archives

Partial vectorisation of Haskell programs

Partial vectorisation of Haskell programs. Manuel M. T. Chakravarty, Roman Leshchinskiy, Simon Peyton Jones, and Gabriele Keller, Proc ACM Workshop on Declarative Aspects of Multicore Programming, San Francisco, Jan 2008.

Vectorisation for functional programs, also called the flattening transformation, relies on drastically reordering computations and restructuring the representation of data types. As a result, it only applies to the purely functional core of a fully-fledged functional language, such as Haskell or ML. A concrete implementation needs to apply vectorisation selectively and integrate vectorised with unvectorised code. This is challenging, as vectorisation alters the data representation, which must be suitably converted between vectorised and unvectorised code. In this paper, we present an approach to partial vectorisation that selectively vectorises sub-expressions and data types, and also, enables linking vectorised with unvectorised modules.

The idea is fairly simple, and utilizes conversion between vectorized and unvectorized representations of the datatypes. A formal translation scheme is provided.

Data Parallel Haskell papers are here.