Parallelize synaptic effects application in `target` mode for heterogeneous delays WITHOUT bundle mode and WITHOUT atomics about brian2cuda HOT 5 OPEN

We are using the bundle mode by default now as it is faster in most use cases. We are keeping the synapse push mode as a user preference, therefore this issue stays open. But it is only relevant it we don't use bundles AND can't use atomics (#78). It has therefore low priority for now.

from brian2cuda.

first approach: not bitonic sort but instead:
direct paralellization by threadid:postneuronid assignment (and each thread runs through all synapses in the current queue) and use shared memory for the required data, i.e., postneuron index and synapse index

from brian2cuda.

for the finally targeted more advanced approach of sorting of synapses (value array in shared memory) according to postsynaptic neurons (key array in shared memory) for coalesced application of synaptic effects the following code snippet might be very helpful (actually a radixsort instead of a bitonic sort is performed which should be fast enough). note the code relies on https://nvlabs.github.io/cub/classcub_1_1_block_radix_sort.html where also simple/more readable examples are given: http://stackoverflow.com/a/34607511

from brian2cuda.

do this approach onlz if the bundle way is slower... as that one is trivially parallizable in the effect application (given not more than 1 pre post pair)

from brian2cuda.

Check #30, which seems to be the same issue as this one.

from brian2cuda.