Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2016-2019 Arm Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [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 "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 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 42 | template class arm_compute::NEHarrisScoreKernel<3>; |
| 43 | template class arm_compute::NEHarrisScoreKernel<5>; |
| 44 | template class arm_compute::NEHarrisScoreKernel<7>; |
| 45 | template arm_compute::NEHarrisScoreKernel<3>::NEHarrisScoreKernel(); |
| 46 | template arm_compute::NEHarrisScoreKernel<5>::NEHarrisScoreKernel(); |
| 47 | template arm_compute::NEHarrisScoreKernel<7>::NEHarrisScoreKernel(); |
| 48 | |
| 49 | namespace |
| 50 | { |
| 51 | inline float32x4_t harris_score(float32x4_t gx2, float32x4_t gy2, float32x4_t gxgy, float32x4_t sensitivity, float32x4_t strength_thresh) |
| 52 | { |
| 53 | // Trace^2 |
| 54 | float32x4_t trace2 = vaddq_f32(gx2, gy2); |
| 55 | trace2 = vmulq_f32(trace2, trace2); |
| 56 | |
| 57 | // Det(A) |
| 58 | float32x4_t det = vmulq_f32(gx2, gy2); |
| 59 | det = vmlsq_f32(det, gxgy, gxgy); |
| 60 | |
| 61 | // Det(A) - sensitivity * trace^2 |
| 62 | const float32x4_t mc = vmlsq_f32(det, sensitivity, trace2); |
| 63 | |
| 64 | // mc > strength_thresh |
| 65 | const uint32x4_t mask = vcgtq_f32(mc, strength_thresh); |
| 66 | |
| 67 | return vbslq_f32(mask, mc, vdupq_n_f32(0.0f)); |
| 68 | } |
| 69 | |
| 70 | 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, |
| 71 | float32x4_t norm_factor) |
| 72 | { |
| 73 | // Normalize |
| 74 | low_gx = vmulq_f32(low_gx, norm_factor); |
| 75 | low_gy = vmulq_f32(low_gy, norm_factor); |
| 76 | high_gx = vmulq_f32(high_gx, norm_factor); |
| 77 | high_gy = vmulq_f32(high_gy, norm_factor); |
| 78 | |
| 79 | const float32x4_t l_gx = low_gx; |
| 80 | const float32x4_t l_gy = low_gy; |
| 81 | const float32x4_t m_gx = vextq_f32(low_gx, high_gx, 1); |
| 82 | const float32x4_t m_gy = vextq_f32(low_gy, high_gy, 1); |
| 83 | const float32x4_t r_gx = vextq_f32(low_gx, high_gx, 2); |
| 84 | const float32x4_t r_gy = vextq_f32(low_gy, high_gy, 2); |
| 85 | |
| 86 | // Gx*Gx |
| 87 | gx2 = vmlaq_f32(gx2, l_gx, l_gx); |
| 88 | gx2 = vmlaq_f32(gx2, m_gx, m_gx); |
| 89 | gx2 = vmlaq_f32(gx2, r_gx, r_gx); |
| 90 | |
| 91 | // Gy*Gy |
| 92 | gy2 = vmlaq_f32(gy2, l_gy, l_gy); |
| 93 | gy2 = vmlaq_f32(gy2, m_gy, m_gy); |
| 94 | gy2 = vmlaq_f32(gy2, r_gy, r_gy); |
| 95 | |
| 96 | // Gx*Gy |
| 97 | gxgy = vmlaq_f32(gxgy, l_gx, l_gy); |
| 98 | gxgy = vmlaq_f32(gxgy, m_gx, m_gy); |
| 99 | gxgy = vmlaq_f32(gxgy, r_gx, r_gy); |
| 100 | } |
| 101 | |
| 102 | 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, |
| 103 | float32x4_t norm_factor) |
| 104 | { |
| 105 | // Normalize |
| 106 | low_gx = vmulq_f32(low_gx, norm_factor); |
| 107 | low_gy = vmulq_f32(low_gy, norm_factor); |
| 108 | high_gx = vmulq_f32(high_gx, norm_factor); |
| 109 | high_gy = vmulq_f32(high_gy, norm_factor); |
| 110 | |
| 111 | // L2 values |
| 112 | float32x4_t gx = low_gx; |
| 113 | float32x4_t gy = low_gy; |
| 114 | |
| 115 | // Accumulate |
| 116 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 117 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 118 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 119 | |
| 120 | // L1 values |
| 121 | gx = vextq_f32(low_gx, high_gx, 1); |
| 122 | gy = vextq_f32(low_gy, high_gy, 1); |
| 123 | |
| 124 | // Accumulate |
| 125 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 126 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 127 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 128 | |
| 129 | // M values |
| 130 | gx = vextq_f32(low_gx, high_gx, 2); |
| 131 | gy = vextq_f32(low_gy, high_gy, 2); |
| 132 | |
| 133 | // Accumulate |
| 134 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 135 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 136 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 137 | |
| 138 | // R1 values |
| 139 | gx = vextq_f32(low_gx, high_gx, 3); |
| 140 | gy = vextq_f32(low_gy, high_gy, 3); |
| 141 | |
| 142 | // Accumulate |
| 143 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 144 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 145 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 146 | |
| 147 | // R2 values |
| 148 | gx = high_gx; |
| 149 | gy = high_gy; |
| 150 | |
| 151 | // Accumulate |
| 152 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 153 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 154 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 155 | } |
| 156 | |
| 157 | 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, |
| 158 | float32x4_t &gy2, float32x4_t &gxgy, float32x4_t norm_factor) |
| 159 | { |
| 160 | // Normalize |
| 161 | low_gx = vmulq_f32(low_gx, norm_factor); |
| 162 | low_gy = vmulq_f32(low_gy, norm_factor); |
| 163 | high_gx = vmulq_f32(high_gx, norm_factor); |
| 164 | high_gy = vmulq_f32(high_gy, norm_factor); |
| 165 | |
| 166 | // L3 values |
| 167 | float32x4_t gx = low_gx; |
| 168 | float32x4_t gy = low_gy; |
| 169 | |
| 170 | // Accumulate |
| 171 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 172 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 173 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 174 | |
| 175 | // L2 values |
| 176 | gx = vextq_f32(low_gx, high_gx, 1); |
| 177 | gy = vextq_f32(low_gy, high_gy, 1); |
| 178 | |
| 179 | // Accumulate |
| 180 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 181 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 182 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 183 | |
| 184 | // L1 values |
| 185 | gx = vextq_f32(low_gx, high_gx, 2); |
| 186 | gy = vextq_f32(low_gy, high_gy, 2); |
| 187 | |
| 188 | // Accumulate |
| 189 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 190 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 191 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 192 | |
| 193 | // M values |
| 194 | gx = vextq_f32(low_gx, high_gx, 3); |
| 195 | gy = vextq_f32(low_gy, high_gy, 3); |
| 196 | |
| 197 | // Accumulate |
| 198 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 199 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 200 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 201 | |
| 202 | // R1 values |
| 203 | gx = high_gx; |
| 204 | gy = high_gy; |
| 205 | |
| 206 | // Accumulate |
| 207 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 208 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 209 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 210 | |
| 211 | // Change tmp_low and tmp_high for calculating R2 and R3 values |
| 212 | low_gx = high_gx; |
| 213 | low_gy = high_gy; |
| 214 | high_gx = high_gx1; |
| 215 | high_gy = high_gy1; |
| 216 | |
| 217 | // Normalize |
| 218 | high_gx = vmulq_f32(high_gx, norm_factor); |
| 219 | high_gy = vmulq_f32(high_gy, norm_factor); |
| 220 | |
| 221 | // R2 values |
| 222 | gx = vextq_f32(low_gx, high_gx, 1); |
| 223 | gy = vextq_f32(low_gy, high_gy, 1); |
| 224 | |
| 225 | // Accumulate |
| 226 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 227 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 228 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 229 | |
| 230 | // R3 values |
| 231 | gx = vextq_f32(low_gx, high_gx, 2); |
| 232 | gy = vextq_f32(low_gy, high_gy, 2); |
| 233 | |
| 234 | // Accumulate |
| 235 | gx2 = vmlaq_f32(gx2, gx, gx); |
| 236 | gy2 = vmlaq_f32(gy2, gy, gy); |
| 237 | gxgy = vmlaq_f32(gxgy, gx, gy); |
| 238 | } |
| 239 | |
| 240 | 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, |
| 241 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 242 | |
| 243 | { |
| 244 | const auto gx_ptr_0 = static_cast<const int16_t *__restrict>(input1_ptr) - 1; |
| 245 | const auto gy_ptr_0 = static_cast<const int16_t *__restrict>(input2_ptr) - 1; |
