static void EstBufDelayExtended(aecpc_t* self) {
int reported_delay = self->msInSndCardBuf * sampMsNb * self->rate_factor;
int current_delay = reported_delay - WebRtcAec_system_delay(self->aec);
int delay_difference = 0;
// Before we proceed with the delay estimate filtering we:
// 1) Compensate for the frame that will be read.
// 2) Compensate for drift resampling.
// 3) Compensate for non-causality if needed, since the estimated delay can't
// be negative.
// 1) Compensating for the frame(s) that will be read/processed.
current_delay += FRAME_LEN * self->rate_factor;
// 2) Account for resampling frame delay.
if (self->skewMode == kAecTrue && self->resample == kAecTrue) {
current_delay -= kResamplingDelay;
}
// 3) Compensate for non-causality, if needed, by flushing two blocks.
if (current_delay < PART_LEN) {
current_delay += WebRtcAec_MoveFarReadPtr(self->aec, 2) * PART_LEN;
}
if (self->filtDelay == -1) {
self->filtDelay = WEBRTC_SPL_MAX(0, 0.5 * current_delay);
} else {
self->filtDelay = WEBRTC_SPL_MAX(
0, (short)(0.95 * self->filtDelay + 0.05 * current_delay));
}
delay_difference = self->filtDelay - self->knownDelay;
if (delay_difference > 384) {
if (self->lastDelayDiff < 128) {
self->timeForDelayChange = 0;
} else {
self->timeForDelayChange++;
}
} else if (delay_difference < 128 && self->knownDelay > 0) {
if (self->lastDelayDiff > 384) {
self->timeForDelayChange = 0;
} else {
self->timeForDelayChange++;
}
} else {
self->timeForDelayChange = 0;
}
self->lastDelayDiff = delay_difference;
if (self->timeForDelayChange > 25) {
self->knownDelay = WEBRTC_SPL_MAX((int)self->filtDelay - 256, 0);
}
}
static void EstBufDelayNormal(aecpc_t* aecpc) {
int nSampSndCard = aecpc->msInSndCardBuf * sampMsNb * aecpc->rate_factor;
int current_delay = nSampSndCard - WebRtcAec_system_delay(aecpc->aec);
int delay_difference = 0;
// Before we proceed with the delay estimate filtering we:
// 1) Compensate for the frame that will be read.
// 2) Compensate for drift resampling.
// 3) Compensate for non-causality if needed, since the estimated delay can't
// be negative.
// 1) Compensating for the frame(s) that will be read/processed.
current_delay += FRAME_LEN * aecpc->rate_factor;
// 2) Account for resampling frame delay.
if (aecpc->skewMode == kAecTrue && aecpc->resample == kAecTrue) {
current_delay -= kResamplingDelay;
}
// 3) Compensate for non-causality, if needed, by flushing one block.
if (current_delay < PART_LEN) {
current_delay += WebRtcAec_MoveFarReadPtr(aecpc->aec, 1) * PART_LEN;
}
// We use -1 to signal an initialized state in the "extended" implementation;
// compensate for that.
