1/* 2 * Copyright (c) 2012 The WebRTC project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11#include "webrtc/modules/audio_processing/ns/nsx_core.h" 12 13#include <arm_neon.h> 14#include <assert.h> 15 16// Constants to compensate for shifting signal log(2^shifts). 17const int16_t WebRtcNsx_kLogTable[9] = { 18 0, 177, 355, 532, 710, 887, 1065, 1242, 1420 19}; 20 21const int16_t WebRtcNsx_kCounterDiv[201] = { 22 32767, 16384, 10923, 8192, 6554, 5461, 4681, 4096, 3641, 3277, 2979, 2731, 23 2521, 2341, 2185, 2048, 1928, 1820, 1725, 1638, 1560, 1489, 1425, 1365, 1311, 24 1260, 1214, 1170, 1130, 1092, 1057, 1024, 993, 964, 936, 910, 886, 862, 840, 25 819, 799, 780, 762, 745, 728, 712, 697, 683, 669, 655, 643, 630, 618, 607, 26 596, 585, 575, 565, 555, 546, 537, 529, 520, 512, 504, 496, 489, 482, 475, 27 468, 462, 455, 449, 443, 437, 431, 426, 420, 415, 410, 405, 400, 395, 390, 28 386, 381, 377, 372, 368, 364, 360, 356, 352, 349, 345, 341, 338, 334, 331, 29 328, 324, 321, 318, 315, 312, 309, 306, 303, 301, 298, 295, 293, 290, 287, 30 285, 282, 280, 278, 275, 273, 271, 269, 266, 264, 262, 260, 258, 256, 254, 31 252, 250, 248, 246, 245, 243, 241, 239, 237, 236, 234, 232, 231, 229, 228, 32 226, 224, 223, 221, 220, 218, 217, 216, 214, 213, 211, 210, 209, 207, 206, 33 205, 204, 202, 201, 200, 199, 197, 196, 195, 194, 193, 192, 191, 189, 188, 34 187, 186, 185, 184, 183, 182, 181, 180, 179, 178, 177, 176, 175, 174, 173, 35 172, 172, 171, 170, 169, 168, 167, 166, 165, 165, 164, 163 36}; 37 38const int16_t WebRtcNsx_kLogTableFrac[256] = { 39 0, 1, 3, 4, 6, 7, 9, 10, 11, 13, 14, 16, 17, 18, 20, 21, 40 22, 24, 25, 26, 28, 29, 30, 32, 33, 34, 36, 37, 38, 40, 41, 42, 41 44, 45, 46, 47, 49, 50, 51, 52, 54, 55, 56, 57, 59, 60, 61, 62, 42 63, 65, 66, 67, 68, 69, 71, 72, 73, 74, 75, 77, 78, 79, 80, 81, 43 82, 84, 85, 86, 87, 88, 89, 90, 92, 93, 94, 95, 96, 97, 98, 99, 44 100, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 116, 45 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 46 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 47 147, 148, 149, 150, 151, 152, 153, 154, 155, 155, 156, 157, 158, 159, 160, 48 161, 162, 163, 164, 165, 166, 167, 168, 169, 169, 170, 171, 172, 173, 174, 49 175, 176, 177, 178, 178, 179, 180, 181, 182, 183, 184, 185, 185, 186, 187, 50 188, 189, 190, 191, 192, 192, 193, 194, 195, 196, 197, 198, 198, 199, 200, 51 201, 202, 203, 203, 204, 205, 206, 207, 208, 208, 209, 210, 211, 212, 212, 52 213, 214, 215, 216, 216, 217, 218, 219, 220, 220, 221, 222, 223, 224, 224, 53 225, 226, 227, 228, 228, 229, 230, 231, 231, 232, 233, 234, 234, 235, 236, 54 237, 238, 238, 239, 240, 241, 241, 242, 243, 244, 244, 245, 246, 247, 247, 55 248, 249, 249, 250, 251, 252, 252, 253, 254, 255, 255 56}; 57 58// Update the noise estimation information. 59static void UpdateNoiseEstimateNeon(NsxInst_t* inst, int offset) { 60 const int16_t kExp2Const = 11819; // Q13 61 int16_t* ptr_noiseEstLogQuantile = NULL; 62 int16_t* ptr_noiseEstQuantile = NULL; 63 int16x4_t kExp2Const16x4 = vdup_n_s16(kExp2Const); 64 int32x4_t twentyOne32x4 = vdupq_n_s32(21); 65 int32x4_t constA32x4 = vdupq_n_s32(0x1fffff); 66 int32x4_t constB32x4 = vdupq_n_s32(0x200000); 67 68 int16_t tmp16 = WebRtcSpl_MaxValueW16(inst->noiseEstLogQuantile + offset, 69 inst->magnLen); 70 71 // Guarantee a Q-domain as high as possible and still fit