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/NEHarrisCornersKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Coordinates.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/Helpers.h" |
| 29 | #include "arm_compute/core/TensorInfo.h" |
| 30 | #include "arm_compute/core/Types.h" |
| 31 | #include "arm_compute/core/Utils.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | #include "arm_compute/core/Window.h" |
| 34 | |
| 35 | #include <algorithm> |
| 36 | #include <arm_neon.h> |
| 37 | #include <cmath> |
| 38 | #include <cstddef> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 42 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 43 | |
| 44 | template class arm_compute::NEHarrisScoreFP16Kernel<3>; |
| 45 | template class arm_compute::NEHarrisScoreFP16Kernel<5>; |
| 46 | template class arm_compute::NEHarrisScoreFP16Kernel<7>; |
| 47 | |
| 48 | namespace fp16 |
| 49 | { |
| 50 | inline float16x8_t harris_score(float16x8_t gx2, float16x8_t gy2, float16x8_t gxgy, float sensitivity, float strength_thresh) |
| 51 | { |
| 52 | static const float16x8_t zero = vdupq_n_f16(0.f); |
| 53 | |
| 54 | // Trace^2 |
| 55 | float16x8_t trace2 = vaddq_f16(gx2, gy2); |
| 56 | trace2 = vmulq_f16(trace2, trace2); |
| 57 | |
| 58 | // Det(A) |
| 59 | float16x8_t det = vmulq_f16(gx2, gy2); |
| 60 | det = vfmsq_f16(det, gxgy, gxgy); |
| 61 | |
| 62 | // Det(A) - sensitivity * trace^2 |
| 63 | const float16x8_t mc = vfmsq_f16(det, vdupq_n_f16(sensitivity), trace2); |
| 64 | |
| 65 | // mc > strength_thresh |
| 66 | const uint16x8_t mask = vcgtq_f16(mc, vdupq_n_f16(strength_thresh)); |
| 67 | |
| 68 | return vbslq_f16(mask, mc, zero); |
| 69 | } |
| 70 | |
| 71 | template <size_t block_size> |
| 72 | inline void harris_score1xN_FLOAT_FLOAT_FLOAT(float16x8_t low_gx, float16x8_t low_gy, float16x8_t high_gx, float16x8_t high_gy, float16x8_t &gx2, float16x8_t &gy2, float16x8_t &gxgy, |
| 73 | float norm_factor) |
| 74 | { |
| 75 | const float16x8_t norm_factor_fp16 = vdupq_n_f16(norm_factor); |
| 76 | |
| 77 | // Normalize |
| 78 | low_gx = vmulq_f16(low_gx, norm_factor_fp16); |
| 79 | low_gy = vmulq_f16(low_gy, norm_factor_fp16); |
| 80 | high_gx = vmulq_f16(high_gx, norm_factor_fp16); |
| 81 | high_gy = vmulq_f16(high_gy, norm_factor_fp16); |
| 82 | |
| 83 | float16x8_t gx = vextq_f16(low_gx, high_gx, 0); |
| 84 | float16x8_t gy = vextq_f16(low_gy, high_gy, 0); |
| 85 | |
| 86 | gx2 = vfmaq_f16(gx2, gx, gx); |
| 87 | gy2 = vfmaq_f16(gy2, gy, gy); |
| 88 | gxgy = vfmaq_f16(gxgy, gx, gy); |
| 89 | |
| 90 | gx = vextq_f16(low_gx, high_gx, 1); |
| 91 | gy = vextq_f16(low_gy, high_gy, 1); |
| 92 | |
| 93 | gx2 = vfmaq_f16(gx2, gx, gx); |
| 94 | gy2 = vfmaq_f16(gy2, gy, gy); |
| 95 | gxgy = vfmaq_f16(gxgy, gx, gy); |
| 96 | |
| 97 | gx = vextq_f16(low_gx, high_gx, 2); |
| 98 | gy = vextq_f16(low_gy, high_gy, 2); |
| 99 | |
| 100 | gx2 = vfmaq_f16(gx2, gx, gx); |
| 101 | gy2 = vfmaq_f16(gy2, gy, gy); |
| 102 | gxgy = vfmaq_f16(gxgy, gx, gy); |
| 103 | |
| 104 | if(block_size > 3) |
| 105 | { |
| 106 | gx = vextq_f16(low_gx, high_gx, 3); |
| 107 | gy = vextq_f16(low_gy, high_gy, 3); |
| 108 | |
| 109 | gx2 = vfmaq_f16(gx2, gx, gx); |
| 110 | gy2 = vfmaq_f16(gy2, gy, gy); |
| 111 | gxgy = vfmaq_f16(gxgy, gx, gy); |
| 112 | |
| 113 | gx = vextq_f16(low_gx, high_gx, 4); |
| 114 | gy = vextq_f16(low_gy, high_gy, 4); |
| 115 | |
| 116 | gx2 = vfmaq_f16(gx2, gx, gx); |
| 117 | gy2 = vfmaq_f16(gy2, gy, gy); |
| 118 | gxgy = vfmaq_f16(gxgy, gx, gy); |
| 119 | } |
| 120 | |
| 121 | if(block_size == 7) |
| 122 | { |
| 123 | gx = vextq_f16(low_gx, high_gx, 5); |
| 124 | gy = vextq_f16(low_gy, high_gy, 5); |
| 125 | |
| 126 | gx2 = vfmaq_f16(gx2, gx, gx); |
| 127 | gy2 = vfmaq_f16(gy2, gy, gy); |
| 128 | gxgy = vfmaq_f16(gxgy, gx, gy); |
| 129 | |
| 130 | gx = vextq_f16(low_gx, high_gx, 6); |
| 131 | gy = vextq_f16(low_gy, high_gy, 6); |
| 132 | |
| 133 | gx2 = vfmaq_f16(gx2, gx, gx); |
| 134 | gy2 = vfmaq_f16(gy2, gy, gy); |
| 135 | gxgy = vfmaq_f16(gxgy, gx, gy); |
| 136 | } |
| 137 | } |
| 138 | |
| 139 | template <size_t block_size> |
| 140 | inline void harris_score_S16_S16_FLOAT(const void *__restrict in1_ptr, const void *__restrict in2_ptr, void *__restrict out_ptr, int32_t in_stride, float norm_factor, float sensitivity, |
| 141 | float strength_thresh) |
| 142 | { |
| 143 | auto gx_ptr_0 = static_cast<const int16_t *__restrict>(in1_ptr) - (block_size / 2) * (in_stride + 1); |
| 144 | auto gy_ptr_0 = static_cast<const int16_t *__restrict>(in2_ptr) - (block_size / 2) * (in_stride + 1); |
| 145 | const int16_t *gx_ptr_1 = gx_ptr_0 + 8; |
| 146 | const int16_t *gy_ptr_1 = gy_ptr_0 + 8; |
| 147 | const auto output = static_cast<float *__restrict>(out_ptr); |
| 148 | |
| 149 | // Gx^2, Gy^2 and Gx*Gy |
| 150 | float16x8_t gx2 = vdupq_n_f16(0.0f); |
| 151 | float16x8_t gy2 = vdupq_n_f16(0.0f); |
| 152 | float16x8_t gxgy = vdupq_n_f16(0.0f); |
| 153 | |
| 154 | for(size_t i = 0; i < block_size; ++i) |
| 155 | { |
| 156 | const float16x8_t low_gx = vcvtq_f16_s16(vld1q_s16(gx_ptr_0)); |
| 157 | const float16x8_t high_gx = vcvtq_f16_s16(vld1q_s16(gx_ptr_1)); |
| 158 | const float16x8_t low_gy = vcvtq_f16_s16(vld1q_s16(gy_ptr_0)); |
| 159 | const float16x8_t high_gy = vcvtq_f16_s16(vld1q_s16(gy_ptr_1)); |
| 160 | harris_score1xN_FLOAT_FLOAT_FLOAT<block_size>(low_gx, low_gy, high_gx, high_gy, gx2, gy2, gxgy, norm_factor); |
| 161 | |
| 162 | // Update gx and gy pointer |
| 163 | gx_ptr_0 += in_stride; |
| 164 | gy_ptr_0 += in_stride; |
| 165 | gx_ptr_1 += in_stride; |
| 166 | gy_ptr_1 += in_stride; |
| 167 | } |
| 168 | |
| 169 | // Calculate harris score |
| 170 | const float16x8_t mc = harris_score(gx2, gy2, gxgy, sensitivity, strength_thresh); |
| 171 | |
| 172 | // Store score |
| 173 | vst1q_f32(output + 0, vcvt_f32_f16(vget_low_f16(mc))); |
| 174 | vst1q_f32(output + 4, vcvt_f32_f16(vget_high_f16(mc))); |
| 175 | } |
| 176 | |
