futhark-0.25.30: src/Futhark/CodeGen/ImpGen/GPU/SegHist.hs
{-# LANGUAGE TypeFamilies #-}
-- | Our compilation strategy for 'SegHist' is based around avoiding
-- bin conflicts. We do this by splitting the input into chunks, and
-- for each chunk computing a single subhistogram. Then we combine
-- the subhistograms using an ordinary segmented reduction ('SegRed').
--
-- There are some branches around to efficiently handle the case where
-- we use only a single subhistogram (because it's large), so that we
-- respect the asymptotics, and do not copy the destination array.
--
-- We also use a heuristic strategy for computing subhistograms in
-- shared memory when possible. Given:
--
-- H: total size of histograms in bytes, including any lock arrays.
--
-- G: block size
--
-- T: number of bytes of shared memory each thread can be given without
-- impacting occupancy (determined experimentally, e.g. 32).
--
-- LMAX: maximum amount of shared memory per threadblock (hard limit).
--
-- We wish to compute:
--
-- COOP: cooperation level (number of threads per subhistogram)
--
-- LH: number of shared memory subhistograms
--
-- We do this as:
--
-- COOP = ceil(H / T)
-- LH = ceil((G*T)/H)
-- if COOP <= G && H <= LMAX then
-- use shared memory
-- else
-- use global memory
module Futhark.CodeGen.ImpGen.GPU.SegHist (compileSegHist) where
import Control.Monad
import Data.List qualified as L
import Data.Map qualified as M
import Data.Maybe
import Futhark.CodeGen.ImpCode.GPU qualified as Imp
import Futhark.CodeGen.ImpGen
import Futhark.CodeGen.ImpGen.GPU.Base
import Futhark.CodeGen.ImpGen.GPU.SegRed (compileSegRed')
import Futhark.Construct (fullSliceNum)
import Futhark.IR.GPUMem
import Futhark.IR.Mem.LMAD qualified as LMAD
import Futhark.Pass.ExplicitAllocations ()
import Futhark.Transform.Substitute
import Futhark.Util (chunks, mapAccumLM, maxinum, splitFromEnd, takeLast)
import Futhark.Util.IntegralExp (divUp, quot, rem)
import Prelude hiding (quot, rem)
data SubhistosInfo = SubhistosInfo
{ subhistosArray :: VName,
subhistosAlloc :: CallKernelGen ()
}
data SegHistSlug = SegHistSlug
{ slugOp :: HistOp GPUMem,
slugNumSubhistos :: TV Int64,
slugSubhistos :: [SubhistosInfo],
slugAtomicUpdate :: AtomicUpdate GPUMem KernelEnv
}
histSpaceUsage ::
HistOp GPUMem ->
Imp.Count Imp.Bytes (Imp.TExp Int64)
histSpaceUsage op =
sum . map (typeSize . (`arrayOfShape` (histShape op <> histOpShape op))) $
lambdaReturnType $
histOp op
histSize :: HistOp GPUMem -> Imp.TExp Int64
histSize = product . map pe64 . shapeDims . histShape
histRank :: HistOp GPUMem -> Int
histRank = shapeRank . histShape
-- | Figure out how much memory is needed per histogram, both
-- segmented and unsegmented, and compute some other auxiliary
-- information.
computeHistoUsage ::
SegSpace ->
HistOp GPUMem ->
CallKernelGen
( Imp.Count Imp.Bytes (Imp.TExp Int64),
Imp.Count Imp.Bytes (Imp.TExp Int64),
SegHistSlug
)
computeHistoUsage space op = do
let segment_dims = init $ unSegSpace space
num_segments = length segment_dims
-- Create names for the intermediate array memory blocks,
-- memory block sizes, arrays, and number of subhistograms.
num_subhistos <- dPrim "num_subhistos"
subhisto_infos <- forM (zip (histDest op) (histNeutral op)) $ \(dest, ne) -> do
dest_t <- lookupType dest
dest_mem <- entryArrayLoc <$> lookupArray dest
subhistos_mem <-
sDeclareMem (baseString dest ++ "_subhistos_mem") (Space "device")
let subhistos_shape =
Shape (map snd segment_dims ++ [tvSize num_subhistos])
<> stripDims num_segments (arrayShape dest_t)
subhistos <-
sArray
(baseString dest ++ "_subhistos")
(elemType dest_t)
subhistos_shape
subhistos_mem
$ LMAD.iota 0
$ map pe64
$ shapeDims subhistos_shape
pure $
SubhistosInfo subhistos $ do
let unitHistoCase =
emit $
Imp.SetMem subhistos_mem (memLocName dest_mem) $
Space "device"
multiHistoCase = do
let num_elems = product $ map pe64 $ shapeDims subhistos_shape
subhistos_mem_size =
Imp.bytes $
Imp.unCount (Imp.elements num_elems `Imp.withElemType` elemType dest_t)
sAlloc_ subhistos_mem subhistos_mem_size $ Space "device"
sReplicate subhistos ne
subhistos_t <- lookupType subhistos
let slice =
fullSliceNum (map pe64 $ arrayDims subhistos_t) $
map (unitSlice 0 . pe64 . snd) segment_dims
++ [DimFix 0]
sUpdate subhistos slice $ Var dest
sIf (tvExp num_subhistos .==. 1) unitHistoCase multiHistoCase
let h = histSpaceUsage op
segmented_h = h * product (map (Imp.bytes . pe64) $ init $ segSpaceDims space)
atomics <- hostAtomics <$> askEnv
pure
( h,
segmented_h,
SegHistSlug op num_subhistos subhisto_infos $
atomicUpdateLocking atomics $
histOp op
)
prepareAtomicUpdateGlobal ::
Maybe Locking ->
Shape ->
[VName] ->
SegHistSlug ->
CallKernelGen
( Maybe Locking,
[Imp.TExp Int64] -> InKernelGen ()
)
prepareAtomicUpdateGlobal l segments dests slug =
-- We need a separate lock array if the operators are not all of a
-- particularly simple form that permits pure atomic operations.
case (l, slugAtomicUpdate slug) of
(_, AtomicPrim f) -> pure (l, f (Space "global") dests)
(_, AtomicCAS f) -> pure (l, f (Space "global") dests)
(Just l', AtomicLocking f) -> pure (l, f l' (Space "global") dests)
(Nothing, AtomicLocking f) -> do
-- The number of locks used here is too low, but since we are
-- currently forced to inline a huge list, I'm keeping it down
-- for now. Some quick experiments suggested that it has little
-- impact anyway (maybe the locking case is just too slow).
