[pytorch source read][1-2]setup.py

2024年1月27日 198次阅读 来源: Nick.Zxx

这篇文章是接着
上一篇文章的，由于知乎字数的限制我被迫分成了两个文章

################################################################################
# Configure compile flags
################################################################################

include_dirs = []
library_dirs = []
extra_link_args = []

if IS_WINDOWS:
    extra_compile_args = ['/Z7', '/EHa', '/DNOMINMAX', '/wd4267', '/wd4251', '/wd4522',
                          '/wd4522', '/wd4838', '/wd4305', '/wd4244', '/wd4190',
                          '/wd4101', '/wd4996', '/wd4275'
                          # /Z7 turns on symbolic debugging information in .obj files
                          # /EHa is about native C++ catch support for asynchronous
                          # structured exception handling (SEH)
                          # /DNOMINMAX removes builtin min/max functions
                          # /wdXXXX disables warning no. XXXX
                          ]
    if sys.version_info[0] == 2:
        # /bigobj increases number of sections in .obj file, which is needed to link
        # against libaries in Python 2.7 under Windows
        extra_compile_args.append('/bigobj')
else:
    extra_compile_args = [
        '-std=c++11',
        '-Wall',
        '-Wextra',
        '-Wno-unused-parameter',
        '-Wno-missing-field-initializers',
        '-Wno-write-strings',
        '-Wno-zero-length-array',
        # This is required for Python 2 declarations that are deprecated in 3.
        '-Wno-deprecated-declarations',
        # Python 2.6 requires -fno-strict-aliasing, see
        # http://legacy.python.org/dev/peps/pep-3123/
        # We also depend on it in our code (even Python 3).
        '-fno-strict-aliasing',
        # Clang has an unfixed bug leading to spurious missing
        # braces warnings, see
        # https://bugs.llvm.org/show_bug.cgi?id=21629
        '-Wno-missing-braces'
    ]
    if check_env_flag('WERROR'):
        extra_compile_args.append('-Werror')

cwd = os.path.dirname(os.path.abspath(__file__))
lib_path = os.path.join(cwd, "torch", "lib")
third_party_path = os.path.join(cwd, "third_party")


tmp_install_path = lib_path + "/tmp_install"
include_dirs += [
    cwd,
    os.path.join(cwd, "torch", "csrc"),
    third_party_path + "/pybind11/include",
    tmp_install_path + "/include",
    tmp_install_path + "/include/TH",
    tmp_install_path + "/include/THNN",
    tmp_install_path + "/include/ATen",
]

library_dirs.append(lib_path)

# we specify exact lib names to avoid conflict with lua-torch installs
ATEN_LIB = os.path.join(lib_path, 'libATen.so')
THD_LIB = os.path.join(lib_path, 'libTHD.a')
NCCL_LIB = os.path.join(lib_path, 'libnccl.so.1')

# static library only
NANOPB_STATIC_LIB = os.path.join(lib_path, 'libprotobuf-nanopb.a')

if IS_DARWIN:
    ATEN_LIB = os.path.join(lib_path, 'libATen.dylib')
    NCCL_LIB = os.path.join(lib_path, 'libnccl.1.dylib')

if IS_WINDOWS:
    ATEN_LIB = os.path.join(lib_path, 'ATen.lib')
    if DEBUG:
        NANOPB_STATIC_LIB = os.path.join(lib_path, 'protobuf-nanopbd.lib')
    else:
        NANOPB_STATIC_LIB = os.path.join(lib_path, 'protobuf-nanopb.lib')

这一部分比较荣昌, 就是对编译的configure以及一些库的位置的说明

普及一下.so文件和.a文件

.so文件是 动态链接库, 动态链接所调用的函数代码并没有被拷贝到应用程序的可执行文件中去，而是仅仅在其中加入了所调用函数的描述信息（往往是一些重定位信息），仅当应用程序被装入内存开始运行时，在操作系统的管理下，才在应用程序与相应的.so之间建立链接关系

.a文件是 静态链接库文件, 指把要调用的函数或者过程链接到可执行文件中，成为可执行文件的一部分。当多个程序都调用相同函数时，内存中就会存在这个函数的多个拷贝，这样就浪费了宝贵的内存资源

可以这么理解.a文件是多个.o文件的组合

main_compile_args = ['-D_THP_CORE']
main_libraries = ['shm']
main_link_args = [ATEN_LIB, NANOPB_STATIC_LIB]
main_sources = [
    "torch/csrc/PtrWrapper.cpp",
    "torch/csrc/Module.cpp",
    "torch/csrc/Generator.cpp",
    "torch/csrc/Size.cpp",
    "torch/csrc/Dtype.cpp",
    "torch/csrc/Device.cpp",
    "torch/csrc/Exceptions.cpp",
    "torch/csrc/Layout.cpp",
    "torch/csrc/Storage.cpp",
    "torch/csrc/DataLoader.cpp",
    "torch/csrc/DynamicTypes.cpp",
    "torch/csrc/assertions.cpp",
    "torch/csrc/byte_order.cpp",
    "torch/csrc/torch.cpp",
    "torch/csrc/utils.cpp",
    "torch/csrc/utils/cuda_lazy_init.cpp",
    "torch/csrc/utils/device.cpp",
    "torch/csrc/utils/invalid_arguments.cpp",
    "torch/csrc/utils/object_ptr.cpp",
    "torch/csrc/utils/python_arg_parser.cpp",
    "torch/csrc/utils/tensor_list.cpp",
    "torch/csrc/utils/tensor_new.cpp",
    "torch/csrc/utils/tensor_numpy.cpp",
    "torch/csrc/utils/tensor_dtypes.cpp",
    "torch/csrc/utils/tensor_layouts.cpp",
    "torch/csrc/utils/tensor_types.cpp",
    "torch/csrc/utils/tuple_parser.cpp",
    "torch/csrc/utils/tensor_apply.cpp",
    "torch/csrc/utils/tensor_conversion_dispatch.cpp",
    "torch/csrc/utils/tensor_flatten.cpp",
    "torch/csrc/utils/variadic.cpp",
    "torch/csrc/allocators.cpp",
    "torch/csrc/serialization.cpp",
    "torch/csrc/jit/init.cpp",
    "torch/csrc/jit/interpreter.cpp",
    "torch/csrc/jit/ir.cpp",
    "torch/csrc/jit/fusion_compiler.cpp",
    "torch/csrc/jit/graph_executor.cpp",
    "torch/csrc/jit/python_ir.cpp",
    "torch/csrc/jit/test_jit.cpp",
    "torch/csrc/jit/tracer.cpp",
    "torch/csrc/jit/tracer_state.cpp",
    "torch/csrc/jit/python_tracer.cpp",
    "torch/csrc/jit/passes/shape_analysis.cpp",
    "torch/csrc/jit/interned_strings.cpp",
    "torch/csrc/jit/type.cpp",
    "torch/csrc/jit/export.cpp",
    "torch/csrc/jit/import.cpp",
    "torch/csrc/jit/autodiff.cpp",
    "torch/csrc/jit/interpreter_autograd_function.cpp",
    "torch/csrc/jit/python_arg_flatten.cpp",
    "torch/csrc/jit/python_compiled_function.cpp",
    "torch/csrc/jit/variable_flags.cpp",
    "torch/csrc/jit/passes/create_autodiff_subgraphs.cpp",
    "torch/csrc/jit/passes/graph_fuser.cpp",
    "torch/csrc/jit/passes/onnx.cpp",
    "torch/csrc/jit/passes/dead_code_elimination.cpp",
    "torch/csrc/jit/passes/lower_tuples.cpp",
    "torch/csrc/jit/passes/common_subexpression_elimination.cpp",
    "torch/csrc/jit/passes/peephole.cpp",
    "torch/csrc/jit/passes/inplace_check.cpp",
    "torch/csrc/jit/passes/canonicalize.cpp",
    "torch/csrc/jit/passes/batch_mm.cpp",
    "torch/csrc/jit/passes/onnx/peephole.cpp",
    "torch/csrc/jit/passes/onnx/fixup_onnx_loop.cpp",
    "torch/csrc/jit/generated/aten_dispatch.cpp",
    "torch/csrc/jit/script/lexer.cpp",
    "torch/csrc/jit/script/compiler.cpp",
    "torch/csrc/jit/script/module.cpp",
    "torch/csrc/jit/script/init.cpp",
    "torch/csrc/jit/script/python_tree_views.cpp",
    "torch/csrc/autograd/init.cpp",
    "torch/csrc/autograd/grad_mode.cpp",
    "torch/csrc/autograd/engine.cpp",
    "torch/csrc/autograd/function.cpp",
    "torch/csrc/autograd/variable.cpp",
    "torch/csrc/autograd/saved_variable.cpp",
    "torch/csrc/autograd/input_buffer.cpp",
    "torch/csrc/autograd/profiler.cpp",
    "torch/csrc/autograd/python_function.cpp",
    "torch/csrc/autograd/python_cpp_function.cpp",
    "torch/csrc/autograd/python_variable.cpp",
    "torch/csrc/autograd/python_variable_indexing.cpp",
    "torch/csrc/autograd/python_legacy_variable.cpp",
    "torch/csrc/autograd/python_engine.cpp",
    "torch/csrc/autograd/python_hook.cpp",
    "torch/csrc/autograd/generated/VariableType.cpp",
    "torch/csrc/autograd/generated/Functions.cpp",
    "torch/csrc/autograd/generated/python_torch_functions.cpp",
    "torch/csrc/autograd/generated/python_variable_methods.cpp",
    "torch/csrc/autograd/generated/python_functions.cpp",
    "torch/csrc/autograd/generated/python_nn_functions.cpp",
    "torch/csrc/autograd/functions/basic_ops.cpp",
    "torch/csrc/autograd/functions/tensor.cpp",
    "torch/csrc/autograd/functions/accumulate_grad.cpp",
    "torch/csrc/autograd/functions/special.cpp",
    "torch/csrc/autograd/functions/utils.cpp",
    "torch/csrc/autograd/functions/init.cpp",
    "torch/csrc/nn/THNN.cpp",
    "torch/csrc/tensor/python_tensor.cpp",
    "torch/csrc/onnx/onnx.pb.cpp",
    "torch/csrc/onnx/onnx.cpp",
    "torch/csrc/onnx/init.cpp",
]

