[mlir][linalg] Add bufferization patterns for non-DPS ops

[springerm]

Summary

These patterns bufferize non-destination-style ops into a newly materialized buffer allocation. The provided helper function can be used to materialize allocations of tensor.pad when folding is not desirable and an allocation must happen.

The resulting IR can be further bufferized with One-Shot Bufferize.

Depends On: D144021

[nicolasvasilache]

mlir/test/Dialect/Linalg/bufferize-non-dps-ops.mlir

2

Since this is a new file, we do not need to continue using a test pass to test this.
Can we use a new transform op instead?

[springerm]

update

mlir/include/mlir/Dialect/Linalg/Transforms/Transforms.h

 struct LinalgFusionOptions;
 struct LinalgTilingOptions;
 
 //===----------------------------------------------------------------------===//
 // Transformations exposed as function calls.
 //===----------------------------------------------------------------------===//
 using LinalgLoops = SmallVector<Operation *, 4>;
 
+/// Materialize a buffer allocation for the given tensor.pad op and lower the
+/// op to linalg.fill/linalg.generic + memref.tensor_store. E.g.:
+///
+/// %0 = tensor.pad low[%l] high[%h] %t ...
+///
+/// is lowered to:
+///
+/// %alloc = memref.alloc
+/// linalg.fill ... outs(%alloc)
+/// %subview = memref.subview %alloc [%l] [...] [1]
+/// memref.tensor_store %t, %subview
+/// %0 = bufferization.to_tensor %alloc restrict writable
+///
+/// In addition to rewriting the IR as shown above, the result of the
+/// bufferization.to_tensor op is returned.
+Value bufferizePadOp(RewriterBase &rewriter, tensor::PadOp padOp,
+                     Attribute memorySpace);
+
 void populatePadTensorTilingPatterns(RewritePatternSet &patterns,
                                      const LinalgTilingOptions &options);
 
 /// Populate patterns for splitting a `LinalgOp` with multiple statements within
 /// its payload into multiple `GenericOp` that have a single statement.
 /// The option `removeDeadArgsAndResults` adds patterns to remove dead arguments
 /// and results from the generated decomposed ops. This is default `true` since
 /// the core decomposition patterns relies on these clean up patterns. It is set
 /// to false only for testing purposes.
 void populateDecomposeLinalgOpsPattern(RewritePatternSet &patterns,
                                        bool removeDeadArgsAndResults = true);
 
 /// Populate patterns that convert non-destination-style ops to destination
 /// style ops.
 void populateConvertToDestinationStylePatterns(RewritePatternSet &patterns);
 
+/// Populate patterns that bufferize and materialize allocations for
+/// non-destination-style ops.
+void populateBufferizeNonDpsOpsPatterns(RewritePatternSet &patterns,
+                                        Attribute memorySpace = {});
+
 /// Populate patterns for vectorizing low-D convolution ops. This is a step in
 /// progressive lowering for convolution ops, it assume high-D convolution ops
 /// were decomposed previously.
 void populateConvolutionVectorizationPatterns(RewritePatternSet &patterns,
                                               PatternBenefit benefit = 1);
 
 /// Populate patterns that convert `ElementwiseMappable` ops to linalg
 /// parallel loops.

mlir/lib/Dialect/Linalg/Transforms/ConvertToDestinationStyle.cpp

 //===- ConvertToDestinationStyle.cpp - Convert non-DPS to DPS ops ---------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This file contains patterns to convert non-DPS ops to DPS ops. New
 // tensor.empty ops are inserted as a destination. Such tensor.empty can be
 // eliminated with "empty tensor elimination", allowing them to bufferize
 // without an allocation (assuming there are no further conflicts).
 //
 //===----------------------------------------------------------------------===//
 //
 #include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h"
+#include "mlir/Dialect/Bufferization/IR/Bufferization.h"
 #include "mlir/Dialect/Linalg/IR/Linalg.h"
 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/IR/Matchers.h"
 #include "mlir/IR/PatternMatch.h"
 #include "llvm/Support/Debug.h"
 
 using namespace mlir;
LogicalResult matchAndRewrite(GenerateOp generateOp,
     auto yieldOp = cast<tensor::YieldOp>(body->getTerminator());
     rewriter.replaceOpWithNewOp<linalg::YieldOp>(yieldOp, yieldOp.getValue());
 
