| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // SPDX-FileCopyrightText: 2023 - 2025 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/sigma.h" | ||
| 18 | #include "kleidicv/workspace/separable.h" | ||
| 19 | |||
| 20 | #if KLEIDICV_TARGET_SME || KLEIDICV_TARGET_SME2 | ||
| 21 | #include <arm_sme.h> | ||
| 22 | #endif | ||
| 23 | |||
| 24 | namespace KLEIDICV_TARGET_NAMESPACE { | ||
| 25 | |||
| 26 | // Primary template for Gaussian Blur filters. | ||
| 27 | template <typename ScalarType, size_t KernelSize, bool IsBinomial> | ||
| 28 | class GaussianBlur; | ||
| 29 | |||
| 30 | // Template for 3x3 Gaussian Blur binomial filters. | ||
| 31 | // | ||
| 32 | // [ 1, 2, 1 ] [ 1 ] | ||
| 33 | // F = 1/16 * [ 2, 4, 2 ] = 1/16 * [ 2 ] * [ 1, 2, 1 ] | ||
| 34 | // [ 1, 2, 1 ] [ 1 ] | ||
| 35 | template <> | ||
| 36 | class GaussianBlur<uint8_t, 3, true> { | ||
| 37 | public: | ||
| 38 | using SourceType = uint8_t; | ||
| 39 | using BufferType = uint16_t; | ||
| 40 | using DestinationType = uint8_t; | ||
| 41 | |||
| 42 | // Applies vertical filtering vector using SIMD operations. | ||
| 43 | // | ||
| 44 | // DST = [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 45 | 504 | void vertical_vector_path(svbool_t pg, | |
| 46 | std::reference_wrapper<svuint8_t> src[3], | ||
| 47 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 48 | 504 | svuint16_t acc_0_2_b = svaddlb_u16(src[0], src[2]); | |
| 49 | 504 | svuint16_t acc_0_2_t = svaddlt_u16(src[0], src[2]); | |
| 50 | |||
| 51 | 504 | svuint16_t acc_1_b = svshllb_n_u16(src[1], 1); | |
| 52 | 504 | svuint16_t acc_1_t = svshllt_n_u16(src[1], 1); | |
| 53 | |||
| 54 | 504 | svuint16_t acc_u16_b = svadd_u16_x(pg, acc_0_2_b, acc_1_b); | |
| 55 | 504 | svuint16_t acc_u16_t = svadd_u16_x(pg, acc_0_2_t, acc_1_t); | |
| 56 | |||
| 57 | 504 | svuint16x2_t interleaved = svcreate2(acc_u16_b, acc_u16_t); | |
| 58 | 504 | svst2(pg, &dst[0], interleaved); | |
| 59 | 504 | } | |
| 60 | |||
| 61 | // Applies horizontal filtering vector using SIMD operations. | ||
| 62 | // | ||
| 63 | // DST = 1/16 * [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 64 | 400 | void horizontal_vector_path(svbool_t pg, | |
| 65 | std::reference_wrapper<svuint16_t> src[3], | ||
| 66 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 67 | 400 | svuint16_t acc_0_2 = svhadd_u16_x(pg, src[0], src[2]); | |
| 68 | |||
| 69 | 400 | svuint16_t acc = svadd_u16_x(pg, acc_0_2, src[1]); | |
| 70 | 400 | acc = svrshr_x(pg, acc, 3); | |
| 71 | |||
| 72 | 400 | svst1b(pg, &dst[0], acc); | |
| 73 | 400 | } | |
| 74 | |||
| 75 | // Applies horizontal filtering vector using scalar operations. | ||
