frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

Revert "[Reland2] Update NVTX to NVTX3 (#109843)"

Browse Source 
This reverts commit dcb486232d3eb61024ad9e76cca367c60019c84c.

Reverted https://github.com/pytorch/pytorch/pull/109843 on behalf of https://github.com/atalman due to Diff broke internal builds and tests ([comment](https://github.com/pytorch/pytorch/pull/109843#issuecomment-1841105398))
This commit is contained in:
PyTorch MergeBot
2023-12-05 16:10:20 +00:00
parent e06bff8bbe
commit ee96399bb4
16 changed files with 56 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
.ci/pytorch/win-test-helpers/build_pytorch.bat
View File

@ -66,6 +66,13 @@ set CUDA_PATH_V%VERSION_SUFFIX%=%CUDA_PATH%
set CUDNN_LIB_DIR=%CUDA_PATH%\lib\x64
set CUDA_TOOLKIT_ROOT_DIR=%CUDA_PATH%
set CUDNN_ROOT_DIR=%CUDA_PATH%
set NVTOOLSEXT_PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt
set PATH=%CUDA_PATH%\bin;%CUDA_PATH%\libnvvp;%PATH%
set CUDNN_LIB_DIR=%CUDA_PATH%\lib\x64
set CUDA_TOOLKIT_ROOT_DIR=%CUDA_PATH%
set CUDNN_ROOT_DIR=%CUDA_PATH%
set NVTOOLSEXT_PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt
set PATH=%CUDA_PATH%\bin;%CUDA_PATH%\libnvvp;%PATH%
:cuda_build_end

1
.ci/pytorch/win-test-helpers/setup_pytorch_env.bat
View File

@ -40,6 +40,7 @@ set CUDA_PATH_V%VERSION_SUFFIX%=%CUDA_PATH%
set CUDNN_LIB_DIR=%CUDA_PATH%\lib\x64
set CUDA_TOOLKIT_ROOT_DIR=%CUDA_PATH%
set CUDNN_ROOT_DIR=%CUDA_PATH%
set NVTOOLSEXT_PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt
set PATH=%CUDA_PATH%\bin;%CUDA_PATH%\libnvvp;%PATH%
set NUMBAPRO_CUDALIB=%CUDA_PATH%\bin
set NUMBAPRO_LIBDEVICE=%CUDA_PATH%\nvvm\libdevice

2
.ci/pytorch/win-test-helpers/test_custom_backend.bat
View File

@ -31,6 +31,6 @@ if ERRORLEVEL 1 exit /b 1
:: Run tests C++-side and load the exported script module.
cd build
set PATH=%TMP_DIR_WIN%\build\torch\lib;%PATH%
set PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt\bin\x64;%TMP_DIR_WIN%\build\torch\lib;%PATH%
test_custom_backend.exe model.pt
if ERRORLEVEL 1 exit /b 1

2
.ci/pytorch/win-test-helpers/test_custom_script_ops.bat
View File

@ -31,6 +31,6 @@ if ERRORLEVEL 1 exit /b 1
:: Run tests C++-side and load the exported script module.
cd build
set PATH=%TMP_DIR_WIN%\build\torch\lib;%PATH%
set PATH=C:\Program Files\NVIDIA Corporation\NvToolsExt\bin\x64;%TMP_DIR_WIN%\build\torch\lib;%PATH%
test_custom_ops.exe model.pt
if ERRORLEVEL 1 exit /b 1

3
.gitmodules vendored
View File

@ -149,6 +149,3 @@
[submodule "third_party/mimalloc"]
path = third_party/mimalloc
url = https://github.com/microsoft/mimalloc.git
[submodule "third_party/NVTX"]
path = third_party/NVTX
url = https://github.com/NVIDIA/NVTX.git

5
caffe2/CMakeLists.txt
View File

@ -1548,8 +1548,7 @@ if(USE_CUDA)
target_link_libraries(torch_cpu PRIVATE torch::cudart)
endif()
target_link_libraries(torch_cuda INTERFACE torch::cudart)
target_link_libraries(torch_cuda PUBLIC c10_cuda)
target_link_libraries(torch_cuda PRIVATE torch::nvtoolsext)
target_link_libraries(torch_cuda PUBLIC c10_cuda torch::nvtoolsext)
target_include_directories(
torch_cuda INTERFACE $<INSTALL_INTERFACE:include>)
@ -1606,7 +1605,7 @@ if(BUILD_SHARED_LIBS)
# not find them, because they're usually in non-standard locations)
if(USE_CUDA)
target_link_libraries(torch_global_deps ${Caffe2_PUBLIC_CUDA_DEPENDENCY_LIBS})
target_link_libraries(torch_global_deps torch::cudart)
target_link_libraries(torch_global_deps torch::cudart torch::nvtoolsext)
endif()
if(USE_TBB)
target_link_libraries(torch_global_deps TBB::tbb)

9
caffe2/python/__init__.py
View File

@ -26,6 +26,13 @@ if sys.platform == "win32":
th_root = os.path.join(os.path.dirname(os.path.dirname(os.path.dirname(__file__))), 'torch')
th_dll_path = os.path.join(th_root, 'lib')
if not os.path.exists(os.path.join(th_dll_path, 'nvToolsExt64_1.dll')) and \
not os.path.exists(os.path.join(py_dll_path, 'nvToolsExt64_1.dll')):
nvtoolsext_dll_path = os.path.join(
os.getenv('NVTOOLSEXT_PATH', 'C:\\Program Files\\NVIDIA Corporation\\NvToolsExt'), 'bin', 'x64')
else:
nvtoolsext_dll_path = ''
import importlib.util
import glob
spec = importlib.util.spec_from_file_location('torch_version', os.path.join(th_root, 'version.py'))
@ -43,7 +50,7 @@ if sys.platform == "win32":
import ctypes
kernel32 = ctypes.WinDLL('kernel32.dll', use_last_error=True)
dll_paths = list(filter(os.path.exists, [th_dll_path, py_dll_path, cuda_path]))
dll_paths = list(filter(os.path.exists, [th_dll_path, py_dll_path, nvtoolsext_dll_path, cuda_path]))
with_load_library_flags = hasattr(kernel32, 'AddDllDirectory')
prev_error_mode = kernel32.SetErrorMode(0x0001)

14
cmake/TorchConfig.cmake.in
View File

@ -129,18 +129,30 @@ endif()
if(@USE_CUDA@)
if(MSVC)
set(TORCH_CUDA_LIBRARIES ${CUDA_LIBRARIES})
if(NOT NVTOOLEXT_HOME)
set(NVTOOLEXT_HOME "C:/Program Files/NVIDIA Corporation/NvToolsExt")
endif()
if(DEFINED ENV{NVTOOLSEXT_PATH})
set(NVTOOLEXT_HOME $ENV{NVTOOLSEXT_PATH})
endif()
set(TORCH_CUDA_LIBRARIES
${NVTOOLEXT_HOME}/lib/x64/nvToolsExt64_1.lib
${CUDA_LIBRARIES})
list(APPEND TORCH_INCLUDE_DIRS ${NVTOOLEXT_HOME}/include)
find_library(CAFFE2_NVRTC_LIBRARY caffe2_nvrtc PATHS "${TORCH_INSTALL_PREFIX}/lib")
list(APPEND TORCH_CUDA_LIBRARIES ${CAFFE2_NVRTC_LIBRARY})
elseif(APPLE)
set(TORCH_CUDA_LIBRARIES
${CUDA_TOOLKIT_ROOT_DIR}/lib/libcudart.dylib
${CUDA_TOOLKIT_ROOT_DIR}/lib/libnvrtc.dylib
${CUDA_TOOLKIT_ROOT_DIR}/lib/libnvToolsExt.dylib
${CUDA_LIBRARIES})
else()
find_library(LIBNVTOOLSEXT libnvToolsExt.so PATHS ${CUDA_TOOLKIT_ROOT_DIR}/lib64/)
set(TORCH_CUDA_LIBRARIES
${CUDA_CUDA_LIB}
${CUDA_NVRTC_LIB}
${LIBNVTOOLSEXT}
${CUDA_LIBRARIES})
endif()
if(@BUILD_SHARED_LIBS@)

17
cmake/public/cuda.cmake
View File

@ -66,6 +66,10 @@ if(NOT CMAKE_CUDA_COMPILER_VERSION VERSION_EQUAL CUDAToolkit_VERSION)
"V${CUDAToolkit_VERSION} in '${CUDAToolkit_INCLUDE_DIRS}'")
endif()
if(NOT TARGET CUDA::nvToolsExt)
message(FATAL_ERROR "Failed to find nvToolsExt")
endif()
message(STATUS "Caffe2: CUDA detected: " ${CUDA_VERSION})
message(STATUS "Caffe2: CUDA nvcc is: " ${CUDA_NVCC_EXECUTABLE})
message(STATUS "Caffe2: CUDA toolkit directory: " ${CUDA_TOOLKIT_ROOT_DIR})
@ -210,15 +214,10 @@ else()
endif()
# nvToolsExt
find_path(nvtx3_dir NAMES nvtx3 PATHS "${CUDA_INCLUDE_DIRS}" "${PROJECT_SOURCE_DIR}/third_party/NVTX/c/include" NO_DEFAULT_PATH)
find_package_handle_standard_args(nvtx3 DEFAULT_MSG nvtx3_dir)
if(nvtx3_FOUND)
add_library(torch::nvtoolsext INTERFACE IMPORTED)
target_include_directories(torch::nvtoolsext INTERFACE "${nvtx3_dir}")
else()
message(WARNING "Cannot find NVTX3")
endif()
add_library(torch::nvtoolsext INTERFACE IMPORTED)
set_property(
TARGET torch::nvtoolsext PROPERTY INTERFACE_LINK_LIBRARIES
CUDA::nvToolsExt)
# cublas
add_library(caffe2::cublas INTERFACE IMPORTED)

