onnx-mlir/src/Dialect/ONNX/ONNXOps.cpp

//===------------------ ONNXOps.cpp - ONNX Operations ---------------------===//
//
// Copyright 2019-2020 The IBM Research Authors.
//
// =============================================================================
//
// This file provides definition of ONNX dialect operations.
//
//===----------------------------------------------------------------------===//

#include "mlir/Dialect/Traits.h"
#include "mlir/IR/Block.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/Function.h"
#include "mlir/IR/IntegerSet.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/Module.h"
#include "mlir/IR/OpImplementation.h"
#include "mlir/IR/PatternMatch.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallBitVector.h"

#include "ONNXOps.hpp"

using namespace mlir;
using namespace mlir::OpTrait::util;

//===----------------------------------------------------------------------===//
// ONNX Helper functions
//===----------------------------------------------------------------------===//

static size_t ArrayAttrSize(ArrayAttr a) { return a.size(); }

static size_t ArrayAttrSize(Optional<ArrayAttr> a) {
  return a.getValue().size();
}

static int64_t ArrayAttrIntVal(ArrayAttr a, int i) {
  return (a.getValue()[i]).cast<IntegerAttr>().getInt();
}

static int64_t ArrayAttrIntVal(Optional<ArrayAttr> a, int i) {
  return (a.getValue().getValue()[i]).cast<IntegerAttr>().getInt();
}

// Returns the ConstantOp which defines an MLIR Value or null.
static mlir::ONNXConstantOp getONNXConstantOp(Value value) {
  return dyn_cast_or_null<mlir::ONNXConstantOp>(value.getDefiningOp());
}

//===----------------------------------------------------------------------===//
// Get reduction type
//===----------------------------------------------------------------------===//
RankedTensorType getReductionOutputType(
    RankedTensorType operandTy, Optional<ArrayAttr> axesAttrs, APInt keepdims) {
  int64_t rank = operandTy.getRank();

  SmallVector<int64_t, 4> axes;
  if (axesAttrs != llvm::None) {
    for (auto axisAttr : axesAttrs.getValue()) {
      int64_t axis = axisAttr.cast<IntegerAttr>().getInt();
      axis = axis >= 0 ? axis : (rank + axis);
      assert(axis >= -rank && axis <= rank - 1);
      if (std::find(axes.begin(), axes.end(), axis) == axes.end())
        axes.emplace_back(axis);
    }
  } else {
    for (decltype(rank) i = 0; i < rank; ++i) {
      axes.emplace_back(i);
    }
  }

  // Mark reduction axes.
  SmallVector<bool, 4> isReductionAxis;
  for (decltype(rank) i = 0; i < rank; ++i) {
    if (std::find(axes.begin(), axes.end(), i) != axes.end())
      isReductionAxis.emplace_back(true);
    else
      isReductionAxis.emplace_back(false);
  }

  // KeepDims
  bool isKeepdims = (keepdims == 1) ? true : false;

  SmallVector<int64_t, 4> dims;
  for (decltype(rank) i = 0; i < rank; ++i) {
    if (isReductionAxis[i]) {
      if (isKeepdims)
        dims.emplace_back(1); // reduction dimension
    } else {
      dims.emplace_back(operandTy.getShape()[i]);
    }
  }

  return RankedTensorType::get(dims, operandTy.getElementType());
}

//===----------------------------------------------------------------------===//
// Support function that computes default values for dilations.
//
template <class T>
static void processConvDilationParam(T *op, Optional<ArrayAttr> kernelShape) {
  auto builder = mlir::Builder(op->getContext());
  auto kernelRank = ArrayAttrSize(kernelShape);

  auto dilationsOpt = op->dilations();
  if (dilationsOpt.hasValue()) {
    if (ArrayAttrSize(dilationsOpt) != kernelRank)
      op->emitError("dialation rank is not the same as the spatial rank");
    // Test values to be greater than 0.
    for (int i = 0; i < kernelRank; ++i) {
      if (ArrayAttrIntVal(dilationsOpt, i) < 1)
        op->emitError("dialation value must be nonzero positive");
    }
  } else {
    // Default dilatation is needed, all dimensions init with 1.
    SmallVector<int64_t, 4> defaultVals(kernelRank, 1);
    // Convert to ArrayRef, then build attribute, then store attribute.
    ArrayRef<int64_t> defaultRefs(defaultVals);
    op->dilationsAttr(builder.getI64ArrayAttr(defaultRefs));
  }
}

//===----------------------------------------------------------------------===//
// Support function that computes default values for strides.
//
template <class T>
static void processConvStrideParam(T *op, Optional<ArrayAttr> kernelShape) {
  auto builder = mlir::Builder(op->getContext());
  auto kernelRank = ArrayAttrSize(kernelShape);

  auto stridesOpt = op->strides();
  if (stridesOpt.hasValue()) {
    if (ArrayAttrSize(stridesOpt) != kernelRank)
      op->emitError("strides rank is not the same as the spatial rank");
    // Check values to be greater than 0.
    for (int i = 0; i < kernelRank; ++i) {
      if (ArrayAttrIntVal(stridesOpt, i) < 1)
        op->emitError("strides value must be nonzero positive");
    }
  } else {
    // Default stride is needed, all dimensions init with 1.
    SmallVector<int64_t, 4> defaultVals(kernelRank, 1);
    // Convert to ArrayRef, then build attribute, then store attribute.
    ArrayRef<int64_t> defaultRefs(defaultVals);
    op->stridesAttr(builder.getI64ArrayAttr(defaultRefs));
  }
}

//===----------------------------------------------------------------------===//
// Support function that computes default values for pads.
//
template <class T>
static void processConvPadParam(T *op,
    ArrayRef<int64_t> inputShape, Optional<ArrayAttr> kernelShape,
    Optional<ArrayAttr> stridesOpt,
    Optional<ArrayAttr> dilationsOpt = llvm::None) {
  auto builder = mlir::Builder(op->getContext());

  auto inputRank = inputShape.size();
  auto kernelRank = ArrayAttrSize(kernelShape);
  auto kernelOffset = inputRank - kernelRank;

