Accelerate is a language for data-parallel array computations embedded within the programming language Haskell. More specifically, it is a deeply embedded language. This means that when you write programs with Accelerate, you are writing a Haskell program using operations from the Accelerate library, but the method by which the program runs is different from a conventional Haskell program. A program written in Accelerate is actually a Haskell program that generates, optimises, and compiles code for the GPU or CPU on-the-fly at program runtime.
To get started you will need to set up a Haskell environment as well as a few external libraries.
Select your operating system:
The Haskell ecosystem has two tools to help with building and installing packages:
cabal (the default) which installs packages to a global location, and
stack, which has a more project-centric focus.
We can now install the core Accelerate library:
cabal install accelerate
This will install the current stable release of Accelerate from Hackage. If you would like to instead install the latest in-development version, see how to install from GitHub.
This is sufficient to write programs in Accelerate as well as execute them using the included interpreter backend.1 For good performance however we also need to install one (or both) of the LLVM backends, which will compile Accelerate programs to native code.
Install a version of the
llvm-hs package suitable for the version of LLVM installed on your system. The first two numbers of the version of LLVM and the
llvm-hs package must match. We must also install with shared library support so that we can use
llvm-hs from within
ghci and Template Haskell. For example, if you have LLVM-4.0 installed:
cabal install llvm-hs -fshared-llvm --constraint="llvm-hs==4.0.*"
Install the Accelerate LLVM backend for multicore CPUs:
cabal install accelerate-llvm-native
(Optional) If you have a CUDA capable GPU, you can also install the Accelerate backend for NVIDIA GPUs:
cabal install accelerate-llvm-ptx
You can use Accelerate in a stack-based workflow by including the following (or similar) into the
stack.yaml file of your project:
resolver: lts-9.0 extra-deps: - 'accelerate-llvm-188.8.131.52' - 'accelerate-llvm-native-184.108.40.206' - 'accelerate-llvm-ptx-220.127.116.11' - 'cuda-0.7.5.3' - 'llvm-hs-18.104.22.168' - 'llvm-hs-pure-22.214.171.124' flags: llvm-hs: shared-llvm: true
Copy the following content into a file called
Dotp.hs. This simple example computes the dot product of two vectors of single-precision floating-point numbers. If you installed the GPU backend in step 2, you can uncomment the third line (delete the leading
--) to enable both the CPU and GPU backends.
import Data.Array.Accelerate as A import Data.Array.Accelerate.LLVM.Native as CPU -- import Data.Array.Accelerate.LLVM.PTX as GPU dotp :: Acc (Vector Float) -> Acc (Vector Float) -> Acc (Scalar Float) dotp xs ys = A.fold (+) 0 (A.zipWith (*) xs ys)
Open up a terminal and load the file into the Haskell interpreter with
Create some arrays to feed into the computation. See the documentation for more information, as well as additional ways to get data into the program.
ghci> let xs = fromList (Z:.10) [0..] :: Vector Float ghci> let ys = fromList (Z:.10) [1,3..] :: Vector Float
Run the computation:
ghci> CPU.run $ dotp (use xs) (use ys) Scalar Z [615.0]
This will convert the Accelerate program into LLVM code, optimise, compile, and execute it on the CPU. If your computer has multiple CPU cores, you can execute using multiple CPU cores by launching
ghci (or running a compiled program) with the additional command line options
+RTS -Nx -RTS, to use x CPU cores (or omit x to use as many cores as your machine has).
(Optional) If you installed the
accelerate-llvm-ptx backend, you can also execute the computation on the GPU simply by:
ghci> GPU.run $ dotp (use xs) (use ys) Scalar Z [615.0]
This will instead convert the Accelerate program into LLVM code suitable for the GPU, optimise, compile, and execute it on the GPU, as well as copy the input arrays into GPU memory and copy the result back into CPU memory.
Congratulations, you are set up to use Accelerate! Now you are ready to:
Although the core
accelerate package includes an interpreter that can be used to run Accelerate programs, its performance is fairly poor as it is designed as a reference implementation of the language semantics, rather than for performance.↩