| 76 | // | ||
| 77 | // DST = 1/16 * [ SRC0, SRC1, SRC2 ] * [ 1, 2, 1 ]T | ||
| 78 | 1008 | void horizontal_scalar_path(const BufferType src[3], | |
| 79 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 80 | 1008 | auto acc = src[0] + 2 * src[1] + src[2]; | |
| 81 | 1008 | dst[0] = rounding_shift_right(acc, 4); | |
| 82 | 1008 | } | |
| 83 | }; // end of class GaussianBlur<uint8_t, 3, true> | ||
| 84 | |||
| 85 | // Template for 5x5 Gaussian Blur binomial filters. | ||
| 86 | // | ||
| 87 | // [ 1, 4, 6, 4, 1 ] [ 1 ] | ||
| 88 | // [ 4, 16, 24, 16, 4 ] [ 4 ] | ||
| 89 | // F = 1/256 * [ 6, 24, 36, 24, 6 ] = 1/256 * [ 6 ] * [ 1, 4, 6, 4, 1 ] | ||
| 90 | // [ 4, 16, 24, 16, 4 ] [ 4 ] | ||
| 91 | // [ 1, 4, 6, 4, 1 ] [ 1 ] | ||
| 92 | template <> | ||
| 93 | class GaussianBlur<uint8_t, 5, true> { | ||
| 94 | public: | ||
| 95 | using SourceType = uint8_t; | ||
| 96 | using BufferType = uint16_t; | ||
| 97 | using DestinationType = uint8_t; | ||
| 98 | |||
| 99 | // Applies vertical filtering vector using SIMD operations. | ||
| 100 | // | ||
| 101 | // DST = [ SRC0, SRC1, SRC2, SRC3, SRC4 ] * [ 1, 4, 6, 4, 1 ]T | ||
| 102 | 1524 | void vertical_vector_path(svbool_t pg, | |
| 103 | std::reference_wrapper<svuint8_t> src[5], | ||
| 104 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 105 | 1524 | svuint16_t acc_0_4_b = svaddlb_u16(src[0], src[4]); | |
| 106 | 1524 | svuint16_t acc_0_4_t = svaddlt_u16(src[0], src[4]); | |
| 107 | 1524 | svuint16_t acc_1_3_b = svaddlb_u16(src[1], src[3]); | |
| 108 | 1524 | svuint16_t acc_1_3_t = svaddlt_u16(src[1], src[3]); | |
| 109 | |||
| 110 | 1524 | svuint16_t acc_u16_b = svmlalb_n_u16(acc_0_4_b, src[2], 6); | |
| 111 | 1524 | svuint16_t acc_u16_t = svmlalt_n_u16(acc_0_4_t, src[2], 6); | |
| 112 | 1524 | acc_u16_b = svmla_n_u16_x(pg, acc_u16_b, acc_1_3_b, 4); | |
| 113 | 1524 | acc_u16_t = svmla_n_u16_x(pg, acc_u16_t, acc_1_3_t, 4); | |
| 114 | |||
| 115 | 1524 | svuint16x2_t interleaved = svcreate2(acc_u16_b, acc_u16_t); | |
| 116 | 1524 | svst2(pg, &dst[0], interleaved); | |
| 117 | 1524 | } | |
| 118 | |||
| 119 | // Applies horizontal filtering vector using SIMD operations. | ||
| 120 | // | ||
| 121 | // DST = 1/256 * [ SRC0, SRC1, SRC2, SRC3, SRC4 ] * [ 1, 4, 6, 4, 1 ]T | ||
| 122 | 1316 | void horizontal_vector_path(svbool_t pg, | |
| 123 | std::reference_wrapper<svuint16_t> src[5], | ||
| 124 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 125 | 1316 | svuint16_t acc_0_4 = svadd_x(pg, src[0], src[4]); | |
| 126 | 1316 | svuint16_t acc_1_3 = svadd_x(pg, src[1], src[3]); | |
