Merge pull request #1316 from axch/vectorize-user-index-sets

Vectorize through user-defined index sets

Author: axch

src/lib/ImpToLLVM.hs

@@ -902,16 +902,33 @@ withWidthOfFP x template = case typeOf template of
   L.FloatingPointType L.FloatFP  -> litVal $ Float32Lit $ realToFrac x
   _ -> error $ "Unsupported floating point type: " ++ show (typeOf template)
 
+-- If we are accessing a `L.Type` from a Dex array, what memory alignment (in
+-- bytes) can we guarantee?  This is probably better expressed in Dex types, but
+-- we would need to plumb them to do it that way.  1-byte alignment should
+-- always be safe, but we can promise higher-performance alignments for some
+-- types.
+dexAlignment :: L.Type -> Word32
+dexAlignment = \case
+  L.IntegerType bits | bits `mod` 8 == 0 -> bits `div` 8
+  L.IntegerType _ -> 1
+  L.PointerType _ _ -> 4
+  L.FloatingPointType L.FloatFP -> 4
+  L.FloatingPointType L.DoubleFP -> 8
+  L.VectorType _ eltTy -> dexAlignment eltTy
+  _ -> 1
+
 store :: LLVMBuilder m => Operand -> Operand -> m ()
-store ptr x =  addInstr $ L.Do $ L.Store False ptr x Nothing 0 []
+store ptr x = addInstr $ L.Do $ L.Store False ptr x Nothing alignment [] where
+  alignment = dexAlignment $ typeOf x
 
 load :: LLVMBuilder m => L.Type -> Operand -> m Operand
 load pointeeTy ptr =
 #if MIN_VERSION_llvm_hs(15,0,0)
-  emitInstr pointeeTy $ L.Load False pointeeTy ptr Nothing 0 []
+  emitInstr pointeeTy $ L.Load False pointeeTy ptr Nothing alignment []
 #else
-  emitInstr pointeeTy $ L.Load False ptr Nothing 0 []
+  emitInstr pointeeTy $ L.Load False ptr Nothing alignment []
 #endif
+  where alignment = dexAlignment pointeeTy
 
 ilt :: LLVMBuilder m => Operand -> Operand -> m Operand
 ilt x y = emitInstr i1 $ L.ICmp IP.SLT x y []

src/lib/Vectorize.hs

@@ -49,9 +49,13 @@ import Util (allM, zipWithZ)
 -- TODO: Local vector values? We might want to pack short and pure for loops into vectors,
 -- to support things like float3 etc.
 data Stability
-  = Uniform    -- constant across vectorized dimension
-  | Varying    -- varying across vectorized dimension
-  | Contiguous -- varying, but contiguous across vectorized dimension
+  -- Constant across vectorized dimension, represented as a scalar
+  = Uniform
+  -- Varying across vectorized dimension, represented as a vector
+  | Varying
+  -- Varying, but contiguous across vectorized dimension; represented as a
+  -- scalar carrying the first value
+  | Contiguous
   | ProdStability [Stability]
   deriving (Eq, Show)
 
@@ -168,25 +172,27 @@ vectorizeLoopsExpr expr = do
   narrowestTypeByteWidth <- getNarrowestTypeByteWidth =<< renameM expr
   let loopWidth = vectorByteWidth `div` narrowestTypeByteWidth
   case expr of
-    PrimOp (DAMOp (Seq effs dir (IxType IdxRepTy (IxDictRawFin (IdxRepVal n))) dest body))
-      | n `mod` loopWidth == 0 -> (do
-          safe <- vectorSafeEffect effs
-          if safe
-            then (do
-              Distinct <- getDistinct
-              let vn = n `div` loopWidth
-              body' <- vectorizeSeq loopWidth body
-              dest' <- renameM dest
-              seqOp <- mkSeq dir (IxType IdxRepTy (IxDictRawFin (IdxRepVal vn))) dest' body'
-              return $ PrimOp $ DAMOp seqOp)
-            else renameM expr)
-        `catchErr` \errs -> do
-            let msg = "In `vectorizeLoopsDecls`:\nExpr:\n" ++ pprint expr
-                ctx = mempty { messageCtx = [msg] }
-                errs' = prependCtxToErrs ctx errs
-            modify (<> LiftE errs')
-            recurSeq expr
-    PrimOp (DAMOp (Seq _ _ _ _ _)) -> recurSeq expr
+    PrimOp (DAMOp (Seq effs dir ixty dest body)) -> do
+      sz <- simplifyIxSize =<< renameM ixty
+      case sz of
+        Just n | n `mod` loopWidth == 0 -> (do
+            safe <- vectorSafeEffect effs
+            if safe
+              then (do
+                Distinct <- getDistinct
+                let vn = n `div` loopWidth
+                body' <- vectorizeSeq loopWidth ixty body
+                dest' <- renameM dest
+                seqOp <- mkSeq dir (IxType IdxRepTy (IxDictRawFin (IdxRepVal vn))) dest' body'
+                return $ PrimOp $ DAMOp seqOp)
+              else renameM expr)
+          `catchErr` \errs -> do
+              let msg = "In `vectorizeLoopsDecls`:\nExpr:\n" ++ pprint expr
+                  ctx = mempty { messageCtx = [msg] }
+                  errs' = prependCtxToErrs ctx errs
+              modify (<> LiftE errs')
+              recurSeq expr
+        _ -> recurSeq expr
     PrimOp (Hof (TypedHof _ (RunReader item (BinaryLamExpr hb' refb' body)))) -> do
       item' <- renameM item
       itemTy <- return $ getType item'
@@ -218,6 +224,15 @@ vectorizeLoopsExpr expr = do
       return $ PrimOp $ DAMOp $ Seq effs' dir ixty' dest' body'
     recurSeq _ = error "Impossible"
 
