Image grayscale in OpenCL - c++

Question

I want to transform an RGB-Image to a grayscale image.
My problem when I copy back the data the kernel returns zeros.
OpenCL code:
__kernel void grayscale(__global uchar * input, __global uchar * output)
{
int gid = get_global_id(0);
output[gid] = 0.0722 * input[gid][0] + 0.7152 * input[gid][1] + 0.2126 * input[gid][2];
}
Host code:
void RunKernel(char fileName[], char methodName[], Mat inputImg, Mat outputImg,
char outputLoc[], int mem_size){
/*
Initialisation of the device and read the kernel source.
*/
//Creating cl_mem objects for input and output. men_size is the image width*height
imgInMem = clCreateBuffer(img_context, CL_MEM_READ_ONLY,
mem_size * sizeof(uchar), NULL, &err);
imgOutMem = clCreateBuffer(img_context, CL_MEM_WRITE_ONLY,
mem_size * sizeof(uchar), NULL, &err);
//copy the data into cl_mem input
err = clEnqueueWriteBuffer(img_cmd_queue, imgInMem, CL_TRUE, 0, mem_size *sizeof(uchar),
&inputImg.data, 0, NULL, NULL);
//Create the program and load the kernel source to it
img_program = clCreateProgramWithSource(img_context, 1, (const char **) &kernel_src_str,
(const size_t *) &kernel_size, &err);
err = clBuildProgram(img_program, 1, &dev_id, NULL, NULL, NULL);
img_kernel = clCreateKernel(img_program, methodName, &err);
//Setting the kernel args
err = clSetKernelArg(img_kernel, 0, sizeof(cl_mem), (void *) &imgInMem);
err = clSetKernelArg(img_kernel, 1, sizeof(cl_mem), (void *) &imgOutMem);
//define the global size and local size
size_t global_work_size = mem_size;
size_t local_work_size = 256;
//Enqueue a command to execute a kernel on a device ("1" indicates 1-dim work)
err = clEnqueueNDRangeKernel(img_cmd_queue, img_kernel, 1, NULL, &global_work_size,
&local_work_size, 0, NULL, NULL);
err = clFinish(img_cmd_queue);
//Read back the result from device
err = clEnqueueReadBuffer(img_cmd_queue, imgOutMem, CL_TRUE, 0,
mem_size *sizeof(uchar), outputImg.data, 0, NULL, NULL);
/*
Release the necessary objects.
*/
}
After the clEnqueueReadBuffer if I write the values to the console it is all zeros. My outputImg is declared like this in the main:
Mat outImg(height,width,CV_8UC1,Scalar(0));
and call the method with this:
RunKernel("kernels/grayscale.cl","grayscale", inImg, outImg,"resources/grayscale_car_gpu.jpg", MEM_SIZE);

Answer 1

The problem is likely the 2D array syntax you're using:
0.0722 * input[gid][0] + 0.7152 * input[gid][1] + 0.2126 * input[gid][2]
What addresses do you think that is accessing exactly?
Instead, assuming you're trying to access sequential bytes as RGB (in BGR order, judging by the coefficient value), try:
0.0722 * input[3*gid+0] + 0.7152 * input[3*gid+1] + 0.2126 * input[3*gid+2]
You should add an "f" to the float constants (otherwise they are doubles, which are not supported on all devices).
You should add rounding from float back to uchar. So, together, something like:
convert_uchar_sat_rte(0.0722f * input[3*gid+0] +
0.7152f * input[3*gid+1] +
0.2126f * input[3*gid+2])
Finally, you're passing the same size buffer for the input and output images, but seemingly treating the input buffer as RGB, which is 3x larger than a single byte of monochrome. So you'll need to fix that in the host code.
Any time you're getting incorrect output from a kernel, simplify it to see if it is an input problem, a calculation problem, an output problem, or host cost issues. Keep narrowing it down until you've found your problem.