path: root/webrtc/modules/audio_processing/aecm/aecm_core_c.c

/*
 *  Copyright (c) 2013 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 "webrtc/modules/audio_processing/aecm/aecm_core.h"

#include <assert.h>
#include <stddef.h>
#include <stdlib.h>

#include "webrtc/common_audio/ring_buffer.h"
#include "webrtc/common_audio/signal_processing/include/real_fft.h"
#include "webrtc/modules/audio_processing/aecm/echo_control_mobile.h"
#include "webrtc/modules/audio_processing/utility/delay_estimator_wrapper.h"
#include "webrtc/system_wrappers/include/compile_assert_c.h"
#include "webrtc/system_wrappers/include/cpu_features_wrapper.h"
#include "webrtc/typedefs.h"

// Square root of Hanning window in Q14.
#if defined(WEBRTC_DETECT_NEON) || defined(WEBRTC_HAS_NEON)
// Table is defined in an ARM assembly file.
extern const ALIGN8_BEG int16_t WebRtcAecm_kSqrtHanning[] ALIGN8_END;
#else
static const ALIGN8_BEG int16_t WebRtcAecm_kSqrtHanning[] ALIGN8_END = {
  0, 399, 798, 1196, 1594, 1990, 2386, 2780, 3172,
  3562, 3951, 4337, 4720, 5101, 5478, 5853, 6224,
  6591, 6954, 7313, 7668, 8019, 8364, 8705, 9040,
  9370, 9695, 10013, 10326, 10633, 10933, 11227, 11514,
  11795, 12068, 12335, 12594, 12845, 13089, 13325, 13553,
  13773, 13985, 14189, 14384, 14571, 14749, 14918, 15079,
  15231, 15373, 15506, 15631, 15746, 15851, 15947, 16034,
  16111, 16179, 16237, 16286, 16325, 16354, 16373, 16384
};
#endif

#ifdef AECM_WITH_ABS_APPROX
//Q15 alpha = 0.99439986968132  const Factor for magnitude approximation
static const uint16_t kAlpha1 = 32584;
//Q15 beta = 0.12967166976970   const Factor for magnitude approximation
static const uint16_t kBeta1 = 4249;
//Q15 alpha = 0.94234827210087  const Factor for magnitude approximation
static const uint16_t kAlpha2 = 30879;
//Q15 beta = 0.33787806009150   const Factor for magnitude approximation
static const uint16_t kBeta2 = 11072;
//Q15 alpha = 0.82247698684306  const Factor for magnitude approximation
static const uint16_t kAlpha3 = 26951;
//Q15 beta = 0.57762063060713   const Factor for magnitude approximation
static const uint16_t kBeta3 = 18927;
#endif

static const int16_t kNoiseEstQDomain = 15;
static const int16_t kNoiseEstIncCount = 5;

static void ComfortNoise(AecmCore* aecm,
                         const uint16_t* dfa,
                         ComplexInt16* out,
                         const int16_t* lambda);

static void WindowAndFFT(AecmCore* aecm,
                         int16_t* fft,
                         const int16_t* time_signal,
                         ComplexInt16* freq_signal,
                         int time_signal_scaling) {
  int i = 0;

  // FFT of signal
  for (i = 0; i < PART_LEN; i++) {
    // Window time domain signal and insert into real part of
    // transformation array |fft|
    int16_t scaled_time_signal = time_signal[i] << time_signal_scaling;
    fft[i] = (int16_t)((scaled_time_signal * WebRtcAecm_kSqrtHanning[i]) >> 14);
    scaled_time_signal = time_signal[i + PART_LEN] << time_signal_scaling;
    fft[PART_LEN + i] = (int16_t)((
        scaled_time_signal * WebRtcAecm_kSqrtHanning[PART_LEN - i]) >> 14);
  }

  // Do forward FFT, then take only the first PART_LEN complex samples,
  // and change signs of the imaginary parts.
  WebRtcSpl_RealForwardFFT(aecm->real_fft, fft, (int16_t*)freq_signal);
  for (i = 0; i < PART_LEN; i++) {
    freq_signal[i].imag = -freq_signal[i].imag;
  }
}

static void InverseFFTAndWindow(AecmCore* aecm,
                                int16_t* fft,
                                ComplexInt16* efw,
                                int16_t* output,
                                const int16_t* nearendClean) {
  int i, j, outCFFT;
  int32_t tmp32no1;
  // Reuse |efw| for the inverse FFT output after transferring
  // the contents to |fft|.
  int16_t* ifft_out = (int16_t*)efw;