| 177 | template <size_t block_size> |
| 178 | inline void harris_score_S32_S32_FLOAT(const void *__restrict in1_ptr, const void *__restrict in2_ptr, void *__restrict out_ptr, int32_t in_stride, float norm_factor, float sensitivity, |
| 179 | float strength_thresh) |
| 180 | { |
| 181 | static const float16x8_t zero = vdupq_n_f16(0.0f); |
| 182 | |
| 183 | auto gx_ptr_0 = static_cast<const int32_t *__restrict>(in1_ptr) - (block_size / 2) * (in_stride + 1); |
| 184 | auto gy_ptr_0 = static_cast<const int32_t *__restrict>(in2_ptr) - (block_size / 2) * (in_stride + 1); |
| 185 | const int32_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 186 | const int32_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 187 | const int32_t *gx_ptr_2 = gx_ptr_0 + 8; |
| 188 | const int32_t *gy_ptr_2 = gy_ptr_0 + 8; |
| 189 | const auto output = static_cast<float *__restrict>(out_ptr); |
| 190 | |
| 191 | // Gx^2, Gy^2 and Gx*Gy |
| 192 | float16x8_t gx2 = zero; |
| 193 | float16x8_t gy2 = zero; |
| 194 | float16x8_t gxgy = zero; |
| 195 | |
| 196 | for(size_t i = 0; i < block_size; ++i) |
| 197 | { |
| 198 | const float16x8_t low_gx = vcombine_f16(vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gx_ptr_0))), |
| 199 | vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gx_ptr_1)))); |
| 200 | const float16x8_t high_gx = vcombine_f16(vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gx_ptr_2))), |
| 201 | vget_low_f16(zero)); |
| 202 | const float16x8_t low_gy = vcombine_f16(vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gy_ptr_0))), |
| 203 | vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gy_ptr_1)))); |
| 204 | const float16x8_t high_gy = vcombine_f16(vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gy_ptr_2))), |
| 205 | vget_low_f16(zero)); |
| 206 | harris_score1xN_FLOAT_FLOAT_FLOAT<block_size>(low_gx, low_gy, high_gx, high_gy, gx2, gy2, gxgy, norm_factor); |
| 207 | |
| 208 | // Update gx and gy pointer |
| 209 | gx_ptr_0 += in_stride; |
| 210 | gy_ptr_0 += in_stride; |
| 211 | gx_ptr_1 += in_stride; |
| 212 | gy_ptr_1 += in_stride; |
| 213 | gx_ptr_2 += in_stride; |
| 214 | gy_ptr_2 += in_stride; |
| 215 | } |
| 216 | |
| 217 | // Calculate harris score |
| 218 | const float16x8_t mc = harris_score(gx2, gy2, gxgy, sensitivity, strength_thresh); |
| 219 | |
| 220 | // Store score |
| 221 | vst1q_f32(output + 0, vcvt_f32_f16(vget_low_f16(mc))); |
| 222 | vst1q_f32(output + 4, vcvt_f32_f16(vget_high_f16(mc))); |
| 223 | } |
| 224 | |
| 225 | template <> |
| 226 | inline void harris_score_S32_S32_FLOAT<7>(const void *__restrict in1_ptr, const void *__restrict in2_ptr, void *__restrict out_ptr, int32_t in_stride, float norm_factor, float sensitivity, |
| 227 | float strength_thresh) |
| 228 | { |
| 229 | static const float16x8_t zero = vdupq_n_f16(0.0f); |
| 230 | |
| 231 | auto gx_ptr_0 = static_cast<const int32_t *__restrict>(in1_ptr) - 3 * (in_stride + 1); |
| 232 | auto gy_ptr_0 = static_cast<const int32_t *__restrict>(in2_ptr) - 3 * (in_stride + 1); |
| 233 | const int32_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 234 | const int32_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 235 | const int32_t *gx_ptr_2 = gx_ptr_0 + 8; |
| 236 | const int32_t *gy_ptr_2 = gy_ptr_0 + 8; |
| 237 | const int32_t *gx_ptr_3 = gx_ptr_0 + 12; |
| 238 | const int32_t *gy_ptr_3 = gy_ptr_0 + 12; |
| 239 | const auto output = static_cast<float *__restrict>(out_ptr); |
| 240 | |
| 241 | // Gx^2, Gy^2 and Gx*Gy |
| 242 | float16x8_t gx2 = zero; |
| 243 | float16x8_t gy2 = zero; |
| 244 | float16x8_t gxgy = zero; |
| 245 | |
| 246 | for(size_t i = 0; i < 7; ++i) |
| 247 | { |
| 248 | const float16x8_t low_gx = vcombine_f16(vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gx_ptr_0))), |
| 249 | vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gx_ptr_1)))); |
| 250 | const float16x8_t high_gx = vcombine_f16(vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gx_ptr_2))), |
| 251 | vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gx_ptr_3)))); |
| 252 | const float16x8_t low_gy = vcombine_f16(vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gy_ptr_0))), |
| 253 | vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gy_ptr_1)))); |
| 254 | const float16x8_t high_gy = vcombine_f16(vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gy_ptr_2))), |
| 255 | vcvt_f16_f32(vcvtq_f32_s32(vld1q_s32(gy_ptr_3)))); |
| 256 | harris_score1xN_FLOAT_FLOAT_FLOAT<7>(low_gx, low_gy, high_gx, high_gy, gx2, gy2, gxgy, norm_factor); |
| 257 | |
| 258 | // Update gx and gy pointer |
| 259 | gx_ptr_0 += in_stride; |
| 260 | gy_ptr_0 += in_stride; |
| 261 | gx_ptr_1 += in_stride; |
| 262 | gy_ptr_1 += in_stride; |
| 263 | gx_ptr_2 += in_stride; |
| 264 | gy_ptr_2 += in_stride; |
| 265 | } |
| 266 | |
| 267 | // Calculate harris score |
| 268 | const float16x8_t mc = harris_score(gx2, gy2, gxgy, sensitivity, strength_thresh); |
| 269 | |
| 270 | // Store score |
| 271 | vst1q_f32(output + 0, vcvt_f32_f16(vget_low_f16(mc))); |
| 272 | vst1q_f32(output + 4, vcvt_f32_f16(vget_high_f16(mc))); |
| 273 | } |
| 274 | |
| 275 | } // namespace fp16 |
| 276 | |
| 277 | template <int32_t block_size> |
| 278 | BorderSize NEHarrisScoreFP16Kernel<block_size>::border_size() const |
| 279 | { |
| 280 | return _border_size; |
| 281 | } |
| 282 | |
| 283 | template <int32_t block_size> |
| 284 | NEHarrisScoreFP16Kernel<block_size>::NEHarrisScoreFP16Kernel() |
| 285 | : INEHarrisScoreKernel(), _func(nullptr) |
| 286 | { |
| 287 | } |
| 288 | |
| 289 | template <int32_t block_size> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 290 | void NEHarrisScoreFP16Kernel<block_size>::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 291 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 292 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 293 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 294 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 295 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 296 | |
