Numeric Haskell: A Repa Tutorial

From HaskellWiki
Revision as of 10:59, 5 October 2012 by Killy9999 (talk | contribs) (Copying arrays from pointers: - explanation of runIL)

Jump to: navigation, search

Note: This tutorial is for an old version of Repa. The current version (Repa 3.1) has a slightly different API. You can read more about Repa 3 in this paper.

Repa is a Haskell library for high performance, regular, multi-dimensional parallel arrays. All numeric data is stored unboxed and functions written with the Repa combinators are automatically parallel (provided you supply "+RTS -N" on the command line when running the program).


This document provides a tutorial on array programming in Haskell using the repa package.

Note: a companion tutorial to this is provided as the vector tutorial, and is based on the NumPy tutorial.

Authors: Don Stewart.

Quick Tour

Repa (REgular PArallel arrays) is an advanced, multi-dimensional parallel arrays library for Haskell, with a number of distinct capabilities:

  • The arrays are "regular" (i.e. dense, rectangular and store elements all of the same type); and
  • Functions may be written that are polymorphic in the shape of the array;
  • Many operations on arrays are accomplished by changing only the shape of the array (without copying elements);
  • The library will automatically parallelize operations over arrays.

This is a quick start guide for the package. For further information, consult:

Importing the library

Download the `repa` package:

$ cabal install repa

and import it qualified:

import qualified Data.Array.Repa as R

The library needs to be imported qualified as it shares the same function names as list operations in the Prelude.

Note: Operations that involve writing new index types for Repa arrays will require the '-XTypeOperators' language extension.

For non-core functionality, a number of related packages are available:

and example algorithms in:

Index types and shapes

Before we can get started manipulating arrays, we need a grasp of repa's notion of array shape. Much like the classic 'array' library in Haskell, repa-based arrays are parameterized via a type which determines the dimension of the array, and the type of its index. However, while classic arrays take tuples to represent multiple dimensions, Repa arrays use a richer type language for describing multi-dimensional array indices and shapes (technically, a heterogeneous snoc list).

Shape types are built somewhat like lists. The constructor Z corresponds to a rank zero shape, and is used to mark the end of the list. The :. constructor adds additional dimensions to the shape. So, for example, the shape:

   (Z :. 3 :. 2 :. 3)

is the shape of a small 3D array, with shape type

  (Z :. Int :. Int :. Int)

The most common dimensions are given by the shorthand names:

type DIM0 = Z
type DIM1 = DIM0 :. Int
type DIM2 = DIM1 :. Int
type DIM3 = DIM2 :. Int
type DIM4 = DIM3 :. Int
type DIM5 = DIM4 :. Int


Array U DIM2 Double

is the type of a two-dimensional array of unboxed doubles, indexed via Int keys, while

Array U Z Double

is a zero-dimension object (i.e. a point) holding an unboxed Double.

Many operations over arrays are polymorphic in the shape / dimension component. Others require operating on the shape itself, rather than the array. A typeclass, Shape, lets us operate uniformly over arrays with different shape.

Building shapes

To build values of shape type, we can use the Z and :. constructors. Open the ghci and import Repa:

Prelude> :m +Data.Array.Repa
Repa> Z         -- the zero-dimension

For arrays of non-zero dimension, we must give a size. Note: a common error is to leave off the type of the size.

Repa> :t Z :. 10
Z :. 10 :: Num head => Z :. head

leading to annoying type errors about unresolved instances, such as:

   No instance for (Shape (Z :. head0))

To select the correct instance, you will need to annotate the size literals with their concrete type:

Repa> :t Z :. (10 :: Int)
Z :. (10 :: Int) :: Z :. Int

is the shape of 1D arrays of length 10, indexed via Ints.

Given an array, you can always find its shape by calling extent.

Additional convenience types for selecting particular parts of a shape are also provided (All, Any, Slice etc.) which are covered later in the tutorial.

Working with shapes

That one key operation, extent, gives us many attributes of an array:

     -- Extract the shape of the array
     extent :: (Shape sh, Source r e) => Array r sh e -> sh

So, given a 3x3x3 array, of type Array U DIM3 Int, we can:

-- build an array
Repa> let x :: Array U DIM3 Int; x = fromListUnboxed (Z :. (3::Int) :. (3::Int) :. (3::Int)) [1..27]
Repa> :t x
x :: Array U DIM3 Int

-- query the extent
Repa> extent x
((Z :. 3) :. 3) :. 3

-- compute the rank (number of dimensions)
Repa> let sh = extent x
Repa> rank sh

-- compute the size (total number of elements)
> size sh

-- extract the elements of the array as a flat vector
Repa> toUnboxed x
fromList [1,2,3,4,5,6,7,8,9,10
            ,20,21,22,23,24,25,26,27] :: Data.Vector.Unboxed.Base.Vector Int

Generating arrays

New repa arrays ("arrays" from here on) can be generated in many ways, and we always begin by importing the Data.Array.Repa module:

$ ghci
GHCi, version 7.4.1:  :? for help
Loading package ghc-prim ... linking ... done.
Loading package integer-gmp ... linking ... done.
Loading package base ... linking ... done.
Prelude > :m + Data.Array.Repa

They may be constructed from lists, for example. Here is a one dimensional array of length 10, here, given the shape (Z :. 10):

Repa> let inputs = [1..10] :: [Double]
Repa> let x = fromListUnboxed (Z :. (10::Int)) inputs
Repa> x
AUnboxed (Z :. 10) (fromList [1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0,9.0,10.0])

The type of x is inferred as:

Repa> :t x
x :: Array U (Z :. Int) Double

which we can read as "an array of dimension 1, indexed via Int keys, holding elements of type Double stored using unboxed vectors"

We could also have written the type as:

Repa> let x' = fromListUnboxed (Z :. 10 :: DIM1) inputs
Repa> :t x'
x' :: Array U DIM1 Double

The same data may also be treated as a two dimensional array, by changing the shape parameter:

Repa> let x2 = fromListUnboxed (Z :. (5::Int) :. (2::Int)) inputs
Repa> x2
AUnboxed ((Z :. 5) :. 2) (fromList [1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0,9.0,10.0])

which has the type:

Repa> :t x2
x2 :: Array U ((Z :. Int) :. Int) Double

or, as above, if we define it with the type synonym for 2 dimensional Int- indexed arrays, DIM2:

Repa> let x2' = fromListUnboxed (Z :. 5 :. 2 :: DIM2) inputs
Repa> x2'
AUnboxed ((Z :. 5) :. 2) (fromList [1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0,9.0,10.0])
Repa> :t x2'
x2' :: Array U DIM2 Double

Building arrays from vectors

It is also possible to build arrays from unboxed vectors, from the 'vector' package:

fromUnboxed :: (Shape sh, Unbox e) => sh -> Vector e -> Array U sh e

New arrays are built by applying a shape to the vector. For example:

Repa> :m + Data.Vector.Unboxed
Repa Unboxed> let x = fromUnboxed (Z :. (10::Int)) (enumFromN 0 10)
Repa Unboxed> x
AUnboxed (Z :. 10) (fromList [0.0,1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0,9.0])

is a one-dimensional array of doubles. As usual, we can also impose other shapes:

Repa Unboxed> let x = fromUnboxed (Z :. (3::Int) :. (3::Int)) (enumFromN 0 9)
Repa Unboxed> x
AUnboxed ((Z :. 3) :. 3) (fromList [0.0,1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0])
Repa Unboxed> :t x
x :: Array U ((Z :. Int) :. Int) Double

to create a 3x3 array.

Generating random arrays

The repa-algorithms package lets us generate new arrays with random data:

-- 3d array of Ints, bounded between 0 and 255.
Repa Randomish> randomishIntArray (Z :. (3::Int) :. (3::Int) :. (3::Int)) 0 255 1
AUnboxed (((Z :. 3) :. 3) :. 3) (fromList [217,42,130,200,216,254,67,77,152,

Reading arrays from files

Using the repa-io package, arrays may be written and read from files in a number of formats:

  • as BMP files; and
  • in a number of text formats.

with other formats rapidly appearing. An example: to write an 2D array to an ascii file:

Repa> :m +Data.Array.Repa.IO.Matrix
Repa Matrix> let x = fromList (Z :. 5 :. 2 :: DIM2) [1..10]
Repa Matrix> writeMatrixToTextFile "test.dat" x

This will result in a file containing:

5 2

In turn, this file may be read back in via readMatrixFromTextFile.

