| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // SPDX-FileCopyrightText: 2023 - 2026 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 2 | // | ||
| 3 | // SPDX-License-Identifier: Apache-2.0 | ||
| 4 | |||
| 5 | #ifndef KLEIDICV_GAUSSIAN_BLUR_SC_H | ||
| 6 | #define KLEIDICV_GAUSSIAN_BLUR_SC_H | ||
| 7 | |||
| 8 | #include <array> | ||
| 9 | #include <cassert> | ||
| 10 | |||
| 11 | #include "kleidicv/filters/gaussian_blur.h" | ||
| 12 | #include "kleidicv/filters/separable_filter_15x15_sc.h" | ||
| 13 | #include "kleidicv/filters/separable_filter_21x21_sc.h" | ||
| 14 | #include "kleidicv/filters/separable_filter_3x3_sc.h" | ||
| 15 | #include "kleidicv/filters/separable_filter_5x5_sc.h" | ||
| 16 | #include "kleidicv/filters/separable_filter_7x7_sc.h" | ||
| 17 | #include "kleidicv/filters/separable_filter_9x9_sc.h" | ||
| 18 | #include "kleidicv/filters/sigma.h" | ||
| 19 | #include "kleidicv/workspace/separable.h" | ||
| 20 | |||
| 21 | #if KLEIDICV_TARGET_SME || KLEIDICV_TARGET_SME2 | ||
| 22 | #include <arm_sme.h> | ||
| 23 | #endif | ||
| 24 | |||
| 25 | namespace KLEIDICV_TARGET_NAMESPACE { | ||
| 26 | |||
| 27 | // Primary template for Gaussian Blur filters. | ||
| 28 | template <typename ScalarType, size_t KernelSize, bool IsBinomial> | ||
| 29 | class GaussianBlur; | ||
| 30 | |||
| 31 | // Template for 3x3 Gaussian Blur binomial filters. | ||
| 32 | // | ||
| 33 | // [ 1, 2, 1 ] [ 1 ] | ||
| 34 | // F = 1/16 * [ 2, 4, 2 ] = 1/16 * [ 2 ] * [ 1, 2, 1 ] | ||
| 35 | // [ 1, 2, 1 ] [ 1 ] | ||
| 36 | template <> | ||
| 37 | class GaussianBlur<uint8_t, 3, true> { | ||
| 38 | public: | ||
| 39 | using SourceType = uint8_t; | ||
| 40 | using BufferType = uint16_t; | ||
| 41 | using DestinationType = uint8_t; | ||
| 42 | |||
| 43 | // Applies vertical filtering vector using SIMD operations. | ||
| 44 | // | ||
| 45 | // DST = [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 46 | 804 | void vertical_vector_path(svbool_t pg, | |
| 47 | std::reference_wrapper<svuint8_t> src[3], | ||
| 48 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 49 | 804 | svuint16_t acc_0_2_b = svaddlb_u16(src[0], src[2]); | |
| 50 | 804 | svuint16_t acc_0_2_t = svaddlt_u16(src[0], src[2]); | |
| 51 | |||
| 52 | 804 | svuint16_t acc_1_b = svshllb_n_u16(src[1], 1); | |
| 53 | 804 | svuint16_t acc_1_t = svshllt_n_u16(src[1], 1); | |
| 54 | |||
| 55 | 804 | svuint16_t acc_u16_b = svadd_u16_x(pg, acc_0_2_b, acc_1_b); | |
| 56 | 804 | svuint16_t acc_u16_t = svadd_u16_x(pg, acc_0_2_t, acc_1_t); | |
| 57 | |||
| 58 | 804 | svuint16x2_t interleaved = svcreate2(acc_u16_b, acc_u16_t); | |
| 59 | 804 | svst2(pg, &dst[0], interleaved); | |
| 60 | 804 | } | |
| 61 | |||
| 62 | // Applies horizontal filtering vector using SIMD operations. | ||
| 63 | // | ||
| 64 | // DST = 1/16 * [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 65 | 604 | void horizontal_vector_path(svbool_t pg, | |
| 66 | std::reference_wrapper<svuint16_t> src[3], | ||
| 67 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 68 | 604 | svuint16_t acc_0_2 = svhadd_u16_x(pg, src[0], src[2]); | |
| 69 | |||
| 70 | 604 | svuint16_t acc = svadd_u16_x(pg, acc_0_2, src[1]); | |
| 71 | 604 | acc = svrshr_x(pg, acc, 3); | |
| 72 | |||
| 73 | 604 | svst1b(pg, &dst[0], acc); | |
| 74 | 604 | } | |
| 75 | |||
| 76 | // Applies horizontal filtering vector using scalar operations. | ||
| 77 | // | ||
| 78 | // DST = 1/16 * [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 79 | 1608 | void horizontal_scalar_path(const BufferType src[3], | |
| 80 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 81 | 1608 | auto acc = src[0] + 2 * src[1] + src[2]; | |
| 82 | 1608 | dst[0] = rounding_shift_right(acc, 4); | |
| 83 | 1608 | } | |
| 84 | }; // end of class GaussianBlur<uint8_t, 3, true> | ||
| 85 | |||
| 86 | // Template for 5x5 Gaussian Blur binomial filters. | ||
| 87 | // | ||
| 88 | // [ 1, 4, 6, 4, 1 ] [ 1 ] | ||
| 89 | // [ 4, 16, 24, 16, 4 ] [ 4 ] | ||
| 90 | // F = 1/256 * [ 6, 24, 36, 24, 6 ] = 1/256 * [ 6 ] * [ 1, 4, 6, 4, 1 ] | ||
| 91 | // [ 4, 16, 24, 16, 4 ] [ 4 ] | ||
| 92 | // [ 1, 4, 6, 4, 1 ] [ 1 ] | ||
| 93 | template <> | ||
| 94 | class GaussianBlur<uint8_t, 5, true> { | ||
| 95 | public: | ||
| 96 | using SourceType = uint8_t; | ||
| 97 | using BufferType = uint16_t; | ||
| 98 | using DestinationType = uint8_t; | ||
| 99 | |||
| 100 | // Applies vertical filtering vector using SIMD operations. | ||
| 101 | // | ||
| 102 | // DST = [ SRC0, SRC1, SRC2, SRC3, SRC4 ] * [ 1, 4, 6, 4, 1 ]T | ||
| 103 | 2218 | void vertical_vector_path(svbool_t pg, | |
| 104 | std::reference_wrapper<svuint8_t> src[5], | ||
| 105 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 106 | 2218 | svuint16_t acc_0_4_b = svaddlb_u16(src[0], src[4]); | |
| 107 | 2218 | svuint16_t acc_0_4_t = svaddlt_u16(src[0], src[4]); | |
| 108 | 2218 | svuint16_t acc_1_3_b = svaddlb_u16(src[1], src[3]); | |
| 109 | 2218 | svuint16_t acc_1_3_t = svaddlt_u16(src[1], src[3]); | |
| 110 | |||
| 111 | 2218 | svuint16_t acc_u16_b = svmlalb_n_u16(acc_0_4_b, src[2], 6); | |
| 112 | 2218 | svuint16_t acc_u16_t = svmlalt_n_u16(acc_0_4_t, src[2], 6); | |
| 113 | 2218 | acc_u16_b = svmla_n_u16_x(pg, acc_u16_b, acc_1_3_b, 4); | |
| 114 | 2218 | acc_u16_t = svmla_n_u16_x(pg, acc_u16_t, acc_1_3_t, 4); | |
| 115 | |||
| 116 | 2218 | svuint16x2_t interleaved = svcreate2(acc_u16_b, acc_u16_t); | |
| 117 | 2218 | svst2(pg, &dst[0], interleaved); | |
| 118 | 2218 | } | |
| 119 | |||
| 120 | // Applies horizontal filtering vector using SIMD operations. | ||
| 121 | // | ||
| 122 | // DST = 1/256 * [ SRC0, SRC1, SRC2, SRC3, SRC4 ] * [ 1, 4, 6, 4, 1 ]T | ||
| 123 | 1872 | void horizontal_vector_path(svbool_t pg, | |
| 124 | std::reference_wrapper<svuint16_t> src[5], | ||
| 125 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 126 | 1872 | svuint16_t acc_0_4 = svadd_x(pg, src[0], src[4]); | |
| 127 | 1872 | svuint16_t acc_1_3 = svadd_x(pg, src[1], src[3]); | |
| 128 | 1872 | svuint16_t acc = svmla_n_u16_x(pg, acc_0_4, src[2], 6); | |
| 129 | 1872 | acc = svmla_n_u16_x(pg, acc, acc_1_3, 4); | |
| 130 | 1872 | acc = svrshr_x(pg, acc, 8); | |
| 131 | 1872 | svst1b(pg, &dst[0], acc); | |
| 132 | 1872 | } | |
| 133 | |||
| 134 | // Applies horizontal filtering vector using scalar operations. | ||
| 135 | // | ||
| 136 | // DST = 1/256 * [ SRC0, SRC1, SRC2, SRC3, SRC4 ] * [ 1, 4, 6, 4, 1 ]T | ||
| 137 | 8328 | void horizontal_scalar_path(const BufferType src[5], | |
| 138 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 139 | 8328 | auto acc = src[0] + src[4] + 4 * (src[1] + src[3]) + 6 * src[2]; | |
| 140 | 8328 | dst[0] = rounding_shift_right(acc, 8); | |
| 141 | 8328 | } | |
| 142 | }; // end of class GaussianBlur<uint8_t, 5, true> | ||
| 143 | |||
| 144 | // Template for 7x7 Gaussian Blur binomial filters. | ||
| 145 | // | ||
| 146 | // [ 4, 14, 28, 36, 28, 14, 4 ] | ||
| 147 | // [ 14, 49, 98, 126, 98, 49, 14 ] | ||
| 148 | // [ 28, 98, 196, 252, 196, 98, 28 ] | ||
| 149 | // F = 1/4096 * [ 36, 126, 252, 324, 252, 126, 36 ] = | ||
| 150 | // [ 28, 98, 196, 252, 196, 98, 28 ] | ||
| 151 | // [ 14, 49, 98, 126, 98, 49, 14 ] | ||
| 152 | // [ 4, 14, 28, 36, 28, 14, 4 ] | ||
| 153 | // | ||
| 154 | // [ 2 ] | ||
| 155 | // [ 7 ] | ||
| 156 | // [ 14 ] | ||
| 157 | // = 1/4096 * [ 18 ] * [ 2, 7, 14, 18, 14, 7, 2 ] | ||
| 158 | // [ 14 ] | ||
| 159 | // [ 7 ] | ||
| 160 | // [ 2 ] | ||
| 161 | template <> | ||
| 162 | class GaussianBlur<uint8_t, 7, true> { | ||
| 163 | public: | ||
| 164 | using SourceType = uint8_t; | ||
| 165 | using BufferType = uint16_t; | ||
| 166 | using DestinationType = uint8_t; | ||
| 167 | |||
| 168 | // Applies vertical filtering vector using SIMD operations. | ||
| 169 | // | ||
| 170 | // DST = [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6 ] * | ||
| 171 | // * [ 2, 7, 14, 18, 14, 7, 2 ]T | ||
| 172 | 870 | void vertical_vector_path(svbool_t pg, | |
| 173 | std::reference_wrapper<svuint8_t> src[7], | ||
| 174 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 175 | 870 | svuint16_t acc_0_6_b = svaddlb_u16(src[0], src[6]); | |
| 176 | 870 | svuint16_t acc_0_6_t = svaddlt_u16(src[0], src[6]); | |
| 177 | |||
| 178 | 870 | svuint16_t acc_1_5_b = svaddlb_u16(src[1], src[5]); | |
| 179 | 870 | svuint16_t acc_1_5_t = svaddlt_u16(src[1], src[5]); | |
