xref: /aosp_15_r20/external/ComputeLibrary/arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2023 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #ifndef ARM_COMPUTE_NEFULLYCONNECTEDLAYER_H
25 #define ARM_COMPUTE_NEFULLYCONNECTEDLAYER_H
26 
27 #include "arm_compute/runtime/IFunction.h"
28 #include "arm_compute/runtime/IMemoryManager.h"
29 #include "arm_compute/runtime/IWeightsManager.h"
30 
31 #include "arm_compute/runtime/NEON/functions/NETranspose.h"
32 #include "arm_compute/runtime/Tensor.h"
33 
34 #include <memory>
35 
36 namespace arm_compute
37 {
38 namespace weights_transformations
39 {
40 /** Basic function to manage the reshape weights generated from @ref NETranspose */
41 class NEFullyConnectedLayerReshapeWeightsManaged : public ITransformWeights
42 {
43 public:
run()44     void run() override
45     {
46         _output.allocator()->allocate();
47         _func.run();
48         _reshape_run = true;
49     }
50 
release()51     void release() override
52     {
53         _output.allocator()->free();
54     }
55 
get_weights()56     ITensor *get_weights() override
57     {
58         return &_output;
59     }
60 
uid()61     uint32_t uid() override
62     {
63         return _uid;
64     }
65 
configure(const ITensor * input)66     void configure(const ITensor *input)
67     {
68         _func.configure(input, &_output);
69     }
70 
71 private:
72     static constexpr uint32_t _uid = 0x0;
73     Tensor                    _output{};
74     NETranspose               _func{};
75 };
76 } // namespace weights_transformations
77 
78 /** Basic function to compute a Fully Connected layer. This function calls the following kernels:
79  *  -# @ref cpu::kernels::CpuIm2ColKernel (called when the input comes from a convolutional layer)
80  *  -# @ref NETranspose (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
81  *  -# @ref NEGEMM or @ref NEGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
82  *  -# @ref cpu::kernels::CpuGemmMatrixAdditionKernel or @ref NEGEMMLowpOutputStage (if quantized asymmetric) (if @p biases is not equal to nullptr)
83  *
84  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
85  */
86 class NEFullyConnectedLayer : public IFunction
87 {
88 public:
89     /** Constructor */
90     NEFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr);
91     /** Prevent instances of this class from being copied (As this class contains pointers) */
92     NEFullyConnectedLayer(const NEFullyConnectedLayer &) = delete;
93     /** Prevent instances of this class from being moved (As this class contains pointers) */
94     NEFullyConnectedLayer(NEFullyConnectedLayer &&) = delete;
95     /** Prevent instances of this class from being copied (As this class contains pointers) */
96     NEFullyConnectedLayer &operator=(const NEFullyConnectedLayer &) = delete;
97     /** Prevent instances of this class from being moved (As this class contains pointers) */
98     NEFullyConnectedLayer &operator=(NEFullyConnectedLayer &&) = delete;
99     /** Default destructor */
100     ~NEFullyConnectedLayer();
101     /** Set the input and output tensors.
102      *
103      * Valid data layouts:
104      * - NHWC
105      * - NCHW
106      *
107      * Valid data type configurations:
108      * |src0           |src1               |src2   |dst            |
109      * |:--------------|:------------------|:------|:--------------|
110      * |F16            |F16                |F16    |F16            |
111      * |F32            |F32                |F32    |F32            |
112      * |QASYMM8        |QASYMM8            |S32    |QASYMM8        |
113      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32    |QASYMM8_SIGNED |
114      *
115      * @param[in]  input        Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
116      * @param[in]  weights      Weights tensor. The weights must be 2 dimensional.
117      *                          If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions.
118      *                          If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
119      *                          Data type supported: Same as @p input.
120      * @param[in]  biases       Bias tensor. Can be nullptr. Data type supported: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED.
121      * @param[out] output       Destination tensor. Its shape should be equal to the output of a matrix multiplication between:
122      *                          - The output of im2col on the input and the (transposed) 2D weights, if the function is called after a Convolution Layer
123      *                          - The input tensor and the (transposed) 2D weights, if the function is called after another FullyConnected Layer.
124      *                          Data type supported: Same as @p input.
125      * @param[in]  fc_info      (Optional) Fully connected layer additional info
126      * @param[in]  weights_info (Optional) Stores neccessary compute information when weights are already reshaped
127      */
128     void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output,
129                    FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo(), const WeightsInfo &weights_info = WeightsInfo());
130     /** Static function to check if given info will lead to a valid configuration of @ref NEFullyConnectedLayer
131      *
132      * Similar to @ref NEFullyConnectedLayer::configure()
133      *
134      * @return a status
135      */
136     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
137                            FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo(), const WeightsInfo &weights_info = WeightsInfo());
138 
139     /** Static function that queries whether fixed-format kernel exists for a given problem description
140      *
141      * @param[out] expected_weight_format Format in which weights should be for found fixed format kernel
142      * @param[in]  input                  Source tensor
143      * @param[in]  weights                Weights tensor.
144      * @param[in]  biases                 Bias tensor. Can be nullptr. Data type supported: Same as @p weights, S32 if @p weights is QASYMM8/QASYMM8_SIGNED.
145      * @param[in]  output                 Destination tensor
146      * @param[in]  fc_info                Fully connected layer additional info
147      * @param[in]  weights_info           Describes weights shape
148      *
149      * @return a status
150      */
151     static Status has_opt_impl(arm_compute::WeightFormat &expected_weight_format, const ITensorInfo *input, const ITensorInfo *weights,
152                                const ITensorInfo *biases, const ITensorInfo *output, const FullyConnectedLayerInfo &fc_info, const WeightsInfo &weights_info);
153 
154     //Inherited methods override
155     void run() override;
156     void prepare() override;
157 
158 private:
159     struct Impl;
160     std::unique_ptr<Impl> _impl;
161 };
162 } // namespace arm_compute
163 #endif /* ARM_COMPUTE_NEFULLYCONNECTEDLAYER_H */
164