Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017-2018 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_CLGEMMCONVOLUTIONLAYER_H__ |
| 25 | #define __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ |
| 26 | |
| 27 | #include "arm_compute/runtime/IFunction.h" |
| 28 | |
| 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" |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 39 | #include "arm_compute/runtime/CL/functions/CLActivationLayer.h" |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 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" |
| 44 | |
| 45 | #include <memory> |
| 46 | |
| 47 | namespace arm_compute |
| 48 | { |
| 49 | class ICLTensor; |
| 50 | |
| 51 | /** Function to reshape and transpose the weights. This function calls the following kernels: |
| 52 | * -# @ref CLWeightsReshapeKernel |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 53 | */ |
| 54 | class CLConvolutionLayerReshapeWeights : public IFunction |
| 55 | { |
| 56 | public: |
| 57 | /** Constructor */ |
Georgios Pinitas | d8734b5 | 2017-12-22 15:27:52 +0000 | [diff] [blame] | 58 | CLConvolutionLayerReshapeWeights(); |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 59 | /** Set the input and output tensors. |
| 60 | * |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 61 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame^] | 62 | * Data type supported: QASYMM8/F16/F32. |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 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. |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 65 | */ |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 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 |
| 68 | * |
| 69 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame^] | 70 | * Data type supported: QASYMM8/F16/F32. |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 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. |
| 73 | * |
| 74 | * @return a status |
| 75 | */ |
| 76 | static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output); |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 77 | // Inherited methods overridden: |
| 78 | void run() override; |
| 79 | |
| 80 | private: |
Georgios Pinitas | d8734b5 | 2017-12-22 15:27:52 +0000 | [diff] [blame] | 81 | CLWeightsReshapeKernel _weights_reshape_kernel; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 82 | }; |
| 83 | |
| 84 | /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: |
| 85 | * |
| 86 | * Note: weights already reshaped for quantized asymmetric is not supported |
| 87 | * |
| 88 | * -# @ref CLIm2ColKernel |
| 89 | * -# @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric) |
Isabella Gottardi | a1667fb | 2018-04-11 11:53:49 +0100 | [diff] [blame] | 90 | * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint (if quantized asymmetric) |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 91 | * -# @ref CLCol2ImKernel |
| 92 | * |
| 93 | * if the weights are already reshaped: |
| 94 | * -# @ref CLGEMMInterleave4x4Kernel |
| 95 | * -# @ref CLGEMMMatrixMultiplyKernel |
| 96 | * else |
| 97 | * -# @ref CLGEMM |
| 98 | */ |
| 99 | class CLGEMMConvolutionLayer : public IFunction |
| 100 | { |
| 101 | public: |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 102 | /** Default constructor |
| 103 | * |
| 104 | * @param[in] memory_manager (Optional) Memory manager. |
| 105 | */ |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 106 | CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); |
Georgios Pinitas | 1562be3 | 2018-03-08 19:09:19 +0000 | [diff] [blame] | 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; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 115 | /** Set the input and output tensors. |
| 116 | * |
| 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. |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame^] | 119 | * Data types supported: QASYMM8/F16/F32. |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 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. |
Alex Gilday | 7da29b6 | 2018-03-23 14:16:00 +0000 | [diff] [blame] | 128 | * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 129 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 130 | */ |
Alex Gilday | 7da29b6 | 2018-03-23 14:16:00 +0000 | [diff] [blame] | 131 | void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(), |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 132 | const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo()); |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 133 | /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMConvolutionLayer. |
| 134 | * |
| 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. |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame^] | 137 | * Data types supported: QASYMM8/F16/F32. |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 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. |
Alex Gilday | 7da29b6 | 2018-03-23 14:16:00 +0000 | [diff] [blame] | 146 | * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 147 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 148 | * |
| 149 | * @return a status |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 150 | */ |
| 151 | static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 152 | const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo()); |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 153 | |
| 154 | // Inherited methods overridden: |
| 155 | void run() override; |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 156 | void prepare() override; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 157 | |
| 158 | private: |
| 159 | /** Configures the appropriate matrix multiply routine |
| 160 | * |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame^] | 161 | * @param[in] input Input tensor. Data types supported: QASYMM8/F16/F32. |
Georgios Pinitas | 19ea419 | 2018-06-19 13:09:53 +0100 | [diff] [blame] | 162 | * @param[in] weights Weights tensor. Data type supported: Same as @p input. |
| 163 | * @param[in, out] output Output tensor. Data types supported: Same as @p input, |
| 164 | * except for input of QASYMM8 type where output should be of S32 type. |
| 165 | * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1) |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 166 | */ |
Georgios Pinitas | 19ea419 | 2018-06-19 13:09:53 +0100 | [diff] [blame] | 167 | void configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, int gemm_3d_depth = 1); |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 168 | /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMConvolutionLayer matrix multiply routines |
| 169 | * |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame^] | 170 | * @param[in] input Input tensor. Data types supported: QASYMM8/F16/F32. |
Georgios Pinitas | 19ea419 | 2018-06-19 13:09:53 +0100 | [diff] [blame] | 171 | * @param[in] weights Weights tensor. Data type supported: Same as @p input. |
| 172 | * @param[in] output Output tensor. Data types supported: Same as @p input, |
| 173 | * except for input of QASYMM8 type where output should be of S32 type. |
| 174 | * @param[in] gemm_3d_depth (Optional) Depth of GEMM 3D (Defaults to 1) |
Georgios Pinitas | 78c0090 | 2018-01-09 17:33:11 +0000 | [diff] [blame] | 175 | * |
| 176 | * @return a status |
| 177 | */ |
Georgios Pinitas | 19ea419 | 2018-06-19 13:09:53 +0100 | [diff] [blame] | 178 | static Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, int gemm_3d_depth = 1); |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 179 | |
| 180 | private: |
| 181 | CLMemoryGroup _memory_group; |
| 182 | CLConvolutionLayerReshapeWeights _reshape_weights; |
| 183 | CLIm2ColKernel _im2col_kernel; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 184 | CLGEMM _mm_gemm; |
| 185 | CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp; |
| 186 | CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage; |
| 187 | CLCol2ImKernel _col2im_kernel; |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 188 | CLActivationLayer _activationlayer_function; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 189 | |
Georgios Pinitas | 1562be3 | 2018-03-08 19:09:19 +0000 | [diff] [blame] | 190 | const ICLTensor *_original_weights; |
| 191 | |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 192 | CLTensor _im2col_output; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 193 | CLTensor _weights_reshaped; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 194 | CLTensor _gemm_output; |
| 195 | CLTensor _tmp_output; |
| 196 | |
Georgios Pinitas | 19ea419 | 2018-06-19 13:09:53 +0100 | [diff] [blame] | 197 | DataLayout _data_layout; |
| 198 | |
| 199 | bool _skip_im2col; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 200 | bool _is_quantized; |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 201 | bool _is_activationlayer_enabled; |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 202 | bool _is_prepared; |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 203 | }; |
Georgios Pinitas | 19ea419 | 2018-06-19 13:09:53 +0100 | [diff] [blame] | 204 | } // namespace arm_compute |
Isabella Gottardi | f07d28d | 2018-02-06 14:52:43 +0000 | [diff] [blame] | 205 | #endif /* __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ */ |