LLNL / RAJA

RAJA Performance Portability Layer (C++)
BSD 3-Clause "New" or "Revised" License
491 stars 103 forks source link

Improving RAJA integration with perf tools #1664

Open artv3 opened 5 months ago

artv3 commented 5 months ago

PR: https://github.com/LLNL/RAJA/pull/1662 propagates the kernel naming API to all back-ends. For CUDA and HIP it is well defined as we can use the ROCTX and NVTX tools to name our kernels. We currently don't have tool support for the other back-ends. An area of future investigation is how we can better integrate with tools and provide kernel naming when running on other backends.

rhornung67 commented 5 months ago

Maybe using caliper for other back-ends is a good option. Look at what we have and other options (plug-in interface to Sandia Perf Tools)