Reworking the placement of metatype groups (#2094)

### Changes

Create a separate module to store metatype groups in one location.

### Reason for changes

Ref: 113580

### Related tickets

Ref: 113580

### Tests

N/A

nncf/onnx/graph/metatypes/groups.py

@@ -0,0 +1,122 @@
+# Copyright (c) 2023 Intel Corporation
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#      http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from nncf.onnx.graph.metatypes import onnx_metatypes
+
+QUANTIZE_AGNOSTIC_OPERATIONS = [
+    onnx_metatypes.ONNXMaxPoolMetatype,
+    onnx_metatypes.ONNXReduceMaxMetatype,
+    onnx_metatypes.ONNXReshapeMetatype,
+    onnx_metatypes.ONNXTransposeMetatype,
+    onnx_metatypes.ONNXSqueezeMetatype,
+    onnx_metatypes.ONNXUnsqueezeMetatype,
+    onnx_metatypes.ONNXSplitMetatype,
+    onnx_metatypes.ONNXTileMetatype,
+    onnx_metatypes.ONNXCenterCropPadMetatype,
+    onnx_metatypes.ONNXSliceMetatype,
+    onnx_metatypes.ONNXPadMetatype,
+    onnx_metatypes.ONNXGatherMetatype,
+    onnx_metatypes.ONNXGatherNDMetatype,
+    onnx_metatypes.ONNXGatherElementsMetatype,
+    onnx_metatypes.ONNXDepthToSpaceMetatype,
+    onnx_metatypes.ONNXSpaceToDepthMetatype,
+    onnx_metatypes.ONNXScatterElementsMetatype,
+    onnx_metatypes.ONNXScatterNDMetatype,
+    onnx_metatypes.ONNXScatterMetatype,
+    onnx_metatypes.ONNXCastLikeMetatype,
+    onnx_metatypes.ONNXDropoutMetatype,
+    onnx_metatypes.ONNXFlattenMetatype,
+    onnx_metatypes.ONNXExpandMetatype,
+    onnx_metatypes.ONNXIdentityMetatype,
+    # ONNXReluMetatype is not considered to be QUANTIZATION_AGNOSTIC, because:
+    # 1. Runtime doesn't provide performance benefits by quantizing the stand-alone RELU's (ticket: 59548)
+    # 2. It's frequently better for the end accuracy to have quantizers set up after the RELU
+    # so that the input distribution to the quantizer is non-negative
+    # and we can therefore have better quantization resolution while preserving the original dynamic range
+]
+
+
+MATMUL_METATYPES = [onnx_metatypes.ONNXGemmMetatype, onnx_metatypes.ONNXMatMulMetatype]
+
+
+INPUTS_QUANTIZABLE_OPERATIONS = [
+    onnx_metatypes.ONNXConvolutionMetatype,
+    onnx_metatypes.ONNXDepthwiseConvolutionMetatype,
+    onnx_metatypes.ONNXConvolutionTransposeMetatype,
+    *MATMUL_METATYPES,
+    onnx_metatypes.ONNXAveragePoolMetatype,
+    onnx_metatypes.ONNXGlobalAveragePoolMetatype,
+    onnx_metatypes.ONNXAddLayerMetatype,
+    onnx_metatypes.ONNXSubMetatype,
+    onnx_metatypes.ONNXMulLayerMetatype,
+    onnx_metatypes.ONNXBatchNormMetatype,
+    onnx_metatypes.ONNXHardSigmoidMetatype,
+    onnx_metatypes.ONNXResizeMetatype,
+    onnx_metatypes.ONNXPowMetatype,
+    onnx_metatypes.ONNXReciprocalMetatype,
+    onnx_metatypes.ONNXMaximumMetatype,
+    onnx_metatypes.ONNXMinimumMetatype,
+]
+
+
+CONSTANT_WEIGHT_LAYER_METATYPES = [
+    onnx_metatypes.ONNXConvolutionMetatype,
+    onnx_metatypes.ONNXDepthwiseConvolutionMetatype,
+    onnx_metatypes.ONNXConvolutionTransposeMetatype,
+    onnx_metatypes.ONNXEmbeddingMetatype,
+]
+
+
+LINEAR_OPERATIONS = [
+    onnx_metatypes.ONNXConvolutionMetatype,
+    onnx_metatypes.ONNXDepthwiseConvolutionMetatype,
+    onnx_metatypes.ONNXConvolutionTransposeMetatype,
+    onnx_metatypes.ONNXDeformableConvolutionMetatype,
+    *MATMUL_METATYPES,
+]
+
+
+ATOMIC_ACTIVATIONS_OPERATIONS = [
+    onnx_metatypes.ONNXReluMetatype,
+    onnx_metatypes.ONNXLeakyReluMetatype,
+    onnx_metatypes.ONNXThresholdedReluMetatype,
+    onnx_metatypes.ONNXEluMetatype,
+    onnx_metatypes.ONNXPReluMetatype,
+    onnx_metatypes.ONNXSigmoidMetatype,
+    onnx_metatypes.ONNXHardSigmoidMetatype,
+    onnx_metatypes.ONNXHardSwishMetatype,
+]
+
+
+ARITHMETIC_OPERATIONS = [
+    onnx_metatypes.ONNXAddLayerMetatype,
+    onnx_metatypes.ONNXSubMetatype,
+    onnx_metatypes.ONNXMulLayerMetatype,
+    onnx_metatypes.ONNXDivLayerMetatype,
+]
+
+
+OPERATIONS_WITH_WEIGHTS = [
+    *CONSTANT_WEIGHT_LAYER_METATYPES,
+    *MATMUL_METATYPES,
+]
+
+
+BATCH_NORMALIZATION_OPERATIONS = [
+    onnx_metatypes.ONNXBatchNormMetatype,
+]
+
+
+# Contains the operation metatypes for which bias can be applied.
+OPERATIONS_WITH_BIAS = [
+    onnx_metatypes.ONNXConvolutionMetatype,
+    onnx_metatypes.ONNXDepthwiseConvolutionMetatype,
+]

