llvm-project/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/Arith/Utils/Utils.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/Dialect/Utils/StaticValueUtils.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/PatternMatch.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/Debug.h"

using namespace mlir;
using namespace mlir::tensor;

// Implements backtracking to traverse indices of the output buffer while
// iterating over op.elements().
static Value createInserts(RewriterBase &rewriter, Location loc, int dim,
                           Value destination, ArrayRef<int64_t> shape,
                           ArrayRef<Value> constants,
                           OperandRange::iterator &elementIt,
                           SmallVectorImpl<Value> &indices) {
  if (dim == static_cast<int>(shape.size()) - 1) {
    for (int i = 0; i < shape.back(); ++i) {
      indices.back() = constants[i];
      destination = rewriter.create<tensor::InsertOp>(loc, *elementIt,
                                                      destination, indices);
      ++elementIt;
    }
    return destination;
  }
  for (int i = 0; i < shape[dim]; ++i) {
    indices[dim] = constants[i];
    destination = createInserts(rewriter, loc, dim + 1, destination, shape,
                                constants, elementIt, indices);
  }
  return destination;
}

static Operation *movePaddingToFillOrGenericOp(RewriterBase &rewriter,
                                               Location loc, PadOp padOp,
                                               Value dest) {
  OpBuilder::InsertionGuard g(rewriter);
  RankedTensorType resultType = padOp.getResultType();

  // Examine the yielded value to decide if a linalg.generic is neede or a
  // linalg.fill is sufficient.
  Value yieldedValue =
      cast<tensor::YieldOp>(padOp.getBody()->getTerminator()).getValue();
  Attribute constYieldedValue;
  // Is the yielded value a bbArg defined outside of the PadOp?
  bool outsideBbArg =
      isa<BlockArgument>(yieldedValue) &&
      cast<BlockArgument>(yieldedValue).getOwner()->getParentOp() !=
          padOp.getOperation();
  // Is the yielded value an OpResult defined outside of the PadOp?
  bool outsideOpResult =
      isa<OpResult>(yieldedValue) &&
      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(dest));
    return fillOp;
  }

  if (invariantYieldedValue) {
    // Padding with an invariant value.
    auto fillOp = rewriter.create<linalg::FillOp>(loc, ValueRange(yieldedValue),
                                                  ValueRange(dest));
    return fillOp;
  }

  // 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{dest}, /*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);

  // Update terminator.
  auto yieldOp = cast<tensor::YieldOp>(body->getTerminator());
  rewriter.replaceOpWithNewOp<linalg::YieldOp>(yieldOp, yieldOp.getValue());
  return genericOp;
}

static SmallVector<Value> reifyOrComputeDynamicSizes(OpBuilder &b,
                                                     Value value) {
  auto tensorType = cast<RankedTensorType>(value.getType());
  if (tensorType.hasStaticShape())
    return {};

  // Try to reify dynamic sizes.
  ReifiedRankedShapedTypeDims reifiedShape;
  if (isa<OpResult>(value) &&
      succeeded(reifyResultShapes(b, value.getDefiningOp(), reifiedShape))) {
    SmallVector<Value> dynSizes;
    for (int64_t i = 0; i < tensorType.getRank(); ++i) {
      if (tensorType.isDynamicDim(i))
        dynSizes.push_back(
            reifiedShape[cast<OpResult>(value).getResultNumber()][i]
                .get<Value>());
    }
    return dynSizes;
  }

  // Create tensor.dim ops.
  SmallVector<Value> dynSizes;
  for (int64_t i = 0; i < tensorType.getRank(); ++i) {
    if (tensorType.isDynamicDim(i))
      dynSizes.push_back(
          b.create<DimOp>(value.getLoc(), value,
                          b.create<arith::ConstantIndexOp>(value.getLoc(), i)));
  }
  return dynSizes;
}

static Value createAllocationForTensor(RewriterBase &rewriter, Location loc,
                                       Value value,
                                       Attribute memorySpace = {}) {
  OpBuilder::InsertionGuard g(rewriter);
  auto tensorType = cast<RankedTensorType>(value.getType());

  // Create buffer allocation.
  auto memrefType =
      cast<MemRefType>(bufferization::getMemRefTypeWithStaticIdentityLayout(
          tensorType, memorySpace));
  SmallVector<Value> dynamicSizes = reifyOrComputeDynamicSizes(rewriter, value);
  Value alloc = rewriter.create<memref::AllocOp>(loc, memrefType, dynamicSizes);

  // Place deallocation at the end of the block.
  rewriter.setInsertionPoint(rewriter.getInsertionBlock()->getTerminator());
  rewriter.create<memref::DeallocOp>(loc, alloc);

  return alloc;
}

Value linalg::bufferizeToAllocation(RewriterBase &rewriter, PadOp padOp,
                                    Attribute memorySpace) {
  OpBuilder::InsertionGuard g(rewriter);
  rewriter.setInsertionPoint(padOp);
  Location loc = padOp.getLoc();

  // Create buffer allocation.
  Value alloc =
      createAllocationForTensor(rewriter, loc, padOp.getResult(), memorySpace);
  rewriter.setInsertionPointAfter(alloc.getDefiningOp());

  // 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(padOp.getResultType().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;
}

/// Lower tensor.from_elements to a sequence of chained tensor.insert.
FailureOr<Operation *> mlir::linalg::rewriteInDestinationPassingStyle(
    RewriterBase &rewriter, tensor::FromElementsOp fromElementsOp) {
  Location loc = fromElementsOp.getLoc();
  RankedTensorType tensorType =
      cast<RankedTensorType>(fromElementsOp.getType());
  auto shape = tensorType.getShape();

  // Create tensor.empty.
  auto emptyOp = rewriter.create<EmptyOp>(loc, tensorType, ValueRange());

  // Case: tensor<elem_type>.
  if (shape.empty()) {
    Operation *res = rewriter.replaceOpWithNewOp<tensor::InsertOp>(
        fromElementsOp, fromElementsOp.getElements().front(),
        emptyOp.getResult(), ValueRange());
    return res;
  }

  // Create constants for the range of possible indices [0, max{shape_i}).
  auto maxDim = *std::max_element(shape.begin(), shape.end());
  SmallVector<Value, 2> constants;
  constants.reserve(maxDim);
  for (int i = 0; i < maxDim; ++i)
    constants.push_back(rewriter.create<arith::ConstantIndexOp>(loc, i));

  // Traverse all elements and create tensor.insert ops.
  auto elementIt = fromElementsOp.getElements().begin();
  SmallVector<Value, 2> indices(tensorType.getRank(), constants[0]);
  Value result = createInserts(rewriter, loc, /*dim=*/0, emptyOp.getResult(),
                               shape, constants, elementIt, indices);

  // Replace tensor.from_elements.
  rewriter.replaceOp(fromElementsOp, result);
  return result.getDefiningOp();
}

/// Lower tensor.generate to linalg.generic.
FailureOr<Operation *>
mlir::linalg::rewriteInDestinationPassingStyle(RewriterBase &rewriter,
                                               tensor::GenerateOp generateOp) {
  // Only ops with exactly one block are supported.
  if (!generateOp.getBody().hasOneBlock())
    return failure();

  Location loc = generateOp.getLoc();
  RankedTensorType tensorType = cast<RankedTensorType>(generateOp.getType());

  // Create tensor.empty.
  auto emptyOp =
      rewriter.create<EmptyOp>(loc, tensorType, generateOp.getDynamicExtents());

  // Create linalg.generic.
  SmallVector<utils::IteratorType> iteratorTypes(tensorType.getRank(),
                                                 utils::IteratorType::parallel);
  SmallVector<AffineMap> indexingMaps(
      1, rewriter.getMultiDimIdentityMap(tensorType.getRank()));
  auto genericOp = rewriter.create<linalg::GenericOp>(
      loc, tensorType, /*inputs=*/ValueRange(),
      /*outputs=*/ValueRange{emptyOp.getResult()}, /*indexingMaps=*/
      indexingMaps, iteratorTypes);
  Block *body = rewriter.createBlock(&genericOp->getRegion(0), {},
                                     tensorType.getElementType(), loc);
  rewriter.setInsertionPointToStart(body);
  SmallVector<Value> bbArgReplacements;
  for (int64_t i = 0; i < tensorType.getRank(); ++i)
    bbArgReplacements.push_back(rewriter.create<linalg::IndexOp>(loc, i));
  rewriter.mergeBlocks(&generateOp.getBody().front(), body, bbArgReplacements);