| 127 | 1316 | svuint16_t acc = svmla_n_u16_x(pg, acc_0_4, src[2], 6); | |
| 128 | 1316 | acc = svmla_n_u16_x(pg, acc, acc_1_3, 4); | |
| 129 | 1316 | acc = svrshr_x(pg, acc, 8); | |
| 130 | 1316 | svst1b(pg, &dst[0], acc); | |
| 131 | 1316 | } | |
| 132 | |||
| 133 | // Applies horizontal filtering vector using scalar operations. | ||
| 134 | // | ||
| 135 | // DST = 1/256 * [ SRC0, SRC1, SRC2, SRC3, SRC4 ] * [ 1, 4, 6, 4, 1 ]T | ||
| 136 | 5552 | void horizontal_scalar_path(const BufferType src[5], | |
| 137 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 138 | 5552 | auto acc = src[0] + src[4] + 4 * (src[1] + src[3]) + 6 * src[2]; | |
| 139 | 5552 | dst[0] = rounding_shift_right(acc, 8); | |
| 140 | 5552 | } | |
| 141 | }; // end of class GaussianBlur<uint8_t, 5, true> | ||
| 142 | |||
| 143 | // Template for 7x7 Gaussian Blur binomial filters. | ||
| 144 | // | ||
| 145 | // [ 4, 14, 28, 36, 28, 14, 4 ] | ||
| 146 | // [ 14, 49, 98, 126, 98, 49, 14 ] | ||
| 147 | // [ 28, 98, 196, 252, 196, 98, 28 ] | ||
| 148 | // F = 1/4096 * [ 36, 126, 252, 324, 252, 126, 36 ] = | ||
| 149 | // [ 28, 98, 196, 252, 196, 98, 28 ] | ||
| 150 | // [ 14, 49, 98, 126, 98, 49, 14 ] | ||
| 151 | // [ 4, 14, 28, 36, 28, 14, 4 ] | ||
| 152 | // | ||
| 153 | // [ 2 ] | ||
| 154 | // [ 7 ] | ||
| 155 | // [ 14 ] | ||
| 156 | // = 1/4096 * [ 18 ] * [ 2, 7, 14, 18, 14, 7, 2 ] | ||
| 157 | // [ 14 ] | ||
| 158 | // [ 7 ] | ||
| 159 | // [ 2 ] | ||
| 160 | template <> | ||
| 161 | class GaussianBlur<uint8_t, 7, true> { | ||
| 162 | public: | ||
| 163 | using SourceType = uint8_t; | ||
| 164 | using BufferType = uint16_t; | ||
| 165 | using DestinationType = uint8_t; | ||
| 166 | |||
| 167 | // Applies vertical filtering vector using SIMD operations. | ||
| 168 | // | ||
| 169 | // DST = [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6 ] * | ||
| 170 | // * [ 2, 7, 14, 18, 14, 7, 2 ]T | ||
| 171 | 596 | void vertical_vector_path(svbool_t pg, | |
| 172 | std::reference_wrapper<svuint8_t> src[7], | ||
| 173 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 174 | 596 | svuint16_t acc_0_6_b = svaddlb_u16(src[0], src[6]); | |
| 175 | 596 | svuint16_t acc_0_6_t = svaddlt_u16(src[0], src[6]); | |
| 176 | |||
| 177 | 596 | svuint16_t acc_1_5_b = svaddlb_u16(src[1], src[5]); | |
| 178 | 596 | svuint16_t acc_1_5_t = svaddlt_u16(src[1], src[5]); | |
| 179 | |||
| 180 | 596 | svuint16_t acc_2_4_b = svaddlb_u16(src[2], src[4]); | |
| 181 | 596 | svuint16_t acc_2_4_t = svaddlt_u16(src[2], src[4]); | |
| 182 | |||
| 183 | 596 | svuint16_t acc_3_b = svmovlb_u16(src[3]); | |
| 184 | 596 | svuint16_t acc_3_t = svmovlt_u16(src[3]); | |
| 185 | |||
