我正在编写一个在运行时使用 NVRTC（CUDA 9.2 版和 NVRTC 7.5 版）编译的 CUDA 内核，它需要stdint.h标头，以便拥有int32_t等类型。
如果我编写没有包含的内核源代码，它可以正常工作。例如内核
extern "C" __global__ void f() { ... }
编译为 PTX 代码，其中 f 定义为.visible .entry f。
但是如果内核源代码是
#include <stdint.h>
extern "C" __global__ void f() { ... }
它报告A function without execution space annotations (__host__/__device__/__global__) is considered a host function, and host functions are not allowed in JIT mode.（也没有extern "C"）。
传递-default-device使 PTX 代码.visible .func f，因此无法从主机调用该函数。
有没有办法在源代码中包含标头，并且仍然具有__global__入口功能？或者，一种方法可以知道 NVRTC 编译器使用哪种整数大小约定，以便int32_t可以手动定义等类型？
编辑： 显示问题的示例程序：
#include <cstdlib>
#include <string>
#include <vector>
#include <memory>
#include <cassert>
#include <iostream>
#include <cuda.h>
#include <cuda_runtime.h>
#include <nvrtc.h>
[[noreturn]] void fail(const std::string& msg, int code) {
std::cerr << "error: " << msg << " (" << code << ')' << std::endl;
std::exit(EXIT_FAILURE);
}
std::unique_ptr<char[]> compile_to_ptx(const char* program_source) {
nvrtcResult rv;
// create nvrtc program
nvrtcProgram prog;
rv = nvrtcCreateProgram(
&prog,
program_source,
"program.cu",
0,
nullptr,
nullptr
);
if(rv != NVRTC_SUCCESS) fail("nvrtcCreateProgram", rv);
// compile nvrtc program
std::vector<const char*> options = {
"--gpu-architecture=compute_30"
};
//options.push_back("-default-device");
rv = nvrtcCompileProgram(prog, options.size(), options.data());
if(rv != NVRTC_SUCCESS) {
std::size_t log_size;
rv = nvrtcGetProgramLogSize(prog, &log_size);
if(rv != NVRTC_SUCCESS) fail("nvrtcGetProgramLogSize", rv);
auto log = std::make_unique<char[]>(log_size);
rv = nvrtcGetProgramLog(prog, log.get());
if(rv != NVRTC_SUCCESS) fail("nvrtcGetProgramLog", rv);
assert(log[log_size - 1] == '\0');
std::cerr << "Compile error; log:\n" << log.get() << std::endl;
fail("nvrtcCompileProgram", rv);
}
// get ptx code
std::size_t ptx_size;
rv = nvrtcGetPTXSize(prog, &ptx_size);
if(rv != NVRTC_SUCCESS) fail("nvrtcGetPTXSize", rv);
auto ptx = std::make_unique<char[]>(ptx_size);
rv = nvrtcGetPTX(prog, ptx.get());
if(rv != NVRTC_SUCCESS) fail("nvrtcGetPTX", rv);
assert(ptx[ptx_size - 1] == '\0');
nvrtcDestroyProgram(&prog);
return ptx;
}
const char program_source[] = R"%%%(
//#include <stdint.h>
extern "C" __global__ void f(int* in, int* out) {
out[threadIdx.x] = in[threadIdx.x];
}
)%%%";
int main() {
CUresult rv;
// initialize CUDA
rv = cuInit(0);
if(rv != CUDA_SUCCESS) fail("cuInit", rv);
// compile program to ptx
auto ptx = compile_to_ptx(program_source);
std::cout << "PTX code:\n" << ptx.get() << std::endl;
}
当//#include <stdint.h>内核源代码被取消注释时，它不再编译。取消注释时//options.push_back("-default-device");，它会编译但不会将函数标记f为.entry.
CMakeLists.txt 进行编译（需要 CUDA 驱动 API + NVRTC）
cmake_minimum_required(VERSION 3.4)
project(cudabug CXX)
find_package(CUDA REQUIRED)
set(CMAKE_CXX_STANDARD 14)
set(CMAKE_CXX_STANDARD_REQUIRED 14)
add_executable(cudabug cudabug.cc)
include_directories(SYSTEM ${CUDA_INCLUDE_DIRS})
link_directories(${CUDA_LIBRARY_DIRS})
target_link_libraries(cudabug PUBLIC ${CUDA_LIBRARIES} nvrtc cuda)