Nsight cuda trace does not work

Developer Tools Nsight Visual Studio Edition
1

Hi, I am trying to use the cuda trace in nsight but when ever i try to do it, it only displays the api calls and nothing else. I am running the latest version of nsight and toolkit 5.0. I am also running on a K20. Is there something special I have to do to get cuda trace to work?

thnx

2

Oh also, running in visual studio 10

3

CUDA kernel launches and memory copies are asynchronous to the CPU. Nsight trace tool maintains a list of outstanding operations and writes out the event information when the following occurs:

  • CUDA context event buffer is full (1000s of launches?)
  • program calls cudaDeviceSynchronize()
  • program calls cudaDeviceReset()

If the application is not calling cudaDeviceSynchronize or cudaDeviceReset (at completion) then it is likely the data has been collected but the tool was not presented an opportunity to write out the data.

4

A cuda dev sync is called after each kernel lunch

5

Adding a cudaDeviceReset tot the end of my program fix the problem, thanks