nncf/onnx/graph/metatypes/onnx_metatypes.py

@@ -610,24 +610,6 @@ class ONNXDeformableConvolutionMetatype(ONNXOpMetatype):
     op_names = ["DeformConv"]
 
 
-CONSTANT_WEIGHT_LAYER_METATYPES = [
-    ONNXConvolutionMetatype,
-    ONNXDepthwiseConvolutionMetatype,
-    ONNXConvolutionTransposeMetatype,
-    ONNXEmbeddingMetatype,
-]
-
-MATMUL_METATYPES = [ONNXGemmMetatype, ONNXMatMulMetatype]
-
-GENERAL_WEIGHT_LAYER_METATYPES = CONSTANT_WEIGHT_LAYER_METATYPES + MATMUL_METATYPES
-
-# Contains the operation metatypes for which bias can be applied.
-OPERATIONS_WITH_BIAS_METATYPES = [
-    ONNXConvolutionMetatype,
-    ONNXDepthwiseConvolutionMetatype,
-]
-
-
 def get_operator_metatypes() -> List[Type[OperatorMetatype]]:
     """
     Returns a list of the operator metatypes.
@@ -653,43 +635,6 @@ def get_metatype(model: onnx.ModelProto, node: onnx.NodeProto) -> ONNXOpMetatype
     return metatype
 
 
-def get_constant_weight_port_ids(metatype: ONNXOpMetatype) -> List[int]:
-    """
-    Returns port ids on which metatype must have a weight based on Operation definition.
-
-    :param metatype: Metatype.
-    :return: Port ids.
-    """
-    if metatype in CONSTANT_WEIGHT_LAYER_METATYPES:
-        return metatype.weight_port_ids
-    return []
-
-
-def get_possible_weight_port_ids(metatype: ONNXOpMetatype) -> List[int]:
-    """
-    Returns weight port ids on which metatype could have a weight.
-    Example: ONNXMatMulMetatype could have activations or weights on input port ids: 0, 1
-
-    :param metatype: Metatype.
-    :return: Port ids.
-    """
-    if metatype in MATMUL_METATYPES:
-        return metatype.possible_weight_ports
-    return []
-
-
-def get_bias_tensor_port_id(metatype: ONNXOpWithWeightsMetatype) -> Optional[int]:
-    """
-    Returns input port id, where a bias tensor should output.
-
-    :param node: Node, for which input port id is returned,
-    :return: Input port id, where a weight bias should output or None if node can not have bias.
-    """
-    if metatype in OPERATIONS_WITH_BIAS_METATYPES:
-        return metatype.bias_port_id
-    return None
-
-
 def get_tensor_edge_name(onnx_graph: ONNXGraph, node: onnx.NodeProto, port_id: int) -> Optional[str]:
     """
     Returns an edge name associated with a weight of a node laying on  an input port_id.

nncf/onnx/graph/nncf_graph_builder.py

@@ -9,7 +9,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from collections import Counter
-from typing import Any, Dict, Optional, Set
+from typing import Any, Dict, List, Optional, Set
 
 import onnx
 
@@ -21,11 +21,13 @@
 from nncf.common.graph.layer_attributes import Dtype
 from nncf.common.graph.operator_metatypes import InputNoopMetatype
 from nncf.common.graph.operator_metatypes import OutputNoopMetatype
+from nncf.onnx.graph.metatypes.groups import CONSTANT_WEIGHT_LAYER_METATYPES
+from nncf.onnx.graph.metatypes.groups import MATMUL_METATYPES
+from nncf.onnx.graph.metatypes.groups import OPERATIONS_WITH_BIAS
 from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXGemmMetatype
-from nncf.onnx.graph.metatypes.onnx_metatypes import get_bias_tensor_port_id
-from nncf.onnx.graph.metatypes.onnx_metatypes import get_constant_weight_port_ids
+from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpMetatype
+from nncf.onnx.graph.metatypes.onnx_metatypes import ONNXOpWithWeightsMetatype
 from nncf.onnx.graph.metatypes.onnx_metatypes import get_metatype
-from nncf.onnx.graph.metatypes.onnx_metatypes import get_possible_weight_port_ids
 from nncf.onnx.graph.metatypes.onnx_metatypes import get_tensor_edge_name
 from nncf.onnx.graph.onnx_graph import ONNXGraph
 
