which can be landed before this one. Some things of note: * torch/csrc/cuda/nccl.cpp was added to the wrong list of SRCS, now fixed (this didn't matter before because previously they were all in the same library) * The dummy file for libtorch was brought back from the dead; it was previously deleted in #20774 In an initial version of the patch, I forgot to make torch_cuda explicitly depend on torch_cpu. This lead to some very odd errors, most notably "bin/blob_test: hidden symbol `_ZNK6google8protobuf5Arena17OnArenaAllocationEPKSt9type_infom' in lib/libprotobuf.a(arena.cc.o) is referenced by DSO" * A number of places in Android/iOS builds have to add torch_cuda explicitly as a library, as they do not have transitive dependency calculation working correctly * I had to torch_cpu/torch_cuda caffe2_interface_library so that they get whole-archived linked into torch when you statically link. And I had to do this in an *exported* fashion because torch needs to depend on torch_cpu_library. In the end I exported everything and removed the redefinition in the Caffe2Config.cmake. However, I am not too sure why the old code did it in this way in the first place; however, it doesn't seem to have broken anything to switch it this way. * There's some uses of `__HIP_PLATFORM_HCC__` still in `torch_cpu` code, so I had to apply it to that library too (UGH). This manifests as a failer when trying to run the CUDA fuser. This doesn't really matter substantively right now because we still in-place HIPify, but it would be good to fix eventually. This was a bit difficult to debug because of an unrelated HIP bug, see https://github.com/ROCm-Developer-Tools/HIP/issues/1706 Fixes #27215 (as our libraries are smaller), and executes on part of the plan in #29235. Signed-off-by: Edward Z. Yang <ezyang@fb.com> Test Plan: Imported from OSS Differential Revision: D18790941 Pulled By: ezyang fbshipit-source-id: 01296f6089d3de5e8365251b490c51e694f2d6c7
This commit is contained in:
Edward Yang
Facebook Github Bot
parent
1189595875
commit
38986e1dea
|
|
@ -14,7 +14,7 @@ mkdir -p ${ZIP_DIR}/src
|
|||
cp -R ${ARTIFACTS_DIR}/arm64/include ${ZIP_DIR}/install/
|
||||
# build a FAT bianry
|
||||
cd ${ZIP_DIR}/install/lib
|
||||
target_libs=(libc10.a libclog.a libcpuinfo.a libeigen_blas.a libpytorch_qnnpack.a libtorch.a)
|
||||
target_libs=(libc10.a libclog.a libcpuinfo.a libeigen_blas.a libpytorch_qnnpack.a libtorch_cpu.a libtorch.a)
|
||||
for lib in ${target_libs[*]}
|
||||
do
|
||||
libs=(${ARTIFACTS_DIR}/x86_64/lib/${lib} ${ARTIFACTS_DIR}/arm64/lib/${lib})
|
||||
|
|
|
|||
|
|
@ -72,6 +72,7 @@ if (ANDROID_ABI)
|
|||
endfunction(import_static_lib)
|
||||
|
||||
import_static_lib(libtorch)
|
||||
import_static_lib(libtorch_cpu)
|
||||
import_static_lib(libc10)
|
||||
import_static_lib(libnnpack)
|
||||
import_static_lib(libpytorch_qnnpack)
|
||||
|
|
@ -85,6 +86,7 @@ if (ANDROID_ABI)
|
|||
-Wl,--gc-sections
|
||||
-Wl,--whole-archive
|
||||
libtorch
|
||||
libtorch_cpu
|
||||
-Wl,--no-whole-archive
|
||||
libc10
|
||||
libnnpack
|
||||
|
|
@ -100,6 +102,7 @@ else()
|
|||
target_link_libraries(pytorch_jni
|
||||
fbjni
|
||||
torch
|
||||
torch_cpu
|
||||
c10
|
||||
nnpack
|
||||
pytorch_qnnpack
|
||||
|
|
|
|||
|
|
@ -99,12 +99,18 @@
|
|||
#define CAFFE2_API C10_IMPORT
|
||||
#endif
|
||||
|
||||
// This one will eventually be used by libtorch_cuda.so, but for
|
||||
// now it has the same function as CAFFE2_API
