Fix OpenCL gather reduction
authorSzilárd Páll <pall.szilard@gmail.com>
Fri, 15 Feb 2019 19:14:16 +0000 (20:14 +0100)
committerMark Abraham <mark.j.abraham@gmail.com>
Thu, 15 Aug 2019 19:46:18 +0000 (21:46 +0200)
commita19dd7d50c7cedabf746272c2d378dd4b9dbcefb
tree5f37479bba9c4e5366b7460f1f57df58dd4c00c4
parent0a4ca2c479237e915d96bcff8500c282a6ec3050
Fix OpenCL gather reduction

On >=16-wide execution it is correct (narrower is checked and excluded
during compilation).

TODO: Consider changing the default on NVIDIA & Intel where offloading
PME is generally not advantageous to performance.

Addresses part of #2519
Refs #2453 #2516

Change-Id: I24beaaeea096954ba32b3a80251945a9d82a3c05
docs/release-notes/2020/major/portability.rst
docs/user-guide/mdrun-performance.rst
src/gromacs/ewald/pme.cpp
src/gromacs/ewald/pme_gather.clh
src/gromacs/gpu_utils/gpu_utils.h
src/gromacs/gpu_utils/gpu_utils_ocl.cpp