in int16 72 inst->qNoise = 14 - (int) WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(kExp2Const, 73 tmp16, 74 21); 75 76 int32x4_t qNoise32x4 = vdupq_n_s32(inst->qNoise); 77 78 for (ptr_noiseEstLogQuantile = &inst->noiseEstLogQuantile[offset], 79 ptr_noiseEstQuantile = &inst->noiseEstQuantile[0]; 80 ptr_noiseEstQuantile < &inst->noiseEstQuantile[inst->magnLen - 3]; 81 ptr_noiseEstQuantile += 4, ptr_noiseEstLogQuantile += 4) { 82 83 // tmp32no2 = WEBRTC_SPL_MUL_16_16(kExp2Const, 84 // inst->noiseEstLogQuantile[offset + i]); 85 int16x4_t v16x4 = vld1_s16(ptr_noiseEstLogQuantile); 86 int32x4_t v32x4B = vmull_s16(v16x4, kExp2Const16x4); 87 88 // tmp32no1 = (0x00200000 | (tmp32no2 & 0x001FFFFF)); // 2^21 + frac 89 int32x4_t v32x4A = vandq_s32(v32x4B, constA32x4); 90 v32x4A = vorrq_s32(v32x4A, constB32x4); 91 92 // tmp16 = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmp32no2, 21); 93 v32x4B = vshrq_n_s32(v32x4B, 21); 94 95 // tmp16 -= 21;// shift 21 to get result in Q0 96 v32x4B = vsubq_s32(v32x4B, twentyOne32x4); 97 98 // tmp16 += (int16_t) inst->qNoise; 99 // shift to get result in Q(qNoise) 100 v32x4B = vaddq_s32(v32x4B, qNoise32x4); 101 102 // if (tmp16 < 0) { 103 // tmp32no1 = WEBRTC_SPL_RSHIFT_W32(tmp32no1, -tmp16); 104 // } else { 105 // tmp32no1 = WEBRTC_SPL_LSHIFT_W32(tmp32no1, tmp16); 106 // } 107 v32x4B = vshlq_s32(v32x4A, v32x4B); 108 109 // tmp16 = WebRtcSpl_SatW32ToW16(tmp32no1); 110 v16x4 = vqmovn_s32(v32x4B); 111 112 //inst->noiseEstQuantile[i] = tmp16; 113 vst1_s16(ptr_noiseEstQuantile, v16x4); 114 } 115 116 // Last iteration: 117 118 // inst->quantile[i]=exp(inst->lquantile[offset+i]); 119 // in Q21 120 int32_t tmp32no2 = WEBRTC_SPL_MUL_16_16(kExp2Const, 121 *ptr_noiseEstLogQuantile); 122 int32_t tmp32no1 = (0x00200000 | (tmp32no2 & 0x001FFFFF)); // 2^21 + frac 123 124 tmp16 = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmp32no2, 21); 125 tmp16 -= 21;// shift 21 to get result in Q0 126 tmp16 += (int16_t) inst->qNoise; //shift to get result in Q(qNoise) 127 if (tmp16 < 0) { 128 tmp32no1 = WEBRTC_SPL_RSHIFT_W32(tmp32no1, -tmp16); 129 } else { 130 tmp32no1 = WEBRTC_SPL_LSHIFT_W32(tmp32no1, tmp16); 131 } 132 *ptr_noiseEstQuantile = WebRtcSpl_SatW32ToW16(tmp32no1); 133} 134 135// Noise Estimation 136void WebRtcNsx_NoiseEstimationNeon(NsxInst_t* inst, 137 uint16_t* magn, 138 uint32_t* noise, 139 int16_t* q_noise) { 140 int16_t lmagn[HALF_ANAL_BLOCKL], counter, countDiv; 141 int16_t countProd, delta, zeros, frac; 142 int16_t log2, tabind, logval, tmp16, tmp16no1, tmp16no2; 143 const int16_t log2_const = 22713; 144 const int16_t width_factor = 21845; 145 146 int i, s, offset; 147 148 tabind = inst->stages - inst->normData; 149 assert(tabind < 9); 150 assert(tabind > -9); 151 if (tabind < 0) { 152 logval = -WebRtcNsx_kLogTable[-tabind]; 153 } else { 154 logval = WebRtcNsx_kLogTable[tabind]; 155 } 156 157 int16x8_t logval_16x8 = vdupq_n_s16(logval); 158 159 // lmagn(i)=log(magn(i))=log(2)*log2(magn(i)) 160 // magn is in Q(-stages), and the real lmagn values are: 161 // real_lmagn(i)=log(magn(i)*2^stages)=log(magn(i))+log(2^stages) 162 // lmagn in Q8 163 for (i = 0; i < inst->magnLen; i++) { 164 if (magn[i]) { 165 zeros = WebRtcSpl_NormU32((uint32_t)magn[i]); 166 frac = (int16_t)((((uint32_t)magn[i] << zeros) 167 & 0x7FFFFFFF) >> 23); 168 assert(frac < 256); 169 // log2(magn(i)) 170 log2 = (int16_t)(((31 - zeros) << 8) 171 + WebRtcNsx_kLogTableFrac[frac]); 172 // log2(magn(i))*log(2) 173 lmagn[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(log2, log2_const, 15); 174 // + log(2^stages) 175 lmagn[i] += logval; 176 } else { 177 lmagn[i] = logval; 178 } 179 } 180 181 int16x4_t Q3_16x4 = vdup_n_s16(3); 182 int16x8_t WIDTHQ8_16x8 = vdupq_n_s16(WIDTH_Q8); 183 int16x8_t WIDTHFACTOR_16x8 = vdupq_n_s16(width_factor); 184 185 int16_t factor = FACTOR_Q7; 186 if (inst->blockIndex < END_STARTUP_LONG) 187 factor = FACTOR_Q7_STARTUP; 188 189 // Loop over simultaneous estimates 190 for (s = 0; s < SIMULT; s++) { 191 offset = s * inst->magnLen; 192 193 // Get counter values from state 194 counter = inst->noiseEstCounter[s]; 195 assert(counter < 201); 196 countDiv = WebRtcNsx_kCounterDiv[counter]; 197 countProd = (int16_t)WEBRTC_SPL_MUL_16_16(counter, countDiv); 198 199 // quant_est(...) 200 int16_t deltaBuff[8]; 201 int16x4_t tmp16x4_0; 202 int16x4_t tmp16x4_1; 203 int16x4_t countDiv_16x4 = vdup_n_s16(countDiv); 204 int16x8_t countProd_16x8 = vdupq_n_s16(countProd); 205 int16x8_t tmp16x8_0 = vdupq_n_s16(countDiv); 206 int16x8_t prod16x8 = vqrdmulhq_s16(WIDTHFACTOR_16x8, tmp16x8_0); 207 int16x8_t tmp16x8_1; 208 int16x8_t tmp16x8_2; 209 int16x8_t tmp16x8_3; 210 // Initialize tmp16x8_4 to zero to avoid compilaton error. 211 int16x8_t tmp16x8_4 = vdupq_n_s16(0); 212 int16x8_t tmp16x8_5; 213 int32x4_t tmp32x4; 214 215 for (i = 0; i < inst->magnLen - 7; i += 8) { 216 // Compute delta. 217 // Smaller step size during startup. This prevents from using 218 // unrealistic values causing overflow. 219 tmp16x8_0 = vdupq_n_s16(factor); 220 vst1q_s16(deltaBuff, tmp16x8_0); 221 222 int j; 223 for (j = 0; j < 8; j++) { 224 if (inst->noiseEstDensity[offset + i + j] > 512) { 225 // Get values for deltaBuff by shifting intead of dividing. 226 int factor = WebRtcSpl_NormW16(inst->noiseEstDensity[offset + i + j]); 227 deltaBuff[j] = (int16_t)(FACTOR_Q16 >> (14 - factor)); 228 } 229 } 230 231 // Update log quantile estimate 232 233 // tmp16 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(delta, countDiv, 14); 234 tmp32x4 = vmull_s16(vld1_s16(&deltaBuff[0]), countDiv_16x4); 235 tmp16x4_1 = vshrn_n_s32(tmp32x4, 14); 236 tmp32x4 = vmull_s16(vld1_s16(&deltaBuff[4]), countDiv_16x4); 237 tmp16x4_0 = vshrn_n_s32(tmp32x4, 14); 238 tmp16x8_0 = vcombine_s16(tmp16x4_1, tmp16x4_0); // Keep for several lines. 239 240 // prepare for the "if" branch 241 // tmp16 += 2; 242 // tmp16_1 = (Word16)(tmp16>>2); 243 tmp16x8_1 = vrshrq_n_s16(tmp16x8_0, 2); 244 245 // inst->noiseEstLogQuantile[offset+i] + tmp16_1; 246 tmp16x8_2 = vld1q_s16(&inst->noiseEstLogQuantile[offset + i]); // Keep 247 tmp16x8_1 = vaddq_s16(tmp16x8_2, tmp16x8_1); // Keep for several lines 248 249 // Prepare for the "else" branch 250 // tmp16 += 1; 251 // tmp16_1 = (Word16)(tmp16>>1); 252 tmp16x8_0 = vrshrq_n_s16(tmp16x8_0, 1); 253 254 // tmp16_2 = (Word16)WEBRTC_SPL_MUL_16_16_RSFT(tmp16_1,3,1); 255 tmp32x4 = vmull_s16(vget_low_s16(tmp16x8_0), Q3_16x4); 256 tmp16x4_1 = vshrn_n_s32(tmp32x4, 1); 257 258 // tmp16_2 = (Word16)WEBRTC_SPL_MUL_16_16_RSFT(tmp16_1,3,1); 259 tmp32x4 = vmull_s16(vget_high_s16(tmp16x8_0), Q3_16x4); 260 tmp16x4_0 = vshrn_n_s32(tmp32x4, 1); 261 262 // inst->noiseEstLogQuantile[offset + i] - tmp16_2; 263 tmp16x8_0 = vcombine_s16(tmp16x4_1, tmp16x4_0); // keep 264 tmp16x8_0 = vsubq_s16(tmp16x8_2, tmp16x8_0); 265 266 // logval is the smallest fixed point representation we can have. Values 267 // below that will correspond to values in the interval [0, 1], which 268 // can't possibly occur. 