  // Update terminator.
  auto yieldOp = cast<tensor::YieldOp>(body->getTerminator());
  rewriter.replaceOpWithNewOp<linalg::YieldOp>(yieldOp, yieldOp.getValue());

  // Replace tensor.generate.
  rewriter.replaceOp(generateOp, genericOp->getResult(0));
  return genericOp.getOperation();
}

/// Lower tensor.pad to linalg.generic + tensor.insert_slice.
FailureOr<Operation *>
mlir::linalg::rewriteInDestinationPassingStyle(RewriterBase &rewriter,
                                               tensor::PadOp padOp) {
  // 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(reifyResultShapes(rewriter, padOp, 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].get<Value>());

  // If the `padOp` has a nofold attribute and all paddings are known to be 0,
  // explicitly insert a `linalg.copy`.
  if (padOp.getNofoldAttr() &&
      llvm::all_of(padOp.getMixedLowPad(), isZeroIndex) &&
      llvm::all_of(padOp.getMixedHighPad(), isZeroIndex)) {
    using bufferization::AllocTensorOp;
    Value allocated =
        rewriter.create<AllocTensorOp>(loc, resultType, dynamicSizes);
    auto copyOp = rewriter.replaceOpWithNewOp<linalg::CopyOp>(
        padOp, padOp.getSource(), allocated);
    return copyOp.getOperation();
  }

  Value empty = rewriter.create<EmptyOp>(loc, resultType, dynamicSizes);
  // Create linalg.fill or linalg.generic.
  Operation *fillOp = movePaddingToFillOrGenericOp(rewriter, loc, padOp, empty);
  rewriter.setInsertionPointAfter(fillOp);

  // Create tensor::InsertSliceOp.
  SmallVector<OpFoldResult> sliceSizes =
      getMixedSizes(rewriter, loc, padOp.getSource());
  SmallVector<OpFoldResult> sliceStrides(resultType.getRank(),
                                         rewriter.getIndexAttr(1));
  auto insertSliceOp = rewriter.replaceOpWithNewOp<tensor::InsertSliceOp>(
      padOp, padOp.getSource(), fillOp->getResult(0),
      /*offsets=*/padOp.getMixedLowPad(), sliceSizes, sliceStrides);
  return insertSliceOp.getOperation();
}

Value linalg::bufferizeToAllocation(RewriterBase &rewriter, Value value,
                                    Attribute memorySpace) {
  // Call specialized overload for certain ops.
  if (auto padOp = value.getDefiningOp<PadOp>())
    return bufferizeToAllocation(rewriter, padOp, memorySpace);

  // Collect all uses.
  SmallVector<OpOperand *> uses = llvm::to_vector(
      llvm::map_range(value.getUses(), [](OpOperand &use) { return &use; }));

  OpBuilder::InsertionGuard g(rewriter);
  if (auto bbArg = dyn_cast<BlockArgument>(value)) {
    rewriter.setInsertionPointToStart(bbArg.getOwner());
  } else {
    rewriter.setInsertionPointAfter(value.getDefiningOp());
  }
  Location loc = value.getLoc();

  // Create buffer allocation.
  Value alloc = createAllocationForTensor(rewriter, loc, value, memorySpace);

  // Create memref.tensor_store.
  rewriter.setInsertionPointAfter(alloc.getDefiningOp());
  rewriter.create<memref::TensorStoreOp>(loc, value, alloc);

  // 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);
  for (OpOperand *use : uses) {
    rewriter.updateRootInPlace(use->getOwner(),
                               [&]() { use->set(toTensorOp); });
  }

  return toTensorOp;
}

namespace {

template <typename OpTy>
LogicalResult rewriteOpInDestinationPassingStyle(OpTy op,
                                                 PatternRewriter &rewriter) {
  return linalg::rewriteInDestinationPassingStyle(rewriter, op);
}

} // namespace

void linalg::populateConvertToDestinationStylePatterns(
    RewritePatternSet &patterns) {
  patterns.add(rewriteOpInDestinationPassingStyle<tensor::FromElementsOp>);
  patterns.add(rewriteOpInDestinationPassingStyle<tensor::GenerateOp>);
  patterns.add(rewriteOpInDestinationPassingStyle<tensor::PadOp>);
}