     // Replace tensor.generate.
     rewriter.replaceOp(generateOp, genericOp->getResult(0));
     return success();
   }
 };
+} // namespace
 
-/// Lower tensor.pad to linalg.generic + tensor.insert_slice.
-struct PadOpConverter : public OpRewritePattern<PadOp> {
-  using OpRewritePattern<PadOp>::OpRewritePattern;
-
-  LogicalResult matchAndRewrite(PadOp padOp,
-                                PatternRewriter &rewriter) const override {
-    // Only ops with exactly one block are supported.
-    if (!padOp.getBodyRegion().hasOneBlock())
-      return failure();
-
-    // Create tensor.empty.
-    Location loc = padOp.getLoc();
-    RankedTensorType resultType = padOp.getResultType();
-    ReifiedRankedShapedTypeDims reifiedShape;
-    if (failed(cast<ReifyRankedShapedTypeOpInterface>(padOp.getOperation())
-                   .reifyResultShapes(rewriter, reifiedShape)))
-      return rewriter.notifyMatchFailure(
-          padOp, "failed to reify tensor.pad op result shape");
-    SmallVector<Value> dynamicSizes;
-    for (int64_t i = 0; i < resultType.getRank(); ++i)
-      if (resultType.isDynamicDim(i))
-        dynamicSizes.push_back(reifiedShape[0][i]);
-    auto emptyOp = rewriter.create<EmptyOp>(loc, resultType, dynamicSizes);
+static Operation *movePaddingToFillOrGenericOp(RewriterBase &rewriter,
+                                               Location loc, PadOp padOp,
+                                               Value dest) {
+  OpBuilder::InsertionGuard g(rewriter);
 
-    // Examine the yielded value to decide if a linalg.generic is neede or a
-    // linalg.fill is sufficient.
-    Value filled;
-    Value yieldedValue =
-        cast<tensor::YieldOp>(padOp.getBody()->getTerminator()).getValue();
-    Attribute constYieldedValue;
-    // Is the yielded value a bbArg defined outside of the PadOp?
-    bool outsideBbArg =
-        yieldedValue.isa<BlockArgument>() &&
-        yieldedValue.cast<BlockArgument>().getOwner()->getParentOp() !=
-            padOp.getOperation();
-    // Is the yielded value an OpResult defined outside of the PadOp?
-    bool outsideOpResult =
-        yieldedValue.isa<OpResult>() &&
-        yieldedValue.getDefiningOp()->getParentOp() != padOp.getOperation();
-    bool invariantYieldedValue = outsideBbArg || outsideOpResult;
-    if (matchPattern(yieldedValue, m_Constant(&constYieldedValue))) {
-      // Padding with a constant: Create linalg.fill.
-      Dialect *arithDialect =
-          rewriter.getContext()->getLoadedDialect<arith::ArithDialect>();
-      Value fillValue = arithDialect
-                            ->materializeConstant(rewriter, constYieldedValue,
-                                                  yieldedValue.getType(),
-                                                  yieldedValue.getLoc())
-                            ->getResult(0);
-      auto fillOp = rewriter.create<linalg::FillOp>(
-          loc, ValueRange(fillValue), ValueRange(emptyOp.getResult()));
-      rewriter.setInsertionPointAfter(fillOp);
-      filled = fillOp.getResult(0);
-    } else if (invariantYieldedValue) {
-      // Padding with an invariant value.
-      auto fillOp = rewriter.create<linalg::FillOp>(
-          loc, ValueRange(yieldedValue), ValueRange(emptyOp.getResult()));
-      rewriter.setInsertionPointAfter(fillOp);
-      filled = fillOp.getResult(0);
-    } else {
-      // Create linalg.generic.
-      SmallVector<utils::IteratorType> iteratorTypes(
-          resultType.getRank(), utils::IteratorType::parallel);
-      SmallVector<AffineMap> indexingMaps(
-          1, rewriter.getMultiDimIdentityMap(resultType.getRank()));
-      auto genericOp = rewriter.create<linalg::GenericOp>(
-          loc, resultType, /*inputs=*/ValueRange(),
-          /*outputs=*/ValueRange{emptyOp.getResult()}, /*indexingMaps=*/
-          indexingMaps, iteratorTypes);
-      Block *body = rewriter.createBlock(&genericOp->getRegion(0), {},
-                                         resultType.getElementType(), loc);
-      rewriter.setInsertionPointToStart(body);
-      SmallVector<Value> bbArgReplacements;
-      for (int64_t i = 0; i < resultType.getRank(); ++i)
-        bbArgReplacements.push_back(rewriter.create<linalg::IndexOp>(loc, i));
-      rewriter.mergeBlocks(padOp.getBody(), body, bbArgReplacements);
+    Value fillValue =
+        arithDialect
+                                  yieldedValue.getType(), yieldedValue.getLoc())
+    auto fillOp = rewriter.create<linalg::FillOp>(loc, ValueRange(fillValue),
+                                                  ValueRange(dest));
+    return fillOp;
+  }
+
+  if (invariantYieldedValue) {
+    auto fillOp = rewriter.create<linalg::FillOp>(loc, ValueRange(yieldedValue),
+                                                  ValueRange(dest));
+    return fillOp;
+  }
+
+  SmallVector<utils::IteratorType> iteratorTypes(resultType.getRank(),
+                                                 utils::IteratorType::parallel);
+      /*outputs=*/ValueRange{dest}, /*indexingMaps=*/
 
