1
#include <iostream>
2
#include <cuda_runtime.h>
3

4
// CUDA 核函数
5
__global__ void vectorAdd(const float* A, const float* B, float* C, int n) {
6
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
7
    if (idx < n) {
8
        C[idx] = A[idx] + B[idx];
9
    }
10
}
11

12
// 错误检查宏
13
#define CUDA_CHECK(call) \
14
    do { \
15
        cudaError_t err = call; \
16
        if (err != cudaSuccess) { \
17
            std::cerr << "CUDA 错误：" << cudaGetErrorString(err) \
18
                      << " 在 " << __FILE__ << ":" << __LINE__ << std::endl; \
19
            exit(1); \
20
        } \
21
    } while(0)
22

23
int main() {
24
    // 1. 检查设备
25
    int deviceCount;
26
    CUDA_CHECK(cudaGetDeviceCount(&deviceCount));
27
    if (deviceCount == 0) {
28
        std::cerr << "没有找到支持 CUDA 的设备！" << std::endl;
29
        return 1;
30
    }
31
    std::cout << "找到 " << deviceCount << " 个 CUDA 设备" << std::endl;
32

33
    // 2. 数据大小
34
    int n = 1 << 20;
35
    size_t size = n * sizeof(float);
36

37
    // 3. 主机内存分配与初始化
38
    float* h_A = new float[n];
39
    float* h_B = new float[n];
40
    float* h_C = new float[n];
41

42
    for (int i = 0; i < n; ++i) {
43
        h_A[i] = static_cast<float>(i);
44
        h_B[i] = static_cast<float>(i * 2.0f);
45
    }
46

47
    // 4. 设备内存分配
48
    float *d_A, *d_B, *d_C;
49
    CUDA_CHECK(cudaMalloc(&d_A, size));
50
    CUDA_CHECK(cudaMalloc(&d_B, size));
51
    CUDA_CHECK(cudaMalloc(&d_C, size));
52

53
    // 5. 拷贝数据到设备
54
    CUDA_CHECK(cudaMemcpy(d_A, h_A, size, cudaMemcpyHostToDevice));
55
    CUDA_CHECK(cudaMemcpy(d_B, h_B, size, cudaMemcpyHostToDevice));
56

57
    // 6. 配置核函数
58
    int threadsPerBlock = 256;
59
    int blocksPerGrid = (n + threadsPerBlock - 1) / threadsPerBlock;
60

61
    // 7. 启动核函数
62
    vectorAdd<<<blocksPerGrid, threadsPerBlock>>>(d_A, d_B, d_C, n);
63

64
    // 检查核函数启动错误
65
    CUDA_CHECK(cudaGetLastError());
66

67
    // 等待核函数完成
68
    CUDA_CHECK(cudaDeviceSynchronize());
69

70
    // 8. 拷贝结果回主机
71
    CUDA_CHECK(cudaMemcpy(h_C, d_C, size, cudaMemcpyDeviceToHost));
72

73
    // 9. 输出前10个结果验证
74
    std::cout << "计算结果示例（前10个元素）：" << std::endl;
75
    for (int i = 0; i < 10 && i < n; ++i) {
76
        std::cout << "C[" << i << "] = " << h_C[i]
77
                  << " (期望 " << h_A[i] + h_B[i] << ")" << std::endl;
78
    }
79

80
    // 10. 释放内存
81
    CUDA_CHECK(cudaFree(d_A));
82
    CUDA_CHECK(cudaFree(d_B));
83
    CUDA_CHECK(cudaFree(d_C));
84
    delete[] h_A;
85
    delete[] h_B;
86
    delete[] h_C;
87

88
    return 0;
89
}