Hello world for using cuda for image processing

Question

I'm trying to get cuda to execute a kernel on my webcam video.

I would like it to, retrieve data from my webcam, send it to my GPU, process a kernel, then send the resulting image back to be dislpayed.

#include "cuda.h"
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <Windows.h>
#include "Bitmap.h"

#include "OpenCVTest.h"

#include "OpenCVTest.h"
#include <opencv2/opencv.hpp>

using namespace cv;

#define Pixel unsigned char


__global__ void TestKernel(unsigned char * img)
{
    int index = threadIdx.x + blockIdx.x * blockDim.x;
    img[index] = 100;
}

int main(void) 
{
    VideoCapture cap(0); 
    Mat input;
    Mat frame;
    Mat Output;
    cap >> frame;
    //cap >> Output;
    cvtColor(frame, Output, CV_BGR2GRAY);
    uchar *d_frame;
    size_t size = (int) (640 * 480);
    cudaMalloc((void **)&d_frame, size);

    namedWindow("Window",1);
    for(;;)
    {
        cap >> input; 
        cvtColor(input, frame, CV_BGR2GRAY);        

        cudaMemcpy(d_frame, frame.data, size, cudaMemcpyHostToDevice);

        TestKernel<<<640 * 480, 1>>>( d_frame );

        cudaMemcpy(Output.data, d_frame, size, cudaMemcpyDeviceToHost);

        imshow("Window", Output);
        if(waitKey(30) >= 0) break;
    }

    cudaFree(d_frame);

    return 0;
}

I've just written a very basic test kernel to start with. But it appears that the kernel is not being executed as the image that I get displays is simply the gray scale video from my webcam.

EDIT

As Robert sugested I've added some error checking by adding

gpuErrchk( cudaPeekAtLastError() );

After calling the kernel

Where gpuErrchk is

#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, char *file, int line, bool abort=true)
{
   if (code != cudaSuccess) 
   {
      fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
      if (abort) exit(code);
   }
}

GPUError

"invalid argument" is the error you would get if you tried to launch a kernel with grid size too large. — Robert Crovella, Jun 05 '14 at 05:25

score 3 · Accepted Answer · edited May 23 '17 at 12:11

3

640 * 480 = 307200

Unless you have compiled the code for, and are running on a cc 3.0 or higher GPU, this is not an acceptable choice for the first configuration parameter of the kernel:

    TestKernel<<<640 * 480, 1>>>( d_frame );

For pre-cc3.0 devices, the first parameter (i.e. Maximum x-dimension of a grid of thread blocks) is limited to 65535 for the first 2 dimensions of the dim3 quantity.

If you did proper cuda error checking you would find out that your kernel is not running (and/or other errors.) You can also try running your code with cuda-memcheck as a quick test.

edited May 23 '17 at 12:11

Community

1
1

answered Jun 05 '14 at 04:55

Robert Crovella

143,785
11
213
257

did you compile your code with an appropriate architecture switch for Quadro K4000 ? Did you add proper cuda error checking? Did you run your code with cuda-memcheck as a quick test? – Robert Crovella Jun 05 '14 at 05:08
After running the program in NVIDIA's visual profiler I get that same Ivalid argunment error message. – sav Jun 05 '14 at 05:32
Your code is broken if you don't compile for a cc3.0 or newer architecture. Why would you expect something different if you run it under visual profiler? Quadro K4000 is a cc3.0 GPU, so add something like `-arch=sm_30` to your compile command line, and restart your testing. – Robert Crovella Jun 05 '14 at 05:35
adding -arch=sm_30 to the command line seemed to do it, thanks :) – sav Jun 05 '14 at 05:47

Hello world for using cuda for image processing

1 Answers1