Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 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 "arm_compute/core/CL/kernels/CLGaussianPyramidKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 27 | #include "arm_compute/core/CL/ICLTensor.h" |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/Utils.h" |
| 30 | #include "arm_compute/core/Validate.h" |
| 31 | |
| 32 | using namespace arm_compute; |
| 33 | |
| 34 | CLGaussianPyramidHorKernel::CLGaussianPyramidHorKernel() |
| 35 | : _border_size(0), _l2_load_offset(0) |
| 36 | { |
| 37 | } |
| 38 | |
| 39 | BorderSize CLGaussianPyramidHorKernel::border_size() const |
| 40 | { |
| 41 | return _border_size; |
| 42 | } |
| 43 | |
| 44 | void CLGaussianPyramidHorKernel::configure(const ICLTensor *input, ICLTensor *output, bool border_undefined) |
| 45 | { |
| 46 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 47 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U16); |
| 48 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != 2 * output->info()->dimension(0)); |
| 49 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != output->info()->dimension(1)); |
| 50 | |
| 51 | for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) |
| 52 | { |
| 53 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i)); |
| 54 | } |
| 55 | |
| 56 | _input = input; |
| 57 | _output = output; |
| 58 | _border_size = BorderSize(border_undefined ? 0 : 2, 2); |
| 59 | |
| 60 | // Create kernel |
| 61 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gaussian1x5_sub_x")); |
| 62 | |
| 63 | // Configure kernel window |
| 64 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 65 | constexpr unsigned int num_elems_read_per_iteration = 20; |
| 66 | constexpr unsigned int num_elems_written_per_iteration = 8; |
| 67 | constexpr float scale_x = 0.5f; |
| 68 | |
| 69 | Window win = calculate_max_window_horizontal(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 70 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration, scale_x); |
| 71 | |
| 72 | // Sub sampling selects odd pixels (1, 3, 5, ...) for images with even |
| 73 | // width and even pixels (0, 2, 4, ...) for images with odd width. (Whether |
| 74 | // a pixel is even or odd is determined based on the tensor shape not the |
| 75 | // valid region!) |
| 76 | // Thus the offset from which the first pixel (L2) for the convolution is |
| 77 | // loaded depends on the anchor and shape of the valid region. |
| 78 | // In the case of an even shape (= even image width) we need to load L2 |
| 79 | // from -2 if the anchor is odd and from -1 if the anchor is even. That |
| 80 | // makes sure that L2 is always loaded from an odd pixel. |
| 81 | // On the other hand, for an odd shape (= odd image width) we need to load |
| 82 | // L2 from -1 if the anchor is odd and from -2 if the anchor is even to |
| 83 | // achieve the opposite effect. |
| 84 | // The condition can be simplified to checking whether anchor + shape is |
| 85 | // odd (-2) or even (-1) as only adding an odd and an even number will have |
| 86 | // an odd result. |
| 87 | _l2_load_offset = -border_size().left; |
| 88 | |
| 89 | if((_input->info()->valid_region().anchor[0] + _input->info()->valid_region().shape[0]) % 2 == 0) |
| 90 | { |
| 91 | _l2_load_offset += 1; |
| 92 | } |
| 93 | |
| 94 | update_window_and_padding(win, |
| 95 | AccessWindowHorizontal(input->info(), _l2_load_offset, num_elems_read_per_iteration), |
| 96 | output_access); |
| 97 | |
| 98 | ValidRegion valid_region = input->info()->valid_region(); |
| 99 | valid_region.anchor.set(0, std::ceil((valid_region.anchor[0] + (border_undefined ? border_size().left : 0)) / 2.f)); |
| 100 | valid_region.shape.set(0, (valid_region.shape[0] - (border_undefined ? border_size().right : 0)) / 2 - valid_region.anchor[0]); |
| 101 | |
| 102 | output_access.set_valid_region(win, valid_region); |
| 103 | |
| 104 | ICLKernel::configure(win); |
| 105 | } |
| 106 | |
| 107 | void CLGaussianPyramidHorKernel::run(const Window &window, cl::CommandQueue &queue) |
| 108 | { |
| 109 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 110 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 111 | |
| 112 | Window win_in(window); |
| 113 | win_in.shift(Window::DimX, _l2_load_offset); |
| 114 | |
| 115 | //The output is half the width of the input: |
| 116 | Window win_out(window); |
| 117 | win_out.scale(Window::DimX, 0.5f); |
| 118 | |
| 119 | Window slice_in = win_in.first_slice_window_2D(); |
| 120 | Window slice_out = win_out.first_slice_window_2D(); |
| 121 | |
| 122 | do |
| 123 | { |
| 124 | unsigned int idx = 0; |
| 125 | add_2D_tensor_argument(idx, _input, slice_in); |
| 126 | add_2D_tensor_argument(idx, _output, slice_out); |
| 127 | enqueue(queue, *this, slice_out); |
| 128 | } |
| 129 | while(win_in.slide_window_slice_2D(slice_in) && win_out.slide_window_slice_2D(slice_out)); |
| 130 | } |
| 131 | |
| 132 | CLGaussianPyramidVertKernel::CLGaussianPyramidVertKernel() |
| 133 | : _t2_load_offset(0) |
| 134 | { |
| 135 | } |
| 136 | |
| 137 | BorderSize CLGaussianPyramidVertKernel::border_size() const |
| 138 | { |
| 139 | return BorderSize(2, 0); |
| 140 | } |
| 141 | |
| 142 | void CLGaussianPyramidVertKernel::configure(const ICLTensor *input, ICLTensor *output, bool border_undefined) |
| 143 | { |
| 144 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U16); |
| 145 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8); |
| 146 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != output->info()->dimension(0)); |
| 147 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != 2 * output->info()->dimension(1)); |
| 148 | |
| 149 | for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) |
| 150 | { |
| 151 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i)); |
| 152 | } |
| 153 | |
| 154 | _input = input; |
| 155 | _output = output; |
| 156 | |
| 157 | // Create kernel |
| 158 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gaussian5x1_sub_y")); |
| 159 | |
| 160 | // Configure kernel window |
| 161 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 162 | constexpr unsigned int num_rows_processed_per_iteration = 2; |
| 163 | constexpr unsigned int num_elems_written_per_iteration = 8; |
| 164 | constexpr unsigned int num_elems_read_per_iteration = 8; |
| 165 | constexpr unsigned int num_rows_per_iteration = 5; |
| 166 | constexpr float scale_y = 0.5f; |
| 167 | |
| 168 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration, num_rows_processed_per_iteration), |
| 169 | border_undefined, border_size()); |
| 170 | AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration, num_rows_per_iteration, 1.f, scale_y); |
| 171 | |
| 172 | // Determine whether we need to load even or odd rows. See above for a |
| 173 | // detailed explanation. |
| 174 | _t2_load_offset = -border_size().top; |
| 175 | |
| 176 | if((_input->info()->valid_region().anchor[1] + _input->info()->valid_region().shape[1]) % 2 == 0) |
| 177 | { |
| 178 | _t2_load_offset += 1; |
| 179 | } |
| 180 | |
| 181 | update_window_and_padding(win, |
| 182 | AccessWindowRectangle(input->info(), 0, _t2_load_offset, num_elems_read_per_iteration, num_rows_per_iteration), |
| 183 | output_access); |
| 184 | |
| 185 | ValidRegion valid_region = input->info()->valid_region(); |
| 186 | valid_region.anchor.set(1, std::ceil((valid_region.anchor[1] + (border_undefined ? border_size().top : 0)) / 2.f)); |
| 187 | valid_region.shape.set(1, (valid_region.shape[1] - (border_undefined ? border_size().bottom : 0)) / 2 - valid_region.anchor[1]); |
| 188 | |
| 189 | output_access.set_valid_region(win, valid_region); |
| 190 | |
| 191 | ICLKernel::configure(win); |
| 192 | } |
| 193 | |
| 194 | void CLGaussianPyramidVertKernel::run(const Window &window, cl::CommandQueue &queue) |
| 195 | { |
| 196 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 197 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 198 | ARM_COMPUTE_ERROR_ON(window.x().step() != 8); |
| 199 | ARM_COMPUTE_ERROR_ON(window.y().step() % 2); |
| 200 | |
| 201 | Window win_in(window); |
| 202 | win_in.shift(Window::DimY, _t2_load_offset); |
| 203 | |
| 204 | Window win_out(window); |
| 205 | win_out.scale(Window::DimY, 0.5f); |
| 206 | |
| 207 | Window slice_in = win_in.first_slice_window_2D(); |
| 208 | Window slice_out = win_out.first_slice_window_2D(); |
| 209 | |
| 210 | do |
| 211 | { |
| 212 | unsigned int idx = 0; |
| 213 | add_2D_tensor_argument(idx, _input, slice_in); |
| 214 | add_2D_tensor_argument(idx, _output, slice_out); |
| 215 | enqueue(queue, *this, slice_out); |
| 216 | } |
| 217 | while(win_in.slide_window_slice_2D(slice_in) && win_out.slide_window_slice_2D(slice_out)); |
| 218 | } |