# Calling a Haskell function in R - a float expansion example

Posted on June 3, 2017 by Stéphane Laurent

In the previous article, I wrote a R function returning the binary expansion of a real number in $$[0,1]$$. In the present article, I will:

• write a similar function in Haskell;
• write this function in a way compatible with R, inside a module;
• compile this module in a dynamic linker suitable for R (dll for Windows, so for Linux);
• call the function from R through the dynamic linker.

The creation of a Haskell function compatible with R is allowed by the Foreign Function Interface (FFI), in other words the Foreign module.

I learnt how to do such things with the help of this blog post by Neil Mitchell.

## Binary (and more) expansion in Haskell

Let’s go to Haskell. The floatExpansion function below is obtained by a small modification of the floatToDigits function of the Numeric module. It returns the expansion of a real number $$u \in [0,1]$$ in a given integer base.

import Numeric (floatToDigits)
:{
let floatExpansion :: RealFloat a => Integer -> a -> [Int];
floatExpansion base u = replicate (- snd expansion) 0 ++ fst expansion
where expansion = floatToDigits base u
:}
floatExpansion 2 0.125
## [0,0,1]

## First dynamic linker: string output

Firstly, I show how to make this function compatible with R when its output is a string instead of a list. It is easy to convert a list to a string in Haskell:

show [0, 0, 1]
## "[0,0,1]"

To get the output as a vector in R, more work is needed, and I will do it in the next section.

### Make the function compatible with R

To make the function compatible with R, there are two rules:

• Every argument must be a pointer (Ptr) to a C compatible type: CInt, CDouble or CString.

• The result must be IO ().

A value of type Ptr represents a pointer to an object. This type is provided by the Foreign.Ptr module, which is imported via the Foreign module. The types CInt, CDouble and CString are provided by the Foreign.C module.

We end up with this module:

-- FloatExpansion1.hs
{-# LANGUAGE ForeignFunctionInterface #-}

module FloatExpansion where

import Foreign
import Foreign.C
import Numeric (floatToDigits)

foreign export ccall floatExpansion :: Ptr CInt -> Ptr CDouble -> Ptr CString
-> IO ()

floatExpansion :: Ptr CInt -> Ptr CDouble -> Ptr CString -> IO ()
floatExpansion base u result = do
base <- peek base
u <- peek u
expansion <- newCString $show$ floatExpansion' (toInteger base) u
poke result expansion

floatExpansion' :: RealFloat a => Integer -> a -> [Int]
floatExpansion' base u = replicate (- snd expansion) 0 ++ fst expansion
where expansion = floatToDigits base u

### Compilation

We need the following C file to do the compilation, as explained in the GHC users guide.

// StartEnd.c
#include <Rts.h>

void HsStart()
{
int argc = 1;
char* argv[] = {"ghcDll", NULL}; // argv must end with NULL

char** args = argv;
hs_init(&argc, &args);
}

void HsEnd()
{
hs_exit();
}

Then we compile the library with this command on Linux:

ghc -shared -fPIC -dynamic -lHSrts-ghc8.0.2 FloatExpansion1.hs StartEnd.c -o FloatExpansion1.so

and this command on Windows:

ghc -shared -fPIC FloatExpansion1.hs StartEnd.c -o FloatExpansion1.dll

This creates the dynamic linker FloatExpansion1.so on Linux, FloatExpansion1.dll on Windows.

In a cabal file, assuming StartEnd.c is in the project directory, we can do:

library
hs-source-dirs:      src
exposed-modules:     FloatExpansion
build-depends:       base >= 4.7 && < 5
if os(windows)
ghc-options:       -O2 -shared -fPIC StartEnd.c -o FloatExpansion1.dll
else
ghc-options:       -O2 -shared -fPIC -dynamic StartEnd.c -o FloatExpansion1.so
extra-libraries:   HSrts-ghc8.0.2

### Call in R

We firstly load the library with:

dll <- "Haskell/DLLs/FloatExpansion1.so"
.C("HsStart")
## list()

And we invoke the function with the help of the .C function, as follows:

