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) 2017-2019 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
41 #ifndef __ARM_COMPUTE_CLFULLYCONNECTEDLAYEREX_H__
42 #define __ARM_COMPUTE_CLFULLYCONNECTEDLAYEREX_H__
44 #include "arm_compute/runtime/CL/ICLSimpleFunction.h"
46 #include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h"
47 #include "arm_compute/core/CL/kernels/CLTransposeKernel.h"
48 #include "arm_compute/runtime/CL/CLTensor.h"
49 #include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h"
50 #include "arm_compute/runtime/CL/functions/CLFlattenLayer.h"
51 #include "arm_compute/runtime/CL/functions/CLGEMM.h"
52 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
53 #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
54 #include "arm_compute/runtime/IWeightsManager.h"
55 #include "arm_compute/runtime/MemoryGroup.h"
59 /** Basic function to reshape the weights of Fully Connected layer with OpenCL. This function calls
60 * the following kernels:
62 * -# @ref CLTransposeKernel
64 * @note The fully connected layer accepts "weights" tensors only with 2 dimensions.
66 class CLFullyConnectedLayerReshapeWeightsEx : public ICLSimpleFunction
69 /** Set the input and output tensors.
71 * @param[in] input Weights tensor. The weights must be 2 dimensional. Data types supported:
73 * @param[out] output Destination tensor which stores the transposed input tensor. Data type
74 * supported: Same as @p input.
76 void configure(const ICLTensor *input, ICLTensor *output);
77 /** Static function to check if given info will lead to a valid configuration of @ref
78 * CLFullyConnectedLayerReshapeWeightsEx
80 * @param[in] input Weights tensor. The weights must be 2 dimensional. Data types supported:
82 * @param[in] output Destination tensor which stores the transposed input tensor. Data type
83 * supported: Same as @p input.
87 static Status validate(const ITensorInfo *input, const ITensorInfo *output);
90 namespace weights_transformations
92 /** Basic function to manage the reshape weights generated from @ref
93 * CLFullyConnectedLayerReshapeWeightsEx */
94 class CLFullyConnectedLayerReshapeWeightsExManaged : public ITransformWeights
97 // Inherited method override
100 _output.allocator()->allocate();
105 // Inherited method override
106 void release() override { _output.allocator()->free(); }
108 // Inherited method override
109 ICLTensor *get_weights() override { return &_output; }
111 // Inherited method override
112 uint32_t uid() override { return _uid; }
114 /** Configures the @ref CLFullyConnectedLayerReshapeWeightsEx function
116 * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32.
118 void configure(const ICLTensor *input) { _func.configure(input, &_output); }
121 static constexpr uint32_t _uid = 0x0;
123 CLFullyConnectedLayerReshapeWeightsEx _func{};
125 } // namespace weights_transformations
127 /** Basic function to compute a Fully Connected layer on OpenCL. This function calls the following
130 * -# @ref CLIm2ColKernel (called when the input comes from a convolutional layer)
131 * -# @ref CLFullyConnectedLayerReshapeWeightsEx (if @p are_weights_reshaped is set to false and
132 * transpose_weights is set to true ) (called once)
133 * -# @ref CLGEMMMatrixMultiplyKernel or @ref CLGEMMLowpMatrixMultiplyCore (if quantized
135 * -# @ref CLGEMMMatrixAccumulateBiasesKernel or @ref
136 * CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if quantized asymmetric) (if @p biases is
137 * not equal to nullptr)
139 * @note The fully connected layer accepts "weights" tensors only with 2 dimensions.
141 class CLFullyConnectedLayerEx : public IFunction
145 CLFullyConnectedLayerEx(std::shared_ptr<IMemoryManager> memory_manager = nullptr,
146 IWeightsManager *weights_manager = nullptr);
147 /** Prevent instances of this class from being copied (As this class contains pointers) */
148 CLFullyConnectedLayerEx(const CLFullyConnectedLayerEx &) = delete;
149 /** Default move constructor */
150 CLFullyConnectedLayerEx(CLFullyConnectedLayerEx &&) = default;
151 /** Prevent instances of this class from being copied (As this class contains pointers) */
152 CLFullyConnectedLayerEx &operator=(const CLFullyConnectedLayerEx &) = delete;
153 /** Default move assignment operator */
154 CLFullyConnectedLayerEx &operator=(CLFullyConnectedLayerEx &&) = default;
155 /** Set the input and output tensors.
