I am experiencing unexpected behavior when using clWaitForEvents or clFinish. Following is my structure of host code during kernel launch.
for (i = 0; i<10;i++) {
Set kernel Args
...
..
..
clEnqueueNDRange(queue1, Kernel1, ....., &event1);
clEnqueueTask(queue2, kernel2,............&event2);
clEnqueueTask(queue3,kernel3,..........&event3);
start = time();
clFinish(queue3); or clWaitforevents(..., event3);
end = time();
Run time
= end() - start();
event1_time += getKernelStartEndTime(event1);
event2_time
+= getkernelstartendtime(event2);
event3_time += getKernelStartEndTime(event3);
Release events...
....
...
}
I obtain Runtime
(8ms) more than event3_time (4ms) in every iteration (factor of 2). I have tried using clFinish() as recommended in
https://www.alteraforum.com/forum/showthread.php?t=56633. But I obtain the same results. This doesnt let me launch the kernels of the next iteration just after the previous. There is lot of time gap between kernel launch of one iteration and next. Is it common to obtain high overhead?
Am I missing anything while launching kernels?