| 180 | |||
| 181 | 870 | svuint16_t acc_2_4_b = svaddlb_u16(src[2], src[4]); | |
| 182 | 870 | svuint16_t acc_2_4_t = svaddlt_u16(src[2], src[4]); | |
| 183 | |||
| 184 | 870 | svuint16_t acc_3_b = svmovlb_u16(src[3]); | |
| 185 | 870 | svuint16_t acc_3_t = svmovlt_u16(src[3]); | |
| 186 | |||
| 187 | 870 | svuint16_t acc_0_2_4_6_b = svmla_n_u16_x(pg, acc_0_6_b, acc_2_4_b, 7); | |
| 188 | 870 | svuint16_t acc_0_2_4_6_t = svmla_n_u16_x(pg, acc_0_6_t, acc_2_4_t, 7); | |
| 189 | |||
| 190 | 870 | svuint16_t acc_0_2_3_4_6_b = svmla_n_u16_x(pg, acc_0_2_4_6_b, acc_3_b, 9); | |
| 191 | 870 | svuint16_t acc_0_2_3_4_6_t = svmla_n_u16_x(pg, acc_0_2_4_6_t, acc_3_t, 9); | |
| 192 | 870 | acc_0_2_3_4_6_b = svlsl_n_u16_x(pg, acc_0_2_3_4_6_b, 1); | |
| 193 | 870 | acc_0_2_3_4_6_t = svlsl_n_u16_x(pg, acc_0_2_3_4_6_t, 1); | |
| 194 | |||
| 195 | 1740 | svuint16_t acc_0_1_2_3_4_5_6_b = | |
| 196 | 870 | svmla_n_u16_x(pg, acc_0_2_3_4_6_b, acc_1_5_b, 7); | |
| 197 | 1740 | svuint16_t acc_0_1_2_3_4_5_6_t = | |
| 198 | 870 | svmla_n_u16_x(pg, acc_0_2_3_4_6_t, acc_1_5_t, 7); | |
| 199 | |||
| 200 | 1740 | svuint16x2_t interleaved = | |
| 201 | 870 | svcreate2(acc_0_1_2_3_4_5_6_b, acc_0_1_2_3_4_5_6_t); | |
| 202 | 870 | svst2(pg, &dst[0], interleaved); | |
| 203 | 870 | } | |
| 204 | |||
| 205 | // Applies horizontal filtering vector using SIMD operations. | ||
| 206 | // | ||
| 207 | // DST = 1/4096 * [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6 ] * | ||
| 208 | // * [ 2, 7, 14, 18, 14, 7, 2 ]T | ||
| 209 | 684 | void horizontal_vector_path(svbool_t pg, | |
| 210 | std::reference_wrapper<svuint16_t> src[7], | ||
| 211 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 212 | 684 | svuint32_t acc_0_6_b = svaddlb_u32(src[0], src[6]); | |
| 213 | 684 | svuint32_t acc_0_6_t = svaddlt_u32(src[0], src[6]); | |
| 214 | |||
| 215 | 684 | svuint32_t acc_1_5_b = svaddlb_u32(src[1], src[5]); | |
| 216 | 684 | svuint32_t acc_1_5_t = svaddlt_u32(src[1], src[5]); | |
| 217 | |||
| 218 | 684 | svuint16_t acc_2_4 = svadd_u16_x(pg, src[2], src[4]); | |
| 219 | |||
| 220 | 684 | svuint32_t acc_0_2_4_6_b = svmlalb_n_u32(acc_0_6_b, acc_2_4, 7); | |
| 221 | 684 | svuint32_t acc_0_2_4_6_t = svmlalt_n_u32(acc_0_6_t, acc_2_4, 7); | |
| 222 | |||
| 223 | 684 | svuint32_t acc_0_2_3_4_6_b = svmlalb_n_u32(acc_0_2_4_6_b, src[3], 9); | |
| 224 | 684 | svuint32_t acc_0_2_3_4_6_t = svmlalt_n_u32(acc_0_2_4_6_t, src[3], 9); | |
| 225 | |||
| 226 | 684 | acc_0_2_3_4_6_b = svlsl_n_u32_x(pg, acc_0_2_3_4_6_b, 1); | |
| 227 | 684 | acc_0_2_3_4_6_t = svlsl_n_u32_x(pg, acc_0_2_3_4_6_t, 1); | |
| 228 | |||
| 229 | 1368 | svuint32_t acc_0_1_2_3_4_5_6_b = | |
| 230 | 684 | svmla_n_u32_x(pg, acc_0_2_3_4_6_b, acc_1_5_b, 7); | |
| 231 | 1368 | svuint32_t acc_0_1_2_3_4_5_6_t = | |
| 232 | 684 | svmla_n_u32_x(pg, acc_0_2_3_4_6_t, acc_1_5_t, 7); | |
| 233 | |||
| 234 | 1368 | svuint16_t acc_0_1_2_3_4_5_6_u16_b = | |
| 235 | 684 | svrshrnb_n_u32(acc_0_1_2_3_4_5_6_b, 12); | |
| 236 | 1368 | svuint16_t acc_0_1_2_3_4_5_6_u16 = | |
| 237 | 684 | svrshrnt_n_u32(acc_0_1_2_3_4_5_6_u16_b, acc_0_1_2_3_4_5_6_t, 12); | |
| 238 | |||
| 239 | 684 | svst1b(pg, &dst[0], acc_0_1_2_3_4_5_6_u16); | |
| 240 | 684 | } | |
| 241 | |||
| 242 | // Applies horizontal filtering vector using scalar operations. | ||
| 243 | // | ||
| 244 | // DST = 1/4096 * [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6 ] * | ||
| 245 | // * [ 2, 7, 14, 18, 14, 7, 2 ]T | ||
| 246 | 4932 | void horizontal_scalar_path(const BufferType src[7], | |
| 247 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 248 | 14796 | uint32_t acc = src[0] * 2 + src[1] * 7 + src[2] * 14 + src[3] * 18 + | |
| 249 | 9864 | src[4] * 14 + src[5] * 7 + src[6] * 2; | |
| 250 | 4932 | dst[0] = rounding_shift_right(acc, 12); | |
| 251 | 4932 | } | |
| 252 | }; // end of class GaussianBlur<uint8_t, 7, true> | ||
| 253 | |||
| 254 | // Template for 9x9 Gaussian Blur binomial filters. | ||
| 255 | // | ||
| 256 | // [ 16, 52, 120, 204, 240, 204, 120, 52, 16 ] | ||
| 257 | // [ 52, 169, 390, 663, 780, 663, 390, 169, 52 ] | ||
| 258 | // [ 120, 390, 900, 1530, 1800, 1530, 900, 390, 120 ] | ||
| 259 | // F = 1/65536 * [ 204, 663, 1530, 2601, 3060, 2601, 1530, 663, 204 ] = | ||
| 260 | // [ 240, 780, 1800, 3060, 3600, 3060, 1800, 780, 240 ] | ||
| 261 | // [ 204, 663, 1530, 2601, 3060, 2601, 1530, 663, 204 ] | ||
| 262 | // [ 120, 390, 900, 1530, 1800, 1530, 900, 390, 120 ] | ||
| 263 | // [ 52, 169, 390, 663, 780, 663, 390, 169, 52 ] | ||
| 264 | // [ 16, 52, 120, 204, 240, 204, 120, 52, 16 ] | ||
| 265 | // | ||
| 266 | // [ 4 ] | ||
| 267 | // [ 13 ] | ||
| 268 | // [ 30 ] | ||
| 269 | // = 1/65536 * [ 51 ] * [ 4, 13, 30, 51, 60, 51, 30, 13, 4 ] | ||
| 270 | // [ 60 ] | ||
| 271 | // [ 51 ] | ||
| 272 | // [ 30 ] | ||
| 273 | // [ 13 ] | ||
| 274 | // [ 4 ] | ||
| 275 | template <> | ||
| 276 | class GaussianBlur<uint8_t, 9, true> { | ||
| 277 | public: | ||
| 278 | using SourceType = uint8_t; | ||
| 279 | using BufferType = uint16_t; | ||
| 280 | using DestinationType = uint8_t; | ||
| 281 | |||
| 282 | // Applies vertical filtering vector using SIMD operations. | ||
| 283 | // | ||
| 284 | // DST = [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6, SRC7, SRC8 ] * | ||
| 285 | // * [ 4, 13, 30, 51, 60, 51, 30, 13, 4 ]T | ||
| 286 | 1904 | void vertical_vector_path(svbool_t pg, | |
| 287 | std::reference_wrapper<svuint8_t> src[9], | ||
| 288 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 289 | // Lane-level split after widening: *_lo/*_hi are lower/upper lanes. | ||
| 290 | 1904 | svuint16_t acc_0_8_lo = svaddlb_u16(src[0], src[8]); | |
| 291 | 1904 | svuint16_t acc_0_8_hi = svaddlt_u16(src[0], src[8]); | |
| 292 | |||
| 293 | 1904 | svuint16_t acc_1_7_lo = svaddlb_u16(src[1], src[7]); | |
| 294 | 1904 | svuint16_t acc_1_7_hi = svaddlt_u16(src[1], src[7]); | |
| 295 | |||
| 296 | 1904 | svuint16_t acc_2_6_lo = svaddlb_u16(src[2], src[6]); | |
| 297 | 1904 | svuint16_t acc_2_6_hi = svaddlt_u16(src[2], src[6]); | |
| 298 | |||
| 299 | 1904 | svuint16_t acc_3_5_lo = svaddlb_u16(src[3], src[5]); | |
| 300 | 1904 | svuint16_t acc_3_5_hi = svaddlt_u16(src[3], src[5]); | |
| 301 | |||
| 302 | 1904 | svuint16_t acc_4_lo = svmovlb_u16(src[4]); | |
| 303 | 1904 | svuint16_t acc_4_hi = svmovlt_u16(src[4]); | |
| 304 | |||
| 305 | // Window-level grouping: *_tap_even/*_tap_odd are even/odd taps | ||
| 306 | // (0, 2, 4, 6, 8 vs 1, 3, 5, 7). | ||
| 307 | 1904 | svuint16_t acc_lo_tap_even = svlsl_n_u16_x(pg, acc_0_8_lo, 2); | |
| 308 | 1904 | svuint16_t acc_hi_tap_even = svlsl_n_u16_x(pg, acc_0_8_hi, 2); | |
| 309 | 1904 | acc_lo_tap_even = svmla_n_u16_x(pg, acc_lo_tap_even, acc_2_6_lo, 30); | |
| 310 | 1904 | acc_hi_tap_even = svmla_n_u16_x(pg, acc_hi_tap_even, acc_2_6_hi, 30); | |
| 311 | 1904 | acc_lo_tap_even = svmla_n_u16_x(pg, acc_lo_tap_even, acc_4_lo, 60); | |
| 312 | 1904 | acc_hi_tap_even = svmla_n_u16_x(pg, acc_hi_tap_even, acc_4_hi, 60); | |
| 313 | |||
| 314 | 1904 | svuint16_t acc_lo_tap_odd = svmul_n_u16_x(pg, acc_1_7_lo, 13); | |
| 315 | 1904 | svuint16_t acc_hi_tap_odd = svmul_n_u16_x(pg, acc_1_7_hi, 13); | |
| 316 | 1904 | acc_lo_tap_odd = svmla_n_u16_x(pg, acc_lo_tap_odd, acc_3_5_lo, 51); | |
| 317 | 1904 | acc_hi_tap_odd = svmla_n_u16_x(pg, acc_hi_tap_odd, acc_3_5_hi, 51); | |
| 318 | |||
| 319 | 1904 | svuint16_t acc_lo = svadd_u16_x(pg, acc_lo_tap_even, acc_lo_tap_odd); | |
| 320 | 1904 | svuint16_t acc_hi = svadd_u16_x(pg, acc_hi_tap_even, acc_hi_tap_odd); | |
| 321 | |||
| 322 | 1904 | svuint16x2_t interleaved = svcreate2(acc_lo, acc_hi); | |
| 323 | 1904 | svst2(pg, &dst[0], interleaved); | |
| 324 | 1904 | } | |
| 325 | |||
| 326 | // Applies horizontal filtering vector using SIMD operations. | ||
| 327 | // | ||
| 328 | // DST = 1/65536 * [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6, SRC7, SRC8 ] * | ||
| 329 | // * [ 4, 13, 30, 51, 60, 51, 30, 13, 4 ]T | ||
| 330 | 1628 | void horizontal_vector_path(svbool_t pg, | |
| 331 | std::reference_wrapper<svuint16_t> src[9], | ||
| 332 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 333 | // Lane-level split after widening: *_lo/*_hi are lower/upper lanes. | ||
| 334 | 1628 | svuint32_t acc_0_8_lo = svaddlb_u32(src[0], src[8]); | |
| 335 | 1628 | svuint32_t acc_0_8_hi = svaddlt_u32(src[0], src[8]); | |
| 336 | |||
