Shape.hs

{-# LANGUAGE TypeFamilies #-}
module Numeric.LAPACK.Shape where

import qualified Data.Array.Comfort.Shape as Shape
import Data.Tagged (Tagged)
import Data.Ix (Ix)

import Control.DeepSeq (NFData, rnf)


{- |
Class of shapes where indices still make sense if we permute elements.
We use this for all matrix factorisations
involving permutations or more generally orthogonal transformations,
e.g. eigenvalue and singular value, LU and QR decompositions.
E.g. say, we have a square matrix with dimension 'Shape.Enumeration Ordering'.
Its vector of eigenvalues has the same dimension,
but it does not make sense to access the eigenvalues
with indices like 'LT' or 'EQ'.
Thus 'Shape.Enumeration' is no instance of 'Permutable'
(and you should not add an orphan instance).

If you want to factor a matrix with a non-permutable shape,
you should convert it temporarily to a permutable one,
like 'Shape.ZeroBased' (i.e. 'Matrix.ShapeInt') or 'IntIndexed'.

The 'Permutable' class has no method, so you could add any shape to it.
However, you should use good taste when adding an instance.
There is no strict criterion which shape type to add.

We tried to use 'ShapeInt' for eigenvalue vectors
and 'LiberalSquare's as transformation matrices of eigenvalue decompositions.
However, this way, the type checker cannot infer
that the product of the factorisation is a strict square.

We also tried to use 'Shape.IntIndexed' for eigenvalue vectors
with according 'LiberalSquare's transformations.
This has also the problem of inferring squares.
Additionally,
more such transformations lead to nested 'Shape.IntIndexed' wrappers
and for 'Matrix.ShapeInt' even the first wrapper is unnecessary.
-}
class (Shape.C shape) => Permutable shape where
instance Permutable Shape.Zero where
instance Permutable () where
instance (Ix n) => Permutable (Shape.Range n) where
instance (Integral n) => Permutable (Shape.Shifted n) where
instance (Integral n) => Permutable (Shape.ZeroBased n) where
instance (Integral n) => Permutable (Shape.OneBased n) where
instance (Integral n) => Permutable (Shape.Cyclic n) where
instance (Permutable sh) => Permutable (Shape.Deferred sh) where
instance (Permutable sh) => Permutable (Tagged s sh) where


{- |
This shape type wraps any other array shape type.
However, its 'Shape.Indexed' instance just uses zero-based 'Int' indices.
Thus it can turn any shape type into a 'Shape.Indexed' one.
The main usage is to make an arbitrary shape 'Permutable'.
-}
newtype IntIndexed sh = IntIndexed {deconsIntIndexed :: sh}
   deriving (Eq, Show)


instance (NFData sh) => NFData (IntIndexed sh) where
   rnf (IntIndexed sh) = rnf sh

instance (Shape.C sh) => Shape.C (IntIndexed sh) where
   size (IntIndexed sh) = Shape.size sh

shapeInt :: (Shape.C sh) => sh -> Shape.ZeroBased Int
shapeInt = Shape.ZeroBased . Shape.size

instance (Shape.C sh) => Shape.Indexed (IntIndexed sh) where
   type Index (IntIndexed sh) = Int
   indices (IntIndexed sh) = take (Shape.size sh) [0 ..]
   unifiedOffset (IntIndexed sh) = Shape.unifiedOffset (shapeInt sh)
   unifiedSizeOffset (IntIndexed sh) =
      Shape.unifiedSizeOffset (shapeInt sh)
   inBounds (IntIndexed sh) k =
      Shape.inBounds (Shape.ZeroBased $ Shape.size sh) k

instance (Shape.C sh) => Shape.InvIndexed (IntIndexed sh) where
   unifiedIndexFromOffset (IntIndexed sh) =
      Shape.unifiedIndexFromOffset (shapeInt sh)

instance (Shape.C sh) => Permutable (IntIndexed sh) where