How to measure time in cuda kernel ...? [CUDA 4.0]

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I am using CUDA 4.0 and I need to accurately measure a time difference in CUDA kernel.

Currently I am using clock() function and note it is not accurate enough.

How to measure time in kernels accurately?

Note that question is not about measuring time from main() function … but from kernel …

Thank you.

2

What do you mean it is not accurate enough?

3

Your best bet if you are trying to profile your kernels will depend on what environment you are using. For example if you are using Windows/Visual Studio, then you can use the NSight tool (https://developer.nvidia.com/nvidia-nsight-visual-studio-edition)
to provide incredibly accurate timings of your kernels as well as useful profiling information such as GFLOPs, memory accesses and other metrics for finding performance bottlenecks.