futhark-0.20.1: src/Futhark/Transform/FirstOrderTransform.hs
{-# LANGUAGE ConstraintKinds #-}
{-# LANGUAGE FlexibleContexts #-}
{-# LANGUAGE TypeFamilies #-}
-- | The code generator cannot handle the array combinators (@map@ and
-- friends), so this module was written to transform them into the
-- equivalent do-loops. The transformation is currently rather naive,
-- and - it's certainly worth considering when we can express such
-- transformations in-place.
module Futhark.Transform.FirstOrderTransform
( transformFunDef,
transformConsts,
FirstOrderRep,
Transformer,
transformStmRecursively,
transformLambda,
transformSOAC,
)
where
import Control.Monad.Except
import Control.Monad.State
import Data.List (find, zip4)
import qualified Data.Map.Strict as M
import qualified Futhark.Analysis.Alias as Alias
import qualified Futhark.IR as AST
import Futhark.IR.Prop.Aliases
import Futhark.IR.SOACS
import Futhark.MonadFreshNames
import Futhark.Tools
import Futhark.Util (chunks, splitAt3)
-- | The constraints that must hold for a rep in order to be the
-- target of first-order transformation.
type FirstOrderRep rep =
( Buildable rep,
BuilderOps rep,
LetDec SOACS ~ LetDec rep,
LParamInfo SOACS ~ LParamInfo rep,
CanBeAliased (Op rep)
)
-- | First-order-transform a single function, with the given scope
-- provided by top-level constants.
transformFunDef ::
(MonadFreshNames m, FirstOrderRep torep) =>
Scope torep ->
FunDef SOACS ->
m (AST.FunDef torep)
transformFunDef consts_scope (FunDef entry attrs fname rettype params body) = do
(body', _) <- modifyNameSource $ runState $ runBuilderT m consts_scope
return $ FunDef entry attrs fname rettype params body'
where
m = localScope (scopeOfFParams params) $ transformBody body
-- | First-order-transform these top-level constants.
transformConsts ::
(MonadFreshNames m, FirstOrderRep torep) =>
Stms SOACS ->
m (AST.Stms torep)
transformConsts stms =
fmap snd $ modifyNameSource $ runState $ runBuilderT m mempty
where
m = mapM_ transformStmRecursively stms
-- | The constraints that a monad must uphold in order to be used for
-- first-order transformation.
type Transformer m =
( MonadBuilder m,
LocalScope (Rep m) m,
Buildable (Rep m),
BuilderOps (Rep m),
LParamInfo SOACS ~ LParamInfo (Rep m),
CanBeAliased (Op (Rep m))
)
transformBody ::
(Transformer m, LetDec (Rep m) ~ LetDec SOACS) =>
Body ->
m (AST.Body (Rep m))
transformBody (Body () stms res) = buildBody_ $ do
mapM_ transformStmRecursively stms
pure res
-- | First transform any nested t'Body' or t'Lambda' elements, then
-- apply 'transformSOAC' if the expression is a SOAC.
transformStmRecursively ::
(Transformer m, LetDec (Rep m) ~ LetDec SOACS) =>
Stm ->
m ()
transformStmRecursively (Let pat aux (Op soac)) =
auxing aux $ transformSOAC pat =<< mapSOACM soacTransform soac
where
soacTransform = identitySOACMapper {mapOnSOACLambda = transformLambda}
transformStmRecursively (Let pat aux e) =
auxing aux $ letBind pat =<< mapExpM transform e
where
transform =
identityMapper
{ mapOnBody = \scope -> localScope scope . transformBody,
mapOnRetType = return,
mapOnBranchType = return,
mapOnFParam = return,
mapOnLParam = return,
mapOnOp = error "Unhandled Op in first order transform"
}
-- Produce scratch "arrays" for the Map and Scan outputs of Screma.
-- "Arrays" is in quotes because some of those may be accumulators.
resultArray :: Transformer m => [VName] -> [Type] -> m [VName]
resultArray arrs ts = do
arrs_ts <- mapM lookupType arrs
let oneArray t@Acc {}
| Just (v, _) <- find ((== t) . snd) (zip arrs arrs_ts) =
pure v
oneArray t =
letExp "result" =<< eBlank t
mapM oneArray ts
-- | Transform a single 'SOAC' into a do-loop. The body of the lambda
-- is untouched, and may or may not contain further 'SOAC's depending
-- on the given rep.
transformSOAC ::
Transformer m =>
AST.Pat (Rep m) ->
SOAC (Rep m) ->
m ()
transformSOAC pat (Screma w arrs form@(ScremaForm scans reds map_lam)) = do
-- See Note [Translation of Screma].
