基于灰度值的匹配方法研究（二）

//---------------------------------------------------------------------------------------
// 初始化GPU核函数。Up<>其实就是std::unique_ptr<>
//---------------------------------------------------------------------------------------
Up<KernelFunc> GrayMatch::initKernel()
{
    string fillAllScoreStr(R"CLC(
        typedef struct
        {
            int x;
            int y;
        } Point2i;

        typedef struct
        {
            Point2i pos;
            float gray;
        } Gray2i;
        
        kernel void fillAllScore(global const uchar* imageData, int rows, int cols, 
            global const Gray2i *samples, int count, 
            int minx, int maxx, int miny, int maxy, 
            global float* output)
        {
            int x = minx + get_global_id(0);
            int y = miny + get_global_id(1);

            float iavr = 0;
            int last = count - 1;
            for (int i = 0; i < last; i++)
            {
                Point2i pos = { samples[i].pos.x + x, samples[i].pos.y + y };
                iavr += imageData[pos.y * cols + pos.x];
            }
            iavr /= last;
            float hcor = 0, lcor = 0;
            for (int i = 0; i < last; i++)
            {
                Point2i pos = { samples[i].pos.x + x, samples[i].pos.y + y };
                float s = fabs(imageData[pos.y * cols + pos.x] - iavr);
                hcor += s * samples[i].gray;
                lcor += s * s;
            }
            float ccor = samples[last].gray;
            float score = hcor / (sqrt(lcor) * ccor);
            output[y * cols + x] = score;
        }
    )CLC");

    cl::Program program(fillAllScoreStr);
    if (program.build("-cl-std=CL2.0") != CL_SUCCESS)
    {
        cl_int buildErr = CL_SUCCESS;
        auto buildInfo = program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(&buildErr);
        for (auto &pair : buildInfo)
        {
            qDebug() << TQ(pair.second);
        }
        ASSERT(false, "核函数有错务必修改");
    }
    return std::make_unique<KernelFunc>(program, "fillAllScore");
}

//---------------------------------------------------------------------------------------
// 入口函数大致流程
//---------------------------------------------------------------------------------------
void GrayMatch::execute()
{
    /* 第一层遍历搜索的大致流程 */
    /* tplSample是第N层的模板数据，数据类型是vector<Gray2f> */
    /* runtime是第N层的运行时图像，数据类型是Mat */
    /* myGpuFunc的数据类型是Up<KernelFunc>，它在构造函数中初始化 */
    cl::Buffer inputImage(std::begin(runtime), std::end(runtime), true);
    while (每一个旋转角度)
    {
        Mat scoreMap = Mat::zeros(runtime.rows, runtime.cols, CV_32FC1);
        vector<Gray2i> rotatedPoints;
        rotateGrayBorder(tplSample, angle, rotatedPoints); /* 旋转模板数据 */
        ...
        int miny, maxy; /* 它是滑动窗口y轴的范围 */
        int minx, maxx; /* 它是滑动窗口x轴的范围 */
        cl::Buffer inputSample(rotatedPoints.begin(), rotatedPoints.end(), true);
        cl::Buffer output(std::begin<float>(scoreMap), std::end<float>(scoreMap), false);

        (*myGpuFunc)(cl::EnqueueArgs(cl::NDRange(maxx - minx + 1, maxy - miny + 1), cl::NDRange(/*maxx - minx, 1*/)),
            inputImage, runtime.rows, runtime.cols, inputSample, (int)rotatedPoints.size(), minx, maxx, miny, maxy, output);
        cl::copy(output, std::begin<float>(scoreMap), std::end<float>(scoreMap));
        ...
        /* 从scoreMap中找最大相似度的姿势 */
        /* 保存全局最优姿势供后续使用 */
    }
    ...
    /* 根据初始姿势逐层求精 */
}