linear_squeeze_arithmetical_activations pattern

nncf/common/graph/patterns/patterns.py

@@ -347,6 +347,7 @@ class HWFusedPatternNames(Enum):
     LINEAR_SCALE_SHIFT_ACTIVATIONS = PatternDesc("linear_scale_shift_activations")
     LINEAR_CONST_MULTIPLY = PatternDesc("linear_const_multiply")
     LINEAR_SQUEEZE_ACTIVATIONS = PatternDesc("linear_squeeze_activations")
+    LINEAR_SQUEEZE_ARITHMETIC_ACTIVATIONS = PatternDesc("linear_squeeze_arithmetic_activations")
     LINEAR_ACTIVATIONS_UNSQUEEZE_BN_SQUEEZE = PatternDesc("linear_activations_unsqueeze_bn_squeeze")
     SCALE_SHIFT_ACTIVATIONS = PatternDesc("scale_shift_activations")
     MVN_SCALE_SHIFT_ACTIVATIONS = PatternDesc("mvn_scale_shift_activations")

nncf/onnx/graph/metatypes/groups.py

@@ -12,6 +12,7 @@
 from nncf.onnx.graph.metatypes import onnx_metatypes
 
 QUANTIZE_AGNOSTIC_OPERATIONS = [
+    onnx_metatypes.ONNXGlobalMaxPoolMetatype,
     onnx_metatypes.ONNXMaxPoolMetatype,
     onnx_metatypes.ONNXReduceMaxMetatype,
     onnx_metatypes.ONNXReshapeMetatype,

nncf/onnx/graph/metatypes/onnx_metatypes.py

@@ -190,6 +190,13 @@ class ONNXAveragePoolMetatype(ONNXOpMetatype):
     hw_config_names = [HWConfigOpName.AVGPOOL]
 
 
+@ONNX_OPERATION_METATYPES.register()
+class ONNXGlobalMaxPoolMetatype(ONNXOpMetatype):
+    name = "GlobalMaxPoolOp"
+    op_names = ["GlobalMaxPool"]
+    hw_config_names = [HWConfigOpName.MAXPOOL]
+
+
 @ONNX_OPERATION_METATYPES.register()
 class ONNXMaxPoolMetatype(ONNXOpMetatype):
     name = "MaxPoolOp"

nncf/onnx/hardware/fused_patterns.py

@@ -339,6 +339,17 @@ def create_linear_squeeze_activation() -> GraphPattern:
     return linear
 
 
+@ONNX_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_SQUEEZE_ARITHMETIC_ACTIVATIONS)
+def create_linear_squeeze_arithmetic_activation() -> GraphPattern:
+    linear = linear_operations()
+    squeeze = squeeze_operation()
+    arithmetic_activations = create_arithmetic_activations()
+
+    linear.join_patterns(squeeze)
+    linear.join_patterns(arithmetic_activations)
+    return linear
+
+
 @ONNX_HW_FUSED_PATTERNS.register(HWFusedPatternNames.BATCH_NORM_SCALE_SHIFT_ACTIVATIONS)
 def create_bn_scale_shift_activation() -> GraphPattern:
     batch_norm = batch_normalization_operations()

nncf/openvino/hardware/fused_patterns.py

@@ -595,6 +595,17 @@ def create_linear_squeeze_activation() -> GraphPattern:
     return linear
 
 
+@OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_SQUEEZE_ARITHMETIC_ACTIVATIONS)
+def create_linear_squeeze_arithmetic_activation() -> GraphPattern:
+    linear = linear_operations()
+    squeeze = squeeze_operation()
+    arithmetic_activations = create_arithmetic_activations()
+
+    linear.join_patterns(squeeze)
+    linear.join_patterns(arithmetic_activations)
+    return linear
+
+
 @OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.MVN_SCALE_SHIFT_ACTIVATIONS)
 def create_mvn_scale_shift_activations() -> GraphPattern:
     pattern = GraphPattern()

tests/torch/test_pattern_manager.py

@@ -63,6 +63,7 @@
     HWFusedPatternNames.LINEAR_BIASED_ACTIVATION_ELEMENTWISE: "Not relevant for Torch.",
     HWFusedPatternNames.MVN_SCALE_SHIFT_ACTIVATIONS: "Not relevant for Torch.",
     HWFusedPatternNames.LINEAR_SQUEEZE_ACTIVATIONS: "Not relevant for Torch.",
+    HWFusedPatternNames.LINEAR_SQUEEZE_ARITHMETIC_ACTIVATIONS: "Not relevant for Torch.",
     HWFusedPatternNames.LINEAR_ACTIVATIONS_UNSQUEEZE_BN_SQUEEZE: "Not relevant for Torch.",
 }