| 297 | Iterator input1(_input1, window); |
| 298 | Iterator input2(_input2, window); |
| 299 | Iterator output(_output, window); |
| 300 | |
| 301 | const size_t input_stride = _input1->info()->strides_in_bytes()[1] / element_size_from_data_type(_input1->info()->data_type()); |
| 302 | |
| 303 | execute_window_loop(window, [&](const Coordinates & id) |
| 304 | { |
| 305 | (*_func)(input1.ptr(), input2.ptr(), output.ptr(), input_stride, _norm_factor, _sensitivity, _strength_thresh); |
| 306 | }, |
| 307 | input1, input2, output); |
| 308 | } |
| 309 | |
| 310 | template <int32_t block_size> |
| 311 | void NEHarrisScoreFP16Kernel<block_size>::configure(const IImage *input1, const IImage *input2, IImage *output, float norm_factor, float strength_thresh, float sensitivity, |
| 312 | bool border_undefined) |
| 313 | { |
| 314 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input1); |
| 315 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input2); |
| 316 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(output); |
| 317 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::S16, DataType::S32); |
| 318 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::S16, DataType::S32); |
| 319 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); |
| 320 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2); |
| 321 | ARM_COMPUTE_ERROR_ON(0.0f == norm_factor); |
| 322 | |
| 323 | _input1 = input1; |
| 324 | _input2 = input2; |
| 325 | _output = output; |
| 326 | _sensitivity = sensitivity; |
| 327 | _strength_thresh = strength_thresh; |
| 328 | _norm_factor = norm_factor; |
| 329 | _border_size = BorderSize(block_size / 2); |
| 330 | |
| 331 | if(input1->info()->data_type() == DataType::S16) |
| 332 | { |
| 333 | _func = &fp16::harris_score_S16_S16_FLOAT<block_size>; |
| 334 | } |
| 335 | else |
| 336 | { |
| 337 | _func = &fp16::harris_score_S32_S32_FLOAT<block_size>; |
| 338 | } |
| 339 | |
| 340 | ARM_COMPUTE_ERROR_ON(nullptr == _func); |
| 341 | |
| 342 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 343 | constexpr unsigned int num_elems_read_per_iteration = 16; |
| 344 | constexpr unsigned int num_elems_written_per_iteration = 8; |
| 345 | constexpr unsigned int num_rows_read_per_iteration = block_size; |
| 346 | |
| 347 | // Configure kernel window |
| 348 | Window win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 349 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
| 350 | |
| 351 | update_window_and_padding(win, |
| 352 | AccessWindowRectangle(input1->info(), -_border_size.left, -_border_size.top, num_elems_read_per_iteration, num_rows_read_per_iteration), |
| 353 | AccessWindowRectangle(input2->info(), -_border_size.left, -_border_size.top, num_elems_read_per_iteration, num_rows_read_per_iteration), |
| 354 | output_access); |
| 355 | |
| 356 | ValidRegion valid_region = intersect_valid_regions(input1->info()->valid_region(), |
| 357 | input2->info()->valid_region()); |
| 358 | |
| 359 | output_access.set_valid_region(win, valid_region, border_undefined, border_size()); |
| 360 | |
| 361 | INEKernel::configure(win); |
| 362 | } |
| 363 | |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 364 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 365 | |
| 366 | template class arm_compute::NEHarrisScoreKernel<3>; |
| 367 | template class arm_compute::NEHarrisScoreKernel<5>; |
| 368 | template class arm_compute::NEHarrisScoreKernel<7>; |
| 369 | template arm_compute::NEHarrisScoreKernel<3>::NEHarrisScoreKernel(); |
| 370 | template arm_compute::NEHarrisScoreKernel<5>::NEHarrisScoreKernel(); |
| 371 | template arm_compute::NEHarrisScoreKernel<7>::NEHarrisScoreKernel(); |
| 372 | |
| 373 | namespace |
| 374 | { |
| 375 | inline float32x4_t harris_score(float32x4_t gx2, float32x4_t gy2, float32x4_t gxgy, float32x4_t sensitivity, float32x4_t strength_thresh) |
| 376 | { |
| 377 | // Trace^2 |
| 378 | float32x4_t trace2 = vaddq_f32(gx2, gy2); |
| 379 | trace2 = vmulq_f32(trace2, trace2); |
| 380 | |
| 381 | // Det(A) |
| 382 | float32x4_t det = vmulq_f32(gx2, gy2); |
| 383 | det = vmlsq_f32(det, gxgy, gxgy); |
| 384 | |
| 385 | // Det(A) - sensitivity * trace^2 |
| 386 | const float32x4_t mc = vmlsq_f32(det, sensitivity, trace2); |
| 387 | |
| 388 | // mc > strength_thresh |
| 389 | const uint32x4_t mask = vcgtq_f32(mc, strength_thresh); |
| 390 | |
| 391 | return vbslq_f32(mask, mc, vdupq_n_f32(0.0f)); |
| 392 | } |
| 393 | |
| 394 | inline void harris_score1x3_FLOAT_FLOAT_FLOAT(float32x4_t low_gx, float32x4_t low_gy, float32x4_t high_gx, float32x4_t high_gy, float32x4_t &gx2, float32x4_t &gy2, float32x4_t &gxgy, |
| 395 | float32x4_t norm_factor) |
| 396 | { |
| 397 | // Normalize |
| 398 | low_gx = vmulq_f32(low_gx, norm_factor); |
| 399 | low_gy = vmulq_f32(low_gy, norm_factor); |
| 400 | high_gx = vmulq_f32(high_gx, norm_factor); |
| 401 | high_gy = vmulq_f32(high_gy, norm_factor); |
| 402 | |
| 403 | const float32x4_t l_gx = low_gx; |
| 404 | const float32x4_t l_gy = low_gy; |
| 405 | const float32x4_t m_gx = vextq_f32(low_gx, high_gx, 1); |
| 406 | const float32x4_t m_gy = vextq_f32(low_gy, high_gy, 1); |
| 407 | const float32x4_t r_gx = vextq_f32(low_gx, high_gx, 2); |
| 408 | const float32x4_t r_gy = vextq_f32(low_gy, high_gy, 2); |
| 409 | |
| 410 | // Gx*Gx |
| 411 | gx2 = vmlaq_f32(gx2, l_gx, l_gx); |
| 412 | gx2 = vmlaq_f32(gx2, m_gx, m_gx); |
| 413 | gx2 = vmlaq_f32(gx2, r_gx, r_gx); |
| 414 | |
| 415 | // Gy*Gy |
| 416 | gy2 = vmlaq_f32(gy2, l_gy, l_gy); |
| 417 | gy2 = vmlaq_f32(gy2, m_gy, m_gy); |
| 418 | gy2 = vmlaq_f32(gy2, r_gy, r_gy); |
| 419 | |
| 420 | // Gx*Gy |
| 421 | gxgy = vmlaq_f32(gxgy, l_gx, l_gy); |
| 422 | gxgy = vmlaq_f32(gxgy, m_gx, m_gy); |
| 423 | gxgy = vmlaq_f32(gxgy, r_gx, r_gy); |
| 424 | } |
| 425 | |
