Improve speed of dump_connections() in layer_impl.h #3160
Conversation
Instead of calling get_connections() inside a loop for every source node, it calls get_connections() once for all connections between layers. The code considers that: 1) source nodes positions (src_vec) is ordered by source node. 2) connectome is ordered by source node.
|
This is a proposed solution for issue #3142 |
xavierotazuGDS
changed the title
|
This implementation only works for non-MPI and in some cases (not all) of MPI. Hence, DO NOT USE !!!! |
heplesser
added
S: Normal
|
@xavierotazuGDS Thank you for your PR! Since there still seems to be a problem with the code as of your last message, I have converted it to "Draft" status. You can change that back again later. |
New version. No bugs. It is a more "conservative" approach, since it does not consider nodes are ordered.
|
I have uploaded a new version. No bugs now on MPI. I do not change the Draft status because I am not sure how to do it. |
|
Status changed from "Draft" to "Open". |
Code style modified to agree with clang-format
|
Thanks for the update, xavierotazuGDS! There are/were still some formatting issues (mostly regarding white-spaces as far as I see), but they need to be fixed to keep code in a consistently formatted state (see Contribute to NEST). Since you seem to be a first time contributor, please also send a Contributor Agreement to info@nest-initiative.org so we can accept and correctly acknowledge your contribution. |
Automatic clang-format checking complains about functions I have not modified.
|
I just saw, that you already did send the CA! My appologies. |
|
terhorstd, I believe I solved all my code style issues. The other issues are in methods I have not modified (they are clang-format issues from the original code). Anyway, I could try to solve them (I will need some time). |
|
I took a look at clang-issues in methods I have not modified, but suggested changes are so big code blocks that I would not know what I am doing. Hence, I prefer not to modify clang-format issues outside the method I modifed. |
|
The version of the code formatter changed some time back, this is why there are complaints. Everything is already fixed in the main branch, which your branch is several commits behind (see on the main page of your repo "… behind"). Try to merge those into your branch, that should fix all spots that you didn't touch. |
| const long tnode_lid = tgt_layer->node_collection_->get_lid( target_node_id ); | ||
| assert( tnode_lid >= 0 ); | ||
| tgt_layer->compute_displacement( source_pos, tnode_lid ).print( out ); | ||
| out << '\n'; | ||
| } |
There was a problem hiding this comment.
Missing closing bracket here
|
I ran some basic benchmarks on the JURECA cluster with the following setup:
This seems to indicate that it scales pretty linearly, not even increasing memory usage by much
|
Instead of calling get_connections() inside a loop for every source node, it calls get_connections() once for all connections between layers.
The code considers that:
fixes #3142