2
setup.py
View File

@ -189,6 +189,8 @@
# NCCL_INCLUDE_DIR
# specify where nccl is installed
#
# NVTOOLSEXT_PATH (Windows only)
# specify where nvtoolsext is installed
#
# ACL_ROOT_DIR
# specify where Compute Library is installed

1
third_party/NVTX vendored

Submodule third_party/NVTX deleted from e170594ac7

1
torch/CMakeLists.txt
View File

@ -130,6 +130,7 @@ if(USE_CUDA)
list(APPEND TORCH_PYTHON_LINK_LIBRARIES torch::cudnn)
list(APPEND TORCH_PYTHON_COMPILE_DEFINITIONS USE_CUDNN)
endif()
list(APPEND TORCH_PYTHON_LINK_LIBRARIES torch::nvtoolsext)
endif()

9
torch/__init__.py
View File

@ -80,6 +80,12 @@ if sys.platform == 'win32':
dll_paths = list(filter(os.path.exists, [th_dll_path, py_dll_path, base_py_dll_path]))
if all(not os.path.exists(os.path.join(p, 'nvToolsExt64_1.dll')) for p in dll_paths):
nvtoolsext_dll_path = os.path.join(
os.getenv('NVTOOLSEXT_PATH', os.path.join(pfiles_path, 'NVIDIA Corporation', 'NvToolsExt')), 'bin', 'x64')
else:
nvtoolsext_dll_path = ''
from .version import cuda as cuda_version
import glob
if cuda_version and all(not glob.glob(os.path.join(p, 'cudart64*.dll')) for p in dll_paths):
@ -90,7 +96,7 @@ if sys.platform == 'win32':
else:
cuda_path = ''
dll_paths.extend(filter(os.path.exists, [cuda_path]))
dll_paths.extend(filter(os.path.exists, [nvtoolsext_dll_path, cuda_path]))
kernel32 = ctypes.WinDLL('kernel32.dll', use_last_error=True)
with_load_library_flags = hasattr(kernel32, 'AddDllDirectory')
@ -182,6 +188,7 @@ def _load_global_deps() -> None:
'cusolver': 'libcusolver.so.*[0-9]',
'cusparse': 'libcusparse.so.*[0-9]',
'nccl': 'libnccl.so.*[0-9]',
'nvtx': 'libnvToolsExt.so.*[0-9]',
}
is_cuda_lib_err = [lib for lib in cuda_libs.values() if(lib.split('.')[0] in err.args[0])]
if not is_cuda_lib_err:

4
torch/csrc/cuda/shared/nvtx.cpp
View File

@ -1,7 +1,7 @@
#ifdef _WIN32
#include <wchar.h> // _wgetenv for nvtx
#endif
#include <nvtx3/nvToolsExt.h>
#include <nvToolsExt.h>
#include <torch/csrc/utils/pybind.h>
namespace torch::cuda::shared {
@ -9,7 +9,7 @@ namespace torch::cuda::shared {
void initNvtxBindings(PyObject* module) {
auto m = py::handle(module).cast<py::module>();
auto nvtx = m.def_submodule("_nvtx", "nvtx3 bindings");
auto nvtx = m.def_submodule("_nvtx", "libNvToolsExt.so bindings");
nvtx.def("rangePushA", nvtxRangePushA);
nvtx.def("rangePop", nvtxRangePop);
nvtx.def("rangeStartA", nvtxRangeStartA);

2
torch/csrc/profiler/stubs/cuda.cpp
View File

@ -1,6 +1,6 @@
#include <sstream>
#include <nvtx3/nvToolsExt.h>
#include <nvToolsExt.h>
#include <c10/cuda/CUDAGuard.h>
#include <c10/util/ApproximateClock.h>

2
torch/utils/hipify/cuda_to_hip_mappings.py
View File

@ -647,7 +647,7 @@ CUDA_INCLUDE_MAP = collections.OrderedDict(
("cub/device/device_reduce.cuh", ("hipcub/hipcub.hpp", CONV_INCLUDE, API_BLAS)),
("cub/device/device_scan.cuh", ("hipcub/hipcub.hpp", CONV_INCLUDE, API_BLAS)),
("cub/device/device_select.cuh", ("hipcub/hipcub.hpp", CONV_INCLUDE, API_BLAS)),
("nvtx3/nvToolsExt.h", ("roctracer/roctx.h", CONV_INCLUDE, API_ROCTX)),
("nvToolsExt.h", ("roctracer/roctx.h", CONV_INCLUDE, API_ROCTX)),
("nvml.h", ("rocm_smi/rocm_smi.h", CONV_INCLUDE, API_ROCMSMI)),
]
)