{-# LANGUAGE BangPatterns #-} {-# LANGUAGE DataKinds #-} {-# LANGUAGE TypeFamilies #-} {-# LANGUAGE TypeOperators #-} {-# LANGUAGE ScopedTypeVariables#-} {-# LANGUAGE DeriveDataTypeable #-} {-# LANGUAGE MultiParamTypeClasses #-} {-# LANGUAGE ConstraintKinds #-} {-# LANGUAGE GADTs #-} {-# LANGUAGE FlexibleInstances #-} {-# LANGUAGE FlexibleContexts #-} {-# LANGUAGE CPP #-} {-# LANGUAGE FunctionalDependencies #-} module Numerical.Array.Mutable( MArray(..) ,Array(..) ,RectilinearArray(..) ,DenseArrayBuilder(..) ,DenseArray(..) ,Boxed ,Unboxed ,Stored --,module Numerical.Array.Layout ,module Numerical.Array.Shape ) where import Control.Monad.Primitive ( PrimMonad, PrimState ) --import qualified Numerical.Array.DenseLayout as L import Numerical.Array.Address import qualified Numerical.Array.Layout as L import Numerical.Array.Layout (Layout,Locality(..),LayoutAddress,Format(..),Range(..),AffineRange(..)) import Numerical.Array.Shape --import Numerical.Nat --import GHC.Prim(Constraint) import Numerical.World --import Numerical.Array.Range --import Numerical.Array.Storage(Boxed,Unboxed,Stored) --import Numerical.Array.Locality import qualified Numerical.Array.Pure as P import qualified Numerical.Array.Storage as S import Numerical.Array.Storage (Buffer,Boxed,Unboxed,Stored) import qualified Data.Vector.Generic as VG import qualified Data.Vector.Generic.Mutable as VGM import Control.Monad (liftM) --import qualified Data.Vector.Storable.Mutable as SM --import qualified Data.Vector.Unboxed.Mutable as UM --import qualified Data.Vector.Mutable as BM {- For now we're going to just crib the vector style api and Lift it up into a multi dimensional setting. the tentative design is to have something like you'd think that the following array type is ``right'' but then you'll hit problems supporting -} -- data MArray world rep lay (view:: Locality) rank elm where -- MArray -- {_marrBuffer :: {-# UNPACK #!(MBuffer world rep elm) -- ,_marrForm :: {-# UNPACK #- } !(Form lay loc rank) -- --,_marrShift :: {-# UNPACK #- } !Address -- } -- shift will be zero for most reps i'll ever care about, but in certain cases, -- might not be. So for now not including it, but might be needed later, -- though likely in regards to some sparse format of some sort. --Omitting it for now, but may need to revisit later! -- --For now any 'Address' shift will need to be via the buffer -- -- One ssue in the formats is ``logical'' vs ``manifest'' Address. -- -- --we eedto have 'RepConstraint' be decoupled from the type class instances -- because we to sometimes have things that are world parametric -- -- indexing should be oblivious to locality, --NB: one important assumption we'll have for now, is that every -- dsfdf --type family RepConstraint world rep el :: Constraint --type instance MArrayElem {- | 'MArray' is the generic data family that -} data family MArray world rep lay (view::Locality) (rank :: Nat ) st el data instance MArray Native rep lay locality rank st el = MutableNativeArray { nativeBuffer :: ! (S.BufferMut rep st el ) ,nativeFormat :: ! (Format lay locality rank rep) } -- | Every 'MutableArray' instance has a contiguous version -- of itself, This contiguous version will ALWAYS have a Builder instance. type family MutableArrayContiguous (marr :: * -> * -> *) :: * -> * -> * type instance MutableArrayContiguous (MArray world rep layout locality rank)= MArray world rep layout 'Contiguous rank -- | Sadly 'ArrMutable' will have to have instances written by hand for now -- May later migrate the freeze / thaw machinery to Array.Phased, but lets type family ArrMutable ( arr :: * -> * ) :: * -> * -> * class P.PureArray (ArrPure marr) rank a => Array marr (rank:: Nat) a | marr -> rank where type ArrPure (marr :: * -> * -> * ) :: * -> * -- the type of the underlying