An embedded language for accelerated array processing

henlo, my name is Theia

High-performance parallel arrays for Haskell

GitHub CI Gitter
Stackage LTS Stackage Nightly Hackage

Data.Array.Accelerate defines an embedded language of array computations for high-performance computing in Haskell. Computations on multi-dimensional, regular arrays are expressed in the form of parameterised collective operations (such as maps, reductions, and permutations). These computations are online-compiled and executed on a range of architectures.

For more details, see our papers:

There are also slides from some fairly recent presentations:

Chapter 6 of Simon Marlow's book Parallel and Concurrent Programming in Haskell contains a tutorial introduction to Accelerate.

Trevor's PhD thesis details the design and implementation of frontend optimisations and CUDA backend.

Table of Contents

A simple example

As a simple example, consider the computation of a dot product of two vectors of single-precision floating-point numbers:

dotp :: Acc (Vector Float) -> Acc (Vector Float) -> Acc (Scalar Float)
dotp xs ys = fold (+) 0 (zipWith (*) xs ys)

Except for the type, this code is almost the same as the corresponding Haskell code on lists of floats. The types indicate that the computation may be online-compiled for performance; for example, using Data.Array.Accelerate.LLVM.PTX.run it may be on-the-fly off-loaded to a GPU.


Package accelerate is available from

  • Hackage: accelerate - install with cabal install accelerate
  • GitHub: AccelerateHS/accelerate - get the source with git clone https://github.com/AccelerateHS/accelerate.git. The easiest way to compile the source distributions is via the Haskell stack tool.

Additional components

The following supported add-ons are available as separate packages:

Install them from Hackage with cabal install PACKAGENAME.


  • Haddock documentation is included and linked with the individual package releases on Hackage.
  • Haddock documentation for in-development components can be found here.
  • The idea behind the HOAS (higher-order abstract syntax) to de-Bruijn conversion used in the library is described separately.



The accelerate-examples package provides a range of computational kernels and a few complete applications. To install these from Hackage, issue cabal install accelerate-examples. The examples include:

  • An implementation of canny edge detection
  • An interactive mandelbrot set generator
  • An N-body simulation of gravitational attraction between solid particles
  • An implementation of the PageRank algorithm
  • A simple ray-tracer
  • A particle based simulation of stable fluid flows
  • A cellular automata simulation
  • A "password recovery" tool, for dictionary lookup of MD5 hashes

Mandelbrot Raytracer


LULESH-accelerate is in implementation of the Livermore Unstructured Lagrangian Explicit Shock Hydrodynamics (LULESH) mini-app. LULESH represents a typical hydrodynamics code such as ALE3D, but is a highly simplified application, hard-coded to solve the Sedov blast problem on an unstructured hexahedron mesh.


Additional examples

Accelerate users have also built some substantial applications of their own. Please feel free to add your own examples!

  • Jonathan Fraser, GPUVAC: An explicit advection magnetohydrodynamics simulation
  • David van Balen, Sudokus: A sudoku solver
  • Trevor L. McDonell, lol-accelerate: A backend to the Λ ○ λ (Lol) library for ring-based lattice cryptography
  • Henning Thielemann, patch-image: Combine a collage of overlapping images
  • apunktbau, bildpunkt: A ray-marching distance field renderer
  • klarh, hasdy: Molecular dynamics in Haskell using Accelerate
  • Alexandros Gremm used Accelerate as part of the 2014 CSCS summer school (code)

Who are we?

The Accelerate team (past and present) consists of:

The maintainer and principal developer of Accelerate is Trevor L. McDonell trevor.mcdonell@gmail.com.

Mailing list and contacts

Citing Accelerate

If you use Accelerate for academic research, you are encouraged (though not required) to cite the following papers:

  • Manuel M. T. Chakravarty, Gabriele Keller, Sean Lee, Trevor L. McDonell, and Vinod Grover. Accelerating Haskell Array Codes with Multicore GPUs. In DAMP '11: Declarative Aspects of Multicore Programming, ACM, 2011.

  • Trevor L. McDonell, Manuel M. T. Chakravarty, Gabriele Keller, and Ben Lippmeier. Optimising Purely Functional GPU Programs. In ICFP '13: The 18th ACM SIGPLAN International Conference on Functional Programming, ACM, 2013.

  • Robert Clifton-Everest, Trevor L. McDonell, Manuel M. T. Chakravarty, and Gabriele Keller. Embedding Foreign Code. In PADL '14: The 16th International Symposium on Practical Aspects of Declarative Languages, Springer-Verlag, LNCS, 2014.