  // Try to find padding, getting auto_pad attribute first.
  auto autoPad = op->auto_pad();
  // And then investigate the various different cases. Prefill pad values with
  // zeros, the most common case.
  SmallVector<int64_t, 4> actualPads(2 * kernelRank, 0);
  bool updatedPad = false;
  if (autoPad == "NOTSET") {
    auto padsOpt = op->pads();
    if (padsOpt.hasValue()) {
      // Only option where pads are not updated. Pads consists of two entries
      // for each spatial axis.
      if (ArrayAttrSize(padsOpt) != 2 * kernelRank)
        op->emitError("pads rank is not twice the spatial rank");
      // Check values, pads cannot be negative.
      for (int i = 0; i < 2 * kernelRank; ++i) {
        if (ArrayAttrIntVal(padsOpt, i) < 0)
          op->emitError("pads value must be nonnegative");
      }
    } else {
      // We have notset with no pads, they are assumed to be all zero.
      updatedPad = true;
    }
  } else if (autoPad == "SAME_UPPER" || autoPad == "SAME_LOWER") {
    // Reload dialtion and strides as they may have gotten default values.
    updatedPad = true;
    int64_t dilationVal = 1;
    for (int i = 0; i < kernelRank; ++i) {
      auto inputSize = inputShape[kernelOffset + i];
      auto kernelSize = ArrayAttrIntVal(kernelShape, i);
      if (dilationsOpt.hasValue())
        dilationVal = ArrayAttrIntVal(dilationsOpt, i);
      auto strideVal = ArrayAttrIntVal(stridesOpt, i);
      // Output size is input size divided by stride. When stride is 1, then
      // input and output are the same size, which is the usual case. When
      // stride is greater than 1, take the ceil to be sure to have each input
      // value used, as padding will be used to fill the gaps.
      int64_t outputSize = ceil((1.0 * inputSize) / (1.0 * strideVal));
      // Forumla is from ONNX MaxPool, and can be explained as follows. Pads is
      // the difference between the needed values for the computations, minus
      // the input values. The needed values for the computation is the
      // effective side of the kernel plus the number of times we jump to the
      // next kernel. Number of time we jump is (outputSize - 1). That number is
      // multiplied with the size of the jump, namely strideVal. Now for the
      // effective kernel size. It is the kernelSize + the number of times we
      // have dilation holes time the dialtion. The number of dialtion holes is
      // (kernelSize -1). Thus the effective size is "kernelSize +
      // (kernelSize-1)*dialation". This simplifies to "(kernelSize
      // -1)*dialation + 1".
      auto sumOfPad = (outputSize - 1) * strideVal +
                      ((kernelSize - 1) * dilationVal + 1) - inputSize;
      // Pad values are assumed equal on both size, at half the total value.
      actualPads[i] = actualPads[kernelRank + i] = sumOfPad / 2;
      // But if the total pad value is odd, we add 1 to begining or end
      // depending on autoPad value.
      if (sumOfPad % 2 != 0) {
        if (autoPad == "SAME_UPPER") {
          actualPads[kernelRank + i] += 1;
        } else {
          actualPads[i] += 1;
        }
      }
    }
  } else if (autoPad == "VALID") {
    // No pad, default value was set to zero, we are all set.
    updatedPad = true;
  } else {
    op->emitError("auto_pad of unknown / unsupported value");
  }
  // Set pads values in attributes, if it is needed.
  if (updatedPad) {
    ArrayRef<int64_t> defaultRefs(actualPads);
    op->padsAttr(builder.getI64ArrayAttr(defaultRefs));
  }
  // In all cases now, the acutal pad values are found in the pads attribute.
  op->auto_padAttr(builder.getStringAttr("NOTSET"));
}

//===----------------------------------------------------------------------===//
// Support function that computes default values for dilations, strides, and
// pads.
template <class T>
static void processConvTypeParams(T *op, Value inputOperand) {
  auto builder = mlir::Builder(op->getContext());

  // 1) Get shape of input.
  auto inputShape = inputOperand.getType().cast<RankedTensorType>().getShape();
  auto inputRank = inputShape.size();

  // 2) Get kernel_shape attribute.
  auto kernelShape = op->kernel_shape();

  // Dilation.
  processConvDilationParam<T>(op, kernelShape);
  auto dilationsOpt = op->dilations();

 // Strides.
  processConvStrideParam<T>(op, kernelShape);
  auto stridesOpt = op->strides();

  // Pads.
  processConvPadParam<T>(op, inputShape, kernelShape, stridesOpt, dilationsOpt);
}

//===----------------------------------------------------------------------===//
// Compute spatial dimensions given dilations, strides, pads, and ceil mode.
//
static void insertConvSpatialDim(SmallVector<int64_t, 4> *outputDims,
    ArrayRef<int64_t> xShape, Optional<ArrayAttr> kernelShape,
    Optional<ArrayAttr> padsOpt, Optional<ArrayAttr> stridesOpt,
    Optional<ArrayAttr> dilationsOpt = llvm::None, bool ceilMode = false) {
  auto xRank = xShape.size();
  auto spatialRank = ArrayAttrSize(kernelShape);
  auto spatialOffset = xRank - spatialRank;

  int64_t dilationVal = 1;
  for (int i = 0; i < spatialRank; ++i) {
    auto inputSize = xShape[spatialOffset + i];
    auto sumOfPads =
        ArrayAttrIntVal(padsOpt, i) + ArrayAttrIntVal(padsOpt, spatialRank + i);
    auto kernelSize = ArrayAttrIntVal(kernelShape, i);
    if (dilationsOpt.hasValue())
      dilationVal = ArrayAttrIntVal(dilationsOpt, i);
    auto strideVal = ArrayAttrIntVal(stridesOpt, i);
    // Number of useful values: input plus pad - effective size of kernel (see
    // processConvTypeParams comments to see how this value is derived).
    double numerator =
        inputSize + sumOfPads - ((kernelSize - 1) * dilationVal + 1);
    // Useful number is divided by the strides.
    double denominator = strideVal;
    int64_t res;
    if (ceilMode) {
      res = ceil(numerator / denominator) + 1;
    } else {
      res = floor(numerator / denominator) + 1;
    }
    outputDims->emplace_back(res);
  }
}

//===----------------------------------------------------------------------===//
// ONNXOpsDialect
//===----------------------------------------------------------------------===//