| 337 | 1628 | svuint32_t acc_1_7_lo = svaddlb_u32(src[1], src[7]); | |
| 338 | 1628 | svuint32_t acc_1_7_hi = svaddlt_u32(src[1], src[7]); | |
| 339 | |||
| 340 | 1628 | svuint32_t acc_2_6_lo = svaddlb_u32(src[2], src[6]); | |
| 341 | 1628 | svuint32_t acc_2_6_hi = svaddlt_u32(src[2], src[6]); | |
| 342 | |||
| 343 | 1628 | svuint32_t acc_3_5_lo = svaddlb_u32(src[3], src[5]); | |
| 344 | 1628 | svuint32_t acc_3_5_hi = svaddlt_u32(src[3], src[5]); | |
| 345 | |||
| 346 | 1628 | svuint32_t acc_4_lo = svmovlb_u32(src[4]); | |
| 347 | 1628 | svuint32_t acc_4_hi = svmovlt_u32(src[4]); | |
| 348 | |||
| 349 | // Window-level grouping: *_tap_even/*_tap_odd are even/odd taps | ||
| 350 | // (0, 2, 4, 6, 8 vs 1, 3, 5, 7). | ||
| 351 | 1628 | svuint32_t acc_lo_tap_even = svlsl_n_u32_x(pg, acc_0_8_lo, 2); | |
| 352 | 1628 | svuint32_t acc_hi_tap_even = svlsl_n_u32_x(pg, acc_0_8_hi, 2); | |
| 353 | 1628 | acc_lo_tap_even = svmla_n_u32_x(pg, acc_lo_tap_even, acc_2_6_lo, 30); | |
| 354 | 1628 | acc_hi_tap_even = svmla_n_u32_x(pg, acc_hi_tap_even, acc_2_6_hi, 30); | |
| 355 | 1628 | acc_lo_tap_even = svmla_n_u32_x(pg, acc_lo_tap_even, acc_4_lo, 60); | |
| 356 | 1628 | acc_hi_tap_even = svmla_n_u32_x(pg, acc_hi_tap_even, acc_4_hi, 60); | |
| 357 | |||
| 358 | 1628 | svuint32_t acc_lo_tap_odd = svmul_n_u32_x(pg, acc_1_7_lo, 13); | |
| 359 | 1628 | svuint32_t acc_hi_tap_odd = svmul_n_u32_x(pg, acc_1_7_hi, 13); | |
| 360 | 1628 | acc_lo_tap_odd = svmla_n_u32_x(pg, acc_lo_tap_odd, acc_3_5_lo, 51); | |
| 361 | 1628 | acc_hi_tap_odd = svmla_n_u32_x(pg, acc_hi_tap_odd, acc_3_5_hi, 51); | |
| 362 | |||
| 363 | 1628 | svuint32_t acc_lo = svadd_u32_x(pg, acc_lo_tap_even, acc_lo_tap_odd); | |
| 364 | 1628 | svuint32_t acc_hi = svadd_u32_x(pg, acc_hi_tap_even, acc_hi_tap_odd); | |
| 365 | |||
| 366 | 1628 | svuint16_t acc_u16_lo = svrshrnb_n_u32(acc_lo, 16); | |
| 367 | 1628 | svuint16_t acc_u16 = svrshrnt_n_u32(acc_u16_lo, acc_hi, 16); | |
| 368 | |||
| 369 | 1628 | svst1b(pg, &dst[0], acc_u16); | |
| 370 | 1628 | } | |
| 371 | |||
| 372 | // Applies horizontal filtering vector using scalar operations. | ||
| 373 | // | ||
| 374 | // DST = 1/65536 * [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6, SRC7, SRC8 ] * | ||
| 375 | // * [ 4, 13, 30, 51, 60, 51, 30, 13, 4 ]T | ||
| 376 | 13632 | void horizontal_scalar_path(const BufferType src[9], | |
| 377 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 378 | 40896 | uint32_t acc = src[0] * 4 + src[1] * 13 + src[2] * 30 + src[3] * 51 + | |
| 379 | 40896 | src[4] * 60 + src[5] * 51 + src[6] * 30 + src[7] * 13 + | |
| 380 | 13632 | src[8] * 4; | |
| 381 | 13632 | dst[0] = rounding_shift_right(acc, 16); | |
| 382 | 13632 | } | |
| 383 | }; // end of class GaussianBlur<uint8_t, 9, true> | ||
| 384 | |||
| 385 | // CustomSigma variant | ||
| 386 | template <size_t KernelSize> | ||
| 387 | class GaussianBlur<uint8_t, KernelSize, false> { | ||
| 388 | public: | ||
| 389 | using SourceType = uint8_t; | ||
| 390 | using BufferType = uint8_t; | ||
| 391 | using DestinationType = uint8_t; | ||
| 392 | using SourceVecTraits = | ||
| 393 | typename ::KLEIDICV_TARGET_NAMESPACE::VecTraits<SourceType>; | ||
| 394 | using SourceVectorType = typename SourceVecTraits::VectorType; | ||
| 395 | |||
| 396 | static constexpr size_t kHalfKernelSize = get_half_kernel_size(KernelSize); | ||
| 397 | |||
| 398 | 396 | explicit GaussianBlur(const uint8_t *half_kernel) | |
| 399 | 396 | : half_kernel_(half_kernel) {} | |
| 400 | |||
| 401 | 9100 | void vertical_vector_path( | |
| 402 | svbool_t pg, std::reference_wrapper<SourceVectorType> src[KernelSize], | ||
| 403 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 404 | 9100 | common_vector_path(pg, src, dst); | |
| 405 | 9100 | } | |
| 406 | |||
| 407 | 111624 | void vertical_scalar_path(const SourceType src[KernelSize], | |
| 408 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 409 | 111624 | uint32_t acc = src[kHalfKernelSize - 1] * half_kernel_[kHalfKernelSize - 1]; | |
| 410 | |||
| 411 | // Optimization to avoid unnecessary branching in vector code. | ||
| 412 | KLEIDICV_FORCE_LOOP_UNROLL | ||
| 413 |
12/12✓ Branch 0 taken 33264 times.
✓ Branch 1 taken 232848 times.
✓ Branch 2 taken 67680 times.
✓ Branch 3 taken 676800 times.
✓ Branch 4 taken 372 times.
✓ Branch 5 taken 372 times.
✓ Branch 6 taken 1440 times.
✓ Branch 7 taken 2880 times.
✓ Branch 8 taken 3204 times.
✓ Branch 9 taken 9612 times.
✓ Branch 10 taken 5664 times.
✓ Branch 11 taken 22656 times.
|
1056792 | for (size_t i = 0; i < kHalfKernelSize - 1; i++) { |
| 414 | 945168 | acc += (src[i] + src[KernelSize - i - 1]) * half_kernel_[i]; | |
| 415 | 945168 | } | |
| 416 | |||
| 417 | 111624 | dst[0] = static_cast<BufferType>(rounding_shift_right(acc, 8)); | |
| 418 | 111624 | } | |
| 419 | |||
| 420 | 6432 | void horizontal_vector_path( | |
| 421 | svbool_t pg, std::reference_wrapper<SourceVectorType> src[KernelSize], | ||
| 422 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 423 | 6432 | common_vector_path(pg, src, dst); | |
| 424 | 6432 | } | |
| 425 | |||
| 426 | 111624 | void horizontal_scalar_path(const BufferType src[KernelSize], | |
| 427 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 428 | 111624 | vertical_scalar_path(src, dst); | |
| 429 | 111624 | } | |
| 430 | |||
| 431 | private: | ||
| 432 | 15532 | void common_vector_path( | |
| 433 | svbool_t pg, std::reference_wrapper<SourceVectorType> src[KernelSize], | ||
| 434 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 435 | 15532 | svbool_t pg16_all = svptrue_b16(); | |
| 436 | 31064 | svuint16_t acc_b = svmullb_n_u16(src[kHalfKernelSize - 1], | |
| 437 | 15532 | half_kernel_[kHalfKernelSize - 1]); | |
| 438 | 31064 | svuint16_t acc_t = svmullt_n_u16(src[kHalfKernelSize - 1], | |
| 439 | 15532 | half_kernel_[kHalfKernelSize - 1]); | |
| 440 | |||
| 441 | // Optimization to avoid unnecessary branching in vector code. | ||
| 442 | KLEIDICV_FORCE_LOOP_UNROLL | ||
| 443 |
12/12✓ Branch 0 taken 4752 times.
✓ Branch 1 taken 33264 times.
✓ Branch 2 taken 7416 times.
✓ Branch 3 taken 74160 times.
✓ Branch 4 taken 342 times.
✓ Branch 5 taken 342 times.
✓ Branch 6 taken 660 times.
✓ Branch 7 taken 1320 times.
✓ Branch 8 taken 1002 times.
✓ Branch 9 taken 3006 times.
✓ Branch 10 taken 1360 times.
✓ Branch 11 taken 5440 times.
|
133064 | for (size_t i = 0; i < kHalfKernelSize - 1; i++) { |
| 444 | 117532 | const size_t j = KernelSize - i - 1; | |
| 445 | 117532 | svuint16_t vec_b = svaddlb_u16(src[i], src[j]); | |
| 446 | 117532 | svuint16_t vec_t = svaddlt_u16(src[i], src[j]); | |
| 447 | |||
| 448 | 117532 | acc_b = svmla_n_u16_x(pg16_all, acc_b, vec_b, half_kernel_[i]); | |
| 449 | 117532 | acc_t = svmla_n_u16_x(pg16_all, acc_t, vec_t, half_kernel_[i]); | |
| 450 | 117532 | } | |
| 451 | |||
| 452 | // Rounding before narrowing | ||
| 453 | 15532 | acc_b = svqadd_n_u16(acc_b, 128); | |
| 454 | 15532 | acc_t = svqadd_n_u16(acc_t, 128); | |
| 455 | // Keep only the highest 8 bits | ||
| 456 | 31064 | svuint8_t result = | |
| 457 | 15532 | svtrn2_u8(svreinterpret_u8_u16(acc_b), svreinterpret_u8_u16(acc_t)); | |
| 458 | 15532 | svst1(pg, &dst[0], result); | |
| 459 | 15532 | } | |
| 460 | |||
| 461 | const uint8_t *half_kernel_; | ||
| 462 | }; // end of class GaussianBlur<uint8_t, KernelSize, false> | ||
| 463 | |||
| 464 | template <size_t KernelSize, bool IsBinomial, typename ScalarType> | ||
| 465 | 1290 | static kleidicv_error_t gaussian_blur_fixed_kernel_size( | |
| 466 | const ScalarType *src, size_t src_stride, ScalarType *dst, | ||
| 467 | size_t dst_stride, Rectangle &rect, size_t y_begin, size_t y_end, | ||
| 468 | size_t channels, float sigma, | ||
| 469 | FixedBorderType border_type) KLEIDICV_STREAMING { | ||
| 470 | using GaussianBlurFilter = GaussianBlur<ScalarType, KernelSize, IsBinomial>; | ||
| 471 | 1290 | constexpr size_t intermediate_size{ | |
| 472 | sizeof(typename GaussianBlurFilter::BufferType)}; | ||
| 473 | |||
| 474 | 1290 | auto workspace_variant = | |
| 475 | 1290 | SeparableFilterWorkspace::create(rect, channels, intermediate_size); | |
| 476 |
40/40✓ Branch 0 taken 3 times.
