onnx--onnx/onnx/reference/ops/_op_list.py

# Copyright (c) ONNX Project Contributors

# SPDX-License-Identifier: Apache-2.0
"""Every class imported in this module defines an implementation of
an operator of the main domain. Any class name uses `_` to specify a
version defined in a specific opset. The class name without `_`
defines the current implementation. If an operator has no class
with `_`, it means the implementation is valid for every opset.
The operator may have been updated to support more types but that
did not change the implementation.
"""

from __future__ import annotations

__all__ = [
    "load_op",
    "Abs",
    "Acos",
    "Acosh",
    "Add",
    "AffineGrid",
    "And",
    "ArgMax_1",
    "ArgMax_12",
    "ArgMin_1",
    "ArgMin_12",
    "Asin",
    "Asinh",
    "Atan",
    "Atanh",
    "Attention",
    "AttributeHasValue",
    "AveragePool_1",
    "AveragePool_7",
    "AveragePool_11",
    "AveragePool_19",
    "BatchNormalization_6",
    "BatchNormalization_9",
    "BatchNormalization_14",
    "Bernoulli",
    "BitCast",
    "BitShift",
    "BitwiseAnd",
    "BitwiseNot",
    "BitwiseOr",
    "BitwiseXor",
    "BlackmanWindow",
    "Cast_1",
    "Cast_19",
    "Cast_24",
    "CastLike_15",
    "CastLike_19",
    "CausalConvWithState",
    "Ceil",
    "Celu",
    "CenterCropPad",
    "Clip_6",
    "Clip_11",
    "Col2Im",
    "Compress",
    "Concat",
    "ConcatFromSequence",
    "Constant_1",
    "Constant_9",
    "Constant_11",
    "Constant_12",
    "ConstantOfShape",
    "Conv",
    "ConvInteger",
    "ConvTranspose",
    "Cos",
    "Cosh",
    "CumProd",
    "CumSum",
    "DeformConv",
    "DepthToSpace",
    "DequantizeLinear_19",
    "DequantizeLinear_21",
    "DequantizeLinear_23",
    "DequantizeLinear_25",
    "Det",
    "DFT_17",
    "DFT_20",
    "Div",
    "Dropout_7",
    "Dropout_12",
    "DynamicQuantizeLinear",
    "Einsum",
    "Elu",
    "Equal",
    "Erf",
    "Exp",
    "Expand",
    "EyeLike",
    "Flatten",
    "Floor",
    "Gather",
    "GatherElements",
    "GatherND",
    "Gemm_6",
    "Gemm_7",
    "GlobalAveragePool",
    "GlobalMaxPool",
    "Greater",
    "GreaterOrEqual",
    "GridSample",
    "GRU",
    "HammingWindow",
    "HannWindow",
    "HardSigmoid",
    "Hardmax",
    "Identity",
    "If",
    "ImageDecoder",
    "InstanceNormalization",
    "IsInf",
    "IsNaN",
    "LayerNormalization",
    "LeakyRelu",
    "Less",
    "LessOrEqual",
    "LinearAttention",
    "Log",
    "LogSoftmax",
    "Loop",
    "LpNormalization",
    "LpPool",
    "LRN",
    "LSTM",
    "MatMul",
    "MatMulInteger",
    "Max",
    "MaxPool",
    "MaxUnpool",
    "Mean",
    "MelWeightMatrix",
    "Min",
    "Mod",
    "Mul",
    "Neg",
    "NegativeLogLikelihoodLoss",
    "NonMaxSuppression",
    "NonZero",
    "Not",
    "OneHot",
    "Optional",
    "OptionalGetElement",
    "OptionalHasElement",
    "Or",
    "Pad_1",
    "Pad_2",
    "Pad_11",
    "Pad_18",
    "Pow",
    "PRelu",
    "QLinearConv",
    "QLinearMatMul",
    "QuantizeLinear_10",
    "QuantizeLinear_19",
    "QuantizeLinear_21",
    "QuantizeLinear_23",
    "QuantizeLinear_25",
    "RandomNormal",
    "RandomNormalLike",
    "RandomUniform",
    "RandomUniformLike",
    "Range",
    "Reciprocal",
    "ReduceL1_1",
    "ReduceL1_18",
    "ReduceL2_1",
    "ReduceL2_18",
    "ReduceLogSum_1",
    "ReduceLogSum_18",
    "ReduceLogSumExp_1",
    "ReduceLogSumExp_18",
    "ReduceMax_1",
    "ReduceMax_18",
    "ReduceMean_1",
    "ReduceMean_18",
    "ReduceMin_1",
    "ReduceMin_18",
    "ReduceProd_1",
    "ReduceProd_18",
    "ReduceSum_1",
    "ReduceSum_13",
    "ReduceSumSquare_1",
    "ReduceSumSquare_18",
    "RegexFullMatch",
    "Relu",
    "Reshape_5",
    "Reshape_14",
    "Resize",
    "ReverseSequence",
    "RMSNormalization",
    "RNN_7",
    "RNN_14",
    "RoiAlign",
    "RotaryEmbedding",
    "Round",
    "Scan",
    "ScatterElements",
    "ScatterND",
    "Selu",
    "SequenceAt",
    "SequenceConstruct",
    "SequenceEmpty",
    "SequenceErase",
    "SequenceInsert",
    "SequenceLength",
    "SequenceMap",
    "Shape_1",
    "Shape_15",
    "Shrink",
    "Sigmoid",
    "Sign",
    "Sin",
    "Sinh",
    "Size",
    "Slice_1",
    "Slice_10",
    "Softmax",
    "SoftmaxCrossEntropyLoss",
    "Softplus",
    "Softsign",
    "Swish",
    "SpaceToDepth",
    "Split_2",
    "Split_11",
    "Split_13",
    "Split_18",
    "SplitToSequence",
    "Sqrt",
    "Squeeze_1",
    "Squeeze_11",
    "Squeeze_13",
    "STFT",
    "StringConcat",
    "StringNormalizer",
    "StringSplit",
    "Sub",
    "Sum",
    "Tan",
    "Tanh",
    "TensorScatter",
    "TfIdfVectorizer",
    "ThresholdedRelu",
    "Tile",
    "TopK_1",
    "TopK_10",
    "TopK_11",
    "Transpose",
    "Trilu",
    "Unique",
    "Unsqueeze_1",
    "Unsqueeze_11",
    "Unsqueeze_13",
    "Upsample",
    "Where",
    "Xor",
]