| 426 | inline void harris_score1x5_FLOAT_FLOAT_FLOAT(float32x4_t low_gx, float32x4_t low_gy, float32x4_t high_gx, float32x4_t high_gy, float32x4_t &gx2, float32x4_t &gy2, float32x4_t &gxgy, |
| 427 | float32x4_t norm_factor) |
| 428 | { |
| 429 | // Normalize |
| 430 | low_gx = vmulq_f32(low_gx, norm_factor); |
| 431 | low_gy = vmulq_f32(low_gy, norm_factor); |
| 432 | high_gx = vmulq_f32(high_gx, norm_factor); |
| 433 | high_gy = vmulq_f32(high_gy, norm_factor); |
| 434 | |
| 435 | // L2 values |
| 436 | float32x4_t gx = low_gx; |
| 437 | float32x4_t gy = low_gy; |
| 438 | |
| 439 | // Accumulate |
| 440 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 441 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 442 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 443 | |
| 444 | // L1 values |
| 445 | gx = vextq_f32(low_gx, high_gx, 1); |
| 446 | gy = vextq_f32(low_gy, high_gy, 1); |
| 447 | |
| 448 | // Accumulate |
| 449 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 450 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 451 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 452 | |
| 453 | // M values |
| 454 | gx = vextq_f32(low_gx, high_gx, 2); |
| 455 | gy = vextq_f32(low_gy, high_gy, 2); |
| 456 | |
| 457 | // Accumulate |
| 458 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 459 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 460 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 461 | |
| 462 | // R1 values |
| 463 | gx = vextq_f32(low_gx, high_gx, 3); |
| 464 | gy = vextq_f32(low_gy, high_gy, 3); |
| 465 | |
| 466 | // Accumulate |
| 467 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 468 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 469 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 470 | |
| 471 | // R2 values |
| 472 | gx = high_gx; |
| 473 | gy = high_gy; |
| 474 | |
| 475 | // Accumulate |
| 476 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 477 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 478 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 479 | } |
| 480 | |
| 481 | inline void harris_score1x7_FLOAT_FLOAT_FLOAT(float32x4_t low_gx, float32x4_t low_gy, float32x4_t high_gx, float32x4_t high_gy, float32x4_t high_gx1, float32x4_t high_gy1, float32x4_t &gx2, |
| 482 | float32x4_t &gy2, float32x4_t &gxgy, float32x4_t norm_factor) |
| 483 | { |
| 484 | // Normalize |
| 485 | low_gx = vmulq_f32(low_gx, norm_factor); |
| 486 | low_gy = vmulq_f32(low_gy, norm_factor); |
| 487 | high_gx = vmulq_f32(high_gx, norm_factor); |
| 488 | high_gy = vmulq_f32(high_gy, norm_factor); |
| 489 | |
| 490 | // L3 values |
| 491 | float32x4_t gx = low_gx; |
| 492 | float32x4_t gy = low_gy; |
| 493 | |
| 494 | // Accumulate |
| 495 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 496 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 497 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 498 | |
| 499 | // L2 values |
| 500 | gx = vextq_f32(low_gx, high_gx, 1); |
| 501 | gy = vextq_f32(low_gy, high_gy, 1); |
| 502 | |
| 503 | // Accumulate |
| 504 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 505 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 506 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 507 | |
| 508 | // L1 values |
| 509 | gx = vextq_f32(low_gx, high_gx, 2); |
| 510 | gy = vextq_f32(low_gy, high_gy, 2); |
| 511 | |
| 512 | // Accumulate |
| 513 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 514 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 515 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 516 | |
| 517 | // M values |
| 518 | gx = vextq_f32(low_gx, high_gx, 3); |
| 519 | gy = vextq_f32(low_gy, high_gy, 3); |
| 520 | |
| 521 | // Accumulate |
| 522 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 523 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 524 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 525 | |
| 526 | // R1 values |
| 527 | gx = high_gx; |
| 528 | gy = high_gy; |
| 529 | |
| 530 | // Accumulate |
| 531 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 532 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 533 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 534 | |
| 535 | // Change tmp_low and tmp_high for calculating R2 and R3 values |
| 536 | low_gx = high_gx; |
| 537 | low_gy = high_gy; |
| 538 | high_gx = high_gx1; |
| 539 | high_gy = high_gy1; |
| 540 | |
| 541 | // Normalize |
| 542 | high_gx = vmulq_f32(high_gx, norm_factor); |
| 543 | high_gy = vmulq_f32(high_gy, norm_factor); |
| 544 | |
| 545 | // R2 values |
| 546 | gx = vextq_f32(low_gx, high_gx, 1); |
| 547 | gy = vextq_f32(low_gy, high_gy, 1); |
| 548 | |
| 549 | // Accumulate |
| 550 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 551 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 552 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 553 | |
| 554 | // R3 values |
| 555 | gx = vextq_f32(low_gx, high_gx, 2); |
| 556 | gy = vextq_f32(low_gy, high_gy, 2); |
| 557 | |
| 558 | // Accumulate |
| 559 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 560 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 561 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 562 | } |
| 563 | |
| 564 | inline void harris_score3x3_S16_S16_FLOAT(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, int32_t input_stride, |
| 565 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 566 | |
| 567 | { |
| 568 | const auto gx_ptr_0 = static_cast<const int16_t *__restrict>(input1_ptr) - 1; |
| 569 | const auto gy_ptr_0 = static_cast<const int16_t *__restrict>(input2_ptr) - 1; |
| 570 | const int16_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 571 | const int16_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 572 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 573 | |
| 574 | // Gx^2, Gy^2 and Gx*Gy |
| 575 | float32x4x2_t gx2 = |
| 576 | { |
| 577 | { |
| 578 | vdupq_n_f32(0.0f), |
| 579 | vdupq_n_f32(0.0f) |
| 580 | } |
| 581 | }; |
| 582 | float32x4x2_t gy2 = |
