2 * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 * Copyright (c) 2019-2020 ARM Limited.
20 * SPDX-License-Identifier: MIT
22 * Permission is hereby granted, free of charge, to any person obtaining a copy
23 * of this software and associated documentation files (the "Software"), to
24 * deal in the Software without restriction, including without limitation the
25 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
26 * sell copies of the Software, and to permit persons to whom the Software is
27 * furnished to do so, subject to the following conditions:
29 * The above copyright notice and this permission notice shall be included in all
30 * copies or substantial portions of the Software.
32 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
33 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
34 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
35 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
36 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
37 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
40 #ifndef __ARM_COMPUTE_CLDIRECTTRANSPOSECONVLAYER_H__
41 #define __ARM_COMPUTE_CLDIRECTTRANSPOSECONVLAYER_H__
43 #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h"
44 #include "arm_compute/runtime/CL/functions/CLDeconvolutionLayerUpsample.h"
45 #include "arm_compute/runtime/CL/functions/CLReverse.h"
46 #include "arm_compute/runtime/CL/functions/CLTranspose.h"
48 #include "arm_compute/runtime/CL/CLTensor.h"
49 #include "arm_compute/runtime/IFunction.h"
50 #include "arm_compute/runtime/IMemoryManager.h"
51 #include "arm_compute/runtime/MemoryGroup.h"
58 /** Function to run the deconvolution layer.
60 * Deconvolution Layer is the backward pass of Convolution Layer. First we transform the input
61 * depending on the stride and pad info and then perform a 1x1
62 * convolution pass. Input stride defines how many zeroes we should put between each element of the
63 * input and pad is the amount of padding.
65 * The relation between input to output is as follows:
67 * width\_output = (width\_input - 1) \cdot stride\_x - 2 \cdot padding\_x + kernel\_x
70 * height\_output = (height\_input - 1) \cdot stride\_y - 2 \cdot padding\_y + kernel\_y
74 * width_input is the size of the first input dimension.
75 * height_input is the size of the second input dimension.
76 * width_output is the size of the first output dimension.
77 * height_output is the size of the second output dimension.
78 * kernel_x and kernel_y are the convolution sizes in x and y.
79 * stride_x and stride_y is the input stride of the first and second dimension.
81 * The weights used by Deconvolution are supposed to be the same as the ones used for Convolution.
82 * Therefore, it will be necessary to use the weights in the
83 * reverse order to perform an actual convolution. This is achieved by using @ref CLReverse.
85 * This function calls the following OpenCL kernels/functions:
87 * -# @ref CLDeconvolutionLayerUpsample
88 * -# @ref CLConvolutionLayer
90 * And the following CPP kernels:
94 class CLDirectTransposeConvLayer : public IFunction
98 CLDirectTransposeConvLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
99 /** Prevent instances of this class from being copied (As this class contains pointers) */
100 CLDirectTransposeConvLayer(const CLDirectTransposeConvLayer &) = delete;
101 /** Default move constructor */
102 CLDirectTransposeConvLayer(CLDirectTransposeConvLayer &&) = default;
103 /** Prevent instances of this class from being copied (As this class contains pointers) */
104 CLDirectTransposeConvLayer &operator=(const CLDirectTransposeConvLayer &) = delete;
105 /** Default move assignment operator */
106 CLDirectTransposeConvLayer &operator=(CLDirectTransposeConvLayer &&) = default;
107 /** Set the input, weights, biases and output tensors.
109 * @param[in,out] input Input tensor. 3 lower dimensions represent a single input, and an
110 * optional 4th dimension for batch of inputs.
111 * Data types supported: QASYMM8_SIGNED/QASYMM8/F16/F32.
112 * @param[in] weights The 4d weights with dimensions [width, height, IFM, OFM]. Data type
113 * supported: Same as @p input.
114 * @param[in] bias (Optional) The biases have one dimension.
115 * Data type supported: Should match @p input data type, except for
116 * input of QASYMM8 and QASYMM8_SIGNED type where biases should be of S32 type
117 * @param[out] output Output tensor. The output has the same number of dimensions as the
119 * @param[in] info Contains padding and policies to be used in the deconvolution, this
120 * is decribed in @ref PadStrideInfo.
