Merging of compatible postsynaptic models#201
Conversation
# Conflicts: # lib/src/generateCPU.cc # lib/src/generateInit.cc # lib/src/generateKernels.cc # lib/src/generateRunner.cc # lib/src/synapseGroup.cc
# Conflicts: # lib/include/neuronGroup.h # lib/src/generateCPU.cc # lib/src/generateKernels.cc # lib/src/generateRunner.cc # lib/src/modelSpec.cc
* Added yet another ``GENN_PREFERENCES`` flag defaulting to turn off postsynaptic model merging * Changed flags in ``SynapseGroup`` to allow whether group has been merged (as an origin or target) as well as what the target postsynaptic group is (otherwise it is ambiguous whether a group is the source of a merge or unmerged) * Fixed bug in runner where methods to push and pull state of synapse groups included inSyn and denDelay variables which had been merged away
tnowotny
left a comment
tnowotny
left a comment
There was a problem hiding this comment.
I assume you have tested that this works on at least the model you mentioned? In principle it should work of course ...
You said it was messy but I think it is the kind of thing we should be doing to improve GeNN performance! Of course, I am not sure how many of these kind of ideas can be put in before things get out of hand and there is just too many code paths/options and everything becomes unmanageable. So far I would say it still looks fairly tidy given what it does.
I have no further concerns - if you are confident it's not containing a horrible bug, I think we can merge it.
| for(unsigned int i = 0; !inSyn.empty(); i++) { | ||
| // Remove last element from vector | ||
| SynapseGroup *a = inSyn.back(); | ||
| inSyn.pop_back(); |
There was a problem hiding this comment.
isn't pop_back returning the last element of the vector, i.e. allowing
SynapseGroup *a = inSyn.pop_back();
?
There was a problem hiding this comment.
Sadly it doesn't - it must have for some programming language though because in my head it should too
2 participants
I think I've talked to you about this before but what this change does is, if the GENN_PREFERENCE is set, merge together the postsynaptic models of a neuron population's incoming synapse populations if
When postsynaptic models are merged:
Overall, on a K40, this change more than halves the time spent by the neuron kernel on the Potjans microcircuit model. It's a little messy but I feel worth it!