Task #2528
PME GPU tuning
Description
Currently PME kernels are preferring block sizes which were set long ago (c_[spread/gather/solve]MaxWarpsPerBlock).
These should be specialized for OpenCL. Actually, could be looked at again for CUDA as well. Actually, it should be a reocurring pre-release task, not just for PME :-) As long as anyone steps up.
Associated revisions
History
#1 Updated by Mark Abraham over 2 years ago
I suggest we re-work this redmine into that set of recurring tasks
#2 Updated by Gerrit Code Review Bot over 2 years ago
Gerrit received a related patchset '6' for Issue #2528.
Uploader: Mark Abraham (mark.j.abraham@gmail.com)
Change-Id: gromacs~master~I03062c908db6dad9bbc8c62accdc4707b03ff527
Gerrit URL: https://gerrit.gromacs.org/7861
#3 Updated by Paul Bauer about 2 years ago
Is this still targeted at 2019?
#4 Updated by Paul Bauer about 2 years ago
- Target version changed from 2019 to future
retargeted to future
Gather the PME GPU constants/macros in a single header
Those will be translated to defines by the OpenCL kernel compiler.
Refs #2453, #2528
Change-Id: I03062c908db6dad9bbc8c62accdc4707b03ff527