Repa Matrix> xx <- readMatrixFromTextFile "test.dat" 
Repa Matrix> xx
AUnboxed ((Z :. 5) :. 2) (fromList [1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0,9.0,10.0])
Repa Matrix> :t xx
xx :: Array U DIM2 Double

To process .bmp files, use Data.Array.Repa.IO.BMP, as follows (currently reading only works for 24 bit .bmp):

Data.Array.Repa.IO.BMP> x <- readImageFromBMP "/tmp/test24.bmp"

Reads this .bmp image:


as a 3D array of Word8, which can be further processed.

For image IO in many, many formats, use the repa-devil library.

Copying arrays from pointers

You can also generate new repa arrays by copying data from a pointer. Here is an example, using fromForeignPtr:

import Data.Word
import Foreign.Ptr
import qualified Foreign.ForeignPtr.Safe         as FPS

import qualified Data.Vector.Storable            as V
import qualified Data.Array.Repa                 as R
import qualified Data.Array.Repa.Repr.ForeignPtr as RFP
import Data.Array.Repa

import Data.Array.Repa.IO.DevIL

i, j, k :: Int
(i, j, k) = (255, 255, 4) -- RGBA

-- 1D vector, filled with pretty colors
v :: V.Vector Word8
v = V.fromList . take (i * j * k) . cycle $ concat
        [ [ r, g, b, 255 ]
          | r <- [0 .. 255]
          , g <- [0 .. 255]
          , b <- [0 .. 255]

ptr2repa :: Ptr Word8 -> IO (R.Array RFP.F R.DIM3 Word8)
ptr2repa p = do
    fp <- FPS.newForeignPtr_ p
    return $ RFP.fromForeignPtr (Z :. i :. j :. k) fp

main = do
    -- copy our 1d vector to a repa 3d array, via a pointer
    r <- V.unsafeWith v ptr2repa
    runIL $ writeImage "test.png" (RGBA r)
    return ()

The v function creates a vector filled with pixel data (RGBA values); ptr2repa takes a Ptr, converts it to ForeignPtr which is then used to copy data into a 3D REPA array (first two dimensions correspond to pixel location, the third dimension corresponds to colour channel). In the main function we use unsafeWith function operating on storable vectors. It passes Ptr to vector's internal data to a function that operates on that Ptr (it may not modify the data). In our case this is the ptr2repa array, which returns image data wrapped in an array. That array is then passed as data to writeImage function. RGBA is a data constructor for Image type. This allows writeImage to interpret array data correctly. runIL is a wrapper around IO monad that guarantees that IL library will be properly initialized. The result is written to disk as this image:


Indexing arrays

To access elements in repa arrays, you provide an array and a shape, to access the element:

(!) :: (Shape sh, Elt a) => Array sh a -> sh -> a

Indices start with 0. So:

> let x = fromList (Z :. (10::Int)) [1..10]
> x ! (Z :. 2)

Note that we can't give just a bare literal as the shape, even for one-dimensional arrays, :

> x ! 2

No instance for (Num (Z :. Int))
   arising from the literal `2'

as the Z type isn't in the Num class, and Haskell's numeric literals are overloaded.

What if the index is out of bounds, though?

> x ! (Z :. 11)
*** Exception: ./Data/Vector/Generic.hs:222 ((!)): index out of bounds (11,10)

an exception is thrown. An alternative is to use indexing functions that return a Maybe:

(!?) :: (Shape sh, Elt a) => Array sh a -> sh -> Maybe a

An example:

> x !? (Z :. 9)
Just 10.0

> x !? (Z :. 11)

Operations on arrays

Besides indexing, there are many regular, list-like operations on arrays. Since many of the names parallel those in the Prelude, we import Repa qualified:

Repa> import qualified Data.Array.Repa as Repa

Maps, zips, filters and folds

We can map over multi-dimensional arrays:

Repa> let x = fromList (Z :. (3::Int) :. (3::Int)) [1..9]
Repa> x

since this map conflicts with the definition in the Prelude, we have to use it with the qualifier we requested:

Repa> (^2) x

Repa's map leaves the dimension unchanged:

