torch-mlir/lib/Conversion/TorchToMhlo/ViewLikeOps.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
// Also available under a BSD-style license. See LICENSE.
//
//===----------------------------------------------------------------------===//

#include "torch-mlir/Conversion/TorchToMhlo/TorchToMhlo.h"

#include "../PassDetail.h"
#include "./PopulatePatterns.h"
#include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir-hlo/Dialect/mhlo/IR/hlo_ops.h"
#include "torch-mlir/Conversion/Utils/Utils.h"
#include "torch-mlir/Dialect/Torch/IR/TorchDialect.h"
#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
#include "torch-mlir/Dialect/Torch/Utils/TorchUpstream.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include "torch-mlir/Dialect/TorchConversion/IR/TorchConversionDialect.h"
#include "torch-mlir/Dialect/TorchConversion/IR/TorchConversionOps.h"
#include <numeric>

using namespace mlir;
using namespace mlir::torch;
using namespace mlir::torch::Torch;
using namespace mlir::torch::TorchConversion;

#ifdef TORCH_MLIR_ENABLE_MHLO_TRUNC_DIMSIZE_TO_I32
static constexpr size_t kMhloDimSizeBits = 32;
#else
static constexpr size_t kMhloDimSizeBits = 64;
#endif

namespace {

SmallVector<size_t> toPositiveDims(ArrayRef<int64_t> dims, int64_t rank) {
  SmallVector<size_t> posDims;
  posDims.reserve(rank);
  std::transform(
      dims.begin(), dims.end(), std::back_inserter(posDims),
      [rank](int64_t d) -> size_t { return toPositiveDim(d, rank); });
  return posDims;
}

FailureOr<SmallVector<Value, 4>>
getDimSizesOfTensor(PatternRewriter &rewriter, Operation *op, Value value,
                    ArrayRef<int64_t> inpDims) {
  auto valueTy = value.getType().dyn_cast<RankedTensorType>();
  if (!valueTy) {
    return rewriter.notifyMatchFailure(
        op, "getDimSizesOfTensor(): the input is not a ranked tensor");
  }

  auto rank = valueTy.getRank();
  auto dims = toPositiveDims(inpDims, rank);
  SmallVector<Value, 4> dimSizes;
  dimSizes.reserve(dims.size());

  auto loc = op->getLoc();
  for (auto d : dims) {
    dimSizes.emplace_back(rewriter.create<arith::IndexCastOp>(
        loc, rewriter.getIntegerType(kMhloDimSizeBits),
        rewriter.create<tensor::DimOp>(loc, value, d)));
  }
  return dimSizes;
}

FailureOr<SmallVector<Value, 4>>
getDimSizesOfTensor(PatternRewriter &rewriter, Operation *op, Value value) {
  auto valueTy = value.getType().dyn_cast<RankedTensorType>();
  if (!valueTy) {
    return rewriter.notifyMatchFailure(
        op, "getDimSizesOfTensor(): the input is not a ranked tensor");
  }

  auto rank = valueTy.getRank();
  // Get int vector [0, 1, ..., rank-1]
  std::vector<int64_t> dims(rank);
  std::iota(dims.begin(), dims.end(), 0);
  return getDimSizesOfTensor(rewriter, op, value, dims);
}

// A dimension index from torch.dialect might outside the range [0, dimSize].
// The function is used to normalize the input index into the range.
Value getNormalizedDimSizeInternal(PatternRewriter &rewriter, Operation *op,
                                   Value index, Value dimSize) {
  auto loc = op->getLoc();
  Value zero = rewriter.create<arith::ConstantOp>(
      loc, rewriter.getIntegerAttr(rewriter.getI64Type(), 0));

  // To normalize index into range [-dimSize, dimSize]
  // index = min(max(-dimSize, index), dimSize)
  auto negDimSize = rewriter.create<arith::SubIOp>(loc, zero, dimSize);
  index = rewriter.create<arith::MaxSIOp>(loc, negDimSize, index);
  index = rewriter.create<arith::MinSIOp>(loc, dimSize, index);

  auto dimSizePlusIndex = rewriter.create<arith::AddIOp>(loc, dimSize, index);
  auto indexPositive = rewriter.create<arith::CmpIOp>(
      loc, arith::CmpIPredicate::sge, index, zero);
  // get positive index: (index >=0) ? index: index + dimSize
  return rewriter.create<arith::SelectOp>(loc, indexPositive, index,
                                          dimSizePlusIndex);
}

