path: root/third_party/libwebrtc/modules/audio_processing/agc/legacy/digital_agc.cc

/*
 *  Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include "modules/audio_processing/agc/legacy/digital_agc.h"

#include <string.h>

#include "modules/audio_processing/agc/legacy/gain_control.h"
#include "rtc_base/checks.h"

namespace webrtc {

namespace {

// To generate the gaintable, copy&paste the following lines to a Matlab window:
// MaxGain = 6; MinGain = 0; CompRatio = 3; Knee = 1;
// zeros = 0:31; lvl = 2.^(1-zeros);
// A = -10*log10(lvl) * (CompRatio - 1) / CompRatio;
// B = MaxGain - MinGain;
// gains = round(2^16*10.^(0.05 * (MinGain + B * (
// log(exp(-Knee*A)+exp(-Knee*B)) - log(1+exp(-Knee*B)) ) /
// log(1/(1+exp(Knee*B))))));
// fprintf(1, '\t%i, %i, %i, %i,\n', gains);
// % Matlab code for plotting the gain and input/output level characteristic
// (copy/paste the following 3 lines):
// in = 10*log10(lvl); out = 20*log10(gains/65536);
// subplot(121); plot(in, out); axis([-30, 0, -5, 20]); grid on; xlabel('Input
// (dB)'); ylabel('Gain (dB)');
// subplot(122); plot(in, in+out); axis([-30, 0, -30, 5]); grid on;
// xlabel('Input (dB)'); ylabel('Output (dB)');
// zoom on;

// Generator table for y=log2(1+e^x) in Q8.
enum { kGenFuncTableSize = 128 };
static const uint16_t kGenFuncTable[kGenFuncTableSize] = {
    256,   485,   786,   1126,  1484,  1849,  2217,  2586,  2955,  3324,  3693,
    4063,  4432,  4801,  5171,  5540,  5909,  6279,  6648,  7017,  7387,  7756,
    8125,  8495,  8864,  9233,  9603,  9972,  10341, 10711, 11080, 11449, 11819,
    12188, 12557, 12927, 13296, 13665, 14035, 14404, 14773, 15143, 15512, 15881,
    16251, 16620, 16989, 17359, 17728, 18097, 18466, 18836, 19205, 19574, 19944,
    20313, 20682, 21052, 21421, 21790, 22160, 22529, 22898, 23268, 23637, 24006,
    24376, 24745, 25114, 25484, 25853, 26222, 26592, 26961, 27330, 27700, 28069,
    28438, 28808, 29177, 29546, 29916, 30285, 30654, 31024, 31393, 31762, 32132,
    32501, 32870, 33240, 33609, 33978, 34348, 34717, 35086, 35456, 35825, 36194,
    36564, 36933, 37302, 37672, 38041, 38410, 38780, 39149, 39518, 39888, 40257,
    40626, 40996, 41365, 41734, 42104, 42473, 42842, 43212, 43581, 43950, 44320,
    44689, 45058, 45428, 45797, 46166, 46536, 46905};

static const int16_t kAvgDecayTime = 250;  // frames; < 3000

// the 32 most significant bits of A(19) * B(26) >> 13
#define AGC_MUL32(A, B) (((B) >> 13) * (A) + (((0x00001FFF & (B)) * (A)) >> 13))
// C + the 32 most significant bits of A * B
#define AGC_SCALEDIFF32(A, B, C) \
  ((C) + ((B) >> 16) * (A) + (((0x0000FFFF & (B)) * (A)) >> 16))

}  // namespace

int32_t WebRtcAgc_CalculateGainTable(int32_t* gainTable,       // Q16
                                     int16_t digCompGaindB,    // Q0
                                     int16_t targetLevelDbfs,  // Q0
                                     uint8_t limiterEnable,
                                     int16_t analogTarget) {  // Q0
  // This function generates the compressor gain table used in the fixed digital
  // part.
  uint32_t tmpU32no1, tmpU32no2, absInLevel, logApprox;
  int32_t inLevel, limiterLvl;
  int32_t tmp32, tmp32no1, tmp32no2, numFIX, den, y32;
  const uint16_t kLog10 = 54426;    // log2(10)     in Q14
  const uint16_t kLog10_2 = 49321;  // 10*log10(2)  in Q14
  const uint16_t kLogE_1 = 23637;   // log2(e)      in Q14
  uint16_t constMaxGain;
  uint16_t tmpU16, intPart, fracPart;
  const int16_t kCompRatio = 3;
  int16_t limiterOffset = 0;  // Limiter offset
  int16_t limiterIdx, limiterLvlX;
  int16_t constLinApprox, maxGain, diffGain;
  int16_t i, tmp16, tmp16no1;
  int zeros, zerosScale;