| 246 | const int16_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 247 | const int16_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 248 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 249 | |
| 250 | // Gx^2, Gy^2 and Gx*Gy |
| 251 | float32x4x2_t gx2 = |
| 252 | { |
| 253 | { |
| 254 | vdupq_n_f32(0.0f), |
| 255 | vdupq_n_f32(0.0f) |
| 256 | } |
| 257 | }; |
| 258 | float32x4x2_t gy2 = |
| 259 | { |
| 260 | { |
| 261 | vdupq_n_f32(0.0f), |
| 262 | vdupq_n_f32(0.0f) |
| 263 | } |
| 264 | }; |
| 265 | float32x4x2_t gxgy = |
| 266 | { |
| 267 | { |
| 268 | vdupq_n_f32(0.0f), |
| 269 | vdupq_n_f32(0.0f) |
| 270 | } |
| 271 | }; |
| 272 | |
| 273 | // Row0 |
| 274 | int16x8x2_t tmp_gx = |
| 275 | { |
| 276 | { |
| 277 | vld1q_s16(gx_ptr_0 - input_stride), |
| 278 | vld1q_s16(gx_ptr_1 - input_stride) |
| 279 | } |
| 280 | }; |
| 281 | int16x8x2_t tmp_gy = |
| 282 | { |
| 283 | { |
| 284 | vld1q_s16(gy_ptr_0 - input_stride), |
| 285 | vld1q_s16(gy_ptr_1 - input_stride) |
| 286 | } |
| 287 | }; |
| 288 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 289 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 290 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 291 | |
| 292 | float32x4_t low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[0]))); |
| 293 | float32x4_t low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[0]))); |
| 294 | float32x4_t high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[0]))); |
| 295 | float32x4_t high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[0]))); |
| 296 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 297 | |
| 298 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[1]))); |
| 299 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[1]))); |
| 300 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[1]))); |
| 301 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[1]))); |
| 302 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 303 | |
| 304 | // Row1 |
| 305 | tmp_gx.val[0] = vld1q_s16(gx_ptr_0); |
| 306 | tmp_gy.val[0] = vld1q_s16(gy_ptr_0); |
| 307 | tmp_gx.val[1] = vld1q_s16(gx_ptr_1); |
| 308 | tmp_gy.val[1] = vld1q_s16(gy_ptr_1); |
| 309 | |
| 310 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[0]))); |
| 311 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[0]))); |
| 312 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[0]))); |
| 313 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[0]))); |
| 314 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 315 | |
| 316 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[1]))); |
| 317 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[1]))); |
| 318 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[1]))); |
| 319 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[1]))); |
| 320 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 321 | |
| 322 | // Row2 |
| 323 | tmp_gx.val[0] = vld1q_s16(gx_ptr_0 + input_stride); |
| 324 | tmp_gy.val[0] = vld1q_s16(gy_ptr_0 + input_stride); |
| 325 | tmp_gx.val[1] = vld1q_s16(gx_ptr_1 + input_stride); |
| 326 | tmp_gy.val[1] = vld1q_s16(gy_ptr_1 + input_stride); |
| 327 | |
| 328 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[0]))); |
| 329 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[0]))); |
| 330 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[0]))); |
| 331 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[0]))); |
| 332 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 333 | |
| 334 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[1]))); |
| 335 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[1]))); |
| 336 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[1]))); |
| 337 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[1]))); |
| 338 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 339 | |
| 340 | // Calculate harris score |
| 341 | const float32x4x2_t mc = |
| 342 | { |
| 343 | { |
| 344 | harris_score(gx2.val[0], gy2.val[0], gxgy.val[0], sensitivity, strength_thresh), |