| 186 | 596 | svuint16_t acc_0_2_4_6_b = svmla_n_u16_x(pg, acc_0_6_b, acc_2_4_b, 7); | |
| 187 | 596 | svuint16_t acc_0_2_4_6_t = svmla_n_u16_x(pg, acc_0_6_t, acc_2_4_t, 7); | |
| 188 | |||
| 189 | 596 | svuint16_t acc_0_2_3_4_6_b = svmla_n_u16_x(pg, acc_0_2_4_6_b, acc_3_b, 9); | |
| 190 | 596 | svuint16_t acc_0_2_3_4_6_t = svmla_n_u16_x(pg, acc_0_2_4_6_t, acc_3_t, 9); | |
| 191 | 596 | acc_0_2_3_4_6_b = svlsl_n_u16_x(pg, acc_0_2_3_4_6_b, 1); | |
| 192 | 596 | acc_0_2_3_4_6_t = svlsl_n_u16_x(pg, acc_0_2_3_4_6_t, 1); | |
| 193 | |||
| 194 | 1192 | svuint16_t acc_0_1_2_3_4_5_6_b = | |
| 195 | 596 | svmla_n_u16_x(pg, acc_0_2_3_4_6_b, acc_1_5_b, 7); | |
| 196 | 1192 | svuint16_t acc_0_1_2_3_4_5_6_t = | |
| 197 | 596 | svmla_n_u16_x(pg, acc_0_2_3_4_6_t, acc_1_5_t, 7); | |
| 198 | |||
| 199 | 1192 | svuint16x2_t interleaved = | |
| 200 | 596 | svcreate2(acc_0_1_2_3_4_5_6_b, acc_0_1_2_3_4_5_6_t); | |
| 201 | 596 | svst2(pg, &dst[0], interleaved); | |
| 202 | 596 | } | |
| 203 | |||
| 204 | // Applies horizontal filtering vector using SIMD operations. | ||
| 205 | // | ||
| 206 | // DST = 1/4096 * [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6 ] * | ||
| 207 | // * [ 2, 7, 14, 18, 14, 7, 2 ]T | ||
| 208 | 464 | void horizontal_vector_path(svbool_t pg, | |
| 209 | std::reference_wrapper<svuint16_t> src[7], | ||
| 210 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 211 | 464 | svuint32_t acc_0_6_b = svaddlb_u32(src[0], src[6]); | |
| 212 | 464 | svuint32_t acc_0_6_t = svaddlt_u32(src[0], src[6]); | |
| 213 | |||
| 214 | 464 | svuint32_t acc_1_5_b = svaddlb_u32(src[1], src[5]); | |
| 215 | 464 | svuint32_t acc_1_5_t = svaddlt_u32(src[1], src[5]); | |
| 216 | |||
| 217 | 464 | svuint16_t acc_2_4 = svadd_u16_x(pg, src[2], src[4]); | |
| 218 | |||
| 219 | 464 | svuint32_t acc_0_2_4_6_b = svmlalb_n_u32(acc_0_6_b, acc_2_4, 7); | |
| 220 | 464 | svuint32_t acc_0_2_4_6_t = svmlalt_n_u32(acc_0_6_t, acc_2_4, 7); | |
| 221 | |||
| 222 | 464 | svuint32_t acc_0_2_3_4_6_b = svmlalb_n_u32(acc_0_2_4_6_b, src[3], 9); | |
| 223 | 464 | svuint32_t acc_0_2_3_4_6_t = svmlalt_n_u32(acc_0_2_4_6_t, src[3], 9); | |
| 224 | |||
| 225 | 464 | acc_0_2_3_4_6_b = svlsl_n_u32_x(pg, acc_0_2_3_4_6_b, 1); | |
| 226 | 464 | acc_0_2_3_4_6_t = svlsl_n_u32_x(pg, acc_0_2_3_4_6_t, 1); | |
| 227 | |||
| 228 | 928 | svuint32_t acc_0_1_2_3_4_5_6_b = | |
| 229 | 464 | svmla_n_u32_x(pg, acc_0_2_3_4_6_b, acc_1_5_b, 7); | |
| 230 | 928 | svuint32_t acc_0_1_2_3_4_5_6_t = | |
| 231 | 464 | svmla_n_u32_x(pg, acc_0_2_3_4_6_t, acc_1_5_t, 7); | |
| 232 | |||
| 233 | 928 | svuint16_t acc_0_1_2_3_4_5_6_u16_b = | |
| 234 | 464 | svrshrnb_n_u32(acc_0_1_2_3_4_5_6_b, 12); | |
| 235 | 928 | svuint16_t acc_0_1_2_3_4_5_6_u16 = | |
| 236 | 464 | svrshrnt_n_u32(acc_0_1_2_3_4_5_6_u16_b, acc_0_1_2_3_4_5_6_t, 12); | |
| 237 | |||
| 238 | 464 | svst1b(pg, &dst[0], acc_0_1_2_3_4_5_6_u16); | |
| 239 | 464 | } | |
| 240 | |||
| 241 | // Applies horizontal filtering vector using scalar operations. | ||
| 242 | // | ||
| 243 | // DST = 1/4096 * [ SRC0, SRC1, SRC2, SRC3, SRC4, SRC5, SRC6 ] * | ||
| 244 | // * [ 2, 7, 14, 18, 14, 7, 2 ]T | ||
| 245 | 3288 | void horizontal_scalar_path(const BufferType src[7], | |
| 246 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 247 | 9864 | uint32_t acc = src[0] * 2 + src[1] * 7 + src[2] * 14 + src[3] * 18 + | |
| 248 | 6576 | src[4] * 14 + src[5] * 7 + src[6] * 2; | |
| 249 | 3288 | dst[0] = rounding_shift_right(acc, 12); | |
| 250 | 3288 | } | |
| 251 | }; // end of class GaussianBlur<uint8_t, 7, true> | ||
| 252 | |||
| 253 | // CustomSigma variant | ||
| 254 | template <size_t KernelSize> | ||
| 255 | class GaussianBlur<uint8_t, KernelSize, false> { | ||
| 256 | public: | ||
| 257 | using SourceType = uint8_t; | ||
| 258 | using BufferType = uint8_t; | ||
| 259 | using DestinationType = uint8_t; | ||
| 260 | using SourceVecTraits = | ||
| 261 | typename ::KLEIDICV_TARGET_NAMESPACE::VecTraits<SourceType>; | ||
| 262 | using SourceVectorType = typename SourceVecTraits::VectorType; | ||
| 263 | |||
| 264 | static constexpr size_t kHalfKernelSize = get_half_kernel_size(KernelSize); | ||
| 265 | |||
| 266 | 230 | explicit GaussianBlur(const uint16_t *half_kernel) | |
| 267 | 230 | : half_kernel_(half_kernel) {} | |
| 268 | |||
| 269 | 6048 | void vertical_vector_path( | |
| 270 | svbool_t pg, std::reference_wrapper<SourceVectorType> src[KernelSize], | ||
| 271 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 272 | 6048 | common_vector_path(pg, src, dst); | |
| 273 | 6048 | } | |
| 274 | |||
| 275 | 70640 | void vertical_scalar_path(const SourceType src[KernelSize], | |
| 276 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 277 | 141280 | uint32_t acc = static_cast<uint32_t>(src[kHalfKernelSize - 1]) * | |
| 278 | 70640 | half_kernel_[kHalfKernelSize - 1]; | |
| 279 | |||
| 280 | // Optimization to avoid unnecessary branching in vector code. | ||
| 281 | KLEIDICV_FORCE_LOOP_UNROLL | ||
| 282 |
10/10✓ Branch 0 taken 22176 times.
✓ Branch 1 taken 155232 times.
✓ Branch 2 taken 45120 times.
✓ Branch 3 taken 451200 times.
✓ Branch 4 taken 248 times.
✓ Branch 5 taken 248 times.
✓ Branch 6 taken 960 times.
✓ Branch 7 taken 1920 times.
✓ Branch 8 taken 2136 times.
✓ Branch 9 taken 6408 times.
|
685648 | for (size_t i = 0; i < kHalfKernelSize - 1; i++) { |
| 283 | 1845024 | acc += (static_cast<uint32_t>(src[i]) + | |
| 284 | 1230016 | static_cast<uint32_t>(src[KernelSize - i - 1])) * | |
| 285 | 615008 | half_kernel_[i]; | |
| 286 | 615008 | } | |
| 287 | |||
| 288 | 70640 | dst[0] = static_cast<BufferType>(rounding_shift_right(acc, 8)); | |
| 289 | 70640 | } | |
| 290 | |||
| 291 | 3896 | void horizontal_vector_path( | |
| 292 | svbool_t pg, std::reference_wrapper<SourceVectorType> src[KernelSize], | ||
| 293 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 294 | 3896 | common_vector_path(pg, src, dst); | |
| 295 | 3896 | } | |
| 296 | |||
| 297 | 70640 | void horizontal_scalar_path(const BufferType src[KernelSize], | |
| 298 | DestinationType *dst) const KLEIDICV_STREAMING { | ||
| 299 | 70640 | vertical_scalar_path(src, dst); | |
| 300 | 70640 | } | |
| 301 | |||
| 302 | private: | ||
| 303 | 9944 | void common_vector_path( | |
| 304 | svbool_t pg, std::reference_wrapper<SourceVectorType> src[KernelSize], | ||
| 305 | BufferType *dst) const KLEIDICV_STREAMING { | ||
| 306 | 9944 | svbool_t pg16_all = svptrue_b16(); | |
| 307 | 19888 | svuint16_t acc_b = svmullb_n_u16(src[kHalfKernelSize - 1], | |
| 308 | 9944 | half_kernel_[kHalfKernelSize - 1]); | |
| 309 | 19888 | svuint16_t acc_t = svmullt_n_u16(src[kHalfKernelSize - 1], | |
| 310 | 9944 | half_kernel_[kHalfKernelSize - 1]); | |
| 311 | |||
| 312 | // Optimization to avoid unnecessary branching in vector code. | ||
| 313 | KLEIDICV_FORCE_LOOP_UNROLL | ||
| 314 |
10/10✓ Branch 0 taken 3280 times.
✓ Branch 1 taken 22960 times.
✓ Branch 2 taken 5320 times.
✓ Branch 3 taken 53200 times.
✓ Branch 4 taken 228 times.
✓ Branch 5 taken 228 times.
✓ Branch 6 taken 440 times.
✓ Branch 7 taken 880 times.
✓ Branch 8 taken 676 times.
✓ Branch 9 taken 2028 times.
|
89240 | for (size_t i = 0; i < kHalfKernelSize - 1; i++) { |
| 315 | 79296 | const size_t j = KernelSize - i - 1; | |
| 316 | 79296 | svuint16_t vec_b = svaddlb_u16(src[i], src[j]); | |
| 317 | 79296 | svuint16_t vec_t = svaddlt_u16(src[i], src[j]); | |
| 318 | |||
| 319 | 79296 | acc_b = svmla_n_u16_x(pg16_all, acc_b, vec_b, half_kernel_[i]); | |
| 320 | 79296 | acc_t = svmla_n_u16_x(pg16_all, acc_t, vec_t, half_kernel_[i]); | |
| 321 | 79296 | } | |
| 322 | |||
| 323 | // Rounding before narrowing | ||
| 324 | 9944 | acc_b = svqadd_n_u16(acc_b, 128); | |
| 325 | 9944 | acc_t = svqadd_n_u16(acc_t, 128); | |
| 326 | // Keep only the highest 8 bits | ||
| 327 | 19888 | svuint8_t result = | |
| 328 | 9944 | svtrn2_u8(svreinterpret_u8_u16(acc_b), svreinterpret_u8_u16(acc_t)); | |
| 329 | 9944 | svst1(pg, &dst[0], result); | |
| 330 | 9944 | } | |
| 331 | |||
| 332 | const uint16_t *half_kernel_; | ||
| 333 | }; // end of class GaussianBlur<uint8_t, KernelSize, false> | ||
| 334 | |||
| 335 | template <size_t KernelSize, bool IsBinomial, typename ScalarType> | ||
| 336 | 652 | static kleidicv_error_t gaussian_blur_fixed_kernel_size( | |
| 337 | const ScalarType *src, size_t src_stride, ScalarType *dst, | ||
| 338 | size_t dst_stride, Rectangle &rect, size_t y_begin, size_t y_end, | ||
| 339 | size_t channels, float sigma, FixedBorderType border_type, | ||
| 340 | SeparableFilterWorkspace *workspace) KLEIDICV_STREAMING { | ||
| 341 | using GaussianBlurFilter = GaussianBlur<ScalarType, KernelSize, IsBinomial>; | ||
| 342 | |||
| 343 | 652 | Rows<const ScalarType> src_rows{src, src_stride, channels}; | |
| 344 | 652 | Rows<ScalarType> dst_rows{dst, dst_stride, channels}; | |
| 345 | |||
| 346 | if constexpr (IsBinomial) { | ||
| 347 | 262 | GaussianBlurFilter blur; | |
| 348 | 262 | SeparableFilter<GaussianBlurFilter, KernelSize> filter{blur}; | |
| 349 | 524 | workspace->process(rect, y_begin, y_end, src_rows, dst_rows, channels, | |
| 350 | 262 | border_type, filter); | |
| 351 | |||
| 352 | 262 | return KLEIDICV_OK; | |
| 353 | 262 | } else { | |
| 354 | 390 | constexpr size_t kHalfKernelSize = get_half_kernel_size(KernelSize); | |
| 355 | 390 | uint16_t half_kernel[128]; | |
| 356 | 390 | generate_gaussian_half_kernel(half_kernel, kHalfKernelSize, sigma); | |
| 357 | // If sigma is so small that the middle point gets all the weights, it's | ||
| 358 | // just a copy | ||
| 359 |
10/10✓ Branch 0 taken 64 times.