Value getDynamicSliceInternal(PatternRewriter &rewriter, Operation *op,
                              Value input, Value startIndex, Value endIndex,
                              Value step, size_t dimIndex,
                              ArrayRef<Value> dimSizes) {
  auto loc = op->getLoc();
  // startIndex & endIndex has been normailized into range [0, dSize]
  Type intType = rewriter.getIntegerType(kMhloDimSizeBits);
  Value zero = rewriter.create<arith::ConstantOp>(
      loc, rewriter.getIntegerAttr(intType, 0));
  Value one = rewriter.create<arith::ConstantOp>(
      loc, rewriter.getIntegerAttr(intType, 1));

  SmallVector<Value, 4> startIndices;
  SmallVector<Value, 4> endIndices;
  SmallVector<Value, 4> strides;

  auto inputTy = input.getType().dyn_cast<RankedTensorType>();
  size_t rank = inputTy.getRank();
  startIndices.reserve(rank);
  endIndices.reserve(rank);
  strides.reserve(rank);

  auto endIndexIsZero = rewriter.create<arith::CmpIOp>(
      loc, arith::CmpIPredicate::eq, endIndex, zero);
  endIndex = rewriter.create<arith::SelectOp>(loc, endIndexIsZero,
                                              dimSizes[dimIndex], endIndex);

  for (size_t r = 0; r < rank; ++r) {
    if (r == dimIndex) {
      startIndices.push_back(startIndex);
      endIndices.push_back(endIndex);
      strides.push_back(step);
    } else {
      startIndices.push_back(zero);
      endIndices.push_back(dimSizes[r]);
      strides.push_back(one);
    }
  }

  auto startTensor =
      rewriter.create<tensor::FromElementsOp>(loc, startIndices).getResult();
  auto endTensor =
      rewriter.create<tensor::FromElementsOp>(loc, endIndices).getResult();
  auto stridesTensor =
      rewriter.create<tensor::FromElementsOp>(loc, strides).getResult();

  auto inputShape = inputTy.getShape();
  SmallVector<int64_t, 4> sliceShape(inputShape.begin(), inputShape.end());
  sliceShape[dimIndex] = ShapedType::kDynamicSize;
  auto sliceoutputTy =
      RankedTensorType::get(sliceShape, inputTy.getElementType());
  return rewriter.create<mhlo::RealDynamicSliceOp>(
      loc, sliceoutputTy, input, startTensor, endTensor, stridesTensor);
}

// Get a dynamic slice of the tensor from startIndex to endIndex with stride
// step on the specifed dimension. The input startIndex(default to 0),
// endIndex(default to dimSize), and step(default to 1) can be optional.
FailureOr<Value> getDynamicSlice(PatternRewriter &rewriter, Operation *op,
                                 Value input,
                                 llvm::Optional<Value> startIndexOpt,
                                 llvm::Optional<Value> endIndexOpt,
                                 llvm::Optional<Value> stepOpt, int64_t dim) {
  auto loc = op->getLoc();
  auto inputTy = input.getType().dyn_cast<RankedTensorType>();
  auto rank = inputTy.getRank();

  dim = (dim + rank) % rank;
  Value dimSize = rewriter.create<arith::IndexCastOp>(
      loc, rewriter.getI64Type(),
      rewriter.create<tensor::DimOp>(loc, input, dim));

  Value normStartIndex =
      startIndexOpt
          ? getNormalizedDimSizeInternal(rewriter, op, *startIndexOpt, dimSize)
          : rewriter.create<arith::ConstantOp>(
                loc, rewriter.getIntegerAttr(rewriter.getI64Type(), 0));
  Value normEndIndex =
      endIndexOpt
          ? getNormalizedDimSizeInternal(rewriter, op, *endIndexOpt, dimSize)
          : dimSize;
  Value step =
      stepOpt ? *stepOpt
              : rewriter.create<arith::ConstantOp>(
                    loc, rewriter.getIntegerAttr(rewriter.getI64Type(), 1));

#ifdef TORCH_MLIR_ENABLE_MHLO_TRUNC_DIMSIZE_TO_I32
  auto i32Type = rewriter.getIntegerType(kMhloDimSizeBits);
  normStartIndex =
      rewriter.create<arith::TruncIOp>(loc, i32Type, normStartIndex);
  normEndIndex = rewriter.create<arith::TruncIOp>(loc, i32Type, normEndIndex);
  step = rewriter.create<arith::TruncIOp>(loc, i32Type, step);
#endif
  FailureOr<SmallVector<Value, 4>> dimSizesInfo =
      getDimSizesOfTensor(rewriter, op, input);
  if (failed(dimSizesInfo))
    return rewriter.notifyMatchFailure(
        op, "failed to get dimension sizes of the input");

  auto dimSizes = *dimSizesInfo;
  return getDynamicSliceInternal(rewriter, op, input, normStartIndex,
                                 normEndIndex, step, dim, dimSizes);
}

template <typename AtenOpT>
class ConvertAtenOp : public OpConversionPattern<AtenOpT> {
public:
  using OpConversionPattern<AtenOpT>::OpConversionPattern;
  using OpAdaptor = typename AtenOpT::Adaptor;
  LogicalResult
  matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override;
};

template <>
LogicalResult ConvertAtenOp<AtenSliceTensorOp>::matchAndRewrite(
    AtenSliceTensorOp op, OpAdaptor adaptor,
    ConversionPatternRewriter &rewriter) const {
  auto self = adaptor.self();
  auto selfTy = self.getType().template cast<RankedTensorType>();
  if (!selfTy)
    return op.emitError("only ranked tensor types are supported");
  int64_t dim;
  if (!matchPattern(op.dim(), m_TorchConstantInt(&dim)))
    return rewriter.notifyMatchFailure(
        op, "only constant dim is currently supported");

  auto getOptionalVal = [&](Value val) -> llvm::Optional<Value> {
    if (val.getType().isa<Torch::NoneType>()) {
      return llvm::None;
    } else {
      return val;
    }
  };

  llvm::Optional<Value> start = getOptionalVal(adaptor.start());
  llvm::Optional<Value> end = getOptionalVal(adaptor.end());
  llvm::Optional<Value> step = getOptionalVal(adaptor.step());

  FailureOr<Value> sliceInfo =
      getDynamicSlice(rewriter, op, self, start, end, step, dim);
  if (failed(sliceInfo))
    return op.emitError("can not create a dynmaic slice");

  auto slice = *sliceInfo;
  rewriter.replaceOpWithNewOp<mhlo::ConvertOp>(
      op, getTypeConverter()->convertType(op.getType()), slice);

  return success();
}

// This defines a template to construct ops whose legalizations are
// specialized.
template <typename AtenOpT>
class ConvertAtenViewOp : public OpConversionPattern<AtenOpT> {
 public:
  using OpConversionPattern<AtenOpT>::OpConversionPattern;
  using OpAdaptor = typename AtenOpT::Adaptor;

  LogicalResult matchAndRewrite(
      AtenOpT op,
      OpAdaptor adaptor,
      ConversionPatternRewriter& rewriter) const override {
    auto rankType =
        adaptor.self().getType().template dyn_cast<RankedTensorType>();
    if (!rankType)
      return op.emitError("Only ranked tensor types are currently supported");

    SmallVector<Value, 4> dimSizes;
    if (!getAtenViewOpSizes(op, adaptor, rewriter, dimSizes)) {
      return op.emitError("Dims size must be a list of Scalar");
    }

    auto loc = op.getLoc();
    auto newRank = dimSizes.size();
    if (newRank == 0 || rankType.getRank() == 0) {
      rewriter.replaceOpWithNewOp<mhlo::ReshapeOp>(
          op,
          OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
              op.getType()),
          adaptor.self());
      return success();
    }

    std::for_each(dimSizes.begin(), dimSizes.end(), [&](Value& dSize) {
      dSize = rewriter.create<ToI64Op>(loc, dSize).getResult();
      return dSize;
    });

#ifdef TORCH_MLIR_ENABLE_MHLO_TRUNC_DIMSIZE_TO_I32
    // The i64 calculation is much slower than i32 on some devices, such as Nvidia GPU.
    // One can truncate from i64 to i32 since dimension sizes are unlikely to exceed
    // the range of i32(4GiB)
    std::for_each(dimSizes.begin(), dimSizes.end(), [&](Value& dSize) {
      // dimSize: cast i64 -> i32
      dSize = rewriter.create<arith::TruncIOp>(loc, rewriter.getI32Type(), dSize);
      return dSize;
    });
#endif

    Type intType = rewriter.getIntegerType(kMhloDimSizeBits);
    Value numel = rewriter.create<arith::ConstantOp>(
        loc, rewriter.getIntegerAttr(intType, 1));
    for (auto d : dimSizes) {
      numel = rewriter.create<arith::MulIOp>(loc, numel, d);
    }
    numel = rewriter.create<arith::IndexCastOp>(loc, rewriter.getIndexType(),
                                                numel);