这部分没有太多意思, 就是src文件的罗列

try:
    import numpy as np
    include_dirs.append(np.get_include())
    extra_compile_args.append('-DWITH_NUMPY')
    WITH_NUMPY = True
except ImportError:
    WITH_NUMPY = False

if WITH_DISTRIBUTED:
    extra_compile_args += ['-DWITH_DISTRIBUTED']
    main_sources += [
        "torch/csrc/distributed/Module.cpp",
    ]
    if WITH_DISTRIBUTED_MW:
        main_sources += [
            "torch/csrc/distributed/Tensor.cpp",
            "torch/csrc/distributed/Storage.cpp",
        ]
        extra_compile_args += ['-DWITH_DISTRIBUTED_MW']
    include_dirs += [tmp_install_path + "/include/THD"]
    main_link_args += [THD_LIB]

if WITH_CUDA:
    nvtoolext_lib_name = None
    if IS_WINDOWS:
        cuda_lib_path = CUDA_HOME + '/lib/x64/'
        nvtoolext_lib_path = NVTOOLEXT_HOME + '/lib/x64/'
        nvtoolext_include_path = os.path.join(NVTOOLEXT_HOME, 'include')

        library_dirs.append(nvtoolext_lib_path)
        include_dirs.append(nvtoolext_include_path)

        nvtoolext_lib_name = 'nvToolsExt64_1'

