Cause of "Dropped Invalid Data" and "Unable to Calculate All Metrics"

I’m receiving a couple warning pop-ups, and I can’t figure out their cause. When I first run my program in nvvp, I get the following dialog:

Dropped Invalid Data
The start and end timestamps on 5 kernels, memcpys, and other collected profile data are invalid. Those profiling records have been dropped and will not be displayed in the timeline.

Ok, I say. Then when I run Examine Kernel Performance, I get a different dialog after the program finishes:

Unable to Calculate All Metrics
Some collected events, metrics, or source-level results could not be associated with the session timeline. This may prevent event, metric and source-level results from being assigned to some kernels.

What could cause these errors?

I’ve read this related post, but I’m already using CUDA 5.5:
https://devtalk.nvidia.com/default/topic/524136/unable-to-calculate-all-metrics/

Other details:

  • I’m using a GeForce GTX 650 Ti.
  • The program that I’m running is the cfd benchmark for the Rodinia benchmark suite. I’ve narrowed down the cause to one of the four kernels, cuda_compute_flux. When this kernel is not called, the pop-ups don’t occur.
  • -lineinfo and -arch=sm_20 are included in the compile flags.

Thank you in advance for any suggestions. Let me know if other details would help. In general, though, I don’t even know what can cause these errors – my particular context aside.

I have the same problem.

In fact, I added cudaDeviceReset() before exiting and the problem was gone.

hello vvolkov,I have the same problem and as you sayed,I added cudaDeviceReset() before exiting,but the problem is still.
could you help me?

Hi jockey,How did you resolve your problem?

In my case, there are too many kernels that nvprof can handle. If I take the kernels out and profile them individually, there is no error like

Some collected events, metrics, or source-level results could not be associated with the session timeline. This may prevent event, metric and source-level results from being assigned to some kernels.