Tensor/IR/TensorOps.cpp

//===----------------------------------------------------------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//

#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/IR/BlockAndValueMapping.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/IR/TypeUtilities.h"
#include "llvm/ADT/STLExtras.h"

using namespace mlir;
using namespace mlir::tensor;

//===----------------------------------------------------------------------===//
// CastOp
//===----------------------------------------------------------------------===//

/// Determines whether tensor::CastOp casts to a more dynamic version of the
/// source tensor. This is useful to fold a tensor.cast into a consuming op and
/// implement canonicalization patterns for ops in different dialects that may
/// consume the results of tensor.cast operations. Such foldable tensor.cast
/// operations are typically inserted as `subtensor` ops and are canonicalized,
/// to preserve the type compatibility of their uses.
///
/// Returns true when all conditions are met:
/// 1. source and result are ranked tensors with same element type and rank.
/// 2. the tensor type has more static information than the result
///
/// Example:
/// ```mlir
///   %1 = tensor.cast %0 : tensor<8x16xf32> to tensor<?x?xf32>
///   %2 = consumer %1 ... : tensor<?x?xf32> ...
/// ```
///
/// folds into:
///
/// ```mlir
///   %2 = consumer %0 ... : tensor<8x16xf32> ...
/// ```
bool mlir::tensor::canFoldIntoConsumerOp(CastOp castOp) {
  if (!castOp)
    return false;

  RankedTensorType sourceType =
      castOp.source().getType().dyn_cast<RankedTensorType>();
  RankedTensorType resultType = castOp.getType().dyn_cast<RankedTensorType>();

  // Requires RankedTensorType.
  if (!sourceType || !resultType)
    return false;

  // Requires same elemental type.
  if (sourceType.getElementType() != resultType.getElementType())
    return false;

  // Requires same rank.
  if (sourceType.getRank() != resultType.getRank())
    return false;

  // If cast is towards more static sizes along any dimension, don't fold.
  for (auto t : llvm::zip(sourceType.getShape(), resultType.getShape())) {
    if (ShapedType::isDynamic(std::get<0>(t)) &&
        !ShapedType::isDynamic(std::get<1>(t)))
      return false;
  }

  return true;
}

bool CastOp::areCastCompatible(TypeRange inputs, TypeRange outputs) {
  if (inputs.size() != 1 || outputs.size() != 1)
    return false;
  Type a = inputs.front(), b = outputs.front();
  auto aT = a.dyn_cast<TensorType>();
  auto bT = b.dyn_cast<TensorType>();
  if (!aT || !bT)
    return false;

  if (aT.getElementType() != bT.getElementType())
    return false;

  return succeeded(verifyCompatibleShape(aT, bT));
}

/// Compute a TensorType that has the joined shape knowledge of the two
/// given TensorTypes. The element types need to match.
static TensorType joinShapes(TensorType one, TensorType two) {
  assert(one.getElementType() == two.getElementType());

  if (!one.hasRank())
    return two;
  if (!two.hasRank())
    return one;

  int64_t rank = one.getRank();
  if (rank != two.getRank())
    return {};

  SmallVector<int64_t, 4> join;
  join.reserve(rank);
  for (int64_t i = 0; i < rank; ++i) {
    if (one.isDynamicDim(i)) {
      join.push_back(two.getDimSize(i));
      continue;
    }
    if (two.isDynamicDim(i)) {
      join.push_back(one.getDimSize(i));
      continue;
    }
    if (one.getDimSize(i) != two.getDimSize(i))
      return {};
    join.push_back(one.getDimSize(i));
  }
  return RankedTensorType::get(join, one.getElementType());
}

namespace {

/// Replaces chains of two tensor.cast operations by a single tensor.cast
/// operation if doing so does not remove runtime constraints.
struct ChainedTensorCast : public OpRewritePattern<CastOp> {
  using OpRewritePattern<CastOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(CastOp tensorCast,
                                PatternRewriter &rewriter) const final {
    auto tensorCastOperand = tensorCast.getOperand().getDefiningOp<CastOp>();

    if (!tensorCastOperand)
      return failure();

    auto sourceType =
        tensorCastOperand.getOperand().getType().cast<TensorType>();
    auto intermediateType = tensorCastOperand.getType().cast<TensorType>();
    auto resultType = tensorCast.getType().cast<TensorType>();

    // We can remove the intermediate cast if joining all three produces the
    // same result as just joining the source and result shapes.
    auto firstJoin =
        joinShapes(joinShapes(sourceType, intermediateType), resultType);

    // The join might not exist if the cast sequence would fail at runtime.
    if (!firstJoin)
      return failure();

    // The newJoin always exists if the above join exists, it might just contain
    // less information. If so, we cannot drop the intermediate cast, as doing
    // so would remove runtime checks.
    auto newJoin = joinShapes(sourceType, resultType);
    if (firstJoin != newJoin)
      return failure();

    rewriter.replaceOpWithNewOp<CastOp>(tensorCast, resultType,
                                        tensorCastOperand.getOperand());
    return success();
  }
};

} // namespace

void CastOp::getCanonicalizationPatterns(OwningRewritePatternList &results,
                                         MLIRContext *context) {
  results.insert<ChainedTensorCast>(context);
}

//===----------------------------------------------------------------------===//
// ExtractOp
//===----------------------------------------------------------------------===//

static LogicalResult verify(ExtractOp op) {
  // Verify the # indices match if we have a ranked type.
  if (auto tensorType = op.tensor().getType().dyn_cast<RankedTensorType>())
    if (tensorType.getRank() != static_cast<int64_t>(op.indices().size()))
      return op.emitOpError("incorrect number of indices for extract_element");

  return success();
}

OpFoldResult ExtractOp::fold(ArrayRef<Attribute> operands) {
  // The tensor operand must be a known constant.
  Attribute tensor = operands.front();
  if (!tensor)
    return {};
  // If this is a splat elements attribute, simply return the value. All of the
  // elements of a splat attribute are the same.
  if (auto splatTensor = tensor.dyn_cast<SplatElementsAttr>())
    return splatTensor.getSplatValue();

  // Otherwise, collect the constant indices into the tensor.
  SmallVector<uint64_t, 8> indices;
  for (Attribute indice : llvm::drop_begin(operands, 1)) {
    if (!indice || !indice.isa<IntegerAttr>())
      return {};
    indices.push_back(indice.cast<IntegerAttr>().getInt());
  }

  // If this is an elements attribute, query the value at the given indices.
  auto elementsAttr = tensor.dyn_cast<ElementsAttr>();
  if (elementsAttr && elementsAttr.isValidIndex(indices))
    return elementsAttr.getValue(indices);
  return {};
}

//===----------------------------------------------------------------------===//
// FromElementsOp
//===----------------------------------------------------------------------===//

void FromElementsOp::build(OpBuilder &builder, OperationState &result,
                           Type elementType, ValueRange elements) {
  Type resultTy = RankedTensorType::get({static_cast<int64_t>(elements.size())},
                                        elementType);
  result.addOperands(elements);
  result.addTypes(resultTy);
}

void FromElementsOp::build(OpBuilder &builder, OperationState &result,
                           ValueRange elements) {
  assert(!elements.empty() && "expected at least one element");
  build(builder, result, elements.front().getType(), elements);
}

namespace {

// Canonicalizes the pattern of the form
//
// %tensor = tensor.from_elements(%element) : (i32) -> tensor<1xi32>
// %extracted_element = tensor.extract %tensor[%c0] : tensor<1xi32>
//
// to just %element.
struct ExtractElementFromTensorFromElements
    : public OpRewritePattern<tensor::ExtractOp> {
  using OpRewritePattern<tensor::ExtractOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(tensor::ExtractOp extract,
                                PatternRewriter &rewriter) const final {
    if (extract.indices().size() != 1)
      return failure();

    auto tensorFromElements = extract.tensor().getDefiningOp<FromElementsOp>();
    if (tensorFromElements == nullptr)
      return failure();

    APInt index;
    if (!matchPattern(*extract.indices().begin(), m_ConstantInt(&index)))
      return failure();
    // Prevent out of bounds accesses. This can happen in invalid code that will
    // never execute.
    if (tensorFromElements->getNumOperands() <= index.getZExtValue() ||
        index.getSExtValue() < 0)
      return failure();
    rewriter.replaceOp(extract,
                       tensorFromElements.getOperand(index.getZExtValue()));
    return success();
  }
};

} // namespace

void FromElementsOp::getCanonicalizationPatterns(
    OwningRewritePatternList &results, MLIRContext *context) {
  results.insert<ExtractElementFromTensorFromElements>(context);
}

//===----------------------------------------------------------------------===//
// GenerateOp
//===----------------------------------------------------------------------===//

static LogicalResult verify(GenerateOp op) {
  // Ensure that the tensor type has as many dynamic dimensions as are specified
  // by the operands.
  RankedTensorType resultTy = op.getType().cast<RankedTensorType>();
  if (op.getNumOperands() != resultTy.getNumDynamicDims())
    return op.emitError("must have as many index operands as dynamic extents "
                        "in the result type");

  // Ensure that region arguments span the index space.
  if (!llvm::all_of(op.body().getArgumentTypes(),
                    [](Type ty) { return ty.isIndex(); }))
    return op.emitError("all body arguments must be index");
  if (op.body().getNumArguments() != resultTy.getRank())
    return op.emitError("must have one body argument per input dimension");

  // Ensure that the region yields an element of the right type.
  auto yieldOp =
      llvm::cast<YieldOp>(op.body().getBlocks().front().getTerminator());
  if (yieldOp.value().getType() != resultTy.getElementType())
    return op.emitOpError(
        "body must be terminated with a `yield` operation of the tensor "
        "element type");

  return success();
}

void GenerateOp::build(
    OpBuilder &b, OperationState &result, Type resultTy,
    ValueRange dynamicExtents,
    function_ref<void(OpBuilder &, Location, ValueRange)> bodyBuilder) {
  build(b, result, resultTy, dynamicExtents);

