使用 cmake 使用 cpp 和 cuda 源构建 pybind11 模块

Question

我正在尝试为需要使用支持 cuda 的编译器进行编译的虚拟类生成 python 绑定。我正在使用 cmake 3.12.0、pybind11 v2.2.3 和 nvcc 7.5.17。编译失败，因为 和 等选项-flto直接-fno-fat-lto-objects传递给 nvcc，而 nvcc 无法识别它们。

这是一个（最小）示例：
Cuda 代码：

//Adder.hpp
#include <thrust/host_vector.h>
struct Adder {
    thrust::host_vector<float> a_h;
    thrust::host_vector<float> b_h;
    thrust::host_vector<float> r_h;
    int N;

    Adder(int N);
    void set_a(float const * const in);
    void set_b(float const * const in);
    void calc();
    void calc_gpu();
};

//Adder.cu
#include "Adder.hpp"
#include <thrust/device_vector.h>

Adder::Adder(int N): N(N),a_h(N),b_h(N),r_h(N) {}
void Adder::set_a(float const * const in) {
    for (int i=0; i<N; ++i) {
        a_h[i] = in[i];
    }
}
void Adder::set_b(float const * const in) {
    for (int i=0; i<N; ++i) {
        b_h[i] = in[i];
    }
}
void Adder::calc() {
    for (int i=0; i<N; ++i) {
        r_h[i] = a_h[i]+b_h[i];
    }
}
void Adder::calc_gpu() {
    thrust::device_vector<float> a_d(a_h);
    thrust::device_vector<float> b_d(b_h);
    thrust::device_vector<float> r_d(r_h);

    thrust::transform(a_d.begin(), a_d.end(), b_d.begin(), r_d.begin(),thrust::plus<float>());
    r_h = r_d;
}

绑定代码：

#include "Adder.hpp"
#include "lib/include/pybind11/pybind11.h"
#include "lib/include/pybind11/numpy.h"
#include <stdexcept>

namespace py = pybind11;

void bind_Adder(py::module& m) {
    py::class_<Adder>(m,"Adder","Module docstring")
        .def(py::init<int>(), py::arg("N"), "Init Adder")
        .def(
            "set_a"
            , [](Adder& self, py::array_t<float, py::array::c_style | py::array::forcecast> in) {
                py::buffer_info ai = in.request();
                if (ai.ndim!=1 || ai.shape[0]!=self.N || ai.strides[0]!=sizeof(float)) {
                    throw std::runtime_error("Shape of given numpy array must be (N,)! Type must be float.");
                }
                self.set_a(static_cast<float const * const>(ai.ptr));
            }
            , py::arg("in")
            , "Set a."
        )
        .def(
            "set_b"
            , [](Adder& self, py::array_t<float, py::array::c_style | py::array::forcecast> in) {
                py::buffer_info ai = in.request();
                if (ai.ndim!=1 || ai.shape[0]!=self.N || ai.strides[0]!=sizeof(float))
                    throw std::runtime_error("Shape of given numpy array must be (N,)! Type must be float.");
                self.set_b(static_cast<float const * const>(ai.ptr));
            }
            , py::arg("in")
            , "Set b."
        )
        .def(
            "get_r"
            , [](Adder& self, py::array_t<float> x) {
                auto r = x.mutable_unchecked<1>(); 
                for (ssize_t i = 0; i < r.shape(0); i++) {
                    r(i) = self.r_h[i];
                }
            }
            , py::arg("x").noconvert())
        .def("calc", &Adder::calc, "Calculate on CPU.")
        .def("calc_gpu", &Adder::calc_gpu, "Calculate on GPU.");
}

PYBIND11_MODULE(dummy, m) {
    bind_Adder(m);
}

CMakeLists.txt：

CMAKE_MINIMUM_REQUIRED(VERSION 3.11)
project(dummy LANGUAGES CXX CUDA)

set(PYBIND11_CPP_STANDARD -std=c++11)
add_subdirectory(lib/pybind11)

pybind11_add_module(dummy 
    Adder.pybind.cpp
    Adder.cu
)

