updating buffer in opencl after each execution - c

Question

So i have written a code and it works, now what i want to do is call clEnqueueNDRangeKernel() multiple times and after each execution i want to update a buffer (buffer Y) with that output. i have written the following code and i want to know if its correct for that function.
i did not write a seperate setkernelArg() command for it.
for (int a = 0; a < 100; a++)
{
ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, globalws, NULL, 0, NULL, NULL);
if (ret != CL_SUCCESS) {
printf("Failed to enqueueNDRangeKernel.\n");
exit(1);
}
clEnqueueReadBuffer(command_queue, bufferC, CL_TRUE, 0, M*N * sizeof(float), (void *)C, 0, NULL, NULL);
clEnqueueWriteBuffer(command_queue, bufferY, CL_TRUE, 0, 1 * N * sizeof(float), (void *)C, 0, NULL, NULL);
for (int i = 0; i < N; i++) {
printf("%f, ", C[i]);
}
}

Answer 1

You should wait for every OpenCL Api call. Create Events for every call. So you are sure that every execution has finished before you start the next one. For example its possible that the kernel do some computation on the GPU, but at the same time you start the clEnqueueReadBuffer and you read the Output-Buffer before the Kernel has finished to write to the Output-Buffer. It's also possible that you write to the GPU before clEnqueueReadBuffer has finished.
An OpenCL Api call starts the execution on the GPU, but the Host Programm also goes on.
With the Events, your programm could look like this:
cl_event evKernel, evReadBuf, evWriteBuf;
for(int a = 0; a < 100; a++)
{
ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL, globalws, NULL, 0, NULL, &evKernel);
if (ret != CL_SUCCESS) {
printf("Failed to enqueueNDRangeKernel.\n");
exit(1);
}
clWaitForEvents(1, &evKernel);
clEnqueueReadBuffer(command_queue, bufferC, CL_TRUE, 0, M*N * sizeof(float), (void *)C, 0, NULL, &evReadBuf);
clWaitForEvents(1, &evReadBuf);
clEnqueueWriteBuffer(command_queue, bufferY, CL_TRUE, 0, 1 * N * sizeof(float), (void *)C, 0, NULL, &evWriteBuf);
clWaitForEvents(1, &evWriteBuf);
for (int i = 0; i < N; i++) {
printf("%f, ", C[i]);
}
}
With the events the execution time for the loop will grow.
In the clEnqueueWriteBuffer call you write from Host-Memory ( M*N*sizeof(float) ) that is bigger than the Device-Buffer (1*N*sizeof(float) ).(Maybe you mean(M*N*sizeof(float))?) A programm on the Host-Side would crash (invalid memory access), but OpenCL does not complain about it and copy data. I'm not sure, but this could cause problems in the future.
I don't know what your kernel does, but it would be better if the kernel only write data to the Output-Buffer that you needed on the Host-Side. You copy bufferC to C, but only a part of C to bufferY, which seems to be the input for the next kernel. Maybe you can make changes in the kernel.
Copy data from Host to Device or Device to Host is a expensive part. So for Performance reasons you should not copy data that you don't need for further computations.