269 tmp16x8_0 = vmaxq_s16(tmp16x8_0, logval_16x8); 270 271 // Do the if-else branches: 272 tmp16x8_3 = vld1q_s16(&lmagn[i]); // keep for several lines 273 tmp16x8_5 = vsubq_s16(tmp16x8_3, tmp16x8_2); 274 __asm__("vcgt.s16 %q0, %q1, #0"::"w"(tmp16x8_4), "w"(tmp16x8_5)); 275 __asm__("vbit %q0, %q1, %q2":: 276 "w"(tmp16x8_2), "w"(tmp16x8_1), "w"(tmp16x8_4)); 277 __asm__("vbif %q0, %q1, %q2":: 278 "w"(tmp16x8_2), "w"(tmp16x8_0), "w"(tmp16x8_4)); 279 vst1q_s16(&inst->noiseEstLogQuantile[offset + i], tmp16x8_2); 280 281 // Update density estimate 282 // tmp16_1 + tmp16_2 283 tmp16x8_1 = vld1q_s16(&inst->noiseEstDensity[offset + i]); 284 tmp16x8_0 = vqrdmulhq_s16(tmp16x8_1, countProd_16x8); 285 tmp16x8_0 = vaddq_s16(tmp16x8_0, prod16x8); 286 287 // lmagn[i] - inst->noiseEstLogQuantile[offset + i] 288 tmp16x8_3 = vsubq_s16(tmp16x8_3, tmp16x8_2); 289 tmp16x8_3 = vabsq_s16(tmp16x8_3); 290 tmp16x8_4 = vcgtq_s16(WIDTHQ8_16x8, tmp16x8_3); 291 __asm__("vbit %q0, %q1, %q2":: 292 "w"(tmp16x8_1), "w"(tmp16x8_0), "w"(tmp16x8_4)); 293 vst1q_s16(&inst->noiseEstDensity[offset + i], tmp16x8_1); 294 } // End loop over magnitude spectrum 295 296 // Last iteration over magnitude spectrum: 297 // compute delta 298 if (inst->noiseEstDensity[offset + i] > 512) { 299 // Get values for deltaBuff by shifting intead of dividing. 300 int factor = WebRtcSpl_NormW16(inst->noiseEstDensity[offset + i]); 301 delta = (int16_t)(FACTOR_Q16 >> (14 - factor)); 302 } else { 303 delta = FACTOR_Q7; 304 if (inst->blockIndex < END_STARTUP_LONG) { 305 // Smaller step size during startup. This prevents from using 306 // unrealistic values causing overflow. 307 delta = FACTOR_Q7_STARTUP; 308 } 309 } 310 // update log quantile estimate 311 tmp16 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(delta, countDiv, 14); 312 if (lmagn[i] > inst->noiseEstLogQuantile[offset + i]) { 313 // +=QUANTILE*delta/(inst->counter[s]+1) QUANTILE=0.25, =1 in Q2 314 // CounterDiv=1/(inst->counter[s]+1) in Q15 315 tmp16 += 2; 316 tmp16no1 = WEBRTC_SPL_RSHIFT_W16(tmp16, 2); 317 inst->noiseEstLogQuantile[offset + i] += tmp16no1; 318 } else { 319 tmp16 += 1; 320 tmp16no1 = WEBRTC_SPL_RSHIFT_W16(tmp16, 1); 321 // *(1-QUANTILE), in Q2 QUANTILE=0.25, 1-0.25=0.75=3 in Q2 322 tmp16no2 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(tmp16no1, 3, 1); 323 inst->noiseEstLogQuantile[offset + i] -= tmp16no2; 324 if (inst->noiseEstLogQuantile[offset + i] < logval) { 325 // logval is the smallest fixed point representation we can have. 326 // Values below that will correspond to values in the interval 327 // [0, 1], which can't possibly occur. 