CUDA 函数指针

Question

我试图在 CUDA 中做这样的事情(实际上我需要写一些集成函数)

#include <iostream>
using namespace std;

float f1(float x) {
    return x * x;
}

float f2(float x) {
    return x;
}

void tabulate(float p_f(float)) {
    for (int i = 0; i != 10; ++i) {
        std::cout << p_f(i) << ' ';
    }
    std::cout << std::endl;
}

int main() {
    tabulate(f1);
    tabulate(f2);
    return 0;
}

输出:

0 1 4 9 16 25 36 49 64 81
0 1 2 3 4 5 6 7 8 9

---

我尝试了以下但只得到错误

Error: Function pointers and function template parameters are not supported in sm_1x.

float f1(float x) {
    return x;
}

__global__ void tabulate(float lower, float upper, float p_function(float), float* result) {
    for (lower; lower < upper; lower++) {
        *result = *result + p_function(lower);
    }
}

int main() {
    float res;
    float* dev_res;

    cudaMalloc( (void**)&dev_res, sizeof(float) ) ;

    tabulate<<<1,1>>>(0.0, 5.0, f1, dev_res);
    cudaMemcpy(&res, dev_res, sizeof(float), cudaMemcpyDeviceToHost);

    printf("%f\n", res);
    /************************************************************************/
    scanf("%s");

    return 0;
}

Answer 1

要消除编译错误，您必须在编译代码时使用 -gencode arch=compute_20,code=sm_20 作为编译器参数。但是你可能会遇到一些运行时问题:

摘自 CUDA 编程指南 http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#functions

Function pointers to __global__ functions are supported in host code, but not in device code. Function pointers to __device__ functions are only supported in device code compiled for devices of compute capability 2.x and higher.

It is not allowed to take the address of a __device__ function in host code.

所以你可以有这样的东西(改编自“FunctionPointers”样本):

//your function pointer type - returns unsigned char, takes parameters of type unsigned char and float
typedef unsigned char(*pointFunction_t)(unsigned char, float);

//some device function to be pointed to
__device__ unsigned char
Threshold(unsigned char in, float thresh)
{
   ...
}

//pComputeThreshold is a device-side function pointer to your __device__ function
__device__ pointFunction_t pComputeThreshold = Threshold;
//the host-side function pointer to your __device__ function
pointFunction_t h_pointFunction;

//in host code: copy the function pointers to their host equivalent
cudaMemcpyFromSymbol(&h_pointFunction, pComputeThreshold, sizeof(pointFunction_t))

然后您可以将 h_pointFunction 作为参数传递给您的内核，内核可以使用它来调用您的 __device__ 函数。

//your kernel taking your __device__ function pointer as a parameter
__global__ void kernel(pointFunction_t pPointOperation)
{
    unsigned char tmp;
    ...
    tmp = (*pPointOperation)(tmp, 150.0)
    ...
}

//invoke the kernel in host code, passing in your host-side __device__ function pointer
kernel<<<...>>>(h_pointFunction);

希望这是有道理的。总之，看起来您必须将 f1 函数更改为 __device__ 函数并遵循类似的过程(typedef 不是必需的，但它们确实使代码更好)才能获得它作为主机端的有效函数指针传递给您的内核。我还建议您看一下 FunctionPointers CUDA 示例