1
// 输入图像在显存中是按 uchar3 (RGB三通道) 紧密排列的
2
__global__ void rgbToGray(const uchar3* d_img, unsigned char* d_gray, int width, int height) {
3
    // 计算全局列和行
4
    int x = blockIdx.x * blockDim.x + threadIdx.x;
5
    int y = blockIdx.y * blockDim.y + threadIdx.y;
6

7
    // 边界检查
8
    if (x < width && y < height) {
9
        // 二维转一维索引
10
        int idx = y * width + x;
11

12
        // 读取原始 RGB 像素
13
        uchar3 pixel = d_img[idx];
14

15
        // 计算灰度值 (固定加权公式)
16
        // Y = 0.299R + 0.587G + 0.114B
17
        unsigned char gray = static_cast<unsigned char>(0.299f * pixel.x + 0.587f * pixel.y + 0.114f * pixel.z);
18

19
        // 写回显存
20
        d_gray[idx] = gray;
21
    }
22
}

1
int main() {
2
    // ... 读取图像，获取 width 和 height ...
3

4
    // 定义线程块大小：16x16 是常见且高效的选择
5
    dim3 blockSize(16, 16); // dim3是CUDA中的数据结构
6

7
    // 计算网格大小：用图像尺寸除以块大小并向上取整
8
    // 公式：(N + M - 1) / M
9
    dim3 gridSize((width + blockSize.x - 1) / blockSize.x,
10
                  (height + blockSize.y - 1) / blockSize.y);
11

12
    // 启动 Kernel
13
    rgbToGray<<<gridSize, blockSize>>>(d_img, d_gray, width, height);
14

15
    // ... 后续同步与结果保存 ...
16
}

1
#include <iostream>
2
#include <opencv2/opencv.hpp>
3
#include <cuda_runtime.h>
4
#include <string>
5

6
__global__ void rgbToGray(const uchar3* d_img, unsigned char* d_gray, int width, int height) {
7
    int x = blockIdx.x * blockDim.x + threadIdx.x;
8
    int y = blockIdx.y * blockDim.y + threadIdx.y;
9

10
    if (x < width && y < height) {
11
        int idx = y * width + x;
12
        uchar3 pixel = d_img[idx];
13
        // 使用加权平均法将RGB转换为灰度
14
        d_gray[idx] = static_cast<unsigned char>(0.299f * pixel.x + 0.587f * pixel.y + 0.114f * pixel.z);
15
    }
16
}
17

18
using namespace std;
19
int main() {
20
    string imagePath = "input.png";
21
    cv::Mat img = cv::imread(imagePath);
22

23
    // 检查图像是否加载成功
24
    if (img.empty()) {
25
        cerr << "无法加载图像: " << imagePath << endl;
26
        return -1;
27
    }
28

29
    int width = img.cols;
30
    int height = img.rows;
31
    int channels = img.channels();
32

33
    uchar* d_img;           // GPU上的原始图像数据
34
    unsigned char* d_gray;  // GPU上的灰度图像
35
    cout << "图像尺寸: " << width << "x" << height << ", 通道数: " << channels << endl;
36

37

38
    // 定义CUDA内核的块和网格大小
39
    dim3 blockSize(1, 256);
40
    const int iterations = 10000;
41

42
    size_t imgSize = width * height * sizeof(uchar3);
43
    size_t graySize = width * height * sizeof(unsigned char);
44

45
    // 在GPU上分配内存
46
    cudaMalloc(&d_img, imgSize);
47
    cudaMalloc(&d_gray, graySize);
48

49
    // 将图像数据从CPU复制到GPU
50
    cudaMemcpy(d_img, img.data, imgSize, cudaMemcpyHostToDevice);
51

52

53
    // 取上整除以确保覆盖所有像素
54
    dim3 gridSize((width + blockSize.x - 1) / blockSize.x, (height + blockSize.y - 1) / blockSize.y);
55

56

57
    // 启动CUDA内核进行RGB到灰度的转换
58
    rgbToGray<<<gridSize, blockSize>>>(reinterpret_cast<uchar3*>(d_img), d_gray, width, height);
59
    // 检查内核启动是否成功
60
    cudaGetLastError();
61
    // 同步
62
    cudaDeviceSynchronize();
63

64

65
    // 测试时间
66
    cudaEvent_t start, stop;
67
    float total_time = 0.0f;
68

69
    cudaEventCreate(&start);
70
    cudaEventCreate(&stop);
71
    cout << "正在测试GPU性能..." << endl;
72

73
    for (int i = 0; i < iterations; i++) {
74
        cudaEventRecord(start);
75
        rgbToGray<<<gridSize, blockSize>>>(reinterpret_cast<uchar3*>(d_img), d_gray, width, height);
76
        cudaEventRecord(stop);
77
        cudaEventSynchronize(stop);
78

79
        float single_time = 0.0f;
80
        cudaEventElapsedTime(&single_time, start, stop);
81
        total_time += single_time;
82
    }
83

84
    float avg_time = total_time / iterations;
85
    cout << "平均每次转换时间: " << avg_time << " ms" << endl;
86

87

88

89
    // 将结果从GPU复制回CPU
90
    cv::Mat grayImg(height, width, CV_8UC1);
91
    cudaMemcpy(grayImg.data, d_gray, graySize, cudaMemcpyDeviceToHost);
92
    // 保存灰度图像
93
    cv::imwrite("output.png", grayImg);
94
    // 释放GPU内存
95
    cudaFree(d_img);
96
    cudaFree(d_gray);
97

98

99
    return 0;
100
}