aecpc->filtDelay = aecpc->filtDelay < 0 ? 0 : aecpc->filtDelay;
aecpc->filtDelay =
WEBRTC_SPL_MAX(0, (short)(0.8 * aecpc->filtDelay + 0.2 * current_delay));
delay_difference = aecpc->filtDelay - aecpc->knownDelay;
if (delay_difference > 224) {
if (aecpc->lastDelayDiff < 96) {
aecpc->timeForDelayChange = 0;
} else {
aecpc->timeForDelayChange++;
}
} else if (delay_difference < 96 && aecpc->knownDelay > 0) {
if (aecpc->lastDelayDiff > 224) {
aecpc->timeForDelayChange = 0;
} else {
aecpc->timeForDelayChange++;
}
} else {
aecpc->timeForDelayChange = 0;
}
aecpc->lastDelayDiff = delay_difference;
if (aecpc->timeForDelayChange > 25) {
aecpc->knownDelay = WEBRTC_SPL_MAX((int)aecpc->filtDelay - 160, 0);
}
}
int16_t WebRtcIlbcfix_GainDequant(
/* (o) quantized gain value (Q14) */
int16_t index, /* (i) quantization index */
int16_t maxIn, /* (i) maximum of unquantized gain (Q14) */
int16_t stage /* (i) The stage of the search */
){
int16_t scale;
const int16_t *gain;
/* obtain correct scale factor */
scale=WEBRTC_SPL_ABS_W16(maxIn);
scale = WEBRTC_SPL_MAX(1638, scale); /* if lower than 0.1, set it to 0.1 */
/* select the quantization table and return the decoded value */
gain = WebRtcIlbcfix_kGain[stage];
return((int16_t)((WEBRTC_SPL_MUL_16_16(scale, gain[index])+8192)>>14));
}
/* Compute the energy of the rest of the cb memory
* by step wise adding and subtracting the next
* sample and the last sample respectively */
void WebRtcIlbcfix_CbMemEnergyCalc(
WebRtc_Word32 energy, /* (i) input start energy */
WebRtc_Word16 range, /* (i) number of iterations */
WebRtc_Word16 *ppi, /* (i) input pointer 1 */
WebRtc_Word16 *ppo, /* (i) input pointer 2 */
WebRtc_Word16 *energyW16, /* (o) Energy in the CB vectors */
WebRtc_Word16 *energyShifts, /* (o) Shift value of the energy */
WebRtc_Word16 scale, /* (i) The scaling of all energy values */
WebRtc_Word16 base_size /* (i) Index to where the energy values should be stored */
)
{
WebRtc_Word16 j,shft;
WebRtc_Word32 tmp;
WebRtc_Word16 *eSh_ptr;
WebRtc_Word16 *eW16_ptr;
eSh_ptr = &energyShifts[1+base_size];
eW16_ptr = &energyW16[1+base_size];
for(j=0;j<range-1;j++) {
/* Calculate next energy by a +/-
operation on the edge samples */
tmp = WEBRTC_SPL_MUL_16_16(*ppi, *ppi);
tmp -= WEBRTC_SPL_MUL_16_16(*ppo, *ppo);
energy += WEBRTC_SPL_RSHIFT_W32(tmp, scale);
energy = WEBRTC_SPL_MAX(energy, 0);
ppi--;
ppo--;
/* Normalize the energy into a WebRtc_Word16 and store
the number of shifts */
shft = (WebRtc_Word16)WebRtcSpl_NormW32(energy);
*eSh_ptr++ = shft;
tmp = WEBRTC_SPL_LSHIFT_W32(energy, shft);
*eW16_ptr++ = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp, 16);
}
}
/* Compute the energy of the rest of the cb memory
* by step wise adding and subtracting the next
* sample and the last sample respectively */
void WebRtcIlbcfix_CbMemEnergyCalc(
int32_t energy, /* (i) input start energy */
size_t range, /* (i) number of iterations */
int16_t *ppi, /* (i) input pointer 1 */
int16_t *ppo, /* (i) input pointer 2 */
int16_t *energyW16, /* (o) Energy in the CB vectors */
int16_t *energyShifts, /* (o) Shift value of the energy */
int scale, /* (i) The scaling of all energy values */
size_t base_size /* (i) Index to where energy values should be stored */
)
{
size_t j;
int16_t shft;
int32_t tmp;
int16_t *eSh_ptr;
int16_t *eW16_ptr;
eSh_ptr = &energyShifts[1+base_size];
eW16_ptr = &energyW16[1+base_size];
for (j = 0; j + 1 < range; j++) {
/* Calculate next energy by a +/-
operation on the edge samples */
tmp = (*ppi) * (*ppi) - (*ppo) * (*ppo);
