Including C standard headers in CUDA NVRTC code

I'm writing a CUDA kernel that is compiled at runtime using NVRTC (CUDA version 9.2 with NVRTC version 7.5), which needs the stdint.h header, in order to have the int32_t etc. types.

If I write the kernel source code without the include, it works correctly. For example the kernel

extern "C" __global__ void f() { ... }

Compiles to PTX code where f is defined as .visible .entry f.

But if the kernel source code is

#include <stdint.h>
extern "C" __global__ void f() { ... }

it reports A function without execution space annotations (__host__/__device__/__global__) is considered a host function, and host functions are not allowed in JIT mode. (also without extern "C").

Passing -default-device makes the PTX code .visible .func f, so the function cannot be called from the host.

Is there a way to include headers in the source code, and still have a __global__ entry function? Or alternately, a way to know which integer size convention is used on the by the NVRTC compiler, so that the int32_t etc. types can be manually defined?

Edit: Example program that shows the problem:

#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;
}

When //#include <stdint.h> in the kernel source is uncommented it no longer compiles. When //options.push_back("-default-device"); is uncommented it compiles but does not mark the function f as .entry.

CMakeLists.txt to compile it (needs CUDA driver 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)

Solution

[Preface: this is a very hacky answer, and is specific to the GNU toolchain (although I suspect the problem in the question is also specific to the GNU toolchain)].

It would appear that the problem here is with the GNU standard header features.h, which gets pulled into stdint.h and then winds up defining a lot of stub functions which have the default __host__ compilation space. This causes nvrtc to blow up. It also seems that the -default-device option will result in a resolved glibC compiler feature set which makes the whole nvrtc compiler fail.

You can defeat this (in a very hacky way) by predefining a feature set for the standard library which excludes all the host functions. Changing your JIT kernel code to

const char program_source[] = R"%%%(
#define __ASSEMBLER__
#define __extension__
#include <stdint.h>
extern "C" __global__ void f(int32_t* in, int32_t* out) {
    out[threadIdx.x] = in[threadIdx.x];
}
)%%%";

got me this:

$ nvcc -std=c++14 -ccbin=g++-7 jit_header.cu -o jitheader -lnvrtc -lcuda
$ ./jitheader 
PTX code:
//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-24330188
// Cuda compilation tools, release 9.2, V9.2.148
// Based on LLVM 3.4svn
//

.version 6.2
.target sm_30
.address_size 64

    // .globl   f

.visible .entry f(
    .param .u64 f_param_0,
    .param .u64 f_param_1
)
{
    .reg .b32   %r<3>;
    .reg .b64   %rd<8>;


    ld.param.u64    %rd1, [f_param_0];
    ld.param.u64    %rd2, [f_param_1];
    cvta.to.global.u64  %rd3, %rd2;
    cvta.to.global.u64  %rd4, %rd1;
    mov.u32     %r1, %tid.x;
    mul.wide.u32    %rd5, %r1, 4;
    add.s64     %rd6, %rd4, %rd5;
    ld.global.u32   %r2, [%rd6];
    add.s64     %rd7, %rd3, %rd5;
    st.global.u32   [%rd7], %r2;
    ret;
}

Big caveat: This worked on the glibC system I tried it on. It probably won't work with other toolchains or libC implementations (if, indeed, they have this problem).