✓ Branch 1 taken 32 times.
✓ Branch 2 taken 64 times.
✓ Branch 3 taken 32 times.
✓ Branch 4 taken 34 times.
✓ Branch 5 taken 32 times.
✓ Branch 6 taken 34 times.
✓ Branch 7 taken 32 times.
✓ Branch 8 taken 34 times.
✓ Branch 9 taken 32 times.
|
390 | if (half_kernel[kHalfKernelSize - 1] < 256) { |
| 360 | 230 | GaussianBlurFilter blur(half_kernel); | |
| 361 | 230 | SeparableFilter<GaussianBlurFilter, KernelSize> filter{blur}; | |
| 362 | 460 | workspace->process(rect, y_begin, y_end, src_rows, dst_rows, channels, | |
| 363 | 230 | border_type, filter); | |
| 364 | 230 | } else { | |
| 365 |
10/10✓ Branch 0 taken 456 times.
✓ Branch 1 taken 32 times.
✓ Branch 2 taken 648 times.
✓ Branch 3 taken 32 times.
✓ Branch 4 taken 72 times.
✓ Branch 5 taken 32 times.
✓ Branch 6 taken 136 times.
✓ Branch 7 taken 32 times.
✓ Branch 8 taken 200 times.
✓ Branch 9 taken 32 times.
|
1672 | for (size_t row = y_begin; row < y_end; ++row) { |
| 366 | #if KLEIDICV_TARGET_SME && defined(__ANDROID__) | ||
| 367 | __arm_sc_memcpy( | ||
| 368 | static_cast<void *>(&dst_rows.at(row)[0]), | ||
| 369 | static_cast<const void *>(&src_rows.at(row)[0]), | ||
| 370 | rect.width() * sizeof(ScalarType) * dst_rows.channels()); | ||
| 371 | #else | ||
| 372 | 3024 | std::memcpy(static_cast<void *>(&dst_rows.at(row)[0]), | |
| 373 | 1512 | static_cast<const void *>(&src_rows.at(row)[0]), | |
| 374 | 1512 | rect.width() * sizeof(ScalarType) * dst_rows.channels()); | |
| 375 | #endif | ||
| 376 | 1512 | } | |
| 377 | } | ||
| 378 | 390 | return KLEIDICV_OK; | |
| 379 | 390 | } | |
| 380 | 652 | } | |
| 381 | |||
| 382 | template <bool IsBinomial, typename ScalarType> | ||
| 383 | 652 | static kleidicv_error_t gaussian_blur( | |
| 384 | size_t kernel_size, const ScalarType *src, size_t src_stride, | ||
| 385 | ScalarType *dst, size_t dst_stride, Rectangle &rect, size_t y_begin, | ||
| 386 | size_t y_end, size_t channels, float sigma, FixedBorderType border_type, | ||
| 387 | SeparableFilterWorkspace *workspace) KLEIDICV_STREAMING { | ||
| 388 |
10/12✓ Branch 0 taken 94 times.
✓ Branch 1 taken 102 times.
✓ Branch 2 taken 66 times.
✓ Branch 3 taken 32 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 32 times.
✓ Branch 6 taken 66 times.
✓ Branch 7 taken 66 times.
✓ Branch 8 taken 66 times.
