从C调用cuda函数

#include <cuda_runtime.h>

__global__ void my_cuda_function(int *input, int *output) {
    int index = blockIdx.x * blockDim.x + threadIdx.x;
    output[index] = input[index] * 2;
}

extern "C" {
    void launch_my_cuda_function(int *input, int *output, int size) {
        int blockSize = 256;
        int gridSize = (size + blockSize - 1) / blockSize;
        my_cuda_function<<<gridSize, blockSize>>>(input, output);
    }
}

nvcc -c my_cuda_functions.cu -o my_cuda_functions.o

#include<stdio.h>

void launch_my_cuda_function(int *input, int *output, int size);

int main() {
    int size = 1024;
    int input[size], output[size];

    // 初始化输入数组
    for (int i = 0; i< size; i++) {
        input[i] = i;
    }

    // 调用CUDA函数
    launch_my_cuda_function(input, output, size);

    // 打印输出数组
    for (int i = 0; i< size; i++) {
        printf("%d ", output[i]);
    }

    return 0;
}

gcc main.c my_cuda_functions.o -o my_program -L/usr/local/cuda/lib64 -lcudart

./my_program