Futhask is a code generator that aims to create safe, Haskell friendly wrappers for Futhark libraries.

stack install

futhask [Backend] [Futhark.h] [HaskellSourceDir] [ModuleName]

futhark opencl --library myprogram.fut
futhask opencl myprogram.h src MyLibrary

For a simple example of how generated haskell code can be used, see FuthaskExample

import [ModuleName]
import [ModuleName].Entries

If using stack add c-sources: [Futhark.c] to the library section of package.yaml

extra-libraries: OpenCL

include-dirs: /opt/cuda/include
extra-lib-dirs: /opt/cuda/lib
extra-libraries: cuda cudart nvrtc

transformers and massiv are required for all backends. The codes generated for OpenCL and CUDA, both refer to types from the OpenCL and cuda packages respectively. This is only relevant if one wants to use certain functions in the raw interface, but, without modification, the generated code will not compile without these dependencies.

The generated code can be split in two main parts, raw and wrapped. The raw interface is simply the C-functions wrapped in the IO-monad, providing no added safety and requiring manual memory management. The wrapped interface uses newForeignPtr to introduce all Futhark pointers to the GC, and provides function types closer to those used within Futhark, returning tuples instead of writing to pointers.

getContext :: [ContextOption] -> IO Context

Available context options will depend on backend used.

To make the wrappers safe, and reduce clutter from explicitly passing around the context, the Fut monad is introduced. The Fut monad is an environment monad that implicitly passes the context around as necessary.

To run computations, the function

runFutIn :: Context -> Fut a -> a

is used. Additionally

runFutWith :: [ContextOption] -> Fut a -> a
runFut :: Fut a -> a

are defined for convienience for cases where the context doesn't need to be reused.

For more flexibility, the FutT monad transformer can be used. For convenience the type synonyms

type Fut = FutT Identity
type FutIO = FutT IO

are defined, but entry-points are in the general Monad m => FutT m.

To run the transformer

runFutTIn :: Context -> FutT m a -> m a
runFutTWith :: [ContextOption] -> FutT  m a -> m a
runFutT :: FutT m a -> m a

For lifting

mapFutT :: (m a -> n b) -> FutT m a -> FutT n b
map2FutT :: (m a -> n b -> k c) -> FutT m a -> FutT n b -> FutT k c
pureFut :: Monad m => Fut a -> FutT m a
unsafeFromFutIO :: FutIO a -> Fut a

For conversion between Futhark values and Haskell values, two classes are defined.

class Input fo ho where
    toFuthark :: Monad m => ho -> FutT m fo

class Output fo ho where
    fromFuthark :: Monad m => fo -> FutT m ho

Instances of Input and Output are generated for all transparent Futhark-arrays. The Haskell representation is Array S from Data.Massiv.Array. The absence of functional dependencies in the definitions might require more explicit type signatures, but gives more flexibility to define new instances. For tuples of instances, functions on the form fromFutharkTN, where N is the tuple size, are defined.

All of the wrapped values have finalizers, and should eventually be garbage collected. However, GHCs GC does not know how much memory the context is using, and so collection will not always be triggered frequently enough. This is primarily an issue when the program iterates on Futhark values, without any Haskell-side allocations.

One way to deal with this is to manually manage the memory using

finalizeFO :: (MonadIO m, FutharkObject wrapped raw) => wrapped -> FutT m ()

As with any manual memory management, the programmer is responsible for ensuring that the finalized value will not be used afterwards. For cases where the object is used in more than one thread without synchronisation,

addReferenceFO :: (MonadIO m, FutharkObject wrapped raw) => wrapped -> FutT m ()

can be used. addReferenceFO increments the reference counter of the object and finalizeFO will just decrement this counter until it's 0.