Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2019 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 | |
| 25 | #pragma once |
| 26 | |
| 27 | #include <deque> |
| 28 | #include <functional> |
| 29 | #include <memory> |
| 30 | |
| 31 | #include "depthwise.hpp" |
| 32 | |
| 33 | namespace depthwise |
| 34 | { |
| 35 | |
| 36 | template < |
| 37 | unsigned int OutputTileRows, unsigned int OutputTileCols, |
| 38 | unsigned int KernelRows, unsigned int KernelCols, |
| 39 | unsigned int StrideRows, unsigned int StrideCols, |
| 40 | typename TIn, typename TBias, typename TOut |
| 41 | > |
| 42 | class DilatedDepthwiseConvolution : public IDepthwiseConvolution |
| 43 | { |
| 44 | public: |
| 45 | /** Create a new dilated depthwise convolution engine. |
| 46 | */ |
| 47 | DilatedDepthwiseConvolution( |
| 48 | int n_batches, int n_input_rows, int n_input_cols, int n_channels, |
| 49 | int dilation_factor, |
| 50 | nck::ActivationFunction activation, |
| 51 | unsigned int padding_top, |
| 52 | unsigned int padding_left, |
| 53 | unsigned int padding_bottom, |
| 54 | unsigned int padding_right |
| 55 | ); |
| 56 | |
| 57 | /** Create a new dilated depthwise convolution engine. |
| 58 | */ |
| 59 | DilatedDepthwiseConvolution( |
| 60 | int n_batches, int n_input_rows, int n_input_cols, int n_channels, |
| 61 | int dilation_factor, int n_output_rows, int n_output_cols, |
| 62 | nck::ActivationFunction activation, |
| 63 | unsigned int padding_top, |
| 64 | unsigned int padding_left, |
| 65 | unsigned int padding_bottom, |
| 66 | unsigned int padding_right |
| 67 | ); |
| 68 | |
| 69 | // Cannot copy or move a DilatedDepthwiseConvolution. |
| 70 | DilatedDepthwiseConvolution(DilatedDepthwiseConvolution&) = delete; |
| 71 | DilatedDepthwiseConvolution operator=(DilatedDepthwiseConvolution&) = delete; |
| 72 | |
| 73 | /* Set input tensor and stride. */ |
| 74 | void set_input(const void *inptr) override; |
| 75 | void set_input(const void *inptr, int column_stride) override; |
| 76 | void set_input(const void *inptr, int row_stride, int column_stride) override; |
| 77 | void set_input(const void *inptr, int batch_stride, int row_stride, int column_stride) override; |
| 78 | |
| 79 | /* Set output tensor and stride. */ |
| 80 | void set_output(void *outptr) override; |
| 81 | void set_output(void *outptr, int column_stride) override; |
| 82 | void set_output(void *outptr, int row_stride, int column_stride) override; |
| 83 | void set_output(void *outptr, int batch_stride, int row_stride, int column_stride) override; |
| 84 | |
| 85 | static int get_output_size( |
| 86 | int dim_size, |
| 87 | unsigned int padding_before, |
| 88 | unsigned int padding_after, |
| 89 | int dilation_factor |
| 90 | ); |
| 91 | |
| 92 | int output_size( |
| 93 | int dim_size, unsigned int padding_before, unsigned int padding_after |
| 94 | ) const override; |
| 95 | |
| 96 | /* Weights and biases are re-ordered to improve memory access patterns. Use |
| 97 | * these methods to determine the size of the re-pack buffer and to set the |
| 98 | * address (and implicitly reorder the weights and biases into) the buffer. |
| 99 | */ |
| 100 | size_t get_packed_params_size(void) const override; |
| 101 | void set_packed_params_buffer(void *) override; |
| 102 | |
| 103 | void pack_params(const void *weights, const void *biases=nullptr) const override; |
| 104 | void pack_params(void *buffer, const void *weights, const void *biases=nullptr) const override; |
| 105 | void pack_params( |
| 106 | void *buffer, |
| 107 | const void* weights, |
| 108 | unsigned int weight_row_stride, |
| 109 | unsigned int weight_col_stride, |
| 110 | const void *biases=nullptr |
| 111 | ) const override; |
| 112 | |
| 113 | /* Working space is used to pad tensors on the fly. Before running any |
| 114 | * inference check the amount of space required, allocate and provide a |
| 115 | * pointer to the convolution engine. |
| 116 | */ |
| 117 | size_t get_working_space_size(unsigned int nthreads=1) const override; |
| 118 | void set_working_space(void *) override; |
| 119 | |
| 120 | unsigned int get_window(void) const override; |
| 121 | void run(unsigned int start, unsigned int stop, unsigned int threadid=0) override; |
| 122 | |
| 123 | protected: |
| 124 | /** Protected constructor which also accepts a function to construct a new |
| 125 | * subconvolution |
| 126 | */ |
| 127 | DilatedDepthwiseConvolution( |
| 128 | int n_batches, int n_input_rows, int n_input_cols, int n_channels, |
| 129 | int dilation_factor, int n_output_rows, int n_output_cols, |
| 130 | nck::ActivationFunction activation, |
| 131 | unsigned int padding_top, |
| 132 | unsigned int padding_left, |
| 133 | unsigned int padding_bottom, |
| 134 | unsigned int padding_right, |
| 135 | std::function<IDepthwiseConvolution *(int, int, int, int, int, int, nck::ActivationFunction, unsigned int, unsigned int, unsigned int, unsigned int)> subconvfn |
| 136 | ); |
| 137 | |
| 138 | const int _dilation_factor; |
| 139 | const int _n_input_rows, _n_input_cols, _n_channels; |
| 140 | const int _padding_top, _padding_left; |
| 141 | const int _n_output_rows, _n_output_cols; |
| 142 | |
| 143 | /* Dilated depthwise convolution is performed through repeated calls to |
| 144 | * non-dilated convolutions. If the dilation factor is $n$, then we perform |
| 145 | * $(n + 1)^2$ depthwise convolutions. |
| 146 | */ |
| 147 | using BaseDepthwise = DepthwiseConvolution< |
| 148 | OutputTileRows, OutputTileCols, |
| 149 | KernelRows, KernelCols, |
| 150 | StrideRows, StrideCols, |
| 151 | TIn, TBias, TOut |
| 152 | >; |
| 153 | std::deque<std::deque<std::unique_ptr<IDepthwiseConvolution>>> _convs; |
| 154 | }; |
| 155 | |
| 156 | } // namespace depthwise |