You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Apparently, CUPTI provides a C++/CUDA interface for retrieving various GPU metrics per kernel. In particular, we are interested in checking the achieved occupancy for a given workload, which is fortunately included in the list of metrics CUPTI covers. Sadly, CUPTI has no Python bindings so we need to manually construct a Python-C bridge to enable/disable CUPTI and fetch the correct metric values (unless we re-write the whole repo in C++... which is probably not going to happen).
The text was updated successfully, but these errors were encountered:
Apparently, CUPTI provides a C++/CUDA interface for retrieving various GPU metrics per kernel. In particular, we are interested in checking the achieved occupancy for a given workload, which is fortunately included in the list of metrics CUPTI covers. Sadly, CUPTI has no Python bindings so we need to manually construct a Python-C bridge to enable/disable CUPTI and fetch the correct metric values (unless we re-write the whole repo in C++... which is probably not going to happen).
The text was updated successfully, but these errors were encountered: