{- Copyright 2016, Dominic Orchard, Andrew Rice, Mistral Contrastin, Matthew Danish Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -} {-# LANGUAGE DataKinds #-} {-# LANGUAGE GADTs #-} {-# LANGUAGE KindSignatures #-} {-# LANGUAGE TypeFamilies #-} {-# LANGUAGE FlexibleInstances #-} {-# LANGUAGE FlexibleContexts #-} {-# LANGUAGE NoMonomorphismRestriction #-} {-# LANGUAGE ScopedTypeVariables #-} {-# LANGUAGE PolyKinds #-} module Camfort.Specification.Stencils.InferenceBackend where import Prelude hiding (sum) import Data.Generics.Uniplate.Operations import Data.List hiding (sum) import Data.Data import Control.Arrow ((***)) import Data.Function import Camfort.Specification.Stencils.Model import Camfort.Helpers import Camfort.Helpers.Vec import Debug.Trace import Unsafe.Coerce import Camfort.Specification.Stencils.Syntax {- Spans are a pair of a lower and upper bound -} type Span a = (a, a) mkTrivialSpan a = (a, a) inferFromIndices :: VecList Int -> Specification inferFromIndices (VL ixs) = Specification $ case fromBool mult of Linear -> Single $ inferCore ixs' NonLinear -> Multiple $ inferCore ixs' where (ixs', mult) = hasDuplicates ixs -- Same as inferFromIndices but don't do any linearity checking -- (defaults to NonLinear). This is used when the front-end does -- the linearity check first as an optimimsation. inferFromIndicesWithoutLinearity :: VecList Int -> Specification inferFromIndicesWithoutLinearity (VL ixs) = Specification . Multiple . inferCore $ ixs inferCore :: (IsNatural n, Permutable n) => [Vec n Int] -> Approximation Spatial inferCore = simplify . fromRegionsToSpec . inferMinimalVectorRegions simplify :: Approximation Spatial -> Approximation Spatial simplify = fmap simplifySpatial simplifySpatial :: Spatial -> Spatial simplifySpatial (Spatial (Sum ps)) = Spatial (Sum ps') where ps' = order (reducor ps normaliseNoSort size) order = sort . (map (Product . sort . unProd)) size :: [RegionProd] -> Int size = foldr (+) 0 . map (length . unProd) -- Given a list, a list->list transofmer, a size function -- find the minimal transformed list by applying the transformer -- to every permutation of the list and when a smaller list is found -- iteratively apply to permutations on the smaller list reducor :: [a] -> ([a] -> [a]) -> ([a] -> Int) -> [a] reducor xs f size = reducor' (permutations xs) where reducor' [y] = f y reducor' (y:ys) = if (size y' < size y) then reducor' (permutations y') else reducor' ys where y' = f y fromRegionsToSpec :: IsNatural n => [Span (Vec n Int)] -> Approximation Spatial fromRegionsToSpec = foldr (\x y -> sum (toSpecND x) y) zero -- toSpecND converts an n-dimensional region into an exact -- spatial specification or a bound of spatial specifications toSpecND :: Span (Vec n Int) -> Approximation Spatial toSpecND = toSpecPerDim 1 where -- convert the region one dimension at a time. toSpecPerDim :: Int -> Span (Vec n Int) -> Approximation Spatial toSpecPerDim d (Nil, Nil) = one toSpecPerDim d (Cons l ls, Cons u us) = prod (toSpec1D d l u) (toSpecPerDim (d + 1) (ls, us)) -- toSpec1D takes a dimension identifier, a lower and upper bound of a region in -- that dimension, and builds the simple directional spec. toSpec1D :: Dimension -> Int -> Int -> Approximation Spatial toSpec1D dim l u | l == absoluteRep || u == absoluteRep = Exact $ Spatial (Sum [Product []]) | l == 0 && u == 0 = Exact $ Spatial (Sum [Product [Centered 0 dim True]]) | l < 0 && u == 0 = Exact $ Spatial (Sum [Product [Backward (abs l) dim True]]) | l < 0 && u == (-1) = Exact $ Spatial (Sum [Product [Backward (abs l) dim False]]) | l == 0 && u > 0 = Exact $ Spatial (Sum [Product [Forward u dim True]]) | l == 1 && u > 0 = Exact $ Spatial (Sum [Product [Forward u dim False]]) | l < 0 && u > 0 && (abs l == u) = Exact $ Spatial (Sum [Product [Centered u dim True]]) | l < 0 && u > 0 && (abs l /= u) = Exact $ Spatial (Sum [Product [Backward (abs l) dim True], Product [Forward u dim True]]) -- Represents a non-contiguous region | otherwise = upperBound $ Spatial (Sum [Product [if l > 0 then Forward u dim True else Backward (abs l) dim True]]) {- Normalise a span into the form (lower, upper) based on the first index -} normaliseSpan :: Span (Vec n Int) -> Span (Vec n Int) normaliseSpan (Nil, Nil) = (Nil, Nil) normaliseSpan (a@(Cons l1 ls1), b@(Cons u1 us1)) | l1 <= u1 = (a, b) | otherwise = (b, a) -- DEPRECATED {- `spanBoundingBox` creates a span which is a bounding box over two spans -} spanBoundingBox :: Span (Vec n Int) -> Span (Vec n Int) -> Span (Vec n Int) spanBoundingBox a b = boundingBox' (normaliseSpan a) (normaliseSpan b) where boundingBox' :: Span (Vec n Int) -> Span (Vec n Int) -> Span (Vec n Int) boundingBox' (Nil, Nil) (Nil, Nil) = (Nil, Nil) boundingBox' (Cons l1 ls1, Cons u1 us1) (Cons l2 ls2, Cons u2 us2) = let (ls', us') = boundingBox' (ls1, us1) (ls2, us2) in (Cons (min l1 l2) ls', Cons (max u1 u2) us') {-| Given two spans, if they are consecutive (i.e., (lower1, upper1) (lower2, upper2) where lower2 = upper1 + 1) then compose together returning Just of the new span. Otherwise Nothing -} composeConsecutiveSpans :: Span (Vec n Int) -> Span (Vec n Int) -> [Span (Vec n Int)] composeConsecutiveSpans (Nil, Nil) (Nil, Nil) = [(Nil, Nil)] composeConsecutiveSpans (Cons l1 ls1, Cons u1 us1) (Cons l2 ls2, Cons u2 us2) | (ls1 == ls2) && (us1 == us2) && (u1 + 1 == l2) = [(Cons l1 ls1, Cons u2 us2)] | otherwise = [] {-| |inferMinimalVectorRegions| a key part of the algorithm, from a list of n-dimensional relative indices it infers a list of (possibly overlapping) 1-dimensional spans (vectors) within the n-dimensional space. Built from |minimalise| and |allRegionPermutations| -} inferMinimalVectorRegions :: (Permutable n) => [Vec n Int] -> [Span (Vec n Int)] inferMinimalVectorRegions = fixCoalesce . map mkTrivialSpan where fixCoalesce spans = let spans' = minimaliseRegions . allRegionPermutations $ spans in if spans' == spans then spans' else fixCoalesce spans' {-| Map from a lists of n-dimensional spans of relative indices into all possible contiguous spans within the n-dimensional space (individual pass)-} allRegionPermutations :: (Permutable n) => [Span (Vec n Int)] -> [Span (Vec n Int)] allRegionPermutations = nub . concat . unpermuteIndices . map (coalesceRegions >< id) . groupByPerm . map permutationss where {- Permutations of a indices in a span (independently permutes the lower and upper bounds in the same way) -} permutationss :: Permutable n => Span (Vec n Int) -> [(Span (Vec n Int), Vec n Int -> Vec n Int)] -- Since the permutation ordering is identical for lower & upper bound, -- reuse the same unpermutation permutationss (l, u) = map (\((l', un1), (u', un2)) -> ((l', u'), un1)) $ zip (permutationsV l) (permutationsV u) sortByFst = sortBy (\(l1, u1) (l2, u2) -> compare l1 l2) groupByPerm :: [[(Span (Vec n Int), Vec n Int -> Vec n Int)]] -> [( [Span (Vec n Int)] , Vec n Int -> Vec n Int)] groupByPerm = map (\ixP -> let unPerm = snd $ head ixP in (map fst ixP, unPerm)) . transpose coalesceRegions :: [Span (Vec n Int)] -> [Span (Vec n Int)] coalesceRegions = nub . foldL composeConsecutiveSpans . sortByFst unpermuteIndices :: [([Span (Vec n Int)], Vec n Int -> Vec n Int)] -> [[Span (Vec n Int)]] unpermuteIndices = nub . map (\(rs, unPerm) -> map (unPerm *** unPerm) rs) -- Helper function, reduces a list two elements at a time with a non-determistic operation foldL :: (a -> a -> [a]) -> [a] -> [a] foldL f [] = [] foldL f [a] = [a] foldL f (a:(b:xs)) = case f a b of [] -> a : foldL f (b : xs) cs -> foldL f (cs ++ xs) {-| Collapses the regions into a small set by looking for