  • Trevor L. McDonell, Manuel M. T. Chakravarty, Vinod Grover, and Ryan R. Newton. Type-safe Runtime Code Generation: Accelerate to LLVM. In Haskell '15: The 8th ACM SIGPLAN Symposium on Haskell, ACM, 2015.

  • Robert Clifton-Everest, Trevor L. McDonell, Manuel M. T. Chakravarty, and Gabriele Keller. Streaming Irregular Arrays. In Haskell '17: The 10th ACM SIGPLAN Symposium on Haskell, ACM, 2017.

Accelerate is primarily developed by academics, so citations matter a lot to us. As an added benefit, you increase Accelerate's exposure and potential user (and developer!) base, which is a benefit to all users of Accelerate. Thanks in advance!

What's missing?

Here is a list of features that are currently missing:

  • Preliminary API (parts of the API may still change in subsequent releases)
  • Many more features... contact us!
  • Installation

  • Tested Compilers

  • Dependencies (36)

  • Dependents (47)

    @hackage/accelerate-io-repa, @hackage/algebraic, @hackage/accelerate-bignum, @hackage/containers-accelerate, @hackage/accelerate-fourier, @hackage/accelerate-llvm, Show all…
  • Package Flags

       (off by default)

      Enable debug tracing messages. The following options are read from the environment variable ACCELERATE_FLAGS, and via the command-line as:

      ./program +ACC ... -ACC

      Note that a backend may not implement (or be applicable to) all options.

      The following flags control phases of the compiler. The are enabled with -f<flag> and can be reveresed with -fno-<flag>:

      • acc-sharing: Enable sharing recovery of array expressions (True).

      • exp-sharing: Enable sharing recovery of scalar expressions (True).

      • fusion: Enable array fusion (True).

      • simplify: Enable program simplification phase (True).

      • inplace: Enable in-place array updates (True).

      • flush-cache: Clear any persistent caches on program startup (False).

      • force-recomp: Force recompilation of array programs (False).

      • fast-math: Allow algebraically equivalent transformations which may change floating point results (e.g., reassociate) (True).

      • fast-permute-const: Allow non-atomic `permute const` for product types (True).

      The following options control debug message output, and are enabled with -d<flag>.

      • verbose: Be extra chatty.

      • dump-phases: Print timing information about each phase of the compiler. Enable GC stats (+RTS -t or otherwise) for memory usage information.

      • dump-sharing: Print information related to sharing recovery.

      • dump-simpl-stats: Print statistics related to fusion & simplification.

      • dump-simpl-iterations: Print a summary after each simplifier iteration.

      • dump-vectorisation: Print information related to the vectoriser.

      • dump-dot: Generate a representation of the program graph in Graphviz DOT format.

      • dump-simpl-dot: Generate a more compact representation of the program graph in Graphviz DOT format. In particular, scalar expressions are elided.

      • dump-gc: Print information related to the Accelerate garbage collector.

      • dump-gc-stats: Print aggregate garbage collection information at the end of program execution.

      • dubug-cc: Include debug symbols in the generated and compiled kernels.

      • dump-cc: Print information related to kernel code generation/compilation. Print the generated code if verbose.

      • dump-ld: Print information related to runtime linking.

      • dump-asm: Print information related to kernel assembly. Print the assembled code if verbose.

      • dump-exec: Print information related to program execution.

      • dump-sched: Print information related to execution scheduling.

       (off by default)

      Enable hooks for monitoring the running application using EKG. Implies debug mode. In order to view the metrics, your application will need to call Data.Array.Accelerate.Debug.beginMonitoring before running any Accelerate computations. This will launch the server on the local machine at port 8000.

      Alternatively, if you wish to configure the EKG monitoring server you can initialise it like so:

      import Data.Array.Accelerate.Debug
      import System.Metrics
      import System.Remote.Monitoring
      main :: IO ()
      main = do
        store  <- initAccMetrics
        registerGcMetrics store      -- optional
        server <- forkServerWith store "localhost" 8000

      Note that, as with any program utilising EKG, in order to collect Haskell GC statistics, you must either run the program with:

      +RTS -T -RTS

      or compile it with:

       (on by default)

      Enable bounds checking

       (off by default)

      Enable bounds checking in unsafe operations

       (off by default)

      Enable internal consistency checks

       (off by default)

      Build the nofib test suite (required for backend testing)