| 345 | harris_score(gx2.val[1], gy2.val[1], gxgy.val[1], sensitivity, strength_thresh) |
| 346 | } |
| 347 | }; |
| 348 | |
| 349 | // Store score |
| 350 | vst1q_f32(output + 0, mc.val[0]); |
| 351 | vst1q_f32(output + 4, mc.val[1]); |
| 352 | } |
| 353 | |
| 354 | 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, |
| 355 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 356 | { |
| 357 | auto gx_ptr_0 = static_cast<const int32_t *__restrict>(input1_ptr) - 1; |
| 358 | auto gy_ptr_0 = static_cast<const int32_t *__restrict>(input2_ptr) - 1; |
| 359 | const int32_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 360 | const int32_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 361 | const int32_t *gx_ptr_2 = gx_ptr_0 + 8; |
| 362 | const int32_t *gy_ptr_2 = gy_ptr_0 + 8; |
| 363 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 364 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 365 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 366 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 367 | |
| 368 | // Gx^2, Gy^2 and Gx*Gy |
| 369 | float32x4x2_t gx2 = |
| 370 | { |
| 371 | { |
| 372 | vdupq_n_f32(0.0f), |
| 373 | vdupq_n_f32(0.0f) |
| 374 | } |
| 375 | }; |
| 376 | float32x4x2_t gy2 = |
| 377 | { |
| 378 | { |
| 379 | vdupq_n_f32(0.0f), |
| 380 | vdupq_n_f32(0.0f) |
| 381 | } |
| 382 | }; |
| 383 | float32x4x2_t gxgy = |
| 384 | { |
| 385 | { |
| 386 | vdupq_n_f32(0.0f), |
| 387 | vdupq_n_f32(0.0f) |
| 388 | } |
| 389 | }; |
| 390 | |
| 391 | // Row0 |
| 392 | float32x4_t low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_0 - input_stride)); |
| 393 | float32x4_t low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_0 - input_stride)); |
| 394 | float32x4_t high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1 - input_stride)); |
| 395 | float32x4_t high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1 - input_stride)); |
| 396 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 397 | |
| 398 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1 - input_stride)); |
| 399 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1 - input_stride)); |
| 400 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_2 - input_stride)); |
| 401 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_2 - input_stride)); |
| 402 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 403 | |
| 404 | // Row1 |
| 405 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_0)); |
| 406 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_0)); |
| 407 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 408 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 409 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 410 | |
| 411 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 412 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 413 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_2)); |
| 414 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_2)); |
| 415 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 416 | |
| 417 | // Row2 |
| 418 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_0 + input_stride)); |
| 419 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_0 + input_stride)); |
| 420 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1 + input_stride)); |
| 421 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1 + input_stride)); |
| 422 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 423 | |
| 424 | low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1 + input_stride)); |
| 425 | low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1 + input_stride)); |
| 426 | high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_2 + input_stride)); |
| 427 | high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_2 + input_stride)); |
| 428 | harris_score1x3_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 429 | |
| 430 | // Calculate harris score |
| 431 | const float32x4x2_t mc = |
| 432 | { |
| 433 | { |
