C++ 使用?__global__?聲明說明符定義內核，并使用新的?<<<...>>>?執行配置語法指定內核調用的 CUDA 線程數（請參閱?C++ 語言擴展）。 每個執行內核的線程都有一個唯一的線程 ID，可以通過內置變量在內核中訪問。

示例代碼使用內置變量?threadIdx?將兩個大小為 N 的向量 A 和 B 相加，并將結果存儲到向量 C 中：

#include <cuda_runtime_api.h>
#include <iostream>#define RANDOM(x) (rand() % x)
#define MAX 10// single block multiple threads
__global__ void vector_add_gpu_2(int *d_a, int *d_b, int *d_c, int n){int tid = threadIdx.x;const int t_n = blockDim.x;while(tid < n){d_c[tid] = d_a[tid] + d_b[tid];tid+=t_n;}
}int main(){/***向量相加的實現***/int n = 5;int *a = (int *)malloc(sizeof(int)*n);int *b = (int *)malloc(sizeof(int)*n);int *c = (int *)malloc(sizeof(int)*n);for (size_t i = 0; i<n; i++){a[i] = RANDOM(MAX);b[i] = RANDOM(MAX);std::cout << a[i] << "   " << b[i] << std::endl;}cudaError_t  cudaStatus;// GPU memory allocateint *d_a, *d_b, *d_c;cudaMalloc((void **)&d_a, sizeof(int)*n);cudaMalloc((void **)&d_b, sizeof(int)*n);cudaMalloc((void **)&d_c, sizeof(int)*n);// data a and b copy to GPUcudaStatus = cudaMemcpy(d_a, a, sizeof(int)*n, cudaMemcpyHostToDevice);if (cudaStatus != cudaSuccess) {std::cout << ("Memory copy failed! error code: %s", cudaGetErrorString(cudaStatus)) << std::endl;}cudaStatus = cudaMemcpy(d_b, b, sizeof(int)*n, cudaMemcpyHostToDevice);if (cudaStatus != cudaSuccess) {std::cout << ("Memory copy failed! error code: %s", cudaGetErrorString(cudaStatus)) << std::endl;}vector_add_gpu_2<<<1, 3>>>(d_a, d_b, d_c, n);// result copy back to CPUcudaMemcpy(c, d_c, sizeof(int)*n, cudaMemcpyDeviceToHost);std::cout << "the result of add is: " << std::endl;for (size_t i = 0; i<n; i++){std::cout << c[i] << " ";}std::cout << std::endl;// GPU memory freecudaFree(d_a);cudaFree(d_b);cudaFree(d_c);free(a);free(b);free(c);return 0;
}

思考：

1 VecAdd 被執行幾次？

2 怎么被調用的？

3 有多少個線程在執行？線程間數據相互影響嗎？