Optimising purely functional GPU programs

Abstract

Purely functional, embedded array programs are a good match for SIMD hardware, such as GPUs. However, the naive compilation of such programs quickly leads to both code explosion and an excessive use of intermediate data structures. The resulting slowdown is not acceptable on target hardware that is usually chosen to achieve high performance.In this paper, we discuss two optimisation techniques, sharing recovery and array fusion, that tackle code explosion and eliminate superfluous intermediate structures. Both techniques are well known from other contexts, but they present unique challenges for an embedded language compiled for execution on a GPU. We present novel methods for implementing sharing recovery and array fusion, and demonstrate their effectiveness on a set of benchmarks.

Using circular programs for higher-order syntax

“…There are many EDSLs that rely on a higher-order interface towards the user and a first-order representation for analysis and code generation: Lava [3], Pan [6], Nikola [8], Accelerate [10], Obsidian [12] and Feldspar [1], to name some. All of these EDSLs employ some kind of higher-order to first-order conversion.…”

Foreign inline code

“…Mainland [2] recently extended Template Haskell with support for quasiquoting arbitrary programming languages, which greatly simplifies writing code generators that produce complex C, CUDA, OpenCL, or Objective-C code by writing code templates in the syntax of the generated language-for example, Accelerate, an embedded language for GPU programming, makes extensive use of that facility to generate CUDA GPU code [3].…”

“…Mainland [2] recently extended Template Haskell with support for quasiquoting arbitrary programming languages, which greatly simplifies writing code generators that produce complex C, CUDA, OpenCL, or Objective-C code by writing code templates in the syntax of the generated language-for example, Accelerate, an embedded language for GPU programming, makes extensive use of that facility to generate CUDA GPU code [3].In this demo, I will show that quasiquoting also enables a new form of language interoperability. Here, a simple example using Objective-C: nslog :: String -> IO () nslog msg = $(objc ['msg :> ''String] (void [cexp| NSLog(@"A message from Haskell: %@", msg) |]))…”

A T2 graph-reduction approach to fusion

“…Accelerate [23] uses an elaboration of the delayed arrays representation from Repa, and in particular manages to avoid duplicating work. All array operations have a uniform representation as constructors for delayed arrays, on which fusion is performed by tree contraction.…”