gen_onnx_mlir.py

#!/usr/bin/env python3

# After modifying this file, the script will need to run to rebuild the
# onnx-mlir ONNX Dialect. This is performed by calling
# `make OMONNXOpsIncTranslation` in the build dir.
# If the changes are not seen, then you need to rebuild the entire onnx-mlir.

# After changes that impact the documentation of the ops, run
# "make onnx-mlir-docs".

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

from collections import defaultdict, OrderedDict
from io import StringIO
import io
import os
import sys
import datetime
import argparse

import numpy as np  # type: ignore

from onnx import defs, FunctionProto, helper, OperatorStatus
from onnx.defs import OpSchema, ONNX_DOMAIN, ONNX_ML_DOMAIN
from onnx.backend.test.case import collect_snippets
from onnx.backend.sample.ops import collect_sample_implementations
from typing import Any, Text, Sequence, Dict, List, Type, Set, Tuple

import pprint
import onnx

parser = argparse.ArgumentParser()
parser.add_argument(
    "--dry-run-onnx-ops",
    help="Output ONNXOps.td.inc content to stdout.",
    action="store_true",
    default=False,
)
parser.add_argument(
    "--dry-run-op-build-table",
    help="Output OpBuildTable.inc content to stdout.",
    action="store_true",
    default=False,
)
parser.add_argument(
    "--check-operation-version",
    help="check whether the imported onnx package has new operation or "
    " newer version of operation compared with version stored in version_dict",
    action="store_true",
    default=False,
)
parser.add_argument(
    "--list-operation-version",
    help="list the version stored in version_dict without performing checks",
    action="store_true",
    default=False,
)

args = parser.parse_args()

check_operation_version = args.check_operation_version
list_operation_version = args.list_operation_version

# ==UPDATE_ONNX_VERSION_OPSET==
# Look for tag above and update all references when upgrading the ONNX support within ONNX-MLIR.
current_onnx_version = "1.17.0"

# Check the version of onnx package being used.
if (
    not check_operation_version and not list_operation_version
) and current_onnx_version != onnx.__version__:
    print(
        "version of expected onnx is {}, ".format(current_onnx_version)
        + "while onnx package being used is {}".format(onnx.__version__)
    )
    quit()

# Record the version of each operation that is treated as the current version.
# To check whether the onnx package being used has newer version operation,
# run this script with --check-operation-version flag.
# Update this dictionary when a newer version is implemented
# TODO: how to keep the old version

version_dict = {
    "Abs": [13],
    "Acos": [7],
    "Acosh": [9],
    "Adagrad": [1],
    "Adam": [1],
    "Add": [14],
    "And": [7],
    "ArgMax": [13],
    "ArgMin": [13],
    "ArrayFeatureExtractor": [1],
    "Asin": [7],
    "Asinh": [9],
    "Atan": [7],
    "Atanh": [9],
    "AveragePool": [19],
    "BatchNormalization": [15],
    "Bernoulli": [15],
    "Binarizer": [1],
    "BitShift": [11],
    "BitwiseAnd": [18],
    "BitwiseNot": [18],
    "BitwiseOr": [18],
    "BitwiseXor": [18],
    "BlackmanWindow": [17],
    "Cast": [19],
    "CastLike": [19],
    "CastMap": [1],
    "CategoryMapper": [1],
    "Ceil": [13],
    "Celu": [12],
    "CenterCropPad": [18],
    "Clip": [13, 12, 11, 6],
    "Compress": [11],
    "Concat": [13],
    "ConcatFromSequence": [11],
    "Constant": [19],
    "ConstantOfShape": [20],
    "Conv": [11],
    "ConvInteger": [10],
    "ConvTranspose": [11],
    "Cos": [7],
    "Cosh": [9],
    "Col2Im": [18],
    "CumSum": [14],
    "DeformConv": [19],
    "DepthToSpace": [13],
    "DequantizeLinear": [19],
    "Det": [11],
    "DFT": [20, 17],
    "DictVectorizer": [1],
    "Div": [14],
    "Dropout": [13],
    "DynamicQuantizeLinear": [11],
    "Einsum": [12],
    "Elu": [6],
    "Equal": [19],
    "Erf": [13],
    "Exp": [13],
    "Expand": [13],
    "EyeLike": [9],
    "FeatureVectorizer": [1],
    "Flatten": [13],
    "Floor": [13],
    "GRU": [14],
    "Gather": [13],
    "GatherElements": [13],
    "GatherND": [13],
    "Gelu": [20],
    "Gemm": [13],
    "GlobalAveragePool": [1],
    "GlobalLpPool": [2],
    "GlobalMaxPool": [1],
    "Gradient": [1],
    "Greater": [13],
    "GreaterOrEqual": [16],
    "GridSample": [16],
    "GroupNormalization": [21, 18],
    "HammingWindow": [17],
    "HannWindow": [17],
    "HardSigmoid": [6],
    "Hardmax": [13],
    "HardSwish": [14],
    "Identity": [19],
    "If": [19],
    "Imputer": [1],
    "InstanceNormalization": [6],
    "IsInf": [20],
    "IsNaN": [20],
    "LayerNormalization": [17],
    "LRN": [13],
    "LSTM": [14],
    "LabelEncoder": [2],
    "LeakyRelu": [16],
    "Less": [13],
    "LessOrEqual": [16],
    "LinearClassifier": [1],
    "LinearRegressor": [1],
    "Log": [13],
    "LogSoftmax": [13],
    "Loop": [19],
    "LpNormalization": [1],
    "LpPool": [18],
    "MatMul": [13],
    "MatMulInteger": [10],
    "Max": [13],
    "MaxPool": [12],
    "MaxRoiPool": [1],
    "MaxUnpool": [11],
    "Mean": [13],
    "MeanVarianceNormalization": [13],
    "MelWeightMatrix": [17],
    "Min": [13],
    "Mish": [18],
    "Mod": [13],
    "Momentum": [1],
    "Mul": [14],
    "Multinomial": [7],
    "Neg": [13],
    "NegativeLogLikelihoodLoss": [13],
    "NonMaxSuppression": [11],
    "NonZero": [13],
    "Normalizer": [1],
    "Not": [1],
    "OneHot": [11],
    "OneHotEncoder": [1],
    "Optional": [15],
    "OptionalGetElement": [18],
    "OptionalHasElement": [18],
    "Or": [7],
    "PRelu": [16],
    "Pad": [19, 18, 13, 11, 2],
    "Pow": [15],
    "QLinearConv": [10],
    "QLinearMatMul": [10],
    "QuantizeLinear": [19],
    "RNN": [14],
    "RandomNormal": [1],
    "RandomNormalLike": [1],
    "RandomUniform": [1],
    "RandomUniformLike": [1],
    "Range": [11],
    "Reciprocal": [13],
    "ReduceL1": [18, 13],
    "ReduceL2": [18, 13],
    "ReduceLogSum": [18, 13],
    "ReduceLogSumExp": [18, 13],
    "ReduceMax": [20, 18, 13],
    "ReduceMean": [18, 13],
    "ReduceMin": [20, 18, 13],
    "ReduceProd": [18, 13],
    "ReduceSum": [13, 11],
    "ReduceSumSquare": [18, 13],
    "Relu": [14],
    "Reshape": [19],
    "Resize": [19, 18, 13, 11, 10],
    "ReverseSequence": [10],
    "RoiAlign": [16],
    "Round": [11],
    "SVMClassifier": [1],
    "SVMRegressor": [1],
    "Scaler": [1],
    "Scan": [19],
    "Scatter": [11],
    "ScatterElements": [18],
    "ScatterND": [18],
    "Selu": [6],
    "SequenceAt": [11],
    "SequenceConstruct": [11],
    "SequenceEmpty": [11],
    "SequenceErase": [11],
    "SequenceInsert": [11],
    "SequenceLength": [11],
    "SequenceMap": [17],
    "Shape": [19],
    "Shrink": [9],
    "Sigmoid": [13],
    "Sign": [13],
    "Sin": [7],
    "Sinh": [9],
    "Size": [19],
    "Slice": [13],
    "Softmax": [13, 11],
    "SoftmaxCrossEntropyLoss": [13],
    "Softplus": [1],
    "Softsign": [1],
    "SpaceToDepth": [13],
    "Split": [18, 13, 11],
    "SplitToSequence": [11],
    "Sqrt": [13],
    "Squeeze": [13, 11],
    "StringNormalizer": [10],
    "STFT": [17],
    "Sub": [14],
    "Sum": [13],
    "Tan": [7],
    "Tanh": [13],
    "TfIdfVectorizer": [9],
    "ThresholdedRelu": [10],
    "Tile": [13],
    "TopK": [11],
    "Transpose": [13],
    "Trilu": [14],
    "TreeEnsembleClassifier": [1],
    "TreeEnsembleRegressor": [1],
    "Unique": [11],
    "Unsqueeze": [13, 11],
    "Upsample": [9, 7],
    "Where": [16],
    "Xor": [7],
    "ZipMap": [1],
}

# Manual specification of attribute type.
special_attr_types = dict([("Cast.to", "type")])

# Manual specification of attribute order:
# The names in each tuple will be ordered in that sequence.
special_attr_order = {
    ("then_branch", "else_branch"),
}

# Special operation importing handlers.
special_op_handler = dict(
    [
        ("BatchNormalization", "ImportNodeBatchNormalization"),
        ("CategoryMapper", "ImportCategoryMapper"),
        ("Dropout", "ImportNodeDropout"),
        ("Cast", "ImportNodeCast"),
        ("MaxPool", "ImportNodeMaxPool"),
        ("Pad", "ImportNodePad"),
        ("Slice", "ImportNodeSlice"),
    ]
)

# Operations with custom assembly format (alphabetical order).
OpsWithCustomAssemblyFormat = [
    "Constant",
    "ConstantOfShape",
]

# Operations supporting canonicalization (alphabetical order).
OpsWithCanonicalizer = [
    "Add",
    "And",
    "Cast",
    "Constant",
    "DepthToSpace",
    "DequantizeLinear",
    "Div",
    "Dropout",
    "Equal",
    "GlobalAveragePool",
    "GlobalMaxPool",
    "Greater",
    "GRU",
    "Identity",
    "Less",
    "Loop",
    "LSTM",
    "Mul",
    "Or",
    "Pow",
    "Reshape",
    "Resize",
    "RNN",
    "Shape",
    "Size",
    "SoftmaxV11",
    "SpaceToDepth",
    "Squeeze",
    "SqueezeV11",
    "Sub",
    "Tile",
    "Transpose",
    "Unsqueeze",
    "UnsqueezeV11",
    "Where",
    "Xor",
]

# Operations with custom verifiers (alphabetical order).
OpsWithVerifier = [
    "Add",
    "And",
    "ArgMax",
    "ArgMin",
    "AveragePool",
    "BitShift",
    "BitwiseAnd",
    "BitwiseOr",
    "BitwiseXor",
    "CategoryMapper",
    "Compress",
    "Concat",
    "ConcatFromSequence",
    "ConstantOfShape",
    "Conv",
    "ConvTranspose",
    "DepthToSpace",
    "DequantizeLinear",
    "Div",
    "Einsum",
    "Equal",
    "Expand",
    "Flatten",
    "Gather",
    "GatherElements",
    "GatherND",
    "Gelu",
    "Greater",
    "GreaterOrEqual",
    "GroupNormalizationV18",
    "Hardmax",
    "If",
    "IsInf",
    "InstanceNormalization",
    "LayerNormalization",
    "Less",
    "LessOrEqual",
    "LogSoftmax",
    "Max",
    "MatMulInteger",
    "Mean",
    "Min",
    "Mod",
    "Mul",
    "NonMaxSuppression",
    "OneHot",
    "OneHotEncoder",
    "Optional",
    "OptionalGetElement",
    "OptionalHasElement",
    "Or",
    "PRelu",
    "Pad",
    "Pow",
    "RandomNormalLike",
    "Range",
    "Reshape",
    "Resize",
    "ReverseSequence",
    "RoiAlign",
    "ScatterElements",
    "ScatterND",
    "SequenceEmpty",
    "SequenceInsert",
    "Shape",
    "SpaceToDepth",
    "Split",
    "SplitToSequence",
    "Sub",
    "Sum",
    "TopK",
    "Unique",
    "Upsample",
    "Where",
    "Xor",
]

# Op with fold function
OpsWithFolder = ["Constant", "Squeeze", "SqueezeV11"]

# Op with ConstantLike trait
OpsWithConstantLike = ["Constant"]

# Op with Helper functions
# Here the functions are for data flow analysis.
OpsWithHelpers = {
    "Loop": """
    mlir::Operation::result_range v_final();
    mlir::Operation::result_range scan_outputs();
  """,
    "Scan": """
    mlir::Operation::operand_range getVInitial();
    mlir::Operation::result_range v_final();
    mlir::Operation::operand_range scan_inputs();
    mlir::Operation::result_range scan_outputs();
  """,
}

# Type inference are usually done with the type string for Op definition.
# This dictionary provides special code for type inference for some Ops.
# The type inference is used only in Builder before constant canonicalization.
OpsWithResultTypeInference = [
    "Constant",
    "Cast",
    "ConstantOfShape",
    "If",
    "Loop",
    "RandomNormal",
    "Scan",
]

# Add an Op in this list if the Op needs result type deduction which is required
# when writing declarative rewriting rules. Deduced type is always
# an UnrankedTensorType whose element type is the same as the first operand's
# element type.
#
# Currently, there are only two build methods generated:
#  - one with operands and attributes having a separate parameter, and
#  - one with operands and attributes having aggregated parameters.
custom_builder_unranked_ops_list = [
    "Abs",
    "Conv",
    "Exp",
    "Identity",
    "Neg",
    "Pad",
    "ReduceLogSum",
    "ReduceMaxV13",
    "ReduceMaxV18",
    "ReduceMax",
    "ReduceSum",
    "ReduceSumSquare",
    "ReduceSumV11",
    "Softmax",
    "Split",
    "SplitV13",
    "Sqrt",
    "Squeeze",
    "SqueezeV11",
    "Unsqueeze",
    "UnsqueezeV11",
]
# Custom builder op list for operations with broadcast; we can deduce the right
# output type, no need to leave it undef as in the above list.
# Ops must have two operands, not one, not three... And there shall be two.
# TODO: handle variadic ops omitted here: Max, Min, Min, Sum.
custom_builder_broadcast_to_same_type_ops_list = [
    "Add",
    "And",
    "Div",
    "Mul",
    "Or",
    "Pow",
    "Sub",
    "Xor",
]
custom_builder_broadcast_to_bool_ops_list = [
    "Equal",
    "Greater",
    "GreaterOrEqual",
    "Less",
    "LessOrEqual",
]
custom_builder_broadcast_ops_list = (
    custom_builder_broadcast_to_same_type_ops_list
    + custom_builder_broadcast_to_bool_ops_list
)
# Union of both.
custom_builder_ops_list = (
    custom_builder_unranked_ops_list + custom_builder_broadcast_ops_list
)

# A dictionary to add any special definition for an operation.
custom_definition_misc = dict(
    [
        (
            "Constant",
            """  let builders = [
  OpBuilder<(ins "Attribute":$sparse_value, "Attribute":$value), [{
   if (value) {
    auto tensorType = mlir::cast<TypedAttr>(value).getType();
    build($_builder, $_state, tensorType, sparse_value, value,
      FloatAttr(), ArrayAttr(), IntegerAttr(), ArrayAttr(), StringAttr(), ArrayAttr());
   } else {
    auto tensorType = mlir::cast<TypedAttr>(sparse_value).getType();
    build($_builder, $_state, tensorType, sparse_value, value,
      FloatAttr(), ArrayAttr(), IntegerAttr(), ArrayAttr(), StringAttr(), ArrayAttr());
   }
  }]>
  ];""",
        ),
        (
            "Cast",
            """   let builders = [
  OpBuilder<(ins "Value":$input, "IntegerAttr":$saturate, "TypeAttr":$to), [{
   auto resultType = mlir::UnrankedTensorType::get(to.getValue());
   build($_builder, $_state, resultType, input, saturate, to);
  }] >
  ];""",
        ),
    ]
)

# Get this order from TensorProto in https://github.com/onnx/onnx/blob/main/onnx/onnx.in.proto#L481.
# enum DataType {
#     UNDEFINED = 0;
#     // Basic types.
#     FLOAT = 1;   // float
#     UINT8 = 2;   // uint8_t
#     INT8 = 3;    // int8_t
#     UINT16 = 4;  // uint16_t
#     INT16 = 5;   // int16_t
#     INT32 = 6;   // int32_t
#     INT64 = 7;   // int64_t
#     STRING = 8;  // string
#     BOOL = 9;    // bool
#
#     // IEEE754 half-precision floating-point format (16 bits wide).
#     // This format has 1 sign bit, 5 exponent bits, and 10 mantissa bits.
#     FLOAT16 = 10;
#
#     DOUBLE = 11;
#     UINT32 = 12;
#     UINT64 = 13;
#     COMPLEX64 = 14;     // complex with float32 real and imaginary components
#     COMPLEX128 = 15;    // complex with float64 real and imaginary components
#
#     // Non-IEEE floating-point format based on IEEE754 single-precision
#     // floating-point number truncated to 16 bits.
#     // This format has 1 sign bit, 8 exponent bits, and 7 mantissa bits.
#     BFLOAT16 = 16;
#
#     // Non-IEEE floating-point format based on papers
#     // FP8 Formats for Deep Learning, https://arxiv.org/abs/2209.05433,
#     // 8-bit Numerical Formats For Deep Neural Networks, https://arxiv.org/pdf/2206.02915.pdf.
#     // Operators supported FP8 are Cast, CastLike, QuantizeLinear, DequantizeLinear.
#     // The computation usually happens inside a block quantize / dequantize
#     // fused by the runtime.
#     FLOAT8E4M3FN = 17;    // float 8, mostly used for coefficients, supports nan, not inf
#     FLOAT8E4M3FNUZ = 18;  // float 8, mostly used for coefficients, supports nan, not inf, no negative zero
#     FLOAT8E5M2 = 19;      // follows IEEE 754, supports nan, inf, mostly used for gradients
#     FLOAT8E5M2FNUZ = 20;  // follows IEEE 754, supports nan, inf, mostly used for gradients, no negative zero
#
#     // Future extensions go here.
#   }
onnx_types = (
    "undefined",
    "float",
    "uint8",
    "int8",
    "uint16",
    "int16",
    "int32",
    "int64",
    "string",
    "bool",
    "float16",
    "double",
    "uint32",
    "uint64",
    "complex64",
    "complex128",
    "bfloat16",
    "float8e4m3fn",
    "float8e4m3fnuz",
    "float8e5m2",
    "float8e5m2fnuz",
)
tblgen_types = (
    "BF16",
    "F32",
    "AnyUI8",
    "AnyI8",
    "AnyUI16",
    "AnyI16",
    "AnyI32",
    "AnyI64",
    "StringType",
    "AnyI1",
    "F16",
    "F64",
    "AnyUI32",
    "AnyUI64",
    "Complex<F32>",
    "Complex<F64>",
    "BF16",
    "F8E4M3FN",
    "F8E4M3FNUZ",
    "F8E5M2",
    "F8E5M2FNUZ",
)

# Maximum count for actual type. Number more than MAX_NUM_TYPES will be used to encode
# the mapping method. MAX_NUM_TYPES should be greater than the length of onnx_types
# This value has to be kept the same with MAX_TYPE in FrontendDialectTransformer.cpp
MAX_NUM_TYPES = 30


# Attribute names are ordered alphabetically except for the
# manually specified special orderings in special_attr_order.
def order_attr_names(attrNames):
    attrNames = sorted(attrNames)
    for namesOrder in special_attr_order:
        # If attrNames includes all the namesOrder names, then reorder
        # those names in attrNames to their order in namesOrder,
        if set(namesOrder).issubset(attrNames):
            # The namesIndexes are where the namesOrder names appear in attrNames.
            namesIndexes = (attrNames.index(name) for name in namesOrder)
            # Write the namesOrder names into those indexes in the correct order.
            for name, index in zip(namesOrder, sorted(namesIndexes)):
                attrNames[index] = name
    return attrNames


def should_render_domain(domain):  # type: (Text) -> bool
    return True


def display_attr_type(v):  # type: (OpSchema.AttrType) -> Text
    assert isinstance(v, OpSchema.AttrType)
    s = Text(v)
    s = s[s.rfind(".") + 1 :].lower()
    if s[-1] == "s":
        s = "list of " + s
    return s


# ONNX allows input and output to have the same name. But MLIR does not
# When the conflict occurs, add 'out_' to the output name
def get_unique_output_name(schema, name):
    for input in schema.inputs:
        if input.name == name:
            return "out_" + name
    return name


def onnx_attr_type_to_mlir_attr_type(t):
    onnx_attr_type = Text(t)
    onnx_attr_type = onnx_attr_type[onnx_attr_type.rfind(".") + 1 :].lower()

    if onnx_attr_type == "int":
        mlir_attr_type = "SI64Attr"
    elif onnx_attr_type == "float":
        mlir_attr_type = "F32Attr"
    elif onnx_attr_type == "ints":
        mlir_attr_type = "I64ArrayAttr"
    elif onnx_attr_type == "floats":
        mlir_attr_type = "F32ArrayAttr"
    elif onnx_attr_type == "string":
        mlir_attr_type = "StrAttr"
    elif onnx_attr_type == "strings":
        mlir_attr_type = "StrArrayAttr"
    elif onnx_attr_type in {"type", "type_proto"}:
        # 'type' is the attribute type used in special_attr_types,
        # 'type_proto' is Optional op's type attribute's type
        mlir_attr_type = "TypeAttr"
    else:
        mlir_attr_type = "AnyAttr"
    # TODO: tensor and sparse tensor.
    return mlir_attr_type


# TODO: any better way to do this.
def tblgen_attr_type_to_cpp_type(t):
    if "I64Attr" in t:
        cpp_type = "IntegerAttr"
    elif "F32Attr" in t:
        cpp_type = "FloatAttr"
    elif "I64ArrayAttr" in t or "F32ArrayAttr" in t:
        cpp_type = "ArrayAttr"
    elif "StrAttr" in t:
        cpp_type = "StringAttr"
    elif "strings" in t:
        cpp_type = "ArrayAttr"
    else:
        cpp_type = "Attribute"
    return cpp_type


def tblgen_operand_type_to_cpp_type(op_type):
    if op_type.startswith("Variadic"):
        mytype = "ValueRange"
    else:
        mytype = "Value"
    return mytype


def np_type_to_tblgen_attr_type(tstr):
    index = -1
    for i in range(len(onnx_types)):
        if onnx_types[i] in tstr:
            index = i
            break
    if index == -1:
        return None
    else:
        return tblgen_types[i]


def get_tblgen_type_index(type_str):
    return tblgen_types.index(type_str)


# The possible data structures are tensor, map and seq(tensor()).
def get_data_structure_element(allowed_type_str):
    structure_list = ["tensor", "seq", "map"]
    for structure in structure_list:
        if allowed_type_str.startswith(structure):
            element = allowed_type_str.replace(structure + "(", "", 1).replace(
                ")", "", 1
            )
            return (structure, element)
    return (None, None)


def get_allowed_elem_types(schema, input):
    # allowed_types_str = None
    # return allowed_types_str
    # TODO: enable type constraints.
    if input.type_str:
        tstr = input.type_str
        structure, element = get_data_structure_element(tstr)
        # In case the type is directly specified.
        if structure and element:
            t = np_type_to_tblgen_attr_type(element)
            if t == None:
                return allowed_structure, None
            else:
                return structure, [t]
    else:
        return None
    if schema.type_constraints:
        for type_constraint in schema.type_constraints:
            if type_constraint.type_param_str != tstr:
                continue
            allowed_type_list = []
            allowedTypes = type_constraint.allowed_type_strs
            allowed_structure = None
            for allowedType in allowedTypes:
                structure, element = get_data_structure_element(allowedType)
                if structure == None or element == None:
                    return None, None

                if allowed_structure != None and allowed_structure != structure:
                    return None, None
                allowed_structure = structure
                t = np_type_to_tblgen_attr_type(element)
                if t == None:
                    return allowed_structure, None
                if not t in allowed_type_list:
                    allowed_type_list.append(t)

            return allowed_structure, allowed_type_list

    return None, None


def inc_indent(indent=None):
    return "" if indent is None else indent + " " * 2


def dec_indent(indent):
    return indent[:-2]


def join_args(args):
    return ", ".join(args)


def get_operands_or_results(schema, type_str_dict, op_name, is_input):
    value_list = schema.inputs if is_input else schema.outputs
    if not value_list:
        return OrderedDict()

    def any_type_of(types):
        assert isinstance(types, list)
        if len(types) == 1:
            return types[0]
        else:
            return "AnyTypeOf<[{}]>".format(", ".join(types))

    name_to_types = OrderedDict()
    for i, value in enumerate(value_list):
        str_types = get_onnx_mlir_types(schema, type_str_dict, value)

        # In case the type string is used more than once.
        types = str_types.copy()

        # No need to add AnyMemRef type. Keep the code in case.
        # types.append("AnyMemRef")

        if OpSchema.FormalParameterOption.Optional == value.option:
            types.append("NoneType")

        if OpSchema.FormalParameterOption.Variadic == value.option:
            if value.is_homogeneous:
                types = ["Variadic<{}>".format(any_type_of(types))]
            else:
                # TODO handle(variadic, heterogeneous) "
                types = ["Variadic<{}>".format(any_type_of(types))]
                sys.stderr.write(
                    "warning: (variadic, heterogeneous) for "
                    + schema.name
                    + " "
                    + value.name
                    + "\n"
                )

        # Since output name can coincide with that of an input, we explicitly
        # append a suffix "_out" to such names for disambiguation.
        if is_input:
            value_name = value.name
        else:
            value_name = get_unique_output_name(schema, value.name)

        name_to_types[value_name] = any_type_of(types)
    return name_to_types


def get_attrs(schema):
    def get_attr_type_optional(attr_type):
        return "OptionalAttr<{}>".format(onnx_attr_type_to_mlir_attr_type(attr_type))

    def get_attr_type_with_default(attr_type, attr_default):
        if attr_type == OpSchema.AttrType.STRING:
            return 'DefaultValuedStrAttr<{}, "{}">'.format(
                onnx_attr_type_to_mlir_attr_type(attr_type), attr_default
            )
        else:
            return 'DefaultValuedAttr<{}, "{}">'.format(
                onnx_attr_type_to_mlir_attr_type(attr_type), attr_default
            )

    if not schema.attributes:
        return OrderedDict()

    name_to_type = OrderedDict()
    for _, attr in sorted(schema.attributes.items()):
        if attr.type == OpSchema.AttrType.GRAPH:
            continue

        qualified_attr_name = "{}.{}".format(schema.name, attr.name)
        if qualified_attr_name in special_attr_types:
            name_to_type[attr.name] = onnx_attr_type_to_mlir_attr_type(
                special_attr_types[qualified_attr_name]
            )
        # Option holds either required or default value.
        elif attr.required:
            name_to_type[attr.name] = onnx_attr_type_to_mlir_attr_type(attr.type)
        elif attr.default_value.name:

            def format_value(value):  # type: (Any) -> Text
                if isinstance(value, float):
                    formatted = str(np.round(value, 5))
                    # Use default formatting, unless too long.
                    if len(formatted) > 10:
                        formatted = str("({:e})".format(value))
                    return formatted
                elif isinstance(value, (bytes, bytearray)) and sys.version_info[0] == 3:
                    return str(value.decode("utf-8"))
                return str(value)

            default_value = helper.get_attribute_value(attr.default_value)
            if isinstance(default_value, list):
                default_value = [format_value(val) for val in default_value]
                default_value_str = "{}".format(default_value)
                default_value_str = default_value_str.replace("[", "{", 1)
                default_value_str = default_value_str.replace("]", "}", 1)
                if Text(attr.type) == "AttrType.STRINGS":
                    default_value_str = default_value_str.replace("'", '\\"')
                else:
                    default_value_str = default_value_str.replace("'", "")
            else:
                default_value = format_value(default_value)
                default_value_str = default_value

            name_to_type[attr.name] = get_attr_type_with_default(
                attr.type, default_value_str
            )
        else:
            name_to_type[attr.name] = get_attr_type_optional(attr.type)
    return name_to_type


def get_numberof_list(my_list):
    expected_num = len(my_list)
    for element in my_list:
        if OpSchema.FormalParameterOption.Variadic == element.option:
            expected_num = -1
    return expected_num


def get_output_type_mapping(schema):
    mapping = []
    for output in schema.outputs:
        # If only one type is allowed, just set that.
        structure, allowed_elem_types = get_allowed_elem_types(schema, output)
        if allowed_elem_types != None and len(allowed_elem_types) == 1:
            mapping.append(str(get_tblgen_type_index(allowed_elem_types[0])))
            continue

        # Map the type string.
        if output.type_str:
            tstr = output.type_str
            found = False
            for i, input in enumerate(schema.inputs):
                if input.type_str and input.type_str == tstr:
                    mapping.append(str(i + MAX_NUM_TYPES))
                    found = True
                    break
            if found:
                continue

        # Unknown output type.
        mapping.append(str(-1))

    return mapping


def get_numberof_inout(s, indent, schema):
    expected_num_operands = get_numberof_list(schema.inputs)
    indent = inc_indent(indent)
    s += indent + "static int getNumberOfOperands() {\n"
    indent = inc_indent(indent)
    s += indent + "return {};\n".format(expected_num_operands)
    indent = dec_indent(indent)
    s += indent + "}\n"

    expected_num_results = get_numberof_list(schema.outputs)
    s += indent + "static int getNumberOfResults() {\n"
    indent = inc_indent(indent)
    s += indent + "return {};\n".format(expected_num_results)
    indent = dec_indent(indent)
    s += indent + "}\n"

    s += indent + "static std::vector<int> getTypeMap() {\n"
    mapping = get_output_type_mapping(schema)
    indent = inc_indent(indent)
    s += indent + "return {" + ",".join(mapping) + "};\n"
    indent = dec_indent(indent)
    s += indent + "}\n"

    return s


def get_promotable_const_operands_func(s, indent, const_operands_name_to_idx):
    cpp_name_to_idx_literal = (
        "{"
        + ", ".join(
            [
                '{{"{}", {}}}'.format(*name_to_idx)
                for name_to_idx in const_operands_name_to_idx
            ]
        )
        + "}"
    )

    # s += indent + "let extraClassDeclaration = [{\n"
    indent = inc_indent(indent)
    s += indent + "std::map<std::string, size_t> promotableConstOperands() {\n"
    indent = inc_indent(indent)
    s += indent + "return {};\n".format(cpp_name_to_idx_literal)
    indent = dec_indent(indent)
    s += indent + "}\n"
    # indent = dec_indent(indent)
    # s += indent + "}];\n"

    return s


def parse_type_str(allowedType):
    # AnyI may be used for uint because the onnx_mlir is not generating uint output.
    # This will be fixed later and UI will be replace AnyI.
    onnx_to_mlir_type_dict = {
        "(": "<[",
        ")": "]>",
        "tensor": "TensorOf",
        "seq": "SeqOf",
        "map": "TupleOf",
        "bool": "I1",
        "uint8": "UI8",
        "uint16": "UI16",
        "uint32": "UI32",
        "uint64": "UI64",
        "int8": "I8",
        "int16": "I16",
        "int32": "I32",
        "int64": "I64",
        "double": "F64",
        "float": "F32",
        "float16": "F16",
        "bfloat16": "BF16",
        "float8e4m3fn": "F8E4M3FN",
        "float8e4m3fnuz": "F8E4M3FNUZ",
        "float8e5m2": "F8E5M2",
        "float8e5m2fnuz": "F8E5M2FNUZ",
        "complex64": "Complex<F32>",
        "complex128": "Complex<F64>",
        "string": "StringType",
    }

    # Optional(...) always appears outermost.
    if allowedType.find("optional") == 0:
        allowedType = allowedType.replace("optional(", "OptOf<", 1)
        allowedType = allowedType[:-1] + ">"

    # Apply substitutions in decreasing order of key-length, so that float16
    # is replaced before float, and uint16 is replaced before int16, etc.
    mapping = list(onnx_to_mlir_type_dict.items())
    mapping.sort(key=lambda pair: len(pair[0]), reverse=True)
    for key, item in mapping:
        allowedType = allowedType.replace(key, item)
    return allowedType


def parse_a_type_constraint(constraint):
    allowedTypes = constraint.allowed_type_strs
    mlirTypes = []
    for allowedType in allowedTypes:
        mlirType = parse_type_str(allowedType)
        mlirTypes.append(mlirType)

    # Remove redundant and sort.
    # However onnx keeps a consistently meaningful order
    # There is no redundancy as long as each onnx type is mapped uniquely
    # mlirTypes = sorted(list(set(mlirTypes)))

    return mlirTypes


def parse_type_constraints(schema):
    type_str_dict = dict()
    for type_constraint in schema.type_constraints:
        type_str_dict[type_constraint.type_param_str] = parse_a_type_constraint(
            type_constraint
        )
    return type_str_dict


def get_onnx_mlir_types(schema, type_str_dict, input):
    if input.type_str:
        if not input.type_str in type_str_dict:
            # Some arguments use type description directly
            # instead of constraint.
            type_str = parse_type_str(input.type_str)
            return [type_str]
        else:
            return type_str_dict[input.type_str]
    else:
        print("No type_str ", schema.name)
        return []


# Generate extra class declaration for shape helper.
def gen_shape_helper_code(s, indent, opName):
    # Print getShapeHelper.
    indent = inc_indent(indent)
    s += (
        indent
        + "onnx_mlir::ONNXOpShapeHelper * $cppClass::getShapeHelper(mlir::Operation *op, llvm::ArrayRef<mlir::Value> oper, \n"
    )
    indent = inc_indent(indent)
    indent = inc_indent(indent)
    s += (
        indent
        + "onnx_mlir::IndexExprBuilder *ieb, onnx_mlir::IndexExprScope *scope) {\n"
    )
    indent = dec_indent(indent)
    s += (
        indent
        + "onnx_mlir::ONNXOpShapeHelper *sh = new onnx_mlir::ONNX{0}OpShapeHelper(op, oper, ieb, scope);\n".format(
            opName
        )
    )
    s += indent + 'assert(sh && "failed to allocate shape helper");\n'
    s += indent + "return sh;\n"
    indent = dec_indent(indent)
    s += indent + "}\n"
    return s


def gen_op_name(schema, with_version):
    if with_version:
        return schema.name + "V" + str(schema.since_version)
    return schema.name


# Generate entry for a given operation given by opName (from schema).
def gen_op_def(schema, with_version=False):
    indent = inc_indent()
    opName = gen_op_name(schema, with_version)
    s = 'def ONNX{0}Op:ONNX_Op<"{0}",\n'.format(opName)

    regions = OrderedDict()
    for name in order_attr_names(schema.attributes.keys()):
        attr = schema.attributes[name]
        if attr.type == OpSchema.AttrType.GRAPH:
            if attr.required:
                regions[attr.name] = "SizedRegion<1>"
            else:
                regions[attr.name] = "AnyRegion"

    # Generate decl for op traits.
    traits = ["Pure"]

    # Generate ConstantLike traits.
    if opName in OpsWithConstantLike:
        traits.append("ConstantLike")

    # OpsWithShapeInference:
    # Now the ShapeInference traits are added to all operation.
    # Dummy implementations are added to ONNXOps.cpp.
    # Error will be report if these operations are encountered at runtime.
    traits.append("DeclareOpInterfaceMethods<ShapeInferenceOpInterface>")
    traits.append("DeclareOpInterfaceMethods<ShapeHelperOpInterface>")
    if opName in OpsWithResultTypeInference:
        traits.append("DeclareOpInterfaceMethods<ResultTypeInferenceOpInterface>")
    if len(regions):
        traits.append('OpInterface<"HasOnnxSubgraphOpInterface">')
    s += inc_indent(indent) + "[{}]> {{\n".format(join_args(traits))

    indent = inc_indent(indent)

    # Generate decl for custom assembly format.
    if opName in OpsWithCustomAssemblyFormat:
        s += indent + "let hasCustomAssemblyFormat = 1;\n"

    # Generate decl for canonicalizer.
    if opName in OpsWithCanonicalizer:
        s += indent + "let hasCanonicalizer = 1;\n"

    # Generate decl for summary.
    s += indent + 'let summary = "ONNX {} operation";\n'.format(schema.name)

    # Generate description.
    s += indent + "let description = [{\n"
    if schema.doc:
        lines = schema.doc.lstrip().splitlines()
        for line in lines:
            escaped_line = line.replace('"', '\\"').replace("}]", "\\}\\]")
            # Description does not really need to have "" for each line.
            # s += indent + '"{}"\n'.format(escaped_line)
            s += indent + "{}\n".format(escaped_line)
    s += indent + "}];\n"

    # Handle the type constraint for input and output.
    # Parse type constraint into onnx-mlir type string list.
    type_str_dict = parse_type_constraints(schema)

    ###########################################
    # Generate ins (consisting of operands and attributes).
    ins = get_operands_or_results(schema, type_str_dict, opName, is_input=True)
    ins.update(get_attrs(schema))

    ins_strs = ["{1}:${0}".format(*i) for i in ins.items()]
    s += indent + "let arguments = (ins {});\n".format(
        (",\n" + inc_indent(indent)).join(ins_strs)
    )

    ###########################################
    # Generate outs (operation results).
    outs = get_operands_or_results(schema, type_str_dict, opName, is_input=False)
    outs_strs = ["{1}:${0}".format(*i) for i in outs.items()]
    s += indent + "let results = (outs {});\n".format(
        (",\n" + inc_indent(indent)).join(outs_strs)
    )

    regions_strs = ["{1}:${0}".format(*i) for i in regions.items()]

    if len(regions):
        s += indent + "let regions = (region {});\n".format(
            (",\n" + inc_indent(indent)).join(regions_strs)
        )

    # custom_builder_broadcast_ops_list

    ###########################################
    # Add custom builders.
    # Use element type of the first operand to construct an UnrankedTensorType
    # for the output.
    if opName in custom_builder_ops_list:
        if len(ins) == 0:
            raise RuntimeWarning(
                "warning: not generate custom build methods for "
                + schema.name
                + " since it does not have operands."
            )

        r = ""  # r is the resultType, use it with r.format(*operands, indent=indent)
        if opName in custom_builder_broadcast_ops_list:
            numOperands = 2
            r += "{indent}auto lhsTy = {0}.getType();\n"
            r += "{indent}auto rhsTy = {1}.getType();\n"
            if opName in custom_builder_broadcast_to_bool_ops_list:
                r += "{indent}auto elTy = $_builder.getI1Type();\n"
                elTy = "elTy"
            else:
                elTy = ""
            r += (
                "{indent}auto resultType = getBroadcastedRankedType(lhsTy, rhsTy"
                + (", " + elTy if elTy else "")
                + ");\n"
            )
            r += "{indent}auto shapedType = mlir::dyn_cast_or_null<ShapedType>(resultType);\n"
            r += "{indent}if (!shapedType || !shapedType.hasStaticShape())\n"
            r += (
                "{indent}  resultType = UnrankedTensorType::get("
                + (elTy if elTy else "mlir::cast<ShapedType>(lhsTy).getElementType()")
                + ");\n"
            )
        else:
            numOperands = 1
            r += (
                "{indent}auto resultType = UnrankedTensorType::get("
                + "mlir::cast<ShapedType>({0}.getType()).getElementType());\n"
            )
        resultType = r

        s += indent + "let builders = [\n"
        # Custom builders with operands and attributes having a separate parameter.
        # E.g. OpBuilder<(ins "Value":$X, "Value":$Y, "Attribute":$A), [{}]>
        indent = inc_indent(indent)
        s += indent + "OpBuilder<(ins "
        operands_dict = get_operands_or_results(
            schema, type_str_dict, opName, is_input=True
        )
        attrs_dict = get_attrs(schema)
        s += ", ".join(
            '"{}":${}'.format(tblgen_operand_type_to_cpp_type(ty), name)
            for name, ty in operands_dict.items()
        )
        if operands_dict and attrs_dict:
            s += ", "
        s += ", ".join(
            '"{}":${}'.format(tblgen_attr_type_to_cpp_type(ty), name)
            for name, ty in attrs_dict.items()
        )
        s += "), [{\n"
        indent = inc_indent(indent)
        # Get output type from first operand's type.
        operands = operands_dict.keys()
        s += resultType.format(*operands, indent=indent)
        s += indent + "build($_builder, $_state, resultType"
        for name, _ in ins.items():
            s += ", " + name
        s += ");\n"
        indent = dec_indent(indent)
        s += indent + "}]>,\n"

        # Custom builders with all operands and attributes having aggregate parameters.
        # E.g. OpBuilder<(ins "ValueRange operands,
        #    ArrayRef<NamedAttribute> attributes", [{}]>'
        s += (
            indent
            + "OpBuilder<(ins "
            + '"ValueRange":$operands, "ArrayRef<NamedAttribute>":$attributes), [{\n'
        )
        indent = inc_indent(indent)
        operands = (f"operands[{i}]" for i in range(numOperands))
        s += resultType.format(*operands, indent=indent)
        s += indent + "build($_builder, $_state, {resultType}, operands, attributes);\n"
        indent = dec_indent(indent)
        s += indent + "}]>"

        s += "\n" + indent + "];\n"

    ###########################################
    # Generate extraClassDeclaration.
    s += indent + "let extraClassDeclaration = [{\n"
    # indent = inc_indent(indent)

    # Generate input/output number and output type mapping
    s = get_numberof_inout(s, indent, schema)

    if opName in OpsWithHelpers:
        s += OpsWithHelpers[opName]

    if len(regions):
        s += indent + "int64_t getSubgraphRegionIdx(const std::string& name) {\n"
        indent = inc_indent(indent)
        for idx, region_name in enumerate(regions.keys()):
            s += indent + 'if (name == "{}") return {};\n'.format(region_name, idx)
        s += (
            indent
            + 'llvm_unreachable("region with the specified name does not exist");\n'
        )
        indent = dec_indent(indent)
        s += indent + "}\n"
    s += indent + "}];\n"

    ###########################################
    # Generate extraClassDefinition.
    s += indent + "let extraClassDefinition = [{\n"

    # Generate shape helper code
    s = gen_shape_helper_code(s, indent, opName)

    s += indent + "}];\n"

    if opName in custom_definition_misc:
        s += custom_definition_misc[opName] + "\n"

    ###########################################
    # Generate decl for verifier.
    if opName in OpsWithVerifier:
        s += indent + "let hasVerifier = 1;\n"
    if opName in OpsWithFolder:
        s += indent + "let hasFolder = 1;\n"
    s += "}\n\n"
    return s


def gen_op_versions(file):
    indent = inc_indent()
    s = ""
    for key, item in version_dict.items():
        s += indent + 'op_dialect_version_map_["' + key + '"] = '
        s += "{" + "{}".format(", ".join(str(x) for x in item)) + "};\n"
    file.write(s)


"""
special cases:
* Split: attr split default value: sizeof(output1) namely 1
* Conv: attr dilations default value is {num_dim of first input - 2, 1}
* Conv: attr kernel_shape type is ints
* Transpose: attr perm default value is {} empty int list
"""


def gen_op_importer(schema, file, with_version=False):
    indent = inc_indent()
    if with_version:
        opName = schema.name + "V" + str(schema.since_version)
    else:
        opName = schema.name
    s = indent + 'import_handler_map_["' + opName + '"] = \n '

    expected_num_operands = len(schema.inputs)
    expected_num_results = len(schema.outputs)
    for input in schema.inputs:
        if OpSchema.FormalParameterOption.Variadic == input.option:
            expected_num_operands = -1
    for output in schema.outputs:
        if OpSchema.FormalParameterOption.Variadic == output.option:
            expected_num_results = -1

    # Only support special op handler for the op without version.
    if with_version:
        handler_func = "buildOperation<mlir::ONNX{}Op>".format(opName)
    else:
        handler_func = special_op_handler.get(
            schema.name, "buildOperation<mlir::ONNX{}Op>".format(opName)
        )

    # Special handlers currently require expected num operands/results to be specified.
    # TODO: remove special handlers.
    args = ["node"]
    """
    if expected_num_operands != -1 or expected_num_results != -1 or "buildOperation" not in handler_func:
        args.append(
            "/* expected_num_operands = */ {}".format(expected_num_operands))
        args.append(
            '/* expected_num_results = */ {}'.format(expected_num_results))
    """
    s += inc_indent(indent) + "&onnx_mlir::detail::FrontendGenImpl::"
    s += handler_func + ";\n"

    file.write(s)


def build_operator_schemas():
    # domain -> support level -> name -> [schema]
    index = defaultdict(
        lambda: defaultdict(lambda: defaultdict(list))
    )  # type: Dict[Text, Dict[int, Dict[Text, List[OpSchema]]]]
    for schema in defs.get_all_schemas_with_history():
        index[schema.domain][int(schema.support_level)][schema.name].append(schema)

    # Preprocess the Operator Schemas:
    # [(domain, [(support_level, [(schema name, current schema, all versions schemas)])])]
    operator_schemas = (
        list()
    )  # type: List[Tuple[Text, List[Tuple[int, List[Tuple[Text, OpSchema, List[OpSchema]]]]]]]
    existing_ops = set()  # type: Set[Text]
    for domain, _support_map in sorted(index.items()):
        if not should_render_domain(domain):
            continue
        processed_support_map = list()
        for _support, _name_map in sorted(_support_map.items()):
            processed_name_map = list()
            for n, unsorted_versions in sorted(_name_map.items()):
                versions = sorted(unsorted_versions, key=lambda s: s.since_version)
                schema = versions[-1]
                if schema.name in existing_ops:
                    continue

                if check_operation_version:
                    # Generate operation of the latest version of your onnx.
                    existing_ops.add(schema.name)
                    processed_name_map.append((n, schema, versions))

                    # Add checks against version_dict.
                    if schema.name not in version_dict:
                        print(
                            "Check-operation-version: Operation {} is new  with version {}".format(
                                schema.name, schema.since_version
                            )
                        )
                    elif schema.since_version > version_dict[schema.name][0]:
                        print(
                            "Check-operation-version: Operation {}".format(schema.name)
                            + " has a newer version {} over old version {}".format(
                                schema.since_version, version_dict[schema.name][0]
                            )
                        )
                else:
                    # Generate operation according to the version in version_dict.
                    if schema.name not in version_dict:
                        continue
                    found = False
                    v_counter = 0
                    for schema in reversed(versions):
                        # Check the version number against the version_dict.
                        specified_version = version_dict[schema.name][v_counter]
                        if schema.since_version == specified_version:
                            existing_ops.add(schema.name)
                            processed_name_map.append((n, schema, versions))
                            found = True
                            v_counter += 1
                            if len(version_dict[schema.name]) == v_counter:
                                break
                    if not found:
                        print(
                            "Your onnx installation may be too old. "
                            "The desired version for operation {} is not found.".format(
                                schema.name
                            )
                        )
                        sys.exit()
            processed_support_map.append((_support, processed_name_map))
        operator_schemas.append((domain, processed_support_map))
    return operator_schemas


def main(args):  # type: (Type[Args]) -> None
    if list_operation_version:
        pprint.pprint(version_dict)
        return

    curr_utc_time = datetime.datetime.now(datetime.timezone.utc).strftime(
        "%m/%d/%Y, %H:%M:%S"
    )
    autogen_warning = (
        "//********************************************************\n"
        "//   Do not modify this file directly.\n"
        "//   This file is automatically generated via script.\n"
        "//   Details can be found in docs/ImportONNXDefs.md .\n"
        "//********************************************************\n\n"
    )
    autogen_warning = autogen_warning.format(curr_utc_time)

    op_def = args.op_def
    op_def.write(autogen_warning)

    op_importer = args.op_importer
    op_importer.write(autogen_warning)
    gen_op_versions(op_importer)

    new_version_dict = dict()
    for domain, support_map in build_operator_schemas():
        for _, name_map in support_map:
            # Generate Op with version number if not the latest version.
            previous_name = ""
            for op_type, schema, versions in name_map:
                new_version_dict[schema.name] = [schema.since_version]
                if not check_operation_version:
                    with_version = previous_name == schema.name
                    gen_op_importer(schema, op_importer, with_version)
                    r = gen_op_def(schema, with_version)
                    op_def.write(r)
                    previous_name = schema.name

    if check_operation_version:
        for key in version_dict:
            if not key in new_version_dict:
                print("op {} is not in the version".format(key))
            # Assume the top version will be upgraded to the latest version.
            # The existing extra version (from index 1) will be kept.
            for x in version_dict[key][1:]:
                new_version_dict[key].append(x)
        pprint.pprint(new_version_dict)


if __name__ == "__main__":
    curr_dir = os.path.dirname(os.path.realpath(__file__))

    class Args(object):
        dry_run = args.dry_run_onnx_ops or args.dry_run_op_build_table

        # If either dry_run_onnx_ops or dry_run_op_build_table is true, then treat
        # both of them as true. Otherwise, one of them runs as a dry-run and one
        # of them runs as a real run creating unnecessary artifacts in the wrong
        # locations in the build tree.
        if dry_run:
            op_def = StringIO()
            op_importer = StringIO()
        else:
            op_def_file_path = os.path.join(curr_dir, "ONNXOps.td.inc")
            op_def = io.open(op_def_file_path, "w", newline="")
            op_importer_file_path = os.path.join(curr_dir, "OpBuildTable.inc")
            op_importer = io.open(op_importer_file_path, "w", newline="")

    main(Args)

    # This is based on diff.py from llvm-project (llvm\utils\lit\lit\builtin_commands\diff.py).
    # On Windows, by default, stdout uses \r\n for newlines, however, all the
    # files we compare against use \n. This piece of code forces the windows stdout
    # to use \n for newlines.
    if sys.platform == "win32":
        if hasattr(sys.stdout, "buffer"):
            # python 3
            sys.stdout = io.TextIOWrapper(sys.stdout.buffer, newline="\n")
        else:
            # python 2.7
            import msvcrt

            msvcrt.setmode(sys.stdout.fileno(), os.O_BINARY)

    # Only output the generated values for the specifically requested dry run.
    if args.dry_run_onnx_ops:
        sys.stdout.write(Args.op_def.getvalue())
    if args.dry_run_op_build_table:
        sys.stdout.write(Args.op_importer.getvalue())