Failing brian2 test: test_synapses.test_vectorisation_STDP_like

[brian-team]

Failing brian2 test: test_synapses.test_vectorisation_STDP_like

[thesamovar]

We actually do have a check for potentially undefined behaviour, and emit a warning or error (can't remember which) in that case. I seem to remember in an earlier version of brian2genn we had an explicit check for code of the form neuron_var += synaptic_var that was treated specially. Are we talking about some generalisation of that?

Also, atomicAdd may be very inefficient on medium density synaptic operations (but is good for very low density ones, see mine and Romain's paper on GPU spiking neural networks).

[tnowotny]

This problem arises where the "on_pre" modifies pre-synaptic neuron variables and "on_post" modifies post-synaptic neurons variables (somewhat unusual, the case we covered was "on_pre" changing post-synaptic vars).
This exposes a bit more general problem, where just giving full access to pre- and post-synaptic neuron variables in GeNN can lead to undefined behaviour due to race conditions. If you exclude truly undefined behaviour within Brian, then the expected code should only be +=, *=, -=, /= and equivalent, right? However, I am not sure how one would do a "atomicDivide" ...
As for efficiency: If one wants to allow these manipulations, there is no choice but to use atomic operations I believe.

[thesamovar]

Ah I get you. At a conference at the moment so can't think too deeply about it, so just a couple of points for now.

Yes, I think ultimately you'll only allow += etc. however there could be relatively complex expressions on the RHS of the += so you can't just consider v += w type cases.

There is a way to do atomic divide although I've forgotten how you do it now. There's not a single operation, but I think you can put an atomic lock on a memory location? Something like that anyway.

Do you mean that if you want to allow modifications of both pre AND postsynaptic variables you have to use atomic operations? Because if you're just going in one direction there's all sorts of tricks, right? (I guess probably GeNN is using some of them, and NeMo uses a whole lot of tricks for sparse propagation.)