| 1 |
{-# LANGUAGE TypeFamilies #-} |
| 2 |
{-# OPTIONS_GHC -fno-warn-orphans #-} |
| 3 |
|
| 4 |
-- | Facilities for converting a 'GPU' program to 'GPUMem'. |
| 5 |
module Futhark.Pass.ExplicitAllocations.GPU |
| 6 |
( explicitAllocations, |
| 7 |
explicitAllocationsInStms, |
| 8 |
) |
| 9 |
where |
| 10 |
|
| 11 |
import Control.Monad |
| 12 |
import Data.Set qualified as S |
| 13 |
import Futhark.IR.GPU |
| 14 |
import Futhark.IR.GPUMem |
| 15 |
import Futhark.IR.Mem.LMAD qualified as LMAD |
| 16 |
import Futhark.Pass.ExplicitAllocations |
| 17 |
import Futhark.Pass.ExplicitAllocations.SegOp |
| 18 |
|
| 19 |
instance SizeSubst (HostOp rep op) where |
| 20 |
opIsConst (SizeOp GetSize {}) = True |
| 21 |
opIsConst (SizeOp GetSizeMax {}) = True |
| 22 |
opIsConst _ = False |
| 23 |
|
| 24 |
allocAtLevel :: SegLevel -> AllocM GPU GPUMem a -> AllocM GPU GPUMem a |
| 25 |
allocAtLevel lvl = local $ \env -> |
| 26 |
env |
| 27 |
{ allocSpace = space, |
| 28 |
allocInOp = handleHostOp (Just lvl) |
| 29 |
} |
| 30 |
where |
| 31 |
space = case lvl of |
| 32 |
SegBlock {} -> Space "shared" |
| 33 |
SegThread {} -> Space "device" |
| 34 |
SegThreadInBlock {} -> Space "device" |
| 35 |
|
| 36 |
handleSegOp :: |
| 37 |
Maybe SegLevel -> |
| 38 |
SegOp SegLevel GPU -> |
| 39 |
AllocM GPU GPUMem (SegOp SegLevel GPUMem) |
| 40 |
handleSegOp outer_lvl op = do |
| 41 |
num_threads <- |
| 42 |
case (outer_lvl, segLevel op) of |
| 43 |
-- This implies we are in the intragroup parallelism situation. |
| 44 |
-- Just allocate for a single group; memory expansion will |
| 45 |
-- handle the rest later. |
| 46 |
(Just (SegBlock _ (Just grid)), _) -> pure $ unCount $ gridBlockSize grid |
| 47 |
_ -> |
| 48 |
letSubExp "num_threads" |
| 49 |
=<< case maybe_grid of |
| 50 |
Just grid -> |
| 51 |
pure . BasicOp $ |
| 52 |
BinOp |
| 53 |
(Mul Int64 OverflowUndef) |
| 54 |
(unCount (gridNumBlocks grid)) |
| 55 |
(unCount (gridBlockSize grid)) |
| 56 |
Nothing -> |
| 57 |
foldBinOp |
| 58 |
(Mul Int64 OverflowUndef) |
| 59 |
(intConst Int64 1) |
| 60 |
(segSpaceDims $ segSpace op) |
| 61 |
allocAtLevel (segLevel op) $ mapSegOpM (mapper num_threads) op |
| 62 |
where |
| 63 |
maybe_grid = |
| 64 |
case (outer_lvl, segLevel op) of |
| 65 |
(Just (SegThread _ (Just grid)), _) -> Just grid |
| 66 |
(Just (SegBlock _ (Just grid)), _) -> Just grid |
| 67 |
(_, SegThread _ (Just grid)) -> Just grid |
| 68 |
(_, SegBlock _ (Just grid)) -> Just grid |
| 69 |
_ -> Nothing |
| 70 |
scope = scopeOfSegSpace $ segSpace op |
| 71 |
mapper num_threads = |
| 72 |
identitySegOpMapper |
| 73 |
{ mapOnSegOpBody = |
| 74 |
localScope scope . local f . allocInKernelBody, |
| 75 |
