2 * Copyright (c) 2019 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_NEFULLYCONNECTEDHYBRIDLAYER_H__
42 #define __ARM_COMPUTE_NEFULLYCONNECTEDHYBRIDLAYER_H__
44 #include "arm_compute/core/NEON/kernels/NEQuantizationSymmetricKernel.h"
45 #include "arm_compute/core/NEON/kernels/NEGEMMMatrixAccumulateBiasesKernel.h"
46 #include "arm_compute/core/NEON/kernels/NEMuliplyScaleFactorKernel.h"
47 #include "arm_compute/core/NEON/kernels/NETransposeKernel.h"
48 #include "arm_compute/runtime/MemoryGroup.h"
49 #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h"
50 #include "arm_compute/runtime/NEON/INESimpleFunctionNoBorder.h"
51 #include "arm_compute/runtime/Tensor.h"
55 /** Basic function to reshape the weights of Fully Connected layer with NEON. This function calls
56 * the following kernels:
58 * -# @ref NETransposeKernel
60 * @note The fully connected layer accepts "weights" tensors only with 2 dimensions.
62 class NEFullyConnectedHybridLayerReshapeWeights : public INESimpleFunctionNoBorder
65 /** Set the input and output tensors.
67 * @param[in] input Weights tensor. The weights must be 2 dimensional. Data types supported:
69 * @param[out] output Destination tensor. Data type supported: Same as @p input.
71 void configure(const ITensor *input, ITensor *output);
72 /** Static function to check if given info will lead to a valid configuration of @ref
73 * NEFullyConnectedHybridLayerReshapeWeights
75 * @param[in] input Weights tensor info. The weights must be 2 dimensional. Data types supported:
77 * @param[in] output Destination tensor info. Data type supported: Same as @p input.
81 static Status validate(const ITensorInfo *input, const ITensorInfo *output);
84 /** Basic function to compute a Fully Connected layer on NEON. This function calls the following
86 * -# @ref NEIm2ColKernel (called when the input comes from a convolutional layer)
87 * -# @ref NEFullyConnectedHybridLayerReshapeWeights (if @p are_weights_reshaped is set to false
88 * and transpose_weights is set to true ) (called once)
89 * -# @ref NEGEMMMatrixMultiplyKernel or @ref NEGEMMLowpMatrixMultiplyCore (if quantized
91 * -# @ref NEGEMMMatrixAccumulateBiasesKernel or @ref
92 * NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if quantized asymmetric) (if @p biases is
93 * not equal to nullptr)
95 * @note The fully connected layer accepts "weights" tensors only with 2 dimensions.
97 class NEFullyConnectedHybridLayer : public IFunction
101 NEFullyConnectedHybridLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
102 /** Prevent instances of this class from being copied (As this class contains pointers) */
103 NEFullyConnectedHybridLayer(const NEFullyConnectedHybridLayer &) = delete;
104 /** Default move constructor */
105 NEFullyConnectedHybridLayer(NEFullyConnectedHybridLayer &&) = default;
106 /** Prevent instances of this class from being copied (As this class contains pointers) */
107 NEFullyConnectedHybridLayer &operator=(const NEFullyConnectedHybridLayer &) = delete;
108 /** Default move assignment operator */
109 NEFullyConnectedHybridLayer &operator=(NEFullyConnectedHybridLayer &&) = default;
110 /** Set the input and output tensors.
112 * @param[in] input Source tensor. Data type supported: F16/F32.
113 * @param[in] weights Weights tensor. The weights must be 2 dimensional.
114 * If this function is called after a Convolution Layer, the (transposed)
115 * weights will have as many rows as the product of the first 3 input's dimensions.
116 * If it is called after another FullyConnected Layer, the (transposed)
117 * weights will have as many rows as the input's first dimension.
118 * Data type supported: S8.
119 * @param[in] biases Bias tensor. Can be nullptr. Data type supported:Same as @p input.
120 * @param[out] output Destination tensor. Its shape should be equal to the output of a matrix
121 * multiplication between:
122 * - The output of im2col on the input and the (transposed) 2D weights, if the
123 * function is called after a Convolution Layer
124 * - The input tensor and the (transposed) 2D weights, if the function is
125 * called after another FullyConnected Layer.
126 * Data type supported: Same as @p input.
127 * @param[in] fc_info (Optional) Fully connected layer additional info
129 void configure(const ITensor *input, const ITensor *weights, const ITensor *biases,
130 ITensor *output, FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
131 /** Static function to check if given info will lead to a valid configuration of @ref
132 * NEFullyConnectedHybridLayer
134 * @param[in] input Source tensor info. Data type supported: F16/F32.
135 * @param[in] weights Weights tensor info. The weights must be 2 dimensional.
136 * If this function is called after a Convolution Layer, the (transposed)
137 * weights will have as many rows as the product of the first 3 input's dimensions.
138 * If it is called after another FullyConnected Layer, the (transposed)
139 * weights will have as many rows as the input's first dimension.
140 * Data type supported: S8.
141 * @param[in] biases Bias tensor info. Can be nullptr. Data type supported:Same as @p input.
142 * @param[out] output Destination tensor info. Its shape should be equal to the output of a
143 * matrix multiplication between:
144 * - The output of im2col on the input and the (transposed) 2D weights, if the
145 * function is called after a Convolution Layer
146 * - The input tensor and the (transposed) 2D weights, if the function is
147 * called after another FullyConnected Layer.
148 * Data type supported: Same as @p input.
149 * @param[in] fc_info (Optional) Fully connected layer additional info
153 static Status validate(const ITensorInfo *input, const ITensorInfo *weights,
154 const ITensorInfo *biases, const ITensorInfo *output,
155 FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo());
157 // Inherited methods override
159 void prepare() override;
162 void configure_mm(const ITensor *input, const ITensor *weights, ITensor *output);
164 MemoryGroup _memory_group;
165 NEFullyConnectedHybridLayerReshapeWeights _reshape_weights_function;
166 NEQuantizationSymmetricKernel _quant_input_kernel;
167 NEGEMMLowpMatrixMultiplyCore _mm_gemmlowp;
168 NEMultiplyScaleFactorKernel _multiply_scale_kernel;
169 NEGEMMMatrixAccumulateBiasesKernel _accumulate_biases_kernel;
170 Tensor _reshape_weights_output;
171 Tensor _quantized_input;
172 Tensor _scale_factor;
173 Tensor _gemmlowp_output;
174 const ITensor *_original_weights;
175 bool _are_weights_reshaped;
176 bool _accumulate_biases;
179 } // namespace arm_compute
180 #endif /* __ARM_COMPUTE_NEFULLYCONNECTEDHYBRIDLAYER_H__ */