--
-- A fun solution would also be to use a simple hashing
-- algorithm to ensure good distribution of locks.
let num_locks = 100151
dims =
map pe64 $
shapeDims segments
++ shapeDims (histOpShape (slugOp slug))
++ [tvSize (slugNumSubhistos slug)]
++ shapeDims (histShape (slugOp slug))
locks <- genZeroes "hist_locks" num_locks
let l' = Locking locks 0 1 0 (pure . (`rem` fromIntegral num_locks) . flattenIndex dims)
pure (Just l', f l' (Space "global") dests)
-- | Some kernel bodies are not safe (or efficient) to execute
-- multiple times.
data Passage = MustBeSinglePass | MayBeMultiPass deriving (Eq, Ord)
bodyPassage :: KernelBody GPUMem -> Passage
bodyPassage kbody
| mempty == consumedInKernelBody (aliasAnalyseKernelBody mempty kbody) =
MayBeMultiPass
| otherwise =
MustBeSinglePass
prepareIntermediateArraysGlobal ::
Passage ->
Shape ->
Imp.TExp Int32 ->
Imp.TExp Int64 ->
[SegHistSlug] ->
CallKernelGen
( Imp.TExp Int32,
[[Imp.TExp Int64] -> InKernelGen ()]
)
prepareIntermediateArraysGlobal passage segments hist_T hist_N slugs = do
-- The paper formulae assume there is only one histogram, but in our
-- implementation there can be multiple that have been horisontally
-- fused. We do a bit of trickery with summings and averages to
-- pretend there is really only one. For the case of a single
-- histogram, the actual calculations should be the same as in the
-- paper.
-- The sum of all Hs.
hist_H <- dPrimVE "hist_H" $ sum $ map (histSize . slugOp) slugs
hist_RF <-
dPrimVE "hist_RF" $
sum (map (r64 . pe64 . histRaceFactor . slugOp) slugs)
/ L.genericLength slugs
hist_el_size <- dPrimVE "hist_el_size" $ sum $ map slugElAvgSize slugs
hist_C_max <-
dPrimVE "hist_C_max" $
fMin64 (r64 hist_T) $
r64 hist_H / hist_k_ct_min
hist_M_min <-
dPrimVE "hist_M_min" $
sMax32 1 $
sExt32 $
t64 $
r64 hist_T / hist_C_max
-- Equivalent to F_L2*L2 in paper.
hist_L2 <- getSize "hist_L2" Imp.SizeCache
let hist_L2_ln_sz = 16 * 4 -- L2 cache line size approximation
hist_RACE_exp <-
dPrimVE "hist_RACE_exp" $
fMax64 1 $
(hist_k_RF * hist_RF)
/ (hist_L2_ln_sz / r64 hist_el_size)
hist_S <- dPrim "hist_S"
-- For sparse histograms (H exceeds N) we only want a single chunk.
sIf
(hist_N .<. hist_H)
(hist_S <-- (1 :: Imp.TExp Int32))
$ hist_S
<-- case passage of
MayBeMultiPass ->
sExt32 $
(sExt64 hist_M_min * hist_H * sExt64 hist_el_size)
`divUp` t64 (hist_F_L2 * r64 (tvExp hist_L2) * hist_RACE_exp)
MustBeSinglePass ->
1
emit $ Imp.DebugPrint "Race expansion factor (RACE^exp)" $ Just $ untyped hist_RACE_exp
emit $ Imp.DebugPrint "Number of chunks (S)" $ Just $ untyped $ tvExp hist_S
histograms <-
snd
<$> mapAccumLM
(onOp (tvExp hist_L2) hist_M_min (tvExp hist_S) hist_RACE_exp)
Nothing
slugs
pure (tvExp hist_S, histograms)
where
hist_k_ct_min = 2 -- Chosen experimentally
hist_k_RF = 0.75 -- Chosen experimentally
hist_F_L2 = 0.4 -- Chosen experimentally
r64 = isF64 . ConvOpExp (SIToFP Int32 Float64) . untyped
t64 = isInt64 . ConvOpExp (FPToSI Float64 Int64) . untyped
-- "Average element size" as computed by a formula that also takes
-- locking into account.
slugElAvgSize slug@(SegHistSlug op _ _ do_op) =
case do_op of
AtomicLocking {} ->
slugElSize slug `quot` (1 + L.genericLength (lambdaReturnType (histOp op)))
_ ->
slugElSize slug `quot` L.genericLength (lambdaReturnType (histOp op))
-- "Average element size" as computed by a formula that also takes
-- locking into account.