import textwrap
from typing import Any

from onnx import FunctionProto, NodeProto, TypeProto
from onnx.defs import get_schema, onnx_opset_version
from onnx.onnx_cpp2py_export.defs import SchemaError
from onnx.reference.op_run import (
    OpFunction,
    OpRun,
    RuntimeContextError,
    RuntimeImplementationError,
)
from onnx.reference.ops._helpers import build_registered_operators_any_domain
from onnx.reference.ops.op_abs import Abs
from onnx.reference.ops.op_acos import Acos
from onnx.reference.ops.op_acosh import Acosh
from onnx.reference.ops.op_add import Add
from onnx.reference.ops.op_affine_grid import AffineGrid
from onnx.reference.ops.op_and import And
from onnx.reference.ops.op_argmax import ArgMax_1, ArgMax_12
from onnx.reference.ops.op_argmin import ArgMin_1, ArgMin_12
from onnx.reference.ops.op_asin import Asin
from onnx.reference.ops.op_asinh import Asinh
from onnx.reference.ops.op_atan import Atan
from onnx.reference.ops.op_atanh import Atanh
from onnx.reference.ops.op_attention import Attention
from onnx.reference.ops.op_attribute_has_value import AttributeHasValue
from onnx.reference.ops.op_average_pool import (
    AveragePool_1,
    AveragePool_7,
    AveragePool_11,
    AveragePool_19,
)
from onnx.reference.ops.op_batch_normalization import (
    BatchNormalization_6,
    BatchNormalization_9,
    BatchNormalization_14,
)
from onnx.reference.ops.op_bernoulli import Bernoulli
from onnx.reference.ops.op_bitcast import BitCast
from onnx.reference.ops.op_bitshift import BitShift
from onnx.reference.ops.op_bitwise_and import BitwiseAnd
from onnx.reference.ops.op_bitwise_not import BitwiseNot
from onnx.reference.ops.op_bitwise_or import BitwiseOr
from onnx.reference.ops.op_bitwise_xor import BitwiseXor
from onnx.reference.ops.op_blackman_window import BlackmanWindow
from onnx.reference.ops.op_cast import Cast_1, Cast_19, Cast_24
from onnx.reference.ops.op_cast_like import CastLike_15, CastLike_19
from onnx.reference.ops.op_causal_conv_with_state import CausalConvWithState
from onnx.reference.ops.op_ceil import Ceil
from onnx.reference.ops.op_celu import Celu
from onnx.reference.ops.op_center_crop_pad import CenterCropPad
from onnx.reference.ops.op_clip import Clip_6, Clip_11
from onnx.reference.ops.op_col2im import Col2Im
from onnx.reference.ops.op_compress import Compress
from onnx.reference.ops.op_concat import Concat
from onnx.reference.ops.op_concat_from_sequence import ConcatFromSequence
from onnx.reference.ops.op_constant import (
    Constant_1,
    Constant_9,
    Constant_11,
    Constant_12,
)
from onnx.reference.ops.op_constant_of_shape import ConstantOfShape
from onnx.reference.ops.op_conv import Conv
from onnx.reference.ops.op_conv_integer import ConvInteger
from onnx.reference.ops.op_conv_transpose import ConvTranspose
from onnx.reference.ops.op_cos import Cos
from onnx.reference.ops.op_cosh import Cosh
from onnx.reference.ops.op_cum_prod import CumProd
from onnx.reference.ops.op_cum_sum import CumSum
from onnx.reference.ops.op_deform_conv import DeformConv
from onnx.reference.ops.op_depth_to_space import DepthToSpace
from onnx.reference.ops.op_dequantize_linear import (
    DequantizeLinear_19,
    DequantizeLinear_21,
    DequantizeLinear_23,
    DequantizeLinear_25,
)
from onnx.reference.ops.op_det import Det
from onnx.reference.ops.op_dft import DFT_17, DFT_20
from onnx.reference.ops.op_div import Div
from onnx.reference.ops.op_dropout import Dropout_7, Dropout_12
from onnx.reference.ops.op_dynamic_quantize_linear import DynamicQuantizeLinear
from onnx.reference.ops.op_einsum import Einsum
from onnx.reference.ops.op_elu import Elu
from onnx.reference.ops.op_equal import Equal
from onnx.reference.ops.op_erf import Erf
from onnx.reference.ops.op_exp import Exp
from onnx.reference.ops.op_expand import Expand
from onnx.reference.ops.op_eyelike import EyeLike
from onnx.reference.ops.op_flatten import Flatten
from onnx.reference.ops.op_floor import Floor
from onnx.reference.ops.op_gather import Gather
from onnx.reference.ops.op_gather_elements import GatherElements
from onnx.reference.ops.op_gathernd import GatherND
from onnx.reference.ops.op_gemm import Gemm_6, Gemm_7
from onnx.reference.ops.op_global_average_pool import GlobalAveragePool
from onnx.reference.ops.op_global_max_pool import GlobalMaxPool
from onnx.reference.ops.op_greater import Greater
from onnx.reference.ops.op_greater_or_equal import GreaterOrEqual
from onnx.reference.ops.op_grid_sample import GridSample
from onnx.reference.ops.op_gru import GRU
from onnx.reference.ops.op_hamming_window import HammingWindow
from onnx.reference.ops.op_hann_window import HannWindow
from onnx.reference.ops.op_hard_sigmoid import HardSigmoid
from onnx.reference.ops.op_hardmax import Hardmax
from onnx.reference.ops.op_identity import Identity
from onnx.reference.ops.op_if import If
from onnx.reference.ops.op_image_decoder import ImageDecoder
from onnx.reference.ops.op_instance_normalization import InstanceNormalization
from onnx.reference.ops.op_isinf import IsInf
from onnx.reference.ops.op_isnan import IsNaN
from onnx.reference.ops.op_layer_normalization import LayerNormalization
from onnx.reference.ops.op_leaky_relu import LeakyRelu
from onnx.reference.ops.op_less import Less
from onnx.reference.ops.op_less_or_equal import LessOrEqual
from onnx.reference.ops.op_linear_attention import LinearAttention
from onnx.reference.ops.op_log import Log
from onnx.reference.ops.op_log_softmax import LogSoftmax
from onnx.reference.ops.op_loop import Loop
from onnx.reference.ops.op_lp_normalization import LpNormalization
from onnx.reference.ops.op_lp_pool import LpPool
from onnx.reference.ops.op_lrn import LRN
from onnx.reference.ops.op_lstm import LSTM
from onnx.reference.ops.op_matmul import MatMul
from onnx.reference.ops.op_matmul_integer import MatMulInteger
from onnx.reference.ops.op_max import Max
from onnx.reference.ops.op_max_pool import MaxPool
from onnx.reference.ops.op_max_unpool import MaxUnpool
from onnx.reference.ops.op_mean import Mean
from onnx.reference.ops.op_mel_weight_matrix import MelWeightMatrix
from onnx.reference.ops.op_min import Min
from onnx.reference.ops.op_mod import Mod
from onnx.reference.ops.op_mul import Mul
from onnx.reference.ops.op_neg import Neg
from onnx.reference.ops.op_negative_log_likelihood_loss import NegativeLogLikelihoodLoss
from onnx.reference.ops.op_non_max_suppression import NonMaxSuppression
from onnx.reference.ops.op_non_zero import NonZero
from onnx.reference.ops.op_not import Not
from onnx.reference.ops.op_one_hot import OneHot
from onnx.reference.ops.op_optional import Optional
from onnx.reference.ops.op_optional_get_element import OptionalGetElement
from onnx.reference.ops.op_optional_has_element import OptionalHasElement
from onnx.reference.ops.op_or import Or
from onnx.reference.ops.op_pad import Pad_1, Pad_2, Pad_11, Pad_18
from onnx.reference.ops.op_pow import Pow
from onnx.reference.ops.op_prelu import PRelu
from onnx.reference.ops.op_qlinear_conv import QLinearConv
from onnx.reference.ops.op_qlinear_matmul import QLinearMatMul
from onnx.reference.ops.op_quantize_linear import (
    QuantizeLinear_10,
    QuantizeLinear_19,
    QuantizeLinear_21,
    QuantizeLinear_23,
    QuantizeLinear_25,
)
from onnx.reference.ops.op_random_normal import RandomNormal
from onnx.reference.ops.op_random_normal_like import RandomNormalLike
from onnx.reference.ops.op_random_uniform import RandomUniform
from onnx.reference.ops.op_random_uniform_like import RandomUniformLike
from onnx.reference.ops.op_range import Range
from onnx.reference.ops.op_reciprocal import Reciprocal
from onnx.reference.ops.op_reduce_l1 import ReduceL1_1, ReduceL1_18
from onnx.reference.ops.op_reduce_l2 import ReduceL2_1, ReduceL2_18
from onnx.reference.ops.op_reduce_log_sum import ReduceLogSum_1, ReduceLogSum_18
from onnx.reference.ops.op_reduce_log_sum_exp import (
    ReduceLogSumExp_1,
    ReduceLogSumExp_18,
)
from onnx.reference.ops.op_reduce_max import ReduceMax_1, ReduceMax_18
from onnx.reference.ops.op_reduce_mean import ReduceMean_1, ReduceMean_18
from onnx.reference.ops.op_reduce_min import ReduceMin_1, ReduceMin_18
from onnx.reference.ops.op_reduce_prod import ReduceProd_1, ReduceProd_18
from onnx.reference.ops.op_reduce_sum import ReduceSum_1, ReduceSum_13
from onnx.reference.ops.op_reduce_sum_square import (
    ReduceSumSquare_1,
    ReduceSumSquare_18,
)
from onnx.reference.ops.op_regex_full_match import RegexFullMatch
from onnx.reference.ops.op_relu import Relu
from onnx.reference.ops.op_reshape import Reshape_5, Reshape_14
from onnx.reference.ops.op_resize import Resize
from onnx.reference.ops.op_reverse_sequence import ReverseSequence
from onnx.reference.ops.op_rms_normalization import RMSNormalization
from onnx.reference.ops.op_rnn import RNN_7, RNN_14
from onnx.reference.ops.op_roi_align import RoiAlign
from onnx.reference.ops.op_rotary_embedding import RotaryEmbedding
from onnx.reference.ops.op_round import Round
from onnx.reference.ops.op_scan import Scan
from onnx.reference.ops.op_scatter_elements import ScatterElements
from onnx.reference.ops.op_scatternd import ScatterND
from onnx.reference.ops.op_selu import Selu
from onnx.reference.ops.op_sequence_at import SequenceAt
from onnx.reference.ops.op_sequence_construct import SequenceConstruct
from onnx.reference.ops.op_sequence_empty import SequenceEmpty
from onnx.reference.ops.op_sequence_erase import SequenceErase
from onnx.reference.ops.op_sequence_insert import SequenceInsert
from onnx.reference.ops.op_sequence_length import SequenceLength
from onnx.reference.ops.op_sequence_map import SequenceMap
from onnx.reference.ops.op_shape import Shape_1, Shape_15
from onnx.reference.ops.op_shrink import Shrink
from onnx.reference.ops.op_sigmoid import Sigmoid
from onnx.reference.ops.op_sign import Sign
from onnx.reference.ops.op_sin import Sin
from onnx.reference.ops.op_sinh import Sinh
from onnx.reference.ops.op_size import Size
from onnx.reference.ops.op_slice import Slice_1, Slice_10
from onnx.reference.ops.op_softmax import Softmax
from onnx.reference.ops.op_softmax_cross_entropy_loss import SoftmaxCrossEntropyLoss
from onnx.reference.ops.op_softplus import Softplus
from onnx.reference.ops.op_softsign import Softsign
from onnx.reference.ops.op_space_to_depth import SpaceToDepth
from onnx.reference.ops.op_split import Split_2, Split_11, Split_13, Split_18
from onnx.reference.ops.op_split_to_sequence import SplitToSequence
from onnx.reference.ops.op_sqrt import Sqrt
from onnx.reference.ops.op_squeeze import Squeeze_1, Squeeze_11, Squeeze_13
from onnx.reference.ops.op_stft import STFT
from onnx.reference.ops.op_string_concat import StringConcat
from onnx.reference.ops.op_string_normalizer import StringNormalizer
from onnx.reference.ops.op_string_split import StringSplit
from onnx.reference.ops.op_sub import Sub
from onnx.reference.ops.op_sum import Sum
from onnx.reference.ops.op_swish import Swish
from onnx.reference.ops.op_tan import Tan
from onnx.reference.ops.op_tanh import Tanh
from onnx.reference.ops.op_tensor_scatter import TensorScatter
from onnx.reference.ops.op_tfidf_vectorizer import TfIdfVectorizer
from onnx.reference.ops.op_thresholded_relu import ThresholdedRelu
from onnx.reference.ops.op_tile import Tile
from onnx.reference.ops.op_topk import TopK_1, TopK_10, TopK_11
from onnx.reference.ops.op_transpose import Transpose
from onnx.reference.ops.op_trilu import Trilu
from onnx.reference.ops.op_unique import Unique
from onnx.reference.ops.op_unsqueeze import Unsqueeze_1, Unsqueeze_11, Unsqueeze_13
from onnx.reference.ops.op_upsample import Upsample
from onnx.reference.ops.op_where import Where
from onnx.reference.ops.op_xor import Xor


def _build_registered_operators() -> dict[str, dict[int | None, type[OpRun]]]:
    return build_registered_operators_any_domain(globals().copy())


def load_op(
    domain: str,
    op_type: str,
    version: None | int = None,
    custom: Any = None,
    node: None | NodeProto = None,
    input_types: None | list[TypeProto] = None,
    expand: bool = False,
    evaluator_cls: type | None = None,
) -> Any:
    """Loads the implemented for a specified operator.

    Args:
        domain: domain
        op_type: operator type
        version: requested version
        custom: custom implementation (like a function)
        node: used if no implementation was found and the operator
            defines a function which is context dependent
        input_types: used if no implementation was found and the
            operator defines a function which is context dependent
        expand: use the function implemented in the schema instead of
            its reference implementation
        evaluator_cls: evaluator to use

    Returns:
        class
    """
    global _registered_operators  # noqa: PLW0603
    schema = None
    if _registered_operators is None:
        _registered_operators = _build_registered_operators()  # type: ignore[assignment]
    assert _registered_operators is not None
    if custom is not None:
        return lambda *args: OpFunction(*args, impl=custom)
    if version is None:
        version = onnx_opset_version()
    if domain != "":
        raise ValueError(f"Domain must be '' not {domain!r}.")
    if op_type in _registered_operators and not expand:
        found = True
    else:
        # maybe the operator can be replaced by a function
        try:
            schema = get_schema(op_type, version, domain)
        except SchemaError:
            raise NotImplementedError(
                f"No registered schema for operator {op_type!r} "
                f"and domain {domain!r}. Did you recompile the sources after updating the repository?"
            ) from None
        if schema.has_function:
            body = schema.function_body
            assert evaluator_cls is not None, (
                f"evaluator_cls must be specified to implement operator {op_type!r} from domain {domain!r}"
            )
            sess = evaluator_cls(body)
            return lambda *args, sess=sess: OpFunction(*args, impl=sess)
        if schema.has_context_dependent_function:
            if node is None or input_types is None:
                raise RuntimeContextError(
                    f"No registered implementation for operator {op_type!r} "
                    f"and domain {domain!r}, the operator has a context dependent function. "
                    f"but argument node or input_types is not defined (input_types={input_types})."
                )
            body = schema.get_context_dependent_function(
                node.SerializeToString(), [it.SerializeToString() for it in input_types]
            )
            proto = FunctionProto()
            proto.ParseFromString(body)
            assert evaluator_cls is not None, (
                f"evaluator_cls must be specified to evaluate function {proto.name!r}"
            )
            sess = evaluator_cls(proto)
            return lambda *args, sess=sess: OpFunction(*args, impl=sess)
        found = False
    if not found:
        available = "\n".join(textwrap.wrap(", ".join(sorted(_registered_operators))))
        has_function = schema.has_function if schema else None
        has_context_dependent_function = (
            schema.has_context_dependent_function if schema else None
        )
        raise RuntimeImplementationError(
            f"No registered implementation for operator {op_type!r} "
            f"and domain {domain!r}, schema.has_function is {has_function}, "
            f"schema.has_context_dependent_function is {has_context_dependent_function}. "
            f"You may either add one or skip the test in "
            f"'test_backend_reference.py'. Available implementations:\n{available}"
        )
    impl = _registered_operators[op_type]
    if None not in impl:
        raise RuntimeError(
            f"No default implementation for operator {op_type!r} "
            f"and domain {domain!r}, found "
            f"{', '.join(map(str, impl))}."
        )
    if version is None or len(impl) == 1:
        cl = impl[None]
    else:
        best = -1
        for v in impl:
            if v is None:
                continue
            if best < v <= version:
                best = v
        if best == -1:
            raise RuntimeError(
                f"No implementation for operator {op_type!r} "
                f"domain {domain!r} and version {version!r}, found "
                f"{', '.join(map(str, impl))}."
            )
        cl = impl[best]
    if cl is None:
        available = "\n".join(textwrap.wrap(", ".join(sorted(_registered_operators))))
        raise ValueError(
            f"Not registered implementation for operator {op_type!r}, "
            f"domain {domain!r}, and {version!r} in\n{available}"
        )
    return cl


_registered_operators: dict[str, dict[int | None, OpRun]] | None = None