    Value mhloShape = rewriter.create<tensor::FromElementsOp>(loc, dimSizes);
    Value computedShape = rewriter.create<mhlo::ComputeReshapeShapeOp>(
        loc, mhloShape.getType(), numel, mhloShape);
    rewriter.replaceOpWithNewOp<mhlo::DynamicReshapeOp>(
        op,
        OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
            op.getType()),
        adaptor.self(), computedShape);
    return success();
  }

  bool getAtenViewOpSizes(
      AtenOpT op,
      OpAdaptor adaptor,
      ConversionPatternRewriter& rewriter,
      SmallVector<Value, 4>& dimSizes) const;
};

template <>
bool ConvertAtenViewOp<AtenViewOp>::getAtenViewOpSizes(
    AtenViewOp op,
    OpAdaptor adaptor,
    ConversionPatternRewriter& rewriter,
    SmallVector<Value, 4>& dimSizes) const {
  return getListConstructElements(adaptor.size(), dimSizes);
}

template <>
bool ConvertAtenViewOp<AtenReshapeOp>::getAtenViewOpSizes(
    AtenReshapeOp op,
    OpAdaptor adaptor,
    ConversionPatternRewriter& rewriter,
    SmallVector<Value, 4>& dimSizes) const {
  return getListConstructElements(adaptor.shape(), dimSizes);
}

FailureOr<Value> unsqueezeTensor(PatternRewriter &rewriter, Operation *op,
                                 Value tensor,
                                 ArrayRef<int64_t> inputUnsqzDims) {
  // Returns a new tensor with dims of size 1 inserted at the specified
  // position.
  //
  // The position indices (must be high to low dimension number of the returned
  // tensor) are specified with unsqzDims. Indices must be in-order, and in
  // range of tensor rank. Thus, unsqueeze a rank 1 tensor with {0, 2}, {0, 1,
  // 3}, {0, 1, 2} are all valid dimension sets, but {0, 3}, {2} are not.
  auto dimSizesInfo = getDimSizesOfTensor(rewriter, op, tensor);
  if (failed(dimSizesInfo))
    return rewriter.notifyMatchFailure(
        op, "failed to get dimension sizes of the input");

  auto dimSizes = *dimSizesInfo;
  auto rank = dimSizes.size();
  size_t newRank = rank + inputUnsqzDims.size();
  auto unsqzDims = toPositiveDims(inputUnsqzDims, newRank);
  for (size_t k = 0, sz = unsqzDims.size(); k < sz; ++k)
    if (k > 1 && unsqzDims[k] <= unsqzDims[k - 1])
      return rewriter.notifyMatchFailure(
          op, "unsqueeze dimensions must be specified in order");

  auto loc = op->getLoc();
  auto rankTy = tensor.getType().dyn_cast<RankedTensorType>();
  auto oldShape = rankTy.getShape();
  Type intType = rewriter.getIntegerType(kMhloDimSizeBits);
  auto one = rewriter.create<arith::ConstantOp>(
      loc, rewriter.getIntegerAttr(intType, 1));

  std::vector<Value> newDimSizes;
  std::vector<int64_t> newShape;
  newDimSizes.reserve(newRank);
  newShape.reserve(newRank);
  for (size_t k = 0, i = 0, j = 0; k < newRank; ++k) {
    if (j < unsqzDims.size() && unsqzDims[j] == k) {
      newDimSizes.push_back(one);
      newShape.push_back(1);
      j++;
    } else {
      newDimSizes.push_back(dimSizes[i]);
      newShape.push_back(oldShape[i]);
      i++;
    }
  }

  auto outTy = RankedTensorType::get(newShape, rankTy.getElementType());
  auto mhloShape = rewriter.create<tensor::FromElementsOp>(loc, newDimSizes);
  return rewriter.create<mhlo::DynamicReshapeOp>(loc, outTy, tensor, mhloShape)
      .getResult();
}

template <>
LogicalResult ConvertAtenOp<AtenSqueezeOp>::matchAndRewrite(
    AtenSqueezeOp op, OpAdaptor adaptor,
    ConversionPatternRewriter &rewriter) const {
  auto self = adaptor.self();
  auto selfTy = self.getType().template cast<RankedTensorType>();
  if (!selfTy)
    return op.emitError("only ranked tensor types are supported");

  auto rank = selfTy.getRank();
  if (rank == 0)
    return rewriter.notifyMatchFailure(
        op, "The rank of tensor must be greater than 0");