  // Synthesis
  for (i = 1, j = 2; i < PART_LEN; i += 1, j += 2) {
    fft[j] = efw[i].real;
    fft[j + 1] = -efw[i].imag;
  }
  fft[0] = efw[0].real;
  fft[1] = -efw[0].imag;

  fft[PART_LEN2] = efw[PART_LEN].real;
  fft[PART_LEN2 + 1] = -efw[PART_LEN].imag;

  // Inverse FFT. Keep outCFFT to scale the samples in the next block.
  outCFFT = WebRtcSpl_RealInverseFFT(aecm->real_fft, fft, ifft_out);
  for (i = 0; i < PART_LEN; i++) {
    ifft_out[i] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(
                    ifft_out[i], WebRtcAecm_kSqrtHanning[i], 14);
    tmp32no1 = WEBRTC_SPL_SHIFT_W32((int32_t)ifft_out[i],
                                     outCFFT - aecm->dfaCleanQDomain);
    output[i] = (int16_t)WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX,
                                        tmp32no1 + aecm->outBuf[i],
                                        WEBRTC_SPL_WORD16_MIN);

    tmp32no1 = (ifft_out[PART_LEN + i] *
        WebRtcAecm_kSqrtHanning[PART_LEN - i]) >> 14;
    tmp32no1 = WEBRTC_SPL_SHIFT_W32(tmp32no1,
                                    outCFFT - aecm->dfaCleanQDomain);
    aecm->outBuf[i] = (int16_t)WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX,
                                                tmp32no1,
                                                WEBRTC_SPL_WORD16_MIN);
  }

  // Copy the current block to the old position
  // (aecm->outBuf is shifted elsewhere)
  memcpy(aecm->xBuf, aecm->xBuf + PART_LEN, sizeof(int16_t) * PART_LEN);
  memcpy(aecm->dBufNoisy,
         aecm->dBufNoisy + PART_LEN,
         sizeof(int16_t) * PART_LEN);
  if (nearendClean != NULL)
  {
    memcpy(aecm->dBufClean,
           aecm->dBufClean + PART_LEN,
           sizeof(int16_t) * PART_LEN);
  }
}

// Transforms a time domain signal into the frequency domain, outputting the
// complex valued signal, absolute value and sum of absolute values.
//
// time_signal          [in]    Pointer to time domain signal
// freq_signal_real     [out]   Pointer to real part of frequency domain array
// freq_signal_imag     [out]   Pointer to imaginary part of frequency domain
//                              array
// freq_signal_abs      [out]   Pointer to absolute value of frequency domain
//                              array
// freq_signal_sum_abs  [out]   Pointer to the sum of all absolute values in
//                              the frequency domain array
// return value                 The Q-domain of current frequency values
//
static int TimeToFrequencyDomain(AecmCore* aecm,
                                 const int16_t* time_signal,
                                 ComplexInt16* freq_signal,
                                 uint16_t* freq_signal_abs,
                                 uint32_t* freq_signal_sum_abs) {
  int i = 0;
  int time_signal_scaling = 0;

  int32_t tmp32no1 = 0;
  int32_t tmp32no2 = 0;

  // In fft_buf, +16 for 32-byte alignment.
  int16_t fft_buf[PART_LEN4 + 16];
  int16_t *fft = (int16_t *) (((uintptr_t) fft_buf + 31) & ~31);

  int16_t tmp16no1;
#ifndef WEBRTC_ARCH_ARM_V7
  int16_t tmp16no2;
#endif
#ifdef AECM_WITH_ABS_APPROX
  int16_t max_value = 0;
  int16_t min_value = 0;
  uint16_t alpha = 0;
  uint16_t beta = 0;
#endif

#ifdef AECM_DYNAMIC_Q
  tmp16no1 = WebRtcSpl_MaxAbsValueW16(time_signal, PART_LEN2);
  time_signal_scaling = WebRtcSpl_NormW16(tmp16no1);
#endif

  WindowAndFFT(aecm, fft, time_signal, freq_signal, time_signal_scaling);

  // Extract imaginary and real part, calculate the magnitude for
  // all frequency bins
  freq_signal[0].imag = 0;
  freq_signal[PART_LEN].imag = 0;
  freq_signal_abs[0] = (uint16_t)WEBRTC_SPL_ABS_W16(freq_signal[0].real);
  freq_signal_abs[PART_LEN] = (uint16_t)WEBRTC_SPL_ABS_W16(
                                freq_signal[PART_LEN].real);
  (*freq_signal_sum_abs) = (uint32_t)(freq_signal_abs[0]) +
                           (uint32_t)(freq_signal_abs[PART_LEN]);

  for (i = 1; i < PART_LEN; i++)
  {
    if (freq_signal[i].real == 0)
    {
      freq_signal_abs[i] = (uint16_t)WEBRTC_SPL_ABS_W16(freq_signal[i].imag);
    }
    else if (freq_signal[i].imag == 0)
    {
      freq_signal_abs[i] = (uint16_t)WEBRTC_SPL_ABS_W16(freq_signal[i].real);
    }
    else
    {
      // Approximation for magnitude of complex fft output
      // magn = sqrt(real^2 + imag^2)
      // magn ~= alpha * max(|imag|,|real|) + beta * min(|imag|,|real|)
      //
      // The parameters alpha and beta are stored in Q15

#ifdef AECM_WITH_ABS_APPROX
      tmp16no1 = WEBRTC_SPL_ABS_W16(freq_signal[i].real);
      tmp16no2 = WEBRTC_SPL_ABS_W16(freq_signal[i].imag);

      if(tmp16no1 > tmp16no2)
      {
        max_value = tmp16no1;
        min_value = tmp16no2;
      } else
      {
        max_value = tmp16no2;
        min_value = tmp16no1;
      }

      // Magnitude in Q(-6)
      if ((max_value >> 2) > min_value)
      {
        alpha = kAlpha1;
        beta = kBeta1;
      } else if ((max_value >> 1) > min_value)
      {
        alpha = kAlpha2;
        beta = kBeta2;
      } else
      {
        alpha = kAlpha3;
        beta = kBeta3;
      }
      tmp16no1 = (int16_t)((max_value * alpha) >> 15);
      tmp16no2 = (int16_t)((min_value * beta) >> 15);
      freq_signal_abs[i] = (uint16_t)tmp16no1 + (uint16_t)tmp16no2;
#else
#ifdef WEBRTC_ARCH_ARM_V7
      __asm __volatile(
        "smulbb %[tmp32no1], %[real], %[real]\n\t"
        "smlabb %[tmp32no2], %[imag], %[imag], %[tmp32no1]\n\t"
        :[tmp32no1]"+&r"(tmp32no1),
         [tmp32no2]"=r"(tmp32no2)
        :[real]"r"(freq_signal[i].real),
         [imag]"r"(freq_signal[i].imag)
      );
#else
      tmp16no1 = WEBRTC_SPL_ABS_W16(freq_signal[i].real);
      tmp16no2 = WEBRTC_SPL_ABS_W16(freq_signal[i].imag);
      tmp32no1 = tmp16no1 * tmp16no1;
      tmp32no2 = tmp16no2 * tmp16no2;
      tmp32no2 = WebRtcSpl_AddSatW32(tmp32no1, tmp32no2);
#endif // WEBRTC_ARCH_ARM_V7
      tmp32no1 = WebRtcSpl_SqrtFloor(tmp32no2);

      freq_signal_abs[i] = (uint16_t)tmp32no1;
#endif // AECM_WITH_ABS_APPROX
    }
    (*freq_signal_sum_abs) += (uint32_t)freq_signal_abs[i];
  }

  return time_signal_scaling;
}

int WebRtcAecm_ProcessBlock(AecmCore* aecm,
                            const int16_t* farend,
                            const int16_t* nearendNoisy,
                            const int16_t* nearendClean,
                            int16_t* output) {
  int i;

  uint32_t xfaSum;
  uint32_t dfaNoisySum;
  uint32_t dfaCleanSum;
  uint32_t echoEst32Gained;
  uint32_t tmpU32;

  int32_t tmp32no1;

  uint16_t xfa[PART_LEN1];
  uint16_t dfaNoisy[PART_LEN1];
  uint16_t dfaClean[PART_LEN1];
  uint16_t* ptrDfaClean = dfaClean;
  const uint16_t* far_spectrum_ptr = NULL;

  // 32 byte aligned buffers (with +8 or +16).
  // TODO(kma): define fft with ComplexInt16.
  int16_t fft_buf[PART_LEN4 + 2 + 16]; // +2 to make a loop safe.
  int32_t echoEst32_buf[PART_LEN1 + 8];
  int32_t dfw_buf[PART_LEN2 + 8];
  int32_t efw_buf[PART_LEN2 + 8];

  int16_t* fft = (int16_t*) (((uintptr_t) fft_buf + 31) & ~ 31);
  int32_t* echoEst32 = (int32_t*) (((uintptr_t) echoEst32_buf + 31) & ~ 31);
  ComplexInt16* dfw = (ComplexInt16*)(((uintptr_t)dfw_buf + 31) & ~31);
  ComplexInt16* efw = (ComplexInt16*)(((uintptr_t)efw_buf + 31) & ~31);

  int16_t hnl[PART_LEN1];
  int16_t numPosCoef = 0;
  int16_t nlpGain = ONE_Q14;
  int delay;
  int16_t tmp16no1;
  int16_t tmp16no2;
  int16_t mu;
  int16_t supGain;
  int16_t zeros32, zeros16;
  int16_t zerosDBufNoisy, zerosDBufClean, zerosXBuf;
  int far_q;
  int16_t resolutionDiff, qDomainDiff, dfa_clean_q_domain_diff;

  const int kMinPrefBand = 4;
  const int kMaxPrefBand = 24;
  int32_t avgHnl32 = 0;

  // Determine startup state. There are three states:
  // (0) the first CONV_LEN blocks
  // (1) another CONV_LEN blocks
  // (2) the rest

  if (aecm->startupState < 2)
  {
    aecm->startupState = (aecm->totCount >= CONV_LEN) +
                         (aecm->totCount >= CONV_LEN2);
  }
  // END: Determine startup state

  // Buffer near and far end signals
  memcpy(aecm->xBuf + PART_LEN, farend, sizeof(int16_t) * PART_LEN);
  memcpy(aecm->dBufNoisy + PART_LEN, nearendNoisy, sizeof(int16_t) * PART_LEN);
  if (nearendClean != NULL)
  {
    memcpy(aecm->dBufClean + PART_LEN,
           nearendClean,
           sizeof(int16_t) * PART_LEN);
  }

  // Transform far end signal from time domain to frequency domain.
  far_q = TimeToFrequencyDomain(aecm,
                                aecm->xBuf,
                                dfw,
                                xfa,
                                &xfaSum);

  // Transform noisy near end signal from time domain to frequency domain.
  zerosDBufNoisy = TimeToFrequencyDomain(aecm,
                                         aecm->dBufNoisy,
                                         dfw,
                                         dfaNoisy,
                                         &dfaNoisySum);
  aecm->dfaNoisyQDomainOld = aecm->dfaNoisyQDomain;
  aecm->dfaNoisyQDomain = (int16_t)zerosDBufNoisy;


  if (nearendClean == NULL)
  {
    ptrDfaClean = dfaNoisy;
    aecm->dfaCleanQDomainOld = aecm->dfaNoisyQDomainOld;
    aecm->dfaCleanQDomain = aecm->dfaNoisyQDomain;
    dfaCleanSum = dfaNoisySum;
  } else
  {
    // Transform clean near end signal from time domain to frequency domain.
    zerosDBufClean = TimeToFrequencyDomain(aecm,
                                           aecm->dBufClean,
                                           dfw,
                                           dfaClean,
                                           &dfaCleanSum);
    aecm->dfaCleanQDomainOld = aecm->dfaCleanQDomain;
    aecm->dfaCleanQDomain = (int16_t)zerosDBufClean;
  }

  // Get the delay
  // Save far-end history and estimate delay
  WebRtcAecm_UpdateFarHistory(aecm, xfa, far_q);
  if (WebRtc_AddFarSpectrumFix(aecm->delay_estimator_farend,
                               xfa,
                               PART_LEN1,
                               far_q) == -1) {
    return -1;
  }
  delay = WebRtc_DelayEstimatorProcessFix(aecm->delay_estimator,
                                          dfaNoisy,
                                          PART_LEN1,
                                          zerosDBufNoisy);
  if (delay == -1)
  {
    return -1;
  }
  else if (delay == -2)
  {
    // If the delay is unknown, we assume zero.
    // NOTE: this will have to be adjusted if we ever add lookahead.
    delay = 0;
  }

  if (aecm->fixedDelay >= 0)
  {
    // Use fixed delay
    delay = aecm->fixedDelay;
  }