|
||||
#ifdef CAFFE2_BUILD_MAIN_LIB
|
||||
// NB: For now, HIP is overloaded to use the same macro, but ideally
|
||||
// HIPify should translate TORCH_CUDA_API to TORCH_HIP_API
|
||||
#if defined(TORCH_CUDA_BUILD_MAIN_LIB) || defined(TORCH_HIP_BUILD_MAIN_LIB)
|
||||
#define TORCH_CUDA_API C10_EXPORT
|
||||
#else
|
||||
#define TORCH_CUDA_API C10_IMPORT
|
||||
#endif
|
||||
|
||||
#if defined(TORCH_HIP_BUILD_MAIN_LIB)
|
||||
#define TORCH_HIP_API C10_EXPORT
|
||||
#else
|
||||
#define TORCH_HIP_API C10_IMPORT
|
||||
#endif
|
||||
|
||||
#endif // C10_MACROS_MACROS_H_
|
||||
|
|
|
|||
|
|
@ -477,11 +477,6 @@ if (NOT INTERN_BUILD_MOBILE OR NOT BUILD_CAFFE2_MOBILE)
|
|||
)
|
||||
endif()
|
||||
|
||||
if (USE_NCCL)
|
||||
list(APPEND TORCH_SRCS
|
||||
${TORCH_SRC_DIR}/csrc/cuda/nccl.cpp)
|
||||
endif()
|
||||
|
||||
if (NOT INTERN_BUILD_MOBILE)
|
||||
list(APPEND TORCH_SRCS
|
||||
${TORCH_SRC_DIR}/csrc/api/src/jit.cpp
|
||||
|
|
@ -535,6 +530,10 @@ if (NOT INTERN_BUILD_MOBILE OR NOT BUILD_CAFFE2_MOBILE)
|
|||
target_link_libraries(caffe2_nvrtc ${CUDA_NVRTC} ${CUDA_CUDA_LIB} ${CUDA_NVRTC_LIB})
|
||||
target_include_directories(caffe2_nvrtc PRIVATE ${CUDA_INCLUDE_DIRS})
|
||||
install(TARGETS caffe2_nvrtc DESTINATION "${TORCH_INSTALL_LIB_DIR}")
|
||||
if (USE_NCCL)
|
||||
list(APPEND Caffe2_GPU_SRCS
|
||||
${TORCH_SRC_DIR}/csrc/cuda/nccl.cpp)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
if (USE_ROCM)
|
||||
|
|
@ -544,6 +543,10 @@ if (NOT INTERN_BUILD_MOBILE OR NOT BUILD_CAFFE2_MOBILE)
|
|||
${TORCH_SRC_DIR}/csrc/autograd/functions/comm.cpp
|
||||
${TORCH_SRC_DIR}/csrc/cuda/comm.cpp
|
||||
)
|
||||
if (USE_NCCL)
|
||||
list(APPEND Caffe2_HIP_SRCS
|
||||
${TORCH_SRC_DIR}/csrc/cuda/nccl.cpp)
|
||||
endif()
|
||||
# caffe2_nvrtc's stubs to driver APIs are useful for HIP.
|
||||
# See NOTE [ ATen NVRTC Stub and HIP ]
|
||||
add_library(caffe2_nvrtc SHARED ${ATen_NVRTC_STUB_SRCS})
|
||||
|
|
@ -616,41 +619,52 @@ endif()
|
|||
# ==========================================================
|
||||
|
||||
|
||||
# Instead of separate .so libraries, GPU sources are now conditionally
|
||||
# compiled into the main torch.so library.
|
||||
if(USE_CUDA)
|
||||
list(APPEND Caffe2_CPU_SRCS ${Caffe2_GPU_SRCS})
|
||||
foreach(tmp ${Caffe2_GPU_SRCS})
|
||||
message(STATUS " " ${tmp})
|
||||
endforeach()
|
||||
add_library(torch_cpu ${Caffe2_CPU_SRCS})
|
||||
torch_compile_options(torch_cpu) # see cmake/public/utils.cmake
|
||||
|
||||
# This is required for older versions of CMake, which don't allow
|
||||
# specifying add_library() without a list of source files
|
||||
set(DUMMY_EMPTY_FILE ${CMAKE_BINARY_DIR}/empty.cpp)
|
||||
|
||||
if (MSVC)
|
||||
set(DUMMY_FILE_CONTENT "__declspec(dllexport) int ignore_this_library_placeholder(){return 0\\;}")
|
||||
else()
|
||||
set(DUMMY_FILE_CONTENT "")
|
||||
endif()
|
||||
|
||||
file(WRITE ${DUMMY_EMPTY_FILE} ${DUMMY_FILE_CONTENT})
|
||||
|
||||
# Wrapper library for people who link against torch and expect both CPU and CUDA support
|
||||
# Contains "torch_cpu" and "torch_cuda"
|
||||
add_library(torch ${DUMMY_EMPTY_FILE})
|
||||
|
||||
if(USE_ROCM)
|
||||
filter_list(__caffe2_hip_srcs_cpp Caffe2_HIP_SRCS "\\.(cu|hip)$")
|
||||
set_source_files_properties(${__caffe2_hip_srcs_cpp} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
|
||||
|
||||
list(APPEND Caffe2_CPU_SRCS ${Caffe2_HIP_SRCS})
|
||||
endif()
|
||||
|
||||
|
||||
# Compile exposed libraries.
|
||||
IF (USE_ROCM)
|
||||
set(CUDA_LINK_LIBRARIES_KEYWORD PRIVATE)
|
||||
hip_add_library(torch ${Caffe2_CPU_SRCS})
|
||||
hip_add_library(torch_hip ${Caffe2_HIP_SRCS})
|
||||
set(CUDA_LINK_LIBRARIES_KEYWORD)
|
||||
torch_compile_options(torch_hip) # see cmake/public/utils.cmake
|
||||
# TODO: Not totally sure if this is live or not
|
||||
if (USE_NCCL)
|
||||
target_link_libraries(torch_hip PRIVATE __caffe2_nccl)
|
||||
target_compile_definitions(torch_hip PRIVATE USE_NCCL)
|
||||
endif()
|
||||
ELSEIF(USE_CUDA)
|
||||
set(CUDA_LINK_LIBRARIES_KEYWORD PRIVATE)
|
||||
cuda_add_library(torch ${Caffe2_CPU_SRCS})
|
||||
cuda_add_library(torch_cuda ${Caffe2_GPU_SRCS})
|
||||
set(CUDA_LINK_LIBRARIES_KEYWORD)
|
||||
ELSE()
|
||||
add_library(torch ${Caffe2_CPU_SRCS})
|
||||
torch_compile_options(torch_cuda) # see cmake/public/utils.cmake
|
||||
if (USE_NCCL)
|
||||
target_link_libraries(torch_cuda PRIVATE __caffe2_nccl)
|
||||
target_compile_definitions(torch_cuda PRIVATE USE_NCCL)
|
||||
endif()
|
||||
ENDIF()
|
||||
|
||||
if (USE_NCCL)
|
||||
target_link_libraries(torch PRIVATE __caffe2_nccl)
|
||||
target_compile_definitions(torch PRIVATE USE_NCCL)
|
||||
endif()
|
||||
|
||||
|
||||
# ==========================================================
|
||||
|
|
@ -662,14 +676,12 @@ if (NOT INTERN_BUILD_MOBILE)
|
|||
# Adding the generated header file to the ${TORCH_SRCS} list is not sufficient
|
||||
# to establish the dependency, since the generation procedure is declared in a different CMake file.
|
||||
# See https://samthursfield.wordpress.com/2015/11/21/cmake-dependencies-between-targets-and-files-and-custom-commands/#custom-commands-in-different-directories
|
||||
add_dependencies(torch Caffe2_PROTO)
|
||||
add_dependencies(torch_cpu Caffe2_PROTO)
|
||||
endif()
|
||||
|
||||
torch_compile_options(torch) # see cmake/public/utils.cmake
|
||||
|
||||
if (NOT INTERN_BUILD_MOBILE OR NOT BUILD_CAFFE2_MOBILE)
|
||||
if (NOT NO_API)
|
||||
target_include_directories(torch PRIVATE
|
||||
target_include_directories(torch_cpu PRIVATE
|
||||
${TORCH_SRC_DIR}/csrc/api
|
||||
${TORCH_SRC_DIR}/csrc/api/include)
|
||||
endif()
|
||||
|
|
@ -686,7 +698,7 @@ if (NOT INTERN_BUILD_MOBILE OR NOT BUILD_CAFFE2_MOBILE)
|
|||
set(TORCH_CUDA_LIBRARIES
|
||||
${NVTOOLEXT_HOME}/lib/x64/nvToolsExt64_1.lib
|
||||
${CUDA_LIBRARIES})
|
||||
target_include_directories(torch PUBLIC "${NVTOOLEXT_HOME}/include")
|
||||
target_include_directories(torch_cuda PUBLIC "${NVTOOLEXT_HOME}/include")
|
||||
|
||||
elseif(APPLE)
|
||||
set(TORCH_CUDA_LIBRARIES
|
||||
|
|
@ -701,8 +713,6 @@ if (NOT INTERN_BUILD_MOBILE OR NOT BUILD_CAFFE2_MOBILE)
|
|||
${LIBNVTOOLSEXT}
|
||||
${CUDA_LIBRARIES})
|
||||
endif()
|
||||
|
||||
target_compile_definitions(torch PRIVATE USE_CUDA)
|
||||
endif()
|
||||
|
||||
|
||||
|
|
@ -713,7 +723,7 @@ if (NOT INTERN_BUILD_MOBILE OR NOT BUILD_CAFFE2_MOBILE)
|
|||
${TORCH_ROOT}/aten/src
|
||||
${CMAKE_CURRENT_BINARY_DIR}/aten/src
|
||||
${CMAKE_BINARY_DIR}/aten/src)
|
||||
target_include_directories(torch PRIVATE ${TH_CPU_INCLUDE})
|
||||
target_include_directories(torch_cpu PRIVATE ${TH_CPU_INCLUDE})
|
||||
|
||||
set(ATen_CPU_INCLUDE
|
||||
${TORCH_ROOT}/aten/src
|
||||
|
|
@ -723,16 +733,16 @@ if (NOT INTERN_BUILD_MOBILE OR NOT BUILD_CAFFE2_MOBILE)