Repa> extent x
(Z :. 3) :. 3
Repa> extent ( (^2) x)
(Z :. 3) :. 3

A fold reduces the inner dimension of the array:

fold :: (Shape sh, Elt a) 
     => (a -> a -> a) -> a -> Array (sh :. Int) a -> Array sh a

The x defined above was a 2D array:

Repa> extent x
(Z :. 3) :. 3

but if we sum each row:

Repa> Repa.fold (+) 0 x

we get a 1D array instead:

Repa> extent (Repa.fold (+) 0 x)
Z :. 3

Similarly, if y is a (3 x 3 x 3) 3D array:

Repa> let y = fromList ((Z :. 3 :. 3 :. 3) :: DIM3) [1..27]
Repa> y

we can fold over the inner dimension:

Repa> Repa.fold (+) 0 y

yielding a 2D (3 x 3) array in place of our 3D (3 x 3 x 3) array:

Repa> extent y
((Z :. 3) :. 3) :. 3
Repa> extent (Repa.fold (+) 0 y)
(Z :. 3) :. 3

Two arrays may be combined via zipWith:

zipWith :: (Shape sh, Elt b, Elt c, Elt a) =>
           (a -> b -> c) -> Array sh a -> Array sh b -> Array sh c

an example:

Repa> Repa.zipWith (*) x x
Repa> extent it
(Z :. 3) :. 3
Repa> Repa.zipWith (*) y y   
 361.0,400.0,441.0,484.0,529.0,576.0,625.0,676.0,729.0] --reformatted
Repa> extent it
((Z :. 3) :. 3) :. 3

Mapping, with indices

A very powerful operator is traverse, a parallel array traversal which also supplies the current index:

traverse :: (Shape sh, Shape sh', Elt a)	 
         => Array sh a	             
                 -- Source array
         -> (sh -> sh')	             
                 -- Function to produce the extent of the result.
         -> ((sh -> a) -> sh' -> b)  
                 -- Function to produce elements of the result.
                 -- It is passed a lookup function to
                 -- get elements of the source.
         -> Array sh' b

This is quite a compicated type, because it is very general. Let's take it apart. The first argument is the source array, which is obvious. The second argument is a function that transforms the shape of the input array to yield the output array. So if the arrays are the same size, this function is id. It might grow or resize the shape in other ways.

Finally, the 3rd argument is where the magic is. Given an index, return a new element, and you also get a lookup function which when applied yields the current element.

So we see this generalizes map to support indexes, and optional inspection of the current element. Let's try some examples:

$ ghci              
GHCi, version 7.0.3:  :? for help
*> :m + Data.Array.Repa
*> :m + Data.Array.Repa.Algorithms.Randomish
*> let a :: Array DIM3 Int;
       a = fromList (Z :. (3::Int) :. (3::Int) :. (3::Int)) [1..27]
*> a

-- Keeping the shape the same, and just overwriting elements
-- Use `traverse` to set all elements to their `x` axis:
*> traverse a id (\_ (Z :. i :. j :. k) -> i) 

-- Shuffle elements around, based on their index.
-- Rotate elements by swapping elements from rotated locations:
> traverse a id (\f (Z :. i :. j :. k) -> f (Z :. j :. k :. i)) 

The documentation on traverse provides further information.

Numeric operations: negation, addition, subtraction, multiplication

Repa arrays are instances of the Num. This means that operations on numerical elements are lifted automagically onto arrays of such elements. For example, (+) on two double values corresponds to element-wise addition, (+), of the two arrays of doubles:

> let x = fromList (Z :. (10::Int)) [1..10]
> x + x

Other operations from the Num class work just as well:

> -x

> x ^ 3
> x * x

Changing the shape of an array

One of the main advantages of repa-style arrays over other arrays in Haskell is the ability to reshape data without copying. This is achieved via *index-space transformations*.

An example: transposing a 2D array (this example taken from the repa paper). First, the type of the transformation:

transpose2D :: Elt e => Array DIM2 e -> Array DIM2 e 

Note that this transform will work on DIM2 arrays holding any elements. Now, to swap rows and columns, we have to modify the shape:

transpose2D a = backpermute (swap e) swap a
    e = extent a
    swap (Z :. i :. j) = Z :. j :. i

The swap function reorders the index space of the array. To do this, we extract the current shape of the array, and write a function that maps the index space from the old array to the new array. That index space function is then passed to backpermute which actually constructs the new array from the old one.

backpermute generates a new array from an old, when given the new shape, and a function that translates between the index space of each array (i.e. a shape transformer).