  // Get aligned far end spectrum
  far_spectrum_ptr = WebRtcAecm_AlignedFarend(aecm, &far_q, delay);
  zerosXBuf = (int16_t) far_q;
  if (far_spectrum_ptr == NULL)
  {
    return -1;
  }

  // Calculate log(energy) and update energy threshold levels
  WebRtcAecm_CalcEnergies(aecm,
                          far_spectrum_ptr,
                          zerosXBuf,
                          dfaNoisySum,
                          echoEst32);

  // Calculate stepsize
  mu = WebRtcAecm_CalcStepSize(aecm);

  // Update counters
  aecm->totCount++;

  // This is the channel estimation algorithm.
  // It is base on NLMS but has a variable step length,
  // which was calculated above.
  WebRtcAecm_UpdateChannel(aecm,
                           far_spectrum_ptr,
                           zerosXBuf,
                           dfaNoisy,
                           mu,
                           echoEst32);
  supGain = WebRtcAecm_CalcSuppressionGain(aecm);


  // Calculate Wiener filter hnl[]
  for (i = 0; i < PART_LEN1; i++)
  {
    // Far end signal through channel estimate in Q8
    // How much can we shift right to preserve resolution
    tmp32no1 = echoEst32[i] - aecm->echoFilt[i];
    aecm->echoFilt[i] += (tmp32no1 * 50) >> 8;

    zeros32 = WebRtcSpl_NormW32(aecm->echoFilt[i]) + 1;
    zeros16 = WebRtcSpl_NormW16(supGain) + 1;
    if (zeros32 + zeros16 > 16)
    {
      // Multiplication is safe
      // Result in
      // Q(RESOLUTION_CHANNEL+RESOLUTION_SUPGAIN+
      //   aecm->xfaQDomainBuf[diff])
      echoEst32Gained = WEBRTC_SPL_UMUL_32_16((uint32_t)aecm->echoFilt[i],
                                              (uint16_t)supGain);
      resolutionDiff = 14 - RESOLUTION_CHANNEL16 - RESOLUTION_SUPGAIN;
      resolutionDiff += (aecm->dfaCleanQDomain - zerosXBuf);
    } else
    {
      tmp16no1 = 17 - zeros32 - zeros16;
      resolutionDiff = 14 + tmp16no1 - RESOLUTION_CHANNEL16 -
                       RESOLUTION_SUPGAIN;
      resolutionDiff += (aecm->dfaCleanQDomain - zerosXBuf);
      if (zeros32 > tmp16no1)
      {
        echoEst32Gained = WEBRTC_SPL_UMUL_32_16((uint32_t)aecm->echoFilt[i],
                                                supGain >> tmp16no1);
      } else
      {
        // Result in Q-(RESOLUTION_CHANNEL+RESOLUTION_SUPGAIN-16)
        echoEst32Gained = (aecm->echoFilt[i] >> tmp16no1) * supGain;
      }
    }

    zeros16 = WebRtcSpl_NormW16(aecm->nearFilt[i]);
    assert(zeros16 >= 0);  // |zeros16| is a norm, hence non-negative.
    dfa_clean_q_domain_diff = aecm->dfaCleanQDomain - aecm->dfaCleanQDomainOld;
    if (zeros16 < dfa_clean_q_domain_diff && aecm->nearFilt[i]) {
      tmp16no1 = aecm->nearFilt[i] << zeros16;
      qDomainDiff = zeros16 - dfa_clean_q_domain_diff;
      tmp16no2 = ptrDfaClean[i] >> -qDomainDiff;
    } else {
      tmp16no1 = dfa_clean_q_domain_diff < 0
          ? aecm->nearFilt[i] >> -dfa_clean_q_domain_diff
          : aecm->nearFilt[i] << dfa_clean_q_domain_diff;
      qDomainDiff = 0;
      tmp16no2 = ptrDfaClean[i];
    }
    tmp32no1 = (int32_t)(tmp16no2 - tmp16no1);
    tmp16no2 = (int16_t)(tmp32no1 >> 4);
    tmp16no2 += tmp16no1;
    zeros16 = WebRtcSpl_NormW16(tmp16no2);
    if ((tmp16no2) & (-qDomainDiff > zeros16)) {
      aecm->nearFilt[i] = WEBRTC_SPL_WORD16_MAX;
    } else {
      aecm->nearFilt[i] = qDomainDiff < 0 ? tmp16no2 << -qDomainDiff
                                          : tmp16no2 >> qDomainDiff;
    }