| 583 | { |
| 584 | { |
| 585 | vdupq_n_f32(0.0f), |
| 586 | vdupq_n_f32(0.0f) |
| 587 | } |
| 588 | }; |
| 589 | float32x4x2_t gxgy = |
| 590 | { |
| 591 | { |
| 592 | vdupq_n_f32(0.0f), |
| 593 | vdupq_n_f32(0.0f) |
| 594 | } |
| 595 | }; |
| 596 | |
| 597 | // Row0 |
| 598 | int16x8x2_t tmp_gx = |
| 599 | { |
| 600 | { |
| 601 | vld1q_s16(gx_ptr_0 - input_stride), |
| 602 | vld1q_s16(gx_ptr_1 - input_stride) |
| 603 | } |
| 604 | }; |
| 605 | int16x8x2_t tmp_gy = |
| 606 | { |
| 607 | { |
| 608 | vld1q_s16(gy_ptr_0 - input_stride), |
| 609 | vld1q_s16(gy_ptr_1 - input_stride) |
| 610 | } |
| 611 | }; |
| 612 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 613 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 614 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 615 | |
| 616 | float32x4_t low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[0]))); |
| 617 | float32x4_t low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[0]))); |
| 618 | float32x4_t high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[0]))); |
| 619 | float32x4_t high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[0]))); |
| 620 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 621 | |
| 622 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[1]))); |
| 623 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[1]))); |
| 624 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[1]))); |
| 625 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[1]))); |
| 626 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 627 | |
| 628 | // Row1 |
| 629 | tmp_gx.val[0] = vld1q_s16(gx_ptr_0); |
| 630 | tmp_gy.val[0] = vld1q_s16(gy_ptr_0); |
| 631 | tmp_gx.val[1] = vld1q_s16(gx_ptr_1); |
| 632 | tmp_gy.val[1] = vld1q_s16(gy_ptr_1); |
| 633 | |
| 634 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[0]))); |
| 635 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[0]))); |
| 636 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[0]))); |
| 637 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[0]))); |
| 638 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 639 | |
| 640 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[1]))); |
| 641 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[1]))); |
| 642 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[1]))); |
| 643 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[1]))); |
| 644 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 645 | |
| 646 | // Row2 |
| 647 | tmp_gx.val[0] = vld1q_s16(gx_ptr_0 + input_stride); |
| 648 | tmp_gy.val[0] = vld1q_s16(gy_ptr_0 + input_stride); |
| 649 | tmp_gx.val[1] = vld1q_s16(gx_ptr_1 + input_stride); |
| 650 | tmp_gy.val[1] = vld1q_s16(gy_ptr_1 + input_stride); |
| 651 | |
| 652 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[0]))); |
| 653 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[0]))); |
| 654 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[0]))); |
| 655 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[0]))); |
| 656 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 657 | |
| 658 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[1]))); |
| 659 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[1]))); |
| 660 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[1]))); |
| 661 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[1]))); |
| 662 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 663 | |
| 664 | // Calculate harris score |
| 665 | const float32x4x2_t mc = |
| 666 | { |
| 667 | { |
| 668 | harris_score(gx2.val[0], gy2.val[0], gxgy.val[0], sensitivity, strength_thresh), |
| 669 | harris_score(gx2.val[1], gy2.val[1], gxgy.val[1], sensitivity, strength_thresh) |
| 670 | } |
| 671 | }; |
| 672 | |
| 673 | // Store score |
| 674 | vst1q_f32(output + 0, mc.val[0]); |
| 675 | vst1q_f32(output + 4, mc.val[1]); |
| 676 | } |
| 677 | |
| 678 | inline void harris_score3x3_S32_S32_FLOAT(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, int32_t input_stride, |
| 679 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 680 | { |
| 681 | auto gx_ptr_0 = static_cast<const int32_t *__restrict>(input1_ptr) - 1; |
| 682 | auto gy_ptr_0 = static_cast<const int32_t *__restrict>(input2_ptr) - 1; |
| 683 | const int32_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 684 | const int32_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 685 | const int32_t *gx_ptr_2 = gx_ptr_0 + 8; |
| 686 | const int32_t *gy_ptr_2 = gy_ptr_0 + 8; |
| 687 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 688 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 689 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 690 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 691 | |
| 692 | // Gx^2, Gy^2 and Gx*Gy |
| 693 | float32x4x2_t gx2 = |
| 694 | { |
| 695 | { |
| 696 | vdupq_n_f32(0.0f), |
| 697 | vdupq_n_f32(0.0f) |
| 698 | } |
| 699 | }; |
| 700 | float32x4x2_t gy2 = |
| 701 | { |
| 702 | { |
| 703 | vdupq_n_f32(0.0f), |
| 704 | vdupq_n_f32(0.0f) |
| 705 | } |
| 706 | }; |
| 707 | float32x4x2_t gxgy = |
| 708 | { |
| 709 | { |
| 710 | vdupq_n_f32(0.0f), |
| 711 | vdupq_n_f32(0.0f) |
| 712 | } |
| 713 | }; |
| 714 | |
| 715 | // Row0 |
| 716 | float32x4_t low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_0 - input_stride)); |
| 717 | float32x4_t low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_0 - input_stride)); |
| 718 | float32x4_t high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1 - input_stride)); |
| 719 | float32x4_t high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1 - input_stride)); |
| 720 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 721 | |
| 722 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1 - input_stride)); |
| 723 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1 - input_stride)); |
| 724 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_2 - input_stride)); |