121 * @param[in] invalid_right The number of zeros added to right edge of the output.
122 * @param[in] invalid_bottom The number of zeros added to bottom edge of the output.
123 * @param[in] weights_info (Optional) Weights information needed for @ref CLConvolutionLayer,
124 * specifies if the weights tensor has been reshaped with @ref CLWeightsReshapeKernel.
127 void configure(ICLTensor *input, ICLTensor *weights, const ICLTensor *bias, ICLTensor *output,
128 const PadStrideInfo &info, unsigned int invalid_right, unsigned int invalid_bottom,
129 const WeightsInfo &weights_info = WeightsInfo());
130 /** Set the input, weights, biases and output tensors.
132 * @param[in] compile_context The compile context to be used.
133 * @param[in,out] input Input tensor. 3 lower dimensions represent a single input, and
134 * an optional 4th dimension for batch of inputs.
135 * Data types supported: QASYMM8_SIGNED/QASYMM8/F16/F32.
136 * @param[in] weights The 4d weights with dimensions [width, height, IFM, OFM]. Data
137 * type supported: Same as @p input.
138 * @param[in] bias (Optional) The biases have one dimension.
139 * Data type supported: Should match @p input data type, except for
140 * input of QASYMM8 and QASYMM8_SIGNED type where biases should be of S32 type
141 * @param[out] output Output tensor. The output has the same number of dimensions as
143 * @param[in] info Contains padding and policies to be used in the deconvolution,
144 * this is decribed in @ref PadStrideInfo.
145 * @param[in] invalid_right The number of zeros added to right edge of the output.
146 * @param[in] invalid_bottom The number of zeros added to bottom edge of the output.
147 * @param[in] weights_info (Optional) Weights information needed for @ref
148 * CLConvolutionLayer, specifies if the weights tensor has been reshaped with @ref
149 * CLWeightsReshapeKernel.
152 void configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *weights,
153 const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &info,
154 unsigned int invalid_right, unsigned int invalid_bottom,
155 const WeightsInfo &weights_info = WeightsInfo());
156 /** Static function to check if given info will lead to a valid configuration of @ref
157 * CLDirectTransposeConvLayer
159 * @param[in] input Input tensor info. 3 lower dimensions represent a single input, and an
160 * optional 4th dimension for batch of inputs.
161 * Data types supported: QASYMM8_SIGNED/QASYMM8/F16/F32.
162 * @param[in] weights The 4d weights info with dimensions [width, height, IFM, OFM]. Data
163 * type supported: Same as @p input.
164 * @param[in] bias (Optional) The biases have one dimension.
165 * Data type supported: Should match @p input data type, except for input
166 * of QASYMM8 and QASYMM8_SIGNED type where biases should be of S32 type
167 * @param[in] output Output tensor info. The output has the same number of dimensions as the
169 * @param[in] info Contains padding and policies to be used in the deconvolution, this is
170 * decribed in @ref PadStrideInfo.
171 * @param[in] invalid_right The number of zeros added to right edge of the output.
172 * @param[in] invalid_bottom The number of zeros added to bottom edge of the output.
173 * @param[in] weights_info (Optional) Weights information needed for @ref CLConvolutionLayer,
174 * specifies if the weights tensor has been reshaped with @ref CLWeightsReshapeKernel.
178 static Status validate(const ITensorInfo *input, const ITensorInfo *weights,
179 const ITensorInfo *bias, ITensorInfo *output, const PadStrideInfo &info,
180 unsigned int invalid_right, unsigned int invalid_bottom,
181 const WeightsInfo &weights_info = WeightsInfo());
183 // Inherited methods overridden:
185 void prepare() override;
188 MemoryGroup _memory_group;
189 CLDeconvolutionLayerUpsample _scale_f;
190 CLConvolutionLayer _conv_f;
191 CLReverse _flip_weights;
193 CLTensor _scaled_output;
194 ICLTensor *_original_weights;
195 CLTensor _weights_flipped;
200 } // namespace arm_compute
201 #endif /* __ARM_COMPUTE_CLDIRECTTRANSPOSECONVLAYER_H__ */