/// Dialect creation, the instance will be owned by the context. This is the
/// point of registration of custom types and operations for the dialect.
ONNXOpsDialect::ONNXOpsDialect(mlir::MLIRContext *ctx)
    : mlir::Dialect(getDialectNamespace(), ctx) {
  addOperations<
#define GET_OP_LIST
#include "src/onnx.cpp.inc"
      >();
}

void ONNXEntryPointOp::build(mlir::Builder *builder,
    mlir::OperationState &state, mlir::FuncOp function, int numInputs,
    int numOutputs) {
  state.addAttribute(ONNXEntryPointOp::getEntryPointFuncAttrName(),
      builder->getSymbolRefAttr(function));
  state.addAttribute(ONNXEntryPointOp::getNumInputsAttrName(),
      builder->getI32IntegerAttr(numInputs));
  state.addAttribute(ONNXEntryPointOp::getNumOutputsAttrName(),
      builder->getI32IntegerAttr(numOutputs));
}

ONNXEntryPointOp ONNXEntryPointOp::create(mlir::Location location,
    mlir::FuncOp &func, int numInputs, int numOutputs) {
  mlir::OperationState state(location, "onnx.EntryPoint");
  Builder builder(location->getContext());
  mlir::ONNXEntryPointOp::build(&builder, state, func, numInputs, numOutputs);
  Operation *op = mlir::Operation::create(state);
  auto onnxEntryOp = llvm::cast<mlir::ONNXEntryPointOp>(op);
  return onnxEntryOp;
}

//===----------------------------------------------------------------------===//
// ONNX Operations
//===----------------------------------------------------------------------===//
// Exp
/// Infer the output shape of the ONNXExpOp. This method is required by the
/// shape inference interface.
void ONNXExpOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Tanh
/// Infer the output shape of the ONNXTanhOp. This method is required by the
/// shape inference interface.
void ONNXTanhOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Sinh
/// Infer the output shape of the ONNXSinhOp. This method is required by the
/// shape inference interface.
void ONNXSinhOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Cosh
/// Infer the output shape of the ONNXCoshOp. This method is required by the
/// shape inference interface.
void ONNXCoshOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Cos
/// Infer the output shape of the ONNXCosOp. This method is required by the
/// shape inference interface.
void ONNXCosOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Log
/// Infer the output shape of the ONNXLogOp. This method is required by the
/// shape inference interface.
void ONNXLogOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// HardSigmoid
/// Infer the output shape of the ONNXHardSigmoidOp. This method is required by
/// the shape inference interface.
void ONNXHardSigmoidOp::inferShapes() {
  getResult().setType(getOperand().getType());
}

//===----------------------------------------------------------------------===//
// Sigmoid
/// Infer the output shape of the ONNXSigmoidOp. This method is required by the
/// shape inference interface.
void ONNXSigmoidOp::inferShapes() {
  getResult().setType(getOperand().getType());
}

//===----------------------------------------------------------------------===//
// Elu
/// Infer the output shape of the ONNXEluOp. This method is required by the
/// shape inference interface.
void ONNXEluOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Relu
/// Infer the output shape of the ONNXReluOp. This method is required by the
/// shape inference interface.
void ONNXReluOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// LeakyRelu
/// Infer the output shape of the ONNXLeakyReluOp. This method is required by
/// the shape inference interface.
void ONNXLeakyReluOp::inferShapes() {
  getResult().setType(getOperand().getType());
}

//===----------------------------------------------------------------------===//
// Selu
/// Infer the output shape of the ONNXSeluOp. This method is required by
/// the shape inference interface.
void ONNXSeluOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Reciprocal
/// Infer the output shape of the ONNXReciprocalOp. This method is required by
/// the shape inference interface.
void ONNXReciprocalOp::inferShapes() {
  getResult().setType(getOperand().getType());
}

//===----------------------------------------------------------------------===//
// Softmax
/// Infer the output shape of the ONNXSoftmaxOp. This method is required by
/// the shape inference interface.
void ONNXSoftmaxOp::inferShapes() {
  getResult().setType(getOperand().getType());
}

//===----------------------------------------------------------------------===//
// Softplus
/// Infer the output shape of the ONNXSoftplusOp. This method is required by
/// the shape inference interface.
void ONNXSoftplusOp::inferShapes() {
  getResult().setType(getOperand().getType());
}

//===----------------------------------------------------------------------===//
// Softsign
/// Infer the output shape of the ONNXSoftsignOp. This method is required by
/// the shape inference interface.
void ONNXSoftsignOp::inferShapes() {
  getResult().setType(getOperand().getType());
}

//===----------------------------------------------------------------------===//
// Sqrt
/// Infer the output shape of the ONNXSqrtOp. This method is required by
/// the shape inference interface.
void ONNXSqrtOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Sign
/// Infer the output shape of the ONNXSignOp. This method is required by
/// the shape inference interface.
void ONNXSignOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Abs
/// Infer the output shape of the ONNXAbsOp. This method is required by the
/// shape inference interface.
void ONNXAbsOp::inferShapes() { getResult().setType(getOperand().getType()); }

//===----------------------------------------------------------------------===//
// Add
/// Infer the output shape of the ONNXAddOp. This method is required by the
/// shape inference interface.
void ONNXAddOp::inferShapes() {
  if (!getOperand(0).getType().isa<RankedTensorType>() ||
      !getOperand(1).getType().isa<RankedTensorType>())
    return;
  auto lhsTy = getOperand(0).getType().cast<RankedTensorType>();
  auto rhsTy = getOperand(1).getType().cast<RankedTensorType>();
  getResult().setType(getBroadcastedType(lhsTy, rhsTy));
}

//===----------------------------------------------------------------------===//
// Mul
/// Infer the output shape of the ONNXMulOp. This method is required by the
/// shape inference interface.
void ONNXMulOp::inferShapes() {
  if (!getOperand(0).getType().isa<RankedTensorType>() ||
      !getOperand(1).getType().isa<RankedTensorType>())
    return;
  auto lhsTy = getOperand(0).getType().cast<RankedTensorType>();
  auto rhsTy = getOperand(1).getType().cast<RankedTensorType>();
  getResult().setType(getBroadcastedType(lhsTy, rhsTy));
}

//===----------------------------------------------------------------------===//
// Div
/// Infer the output shape of the ONNXDivOp. This method is required by the
/// shape inference interface.
void ONNXDivOp::inferShapes() {
  if (!getOperand(0).getType().isa<RankedTensorType>() ||
      !getOperand(1).getType().isa<RankedTensorType>())
    return;
  auto lhsTy = getOperand(0).getType().cast<RankedTensorType>();
  auto rhsTy = getOperand(1).getType().cast<RankedTensorType>();
  getResult().setType(getBroadcastedType(lhsTy, rhsTy));
}