| 725 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_2 - input_stride)); |
| 726 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 727 | |
| 728 | // Row1 |
| 729 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_0)); |
| 730 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_0)); |
| 731 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 732 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 733 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 734 | |
| 735 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 736 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 737 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_2)); |
| 738 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_2)); |
| 739 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 740 | |
| 741 | // Row2 |
| 742 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_0 + input_stride)); |
| 743 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_0 + input_stride)); |
| 744 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1 + input_stride)); |
| 745 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1 + input_stride)); |
| 746 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 747 | |
| 748 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1 + input_stride)); |
| 749 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1 + input_stride)); |
| 750 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_2 + input_stride)); |
| 751 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_2 + input_stride)); |
| 752 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 753 | |
| 754 | // Calculate harris score |
| 755 | const float32x4x2_t mc = |
| 756 | { |
| 757 | { |
| 758 | harris_score(gx2.val[0], gy2.val[0], gxgy.val[0], sensitivity, strength_thresh), |
| 759 | harris_score(gx2.val[1], gy2.val[1], gxgy.val[1], sensitivity, strength_thresh) |
| 760 | } |
| 761 | }; |
| 762 | |
| 763 | // Store score |
| 764 | vst1q_f32(output + 0, mc.val[0]); |
| 765 | vst1q_f32(output + 4, mc.val[1]); |
| 766 | } |
| 767 | |
| 768 | inline void harris_score5x5_S16_S16_FLOAT(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, int32_t input_stride, |
| 769 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 770 | { |
| 771 | auto gx_ptr_0 = static_cast<const int16_t *__restrict>(input1_ptr) - 2 - 2 * input_stride; |
| 772 | auto gy_ptr_0 = static_cast<const int16_t *__restrict>(input2_ptr) - 2 - 2 * input_stride; |
| 773 | const int16_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 774 | const int16_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 775 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 776 | |
| 777 | // Gx^2, Gy^2 and Gx*Gy |
| 778 | float32x4x2_t gx2 = |
| 779 | { |
| 780 | { |
| 781 | vdupq_n_f32(0.0f), |
| 782 | vdupq_n_f32(0.0f) |
| 783 | } |
| 784 | }; |
| 785 | float32x4x2_t gy2 = |
| 786 | { |
| 787 | { |
| 788 | vdupq_n_f32(0.0f), |
| 789 | vdupq_n_f32(0.0f) |
| 790 | } |
| 791 | }; |
| 792 | float32x4x2_t gxgy = |
| 793 | { |
| 794 | { |
| 795 | vdupq_n_f32(0.0f), |
| 796 | vdupq_n_f32(0.0f) |
| 797 | } |
| 798 | }; |
| 799 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 800 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 801 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 802 | |
| 803 | for(int i = 0; i < 5; ++i) |
| 804 | { |
| 805 | const int16x8x2_t tmp_gx = |
| 806 | { |
| 807 | { |
| 808 | vld1q_s16(gx_ptr_0), |
| 809 | vld1q_s16(gx_ptr_1) |
| 810 | } |
| 811 | }; |
| 812 | const int16x8x2_t tmp_gy = |
| 813 | { |
| 814 | { |
| 815 | vld1q_s16(gy_ptr_0), |
| 816 | vld1q_s16(gy_ptr_1) |
| 817 | } |
| 818 | }; |
| 819 | |
| 820 | float32x4_t low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[0]))); |
| 821 | float32x4_t low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[0]))); |
| 822 | float32x4_t high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[0]))); |
| 823 | float32x4_t high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[0]))); |
| 824 | harris_score1x5_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 825 | |
| 826 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[1]))); |
| 827 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[1]))); |
| 828 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[1]))); |
| 829 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[1]))); |
| 830 | harris_score1x5_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 831 | |
| 832 | // Update gx and gy pointer |
| 833 | gx_ptr_0 += input_stride; |
| 834 | gy_ptr_0 += input_stride; |
| 835 | gx_ptr_1 += input_stride; |
| 836 | gy_ptr_1 += input_stride; |
| 837 | } |
| 838 | |
| 839 | // Calculate harris score |
| 840 | const float32x4x2_t mc = |
| 841 | { |
| 842 | { |
| 843 | harris_score(gx2.val[0], gy2.val[0], gxgy.val[0], sensitivity, strength_thresh), |
| 844 | harris_score(gx2.val[1], gy2.val[1], gxgy.val[1], sensitivity, strength_thresh) |
| 845 | } |
| 846 | }; |
| 847 | |
| 848 | // Store score |
| 849 | vst1q_f32(output + 0, mc.val[0]); |
| 850 | vst1q_f32(output + 4, mc.val[1]); |
| 851 | } |
| 852 | |
| 853 | inline void harris_score5x5_S32_S32_FLOAT(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, int32_t input_stride, |
| 854 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 855 | |
| 856 | { |
| 857 | auto gx_ptr_0 = static_cast<const int32_t *__restrict>(input1_ptr) - 2 - 2 * input_stride; |
| 858 | auto gy_ptr_0 = static_cast<const int32_t *__restrict>(input2_ptr) - 2 - 2 * input_stride; |
| 859 | const int32_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 860 | const int32_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 861 | const int32_t *gx_ptr_2 = gx_ptr_0 + 8; |
| 862 | const int32_t *gy_ptr_2 = gy_ptr_0 + 8; |
| 863 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 864 | |
| 865 | // Gx^2, Gy^2 and Gx*Gy |
| 866 | float32x4x2_t gx2 = |
| 867 | { |
| 868 | { |
| 869 | vdupq_n_f32(0.0f), |
| 870 | vdupq_n_f32(0.0f) |
| 871 | } |
| 872 | }; |
| 873 | float32x4x2_t gy2 = |
| 874 | { |
| 875 | { |
| 876 | vdupq_n_f32(0.0f), |
| 877 | vdupq_n_f32(0.0f) |
| 878 | } |
| 879 | }; |
| 880 | float32x4x2_t gxgy = |
| 881 | { |
| 882 | { |
| 883 | vdupq_n_f32(0.0f), |
| 884 | vdupq_n_f32(0.0f) |
| 885 | } |
| 886 | }; |
| 887 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 888 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 889 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 890 | |
| 891 | for(int i = 0; i < 5; ++i) |
| 892 | { |
| 893 | const float32x4_t low_gx_0 = vcvtq_f32_s32(vld1q_s32(gx_ptr_0)); |
| 894 | const float32x4_t low_gy_0 = vcvtq_f32_s32(vld1q_s32(gy_ptr_0)); |
| 895 | const float32x4_t high_gx_0 = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 896 | const float32x4_t high_gy_0 = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 897 | harris_score1x5_FLOAT_FLOAT_FLOAT(low_gx_0, low_gy_0, high_gx_0, high_gy_0, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 898 | |
| 899 | const float32x4_t low_gx_1 = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 900 | const float32x4_t low_gy_1 = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 901 | const float32x4_t high_gx_1 = vcvtq_f32_s32(vld1q_s32(gx_ptr_2)); |
| 902 | const float32x4_t high_gy_1 = vcvtq_f32_s32(vld1q_s32(gy_ptr_2)); |
| 903 | harris_score1x5_FLOAT_FLOAT_FLOAT(low_gx_1, low_gy_1, high_gx_1, high_gy_1, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 904 | |
| 905 | // Update gx and gy pointer |
| 906 | gx_ptr_0 += input_stride; |
| 907 | gy_ptr_0 += input_stride; |
| 908 | gx_ptr_1 += input_stride; |
| 909 | gy_ptr_1 += input_stride; |
| 910 | gx_ptr_2 += input_stride; |
| 911 | gy_ptr_2 += input_stride; |
| 912 | } |
| 913 | |
| 914 | // Calculate harris score |
| 915 | const float32x4x2_t mc = |
| 916 | { |
| 917 | { |
| 918 | harris_score(gx2.val[0], gy2.val[0], gxgy.val[0], sensitivity, strength_thresh), |
| 919 | harris_score(gx2.val[1], gy2.val[1], gxgy.val[1], sensitivity, strength_thresh) |
| 920 | } |
| 921 | }; |
| 922 | |
| 923 | // Store score |
| 924 | vst1q_f32(output + 0, mc.val[0]); |
| 925 | vst1q_f32(output + 4, mc.val[1]); |
| 926 | } |
| 927 | |
| 928 | inline void harris_score7x7_S16_S16_FLOAT(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, int32_t input_stride, |
| 929 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 930 | { |
| 931 | auto gx_ptr_0 = static_cast<const int16_t *__restrict>(input1_ptr) - 3 - 3 * input_stride; |
| 932 | auto gy_ptr_0 = static_cast<const int16_t *__restrict>(input2_ptr) - 3 - 3 * input_stride; |
| 933 | const int16_t *gx_ptr_1 = gx_ptr_0 + 8; |
| 934 | const int16_t *gy_ptr_1 = gy_ptr_0 + 8; |
| 935 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 936 | |
| 937 | // Gx^2, Gy^2 and Gx*Gy |
| 938 | float32x4_t gx2 = vdupq_n_f32(0.0f); |
| 939 | float32x4_t gy2 = vdupq_n_f32(0.0f); |
| 940 | float32x4_t gxgy = vdupq_n_f32(0.0f); |
| 941 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 942 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 943 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 944 | |
| 945 | for(int i = 0; i < 7; ++i) |
| 946 | { |
| 947 | const int16x8_t tmp0_gx = vld1q_s16(gx_ptr_0); |
| 948 | const int16x8_t tmp0_gy = vld1q_s16(gy_ptr_0); |
| 949 | const int16x4_t tmp1_gx = vld1_s16(gx_ptr_1); |
| 950 | const int16x4_t tmp1_gy = vld1_s16(gy_ptr_1); |
| 951 | |
| 952 | float32x4_t low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp0_gx))); |
| 953 | float32x4_t low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp0_gy))); |
| 954 | float32x4_t high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp0_gx))); |
| 955 | float32x4_t high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp0_gy))); |
| 956 | float32x4_t high_gx1 = vcvtq_f32_s32(vmovl_s16(tmp1_gx)); |
| 957 | float32x4_t high_gy1 = vcvtq_f32_s32(vmovl_s16(tmp1_gy)); |
| 958 | harris_score1x7_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, high_gx1, high_gy1, gx2, gy2, gxgy, norm_factor); |
| 959 | |
| 960 | // Update gx and gy pointer |
| 961 | gx_ptr_0 += input_stride; |
| 962 | gy_ptr_0 += input_stride; |
| 963 | gx_ptr_1 += input_stride; |
| 964 | gy_ptr_1 += input_stride; |
| 965 | } |
| 966 | |
| 967 | // Calculate harris score |
| 968 | const float32x4_t mc = harris_score(gx2, gy2, gxgy, sensitivity, strength_thresh); |
| 969 | |
| 970 | // Store score |
| 971 | vst1q_f32(output, mc); |
| 972 | } |
| 973 | |
| 974 | inline void harris_score7x7_S32_S32_FLOAT(const void *__restrict input1_ptr, const void *__restrict input2_ptr, void *__restrict output_ptr, int32_t input_stride, |
| 975 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 976 | { |
| 977 | auto gx_ptr_0 = static_cast<const int32_t *__restrict>(input1_ptr) - 3 - 3 * input_stride; |
| 978 | auto gy_ptr_0 = static_cast<const int32_t *__restrict>(input2_ptr) - 3 - 3 * input_stride; |
| 979 | const int32_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 980 | const int32_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 981 | const int32_t *gx_ptr_2 = gx_ptr_1 + 4; |
| 982 | const int32_t *gy_ptr_2 = gy_ptr_1 + 4; |
| 983 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 984 | |
| 985 | // Gx^2, Gy^2 and Gx*Gy |
| 986 | float32x4_t gx2 = vdupq_n_f32(0.0f); |
| 987 | float32x4_t gy2 = vdupq_n_f32(0.0f); |
| 988 | float32x4_t gxgy = vdupq_n_f32(0.0f); |
| 989 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 990 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 991 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 992 | |
| 993 | for(int i = 0; i < 7; ++i) |
| 994 | { |
| 995 | const float32x4_t low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_0)); |
| 996 | const float32x4_t low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_0)); |
| 997 | const float32x4_t high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 998 | const float32x4_t high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 999 | const float32x4_t high_gx1 = vcvtq_f32_s32(vld1q_s32(gx_ptr_2)); |
| 1000 | const float32x4_t high_gy1 = vcvtq_f32_s32(vld1q_s32(gy_ptr_2)); |
| 1001 | harris_score1x7_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, high_gx1, high_gy1, gx2, gy2, gxgy, norm_factor); |
| 1002 | |
| 1003 | // Update gx and gy pointer |
| 1004 | gx_ptr_0 += input_stride; |
| 1005 | gy_ptr_0 += input_stride; |
| 1006 | gx_ptr_1 += input_stride; |
| 1007 | gy_ptr_1 += input_stride; |
| 1008 | gx_ptr_2 += input_stride; |
| 1009 | gy_ptr_2 += input_stride; |
| 1010 | } |
| 1011 | |
| 1012 | // Calculate harris score |
| 1013 | const float32x4_t mc = harris_score(gx2, gy2, gxgy, sensitivity, strength_thresh); |
| 1014 | |
| 1015 | // Store score |
| 1016 | vst1q_f32(output, mc); |
| 1017 | } |
| 1018 | |
| 1019 | } // namespace |
| 1020 | |
| 1021 | INEHarrisScoreKernel::INEHarrisScoreKernel() |
| 1022 | : _input1(nullptr), _input2(nullptr), _output(nullptr), _sensitivity(0.0f), _strength_thresh(0.0f), _norm_factor(0.0f), _border_size() |
| 1023 | { |
| 1024 | } |
| 1025 | |
| 1026 | template <int32_t block_size> |
| 1027 | NEHarrisScoreKernel<block_size>::NEHarrisScoreKernel() |
| 1028 | : INEHarrisScoreKernel(), _func(nullptr) |
| 1029 | { |
| 1030 | } |
| 1031 | |
| 1032 | template <int32_t block_size> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1033 | void NEHarrisScoreKernel<block_size>::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1034 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 1035 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1036 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 1037 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 1038 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 1039 | |
| 1040 | Iterator input1(_input1, window); |
| 1041 | Iterator input2(_input2, window); |
| 1042 | Iterator output(_output, window); |
| 1043 | |
| 1044 | const size_t input_stride = _input1->info()->strides_in_bytes()[1] / element_size_from_data_type(_input1->info()->data_type()); |
| 1045 | |
| 1046 | execute_window_loop(window, [&](const Coordinates & id) |
| 1047 | { |
| 1048 | (*_func)(input1.ptr(), input2.ptr(), output.ptr(), input_stride, _norm_factor, _sensitivity, _strength_thresh); |
| 1049 | }, |
| 1050 | input1, input2, output); |
| 1051 | } |
| 1052 | |
| 1053 | template <int32_t block_size> |
| 1054 | BorderSize NEHarrisScoreKernel<block_size>::border_size() const |
| 1055 | { |
| 1056 | return _border_size; |
| 1057 | } |
| 1058 | |
| 1059 | template <int32_t block_size> |
| 1060 | void NEHarrisScoreKernel<block_size>::configure(const IImage *input1, const IImage *input2, IImage *output, float norm_factor, float strength_thresh, float sensitivity, |
| 1061 | bool border_undefined) |
| 1062 | { |
| 1063 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input1); |
| 1064 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input2); |
| 1065 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(output); |
| 1066 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::S16, DataType::S32); |
| 1067 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::S16, DataType::S32); |
| 1068 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); |
| 1069 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2); |
| 1070 | ARM_COMPUTE_ERROR_ON(0.0f == norm_factor); |
| 1071 | |
| 1072 | _input1 = input1; |
| 1073 | _input2 = input2; |
| 1074 | _output = output; |
| 1075 | _sensitivity = sensitivity; |
| 1076 | _strength_thresh = strength_thresh; |
| 1077 | _norm_factor = norm_factor; |
| 1078 | _border_size = BorderSize(block_size / 2); |
| 1079 | |
| 1080 | if(input1->info()->data_type() == DataType::S16) |
| 1081 | { |
| 1082 | switch(block_size) |
| 1083 | { |
| 1084 | case 3: |
| 1085 | _func = &harris_score3x3_S16_S16_FLOAT; |
| 1086 | break; |
| 1087 | case 5: |
| 1088 | _func = &harris_score5x5_S16_S16_FLOAT; |
| 1089 | break; |
| 1090 | case 7: |
| 1091 | _func = &harris_score7x7_S16_S16_FLOAT; |
| 1092 | break; |
| 1093 | default: |
| 1094 | ARM_COMPUTE_ERROR("Invalid block size"); |
| 1095 | break; |
| 1096 | } |
| 1097 | } |
| 1098 | else |
| 1099 | { |
| 1100 | switch(block_size) |
| 1101 | { |
| 1102 | case 3: |
| 1103 | _func = &harris_score3x3_S32_S32_FLOAT; |
| 1104 | break; |
| 1105 | case 5: |
| 1106 | _func = &harris_score5x5_S32_S32_FLOAT; |
| 1107 | break; |
| 1108 | case 7: |
| 1109 | _func = &harris_score7x7_S32_S32_FLOAT; |
| 1110 | break; |
| 1111 | default: |
| 1112 | ARM_COMPUTE_ERROR("Invalid block size"); |
| 1113 | break; |
| 1114 | } |
| 1115 | } |
| 1116 | |
| 1117 | ARM_COMPUTE_ERROR_ON(nullptr == _func); |
| 1118 | |
| 1119 | constexpr unsigned int num_elems_processed_per_iteration = block_size != 7 ? 8 : 4; |
| 1120 | constexpr unsigned int num_elems_read_per_iteration = block_size != 7 ? 16 : 12; |
| 1121 | constexpr unsigned int num_elems_written_per_iteration = block_size != 7 ? 8 : 4; |
| 1122 | constexpr unsigned int num_rows_read_per_iteration = block_size; |
| 1123 | |
| 1124 | // Configure kernel window |
| 1125 | Window win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 1126 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
| 1127 | |
| 1128 | update_window_and_padding(win, |
| 1129 | AccessWindowRectangle(input1->info(), -_border_size.left, -_border_size.top, num_elems_read_per_iteration, num_rows_read_per_iteration), |
| 1130 | AccessWindowRectangle(input2->info(), -_border_size.left, -_border_size.top, num_elems_read_per_iteration, num_rows_read_per_iteration), |
| 1131 | output_access); |
| 1132 | |
| 1133 | ValidRegion valid_region = intersect_valid_regions(input1->info()->valid_region(), |
| 1134 | input2->info()->valid_region()); |
| 1135 | |
| 1136 | output_access.set_valid_region(win, valid_region, border_undefined, border_size()); |
| 1137 | |
| 1138 | INEKernel::configure(win); |
| 1139 | } |