Measuring GFLOPS for a kernel

How do I measure the number of GFLOPS for a kernel? I can measure the time but how do I measure the number of calculations? Do I have to calculate it myself?

How do I measure the used memory bandwidth?

Hi,

Yes you have to calculate this by yourself. The new visual profiler from 2.2beta should help with the FLOPs and it also gives you the used bandwidth per kernel - which

is very useful.

eyal