+simplifyIxSize :: (EnvReader m, ScopableBuilder SimpIR m)
+  => IxType SimpIR n -> m n (Maybe Word32)
+simplifyIxSize ixty = do
+  sizeMethod <- buildBlock $ applyIxMethod (sink $ ixTypeDict ixty) Size []
+  cheapReduce sizeMethod >>= \case
+    Just (IdxRepVal n) -> return $ Just n
+    _ -> return Nothing
+{-# INLINE simplifyIxSize #-}
+
 -- Really we should check this by seeing whether there is an instance for a
 -- `Commutative` class, or something like that, but for now just pattern-match
 -- to detect scalar addition as the only monoid we recognize as commutative.
@@ -300,22 +315,27 @@ vectorSafeEffect (EffectRow effs NoTail) = allM safe $ eSetToList effs where
       Nothing -> error $ "Handle " ++ pprint h ++ " not present in commute map?"
   safe _ = return False
 
-vectorizeSeq :: Word32 -> LamExpr SimpIR i -> TopVectorizeM i o (LamExpr SimpIR o)
-vectorizeSeq loopWidth (UnaryLamExpr (b:>ty) body) = do
-  (_, ty') <- case ty of
-    ProdTy [ixTy, ref] -> do
-      ixTy' <- renameM ixTy
+vectorizeSeq :: Word32 -> IxType SimpIR i -> LamExpr SimpIR i
+  -> TopVectorizeM i o (LamExpr SimpIR o)
+vectorizeSeq loopWidth ixty (UnaryLamExpr (b:>ty) body) = do
+  newLoopTy <- case ty of
+    ProdTy [_ixType, ref] -> do
       ref' <- renameM ref
-      return (ixTy', ProdTy [IdxRepTy, ref'])
+      return $ ProdTy [IdxRepTy, ref']
     _ -> error "Unexpected seq binder type"
+  ixty' <- renameM ixty
   liftVectorizeM loopWidth $
-    buildUnaryLamExpr (getNameHint b) ty' \ci -> do
-      -- XXX: we're assuming `Fin n` here
+    buildUnaryLamExpr (getNameHint b) newLoopTy \ci -> do
+      -- The per-tile loop iterates on `Fin`
       (viOrd, dest) <- fromPair $ Var ci
       iOrd <- imul viOrd $ IdxRepVal loopWidth
-      extendSubst (b @> VVal (ProdStability [Contiguous, ProdStability [Uniform]]) (PairVal iOrd dest)) $
+      -- TODO: It would be nice to cancel this UnsafeFromOrdinal with the
+      -- Ordinal that will be taken later when indexing, but that should
+      -- probably be a separate pass.
+      i <- applyIxMethod (sink $ ixTypeDict ixty') UnsafeFromOrdinal [iOrd]
+      extendSubst (b @> VVal (ProdStability [Contiguous, ProdStability [Uniform]]) (PairVal i dest)) $
         vectorizeBlock body $> UnitVal
-vectorizeSeq _ _ = error "expected a unary lambda expression"
+vectorizeSeq _ _ _ = error "expected a unary lambda expression"
 
