setup.py

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import io
import re
import glob
import os
import subprocess

import numpy as np
import torch
from setuptools import Extension
from setuptools import find_packages
from setuptools import setup
from torch.utils.cpp_extension import CppExtension
from torch.utils.cpp_extension import CUDAExtension

def _find_cuda_home():
    # guess rule 3 of torch.utils.cpp_extension
    nvcc = subprocess.check_output(['which', 'nvcc']).decode().rstrip('\r\n')
    cuda_home = os.path.dirname(os.path.dirname(nvcc))
    print(f'find cuda home:{cuda_home}')
    return cuda_home


# remember to overwrite PyTorch auto-detected cuda home which
# may not be our expected
torch.utils.cpp_extension.CUDA_HOME = _find_cuda_home()
CUDA_HOME = torch.utils.cpp_extension.CUDA_HOME

CORE_DIR = 'x_temporal'
EXT_DIR = 'cuda_shift'

def recursive_glob(base_dir, pattern):
    files = []
    for root, subdirs, subfiles in os.walk(base_dir):
        files += glob.glob(os.path.join(root, pattern))
    return files

def get_extensions():
    # this_dir = os.path.dirname(os.path.abspath(__file__))
    sources_dir = os.path.join(CORE_DIR, EXT_DIR, "src")

    source_cpu = recursive_glob(sources_dir, "*.cpp")
    source_cuda = recursive_glob(sources_dir, "*.cu")

    sources = source_cpu
    extension = CppExtension

    extra_compile_args = {"cxx": []}
    define_macros = []
    assert torch.cuda.is_available() and CUDA_HOME is not None

    if torch.cuda.is_available() and CUDA_HOME is not None:
        extension = CUDAExtension
        sources += source_cuda
        define_macros += [("WITH_CUDA", None)]
        extra_compile_args["nvcc"] = [
            "-DCUDA_HAS_FP16=1",
            "-D__CUDA_NO_HALF_OPERATORS__",
            "-D__CUDA_NO_HALF_CONVERSIONS__",
            "-D__CUDA_NO_HALF2_OPERATORS__",
        ]

    # sources = [os.path.join(extensions_dir, s) for s in sources]
    print(f'sources:{sources}')

    include_dirs = [sources_dir]

    ext_modules = [
        extension(
            f"{CORE_DIR}.{EXT_DIR}._C",
            sources,
            include_dirs=include_dirs,
            define_macros=define_macros,
            extra_compile_args=extra_compile_args,
        )
    ]

    return ext_modules

with io.open("x_temporal/__init__.py", "rt", encoding="utf8") as f:
    version = re.search(r'__version__ = "(\D*)(.*?)"', f.read(), re.M).group(2)

setup(
    name="x_temporal",
    version=version,
    author="X-Lab Temporal Team",
    url="http://gitlab.bj.sensetime.com/spring-ce/element/x-temporal",
    description="Video Understanding Framework in Distributed PyTorch",
    author_email="spring-support@senstime.com",
    package_data={
    },
    packages=find_packages(exclude=(
        "scripts",
        "experiments"
    )),
    ext_modules=get_extensions(),
    cmdclass={"build_ext": torch.utils.cpp_extension.BuildExtension},
)