|
|||
|
||||
IF (USE_TBB)
|
||||
list(APPEND ATen_CPU_INCLUDE ${TBB_ROOT_DIR}/include)
|
||||
target_link_libraries(torch PUBLIC tbb)
|
||||
target_link_libraries(torch_cpu PUBLIC tbb)
|
||||
ENDIF()
|
||||
|
||||
|
||||
target_include_directories(torch PRIVATE ${ATen_CPU_INCLUDE})
|
||||
target_include_directories(torch_cpu PRIVATE ${ATen_CPU_INCLUDE})
|
||||
|
||||
target_include_directories(torch PRIVATE
|
||||
target_include_directories(torch_cpu PRIVATE
|
||||
${TORCH_SRC_DIR}/csrc)
|
||||
|
||||
target_include_directories(torch PRIVATE
|
||||
target_include_directories(torch_cpu PRIVATE
|
||||
${TORCH_ROOT}/third_party/miniz-2.0.8)
|
||||
|
||||
|
||||
|
|
@ -818,7 +828,7 @@ endif()
|
|||
|
||||
|
||||
if (NOT NO_API)
|
||||
target_include_directories(torch PUBLIC
|
||||
target_include_directories(torch_cpu PUBLIC
|
||||
$<BUILD_INTERFACE:${TORCH_SRC_DIR}/csrc/api>
|
||||
$<BUILD_INTERFACE:${TORCH_SRC_DIR}/csrc/api/include>)
|
||||
endif()
|
||||
|
|
@ -831,17 +841,31 @@ if(USE_OPENMP AND OPENMP_FOUND)
|
|||
message(STATUS "pytorch is compiling with OpenMP. \n"
|
||||
"OpenMP CXX_FLAGS: ${OpenMP_CXX_FLAGS}. \n"
|
||||
"OpenMP libraries: ${OpenMP_CXX_LIBRARIES}.")
|
||||
target_compile_options(torch INTERFACE ${OpenMP_CXX_FLAGS})
|
||||
target_link_libraries(torch PRIVATE ${OpenMP_CXX_LIBRARIES})
|
||||
target_compile_options(torch_cpu INTERFACE ${OpenMP_CXX_FLAGS})
|
||||
target_link_libraries(torch_cpu PRIVATE ${OpenMP_CXX_LIBRARIES})
|
||||
endif()
|
||||
|
||||
|
||||
if(USE_ROCM)
|
||||
target_compile_definitions(torch_hip PRIVATE
|
||||
USE_ROCM
|
||||
__HIP_PLATFORM_HCC__
|
||||
)
|
||||
# NB: Massive hack. torch/csrc/jit/fuser/codegen.cpp includes
|
||||
# torch/csrc/jit/fuser/cuda/resource_strings.h which changes the
|
||||
# strings depending on if you're __HIP_PLATFORM_HCC__ or not.
|
||||
# But that file is in torch_cpu! So, against all odds, this macro
|
||||
# has to be set on torch_cpu too. I also added it to torch for
|
||||
# better luck
|
||||
target_compile_definitions(torch_cpu PRIVATE
|
||||
USE_ROCM
|
||||
__HIP_PLATFORM_HCC__
|
||||
)
|
||||
target_compile_definitions(torch PRIVATE
|
||||
USE_ROCM
|
||||
__HIP_PLATFORM_HCC__
|
||||
)
|
||||
target_include_directories(torch PRIVATE
|
||||
target_include_directories(torch_hip PRIVATE
|
||||
/opt/rocm/include
|
||||
/opt/rocm/hcc/include
|
||||
/opt/rocm/rocblas/include
|
||||
|
|
@ -851,11 +875,11 @@ endif()
|
|||
|
||||
if (NOT INTERN_BUILD_MOBILE OR BUILD_CAFFE2_MOBILE)
|
||||
caffe2_interface_library(caffe2_protos caffe2_protos_whole)
|
||||
target_link_libraries(torch PRIVATE caffe2_protos_whole)
|
||||
target_link_libraries(torch_cpu PRIVATE caffe2_protos_whole)
|
||||
if (${CAFFE2_LINK_LOCAL_PROTOBUF})
|
||||
target_link_libraries(torch INTERFACE protobuf::libprotobuf)
|
||||
target_link_libraries(torch_cpu INTERFACE protobuf::libprotobuf)
|
||||
else()
|
||||
target_link_libraries(torch PUBLIC protobuf::libprotobuf)
|
||||
target_link_libraries(torch_cpu PUBLIC protobuf::libprotobuf)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
|
|
@ -863,24 +887,30 @@ if (USE_OPENMP AND OPENMP_FOUND)
|
|||
message(STATUS "Caffe2 is compiling with OpenMP. \n"
|
||||
"OpenMP CXX_FLAGS: ${OpenMP_CXX_FLAGS}. \n"
|
||||
"OpenMP libraries: ${OpenMP_CXX_LIBRARIES}.")
