onnx-mlir/src/dialect/onnx/onnx_ops.cpp

//===- onnx_ops.cpp - MLIR ONNX Operations --------------------------------===//
//
// Copyright 2019 The IBM Research Authors.
//
// =============================================================================
//
// This file defines ONNX operations in the MLIR operation set.
//
//===----------------------------------------------------------------------===//
#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 "onnx_ops.hpp"

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

//===----------------------------------------------------------------------===//
// 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());
}

//===----------------------------------------------------------------------===//
// 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 (!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>();
  SmallVector<int64_t, 2> dims;
  dims.emplace_back(lhsTy.getShape()[0]);
  dims.emplace_back(rhsTy.getShape()[1]);
  getResult().setType(RankedTensorType::get(dims, lhsTy.getElementType()));
}

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

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

// Gemm

void ONNXGemmOp::inferShapes() {
  // Cannot infer shape if no shape exists.
  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>();
  SmallVector<int64_t, 2> dims;
  dims.emplace_back(lhsTy.getShape()[0]);
  dims.emplace_back(rhsTy.getShape()[1]);
  getResult().setType(RankedTensorType::get(dims, lhsTy.getElementType()));
}

// GemmNoBias

void ONNXGemmNoBiasOp::inferShapes() {
  // Cannot infer shape if no shape exists.
  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>();
  SmallVector<int64_t, 2> dims;
  dims.emplace_back(lhsTy.getShape()[0]);
  dims.emplace_back(rhsTy.getShape()[1]);
  getResult().setType(RankedTensorType::get(dims, lhsTy.getElementType()));
}

// 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 (!getOperand(1).getType().isa<RankedTensorType>())
    emitError("Shape tensor not ranked.");

  auto inputTensorTy = getOperand(0).getType().cast<RankedTensorType>();
  auto shapeTensorTy = getOperand(1).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.");

  SmallVector<int64_t, 2> dims;
  for (int i = 0; i < outputRank; ++i)
    dims.emplace_back(-1);

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

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

// Transpose

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

  // Naive transposition which handles the default case of
  // reversing the shape of the tensor (similar to numpy.transpose).
  auto arrayTy = getOperand().getType().cast<RankedTensorType>();
  SmallVector<int64_t, 2> dims;

  if (auto permutation = getAttrOfType<ArrayAttr>(
          ONNXTransposeOp::getPermAttrName())) {
    // 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()));
}

LogicalResult verify(ONNXTransposeOp op) {
  auto module = op.getParentOfType<ModuleOp>();
  if (!module)
    op.emitError("Expected to belong to a module.");

  if (auto permutation = op.getAttrOfType<ArrayAttr>(
          ONNXTransposeOp::getPermAttrName())) {
    for (auto perm : permutation.getValue())
      if (perm.cast<IntegerAttr>().getInt() < 0)
        op.emitError("Cannot tranpose, permuation contains negative index.");
  }

  return success();
}

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

// 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.
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 (!getOperand(0).getType().isa<RankedTensorType>() ||
      !getOperand(1).getType().isa<RankedTensorType>())
    return;

  auto dataTy = getOperand(0).getType().cast<RankedTensorType>();
  auto weightTy = getOperand(1).getType().cast<RankedTensorType>();
  auto dataShape = dataTy.getShape();
  auto weightShape = weightTy.getShape();

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

  // Required attribute auto_pad defaults to NOTSET.
  auto autoPad = getAttrOfType<StringAttr>(
      ONNXConvOp::getAutoPadAttrName()).getValue();
  // Group is a required attribute and should have default value of 1.
  int64_t group = getAttrOfType<IntegerAttr>(
      ONNXConvOp::getGroupAttrName()).getInt();
  // Check that the X.shape[1] == (W.shape[1] * group) == C condition holds.
  if (dataShape[1] != (weightShape[1] * group))
    emitError("Channel dimension mismatch.");

  // First two output dimensions consist of the number of batches and the
  // number of kernels being applied.
  //
  SmallVector<int64_t, 2> dims;
  // Insert batch size.
  dims.emplace_back(dataShape[0]);
  // Insert number of filters being applied (number of output channels).
  dims.emplace_back(weightShape[0]);

  // Spatial dimensions of the output are computed using the formula:
  //
  // dim = (inputDim - kernelDim + startPadding + endPadding) / stride + 1
  //
  SmallVector<int64_t, 2> outSpatialDims;
  // Number of spatial dimensions.
  int32_t nDims = dataShape.size() - 2;

  // Initialize dimenions based on the input spatial dimensions.
  for (int i = 2; i < dataShape.size(); ++i)
    outSpatialDims.emplace_back(dataShape[i]);

  // Use kernel_shape attribute if present otherwise use size from weight
  // argument.
  SmallVector<int64_t, 2> kernelDims;
  if (auto kernelShape = getAttrOfType<ArrayAttr>(
          ONNXConvOp::getKernelShapeAttrName())) {
    if (kernelShape.getValue().size() != nDims)
      emitError("kernel_shape length incompatible with spatial dimensions.");
    for (int i = 0; i < nDims; ++i)
      kernelDims.emplace_back(
          (kernelShape.getValue()[i]).cast<IntegerAttr>().getInt());
  } else {
    for (int i = 0; i < nDims; ++i)
      kernelDims.emplace_back(weightShape[i + 2]);
  }

  // Subtract kernel dimensions from input data dimensions.
  for (int i = 0; i < nDims; ++i)
    outSpatialDims[i] -= kernelDims[i];

  // Add padding information.
  if (autoPad == "NOTSET") {
    // Use pads to to determine the padding. If attribute is not
    // present then pads is considered to be all zeros (no padding).
    if (auto pads = getAttrOfType<ArrayAttr>(
            ONNXConvOp::getPadsAttrName())) {
      // pads consists of two entries for each spatial axis.
      if (pads.getValue().size() != 2 * nDims)
        emitError("pads size is not twice the spatial size.");

      for (int i = 0; i < nDims; ++i) {
        // Padding for beginning of axis.
        int32_t p = (pads.getValue()[i]).cast<IntegerAttr>().getInt();
        outSpatialDims[i] += p;
        // Padding for end of axis.
        p = (pads.getValue()[i + nDims]).cast<IntegerAttr>().getInt();
        outSpatialDims[i] += p;
      }
    }
  } else if (autoPad == "SAME_UPPER" || autoPad == "SAME_LOWER") {
    // Pad input so that output size matches input size.
    // Each spatial dimension needs to be padded by:
    //
    // ( K - 1 ) / 2
    //
    // where K is a kernel spatial dimension.
    for (int i = 0; i < nDims; ++i)
      outSpatialDims[i] += floor((kernelDims[i] - 1) / 2);
  } else if (autoPad == "VALID") {
    // No padding
  } else {
    emitError("Unexpected attribute value for auto_pad.");
  }

  // Strides
  if (auto strides = getAttrOfType<ArrayAttr>(
      ONNXConvOp::getStridesAttrName())) {
    if (strides.getValue().size() != nDims)
      emitError("strides length incompatible with spatial dimensions.");
    for (int i = 0; i < nDims; ++i) {
      int64_t stride =
          (strides.getValue()[i]).cast<IntegerAttr>().getInt();
      outSpatialDims[i] = floor(outSpatialDims[i] / stride);
    }
  }

  for (int i = 0; i < nDims; ++i)
    outSpatialDims[i] += 1;

  dims.append(outSpatialDims.begin(), outSpatialDims.end());
  getResult().setType(RankedTensorType::get(dims, dataTy.getElementType()));
}

LogicalResult verify(ONNXConvNoBiasOp op) {
  auto module = op.getParentOfType<ModuleOp>();
  if (!module)
    op.emitError("expected to belong to a module");

  auto autoPadAttr = op.getAttrOfType<StringAttr>(
      ONNXConvOp::getAutoPadAttrName());
  if (!autoPadAttr)
    op.emitError("auto_pad attribute not specified.");
  if (autoPadAttr.getValue() != "NOTSET")
    if (auto pads = op.getAttrOfType<ArrayAttr>(
            ONNXConvOp::getPadsAttrName()))
      op.emitError("auto_pad and pads are both set.");

  auto groupAttr =
      op.getAttrOfType<IntegerAttr>(ONNXConvOp::getGroupAttrName());
  if (!groupAttr)
    op.emitError("group attribute not specified.");

  return success();
}

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

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