C++ GpuMat-访问自定义内核中的2通道浮点数据

据我所知，

cv:：cuda:：PtrStep

用于将

GpuMat

数据直接传递到自定义内核。我找到了单通道访问的示例，但我的示例是双通道mat（

CV_32FC2

）。在这种情况下，我试图获得复数绝对平方值，其中复数值被编码为：实部是第一个平面，虚部是给定

Mat

的第二个平面

我试过：

__global__ void testKernel(const cv::cuda::PtrStepSz<cv::Vec2f> input, cv::cuda::PtrStepf output)
{
    int x = blockIdx.x * blockDim.x + threadIdx.x;
    int y = blockIdx.y * blockDim.y + threadIdx.y;

    if (x <= input.cols - 1 && y <= input.rows - 1 && y >= 0 && x >= 0)
    {
        float val_re = input(x, y)[0];
        float val_im = input(x, y) [1];
        output(x, y) = val_re * val_re + val_im * val_im;
    }
}

---

但是我想知道是否有一个“更干净”的解决方案，像Vec2f一样。

如果您想使用内核的单个输入，您可以

---

现在，您可以将

d_mat_flat

作为

PtrStepSzf

传递到内核。您可以使用原始数据类型访问自定义CUDA内核中的

GpuMat

数据。e、 g.CUDA运行时提供的

float2

类型可用作

cv:：Vec2f

的部分替换。下面是一个示例代码，演示如何使用原始数据类型访问

GpuMat

数据

#include <iostream>
#include <cuda_runtime.h>
#include <opencv2/opencv.hpp>

using std::cout;
using std::endl;

__global__ void kernel_absolute(float2* src, float* dst, int rows, int cols, int iStep, int oStep)
{
    int i = blockIdx.y * blockDim.y + threadIdx.y; //Row number
    int j = blockIdx.x * blockDim.x + threadIdx.x; //Column number

    if (i<rows && j<cols)
    {
        /* Compute linear index from 2D indices */
        int tidIn = i * iStep + j;
        int tidOut = i * oStep + j;

        /* Read input value */
        float2 input = src[tidIn];

        /* Calculate absolute value */
        float output = sqrtf(input.x * input.x + input.y * input.y);

        /* Write output value */
        dst[tidOut] = output;
    }
}

int main(int argc, char** argv)
{
    /* Example to compute absolute value of each element of a complex matrix */
    int rows = 10;
    int cols = 10;
    int input_data_type = CV_32FC2; //input is complex
    int output_data_type = CV_32FC1; //output is real

    /* Create input matrix on host */
    cv::Mat input = cv::Mat::zeros(rows,cols,input_data_type) + cv::Vec2f(1,1) /* Initial value is (1,1) */;

    /* Display input */
    cout<<input<<endl;

    /* Create input matrix on device */
    cv::cuda::GpuMat input_d;
    /* Copy from host to device */
    input_d.upload(input);

    /* Create output matrix on device */
    cv::cuda::GpuMat output_d(rows,cols, output_data_type);

    /* Compute element step value of input and output */
    int iStep = input_d.step / sizeof(float2);
    int oStep = output_d.step / sizeof(float);

    /* Choose appropriate block size */
    dim3 block(8,8);

     /* Compute grid size using input size and block size */
    dim3 grid ( (cols + block.x -1)/block.x, (rows + block.y -1)/block.y );

    /* Launch CUDA kernel to compute absolute value */
    kernel_absolute<<<grid, block>>>( reinterpret_cast<float2*>(input_d.data), reinterpret_cast<float*>(output_d.data), rows, cols, iStep, oStep );

    /* Check kernel launch errors */
    assert( cudaSuccess == cudaDeviceSynchronize() );

    cv::Mat output;

    /* Copy results from device to host */
    output_d.download(output);

    /* Display output */
    cout<<endl<<output<<endl;

    return 0;
}

#包括
#包括
#包括
使用std：：cout；
使用std：：endl；
__全局无效内核绝对值（float2*src、float*dst、int行、int列、int iStep、int oStep）
{
int i=blockIdx.y*blockDim.y+threadIdx.y；//行号
int j=blockIdx.x*blockDim.x+threadIdx.x；//列号
如果（你可以使用<代码>浮点2 < /代码>代替<代码> CV::VEC2F。也<代码>输入（x，y）< /代码>应该是代码>输入（y，x）< /代码>，因为第一个参数是行和第二个列。我喜欢它。没有CV：：拆分为分离的平面-没有数据复制。@迈克尔逊……如果解决了这个问题，你可以考虑接受答案：）。
__global__ void testKernel(const cv::cuda::PtrStepSzf re, const cv::cuda::PtrStepSzf im, cv::cuda::PtrStepf output)

// test image
Mat h_mat(Size(50,50),CV_32FC2,Scalar(0.0));

// Mat::reshape takes number of channels and rows, for your example 1,1
Mat h_mat_flat = h_mat.reshape(1,1);

// to upload to gpu
GpuMat d_mat_flat(h_mat_flat.size(), h_mat_flat.type());
d_mat_flat.upload(h_mat_flat);

#include <iostream>
#include <cuda_runtime.h>
#include <opencv2/opencv.hpp>

using std::cout;
using std::endl;

__global__ void kernel_absolute(float2* src, float* dst, int rows, int cols, int iStep, int oStep)
{
    int i = blockIdx.y * blockDim.y + threadIdx.y; //Row number
    int j = blockIdx.x * blockDim.x + threadIdx.x; //Column number

    if (i<rows && j<cols)
    {
        /* Compute linear index from 2D indices */
        int tidIn = i * iStep + j;
        int tidOut = i * oStep + j;

        /* Read input value */
        float2 input = src[tidIn];

        /* Calculate absolute value */
        float output = sqrtf(input.x * input.x + input.y * input.y);

        /* Write output value */
        dst[tidOut] = output;
    }
}

int main(int argc, char** argv)
{
    /* Example to compute absolute value of each element of a complex matrix */
    int rows = 10;
    int cols = 10;
    int input_data_type = CV_32FC2; //input is complex
    int output_data_type = CV_32FC1; //output is real

    /* Create input matrix on host */
    cv::Mat input = cv::Mat::zeros(rows,cols,input_data_type) + cv::Vec2f(1,1) /* Initial value is (1,1) */;

    /* Display input */
    cout<<input<<endl;

    /* Create input matrix on device */
    cv::cuda::GpuMat input_d;
    /* Copy from host to device */
    input_d.upload(input);

    /* Create output matrix on device */
    cv::cuda::GpuMat output_d(rows,cols, output_data_type);

    /* Compute element step value of input and output */
    int iStep = input_d.step / sizeof(float2);
    int oStep = output_d.step / sizeof(float);

    /* Choose appropriate block size */
    dim3 block(8,8);

     /* Compute grid size using input size and block size */
    dim3 grid ( (cols + block.x -1)/block.x, (rows + block.y -1)/block.y );

    /* Launch CUDA kernel to compute absolute value */
    kernel_absolute<<<grid, block>>>( reinterpret_cast<float2*>(input_d.data), reinterpret_cast<float*>(output_d.data), rows, cols, iStep, oStep );

    /* Check kernel launch errors */
    assert( cudaSuccess == cudaDeviceSynchronize() );

    cv::Mat output;

    /* Copy results from device to host */
    output_d.download(output);

    /* Display output */
    cout<<endl<<output<<endl;

    return 0;
}