-      // Update terminator.
-      auto yieldOp = cast<tensor::YieldOp>(body->getTerminator());
-      rewriter.replaceOpWithNewOp<linalg::YieldOp>(yieldOp, yieldOp.getValue());
-      rewriter.setInsertionPointAfter(genericOp);
-      filled = genericOp->getResult(0);
-    }
+  return genericOp;
 
+Value linalg::bufferizePadOp(RewriterBase &rewriter, PadOp padOp,
+                             Attribute memorySpace) {
+  OpBuilder::InsertionGuard g(rewriter);
+  Location loc = padOp.getLoc();
+  RankedTensorType resultType = padOp.getResultType();
+
+  // Create buffer allocation.
+  auto memrefType = bufferization::getMemRefTypeWithStaticIdentityLayout(
+                        resultType, memorySpace)
+                        .cast<MemRefType>();
+  ReifiedRankedShapedTypeDims reifiedShape;
+  if (failed(cast<ReifyRankedShapedTypeOpInterface>(padOp.getOperation())
+                 .reifyResultShapes(rewriter, reifiedShape)))
+    llvm_unreachable("failed to reify tensor.pad op result shape");
+  SmallVector<Value> dynamicSizes;
+  for (int64_t i = 0; i < resultType.getRank(); ++i)
+    if (resultType.isDynamicDim(i))
+      dynamicSizes.push_back(reifiedShape[0][i]);
+  Value alloc = rewriter.create<memref::AllocOp>(loc, memrefType, dynamicSizes);
+
+  // Create linalg.fill or linalg.generic.
+  Operation *fillOp = movePaddingToFillOrGenericOp(rewriter, loc, padOp, alloc);
+  rewriter.setInsertionPointAfter(fillOp);
+
+  // Create memref.tensor_store.
+  SmallVector<OpFoldResult> sizes =
+      getMixedSizes(rewriter, loc, padOp.getSource());
+  SmallVector<OpFoldResult> strides(resultType.getRank(),
+                                    rewriter.getIndexAttr(1));
+  Value subview = rewriter.create<memref::SubViewOp>(
+      loc, alloc, /*offsets=*/padOp.getMixedLowPad(), sizes, strides);
+  rewriter.create<memref::TensorStoreOp>(loc, padOp.getSource(), subview);
+
+  // Create bufferization.to_tensor with "restrict" and "writable". The returned
+  // tensor is a new buffer allocation, so it does not alias with any buffer.
+  Value toTensorOp = rewriter.create<bufferization::ToTensorOp>(
+      loc, alloc, /*restrict=*/true, /*writable=*/true);
+  rewriter.replaceOp(padOp, toTensorOp);
+  return toTensorOp;
+}
+
+namespace {
+/// Lower tensor.pad to linalg.generic + tensor.insert_slice.
+struct PadOpConverter : public OpRewritePattern<PadOp> {
+  using OpRewritePattern<PadOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(PadOp padOp,
+                                PatternRewriter &rewriter) const override {
+    // Only ops with exactly one block are supported.
+    if (!padOp.getBodyRegion().hasOneBlock())
+      return failure();
+
+    // Create tensor.empty.
+    Location loc = padOp.getLoc();
+    RankedTensorType resultType = padOp.getResultType();
+    ReifiedRankedShapedTypeDims reifiedShape;
+    if (failed(cast<ReifyRankedShapedTypeOpInterface>(padOp.getOperation())
+                   .reifyResultShapes(rewriter, reifiedShape)))
+      return rewriter.notifyMatchFailure(
+          padOp, "failed to reify tensor.pad op result shape");
+    SmallVector<Value> dynamicSizes;
+    for (int64_t i = 0; i < resultType.getRank(); ++i)
+      if (resultType.isDynamicDim(i))
+        dynamicSizes.push_back(reifiedShape[0][i]);
+    auto emptyOp = rewriter.create<EmptyOp>(loc, resultType, dynamicSizes);
+
+    // Create linalg.fill or linalg.generic.
+    Operation *fillOp =
+        movePaddingToFillOrGenericOp(rewriter, loc, padOp, emptyOp.getResult());
+    rewriter.setInsertionPointAfter(fillOp);
+
     // Create tensor::InsertSliceOp.
     SmallVector<OpFoldResult> sliceSizes =
         getMixedSizes(rewriter, loc, padOp.getSource());
     SmallVector<OpFoldResult> sliceStrides(resultType.getRank(),
                                            rewriter.getIndexAttr(1));
     rewriter.replaceOpWithNewOp<tensor::InsertSliceOp>(
-        padOp, padOp.getSource(), filled,
+        padOp, padOp.getSource(), fillOp->getResult(0),
         /*offsets=*/padOp.getMixedLowPad(), sliceSizes, sliceStrides);
 
