Giovanni Pederiva (@GioPede) suggest I look at numba to accelerate the worm, given that the kernel is now completely understood, relatively tight, and the slowest part of sampling.
I'm happy to say that on the little parallel production demo it cuts the whole pipeline from 4 minutes to 2.5 minutes.
Also
For W=3 N=7 κ=0.08 (below critical) the speed-up was ~3.5 times measured on 1000 configurations
Same but N=21 it was closer to 5 times.
Running example/action-comparison.py shows that the hammers still match.
Giovanni Pederiva (@GioPede) suggest I look at numba to accelerate the worm, given that the kernel is now completely understood, relatively tight, and the slowest part of sampling.
I'm happy to say that on the little parallel production demo it cuts the whole pipeline from 4 minutes to 2.5 minutes.
Also
Running
example/action-comparison.pyshows that the hammers still match.