About the reduction of loops

[yger]

Hi all

While doing some benchmarks for the OpenMP implementation, and trying to profile a bit the code, to see where are the main bottlenecks in the cpp standalone module, I came up with a maybe naive but obvious questions : why are we looping multiple times over cells, at every time steps, to first do the stateupdate, then to threshold, then to reset (actually, this loop is done only on spiking neurons, so this is faster) ? It will be more efficient to merge those threee operations in the same loop, no ? Are there any reasons for not doing that ? Will it reduce the flexibility of the code ?

Best

Pierre

[thesamovar]

There are a number of reasons. First, it makes it simpler to code. For example, you may not have a threshold/reset condition. Second, reset happens after synaptic propagation by default, which means a new loop for that.

A third reason is that it may actually be more efficient, although we haven't benchmarked that. The reason being that a threshold condition introduces an if statement which interferes with the instruction cacheing (and memory cacheing too possibly) and I think stops the C++ compiler from using SSE instructions. That may be balanced by the fact that combining them into one loop would reduce the memory bandwidth, so we'd have to benchmark that.

So what I'd suggest is we benchmark the performance in real scenarios, and if there's a substantial speedup to combining then we can add some code complexity to allow for that, but if the speedup is small maybe let's not bother.

Thoughts?