Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Manuel Bottini | 87350f4 | 2020-09-15 13:03:34 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [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 | */ |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 24 | #ifndef ARM_COMPUTE_NEDIRECTCONVOLUTIONLAYERKERNEL_H |
| 25 | #define ARM_COMPUTE_NEDIRECTCONVOLUTIONLAYERKERNEL_H |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 26 | |
Michalis Spyrou | ebcebf1 | 2020-10-21 00:04:14 +0100 | [diff] [blame] | 27 | #include "src/core/NEON/INEKernel.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 28 | |
| 29 | namespace arm_compute |
| 30 | { |
| 31 | class ITensor; |
| 32 | |
| 33 | /** NEON interface for Direct Convolution Layer kernel */ |
| 34 | class NEDirectConvolutionLayerKernel : public INEKernel |
| 35 | { |
| 36 | public: |
Anthony Barbier | e8a4983 | 2018-01-18 10:04:05 +0000 | [diff] [blame] | 37 | const char *name() const override |
| 38 | { |
| 39 | return "NEDirectConvolutionLayerKernel"; |
| 40 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | /** Default constructor */ |
| 42 | NEDirectConvolutionLayerKernel(); |
| 43 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 44 | NEDirectConvolutionLayerKernel(const NEDirectConvolutionLayerKernel &) = delete; |
| 45 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 46 | NEDirectConvolutionLayerKernel &operator=(const NEDirectConvolutionLayerKernel &) = delete; |
| 47 | /** Allow instances of this class to be moved */ |
| 48 | NEDirectConvolutionLayerKernel(NEDirectConvolutionLayerKernel &&) = default; |
| 49 | /** Allow instances of this class to be moved */ |
| 50 | NEDirectConvolutionLayerKernel &operator=(NEDirectConvolutionLayerKernel &&) = default; |
| 51 | /** Default destructor */ |
| 52 | ~NEDirectConvolutionLayerKernel() = default; |
steniu01 | 27b386c | 2017-07-18 17:37:43 +0100 | [diff] [blame] | 53 | /** Set the input, weights, and output tensors. |
| 54 | * |
Gian Marco Iodice | 5cb4d6a | 2017-08-08 10:53:00 +0100 | [diff] [blame] | 55 | * @note: DirectConvolution only works in the following configurations: |
| 56 | * 1x1 convolution with stride_x = 1/2/3, stride_y = 1/2/3 |
| 57 | * 3x3 convolution with stride_x = 1/2/3, stride_y = 1/2/3 |
| 58 | * |
steniu01 | 27b386c | 2017-07-18 17:37:43 +0100 | [diff] [blame] | 59 | * @param[in] input The input tensor to convolve. 3 lower dimensions represent a single input [width, height, IFM], |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 60 | * while every optional dimension from 4 and above represent a batch of inputs. Data types supported: F16/F32. |
steniu01 | 27b386c | 2017-07-18 17:37:43 +0100 | [diff] [blame] | 61 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. |
| 62 | * The 3rd dimension must be the same as the input's volume 3rd dimension. |
| 63 | * Data type supported:Same as @p input. |
| 64 | * @param[out] output Output tensor. |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 65 | * The 3rd dimensions must be equal to the 4th dimension of the @p kernels tensor. Data types supported: F16/F32 |
steniu01 | 27b386c | 2017-07-18 17:37:43 +0100 | [diff] [blame] | 66 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 67 | */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 68 | void configure(const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info); |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 69 | /** Static function to check if given info will lead to a valid configuration of @ref NEDirectConvolutionLayerKernel |
| 70 | * |
| 71 | * @param[in] input The input tensor to convolve. 3 lower dimensions represent a single input [width, height, IFM], |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 72 | * while every optional dimension from 4 and above represent a batch of inputs. Data types supported: F16/F32. |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 73 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. |
| 74 | * The 3rd dimension must be the same as the input's volume 3rd dimension. |
| 75 | * Data type supported:Same as @p input. |
| 76 | * @param[in] output Output tensor. |
Vidhya Sudhan Loganathan | f4cb81b | 2018-07-04 15:13:14 +0100 | [diff] [blame] | 77 | * The 3rd dimensions must be equal to the 4th dimension of the @p kernels tensor. Data types supported: F16/F32 |
Michalis Spyrou | afa5d81 | 2017-11-30 14:25:57 +0000 | [diff] [blame] | 78 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 79 | * |
| 80 | * @return a status |
| 81 | */ |
| 82 | static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 83 | |
| 84 | // Inherited methods overridden: |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 85 | void run(const Window &window, const ThreadInfo &info) override; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 86 | BorderSize border_size() const override; |
| 87 | |
| 88 | private: |
Manuel Bottini | 87350f4 | 2020-09-15 13:03:34 +0100 | [diff] [blame] | 89 | /* Template function for optimized convolution NHWC */ |
| 90 | template <typename T> |
| 91 | void convolve_nhwc_optimized(const Window &window); |
| 92 | |
| 93 | /* Template function for convolution NHWC */ |
| 94 | template <typename T> |
| 95 | void convolve_nhwc(const Window &window); |
| 96 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 97 | const ITensor *_input; |
| 98 | const ITensor *_weights; |
| 99 | ITensor *_output; |
| 100 | PadStrideInfo _conv_info; |
| 101 | BorderSize _border_size; |
| 102 | unsigned int _kernel_size; |
Georgios Pinitas | 898a806 | 2017-09-12 19:19:12 +0100 | [diff] [blame] | 103 | unsigned int _num_weight_elems_read_per_row; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 104 | unsigned int _num_elems_read_per_iteration; |
| 105 | unsigned int _num_elems_written_per_iteration; |
| 106 | }; |
Gian Marco Iodice | 356f643 | 2017-09-22 11:32:21 +0100 | [diff] [blame] | 107 | } // namespace arm_compute |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 108 | #endif /*ARM_COMPUTE_NEDIRECTCONVOLUTIONLAYERKERNEL_H */ |