157 * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32.
158 * @param[in] weights Weights tensor. The weights must be 2 dimensional.
159 * If this function is called after a Convolution Layer, the (transposed)
160 * weights will have as many rows as the product of the first 3 input's dimensions.
161 * If it is called after another FullyConnected Layer, the (transposed)
162 * weights will have as many rows as the input's first dimension.
163 * Data type supported: Same as @p input.
164 * @param[in] biases Bias tensor. Can be nullptr. Data type supported:Same as @p input.
165 * @param[out] output Destination tensor. Its shape should be equal to the output of a matrix
166 * multiplication between:
167 * - The output of im2col on the input and the (transposed) 2D weights, if the
168 * function is called after a Convolution Layer
169 * - The input tensor and the (transposed) 2D weights, if the function is
170 * called after another FullyConnected Layer.
171 * Data type supported: Same as @p input.
172 * @param[in] fc_info (Optional) Fully connected layer additional info
174 void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases,
175 ICLTensor *output, FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
176 /** Static function to check if given info will lead to a valid configuration of @ref
177 * CLFullyConnectedLayerEx
179 * @param[in] input Source tensor info. Data type supported: QASYMM8/F16/F32.
180 * @param[in] weights Weights tensor info. The weights must be 2 dimensional.
181 * If this function is called after a Convolution Layer, the (transposed)
182 * weights will have as many rows as the product of the first 3 input's dimensions.
183 * If it is called after another FullyConnected Layer, the (transposed)
184 * weights will have as many rows as the input's first dimension.
185 * Data type supported: Same as @p input.
186 * @param[in] biases Bias tensor info. Can be nullptr. Data type supported:Same as @p input.
187 * @param[out] output Destination tensor info. Its shape should be equal to the output of a
188 * matrix multiplication between:
189 * - The output of im2col on the input and the (transposed) 2D weights, if the
190 * function is called after a Convolution Layer
191 * - The input tensor and the (transposed) 2D weights, if the function is
192 * called after another FullyConnected Layer.
193 * Data type supported: Same as @p input.
194 * @param[in] fc_info (Optional) Fully connected layer additional info
198 static Status validate(const ITensorInfo *input, const ITensorInfo *weights,
199 const ITensorInfo *biases, const ITensorInfo *output,
200 FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
202 // Inherited methods override
204 void prepare() override;
207 void configure_fc_fc(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias,
208 ICLTensor *output, const FullyConnectedLayerInfo &fc_info);
209 void configure_conv_fc(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias,
210 ICLTensor *output, const FullyConnectedLayerInfo &fc_info);
211 void configure_mm(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias,
212 ICLTensor *output, const FullyConnectedLayerInfo &fc_info);
214 MemoryGroup _memory_group;
215 IWeightsManager *_weights_manager;
216 CLConvertFullyConnectedWeights _convert_weights;
217 weights_transformations::CLConvertFullyConnectedWeightsManaged _convert_weights_managed;
218 weights_transformations::CLFullyConnectedLayerReshapeWeightsExManaged
219 _reshape_weights_managed_function;
220 CLFlattenLayer _flatten_layer;
221 CLFullyConnectedLayerReshapeWeightsEx _reshape_weights_function;
223 CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp;
224 CLTensor _flatten_output;
225 CLTensor _converted_weights_output;
226 CLTensor _reshape_weights_output;
227 bool _are_weights_converted;
228 bool _are_weights_reshaped;
229 bool _is_fc_after_conv;
232 const ICLTensor *_original_weights;
234 } // namespace arm_compute
235 #endif /* __ARM_COMPUTE_CLFULLYCONNECTEDLAYEREX_H__ */