xref: /aosp_15_r20/external/webrtc/modules/audio_processing/agc/legacy/digital_agc.cc (revision d9f758449e529ab9291ac668be2861e7a55c2422)
1 /*
2  *  Copyright (c) 2011 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 "modules/audio_processing/agc/legacy/digital_agc.h"
12 
13 #include <string.h>
14 
15 #include "modules/audio_processing/agc/legacy/gain_control.h"
16 #include "rtc_base/checks.h"
17 
18 namespace webrtc {
19 
20 namespace {
21 
22 // To generate the gaintable, copy&paste the following lines to a Matlab window:
23 // MaxGain = 6; MinGain = 0; CompRatio = 3; Knee = 1;
24 // zeros = 0:31; lvl = 2.^(1-zeros);
25 // A = -10*log10(lvl) * (CompRatio - 1) / CompRatio;
26 // B = MaxGain - MinGain;
27 // gains = round(2^16*10.^(0.05 * (MinGain + B * (
28 // log(exp(-Knee*A)+exp(-Knee*B)) - log(1+exp(-Knee*B)) ) /
29 // log(1/(1+exp(Knee*B))))));
30 // fprintf(1, '\t%i, %i, %i, %i,\n', gains);
31 // % Matlab code for plotting the gain and input/output level characteristic
32 // (copy/paste the following 3 lines):
33 // in = 10*log10(lvl); out = 20*log10(gains/65536);
34 // subplot(121); plot(in, out); axis([-30, 0, -5, 20]); grid on; xlabel('Input
35 // (dB)'); ylabel('Gain (dB)');
36 // subplot(122); plot(in, in+out); axis([-30, 0, -30, 5]); grid on;
37 // xlabel('Input (dB)'); ylabel('Output (dB)');
38 // zoom on;
39 
40 // Generator table for y=log2(1+e^x) in Q8.
41 enum { kGenFuncTableSize = 128 };
42 static const uint16_t kGenFuncTable[kGenFuncTableSize] = {
43     256,   485,   786,   1126,  1484,  1849,  2217,  2586,  2955,  3324,  3693,
44     4063,  4432,  4801,  5171,  5540,  5909,  6279,  6648,  7017,  7387,  7756,
45     8125,  8495,  8864,  9233,  9603,  9972,  10341, 10711, 11080, 11449, 11819,
46     12188, 12557, 12927, 13296, 13665, 14035, 14404, 14773, 15143, 15512, 15881,
47     16251, 16620, 16989, 17359, 17728, 18097, 18466, 18836, 19205, 19574, 19944,
48     20313, 20682, 21052, 21421, 21790, 22160, 22529, 22898, 23268, 23637, 24006,
49     24376, 24745, 25114, 25484, 25853, 26222, 26592, 26961, 27330, 27700, 28069,
50     28438, 28808, 29177, 29546, 29916, 30285, 30654, 31024, 31393, 31762, 32132,
51     32501, 32870, 33240, 33609, 33978, 34348, 34717, 35086, 35456, 35825, 36194,
52     36564, 36933, 37302, 37672, 38041, 38410, 38780, 39149, 39518, 39888, 40257,
53     40626, 40996, 41365, 41734, 42104, 42473, 42842, 43212, 43581, 43950, 44320,
54     44689, 45058, 45428, 45797, 46166, 46536, 46905};
55 
56 static const int16_t kAvgDecayTime = 250;  // frames; < 3000
57 
58 // the 32 most significant bits of A(19) * B(26) >> 13
59 #define AGC_MUL32(A, B) (((B) >> 13) * (A) + (((0x00001FFF & (B)) * (A)) >> 13))
60 // C + the 32 most significant bits of A * B
61 #define AGC_SCALEDIFF32(A, B, C) \
62   ((C) + ((B) >> 16) * (A) + (((0x0000FFFF & (B)) * (A)) >> 16))
63 
64 }  // namespace
65 
WebRtcAgc_CalculateGainTable(int32_t * gainTable,int16_t digCompGaindB,int16_t targetLevelDbfs,uint8_t limiterEnable,int16_t analogTarget)66 int32_t WebRtcAgc_CalculateGainTable(int32_t* gainTable,       // Q16
67                                      int16_t digCompGaindB,    // Q0
68                                      int16_t targetLevelDbfs,  // Q0
69                                      uint8_t limiterEnable,
70                                      int16_t analogTarget) {  // Q0
71   // This function generates the compressor gain table used in the fixed digital
72   // part.
73   uint32_t tmpU32no1, tmpU32no2, absInLevel, logApprox;
74   int32_t inLevel, limiterLvl;
75   int32_t tmp32, tmp32no1, tmp32no2, numFIX, den, y32;
76   const uint16_t kLog10 = 54426;    // log2(10)     in Q14
77   const uint16_t kLog10_2 = 49321;  // 10*log10(2)  in Q14
78   const uint16_t kLogE_1 = 23637;   // log2(e)      in Q14
79   uint16_t constMaxGain;
80   uint16_t tmpU16, intPart, fracPart;
81   const int16_t kCompRatio = 3;
82   int16_t limiterOffset = 0;  // Limiter offset
83   int16_t limiterIdx, limiterLvlX;
84   int16_t constLinApprox, maxGain, diffGain;
85   int16_t i, tmp16, tmp16no1;
86   int zeros, zerosScale;
87 
88   // Constants
89   //    kLogE_1 = 23637; // log2(e)      in Q14
90   //    kLog10 = 54426; // log2(10)     in Q14
91   //    kLog10_2 = 49321; // 10*log10(2)  in Q14
92 
93   // Calculate maximum digital gain and zero gain level
94   tmp32no1 = (digCompGaindB - analogTarget) * (kCompRatio - 1);
95   tmp16no1 = analogTarget - targetLevelDbfs;
96   tmp16no1 +=
97       WebRtcSpl_DivW32W16ResW16(tmp32no1 + (kCompRatio >> 1), kCompRatio);
98   maxGain = WEBRTC_SPL_MAX(tmp16no1, (analogTarget - targetLevelDbfs));
99   tmp32no1 = maxGain * kCompRatio;
100   if ((digCompGaindB <= analogTarget) && (limiterEnable)) {
101     limiterOffset = 0;
102   }
103 
104   // Calculate the difference between maximum gain and gain at 0dB0v
105   tmp32no1 = digCompGaindB * (kCompRatio - 1);
106   diffGain =
107       WebRtcSpl_DivW32W16ResW16(tmp32no1 + (kCompRatio >> 1), kCompRatio);
108   if (diffGain < 0 || diffGain >= kGenFuncTableSize) {
109     RTC_DCHECK(0);
110     return -1;
111   }
112 
113   // Calculate the limiter level and index:
114   //  limiterLvlX = analogTarget - limiterOffset
115   //  limiterLvl  = targetLevelDbfs + limiterOffset/compRatio
116   limiterLvlX = analogTarget - limiterOffset;
117   limiterIdx = 2 + WebRtcSpl_DivW32W16ResW16((int32_t)limiterLvlX * (1 << 13),
118                                              kLog10_2 / 2);
119   tmp16no1 =
120       WebRtcSpl_DivW32W16ResW16(limiterOffset + (kCompRatio >> 1), kCompRatio);
121   limiterLvl = targetLevelDbfs + tmp16no1;
122 
123   // Calculate (through table lookup):
124   //  constMaxGain = log2(1+2^(log2(e)*diffGain)); (in Q8)
125   constMaxGain = kGenFuncTable[diffGain];  // in Q8
126 
127   // Calculate a parameter used to approximate the fractional part of 2^x with a
128   // piecewise linear function in Q14:
129   //  constLinApprox = round(3/2*(4*(3-2*sqrt(2))/(log(2)^2)-0.5)*2^14);
130   constLinApprox = 22817;  // in Q14
131 
132   // Calculate a denominator used in the exponential part to convert from dB to
133   // linear scale:
134   //  den = 20*constMaxGain (in Q8)
135   den = WEBRTC_SPL_MUL_16_U16(20, constMaxGain);  // in Q8
136 
137   for (i = 0; i < 32; i++) {
138     // Calculate scaled input level (compressor):
139     //  inLevel =
140     //  fix((-constLog10_2*(compRatio-1)*(1-i)+fix(compRatio/2))/compRatio)
141     tmp16 = (int16_t)((kCompRatio - 1) * (i - 1));       // Q0
142     tmp32 = WEBRTC_SPL_MUL_16_U16(tmp16, kLog10_2) + 1;  // Q14
143     inLevel = WebRtcSpl_DivW32W16(tmp32, kCompRatio);    // Q14
144 
145     // Calculate diffGain-inLevel, to map using the genFuncTable
146     inLevel = (int32_t)diffGain * (1 << 14) - inLevel;  // Q14
147 
148     // Make calculations on abs(inLevel) and compensate for the sign afterwards.
149     absInLevel = (uint32_t)WEBRTC_SPL_ABS_W32(inLevel);  // Q14
150 
151     // LUT with interpolation
152     intPart = (uint16_t)(absInLevel >> 14);
153     fracPart =
154         (uint16_t)(absInLevel & 0x00003FFF);  // extract the fractional part
155     tmpU16 = kGenFuncTable[intPart + 1] - kGenFuncTable[intPart];  // Q8
156     tmpU32no1 = tmpU16 * fracPart;                                 // Q22
157     tmpU32no1 += (uint32_t)kGenFuncTable[intPart] << 14;           // Q22
158     logApprox = tmpU32no1 >> 8;                                    // Q14
159     // Compensate for negative exponent using the relation:
160     //  log2(1 + 2^-x) = log2(1 + 2^x) - x
161     if (inLevel < 0) {
162       zeros = WebRtcSpl_NormU32(absInLevel);
163       zerosScale = 0;
164       if (zeros < 15) {
165         // Not enough space for multiplication
166         tmpU32no2 = absInLevel >> (15 - zeros);                 // Q(zeros-1)
167         tmpU32no2 = WEBRTC_SPL_UMUL_32_16(tmpU32no2, kLogE_1);  // Q(zeros+13)
168         if (zeros < 9) {
169           zerosScale = 9 - zeros;
170           tmpU32no1 >>= zerosScale;  // Q(zeros+13)
171         } else {
172           tmpU32no2 >>= zeros - 9;  // Q22
173         }
174       } else {
175         tmpU32no2 = WEBRTC_SPL_UMUL_32_16(absInLevel, kLogE_1);  // Q28
176         tmpU32no2 >>= 6;                                         // Q22
177       }
178       logApprox = 0;
179       if (tmpU32no2 < tmpU32no1) {
180         logApprox = (tmpU32no1 - tmpU32no2) >> (8 - zerosScale);  // Q14
181       }
182     }
183     numFIX = (maxGain * constMaxGain) * (1 << 6);  // Q14
184     numFIX -= (int32_t)logApprox * diffGain;       // Q14
185 
186     // Calculate ratio
187     // Shift `numFIX` as much as possible.
188     // Ensure we avoid wrap-around in `den` as well.
189     if (numFIX > (den >> 8) || -numFIX > (den >> 8)) {  // `den` is Q8.
190       zeros = WebRtcSpl_NormW32(numFIX);
191     } else {
192       zeros = WebRtcSpl_NormW32(den) + 8;
193     }
194     numFIX *= 1 << zeros;  // Q(14+zeros)
195 
196     // Shift den so we end up in Qy1
197     tmp32no1 = WEBRTC_SPL_SHIFT_W32(den, zeros - 9);  // Q(zeros - 1)
198     y32 = numFIX / tmp32no1;                          // in Q15
199     // This is to do rounding in Q14.
200     y32 = y32 >= 0 ? (y32 + 1) >> 1 : -((-y32 + 1) >> 1);
201 
202     if (limiterEnable && (i < limiterIdx)) {
203       tmp32 = WEBRTC_SPL_MUL_16_U16(i - 1, kLog10_2);  // Q14
204       tmp32 -= limiterLvl * (1 << 14);                 // Q14
205       y32 = WebRtcSpl_DivW32W16(tmp32 + 10, 20);
206     }
207     if (y32 > 39000) {
208       tmp32 = (y32 >> 1) * kLog10 + 4096;  // in Q27
209       tmp32 >>= 13;                        // In Q14.
210     } else {
211       tmp32 = y32 * kLog10 + 8192;  // in Q28
212       tmp32 >>= 14;                 // In Q14.
213     }
214     tmp32 += 16 << 14;  // in Q14 (Make sure final output is in Q16)
215 
216     // Calculate power
217     if (tmp32 > 0) {
218       intPart = (int16_t)(tmp32 >> 14);
219       fracPart = (uint16_t)(tmp32 & 0x00003FFF);  // in Q14
220       if ((fracPart >> 13) != 0) {
221         tmp16 = (2 << 14) - constLinApprox;
222         tmp32no2 = (1 << 14) - fracPart;
223         tmp32no2 *= tmp16;
224         tmp32no2 >>= 13;
225         tmp32no2 = (1 << 14) - tmp32no2;
226       } else {
227         tmp16 = constLinApprox - (1 << 14);
228         tmp32no2 = (fracPart * tmp16) >> 13;
229       }
230       fracPart = (uint16_t)tmp32no2;
231       gainTable[i] =
232           (1 << intPart) + WEBRTC_SPL_SHIFT_W32(fracPart, intPart - 14);
233     } else {
234       gainTable[i] = 0;
235     }
236   }
237 
238   return 0;
239 }
240 
WebRtcAgc_InitDigital(DigitalAgc * stt,int16_t agcMode)241 int32_t WebRtcAgc_InitDigital(DigitalAgc* stt, int16_t agcMode) {
242   if (agcMode == kAgcModeFixedDigital) {
243     // start at minimum to find correct gain faster
244     stt->capacitorSlow = 0;
245   } else {
246     // start out with 0 dB gain
247     stt->capacitorSlow = 134217728;  // (int32_t)(0.125f * 32768.0f * 32768.0f);
248   }
249   stt->capacitorFast = 0;
250   stt->gain = 65536;
251   stt->gatePrevious = 0;
252   stt->agcMode = agcMode;
253 
254   // initialize VADs
255   WebRtcAgc_InitVad(&stt->vadNearend);
256   WebRtcAgc_InitVad(&stt->vadFarend);
257 
258   return 0;
259 }
260 
WebRtcAgc_AddFarendToDigital(DigitalAgc * stt,const int16_t * in_far,size_t nrSamples)261 int32_t WebRtcAgc_AddFarendToDigital(DigitalAgc* stt,
262                                      const int16_t* in_far,
263                                      size_t nrSamples) {
264   RTC_DCHECK(stt);
265   // VAD for far end
266   WebRtcAgc_ProcessVad(&stt->vadFarend, in_far, nrSamples);
267 
268   return 0;
269 }
270 
271 // Gains is an 11 element long array (one value per ms, incl start & end).
WebRtcAgc_ComputeDigitalGains(DigitalAgc * stt,const int16_t * const * in_near,size_t num_bands,uint32_t FS,int16_t lowlevelSignal,int32_t gains[11])272 int32_t WebRtcAgc_ComputeDigitalGains(DigitalAgc* stt,
273                                       const int16_t* const* in_near,
274                                       size_t num_bands,
275                                       uint32_t FS,
276                                       int16_t lowlevelSignal,
277                                       int32_t gains[11]) {
278   int32_t tmp32;
279   int32_t env[10];
280   int32_t max_nrg;
281   int32_t cur_level;
282   int32_t gain32;
283   int16_t logratio;
284   int16_t lower_thr, upper_thr;
285   int16_t zeros = 0, zeros_fast, frac = 0;
286   int16_t decay;
287   int16_t gate, gain_adj;
288   int16_t k;
289   size_t n, L;
290 
291   // determine number of samples per ms
292   if (FS == 8000) {
293     L = 8;
294   } else if (FS == 16000 || FS == 32000 || FS == 48000) {
295     L = 16;
296   } else {
297     return -1;
298   }
299 
300   // VAD for near end
301   logratio = WebRtcAgc_ProcessVad(&stt->vadNearend, in_near[0], L * 10);
302 
303   // Account for far end VAD
304   if (stt->vadFarend.counter > 10) {
305     tmp32 = 3 * logratio;
306     logratio = (int16_t)((tmp32 - stt->vadFarend.logRatio) >> 2);
307   }
308 
309   // Determine decay factor depending on VAD
310   //  upper_thr = 1.0f;
311   //  lower_thr = 0.25f;
312   upper_thr = 1024;  // Q10
313   lower_thr = 0;     // Q10
314   if (logratio > upper_thr) {
315     // decay = -2^17 / DecayTime;  ->  -65
316     decay = -65;
317   } else if (logratio < lower_thr) {
318     decay = 0;
319   } else {
320     // decay = (int16_t)(((lower_thr - logratio)
321     //       * (2^27/(DecayTime*(upper_thr-lower_thr)))) >> 10);
322     // SUBSTITUTED: 2^27/(DecayTime*(upper_thr-lower_thr))  ->  65
323     tmp32 = (lower_thr - logratio) * 65;
324     decay = (int16_t)(tmp32 >> 10);
325   }
326 
327   // adjust decay factor for long silence (detected as low standard deviation)
328   // This is only done in the adaptive modes
329   if (stt->agcMode != kAgcModeFixedDigital) {
330     if (stt->vadNearend.stdLongTerm < 4000) {
331       decay = 0;
332     } else if (stt->vadNearend.stdLongTerm < 8096) {
333       // decay = (int16_t)(((stt->vadNearend.stdLongTerm - 4000) * decay) >>
334       // 12);
335       tmp32 = (stt->vadNearend.stdLongTerm - 4000) * decay;
336       decay = (int16_t)(tmp32 >> 12);
337     }
338 
339     if (lowlevelSignal != 0) {
340       decay = 0;
341     }
342   }
343   // Find max amplitude per sub frame
344   // iterate over sub frames
345   for (k = 0; k < 10; k++) {
346     // iterate over samples
347     max_nrg = 0;
348     for (n = 0; n < L; n++) {
349       int32_t nrg = in_near[0][k * L + n] * in_near[0][k * L + n];
350       if (nrg > max_nrg) {
351         max_nrg = nrg;
352       }
353     }
354     env[k] = max_nrg;
355   }
356 
357   // Calculate gain per sub frame
358   gains[0] = stt->gain;
359   for (k = 0; k < 10; k++) {
360     // Fast envelope follower
361     //  decay time = -131000 / -1000 = 131 (ms)
362     stt->capacitorFast =
363         AGC_SCALEDIFF32(-1000, stt->capacitorFast, stt->capacitorFast);
364     if (env[k] > stt->capacitorFast) {
365       stt->capacitorFast = env[k];
366     }
367     // Slow envelope follower
368     if (env[k] > stt->capacitorSlow) {
369       // increase capacitorSlow
370       stt->capacitorSlow = AGC_SCALEDIFF32(500, (env[k] - stt->capacitorSlow),
371                                            stt->capacitorSlow);
372     } else {
373       // decrease capacitorSlow
374       stt->capacitorSlow =
375           AGC_SCALEDIFF32(decay, stt->capacitorSlow, stt->capacitorSlow);
376     }
377 
378     // use maximum of both capacitors as current level
379     if (stt->capacitorFast > stt->capacitorSlow) {
380       cur_level = stt->capacitorFast;
381     } else {
382       cur_level = stt->capacitorSlow;
383     }
384     // Translate signal level into gain, using a piecewise linear approximation
385     // find number of leading zeros
386     zeros = WebRtcSpl_NormU32((uint32_t)cur_level);
387     if (cur_level == 0) {
388       zeros = 31;
389     }
390     tmp32 = ((uint32_t)cur_level << zeros) & 0x7FFFFFFF;
391     frac = (int16_t)(tmp32 >> 19);  // Q12.
392     // Interpolate between gainTable[zeros] and gainTable[zeros-1].
393     tmp32 =
394         ((stt->gainTable[zeros - 1] - stt->gainTable[zeros]) * (int64_t)frac) >>
395         12;
396     gains[k + 1] = stt->gainTable[zeros] + tmp32;
397   }
398 
399   // Gate processing (lower gain during absence of speech)
400   zeros = (zeros << 9) - (frac >> 3);
401   // find number of leading zeros
402   zeros_fast = WebRtcSpl_NormU32((uint32_t)stt->capacitorFast);
403   if (stt->capacitorFast == 0) {
404     zeros_fast = 31;
405   }
406   tmp32 = ((uint32_t)stt->capacitorFast << zeros_fast) & 0x7FFFFFFF;
407   zeros_fast <<= 9;
408   zeros_fast -= (int16_t)(tmp32 >> 22);
409 
410   gate = 1000 + zeros_fast - zeros - stt->vadNearend.stdShortTerm;
411 
412   if (gate < 0) {
413     stt->gatePrevious = 0;
414   } else {
415     tmp32 = stt->gatePrevious * 7;
416     gate = (int16_t)((gate + tmp32) >> 3);
417     stt->gatePrevious = gate;
418   }
419   // gate < 0     -> no gate
420   // gate > 2500  -> max gate
421   if (gate > 0) {
422     if (gate < 2500) {
423       gain_adj = (2500 - gate) >> 5;
424     } else {
425       gain_adj = 0;
426     }
427     for (k = 0; k < 10; k++) {
428       if ((gains[k + 1] - stt->gainTable[0]) > 8388608) {
429         // To prevent wraparound
430         tmp32 = (gains[k + 1] - stt->gainTable[0]) >> 8;
431         tmp32 *= 178 + gain_adj;
432       } else {
433         tmp32 = (gains[k + 1] - stt->gainTable[0]) * (178 + gain_adj);
434         tmp32 >>= 8;
435       }
436       gains[k + 1] = stt->gainTable[0] + tmp32;
437     }
438   }
439 
440   // Limit gain to avoid overload distortion
441   for (k = 0; k < 10; k++) {
442     // Find a shift of gains[k + 1] such that it can be squared without
443     // overflow, but at least by 10 bits.
444     zeros = 10;
445     if (gains[k + 1] > 47452159) {
446       zeros = 16 - WebRtcSpl_NormW32(gains[k + 1]);
447     }
448     gain32 = (gains[k + 1] >> zeros) + 1;
449     gain32 *= gain32;
450     // check for overflow
451     while (AGC_MUL32((env[k] >> 12) + 1, gain32) >
452            WEBRTC_SPL_SHIFT_W32((int32_t)32767, 2 * (1 - zeros + 10))) {
453       // multiply by 253/256 ==> -0.1 dB
454       if (gains[k + 1] > 8388607) {
455         // Prevent wrap around
456         gains[k + 1] = (gains[k + 1] / 256) * 253;
457       } else {
458         gains[k + 1] = (gains[k + 1] * 253) / 256;
459       }
460       gain32 = (gains[k + 1] >> zeros) + 1;
461       gain32 *= gain32;
462     }
463   }
464   // gain reductions should be done 1 ms earlier than gain increases
465   for (k = 1; k < 10; k++) {
466     if (gains[k] > gains[k + 1]) {
467       gains[k] = gains[k + 1];
468     }
469   }
470   // save start gain for next frame
471   stt->gain = gains[10];
472 
473   return 0;
474 }
475 
WebRtcAgc_ApplyDigitalGains(const int32_t gains[11],size_t num_bands,uint32_t FS,const int16_t * const * in_near,int16_t * const * out)476 int32_t WebRtcAgc_ApplyDigitalGains(const int32_t gains[11],
477                                     size_t num_bands,
478                                     uint32_t FS,
479                                     const int16_t* const* in_near,
480                                     int16_t* const* out) {
481   // Apply gain
482   // handle first sub frame separately
483   size_t L;
484   int16_t L2;  // samples/subframe
485 
486   // determine number of samples per ms
487   if (FS == 8000) {
488     L = 8;
489     L2 = 3;
490   } else if (FS == 16000 || FS == 32000 || FS == 48000) {
491     L = 16;
492     L2 = 4;
493   } else {
494     return -1;
495   }
496 
497   for (size_t i = 0; i < num_bands; ++i) {
498     if (in_near[i] != out[i]) {
499       // Only needed if they don't already point to the same place.
500       memcpy(out[i], in_near[i], 10 * L * sizeof(in_near[i][0]));
501     }
502   }
503 
504   // iterate over samples
505   int32_t delta = (gains[1] - gains[0]) * (1 << (4 - L2));
506   int32_t gain32 = gains[0] * (1 << 4);
507   for (size_t n = 0; n < L; n++) {
508     for (size_t i = 0; i < num_bands; ++i) {
509       int32_t out_tmp = (int64_t)out[i][n] * ((gain32 + 127) >> 7) >> 16;
510       if (out_tmp > 4095) {
511         out[i][n] = (int16_t)32767;
512       } else if (out_tmp < -4096) {
513         out[i][n] = (int16_t)-32768;
514       } else {
515         int32_t tmp32 = ((int64_t)out[i][n] * (gain32 >> 4)) >> 16;
516         out[i][n] = (int16_t)tmp32;
517       }
518     }
519 
520     gain32 += delta;
521   }
522   // iterate over subframes
523   for (int k = 1; k < 10; k++) {
524     delta = (gains[k + 1] - gains[k]) * (1 << (4 - L2));
525     gain32 = gains[k] * (1 << 4);
526     // iterate over samples
527     for (size_t n = 0; n < L; n++) {
528       for (size_t i = 0; i < num_bands; ++i) {
529         int64_t tmp64 = ((int64_t)(out[i][k * L + n])) * (gain32 >> 4);
530         tmp64 = tmp64 >> 16;
531         if (tmp64 > 32767) {
532           out[i][k * L + n] = 32767;
533         } else if (tmp64 < -32768) {
534           out[i][k * L + n] = -32768;
535         } else {
536           out[i][k * L + n] = (int16_t)(tmp64);
537         }
538       }
539       gain32 += delta;
540     }
541   }
542   return 0;
543 }
544 
WebRtcAgc_InitVad(AgcVad * state)545 void WebRtcAgc_InitVad(AgcVad* state) {
546   int16_t k;
547 
548   state->HPstate = 0;   // state of high pass filter
549   state->logRatio = 0;  // log( P(active) / P(inactive) )
550   // average input level (Q10)
551   state->meanLongTerm = 15 << 10;
552 
553   // variance of input level (Q8)
554   state->varianceLongTerm = 500 << 8;
555 
556   state->stdLongTerm = 0;  // standard deviation of input level in dB
557   // short-term average input level (Q10)
558   state->meanShortTerm = 15 << 10;
559 
560   // short-term variance of input level (Q8)
561   state->varianceShortTerm = 500 << 8;
562 
563   state->stdShortTerm =
564       0;               // short-term standard deviation of input level in dB
565   state->counter = 3;  // counts updates
566   for (k = 0; k < 8; k++) {
567     // downsampling filter
568     state->downState[k] = 0;
569   }
570 }
571 
WebRtcAgc_ProcessVad(AgcVad * state,const int16_t * in,size_t nrSamples)572 int16_t WebRtcAgc_ProcessVad(AgcVad* state,       // (i) VAD state
573                              const int16_t* in,   // (i) Speech signal
574                              size_t nrSamples) {  // (i) number of samples
575   uint32_t nrg;
576   int32_t out, tmp32, tmp32b;
577   uint16_t tmpU16;
578   int16_t k, subfr, tmp16;
579   int16_t buf1[8];
580   int16_t buf2[4];
581   int16_t HPstate;
582   int16_t zeros, dB;
583   int64_t tmp64;
584 
585   // process in 10 sub frames of 1 ms (to save on memory)
586   nrg = 0;
587   HPstate = state->HPstate;
588   for (subfr = 0; subfr < 10; subfr++) {
589     // downsample to 4 kHz
590     if (nrSamples == 160) {
591       for (k = 0; k < 8; k++) {
592         tmp32 = (int32_t)in[2 * k] + (int32_t)in[2 * k + 1];
593         tmp32 >>= 1;
594         buf1[k] = (int16_t)tmp32;
595       }
596       in += 16;
597 
598       WebRtcSpl_DownsampleBy2(buf1, 8, buf2, state->downState);
599     } else {
600       WebRtcSpl_DownsampleBy2(in, 8, buf2, state->downState);
601       in += 8;
602     }
603 
604     // high pass filter and compute energy
605     for (k = 0; k < 4; k++) {
606       out = buf2[k] + HPstate;
607       tmp32 = 600 * out;
608       HPstate = (int16_t)((tmp32 >> 10) - buf2[k]);
609 
610       // Add 'out * out / 2**6' to 'nrg' in a non-overflowing
611       // way. Guaranteed to work as long as 'out * out / 2**6' fits in
612       // an int32_t.
613       nrg += out * (out / (1 << 6));
614       nrg += out * (out % (1 << 6)) / (1 << 6);
615     }
616   }
617   state->HPstate = HPstate;
618 
619   // find number of leading zeros
620   if (!(0xFFFF0000 & nrg)) {
621     zeros = 16;
622   } else {
623     zeros = 0;
624   }
625   if (!(0xFF000000 & (nrg << zeros))) {
626     zeros += 8;
627   }
628   if (!(0xF0000000 & (nrg << zeros))) {
629     zeros += 4;
630   }
631   if (!(0xC0000000 & (nrg << zeros))) {
632     zeros += 2;
633   }
634   if (!(0x80000000 & (nrg << zeros))) {
635     zeros += 1;
636   }
637 
638   // energy level (range {-32..30}) (Q10)
639   dB = (15 - zeros) * (1 << 11);
640 
641   // Update statistics
642 
643   if (state->counter < kAvgDecayTime) {
644     // decay time = AvgDecTime * 10 ms
645     state->counter++;
646   }
647 
648   // update short-term estimate of mean energy level (Q10)
649   tmp32 = state->meanShortTerm * 15 + dB;
650   state->meanShortTerm = (int16_t)(tmp32 >> 4);
651 
652   // update short-term estimate of variance in energy level (Q8)
653   tmp32 = (dB * dB) >> 12;
654   tmp32 += state->varianceShortTerm * 15;
655   state->varianceShortTerm = tmp32 / 16;
656 
657   // update short-term estimate of standard deviation in energy level (Q10)
658   tmp32 = state->meanShortTerm * state->meanShortTerm;
659   tmp32 = (state->varianceShortTerm << 12) - tmp32;
660   state->stdShortTerm = (int16_t)WebRtcSpl_Sqrt(tmp32);
661 
662   // update long-term estimate of mean energy level (Q10)
663   tmp32 = state->meanLongTerm * state->counter + dB;
664   state->meanLongTerm =
665       WebRtcSpl_DivW32W16ResW16(tmp32, WebRtcSpl_AddSatW16(state->counter, 1));
666 
667   // update long-term estimate of variance in energy level (Q8)
668   tmp32 = (dB * dB) >> 12;
669   tmp32 += state->varianceLongTerm * state->counter;
670   state->varianceLongTerm =
671       WebRtcSpl_DivW32W16(tmp32, WebRtcSpl_AddSatW16(state->counter, 1));
672 
673   // update long-term estimate of standard deviation in energy level (Q10)
674   tmp32 = state->meanLongTerm * state->meanLongTerm;
675   tmp32 = (state->varianceLongTerm << 12) - tmp32;
676   state->stdLongTerm = (int16_t)WebRtcSpl_Sqrt(tmp32);
677 
678   // update voice activity measure (Q10)
679   tmp16 = 3 << 12;
680   // TODO(bjornv): (dB - state->meanLongTerm) can overflow, e.g., in
681   // ApmTest.Process unit test. Previously the macro WEBRTC_SPL_MUL_16_16()
682   // was used, which did an intermediate cast to (int16_t), hence losing
683   // significant bits. This cause logRatio to max out positive, rather than
684   // negative. This is a bug, but has very little significance.
685   tmp32 = tmp16 * (int16_t)(dB - state->meanLongTerm);
686   tmp32 = WebRtcSpl_DivW32W16(tmp32, state->stdLongTerm);
687   tmpU16 = (13 << 12);
688   tmp32b = WEBRTC_SPL_MUL_16_U16(state->logRatio, tmpU16);
689   tmp64 = tmp32;
690   tmp64 += tmp32b >> 10;
691   tmp64 >>= 6;
692 
693   // limit
694   if (tmp64 > 2048) {
695     tmp64 = 2048;
696   } else if (tmp64 < -2048) {
697     tmp64 = -2048;
698   }
699   state->logRatio = (int16_t)tmp64;
700 
701   return state->logRatio;  // Q10
702 }
703 
704 }  // namespace webrtc
705