Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 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_NECONVOLUTIONLAYER_H__ |
| 25 | #define __ARM_COMPUTE_NECONVOLUTIONLAYER_H__ |
| 26 | |
| 27 | #include "arm_compute/runtime/IFunction.h" |
| 28 | |
| 29 | #include "arm_compute/core/NEON/kernels/NECol2ImKernel.h" |
| 30 | #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 31 | #include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 32 | #include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" |
| 33 | #include "arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h" |
| 34 | #include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" |
| 35 | #include "arm_compute/core/NEON/kernels/NEIm2ColKernel.h" |
| 36 | #include "arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h" |
| 37 | #include "arm_compute/core/Types.h" |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 38 | #include "arm_compute/runtime/MemoryGroup.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 39 | #include "arm_compute/runtime/Tensor.h" |
| 40 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 41 | #include <memory> |
| 42 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 43 | namespace arm_compute |
| 44 | { |
| 45 | class ITensor; |
| 46 | |
| 47 | /** Function to reshape and perform 1xW transposition on the weights. This function calls the following kernels: |
| 48 | * -# @ref NEWeightsReshapeKernel |
| 49 | * -# @ref NEGEMMTranspose1xWKernel (executed in case GEMM is required for the operation) |
| 50 | */ |
| 51 | class NEConvolutionLayerReshapeWeights : public IFunction |
| 52 | { |
| 53 | public: |
| 54 | /** Constructor */ |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 55 | NEConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager = nullptr); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 56 | /** Set the input and output tensors. |
| 57 | * |
Gian Marco Iodice | 2bbd964 | 2017-07-04 16:46:32 +0100 | [diff] [blame] | 58 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QS8/QS16/F32. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 59 | * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. |
| 60 | * @param[out] output Destination tensor. Data types supported: Same as @p weights. |
| 61 | * @param[in] transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise. |
| 62 | * Data types supported: Same as @p weights. |
| 63 | */ |
| 64 | void configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW); |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 65 | /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayerReshapeWeights |
| 66 | * |
| 67 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QS8/QS16/F16/F32. |
| 68 | * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. |
| 69 | * @param[in] output Destination tensor. Data types supported: Same as @p weights. |
| 70 | * @param[in] transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise. |
| 71 | * Data types supported: Same as @p weights. |
| 72 | * |
| 73 | * @return an error status |
| 74 | */ |
| 75 | static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW); |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 76 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 77 | // Inherited methods overridden: |
| 78 | void run() override; |
| 79 | |
| 80 | private: |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 81 | MemoryGroup _memory_group; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 82 | NEWeightsReshapeKernel _weights_reshape_kernel; |
| 83 | NEGEMMTranspose1xWKernel _weights_transposed_kernel; |
| 84 | Tensor _weights_reshaped; |
| 85 | bool _transpose1xW; |
| 86 | }; |
| 87 | |
| 88 | /** Basic function to simulate a convolution layer. This function calls the following NEON kernels: |
| 89 | * -# @ref NEWeightsReshapeKernel (executed only once for each configuration) |
| 90 | * -# @ref NEIm2ColKernel |
| 91 | * -# @ref NEGEMMInterleave4x4Kernel (executed only in case GEMM is required for the operation) |
| 92 | * -# @ref NEGEMMMatrixMultiplyKernel |
| 93 | * -# @ref NECol2ImKernel |
| 94 | */ |
| 95 | class NEConvolutionLayer : public IFunction |
| 96 | { |
| 97 | public: |
| 98 | /** Constructor */ |
Georgios Pinitas | baf174e | 2017-09-08 19:47:30 +0100 | [diff] [blame] | 99 | NEConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 100 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 101 | /** Set the input and output tensors. |
| 102 | * |
| 103 | * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], |
| 104 | * while every optional dimension from 4 and above represent a batch of inputs. |
Gian Marco Iodice | 2bbd964 | 2017-07-04 16:46:32 +0100 | [diff] [blame] | 105 | * Data types supported: QS8/QS16/F32. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 106 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. |
| 107 | * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input. |
| 108 | * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. |
| 109 | * Data types supported: Same as @p input. |
| 110 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 111 | * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights |
| 112 | * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input. |
| 113 | */ |
| 114 | void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()); |
Giorgio Arena | 7c23ad0 | 2017-11-30 15:08:38 +0000 | [diff] [blame] | 115 | /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayer |
| 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. |
| 119 | * Data types supported: QS8/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]. Data type supported:Same as @p input. |
| 122 | * @param[in] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. |
| 123 | * Data types supported: Same as @p input. |
| 124 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 125 | * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights |
| 126 | * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input. |
| 127 | * |
| 128 | * @return a status |
| 129 | */ |
| 130 | static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
| 131 | const WeightsInfo &weights_info = WeightsInfo()); |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 132 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 133 | // Inherited methods overridden: |
| 134 | void run() override; |
| 135 | |
| 136 | private: |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 137 | MemoryGroup _memory_group; |
| 138 | NEIm2ColKernel _input_im2col_kernel; |
| 139 | NEGEMMInterleave4x4Kernel _input_interleave_kernel; |
| 140 | NEConvolutionLayerReshapeWeights _reshape_weights; |
| 141 | NEGEMMMatrixMultiplyKernel _mm_kernel; |
| 142 | std::unique_ptr<NEGEMMAssemblyBaseKernel> _mm_optimised_kernel; |
| 143 | NECol2ImKernel _output_col2im_kernel; |
| 144 | Tensor _input_im2col_reshaped; |
| 145 | Tensor _input_interleaved_reshaped; |
| 146 | Tensor _weights_reshaped; |
| 147 | Tensor _gemm_output; |
| 148 | Tensor _workspace; |
| 149 | bool _has_bias; |
| 150 | bool _is_fully_connected_convolution; |
| 151 | bool _are_weights_reshaped; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 152 | }; |
| 153 | } |
| 154 | #endif /* __ARM_COMPUTE_NECONVOLUTIONLAYER_H__ */ |