Spikemonitor kernel is performance bottleneck for networks without synapses. about brian2cuda HOT 1 OPEN

Looking at some detailed benchmarking results, the STDP speed test with spikemonitor has a weird performance drop at N = 10^5 neurons. And looking at the profiling measurement the spikemonitor is taking up the most time. This is surely because the spikemonitor records all spikes from all those N neurons, but still the drop I can't quite understand.

The spikemonitor currently just loops through the spikespace and pushes spiking neuron IDs into a cudaVector. So maybe we just hid a point, where we end up having to reallocate a huge cudaVector and that is why we have that drop? Needs some more detailed profiling for N = 10^5 and maybe N = 5 x 10^5

And then I just had a look at the spikemonitor and realized we are serializing the recording of spiking neuron IDs, the kernel is called with 1 thread and 1 block. Can't we just reserve enough memory in the cudaVectors (since we know the number of spiking neurons at each timestep anyways) and then just write the recorded variables in parallel (thread <-> spiking neuron)?

from brian2cuda.