     return success();
   }
 };
 
+struct BufferizePadOp : public OpRewritePattern<PadOp> {
+  BufferizePadOp(MLIRContext *ctx, Attribute memorySpace = {},
+                 PatternBenefit benefit = 1)
+      : OpRewritePattern<PadOp>(ctx, benefit), memorySpace(memorySpace) {}
+
+  LogicalResult matchAndRewrite(PadOp padOp,
+                                PatternRewriter &rewriter) const override {
+    // Only ops with exactly one block are supported.
+    if (!padOp.getBodyRegion().hasOneBlock())
+      return failure();
+    (void)linalg::bufferizePadOp(rewriter, padOp, memorySpace);
+    return success();
+  }
+
+private:
+  const Attribute memorySpace;
+};
 } // namespace
 
+void linalg::populateBufferizeNonDpsOpsPatterns(RewritePatternSet &patterns,
+                                                Attribute memorySpace) {
+  patterns.insert<BufferizePadOp>(patterns.getContext(), memorySpace);
+}
+
 void linalg::populateConvertToDestinationStylePatterns(
     RewritePatternSet &patterns) {
   patterns.insert<FromElementsOpConverter, GenerateOpConverter, PadOpConverter>(
       patterns.getContext());
 }

mlir/test/Dialect/Linalg/bufferize-non-dps-ops.mlir

+// RUN: mlir-opt -split-input-file \
+// RUN:   -test-linalg-transform-patterns=test-bufferize-non-dps-ops-patterns \
+// RUN:   -canonicalize %s | FileCheck %s
+
+// CHECK:       #[[$map:.+]] = affine_map<()[s0, s1] -> (s0 + s1 + 5)>
+// CHECK:       #[[$map1:.+]] = affine_map<()[s0, s1] -> (s0 + s1 + 10)>
+// CHECK-LABEL: func @tensor_pad_constant(
+//  CHECK-SAME:   %[[t:.*]]: tensor<?x10xindex>, %[[l2:.*]]: index, %[[h1:.*]]: index, %[[h2:.*]]: index
+//   CHECK-DAG:   %[[c0:.*]] = arith.constant 0 : index
+//   CHECK-DAG:   %[[c50:.*]] = arith.constant 50 : index
+//   CHECK-DAG:   %[[dim0:.*]] = tensor.dim %[[t]], %[[c0]]
+//   CHECK-DAG:   %[[size0:.*]] = affine.apply #[[$map]]()[%[[h1]], %[[dim0]]]
+//   CHECK-DAG:   %[[size1:.*]] = affine.apply #[[$map1]]()[%[[l2]], %[[h2]]]
+//       CHECK:   %[[alloc:.*]] = memref.alloc(%[[size0]], %[[size1]]) : memref<?x?xindex>
+//       CHECK:   linalg.fill ins(%[[c50]] : index) outs(%[[alloc]] : memref<?x?xindex>)
+//       CHECK:   %[[dim0:.*]] = tensor.dim %[[t]], %[[c0]]
+//       CHECK:   %[[subview:.*]] = memref.subview %[[alloc]][5, %[[l2]]] [%[[dim0]], 10] [1, 1]
+//       CHECK:   memref.tensor_store %[[t]], %[[subview]]
+//       CHECK:   %[[r:.*]] = bufferization.to_tensor %[[alloc]] restrict writable : memref<?x?xindex>
+//       CHECK:   return %[[r]]
+func.func @tensor_pad_constant(%t: tensor<?x10xindex>, %l2: index, %h1: index,
+                               %h2: index) -> tensor<?x?xindex> {
+  %0 = tensor.pad %t low[5, %l2] high[%h1, %h2] {