.C("floatExpansion", base = 2L, x = 0.125, result = "")$result ##  "[0,0,1]" It works. But it would be better to have a vector as output, rather than a string. dyn.unload(dll) ## Second dynamic linker: vector output To get the output as a vector, the additional modules we need are: Foreign.R, Foreign.R.Types and Data.Vector.SEXP. They are provided by the inline-r package. The [Int] type of the output list of the floatExpansion function must be converted to [Int32]. We write a simple function intToInt32 to help us to do the conversion. It works with the help of the Data.Int module which is imported via the Foreign module. We end up with this module: -- FloatExpansion2.hs {-# LANGUAGE ForeignFunctionInterface #-} {-# LANGUAGE DataKinds #-} module FloatExpansion where import Foreign import Foreign.C import Foreign.R (SEXP) import qualified Foreign.R.Type as R import qualified Data.Vector.SEXP as DV import Numeric (floatToDigits) foreign export ccall floatExpansion :: Ptr CInt -> Ptr CDouble -> Ptr (SEXP s R.Int) -> IO () floatExpansion :: Ptr CInt -> Ptr CDouble -> Ptr (SEXP s R.Int) -> IO () floatExpansion base u result = do base <- peek base u <- peek u let expansion = map intToInt32$ floatExpansion' (toInteger base) u
poke result $DV.toSEXP$ DV.fromList expansion

intToInt32 :: Int -> Int32
intToInt32 = fromIntegral

floatExpansion' :: RealFloat a => Integer -> a -> [Int]
floatExpansion' base u = replicate (- snd expansion) 0 ++ fst expansion
where expansion = floatToDigits base u

We compile the library as before. And we load it in R as before:

dll <- "Haskell/DLLs/FloatExpansion2.so"
.C("HsStart")
## list()

And we invoke the function with the help of the .C function, as follows:

.C("floatExpansion", base = 2L, x = 0.125, result = list(0L))$result ## [] ##  0 0 1 In fact, the output is a list with one element, the desired vector. Let’s write a user-friendly function: floatExpand <- function(x, base = 2L){ .C( "floatExpansion", base = as.integer(base), x = as.double(x), result = list(0L) )$result[[1L]]
}

Let’s compare it with my R function num2dyadic:

library(microbenchmark)
microbenchmark(
floatExpand = floatExpand(runif(1)),
times = 5000
)
## Unit: microseconds
##         expr    min      lq     mean  median      uq       max neval cld
##  floatExpand 20.982 25.5930 35.66431 27.9130 37.6595  3504.444  5000  a
##   num2dyadic 26.062 45.5035 67.28613 52.8915 69.0730 19308.745  5000   b

It is faster. And I have checked that the two functions always return the same results.

Moreover the “RHaskell” function allows more than the binary expansion, for example the ternary expansion:

floatExpand(1/3+1/27, base = 3)
##  1 0 1

Quite nice, isn’t it ?

dyn.unload(dll)

## 2020 update: the ‘foreign-library’ stanza

Nowadays, there is a more convenient way to generate a Haskell DLL. I’m using stack now, and here is the contents of my stack project:

FloatExpansion1
├── FloatExpansion1.cabal
├── Setup.hs
├── src
│   └── FloatExpansion.hs
├── src-dll
│   └── FloatExpansionDLL.hs
├── stack.yaml
└── StartEnd.c

The file FloatExpansion1.cabal contains:

library
hs-source-dirs:      src
exposed-modules:     FloatExpansion
build-depends:       base >= 4.7 && < 5
ghc-options:         -Wall

foreign-library FloatExpansion1
buildable:           True
type:                native-shared
if os(Windows)
options: standalone
other-modules:       FloatExpansionDLL
build-depends:       base >=4.7 && < 5
, FloatExpansion1
hs-source-dirs:      src-dll
c-sources:           StartEnd.c
default-language:    Haskell2010

The file FloatExpansion.hs:

module FloatExpansion
where
import Numeric (floatToDigits)

floatExpansion' :: RealFloat a => Integer -> a -> [Int]
floatExpansion' base u = replicate (- snd expansion) 0 ++ fst expansion
where
expansion = floatToDigits base u

The file FloatExpansionDLL.hs:

module FloatExpansionDLL
where
import FloatExpansion
import Foreign
import Foreign.C

foreign export ccall floatExpansion :: Ptr CInt -> Ptr CDouble -> Ptr CString -> IO ()

floatExpansion :: Ptr CInt -> Ptr CDouble -> Ptr CString -> IO ()
floatExpansion base u result = do
base <- peek base
u <- peek u
expansion <- newCString $show$ floatExpansion' (toInteger base) u
poke result expansion

Then, running stack build will generate the DLL.