Eric | 3afc1d1 | 2020-07-23 02:16:46 +0200 | [diff] [blame] | 1 | /*! \file conv_acc_neon_impl.h |
| 2 | * Accelerated Viterbi decoder implementation: |
| 3 | * straight port of SSE to NEON based on Tom Tsous work */ |
| 4 | /* |
| 5 | * (C) 2020 by sysmocom - s.f.m.c. GmbH |
| 6 | * Author: Eric Wild |
| 7 | * |
| 8 | * All Rights Reserved |
| 9 | * |
| 10 | * SPDX-License-Identifier: GPL-2.0+ |
| 11 | * |
| 12 | * This program is free software; you can redistribute it and/or modify |
| 13 | * it under the terms of the GNU General Public License as published by |
| 14 | * the Free Software Foundation; either version 2 of the License, or |
| 15 | * (at your option) any later version. |
| 16 | * |
| 17 | * This program is distributed in the hope that it will be useful, |
| 18 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 19 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 20 | * GNU General Public License for more details. |
Eric | 3afc1d1 | 2020-07-23 02:16:46 +0200 | [diff] [blame] | 21 | */ |
| 22 | |
| 23 | /* Some distributions (notably Alpine Linux) for some strange reason |
| 24 | * don't have this #define */ |
| 25 | #ifndef __always_inline |
| 26 | #define __always_inline inline __attribute__((always_inline)) |
| 27 | #endif |
| 28 | |
| 29 | #define NEON_BUTTERFLY(M0,M1,M2,M3,M4) \ |
| 30 | { \ |
| 31 | M3 = vqaddq_s16(M0, M2); \ |
| 32 | M4 = vqsubq_s16(M1, M2); \ |
| 33 | M0 = vqsubq_s16(M0, M2); \ |
| 34 | M1 = vqaddq_s16(M1, M2); \ |
| 35 | M2 = vmaxq_s16(M3, M4); \ |
| 36 | M3 = vreinterpretq_s16_u16(vcgtq_s16(M3, M4)); \ |
| 37 | M4 = vmaxq_s16(M0, M1); \ |
| 38 | M1 = vreinterpretq_s16_u16(vcgtq_s16(M0, M1)); \ |
| 39 | } |
| 40 | |
| 41 | #define NEON_DEINTERLEAVE_K5(M0,M1,M2,M3) \ |
| 42 | { \ |
| 43 | int16x8x2_t tmp; \ |
| 44 | tmp = vuzpq_s16(M0, M1); \ |
| 45 | M2 = tmp.val[0]; \ |
| 46 | M3 = tmp.val[1]; \ |
| 47 | } |
| 48 | |
| 49 | #define NEON_DEINTERLEAVE_K7(M0,M1,M2,M3,M4,M5,M6,M7,M8,M9,M10,M11,M12,M13,M14,M15) \ |
| 50 | { \ |
| 51 | int16x8x2_t tmp; \ |
| 52 | tmp = vuzpq_s16(M0, M1); \ |
| 53 | M8 = tmp.val[0]; M9 = tmp.val[1]; \ |
| 54 | tmp = vuzpq_s16(M2, M3); \ |
| 55 | M10 = tmp.val[0]; M11 = tmp.val[1]; \ |
| 56 | tmp = vuzpq_s16(M4, M5); \ |
| 57 | M12 = tmp.val[0]; M13 = tmp.val[1]; \ |
| 58 | tmp = vuzpq_s16(M6, M7); \ |
| 59 | M14 = tmp.val[0]; M15 = tmp.val[1]; \ |
| 60 | } |
| 61 | |
| 62 | #define NEON_BRANCH_METRIC_N2(M0,M1,M2,M3,M4,M6,M7) \ |
| 63 | { \ |
| 64 | M0 = vmulq_s16(M4, M0); \ |
| 65 | M1 = vmulq_s16(M4, M1); \ |
| 66 | M2 = vmulq_s16(M4, M2); \ |
| 67 | M3 = vmulq_s16(M4, M3); \ |
