xref: /aosp_15_r20/external/armnn/src/armnn/layers/QuantizedLstmLayer.hpp (revision 89c4ff92f2867872bb9e2354d150bf0c8c502810)
1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #pragma once
6 
7 #include <Layer.hpp>
8 
9 namespace armnn
10 {
11 
12 class ScopedTensorHandle;
13 
14 struct QuantizedLstmParameters
15 {
16     /// A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8).
17     std::shared_ptr<ConstTensorHandle> m_InputToInputWeights;
18     /// A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8).
19     std::shared_ptr<ConstTensorHandle> m_InputToForgetWeights;
20     /// A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8).
21     std::shared_ptr<ConstTensorHandle> m_InputToCellWeights;
22     /// A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8).
23     std::shared_ptr<ConstTensorHandle> m_InputToOutputWeights;
24 
25     /// A unique pointer to represent 2D weights tensor with dimensions [outputSize, outputSize] (QAsymm8).
26     std::shared_ptr<ConstTensorHandle> m_RecurrentToInputWeights;
27     /// A unique pointer to represent 2D weights tensor with dimensions [outputSize, outputSize] (QAsymm8).
28     std::shared_ptr<ConstTensorHandle> m_RecurrentToForgetWeights;
29     /// A unique pointer to represent 2D weights tensor with dimensions [outputSize, outputSize] (QAsymm8).
30     std::shared_ptr<ConstTensorHandle> m_RecurrentToCellWeights;
31     /// A unique pointer to represent 2D weights tensor with dimensions [outputSize, outputSize] (QAsymm8).
32     std::shared_ptr<ConstTensorHandle> m_RecurrentToOutputWeights;
33 
34     /// A unique pointer to represent 1D bias tensor with dimensions [outputSize] (int32).
35     std::shared_ptr<ConstTensorHandle> m_InputGateBias;
36     /// A unique pointer to represent 1D bias tensor with dimensions [outputSize] (int32).
37     std::shared_ptr<ConstTensorHandle> m_ForgetGateBias;
38     /// A unique pointer to represent 1D bias tensor with dimensions [outputSize] (int32).
39     std::shared_ptr<ConstTensorHandle> m_CellBias;
40     /// A unique pointer to represent 1D bias tensor with dimensions [outputSize] (int32).
41     std::shared_ptr<ConstTensorHandle> m_OutputGateBias;
42 };
43 
44 /// This layer represents a QuantizedLstm operation.
45 class QuantizedLstmLayer : public Layer
46 {
47 public:
48 
49     QuantizedLstmParameters m_QuantizedLstmParameters;
50 
51     /// Makes a workload for the QuantizedLstm type.
52     /// @param [in] graph The graph where this layer can be found.
53     /// @param [in] factory The workload factory which will create the workload.
54     /// @return A pointer to the created workload, or nullptr if not created.
55     virtual std::unique_ptr<IWorkload> CreateWorkload(const IWorkloadFactory& factory) const override;
56 
57     /// Creates a dynamically-allocated copy of this layer.
58     /// @param [in] graph The graph into which this layer is being cloned.
59     QuantizedLstmLayer* Clone(Graph& graph) const override;
60 
61     /// Check if the input tensor shape(s)
62     /// will lead to a valid configuration of @ref QuantizedLstmLayer.
63     /// @param [in] shapeInferenceMethod Indicates if output shape shall be overwritten or just validated.
64     void ValidateTensorShapesFromInputs() override;
65 
66     /// By default returns inputShapes if the number of inputs are equal to number of outputs,
67     /// otherwise infers the output shapes from given input shapes and layer properties.
68     /// @param [in] inputShapes The input shapes layer has.
69     /// @return A vector to the inferred output shape.
70     std::vector<TensorShape> InferOutputShapes(const std::vector<TensorShape>& inputShapes) const override;
71 
72     void ExecuteStrategy(IStrategy& strategy) const override;
73 
74 protected:
75     /// Constructor to create a QuantizedLstmLayer.
76     /// @param [in] name Optional name for the layer.
77     QuantizedLstmLayer(const char* name);
78 
79     /// Default destructor
80     ~QuantizedLstmLayer() = default;
81 
82     /// Retrieve the handles to the constant values stored by the layer.
83     /// @return A vector of the constant tensors stored by this layer.
84     Layer::ImmutableConstantTensors GetConstantTensorsByRef() const override;
85 };
86 
87 } // namespace armnn
88