  // Constants
  //    kLogE_1 = 23637; // log2(e)      in Q14
  //    kLog10 = 54426; // log2(10)     in Q14
  //    kLog10_2 = 49321; // 10*log10(2)  in Q14

  // Calculate maximum digital gain and zero gain level
  tmp32no1 = (digCompGaindB - analogTarget) * (kCompRatio - 1);
  tmp16no1 = analogTarget - targetLevelDbfs;
  tmp16no1 +=
      WebRtcSpl_DivW32W16ResW16(tmp32no1 + (kCompRatio >> 1), kCompRatio);
  maxGain = WEBRTC_SPL_MAX(tmp16no1, (analogTarget - targetLevelDbfs));
  tmp32no1 = maxGain * kCompRatio;
  if ((digCompGaindB <= analogTarget) && (limiterEnable)) {
    limiterOffset = 0;
  }

  // Calculate the difference between maximum gain and gain at 0dB0v
  tmp32no1 = digCompGaindB * (kCompRatio - 1);
  diffGain =
      WebRtcSpl_DivW32W16ResW16(tmp32no1 + (kCompRatio >> 1), kCompRatio);
  if (diffGain < 0 || diffGain >= kGenFuncTableSize) {
    RTC_DCHECK(0);
    return -1;
  }

  // Calculate the limiter level and index:
  //  limiterLvlX = analogTarget - limiterOffset
  //  limiterLvl  = targetLevelDbfs + limiterOffset/compRatio
  limiterLvlX = analogTarget - limiterOffset;
  limiterIdx = 2 + WebRtcSpl_DivW32W16ResW16((int32_t)limiterLvlX * (1 << 13),
                                             kLog10_2 / 2);
  tmp16no1 =
      WebRtcSpl_DivW32W16ResW16(limiterOffset + (kCompRatio >> 1), kCompRatio);
  limiterLvl = targetLevelDbfs + tmp16no1;

  // Calculate (through table lookup):
  //  constMaxGain = log2(1+2^(log2(e)*diffGain)); (in Q8)
  constMaxGain = kGenFuncTable[diffGain];  // in Q8

  // Calculate a parameter used to approximate the fractional part of 2^x with a
  // piecewise linear function in Q14:
  //  constLinApprox = round(3/2*(4*(3-2*sqrt(2))/(log(2)^2)-0.5)*2^14);
  constLinApprox = 22817;  // in Q14

  // Calculate a denominator used in the exponential part to convert from dB to
  // linear scale:
  //  den = 20*constMaxGain (in Q8)
  den = WEBRTC_SPL_MUL_16_U16(20, constMaxGain);  // in Q8

  for (i = 0; i < 32; i++) {
    // Calculate scaled input level (compressor):
    //  inLevel =
    //  fix((-constLog10_2*(compRatio-1)*(1-i)+fix(compRatio/2))/compRatio)
    tmp16 = (int16_t)((kCompRatio - 1) * (i - 1));       // Q0
    tmp32 = WEBRTC_SPL_MUL_16_U16(tmp16, kLog10_2) + 1;  // Q14
    inLevel = WebRtcSpl_DivW32W16(tmp32, kCompRatio);    // Q14

    // Calculate diffGain-inLevel, to map using the genFuncTable
    inLevel = (int32_t)diffGain * (1 << 14) - inLevel;  // Q14

    // Make calculations on abs(inLevel) and compensate for the sign afterwards.
    absInLevel = (uint32_t)WEBRTC_SPL_ABS_W32(inLevel);  // Q14

    // LUT with interpolation
    intPart = (uint16_t)(absInLevel >> 14);
    fracPart =
        (uint16_t)(absInLevel & 0x00003FFF);  // extract the fractional part
    tmpU16 = kGenFuncTable[intPart + 1] - kGenFuncTable[intPart];  // Q8
    tmpU32no1 = tmpU16 * fracPart;                                 // Q22
    tmpU32no1 += (uint32_t)kGenFuncTable[intPart] << 14;           // Q22
    logApprox = tmpU32no1 >> 8;                                    // Q14
    // Compensate for negative exponent using the relation:
    //  log2(1 + 2^-x) = log2(1 + 2^x) - x
    if (inLevel < 0) {
      zeros = WebRtcSpl_NormU32(absInLevel);
      zerosScale = 0;
      if (zeros < 15) {
        // Not enough space for multiplication
        tmpU32no2 = absInLevel >> (15 - zeros);                 // Q(zeros-1)
        tmpU32no2 = WEBRTC_SPL_UMUL_32_16(tmpU32no2, kLogE_1);  // Q(zeros+13)
        if (zeros < 9) {
          zerosScale = 9 - zeros;
          tmpU32no1 >>= zerosScale;  // Q(zeros+13)
        } else {
          tmpU32no2 >>= zeros - 9;  // Q22
        }
      } else {
        tmpU32no2 = WEBRTC_SPL_UMUL_32_16(absInLevel, kLogE_1);  // Q28
        tmpU32no2 >>= 6;                                         // Q22
      }
      logApprox = 0;
      if (tmpU32no2 < tmpU32no1) {
        logApprox = (tmpU32no1 - tmpU32no2) >> (8 - zerosScale);  // Q14
      }
    }
    numFIX = (maxGain * constMaxGain) * (1 << 6);  // Q14
    numFIX -= (int32_t)logApprox * diffGain;       // Q14

    // Calculate ratio
    // Shift `numFIX` as much as possible.
    // Ensure we avoid wrap-around in `den` as well.
    if (numFIX > (den >> 8) || -numFIX > (den >> 8)) {  // `den` is Q8.
      zeros = WebRtcSpl_NormW32(numFIX);
    } else {
      zeros = WebRtcSpl_NormW32(den) + 8;
    }
    numFIX *= 1 << zeros;  // Q(14+zeros)

    // Shift den so we end up in Qy1
    tmp32no1 = WEBRTC_SPL_SHIFT_W32(den, zeros - 9);  // Q(zeros - 1)
    y32 = numFIX / tmp32no1;                          // in Q15
    // This is to do rounding in Q14.
    y32 = y32 >= 0 ? (y32 + 1) >> 1 : -((-y32 + 1) >> 1);

    if (limiterEnable && (i < limiterIdx)) {
      tmp32 = WEBRTC_SPL_MUL_16_U16(i - 1, kLog10_2);  // Q14
      tmp32 -= limiterLvl * (1 << 14);                 // Q14
      y32 = WebRtcSpl_DivW32W16(tmp32 + 10, 20);
    }
    if (y32 > 39000) {
      tmp32 = (y32 >> 1) * kLog10 + 4096;  // in Q27
      tmp32 >>= 13;                        // In Q14.
    } else {
      tmp32 = y32 * kLog10 + 8192;  // in Q28
      tmp32 >>= 14;                 // In Q14.
    }
    tmp32 += 16 << 14;  // in Q14 (Make sure final output is in Q16)

    // Calculate power
    if (tmp32 > 0) {
      intPart = (int16_t)(tmp32 >> 14);
      fracPart = (uint16_t)(tmp32 & 0x00003FFF);  // in Q14
      if ((fracPart >> 13) != 0) {
        tmp16 = (2 << 14) - constLinApprox;
        tmp32no2 = (1 << 14) - fracPart;
        tmp32no2 *= tmp16;
        tmp32no2 >>= 13;
        tmp32no2 = (1 << 14) - tmp32no2;
      } else {
        tmp16 = constLinApprox - (1 << 14);
        tmp32no2 = (fracPart * tmp16) >> 13;
      }
      fracPart = (uint16_t)tmp32no2;
      gainTable[i] =
          (1 << intPart) + WEBRTC_SPL_SHIFT_W32(fracPart, intPart - 14);
    } else {
      gainTable[i] = 0;
    }
  }

  return 0;
}

int32_t WebRtcAgc_InitDigital(DigitalAgc* stt, int16_t agcMode) {
  if (agcMode == kAgcModeFixedDigital) {
    // start at minimum to find correct gain faster
    stt->capacitorSlow = 0;
  } else {
    // start out with 0 dB gain
    stt->capacitorSlow = 134217728;  // (int32_t)(0.125f * 32768.0f * 32768.0f);
  }
  stt->capacitorFast = 0;
  stt->gain = 65536;
  stt->gatePrevious = 0;
  stt->agcMode = agcMode;

  // initialize VADs
  WebRtcAgc_InitVad(&stt->vadNearend);
  WebRtcAgc_InitVad(&stt->vadFarend);

  return 0;
}

int32_t WebRtcAgc_AddFarendToDigital(DigitalAgc* stt,
                                     const int16_t* in_far,
                                     size_t nrSamples) {
  RTC_DCHECK(stt);
  // VAD for far end
  WebRtcAgc_ProcessVad(&stt->vadFarend, in_far, nrSamples);

  return 0;
}

// Gains is an 11 element long array (one value per ms, incl start & end).
int32_t WebRtcAgc_ComputeDigitalGains(DigitalAgc* stt,
                                      const int16_t* const* in_near,
                                      size_t num_bands,
                                      uint32_t FS,
                                      int16_t lowlevelSignal,
                                      int32_t gains[11]) {
  int32_t tmp32;
  int32_t env[10];
  int32_t max_nrg;
  int32_t cur_level;
  int32_t gain32;
  int16_t logratio;
  int16_t lower_thr, upper_thr;
  int16_t zeros = 0, zeros_fast, frac = 0;
  int16_t decay;
  int16_t gate, gain_adj;
  int16_t k;
  size_t n, L;

  // determine number of samples per ms
  if (FS == 8000) {
    L = 8;
  } else if (FS == 16000 || FS == 32000 || FS == 48000) {
    L = 16;
  } else {
    return -1;
  }

  // VAD for near end
  logratio = WebRtcAgc_ProcessVad(&stt->vadNearend, in_near[0], L * 10);

  // Account for far end VAD
  if (stt->vadFarend.counter > 10) {
    tmp32 = 3 * logratio;
    logratio = (int16_t)((tmp32 - stt->vadFarend.logRatio) >> 2);
  }

  // Determine decay factor depending on VAD
  //  upper_thr = 1.0f;
  //  lower_thr = 0.25f;
  upper_thr = 1024;  // Q10
  lower_thr = 0;     // Q10
  if (logratio > upper_thr) {
    // decay = -2^17 / DecayTime;  ->  -65
    decay = -65;
  } else if (logratio < lower_thr) {
    decay = 0;
  } else {
    // decay = (int16_t)(((lower_thr - logratio)
    //       * (2^27/(DecayTime*(upper_thr-lower_thr)))) >> 10);
    // SUBSTITUTED: 2^27/(DecayTime*(upper_thr-lower_thr))  ->  65
    tmp32 = (lower_thr - logratio) * 65;
    decay = (int16_t)(tmp32 >> 10);
  }

  // adjust decay factor for long silence (detected as low standard deviation)
  // This is only done in the adaptive modes
  if (stt->agcMode != kAgcModeFixedDigital) {
    if (stt->vadNearend.stdLongTerm < 4000) {
      decay = 0;
    } else if (stt->vadNearend.stdLongTerm < 8096) {
      // decay = (int16_t)(((stt->vadNearend.stdLongTerm - 4000) * decay) >>
      // 12);
      tmp32 = (stt->vadNearend.stdLongTerm - 4000) * decay;
      decay = (int16_t)(tmp32 >> 12);
    }

    if (lowlevelSignal != 0) {
      decay = 0;
    }
  }
  // Find max amplitude per sub frame
  // iterate over sub frames
  for (k = 0; k < 10; k++) {
    // iterate over samples
    max_nrg = 0;
    for (n = 0; n < L; n++) {
      int32_t nrg = in_near[0][k * L + n] * in_near[0][k * L + n];
      if (nrg > max_nrg) {
        max_nrg = nrg;
      }
    }
    env[k] = max_nrg;
  }

  // Calculate gain per sub frame
  gains[0] = stt->gain;
  for (k = 0; k < 10; k++) {
    // Fast envelope follower
    //  decay time = -131000 / -1000 = 131 (ms)
    stt->capacitorFast =
        AGC_SCALEDIFF32(-1000, stt->capacitorFast, stt->capacitorFast);
    if (env[k] > stt->capacitorFast) {
      stt->capacitorFast = env[k];
    }
    // Slow envelope follower
    if (env[k] > stt->capacitorSlow) {
      // increase capacitorSlow
      stt->capacitorSlow = AGC_SCALEDIFF32(500, (env[k] - stt->capacitorSlow),
                                           stt->capacitorSlow);
    } else {
      // decrease capacitorSlow
      stt->capacitorSlow =
          AGC_SCALEDIFF32(decay, stt->capacitorSlow, stt->capacitorSlow);
    }

    // use maximum of both capacitors as current level
    if (stt->capacitorFast > stt->capacitorSlow) {
      cur_level = stt->capacitorFast;
    } else {
      cur_level = stt->capacitorSlow;
    }
    // Translate signal level into gain, using a piecewise linear approximation
    // find number of leading zeros
    zeros = WebRtcSpl_NormU32((uint32_t)cur_level);
    if (cur_level == 0) {
      zeros = 31;
    }
    tmp32 = ((uint32_t)cur_level << zeros) & 0x7FFFFFFF;
    frac = (int16_t)(tmp32 >> 19);  // Q12.
    // Interpolate between gainTable[zeros] and gainTable[zeros-1].
    tmp32 =
        ((stt->gainTable[zeros - 1] - stt->gainTable[zeros]) * (int64_t)frac) >>
        12;
    gains[k + 1] = stt->gainTable[zeros] + tmp32;
  }

  // Gate processing (lower gain during absence of speech)
  zeros = (zeros << 9) - (frac >> 3);
  // find number of leading zeros
  zeros_fast = WebRtcSpl_NormU32((uint32_t)stt->capacitorFast);
  if (stt->capacitorFast == 0) {
    zeros_fast = 31;
  }
  tmp32 = ((uint32_t)stt->capacitorFast << zeros_fast) & 0x7FFFFFFF;
  zeros_fast <<= 9;
  zeros_fast -= (int16_t)(tmp32 >> 22);

  gate = 1000 + zeros_fast - zeros - stt->vadNearend.stdShortTerm;

  if (gate < 0) {
    stt->gatePrevious = 0;
  } else {
    tmp32 = stt->gatePrevious * 7;
    gate = (int16_t)((gate + tmp32) >> 3);
    stt->gatePrevious = gate;
  }
  // gate < 0     -> no gate
  // gate > 2500  -> max gate
  if (gate > 0) {
    if (gate < 2500) {
      gain_adj = (2500 - gate) >> 5;
    } else {
      gain_adj = 0;
    }
    for (k = 0; k < 10; k++) {
      if ((gains[k + 1] - stt->gainTable[0]) > 8388608) {
        // To prevent wraparound
        tmp32 = (gains[k + 1] - stt->gainTable[0]) >> 8;
        tmp32 *= 178 + gain_adj;
      } else {
        tmp32 = (gains[k + 1] - stt->gainTable[0]) * (178 + gain_adj);
        tmp32 >>= 8;
      }
      gains[k + 1] = stt->gainTable[0] + tmp32;
    }
  }

  // Limit gain to avoid overload distortion
  for (k = 0; k < 10; k++) {
    // Find a shift of gains[k + 1] such that it can be squared without
    // overflow, but at least by 10 bits.
    zeros = 10;
    if (gains[k + 1] > 47452159) {
      zeros = 16 - WebRtcSpl_NormW32(gains[k + 1]);
    }
    gain32 = (gains[k + 1] >> zeros) + 1;
    gain32 *= gain32;
    // check for overflow
    while (AGC_MUL32((env[k] >> 12) + 1, gain32) >
           WEBRTC_SPL_SHIFT_W32((int32_t)32767, 2 * (1 - zeros + 10))) {
      // multiply by 253/256 ==> -0.1 dB
      if (gains[k + 1] > 8388607) {
        // Prevent wrap around
        gains[k + 1] = (gains[k + 1] / 256) * 253;
      } else {
        gains[k + 1] = (gains[k + 1] * 253) / 256;
      }
      gain32 = (gains[k + 1] >> zeros) + 1;
      gain32 *= gain32;
    }
  }
  // gain reductions should be done 1 ms earlier than gain increases
  for (k = 1; k < 10; k++) {
    if (gains[k] > gains[k + 1]) {
      gains[k] = gains[k + 1];
    }
  }
  // save start gain for next frame
  stt->gain = gains[10];

  return 0;
}

int32_t WebRtcAgc_ApplyDigitalGains(const int32_t gains[11],
                                    size_t num_bands,
                                    uint32_t FS,
                                    const int16_t* const* in_near,
                                    int16_t* const* out) {
  // Apply gain
  // handle first sub frame separately
  size_t L;
  int16_t L2;  // samples/subframe

  // determine number of samples per ms
  if (FS == 8000) {
    L = 8;
    L2 = 3;
  } else if (FS == 16000 || FS == 32000 || FS == 48000) {
    L = 16;
    L2 = 4;
  } else {
    return -1;
  }

  for (size_t i = 0; i < num_bands; ++i) {
    if (in_near[i] != out[i]) {
      // Only needed if they don't already point to the same place.
      memcpy(out[i], in_near[i], 10 * L * sizeof(in_near[i][0]));
    }
  }

  // iterate over samples
  int32_t delta = (gains[1] - gains[0]) * (1 << (4 - L2));
  int32_t gain32 = gains[0] * (1 << 4);
  for (size_t n = 0; n < L; n++) {
    for (size_t i = 0; i < num_bands; ++i) {
      int32_t out_tmp = (int64_t)out[i][n] * ((gain32 + 127) >> 7) >> 16;
      if (out_tmp > 4095) {
        out[i][n] = (int16_t)32767;
      } else if (out_tmp < -4096) {
        out[i][n] = (int16_t)-32768;
      } else {
        int32_t tmp32 = ((int64_t)out[i][n] * (gain32 >> 4)) >> 16;
        out[i][n] = (int16_t)tmp32;
      }
    }

    gain32 += delta;
  }
  // iterate over subframes
  for (int k = 1; k < 10; k++) {
    delta = (gains[k + 1] - gains[k]) * (1 << (4 - L2));
    gain32 = gains[k] * (1 << 4);
    // iterate over samples
    for (size_t n = 0; n < L; n++) {
      for (size_t i = 0; i < num_bands; ++i) {
        int64_t tmp64 = ((int64_t)(out[i][k * L + n])) * (gain32 >> 4);
        tmp64 = tmp64 >> 16;
        if (tmp64 > 32767) {
          out[i][k * L + n] = 32767;
        } else if (tmp64 < -32768) {
          out[i][k * L + n] = -32768;
        } else {
          out[i][k * L + n] = (int16_t)(tmp64);
        }
      }
      gain32 += delta;
    }
  }
  return 0;
}

void WebRtcAgc_InitVad(AgcVad* state) {
  int16_t k;

  state->HPstate = 0;   // state of high pass filter
  state->logRatio = 0;  // log( P(active) / P(inactive) )
  // average input level (Q10)
  state->meanLongTerm = 15 << 10;

  // variance of input level (Q8)
  state->varianceLongTerm = 500 << 8;

  state->stdLongTerm = 0;  // standard deviation of input level in dB
  // short-term average input level (Q10)
  state->meanShortTerm = 15 << 10;

  // short-term variance of input level (Q8)
  state->varianceShortTerm = 500 << 8;

  state->stdShortTerm =
      0;               // short-term standard deviation of input level in dB
  state->counter = 3;  // counts updates
  for (k = 0; k < 8; k++) {
    // downsampling filter
    state->downState[k] = 0;
  }
}

int16_t WebRtcAgc_ProcessVad(AgcVad* state,       // (i) VAD state
                             const int16_t* in,   // (i) Speech signal
                             size_t nrSamples) {  // (i) number of samples
  uint32_t nrg;
  int32_t out, tmp32, tmp32b;
  uint16_t tmpU16;
  int16_t k, subfr, tmp16;
  int16_t buf1[8];
  int16_t buf2[4];
  int16_t HPstate;
  int16_t zeros, dB;
  int64_t tmp64;

  // process in 10 sub frames of 1 ms (to save on memory)
  nrg = 0;
  HPstate = state->HPstate;
  for (subfr = 0; subfr < 10; subfr++) {
    // downsample to 4 kHz
    if (nrSamples == 160) {
      for (k = 0; k < 8; k++) {
        tmp32 = (int32_t)in[2 * k] + (int32_t)in[2 * k + 1];
        tmp32 >>= 1;
        buf1[k] = (int16_t)tmp32;
      }
      in += 16;