| 68 | M6 = vcombine_s16(vpadd_s16(vget_low_s16(M0), vget_high_s16(M0)), vpadd_s16(vget_low_s16(M1), vget_high_s16(M1))); \ |
| 69 | M7 = vcombine_s16(vpadd_s16(vget_low_s16(M2), vget_high_s16(M2)), vpadd_s16(vget_low_s16(M3), vget_high_s16(M3))); \ |
| 70 | } |
| 71 | |
| 72 | #define NEON_BRANCH_METRIC_N4(M0,M1,M2,M3,M4,M5) \ |
| 73 | { \ |
| 74 | M0 = vmulq_s16(M4, M0); \ |
| 75 | M1 = vmulq_s16(M4, M1); \ |
| 76 | M2 = vmulq_s16(M4, M2); \ |
| 77 | M3 = vmulq_s16(M4, M3); \ |
| 78 | int16x4_t t1 = vpadd_s16(vpadd_s16(vget_low_s16(M0), vget_high_s16(M0)), vpadd_s16(vget_low_s16(M1), vget_high_s16(M1))); \ |
| 79 | int16x4_t t2 = vpadd_s16(vpadd_s16(vget_low_s16(M2), vget_high_s16(M2)), vpadd_s16(vget_low_s16(M3), vget_high_s16(M3))); \ |
| 80 | M5 = vcombine_s16(t1, t2); \ |
| 81 | } |
| 82 | |
| 83 | #define NEON_NORMALIZE_K5(M0,M1,M2,M3) \ |
| 84 | { \ |
| 85 | M2 = vminq_s16(M0, M1); \ |
| 86 | int16x4_t t = vpmin_s16(vget_low_s16(M2), vget_high_s16(M2)); \ |
| 87 | t = vpmin_s16(t, t); \ |
| 88 | t = vpmin_s16(t, t); \ |
| 89 | M2 = vdupq_lane_s16(t, 0); \ |
| 90 | M0 = vqsubq_s16(M0, M2); \ |
| 91 | M1 = vqsubq_s16(M1, M2); \ |
| 92 | } |
| 93 | |
| 94 | #define NEON_NORMALIZE_K7(M0,M1,M2,M3,M4,M5,M6,M7,M8,M9,M10,M11) \ |
| 95 | { \ |
| 96 | M8 = vminq_s16(M0, M1); \ |
| 97 | M9 = vminq_s16(M2, M3); \ |
| 98 | M10 = vminq_s16(M4, M5); \ |
| 99 | M11 = vminq_s16(M6, M7); \ |
| 100 | M8 = vminq_s16(M8, M9); \ |
| 101 | M10 = vminq_s16(M10, M11); \ |
| 102 | M8 = vminq_s16(M8, M10); \ |
| 103 | int16x4_t t = vpmin_s16(vget_low_s16(M8), vget_high_s16(M8)); \ |
| 104 | t = vpmin_s16(t, t); \ |
| 105 | t = vpmin_s16(t, t); \ |
| 106 | M8 = vdupq_lane_s16(t, 0); \ |
| 107 | M0 = vqsubq_s16(M0, M8); \ |
| 108 | M1 = vqsubq_s16(M1, M8); \ |
| 109 | M2 = vqsubq_s16(M2, M8); \ |
| 110 | M3 = vqsubq_s16(M3, M8); \ |
| 111 | M4 = vqsubq_s16(M4, M8); \ |
| 112 | M5 = vqsubq_s16(M5, M8); \ |
| 113 | M6 = vqsubq_s16(M6, M8); \ |
| 114 | M7 = vqsubq_s16(M7, M8); \ |
| 115 | } |
| 116 | |
| 117 | __always_inline void _neon_metrics_k5_n2(const int16_t *val, const int16_t *outa, int16_t *sumsa, int16_t *paths, |
| 118 | int norm) |
| 119 | { |
| 120 | int16_t *__restrict out = __builtin_assume_aligned(outa, 8); |
| 121 | int16_t *__restrict sums = __builtin_assume_aligned(sumsa, 8); |
| 122 | int16x8_t m0, m1, m2, m3, m4, m5, m6; |
| 123 | int16x4_t input; |
| 124 | |
| 125 | /* (BMU) Load and expand 8-bit input out to 16-bits */ |
| 126 | input = vld1_s16(val); |
| 127 | m2 = vcombine_s16(input, input); |
| 128 | |
| 129 | /* (BMU) Load and compute branch metrics */ |
| 130 | m0 = vld1q_s16(&out[0]); |
| 131 | m1 = vld1q_s16(&out[8]); |
| 132 | |
| 133 | m0 = vmulq_s16(m2, m0); |
| 134 | m1 = vmulq_s16(m2, m1); |
| 135 | m2 = vcombine_s16(vpadd_s16(vget_low_s16(m0), vget_high_s16(m0)), |
| 136 | vpadd_s16(vget_low_s16(m1), vget_high_s16(m1))); |
| 137 | |
| 138 | /* (PMU) Load accumulated path matrics */ |
| 139 | m0 = vld1q_s16(&sums[0]); |
| 140 | m1 = vld1q_s16(&sums[8]); |
| 141 | |
| 142 | NEON_DEINTERLEAVE_K5(m0, m1, m3, m4) |
| 143 | |
| 144 | /* (PMU) Butterflies: 0-7 */ |
| 145 | NEON_BUTTERFLY(m3, m4, m2, m5, m6) |
| 146 | |
| 147 | if (norm) |
| 148 | NEON_NORMALIZE_K5(m2, m6, m0, m1) |
| 149 | |
| 150 | vst1q_s16(&sums[0], m2); |
| 151 | vst1q_s16(&sums[8], m6); |
| 152 | vst1q_s16(&paths[0], m5); |
| 153 | vst1q_s16(&paths[8], m4); |
| 154 | } |
| 155 | |
| 156 | __always_inline void _neon_metrics_k5_n4(const int16_t *val, const int16_t *outa, int16_t *sumsa, int16_t *paths, |
| 157 | int norm) |
| 158 | { |
| 159 | int16_t *__restrict out = __builtin_assume_aligned(outa, 8); |
| 160 | int16_t *__restrict sums = __builtin_assume_aligned(sumsa, 8); |
| 161 | int16x8_t m0, m1, m2, m3, m4, m5, m6; |
| 162 | int16x4_t input; |
| 163 | |
| 164 | /* (BMU) Load and expand 8-bit input out to 16-bits */ |
| 165 | input = vld1_s16(val); |
| 166 | m4 = vcombine_s16(input, input); |
| 167 | |
| 168 | /* (BMU) Load and compute branch metrics */ |
| 169 | m0 = vld1q_s16(&out[0]); |
| 170 | m1 = vld1q_s16(&out[8]); |
| 171 | m2 = vld1q_s16(&out[16]); |
| 172 | m3 = vld1q_s16(&out[24]); |
| 173 | |
| 174 | NEON_BRANCH_METRIC_N4(m0, m1, m2, m3, m4, m2) |
| 175 | |
| 176 | /* (PMU) Load accumulated path matrics */ |
| 177 | m0 = vld1q_s16(&sums[0]); |
| 178 | m1 = vld1q_s16(&sums[8]); |
| 179 | |
| 180 | NEON_DEINTERLEAVE_K5(m0, m1, m3, m4) |
| 181 | |
| 182 | /* (PMU) Butterflies: 0-7 */ |
| 183 | NEON_BUTTERFLY(m3, m4, m2, m5, m6) |
| 184 | |
| 185 | if (norm) |
| 186 | NEON_NORMALIZE_K5(m2, m6, m0, m1) |
| 187 | |
| 188 | vst1q_s16(&sums[0], m2); |
| 189 | vst1q_s16(&sums[8], m6); |
| 190 | vst1q_s16(&paths[0], m5); |
| 191 | vst1q_s16(&paths[8], m4); |
| 192 | } |
| 193 | |
| 194 | __always_inline static void _neon_metrics_k7_n2(const int16_t *val, const int16_t *outa, int16_t *sumsa, int16_t *paths, |
| 195 | int norm) |
| 196 | { |
| 197 | int16_t *__restrict out = __builtin_assume_aligned(outa, 8); |
| 198 | int16_t *__restrict sums = __builtin_assume_aligned(sumsa, 8); |
| 199 | int16x8_t m0, m1, m2, m3, m4, m5, m6, m7; |
| 200 | int16x8_t m8, m9, m10, m11, m12, m13, m14, m15; |
| 201 | int16x4_t input; |
| 202 | |
| 203 | /* (PMU) Load accumulated path matrics */ |
| 204 | m0 = vld1q_s16(&sums[0]); |
| 205 | m1 = vld1q_s16(&sums[8]); |
| 206 | m2 = vld1q_s16(&sums[16]); |
| 207 | m3 = vld1q_s16(&sums[24]); |
| 208 | m4 = vld1q_s16(&sums[32]); |
| 209 | m5 = vld1q_s16(&sums[40]); |
| 210 | m6 = vld1q_s16(&sums[48]); |
| 211 | m7 = vld1q_s16(&sums[56]); |
| 212 | |
| 213 | /* (PMU) Deinterleave into even and odd packed registers */ |
| 214 | NEON_DEINTERLEAVE_K7(m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11, m12, m13, m14, m15) |
| 215 | |
| 216 | /* (BMU) Load and expand 8-bit input out to 16-bits */ |
| 217 | input = vld1_s16(val); |
| 218 | m7 = vcombine_s16(input, input); |
| 219 | |
| 220 | /* (BMU) Load and compute branch metrics */ |
| 221 | m0 = vld1q_s16(&out[0]); |
| 222 | m1 = vld1q_s16(&out[8]); |
| 223 | m2 = vld1q_s16(&out[16]); |
| 224 | m3 = vld1q_s16(&out[24]); |
| 225 | |
| 226 | NEON_BRANCH_METRIC_N2(m0, m1, m2, m3, m7, m4, m5) |
| 227 | |
| 228 | m0 = vld1q_s16(&out[32]); |
| 229 | m1 = vld1q_s16(&out[40]); |
| 230 | m2 = vld1q_s16(&out[48]); |
| 231 | m3 = vld1q_s16(&out[56]); |
| 232 | |
| 233 | NEON_BRANCH_METRIC_N2(m0, m1, m2, m3, m7, m6, m7) |
| 234 | |
| 235 | /* (PMU) Butterflies: 0-15 */ |
| 236 | NEON_BUTTERFLY(m8, m9, m4, m0, m1) |
| 237 | NEON_BUTTERFLY(m10, m11, m5, m2, m3) |
| 238 | |
| 239 | vst1q_s16(&paths[0], m0); |
| 240 | vst1q_s16(&paths[8], m2); |
| 241 | vst1q_s16(&paths[32], m9); |
| 242 | vst1q_s16(&paths[40], m11); |
| 243 | |
| 244 | /* (PMU) Butterflies: 17-31 */ |
| 245 | NEON_BUTTERFLY(m12, m13, m6, m0, m2) |
| 246 | NEON_BUTTERFLY(m14, m15, m7, m9, m11) |
| 247 | |
| 248 | vst1q_s16(&paths[16], m0); |
| 249 | vst1q_s16(&paths[24], m9); |
| 250 | vst1q_s16(&paths[48], m13); |
| 251 | vst1q_s16(&paths[56], m15); |
| 252 | |
| 253 | if (norm) |
| 254 | NEON_NORMALIZE_K7(m4, m1, m5, m3, m6, m2, m7, m11, m0, m8, m9, m10) |
| 255 | |
| 256 | vst1q_s16(&sums[0], m4); |
| 257 | vst1q_s16(&sums[8], m5); |
| 258 | vst1q_s16(&sums[16], m6); |
| 259 | vst1q_s16(&sums[24], m7); |
| 260 | vst1q_s16(&sums[32], m1); |
| 261 | vst1q_s16(&sums[40], m3); |
| 262 | vst1q_s16(&sums[48], m2); |
| 263 | vst1q_s16(&sums[56], m11); |
| 264 | } |
| 265 | |
| 266 | __always_inline static void _neon_metrics_k7_n4(const int16_t *val, const int16_t *outa, int16_t *sumsa, int16_t *paths, |
| 267 | int norm) |
| 268 | { |
| 269 | int16_t *__restrict out = __builtin_assume_aligned(outa, 8); |
| 270 | int16_t *__restrict sums = __builtin_assume_aligned(sumsa, 8); |
| 271 | int16x8_t m0, m1, m2, m3, m4, m5, m6, m7; |
| 272 | int16x8_t m8, m9, m10, m11, m12, m13, m14, m15; |
| 273 | int16x4_t input; |
| 274 | |
| 275 | /* (PMU) Load accumulated path matrics */ |
