2 * Copyright (c) 2017-2018 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
24 #ifndef __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__
25 #define __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__
27 #include "arm_compute/runtime/IFunction.h"
29 #include "arm_compute/core/CL/kernels/CLCol2ImKernel.h"
30 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
31 #include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
32 #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h"
33 #include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h"
34 #include "arm_compute/core/CL/kernels/CLIm2ColKernel.h"
35 #include "arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h"
36 #include "arm_compute/core/Types.h"
37 #include "arm_compute/runtime/CL/CLMemoryGroup.h"
38 #include "arm_compute/runtime/CL/CLTensor.h"
39 #include "arm_compute/runtime/CL/functions/CLActivationLayer.h"
40 #include "arm_compute/runtime/CL/functions/CLGEMM.h"
41 #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
42 #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
43 #include "arm_compute/runtime/IMemoryManager.h"
51 /** Function to reshape and transpose the weights. This function calls the following kernels:
52 * -# @ref CLWeightsReshapeKernel
54 class CLConvolutionLayerReshapeWeights : public IFunction
58 CLConvolutionLayerReshapeWeights();
59 /** Set the input and output tensors.
61 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
62 * Data type supported: QS8/QASYMM8/QS16/F16/F32.
63 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
64 * @param[out] output Destination tensor. Data types supported: Same as @p weights.
66 void configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output);
67 /** Static function to check if given info will lead to a valid configuration of @ref CLConvolutionLayerReshapeWeights
69 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
70 * Data type supported: QS8/QASYMM8/QS16/F16/F32.
71 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
72 * @param[in] output Destination tensor. Data types supported: Same as @p weights.
76 static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output);
77 // Inherited methods overridden:
81 CLWeightsReshapeKernel _weights_reshape_kernel;
84 /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions:
86 * Note: weights already reshaped for quantized asymmetric is not supported
88 * -# @ref CLIm2ColKernel
89 * -# @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric)
90 * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if quantized asymmetric)
91 * -# @ref CLCol2ImKernel
93 * if the weights are already reshaped:
94 * -# @ref CLGEMMInterleave4x4Kernel
95 * -# @ref CLGEMMMatrixMultiplyKernel
99 class CLGEMMConvolutionLayer : public IFunction
102 /** Default constructor
104 * @param[in] memory_manager (Optional) Memory manager.
106 CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
107 /** Prevent instances of this class from being copied (As this class contains pointers) */
108 CLGEMMConvolutionLayer(const CLGEMMConvolutionLayer &) = delete;
109 /** Default move constructor */
110 CLGEMMConvolutionLayer(CLGEMMConvolutionLayer &&) = default;
111 /** Prevent instances of this class from being copied (As this class contains pointers) */
112 CLGEMMConvolutionLayer &operator=(const CLGEMMConvolutionLayer &) = delete;
113 /** Default move assignment operator */
114 CLGEMMConvolutionLayer &operator=(CLGEMMConvolutionLayer &&) = default;
115 /** Set the input and output tensors.
117 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
118 * while every optional dimension from 4 and above represent a batch of inputs.
119 * Data types supported: QS8/QASYMM8/QS16/F16/F32.
120 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
121 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
122 * Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
123 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
124 * Data types supported: Same as @p input.
125 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
126 * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. If this is not part of the fully connected layer the weights
127 * tensor has also been transposed with CLGEMMTranspose1xWKernel. Data type supported: Same as @p input.
128 * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
129 * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
131 void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(),
132 const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo());
133 /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMConvolutionLayer.
135 * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
136 * while every optional dimension from 4 and above represent a batch of inputs.
137 * Data types supported: QS8/QASYMM8/QS16/F16/F32.
138 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
139 * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
140 * Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
141 * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
142 * Data types supported: Same as @p input.
143 * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
144 * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. If this is not part of the fully connected layer the weights
145 * tensor has also been transposed with CLGEMMTranspose1xWKernel. Data type supported: Same as @p input.
146 * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
147 * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
151 static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
152 const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo());
154 // Inherited methods overridden:
156 void prepare() override;
159 /** Configures the appropriate matrix multiply routine
161 * @param input Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32.
162 * @param weights Weights tensor. Data type supported: Same as @p input.
163 * @param output Output tensor. Data types supported: Same as @p input,
164 * except for input of QASYMM8 type where output should be of S32 type.
166 void configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output);
167 /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMConvolutionLayer matrix multiply routines
169 * @param[in] input Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32.
170 * @param[in] weights Weights tensor. Data type supported: Same as @p input.
171 * @param[in] output Output tensor. Data types supported: Same as @p input,
172 * except for input of QASYMM8 type where output should be of S32 type.
176 static Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output);
179 CLMemoryGroup _memory_group;
180 CLConvolutionLayerReshapeWeights _reshape_weights;
181 CLIm2ColKernel _im2col_kernel;
183 CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp;
184 CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage;
185 CLCol2ImKernel _col2im_kernel;
186 CLActivationLayer _activationlayer_function;
188 const ICLTensor *_original_weights;
190 CLTensor _im2col_output;
191 CLTensor _weights_reshaped;
192 CLTensor _gemm_output;
193 CLTensor _tmp_output;
196 bool _is_activationlayer_enabled;
200 #endif /* __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ */