• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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).
17 const int16_t WebRtcNsx_kLogTable[9] = {
18   0, 177, 355, 532, 710, 887, 1065, 1242, 1420
19 };
20 
21 const 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 
38 const 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.
UpdateNoiseEstimateNeon(NoiseSuppressionFixedC * inst,int offset)59 static void UpdateNoiseEstimateNeon(NoiseSuppressionFixedC* 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 = kExp2Const * inst->noiseEstLogQuantile[offset + i];
84     int16x4_t v16x4 = vld1_s16(ptr_noiseEstLogQuantile);
85     int32x4_t v32x4B = vmull_s16(v16x4, kExp2Const16x4);
86 
87     // tmp32no1 = (0x00200000 | (tmp32no2 & 0x001FFFFF)); // 2^21 + frac
88     int32x4_t v32x4A = vandq_s32(v32x4B, constA32x4);
89     v32x4A = vorrq_s32(v32x4A, constB32x4);
90 
91     // tmp16 = (int16_t)(tmp32no2 >> 21);
92     v32x4B = vshrq_n_s32(v32x4B, 21);
93 
94     // tmp16 -= 21;// shift 21 to get result in Q0
95     v32x4B = vsubq_s32(v32x4B, twentyOne32x4);
96 
97     // tmp16 += (int16_t) inst->qNoise;
98     // shift to get result in Q(qNoise)
99     v32x4B = vaddq_s32(v32x4B, qNoise32x4);
100 
101     // if (tmp16 < 0) {
102     //   tmp32no1 >>= -tmp16;
103     // } else {
104     //   tmp32no1 <<= tmp16;
105     // }
106     v32x4B = vshlq_s32(v32x4A, v32x4B);
107 
108     // tmp16 = WebRtcSpl_SatW32ToW16(tmp32no1);
109     v16x4 = vqmovn_s32(v32x4B);
110 
111     //inst->noiseEstQuantile[i] = tmp16;
112     vst1_s16(ptr_noiseEstQuantile, v16x4);
113   }
114 
115   // Last iteration:
116 
117   // inst->quantile[i]=exp(inst->lquantile[offset+i]);
118   // in Q21
119   int32_t tmp32no2 = kExp2Const * *ptr_noiseEstLogQuantile;
120   int32_t tmp32no1 = (0x00200000 | (tmp32no2 & 0x001FFFFF)); // 2^21 + frac
121 
122   tmp16 = (int16_t)(tmp32no2 >> 21);
123   tmp16 -= 21;// shift 21 to get result in Q0
124   tmp16 += (int16_t) inst->qNoise; //shift to get result in Q(qNoise)
125   if (tmp16 < 0) {
126     tmp32no1 >>= -tmp16;
127   } else {
128     tmp32no1 <<= tmp16;
129   }
130   *ptr_noiseEstQuantile = WebRtcSpl_SatW32ToW16(tmp32no1);
131 }
132 
133 // Noise Estimation
WebRtcNsx_NoiseEstimationNeon(NoiseSuppressionFixedC * inst,uint16_t * magn,uint32_t * noise,int16_t * q_noise)134 void WebRtcNsx_NoiseEstimationNeon(NoiseSuppressionFixedC* inst,
135                                    uint16_t* magn,
136                                    uint32_t* noise,
137                                    int16_t* q_noise) {
138   int16_t lmagn[HALF_ANAL_BLOCKL], counter, countDiv;
139   int16_t countProd, delta, zeros, frac;
140   int16_t log2, tabind, logval, tmp16, tmp16no1, tmp16no2;
141   const int16_t log2_const = 22713;
142   const int16_t width_factor = 21845;
143 
144   size_t i, s, offset;
145 
146   tabind = inst->stages - inst->normData;
147   assert(tabind < 9);
148   assert(tabind > -9);
149   if (tabind < 0) {
150     logval = -WebRtcNsx_kLogTable[-tabind];
151   } else {
152     logval = WebRtcNsx_kLogTable[tabind];
153   }
154 
155   int16x8_t logval_16x8 = vdupq_n_s16(logval);
156 
157   // lmagn(i)=log(magn(i))=log(2)*log2(magn(i))
158   // magn is in Q(-stages), and the real lmagn values are:
159   // real_lmagn(i)=log(magn(i)*2^stages)=log(magn(i))+log(2^stages)
160   // lmagn in Q8
161   for (i = 0; i < inst->magnLen; i++) {
162     if (magn[i]) {
163       zeros = WebRtcSpl_NormU32((uint32_t)magn[i]);
164       frac = (int16_t)((((uint32_t)magn[i] << zeros)
165                         & 0x7FFFFFFF) >> 23);
166       assert(frac < 256);
167       // log2(magn(i))
168       log2 = (int16_t)(((31 - zeros) << 8)
169                        + WebRtcNsx_kLogTableFrac[frac]);
170       // log2(magn(i))*log(2)
171       lmagn[i] = (int16_t)((log2 * log2_const) >> 15);
172       // + log(2^stages)
173       lmagn[i] += logval;
174     } else {
175       lmagn[i] = logval;
176     }
177   }
178 
179   int16x4_t Q3_16x4  = vdup_n_s16(3);
180   int16x8_t WIDTHQ8_16x8 = vdupq_n_s16(WIDTH_Q8);
181   int16x8_t WIDTHFACTOR_16x8 = vdupq_n_s16(width_factor);
182 
183   int16_t factor = FACTOR_Q7;
184   if (inst->blockIndex < END_STARTUP_LONG)
185     factor = FACTOR_Q7_STARTUP;
186 
187   // Loop over simultaneous estimates
188   for (s = 0; s < SIMULT; s++) {
189     offset = s * inst->magnLen;
190 
191     // Get counter values from state
192     counter = inst->noiseEstCounter[s];
193     assert(counter < 201);
194     countDiv = WebRtcNsx_kCounterDiv[counter];
195     countProd = (int16_t)(counter * countDiv);
196 
197     // quant_est(...)
198     int16_t deltaBuff[8];
199     int16x4_t tmp16x4_0;
200     int16x4_t tmp16x4_1;
201     int16x4_t countDiv_16x4 = vdup_n_s16(countDiv);
202     int16x8_t countProd_16x8 = vdupq_n_s16(countProd);
203     int16x8_t tmp16x8_0 = vdupq_n_s16(countDiv);
204     int16x8_t prod16x8 = vqrdmulhq_s16(WIDTHFACTOR_16x8, tmp16x8_0);
205     int16x8_t tmp16x8_1;
206     int16x8_t tmp16x8_2;
207     int16x8_t tmp16x8_3;
208     uint16x8_t tmp16x8_4;
209     int32x4_t tmp32x4;
210 
211     for (i = 0; i + 7 < inst->magnLen; i += 8) {
212       // Compute delta.
213       // Smaller step size during startup. This prevents from using
214       // unrealistic values causing overflow.
215       tmp16x8_0 = vdupq_n_s16(factor);
216       vst1q_s16(deltaBuff, tmp16x8_0);
217 
218       int j;
219       for (j = 0; j < 8; j++) {
220         if (inst->noiseEstDensity[offset + i + j] > 512) {
221           // Get values for deltaBuff by shifting intead of dividing.
222           int factor = WebRtcSpl_NormW16(inst->noiseEstDensity[offset + i + j]);
223           deltaBuff[j] = (int16_t)(FACTOR_Q16 >> (14 - factor));
224         }
225       }
226 
227       // Update log quantile estimate
228 
229       // tmp16 = (int16_t)((delta * countDiv) >> 14);
230       tmp32x4 = vmull_s16(vld1_s16(&deltaBuff[0]), countDiv_16x4);
231       tmp16x4_1 = vshrn_n_s32(tmp32x4, 14);
232       tmp32x4 = vmull_s16(vld1_s16(&deltaBuff[4]), countDiv_16x4);
233       tmp16x4_0 = vshrn_n_s32(tmp32x4, 14);
234       tmp16x8_0 = vcombine_s16(tmp16x4_1, tmp16x4_0); // Keep for several lines.
235 
236       // prepare for the "if" branch
237       // tmp16 += 2;
238       // tmp16_1 = (Word16)(tmp16>>2);
239       tmp16x8_1 = vrshrq_n_s16(tmp16x8_0, 2);
240 
241       // inst->noiseEstLogQuantile[offset+i] + tmp16_1;
242       tmp16x8_2 = vld1q_s16(&inst->noiseEstLogQuantile[offset + i]); // Keep
243       tmp16x8_1 = vaddq_s16(tmp16x8_2, tmp16x8_1); // Keep for several lines
244 
245       // Prepare for the "else" branch
246       // tmp16 += 1;
247       // tmp16_1 = (Word16)(tmp16>>1);
248       tmp16x8_0 = vrshrq_n_s16(tmp16x8_0, 1);
249 
250       // tmp16_2 = (int16_t)((tmp16_1 * 3) >> 1);
251       tmp32x4 = vmull_s16(vget_low_s16(tmp16x8_0), Q3_16x4);
252       tmp16x4_1 = vshrn_n_s32(tmp32x4, 1);
253 
254       // tmp16_2 = (int16_t)((tmp16_1 * 3) >> 1);
255       tmp32x4 = vmull_s16(vget_high_s16(tmp16x8_0), Q3_16x4);
256       tmp16x4_0 = vshrn_n_s32(tmp32x4, 1);
257 
258       // inst->noiseEstLogQuantile[offset + i] - tmp16_2;
259       tmp16x8_0 = vcombine_s16(tmp16x4_1, tmp16x4_0); // keep
260       tmp16x8_0 = vsubq_s16(tmp16x8_2, tmp16x8_0);
261 
262       // logval is the smallest fixed point representation we can have. Values
263       // below that will correspond to values in the interval [0, 1], which
264       // can't possibly occur.
265       tmp16x8_0 = vmaxq_s16(tmp16x8_0, logval_16x8);
266 
267       // Do the if-else branches:
268       tmp16x8_3 = vld1q_s16(&lmagn[i]); // keep for several lines
269       tmp16x8_4 = vcgtq_s16(tmp16x8_3, tmp16x8_2);
270       tmp16x8_2 = vbslq_s16(tmp16x8_4, tmp16x8_1, tmp16x8_0);
271       vst1q_s16(&inst->noiseEstLogQuantile[offset + i], tmp16x8_2);
272 
273       // Update density estimate
274       // tmp16_1 + tmp16_2
275       tmp16x8_1 = vld1q_s16(&inst->noiseEstDensity[offset + i]);
276       tmp16x8_0 = vqrdmulhq_s16(tmp16x8_1, countProd_16x8);
277       tmp16x8_0 = vaddq_s16(tmp16x8_0, prod16x8);
278 
279       // lmagn[i] - inst->noiseEstLogQuantile[offset + i]
280       tmp16x8_3 = vsubq_s16(tmp16x8_3, tmp16x8_2);
281       tmp16x8_3 = vabsq_s16(tmp16x8_3);
282       tmp16x8_4 = vcgtq_s16(WIDTHQ8_16x8, tmp16x8_3);
283       tmp16x8_1 = vbslq_s16(tmp16x8_4, tmp16x8_0, tmp16x8_1);
284       vst1q_s16(&inst->noiseEstDensity[offset + i], tmp16x8_1);
285     }  // End loop over magnitude spectrum
286 
287     // Last iteration over magnitude spectrum:
288     // compute delta
289     if (inst->noiseEstDensity[offset + i] > 512) {
290       // Get values for deltaBuff by shifting intead of dividing.
291       int factor = WebRtcSpl_NormW16(inst->noiseEstDensity[offset + i]);
292       delta = (int16_t)(FACTOR_Q16 >> (14 - factor));
293     } else {
294       delta = FACTOR_Q7;
295       if (inst->blockIndex < END_STARTUP_LONG) {
296         // Smaller step size during startup. This prevents from using
297         // unrealistic values causing overflow.
298         delta = FACTOR_Q7_STARTUP;
299       }
300     }
301     // update log quantile estimate
302     tmp16 = (int16_t)((delta * countDiv) >> 14);
303     if (lmagn[i] > inst->noiseEstLogQuantile[offset + i]) {
304       // +=QUANTILE*delta/(inst->counter[s]+1) QUANTILE=0.25, =1 in Q2
305       // CounterDiv=1/(inst->counter[s]+1) in Q15
306       tmp16 += 2;
307       inst->noiseEstLogQuantile[offset + i] += tmp16 / 4;
308     } else {
309       tmp16 += 1;
310       // *(1-QUANTILE), in Q2 QUANTILE=0.25, 1-0.25=0.75=3 in Q2
311       // TODO(bjornv): investigate why we need to truncate twice.
312       tmp16no2 = (int16_t)((tmp16 / 2) * 3 / 2);
313       inst->noiseEstLogQuantile[offset + i] -= tmp16no2;
314       if (inst->noiseEstLogQuantile[offset + i] < logval) {
315         // logval is the smallest fixed point representation we can have.
316         // Values below that will correspond to values in the interval
317         // [0, 1], which can't possibly occur.
318         inst->noiseEstLogQuantile[offset + i] = logval;
319       }
320     }
321 
322     // update density estimate
323     if (WEBRTC_SPL_ABS_W16(lmagn[i] - inst->noiseEstLogQuantile[offset + i])
324         < WIDTH_Q8) {
325       tmp16no1 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(
326                    inst->noiseEstDensity[offset + i], countProd, 15);
327       tmp16no2 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(
328                    width_factor, countDiv, 15);
329       inst->noiseEstDensity[offset + i] = tmp16no1 + tmp16no2;
330     }
331 
332 
333     if (counter >= END_STARTUP_LONG) {
334       inst->noiseEstCounter[s] = 0;
335       if (inst->blockIndex >= END_STARTUP_LONG) {
336         UpdateNoiseEstimateNeon(inst, offset);
337       }
338     }
339     inst->noiseEstCounter[s]++;
340 
341   }  // end loop over simultaneous estimates
342 
343   // Sequentially update the noise during startup
344   if (inst->blockIndex < END_STARTUP_LONG) {
345     UpdateNoiseEstimateNeon(inst, offset);
346   }
347 
348   for (i = 0; i < inst->magnLen; i++) {
349     noise[i] = (uint32_t)(inst->noiseEstQuantile[i]); // Q(qNoise)
350   }
351   (*q_noise) = (int16_t)inst->qNoise;
352 }
353 
354 // Filter the data in the frequency domain, and create spectrum.
WebRtcNsx_PrepareSpectrumNeon(NoiseSuppressionFixedC * inst,int16_t * freq_buf)355 void WebRtcNsx_PrepareSpectrumNeon(NoiseSuppressionFixedC* inst,
356                                    int16_t* freq_buf) {
357   assert(inst->magnLen % 8 == 1);
358   assert(inst->anaLen2 % 16 == 0);
359 
360   // (1) Filtering.
361 
362   // Fixed point C code for the next block is as follows:
363   // for (i = 0; i < inst->magnLen; i++) {
364   //   inst->real[i] = (int16_t)((inst->real[i] *
365   //      (int16_t)(inst->noiseSupFilter[i])) >> 14);  // Q(normData-stages)
366   //   inst->imag[i] = (int16_t)((inst->imag[i] *
367   //      (int16_t)(inst->noiseSupFilter[i])) >> 14);  // Q(normData-stages)
368   // }
369 
370   int16_t* preal = &inst->real[0];
371   int16_t* pimag = &inst->imag[0];
372   int16_t* pns_filter = (int16_t*)&inst->noiseSupFilter[0];
373   int16_t* pimag_end = pimag + inst->magnLen - 4;
374 
375   while (pimag < pimag_end) {
376     int16x8_t real = vld1q_s16(preal);
377     int16x8_t imag = vld1q_s16(pimag);
378     int16x8_t ns_filter = vld1q_s16(pns_filter);
379 
380     int32x4_t tmp_r_0 = vmull_s16(vget_low_s16(real), vget_low_s16(ns_filter));
381     int32x4_t tmp_i_0 = vmull_s16(vget_low_s16(imag), vget_low_s16(ns_filter));
382     int32x4_t tmp_r_1 = vmull_s16(vget_high_s16(real),
383                                   vget_high_s16(ns_filter));
384     int32x4_t tmp_i_1 = vmull_s16(vget_high_s16(imag),
385                                   vget_high_s16(ns_filter));
386 
387     int16x4_t result_r_0 = vshrn_n_s32(tmp_r_0, 14);
388     int16x4_t result_i_0 = vshrn_n_s32(tmp_i_0, 14);
389     int16x4_t result_r_1 = vshrn_n_s32(tmp_r_1, 14);
390     int16x4_t result_i_1 = vshrn_n_s32(tmp_i_1, 14);
391 
392     vst1q_s16(preal, vcombine_s16(result_r_0, result_r_1));
393     vst1q_s16(pimag, vcombine_s16(result_i_0, result_i_1));
394     preal += 8;
395     pimag += 8;
396     pns_filter += 8;
397   }
398 
399   // Filter the last element
400   *preal = (int16_t)((*preal * *pns_filter) >> 14);
401   *pimag = (int16_t)((*pimag * *pns_filter) >> 14);
402 
403   // (2) Create spectrum.
404 
405   // Fixed point C code for the rest of the function is as follows:
406   // freq_buf[0] = inst->real[0];
407   // freq_buf[1] = -inst->imag[0];
408   // for (i = 1, j = 2; i < inst->anaLen2; i += 1, j += 2) {
409   //   freq_buf[j] = inst->real[i];
410   //   freq_buf[j + 1] = -inst->imag[i];
411   // }
412   // freq_buf[inst->anaLen] = inst->real[inst->anaLen2];
413   // freq_buf[inst->anaLen + 1] = -inst->imag[inst->anaLen2];
414 
415   preal = &inst->real[0];
416   pimag = &inst->imag[0];
417   pimag_end = pimag + inst->anaLen2;
418   int16_t * freq_buf_start = freq_buf;
419   while (pimag < pimag_end) {
420     // loop unroll
421     int16x8x2_t real_imag_0;
422     int16x8x2_t real_imag_1;
423     real_imag_0.val[1] = vld1q_s16(pimag);
424     real_imag_0.val[0] = vld1q_s16(preal);
425     preal += 8;
426     pimag += 8;
427     real_imag_1.val[1] = vld1q_s16(pimag);
428     real_imag_1.val[0] = vld1q_s16(preal);
429     preal += 8;
430     pimag += 8;
431 
432     real_imag_0.val[1] = vnegq_s16(real_imag_0.val[1]);
433     real_imag_1.val[1] = vnegq_s16(real_imag_1.val[1]);
434     vst2q_s16(freq_buf_start, real_imag_0);
435     freq_buf_start += 16;
436     vst2q_s16(freq_buf_start, real_imag_1);
437     freq_buf_start += 16;
438   }
439   freq_buf[inst->anaLen] = inst->real[inst->anaLen2];
440   freq_buf[inst->anaLen + 1] = -inst->imag[inst->anaLen2];
441 }
442 
443 // For the noise supress process, synthesis, read out fully processed segment,
444 // and update synthesis buffer.
WebRtcNsx_SynthesisUpdateNeon(NoiseSuppressionFixedC * inst,int16_t * out_frame,int16_t gain_factor)445 void WebRtcNsx_SynthesisUpdateNeon(NoiseSuppressionFixedC* inst,
446                                    int16_t* out_frame,
447                                    int16_t gain_factor) {
448   assert(inst->anaLen % 16 == 0);
449   assert(inst->blockLen10ms % 16 == 0);
450 
451   int16_t* preal_start = inst->real;
452   const int16_t* pwindow = inst->window;
453   int16_t* preal_end = preal_start + inst->anaLen;
454   int16_t* psynthesis_buffer = inst->synthesisBuffer;
455 
456   while (preal_start < preal_end) {
457     // Loop unroll.
458     int16x8_t window_0 = vld1q_s16(pwindow);
459     int16x8_t real_0 = vld1q_s16(preal_start);
460     int16x8_t synthesis_buffer_0 = vld1q_s16(psynthesis_buffer);
461 
462     int16x8_t window_1 = vld1q_s16(pwindow + 8);
463     int16x8_t real_1 = vld1q_s16(preal_start + 8);
464     int16x8_t synthesis_buffer_1 = vld1q_s16(psynthesis_buffer + 8);
465 
466     int32x4_t tmp32a_0_low = vmull_s16(vget_low_s16(real_0),
467                                        vget_low_s16(window_0));
468     int32x4_t tmp32a_0_high = vmull_s16(vget_high_s16(real_0),
469                                         vget_high_s16(window_0));
470 
471     int32x4_t tmp32a_1_low = vmull_s16(vget_low_s16(real_1),
472                                        vget_low_s16(window_1));
473     int32x4_t tmp32a_1_high = vmull_s16(vget_high_s16(real_1),
474                                         vget_high_s16(window_1));
475 
476     int16x4_t tmp16a_0_low = vqrshrn_n_s32(tmp32a_0_low, 14);
477     int16x4_t tmp16a_0_high = vqrshrn_n_s32(tmp32a_0_high, 14);
478 
479     int16x4_t tmp16a_1_low = vqrshrn_n_s32(tmp32a_1_low, 14);
480     int16x4_t tmp16a_1_high = vqrshrn_n_s32(tmp32a_1_high, 14);
481 
482     int32x4_t tmp32b_0_low = vmull_n_s16(tmp16a_0_low, gain_factor);
483     int32x4_t tmp32b_0_high = vmull_n_s16(tmp16a_0_high, gain_factor);
484 
485     int32x4_t tmp32b_1_low = vmull_n_s16(tmp16a_1_low, gain_factor);
486     int32x4_t tmp32b_1_high = vmull_n_s16(tmp16a_1_high, gain_factor);
487 
488     int16x4_t tmp16b_0_low = vqrshrn_n_s32(tmp32b_0_low, 13);
489     int16x4_t tmp16b_0_high = vqrshrn_n_s32(tmp32b_0_high, 13);
490 
491     int16x4_t tmp16b_1_low = vqrshrn_n_s32(tmp32b_1_low, 13);
492     int16x4_t tmp16b_1_high = vqrshrn_n_s32(tmp32b_1_high, 13);
493 
494     synthesis_buffer_0 = vqaddq_s16(vcombine_s16(tmp16b_0_low, tmp16b_0_high),
495                                     synthesis_buffer_0);
496     synthesis_buffer_1 = vqaddq_s16(vcombine_s16(tmp16b_1_low, tmp16b_1_high),
497                                     synthesis_buffer_1);
498     vst1q_s16(psynthesis_buffer, synthesis_buffer_0);
499     vst1q_s16(psynthesis_buffer + 8, synthesis_buffer_1);
500 
501     pwindow += 16;
502     preal_start += 16;
503     psynthesis_buffer += 16;
504   }
505 
506   // Read out fully processed segment.
507   int16_t * p_start = inst->synthesisBuffer;
508   int16_t * p_end = inst->synthesisBuffer + inst->blockLen10ms;
509   int16_t * p_frame = out_frame;
510   while (p_start < p_end) {
511     int16x8_t frame_0 = vld1q_s16(p_start);
512     vst1q_s16(p_frame, frame_0);
513     p_start += 8;
514     p_frame += 8;
515   }
516 
517   // Update synthesis buffer.
518   int16_t* p_start_src = inst->synthesisBuffer + inst->blockLen10ms;
519   int16_t* p_end_src = inst->synthesisBuffer + inst->anaLen;
520   int16_t* p_start_dst = inst->synthesisBuffer;
521   while (p_start_src < p_end_src) {
522     int16x8_t frame = vld1q_s16(p_start_src);
523     vst1q_s16(p_start_dst, frame);
524     p_start_src += 8;
525     p_start_dst += 8;
526   }
527 
528   p_start = inst->synthesisBuffer + inst->anaLen - inst->blockLen10ms;
529   p_end = p_start + inst->blockLen10ms;
530   int16x8_t zero = vdupq_n_s16(0);
531   for (;p_start < p_end; p_start += 8) {
532     vst1q_s16(p_start, zero);
533   }
534 }
535 
536 // Update analysis buffer for lower band, and window data before FFT.
WebRtcNsx_AnalysisUpdateNeon(NoiseSuppressionFixedC * inst,int16_t * out,int16_t * new_speech)537 void WebRtcNsx_AnalysisUpdateNeon(NoiseSuppressionFixedC* inst,
538                                   int16_t* out,
539                                   int16_t* new_speech) {
540   assert(inst->blockLen10ms % 16 == 0);
541   assert(inst->anaLen % 16 == 0);
542 
543   // For lower band update analysis buffer.
544   // memcpy(inst->analysisBuffer, inst->analysisBuffer + inst->blockLen10ms,
545   //     (inst->anaLen - inst->blockLen10ms) * sizeof(*inst->analysisBuffer));
546   int16_t* p_start_src = inst->analysisBuffer + inst->blockLen10ms;
547   int16_t* p_end_src = inst->analysisBuffer + inst->anaLen;
548   int16_t* p_start_dst = inst->analysisBuffer;
549   while (p_start_src < p_end_src) {
550     int16x8_t frame = vld1q_s16(p_start_src);
551     vst1q_s16(p_start_dst, frame);
552 
553     p_start_src += 8;
554     p_start_dst += 8;
555   }
556 
557   // memcpy(inst->analysisBuffer + inst->anaLen - inst->blockLen10ms,
558   //     new_speech, inst->blockLen10ms * sizeof(*inst->analysisBuffer));
559   p_start_src = new_speech;
560   p_end_src = new_speech + inst->blockLen10ms;
561   p_start_dst = inst->analysisBuffer + inst->anaLen - inst->blockLen10ms;
562   while (p_start_src < p_end_src) {
563     int16x8_t frame = vld1q_s16(p_start_src);
564     vst1q_s16(p_start_dst, frame);
565 
566     p_start_src += 8;
567     p_start_dst += 8;
568   }
569 
570   // Window data before FFT.
571   int16_t* p_start_window = (int16_t*) inst->window;
572   int16_t* p_start_buffer = inst->analysisBuffer;
573   int16_t* p_start_out = out;
574   const int16_t* p_end_out = out + inst->anaLen;
575 
576   // Load the first element to reduce pipeline bubble.
577   int16x8_t window = vld1q_s16(p_start_window);
578   int16x8_t buffer = vld1q_s16(p_start_buffer);
579   p_start_window += 8;
580   p_start_buffer += 8;
581 
582   while (p_start_out < p_end_out) {
583     // Unroll loop.
584     int32x4_t tmp32_low = vmull_s16(vget_low_s16(window), vget_low_s16(buffer));
585     int32x4_t tmp32_high = vmull_s16(vget_high_s16(window),
586                                      vget_high_s16(buffer));
587     window = vld1q_s16(p_start_window);
588     buffer = vld1q_s16(p_start_buffer);
589 
590     int16x4_t result_low = vrshrn_n_s32(tmp32_low, 14);
591     int16x4_t result_high = vrshrn_n_s32(tmp32_high, 14);
592     vst1q_s16(p_start_out, vcombine_s16(result_low, result_high));
593 
594     p_start_buffer += 8;
595     p_start_window += 8;
596     p_start_out += 8;
597   }
598 }
599