| 276 | m0 = vld1q_s16(&sums[0]); |
| 277 | m1 = vld1q_s16(&sums[8]); |
| 278 | m2 = vld1q_s16(&sums[16]); |
| 279 | m3 = vld1q_s16(&sums[24]); |
| 280 | m4 = vld1q_s16(&sums[32]); |
| 281 | m5 = vld1q_s16(&sums[40]); |
| 282 | m6 = vld1q_s16(&sums[48]); |
| 283 | m7 = vld1q_s16(&sums[56]); |
| 284 | |
| 285 | /* (PMU) Deinterleave into even and odd packed registers */ |
| 286 | NEON_DEINTERLEAVE_K7(m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11, m12, m13, m14, m15) |
| 287 | |
| 288 | /* (BMU) Load and expand 8-bit input out to 16-bits */ |
| 289 | input = vld1_s16(val); |
| 290 | m7 = vcombine_s16(input, input); |
| 291 | |
| 292 | /* (BMU) Load and compute branch metrics */ |
| 293 | m0 = vld1q_s16(&out[0]); |
| 294 | m1 = vld1q_s16(&out[8]); |
| 295 | m2 = vld1q_s16(&out[16]); |
| 296 | m3 = vld1q_s16(&out[24]); |
| 297 | |
| 298 | NEON_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m4) |
| 299 | |
| 300 | m0 = vld1q_s16(&out[32]); |
| 301 | m1 = vld1q_s16(&out[40]); |
| 302 | m2 = vld1q_s16(&out[48]); |
| 303 | m3 = vld1q_s16(&out[56]); |
| 304 | |
| 305 | NEON_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m5) |
| 306 | |
| 307 | m0 = vld1q_s16(&out[64]); |
| 308 | m1 = vld1q_s16(&out[72]); |
| 309 | m2 = vld1q_s16(&out[80]); |
| 310 | m3 = vld1q_s16(&out[88]); |
| 311 | |
| 312 | NEON_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m6) |
| 313 | |
| 314 | m0 = vld1q_s16(&out[96]); |
| 315 | m1 = vld1q_s16(&out[104]); |
| 316 | m2 = vld1q_s16(&out[112]); |
| 317 | m3 = vld1q_s16(&out[120]); |
| 318 | |
| 319 | NEON_BRANCH_METRIC_N4(m0, m1, m2, m3, m7, m7) |
| 320 | |
| 321 | /* (PMU) Butterflies: 0-15 */ |
| 322 | NEON_BUTTERFLY(m8, m9, m4, m0, m1) |
| 323 | NEON_BUTTERFLY(m10, m11, m5, m2, m3) |
| 324 | |
| 325 | vst1q_s16(&paths[0], m0); |
| 326 | vst1q_s16(&paths[8], m2); |
| 327 | vst1q_s16(&paths[32], m9); |
| 328 | vst1q_s16(&paths[40], m11); |
| 329 | |
| 330 | /* (PMU) Butterflies: 17-31 */ |
| 331 | NEON_BUTTERFLY(m12, m13, m6, m0, m2) |
| 332 | NEON_BUTTERFLY(m14, m15, m7, m9, m11) |
| 333 | |
| 334 | vst1q_s16(&paths[16], m0); |
| 335 | vst1q_s16(&paths[24], m9); |
| 336 | vst1q_s16(&paths[48], m13); |
| 337 | vst1q_s16(&paths[56], m15); |
| 338 | |
| 339 | if (norm) |
| 340 | NEON_NORMALIZE_K7(m4, m1, m5, m3, m6, m2, m7, m11, m0, m8, m9, m10) |
| 341 | |
| 342 | vst1q_s16(&sums[0], m4); |
| 343 | vst1q_s16(&sums[8], m5); |
| 344 | vst1q_s16(&sums[16], m6); |
| 345 | vst1q_s16(&sums[24], m7); |
| 346 | vst1q_s16(&sums[32], m1); |
| 347 | vst1q_s16(&sums[40], m3); |
| 348 | vst1q_s16(&sums[48], m2); |
| 349 | vst1q_s16(&sums[56], m11); |
| 350 | } |