slugElSize (SegHistSlug op _ _ do_op) =
sExt32 . unCount . sum $
case do_op of
AtomicLocking {} ->
map (typeSize . (`arrayOfShape` histOpShape op)) $
Prim int32 : lambdaReturnType (histOp op)
_ ->
map (typeSize . (`arrayOfShape` histOpShape op)) $
lambdaReturnType (histOp op)
onOp hist_L2 hist_M_min hist_S hist_RACE_exp l slug = do
let SegHistSlug op num_subhistos subhisto_info do_op = slug
hist_H = histSize op
hist_H_chk <- dPrimVE "hist_H_chk" $ hist_H `divUp` sExt64 hist_S
emit $ Imp.DebugPrint "Chunk size (H_chk)" $ Just $ untyped hist_H_chk
hist_k_max <-
dPrimVE "hist_k_max" $
fMin64
(hist_F_L2 * (r64 hist_L2 / r64 (slugElSize slug)) * hist_RACE_exp)
(r64 hist_N)
/ r64 hist_T
hist_u <- dPrimVE "hist_u" $
case do_op of
AtomicPrim {} -> 2
_ -> 1
hist_C <-
dPrimVE "hist_C" $
fMin64 (r64 hist_T) $
r64 (hist_u * hist_H_chk) / hist_k_max
-- Number of subhistograms per result histogram.
hist_M <- dPrimVE "hist_M" $
case slugAtomicUpdate slug of
AtomicPrim {} -> 1
_ -> sMax32 hist_M_min $ sExt32 $ t64 $ r64 hist_T / hist_C
emit $ Imp.DebugPrint "Elements/thread in L2 cache (k_max)" $ Just $ untyped hist_k_max
emit $ Imp.DebugPrint "Multiplication degree (M)" $ Just $ untyped hist_M
emit $ Imp.DebugPrint "Cooperation level (C)" $ Just $ untyped hist_C
-- num_subhistos is the variable we use to communicate back.
num_subhistos <-- sExt64 hist_M
-- Initialise sub-histograms.
--
-- If hist_M is 1, then we just reuse the original
-- destination. The idea is to avoid a copy if we are writing a
-- small number of values into a very large prior histogram.
dests <- forM (zip (histDest op) subhisto_info) $ \(dest, info) -> do
dest_mem <- entryArrayLoc <$> lookupArray dest
sub_mem <-
fmap memLocName $
entryArrayLoc
<$> lookupArray (subhistosArray info)
let unitHistoCase =
emit $
Imp.SetMem sub_mem (memLocName dest_mem) $
Space "device"
multiHistoCase = subhistosAlloc info
sIf (hist_M .==. 1) unitHistoCase multiHistoCase
pure $ subhistosArray info
(l', do_op') <- prepareAtomicUpdateGlobal l segments dests slug
pure (l', do_op')
histKernelGlobalPass ::
[PatElem LetDecMem] ->
Count NumBlocks SubExp ->
Count BlockSize SubExp ->
SegSpace ->
[SegHistSlug] ->
KernelBody GPUMem ->
[[Imp.TExp Int64] -> InKernelGen ()] ->
Imp.TExp Int32 ->
Imp.TExp Int32 ->
CallKernelGen ()
histKernelGlobalPass map_pes num_tblocks tblock_size space slugs kbody histograms hist_S chk_i = do
let (space_is, space_sizes) = unzip $ unSegSpace space
space_sizes_64 = map (sExt64 . pe64) space_sizes
total_w_64 = product space_sizes_64
hist_H_chks <- forM (map (histSize . slugOp) slugs) $ \w ->
dPrimVE "hist_H_chk" $ w `divUp` sExt64 hist_S
sKernelThread "seghist_global" (segFlat space) (defKernelAttrs num_tblocks tblock_size) $ do
constants <- kernelConstants <$> askEnv
-- Compute subhistogram index for each thread, per histogram.
subhisto_inds <- forM slugs $ \slug ->
dPrimVE "subhisto_ind" $
sExt32 (kernelGlobalThreadId constants)
`quot` ( kernelNumThreads constants
`divUp` sExt32 (tvExp (slugNumSubhistos slug))
)
-- Loop over flat offsets into the input and output. The
-- calculation is done with 64-bit integers to avoid overflow,
-- but the final unflattened segment indexes are 32 bit.
let gtid = sExt64 $ kernelGlobalThreadId constants
num_threads = sExt64 $ kernelNumThreads constants
kernelLoop gtid num_threads total_w_64 $ \offset -> do
-- Construct segment indices.
dIndexSpace (zip space_is space_sizes_64) offset
-- We execute the bucket function once and update each histogram serially.
-- We apply the bucket function if j=offset+ltid is less than
-- num_elements. This also involves writing to the mapout
-- arrays.
