biod.pnpi.spb.ru / alexxy / gromacs.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: 3be6d84) | patch
Remove thread-MPI limitation for GPU direct PME-PP communication
authorGaurav Garg <gaugarg@nvidia.com>
Wed, 14 Apr 2021 05:50:07 +0000 (11:20 +0530)
committerMark Abraham <mark.j.abraham@gmail.com>
Fri, 16 Apr 2021 13:10:08 +0000 (13:10 +0000)
commit7f2aeec0cb3aee77b7be2c14be7ec63315b87045
tree0fe86f0ee12f51617690fccf3adbe23a14f5399ctree | snapshot
parent3be6d84a8af77a4aeb91e41c3110e56d22bd4709commit | diff
Remove thread-MPI limitation for GPU direct PME-PP communication

Allows use of direct-GPU communication for PP-PME communication 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
18 files changed:
docs/install-guide/index.rst diff | blob | history
src/gromacs/ewald/pme_coordinate_receiver_gpu.h diff | blob | history
src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cpp diff | blob | history
src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.cu diff | blob | history
src/gromacs/ewald/pme_coordinate_receiver_gpu_impl.h diff | blob | history
src/gromacs/ewald/pme_force_sender_gpu.h diff | blob | history
src/gromacs/ewald/pme_force_sender_gpu_impl.cpp diff | blob | history
src/gromacs/ewald/pme_force_sender_gpu_impl.cu diff | blob | history
src/gromacs/ewald/pme_force_sender_gpu_impl.h diff | blob | history
src/gromacs/ewald/pme_only.cpp diff | blob | history
src/gromacs/ewald/pme_only.h diff | blob | history
src/gromacs/ewald/pme_pp.cpp diff | blob | history
src/gromacs/ewald/pme_pp_comm_gpu.h diff | blob | history
src/gromacs/ewald/pme_pp_comm_gpu_impl.cpp diff | blob | history
src/gromacs/ewald/pme_pp_comm_gpu_impl.cu diff | blob | history
src/gromacs/ewald/pme_pp_comm_gpu_impl.h diff | blob | history
src/gromacs/ewald/pme_pp_communication.h diff | blob | history
src/gromacs/mdrun/runner.cpp diff | blob | history
Local GROMACS mirror with custom stuff