--
-- Start by combining all the reduction and scan parts into a single
-- operator
let Reduce _ red_lam red_nes = singleReduce reds
Scan scan_lam scan_nes = singleScan scans
(scan_arr_ts, _red_ts, map_arr_ts) =
splitAt3 (length scan_nes) (length red_nes) $ scremaType w form
scan_arrs <- resultArray [] scan_arr_ts
map_arrs <- resultArray arrs map_arr_ts
scanacc_params <- mapM (newParam "scanacc" . flip toDecl Nonunique) $ lambdaReturnType scan_lam
scanout_params <- mapM (newParam "scanout" . flip toDecl Unique) scan_arr_ts
redout_params <- mapM (newParam "redout" . flip toDecl Nonunique) $ lambdaReturnType red_lam
mapout_params <- mapM (newParam "mapout" . flip toDecl Unique) map_arr_ts
arr_ts <- mapM lookupType arrs
let paramForAcc (Acc c _ _ _) = find (f . paramType) mapout_params
where
f (Acc c2 _ _ _) = c == c2
f _ = False
paramForAcc _ = Nothing
let merge =
concat
[ zip scanacc_params scan_nes,
zip scanout_params $ map Var scan_arrs,
zip redout_params red_nes,
zip mapout_params $ map Var map_arrs
]
i <- newVName "i"
let loopform = ForLoop i Int64 w []
lam_cons = consumedByLambda $ Alias.analyseLambda mempty map_lam
loop_body <- runBodyBuilder
. localScope (scopeOfFParams (map fst merge) <> scopeOf loopform)
$ do
-- Bind the parameters to the lambda.
forM_ (zip3 (lambdaParams map_lam) arrs arr_ts) $ \(p, arr, arr_t) ->
case paramForAcc arr_t of
Just acc_out_p ->
letBindNames [paramName p] . BasicOp $
SubExp $ Var $ paramName acc_out_p
Nothing
| paramName p `nameIn` lam_cons -> do
p' <-
letExp (baseString (paramName p)) $
BasicOp $
Index arr $ fullSlice arr_t [DimFix $ Var i]
letBindNames [paramName p] $ BasicOp $ Copy p'
| otherwise ->
letBindNames [paramName p] $
BasicOp $
Index arr $ fullSlice arr_t [DimFix $ Var i]
-- Insert the statements of the lambda. We have taken care to
-- ensure that the parameters are bound at this point.
mapM_ addStm $ bodyStms $ lambdaBody map_lam
-- Split into scan results, reduce results, and map results.
let (scan_res, red_res, map_res) =
splitAt3 (length scan_nes) (length red_nes) $
bodyResult $ lambdaBody map_lam
scan_res' <-
eLambda scan_lam $
map (pure . BasicOp . SubExp) $
map (Var . paramName) scanacc_params ++ map resSubExp scan_res
red_res' <-
eLambda red_lam $
map (pure . BasicOp . SubExp) $
map (Var . paramName) redout_params ++ map resSubExp red_res
-- Write the scan accumulator to the scan result arrays.
scan_outarrs <-
certifying (foldMap resCerts scan_res) $
letwith (map paramName scanout_params) (Var i) $ map resSubExp scan_res'
-- Write the map results to the map result arrays.
map_outarrs <-
certifying (foldMap resCerts map_res) $
letwith (map paramName mapout_params) (Var i) $ map resSubExp map_res
return . mkBody mempty . concat $
[ scan_res',
varsRes scan_outarrs,
red_res',
varsRes map_outarrs
]
-- We need to discard the final scan accumulators, as they are not
-- bound in the original pattern.
names <-
(++ patNames pat)
<$> replicateM (length scanacc_params) (newVName "discard")
letBindNames names $ DoLoop merge loopform loop_body
transformSOAC pat (Stream w arrs _ nes lam) = do
-- Create a loop that repeatedly applies the lambda body to a
-- chunksize of 1. Hopefully this will lead to this outer loop
-- being the only one, as all the innermost one can be simplified
-- array (as they will have one iteration each).
let (chunk_size_param, fold_params, chunk_params) =
partitionChunkedFoldParameters (length nes) $ lambdaParams lam
mapout_merge <- forM (drop (length nes) $ lambdaReturnType lam) $ \t ->
let t' = t `setOuterSize` w
scratch = BasicOp $ Scratch (elemType t') (arrayDims t')
in (,)
<$> newParam "stream_mapout" (toDecl t' Unique)
<*> letSubExp "stream_mapout_scratch" scratch
-- We need to copy the neutral elements because they may be consumed
-- in the body of the Stream.
