Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 1 | /* |
Pablo Tello | 9ceebbe | 2018-01-10 16:44:13 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 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 | */ |
Georgios Pinitas | 9fb1159 | 2018-04-26 20:34:58 +0100 | [diff] [blame] | 24 | #ifndef __ARM_COMPUTE_NEWINOGRADCONVOLUTIONLAYER_H__ |
| 25 | #define __ARM_COMPUTE_NEWINOGRADCONVOLUTIONLAYER_H__ |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 26 | |
| 27 | #include "arm_compute/runtime/IFunction.h" |
| 28 | |
Pablo Tello | f6c572c | 2018-02-14 12:47:30 +0000 | [diff] [blame] | 29 | #include "arm_compute/core/NEON/INEKernel.h" |
Michalis Spyrou | 2b3129e | 2018-04-25 18:10:13 +0100 | [diff] [blame] | 30 | #include "arm_compute/core/NEON/kernels/assembly/arm_gemm.hpp" |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 31 | #include "arm_compute/core/Types.h" |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame] | 32 | #include "arm_compute/runtime/CPP/functions/CPPPermute.h" |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 33 | #include "arm_compute/runtime/MemoryGroup.h" |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 34 | #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h" |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 35 | #include "arm_compute/runtime/Tensor.h" |
| 36 | |
| 37 | #include <memory> |
| 38 | |
| 39 | namespace arm_compute |
| 40 | { |
| 41 | class ITensor; |
| 42 | /** Basic function to simulate a convolution layer. This function calls the following NEON kernels: |
Pablo Tello | f6c572c | 2018-02-14 12:47:30 +0000 | [diff] [blame] | 43 | * -# @ref NEWinogradLayerTransformWeightsKernel (executed only once in the first call to the run() method ) |
| 44 | * -# @ref NEWinogradLayerTransformInputKernel |
| 45 | * -# @ref NEWinogradLayerTransformOutputKernel |
| 46 | * -# @ref NEWinogradLayerBatchedGEMMKernel |
| 47 | * -# @ref CPPPermute (three times: weights, input and output) |
Giorgio Arena | a3221e6 | 2018-05-03 15:57:48 +0100 | [diff] [blame] | 48 | * |
| 49 | * @note Some Winograd configurations (i.e. F(2x2, 5x5), F(4x4, 5x5)) are supported only with enable_fast_math = true |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 50 | */ |
Georgios Pinitas | 9fb1159 | 2018-04-26 20:34:58 +0100 | [diff] [blame] | 51 | class NEWinogradConvolutionLayer : public IFunction |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 52 | { |
| 53 | public: |
| 54 | /** Constructor */ |
Georgios Pinitas | 9fb1159 | 2018-04-26 20:34:58 +0100 | [diff] [blame] | 55 | NEWinogradConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 56 | |
| 57 | /** Set the input and output tensors. |
| 58 | * |
Giorgio Arena | a3221e6 | 2018-05-03 15:57:48 +0100 | [diff] [blame] | 59 | * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], |
| 60 | * while every optional dimension from 4 and above represent a batch of inputs. |
| 61 | * Data types supported: F32. |
| 62 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. |
| 63 | * Currently only 3x3 and 5x5 kernels are supported. |
| 64 | * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. |
| 65 | * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. |
| 66 | * Data types supported: Same as @p input. |
| 67 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported. |
| 68 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. |
| 69 | * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation |
| 70 | * available which may introduce a drop of accuracy as well. Default is false |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 71 | */ |
Giorgio Arena | a3221e6 | 2018-05-03 15:57:48 +0100 | [diff] [blame] | 72 | void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info = ActivationLayerInfo(), |
| 73 | bool enable_fast_math = false); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 74 | |
| 75 | // Inherited methods overridden: |
| 76 | void run() override; |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame^] | 77 | void prepare() override; |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 78 | |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 79 | /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer |
| 80 | * |
Giorgio Arena | a3221e6 | 2018-05-03 15:57:48 +0100 | [diff] [blame] | 81 | * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], |
| 82 | * while every optional dimension from 4 and above represent a batch of inputs. |
| 83 | * Data types supported: F32. |
| 84 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. |
| 85 | * Currently only 3x3 and 5x5 kernels are supported. |
| 86 | * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. |
| 87 | * @param[in] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. |
| 88 | * Data types supported: Same as @p input. |
| 89 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported. |
| 90 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. |
| 91 | * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation |
| 92 | * available which may introduce a drop of accuracy as well. Default is false |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 93 | * |
| 94 | * @return a status |
| 95 | */ |
Vidhya Sudhan Loganathan | 3ca9786 | 2018-04-23 08:20:04 +0100 | [diff] [blame] | 96 | static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
Giorgio Arena | a3221e6 | 2018-05-03 15:57:48 +0100 | [diff] [blame] | 97 | const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 98 | |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 99 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
Georgios Pinitas | 9fb1159 | 2018-04-26 20:34:58 +0100 | [diff] [blame] | 100 | NEWinogradConvolutionLayer(const NEWinogradConvolutionLayer &) = delete; |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 101 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
Georgios Pinitas | 9fb1159 | 2018-04-26 20:34:58 +0100 | [diff] [blame] | 102 | NEWinogradConvolutionLayer &operator=(const NEWinogradConvolutionLayer &) = delete; |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 103 | |
| 104 | private: |
Michalis Spyrou | 2b3129e | 2018-04-25 18:10:13 +0100 | [diff] [blame] | 105 | MemoryGroup _memory_group; |
| 106 | std::unique_ptr<arm_gemm::GemmCommon<float, float>> _arm_gemm; |
| 107 | std::unique_ptr<INEKernel> _gemm_kernel; |
Pablo Tello | f6c572c | 2018-02-14 12:47:30 +0000 | [diff] [blame] | 108 | std::unique_ptr<INEKernel> _transform_input_kernel; |
| 109 | std::unique_ptr<INEKernel> _transform_output_kernel; |
| 110 | std::unique_ptr<INEKernel> _transform_weights_kernel; |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 111 | NEActivationLayer _activationlayer_function; |
Pablo Tello | f6c572c | 2018-02-14 12:47:30 +0000 | [diff] [blame] | 112 | |
Pablo Tello | 52140b4 | 2018-01-30 14:48:11 +0000 | [diff] [blame] | 113 | CPPPermute _permute_input; |
| 114 | CPPPermute _permute_weights; |
| 115 | CPPPermute _permute_output; |
| 116 | Tensor _input_workspace; |
| 117 | Tensor _output_workspace; |
| 118 | Tensor _kernel_storage; |
| 119 | Tensor _input_nhwc; |
| 120 | Tensor _output_nhwc; |
| 121 | Tensor _weights_hwio; |
Michalis Spyrou | 2b3129e | 2018-04-25 18:10:13 +0100 | [diff] [blame] | 122 | Tensor _workspace; |
Pablo Tello | 52140b4 | 2018-01-30 14:48:11 +0000 | [diff] [blame] | 123 | const ITensor *_input; |
| 124 | const ITensor *_weights; |
| 125 | ITensor *_output; |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame^] | 126 | bool _is_prepared; |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 127 | bool _is_activationlayer_enabled; |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 128 | }; |
| 129 | } |
Georgios Pinitas | 9fb1159 | 2018-04-26 20:34:58 +0100 | [diff] [blame] | 130 | #endif /* __ARM_COMPUTE_NEWINOGRADCONVOLUTIONLAYER_H__ */ |