potential overlaps and eliminating those that overlap -} minimaliseRegions :: [Span (Vec n Int)] -> [Span (Vec n Int)] minimaliseRegions [] = [] minimaliseRegions xss = nub . minimalise $ xss where localMin x ys = (filter' x (\y -> containedWithin x y && (x /= y)) xss) ++ ys minimalise = foldr localMin [] -- If nothing is caught by the filter, i.e. no overlaps then return -- the original regions r filter' r f xs = case filter f xs of [] -> [r] ys -> ys {-| Binary predicate on whether the first region containedWithin the second -} containedWithin :: Span (Vec n Int) -> Span (Vec n Int) -> Bool containedWithin (Nil, Nil) (Nil, Nil) = True containedWithin (Cons l1 ls1, Cons u1 us1) (Cons l2 ls2, Cons u2 us2) = (l2 <= l1 && u1 <= u2) && containedWithin (ls1, us1) (ls2, us2) {-| Defines the (total) class of vector sizes which are permutable, along with the permutation function which pairs permutations with the 'unpermute' operation -} class Permutable (n :: Nat) where -- From a Vector of length n to a list of 'selections' -- (triples of a selected element, the rest of the vector, -- a function to 'unselect') selectionsV :: Vec n a -> [Selection n a] -- From a Vector of length n to a list of its permutations paired with the -- 'unpermute' function permutationsV :: Vec n a -> [(Vec n a, Vec n a -> Vec n a)] -- 'Split' is a size-indexed family which gives the type of selections -- for each size: -- Z is trivial -- (S n) provides a triple of the select element, the remaining vector, -- and the 'unselect' function for returning the original value type family Selection n a where Selection Z a = a Selection (S n) a = (a, Vec n a, a -> Vec n a -> Vec (S n) a) instance Permutable Z where selectionsV Nil = [] permutationsV Nil = [(Nil, id)] instance Permutable (S Z) where selectionsV (Cons x xs) = [(x, Nil, Cons)] permutationsV (Cons x Nil) = [(Cons x Nil, id)] instance Permutable (S n) => Permutable (S (S n)) where selectionsV (Cons x xs) = (x, xs, Cons) : [ (y, Cons x ys, unselect unSel) | (y, ys, unSel) <- selectionsV xs ] where unselect :: (a -> Vec n a -> Vec (S n) a) -> (a -> Vec (S n) a -> Vec (S (S n)) a) unselect f y' (Cons x' ys') = Cons x' (f y' ys') permutationsV xs = [ (Cons y zs, \(Cons y' zs') -> unSel y' (unPerm zs')) | (y, ys, unSel) <- selectionsV xs, (zs, unPerm) <- permutationsV ys ] {- Vector list repreentation where the size 'n' is existential quantified -} data VecList a where VL :: (IsNatural n, Permutable n) => [Vec n a] -> VecList a -- Lists existentially quanitify over a vector's size : Exists n . Vec n a data List a where List :: (IsNatural n, Permutable n) => Vec n a -> List a lnil :: List a lnil = List Nil lcons :: a -> List a -> List a lcons x (List Nil) = List (Cons x Nil) lcons x (List (Cons y Nil)) = List (Cons x (Cons y Nil)) lcons x (List (Cons y (Cons z xs))) = List (Cons x (Cons y (Cons z xs))) fromList :: [a] -> List a fromList = foldr lcons lnil -- pre-condition: the input is a 'rectangular' list of lists (i.e. all internal -- lists have the same size) fromLists :: [[Int]] -> VecList Int fromLists [] = VL ([] :: [Vec Z Int]) fromLists (xs:xss) = consList (fromList xs) (fromLists xss) where consList :: List Int -> VecList Int -> VecList Int consList (List vec) (VL []) = VL [vec] consList (List vec) (VL (x:xs)) = let (vec', x') = zipVec vec x in -- Force the pre-condition equality case (preCondition x' xs, preCondition vec' xs) of (ReflEq, ReflEq) -> VL (vec' : (x' : xs)) where -- At the moment the pre-condition is 'assumed', and therefore -- force used unsafeCoerce: TODO, rewrite preCondition :: Vec n a -> [Vec n1 a] -> EqT n n1 preCondition xs x = unsafeCoerce ReflEq -- Equality type data EqT (a :: k) (b :: k) where ReflEq :: EqT a a -- Local variables: -- mode: haskell -- haskell-program-name: "cabal repl" -- End: