OpenCL tasks hang at submitted - opencl

Question

Ich have a strange problem. I just started with openCL development and implemented a simple matrix multiplication:
void MatrixMult(float *M, float *N, float *P, int width) {
cl_context context = core.getContext();
cl_command_queue commandQueue = core.getCommandQueue();
// create the kernel
cl_kernel kernel = core.createKernel("MatrixMultKernel");
// reserve memory for i/o
int size = width * width * sizeof(float);
cl_mem Md = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, size, M, NULL);
cl_mem Nd = clCreateBuffer(context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, size, N, NULL);
cl_mem Pd = clCreateBuffer(context, CL_MEM_WRITE_ONLY , size, NULL, NULL);
// define kernel args
clSetKernelArg(kernel, 0, sizeof(cl_mem), &Md);
clSetKernelArg(kernel, 1, sizeof(cl_mem), &Nd);
clSetKernelArg(kernel, 2, sizeof(cl_mem), &Pd);
clSetKernelArg(kernel, 3, sizeof(int), &width);
// create kernel instances
size_t globalSize[] = {static_cast<size_t>(width), static_cast<size_t>(width)};
cl_int kernelError = clEnqueueNDRangeKernel(commandQueue, kernel, 2, NULL, globalSize, NULL, 0, NULL, NULL);
if (kernelError != CL_SUCCESS) {
NSLog(#"Error executing kernel");
}
// read back results
clEnqueueReadBuffer(commandQueue, Pd, false, 0, size, P, 0, NULL, &readEvent);
clSetEventCallback(readEvent, CL_COMPLETE, &eventCallback, NULL);
}
My problem is that the callback is never called. I test with small inputs which should finish immediately.
The strange this is whenever I call the function (MatrixMult) a second time the callback fires as it should. The same thing for every consecutive call. When I check the event while it is not firing it always has the status CL_SUBMITTED. What is going on here?

Answer 1

You probably need to add a call to clFlush at the end of your code. Though the OpenCL spec may not clearly list that as a requirement, I find it is the case on my system. It is waiting for either clFlush of a command that does an implicit clFlush.