In hierarchical modelling scenarios we often have an expression taking the dot product of each row (or column) of a data matrix and a parameter vector: data{ int r ; int c ; matrix[r,c] X; ... // other data } model{ vector[c] beta ; ... // other parameters } model{ ... …

So I put together a script to automatically piece together the core stan file components in a variety of orders (where are things declared, where are things computed, order of declarations, order of computing, etc), yielding 100 combinations. I compiled cmdstan with an empty makefile and compiled ea…

Here’s the same plots as above but with the optimization configuration noted at the end of the last post: First the model configurations where only one method is computed (note that the configurations with the mike method computed in the model block failed to compile): [opt1_solo_time_values_and_…

[image] mike-lawrence: note that the configurations with the mike method computed in the model block failed to compile Can you upload the error you get somewhere?

Sure. Here’s the error: /tmp/RtmpgEpPax/model-c8f35b59817e.hpp: In member function ‘stan::scalar_type_t<T2> none_model_NA_NA_NA_none_NA_1_model_namespace::none_model_NA_NA_NA_none_NA_1_model::log_prob_impl(VecR&, VecI&, std::ostream*) const’: /tmp/RtmpgEpPax/model-c8f35b59817e.hpp:1000:48: error: ‘…

Huh, that’s a bug in the inliner that is easy to fix. No clue if it will end up impacting the rest of your examples or not, but I’ll ping you tomorrow when I have a patch

[image] WardBrian: My guess would be, though, that adding another set of profiles in the generated quantities block would show the same behavior as my “no stanc optimizations” run above, e.g. mike being slightly faster in all cases. Function inlining in that block should never create vars Oh,…

Stanc patch here: Consistently treat profile blocks as blocks in optimizer by WardBrian · Pull Request #1281 · stan-dev/stanc3 · GitHub For some reason I can no longer run the comparison script myself (getting Error evaluating the log probability at the initial value. Exception: normal_lpdf: Rando…

@WardBrian are there intstructions anywhere for building cmdstan using a branch of stanc? Also, for posterity, here’s (7.5 KB) a zip file (using a .txt extension to upload, you’ll have to change it to .zip) with the automated writing of the various model configurations and sampling thereof.

If you have the ocaml toolchain installed, you can set a make/local variable STANC to the path to a local repo and it will build and use that stanc. I’ve also kicked off a job which will build the binaries here , so once that is complete you can download the one for your system and just place it in $…

@WardBrian So the mike-declared-and-computed-in-model-block bug blocking compilation of those models is fixed, but the performance pattern persists whereby when using -O1 the mike method only performs better than columns_dot_product in that obscure special case while without -O1 the mike method is s…

Improving the efficiency of *_dot_product for trinary-and-redundant matrices

Modeling

stevebronder February 7, 2023, 7:40pm 38

Sorry I’m just getting caught up to speed on this. I’ll try to take a look at the example and code this weekend. Very possible I’m missing a piece somewhere in the SoA optimization where this should get demoted.

Topic		Replies	Views
Rows_dot_product is intended as a shorthand rather than compute optimization, right? Modeling	10	1187	February 15, 2021
Elementwise multiplications with row sums appear to be slow Modeling fitting-issues , specification	32	1495	August 23, 2021
Measuring and comparing computational performance in Stan with different compilation alternatives. Using reduce_sum does not bring any advantage? General performance	30	1657	March 26, 2021
30%-40% drop in sampling time using stanc O1 optimizations General	22	1986	September 1, 2022
Matrix operations (1x1) run 10x slower than using reals RStan performance	8	808	June 3, 2020

Improving the efficiency of *_dot_product for trinary-and-redundant matrices

Related topics