Ubuntu20.04中PVN3D复现过程交叉编译问题

Question

在Ubuntu20.04复现PVN3D的时候，编译Pointnet++，已经安装好了依赖，在执行python3 setup.py build_ext的时候报错，首先报错如下：

据我分析，我的显卡是3090，而这个项目要求的Cuda版本是10.2，torch用的1.5，gcc用的7，所以我认为是算力过高，加入以下命令：

export TORCH_CUDA_ARCH_LIST="7.5"

这个报错解决，但是出现新的报错：

/usr/local/cuda-10.2/bin/nvcc -I/home/ubuntu/anaconda3/envs/cppn/lib/python3.8/site-packages/torch/include -I/home/ubuntu/anaconda3/envs/cppn/lib/python3.8/site-packages/torch/include/torch/csrc/api/include -I/home/ubuntu/anaconda3/envs/cppn/lib/python3.8/site-packages/torch/include/TH -I/home/ubuntu/anaconda3/envs/cppn/lib/python3.8/site-packages/torch/include/THC -I/usr/local/cuda-10.2/include -I/home/ubuntu/anaconda3/envs/cppn/include/python3.8 -c ./pvn3d/_ext-src/src/ball_query_gpu.cu -o build/temp.linux-x86_64-cpython-38/./pvn3d/_ext-src/src/ball_query_gpu.o -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr --compiler-options '-fPIC' -O2 -I./pvn3d/_ext-src/include -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=_ext -D_GLIBCXX_USE_CXX11_ABI=0 -gencode=arch=compute_75,code=sm_75 -std=c++14
cc1plus: error: unrecognized command line option ‘-fstack-clash-protection’
cc1plus: error: unrecognized command line option ‘-fcf-protection’
error: command '/usr/local/cuda-10.2/bin/nvcc' failed with exit code 1

这个真的百思不得其解，我在Ubuntu18.04复现的时候就没有这个问题，目前猜测是20.04的某些东西自动喂给了gcc这两个参数
麻烦不要用AI模型回答我！
serup.py的内容如下：

#!/usr/bin/env python3
import os
import glob
from setuptools import setup
from torch.utils.cpp_extension import BuildExtension, CUDAExtension


_ext_src_root = "./pvn3d/_ext-src"
_ext_sources = glob.glob("{}/src/*.cpp".format(_ext_src_root)) + glob.glob(
    "{}/src/*.cu".format(_ext_src_root)
)
_ext_headers = glob.glob("{}/include/*".format(_ext_src_root))


setup(
    name='pvn3d',
    ext_modules=[
        CUDAExtension(
            name='pointnet2_utils._ext',
            sources=_ext_sources,
            extra_compile_args={
                "cxx": ["-O2", "-I{}".format("{}/include".format(_ext_src_root))],
                "nvcc": [
                    "-O2", "-I{}".format("{}/include".format(_ext_src_root))
                ],
            },
        )
    ],
    cmdclass={
        'build_ext': BuildExtension
    }
)


try:
    src_pth = './build'
    tg_pth = 'pvn3d/lib/pointnet2_utils/'
    fd_lst = os.listdir(src_pth)
    for fd in fd_lst:
        if 'lib' in fd:
            src_pth = os.path.join(src_pth, fd, 'pointnet2_utils')
            f_nm = os.listdir(src_pth)[0]
            src_pth = os.path.join(src_pth, f_nm)
            tg_pth = os.path.join(tg_pth, f_nm)
    os.system('cp {} {}'.format(src_pth, tg_pth))
    print(
        src_pth, '==>', tg_pth,
    )
except:
    print(
        "\n****************************************************************\n",
        "Failed to copy builded .so to ./pvn3d/lib/pointnet2_utils/.\n",
        "Please manually copy the builded .so file (_ext.cpython*.so) in ./build"+\
        " to ./pvn3d/lib/pointnet2_utils/.",
        "\n****************************************************************\n"
    )

# vim: ts=4 sw=4 sts=4 expandtab

Answer 1

这个问题我也遇到过，可以检查一下你安装的torch是用什么编译的，指令:torch.config.show()，看看使用的gcc和这个版本一样不一样，强烈建议从官网上下载torch，不要用轮子文件编译，问题真的很多。另外这个项目可以用更高版本的torch复现，不一定要用10.2