|
| 1 | +# Copyright (c) ONNX Project Contributors |
| 2 | +# SPDX-License-Identifier: Apache-2.0 |
| 3 | +"""Eliminate common subexpression in ONNX graphs.""" |
| 4 | + |
| 5 | +from __future__ import annotations |
| 6 | + |
| 7 | +__all__ = [ |
| 8 | + "CommonSubexpressionEliminationPass", |
| 9 | +] |
| 10 | + |
| 11 | +import logging |
| 12 | +from collections.abc import Sequence |
| 13 | + |
| 14 | +import onnx_ir as ir |
| 15 | + |
| 16 | +logger = logging.getLogger(__name__) |
| 17 | + |
| 18 | + |
| 19 | +class CommonSubexpressionEliminationPass(ir.passes.InPlacePass): |
| 20 | + """Eliminate common subexpression in ONNX graphs.""" |
| 21 | + |
| 22 | + def call(self, model: ir.Model) -> ir.passes.PassResult: |
| 23 | + """Return the same ir.Model but with CSE applied to the graph.""" |
| 24 | + modified = False |
| 25 | + graph = model.graph |
| 26 | + |
| 27 | + modified = _eliminate_common_subexpression(graph, modified) |
| 28 | + |
| 29 | + return ir.passes.PassResult( |
| 30 | + model, |
| 31 | + modified=modified, |
| 32 | + ) |
| 33 | + |
| 34 | + |
| 35 | +def _eliminate_common_subexpression(graph: ir.Graph, modified: bool) -> bool: |
| 36 | + """Eliminate common subexpression in ONNX graphs.""" |
| 37 | + # node to node identifier, length of outputs, inputs, and attributes |
| 38 | + existing_node_info_to_the_node: dict[ |
| 39 | + tuple[ |
| 40 | + ir.OperatorIdentifier, |
| 41 | + int, # len(outputs) |
| 42 | + tuple[int, ...], # input ids |
| 43 | + tuple[tuple[str, object], ...], # attributes |
| 44 | + ], |
| 45 | + ir.Node, |
| 46 | + ] = {} |
| 47 | + |
| 48 | + for node in graph: |
| 49 | + # Skip control flow ops like Loop and If. |
| 50 | + control_flow_op: bool = False |
| 51 | + # Use equality to check if the node is a common subexpression. |
| 52 | + attributes = {} |
| 53 | + for k, v in node.attributes.items(): |
| 54 | + # TODO(exporter team): CSE subgraphs. |
| 55 | + # NOTE: control flow ops like Loop and If won't be CSEd |
| 56 | + # because attribute: graph won't match. |
| 57 | + if v.type in (ir.AttributeType.GRAPH, ir.AttributeType.GRAPHS): |
| 58 | + control_flow_op = True |
| 59 | + logger.debug("Skipping control flow op %s", node) |
| 60 | + # The attribute value could be directly taken from the original |
| 61 | + # protobuf, so we need to make a copy of it. |
| 62 | + value = v.value |
| 63 | + if v.type in ( |
| 64 | + ir.AttributeType.INTS, |
| 65 | + ir.AttributeType.FLOATS, |
| 66 | + ir.AttributeType.STRINGS, |
| 67 | + ): |
| 68 | + # For INT, FLOAT and STRING attributes, we convert them to tuples |
| 69 | + # to ensure they are hashable. |
| 70 | + value = tuple(value) |
| 71 | + attributes[k] = value |
| 72 | + |
| 73 | + if control_flow_op: |
| 74 | + # If the node is a control flow op, we skip it. |
| 75 | + logger.debug("Skipping control flow op %s", node) |
| 76 | + continue |
| 77 | + |
| 78 | + if _is_non_deterministic_op(node): |
| 79 | + # If the node is a non-deterministic op, we skip it. |
| 80 | + logger.debug("Skipping non-deterministic op %s", node) |
| 81 | + continue |
| 82 | + |
| 83 | + node_info = ( |
| 84 | + node.op_identifier(), |
| 85 | + len(node.outputs), |
| 86 | + tuple(id(input) for input in node.inputs), |
| 87 | + tuple(sorted(attributes.items())), |
| 88 | + ) |
| 89 | + # Check if the node is a common subexpression. |
| 90 | + if node_info in existing_node_info_to_the_node: |
| 91 | + # If it is, this node has an existing node with the same |
| 92 | + # operator, number of outputs, inputs, and attributes. |
| 93 | + # We replace the node with the existing node. |
| 94 | + modified = True |
| 95 | + existing_node = existing_node_info_to_the_node[node_info] |
| 96 | + _remove_node_and_replace_values( |
| 97 | + graph, |
| 98 | + remove_node=node, |
| 99 | + remove_values=node.outputs, |
| 100 | + new_values=existing_node.outputs, |
| 101 | + ) |
| 102 | + logger.debug("Reusing node %s", existing_node) |
| 103 | + else: |
| 104 | + # If it is not, add to the mapping. |
| 105 | + existing_node_info_to_the_node[node_info] = node |
| 106 | + return modified |
| 107 | + |
| 108 | + |
| 109 | +def _remove_node_and_replace_values( |
| 110 | + graph: ir.Graph, |
| 111 | + /, |
| 112 | + remove_node: ir.Node, |
| 113 | + remove_values: Sequence[ir.Value], |
| 114 | + new_values: Sequence[ir.Value], |
| 115 | +) -> None: |
| 116 | + """Replaces nodes and values in the graph or function. |
| 117 | +
|
| 118 | + Args: |
| 119 | + graph: The graph to replace nodes and values in. |
| 120 | + remove_node: The node to remove. |
| 121 | + remove_values: The values to replace. |
| 122 | + new_values: The values to replace with. |
| 123 | + """ |
| 124 | + # Reconnect the users of the deleted values to use the new values |
| 125 | + ir.convenience.replace_all_uses_with(remove_values, new_values) |
| 126 | + # Update graph/function outputs if the node generates output |
| 127 | + if any(remove_value.is_graph_output() for remove_value in remove_values): |
| 128 | + replacement_mapping = dict(zip(remove_values, new_values)) |
| 129 | + for idx, graph_output in enumerate(graph.outputs): |
| 130 | + if graph_output in replacement_mapping: |
| 131 | + new_value = replacement_mapping[graph_output] |
| 132 | + if new_value.is_graph_output() or new_value.is_graph_input(): |
| 133 | + # If the new value is also a graph input/output, we need to |
| 134 | + # create a Identity node to preserve the remove_value and |
| 135 | + # prevent from changing new_value name. |
| 136 | + identity_node = ir.node( |
| 137 | + "Identity", |
| 138 | + inputs=[new_value], |
| 139 | + outputs=[ |
| 140 | + ir.Value( |
| 141 | + name=graph_output.name, |
| 142 | + type=graph_output.type, |
| 143 | + shape=graph_output.shape, |
| 144 | + ) |
| 145 | + ], |
| 146 | + ) |
| 147 | + # reuse the name of the graph output |
| 148 | + graph.outputs[idx] = identity_node.outputs[0] |
| 149 | + graph.insert_before( |
| 150 | + remove_node, |
| 151 | + identity_node, |
| 152 | + ) |
| 153 | + else: |
| 154 | + # if new_value is not graph output, we just |
| 155 | + # update it to use old_value name. |
| 156 | + new_value.name = graph_output.name |
| 157 | + graph.outputs[idx] = new_value |
| 158 | + |
| 159 | + graph.remove(remove_node, safe=True) |
| 160 | + |
| 161 | + |
| 162 | +def _is_non_deterministic_op(node: ir.Node) -> bool: |
| 163 | + non_deterministic_ops = frozenset( |
| 164 | + { |
| 165 | + "RandomUniform", |
| 166 | + "RandomNormal", |
| 167 | + "RandomUniformLike", |
| 168 | + "RandomNormalLike", |
| 169 | + "Multinomial", |
| 170 | + } |
| 171 | + ) |
| 172 | + return node.op_type in non_deterministic_ops and _is_onnx_domain(node.domain) |
| 173 | + |
| 174 | + |
| 175 | +def _is_onnx_domain(d: str) -> bool: |
| 176 | + """Check if the domain is the ONNX domain.""" |
| 177 | + return d == "" |
![dependabot[bot]](https://avatars.githubusercontent.com/in/29110?v=4&size=40){kind=avatar}
0 commit comments