  // Build and populate body.
  OpBuilder::InsertionGuard guard(b);
  Region *bodyRegion = result.regions.front().get();
  auto rank = resultTy.cast<RankedTensorType>().getRank();
  SmallVector<Type, 2> argumentTypes(rank, b.getIndexType());
  Block *bodyBlock =
      b.createBlock(bodyRegion, bodyRegion->end(), argumentTypes);
  bodyBuilder(b, result.location, bodyBlock->getArguments());
}

namespace {

/// Canonicalizes tensor.generate operations with a constant
/// operand into the equivalent operation with the operand expressed in the
/// result type, instead. We also insert a type cast to make sure that the
/// resulting IR is still well-typed.
struct StaticTensorGenerate : public OpRewritePattern<GenerateOp> {
  using OpRewritePattern<GenerateOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(GenerateOp tensorFromElements,
                                PatternRewriter &rewriter) const final {
    auto resultType =
        tensorFromElements.getResult().getType().cast<RankedTensorType>();

    if (resultType.hasStaticShape())
      return failure();

    SmallVector<Value, 4> newOperands;
    SmallVector<int64_t, 4> newShape;
    auto operandsIt = tensorFromElements.dynamicExtents().begin();

    for (int64_t dim : resultType.getShape()) {
      if (dim != RankedTensorType::kDynamicSize) {
        newShape.push_back(dim);
        continue;
      }
      APInt index;
      if (!matchPattern(*operandsIt, m_ConstantInt(&index))) {
        newShape.push_back(RankedTensorType::kDynamicSize);
        newOperands.push_back(*operandsIt++);
        continue;
      }
      newShape.push_back(index.getSExtValue());
      operandsIt++;
    }

    if (newOperands.size() == tensorFromElements.dynamicExtents().size())
      return failure();

    auto loc = tensorFromElements.getLoc();
    auto newOp = rewriter.create<GenerateOp>(
        loc, RankedTensorType::get(newShape, resultType.getElementType()),
        newOperands);
    rewriter.inlineRegionBefore(tensorFromElements.body(), newOp.body(),
                                newOp.body().begin());
    rewriter.replaceOpWithNewOp<tensor::CastOp>(tensorFromElements, resultType,
                                                newOp);
    return success();
  }
};

/// Canonicalizes the pattern of the form
///
/// %tensor = tensor.generate %x {
///   ^bb0(%arg0: index):  // no predecessors
///   <computation>
///   yield %1 : index
/// } : tensor<?xindex>
/// %extracted_element = tensor.extract %tensor[%c0] : tensor<?xi32>
///
/// to just <computation> with %arg0 replaced by %c0. We only do this if the
/// tensor.generate operation has no side-effects.
struct ExtractFromTensorGenerate : public OpRewritePattern<tensor::ExtractOp> {
  using OpRewritePattern<tensor::ExtractOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(tensor::ExtractOp extract,
                                PatternRewriter &rewriter) const final {
    auto tensorFromElements = extract.tensor().getDefiningOp<GenerateOp>();
    if (!tensorFromElements || !wouldOpBeTriviallyDead(tensorFromElements))
      return failure();

    BlockAndValueMapping mapping;
    Block *body = tensorFromElements.getBody();
    mapping.map(body->getArguments(), extract.indices());
    for (auto &op : body->without_terminator())
      rewriter.clone(op, mapping);

    auto yield = cast<YieldOp>(body->getTerminator());

    rewriter.replaceOp(extract, mapping.lookupOrDefault(yield.value()));
    return success();
  }
};

/// Canonicalizes the pattern of the form
///
/// %val = tensor.cast %source : : tensor<?xi32> to tensor<2xi32>
/// %extracted_element = tensor.extract %val[%c0] : tensor<2xi32>
///
/// to
///
/// %extracted_element = tensor.extract %source[%c0] : tensor<?xi32>
struct ExtractFromTensorCast : public OpRewritePattern<tensor::ExtractOp> {
  using OpRewritePattern<tensor::ExtractOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(tensor::ExtractOp extract,
                                PatternRewriter &rewriter) const final {
    auto tensorCast = extract.tensor().getDefiningOp<tensor::CastOp>();
    if (!tensorCast)
      return failure();

    rewriter.replaceOpWithNewOp<tensor::ExtractOp>(extract, tensorCast.source(),
                                                   extract.indices());
    return success();
  }
};

} // namespace

void GenerateOp::getCanonicalizationPatterns(OwningRewritePatternList &results,
                                             MLIRContext *context) {
  // TODO: Move extract patterns to tensor::ExtractOp.
  results.insert<ExtractFromTensorGenerate, ExtractFromTensorCast,
                 StaticTensorGenerate>(context);
}

//===----------------------------------------------------------------------===//
// TableGen'd op method definitions
//===----------------------------------------------------------------------===//

#define GET_OP_CLASSES
#include "mlir/Dialect/Tensor/IR/TensorOps.cpp.inc"