energy += tmp >> scale;
energy = WEBRTC_SPL_MAX(energy, 0);
ppi--;
ppo--;
/* Normalize the energy into a int16_t and store
the number of shifts */
shft = (int16_t)WebRtcSpl_NormW32(energy);
*eSh_ptr++ = shft;
tmp = energy << shft;
*eW16_ptr++ = (int16_t)(tmp >> 16);
}
}
void WebRtcIlbcfix_EnergyInverse(
WebRtc_Word16 *energy, /* (i/o) Energy and inverse
energy (in Q29) */
int noOfEnergies) /* (i) The length of the energy
vector */
{
WebRtc_Word32 Nom=(WebRtc_Word32)0x1FFFFFFF;
WebRtc_Word16 *energyPtr;
int i;
/* Set the minimum energy value to 16384 to avoid overflow */
energyPtr=energy;
for (i=0; i<noOfEnergies; i++) {
(*energyPtr)=WEBRTC_SPL_MAX((*energyPtr),16384);
energyPtr++;
}
/* Calculate inverse energy in Q29 */
energyPtr=energy;
for (i=0; i<noOfEnergies; i++) {
(*energyPtr) = (WebRtc_Word16)WebRtcSpl_DivW32W16(Nom, (*energyPtr));
energyPtr++;
}
}
int WebRtcNetEQ_UpdateIatStatistics(AutomodeInst_t *inst, int maxBufLen,
uint16_t seqNumber, uint32_t timeStamp,
int32_t fsHz, int mdCodec, int streamingMode)
{
uint32_t timeIat; /* inter-arrival time */
int i;
int32_t tempsum = 0; /* temp summation */
int32_t tempvar; /* temporary variable */
int retval = 0; /* return value */
int16_t packetLenSamp; /* packet speech length in samples */
/****************/
/* Sanity check */
/****************/
if (maxBufLen <= 1 || fsHz <= 0)
{
/* maxBufLen must be at least 2 and fsHz must both be strictly positive */
return -1;
}
/****************************/
/* Update packet statistics */
/****************************/
/* Try calculating packet length from current and previous timestamps */
if (!WebRtcNetEQ_IsNewerTimestamp(timeStamp, inst->lastTimeStamp) ||
!WebRtcNetEQ_IsNewerSequenceNumber(seqNumber, inst->lastSeqNo))
{
/* Wrong timestamp or sequence order; revert to backup plan */
packetLenSamp = inst->packetSpeechLenSamp; /* use stored value */
}
else
{
/* calculate timestamps per packet */
packetLenSamp = (int16_t) WebRtcSpl_DivU32U16(timeStamp - inst->lastTimeStamp,
seqNumber - inst->lastSeqNo);
}
/* Check that the packet size is positive; if not, the statistics cannot be updated. */
if (inst->firstPacketReceived && packetLenSamp > 0)
{ /* packet size ok */
/* calculate inter-arrival time in integer packets (rounding down) */
timeIat = WebRtcSpl_DivW32W16(inst->packetIatCountSamp, packetLenSamp);
/* Special operations for streaming mode */
if (streamingMode != 0)
{
/*
* Calculate IAT in Q8, including fractions of a packet (i.e., more accurate
* than timeIat).
*/
int16_t timeIatQ8 = (int16_t) WebRtcSpl_DivW32W16(
WEBRTC_SPL_LSHIFT_W32(inst->packetIatCountSamp, 8), packetLenSamp);
/*
* Calculate cumulative sum iat with sequence number compensation (ideal arrival
* times makes this sum zero).
*/
inst->cSumIatQ8 += (timeIatQ8
- WEBRTC_SPL_LSHIFT_W32(seqNumber - inst->lastSeqNo, 8));
/* subtract drift term */
inst->cSumIatQ8 -= CSUM_IAT_DRIFT;
/* ensure not negative */
inst->cSumIatQ8 = WEBRTC_SPL_MAX(inst->cSumIatQ8, 0);
/* remember max */
if (inst->cSumIatQ8 > inst->maxCSumIatQ8)
{
inst->maxCSumIatQ8 = inst->cSumIatQ8;
inst->maxCSumUpdateTimer = 0;
}
/* too long since the last maximum was observed; decrease max value */
if (inst->maxCSumUpdateTimer > (uint32_t) WEBRTC_SPL_MUL_32_16(fsHz,
MAX_STREAMING_PEAK_PERIOD))
{
inst->maxCSumIatQ8 -= 4; /* remove 1000*4/256 = 15.6 ms/s */
}
} /* end of streaming mode */
/* check for discontinuous packet sequence and re-ordering */
if (WebRtcNetEQ_IsNewerSequenceNumber(seqNumber, inst->lastSeqNo + 1))
{
/* Compensate for gap in the sequence numbers.