328 inst->noiseEstLogQuantile[offset + i] = logval; 329 } 330 } 331 332 // update density estimate 333 if (WEBRTC_SPL_ABS_W16(lmagn[i] - inst->noiseEstLogQuantile[offset + i]) 334 < WIDTH_Q8) { 335 tmp16no1 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND( 336 inst->noiseEstDensity[offset + i], countProd, 15); 337 tmp16no2 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND( 338 width_factor, countDiv, 15); 339 inst->noiseEstDensity[offset + i] = tmp16no1 + tmp16no2; 340 } 341 342 343 if (counter >= END_STARTUP_LONG) { 344 inst->noiseEstCounter[s] = 0; 345 if (inst->blockIndex >= END_STARTUP_LONG) { 346 UpdateNoiseEstimateNeon(inst, offset); 347 } 348 } 349 inst->noiseEstCounter[s]++; 350 351 } // end loop over simultaneous estimates 352 353 // Sequentially update the noise during startup 354 if (inst->blockIndex < END_STARTUP_LONG) { 355 UpdateNoiseEstimateNeon(inst, offset); 356 } 357 358 for (i = 0; i < inst->magnLen; i++) { 359 noise[i] = (uint32_t)(inst->noiseEstQuantile[i]); // Q(qNoise) 360 } 361 (*q_noise) = (int16_t)inst->qNoise; 362} 363 364// Filter the data in the frequency domain, and create spectrum. 365void WebRtcNsx_PrepareSpectrumNeon(NsxInst_t* inst, int16_t* freq_buf) { 366 assert(inst->magnLen % 8 == 1); 367 assert(inst->anaLen2 % 16 == 0); 368 369 // (1) Filtering. 370 371 // Fixed point C code for the next block is as follows: 372 // for (i = 0; i < inst->magnLen; i++) { 373 // inst->real[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(inst->real[i], 374 // (int16_t)(inst->noiseSupFilter[i]), 14); // Q(normData-stages) 375 // inst->imag[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(inst->imag[i], 376 // (int16_t)(inst->noiseSupFilter[i]), 14); // Q(normData-stages) 377 // } 378 379 int16_t* preal = &inst->real[0]; 380 int16_t* pimag = &inst->imag[0]; 381 int16_t* pns_filter = (int16_t*)&inst->noiseSupFilter[0]; 382 int16_t* pimag_end = pimag + inst->magnLen - 4; 383 384 while (pimag < pimag_end) { 385 int16x8_t real = vld1q_s16(preal); 386 int16x8_t imag = vld1q_s16(pimag); 387 int16x8_t ns_filter = vld1q_s16(pns_filter); 388 389 int32x4_t tmp_r_0 = vmull_s16(vget_low_s16(real), vget_low_s16(ns_filter)); 390 int32x4_t tmp_i_0 = vmull_s16(vget_low_s16(imag), vget_low_s16(ns_filter)); 391 int32x4_t tmp_r_1 = vmull_s16(vget_high_s16(real), 392 vget_high_s16(ns_filter)); 393 int32x4_t tmp_i_1 = vmull_s16(vget_high_s16(imag), 394 vget_high_s16(ns_filter)); 395 396 int16x4_t result_r_0 = vshrn_n_s32(tmp_r_0, 14); 397 int16x4_t result_i_0 = vshrn_n_s32(tmp_i_0, 14); 398 int16x4_t result_r_1 = vshrn_n_s32(tmp_r_1, 14); 399 int16x4_t result_i_1 = vshrn_n_s32(tmp_i_1, 14); 400 401 vst1q_s16(preal, vcombine_s16(result_r_0, result_r_1)); 402 vst1q_s16(pimag, vcombine_s16(result_i_0, result_i_1)); 403 preal += 8; 404 pimag += 8; 405 pns_filter += 8; 406 } 407 408 // Filter the last element 409 *preal = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(*preal, *pns_filter, 14); 410 *pimag = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(*pimag, *pns_filter, 14); 411 412 // (2) Create spectrum. 413 414 // Fixed point C code for the rest of the function is as follows: 415 // freq_buf[0] = inst->real[0]; 416 // freq_buf[1] = -inst->imag[0]; 417 // for (i = 1, j = 2; i < inst->anaLen2; i += 1, j += 2) { 418 // freq_buf[j] = inst->real[i]; 419 // freq_buf[j + 1] = -inst->imag[i]; 420 // } 421 // freq_buf[inst->anaLen] = inst->real[inst->anaLen2]; 422 // freq_buf[inst->anaLen + 1] = -inst->imag[inst->anaLen2]; 423 424 preal = &inst->real[0]; 425 pimag = &inst->imag[0]; 426 pimag_end = pimag + inst->anaLen2; 427 int16_t * freq_buf_start = freq_buf; 428 while (pimag < pimag_end) { 429 // loop unroll 430 int16x8x2_t real_imag_0; 431 int16x8x2_t real_imag_1; 432 real_imag_0.val[1] = vld1q_s16(pimag); 433 real_imag_0.val[0] = vld1q_s16(preal); 