xref: /aosp_15_r20/external/executorch/backends/qualcomm/builders/op_layer_norm.py (revision 523fa7a60841cd1ecfb9cc4201f1ca8b03ed023a) 
1# Copyright (c) Qualcomm Innovation Center, Inc.
2# All rights reserved
3#
4# This source code is licensed under the BSD-style license found in the
5# LICENSE file in the root directory of this source tree.
6
7import warnings
8from typing import Dict
9
10import executorch.backends.qualcomm.python.PyQnnWrapperAdaptor as PyQnnWrapper
11
12import numpy as np
13import torch
14from executorch.backends.qualcomm.utils.constants import QCOM_DATA
15
16from .node_visitor import NodeVisitor, register_node_visitor
17from .qnn_constants import OpLayerNorm, QNN_OP_PACKAGE_NAME_QTI_AISW
18from .utils import get_parameter
19
20
21@register_node_visitor
22class LayerNormVisitor(NodeVisitor):
23    target = ["aten.native_layer_norm.default"]
24
25    def __init__(self, *args) -> None:
26        super().__init__(*args)
27
28    def define_node(
29        self,
30        node: torch.fx.Node,
31        nodes_to_wrappers: Dict[torch.fx.Node, PyQnnWrapper.TensorWrapper],
32    ) -> PyQnnWrapper.PyQnnOpWrapper:
33        input_node = node.args[0]
34        input_tensor = self.get_tensor(input_node, node)
35        input_tensor_wrapper = self.define_tensor(
36            input_node,
37            input_tensor,
38            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
39            nodes_to_wrappers,
40            is_input_tensor=True,
41        )
42
43        normalized_shapes = node.args[1]
44        if (
45            len(normalized_shapes) != 1
46            and normalized_shapes[0] != input_tensor.shape[-1]
47        ):
48            warnings.warn(
49                "[QNN Delegate Op Builder]: Only supports normalization with last input dimension.",
50                stacklevel=1,
51            )
52            return
53        axis = [len(input_tensor.shape) - 1]
54        axis_shape = [len(axis)]
55
56        weight_node = node.args[2]
57        weight_tensor = get_parameter(weight_node, self.edge_program)
58        weight_tensor_wrapper = self.define_tensor(
59            weight_node,
60            weight_tensor,
61            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_STATIC,
62            nodes_to_wrappers,
63            is_input_tensor=False,
64        )
65
66        bias_node = node.args[3]
67        bias_tensor = get_parameter(bias_node, self.edge_program)
68        bias_tensor_wrapper = self.define_tensor(
69            bias_node,
70            bias_tensor,
71            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_STATIC,
72            nodes_to_wrappers,
73            is_input_tensor=False,
74        )
75
76        epsilon = node.args[4]
77
78        output_tensor = self.get_tensor(node, node, 0)
79        output_tensor_wrapper = self.define_tensor(
80            node,
81            output_tensor,
82            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
83            nodes_to_wrappers,
84            is_input_tensor=False,
85        )
86
87        layer_norm_op = PyQnnWrapper.PyQnnOpWrapper(
88            node.name,
89            QNN_OP_PACKAGE_NAME_QTI_AISW,
90            OpLayerNorm.op_name,
91        )
92        layer_norm_op.AddInputTensors(
93            [input_tensor_wrapper, weight_tensor_wrapper, bias_tensor_wrapper]
94        )
95        layer_norm_op.AddOutputTensors([output_tensor_wrapper])
96        layer_norm_op.AddScalarParam(
97            OpLayerNorm.param_epsilon,
98            PyQnnWrapper.Qnn_DataType_t.QNN_DATATYPE_FLOAT_32,
99            {QCOM_DATA: np.float32(epsilon)},
100        )
101        layer_norm_op.AddTensorParam(
102            OpLayerNorm.param_axes,
103            PyQnnWrapper.Qnn_DataType_t.QNN_DATATYPE_UINT_32,
104            len(axis_shape),
105            axis_shape,
106            np.array(axis, dtype=np.uint32),
107            True,
108        )
109
110        return layer_norm_op
111