|
||||
target_link_libraries(torch PRIVATE ${OpenMP_CXX_LIBRARIES})
|
||||
target_link_libraries(torch_cpu PRIVATE ${OpenMP_CXX_LIBRARIES})
|
||||
endif()
|
||||
|
||||
target_link_libraries(torch PUBLIC c10)
|
||||
target_link_libraries(torch PUBLIC ${Caffe2_PUBLIC_DEPENDENCY_LIBS})
|
||||
target_link_libraries(torch PRIVATE ${Caffe2_DEPENDENCY_LIBS})
|
||||
target_link_libraries(torch PRIVATE ${Caffe2_DEPENDENCY_WHOLE_LINK_LIBS})
|
||||
target_include_directories(torch INTERFACE $<INSTALL_INTERFACE:include>)
|
||||
target_include_directories(torch PRIVATE ${Caffe2_CPU_INCLUDE})
|
||||
target_include_directories(torch SYSTEM PRIVATE "${Caffe2_DEPENDENCY_INCLUDE}")
|
||||
target_link_libraries(torch_cpu PUBLIC c10)
|
||||
target_link_libraries(torch_cpu PUBLIC ${Caffe2_PUBLIC_DEPENDENCY_LIBS})
|
||||
target_link_libraries(torch_cpu PRIVATE ${Caffe2_DEPENDENCY_LIBS})
|
||||
target_link_libraries(torch_cpu PRIVATE ${Caffe2_DEPENDENCY_WHOLE_LINK_LIBS})
|
||||
target_include_directories(torch_cpu INTERFACE $<INSTALL_INTERFACE:include>)
|
||||
target_include_directories(torch_cpu PRIVATE ${Caffe2_CPU_INCLUDE})
|
||||
target_include_directories(torch_cpu SYSTEM PRIVATE "${Caffe2_DEPENDENCY_INCLUDE}")
|
||||
# Set standard properties on the target
|
||||
torch_set_target_props(torch)
|
||||
torch_set_target_props(torch_cpu)
|
||||
|
||||
|
||||
target_compile_options(torch PRIVATE "-DCAFFE2_BUILD_MAIN_LIB")
|
||||
# NB: This must be target_compile_definitions, not target_compile_options,
|
||||
# as the latter is not respected by nvcc
|
||||
target_compile_definitions(torch PRIVATE "-DCAFFE2_BUILD_MAIN_LIB")
|
||||
target_compile_options(torch_cpu PRIVATE "-DCAFFE2_BUILD_MAIN_LIB")
|
||||
if(USE_CUDA)
|
||||
target_compile_options(torch_cuda PRIVATE "-DTORCH_CUDA_BUILD_MAIN_LIB")
|
||||
# NB: This must be target_compile_definitions, not target_compile_options,
|
||||
# as the latter is not respected by nvcc
|
||||
target_compile_definitions(torch_cuda PRIVATE "-DTORCH_CUDA_BUILD_MAIN_LIB")
|
||||
elseif(USE_ROCM)
|
||||
target_compile_options(torch_hip PRIVATE "-DTORCH_HIP_BUILD_MAIN_LIB")
|
||||
target_compile_definitions(torch_hip PRIVATE "-DTORCH_HIP_BUILD_MAIN_LIB")
|
||||
endif()
|
||||
|
||||
|
||||
# ATen parallelism settings
|
||||
|
|
@ -895,21 +925,21 @@ endif()
|
|||
|
||||
message(STATUS "Using ATen parallel backend: ${ATEN_THREADING}")
|
||||
if ("${ATEN_THREADING}" STREQUAL "OMP")
|
||||
target_compile_definitions(torch PUBLIC "-DAT_PARALLEL_OPENMP=1")
|
||||
target_compile_definitions(torch_cpu PUBLIC "-DAT_PARALLEL_OPENMP=1")
|
||||
elseif ("${ATEN_THREADING}" STREQUAL "NATIVE")
|
||||
target_compile_definitions(torch PUBLIC "-DAT_PARALLEL_NATIVE=1")
|
||||
target_compile_definitions(torch_cpu PUBLIC "-DAT_PARALLEL_NATIVE=1")
|
||||
elseif ("${ATEN_THREADING}" STREQUAL "TBB")
|
||||
if (NOT USE_TBB)
|
||||
message(FATAL_ERROR "Using TBB backend but USE_TBB is off")
|
||||
endif()
|
||||
target_compile_definitions(torch PUBLIC "-DAT_PARALLEL_NATIVE_TBB=1")
|
||||
target_compile_definitions(torch_cpu PUBLIC "-DAT_PARALLEL_NATIVE_TBB=1")
|
||||
else()
|
||||
message(FATAL_ERROR "Unknown ATen parallel backend: ${ATEN_THREADING}")
|
||||
endif()
|
||||
set(EXPERIMENTAL_SINGLE_THREAD_POOL "0" CACHE STRING
|
||||
"Experimental option to use a single thread pool for inter- and intra-op parallelism")
|
||||
if ("${EXPERIMENTAL_SINGLE_THREAD_POOL}")
|
||||
target_compile_definitions(torch PUBLIC "-DAT_EXPERIMENTAL_SINGLE_THREAD_POOL=1")
|
||||
target_compile_definitions(torch_cpu PUBLIC "-DAT_EXPERIMENTAL_SINGLE_THREAD_POOL=1")
|
||||
endif()
|
||||
|
||||
if (MSVC AND NOT BUILD_SHARED_LIBS)
|
||||
|
|
@ -979,18 +1009,39 @@ if (MSVC AND NOT BUILD_SHARED_LIBS)
|
|||
#
|
||||
# NB: This must be target_compile_definitions, not target_compile_options,
|
||||
# as the latter is not respected by nvcc
|
||||
target_compile_definitions(torch PUBLIC "AT_CORE_STATIC_WINDOWS=1")
|
||||
target_compile_definitions(torch_cpu PUBLIC "AT_CORE_STATIC_WINDOWS=1")
|
||||
endif()
|
||||
if (MSVC AND BUILD_SHARED_LIBS)
|
||||
# ONNX is linked statically and needs to be exported from this library
|
||||
# to be used externally. Make sure that references match the export.
|
||||
target_compile_options(torch PRIVATE "-DONNX_BUILD_MAIN_LIB")
|
||||
target_compile_options(torch_cpu PRIVATE "-DONNX_BUILD_MAIN_LIB")
|
||||
endif()
|
||||
|
||||
install(TARGETS torch EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
|
||||
caffe2_interface_library(torch_cpu torch_cpu_library)
|
||||
|
||||
if (USE_CUDA)
|
||||
caffe2_interface_library(torch_cuda torch_cuda_library)
|
||||
elseif (USE_ROCM)
|
||||
caffe2_interface_library(torch_hip torch_hip_library)
|
||||
endif()
|
||||
|
||||
caffe2_interface_library(torch torch_library)
|
||||
|
||||
install(TARGETS torch_cpu torch_cpu_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
|
||||
if (USE_CUDA)
|
||||
install(TARGETS torch_cuda torch_cuda_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
|
||||
elseif (USE_ROCM)
|
||||
install(TARGETS torch_hip torch_hip_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
|
||||
endif()
|
||||
install(TARGETS torch torch_library EXPORT Caffe2Targets DESTINATION "${TORCH_INSTALL_LIB_DIR}")
|
||||
|
||||
target_link_libraries(torch PUBLIC torch_cpu_library)
|
||||
if(USE_CUDA)
|
||||
target_link_libraries(torch PUBLIC torch_cuda_library)
|
||||
elseif(USE_ROCM)
|
||||
target_link_libraries(torch PUBLIC torch_hip_library)
|
||||
endif()
|
||||
|
||||
list(APPEND Caffe2_MAIN_LIBS torch_library)
|
||||
if (USE_TBB)
|
||||
list(APPEND Caffe2_MAIN_LIBS tbb)
|
||||
|
|
@ -998,28 +1049,33 @@ endif()
|
|||
|
||||
# Install PDB files for MSVC builds
|
||||
if (MSVC AND BUILD_SHARED_LIBS)
|
||||
install(FILES $<TARGET_PDB_FILE:torch> DESTINATION "${TORCH_INSTALL_LIB_DIR}" OPTIONAL)
|
||||
install(FILES $<TARGET_PDB_FILE:torch_cpu> DESTINATION "${TORCH_INSTALL_LIB_DIR}" OPTIONAL)
|
||||
if(USE_CUDA)
|
||||
install(FILES $<TARGET_PDB_FILE:torch_cuda> DESTINATION "${TORCH_INSTALL_LIB_DIR}" OPTIONAL)
|
||||
elseif(USE_ROCM)
|
||||
install(FILES $<TARGET_PDB_FILE:torch_hip> DESTINATION "${TORCH_INSTALL_LIB_DIR}" OPTIONAL)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
# ---[ CUDA library.
|
||||
if(USE_CUDA)
|
||||
|
||||
target_link_libraries(torch INTERFACE torch::cudart)
|
||||
target_link_libraries(torch PUBLIC c10_cuda)
|
||||
target_link_libraries(torch_cuda INTERFACE torch::cudart)
|
||||
target_link_libraries(torch_cuda PUBLIC c10_cuda)
|
||||
|
||||
target_link_libraries(torch PUBLIC ${TORCH_CUDA_LIBRARIES})
|
||||
target_link_libraries(torch_cuda PUBLIC ${TORCH_CUDA_LIBRARIES})
|
||||
|
||||
target_include_directories(
|
||||
torch INTERFACE $<INSTALL_INTERFACE:include>)
|
||||
torch_cuda INTERFACE $<INSTALL_INTERFACE:include>)
|
||||
target_include_directories(
|
||||
torch PRIVATE ${Caffe2_GPU_INCLUDE})
|
||||
torch_cuda PRIVATE ${Caffe2_GPU_INCLUDE})
|
||||
target_link_libraries(
|
||||
torch PRIVATE ${Caffe2_CUDA_DEPENDENCY_LIBS})
|
||||
torch_cuda PRIVATE ${Caffe2_CUDA_DEPENDENCY_LIBS})
