| commit | 272736bc4e81828bcd9d73e04e4e6e70a1712cf1 | |
| author | Berk Hess <hess@kth.se> | |
| Thu, 12 Jun 2014 13:01:40 +0000 (12 15:01 +0200) | ||
| committer | Gerrit Code Review <gerrit@gerrit.gromacs.org> | |
| Sat, 14 Jun 2014 20:27:46 +0000 (14 22:27 +0200) | ||
| tree | 8a195b6ab5b895b16e06bc569361d86d661a8d51 | treesnapshot (tar.gz zip) |
| parent | 75f84652037affa3e28fcb651a8769c672f736a9 | commitdiff |
Fixed PME bug with #OpenMP-threads a large prime
With hybrid MPI+OpenMP parallelization and the rank local FFT grid
size for a dimension not divisible by the number of OpenMP threads
in that dimension, a very small amount of the FFT grid overlap part
could set/added twice. This would only occur at low-medium MPI
parallelization with OpenMP thread counts with large prime factors,
which is practice means almost never. Even when it occured, no actual
differences in PME energies or forces were observed.
This issue was due to a leftover from when space was uniformly
divided over the grids iso assigning whole grid lines.
Fixes #1388.
Change-Id: I22904c7f55d2e96fc4b8cd1498af2087eaed47ac
With hybrid MPI+OpenMP parallelization and the rank local FFT grid
size for a dimension not divisible by the number of OpenMP threads
in that dimension, a very small amount of the FFT grid overlap part
could set/added twice. This would only occur at low-medium MPI
parallelization with OpenMP thread counts with large prime factors,
which is practice means almost never. Even when it occured, no actual
differences in PME energies or forces were observed.
This issue was due to a leftover from when space was uniformly
divided over the grids iso assigning whole grid lines.
Fixes #1388.
Change-Id: I22904c7f55d2e96fc4b8cd1498af2087eaed47ac
| src/mdlib/pme.c | diffblobblamehistory |