Repa

Repa is a Haskell library for high-performance, regular, shape-polymorphic parallel arrays. It is designed to express computations on large multi-dimensional arrays in a pure functional style while exploiting modern multi-core CPUs. Repa emphasizes data parallelism and fusion to minimize intermediate allocations and maximize throughput.

In Repa, an array is parameterized by its representation and its shape. The representation tags distinguish how and where data is stored or computed, such as a delayed representation in which operations are chained without materialization, and one or more manifest representations in which concrete memory layouts are used for efficient access. Repa supports unboxed representations for primitive element types, enabling efficient memory layout and vectorized access. The shape parameter makes the same code applicable to one-, two-, and higher-dimensional arrays.

A central goal of Repa is fusion: combining successive array operations into a single pass over the data. This avoids creating intermediate arrays and improves cache locality. Repa also provides facilities to execute computations in parallel on multi-core processors, while preserving referential transparency.

Typical usage involves constructing arrays, applying a sequence of transformations such as mapping, zipping, or reductions, and then forcing evaluation to a concrete representation for output or further processing. Repa is often applied to domains such as image processing, numerical simulations, and data analysis where large, regular arrays arise.

The library is part of the Haskell ecosystem of array processing tools and has influenced subsequent work in the field. See also related projects such as Accelerate and the broader array libraries in Haskell.