We need to use both the CPU and GPU of our Ivy Bridge processor at the same time. However, in Intel OpenCL SDK, the kernel launch call is not asynchronous. In other words, clEnqueueNDRangeKernel waits for the GPU execution to be completed, which is against the OpenCL standard. We have tried the latest Intel OpenCL SDKs (2012 and 2013). Is there a workaround?
It is easy to see for any code. For instance, if you comment out clFinish call after the kernel launch in one of the Intel opencl samples (e.g. BitonicSort.cpp:286), the timer shows the same time and the results are still correct. This means that the kernel launch had not been asynchronous.
I think I found the problem of my code. The NDRange kernel launch was inside a loop so the runtime had to flush the queue eventually. I need to reorganize the code. Sorry for taking your time.