let copyIfArray se = do
se_t <- subExpType se
case (se_t, se) of
(Array {}, Var v) -> letSubExp (baseString v) $ BasicOp $ Copy v
_ -> pure se
nes' <- mapM copyIfArray nes
let onType t = t `toDecl` Unique
merge = zip (map (fmap onType) fold_params) nes' ++ mapout_merge
merge_params = map fst merge
mapout_params = map fst mapout_merge
i <- newVName "i"
let loop_form = ForLoop i Int64 w []
letBindNames [paramName chunk_size_param] $
BasicOp $ SubExp $ intConst Int64 1
loop_body <- runBodyBuilder $
localScope (scopeOf loop_form <> scopeOfFParams merge_params) $ do
let slice = [DimSlice (Var i) (Var (paramName chunk_size_param)) (intConst Int64 1)]
forM_ (zip chunk_params arrs) $ \(p, arr) ->
letBindNames [paramName p] . BasicOp . Index arr $
fullSlice (paramType p) slice
(res, mapout_res) <- splitAt (length nes) <$> bodyBind (lambdaBody lam)
mapout_res' <- forM (zip mapout_params mapout_res) $ \(p, SubExpRes cs se) ->
certifying cs . letSubExp "mapout_res" . BasicOp $
Update Unsafe (paramName p) (fullSlice (paramType p) slice) se
mkBodyM mempty $ res ++ subExpsRes mapout_res'
letBind pat $ DoLoop merge loop_form loop_body
transformSOAC pat (Scatter len lam ivs as) = do
iter <- newVName "write_iter"
let (as_ws, as_ns, as_vs) = unzip3 as
ts <- mapM lookupType as_vs
asOuts <- mapM (newIdent "write_out") ts
-- Scatter is in-place, so we use the input array as the output array.
let merge = loopMerge asOuts $ map Var as_vs
loopBody <- runBodyBuilder $
localScope (M.insert iter (IndexName Int64) $ scopeOfFParams $ map fst merge) $ do
ivs' <- forM ivs $ \iv -> do
iv_t <- lookupType iv
letSubExp "write_iv" $ BasicOp $ Index iv $ fullSlice iv_t [DimFix $ Var iter]
ivs'' <- bindLambda lam (map (BasicOp . SubExp) ivs')
let indexes = groupScatterResults (zip3 as_ws as_ns $ map identName asOuts) ivs''
ress <- forM indexes $ \(_, arr, indexes') -> do
arr_t <- lookupType arr
let saveInArray arr' (indexCur, SubExpRes value_cs valueCur) =
certifying (foldMap resCerts indexCur <> value_cs) . letExp "write_out" $
BasicOp $ Update Safe arr' (fullSlice arr_t $ map (DimFix . resSubExp) indexCur) valueCur
foldM saveInArray arr indexes'
return $ resultBody (map Var ress)
letBind pat $ DoLoop merge (ForLoop iter Int64 len []) loopBody
transformSOAC pat (Hist len ops bucket_fun imgs) = do
iter <- newVName "iter"
-- Bind arguments to parameters for the merge-variables.
hists_ts <- mapM lookupType $ concatMap histDest ops
hists_out <- mapM (newIdent "dests") hists_ts
let merge = loopMerge hists_out $ concatMap (map Var . histDest) ops
-- Bind lambda-bodies for operators.
let iter_scope = M.insert iter (IndexName Int64) $ scopeOfFParams $ map fst merge
loopBody <- runBodyBuilder . localScope iter_scope $ do
-- Bind images to parameters of bucket function.
imgs' <- forM imgs $ \img -> do
img_t <- lookupType img
letSubExp "pixel" $ BasicOp $ Index img $ fullSlice img_t [DimFix $ Var iter]
imgs'' <- map resSubExp <$> bindLambda bucket_fun (map (BasicOp . SubExp) imgs')
-- Split out values from bucket function.
let lens = length ops
inds = take lens imgs''
vals = chunks (map (length . lambdaReturnType . histOp) ops) $ drop lens imgs''
hists_out' =
chunks (map (length . lambdaReturnType . histOp) ops) $
map identName hists_out
hists_out'' <- forM (zip4 hists_out' ops inds vals) $ \(hist, op, idx, val) -> do
-- Check whether the indexes are in-bound. If they are not, we
-- return the histograms unchanged.
let outside_bounds_branch = buildBody_ $ pure $ varsRes hist
oob = case hist of
[] -> eSubExp $ constant True
arr : _ -> eOutOfBounds arr [eSubExp idx]
letTupExp "new_histo" <=< eIf oob outside_bounds_branch $
buildBody_ $ do
-- Read values from histogram.
h_val <- forM hist $ \arr -> do
arr_t <- lookupType arr
letSubExp "read_hist" $ BasicOp $ Index arr $ fullSlice arr_t [DimFix idx]
-- Apply operator.
h_val' <-
bindLambda (histOp op) $
map (BasicOp . SubExp) $ h_val ++ val
-- Write values back to histograms.
hist' <- forM (zip hist h_val') $ \(arr, SubExpRes cs v) -> do
arr_t <- lookupType arr
certifying cs . letInPlace "hist_out" arr (fullSlice arr_t [DimFix idx]) $
BasicOp $ SubExp v
pure $ varsRes hist'
return $ resultBody $ map Var $ concat hists_out''
-- Wrap up the above into a for-loop.
letBind pat $ DoLoop merge (ForLoop iter Int64 len []) loopBody
-- | Recursively first-order-transform a lambda.
transformLambda ::
( MonadFreshNames m,
Buildable rep,
BuilderOps rep,
LocalScope somerep m,
SameScope somerep rep,
LetDec rep ~ LetDec SOACS,
CanBeAliased (Op rep)
) =>
Lambda ->
m (AST.Lambda rep)
transformLambda (Lambda params body rettype) = do
body' <-
runBodyBuilder $
localScope (scopeOfLParams params) $
transformBody body
return $ Lambda params body' rettype
letwith :: Transformer m => [VName] -> SubExp -> [SubExp] -> m [VName]
letwith ks i vs = do
let update k v = do
k_t <- lookupType k
case k_t of
Acc {} ->
letExp "lw_acc" $ BasicOp $ SubExp v
_ ->
letInPlace "lw_dest" k (fullSlice k_t [DimFix i]) $ BasicOp $ SubExp v
zipWithM update ks vs
bindLambda ::
Transformer m =>
AST.Lambda (Rep m) ->
[AST.Exp (Rep m)] ->
m Result
bindLambda (Lambda params body _) args = do
forM_ (zip params args) $ \(param, arg) ->
if primType $ paramType param
then letBindNames [paramName param] arg
else letBindNames [paramName param] =<< eCopy (pure arg)
bodyBind body
loopMerge :: [Ident] -> [SubExp] -> [(Param DeclType, SubExp)]
loopMerge vars = loopMerge' $ zip vars $ repeat Unique
loopMerge' :: [(Ident, Uniqueness)] -> [SubExp] -> [(Param DeclType, SubExp)]
loopMerge' vars vals =
[ (Param pname $ toDecl ptype u, val)
| ((Ident pname ptype, u), val) <- zip vars vals
]
-- Note [Translation of Screma]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
-- Screma is the most general SOAC. It is translated by constructing
-- a loop that contains several groups of parameters, in this order:
--
-- (0) Scan accumulator, initialised with neutral element.
-- (1) Scan results, initialised with Scratch.
-- (2) Reduce results (also functioning as accumulators),
-- initialised with neutral element.
-- (3) Map results, mostly initialised with Scratch.
--
-- However, category (3) is a little more tricky in the case where one
-- of the results is an Acc. In that case, the result is not an
-- array, but another Acc. Any Acc result of a Map must correspond to
-- an Acc that is an input to the map, and the result is initialised
-- to be that input. This requires a 1:1 relationship between Acc
-- inputs and Acc outputs, which the type checker should enforce.
-- There is no guarantee that the map results appear in any particular
-- order (e.g. accumulator results before non-accumulator results), so
-- we need to do a little sleuthing to establish the relationship.
--
-- Inside the loop, the non-Acc parameters to map_lam become for-in
-- parameters. Acc parameters refer to the loop parameters for the
-- corresponding Map result instead.
--
-- Intuitively, a Screma(w,
-- (scan_op, scan_ne),
-- (red_op, red_ne),
-- map_fn,
-- {acc_input, arr_input})
--
-- then becomes
--
-- loop (scan_acc, scan_arr, red_acc, map_acc, map_arr) =
-- for i < w, x in arr_input do
-- let (a,b,map_acc',d) = map_fn(map_acc, x)
-- let scan_acc' = scan_op(scan_acc, a)
-- let scan_arr[i] = scan_acc'
-- let red_acc' = red_op(red_acc, b)
-- let map_arr[i] = d
-- in (scan_acc', scan_arr', red_acc', map_acc', map_arr)