      WebRtcSpl_DownsampleBy2(buf1, 8, buf2, state->downState);
    } else {
      WebRtcSpl_DownsampleBy2(in, 8, buf2, state->downState);
      in += 8;
    }

    // high pass filter and compute energy
    for (k = 0; k < 4; k++) {
      out = buf2[k] + HPstate;
      tmp32 = 600 * out;
      HPstate = (int16_t)((tmp32 >> 10) - buf2[k]);

      // Add 'out * out / 2**6' to 'nrg' in a non-overflowing
      // way. Guaranteed to work as long as 'out * out / 2**6' fits in
      // an int32_t.
      nrg += out * (out / (1 << 6));
      nrg += out * (out % (1 << 6)) / (1 << 6);
    }
  }
  state->HPstate = HPstate;

  // find number of leading zeros
  if (!(0xFFFF0000 & nrg)) {
    zeros = 16;
  } else {
    zeros = 0;
  }
  if (!(0xFF000000 & (nrg << zeros))) {
    zeros += 8;
  }
  if (!(0xF0000000 & (nrg << zeros))) {
    zeros += 4;
  }
  if (!(0xC0000000 & (nrg << zeros))) {
    zeros += 2;
  }
  if (!(0x80000000 & (nrg << zeros))) {
    zeros += 1;
  }

  // energy level (range {-32..30}) (Q10)
  dB = (15 - zeros) * (1 << 11);

  // Update statistics

  if (state->counter < kAvgDecayTime) {
    // decay time = AvgDecTime * 10 ms
    state->counter++;
  }

  // update short-term estimate of mean energy level (Q10)
  tmp32 = state->meanShortTerm * 15 + dB;
  state->meanShortTerm = (int16_t)(tmp32 >> 4);

  // update short-term estimate of variance in energy level (Q8)
  tmp32 = (dB * dB) >> 12;
  tmp32 += state->varianceShortTerm * 15;
  state->varianceShortTerm = tmp32 / 16;

  // update short-term estimate of standard deviation in energy level (Q10)
  tmp32 = state->meanShortTerm * state->meanShortTerm;
  tmp32 = (state->varianceShortTerm << 12) - tmp32;
  state->stdShortTerm = (int16_t)WebRtcSpl_Sqrt(tmp32);

  // update long-term estimate of mean energy level (Q10)
  tmp32 = state->meanLongTerm * state->counter + dB;
  state->meanLongTerm =
      WebRtcSpl_DivW32W16ResW16(tmp32, WebRtcSpl_AddSatW16(state->counter, 1));

  // update long-term estimate of variance in energy level (Q8)
  tmp32 = (dB * dB) >> 12;
  tmp32 += state->varianceLongTerm * state->counter;
  state->varianceLongTerm =
      WebRtcSpl_DivW32W16(tmp32, WebRtcSpl_AddSatW16(state->counter, 1));

  // update long-term estimate of standard deviation in energy level (Q10)
  tmp32 = state->meanLongTerm * state->meanLongTerm;
  tmp32 = (state->varianceLongTerm << 12) - tmp32;
  state->stdLongTerm = (int16_t)WebRtcSpl_Sqrt(tmp32);

  // update voice activity measure (Q10)
  tmp16 = 3 << 12;
  // TODO(bjornv): (dB - state->meanLongTerm) can overflow, e.g., in
  // ApmTest.Process unit test. Previously the macro WEBRTC_SPL_MUL_16_16()
  // was used, which did an intermediate cast to (int16_t), hence losing
  // significant bits. This cause logRatio to max out positive, rather than
  // negative. This is a bug, but has very little significance.
  tmp32 = tmp16 * (int16_t)(dB - state->meanLongTerm);
  tmp32 = WebRtcSpl_DivW32W16(tmp32, state->stdLongTerm);
  tmpU16 = (13 << 12);
  tmp32b = WEBRTC_SPL_MUL_16_U16(state->logRatio, tmpU16);
  tmp64 = tmp32;
  tmp64 += tmp32b >> 10;
  tmp64 >>= 6;

  // limit
  if (tmp64 > 2048) {
    tmp64 = 2048;
  } else if (tmp64 < -2048) {
    tmp64 = -2048;
  }
  state->logRatio = (int16_t)tmp64;

  return state->logRatio;  // Q10
}

}  // namespace webrtc