let input_in_bounds = offset .<. total_w_64
sWhen input_in_bounds $
compileStms mempty (kernelBodyStms kbody) $ do
let (red_res, map_res) = splitFromEnd (length map_pes) $ kernelBodyResult kbody
sComment "save map-out results" $
forM_ (zip map_pes map_res) $ \(pe, res) ->
copyDWIMFix
(patElemName pe)
(map (Imp.le64 . fst) $ unSegSpace space)
(kernelResultSubExp res)
[]
let red_res_split =
splitHistResults (map slugOp slugs) $
map kernelResultSubExp red_res
sComment "perform atomic updates" $
forM_ (L.zip5 (map slugOp slugs) histograms red_res_split subhisto_inds hist_H_chks) $
\( HistOp dest_shape _ _ _ shape lam,
do_op,
(bucket, vs'),
subhisto_ind,
hist_H_chk
) -> do
let chk_beg = sExt64 chk_i * hist_H_chk
bucket' = map pe64 bucket
dest_shape' = map pe64 $ shapeDims dest_shape
flat_bucket = flattenIndex dest_shape' bucket'
bucket_in_bounds =
chk_beg
.<=. flat_bucket
.&&. flat_bucket
.<. (chk_beg + hist_H_chk)
.&&. inBounds (Slice (map DimFix bucket')) dest_shape'
vs_params = takeLast (length vs') $ lambdaParams lam
sWhen bucket_in_bounds $ do
let bucket_is =
map Imp.le64 (init space_is)
++ [sExt64 subhisto_ind]
++ unflattenIndex dest_shape' flat_bucket
dLParams $ lambdaParams lam
sLoopNest shape $ \is -> do
forM_ (zip vs_params vs') $ \(p, res) ->
copyDWIMFix (paramName p) [] res is
do_op (bucket_is ++ is)
histKernelGlobal ::
[PatElem LetDecMem] ->
Count NumBlocks SubExp ->
Count BlockSize SubExp ->
SegSpace ->
[SegHistSlug] ->
KernelBody GPUMem ->
CallKernelGen ()
histKernelGlobal map_pes num_tblocks tblock_size space slugs kbody = do
let num_tblocks' = fmap pe64 num_tblocks
tblock_size' = fmap pe64 tblock_size
let (_space_is, space_sizes) = unzip $ unSegSpace space
num_threads = sExt32 $ unCount num_tblocks' * unCount tblock_size'
emit $ Imp.DebugPrint "## Using global memory" Nothing
(hist_S, histograms) <-
prepareIntermediateArraysGlobal
(bodyPassage kbody)
(Shape (init space_sizes))
num_threads
(pe64 $ last space_sizes)
slugs
sFor "chk_i" hist_S $ \chk_i ->
histKernelGlobalPass
map_pes
num_tblocks
tblock_size
space
slugs
kbody
histograms
hist_S
chk_i
type InitLocalHistograms =
[ ( [VName],
SubExp ->
InKernelGen
( [VName],
[Imp.TExp Int64] -> InKernelGen ()
)
)
]
prepareIntermediateArraysLocal ::
TV Int32 ->
Count NumBlocks (Imp.TExp Int64) ->
[SegHistSlug] ->
CallKernelGen InitLocalHistograms
prepareIntermediateArraysLocal num_subhistos_per_block blocks_per_segment =
mapM onOp
where
onOp (SegHistSlug op num_subhistos subhisto_info do_op) = do
num_subhistos <-- sExt64 (unCount blocks_per_segment)
emit $
Imp.DebugPrint "Number of subhistograms in global memory per segment" $
Just $
untyped $
tvExp num_subhistos
mk_op <-
case do_op of
AtomicPrim f -> pure $ const $ pure f
AtomicCAS f -> pure $ const $ pure f
AtomicLocking f -> pure $ \hist_H_chk -> do
let lock_shape =
Shape [tvSize num_subhistos_per_block, hist_H_chk]
let dims = [sExt64 (tvExp num_subhistos_per_block), pe64 hist_H_chk]
locks <- sAllocArray "locks" int32 lock_shape $ Space "shared"
sComment "All locks start out unlocked" $
blockCoverSpace dims $ \is ->
copyDWIMFix locks is (intConst Int32 0) []
pure $ f $ Locking locks 0 1 0 id
-- Initialise local-memory sub-histograms. These are
-- represented as two-dimensional arrays.
let init_local_subhistos hist_H_chk = do
local_subhistos <- forM (histType op) $ \t -> do
let subhisto_shape =
setOuterDims
(arrayShape t)
(histRank op)
(Shape [hist_H_chk])
sAllocArray
"subhistogram_local"
(elemType t)
(Shape [tvSize num_subhistos_per_block] <> subhisto_shape)
(Space "shared")
do_op' <- mk_op hist_H_chk
pure (local_subhistos, do_op' (Space "shared") local_subhistos)
-- Initialise global-memory sub-histograms.
glob_subhistos <- forM subhisto_info $ \info -> do
subhistosAlloc info
pure $ subhistosArray info
pure (glob_subhistos, init_local_subhistos)
histKernelLocalPass ::
TV Int32 ->
Count NumBlocks (Imp.TExp Int64) ->
[PatElem LetDecMem] ->
Count NumBlocks SubExp ->
Count BlockSize SubExp ->
SegSpace ->
[SegHistSlug] ->
KernelBody GPUMem ->
InitLocalHistograms ->
Imp.TExp Int32 ->
Imp.TExp Int32 ->
CallKernelGen ()
histKernelLocalPass
num_subhistos_per_block_var
blocks_per_segment
map_pes
num_tblocks
tblock_size
space
slugs
kbody
init_histograms
hist_S
chk_i = do
let (space_is, space_sizes) = unzip $ unSegSpace space
segment_is = init space_is
segment_dims = init space_sizes
(i_in_segment, segment_size) = last $ unSegSpace space
num_subhistos_per_block = tvExp num_subhistos_per_block_var
segment_size' = pe64 segment_size
num_segments <- dPrimVE "num_segments" $ product $ map pe64 segment_dims
hist_H_chks <- forM (map slugOp slugs) $ \op ->
dPrimV "hist_H_chk" $ histSize op `divUp` sExt64 hist_S
histo_sizes <- forM (zip slugs hist_H_chks) $ \(slug, hist_H_chk) -> do
let histo_dims =
tvExp hist_H_chk : map pe64 (shapeDims (histOpShape (slugOp slug)))
histo_size <-
dPrimVE "histo_size" $ product histo_dims
let block_hists_size = sExt64 num_subhistos_per_block * histo_size
init_per_thread <-
dPrimVE "init_per_thread" $ sExt32 $ block_hists_size `divUp` pe64 (unCount tblock_size)
pure (histo_dims, histo_size, init_per_thread)
let attrs = (defKernelAttrs num_tblocks tblock_size) {kAttrCheckSharedMemory = False}
sKernelThread "seghist_local" (segFlat space) attrs $
virtualiseBlocks SegVirt (sExt32 $ unCount blocks_per_segment * num_segments) $ \tblock_id -> do
constants <- kernelConstants <$> askEnv
flat_segment_id <- dPrimVE "flat_segment_id" $ tblock_id `quot` sExt32 (unCount blocks_per_segment)
gid_in_segment <- dPrimVE "gid_in_segment" $ tblock_id `rem` sExt32 (unCount blocks_per_segment)
-- This pgtid is kind of a "virtualised physical" gtid - not the
-- same thing as the gtid used for the SegHist itself.
