xref: /aosp_15_r20/external/ComputeLibrary/arm_compute/runtime/CL/functions/CLDirectDeconvolutionLayer.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2019-2021 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_CLDIRECTDECONVOLUTIONLAYER_H
25 #define ARM_COMPUTE_CLDIRECTDECONVOLUTIONLAYER_H
26 
27 #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h"
28 #include "arm_compute/runtime/CL/functions/CLDeconvolutionLayerUpsample.h"
29 #include "arm_compute/runtime/CL/functions/CLReverse.h"
30 #include "arm_compute/runtime/CL/functions/CLTranspose.h"
31 
32 #include "arm_compute/runtime/CL/CLTensor.h"
33 #include "arm_compute/runtime/IFunction.h"
34 #include "arm_compute/runtime/IMemoryManager.h"
35 #include "arm_compute/runtime/MemoryGroup.h"
36 
37 #include <memory>
38 
39 namespace arm_compute
40 {
41 class ICLTensor;
42 /** Function to run the deconvolution layer.
43  *
44  * Deconvolution Layer is the backward pass of Convolution Layer. First we transform the input depending on the stride and pad info and then perform a 1x1
45  * convolution pass. Input stride defines how many zeroes we should put between each element of the input and pad is the amount of padding.
46  *
47  *  The relation between input to output is as follows:
48  *  \f[
49  *       width\_output = (width\_input - 1) \cdot stride\_x - 2 \cdot padding\_x + kernel\_x
50  *  \f]
51  *  \f[
52  *       height\_output = (height\_input - 1) \cdot stride\_y - 2 \cdot padding\_y + kernel\_y
53  *  \f]
54  *
55  *  where:
56  *      width_input is the size of the first input dimension.
57  *      height_input is the size of the second input dimension.
58  *      width_output is the size of the first output dimension.
59  *      height_output is the size of the second output dimension.
60  *      kernel_x and kernel_y are the convolution sizes in x and y.
61  *      stride_x and stride_y is the input stride of the first and second dimension.
62  *
63  * The weights used by Deconvolution are supposed to be the same as the ones used for Convolution. Therefore, it will be necessary to use the weights in the
64  * reverse order to perform an actual convolution. This is achieved by using @ref CLReverse.
65  *
66  * This function calls the following OpenCL kernels/functions:
67  *
68  * -# @ref CLDeconvolutionLayerUpsample
69  * -# @ref CLConvolutionLayer
70  *
71  * And the following CPP kernels:
72  * -# @ref CLReverse
73  *
74  */
75 class CLDirectDeconvolutionLayer : public IFunction
76 {
77 public:
78     /** Constructor */
79     CLDirectDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
80     /** Prevent instances of this class from being copied (As this class contains pointers) */
81     CLDirectDeconvolutionLayer(const CLDirectDeconvolutionLayer &) = delete;
82     /** Default move constructor */
83     CLDirectDeconvolutionLayer(CLDirectDeconvolutionLayer &&) = default;
84     /** Prevent instances of this class from being copied (As this class contains pointers) */
85     CLDirectDeconvolutionLayer &operator=(const CLDirectDeconvolutionLayer &) = delete;
86     /** Default move assignment operator */
87     CLDirectDeconvolutionLayer &operator=(CLDirectDeconvolutionLayer &&) = default;
88     /** Set the input, weights, biases and output tensors.
89      *
90      * Valid data layouts:
91      * - NHWC
92      * - NCHW
93      *
94      * Valid data type configurations:
95      * |src0           |src1               |src2   |dst            |
96      * |:--------------|:------------------|:------|:--------------|
97      * |F16            |F16                |F16    |F16            |
98      * |F32            |F32                |F32    |F32            |
99      * |QASYMM8        |QASYMM8            |S32    |QASYMM8        |
100      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32    |QASYMM8_SIGNED |
101      * |QASYMM8        |QSYMM8_PER_CHANNEL |S32    |QASYMM8        |
102      * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32    |QASYMM8_SIGNED |
103      *
104      * @param[in,out] input        Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
105      *                             Data types supported: QASYMM8_SIGNED/QASYMM8/F16/F32.