434 preal += 8; 435 pimag += 8; 436 real_imag_1.val[1] = vld1q_s16(pimag); 437 real_imag_1.val[0] = vld1q_s16(preal); 438 preal += 8; 439 pimag += 8; 440 441 real_imag_0.val[1] = vnegq_s16(real_imag_0.val[1]); 442 real_imag_1.val[1] = vnegq_s16(real_imag_1.val[1]); 443 vst2q_s16(freq_buf_start, real_imag_0); 444 freq_buf_start += 16; 445 vst2q_s16(freq_buf_start, real_imag_1); 446 freq_buf_start += 16; 447 } 448 freq_buf[inst->anaLen] = inst->real[inst->anaLen2]; 449 freq_buf[inst->anaLen + 1] = -inst->imag[inst->anaLen2]; 450} 451 452// Denormalize the input buffer. 453void WebRtcNsx_DenormalizeNeon(NsxInst_t* inst, int16_t* in, int factor) { 454 int16_t* ptr_real = &inst->real[0]; 455 int16_t* ptr_in = &in[0]; 456 457 __asm__ __volatile__("vdup.32 q10, %0" :: 458 "r"((int32_t)(factor - inst->normData)) : "q10"); 459 for (; ptr_real < &inst->real[inst->anaLen];) { 460 461 // Loop unrolled once. Both pointers are incremented. 462 __asm__ __volatile__( 463 // tmp32 = WEBRTC_SPL_SHIFT_W32((int32_t)in[j], 464 // factor - inst->normData); 465 "vld2.16 {d24, d25}, [%[ptr_in]]!\n\t" 466 "vmovl.s16 q12, d24\n\t" 467 "vshl.s32 q12, q10\n\t" 468 // inst->real[i] = WebRtcSpl_SatW32ToW16(tmp32); // Q0 469 "vqmovn.s32 d24, q12\n\t" 470 "vst1.16 d24, [%[ptr_real]]!\n\t" 471 472 // tmp32 = WEBRTC_SPL_SHIFT_W32((int32_t)in[j], 473 // factor - inst->normData); 474 "vld2.16 {d22, d23}, [%[ptr_in]]!\n\t" 475 "vmovl.s16 q11, d22\n\t" 476 "vshl.s32 q11, q10\n\t" 477 // inst->real[i] = WebRtcSpl_SatW32ToW16(tmp32); // Q0 478 "vqmovn.s32 d22, q11\n\t" 479 "vst1.16 d22, [%[ptr_real]]!\n\t" 480 481 // Specify constraints. 482 :[ptr_in]"+r"(ptr_in), 483 [ptr_real]"+r"(ptr_real) 484 : 485 :"d22", "d23", "d24", "d25" 486 ); 487 } 488} 489 490// For the noise supress process, synthesis, read out fully processed segment, 491// and update synthesis buffer. 492void WebRtcNsx_SynthesisUpdateNeon(NsxInst_t* inst, 493 int16_t* out_frame, 494 int16_t gain_factor) { 495 int16_t* ptr_real = &inst->real[0]; 496 int16_t* ptr_syn = &inst->synthesisBuffer[0]; 497 const int16_t* ptr_window = &inst->window[0]; 498 499 // synthesis 500 __asm__ __volatile__("vdup.16 d24, %0" : : "r"(gain_factor) : "d24"); 501 // Loop unrolled once. All pointers are incremented in the assembly code. 502 for (; ptr_syn < &inst->synthesisBuffer[inst->anaLen];) { 503 __asm__ __volatile__( 504 // Load variables. 505 "vld1.16 d22, [%[ptr_real]]!\n\t" 506 "vld1.16 d23, [%[ptr_window]]!\n\t" 507 "vld1.16 d25, [%[ptr_syn]]\n\t" 508 // tmp16a = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND( 509 // inst->window[i], inst->real[i], 14); // Q0, window in Q14 510 "vmull.s16 q11, d22, d23\n\t" 511 "vrshrn.i32 d22, q11, #14\n\t" 512 // tmp32 = WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(tmp16a, gain_factor, 13); 513 "vmull.s16 q11, d24, d22\n\t" 514 // tmp16b = WebRtcSpl_SatW32ToW16(tmp32); // Q0 515 "vqrshrn.s32 d22, q11, #13\n\t" 516 // inst->synthesisBuffer[i] = WebRtcSpl_AddSatW16( 517 // inst->synthesisBuffer[i], tmp16b); // Q0 518 "vqadd.s16 d25, d22\n\t" 519 "vst1.16 d25, [%[ptr_syn]]!\n\t" 520 521 // Load variables. 522 "vld1.16 d26, [%[ptr_real]]!\n\t" 523 "vld1.16 d27, [%[ptr_window]]!\n\t" 524 "vld1.16 d28, [%[ptr_syn]]\n\t" 525 // tmp16a = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND( 526 // inst->window[i], inst->real[i], 14); // Q0, window in Q14 527 "vmull.s16 q13, d26, d27\n\t" 528 "vrshrn.i32 d26, q13, #14\n\t" 529 // tmp32 = WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(tmp16a, gain_factor, 13); 530 "vmull.s16 q13, d24, d26\n\t" 531 // tmp16b = WebRtcSpl_SatW32ToW16(tmp32); // Q0 532 "vqrshrn.s32 d26, q13, #13\n\t" 533 // inst->synthesisBuffer[i] = WebRtcSpl_AddSatW16( 534 // inst->synthesisBuffer[i], tmp16b); // Q0 535 "vqadd.s16 d28, d26\n\t" 536 "vst1.16 d28, [%[ptr_syn]]!\n\t" 537 538 // Specify constraints. 539 :[ptr_real]"+r"(ptr_real), 540 [ptr_window]"+r"(ptr_window), 541 [ptr_syn]"+r"(ptr_syn) 542 : 543 :"d22", "d23", "d24", "d25", "d26", "d27", "d28", "q11", "q12", "q13" 544 ); 545 } 546 547 int16_t* ptr_out = &out_frame[0]; 548 ptr_syn = &inst->synthesisBuffer[0]; 549 // read out fully processed segment 550 for (; ptr_syn < &inst->synthesisBuffer[inst->blockLen10ms];) { 551 // Loop unrolled once. Both pointers are incremented in the assembly code. 552 __asm__ __volatile__( 553 // out_frame[i] = inst->synthesisBuffer[i]; // Q0 554 "vld1.16 {d22, d23}, [%[ptr_syn]]!\n\t" 555 "vld1.16 {d24, d25}, [%[ptr_syn]]!\n\t" 556 "vst1.16 {d22, d23}, [%[ptr_out]]!\n\t" 557 "vst1.16 {d24, d25}, [%[ptr_out]]!\n\t" 558 :[ptr_syn]"+r"(ptr_syn), 559 [ptr_out]"+r"(ptr_out) 560 : 561 :"d22", "d23", "d24", "d25" 562 ); 563 } 564 565 // Update synthesis buffer. 566 // C code: 567 // WEBRTC_SPL_MEMCPY_W16(inst->synthesisBuffer, 568 // inst->synthesisBuffer + inst->blockLen10ms, 569 // inst->anaLen - inst->blockLen10ms); 570 ptr_out = &inst->synthesisBuffer[0], 571 ptr_syn = &inst->synthesisBuffer[inst->blockLen10ms]; 572 for (; ptr_syn < &inst->synthesisBuffer[inst->anaLen];) { 573 // Loop unrolled once. Both pointers are incremented in the assembly code. 574 __asm__ __volatile__( 575 "vld1.16 {d22, d23}, [%[ptr_syn]]!\n\t" 576 "vld1.16 {d24, d25}, [%[ptr_syn]]!\n\t" 577 "vst1.16 {d22, d23}, [%[ptr_out]]!\n\t" 578 "vst1.16 {d24, d25}, [%[ptr_out]]!\n\t" 579 :[ptr_syn]"+r"(ptr_syn), 580 [ptr_out]"+r"(ptr_out) 581 : 582 :"d22", "d23", "d24", "d25" 583 ); 584 } 585 586 // C code: 587 // WebRtcSpl_ZerosArrayW16(inst->synthesisBuffer 588 // + inst->anaLen - inst->blockLen10ms, inst->blockLen10ms); 589 __asm__ __volatile__("vdup.16 q10, %0" : : "r"(0) : "q10"); 590 for (; ptr_out < &inst->synthesisBuffer[inst->anaLen];) { 591 // Loop unrolled once. Pointer is incremented in the assembly code. 592 __asm__ __volatile__( 593 "vst1.16 {d20, d21}, [%[ptr_out]]!\n\t" 594 "vst1.16 {d20, d21}, [%[ptr_out]]!\n\t" 595 :[ptr_out]"+r"(ptr_out) 596 : 597 :"d20", "d21" 598 ); 599 } 600} 601 602// Update analysis buffer for lower band, and window data before FFT. 603void WebRtcNsx_AnalysisUpdateNeon(NsxInst_t* inst, 604 int16_t* out, 605 int16_t* new_speech) { 606 607 int16_t* ptr_ana = &inst->analysisBuffer[inst->blockLen10ms]; 608 int16_t* ptr_out = &inst->analysisBuffer[0]; 609 610 // For lower band update analysis buffer. 611 // WEBRTC_SPL_MEMCPY_W16(inst->analysisBuffer, 612 // inst->analysisBuffer + inst->blockLen10ms, 613 // inst->anaLen - inst->blockLen10ms); 614 for (; ptr_out < &inst->analysisBuffer[inst->anaLen - inst->blockLen10ms];) { 615 // Loop unrolled once, so both pointers are incremented by 8 twice. 616 __asm__ __volatile__( 617 "vld1.16 {d20, d21}, [%[ptr_ana]]!\n\t" 618 "vst1.16 {d20, d21}, [%[ptr_out]]!