Sham: A DSL for Fast DSLs
Authors:
Rajan Walia,
Chung-chieh Shan,
Sam Tobin-Hochstadt
Abstract:
Domain-specific languages (DSLs) are touted as both easy to embed in programs and easy to optimize. Yet these goals are often in tension. Embedded or internal DSLs fit naturally with a host language, while inheriting the host's performance characteristics. External DSLs can use external optimizers and languages but sit apart from the host. We present Sham, a toolkit designed to enable internal DSL…
▽ More
Domain-specific languages (DSLs) are touted as both easy to embed in programs and easy to optimize. Yet these goals are often in tension. Embedded or internal DSLs fit naturally with a host language, while inheriting the host's performance characteristics. External DSLs can use external optimizers and languages but sit apart from the host. We present Sham, a toolkit designed to enable internal DSLs with high performance. Sham provides a domain-specific language (embedded in Racket) for implementing other high-performance DSLs, with transparent compilation to assembly code at runtime. Sham is well suited as both a compilation target for other embedded DSLs and for transparently replacing DSL support code with faster versions. Sham provides seamless inter-operation with its host language without requiring any additional effort from its users. Sham also provides a framework for defining language syntax which implements Sham's own language interface as well. We validate Sham's design on a series of case studies, ranging from Krishnamurthi's classic automata DSL to a sound synthesis DSL and a probabilistic programming language. All of these are existing DSLs where we replaced the backend using Sham, resulting in major performance gains. We present an example-driven description of how Sham can smoothly enhance an existing DSL into a high-performance one. When compared to existing approaches for implementing high-performance DSLs, Sham's design aims for both simplicity and programmer control. This makes it easier to port our techniques to other languages and frameworks, or borrow Sham's innovations "à la carte" without adopting the whole approach. Sham builds a sophisticated and powerful DSL construction toolkit atop fundamental language features including higher-order functions, data structures, and a foreign-function interface (FFI), all readily available in other languages. Furthermore, Sham's approach allows DSL developers to simply write functions, either using Sham or generating Sham, without needing to work through complex staging or partial evaluation systems.
△ Less
Submitted 15 July, 2021; v1 submitted 18 May, 2020;
originally announced May 2020.
From high-level inference algorithms to efficient code
Authors:
Rajan Walia,
Praveen Narayanan,
Jacques Carette,
Sam Tobin-Hochstadt,
Chung-chieh Shan
Abstract:
Probabilistic programming languages are valuable because they allow domain experts to express probabilistic models and inference algorithms without worrying about irrelevant details. However, for decades there remained an important and popular class of probabilistic inference algorithms whose efficient implementation required manual low-level coding that is tedious and error-prone. They are algori…
▽ More
Probabilistic programming languages are valuable because they allow domain experts to express probabilistic models and inference algorithms without worrying about irrelevant details. However, for decades there remained an important and popular class of probabilistic inference algorithms whose efficient implementation required manual low-level coding that is tedious and error-prone. They are algorithms whose idiomatic expression requires random array variables that are latent or whose likelihood is conjugate. Although that is how practitioners communicate and compose these algorithms on paper, executing such expressions requires eliminating the latent variables and recognizing the conjugacy by symbolic mathematics. Moreover, matching the performance of handwritten code requires speeding up loops by more than a constant factor.
We show how probabilistic programs that directly and concisely express these desired inference algorithms can be compiled while maintaining efficiency. We introduce new transformations that turn high-level probabilistic programs with arrays into pure loop code. We then make great use of domain-specific invariants and norms to optimize the code, and to specialize and JIT-compile the code per execution. The resulting performance is competitive with manual implementations.
△ Less
Submitted 1 July, 2019; v1 submitted 16 May, 2018;
originally announced May 2018.