mapOnSegOpLambda = |
| 76 |
local inThread |
| 77 |
. allocInBinOpLambda num_threads (segSpace op) |
| 78 |
} |
| 79 |
f = case segLevel op of |
| 80 |
SegThread {} -> inThread |
| 81 |
SegThreadInBlock {} -> inThread |
| 82 |
SegBlock {} -> inGroup |
| 83 |
inThread env = env {envExpHints = inThreadExpHints} |
| 84 |
inGroup env = env {envExpHints = inGroupExpHints} |
| 85 |
|
| 86 |
handleHostOp :: |
| 87 |
Maybe SegLevel -> |
| 88 |
HostOp SOAC GPU -> |
| 89 |
AllocM GPU GPUMem (MemOp (HostOp NoOp) GPUMem) |
| 90 |
handleHostOp _ (SizeOp op) = |
| 91 |
pure $ Inner $ SizeOp op |
| 92 |
handleHostOp _ (OtherOp op) = |
| 93 |
error $ "Cannot allocate memory in SOAC: " ++ prettyString op |
| 94 |
handleHostOp outer_lvl (SegOp op) = |
| 95 |
Inner . SegOp <$> handleSegOp outer_lvl op |
| 96 |
handleHostOp _ (GPUBody ts (Body _ stms res)) = |
| 97 |
fmap (Inner . GPUBody ts) . buildBody_ . allocInStms stms $ pure res |
| 98 |
|
| 99 |
kernelExpHints :: Exp GPUMem -> AllocM GPU GPUMem [ExpHint] |
| 100 |
kernelExpHints (BasicOp (Manifest v perm)) = do |
| 101 |
dims <- arrayDims <$> lookupType v |
| 102 |
let perm_inv = rearrangeInverse perm |
| 103 |
dims' = rearrangeShape perm dims |
| 104 |
lmad = LMAD.permute (LMAD.iota 0 $ map pe64 dims') perm_inv |
| 105 |
pure [Hint lmad $ Space "device"] |
| 106 |
kernelExpHints (Op (Inner (SegOp (SegMap lvl@(SegThread _ _) space ts body)))) = |
| 107 |
zipWithM (mapResultHint lvl space) ts $ kernelBodyResult body |
| 108 |
kernelExpHints (Op (Inner (SegOp (SegRed lvl@(SegThread _ _) space ts body reds)))) = |
| 109 |
(map (const NoHint) red_res <>) <$> zipWithM (mapResultHint lvl space) (drop num_reds ts) map_res |
| 110 |
where |
| 111 |
num_reds = segBinOpResults reds |
| 112 |
(red_res, map_res) = splitAt num_reds $ kernelBodyResult body |
| 113 |
kernelExpHints e = defaultExpHints e |
| 114 |
|
| 115 |
mapResultHint :: |
| 116 |
SegLevel -> |
| 117 |
SegSpace -> |
| 118 |
Type -> |
| 119 |
KernelResult -> |
| 120 |
AllocM GPU GPUMem ExpHint |
| 121 |
mapResultHint _lvl space = hint |
| 122 |
where |
| 123 |
-- Heuristic: do not rearrange for returned arrays that are |
| 124 |
-- sufficiently small. |
| 125 |
thresholdBytes = 8 |
| 126 |
|
| 127 |
coalesceReturnOfShape _ [] = False |
| 128 |
coalesceReturnOfShape bs [Constant (IntValue (Int64Value d))] = bs * d > thresholdBytes |
| 129 |
coalesceReturnOfShape _ _ = True |
| 130 |
|
| 131 |
hint t Returns {} |
| 132 |
| coalesceReturnOfShape (primByteSize (elemType t)) $ arrayDims t = do |
| 133 |
let space_dims = segSpaceDims space |
| 134 |
pure $ Hint (innermost space_dims (arrayDims t)) $ Space "device" |
| 135 |
hint _ _ = pure NoHint |