| 434 | harris_score(gx2.val[0], gy2.val[0], gxgy.val[0], sensitivity, strength_thresh), |
| 435 | harris_score(gx2.val[1], gy2.val[1], gxgy.val[1], sensitivity, strength_thresh) |
| 436 | } |
| 437 | }; |
| 438 | |
| 439 | // Store score |
| 440 | vst1q_f32(output + 0, mc.val[0]); |
| 441 | vst1q_f32(output + 4, mc.val[1]); |
| 442 | } |
| 443 | |
| 444 | 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, |
| 445 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 446 | { |
| 447 | auto gx_ptr_0 = static_cast<const int16_t *__restrict>(input1_ptr) - 2 - 2 * input_stride; |
| 448 | auto gy_ptr_0 = static_cast<const int16_t *__restrict>(input2_ptr) - 2 - 2 * input_stride; |
| 449 | const int16_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 450 | const int16_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 451 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 452 | |
| 453 | // Gx^2, Gy^2 and Gx*Gy |
| 454 | float32x4x2_t gx2 = |
| 455 | { |
| 456 | { |
| 457 | vdupq_n_f32(0.0f), |
| 458 | vdupq_n_f32(0.0f) |
| 459 | } |
| 460 | }; |
| 461 | float32x4x2_t gy2 = |
| 462 | { |
| 463 | { |
| 464 | vdupq_n_f32(0.0f), |
| 465 | vdupq_n_f32(0.0f) |
| 466 | } |
| 467 | }; |
| 468 | float32x4x2_t gxgy = |
| 469 | { |
| 470 | { |
| 471 | vdupq_n_f32(0.0f), |
| 472 | vdupq_n_f32(0.0f) |
| 473 | } |
| 474 | }; |
| 475 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 476 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 477 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 478 | |
| 479 | for(int i = 0; i < 5; ++i) |
| 480 | { |
| 481 | const int16x8x2_t tmp_gx = |
| 482 | { |
| 483 | { |
| 484 | vld1q_s16(gx_ptr_0), |
| 485 | vld1q_s16(gx_ptr_1) |
| 486 | } |
| 487 | }; |
| 488 | const int16x8x2_t tmp_gy = |
| 489 | { |
| 490 | { |
| 491 | vld1q_s16(gy_ptr_0), |
| 492 | vld1q_s16(gy_ptr_1) |
| 493 | } |
| 494 | }; |
| 495 | |
| 496 | float32x4_t low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[0]))); |
| 497 | float32x4_t low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[0]))); |
| 498 | float32x4_t high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[0]))); |
| 499 | float32x4_t high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[0]))); |
| 500 | harris_score1x5_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[0], gy2.val[0], gxgy.val[0], norm_factor); |
| 501 | |
| 502 | low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gx.val[1]))); |
| 503 | low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp_gy.val[1]))); |
| 504 | high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gx.val[1]))); |
| 505 | high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp_gy.val[1]))); |
| 506 | harris_score1x5_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, gx2.val[1], gy2.val[1], gxgy.val[1], norm_factor); |
| 507 | |
| 508 | // Update gx and gy pointer |
| 509 | gx_ptr_0 += input_stride; |
| 510 | gy_ptr_0 += input_stride; |
| 511 | gx_ptr_1 += input_stride; |
| 512 | gy_ptr_1 += input_stride; |
| 513 | } |
| 514 | |
| 515 | // Calculate harris score |
| 516 | const float32x4x2_t mc = |
| 517 | { |
| 518 | { |
| 519 | harris_score(gx2.val[0], gy2.val[0], gxgy.val[0], sensitivity, strength_thresh), |
| 520 | harris_score(gx2.val[1], gy2.val[1], gxgy.val[1], sensitivity, strength_thresh) |
| 521 | } |
| 522 | }; |
| 523 | |
| 524 | // Store score |
| 525 | vst1q_f32(output + 0, mc.val[0]); |
| 526 | vst1q_f32(output + 4, mc.val[1]); |
| 527 | } |
| 528 | |
| 529 | 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, |
| 530 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 531 | |
| 532 | { |
| 533 | auto gx_ptr_0 = static_cast<const int32_t *__restrict>(input1_ptr) - 2 - 2 * input_stride; |
| 534 | auto gy_ptr_0 = static_cast<const int32_t *__restrict>(input2_ptr) - 2 - 2 * input_stride; |
| 535 | const int32_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 536 | const int32_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 537 | const int32_t *gx_ptr_2 = gx_ptr_0 + 8; |