|
||||
|
||||
# These public dependencies must go after the previous dependencies, as the
|
||||
# order of the libraries in the linker call matters here when statically
|
||||
# linking; libculibos and cublas must be last.
|
||||
target_link_libraries(torch PUBLIC ${Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS})
|
||||
target_link_libraries(torch_cuda PUBLIC torch_cpu_library ${Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS})
|
||||
|
||||
|
||||
endif()
|
||||
|
|
@ -1040,8 +1096,8 @@ if(USE_ROCM)
|
|||
hip_include_directories(${Caffe2_HIP_INCLUDE})
|
||||
|
||||
# Since PyTorch files contain HIP headers, these flags are required for the necessary definitions to be added.
|
||||
target_compile_options(torch PRIVATE ${HIP_CXX_FLAGS})
|
||||
target_link_libraries(torch PUBLIC c10_hip)
|
||||
target_compile_options(torch_hip PUBLIC ${HIP_CXX_FLAGS}) # experiment
|
||||
target_link_libraries(torch_hip PUBLIC c10_hip)
|
||||
|
||||
if(NOT INTERN_BUILD_MOBILE)
|
||||
# TODO: Cut this over to ATEN_HIP_FILES_GEN_LIB. At the moment, we
|
||||
|
|
@ -1049,13 +1105,13 @@ if(USE_ROCM)
|
|||
# NB: This dependency must be PRIVATE, because we don't install
|
||||
# ATEN_CUDA_FILES_GEN_LIB (it's a synthetic target just to get the
|
||||
# correct dependency from generated files.)
|
||||
target_link_libraries(torch PRIVATE ATEN_CUDA_FILES_GEN_LIB)
|
||||
target_link_libraries(torch_hip PRIVATE ATEN_CUDA_FILES_GEN_LIB)
|
||||
endif()
|
||||
target_link_libraries(torch PUBLIC ${Caffe2_HIP_DEPENDENCY_LIBS})
|
||||
target_link_libraries(torch_hip PUBLIC torch_cpu_library ${Caffe2_HIP_DEPENDENCY_LIBS})
|
||||
|
||||
# Since PyTorch files contain HIP headers, this is also needed to capture the includes.
|
||||
target_include_directories(torch PRIVATE ${Caffe2_HIP_INCLUDE})
|
||||
target_include_directories(torch INTERFACE $<INSTALL_INTERFACE:include>)
|
||||
target_include_directories(torch_hip PRIVATE ${Caffe2_HIP_INCLUDE})
|
||||
target_include_directories(torch_hip INTERFACE $<INSTALL_INTERFACE:include>)
|
||||
endif()
|
||||
|
||||
# ---[ Test binaries.
|
||||
|
|
|
|||
|
|
@ -29,40 +29,11 @@
|
|||
#include "c10/cuda/CUDAMathCompat.h"
|
||||
#include <c10/cuda/CUDAGuard.h>
|
||||
|
||||
// Defines CAFFE2_CUDA_EXPORT and CAFFE2_CUDA_IMPORT. On Windows, this
|
||||
// corresponds to different declarations (dllexport and dllimport). On
|
||||
// Linux/Mac, it just resolves to the same "default visibility" setting.
|
||||
#if defined(_MSC_VER)
|
||||
#if defined(CAFFE2_BUILD_SHARED_LIBS)
|
||||
#define CAFFE2_CUDA_EXPORT __declspec(dllexport)
|
||||
#define CAFFE2_CUDA_IMPORT __declspec(dllimport)
|
||||
#else
|
||||
#define CAFFE2_CUDA_EXPORT
|
||||
#define CAFFE2_CUDA_IMPORT
|
||||
#endif
|
||||
#else
|
||||
#if defined(__GNUC__)
|
||||
#define CAFFE2_CUDA_EXPORT __attribute__((__visibility__("default")))
|
||||
#else
|
||||
#define CAFFE2_CUDA_EXPORT
|
||||
#endif
|
||||
#define CAFFE2_CUDA_IMPORT CAFFE2_CUDA_EXPORT
|
||||
#endif
|
||||
#define CAFFE2_CUDA_EXPORT C10_EXPORT
|
||||
#define CAFFE2_CUDA_API TORCH_CUDA_API
|
||||
|
||||
// CAFFE2_CUDA_API is a macro that, depends on whether you are building the
|
||||
// main caffe2 library or not, resolves to either CAFFE2_CUDA_EXPORT or
|
||||
// CAFFE2_CUDA_IMPORT.
|
||||
//
|
||||
// This is used in e.g. Caffe2's protobuf files: when building the main library,
|
||||
// it is defined as CAFFE2_CUDA_EXPORT to fix a Windows global-variable-in-dll
|
||||
// issue, and for anyone dependent on Caffe2 it will be defined as
|
||||
// CAFFE2_CUDA_IMPORT.
|
||||
|
||||
#ifdef CAFFE2_BUILD_MAIN_LIB
|
||||
#define CAFFE2_CUDA_API CAFFE2_CUDA_EXPORT
|
||||
#else
|
||||
#define CAFFE2_CUDA_API CAFFE2_CUDA_IMPORT
|
||||
#endif
|
||||
#define CAFFE2_HIP_EXPORT C10_EXPORT
|
||||
#define CAFFE2_HIP_API TORCH_HIP_API
|
||||
|
||||
// This is a macro defined for cuda fp16 support. In default, cuda fp16 is
|
||||
// supported by NVCC 7.5, but it is also included in the Tegra X1 platform with
|
||||
|
|
|
|||
|
|
@ -118,7 +118,6 @@ include ("${CMAKE_CURRENT_LIST_DIR}/Caffe2Targets.cmake")
|
|||
# Interface libraries, that allows one to build proper link flags.
|
||||
# We will also define a helper variable, Caffe2_MAIN_LIBS, that resolves to
|
||||
# the main caffe2 libraries in cases of cuda presence / absence.
|
||||
caffe2_interface_library(torch torch_library)
|
||||
set(Caffe2_MAIN_LIBS torch_library)
|
||||
|
||||
# include directory.
|
||||
|
|
|
|||
|
|
@ -24,7 +24,7 @@ Pod::Spec.new do |s|
|
|||
end
|
||||
s.user_target_xcconfig = {
|
||||
'HEADER_SEARCH_PATHS' => '$(inherited) "$(PODS_ROOT)/LibTorch/install/include/"',
|
||||
'OTHER_LDFLAGS' => '-force_load "$(PODS_ROOT)/LibTorch/install/lib/libtorch.a"',
|
||||
'OTHER_LDFLAGS' => '-force_load "$(PODS_ROOT)/LibTorch/install/lib/libtorch.a" -force_load "$(PODS_ROOT)/LibTorch/install/lib/libtorch_cpu.a"',
|
||||
'CLANG_CXX_LANGUAGE_STANDARD' => 'c++14',
|
||||
'CLANG_CXX_LIBRARY' => 'libc++'
|
||||
}
|
||||
|
|
|
|||
|
|
@ -63,7 +63,7 @@ targets.each do |target|
|
|||
target.resources_build_phase.add_file_reference(config_file_ref, true)
|
||||
end
|
||||
puts "Linking static libraries..."
|
||||
libs = ['libc10.a', 'libclog.a', 'libnnpack.a', 'libeigen_blas.a', 'libcpuinfo.a', 'libpytorch_qnnpack.a', 'libtorch.a']
|
||||
libs = ['libc10.a', 'libclog.a', 'libnnpack.a', 'libeigen_blas.a', 'libcpuinfo.a', 'libpytorch_qnnpack.a', 'libtorch_cpu.a', 'libtorch.a']
|
||||
targets.each do |target|
|
||||
target.frameworks_build_phases.clear
|
||||
for lib in libs do
|
||||
|
|
|
|||
|
|
@ -51,7 +51,7 @@ end
|
|||
|
||||
# link static libraries
|
||||
target.frameworks_build_phases.clear
|
||||
libs = ['libc10.a', 'libclog.a', 'libnnpack.a', 'libeigen_blas.a', 'libcpuinfo.a', 'libpytorch_qnnpack.a', 'libtorch.a']
|
||||
libs = ['libc10.a', 'libclog.a', 'libnnpack.a', 'libeigen_blas.a', 'libcpuinfo.a', 'libpytorch_qnnpack.a', 'libtorch_cpu.a', 'libtorch.a']
|
||||
for lib in libs do
|
||||
path = "#{install_path}/lib/#{lib}"
|
||||
if File.exist?(path)
|
||||
|
|
|
|||
|
|
@ -458,6 +458,7 @@ def CppExtension(name, sources, *args, **kwargs):
|
|||
libraries = kwargs.get('libraries', [])
|
||||
libraries.append('c10')
|
||||
libraries.append('torch')
|
||||
libraries.append('torch_cpu')
|
||||
libraries.append('torch_python')
|
||||
libraries.append('_C')
|
||||
kwargs['libraries'] = libraries
|
||||
|
|
@ -503,6 +504,8 @@ def CUDAExtension(name, sources, *args, **kwargs):
|
|||
if IS_WINDOWS:
|
||||
libraries.append('c10')
|
||||
libraries.append('c10_cuda')
|
||||
libraries.append('torch_cpu')
|
||||
libraries.append('torch_cuda')
|
||||
libraries.append('torch')
|
||||
libraries.append('torch_python')
|
||||
libraries.append('_C')
|
||||
|
|
@ -943,6 +946,11 @@ def _prepare_ldflags(extra_ldflags, with_cuda, verbose):
|
|||
lib_path = os.path.join(torch_path, 'lib')
|
||||
|
||||
extra_ldflags.append('c10.lib')
|
||||
if with_cuda:
|
||||
extra_ldflags.append('c10_cuda.lib')
|
||||
extra_ldflags.append('torch_cpu.lib')
|
||||
if with_cuda:
|
||||
extra_ldflags.append('torch_cuda.lib')
|
||||
extra_ldflags.append('torch.lib')
|
||||
extra_ldflags.append('torch_python.lib')
|
||||
extra_ldflags.append('_C.lib')
|
||||
|
|
|
|||
Loading…
Reference in New Issue
Block a user