        # MSVC doesn't support runtime symbol resolving, `nvrtc` and `cuda` should be linked
        main_libraries += ['nvrtc', 'cuda']
    else:
        cuda_lib_dirs = ['lib64', 'lib']

        for lib_dir in cuda_lib_dirs:
            cuda_lib_path = os.path.join(CUDA_HOME, lib_dir)
            if os.path.exists(cuda_lib_path):
                break
        extra_link_args.append('-Wl,-rpath,' + cuda_lib_path)

        nvtoolext_lib_name = 'nvToolsExt'

    library_dirs.append(cuda_lib_path)
    cuda_include_path = os.path.join(CUDA_HOME, 'include')
    include_dirs.append(cuda_include_path)
    include_dirs.append(tmp_install_path + "/include/THCUNN")
    extra_compile_args += ['-DWITH_CUDA']
    extra_compile_args += ['-DCUDA_LIB_PATH=' + cuda_lib_path]
    main_libraries += ['cudart', nvtoolext_lib_name]
    main_sources += [
        "torch/csrc/cuda/Module.cpp",
        "torch/csrc/cuda/Storage.cpp",
        "torch/csrc/cuda/Stream.cpp",
        "torch/csrc/cuda/utils.cpp",
        "torch/csrc/cuda/comm.cpp",
        "torch/csrc/cuda/python_comm.cpp",
        "torch/csrc/cuda/serialization.cpp",
        "torch/csrc/nn/THCUNN.cpp",
    ]

if WITH_NCCL:
    if WITH_SYSTEM_NCCL:
        main_link_args += [NCCL_SYSTEM_LIB]
        include_dirs.append(NCCL_INCLUDE_DIR)
    else:
        main_link_args += [NCCL_LIB]
    extra_compile_args += ['-DWITH_NCCL']
    main_sources += [
        "torch/csrc/cuda/nccl.cpp",
        "torch/csrc/cuda/python_nccl.cpp",
    ]
if WITH_CUDNN:
    main_libraries += [CUDNN_LIBRARY]
    # NOTE: these are at the front, in case there's another cuDNN in CUDA path
    include_dirs.insert(0, CUDNN_INCLUDE_DIR)
    if not IS_WINDOWS:
        extra_link_args.insert(0, '-Wl,-rpath,' + CUDNN_LIB_DIR)
    extra_compile_args += ['-DWITH_CUDNN']

if DEBUG:
    if IS_WINDOWS:
        extra_link_args.append('/DEBUG:FULL')
    else:
        extra_compile_args += ['-O0', '-g']
        extra_link_args += ['-O0', '-g']

if os.getenv('PYTORCH_BINARY_BUILD') and platform.system() == 'Linux':
    print('PYTORCH_BINARY_BUILD found. Static linking libstdc++ on Linux')
    # get path of libstdc++ and link manually.
    # for reasons unknown, -static-libstdc++ doesn't fully link some symbols
    CXXNAME = os.getenv('CXX', 'g++')
    STDCPP_LIB = subprocess.check_output([CXXNAME, '-print-file-name=libstdc++.a'])
    STDCPP_LIB = STDCPP_LIB[:-1]
    if type(STDCPP_LIB) != str:  # python 3
        STDCPP_LIB = STDCPP_LIB.decode(sys.stdout.encoding)
    main_link_args += [STDCPP_LIB]
    version_script = os.path.abspath("tools/pytorch.version")
    extra_link_args += ['-Wl,--version-script=' + version_script]


def make_relative_rpath(path):
    if IS_DARWIN:
        return '-Wl,-rpath,@loader_path/' + path
    elif IS_WINDOWS:
        return ''
    else:
        return '-Wl,-rpath,$ORIGIN/' + path

这一部分的代码的作用是 对extra_compile_args和extra_link_args做一些修正

################################################################################
# Declare extensions and package
################################################################################

extensions = []
packages = find_packages(exclude=('tools', 'tools.*', 'caffe2', 'caffe2.*', 'caffe', 'caffe.*'))
C = Extension("torch._C",
              libraries=main_libraries,
              sources=main_sources,
              language='c++',
              extra_compile_args=main_compile_args + extra_compile_args,
              include_dirs=include_dirs,
              library_dirs=library_dirs,
              extra_link_args=extra_link_args + main_link_args + [make_relative_rpath('lib')],
              )
extensions.append(C)

if not IS_WINDOWS:
    DL = Extension("torch._dl",
                   sources=["torch/csrc/dl.c"],
                   language='c',
                   )
    extensions.append(DL)