//===----------------------------------------------------------------------===//
// Sub
/// Infer the output shape of the ONNXSubOp. This method is required by the
/// shape inference interface.
void ONNXSubOp::inferShapes() {
  if (!getOperand(0).getType().isa<RankedTensorType>() ||
      !getOperand(1).getType().isa<RankedTensorType>())
    return;
  auto lhsTy = getOperand(0).getType().cast<RankedTensorType>();
  auto rhsTy = getOperand(1).getType().cast<RankedTensorType>();
  getResult().setType(getBroadcastedType(lhsTy, rhsTy));
}

//===----------------------------------------------------------------------===//
// And
/// Infer the output shape of the ONNXAndOp. This method is required by the
/// shape inference interface.
void ONNXAndOp::inferShapes() {
  if (!getOperand(0).getType().isa<RankedTensorType>() ||
      !getOperand(1).getType().isa<RankedTensorType>())
    return;
  auto lhsTy = getOperand(0).getType().cast<RankedTensorType>();
  auto rhsTy = getOperand(1).getType().cast<RankedTensorType>();
  getResult().setType(getBroadcastedType(lhsTy, rhsTy));
}

//===----------------------------------------------------------------------===//
// Or
/// Infer the output shape of the ONNXOrOp. This method is required by the
/// shape inference interface.
void ONNXOrOp::inferShapes() {
  if (!getOperand(0).getType().isa<RankedTensorType>() ||
      !getOperand(1).getType().isa<RankedTensorType>())
    return;
  auto lhsTy = getOperand(0).getType().cast<RankedTensorType>();
  auto rhsTy = getOperand(1).getType().cast<RankedTensorType>();
  getResult().setType(getBroadcastedType(lhsTy, rhsTy));
}

//===----------------------------------------------------------------------===//
// Xor
/// Infer the output shape of the ONNXXorOp. This method is required by the
/// shape inference interface.
void ONNXXorOp::inferShapes() {
  if (!getOperand(0).getType().isa<RankedTensorType>() ||
      !getOperand(1).getType().isa<RankedTensorType>())
    return;
  auto lhsTy = getOperand(0).getType().cast<RankedTensorType>();
  auto rhsTy = getOperand(1).getType().cast<RankedTensorType>();
  getResult().setType(getBroadcastedType(lhsTy, rhsTy));
}

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

//===----------------------------------------------------------------------===//
// Sum
/// Infer the output shape of the ONNXSumOp. This method is required by the
/// shape inference interface.
void ONNXSumOp::inferShapes() {
  for (int i = 0; i < getNumOperands(); ++i) {
    if (!getOperand(i).getType().cast<RankedTensorType>())
      return;
  }
  Type resultTy = getOperand(0).getType().cast<RankedTensorType>();
  for (int i = 1; i < getNumOperands(); ++i) {
    Type nextTy = getOperand(i).getType().cast<RankedTensorType>();
    resultTy = getBroadcastedType(resultTy, nextTy);
  }
  getResult().setType(resultTy);
}

//===----------------------------------------------------------------------===//
// Max
/// Infer the output shape of the ONNXMaxOp. This method is required by the
/// shape inference interface.
void ONNXMaxOp::inferShapes() {
  for (int i = 0; i < getNumOperands(); ++i) {
    if (!getOperand(i).getType().cast<RankedTensorType>())
      return;
  }
  Type resultTy = getOperand(0).getType().cast<RankedTensorType>();
  for (int i = 1; i < getNumOperands(); ++i) {
    Type nextTy = getOperand(i).getType().cast<RankedTensorType>();
    resultTy = getBroadcastedType(resultTy, nextTy);
  }
  getResult().setType(resultTy);
}

//===----------------------------------------------------------------------===//
// Min
/// Infer the output shape of the ONNXMinOp. This method is required by the
/// shape inference interface.
void ONNXMinOp::inferShapes() {
  for (int i = 0; i < getNumOperands(); ++i) {
    if (!getOperand(i).getType().cast<RankedTensorType>())
      return;
  }
  Type resultTy = getOperand(0).getType().cast<RankedTensorType>();
  for (int i = 1; i < getNumOperands(); ++i) {
    Type nextTy = getOperand(i).getType().cast<RankedTensorType>();
    resultTy = getBroadcastedType(resultTy, nextTy);
  }
  getResult().setType(resultTy);
}

//===----------------------------------------------------------------------===//
// Identity
/// Infer the output shape of the ONNXIdentityOp. This method is required by the
/// shape inference interface.
void ONNXIdentityOp::inferShapes() {
  getResult().setType(getOperand().getType());
}

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

// MatMul

void ONNXMatMulOp::inferShapes() {
  // Cannot infer shape if no shape exists.
  if (!A().getType().isa<RankedTensorType>() ||
      !B().getType().isa<RankedTensorType>())
    return;

  auto lhsTy = A().getType().cast<RankedTensorType>();
  auto rhsTy = B().getType().cast<RankedTensorType>();

  SmallVector<int64_t, 2> dims;
  auto lhsShape = lhsTy.getShape();
  auto rhsShape = rhsTy.getShape();

  if (lhsShape.size() < 1 && rhsShape.size() < 1) {
    // Multiplication by scalars is not allowed.
    emitError("Multiplication by scalar arguments not allowed");
  } else if (lhsShape.size() == 1 && rhsShape.size() == 1) {
    // Special case when both arrays are 1-dimensional and according to
    // numpy rules the types need to be extended to 1xN and Nx1. Helper sizes
    // need to be removed after the multiplication but cannot be removed if all
    // sizes are 1.
    if (lhsShape[0] != -1 && rhsShape[0] != -1 && lhsShape[0] != rhsShape[0])
      emitError("Attempt to multiply incompatible matrices");
    dims.emplace_back(1);
  } else if (lhsShape.size() == 1 && rhsShape.size() >= 2) {
    // If the first argument is 1-D, it is promoted to a matrix by prepending a
    // 1 to its dimensions. After matrix multiplication the prepended 1 is
    // removed.
    //
    // N MATMUL (s1 x s2 x... x sK x N x P)
    // =>
    // (s1 x s2 x... x sK x P)

    // Check legality of matrix multiplication.
    unsigned rhsRank = rhsShape.size();
    if (lhsShape[0] != -1 && rhsShape[rhsRank - 2] != -1 &&
        lhsShape[0] != rhsShape[rhsRank - 2])
      emitError("Attempt to multiply incompatible matrices");

    for (decltype(rhsRank) i = 0; i < rhsRank - 2; ++i)
      dims.emplace_back(rhsShape[i]);
    dims.emplace_back(rhsShape[rhsRank - 1]);
  } else if (lhsShape.size() >= 2 && rhsShape.size() == 1) {
    // If the second argument is 1-D, it is promoted to a matrix by appending a
    // 1 to its dimensions. After matrix multiplication the appended 1 is
    // removed.
    //
    // (s1 x s2 x... x sK x M x N) MATMUL N
    // =>
    // (s1 x s2 x... x sK x M)

    // Check legality of matrix multiplication.
    unsigned lhsRank = lhsShape.size();
    if (lhsShape[lhsRank - 1] != -1 && rhsShape[0] != -1 &&
        lhsShape[lhsRank - 1] != rhsShape[0])
      emitError("Attempt to multiply incompatible matrices");

    for (decltype(lhsRank) i = 0; i < lhsRank - 2; ++i)
      dims.emplace_back(lhsShape[i]);
    dims.emplace_back(lhsShape[lhsRank - 2]);
  } else if (lhsShape.size() > 2 && rhsShape.size() == 2) {
    // (s1 x s2 x... x sK x M x N) MATMUL (N x P)
    // =>
    // (s1 x s2 x... x sK x M x P)

    // Check legality of matrix multiplication.
    unsigned lhsRank = lhsShape.size();
    if (lhsShape[lhsRank - 1] != -1 && rhsShape[0] != -1 &&
        lhsShape[lhsRank - 1] != rhsShape[0])
      emitError("Attempt to multiply incompatible matrices");

    for (decltype(lhsRank) i = 0; i < lhsRank - 1; ++i)
      dims.emplace_back(lhsShape[i]);
    dims.emplace_back(rhsShape[1]);
  } else if (lhsShape.size() == 2 && rhsShape.size() > 2) {
    // (M x N) MATMUL (s1 x s2 x... x sK x N x P)
    // =>
    // (s1 x s2 x... x sK x M x P)

    // Check legality of matrix multiplication.
    unsigned rhsRank = rhsShape.size();
    if (lhsShape[1] != -1 && rhsShape[rhsRank - 2] != -1 &&
        lhsShape[1] != rhsShape[rhsRank - 2])
      emitError("Attempt to multiply incompatible matrices");

    for (decltype(rhsRank) i = 0; i < rhsRank - 2; ++i)
      dims.emplace_back(rhsShape[i]);
    dims.emplace_back(lhsShape[0]);
    dims.emplace_back(rhsShape[rhsRank - 1]);
  } else if (lhsShape.size() > 2 && rhsShape.size() > 2) {
    // (s1 x s2 x... x sK x M x N) MATMUL (t1 x t2 x... x tK x N x P)
    // =>
    // (u1 x u2 x... x uK x M x P)

    // Check legality of matrix multiplication.
    unsigned lhsRank = lhsShape.size();
    unsigned rhsRank = rhsShape.size();
    if (lhsShape[lhsRank - 1] != -1 && rhsShape[rhsRank - 2] != -1 &&
        lhsShape[lhsRank - 1] != rhsShape[rhsRank - 2])
      emitError("Attempt to multiply incompatible matrices");

    // Check and perform broadcasting for the shapes.
    SmallVector<int64_t, 2> lhsBcastShape;
    for (decltype(lhsRank) i = 0; i < lhsRank - 2; ++i)
      lhsBcastShape.emplace_back(lhsShape[i]);
    SmallVector<int64_t, 2> rhsBcastShape;
    for (decltype(rhsRank) i = 0; i < rhsRank - 2; ++i)
      rhsBcastShape.emplace_back(rhsShape[i]);
    if (!getBroadcastedShape(lhsBcastShape, rhsBcastShape, dims))
      emitError("Broadcasted dimensions are incompatible");

    dims.emplace_back(lhsShape[lhsRank - 2]);
    dims.emplace_back(rhsShape[rhsRank - 1]);
  } else {
    // This case covers all remaining combinations of 1 and 2-D matrices.
    int64_t lhsDim = lhsShape[0];
    int64_t rhsDim = rhsShape[0];
    if (lhsShape.size() > 1) {
      lhsDim = lhsShape[1];
      dims.emplace_back(lhsShape[0]);
    }

    // Check legality of matrix multiplication.
    if (lhsDim != -1 && rhsDim != -1 && lhsDim != rhsDim)
      emitError("Attempt to multiply incompatible matrices");

    if (rhsShape.size() > 1)
      dims.emplace_back(rhsShape[1]);
  }

  getResult().setType(RankedTensorType::get(dims, lhsTy.getElementType()));
}

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

// Gemm

void ONNXGemmOp::inferShapes() {
  bool hasBias = !C().getType().isa<NoneType>();
  // Cannot infer shape if no shape exists.
  if (!A().getType().isa<RankedTensorType>() ||
      !B().getType().isa<RankedTensorType>() ||
      (hasBias && !C().getType().isa<RankedTensorType>()))
    return;
  auto lhsTy = A().getType().cast<RankedTensorType>();
  auto rhsTy = B().getType().cast<RankedTensorType>();

  int64_t M, N, K_A, K_B;
  M = (transA() == 0) ? lhsTy.getShape()[0] : lhsTy.getShape()[1];
  K_A = (transA() == 0) ? lhsTy.getShape()[1] : lhsTy.getShape()[0];
  N = (transB() == 0) ? rhsTy.getShape()[1] : rhsTy.getShape()[0];
  K_B = (transB() == 0) ? rhsTy.getShape()[0] : rhsTy.getShape()[1];

  if ((K_A != -1) and (K_B != -1) and (K_A != K_B)) {
    emitError("Tensor shapes mismatched");
  }

  if (hasBias) {
    // Check whether bias is unidirectional broadcasting or not.
    auto biasTy = C().getType().cast<RankedTensorType>();
    auto shape = biasTy.getShape();
    int rank = shape.size();
    if ((rank > 2) ||
        (rank >= 1 && shape[rank - 1] != -1 && N != -1 &&
            N != shape[rank - 1] && shape[rank - 1] != 1) ||
        (rank == 2 && shape[rank - 2] != -1 && M != -1 &&
            M != shape[rank - 2] && shape[rank - 2] != 1)) {
      emitError("Bias shape mismatched");
    }
  }

  SmallVector<int64_t, 2> dims;
  dims.emplace_back(M);
  dims.emplace_back(N);
  getResult().setType(RankedTensorType::get(dims, lhsTy.getElementType()));
}

/// BatchNormalizationTestMode
void ONNXBatchNormalizationTestModeOp::inferShapes() {
  // Cannot infer shape if no shape exists.
  if (!X().getType().isa<RankedTensorType>() ||
      !scale().getType().isa<RankedTensorType>() ||
      !B().getType().isa<RankedTensorType>() ||
      !mean().getType().isa<RankedTensorType>() ||
      !var().getType().isa<RankedTensorType>())
    return;

  auto inputTensorTy = X().getType().cast<RankedTensorType>();
  auto scaleTensorTy = scale().getType().cast<RankedTensorType>();
  auto biasTensorTy = B().getType().cast<RankedTensorType>();
  auto meanTensorTy = mean().getType().cast<RankedTensorType>();
  auto varianceTensorTy = var().getType().cast<RankedTensorType>();

  // Check whether the shapes of scale, bias, mean and variance are valid.
  // Operand's dimensions can be in the form of NxCxD1xD2x...xDn or N.
  // In case of N, C is assumed to be 1.
  // Shapes of scale, bias, mean and variance must be C.
  int64_t c = -1;
  if (inputTensorTy.getShape().size() == 1) {
    c = 1;
  } else if (inputTensorTy.getShape().size() > 2) {
    c = (inputTensorTy.getShape()[1] != -1) ? inputTensorTy.getShape()[1] : -1;
  } else {
    emitError("Wrong rank for the input");
  }

  if (c != -1) {
    auto s = scaleTensorTy.getShape();
    auto b = biasTensorTy.getShape();
    auto m = meanTensorTy.getShape();
    auto v = varianceTensorTy.getShape();

    if ((s.size() != 1) || (s[0] != -1 && s[0] != c))
      emitError("Wrong rank for the scale");
    if ((b.size() != 1) || (b[0] != -1 && b[0] != c))
      emitError("Wrong rank for the bias");
    if ((m.size() != 1) || (m[0] != -1 && m[0] != c))
      emitError("Wrong rank for the mean");
    if ((v.size() != 1) || (v[0] != -1 && v[0] != c))
      emitError("Wrong rank for the variance");
  }

  // The output tensor of the same shape as the input.
  getResult().setType(X().getType());
}

// TODO:
//   Verify that matrix sizes are valid for multiplication and addition.
//   Take into account the dimensionality of the matrix.

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

// Reshape

void ONNXReshapeOp::inferShapes() {
  // Cannot infer shape if no shape tensor is specified.
  if (!shape().getType().isa<RankedTensorType>())
    emitError("Shape tensor not ranked");

  auto inputTensorTy = data().getType().cast<RankedTensorType>();
  auto shapeTensorTy = shape().getType().cast<RankedTensorType>();

  // Only rank 1 shape tensors are supported.
  if (shapeTensorTy.getShape().size() != 1)
    emitError("Shape tensor must have rank one");

  int64_t outputRank = shapeTensorTy.getShape()[0];

  // Shape tensor must have constant shape.
  if (outputRank < 0)
    emitError("Shape tensor must have constant shape");

  // Compute total number of elements.
  int64_t totalInputSize = 1;
  for(auto inputDim : inputTensorTy.getShape())
    totalInputSize *= inputDim;

  // Check if second argument of ReshapeOp is a constant.
  auto constantOp = getONNXConstantOp(shape());

  SmallVector<int64_t, 2> dims(outputRank, -1);
  if (constantOp) {
    DenseElementsAttr valueAttribute =
        constantOp.valueAttr().dyn_cast<DenseElementsAttr>();

    if (!valueAttribute)
      emitError("DenseElementsAttr expected");

    // Get dims from valueAttribute.
    auto valueIt = valueAttribute.getValues<IntegerAttr>().begin();
    for (int i=0; i<outputRank; ++i)
      dims[i] = (*valueIt++).cast<IntegerAttr>().getInt();

    if (valueIt != valueAttribute.getValues<IntegerAttr>().end())
      emitError("Constant value must have same rank as output");

    int64_t numberOfDynamicInputs = 0;
    int64_t totalKnownDimsSize = 1;
    int64_t dynamicValueIndex = -1;
    for (int i=0; i<outputRank; ++i) {
      // Set output dimension.
      if (dims[i] == 0)
        dims[i] = inputTensorTy.getShape()[i];

      if (dims[i] < 0) {
        numberOfDynamicInputs++;
        dynamicValueIndex = i;
      } else {
        totalKnownDimsSize *= dims[i];
      }
    }

    // If the number of dynamic inputs is 1 then deduce the missing value
    // based on the total input size. The total input size must be greater
    // than 0 i.e. all constant dimensions.
    // TODO: Support dynamic input dimensons.
    if (numberOfDynamicInputs == 1 && totalKnownDimsSize > 0 &&
        totalInputSize > 0)
      dims[dynamicValueIndex] = totalInputSize / totalKnownDimsSize;
  }

  getResult().setType(
      RankedTensorType::get(dims, inputTensorTy.getElementType()));
}

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

// Transpose

void ONNXTransposeOp::inferShapes() {
  // Cannot infer shape if no shape exists.
  if (!data().getType().isa<RankedTensorType>())
    return;

  // Naive transposition which handles the default case of
  // reversing the shape of the tensor (similar to numpy.transpose).
  auto arrayTy = data().getType().cast<RankedTensorType>();
  SmallVector<int64_t, 2> dims;
  auto permutation = ONNXTransposeOp::permAttr();
  if (permutation) {
    // Perform transposition according to perm attribute.
    for (auto perm : permutation.getValue())
      dims.emplace_back(arrayTy.getShape()[perm.cast<IntegerAttr>().getInt()]);
  } else {
    // Default
    for (auto dim : llvm::reverse(arrayTy.getShape()))
      dims.emplace_back(dim);
  }

  getResult().setType(RankedTensorType::get(dims, arrayTy.getElementType()));
}

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

// ReduceMax

void ONNXReduceMaxOp::inferShapes() {
  if (!getOperand().getType().isa<RankedTensorType>()) {
    emitError("Shape tensor not ranked");
    return;
  }

  auto operandTy = getOperand().getType().cast<RankedTensorType>();
  getResult().setType(getReductionOutputType(operandTy, axes(), keepdims()));
}

