Can anyone explain the difference in time?

Accelerated Computing CUDA CUDA Programming and Performance
1

I tried to time the duration of a kernel and the memcpy to and from the device, with and without the kernel running.

Here are the results:

with kernel:
duration of memcpy (dynamic array of 8388608 ints) host to device: 33ms
duration of kernel: 0.074ms
duration of memcpy (dynamic array of 1 int) device to host: 30ms;

without kernel:
duration of memcpy (dynamic array of 8388608 ints) host to device: 33ms
duration of memcpy (dynamic array of 1 int) device to host: 0.032ms

How come the duration of memcpy changes so drastically when the kernel is removed?

2

Are you sure you cudaThreadSynchronize() after invoking your kernel? Seems it is not hte case and you’re getting wrong results – your kernel running time is accounted as memcpy because memcpy performs implicit synchronization.

3

Thanks I’ll try with cudaThreadSynchronize().