Weights caching and column to_numpy to speedup histogram filling
The histogram filling has been speed up by caching the broadcasted weights, when the weights are event-level and the data_structure is 1-D.
Moreover profiling the filling I observed that the hist.fill()
method is much faster then we pass numpy arrays instead of awkward arrays. I have included a to_numpy
while caching the variables to fill.
This PR brings a ~30% speedup in the histogram filling phase, which with >10 categories can save quite a lot of time.