| 538 | const int32_t *gy_ptr_2 = gy_ptr_0 + 8; |
| 539 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 540 | |
| 541 | // Gx^2, Gy^2 and Gx*Gy |
| 542 | float32x4x2_t gx2 = |
| 543 | { |
| 544 | { |
| 545 | vdupq_n_f32(0.0f), |
| 546 | vdupq_n_f32(0.0f) |
| 547 | } |
| 548 | }; |
| 549 | float32x4x2_t gy2 = |
| 550 | { |
| 551 | { |
| 552 | vdupq_n_f32(0.0f), |
| 553 | vdupq_n_f32(0.0f) |
| 554 | } |
| 555 | }; |
| 556 | float32x4x2_t gxgy = |
| 557 | { |
| 558 | { |
| 559 | vdupq_n_f32(0.0f), |
| 560 | vdupq_n_f32(0.0f) |
| 561 | } |
| 562 | }; |
| 563 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 564 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 565 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 566 | |
| 567 | for(int i = 0; i < 5; ++i) |
| 568 | { |
| 569 | const float32x4_t low_gx_0 = vcvtq_f32_s32(vld1q_s32(gx_ptr_0)); |
| 570 | const float32x4_t low_gy_0 = vcvtq_f32_s32(vld1q_s32(gy_ptr_0)); |
| 571 | const float32x4_t high_gx_0 = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 572 | const float32x4_t high_gy_0 = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 573 | 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); |
| 574 | |
| 575 | const float32x4_t low_gx_1 = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 576 | const float32x4_t low_gy_1 = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 577 | const float32x4_t high_gx_1 = vcvtq_f32_s32(vld1q_s32(gx_ptr_2)); |
| 578 | const float32x4_t high_gy_1 = vcvtq_f32_s32(vld1q_s32(gy_ptr_2)); |
| 579 | 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); |
| 580 | |
| 581 | // Update gx and gy pointer |
| 582 | gx_ptr_0 += input_stride; |
| 583 | gy_ptr_0 += input_stride; |
| 584 | gx_ptr_1 += input_stride; |
| 585 | gy_ptr_1 += input_stride; |
| 586 | gx_ptr_2 += input_stride; |
| 587 | gy_ptr_2 += input_stride; |
| 588 | } |
| 589 | |
| 590 | // Calculate harris score |
| 591 | const float32x4x2_t mc = |
| 592 | { |
| 593 | { |
| 594 | harris_score(gx2.val[0], gy2.val[0], gxgy.val[0], sensitivity, strength_thresh), |
| 595 | harris_score(gx2.val[1], gy2.val[1], gxgy.val[1], sensitivity, strength_thresh) |
| 596 | } |
| 597 | }; |
| 598 | |
| 599 | // Store score |
| 600 | vst1q_f32(output + 0, mc.val[0]); |
| 601 | vst1q_f32(output + 4, mc.val[1]); |
| 602 | } |
| 603 | |
| 604 | 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, |
| 605 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 606 | { |
| 607 | auto gx_ptr_0 = static_cast<const int16_t *__restrict>(input1_ptr) - 3 - 3 * input_stride; |
| 608 | auto gy_ptr_0 = static_cast<const int16_t *__restrict>(input2_ptr) - 3 - 3 * input_stride; |
| 609 | const int16_t *gx_ptr_1 = gx_ptr_0 + 8; |
| 610 | const int16_t *gy_ptr_1 = gy_ptr_0 + 8; |
| 611 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 612 | |
| 613 | // Gx^2, Gy^2 and Gx*Gy |
| 614 | float32x4_t gx2 = vdupq_n_f32(0.0f); |
| 615 | float32x4_t gy2 = vdupq_n_f32(0.0f); |
| 616 | float32x4_t gxgy = vdupq_n_f32(0.0f); |
| 617 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 618 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 619 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 620 | |
| 621 | for(int i = 0; i < 7; ++i) |
| 622 | { |
| 623 | const int16x8_t tmp0_gx = vld1q_s16(gx_ptr_0); |
| 624 | const int16x8_t tmp0_gy = vld1q_s16(gy_ptr_0); |
| 625 | const int16x4_t tmp1_gx = vld1_s16(gx_ptr_1); |
| 626 | const int16x4_t tmp1_gy = vld1_s16(gy_ptr_1); |
| 627 | |
| 628 | float32x4_t low_gx = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp0_gx))); |
| 629 | float32x4_t low_gy = vcvtq_f32_s32(vmovl_s16(vget_low_s16(tmp0_gy))); |
| 630 | float32x4_t high_gx = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp0_gx))); |
| 631 | float32x4_t high_gy = vcvtq_f32_s32(vmovl_s16(vget_high_s16(tmp0_gy))); |
| 632 | float32x4_t high_gx1 = vcvtq_f32_s32(vmovl_s16(tmp1_gx)); |
| 633 | float32x4_t high_gy1 = vcvtq_f32_s32(vmovl_s16(tmp1_gy)); |
| 634 | harris_score1x7_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, high_gx1, high_gy1, gx2, gy2, gxgy, norm_factor); |
| 635 | |
| 636 | // Update gx and gy pointer |
| 637 | gx_ptr_0 += input_stride; |
| 638 | gy_ptr_0 += input_stride; |
| 639 | gx_ptr_1 += input_stride; |
| 640 | gy_ptr_1 += input_stride; |
| 641 | } |
| 642 | |
| 643 | // Calculate harris score |
| 644 | const float32x4_t mc = harris_score(gx2, gy2, gxgy, sensitivity, strength_thresh); |
| 645 | |
| 646 | // Store score |
| 647 | vst1q_f32(output, mc); |
| 648 | } |
| 649 | |
| 650 | 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, |
| 651 | float in_norm_factor, float in_sensitivity, float in_strength_thresh) |
| 652 | { |
| 653 | auto gx_ptr_0 = static_cast<const int32_t *__restrict>(input1_ptr) - 3 - 3 * input_stride; |
| 654 | auto gy_ptr_0 = static_cast<const int32_t *__restrict>(input2_ptr) - 3 - 3 * input_stride; |
| 655 | const int32_t *gx_ptr_1 = gx_ptr_0 + 4; |
| 656 | const int32_t *gy_ptr_1 = gy_ptr_0 + 4; |
| 657 | const int32_t *gx_ptr_2 = gx_ptr_1 + 4; |
| 658 | const int32_t *gy_ptr_2 = gy_ptr_1 + 4; |
| 659 | const auto output = static_cast<float *__restrict>(output_ptr); |
| 660 | |
| 661 | // Gx^2, Gy^2 and Gx*Gy |
| 662 | float32x4_t gx2 = vdupq_n_f32(0.0f); |
| 663 | float32x4_t gy2 = vdupq_n_f32(0.0f); |
| 664 | float32x4_t gxgy = vdupq_n_f32(0.0f); |
| 665 | float32x4_t sensitivity = vdupq_n_f32(in_sensitivity); |
| 666 | float32x4_t norm_factor = vdupq_n_f32(in_norm_factor); |
| 667 | float32x4_t strength_thresh = vdupq_n_f32(in_strength_thresh); |
| 668 | |
| 669 | for(int i = 0; i < 7; ++i) |
| 670 | { |
| 671 | const float32x4_t low_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_0)); |
| 672 | const float32x4_t low_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_0)); |
| 673 | const float32x4_t high_gx = vcvtq_f32_s32(vld1q_s32(gx_ptr_1)); |
| 674 | const float32x4_t high_gy = vcvtq_f32_s32(vld1q_s32(gy_ptr_1)); |
| 675 | const float32x4_t high_gx1 = vcvtq_f32_s32(vld1q_s32(gx_ptr_2)); |
| 676 | const float32x4_t high_gy1 = vcvtq_f32_s32(vld1q_s32(gy_ptr_2)); |
| 677 | harris_score1x7_FLOAT_FLOAT_FLOAT(low_gx, low_gy, high_gx, high_gy, high_gx1, high_gy1, gx2, gy2, gxgy, norm_factor); |
| 678 | |
| 679 | // Update gx and gy pointer |
| 680 | gx_ptr_0 += input_stride; |
| 681 | gy_ptr_0 += input_stride; |
| 682 | gx_ptr_1 += input_stride; |
| 683 | gy_ptr_1 += input_stride; |
| 684 | gx_ptr_2 += input_stride; |
| 685 | gy_ptr_2 += input_stride; |
| 686 | } |
| 687 | |
| 688 | // Calculate harris score |
| 689 | const float32x4_t mc = harris_score(gx2, gy2, gxgy, sensitivity, strength_thresh); |
| 690 | |
| 691 | // Store score |
| 692 | vst1q_f32(output, mc); |
| 693 | } |
| 694 | |
| 695 | } // namespace |
| 696 | |
| 697 | INEHarrisScoreKernel::INEHarrisScoreKernel() |
| 698 | : _input1(nullptr), _input2(nullptr), _output(nullptr), _sensitivity(0.0f), _strength_thresh(0.0f), _norm_factor(0.0f), _border_size() |
| 699 | { |
| 700 | } |
| 701 | |
| 702 | template <int32_t block_size> |
| 703 | NEHarrisScoreKernel<block_size>::NEHarrisScoreKernel() |
| 704 | : INEHarrisScoreKernel(), _func(nullptr) |
| 705 | { |
| 706 | } |
| 707 | |
| 708 | template <int32_t block_size> |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 709 | void NEHarrisScoreKernel<block_size>::run(const Window &window, const ThreadInfo &info) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 710 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 711 | ARM_COMPUTE_UNUSED(info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 712 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 713 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 714 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 715 | |
| 716 | Iterator input1(_input1, window); |
| 717 | Iterator input2(_input2, window); |
| 718 | Iterator output(_output, window); |
| 719 | |
