xref: /aosp_15_r20/external/executorch/backends/qualcomm/builders/op_softmax.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.
6from typing import cast, Dict
7
8import executorch.backends.qualcomm.python.PyQnnWrapperAdaptor as PyQnnWrapper
9
10import numpy as np
11import torch
12from executorch.backends.qualcomm.utils.constants import QCOM_AXIS_ORDER, QCOM_DATA
13
14from .node_visitor import NodeVisitor, register_node_visitor
15from .qnn_constants import OpSoftmax, QNN_OP_PACKAGE_NAME_QTI_AISW
16
17
18@register_node_visitor
19class Softmax(NodeVisitor):
20    target = ["aten._softmax.default", "aten._safe_softmax.default"]
21
22    def __init__(self, *args) -> None:
23        super().__init__(*args)
24
25    def define_node(
26        self,
27        node: torch.fx.Node,
28        nodes_to_wrappers: Dict[torch.fx.Node, PyQnnWrapper.TensorWrapper],
29    ) -> PyQnnWrapper.PyQnnOpWrapper:
30        input_node = node.args[0]
31        input_tensor = self.get_tensor(input_node, node)
32        softmax_inp_tensor_wrapper = self.define_tensor(
33            input_node,
34            input_tensor,
35            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
36            nodes_to_wrappers,
37            is_input_tensor=True,
38        )
39        softmax_input_tensors = [softmax_inp_tensor_wrapper]
40
41        output_tensor = self.get_tensor(node, node)
42        output_tensor_wrapper = self.define_tensor(
43            node,
44            output_tensor,
45            PyQnnWrapper.Qnn_TensorType_t.QNN_TENSOR_TYPE_NATIVE,
46            nodes_to_wrappers,
47            is_input_tensor=False,
48        )
49        softmax_output_tensors = [output_tensor_wrapper]
50
51        dim = cast(int, node.args[1])
52        if dim < 0:
53            dim = dim % len(input_tensor.shape)
54        if QCOM_AXIS_ORDER in node.meta:
55            dim = node.meta[QCOM_AXIS_ORDER].index(dim)
56
57        # softmax only supports last dimension for now, which is channel in QNN
58        if dim != input_tensor.dim() - 1:
59            return None
60
61        softmax_op = PyQnnWrapper.PyQnnOpWrapper(
62            node.name,
63            QNN_OP_PACKAGE_NAME_QTI_AISW,
64            OpSoftmax.op_name,
65        )
66        softmax_op.AddInputTensors(softmax_input_tensors)
67        softmax_op.AddOutputTensors(softmax_output_tensors)
68
69        softmax_op.AddScalarParam(
70            OpSoftmax.param_axis,
71            PyQnnWrapper.Qnn_DataType_t.QNN_DATATYPE_UINT_32,
72            {QCOM_DATA: np.uint32(dim)},
73        )
74
75        return softmax_op
76