* Reduce IAT with expected extra time due to lost packets, but ensure that
* the IAT is not negative.
*/
timeIat -= WEBRTC_SPL_MIN(timeIat,
(uint16_t) (seqNumber - (uint16_t) (inst->lastSeqNo + 1)));
}
else if (!WebRtcNetEQ_IsNewerSequenceNumber(seqNumber, inst->lastSeqNo))
{
/* compensate for re-ordering */
timeIat += (uint16_t) (inst->lastSeqNo + 1 - seqNumber);
}
/* saturate IAT at maximum value */
timeIat = WEBRTC_SPL_MIN( timeIat, MAX_IAT );
/* update iatProb = forgetting_factor * iatProb for all elements */
for (i = 0; i <= MAX_IAT; i++)
{
int32_t tempHi, tempLo; /* Temporary variables */
/*
* Multiply iatProbFact (Q15) with iatProb (Q30) and right-shift 15 steps
* to come back to Q30. The operation is done in two steps:
*/
/*
* 1) Multiply the high 16 bits (15 bits + sign) of iatProb. Shift iatProb
* 16 steps right to get the high 16 bits in a int16_t prior to
* multiplication, and left-shift with 1 afterwards to come back to
* Q30 = (Q15 * (Q30>>16)) << 1.
*/
tempHi = WEBRTC_SPL_MUL_16_16(inst->iatProbFact,
(int16_t) WEBRTC_SPL_RSHIFT_W32(inst->iatProb[i], 16));
tempHi = WEBRTC_SPL_LSHIFT_W32(tempHi, 1); /* left-shift 1 step */
/*
* 2) Isolate and multiply the low 16 bits of iatProb. Right-shift 15 steps
* afterwards to come back to Q30 = (Q15 * Q30) >> 15.
*/
tempLo = inst->iatProb[i] & 0x0000FFFF; /* sift out the 16 low bits */
tempLo = WEBRTC_SPL_MUL_16_U16(inst->iatProbFact,
(uint16_t) tempLo);
tempLo = WEBRTC_SPL_RSHIFT_W32(tempLo, 15);
/* Finally, add the high and low parts */
inst->iatProb[i] = tempHi + tempLo;
/* Sum all vector elements while we are at it... */
tempsum += inst->iatProb[i];
}
/*
* Increase the probability for the currently observed inter-arrival time
* with 1 - iatProbFact. The factor is in Q15, iatProb in Q30;
* hence, left-shift 15 steps to obtain result in Q30.
*/
inst->iatProb[timeIat] += (32768 - inst->iatProbFact) << 15;
tempsum += (32768 - inst->iatProbFact) << 15; /* add to vector sum */
/*
* Update iatProbFact (changes only during the first seconds after reset)
* The factor converges to IAT_PROB_FACT.
*/
inst->iatProbFact += (IAT_PROB_FACT - inst->iatProbFact + 3) >> 2;
/* iatProb should sum up to 1 (in Q30). */
tempsum -= 1 << 30; /* should be zero */
/* Check if it does, correct if it doesn't. */
if (tempsum > 0)
{
/* tempsum too large => decrease a few values in the beginning */
i = 0;
while (i <= MAX_IAT && tempsum > 0)
{
/* Remove iatProb[i] / 16 from iatProb, but not more than tempsum */
tempvar = WEBRTC_SPL_MIN(tempsum, inst->iatProb[i] >> 4);
inst->iatProb[i++] -= tempvar;
tempsum -= tempvar;
}
}
int WebRtcIlbcfix_XcorrCoef(
WebRtc_Word16 *target, /* (i) first array */
WebRtc_Word16 *regressor, /* (i) second array */
WebRtc_Word16 subl, /* (i) dimension arrays */
WebRtc_Word16 searchLen, /* (i) the search lenght */
WebRtc_Word16 offset, /* (i) samples offset between arrays */
WebRtc_Word16 step /* (i) +1 or -1 */
){
int k;
WebRtc_Word16 maxlag;
WebRtc_Word16 pos;
WebRtc_Word16 max;
WebRtc_Word16 crossCorrScale, Energyscale;