pgtid_in_segment <-
dPrimVE "pgtid_in_segment" $
gid_in_segment * sExt32 (kernelBlockSize constants)
+ kernelLocalThreadId constants
threads_per_segment <-
dPrimVE "threads_per_segment" $
sExt32 $
unCount blocks_per_segment * kernelBlockSize constants
-- Set segment indices.
zipWithM_ dPrimV_ segment_is $
unflattenIndex (map pe64 segment_dims) $
sExt64 flat_segment_id
histograms <- forM (zip init_histograms hist_H_chks) $
\((glob_subhistos, init_local_subhistos), hist_H_chk) -> do
(local_subhistos, do_op) <- init_local_subhistos $ Var $ tvVar hist_H_chk
pure (zip glob_subhistos local_subhistos, hist_H_chk, do_op)
-- Find index of local subhistograms updated by this thread. We
-- try to ensure, as much as possible, that threads in the same
-- warp use different subhistograms, to avoid conflicts.
thread_local_subhisto_i <-
dPrimVE "thread_local_subhisto_i" $
kernelLocalThreadId constants `rem` num_subhistos_per_block
let onSlugs f =
forM_ (zip3 slugs histograms histo_sizes) $
\(slug, (dests, hist_H_chk, _), (histo_dims, histo_size, init_per_thread)) ->
f slug dests (tvExp hist_H_chk) histo_dims histo_size init_per_thread
let onAllHistograms f =
onSlugs $ \slug dests hist_H_chk histo_dims histo_size init_per_thread -> do
let block_hists_size = num_subhistos_per_block * sExt32 histo_size
forM_ (zip dests (histNeutral $ slugOp slug)) $
\((dest_global, dest_local), ne) ->
sFor "local_i" init_per_thread $ \i -> do
j <-
dPrimVE "j" $
i * sExt32 (kernelBlockSize constants)
+ kernelLocalThreadId constants
j_offset <-
dPrimVE "j_offset" $
num_subhistos_per_block * sExt32 histo_size * gid_in_segment + j
local_subhisto_i <- dPrimVE "local_subhisto_i" $ j `quot` sExt32 histo_size
let local_bucket_is = unflattenIndex histo_dims $ sExt64 $ j `rem` sExt32 histo_size
nested_hist_size =
map pe64 $ shapeDims $ histShape $ slugOp slug
global_bucket_is =
unflattenIndex
nested_hist_size
(head local_bucket_is + sExt64 chk_i * hist_H_chk)
++ tail local_bucket_is
global_subhisto_i <- dPrimVE "global_subhisto_i" $ j_offset `quot` sExt32 histo_size
sWhen (j .<. block_hists_size) $
f
dest_local
dest_global
(slugOp slug)
ne
local_subhisto_i
global_subhisto_i
local_bucket_is
global_bucket_is
sComment "initialize histograms in shared memory" $
onAllHistograms $ \dest_local dest_global op ne local_subhisto_i global_subhisto_i local_bucket_is global_bucket_is ->
sComment "First subhistogram is initialised from global memory; others with neutral element." $ do
dest_global_shape <- map pe64 . arrayDims <$> lookupType dest_global
let global_is = map Imp.le64 segment_is ++ [0] ++ global_bucket_is
local_is = sExt64 local_subhisto_i : local_bucket_is
global_in_bounds =
inBounds (Slice (map DimFix global_is)) dest_global_shape
sIf
(global_subhisto_i .==. 0 .&&. global_in_bounds)
(copyDWIMFix dest_local local_is (Var dest_global) global_is)
( sLoopNest (histOpShape op) $ \is ->
copyDWIMFix dest_local (local_is ++ is) ne []
)
sOp $ Imp.Barrier Imp.FenceLocal
kernelLoop (sExt64 pgtid_in_segment) (sExt64 threads_per_segment) segment_size' $ \ie -> do
dPrimV_ i_in_segment ie
-- We execute the bucket function once and update each histogram
-- serially. This also involves writing to the mapout arrays if
-- this is the first chunk.