106      * @param[in]     weights      The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input or QSYMM8_PER_CHANNEL if @p input is QASYMM8/QASYMM8_SIGNED.
107      * @param[in]     bias         (Optional) The biases have one dimension.
108      *                             Data type supported: Should match @p input data type, except for input of QASYMM8 and QASYMM8_SIGNED type where biases should be of S32 type
109      * @param[out]    output       Output tensor. The output has the same number of dimensions as the @p input.
110      * @param[in]     info         Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
111      * @param[in]     weights_info (Optional) Weights information needed for @ref CLConvolutionLayer, specifies if the weights tensor has been reshaped with @ref opencl::kernels::ClWeightsReshapeKernel.
112      *
113      */
114     void configure(ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info, const WeightsInfo &weights_info = WeightsInfo());
115     /** Set the input, weights, biases and output tensors.
116      *
117      * @param[in]     compile_context The compile context to be used.
118      * @param[in,out] input           Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
119      *                                Data types supported: QASYMM8_SIGNED/QASYMM8/F16/F32.
120      * @param[in]     weights         The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input or QSYMM8_PER_CHANNEL if @p input is QASYMM8/QASYMM8_SIGNED.
121      * @param[in]     bias            (Optional) The biases have one dimension.
122      *                                Data type supported: Should match @p input data type, except for input of QASYMM8 and QASYMM8_SIGNED type where biases should be of S32 type
123      * @param[out]    output          Output tensor. The output has the same number of dimensions as the @p input.
124      * @param[in]     info            Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
125      * @param[in]     weights_info    (Optional) Weights information needed for @ref CLConvolutionLayer, specifies if the weights tensor has been reshaped with @ref opencl::kernels::ClWeightsReshapeKernel.
126      *
127      */
128     void configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
129                    const WeightsInfo &weights_info = WeightsInfo());
130     /** Static function to check if given info will lead to a valid configuration of @ref CLDirectDeconvolutionLayer
131      *
132      * @param[in] input        Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
133      *                         Data types supported: QASYMM8_SIGNED/QASYMM8/F16/F32.
134      * @param[in] weights      The 4d weights info with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input or QSYMM8_PER_CHANNEL if @p input is QASYMM8/QASYMM8_SIGNED.
135      * @param[in] bias         (Optional) The biases have one dimension.
136      *                         Data type supported: Should match @p input data type, except for input of QASYMM8 and QASYMM8_SIGNED type where biases should be of S32 type
137      * @param[in] output       Output tensor info. The output has the same number of dimensions as the @p input.
138      * @param[in] info         Contains padding and policies to be used in the deconvolution, this is decribed in @ref PadStrideInfo.
139      * @param[in] weights_info (Optional) Weights information needed for @ref CLConvolutionLayer, specifies if the weights tensor has been reshaped with @ref opencl::kernels::ClWeightsReshapeKernel.
140      *
141      * @return a status
142      */
143     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &info,
144                            const WeightsInfo &weights_info = WeightsInfo());
145 
146     // Inherited methods overridden:
147     void run() override;
148     void prepare() override;
149 
150 private:
151     MemoryGroup                  _memory_group;
152     CLDeconvolutionLayerUpsample _scale_f;
153     CLConvolutionLayer           _conv_f;
154     CLReverse                    _flip_weights;
155 
156     CLTensor   _scaled_output;
157     ICLTensor *_original_weights;
158     CLTensor   _weights_flipped;
159     CLTensor   _flip_axis;
160 
161     bool _is_prepared;
162 };
163 } // namespace arm_compute
164 #endif /* ARM_COMPUTE_CLDECONVOLUTIONLAYER_H */
165