Flop/s measurement

Accelerated Computing CUDA CUDA Programming and Performance
1

I want to measure the performance of my application on a C870 in terms of Flops/seconds. I cannot figure out how to do it with the CUDA 3.1 Compute Profiler. How can I do it?

Also, does the Compute Profiler show the number of floating-point instructions executed?

Thank you in advance.

2

You count up the number of FLOPs in your code, and divide by the time taken for your kernel to execute.

3

You count up the number of FLOPs in your code, and divide by the time taken for your kernel to execute.