Ask Your Question
1

Custom Kernel with GpuMat No Effect

asked 2014-07-03 03:48:27 -0600

Jon1111 gravatar image

Hi so I'm a beginner trying to write a pretty simple kernel. I just want to take the arc cosine of the pixels of an image and then return those. I'm using GpuMat objects but I cannot get the kernel to do anything.

//.cpp code
void gpu_acos(const gpu::GpuMat &src, gpu::GpuMat &dst){
    float* srcptr = (float *)src.data;
    float* dstptr = (float *)dst.data;
    acos_func(srcptr,dstptr, src.step,dst.step, src.cols,src.rows);
    return;

}

bool test_acos(){
    Mat input = imread("corgi.jpg",0);
    int rows = input.rows;
    int cols = input.cols;
    Size in_size(rows,cols);
    gpu::GpuMat src, dst;
    src.upload(input);
    dst.create(in_size,CV_32FC1);
    gpu_acos(src,dst);
    cout << "baack from gpu call\n";

    Mat test_out;
    dst.download(test_out);
    cout << "this is the test: " << test_out << endl;
    return true;

}

 //.cu code
#include "custom_kernels.h"
#include <iostream>
using namespace std;
using namespace cv; 
__global__ void acosKernel(const float* srcptr, float* dstptr, size_t srcstep, size_t dststep, int cols, int rows){
    int rowInd = blockIdx.y*blockDim.y+threadIdx.y;
    int colInd = blockIdx.x*blockDim.x+threadIdx.x;
    if(rowInd >= rows || colInd >= cols)
            return;
    const float* rowsrcPtr = srcptr+rowInd*srcstep;
    float* rowdstPtr = dstptr+rowInd*dststep;

    float pixVal = rowsrcPtr[colInd];

    if( ((int) pixVal % 90)==0)
            rowdstPtr[colInd]=0.0;
    else
            rowdstPtr[colInd] = acos(pixVal);

}

int divUp(int a, int b){ 
    return (a+b-1)/b;

}

void acos_func(const float* srcptr, float* dstptr, size_t srcstep, size_t dststep, int cols, int rows){
    dim3 blDim(32,8);
    dim3 grDim(divUp(cols,blDim.x),divUp(rows,blDim.y));

    acosKernel<<<grDim, blDim>>>(srcptr,dstptr,srcstep,dststep,cols,rows);
    cudaDeviceSynchronize();

} ~

Even when I've replaced the acos logic with something simple (for example all 0's) I get as output the input image. I'm sure I'm missing something obvious, but any help would be greatly appreciated. Thanks in advance!

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted
1

answered 2014-07-06 05:29:23 -0600

You are using a float pointer in the CPP function and the kernel, but the GpuMat src is a char image as you are uploading it from input, which is loaded from a JPG image. Use a char pointer instead of a float for src (not dst which is created as a float image) and it should get fine. This doesn't explain (or does it?) the behavior you get when you change the logic, but this is definitely an issue.

edit flag offensive delete link more

Comments

Hi thanks for the response. I changed the src pointer to a char pointer but it still does not work. I still could not get the dstptr values to be zero. I tried by both defining the GpuMat dst as a CV_8UC1 type and using unsigned char* and also as a CV_32FC1 using float* but neither type had an impact. This seems pretty basic so am I just missing something really obvious?

Jon1111 gravatar imageJon1111 ( 2014-07-07 04:27:10 -0600 )edit

Nevermind I was not stepping correctly when I switched from unsigned chars to floats. Thanks for the tip that GpuMat src is a char image

Jon1111 gravatar imageJon1111 ( 2014-07-08 03:24:33 -0600 )edit

Question Tools

Stats

Asked: 2014-07-03 03:48:27 -0600

Seen: 1,952 times

Last updated: Jul 06 '14