CUPTI unified and driver API

Why aren’t CUPTI unified counters updated when calling cuMemAllocManaged (only cudaMallocManaged)?
Would intercepting cuMemAllocManagedand and calling cudaMallocManaged (eliminating recursions) work?

Hi menny,
CUPTI counters are updated for both cuMemAllocManaged() and cudaMallocManaged() APIs.

Please note that migrations and page faults are triggered only after the managed allocations are accessed on the CPU/GPU.

So that is when CUPTI counters will get updated.

Can you share a snippet of code where you see CUPTI counters not getting updated with cuMemAllocManaged() APIs?

Sorry for not replying - I left this project for a while. Anyway, CUPTI is no longer needed for my work, we just need counters and we get all this information from nvprof and Nsight.