@@ -64,6 +66,43 @@ def has_node_attrs(self) -> bool:
         return bool(self.node_attrs)
 
 
+def get_constant_weight_port_ids(metatype: ONNXOpMetatype) -> List[int]:
+    """
+    Returns port ids on which metatype must have a weight based on Operation definition.
+
+    :param metatype: Metatype.
+    :return: Port ids.
+    """
+    if metatype in CONSTANT_WEIGHT_LAYER_METATYPES:
+        return metatype.weight_port_ids
+    return []
+
+
+def get_possible_weight_port_ids(metatype: ONNXOpMetatype) -> List[int]:
+    """
+    Returns weight port ids on which metatype could have a weight.
+    Example: ONNXMatMulMetatype could have activations or weights on input port ids: 0, 1
+
+    :param metatype: Metatype.
+    :return: Port ids.
+    """
+    if metatype in MATMUL_METATYPES:
+        return metatype.possible_weight_ports
+    return []
+
+
+def get_bias_tensor_port_id(metatype: ONNXOpWithWeightsMetatype) -> Optional[int]:
+    """
+    Returns input port id, where a bias tensor should output.
+
+    :param node: Node, for which input port id is returned,
+    :return: Input port id, where a weight bias should output or None if node can not have bias.
+    """
+    if metatype in OPERATIONS_WITH_BIAS:
+        return metatype.bias_port_id
+    return None
+
+
 def _get_weight_port_ids(node: onnx.NodeProto, onnx_graph: ONNXGraph) -> Set[int]:
     """
     Returns all weight input ports.

nncf/onnx/hardware/fused_patterns.py

@@ -14,10 +14,10 @@
 from nncf.common.graph.patterns import HWFusedPatternNames
 from nncf.common.utils.registry import Registry
 from nncf.onnx.graph.metatypes import onnx_metatypes as om
-from nncf.onnx.hardware.pattern_operations import ARITHMETIC_OPERATIONS
-from nncf.onnx.hardware.pattern_operations import ATOMIC_ACTIVATIONS_OPERATIONS
-from nncf.onnx.hardware.pattern_operations import BATCH_NORMALIZATION_OPERATIONS
-from nncf.onnx.hardware.pattern_operations import LINEAR_OPERATIONS
+from nncf.onnx.graph.metatypes.groups import ARITHMETIC_OPERATIONS
+from nncf.onnx.graph.metatypes.groups import ATOMIC_ACTIVATIONS_OPERATIONS
+from nncf.onnx.graph.metatypes.groups import BATCH_NORMALIZATION_OPERATIONS
+from nncf.onnx.graph.metatypes.groups import LINEAR_OPERATIONS
 
 ONNX_HW_FUSED_PATTERNS = Registry("onnx")
 
@@ -383,19 +383,23 @@ def create_linear_scale_shift() -> GraphPattern:
 
 def linear_operations() -> GraphPattern:
     pattern = GraphPattern()
-    pattern.add_node(**LINEAR_OPERATIONS)
+    pattern.add_node(**{GraphPattern.METATYPE_ATTR: LINEAR_OPERATIONS, GraphPattern.LABEL_ATTR: "LINEAR"})
     return pattern
 
 
 def batch_normalization_operations() -> GraphPattern:
     pattern = GraphPattern()
-    pattern.add_node(**BATCH_NORMALIZATION_OPERATIONS)
+    pattern.add_node(
+        **{GraphPattern.METATYPE_ATTR: BATCH_NORMALIZATION_OPERATIONS, GraphPattern.LABEL_ATTR: "BATCH_NORMALIZATION"}
+    )
     return pattern
 
 
 def atomic_activations_operations() -> GraphPattern:
     pattern = GraphPattern()
-    pattern.add_node(**ATOMIC_ACTIVATIONS_OPERATIONS)
+    pattern.add_node(
+        **{GraphPattern.METATYPE_ATTR: ATOMIC_ACTIVATIONS_OPERATIONS, GraphPattern.LABEL_ATTR: "ATOMIC_ACTIVATIONS"}
+    )
 
     swish_sigmoid = create_swish_with_sigmoid()
     pattern.add_pattern_alternative(swish_sigmoid)
@@ -413,7 +417,7 @@ def atomic_activations_operations() -> GraphPattern:
 
 def arithmetic_operations() -> GraphPattern:
     pattern = GraphPattern()
-    pattern.add_node(**ARITHMETIC_OPERATIONS)
+    pattern.add_node(**{GraphPattern.METATYPE_ATTR: ARITHMETIC_OPERATIONS, GraphPattern.LABEL_ATTR: "ARITHMETIC"})
     return pattern