biod.pnpi.spb.ru / alexxy / gromacs.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: 43b41cb) | patch
make use of CUDA stream priorities
authorSzilard Pall <pall.szilard@gmail.com>
Fri, 4 Oct 2013 00:33:04 +0000 (02:33 +0200)
committerGerrit Code Review <gerrit@gerrit.gromacs.org>
Tue, 15 Oct 2013 15:51:51 +0000 (17:51 +0200)
commit3ee2d276012042dcacb92a83cb0dbcba2888369a
treed9a17ba9c3b6408cec0e5b7116f73266800fd908tree | snapshot
parent43b41cb8cead1a67339e29ae0775dae2c23d0d3ecommit | diff
make use of CUDA stream priorities

CUDA 5.5 introduced steam priorities with 2 levels. We make use of this
feature by launching the non-local non-bonded kernel in a high priority
stream. As a consequence, the non-local kernel will preempt the local
one and finish first. This will improve performance in multi-node runs
by reducing the possibility of late arrival of non-local forces.

Change-Id: I4efc65546e4135f12006c0422e1fca42a788129f
src/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.cu diff | blob | history
Local GROMACS mirror with custom stuff