| 720 | const size_t input_stride = _input1->info()->strides_in_bytes()[1] / element_size_from_data_type(_input1->info()->data_type()); |
| 721 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 722 | execute_window_loop(window, [&](const Coordinates &) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 723 | { |
| 724 | (*_func)(input1.ptr(), input2.ptr(), output.ptr(), input_stride, _norm_factor, _sensitivity, _strength_thresh); |
| 725 | }, |
| 726 | input1, input2, output); |
| 727 | } |
| 728 | |
| 729 | template <int32_t block_size> |
| 730 | BorderSize NEHarrisScoreKernel<block_size>::border_size() const |
| 731 | { |
| 732 | return _border_size; |
| 733 | } |
| 734 | |
| 735 | template <int32_t block_size> |
| 736 | void NEHarrisScoreKernel<block_size>::configure(const IImage *input1, const IImage *input2, IImage *output, float norm_factor, float strength_thresh, float sensitivity, |
| 737 | bool border_undefined) |
| 738 | { |
| 739 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input1); |
| 740 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input2); |
| 741 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(output); |
| 742 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::S16, DataType::S32); |
| 743 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::S16, DataType::S32); |
| 744 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); |
| 745 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2); |
| 746 | ARM_COMPUTE_ERROR_ON(0.0f == norm_factor); |
| 747 | |
| 748 | _input1 = input1; |
| 749 | _input2 = input2; |
| 750 | _output = output; |
| 751 | _sensitivity = sensitivity; |
| 752 | _strength_thresh = strength_thresh; |
| 753 | _norm_factor = norm_factor; |
| 754 | _border_size = BorderSize(block_size / 2); |
| 755 | |
| 756 | if(input1->info()->data_type() == DataType::S16) |
| 757 | { |
| 758 | switch(block_size) |
| 759 | { |
| 760 | case 3: |
| 761 | _func = &harris_score3x3_S16_S16_FLOAT; |
| 762 | break; |
| 763 | case 5: |
| 764 | _func = &harris_score5x5_S16_S16_FLOAT; |
| 765 | break; |
| 766 | case 7: |
| 767 | _func = &harris_score7x7_S16_S16_FLOAT; |
| 768 | break; |
| 769 | default: |
| 770 | ARM_COMPUTE_ERROR("Invalid block size"); |
| 771 | break; |
| 772 | } |
| 773 | } |
| 774 | else |
| 775 | { |
| 776 | switch(block_size) |
| 777 | { |
| 778 | case 3: |
| 779 | _func = &harris_score3x3_S32_S32_FLOAT; |
| 780 | break; |
| 781 | case 5: |
| 782 | _func = &harris_score5x5_S32_S32_FLOAT; |
| 783 | break; |
| 784 | case 7: |
| 785 | _func = &harris_score7x7_S32_S32_FLOAT; |
| 786 | break; |
| 787 | default: |
| 788 | ARM_COMPUTE_ERROR("Invalid block size"); |
| 789 | break; |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | ARM_COMPUTE_ERROR_ON(nullptr == _func); |
| 794 | |
| 795 | constexpr unsigned int num_elems_processed_per_iteration = block_size != 7 ? 8 : 4; |
| 796 | constexpr unsigned int num_elems_read_per_iteration = block_size != 7 ? 16 : 12; |
| 797 | constexpr unsigned int num_elems_written_per_iteration = block_size != 7 ? 8 : 4; |
| 798 | constexpr unsigned int num_rows_read_per_iteration = block_size; |
| 799 | |
| 800 | // Configure kernel window |
| 801 | Window win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); |
| 802 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
| 803 | |
| 804 | update_window_and_padding(win, |
| 805 | AccessWindowRectangle(input1->info(), -_border_size.left, -_border_size.top, num_elems_read_per_iteration, num_rows_read_per_iteration), |
| 806 | AccessWindowRectangle(input2->info(), -_border_size.left, -_border_size.top, num_elems_read_per_iteration, num_rows_read_per_iteration), |
| 807 | output_access); |
| 808 | |
| 809 | ValidRegion valid_region = intersect_valid_regions(input1->info()->valid_region(), |
| 810 | input2->info()->valid_region()); |
| 811 | |
| 812 | output_access.set_valid_region(win, valid_region, border_undefined, border_size()); |
| 813 | |
| 814 | INEKernel::configure(win); |
| 815 | } |