OpenCL kernel execution does not start until clFinish or clWaitForEvents is called

Question

I am trying to run a kernel on the GPU and do additional computation on the host (CPU). I see this effect:

only the kernel needs around 2000 ms:

clEnqueueNDRangeKernel ...

clFinish (or clWaitForEvents, I tried both)

I simulated additional computation on the CPU with sleep(10):

clEnqueueNDRangeKernel ...

sleep(10);

clFinish (or clWaitForEvents)

In theory the kernel should run on GPU and after the 10 sec sleep the kernel should be finished. But time measuring said it all needs 12000ms instead of 10000.

Does clFinish or clWaitForEvents invoke the kernel to start or did I miss something?

I'm using an AMD Fusion CPU/GPU und Linux.

Thanks a lot.

How large is your data? Is it possible that 2000ms represents the time needed to transfer your data from the GPU to the CPU? — user1202136
The data is under 1 MB. But it shouldn't be in my measuring. I copy data to the device, start measuring, execute the kernel + sleep + clFinish, stop measuring, copy data back to host — Tomas
How exactly are you measuring? Do you copy your data to the device synchronously? — reima
yeah, I copy the data with clEnqueueWriteBuffer(command_queue, Ddata, CL_TRUE, ...) and I use the default in order command queue. For time measuring I take gettimeofday(). — Tomas

user1202136 user1202136 · Accepted Answer · 2012-09-20T11:05:24

Try calling clFlush right after clEnqueueNDRangeKernel:

clFlush

Issues all previously queued OpenCL commands in a command-queue to the device associated with the command-queue.

http://www.khronos.org/registry/cl/sdk/1.0/docs/man/xhtml/clFlush.html

OpenCL kernel execution does not start until clFinish or clWaitForEvents is called

2 Answers