| commit | 9112ae6c49e3486a57a6641b018f83b521ec89ec | |
| author | Berk Hess <hess@kth.se> | |
| Wed, 21 Oct 2015 10:04:50 +0000 (21 12:04 +0200) | ||
| committer | David van der Spoel <davidvanderspoel@gmail.com> | |
| Sat, 14 Nov 2015 10:11:33 +0000 (14 11:11 +0100) | ||
| tree | 62ce37ec08f4e161bf10943b979921ee8cb140f6 | treesnapshot (tar.gz zip) |
| parent | dc90d31d051cb245aba5aaa610f9b6048fed3076 | commitdiff |
Optimized the bonded thread force reduction
The bonded thread force reduction now uses fixed size blocks of 32
atoms. Instead of dividing reduction of the whole range of blocks
uniformly over the threads, now only used blocks are divided
(uniformly) over the threads.
This speeds up the reduction by a factor #threads (!) for typical
protein+water systems when not using domain decomposition.
With domain decomposition the speed up is up to a factor of 3.
Change-Id: I255b4336d315ef62ac9bc446c22313bf47baa097
The bonded thread force reduction now uses fixed size blocks of 32
atoms. Instead of dividing reduction of the whole range of blocks
uniformly over the threads, now only used blocks are divided
(uniformly) over the threads.
This speeds up the reduction by a factor #threads (!) for typical
protein+water systems when not using domain decomposition.
With domain decomposition the speed up is up to a factor of 3.
Change-Id: I255b4336d315ef62ac9bc446c22313bf47baa097
| src/gromacs/listed-forces/listed-forces.cpp | diffblobblamehistory | |
| src/gromacs/listed-forces/listed-internal.h | diffblobblamehistory | |
| src/gromacs/listed-forces/manage-threading.cpp | diffblobblamehistory |