  :: (Shape sh, Shape sh', Elt a)    
  => sh'  
  -> (sh' -> sh)  
  -> Array sh a   
  -> Array sh' a

Note that the array created is not actually evaluated (we only modified the index space of the array).

Transposition is such a common operation that it is provided by the library:

transpose :: (Shape sh, Elt a)
          => Array ((sh :. Int) :. Int) a -> Array ((sh :. Int) :. Int)

the type indicate that it works on the lowest two dimensions of the array.

Other operations on index spaces include taking slices and joining arrays into larger ones.


Following are some examples of useful functions that exercise the API.

Example: Rotating an image with backpermute

Flip an image upside down:

import System.Environment
import Data.Word
import Data.Array.Repa hiding ((++))
import Data.Array.Repa.IO.DevIL

main = do
    [f] <- getArgs
    runIL $ do
        v   <- readImage f
        writeImage ("flip-"++f) (rot180 v)
rot180 :: Array DIM3 Word8 -> Array DIM3 Word8
rot180 g = backpermute e flop g
        e@(Z :. x :. y :. _)   = extent g

        flop (Z :. i         :. j         :. k) =
             (Z :. x - i - 1 :. y - j - 1 :. k)

Running this:

   $ ghc -O2 --make A.hs
   $ ./A haskell.jpg 

Results in:


Example: matrix-matrix multiplication

A more advanced example from the Repa paper: matrix-matrix multiplication: the result of matrix multiplication is a matrix whose elements are found by multiplying the elements of each row from the first matrix by the associated elements of the same column from the second matrix and summing the result.

if A=\begin{bmatrix}a&b\\c&d\end{bmatrix} and B=\begin{bmatrix}e&f\\g&h\end{bmatrix}



So we take two, 2D arrays and generate a new array, using our transpose function from earlier:

 mmMult :: (Num e, Elt e)
        => Array DIM2 e
        -> Array DIM2 e
        -> Array DIM2 e

 mmMult a b = sum (zipWith (*) aRepl bRepl)
         t   = transpose2D b
         aRepl = extend (Z :.All :.colsB :.All) a
         bRepl = extend (Z :.rowsA :.All :.All) t
         (Z :.colsA :.rowsA) = extent a
         (Z :.colsB :.rowsB) = extent b

The idea is to expand both 2D argument arrays into 3D arrays by replicating them across a new axis. The front face of the cuboid that results represents the array a, which we replicate as often as b has columns (colsB), producing aRepl.

The top face represents t (the transposed b), which we replicate as often as a has rows (rowsA), producing bRepl,. The two replicated arrays have the same extent, which corresponds to the index space of matrix multiplication

Optimized implementations of this function are available in the repa-algorithms package.

Example: parallel image desaturation

To convert an image from color to greyscale, we can use the luminosity method to average RGB pixels into a common grey value, where the average is weighted for human perception of green.

The formula for luminosity is 0.21 R + 0.71 G + 0.07 B.

We can write a parallel image desaturation tool using repa and the repa-devil image library:

import Data.Array.Repa.IO.DevIL
import Data.Array.Repa hiding ((++))
import Data.Word
import System.Environment

-- Read an image, desaturate, write out with new name.
main = do 
  [f] <- getArgs
  runIL $ do
    a <- readImage f
    let b = traverse a id luminosity 
    writeImage ("grey-" ++ f) b

And now the luminosity transform itself, which averages the 3 RGB colors based on perceived weight:

-- (Parallel) desaturation of an image via the luminosity method.
luminosity :: (DIM3 -> Word8) -> DIM3 -> Word8
luminosity _ (Z :. _ :. _ :. 3) = 255   -- alpha channel
luminosity f (Z :. i :. j :. _) = ceiling $ 0.21 * r + 0.71 * g + 0.07 * b
        r = fromIntegral $ f (Z :. i :. j :. 0)
        g = fromIntegral $ f (Z :. i :. j :. 1)
        b = fromIntegral $ f (Z :. i :. j :. 2)