storage buffer --type MutableArrayBuffer marr :: * -> * -> * -- really shouldnt appear in end user code, will only -- come up in writing new combinators -- the abstraction here is a reflection of the need for type MArrayAddress (marr :: * -> * -> * ) :: * -- | 'basicUnsafeAffineAddressShift' is needed to handle abstracting access in popcount space basicUnsafeAffineAddressShift :: (address ~ MArrayAddress marr) => marr st a -> Int -> address -> address -- question, should the type be -> address or -> Maybe address -- | Unsafely convert a mutable Array to its immutable version without copying. -- The mutable Array may not be used after this operation. Assumed O(1) complexity basicUnsafeFreeze :: (PrimMonad m, arr ~ ArrPure marr, marr ~ ArrMutable arr) => marr (PrimState m) a -> m (arr a) -- | Unsafely convert a pure Array to its mutable version without copying. -- the pure array may not be used after this operation. Assumed O(1) complexity basicUnsafeThaw :: (PrimMonad m, marr ~ ArrMutable arr, arr ~ ArrPure marr ) => arr a -> m (marr (PrimState m) a) -- | gives the shape, a 'rank' length list of the dimensions basicShape :: marr st a -> Index rank -- | 'basicCardinality' reports the number of manifest addresses/entries are -- in the array in a given address sub range. -- This is useful for determining when to switch from a recursive algorithm -- to a direct algorithm. -- Should this be renamed to something like basicPopCount/ basicCardinality ::(address ~ MArrayAddress marr) => marr st a -> Range address -> Int --basicUnsafeRead :: PrimMonad m => marr (PrimState m) a -> Shape rank Int -> m (Maybe a) -- | basicMutableSparseIndexToAddres checks if a index is present or not -- helpful primitive for authoring codes for (un)structured sparse array format basicSparseIndexToAddress :: (address ~ MArrayAddress marr) => marr s a -> Index rank -> Maybe address -- | 'basicMutableAddressToIndex' assumes you only give it legal manifest addresses basicAddressToIndex :: (address ~ MArrayAddress marr) =>marr s a -> address -> Index rank -- | return the smallest and largest valid logical address basicAddressRange :: (address ~ MArrayAddress marr)=> marr st a -> Maybe (Range address) -- | gives the next valid logical address -- undefined on invalid addresses and the greatest valid address. -- Note that for invalid addresses in between minAddress and maxAddress, -- will return the next valid address. basicSparseNextAddress :: (address ~ MArrayAddress marr)=> marr st a -> address -> Maybe address -- I think the case could be made for a basicPreviousAddress opeeration -- | gives the next valid array index, the least valid index that is -- or basicSparseNextIndex ::(address ~ MArrayAddress marr)=> marr st a -> Index rank -> Maybe address -> Maybe ( Index rank, address) -- | for a given valid address, @'basicAddressRegion' addr @ will return an AddressInterval -- that contains @addr@. This will be a singleton when the "maximal uniform stride interval" -- containing @addr@ has strictly less than 3 elements. Otherwise will return an Address range -- covering the maximal interval that will have cardinality at least 3. basicLocalAffineAddressRegion ::(address ~ MArrayAddress marr) => marr st a ->address -> AffineRange address -- | this doesn't quite fit in this class, but thats ok, will deal with that later basicOverlaps :: marr st a -> marr st a -> Bool -- | Reset all elements of the vector to some undefined value, clearing all -- references to external objects. This is usually a noop for unboxed -- vectors. This method should not be called directly, use 'clear' instead. basicClear :: PrimMonad m => marr (PrimState m) a -> m () ---- | Yield the element at the given position. This method should not be ---- called directly, use 'unsafeRead' instead. basicUnsafeAddressRead :: (PrimMonad m ,address ~ MArrayAddress marr) => marr (PrimState m) a -> address-> m a ---- | Replace the element at the given position. This method should not be ---- called directly, use 'unsafeAddressWrite' instead. basicUnsafeAddressWrite :: (PrimMonad m ,address ~ MArrayAddress marr) => marr (PrimState m) a -> address -> a -> m () --note the sparsewrite and sparse read are "fused" versions of basicManifestAddress -- and address read and write. probably needs to be benchmarked! TODO -- | Yield the element at the given position. This method should not be -- called directly, use 'unsafeSparseRead' instead. basicUnsafeSparseRead :: PrimMonad m => marr (PrimState m) a -> Index rank -> m (Maybe a) -- Replace the element at the given position. This method should not be -- called directly, use 'unsafeWrite' instead. -- the following is the type that normal Array indexing, -- as folks are used to, lookslike -- its wrong --basicUnsafeSparseWrite :: PrimMonad m => marr (PrimState m) a -> -- Index rank -> m( Maybe (a -> m ())) -- this might get axed instance (Buffer rep el, Layout (Format lay locality rank rep) rank ) =>Array (MArray Native rep lay locality rank) rank el where type ArrPure (MArray Native rep lay locality rank)= P.ImmArray Native rep lay locality rank type MArrayAddress (MArray Native rep lay locality rank)= LayoutAddress (Format lay locality rank rep) {-# INLINE basicShape #-} basicShape = L.basicLogicalShape . nativeFormat {-# NOINLINE basicUnsafeFreeze #-} basicUnsafeFreeze = \marr -> do pureBuffer <- VG.unsafeFreeze $ nativeBuffer marr return $ P.ImMutableNativeArray pureBuffer $ nativeFormat marr {-# NOINLINE basicUnsafeThaw #-} basicUnsafeThaw = \parr -> do mutBuffer <- VG.unsafeThaw $ P.nativeBufferPure parr return $ MutableNativeArray mutBuffer $ P.nativeFormatPure parr {-# INLINE basicSparseIndexToAddress #-} basicSparseIndexToAddress = \ marr -> L.basicToAddress (nativeFormat marr) {-# INLINE basicAddressToIndex #-} basicAddressToIndex = \ marr -> L.basicToIndex (nativeFormat marr) {-# INLINE basicSparseNextAddress #-} basicSparseNextAddress = \marr -> L.basicNextAddress (nativeFormat marr) {-# INLINE basicSparseNextIndex #-} basicSparseNextIndex = \marr -> L.basicNextIndex (nativeFormat marr) basicOverlaps = \marr1 marr2 -> VGM.overlaps (nativeBuffer marr1) (nativeBuffer marr2) basicClear = \marr -> VGM.clear (nativeBuffer marr) {-# INLINE basicUnsafeAddressRead #-} basicUnsafeAddressRead = \marr addr -> VGM.unsafeRead (nativeBuffer marr) (L.basicAddressAsInt (nativeFormat marr) addr) {-# INLINE basicUnsafeAddressWrite #-} basicUnsafeAddressWrite = \marr addr v-> VGM.unsafeWrite (nativeBuffer marr) (L.basicAddressAsInt (nativeFormat marr) addr) v {-# INLINE basicUnsafeSparseRead #-} basicUnsafeSparseRead = \marr ix -> do maddr <- return $ basicSparseIndexToAddress marr ix maybe (return Nothing) (\addr -> liftM Just $ basicUnsafeAddressRead marr addr ) maddr {-# INLINE basicAddressRange #-} basicAddressRange = \marr -> L.basicAddressRange (nativeFormat marr) basicCardinality = \marr -> L.basicAddressPopCount (nativeFormat marr) basicUnsafeAffineAddressShift = error "carter needs to add this" basicLocalAffineAddressRegion = error "crter needs to add this" {- type ArrPure marr :: * -> * type MArrayAddress marr :: * basicUnsafeAffineAddressShift :: (address ~ MArrayAddress marr) => marr st a -> Int -> address -> address basicLocalAffineAddressRegion :: (address ~ MArrayAddress marr) => marr st a -> address -> AffineRange address -} class ( Array marr rank a, P.PureDenseArray (ArrPure marr) rank a )=> DenseArray marr rank a | marr -> rank where -- | for Dense arrays, it is always easy to check if a given index is valid. -- this operation better have O(1) complexity or else! basicIndexInBounds :: marr st a -> Index rank -> Bool --basicUnsafeAddressDenseRead :: PrimMonad m => marr (PrimState m) a -> Address-> m a -- i already have dense address indexing ? --basicUnsafeAddressDenseWrite :: PrimMonad m => marr (PrimState m) a -> Address -> a -> m () -- | Yield the element at the given position. This method should not be -- called directly, use 'unsafeRead' instead. basicUnsafeDenseRead :: PrimMonad m => marr (PrimState m) a -> Index rank -> m a -- | Replace the element at the given position. This method should not be -- called directly, use 'unsafeWrite' instead. basicUnsafeDenseWrite :: PrimMonad m => marr (PrimState m) a -> Index rank -> a -> m () -- | gives the next valid logical address -- undefined on invalid addresses and the greatest valid address. -- Note that for invalid addresses in between minAddress and maxAddress, -- will return the next valid address. basicNextAddress :: marr st a -> Address -> Address -- I think the case could be made for a basicPreviousAddress opeeration -- | gives the next valid array index -- undefined on invalid indices and the greatest valid index basicNextIndex :: marr st a -> Index rank -> Index rank {- Mutable (Dense) Array Builder will only have contiguous instances and only makes sense for dense arrays afaik BE VERY THOUGHTFUL about what instances you write, or i'll be mad not including the general sparse building in the first release, will include subsequently -} --class MutableArray marr (rank:: Nat) a => MutableArrayBuilder marr rank a where --basicBuildArray:: Index rank -> b class DenseArray marr rank a => DenseArrayBuilder marr rank a where basicUnsafeNew :: PrimMonad m => Index rank -> m (marr (PrimState m) a) basicUnsafeReplicate :: PrimMonad m => Index rank -> a -> m (marr (PrimState m) a) class RectilinearArray marr rank a | marr -> rank where -- | @'MutableRectilinearOrientation' marr@ should equal Row or Column for any sane choice -- of instance, because every MutableRectilinear instance will have a notion of -- what the nominal major axix will be. -- The intended use case is side condition constraints like -- @'MutableRectilinearOrientation' marr~Row)=> marr -> b @ -- for operations where majorAxix projections are correct only for Row -- major formats. Such as Row based forward/backward substitution (triangular solvers) type MutableRectilinearOrientation marr :: * type MutableArrayDownRank marr ( st:: * ) a -- | MutableInnerContigArray is the "meet" (minimum) of the locality level of marr and InnerContiguous. -- Thus both Contiguous and InnerContiguous are made InnerContiguous, and Strided stays Strided -- for now this makes sense to have in the MutableRectilinear class, though that may change. -- This could also be thought of as being the GLB (greatest lower bound) on locality type MutableInnerContigArray (marr :: * -> * -> *) st a --type MutableArrayBuffer --not implementing this .. for now -- | @'basicSliceMajorAxis' arr (x,y)@ returns the sub array of the same rank, -- with the outermost (ie major axis) dimension of arr restricted to the -- (x,y) is an inclusive interval, MUST satisfy x marr (PrimState m) a -> (Int,Int)-> m (marr (PrimState m) a) --but should it be primmonadic? nah, tis pure! -- | semantically, 'basicProjectMajorAxis' arr ix, is the rank reducing version of what -- basicSliceMajorAxis arr (ix,ix) would mean _if_ the (ix,ix) tuple was a legal major axis slice basicMutableProjectMajorAxis :: PrimMonad m =>marr (PrimState m) a -> Int -> m (MutableArrayDownRank marr (PrimState m) a ) -- | @'basicMutableSlice' arr ix1 ix2@ picks out the (hyper) rectangle in dimension @rank@ -- where ix1 is the minimal corner and ix2 basicMutableSlice :: PrimMonad m => marr (PrimState m) a -> Index rank -> Index rank -> m (MutableInnerContigArray marr (PrimState m) a )