Open isaacsas opened 3 weeks ago
(This is also consistent with the performance difference between MassActionJump
s and a hand-coded Direct
method that evaluates propensities via a single function.)
That is rather surprising still, but interesting.
Yeah, I'm not yet convinced of this result, but the conversion would be easy to add and allow us to test this systematically from Catalyst-generated models in SciMLBenchmarks.
In some preliminary benchmarking for a talk I'm finding that when using
Direct
with a small number of jumps (~6), we can actually now get a ~30% speedup usingConstantRateJump
s. This needs more investigation, but it may be that in the case of a small number of jumps, where we don't need function wrappers and use recursion for type-stability, we can now do better thanMassActionJump
s.