And that's it! The result is a parallel image desaturator, when compiled with

$ ghc -O -threaded -rtsopts --make A.hs -fforce-recomp

which we can run, to use two cores:

$ time ./A sunflower.png +RTS -N2 -H                  
./A sunflower.png +RTS -N2 -H  0.19s user 0.03s system 135% cpu 0.165 total

Given an image like this:


The desaturated result from Haskell:


Optimising Repa programs

Fusion, and why you need it

Repa depends critically on array fusion to achieve fast code. Fusion is a fancy name for the combination of inlining and code transformations performed by GHC when it compiles your program. The fusion process merges the array filling loops defined in the Repa library, with the "worker" functions that you write in your own module. If the fusion process fails, then the resulting program will be much slower than it needs to be, often 10x slower an equivalent program using plain Haskell lists. On the other hand, provided fusion works, the resulting code will run as fast as an equivalent cleanly written C program. Making fusion work is not hard once you understand what's going on.

The force function has the loops

Suppose we have the following binding:

 arr' = R.force $ (\x -> x + 1) arr

The right of this binding will compile down to code that first allocates the result array arr', then iterates over the source array arr, reading each element in turn and adding one to it, then writing to the corresponding element in the result.

Importantly, the code that does the allocation, iteration and update is defined as part of the force function. This forcing code has been written to break up the result into several chunks, and evaluate each chunk with a different thread. This is what makes your code run in parallel. If you do not use force then your code will be slow and not run in parallel.

Delayed and Manifest arrays

In the example from the previous section, think of the (\x -> x + 1) arr expression as a specification for a new array. In the library, this specification is referred to as a delayed array. A delayed array is represented as a function that takes an array index, and computes the value of the element at that index.

Applying force to a delayed array causes all elements to be computed in parallel. The result of a force is referred to as a manifest array. A manifest array is a "real" array represented as a flat chunk of memory containing array elements.

All Repa array operators will accept both delayed and manifest arrays. However, if you index into a delayed array without forcing it first, then each indexing operation costs a function call. It also recomputes the value of the array element at that index.

Shells and Springs

Here is another way to think about Repa's approach to array fusion. Suppose we write the following binding:

arr' = R.force $ (\x -> x * 2) $ (\x -> x + 1) arr

Remember from the previous section, that the result of each of the applications of is a delayed array. A delayed array is not a "real", manifest array, it's just a shell that contains a function to compute each element. In this example, the two worker functions correspond to the lambda expressions applied to

When GHC compiles this example, the two worker functions are fused into a fresh unfolding of the parallel loop defined in the code for R.force. Imagine holding R.force in your left hand, and squashing the calls to into it, like a spring. Doing this breaks all the shells, and you end up with the worker functions fused into an unfolding of R.force.

INLINE worker functions

Consider the following example:

f x  = x + 1
arr' = R.force $ R.zipWith (*) ( f arr1) ( f arr2)

During compilation, we need GHC to fuse our worker functions into a fresh unfolding of R.force. In this example, fusion includes inlining the definition of f. If f is not inlined, then the performance of the compiled code will be atrocious. It will perform a function call for each application of f, where it really only needs a single machine instruction to increment the x value.

Now, in general, GHC tries to avoid producing binaries that are "too big". Part of this is a heuristic that controls exactly what functions are inlined. The heuristic says that a function may be inlined only if it is used once, or if its definition is less than some particular size. If neither of these apply, then the function won't be inlined, killing performance.

For Repa programs, as fusion and inlining has such a dramatic effect on performance, we should absolutely not rely on heuristics to control whether or not this inlining takes place. If we rely on a heuristic, then even if our program runs fast today, if this heuristic is ever altered then some functions that used to be inlined may no longer be.

The moral of the story is to attach INLINE pragmas to all of your client functions that compute array values. This ensures that these critical functions will be inlined now, and forever.

{-# INLINE f #-}
f x  = x + 1
arr' = R.force $ R.zipWith (*) ( f arr1) ( f arr2)

Advanced techniques

Repa's parallel programming model

Discussion about the gang threads and hooks to help

Programming with stencils

Discuss the stencil types model