compileStms mempty (kernelBodyStms kbody) $ do
let (red_res, map_res) =
splitFromEnd (length map_pes) $
map kernelResultSubExp $
kernelBodyResult kbody
sWhen (chk_i .==. 0) $
sComment "save map-out results" $
forM_ (zip map_pes map_res) $ \(pe, se) ->
copyDWIMFix
(patElemName pe)
(map Imp.le64 space_is)
se
[]
let red_res_split = splitHistResults (map slugOp slugs) red_res
forM_ (zip3 (map slugOp slugs) histograms red_res_split) $
\( HistOp dest_shape _ _ _ shape lam,
(_, hist_H_chk, do_op),
(bucket, vs')
) -> do
let chk_beg = sExt64 chk_i * tvExp hist_H_chk
bucket' = map pe64 bucket
dest_shape' = map pe64 $ shapeDims dest_shape
flat_bucket = flattenIndex dest_shape' bucket'
bucket_in_bounds =
inBounds (Slice (map DimFix bucket')) dest_shape'
.&&. chk_beg
.<=. flat_bucket
.&&. flat_bucket
.<. (chk_beg + tvExp hist_H_chk)
bucket_is =
[sExt64 thread_local_subhisto_i, flat_bucket - chk_beg]
vs_params = takeLast (length vs') $ lambdaParams lam
sComment "perform atomic updates" $
sWhen bucket_in_bounds $ do
dLParams $ lambdaParams lam
sLoopNest shape $ \is -> do
forM_ (zip vs_params vs') $ \(p, v) ->
copyDWIMFix (paramName p) [] v is
do_op (bucket_is ++ is)
sOp $ Imp.ErrorSync Imp.FenceGlobal
sComment "Compact the multiple shared memory subhistograms to result in global memory" $
onSlugs $ \slug dests hist_H_chk histo_dims _histo_size bins_per_thread -> do
trunc_H <-
dPrimV "trunc_H" . sMin64 hist_H_chk $
histSize (slugOp slug) - sExt64 chk_i * head histo_dims
let trunc_histo_dims =
tvExp trunc_H
: map pe64 (shapeDims (histOpShape (slugOp slug)))
trunc_histo_size <- dPrimVE "histo_size" $ sExt32 $ product trunc_histo_dims
sFor "local_i" bins_per_thread $ \i -> do
j <-
dPrimVE "j" $
i * sExt32 (kernelBlockSize constants)
+ kernelLocalThreadId constants
sWhen (j .<. trunc_histo_size) $ do
-- We are responsible for compacting the flat bin 'j', which
-- we immediately unflatten.
let local_bucket_is = unflattenIndex histo_dims $ sExt64 j
nested_hist_size =
map pe64 $ shapeDims $ histShape $ slugOp slug
global_bucket_is =
unflattenIndex
nested_hist_size
(head local_bucket_is + sExt64 chk_i * hist_H_chk)
++ tail local_bucket_is
dLParams $ lambdaParams $ histOp $ slugOp slug
let (global_dests, local_dests) = unzip dests
(xparams, yparams) =
splitAt (length local_dests) $
lambdaParams $
histOp $
slugOp slug
sComment "Read values from subhistogram 0." $
forM_ (zip xparams local_dests) $ \(xp, subhisto) ->
copyDWIMFix
(paramName xp)
[]
(Var subhisto)
(0 : local_bucket_is)
sComment "Accumulate based on values in other subhistograms." $
sFor "subhisto_id" (num_subhistos_per_block - 1) $ \subhisto_id -> do
forM_ (zip yparams local_dests) $ \(yp, subhisto) ->
copyDWIMFix
(paramName yp)
[]
(Var subhisto)
(sExt64 subhisto_id + 1 : local_bucket_is)
compileBody' xparams $ lambdaBody $ histOp $ slugOp slug
sComment "Put final bucket value in global memory." $ do
let global_is =
map Imp.le64 segment_is
++ [sExt64 tblock_id `rem` unCount blocks_per_segment]
++ global_bucket_is
forM_ (zip xparams global_dests) $ \(xp, global_dest) ->
copyDWIMFix global_dest global_is (Var $ paramName xp) []
histKernelLocal ::
TV Int32 ->
Count NumBlocks (Imp.TExp Int64) ->
[PatElem LetDecMem] ->
Count NumBlocks SubExp ->
Count BlockSize SubExp ->
SegSpace ->
Imp.TExp Int32 ->
[SegHistSlug] ->
KernelBody GPUMem ->
CallKernelGen ()
histKernelLocal num_subhistos_per_block_var blocks_per_segment map_pes num_tblocks tblock_size space hist_S slugs kbody = do
let num_subhistos_per_block = tvExp num_subhistos_per_block_var
emit $
Imp.DebugPrint "Number of local subhistograms per block" $
Just $
untyped num_subhistos_per_block
init_histograms <-
prepareIntermediateArraysLocal num_subhistos_per_block_var blocks_per_segment slugs
sFor "chk_i" hist_S $ \chk_i ->
histKernelLocalPass
num_subhistos_per_block_var
blocks_per_segment
map_pes
num_tblocks
tblock_size
space
slugs
kbody
init_histograms
hist_S
chk_i
-- | The maximum number of passes we are willing to accept for this
-- kind of atomic update.
slugMaxLocalMemPasses :: SegHistSlug -> Int
slugMaxLocalMemPasses slug =
case slugAtomicUpdate slug of
AtomicPrim _ -> 3
AtomicCAS _ -> 4
AtomicLocking _ -> 6
localMemoryCase ::
[PatElem LetDecMem] ->
Imp.TExp Int32 ->
SegSpace ->
Imp.TExp Int64 ->
Imp.TExp Int64 ->
Imp.TExp Int64 ->
Imp.TExp Int32 ->
[SegHistSlug] ->
KernelBody GPUMem ->
CallKernelGen (Imp.TExp Bool, CallKernelGen ())
localMemoryCase map_pes hist_T space hist_H hist_el_size hist_N _ slugs kbody = do
let space_sizes = segSpaceDims space
segment_dims = init space_sizes
segmented = not $ null segment_dims
hist_L <- getSize "hist_L" Imp.SizeSharedMemory
max_tblock_size :: TV Int64 <- dPrim "max_tblock_size"
sOp $ Imp.GetSizeMax (tvVar max_tblock_size) Imp.SizeThreadBlock
-- XXX: we need to record for later use that max_tblock_size is the
-- result of GetSizeMax. This is an ugly hack that reflects our
-- inability to track which variables are actually constants.
