Profiling helped me work out that models are faster when structured to use columns_dot_sum rather than rows_dot_sum (n.b. follow-up explorations on the odd error bar variability over on slack)
@jades can you comment on how much the use of reduce_sum sped up sampling of your hierarchical psychophysical model?
Edit: oh! And back in the spring profiling helped me work out a faster structure for using *_dot_product in hierarchical models.
Edit 2: and profiling helped me work out that Gaussian likelihoods (n.b. only with lots of IID observations) can be sped up via sufficient stats.
Edit 3: and I have a TODO of using profiling to isolate the benefits of my reduced-redundant-computation trick for hierarchical models; may get to that as I work up the notebook for my coming StanConnect talk.