如何在gprbuild中使用nvcc？

Question

我在Ada中有一个代码，它必须使用CUDA而不用Ada绑定。因此，我创建了一个接口，允许Ada程序调用C代码。现在我想编译它。

如何告诉gprbuild不要使用gcc来编译nvcc的.cu文件？如果不可能，也许我必须使用nvcc生成对象，然后将它们与ada代码链接起来？你会怎么做？

编辑:使用Simon Wright提供的链接，我创建了这个gpr文件：

project Cuda_Interface is
  for Languages use ("Ada", "Cuda");

  for Source_Dirs use ("src");
  for Object_Dir  use "obj";
  for Exec_Dir    use ".";

  for Main use ("cuda_interface.adb");

  for Create_Missing_Dirs use "True";

  package Naming is
    for Body_Suffix("Cuda") use ".cu";
    for Spec_Suffix("Cuda") use ".cuh";
  end Naming;

  package Compiler is
    for Driver("Cuda") use "nvcc";
    for Leading_Required_Switches("Cuda") use ("-c");
  end Compiler;

  package Linker is
    for Default_Switches("Ada") use ("-L/usr/local/cuda/lib64", "-lcuda", "-lcudart", "-lm");
  end Linker;
end Cuda_Interface;

编译工作良好，但链接器返回此错误：

/usr/bin/ld : cuda_interface.o : in the function « _ada_cuda_interface » :
cuda_interface.adb:(.text+0x3a5) : undefined reference to « inter_add_two »
collect2: error: ld returned 1 exit status
gprbuild: link of cuda_interface.adb failed

cuda_interface.adb：

with Ada.Text_IO;   use Ada.Text_IO;

procedure Cuda_Interface is
  type Index is range 1 .. 5;
  type Element_Type is new Natural;

  type Array_Type is array (Index) of Element_Type;

  procedure Inter_Add_Two(Arr : in out Array_Type; Length : Index)
    with
      Import        => True,
      Convention    => C,
      External_Name => "inter_add_two";

  A : Array_Type := (1, 2, 3, 4, 5);
begin
  for I in Index loop
    Put_Line("Value at " 
            & Index'Image(I)
            & " is "
            & Element_Type'Image(A(I)));
  end loop;
  New_Line;

  Inter_Add_Two(A, Index'Last);

  for I in Index loop
    Put_Line("Value at " 
            & Index'Image(I)
            & " is "
            & Element_Type'Image(A(I)));
  end loop;
end Cuda_Interface;

kernel.cuh

#ifndef __KERNEL_CUH__
#define __KERNEL_CUH__

#include <cuda.h>

__global__ void kernel_add_two(unsigned int *a, unsigned int length);
void inter_add_two(unsigned int *a, unsigned int length);

#endif // __KERNEL_CUH__

kernel.cu

#include "kernel.cuh"
#include <math.h>

#define THREADS_PER_BLOCK (1024)


__global__ void kernel_add_two(unsigned int *a, unsigned int length)
{
  unsigned int tid = threadIdx.x + blockIdx.x * blockDim.x;
  if (tid < length) a[tid] += 2;
}

void inter_add_two(unsigned int *a, unsigned int length)
{
  unsigned int block_number = ceil(((float)length) / THREADS_PER_BLOCK);
  unsigned int *d_a;
  
  cudaMalloc((void**)&d_a, sizeof(unsigned int) * length);
  cudaMemcpy(d_a, a, sizeof(unsigned int) * length, cudaMemcpyHostToDevice);

  kernel_add_two<<<block_number, THREADS_PER_BLOCK>>>(d_a, length);
  
  cudaMemcpy(a, d_a, sizeof(unsigned int) * length, cudaMemcpyDeviceToHost);
  cudaFree(d_a);
}

Answer 1

感谢这些评论，我成功地编译并运行了一个名为C代码的Ada程序，该程序调用CUDA代码。以下是我编辑的文件：

kernel.cuh

#ifndef __KERNEL_CUH__
#define __KERNEL_CUH__

#include <cuda.h>

void *__gxx_personality_v0;

extern "C"
{
  __global__ void kernel_add_two(unsigned int *a, unsigned int length);
  void inter_add_two(unsigned int *a, unsigned int length);
}

#endif // __KERNEL_CUH__

cuda_interface.gpr

project Cuda_Interface is
  for Languages use ("Ada", "Cuda");

  for Source_Dirs use ("src");
  for Object_Dir  use "obj";
  for Exec_Dir    use ".";

  for Main use ("cuda_interface.adb");

  for Create_Missing_Dirs use "True";

  package Naming is
    for Body_Suffix("Cuda") use ".cu";
    for Spec_Suffix("Cuda") use ".cuh";
  end Naming;

  package Compiler is
    for Driver("Cuda") use "nvcc";
    for Leading_Required_Switches("Cuda") use ("-c");
  end Compiler;

  package Linker is
    for Default_Switches("Ada") use ("-L/usr/local/cuda/lib64", "-lcuda", "-lcudart", "-lcudadevrt", "-lm");
  end Linker;
end Cuda_Interface;