let withSizeMax vtable =
case M.lookup (tvVar max_tblock_size) vtable of
Just (ScalarVar _ se) ->
M.insert
(tvVar max_tblock_size)
(ScalarVar (Just (Op (Inner (SizeOp (GetSizeMax SizeThreadBlock))))) se)
vtable
_ -> vtable
let tblock_size = Imp.Count $ Var $ tvVar max_tblock_size
num_tblocks <-
fmap (Imp.Count . tvSize) $
dPrimV "num_tblocks" $
sExt64 hist_T `divUp` pe64 (unCount tblock_size)
let num_tblocks' = pe64 <$> num_tblocks
tblock_size' = pe64 <$> tblock_size
let r64 = isF64 . ConvOpExp (SIToFP Int64 Float64) . untyped
t64 = isInt64 . ConvOpExp (FPToSI Float64 Int64) . untyped
-- M approximation.
hist_m' <-
dPrimVE "hist_m_prime" $
r64
( sMin64
(sExt64 (tvExp hist_L `quot` hist_el_size))
(hist_N `divUp` sExt64 (unCount num_tblocks'))
)
/ r64 hist_H
let hist_B = unCount tblock_size'
-- M in the paper, but not adjusted for asymptotic efficiency.
hist_M0 <-
dPrimVE "hist_M0" $
sMax64 1 $
sMin64 (t64 hist_m') hist_B
-- Minimal sequential chunking factor.
let q_small = 2
-- The number of segments/histograms produced..
hist_Nout <- dPrimVE "hist_Nout" $ product $ map pe64 segment_dims
hist_Nin <- dPrimVE "hist_Nin" $ pe64 $ last space_sizes
-- Maximum M for work efficiency.
work_asymp_M_max <-
if segmented
then do
hist_T_hist_min <-
dPrimVE "hist_T_hist_min" $
sExt32 $
sMin64 (sExt64 hist_Nin * sExt64 hist_Nout) (sExt64 hist_T)
`divUp` sExt64 hist_Nout
-- Number of blocks, rounded up.
let r = hist_T_hist_min `divUp` sExt32 hist_B
dPrimVE "work_asymp_M_max" $ hist_Nin `quot` (sExt64 r * hist_H)
else
dPrimVE "work_asymp_M_max" $
(hist_Nout * hist_N)
`quot` ( (q_small * unCount num_tblocks' * hist_H)
`quot` L.genericLength slugs
)
-- Number of subhistograms per result histogram.
hist_M <- dPrimV "hist_M" $ sExt32 $ sMin64 hist_M0 work_asymp_M_max
-- hist_M may be zero (which we'll check for below), but we need it
-- for some divisions first, so crudely make a nonzero form.
let hist_M_nonzero = sMax32 1 $ tvExp hist_M
-- "Cooperation factor" - the number of threads cooperatively
-- working on the same (sub)histogram.
hist_C <-
dPrimVE "hist_C" $
hist_B `divUp` sExt64 hist_M_nonzero
emit $ Imp.DebugPrint "local hist_M0" $ Just $ untyped hist_M0
emit $ Imp.DebugPrint "local work asymp M max" $ Just $ untyped work_asymp_M_max
emit $ Imp.DebugPrint "local C" $ Just $ untyped hist_C
emit $ Imp.DebugPrint "local B" $ Just $ untyped hist_B
emit $ Imp.DebugPrint "local M" $ Just $ untyped $ tvExp hist_M
emit $
Imp.DebugPrint "shared memory needed" $
Just $
untyped $
hist_H * hist_el_size * sExt64 (tvExp hist_M)
-- local_mem_needed is what we need to keep a single bucket in local
-- memory - this is an absolute minimum. We can fit anything else
-- by doing multiple passes, although more than a few is
-- (heuristically) not efficient.
local_mem_needed <-
dPrimVE "local_mem_needed" $
hist_el_size * sExt64 (tvExp hist_M)
-- We add one to the memory requirement because if the chunk
-- otherwise *exactly* fits, it might actually *not* fit in the case
-- of a multi-value operator, as we individually round up the sizes
-- of the component arrays. (Very rare edge case.)
hist_S <-
dPrimVE "hist_S" . sExt32 $
(hist_H * local_mem_needed + 1) `divUp` tvExp hist_L
let max_S = case bodyPassage kbody of
MustBeSinglePass -> 1
MayBeMultiPass -> fromIntegral $ maxinum $ map slugMaxLocalMemPasses slugs
blocks_per_segment <-
if segmented
then
fmap Count $
dPrimVE "blocks_per_segment" $
unCount num_tblocks' `divUp` hist_Nout
else pure num_tblocks'
-- We only use shared memory if the number of updates per histogram
-- at least matches the histogram size, as otherwise it is not
-- asymptotically efficient. This mostly matters for the segmented
-- case.
let pick_local =
hist_Nin
.>=. hist_H
.&&. (local_mem_needed .<=. tvExp hist_L)
.&&. (hist_S .<=. max_S)
.&&. hist_C
.<=. hist_B
.&&. tvExp hist_M
.>. 0
run = do
emit $ Imp.DebugPrint "## Using shared memory" Nothing
emit $ Imp.DebugPrint "Histogram size (H)" $ Just $ untyped hist_H
emit $ Imp.DebugPrint "Multiplication degree (M)" $ Just $ untyped $ tvExp hist_M
emit $ Imp.DebugPrint "Cooperation level (C)" $ Just $ untyped hist_C
emit $ Imp.DebugPrint "Number of chunks (S)" $ Just $ untyped hist_S
when segmented $
emit $
Imp.DebugPrint "Blocks per segment" $
Just $
untyped $
unCount blocks_per_segment
localVTable withSizeMax $
histKernelLocal
hist_M
blocks_per_segment
map_pes
num_tblocks
tblock_size
space
hist_S
slugs
kbody
pure (pick_local, run)
-- | Generate code for a segmented histogram called from the host.