\n\t" 619 "vld1.16 {d22, d23}, [%[ptr_ana]]!\n\t" 620 "vst1.16 {d22, d23}, [%[ptr_out]]!\n\t" 621 :[ptr_ana]"+r"(ptr_ana), 622 [ptr_out]"+r"(ptr_out) 623 : 624 :"d20", "d21", "d22", "d23" 625 ); 626 } 627 628 // WEBRTC_SPL_MEMCPY_W16(inst->analysisBuffer 629 // + inst->anaLen - inst->blockLen10ms, new_speech, inst->blockLen10ms); 630 for (ptr_ana = new_speech; ptr_out < &inst->analysisBuffer[inst->anaLen];) { 631 // Loop unrolled once, so both pointers are incremented by 8 twice. 632 __asm__ __volatile__( 633 "vld1.16 {d20, d21}, [%[ptr_ana]]!\n\t" 634 "vst1.16 {d20, d21}, [%[ptr_out]]!\n\t" 635 "vld1.16 {d22, d23}, [%[ptr_ana]]!\n\t" 636 "vst1.16 {d22, d23}, [%[ptr_out]]!\n\t" 637 :[ptr_ana]"+r"(ptr_ana), 638 [ptr_out]"+r"(ptr_out) 639 : 640 :"d20", "d21", "d22", "d23" 641 ); 642 } 643 644 // Window data before FFT 645 const int16_t* ptr_window = &inst->window[0]; 646 ptr_out = &out[0]; 647 ptr_ana = &inst->analysisBuffer[0]; 648 for (; ptr_out < &out[inst->anaLen];) { 649 650 // Loop unrolled once, so all pointers are incremented by 4 twice. 651 __asm__ __volatile__( 652 "vld1.16 d20, [%[ptr_ana]]!\n\t" 653 "vld1.16 d21, [%[ptr_window]]!\n\t" 654 // out[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND( 655 // inst->window[i], inst->analysisBuffer[i], 14); // Q0 656 "vmull.s16 q10, d20, d21\n\t" 657 "vrshrn.i32 d20, q10, #14\n\t" 658 "vst1.16 d20, [%[ptr_out]]!\n\t" 659 660 "vld1.16 d22, [%[ptr_ana]]!\n\t" 661 "vld1.16 d23, [%[ptr_window]]!\n\t" 662 // out[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND( 663 // inst->window[i], inst->analysisBuffer[i], 14); // Q0 664 "vmull.s16 q11, d22, d23\n\t" 665 "vrshrn.i32 d22, q11, #14\n\t" 666 "vst1.16 d22, [%[ptr_out]]!\n\t" 667 668 // Specify constraints. 669 :[ptr_ana]"+r"(ptr_ana), 670 [ptr_window]"+r"(ptr_window), 671 [ptr_out]"+r"(ptr_out) 672 : 673 :"d20", "d21", "d22", "d23", "q10", "q11" 674 ); 675 } 676} 677 678// Create a complex number buffer (out[]) as the intput (in[]) interleaved with 679// zeros, and normalize it. 680void WebRtcNsx_CreateComplexBufferNeon(NsxInst_t* inst, 681 int16_t* in, 682 int16_t* out) { 683 int16_t* ptr_out = &out[0]; 684 int16_t* ptr_in = &in[0]; 685 686 __asm__ __volatile__("vdup.16 d25, %0" : : "r"(0) : "d25"); 687 __asm__ __volatile__("vdup.16 q10, %0" : : "r"(inst->normData) : "q10"); 688 for (; ptr_in < &in[inst->anaLen];) { 689 690 // Loop unrolled once, so ptr_in is incremented by 8 twice, 691 // and ptr_out is incremented by 8 four times. 692 __asm__ __volatile__( 693 // out[j] = WEBRTC_SPL_LSHIFT_W16(in[i], inst->normData); // Q(normData) 694 "vld1.16 {d22, d23}, [%[ptr_in]]!\n\t" 695 "vshl.s16 q11, q10\n\t" 696 "vmov d24, d23\n\t" 697 698 // out[j + 1] = 0; // Insert zeros in imaginary part 699 "vmov d23, d25\n\t" 700 "vst2.16 {d22, d23}, [%[ptr_out]]!\n\t" 701 "vst2.16 {d24, d25}, [%[ptr_out]]!\n\t" 702 703 // out[j] = WEBRTC_SPL_LSHIFT_W16(in[i], inst->normData); // Q(normData) 704 "vld1.16 {d22, d23}, [%[ptr_in]]!\n\t" 705 "vshl.s16 q11, q10\n\t" 706 "vmov d24, d23\n\t" 707 708 // out[j + 1] = 0; // Insert zeros in imaginary part 709 "vmov d23, d25\n\t" 710 "vst2.16 {d22, d23}, [%[ptr_out]]!\n\t" 711 "vst2.16 {d24, d25}, [%[ptr_out]]!\n\t" 712 713 // Specify constraints. 714 :[ptr_in]"+r"(ptr_in), 715 [ptr_out]"+r"(ptr_out) 716 : 717 :"d22", "d23", "d24", "d25", "q10", "q11" 718 ); 719 } 720} 721