 newtype VectorizeM i o a =
   VectorizeM { runVectorizeM ::
@@ -467,9 +487,13 @@ vectorizePrimOp op = case op of
   BinOp opk arg1 arg2 -> do
     sx@(VVal vx x) <- vectorizeAtom arg1
     sy@(VVal vy y) <- vectorizeAtom arg2
-    let v = case (vx, vy) of (Uniform, Uniform) -> Uniform; _ -> Varying
-    x' <- if vx /= v then ensureVarying sx else return x
-    y' <- if vy /= v then ensureVarying sy else return y
+    let v = case (opk, vx, vy) of
+              (_, Uniform, Uniform) -> Uniform
+              (IAdd, Uniform, Contiguous) -> Contiguous
+              (IAdd, Contiguous, Uniform) -> Contiguous
+              _ -> Varying
+    x' <- if v == Varying then ensureVarying sx else return x
+    y' <- if v == Varying then ensureVarying sy else return y
     VVal v <$> emitOp (BinOp opk x' y')
   MiscOp (CastOp tyArg arg) -> do
     ty <- vectorizeType tyArg

tests/opt-tests.dx

@@ -126,13 +126,13 @@ _ = for i:(Fin 20) j:(Fin 4). ordinal j
 "vectorizing int binary op"
 -- CHECK-LABEL: vectorizing int binary op
 %passes vect
-_ = for i:(Fin 256). (n_to_i32 (ordinal i)) + 1
+_ = for i:(Fin 256). (n_to_i32 (ordinal i)) * 2
 -- CHECK: seq (RawFin 0x10)
 -- CHECK: [[i0:v#[0-9]+]]:<16xInt32> = vbroadcast
 -- CHECK: [[i1:v#[0-9]+]]:<16xInt32> = viota
 -- CHECK: [[i2:v#[0-9]+]]:<16xInt32> = %iadd [[i0]] [[i1]]
--- CHECK: [[ones:v#[0-9]+]]:<16xInt32> = vbroadcast 1
--- CHECK: %iadd [[i2]] [[ones]]
+-- CHECK: [[twos:v#[0-9]+]]:<16xInt32> = vbroadcast 2
+-- CHECK: %imul [[i2]] [[twos]]
 
 "vectorizing float binary op"
 -- CHECK-LABEL: vectorizing float binary op
@@ -211,3 +211,38 @@ _ = yield_accum (AddMonoid Int32) \result.
 -- CHECK: [[mat1:v#[0-9]+]]:<16xInt32> = vbroadcast
 -- CHECK: [[prodj:v#[0-9]+]]:<16xInt32> = %imul [[mat1]] [[mat2j]]
 -- CHECK: extend [[refj]] [[prodj]]
+
+"vectorizing through the `tile` combinator and its funny index set"
+-- CHECK-LABEL: vectorizing through the `tile` combinator and its funny index set
+
+%passes vect
+_ = yield_accum (AddMonoid Int32) \result.
+  tile((Fin 256), 32) \set.
+    for_ i:set.
+      ix = inject(i, to=(Fin 256))
+      result!ix += xs[ix]
+-- CHECK: seq (RawFin 0x8)
+-- CHECK: seq (RawFin 0x2)
+-- CHECK: [[refix:v#[0-9]+]]:(Ref {{v#[0-9]+}} <16xInt32>) = vrefslice
+-- CHECK: [[xsix:v#[0-9]+]]:<16xInt32> =
+-- CHECK-NEXT: vslice
+-- CHECK: extend [[refix]] [[xsix]]
+
+"Non-aligned"
+-- CHECK-LABEL: Non-aligned
+
+-- This is a regression test.  We are checking that Dex-side
+-- vectorization does not end up assuming that arrays are aligned on
+-- the size of the vectors, only on the size of the underlying
+-- scalars.
+
+non_aligned = for i:(Fin 7). for j:(Fin 257). +0
+
+%passes llvm
+_ = yield_accum (AddMonoid Int32) \result.
+  tile((Fin 257), 32) \set.
+    for_ i:set.
+      ix = inject(i, to=(Fin 257))
+      result!(6@(Fin 7))!ix += non_aligned[6@_][ix]
+-- CHECK: load <16 x i32>, <16 x i32>* %"v#{{[0-9]+}}", align 4
+-- CHECK: store <16 x i32> %"v#{{[0-9]+}}", <16 x i32>* %"v#{{[0-9]+}}", align 4