if WITH_CUDA:
    thnvrtc_link_flags = extra_link_args + [make_relative_rpath('lib')]
    if IS_LINUX:
        thnvrtc_link_flags = thnvrtc_link_flags + ['-Wl,--no-as-needed']
    # these have to be specified as -lcuda in link_flags because they
    # have to come right after the `no-as-needed` option
    if IS_WINDOWS:
        thnvrtc_link_flags += ['cuda.lib', 'nvrtc.lib']
    else:
        thnvrtc_link_flags += ['-lcuda', '-lnvrtc']
    cuda_stub_path = [cuda_lib_path + '/stubs']
    if IS_DARWIN:
        # on macOS this is where the CUDA stub is installed according to the manual
        cuda_stub_path = ["/usr/local/cuda/lib"]
    THNVRTC = Extension("torch._nvrtc",
                        sources=['torch/csrc/nvrtc.cpp'],
                        language='c++',
                        include_dirs=include_dirs,
                        library_dirs=library_dirs + cuda_stub_path,
                        extra_link_args=thnvrtc_link_flags,
                        )
    extensions.append(THNVRTC)

version = '0.5.0a0'
if os.getenv('PYTORCH_BUILD_VERSION'):
    assert os.getenv('PYTORCH_BUILD_NUMBER') is not None
    build_number = int(os.getenv('PYTORCH_BUILD_NUMBER'))
    version = os.getenv('PYTORCH_BUILD_VERSION')
    if build_number > 1:
        version += '.post' + str(build_number)
else:
    try:
        sha = subprocess.check_output(['git', 'rev-parse', 'HEAD'], cwd=cwd).decode('ascii').strip()
        version += '+' + sha[:7]
    except Exception:
        pass

Building C and C++ Extensions with distutils

这部分是加入extensions

cmdclass = {
    'build': build,
    'build_py': build_py,
    'build_ext': build_ext,
    'build_deps': build_deps,
    'build_module': build_module,
    'develop': develop,
    'install': install,
    'clean': clean,
}
cmdclass.update(build_dep_cmds)
if __name__ == '__main__':
    setup(
        name="torch",
        version=version,
        description=("Tensors and Dynamic neural networks in "
                     "Python with strong GPU acceleration"),
        ext_modules=extensions,
        cmdclass=cmdclass,
        packages=packages,
        package_data={
            'torch': [
                'lib/*.so*',
                'lib/*.dylib*',
                'lib/*.dll',
                'lib/*.lib',
                'lib/torch_shm_manager',
                'lib/*.h',
                'lib/include/ATen/*.h',
                'lib/include/ATen/cuda/*.h',
                'lib/include/ATen/cuda/*.cuh',
                'lib/include/ATen/cudnn/*.h',
                'lib/include/ATen/cuda/detail/*.cuh',
                'lib/include/pybind11/*.h',
                'lib/include/pybind11/detail/*.h',
                'lib/include/TH/*.h',
                'lib/include/TH/generic/*.h',
                'lib/include/THC/*.h',
                'lib/include/THC/*.cuh',
                'lib/include/THC/generic/*.h',
                'lib/include/THCUNN/*.cuh',
                'lib/include/torch/csrc/*.h',
                'lib/include/torch/csrc/autograd/*.h',
                'lib/include/torch/csrc/jit/*.h',
                'lib/include/torch/csrc/utils/*.h',
                'lib/include/torch/csrc/cuda/*.h',
                'lib/include/torch/torch.h',
            ]
        })

最后执行setup


python setup.py install

就可以运行了…

真是不知道读pytorch源码是不是一个正确的选择,

希望能坚持下去吧,

我之所以这么做

是因为

不是很喜欢现在的风气,

做DL的很多人,

连API都看不全,

更何况底层的实现呢?

我怕将来自己也会那些人中的一员

所以 …

    原文作者：Nick.Zxx
    原文地址: https://zhuanlan.zhihu.com/p/36337943
    本文转自网络文章，转载此文章仅为分享知识，如有侵权，请联系博主进行删除。
点赞