Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 1 | /* |
Michalis Spyrou | d1d7722 | 2020-04-08 14:10:15 +0100 | [diff] [blame] | 2 | * Copyright (c) 2018-2020 ARM Limited. |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +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 | */ |
| 24 | #include "Im2Col.h" |
| 25 | |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 26 | #include "arm_compute/core/Types.h" |
| 27 | #include "tests/validation/Helpers.h" |
| 28 | #include "tests/validation/reference/Utils.h" |
| 29 | |
| 30 | namespace arm_compute |
| 31 | { |
| 32 | namespace test |
| 33 | { |
| 34 | namespace validation |
| 35 | { |
| 36 | namespace reference |
| 37 | { |
| 38 | template <typename T> |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 39 | void im2col_nchw(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups) |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 40 | { |
Pablo Tello | 764b1af | 2018-04-23 16:11:45 +0100 | [diff] [blame] | 41 | ARM_COMPUTE_ERROR_ON(src.data_layout() != DataLayout::NCHW); |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 42 | const int stride_x = conv_info.stride().first; |
| 43 | const int stride_y = conv_info.stride().second; |
| 44 | const int kernel_width = kernel_dims.width; |
| 45 | const int kernel_height = kernel_dims.height; |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 46 | const int pad_x = conv_info.pad().first; |
| 47 | const int pad_y = conv_info.pad().second; |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 48 | const int src_width = src.shape().x(); |
| 49 | const int src_height = src.shape().y(); |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 50 | const int src_channels = src.shape().z(); |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 51 | const int batches = src.shape().total_size_upper(3); |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 52 | const int dst_height = dst.shape().y(); |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 53 | const int pad_val = is_data_type_quantized_asymmetric(src.data_type()) ? src.quantization_info().uniform().offset : 0; |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 54 | int dst_idx = 0; |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 55 | |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 56 | // Compute width and height of the convolved tensors |
| 57 | std::pair<unsigned int, unsigned int> convolved_dims = scaled_dimensions(src_width, src_height, kernel_dims.width, kernel_dims.height, conv_info); |
Pablo Tello | 4a626a7 | 2018-04-04 10:01:14 +0100 | [diff] [blame] | 58 | |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 59 | for(int b = 0; b < batches; ++b) |
| 60 | { |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 61 | for(int g = 0; g < static_cast<int>(num_groups); ++g) |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 62 | { |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 63 | const int first_group_ch = g * (src_channels / num_groups); |
| 64 | const int last_group_ch = (g + 1) * (src_channels / num_groups); |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 65 | |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 66 | for(int yo = 0; yo < dst_height; ++yo) |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 67 | { |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 68 | // Compute input spatial coordinates |
| 69 | const int xi = (yo % convolved_dims.first) * stride_x; |
| 70 | const int yi = (yo / convolved_dims.first) * stride_y; |
| 71 | |
| 72 | for(int ci = first_group_ch; ci < last_group_ch; ++ci) |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 73 | { |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 74 | for(int yk = 0; yk < kernel_height; ++yk) |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 75 | { |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 76 | for(int xk = 0; xk < kernel_width; ++xk) |
| 77 | { |
| 78 | dst[dst_idx++] = tensor_elem_at(src, Coordinates(xi + xk - pad_x, yi + yk - pad_y, ci, b), BorderMode::CONSTANT, static_cast<T>(pad_val)); |
| 79 | } |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 80 | } |
| 81 | } |
| 82 | |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 83 | if(has_bias) |
| 84 | { |
| 85 | dst[dst_idx++] = static_cast<T>(1); |
| 86 | } |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 87 | } |
| 88 | } |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | template <typename T> |
Pablo Tello | 764b1af | 2018-04-23 16:11:45 +0100 | [diff] [blame] | 93 | void im2col_nhwc(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias) |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 94 | { |
Pablo Tello | 764b1af | 2018-04-23 16:11:45 +0100 | [diff] [blame] | 95 | ARM_COMPUTE_ERROR_ON(src.data_layout() != DataLayout::NHWC); |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 96 | const int stride_x = conv_info.stride().first; |
| 97 | const int stride_y = conv_info.stride().second; |
| 98 | const int kernel_width = kernel_dims.width; |
| 99 | const int kernel_height = kernel_dims.height; |
| 100 | const int pad_x = conv_info.pad().first; |
| 101 | const int pad_y = conv_info.pad().second; |
| 102 | const int src_width = src.shape().y(); |
| 103 | const int src_height = src.shape().z(); |
| 104 | const int src_channels = src.shape().x(); |
| 105 | const int batches = src.shape().total_size_upper(3); |
| 106 | const int dst_width = has_bias ? dst.shape().x() - 1 : dst.shape().x(); |
| 107 | const int dst_height = dst.shape().y(); |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 108 | const int pad_val = is_data_type_quantized_asymmetric(src.data_type()) ? src.quantization_info().uniform().offset : 0; |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 109 | |
| 110 | // Compute width and height of the convolved tensors |
| 111 | std::pair<unsigned int, unsigned int> convolved_dims = scaled_dimensions(src_width, src_height, kernel_dims.width, kernel_dims.height, conv_info); |
Michalis Spyrou | d1d7722 | 2020-04-08 14:10:15 +0100 | [diff] [blame] | 112 | #if defined(_OPENMP) |
| 113 | #pragma omp parallel for schedule(dynamic, 1) collapse(2) |
| 114 | #endif /* _OPENMP */ |
Gian Marco Iodice | 215b4ea | 2018-06-28 16:29:29 +0100 | [diff] [blame] | 115 | for(int b = 0; b < batches; ++b) |
| 116 | { |
| 117 | for(int yo = 0; yo < dst_height; ++yo) |
| 118 | { |
| 119 | // Compute input spatial coordinates |
| 120 | const int xi = (yo % convolved_dims.first) * stride_x; |
| 121 | const int yi = (yo / convolved_dims.first) * stride_y; |
| 122 | |
| 123 | for(int ci = 0; ci < src_channels; ++ci) |
| 124 | { |
| 125 | for(int yk = 0; yk < kernel_height; ++yk) |
| 126 | { |
| 127 | for(int xk = 0; xk < kernel_width; ++xk) |
| 128 | { |
| 129 | dst[ci + (xk + yk * kernel_width) * src_channels + yo * dst.shape().x() + b * dst.shape().x() * dst.shape().y()] = tensor_elem_at(src, Coordinates(ci, xi + xk - pad_x, yi + yk - pad_y, b), |
| 130 | BorderMode::CONSTANT, static_cast<T>(pad_val)); |
| 131 | } |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | if(has_bias) |
| 136 | { |
| 137 | dst[dst_width + yo * dst.shape().x() + b * dst.shape().x() * dst.shape().y()] = static_cast<T>(1); |
| 138 | } |
| 139 | } |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | template <typename T> |
Georgios Pinitas | 75bde5e | 2019-06-07 11:52:01 +0100 | [diff] [blame] | 144 | void im2col(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups) |
Pablo Tello | 764b1af | 2018-04-23 16:11:45 +0100 | [diff] [blame] | 145 | { |
| 146 | switch(src.data_layout()) |
| 147 | { |
| 148 | case DataLayout::NCHW: |
| 149 | { |
Giorgio Arena | 0f17039 | 2018-07-18 16:13:12 +0100 | [diff] [blame] | 150 | im2col_nchw(src, dst, kernel_dims, conv_info, has_bias, num_groups); |
Pablo Tello | 764b1af | 2018-04-23 16:11:45 +0100 | [diff] [blame] | 151 | break; |
| 152 | } |
| 153 | case DataLayout::NHWC: |
| 154 | { |
Giorgio Arena | fb62908 | 2018-08-20 18:03:27 +0100 | [diff] [blame] | 155 | im2col_nhwc(src, dst, kernel_dims, conv_info, has_bias); |
Pablo Tello | 764b1af | 2018-04-23 16:11:45 +0100 | [diff] [blame] | 156 | break; |
| 157 | } |
| 158 | default: |
| 159 | { |
| 160 | ARM_COMPUTE_ERROR("Not supported."); |
| 161 | break; |
| 162 | } |
| 163 | } |
| 164 | } |
| 165 | |
Giorgio Arena | fb62908 | 2018-08-20 18:03:27 +0100 | [diff] [blame] | 166 | template void im2col(const SimpleTensor<uint8_t> &src, SimpleTensor<uint8_t> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups); |
| 167 | template void im2col(const SimpleTensor<half> &src, SimpleTensor<half> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups); |
| 168 | template void im2col(const SimpleTensor<float> &src, SimpleTensor<float> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups); |
Giorgio Arena | 156fcf3 | 2018-03-09 15:30:43 +0000 | [diff] [blame] | 169 | } // namespace reference |
| 170 | } // namespace validation |
| 171 | } // namespace test |
| 172 | } // namespace arm_compute |