| 136 |
|
| 137 |
innermost :: [SubExp] -> [SubExp] -> LMAD |
| 138 |
innermost space_dims t_dims = |
| 139 |
let r = length t_dims |
| 140 |
dims = space_dims ++ t_dims |
| 141 |
perm = |
| 142 |
[length space_dims .. length space_dims + r - 1] |
| 143 |
++ [0 .. length space_dims - 1] |
| 144 |
perm_inv = rearrangeInverse perm |
| 145 |
dims_perm = rearrangeShape perm dims |
| 146 |
lmad_base = LMAD.iota 0 $ map pe64 dims_perm |
| 147 |
lmad_rearranged = LMAD.permute lmad_base perm_inv |
| 148 |
in lmad_rearranged |
| 149 |
|
| 150 |
semiStatic :: S.Set VName -> SubExp -> Bool |
| 151 |
semiStatic _ Constant {} = True |
| 152 |
semiStatic consts (Var v) = v `S.member` consts |
| 153 |
|
| 154 |
inGroupExpHints :: Exp GPUMem -> AllocM GPU GPUMem [ExpHint] |
| 155 |
inGroupExpHints (Op (Inner (SegOp (SegMap _ space ts body)))) |
| 156 |
| any private $ kernelBodyResult body = do |
| 157 |
consts <- asks envConsts |
| 158 |
pure $ do |
| 159 |
(t, r) <- zip ts $ kernelBodyResult body |
| 160 |
pure $ |
| 161 |
if private r && all (semiStatic consts) (arrayDims t) |
| 162 |
then |
| 163 |
let seg_dims = map pe64 $ segSpaceDims space |
| 164 |
dims = seg_dims ++ map pe64 (arrayDims t) |
| 165 |
nilSlice d = DimSlice 0 d 0 |
| 166 |
in Hint |
| 167 |
( LMAD.slice (LMAD.iota 0 dims) $ |
| 168 |
fullSliceNum dims (map nilSlice seg_dims) |
| 169 |
) |
| 170 |
$ ScalarSpace (arrayDims t) |
| 171 |
$ elemType t |
| 172 |
else NoHint |
| 173 |
where |
| 174 |
private (Returns ResultPrivate _ _) = True |
| 175 |
private _ = False |
| 176 |
inGroupExpHints e = defaultExpHints e |
| 177 |
|
| 178 |
inThreadExpHints :: Exp GPUMem -> AllocM GPU GPUMem [ExpHint] |
| 179 |
inThreadExpHints e = do |
| 180 |
consts <- asks envConsts |
| 181 |
mapM (maybePrivate consts) =<< expExtType e |
| 182 |
where |
| 183 |
maybePrivate consts t |
| 184 |
| Just (Array pt shape _) <- hasStaticShape t, |
| 185 |
all (semiStatic consts) $ shapeDims shape = do |
| 186 |
let lmad = LMAD.iota 0 $ map pe64 $ shapeDims shape |
| 187 |
pure $ Hint lmad $ ScalarSpace (shapeDims shape) pt |
| 188 |
| otherwise = |
| 189 |
pure NoHint |
| 190 |
|
| 191 |
-- | The pass from 'GPU' to 'GPUMem'. |
| 192 |
explicitAllocations :: Pass GPU GPUMem |
| 193 |
explicitAllocations = explicitAllocationsGeneric (Space "device") (handleHostOp Nothing) kernelExpHints |
| 194 |
|
| 195 |
-- | Convert some 'GPU' stms to 'GPUMem'. |
| 196 |
explicitAllocationsInStms :: |
| 197 |
(MonadFreshNames m, HasScope GPUMem m) => |
| 198 |
Stms GPU -> |
| 199 |
m (Stms GPUMem) |
| 200 |
explicitAllocationsInStms = explicitAllocationsInStmsGeneric (Space "device") (handleHostOp Nothing) kernelExpHints |