✓ Branch 9 taken 64 times.
✗ Branch 10 not taken.
✓ Branch 11 taken 64 times.
|
652 | switch (kernel_size) { |
| 389 | case 3: | ||
| 390 | 160 | return gaussian_blur_fixed_kernel_size<3, IsBinomial>( | |
| 391 | 160 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 392 | 160 | sigma, border_type, workspace); | |
| 393 | case 5: | ||
| 394 | 168 | return gaussian_blur_fixed_kernel_size<5, IsBinomial>( | |
| 395 | 168 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 396 | 168 | sigma, border_type, workspace); | |
| 397 | case 7: | ||
| 398 | 132 | return gaussian_blur_fixed_kernel_size<7, IsBinomial>( | |
| 399 | 132 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 400 | 132 | sigma, border_type, workspace); | |
| 401 | case 15: | ||
| 402 | // 15x15 does not have a binomial variant | ||
| 403 | 96 | return gaussian_blur_fixed_kernel_size<15, false>( | |
| 404 | 96 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 405 | 96 | sigma, border_type, workspace); | |
| 406 | case 21: | ||
| 407 | // 21x21 does not have a binomial variant | ||
| 408 | 96 | return gaussian_blur_fixed_kernel_size<21, false>( | |
| 409 | 96 | src, src_stride, dst, dst_stride, rect, y_begin, y_end, channels, | |
| 410 | 96 | sigma, border_type, workspace); | |
| 411 | // gaussian_blur_is_implemented checked the kernel size already. | ||
| 412 | // GCOVR_EXCL_START | ||
| 413 | default: | ||
| 414 | assert(!"kernel size not implemented"); | ||
| 415 | − | return KLEIDICV_ERROR_NOT_IMPLEMENTED; | |
| 416 | // GCOVR_EXCL_STOP | ||
| 417 | } | ||
| 418 | 652 | } | |
| 419 | |||
| 420 | 690 | static kleidicv_error_t gaussian_blur_fixed_stripe_u8_sc( | |
| 421 | const uint8_t *src, size_t src_stride, uint8_t *dst, size_t dst_stride, | ||
| 422 | size_t width, size_t height, size_t y_begin, size_t y_end, size_t channels, | ||
| 423 | size_t kernel_width, size_t /*kernel_height*/, float sigma_x, | ||
| 424 | float /*sigma_y*/, FixedBorderType fixed_border_type, | ||
| 425 | kleidicv_filter_context_t *context) KLEIDICV_STREAMING { | ||
| 426 | 690 | auto *workspace = reinterpret_cast<SeparableFilterWorkspace *>(context); | |
| 427 | 1380 | kleidicv_error_t checks_result = gaussian_blur_checks( | |
| 428 | 690 | src, src_stride, dst, dst_stride, width, height, channels, workspace); | |
| 429 | |||
| 430 |
2/2✓ Branch 0 taken 38 times.
✓ Branch 1 taken 652 times.
|
690 | if (checks_result != KLEIDICV_OK) { |
| 431 | 38 | return checks_result; | |
| 432 | } | ||
| 433 | |||
| 434 | 652 | Rectangle rect{width, height}; | |
| 435 | |||
| 436 |
2/2✓ Branch 0 taken 326 times.
✓ Branch 1 taken 326 times.
|
652 | if (sigma_x == 0.0) { |
| 437 | 652 | return gaussian_blur<true>(kernel_width, src, src_stride, dst, dst_stride, | |
| 438 | 326 | rect, y_begin, y_end, channels, sigma_x, | |
| 439 | 326 | fixed_border_type, workspace); | |
| 440 | } | ||
| 441 | |||
| 442 | 652 | return gaussian_blur<false>(kernel_width, src, src_stride, dst, dst_stride, | |
| 443 | 326 | rect, y_begin, y_end, channels, sigma_x, | |
| 444 | 326 | fixed_border_type, workspace); | |
| 445 | 690 | } | |
| 446 | |||
| 447 | } // namespace KLEIDICV_TARGET_NAMESPACE | ||
| 448 | |||
| 449 | #endif // KLEIDICV_GAUSSIAN_BLUR_SC_H | ||
| 450 |