    // Wiener filter coefficients, resulting hnl in Q14
    if (echoEst32Gained == 0)
    {
      hnl[i] = ONE_Q14;
    } else if (aecm->nearFilt[i] == 0)
    {
      hnl[i] = 0;
    } else
    {
      // Multiply the suppression gain
      // Rounding
      echoEst32Gained += (uint32_t)(aecm->nearFilt[i] >> 1);
      tmpU32 = WebRtcSpl_DivU32U16(echoEst32Gained,
                                   (uint16_t)aecm->nearFilt[i]);

      // Current resolution is
      // Q-(RESOLUTION_CHANNEL+RESOLUTION_SUPGAIN- max(0,17-zeros16- zeros32))
      // Make sure we are in Q14
      tmp32no1 = (int32_t)WEBRTC_SPL_SHIFT_W32(tmpU32, resolutionDiff);
      if (tmp32no1 > ONE_Q14)
      {
        hnl[i] = 0;
      } else if (tmp32no1 < 0)
      {
        hnl[i] = ONE_Q14;
      } else
      {
        // 1-echoEst/dfa
        hnl[i] = ONE_Q14 - (int16_t)tmp32no1;
        if (hnl[i] < 0)
        {
          hnl[i] = 0;
        }
      }
    }
    if (hnl[i])
    {
      numPosCoef++;
    }
  }
  // Only in wideband. Prevent the gain in upper band from being larger than
  // in lower band.
  if (aecm->mult == 2)
  {
    // TODO(bjornv): Investigate if the scaling of hnl[i] below can cause
    //               speech distortion in double-talk.
    for (i = 0; i < PART_LEN1; i++)
    {
      hnl[i] = (int16_t)((hnl[i] * hnl[i]) >> 14);
    }

    for (i = kMinPrefBand; i <= kMaxPrefBand; i++)
    {
      avgHnl32 += (int32_t)hnl[i];
    }
    assert(kMaxPrefBand - kMinPrefBand + 1 > 0);
    avgHnl32 /= (kMaxPrefBand - kMinPrefBand + 1);

    for (i = kMaxPrefBand; i < PART_LEN1; i++)
    {
      if (hnl[i] > (int16_t)avgHnl32)
      {
        hnl[i] = (int16_t)avgHnl32;
      }
    }
  }

  // Calculate NLP gain, result is in Q14
  if (aecm->nlpFlag)
  {
    for (i = 0; i < PART_LEN1; i++)
    {
      // Truncate values close to zero and one.
      if (hnl[i] > NLP_COMP_HIGH)
      {
        hnl[i] = ONE_Q14;
      } else if (hnl[i] < NLP_COMP_LOW)
      {
        hnl[i] = 0;
      }

      // Remove outliers
      if (numPosCoef < 3)
      {
        nlpGain = 0;
      } else
      {
        nlpGain = ONE_Q14;
      }

      // NLP
      if ((hnl[i] == ONE_Q14) && (nlpGain == ONE_Q14))
      {
        hnl[i] = ONE_Q14;
      } else
      {
        hnl[i] = (int16_t)((hnl[i] * nlpGain) >> 14);
      }

      // multiply with Wiener coefficients
      efw[i].real = (int16_t)(WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].real,
                                                                   hnl[i], 14));
      efw[i].imag = (int16_t)(WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].imag,
                                                                   hnl[i], 14));
    }
  }
  else
  {
    // multiply with Wiener coefficients
    for (i = 0; i < PART_LEN1; i++)
    {
      efw[i].real = (int16_t)(WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].real,
                                                                   hnl[i], 14));
      efw[i].imag = (int16_t)(WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(dfw[i].imag,
                                                                   hnl[i], 14));
    }
  }

  if (aecm->cngMode == AecmTrue)
  {
    ComfortNoise(aecm, ptrDfaClean, efw, hnl);
  }

