Ah, ok. _lp functions return the draw, update target and can be used in transformed parameters. Nice.
In my mock example, I don’t need _lpdf and _rng, just _lp in the transformed parameters section. This looks much more familiar.
Will test this, then use map_rect to speed things up. multi_normal_cholesky_truncated_lp isn’t super fast, and testing takes a while.
EDIT: Seems to work, more testing necessary, but it’s slow.
Can someone confirm that map_rect works within transformed parameters (it seems like it per Can map_rect return more than log probability?)? And just realizing how much pain map_rect is vs reduce_sum. map_rect is quite inflexible when it comes to passing on data and parameters. If only I could simply add a #pragma omp parallel ahead of the for (r in 1:S) {...}