CUDA cupti profiler API and compute capability 7.0

[Daniel]

The cupti profiler API metrics are not available anymore for CUDA device with CC 7.0 whereas the profiler API is supported on CC 7.0 CUPTI v11.7 - 2.7. CUPTI Profiling API.
Related to pull request #278

[Giuseppe Congiu]

This is a decision that was made with PR #278, as you pointed out. The CUPTI event API is used up to (and including) CC 7.0. After that, the CUPTI profiler API is used instead. The motivation is that CC 7.0 is the first CC for which the CUPTI profiler API is also available; We wanted to carry the CUPTI event API over to CC 7.0 as that guarantees a smoother transition and (hopefully) provides better stability to the component.

[Giuseppe Congiu]

Is there any major issue with this approach that, maybe, we missed in PR #278?

[Daniel]

Thank you for the explanation, that seems legitimate to me.
There is no major issue that I can think of, it’s just that we will be running our code on GPUs with CC 7.0 and higher, so it’s more convenient to us to use the profiler API in every case.

[Daniel]

Is there any easy way of letting the user choose the API, when running configure for instance ?

[Giuseppe Congiu]

I have “reverted” the change and now the user can select the event vs profiler API through the PAPI_CUDA_ROOT: https://bitbucket.org/icl/papi/pull-requests/298

[Daniel]

Great, thank you!