Efficient compilation of array probabilistic programs
share
Probabilistic programming languages are valuable because they allow us to build, run, and change concise probabilistic models that elide irrelevant details. However, current systems are either inexpressive, failing to support basic features needed to write realistic models, or inefficient, taking orders of magnitude more time to run than hand-coded inference. Without resolving this dilemma, model developers are still required to manually rewrite their high-level models into low-level code to get the needed performance. We tackle this dilemma by presenting an approach for efficient probabilistic programming with arrays. Arrays are a key element of almost any realistic model. Our system extends previous compilation techniques from scalars to arrays. These extensions allow the transformation of high-level programs into known efficient algorithms. We then optimize the resulting code by taking advantage of the domain-specificity of our language. We further JIT-compile the final product using LLVM on a per-execution basis. These steps combined lead to significant new opportunities for specialization. The resulting performance is competitive with manual implementations of the desired algorithms, even though the original program is as concise and expressive as the initial model.
READ FULL TEXT