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

// ReduceMin

void ONNXReduceMinOp::inferShapes() {
  if (!getOperand().getType().isa<RankedTensorType>()) {
    emitError("Shape tensor not ranked");
    return;
  }

  auto operandTy = getOperand().getType().cast<RankedTensorType>();
  getResult().setType(getReductionOutputType(operandTy, axes(), keepdims()));
}

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

// ReduceProd

void ONNXReduceProdOp::inferShapes() {
  if (!getOperand().getType().isa<RankedTensorType>()) {
    emitError("Shape tensor not ranked");
    return;
  }

  auto operandTy = getOperand().getType().cast<RankedTensorType>();
  getResult().setType(getReductionOutputType(operandTy, axes(), keepdims()));
}

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

// ReduceSum

void ONNXReduceSumOp::inferShapes() {
  if (!getOperand().getType().isa<RankedTensorType>()) {
    emitError("Shape tensor not ranked");
    return;
  }

  auto operandTy = getOperand().getType().cast<RankedTensorType>();
  getResult().setType(getReductionOutputType(operandTy, axes(), keepdims()));
}

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

// Conv

// For this operation, we define the attributes once in the original Conv
// operation class. There is no need to redefine the attribute names for the
// other classes based on Conv.
// Conv attributes output:
//   -  auto_pad set to NOTSET;
//   -  dilations, strides: set to 1 if not defined by user;
//   -  kernelShape: inferred from weight matrix if not defined by user;
//   -  pads: set to proper value, 0 if not defined by user.

void ONNXConvNoBiasOp::inferShapes() {
  // Generic shape for data input X and weight tensor W:
  // X: (N x C x D1 x D2 ... x Dn)
  // W: (M x C/group x k1 x k2 x ... x kn)

  // Cannot infer shape if no shape exists.
  if (!X().getType().isa<RankedTensorType>() ||
      !W().getType().isa<RankedTensorType>())
    return;

  auto xTy = X().getType().cast<RankedTensorType>();
  auto xShape = xTy.getShape();
  auto weightTy = W().getType().cast<RankedTensorType>();
  auto weightShape = weightTy.getShape();
  auto builder = mlir::Builder(this->getContext());

  // Lowest supported convolution is a one dimensional convolution.
  if (xShape.size() < 3)
    emitError("Data input shape must be at least (NxCxD1)");

  // Check that shape of weight and data have same length.
  if (xShape.size() != weightShape.size())
    emitError("Weight size not compatible with data size");

  // Group is a required attribute and should have default value of 1.
  int64_t group = ONNXConvNoBiasOp::group().getSExtValue();

  // Check if the attribute actually exists. If it does not then add it.
  if (!groupAttr())
    groupAttr(builder.getI64IntegerAttr(group));

  // Check that the X.shape[1] == (W.shape[1] * group) == C condition holds.
  if (xShape[1] != -1 && weightShape[1] != -1 &&
      xShape[1] != (weightShape[1] * group))
    emitError("Channel dimension mismatch");

  // Note: the value of the group attribut only impacts the way the
  // computation is carried out and not the actual output size.

  // Number of spatial dimensions.
  auto spatialOffset = 2;
  int32_t spatialRank = xShape.size() - spatialOffset;

  // Use kernel_shape attribute if present otherwise use size from weight
  // argument.
  auto kernelShape = kernel_shape();
  if (kernelShape.hasValue()) {
    if (ArrayAttrSize(kernelShape) != spatialRank)
      emitError("kernel_shape length incompatible with spatial dimensions");
    // Have the right number of values, check them.
    for (int i = 0; i < spatialRank; ++i)
      if (ArrayAttrIntVal(kernelShape, i) < 1)
        emitError("bad kernel_shape value");
  } else {
    // Deduce shape from weight input.
    SmallVector<int64_t, 2> defaultVals;
    for (int i = 0; i < spatialRank; ++i)
      defaultVals.emplace_back(weightShape[spatialOffset + i]);
    // Convert to ArrayRef, then build attribute, then store attribute.
    ArrayRef<int64_t> defaultRefs(defaultVals);
    auto builder = mlir::Builder(getContext());
    kernel_shapeAttr(builder.getI64ArrayAttr(defaultRefs));
    kernelShape = kernel_shape();
  }

  // Process strides, dilations, and pads.
  processConvTypeParams<>(this, X());
  auto dilationsOpt = dilations();
  auto stridesOpt = strides();
  auto padsOpt = pads();

  // First two output dimensions consist of the number of batches and the
  // number of kernels being applied.
  SmallVector<int64_t, 4> outputDims;
  // Insert batch size.
  outputDims.emplace_back(xShape[0]);
  // Insert number of filters being applied (number of output channels).
  outputDims.emplace_back(weightShape[0]);
  // Compute and insert spatial dims.
  insertConvSpatialDim(
      &outputDims, xShape, kernelShape, padsOpt, stridesOpt, dilationsOpt);

  getResult().setType(RankedTensorType::get(outputDims, xTy.getElementType()));
}

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

// AveragePool
// Infer shape attributes output:
//   -  auto_pad set to NOTSET;
//   -  strides: set to 1 if not defined by user;
//   -  pads: set to proper value, 0 if not defined by user.

void ONNXAveragePoolOp::inferShapes() {
  // Cannot infer shape if no shape exists.
  if (!X().getType().isa<RankedTensorType>())
    return;

  // Get shape of input.
  auto xTy = X().getType().cast<RankedTensorType>();
  auto xShape = xTy.getShape();

  // Kernel shape.
  auto kernelShape = kernel_shape();
  if (!kernelShape)
    emitError(
        "kernel_shape is a mandatory attribute for which there is no default");

  // Ceil mode.
  auto ceilMode = ceil_mode().getSExtValue();

  // Process strides and pads.
  processConvStrideParam<ONNXAveragePoolOp>(this, kernelShape);
  auto stridesOpt = strides();
  processConvPadParam<ONNXAveragePoolOp>(
      this, xShape, kernelShape, stridesOpt, llvm::None);
  auto padsOpt = pads();