使用eg 构建cmake ../src && make -VERBOSE=1失败。此时目标文件Adder.pybind.cpp已成功生成。编译Adder.cu失败并显示：

/usr/bin/nvcc  -Ddummy_EXPORTS -I/home/user/projects/pybind11_cuda_cmake/lib/pybind11/include -I/home/user/.anaconda3/include/python3.6m  -Xcompiler=-fPIC   -std=c++11 -flto -fno-fat-lto-objects -x cu -c /home/user/projects/pybind11_cuda_cmake/Adder.cu -o CMakeFiles/dummy.dir/Adder.cu.o
nvcc fatal   : Unknown option 'flto'

我尝试禁用到 nvcc 的自动传播，但没有成功。

set(CUDA_PROPAGATE_HOST_FLAGS FALSE)
set(CUDAFLAGS "-Xcompiler -fPIC -Xcompiler -flto -Xcompiler=-fvisibility=hidden")

有人知道如何进行这项工作吗？

Answer 1

更新：使用 target_set_properties 的（也许）更干净的版本，从而让 cmake 整理出实际的编译器/链接器标志。

cmake_minimum_required(VERSION 3.12)
project(dummy LANGUAGES CXX CUDA)

set(PYBIND11_CPP_STANDARD -std=c++11)
add_subdirectory(lib/pybind11)

add_library(dummycu STATIC
    Adder.cu
)

set_target_properties(dummycu PROPERTIES 
    POSITION_INDEPENDENT_CODE ON
    CUDA_VISIBILITY_PRESET "hidden"
    # CUDA_SEPARABLE_COMPILATION ON
)


add_library(dummy MODULE
    Adder.pybind.cpp
)

set_target_properties(dummy PROPERTIES 
    CXX_VISIBILITY_PRESET "hidden"
    INTERPROCEDURAL_OPTIMIZATION TRUE
    PREFIX "${PYTHON_MODULE_PREFIX}"
    SUFFIX "${PYTHON_MODULE_EXTENSION}"
)

target_link_libraries(dummy PRIVATE dummycu)
target_link_libraries(dummy PRIVATE pybind11::module)

我们还可以编译和链接一个目标内的所有内容。这或许也可以用于pybind11_add_module在 pybind11Tools.cmake 中进行修改。
由于 pybind11 无条件地为其目标设置可见性标志module，因此与不带 的 nvcc 一起使用-Xcompiler，如果不想更改 pybind11 CMakeLists.txt 或 pybind11Tools.cmake，则需要进行一些小修改。module为目标和/或函数编写自定义版本可能会更好pybind11_add_module。

cmake_minimum_required(VERSION 3.12)
project(dummy LANGUAGES CXX CUDA)

set(PYBIND11_CPP_STANDARD -std=c++11)
add_subdirectory(lib/pybind11)

##pybind11 sets -fvisibility=hidden in INTERFACE_COMPILE_OPTIONS on it's module target
get_target_property(modifacecopts module INTERFACE_COMPILE_OPTIONS)
list(REMOVE_ITEM modifacecopts "-fvisibility=hidden")
set_target_properties(module PROPERTIES INTERFACE_COMPILE_OPTIONS "${modifacecopts}")

add_library(dummy MODULE
    Adder.pybind.cpp
    Adder.cu
)

set_target_properties(dummy PROPERTIES 
    POSITION_INDEPENDENT_CODE ON
    CUDA_VISIBILITY_PRESET "hidden"
    CXX_VISIBILITY_PRESET "hidden"
    INTERPROCEDURAL_OPTIMIZATION TRUE
    PREFIX "${PYTHON_MODULE_PREFIX}"
    SUFFIX "${PYTHON_MODULE_EXTENSION}"
)

target_link_libraries(dummy PRIVATE pybind11::module)