search

stfc / PSycloneBench

Various benchmarks used to inform PSyclone optimisations
BSD 3-Clause "New" or "Revised" License
6 stars 5 forks source link

NemoLite2D Kokkos implementation maintain Views between PSy-layer invokations. #54

Closed sergisiso closed 3 years ago

sergisiso commented 3 years ago

The Views are stored in the infrastructure (dl_esm_inf) device_pointer fields and can be re-used later (potentially also from different invocations). This solves the main performance issue the Views implementation had and closes issue #31

It still requires to copy back to the host some data at the end of each invoke. I am exploring a solution with an infrastructure callback function in another PR.

The current results are below (all impl now have the momentum calculation as hotspot but the GPUs are still far from saturating the memory bandwidth):

image

sergisiso commented 3 years ago

@arporter @rupertford This is ready for review.

sergisiso commented 3 years ago

@arporter Ready for the next review. I updated the README with the libgfortran note but not much about the general implementation because the description is still valid. Note that some implementation details will change once again the the follow up PR #57 .

arporter commented 3 years ago

Thanks for tweaking. I'm happy now :-)