  SmallVector<int64_t, 4> outputDims;
  // Insert batch size.
  outputDims.emplace_back(xShape[0]);
  outputDims.emplace_back(xShape[1]);
  // Compute and insert spatial dims.
  insertConvSpatialDim(&outputDims, xShape, kernelShape, padsOpt, stridesOpt,
      llvm::None, ceilMode);

  getResult().setType(RankedTensorType::get(outputDims, xTy.getElementType()));
}

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

// MaxPoolSingleOut
// Infer shape attributes output:
//   -  auto_pad set to NOTSET;
//   -  dilations, strides: set to 1 if not defined by user;
//   -  pads: set to proper value, 0 if not defined by user.

void ONNXMaxPoolSingleOutOp::inferShapes() {
  // Cannot infer shape if no shape exists.
  if (!X().getType().isa<RankedTensorType>())
    return;

  // Get shape of input.
  auto xTy = X().getType().cast<RankedTensorType>();
  auto xShape = xTy.getShape();

  // Kernel shape.
  auto kernelShape = kernel_shape();
  if (!kernelShape)
    emitError(
        "kernel_shape is a mandatory attribute for which there is no default");

  // Storage order.
  auto storageOrder = storage_order().getSExtValue();
  if (storageOrder != 0)
    emitError("column major storage order not supported at this time");

  // Process strides, dilations, and pads.
  processConvTypeParams<>(this, X());
  auto dilationsOpt = dilations();
  auto stridesOpt = strides();
  auto padsOpt = pads();

  // Ceil mode.
  auto ceilMode = ceil_mode().getSExtValue();

  SmallVector<int64_t, 4> outputDims;
  // Insert batch size.
  outputDims.emplace_back(xShape[0]);
  outputDims.emplace_back(xShape[1]);
  // Compute and insert spatial dims.
  insertConvSpatialDim(&outputDims, xShape, kernelShape, padsOpt, stridesOpt,
      dilationsOpt, ceilMode);

  getResult().setType(RankedTensorType::get(outputDims, xTy.getElementType()));
}

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

static Type padShapeInferenceHelper(Value data, ArrayAttr padsOpt) {
  // Cannot infer shape if no shape exists.
  if (!data.getType().isa<RankedTensorType>())
    return (Type)NULL;
  auto dataTy = data.getType().cast<RankedTensorType>();
  auto dataShape = dataTy.getShape();
  auto dataRank = dataShape.size();
  SmallVector<int64_t, 4> outputShape(dataShape.begin(), dataShape.end());
  if (padsOpt) {
    auto padsArray = padsOpt.getValue();
    // Pads consists of two values for each axis of data.
    // The two values specify the number of elements padded before and after
    // respectively.
    for (int i = 0; i < dataRank; ++i) {
      int64_t p1 = (padsArray[i]).cast<IntegerAttr>().getInt();
      int64_t p2 = (padsArray[i + dataRank]).cast<IntegerAttr>().getInt();
      // Have to non-negative constant
      if (p1 < 0 || p2 < 0)
        return (Type)NULL;
      if (outputShape[i] != -1) 
        outputShape[i] += p1 + p2;
    }

    return (RankedTensorType::get(outputShape, dataTy.getElementType()));
  } else {
    return (Type)NULL;
  }
}

// PadConstantPad

void ONNXPadConstantPadOp::inferShapes() {
  auto outputType = padShapeInferenceHelper(data(), pads());
  if (outputType) {
    getResult().setType(outputType);
  }
  return;
}

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

// PadConstantValuePad

void ONNXPadConstantValuePadOp::inferShapes() {
  auto outputType = padShapeInferenceHelper(data(), pads());
  if (outputType) {
    getResult().setType(outputType);
  } 
  return;
}

void ONNXPadConstantValuePadOp::build(Builder *builder, OperationState &state,
    Value data, ArrayAttr pads, FloatAttr constant_value, StringAttr mode) {
  Type outputType = padShapeInferenceHelper(data, pads);
  if (!outputType) {
    auto elementType = data.getType().cast<TensorType>().getElementType();
    outputType = UnrankedTensorType::get(elementType);
  }
  build(builder, state, outputType, data, pads, constant_value, mode);
}

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

// Unsqueeze

void ONNXUnsqueezeOp::inferShapes() {
  if (!data().getType().isa<RankedTensorType>())
    return;

  auto operandTy = data().getType().cast<RankedTensorType>();
  int inRank = operandTy.getRank();

  ArrayAttr axisAttrs = axesAttr();
  SmallVector<int, 4> axes;
  int outRank = 0;
  if (axisAttrs) {
    outRank = inRank + axisAttrs.getValue().size();
    for (auto axisAttr : axisAttrs.getValue()) {
      int axis = axisAttr.cast<IntegerAttr>().getInt();
      axis = axis >= 0 ? axis : (outRank + axis);
      // Valid range
      assert(axis >= -outRank && axis <= outRank - 1);
      if (std::find(axes.begin(), axes.end(), axis) == axes.end())
        axes.emplace_back(axis);
      else
        emitError("Duplicated axes");
    }
  } else {
    emitError("Axes attribute is required");
  }

  SmallVector<int64_t, 4> dims;
  for (int i = 0, j = 0; i < outRank || j < inRank; ++i) {
    if (std::find(axes.begin(), axes.end(), i) != axes.end()) {
      dims.emplace_back(1);
    } else {
      dims.emplace_back(operandTy.getShape()[j++]);
    }
  }
  getResult().setType(RankedTensorType::get(dims, operandTy.getElementType()));
}

//===----------------------------------------------------------------------===//
// Constant

void ONNXConstantOp::inferShapes() {
  if ((sparse_value().hasValue() && value().hasValue()) ||
      (!sparse_value().hasValue() && !value().hasValue()))
    emitError("Require exactly one of the two attributes, either value or "
              "sparse_value");

  ElementsAttr valAttr;
  if (sparse_value().hasValue())
    valAttr = sparse_valueAttr().cast<SparseElementsAttr>();
  else
    valAttr = valueAttr().cast<DenseElementsAttr>();
  getResult().setType(valAttr.getType());
}

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

#define GET_OP_CLASSES
#include "src/onnx.cpp.inc"