Giorgio Arena | 1f9ca1d | 2018-03-01 11:13:45 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 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 | #include "Winograd.h" |
| 25 | |
| 26 | #include "tests/validation/Helpers.h" |
| 27 | #include "tests/validation/reference/Utils.h" |
| 28 | |
Gian Marco Iodice | 7e4b239 | 2018-02-22 16:17:20 +0000 | [diff] [blame^] | 29 | #include "arm_compute/core/Types.h" |
| 30 | |
Giorgio Arena | 1f9ca1d | 2018-03-01 11:13:45 +0000 | [diff] [blame] | 31 | namespace arm_compute |
| 32 | { |
| 33 | namespace test |
| 34 | { |
| 35 | namespace validation |
| 36 | { |
| 37 | namespace reference |
| 38 | { |
| 39 | namespace |
| 40 | { |
| 41 | template <typename T> |
| 42 | void winograd_input_transform3x3(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const PadStrideInfo &conv_info) |
| 43 | { |
| 44 | TensorShape shape4x4(4u, 4u); |
| 45 | |
| 46 | // Simple tensor for the 4x4 input tile |
| 47 | SimpleTensor<T> src_tile{ shape4x4, src.data_type() }; |
| 48 | |
| 49 | // Simple tensor for the 4x4 temporary tile |
| 50 | SimpleTensor<T> tmp_tile{ shape4x4, src.data_type() }; |
| 51 | |
| 52 | // Simple tensor for the 4x4 output tile |
| 53 | SimpleTensor<T> dst_tile{ shape4x4, src.data_type() }; |
| 54 | |
| 55 | // Simple tensor for the transformation matrix |
| 56 | SimpleTensor<T> matrix{ shape4x4, src.data_type() }; |
| 57 | |
| 58 | // Simple tensor for the transformation matrix transposed |
| 59 | SimpleTensor<T> matrix_transposed{ shape4x4, src.data_type() }; |
| 60 | |
| 61 | const float matrix_values[] = { 1.f, 0.f, -1.f, 0.f, |
| 62 | 0.f, 1.f, 1.f, 0.f, |
| 63 | 0.f, -1.f, 1.f, 0.f, |
| 64 | 0.f, 1.f, 0.f, -1.f |
| 65 | }; |
| 66 | |
| 67 | for(int i = 0; i < matrix.num_elements(); ++i) |
| 68 | { |
| 69 | matrix[i] = matrix_values[i]; |
| 70 | } |
| 71 | |
| 72 | transpose_matrix(matrix, matrix_transposed); |
| 73 | |
| 74 | const int in_w = src.shape().x(); |
| 75 | const int in_h = src.shape().y(); |
| 76 | const int in_d = src.shape().z(); |
| 77 | const int num_batches = src.shape().total_size() / (in_w * in_h * in_d); |
| 78 | const int num_tiles_x = std::ceil((in_w - 2 + conv_info.pad_left() + conv_info.pad_right()) / 2.0f); |
| 79 | const int num_tiles_y = std::ceil((in_h - 2 + conv_info.pad_top() + conv_info.pad_bottom()) / 2.0f); |
| 80 | |
| 81 | ARM_COMPUTE_ERROR_ON((num_tiles_x * num_tiles_y) != static_cast<int>(dst.shape().y())); |
| 82 | |
| 83 | for(int b = 0; b < num_batches; ++b) |
| 84 | { |
| 85 | for(int z = 0; z < in_d; ++z) |
| 86 | { |
| 87 | for(int y = 0; y < num_tiles_y; ++y) |
| 88 | { |
| 89 | for(int x = 0; x < num_tiles_x; ++x) |
| 90 | { |
| 91 | int xi = x * 2 - conv_info.pad_left(); |
| 92 | int yi = y * 2 - conv_info.pad_top(); |
| 93 | |
| 94 | // Get the 4x4 tile from the input tensor |
| 95 | get_tile(src, src_tile, Coordinates(xi, yi, z, b)); |
| 96 | |
| 97 | // Compute the transformation |
| 98 | matrix_multiply(matrix, src_tile, tmp_tile); |
| 99 | matrix_multiply(tmp_tile, matrix_transposed, dst_tile); |
| 100 | |
| 101 | // Store the 4x4 output tile across the 16 channels |
| 102 | for(int i = 0; i < 16; ++i) |
| 103 | { |
| 104 | int xo = z; |
| 105 | int yo = x + y * num_tiles_x; |
| 106 | dst[coords2index(dst.shape(), Coordinates(xo, yo, i, b))] = dst_tile[i]; |
| 107 | } |
| 108 | } |
| 109 | } |
| 110 | } |
| 111 | } |
| 112 | } |
Gian Marco Iodice | 7e4b239 | 2018-02-22 16:17:20 +0000 | [diff] [blame^] | 113 | |
| 114 | template <typename T> |
| 115 | void winograd_filter_transform3x3(const SimpleTensor<T> &in, SimpleTensor<T> &out) |
| 116 | { |
| 117 | // Simple tensor for the 3x3 input tile |
| 118 | SimpleTensor<T> input_tile{ TensorShape(3u, 3u), in.data_type(), 1 }; |
| 119 | |
| 120 | // Simple tensor for the transformation matrix |
| 121 | SimpleTensor<T> trans_matrix{ TensorShape(3u, 4u), in.data_type(), 1 }; |
| 122 | |
| 123 | // Simple tensor for the transformation matrix transpose |
| 124 | SimpleTensor<T> trans_matrix_transposed{ TensorShape(4u, 3u), in.data_type(), 1 }; |
| 125 | |
| 126 | // Simple tensor for the 4x3 temporary tile |
| 127 | SimpleTensor<T> tmp_tile{ TensorShape(3u, 4u), in.data_type(), 1 }; |
| 128 | |
| 129 | // Simple tensor for the 4x4 output tile |
| 130 | SimpleTensor<T> output_tile{ TensorShape(4u, 4u), in.data_type(), 1 }; |
| 131 | |
| 132 | // Initialize transformation matrix |
| 133 | // 1 | 0 | 0 |
| 134 | // 0.5 | 0.5 | 0.5 |
| 135 | // 0.5 |-0.5 | 0.5 |
| 136 | // 0 | 0 | 1 |
| 137 | trans_matrix[0 + 0 * 3] = 1.0f; |
| 138 | trans_matrix[1 + 0 * 3] = 0.0f; |
| 139 | trans_matrix[2 + 0 * 3] = 0.0f; |
| 140 | trans_matrix[0 + 1 * 3] = 0.5f; |
| 141 | trans_matrix[1 + 1 * 3] = 0.5f; |
| 142 | trans_matrix[2 + 1 * 3] = 0.5f; |
| 143 | trans_matrix[0 + 2 * 3] = 0.5f; |
| 144 | trans_matrix[1 + 2 * 3] = -0.5f; |
| 145 | trans_matrix[2 + 2 * 3] = 0.5f; |
| 146 | trans_matrix[0 + 3 * 3] = 0.0f; |
| 147 | trans_matrix[1 + 3 * 3] = 0.0f; |
| 148 | trans_matrix[2 + 3 * 3] = 1.0f; |
| 149 | |
| 150 | // Transpose the transformation matrix |
| 151 | transpose_matrix(trans_matrix, trans_matrix_transposed); |
| 152 | |
| 153 | const int num_channels = in.shape()[2]; |
| 154 | const int num_filters = in.shape()[3]; |
| 155 | const int num_batches = in.shape().total_size() / (9 * num_channels * num_filters); |
| 156 | |
| 157 | for(int n = 0; n < num_batches; ++n) |
| 158 | { |
| 159 | for(int w = 0; w < num_filters; ++w) |
| 160 | { |
| 161 | for(int z = 0; z < num_channels; ++z) |
| 162 | { |
| 163 | // Load the 3x3 tile from the input tensor |
| 164 | get_tile(in, input_tile, Coordinates(0, 0, z, w, n)); |
| 165 | |
| 166 | // First transformation |
| 167 | matrix_multiply(trans_matrix, input_tile, tmp_tile); |
| 168 | |
| 169 | // Second transformation |
| 170 | matrix_multiply(tmp_tile, trans_matrix_transposed, output_tile); |
| 171 | |
| 172 | // Store the 4x4 output tile across the 16 channels |
| 173 | const int output_offset = w + z * num_filters; |
| 174 | out[output_offset + 0 * num_filters * num_channels] = output_tile[0 + 0 * 4]; |
| 175 | out[output_offset + 1 * num_filters * num_channels] = output_tile[1 + 0 * 4]; |
| 176 | out[output_offset + 2 * num_filters * num_channels] = output_tile[2 + 0 * 4]; |
| 177 | out[output_offset + 3 * num_filters * num_channels] = output_tile[3 + 0 * 4]; |
| 178 | out[output_offset + 4 * num_filters * num_channels] = output_tile[0 + 1 * 4]; |
| 179 | out[output_offset + 5 * num_filters * num_channels] = output_tile[1 + 1 * 4]; |
| 180 | out[output_offset + 6 * num_filters * num_channels] = output_tile[2 + 1 * 4]; |
| 181 | out[output_offset + 7 * num_filters * num_channels] = output_tile[3 + 1 * 4]; |
| 182 | out[output_offset + 8 * num_filters * num_channels] = output_tile[0 + 2 * 4]; |
| 183 | out[output_offset + 9 * num_filters * num_channels] = output_tile[1 + 2 * 4]; |
| 184 | out[output_offset + 10 * num_filters * num_channels] = output_tile[2 + 2 * 4]; |
| 185 | out[output_offset + 11 * num_filters * num_channels] = output_tile[3 + 2 * 4]; |
| 186 | out[output_offset + 12 * num_filters * num_channels] = output_tile[0 + 3 * 4]; |
| 187 | out[output_offset + 13 * num_filters * num_channels] = output_tile[1 + 3 * 4]; |
| 188 | out[output_offset + 14 * num_filters * num_channels] = output_tile[2 + 3 * 4]; |
| 189 | out[output_offset + 15 * num_filters * num_channels] = output_tile[3 + 3 * 4]; |
| 190 | } |
| 191 | } |
| 192 | } |
| 193 | } |
Giorgio Arena | 1f9ca1d | 2018-03-01 11:13:45 +0000 | [diff] [blame] | 194 | } // namespace |
| 195 | |
| 196 | template <typename T> |
| 197 | SimpleTensor<T> winograd_input_transform(const SimpleTensor<T> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims) |
| 198 | { |
| 199 | ARM_COMPUTE_ERROR_ON(kernel_dims.width != kernel_dims.height); |
| 200 | ARM_COMPUTE_ERROR_ON(src.data_layout() != DataLayout::NCHW); |
| 201 | |
| 202 | SimpleTensor<T> dst{ dst_shape, src.data_type() }; |
| 203 | |
| 204 | switch(kernel_dims.width) |
| 205 | { |
| 206 | case 3: |
| 207 | winograd_input_transform3x3(src, dst, conv_info); |
| 208 | break; |
| 209 | default: |
| 210 | ARM_COMPUTE_ERROR("Only 3x3 kernels are supported"); |
| 211 | } |
| 212 | |
| 213 | return dst; |
| 214 | } |
| 215 | |
Gian Marco Iodice | 7e4b239 | 2018-02-22 16:17:20 +0000 | [diff] [blame^] | 216 | template <typename T> |
| 217 | SimpleTensor<T> winograd_filter_transform(const SimpleTensor<T> &in, const TensorShape &output_shape) |
| 218 | { |
| 219 | ARM_COMPUTE_ERROR_ON_MSG(in.data_layout() != DataLayout::NCHW, "Only supported NCHW data format"); |
| 220 | |
| 221 | // Create reference |
| 222 | SimpleTensor<T> out{ output_shape, in.data_type(), 1 }; |
| 223 | |
| 224 | switch(in.shape()[0]) |
| 225 | { |
| 226 | case 3: |
| 227 | winograd_filter_transform3x3(in, out); |
| 228 | break; |
| 229 | default: |
| 230 | ARM_COMPUTE_ERROR("Only supported 3x3 kernel"); |
| 231 | break; |
| 232 | } |
| 233 | |
| 234 | return out; |
| 235 | } |
| 236 | |
Giorgio Arena | 1f9ca1d | 2018-03-01 11:13:45 +0000 | [diff] [blame] | 237 | template SimpleTensor<float> winograd_input_transform(const SimpleTensor<float> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims); |
Gian Marco Iodice | 7e4b239 | 2018-02-22 16:17:20 +0000 | [diff] [blame^] | 238 | template SimpleTensor<float> winograd_filter_transform(const SimpleTensor<float> &in, const TensorShape &output_shape); |
Giorgio Arena | 1f9ca1d | 2018-03-01 11:13:45 +0000 | [diff] [blame] | 239 | } // namespace reference |
| 240 | } // namespace validation |
| 241 | } // namespace test |
| 242 | } // namespace arm_compute |