内核调用会产生错误"error: a host function call cannot be configured"。调用有什么问题?

Kernel invocation produces error "error: a host function call cannot be configured". What is wrong with the invocation?

本文关键字:调用 configured cannot be 问题 什么 call host 错误 error 内核      更新时间:2023-10-16

当使用 nvcc -c mag_cuda.cu 编译以下代码时:

//Standard Libraries
#include <iostream>
#include <math.h>
#include <vector>
//Project Specific Header
#include "mag.hpp"
__global__
void indv_B_components(int *self_coords, int pole_coords[][3], double *indv_B[][3], int No_poles, int counter_1)
{
some code......
}
//----------------------------------------------------------
//------- Function to Calculate B Field at Each Pole -------
//----------------------------------------------------------
void calc_indv_B()
{
//declare namepspace for internal variables
using namespace mag::internal;
int *ppole_coords = &pole_coords[0][0];
double *pindv_B;
int self_coords[3];
int num_threads_in_block = 256;
int num_blocks = 32*2;
cudaMallocManaged(&pindv_B, No_poles*3*sizeof(int));  

//first loop to go over all poles
for(int counter_1 = 0; counter_1 < No_poles; counter_1++)
{
//store coords of the current pole
self_coords[0] = pole_coords[counter_1][0];
self_coords[1] = pole_coords[counter_1][1];
self_coords[2] = pole_coords[counter_1][2];

indv_B_components<<<num_blocks, num_threads_in_block>>>(self_coords, ppole_coords,  pindv_B, No_poles, counter_1); 

cudaDeviceSynchronize();
}
cudaFree(pindv_B);
//return from function
return;
}

返回以下错误:

错误:无法配置主机函数调用

指的是线

indv_B_components<<<num_blocks, num_threads_in_block>>>(self_coords, ppole_coords,  pindv_B, No_poles, counter_1); 

由于定义了所有参数,并且主机设备调用的内核声明为__global__我不知道是什么原因造成的。

头文件 mag.hpp 是:

//make sure MAG_H_ module hasnt been defined multiple times
#ifndef MAG_H_
#define MAG_H_
//standard libraries
#include <iostream>
#include <math.h>
#include <vector>

//Namespace for module
namespace mag
{

//define functions
...

void indv_B_components(int *self_coords, int *pole_coords, double *indv_B, int No_poles, int counter_1);
void calc_indv_B();
...
}
#endif //MAG_H_

有什么帮助吗?

发生此错误的一种情况是,如果在调用内核时只有内核函数的"声明"而没有__global__说明符,例如:

void kernel();
void f()
{
kernel<<<1, 1>>>();
}
__global__
void kernel()
{
}

现场演示在这里

内核函数的声明必须包含__global__说明符:

__global__ void kernel();

否则,声明不会声明内核函数,而只是声明一个普通的主机函数,这就是为什么编译器会抱怨,因为只能在 GPU 上启动内核函数......