0

I have two questions to ask why the API cufftPlan3d(plan, nx,ny,nz,c2c) takes around 300 milliseconds to execute?

is it a gpu operation or cpu operation?

Can anybody give me an explanation in detail. I could not find a good explanation anywhere.

talonmies
  • 70,661
  • 34
  • 192
  • 269
Ankith Reddy
  • 1
  • 3
  • 1
    Is this the first GPU (i.e. cufft, CUDA) call in your application? If not, is it the first cufft call? If you perform cufftPlan3d more than once, are the subsequent calls also taking 300 milliseconds? This is perhaps just an artifact of CUDA or cufft initialization - impacting the first call only. An ordinary cufft planning operation makes GPU memory allocations under the hood, so it involves both CPU activity and GPU activity (unless you take control of the memory allocations yourself within cufft.) – Robert Crovella Mar 15 '18 at 15:25
  • Yeah , am using cufftplan3d only once in my application . However for the second time cufftplan3d call will not take time? – Ankith Reddy Mar 15 '18 at 17:28
  • The first cufft call and the first cuda call in your application are likely to incur additional start-up overheads. Subsequent calls are usually shorter. – Robert Crovella Mar 15 '18 at 17:40
  • thanks robert for the help – Ankith Reddy Mar 20 '18 at 06:31

0 Answers0