FEATURE: fixed dt

[boeschf]

Motivation

So far, multi-compartment cable cells use a variable time step: the user-specified maximum time step $dt$ may be shortened based on incoming events, such as spikes, generators, and (exact) sampling, so that the event delivery is performed exactly at the requested times.

Proposal: All events with time $t_s$ are now delivered at simulation time $t_i = t_0 + i ~dt$ if
$t_s \in \left[ t_i , t_i + dt\right).$

Removing exact delivery increases the speed of the simulation due to elimination of small steps, makes the numerics independent of presence of sampling, and also leads to a number of code simplifications. In particular, we have no more need for

• integration domains,
• event binning,
• sampling policies.

Main Changes

The new event delivery logic is mainly implemented in mc_cell_group::advance and fvm_lowered_cell_impl<Backend>::integrate.

Incidental Changes

• arb_mechanism_ppack now exposes a scalar time and scalar time step: t and dt
  • All cells advance with the same increments, i.e. no separate integration domains
• Deliverable events now organized in vector of vector of events. Outermost vector has one element per mechanism_id, inner vector has one element per time step interval.
  • This vector is then used to initialize a event_stream for each shared_state::mech_storage, while earlier, we had only one multi_event_stream per shared_state.
  • Helps with simpler and faster event delivery, and exposes more parallelism on the GPU
• GPU event_stream is now nearly identical to multi-core implementation (marking etc. is done on CPU) except for optional stable sorting according to mechanism_index, which is done in parallel using a thread_group.
  • Less maintenance
• shared_state gains some responsibility to handle events: register_events, mark_events, deliver_events
  • Better encapsulation

TODO

• More documentation (markdown docs, code docs, and rst)
• Investigate use of grouping within multi_event stream<sampling_event>
• Performance measurements

Performance Measurements

Performance with respect to master branch measured on Piz Daint's GPU partition for the busyring benchmark with the following parameters:

• number of cells: 2000,4000,8000,16000,32000
• ring size: 10,100,1000
• duration: 30
• dt: 0.025
• min_delay: 2
• number of synapses: 2000,4000

Each benchmark run was repeated 20 times with a different node allocation. Error bar show 95% confidence interval.

[thorstenhater]

A few nits have been picked.

[thorstenhater]

Thanks, it's a nice change, simpler and faster.

[thorstenhater]

Get this merged!

[thorstenhater]

Now merge. We'll do the convergence after.