compileSegHist ::
Pat LetDecMem ->
SegLevel ->
SegSpace ->
[HistOp GPUMem] ->
KernelBody GPUMem ->
CallKernelGen ()
compileSegHist (Pat pes) lvl space ops kbody = do
KernelAttrs {kAttrNumBlocks = num_tblocks, kAttrBlockSize = tblock_size} <-
lvlKernelAttrs lvl
-- Most of this function is not the histogram part itself, but
-- rather figuring out whether to use a local or global memory
-- strategy, as well as collapsing the subhistograms produced (which
-- are always in global memory, but their number may vary).
let num_tblocks' = fmap pe64 num_tblocks
tblock_size' = fmap pe64 tblock_size
dims = map pe64 $ segSpaceDims space
num_red_res = length ops + sum (map (length . histNeutral) ops)
(all_red_pes, map_pes) = splitAt num_red_res pes
segment_size = last dims
(op_hs, op_seg_hs, slugs) <- unzip3 <$> mapM (computeHistoUsage space) ops
h <- dPrimVE "h" $ Imp.unCount $ sum op_hs
seg_h <- dPrimVE "seg_h" $ Imp.unCount $ sum op_seg_hs
-- Check for emptyness to avoid division-by-zero.
sUnless (seg_h .==. 0) $ do
-- Maximum block size (or actual, in this case).
let hist_B = unCount tblock_size'
-- Size of a histogram.
hist_H <- dPrimVE "hist_H" $ sum $ map histSize ops
-- Size of a single histogram element. Actually the weighted
-- average of histogram elements in cases where we have more than
-- one histogram operation, plus any locks.
let lockSize slug = case slugAtomicUpdate slug of
AtomicLocking {} -> Just $ primByteSize int32
_ -> Nothing
hist_el_size <-
dPrimVE "hist_el_size" $
L.foldl' (+) (h `divUp` hist_H) $
mapMaybe lockSize slugs
-- Input elements contributing to each histogram.
hist_N <- dPrimVE "hist_N" segment_size
-- Compute RF as the average RF over all the histograms.
hist_RF <-
dPrimVE "hist_RF" $
sExt32 $
sum (map (pe64 . histRaceFactor . slugOp) slugs)
`quot` L.genericLength slugs
let hist_T = sExt32 $ unCount num_tblocks' * unCount tblock_size'
emit $ Imp.DebugPrint "\n# SegHist" Nothing
emit $ Imp.DebugPrint "Number of threads (T)" $ Just $ untyped hist_T
emit $ Imp.DebugPrint "Desired block size (B)" $ Just $ untyped hist_B
emit $ Imp.DebugPrint "Histogram size (H)" $ Just $ untyped hist_H
emit $ Imp.DebugPrint "Input elements per histogram (N)" $ Just $ untyped hist_N
emit $
Imp.DebugPrint "Number of segments" $
Just $
untyped $
product $
map (pe64 . snd) segment_dims
emit $ Imp.DebugPrint "Histogram element size (el_size)" $ Just $ untyped hist_el_size
emit $ Imp.DebugPrint "Race factor (RF)" $ Just $ untyped hist_RF
emit $ Imp.DebugPrint "Memory per set of subhistograms per segment" $ Just $ untyped h
emit $ Imp.DebugPrint "Memory per set of subhistograms times segments" $ Just $ untyped seg_h
(use_shared_memory, run_in_shared_memory) <-
localMemoryCase map_pes hist_T space hist_H hist_el_size hist_N hist_RF slugs kbody
sIf use_shared_memory run_in_shared_memory $
histKernelGlobal map_pes num_tblocks tblock_size space slugs kbody
let pes_per_op = chunks (map (length . histDest) ops) all_red_pes
forM_ (zip3 slugs pes_per_op ops) $ \(slug, red_pes, op) -> do
let num_histos = slugNumSubhistos slug
subhistos = map subhistosArray $ slugSubhistos slug
let unitHistoCase =
-- This is OK because the memory blocks are at least as
-- large as the ones we are supposed to use for the result.
forM_ (zip red_pes subhistos) $ \(pe, subhisto) -> do
pe_mem <-
memLocName . entryArrayLoc
<$> lookupArray (patElemName pe)
subhisto_mem <-
memLocName . entryArrayLoc
<$> lookupArray subhisto
emit $ Imp.SetMem pe_mem subhisto_mem $ Space "device"
sIf (tvExp num_histos .==. 1) unitHistoCase $ do
-- For the segmented reduction, we keep the segment dimensions
-- unchanged. To this, we add two dimensions: one over the number
-- of buckets, and one over the number of subhistograms. This
-- inner dimension is the one that is collapsed in the reduction.
bucket_ids <-
replicateM (shapeRank (histShape op)) (newVName "bucket_id")
subhistogram_id <- newVName "subhistogram_id"
vector_ids <-
replicateM (shapeRank (histOpShape op)) (newVName "vector_id")
flat_gtid <- newVName "flat_gtid"
let grid = KernelGrid num_tblocks tblock_size
segred_space =
SegSpace flat_gtid $
segment_dims
++ zip bucket_ids (shapeDims (histShape op))
++ zip vector_ids (shapeDims $ histOpShape op)
++ [(subhistogram_id, Var $ tvVar num_histos)]
-- The operator may have references to the old flat thread
-- ID, which we must update to point at the new one.
subst = M.singleton (segFlat space) flat_gtid
let segred_op = SegBinOp Commutative (substituteNames subst $ histOp op) (histNeutral op) mempty
compileSegRed' (Pat red_pes) grid segred_space [segred_op] $ \red_cont ->
red_cont . flip map subhistos $ \subhisto ->
( Var subhisto,
map Imp.le64 $
map fst segment_dims
++ [subhistogram_id]
++ bucket_ids
++ vector_ids
)
emit $ Imp.DebugPrint "" Nothing
where
segment_dims = init $ unSegSpace space