  InverseFFTAndWindow(aecm, fft, efw, output, nearendClean);

  return 0;
}

static void ComfortNoise(AecmCore* aecm,
                         const uint16_t* dfa,
                         ComplexInt16* out,
                         const int16_t* lambda) {
  int16_t i;
  int16_t tmp16;
  int32_t tmp32;

  int16_t randW16[PART_LEN];
  int16_t uReal[PART_LEN1];
  int16_t uImag[PART_LEN1];
  int32_t outLShift32;
  int16_t noiseRShift16[PART_LEN1];

  int16_t shiftFromNearToNoise = kNoiseEstQDomain - aecm->dfaCleanQDomain;
  int16_t minTrackShift;

  assert(shiftFromNearToNoise >= 0);
  assert(shiftFromNearToNoise < 16);

  if (aecm->noiseEstCtr < 100)
  {
    // Track the minimum more quickly initially.
    aecm->noiseEstCtr++;
    minTrackShift = 6;
  } else
  {
    minTrackShift = 9;
  }

  // Estimate noise power.
  for (i = 0; i < PART_LEN1; i++)
  {
    // Shift to the noise domain.
    tmp32 = (int32_t)dfa[i];
    outLShift32 = tmp32 << shiftFromNearToNoise;

    if (outLShift32 < aecm->noiseEst[i])
    {
      // Reset "too low" counter
      aecm->noiseEstTooLowCtr[i] = 0;
      // Track the minimum.
      if (aecm->noiseEst[i] < (1 << minTrackShift))
      {
        // For small values, decrease noiseEst[i] every
        // |kNoiseEstIncCount| block. The regular approach below can not
        // go further down due to truncation.
        aecm->noiseEstTooHighCtr[i]++;
        if (aecm->noiseEstTooHighCtr[i] >= kNoiseEstIncCount)
        {
          aecm->noiseEst[i]--;
          aecm->noiseEstTooHighCtr[i] = 0; // Reset the counter
        }
      }
      else
      {
        aecm->noiseEst[i] -= ((aecm->noiseEst[i] - outLShift32)
                              >> minTrackShift);
      }
    } else
    {
      // Reset "too high" counter
      aecm->noiseEstTooHighCtr[i] = 0;
      // Ramp slowly upwards until we hit the minimum again.
      if ((aecm->noiseEst[i] >> 19) > 0)
      {
        // Avoid overflow.
        // Multiplication with 2049 will cause wrap around. Scale
        // down first and then multiply
        aecm->noiseEst[i] >>= 11;
        aecm->noiseEst[i] *= 2049;
      }
      else if ((aecm->noiseEst[i] >> 11) > 0)
      {
        // Large enough for relative increase
        aecm->noiseEst[i] *= 2049;
        aecm->noiseEst[i] >>= 11;
      }
      else
      {
        // Make incremental increases based on size every
        // |kNoiseEstIncCount| block
        aecm->noiseEstTooLowCtr[i]++;
        if (aecm->noiseEstTooLowCtr[i] >= kNoiseEstIncCount)
        {
          aecm->noiseEst[i] += (aecm->noiseEst[i] >> 9) + 1;
          aecm->noiseEstTooLowCtr[i] = 0; // Reset counter
        }
      }
    }
  }

  for (i = 0; i < PART_LEN1; i++)
  {
    tmp32 = aecm->noiseEst[i] >> shiftFromNearToNoise;
    if (tmp32 > 32767)
    {
      tmp32 = 32767;
      aecm->noiseEst[i] = tmp32 << shiftFromNearToNoise;
    }
    noiseRShift16[i] = (int16_t)tmp32;

    tmp16 = ONE_Q14 - lambda[i];
    noiseRShift16[i] = (int16_t)((tmp16 * noiseRShift16[i]) >> 14);
  }

  // Generate a uniform random array on [0 2^15-1].
  WebRtcSpl_RandUArray(randW16, PART_LEN, &aecm->seed);

  // Generate noise according to estimated energy.
  uReal[0] = 0; // Reject LF noise.
  uImag[0] = 0;
  for (i = 1; i < PART_LEN1; i++)
  {
    // Get a random index for the cos and sin tables over [0 359].
    tmp16 = (int16_t)((359 * randW16[i - 1]) >> 15);

    // Tables are in Q13.
    uReal[i] = (int16_t)((noiseRShift16[i] * WebRtcAecm_kCosTable[tmp16]) >>
        13);
    uImag[i] = (int16_t)((-noiseRShift16[i] * WebRtcAecm_kSinTable[tmp16]) >>
        13);
  }
  uImag[PART_LEN] = 0;

  for (i = 0; i < PART_LEN1; i++)
  {
    out[i].real = WebRtcSpl_AddSatW16(out[i].real, uReal[i]);
    out[i].imag = WebRtcSpl_AddSatW16(out[i].imag, uImag[i]);
  }
}