Remove thread-MPI limitation for GPU PP Halo exchange
authorGaurav Garg <gaugarg@nvidia.com>
Thu, 25 Mar 2021 08:19:37 +0000 (13:49 +0530)
committerMark Abraham <mark.j.abraham@gmail.com>
Tue, 13 Apr 2021 09:58:47 +0000 (09:58 +0000)
commit13678441531dc60059df5592e54a2f74900979e6
tree3b267510708c355abd5903d3fdab9d195335b51b
parent7f63c12ad7cd781e8a51079e13608b7cb61dea26
Remove thread-MPI limitation for GPU PP Halo exchange

Allows use of direct-GPU communication for PP halo exchange when
running with "real" MPI, including on multiple compute nodes,
through new CUDA-aware MPI communication code paths.

Implements part of #2891
Refs: #2915 #3960
src/gromacs/domdec/domdec.cpp
src/gromacs/domdec/gpuhaloexchange_impl.cu
src/gromacs/domdec/gpuhaloexchange_impl.cuh
src/gromacs/mdlib/sim_util.cpp
src/gromacs/mdrun/runner.cpp
src/gromacs/taskassignment/decidegpuusage.h