✓ Branch 1 taken 141 times.
✓ Branch 2 taken 3 times.
✓ Branch 3 taken 141 times.
✓ Branch 4 taken 3 times.
✓ Branch 5 taken 183 times.
✓ Branch 6 taken 3 times.
✓ Branch 7 taken 183 times.
✓ Branch 8 taken 3 times.
✓ Branch 9 taken 99 times.
✓ Branch 10 taken 3 times.
✓ Branch 11 taken 99 times.
✓ Branch 12 taken 3 times.
✓ Branch 13 taken 153 times.
✓ Branch 14 taken 3 times.
✓ Branch 15 taken 153 times.
✓ Branch 16 taken 3 times.
✓ Branch 17 taken 144 times.
✓ Branch 18 taken 3 times.
✓ Branch 19 taken 144 times.
✓ Branch 20 taken 3 times.
✓ Branch 21 taken 144 times.
✓ Branch 22 taken 3 times.
✓ Branch 23 taken 144 times.
✓ Branch 24 taken 3 times.
✓ Branch 25 taken 99 times.
✓ Branch 26 taken 3 times.
✓ Branch 27 taken 99 times.
✓ Branch 28 taken 3 times.
✓ Branch 29 taken 99 times.
✓ Branch 30 taken 3 times.
✓ Branch 31 taken 99 times.
✓ Branch 32 taken 3 times.
✓ Branch 33 taken 99 times.
✓ Branch 34 taken 3 times.
✓ Branch 35 taken 99 times.
✓ Branch 36 taken 3 times.
✓ Branch 37 taken 99 times.
✓ Branch 38 taken 3 times.
✓ Branch 39 taken 99 times.
|
1320 | if (auto *err = std::get_if<kleidicv_error_t>(&workspace_variant)) { |
| 477 | 30 | return *err; | |
| 478 | } | ||
| 479 | 1260 | auto &workspace = *std::get_if<SeparableFilterWorkspace>(&workspace_variant); | |
| 480 | |||
| 481 | 1260 | Rows<const ScalarType> src_rows{src, src_stride, channels}; | |
| 482 | 1260 | Rows<ScalarType> dst_rows{dst, dst_stride, channels}; | |
| 483 | |||
| 484 | if constexpr (IsBinomial) { | ||
| 485 | 576 | GaussianBlurFilter blur; | |
| 486 | 576 | SeparableFilter<GaussianBlurFilter, KernelSize> filter{blur}; | |
| 487 | 576 | workspace.process(y_begin, y_end, src_rows, dst_rows, border_type, filter); | |
| 488 | |||
| 489 | 576 | return KLEIDICV_OK; | |
| 490 | 576 | } else { | |
| 491 | 684 | constexpr size_t kHalfKernelSize = get_half_kernel_size(KernelSize); | |
| 492 | 684 | uint8_t half_kernel[128]; | |
| 493 | 1368 | bool success = | |
| 494 | 684 | generate_gaussian_half_kernel(half_kernel, kHalfKernelSize, sigma); | |
| 495 |
12/12✓ Branch 0 taken 96 times.
✓ Branch 1 taken 48 times.
✓ Branch 2 taken 96 times.
✓ Branch 3 taken 48 times.
✓ Branch 4 taken 51 times.
✓ Branch 5 taken 48 times.
✓ Branch 6 taken 51 times.
✓ Branch 7 taken 48 times.
✓ Branch 8 taken 51 times.
✓ Branch 9 taken 48 times.
✓ Branch 10 taken 51 times.
✓ Branch 11 taken 48 times.
|
684 | if (success) { |
| 496 | 396 | GaussianBlurFilter blur(half_kernel); | |
| 497 | 396 | SeparableFilter<GaussianBlurFilter, KernelSize> filter{blur}; | |
| 498 | 792 | workspace.process(y_begin, y_end, src_rows, dst_rows, border_type, | |
| 499 | 396 | filter); | |
| 500 | 396 | } else { | |
| 501 | // Sigma is too small that the middle point would get all the weight | ||
| 502 | // => it's just a copy. | ||
| 503 |
12/12✓ Branch 0 taken 684 times.
✓ Branch 1 taken 48 times.
✓ Branch 2 taken 972 times.
✓ Branch 3 taken 48 times.
✓ Branch 4 taken 108 times.
✓ Branch 5 taken 48 times.
✓ Branch 6 taken 204 times.
✓ Branch 7 taken 48 times.
✓ Branch 8 taken 300 times.
✓ Branch 9 taken 48 times.
✓ Branch 10 taken 396 times.
