Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 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/NEON/kernels/NEHistogramKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/IDistribution1D.h" |
| 29 | #include "arm_compute/core/ITensor.h" |
| 30 | #include "arm_compute/core/TensorInfo.h" |
| 31 | #include "arm_compute/core/Types.h" |
| 32 | #include "arm_compute/core/Window.h" |
| 33 | |
| 34 | #include <algorithm> |
| 35 | #include <arm_neon.h> |
| 36 | #include <array> |
| 37 | |
| 38 | using namespace arm_compute; |
| 39 | |
| 40 | namespace arm_compute |
| 41 | { |
| 42 | class Coordinates; |
| 43 | } // namespace arm_compute |
| 44 | |
| 45 | inline void NEHistogramKernel::merge_histogram(uint32_t *global_hist, const uint32_t *local_hist, size_t bins) |
| 46 | { |
| 47 | std::lock_guard<std::mutex> lock(_hist_mtx); |
| 48 | |
| 49 | const unsigned int v_end = (bins / 4) * 4; |
| 50 | |
| 51 | for(unsigned int b = 0; b < v_end; b += 4) |
| 52 | { |
| 53 | const uint32x4_t tmp_global = vld1q_u32(global_hist + b); |
| 54 | const uint32x4_t tmp_local = vld1q_u32(local_hist + b); |
| 55 | vst1q_u32(global_hist + b, vaddq_u32(tmp_global, tmp_local)); |
| 56 | } |
| 57 | |
| 58 | for(unsigned int b = v_end; b < bins; ++b) |
| 59 | { |
| 60 | global_hist[b] += local_hist[b]; |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | NEHistogramKernel::NEHistogramKernel() |
| 65 | : _func(nullptr), _input(nullptr), _output(nullptr), _local_hist(nullptr), _window_lut(nullptr), _hist_mtx() |
| 66 | { |
| 67 | } |
| 68 | |
| 69 | void NEHistogramKernel::histogram_U8(Window win) |
| 70 | { |
| 71 | ARM_COMPUTE_ERROR_ON(_output->buffer() == nullptr); |
| 72 | |
| 73 | const size_t bins = _output->num_bins(); |
| 74 | const int32_t offset = _output->offset(); |
| 75 | const uint32_t offrange = offset + _output->range(); |
| 76 | const uint32_t *const w_lut = _window_lut; |
| 77 | uint32_t *const local_hist = _local_hist + win.thread_id() * bins; |
| 78 | |
| 79 | // Clear local_histogram |
| 80 | std::fill_n(local_hist, bins, 0); |
| 81 | |
| 82 | auto update_local_hist = [&](uint8_t p) |
| 83 | { |
| 84 | if(offset <= p && p < offrange) |
| 85 | { |
| 86 | ++local_hist[w_lut[p]]; |
| 87 | } |
| 88 | }; |
| 89 | |
| 90 | const unsigned int x_start = win.x().start(); |
| 91 | const unsigned int x_end = win.x().end(); |
| 92 | |
| 93 | // Handle X dimension manually to split into two loops |
| 94 | // First one will use vector operations, second one processes the left over |
| 95 | // pixels |
| 96 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 97 | |
| 98 | Iterator input(_input, win); |
| 99 | |
| 100 | // Calculate local histogram |
| 101 | execute_window_loop(win, [&](const Coordinates &) |
| 102 | { |
| 103 | unsigned int x = x_start; |
| 104 | |
| 105 | // Vector loop |
| 106 | for(; x <= x_end - 8; x += 8) |
| 107 | { |
| 108 | const uint8x8_t pixels = vld1_u8(input.ptr() + x); |
| 109 | |
| 110 | update_local_hist(vget_lane_u8(pixels, 0)); |
| 111 | update_local_hist(vget_lane_u8(pixels, 1)); |
| 112 | update_local_hist(vget_lane_u8(pixels, 2)); |
| 113 | update_local_hist(vget_lane_u8(pixels, 3)); |
| 114 | update_local_hist(vget_lane_u8(pixels, 4)); |
| 115 | update_local_hist(vget_lane_u8(pixels, 5)); |
| 116 | update_local_hist(vget_lane_u8(pixels, 6)); |
| 117 | update_local_hist(vget_lane_u8(pixels, 7)); |
| 118 | } |
| 119 | |
| 120 | // Process leftover pixels |
| 121 | for(; x < x_end; ++x) |
| 122 | { |
| 123 | update_local_hist(input.ptr()[x]); |
| 124 | } |
| 125 | }, |
| 126 | input); |
| 127 | |
| 128 | // Merge histograms |
| 129 | merge_histogram(_output->buffer(), local_hist, bins); |
| 130 | } |
| 131 | |
| 132 | void NEHistogramKernel::histogram_fixed_U8(Window win) |
| 133 | { |
| 134 | ARM_COMPUTE_ERROR_ON(_output->buffer() == nullptr); |
| 135 | |
| 136 | std::array<uint32_t, _max_range_size> local_hist{ { 0 } }; |
| 137 | |
| 138 | const unsigned int x_start = win.x().start(); |
| 139 | const unsigned int x_end = win.x().end(); |
| 140 | |
| 141 | // Handle X dimension manually to split into two loops |
| 142 | // First one will use vector operations, second one processes the left over |
| 143 | // pixels |
| 144 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 145 | |
| 146 | Iterator input(_input, win); |
| 147 | |
| 148 | // Calculate local histogram |
| 149 | execute_window_loop(win, [&](const Coordinates &) |
| 150 | { |
| 151 | unsigned int x = x_start; |
| 152 | |
| 153 | // Vector loop |
| 154 | for(; x <= x_end - 8; x += 8) |
| 155 | { |
| 156 | const uint8x8_t pixels = vld1_u8(input.ptr() + x); |
| 157 | |
| 158 | ++local_hist[vget_lane_u8(pixels, 0)]; |
| 159 | ++local_hist[vget_lane_u8(pixels, 1)]; |
| 160 | ++local_hist[vget_lane_u8(pixels, 2)]; |
| 161 | ++local_hist[vget_lane_u8(pixels, 3)]; |
| 162 | ++local_hist[vget_lane_u8(pixels, 4)]; |
| 163 | ++local_hist[vget_lane_u8(pixels, 5)]; |
| 164 | ++local_hist[vget_lane_u8(pixels, 6)]; |
| 165 | ++local_hist[vget_lane_u8(pixels, 7)]; |
| 166 | } |
| 167 | |
| 168 | // Process leftover pixels |
| 169 | for(; x < x_end; ++x) |
| 170 | { |
| 171 | ++local_hist[input.ptr()[x]]; |
| 172 | } |
| 173 | }, |
| 174 | input); |
| 175 | |
| 176 | // Merge histograms |
| 177 | merge_histogram(_output->buffer(), local_hist.data(), _max_range_size); |
| 178 | } |
| 179 | |
| 180 | void NEHistogramKernel::calculate_window_lut() const |
| 181 | { |
| 182 | const int32_t offset = _output->offset(); |
| 183 | const size_t bins = _output->num_bins(); |
| 184 | const uint32_t range = _output->range(); |
| 185 | |
| 186 | std::fill_n(_window_lut, offset, 0); |
| 187 | |
| 188 | for(unsigned int p = offset; p < _max_range_size; ++p) |
| 189 | { |
| 190 | _window_lut[p] = ((p - offset) * bins) / range; |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | void NEHistogramKernel::configure(const IImage *input, IDistribution1D *output, uint32_t *local_hist, uint32_t *window_lut) |
| 195 | { |
| 196 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input); |
| 197 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 198 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 199 | ARM_COMPUTE_ERROR_ON(nullptr == local_hist); |
| 200 | ARM_COMPUTE_ERROR_ON(nullptr == window_lut); |
| 201 | |
| 202 | _input = input; |
| 203 | _output = output; |
| 204 | _local_hist = local_hist; |
| 205 | _window_lut = window_lut; |
| 206 | |
| 207 | //Check offset |
| 208 | ARM_COMPUTE_ERROR_ON_MSG(0 > _output->offset() || _output->offset() > static_cast<int32_t>(_max_range_size), "Offset is larger than the image value range."); |
| 209 | |
| 210 | //Check range |
| 211 | ARM_COMPUTE_ERROR_ON_MSG(static_cast<int32_t>(_output->range()) > static_cast<int32_t>(_max_range_size) /* max range */, "Range larger than the image value range."); |
| 212 | |
| 213 | // Calculate LUT |
| 214 | calculate_window_lut(); |
| 215 | |
| 216 | // Set appropriate function |
| 217 | _func = &NEHistogramKernel::histogram_U8; |
| 218 | |
| 219 | constexpr unsigned int num_elems_processed_per_iteration = 1; |
| 220 | |
| 221 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); |
| 222 | |
| 223 | INEKernel::configure(win); |
| 224 | } |
| 225 | |
| 226 | void NEHistogramKernel::configure(const IImage *input, IDistribution1D *output) |
| 227 | { |
| 228 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input); |
| 229 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 230 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 231 | |
| 232 | _input = input; |
| 233 | _output = output; |
| 234 | |
| 235 | // Set appropriate function |
| 236 | _func = &NEHistogramKernel::histogram_fixed_U8; |
| 237 | |
| 238 | constexpr unsigned int num_elems_processed_per_iteration = 1; |
| 239 | |
| 240 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); |
| 241 | |
| 242 | INEKernel::configure(win); |
| 243 | } |
| 244 | |
| 245 | void NEHistogramKernel::run(const Window &window) |
| 246 | { |
| 247 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 248 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 249 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 250 | |
| 251 | (this->*_func)(window); |
| 252 | } |