Passing Mat to OpenCL Kernels causes Segmentation fault

Question

I want to pass an OpenCL Mat to a selfwritten OpenCL Kernel for a FGPA (doesnt´t support the OpenCV OpenCL).

Host- Code:

Mat img = imread( "template.jpg", IMREAD_GRAYSCALE );
Mat output(img.rows, img.cols, CV_8UC1);
// Program, Context already declared
// Create Kernel
cl_kernel kernel = NULL;
kernel = clCreateKernel(program, "copy", &status);
// Create Command Queue and associate it with the device you want to execute on
cl_command_queue cmdQueue;
cmdQueue = clCreateCommandQueue(context,devices[0], 0,  &status);

// Buffer, prob i do something wrong here
cl_mem buffer_img = clCreateBuffer(context,CL_MEM_READ_ONLY, sizeof(uint) * img.cols * img.rows,    NULL,&status);
cl_mem buffer_outputimg = clCreateBuffer(context,CL_MEM_WRITE_ONLY, sizeof(uint) * img.cols * img.rows,NULL,&status);

status = clEnqueueWriteBuffer(cmdQueue, buffer_img,CL_FALSE,0,sizeof(uint) * img.cols * img.rows,&img,0,NULL,NULL);
// set kernel arguments
status = clSetKernelArg(kernel,0,sizeof(cl_mem),&buffer_img);
status = clSetKernelArg(kernel,1,sizeof(cl_mem),&buffer_outputimg);

size_t globalWorkSize[2];
globalWorkSize[0] = img.cols;
globalWorkSize[1] = img.rows;
status = clEnqueueNDRangeKernel(cmdQueue,kernel,2,NULL, globalWorkSize, NULL,0, NULL,NULL);
clEnqueueReadBuffer(cmdQueue,buffer_outputimg,CL_TRUE,0,sizeof(uint) * img.cols * img.rows, &output,    0,  NULL,   NULL);

//stop cpu till queue is finish
clFinish(cmdQueue);

Kernel-Code:

__kernel void copy(__global  uchar *  input, __global  uchar *  output) 
{
    const int x = get_global_id(0);
    const int y = get_global_id(1);
    //copy
    output[y * get_global_size(0) + x] = input[y * get_global_size(0) + x] ;
}

When excecuting it on the FPGA i get a Segmentation fault, whichs is propably due the wrong handling with the OpenCV Mat.

EDIT: Edited Host-Code as suggested by api55 solved the problem:

Mat img = imread( "scene.jpg", IMREAD_GRAYSCALE );
Mat output(img.rows, img.cols, CV_8UC1);
// Program, Context already declared
// Create Kernel
cl_kernel kernel = NULL;
kernel = clCreateKernel(program, "copy", &status);
// Create Command Queue and associate it with the device you want to execute on
cl_command_queue cmdQueue;
cmdQueue = clCreateCommandQueue(context,devices[0], 0,  &status);
checkError(status, "Failed to create commadnqueue");

// Buffer
cl_mem buffer_img = clCreateBuffer(context,CL_MEM_READ_ONLY, sizeof(uchar) * img.cols * img.rows,   NULL,&status);
cl_mem buffer_outputimg = clCreateBuffer(context,CL_MEM_WRITE_ONLY, sizeof(uchar) * img.cols * img.rows,NULL,&status);
checkError(status, "Failed to create buffer_mask");

status = clEnqueueWriteBuffer(cmdQueue, buffer_img,CL_FALSE,0,sizeof(uchar) * img.cols * img.rows,img.data,0,NULL,NULL);
checkError(status, "Failed to enqueue buffer_img");


status = clSetKernelArg(kernel,0,sizeof(cl_mem),&buffer_img);
status = clSetKernelArg(kernel,1,sizeof(cl_mem),&buffer_outputimg);

size_t globalWorkSize[2];
globalWorkSize[0] = img.cols;
globalWorkSize[1] = img.rows;
status = clEnqueueNDRangeKernel(cmdQueue,kernel,2,NULL, globalWorkSize, NULL,0, NULL,NULL);
clEnqueueReadBuffer(cmdQueue,buffer_outputimg,CL_TRUE,0,sizeof(uchar) * img.cols * img.rows, output.data,0,NULL,NULL);

imwrite("output.jpg", output);

Answer 1

I do not have much experience with opencl, but i think it is an opencv/c++ problem.

The opencv mat data lies in img.data which is an uchar* of the size sizeof(T) * channels * rows * cols.

Usually, T is uchar when loading images, and channels is 3 (unless that is a greyscale img). 3 channel uchar is 24 bits per pixel and greyscale (as you are loading) is 8 bits per pixel and you are using uint which is size of 32 bits. At some point it will go outside the memory and do the segmentation error. Also, if you do not use the data pointer in the structure, you may be copying the header information and just the pointer to the data and not the data itself.

I suggest you to change &img in:

status = clEnqueueWriteBuffer(cmdQueue, buffer_img,CL_FALSE,0,sizeof(uint) * img.cols * img.rows,&img,0,NULL,NULL);

to img.data

Finally, you need to have the correct data. I am not sure if opencl may use uchar, but if it can't, change the cv::Mat to another type like this:

img.convertTo(img, CV_32S);

After loading the image. This will change it to int... opencv does not support matrices with unsigned int... just make sure to change it accordingly in the other places (i.e. sizeof(uint)) and if you convert the input, remember to create the output with the same type.

If you prefer float, use CV_32F and if you like double CV_64F.