✓ Branch 11 taken 48 times.
|
2952 | for (size_t row = y_begin; row < y_end; ++row) { |
| 504 | #if KLEIDICV_TARGET_SME && defined(__ANDROID__) | ||
| 505 | __arm_sc_memcpy( | ||
| 506 | static_cast<void *>(&dst_rows.at(row)[0]), | ||
| 507 | static_cast<const void *>(&src_rows.at(row)[0]), | ||
| 508 | rect.width() * sizeof(ScalarType) * dst_rows.channels()); | ||
| 509 | #else | ||
| 510 | 5328 | std::memcpy(static_cast<void *>(&dst_rows.at(row)[0]), | |
| 511 | 2664 | static_cast<const void *>(&src_rows.at(row)[0]), | |
| 512 | 2664 | rect.width() * sizeof(ScalarType) * dst_rows.channels()); | |
| 513 | #endif | ||
| 514 | 2664 | } | |
| 515 | } | ||
| 516 | 684 | return KLEIDICV_OK; | |
| 517 | 684 | } | |
| 518 | 1290 | } | |
| 519 | |||
| 520 | template <bool IsBinomial, typename ScalarType> | ||
| 521 | 1290 | static kleidicv_error_t gaussian_blur( | |
| 522 | size_t kernel_size, const ScalarType *src, size_t src_stride, | ||
| 523 | ScalarType *dst, size_t dst_stride, Rectangle &rect, size_t y_begin, | ||
| 524 | size_t y_end, size_t channels, float sigma, | ||
| 525 | FixedBorderType border_type) KLEIDICV_STREAMING { | ||
| 526 |
12/14✓ Branch 0 taken 144 times.
✓ Branch 1 taken 186 times.
✓ Branch 2 taken 102 times.
✓ Branch 3 taken 156 times.
✓ Branch 4 taken 48 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 48 times.
✓ Branch 7 taken 102 times.
✓ Branch 8 taken 102 times.
✓ Branch 9 taken 102 times.
✓ Branch 10 taken 102 times.
✓ Branch 11 taken 99 times.
✗ Branch 12 not taken.
✓ Branch 13 taken 99 times.
|
1290 | switch (kernel_size) { |
| 527 | case 3: | ||
| 528 | 246 | return gaussian_blur_fixed_kernel_size<3, IsBinomial>( | |
| 529 | 246 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 530 | 246 | sigma, border_type); | |
| 531 | case 5: | ||
| 532 | 288 | return gaussian_blur_fixed_kernel_size<5, IsBinomial>( | |
| 533 | 288 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 534 | 288 | sigma, border_type); | |
| 535 | case 7: | ||
| 536 | 204 | return gaussian_blur_fixed_kernel_size<7, IsBinomial>( | |
| 537 | 204 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 538 | 204 | sigma, border_type); | |
| 539 | case 9: | ||
| 540 | 258 | return gaussian_blur_fixed_kernel_size<9, IsBinomial>( | |
| 541 | 258 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 542 | 258 | sigma, border_type); | |
| 543 | case 15: | ||
| 544 | // 15x15 does not have a binomial variant | ||
| 545 | 147 | return gaussian_blur_fixed_kernel_size<15, false>( | |
| 546 | 147 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 547 | 147 | sigma, border_type); | |
| 548 | case 21: | ||
| 549 | // 21x21 does not have a binomial variant | ||
| 550 | 147 | return gaussian_blur_fixed_kernel_size<21, false>( | |
| 551 | 147 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 552 | 147 | sigma, border_type); | |
| 553 | // gaussian_blur_is_implemented checked the kernel size already. | ||
| 554 | // GCOVR_EXCL_START | ||
| 555 | default: | ||
| 556 | assert(!"kernel size not implemented"); | ||
| 557 | − | return KLEIDICV_ERROR_NOT_IMPLEMENTED; | |
| 558 | // GCOVR_EXCL_STOP | ||
| 559 | } | ||
| 560 | 1290 | } | |
| 561 | |||
| 562 | 1320 | static kleidicv_error_t gaussian_blur_fixed_stripe_u8_sc( | |
| 563 | const uint8_t *src, size_t src_stride, uint8_t *dst, size_t dst_stride, | ||
| 564 | size_t width, size_t height, size_t y_begin, size_t y_end, size_t channels, | ||
| 565 | size_t kernel_width, size_t /*kernel_height*/, float sigma_x, | ||
| 566 | float /*sigma_y*/, FixedBorderType fixed_border_type) KLEIDICV_STREAMING { | ||
| 567 |
6/6✓ Branch 0 taken 30 times.
✓ Branch 1 taken 1290 times.
✓ Branch 2 taken 30 times.
✓ Branch 3 taken 1290 times.
✓ Branch 4 taken 30 times.
✓ Branch 5 taken 1290 times.
|
3960 | if (auto result = |
| 568 | 1320 | gaussian_blur_checks(src, src_stride, dst, dst_stride, width, height); | |
| 569 |
2/3✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
✓ Branch 2 taken 1290 times.
|
1350 | result != KLEIDICV_OK) { |
| 570 | 30 | return result; | |
| 571 | } | ||
| 572 | |||
| 573 | 1290 | Rectangle rect{width, height}; | |
| 574 | |||
| 575 |
2/2✓ Branch 0 taken 684 times.
✓ Branch 1 taken 606 times.
|
1290 | if (sigma_x == 0.0) { |
| 576 | 1368 | return gaussian_blur<true>(kernel_width, src, src_stride, dst, dst_stride, | |
| 577 | 684 | rect, y_begin, y_end, channels, sigma_x, | |
| 578 | 684 | fixed_border_type); | |
| 579 | } | ||
| 580 | |||
| 581 | 1212 | return gaussian_blur<false>(kernel_width, src, src_stride, dst, dst_stride, | |
| 582 | 606 | rect, y_begin, y_end, channels, sigma_x, | |
| 583 | 606 | fixed_border_type); | |
| 584 | 1320 | } | |
| 585 | |||
| 586 | } // namespace KLEIDICV_TARGET_NAMESPACE | ||
| 587 | |||
| 588 | #endif // KLEIDICV_GAUSSIAN_BLUR_SC_H | ||
| 589 |