Performance issues on GPU

[smu...@uchicago.edu]

Hi Bogdan,

Thanks for all your help so far!  I have managed to get my split-step integrator working on the GPU, but I seem to be encountering some performance issues.  The split-step loop initially runs very fast, but then becomes 200x slower after about 20 passes through the loop.  It then stays at this speed for the rest of the loop.  I go into more detail about this in my stackoverflow question.  Is this caused by not being careful enough with reikna?  Or is there some way to address this problem with reikna?  Sorry that this question isn't directly related to reikna, but I thought you might be a great resource!

Thanks again,

Seth

[Bogdan Opanchuk]

Hi Seth,

I replied to you on SO, it is quite a long answer because I also had some comments about the code you posted.

Briefly, you were not synchronizing the queue after each iteration (with queue.finish() or thr.synchronize()), so you only measured the time of adding kernels to the queue. When the queue was saturated, you started getting the times that were close to reality.