+  ^bb0(%arg0: index, %arg1: index):
+    %c = arith.constant 50 : index
+    tensor.yield %c : index
+  } : tensor<?x10xindex> to tensor<?x?xindex>
+  return %0 : tensor<?x?xindex>
+}

mlir/test/lib/Dialect/Linalg/TestLinalgTransforms.cpp

struct TestLinalgTransforms
   Option<bool> testEraseUnnecessaryInputs{
       *this, "test-erase-unnecessary-inputs",
       llvm::cl::desc("Test patterns to erase unnecessary inputs"),
       llvm::cl::init(false)};
   Option<bool> testConvertToDestinationStylePatterns{
       *this, "test-convert-to-destination-style-patterns",
       llvm::cl::desc("Test patterns that convert ops to destination style"),
       llvm::cl::init(false)};
+  Option<bool> testBufferizeNonDpsOpsPatterns{
+      *this, "test-bufferize-non-dps-ops-patterns",
+      llvm::cl::desc("Test patterns that bufferize non-DPS ops"),
+      llvm::cl::init(false)};
 };
 } // namespace
 
 static void applyPatterns(func::FuncOp funcOp) {
   MLIRContext *ctx = funcOp.getContext();
   RewritePatternSet patterns(ctx);
 
   //===--------------------------------------------------------------------===//

 }
 
 static void applyConvertToDestinationStylePatterns(Operation *rootOp) {
   RewritePatternSet patterns(rootOp->getContext());
   populateConvertToDestinationStylePatterns(patterns);
   (void)applyPatternsAndFoldGreedily(rootOp, std::move(patterns));
 }
 
+static void applyBufferizeNonDpsOpsPatterns(Operation *rootOp) {
+  RewritePatternSet patterns(rootOp->getContext());
+  populateBufferizeNonDpsOpsPatterns(patterns);
+  (void)applyPatternsAndFoldGreedily(rootOp, std::move(patterns));
+}
+
 /// Apply transformations specified as patterns.
 void TestLinalgTransforms::runOnOperation() {
   if (testPatterns)
     return applyPatterns(getOperation());
   if (testVectorTransferForwardingPatterns)
     return applyVectorTransferForwardingPatterns(getOperation());
   if (testGenericToVectorPattern)
     return applyLinalgToVectorPatterns(getOperation());
void TestLinalgTransforms::runOnOperation() {
   if (testSwapExtractSliceWithFill)
     return applySwapExtractSliceWithFillPattern(getOperation());
   if (testEraseUnusedOperandsAndResults)
     return applyEraseUnusedOperandsAndResultsPatterns(getOperation());
   if (testEraseUnnecessaryInputs)
     return applyEraseUnnecessaryInputs(getOperation());
   if (testConvertToDestinationStylePatterns)
     applyConvertToDestinationStylePatterns(getOperation());
+  if (testBufferizeNonDpsOpsPatterns)
+    applyBufferizeNonDpsOpsPatterns(getOperation());
 }
 
 namespace mlir {
 namespace test {
 void registerTestLinalgTransforms() {
   PassRegistration<TestLinalgTransforms>();
 }
 } // namespace test
 } // namespace mlir