serializing_mode is not set correctly about brian2cuda HOT 4 CLOSED

if the pre(synaptic) spike affects a post(synaptic) variable (as for most examples, e.g. cuba) then to parallelize over the postsynaptic neurons ('post') is correct, not over the synapses ('syn') as one and the same postsynaptic neuron might be affected by two synapses, i.e. we need to serialize per post(synaptic) neuron!

from brian2cuda.

I think a piece of your sentence got deleted?

Well, if the serializing_mode is meant to indicate what to parallelize over, then the 'post' mode is not implemented correctly, since all threads with tid != 0 return without doing anything, effectively resulting in most postsynaptic effects not being applied.

Whereas the 'syn' mode effectively results in parallizing over the synapses in the current spikequeue (one synapse per thread) and applying the effects to the postsynaptic neuron of each synapse.

So for the CUBA example to work, hard coding serializing_mode = 'syn' works right now.

But let's discuss that on Thursday.

from brian2cuda.

there was nothing missing i guess

anyway, the cuba example might work by chance (but even in this example the following problem exists):
consider the very typical situation of a timestep in which two (or more) synapses are active that point onto the same postynaptic neuron. when parallelizing over synapses any of the effects of these synapses onto that particular postneuron might overwrite the respective other -- therefore 'post' serialization is mandatory for this case

from brian2cuda.

The serializing mode was set correclty, the 'post' mode had a bug, which is fixed now. For now the 'post' mode uses only a single thread, parallelization comes later. With this all our examples are producing qualitatively same results in the cpp and cuda version (including STDP).

To reproduce these, clone the repo, initialize brian2 and run the examples:

git submodule update --init frozen_repos/brian2
cd ./examples && ./run_all.sh

This produces the figure ./examples/result_figures.png, comparing cpp and cuda versions.

from brian2cuda.