  SmallVector<int64_t, 4> dims;
  dims.reserve(rank);
  for (int r = 0; r < rank; ++r) {
    auto dSize = selfTy.getShape()[r];
    if (dSize == ShapedType::kDynamicSize)
      return rewriter.notifyMatchFailure(
          op, "the size of the dimension being squeezed can't be unknown");
    if (dSize != 1)
      dims.push_back(r);
  }

  auto newDimSizesInfo = getDimSizesOfTensor(rewriter, op, self, dims);
  if (failed(newDimSizesInfo))
    return rewriter.notifyMatchFailure(
        op, "failed to get dimension sizes of the input");
  auto newDimSizes = *newDimSizesInfo;
  auto mhloShape =
      rewriter.create<tensor::FromElementsOp>(op.getLoc(), newDimSizes);
  rewriter.replaceOpWithNewOp<mhlo::DynamicReshapeOp>(
      op, getTypeConverter()->convertType(op.getType()), self, mhloShape);
  return success();
}

template <>
LogicalResult ConvertAtenOp<AtenSqueezeDimOp>::matchAndRewrite(
    AtenSqueezeDimOp op, OpAdaptor adaptor,
    ConversionPatternRewriter &rewriter) const {
  auto self = adaptor.self();
  auto selfTy = self.getType().template cast<RankedTensorType>();
  if (!selfTy)
    return op.emitError("only ranked tensor types are supported");
  int64_t dim;
  if (!matchPattern(op.dim(), m_TorchConstantInt(&dim)))
    return rewriter.notifyMatchFailure(
        op, "only constant dim is currently supported");

  auto rank = selfTy.getRank();
  if (rank == 0)
    return rewriter.notifyMatchFailure(
        op, "the rank of tensor must be greater than 0");

  dim = toPositiveDim(dim, rank);
  if (selfTy.getShape()[dim] != 1) {
    if (selfTy.getShape()[dim] == ShapedType::kDynamicSize)
      return rewriter.notifyMatchFailure(
          op, "the size of the dimension being squeezed is can't be unknown");

    rewriter.replaceOp(op, adaptor.self());
    return success();
  }

  SmallVector<int64_t, 4> dims(rank);
  std::iota(dims.begin(), dims.end(), 0);
  dims.erase(dims.begin() + dim);
  auto newDimSizesInfo = getDimSizesOfTensor(rewriter, op, self, dims);
  if (failed(newDimSizesInfo))
    return rewriter.notifyMatchFailure(
        op, "failed to get dimension sizes of the input");
  auto newDimSizes = *newDimSizesInfo;
  auto mhloShape =
      rewriter.create<tensor::FromElementsOp>(op.getLoc(), newDimSizes);
  rewriter.replaceOpWithNewOp<mhlo::DynamicReshapeOp>(
      op, getTypeConverter()->convertType(op.getType()), self, mhloShape);
  return success();
}

template <>
LogicalResult ConvertAtenOp<AtenUnsqueezeOp>::matchAndRewrite(
    AtenUnsqueezeOp op, OpAdaptor adaptor,
    ConversionPatternRewriter &rewriter) const {
  auto selfType = adaptor.self().getType().dyn_cast<TensorType>();
  if (!selfType) {
    return op.emitError("only tensor types are currently supported");
  }

  int64_t dim;
  if (!matchPattern(op.dim(), m_TorchConstantInt(&dim)))
    return op->emitError("dim must be a Scalar constant");

  auto unsqzTensorInfo = unsqueezeTensor(rewriter, op, adaptor.self(), {dim});
  if (failed(unsqzTensorInfo))
    return rewriter.notifyMatchFailure(op,
                                       "failed to create unsqueezed tensor");

  rewriter.replaceOp(op, *unsqzTensorInfo);
  return success();
}
} // namespace

void mlir::torch::torch_to_mhlo::populateViewLikeOpPatternsAndLegality(
    TypeConverter &typeConverter, RewritePatternSet &patterns,
    ConversionTarget &target) {
  MLIRContext *context = patterns.getContext();

#define INSERT_ATENOP_PATTERN(AtenOp)                                          \
  target.addIllegalOp<AtenOp>();                                               \
  patterns.add<ConvertAtenOp<AtenOp>>(typeConverter, context);
  INSERT_ATENOP_PATTERN(AtenSliceTensorOp);
  INSERT_ATENOP_PATTERN(AtenSqueezeOp);
  INSERT_ATENOP_PATTERN(AtenSqueezeDimOp);
  INSERT_ATENOP_PATTERN(AtenUnsqueezeOp);
#undef INSERT_ATENOP_PATTERN

#define INSERT_VIEW_OP_PATTERN(AtenOp) \
  target.addIllegalOp<AtenOp>();       \
  patterns.add<ConvertAtenViewOp<AtenOp>>(typeConverter, context);
    INSERT_VIEW_OP_PATTERN(AtenViewOp);
    INSERT_VIEW_OP_PATTERN(AtenReshapeOp);
#undef INSERT_VIEW_OP_PATTERN
}