xref: /aosp_15_r20/external/ComputeLibrary/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2021 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_CLFULLYCONNECTEDLAYER_H
25 #define ARM_COMPUTE_CLFULLYCONNECTEDLAYER_H
26 
27 #include "arm_compute/runtime/IFunction.h"
28 
29 #include "arm_compute/runtime/CL/CLTensor.h"
30 #include "arm_compute/runtime/IWeightsManager.h"
31 #include "arm_compute/runtime/MemoryGroup.h"
32 
33 namespace arm_compute
34 {
35 /** Basic function to compute a Fully Connected layer on OpenCL. This function calls the following OpenCL kernels:
36  *
37  *  -# @ref opencl::kernels::ClIm2ColKernel (called when the input comes from a convolutional layer)
38  *  -# @ref CLTranspose (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once)
39  *  -# @ref opencl::ClGemm or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
40  *
41  * @note  The fully connected layer accepts "weights" tensors only with 2 dimensions.
42  */
43 class CLFullyConnectedLayer : public IFunction
44 {
45 public:
46     /** Constructor */
47     CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr);
48     /** Default destructor */
49     ~CLFullyConnectedLayer();
50     /** Prevent instances of this class from being copied (As this class contains pointers) */
51     CLFullyConnectedLayer(const CLFullyConnectedLayer &) = delete;
52     /** Default move constructor */
53     CLFullyConnectedLayer(CLFullyConnectedLayer &&) = default;
54     /** Prevent instances of this class from being copied (As this class contains pointers) */
55     CLFullyConnectedLayer &operator=(const CLFullyConnectedLayer &) = delete;
56     /** Default move assignment operator */
57     CLFullyConnectedLayer &operator=(CLFullyConnectedLayer &&) = default;
58     /** Set the input and output tensors.
59      *
60      * Valid data layouts:
61      * - NHWC
62      * - NCHW
63      *
64      * Valid data type configurations:
65      * |src0           |src1               |src2   |dst            |
66      * |:--------------|:------------------|:------|:--------------|
67      * |F16            |F16                |F16    |F16            |
68      * |F32            |F32                |F32    |F32            |
69      * |QASYMM8        |QASYMM8            |S32    |QASYMM8        |
70      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32    |QASYMM8_SIGNED |
71      *
72      * @param[in]  compile_context The compile context to be used.
73      * @param[in]  input           Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
74      * @param[in]  weights         Weights tensor. The weights must be 2 dimensional.
75      *                             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.
76      *                             If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension.
77      *                             Data type supported: Same as @p input.
78      * @param[in]  biases          Bias tensor. Can be nullptr. Data type supported:Same as @p input.
79      * @param[out] output          Destination tensor. Its shape should be equal to the output of a matrix multiplication between:
80      *                             - The output of im2col on the input and the (transposed) 2D weights, if the function is called after a Convolution Layer
81      *                             - The input tensor and the (transposed) 2D weights, if the function is called after another FullyConnected Layer.
82      *                             Data type supported: Same as @p input.
83      * @param[in]  fc_info         (Optional) Fully connected layer additional info
84      */
85     void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
86                    FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
87     /** Set the input and output tensors.
88      *
89      * Similar to @ref CLFullyConnectedLayer
90      */
91     void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
92                    FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
93     /** Static function to check if given info will lead to a valid configuration of @ref CLFullyConnectedLayer
94      *
95      * Similar to @ref CLFullyConnectedLayer
96      *
97      * @return a status
98      */
99     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
100                            FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
101 
102     //Inherited methods override
103     void run() override;
104     void prepare() override;
105 
106 private:
107     struct Impl;
108     std::unique_ptr<Impl> _impl;
109 };
110 } // namespace arm_compute
111 #endif /* ARM_COMPUTE_CLFULLYCONNECTEDLAYER_H */
112