Programming Multicore GPUs in Haskell

Trevor L. McDonell
University of New South Wales

Modern graphics processors are increasingly being used for
general-purpose computations, and are supported as such by several
low-level programming interfaces. However, the hardware design does
not lend itself to executing arbitrary code from high-level languages
such as Haskell. Accelerate represents our efforts to overcoming these
challenges; an embedded language of array computations for
high-performance computing, in which computations on multidimensional,
regular arrays are expressed in the form of parameterised collective
operations such as maps, reductions and permutations. In this talk, I
shall briefly introduce the Accelerate EDSL, before focusing on the
implementation of a CUDA backend for parallel execution on NVIDIA
GPUs. I shall discuss the main challenges of this approach, current
status, and preliminary results.