WebRtc_Word16 crossCorrSqMod, crossCorrSqMod_Max;
WebRtc_Word32 crossCorr, Energy;
WebRtc_Word16 crossCorrmod, EnergyMod, EnergyMod_Max;
WebRtc_Word16 *tp, *rp;
WebRtc_Word16 *rp_beg, *rp_end;
WebRtc_Word16 totscale, totscale_max;
WebRtc_Word16 scalediff;
WebRtc_Word32 newCrit, maxCrit;
int shifts;
/* Initializations, to make sure that the first one is selected */
crossCorrSqMod_Max=0;
EnergyMod_Max=WEBRTC_SPL_WORD16_MAX;
totscale_max=-500;
maxlag=0;
pos=0;
/* Find scale value and start position */
if (step==1) {
max=WebRtcSpl_MaxAbsValueW16(regressor, (WebRtc_Word16)(subl+searchLen-1));
rp_beg = regressor;
rp_end = ®ressor[subl];
} else { /* step==-1 */
max=WebRtcSpl_MaxAbsValueW16(®ressor[-searchLen], (WebRtc_Word16)(subl+searchLen-1));
rp_beg = ®ressor[-1];
rp_end = ®ressor[subl-1];
}
/* Introduce a scale factor on the Energy in WebRtc_Word32 in
order to make sure that the calculation does not
overflow */
if (max>5000) {
shifts=2;
} else {
shifts=0;
}
/* Calculate the first energy, then do a +/- to get the other energies */
Energy=WebRtcSpl_DotProductWithScale(regressor, regressor, subl, shifts);
for (k=0;k<searchLen;k++) {
tp = target;
rp = ®ressor[pos];
crossCorr=WebRtcSpl_DotProductWithScale(tp, rp, subl, shifts);
if ((Energy>0)&&(crossCorr>0)) {
/* Put cross correlation and energy on 16 bit word */
crossCorrScale=(WebRtc_Word16)WebRtcSpl_NormW32(crossCorr)-16;
crossCorrmod=(WebRtc_Word16)WEBRTC_SPL_SHIFT_W32(crossCorr, crossCorrScale);
Energyscale=(WebRtc_Word16)WebRtcSpl_NormW32(Energy)-16;
EnergyMod=(WebRtc_Word16)WEBRTC_SPL_SHIFT_W32(Energy, Energyscale);
/* Square cross correlation and store upper WebRtc_Word16 */
crossCorrSqMod=(WebRtc_Word16)WEBRTC_SPL_MUL_16_16_RSFT(crossCorrmod, crossCorrmod, 16);
/* Calculate the total number of (dynamic) right shifts that have
been performed on (crossCorr*crossCorr)/energy
*/
totscale=Energyscale-(crossCorrScale<<1);
/* Calculate the shift difference in order to be able to compare the two
(crossCorr*crossCorr)/energy in the same domain
*/
scalediff=totscale-totscale_max;
scalediff=WEBRTC_SPL_MIN(scalediff,31);
scalediff=WEBRTC_SPL_MAX(scalediff,-31);
/* Compute the cross multiplication between the old best criteria
and the new one to be able to compare them without using a
division */
if (scalediff<0) {
newCrit = ((WebRtc_Word32)crossCorrSqMod*EnergyMod_Max)>>(-scalediff);
maxCrit = ((WebRtc_Word32)crossCorrSqMod_Max*EnergyMod);
} else {
static void NonLinearProcessing(aec_t *aec, int *ip, float *wfft, short *output, short *outputH)
{
float efw[2][PART_LEN1], dfw[2][PART_LEN1];
complex_t xfw[PART_LEN1];
complex_t comfortNoiseHband[PART_LEN1];
float fft[PART_LEN2];
float scale, dtmp;
float nlpGainHband;
int i, j, pos;
// Coherence and non-linear filter
float cohde[PART_LEN1], cohxd[PART_LEN1];
float hNlDeAvg, hNlXdAvg;
float hNl[PART_LEN1];
float hNlPref[PREF_BAND_SIZE];
float hNlFb = 0, hNlFbLow = 0;
const float prefBandQuant = 0.75f, prefBandQuantLow = 0.5f;
const int prefBandSize = PREF_BAND_SIZE / aec->mult;
const int minPrefBand = 4 / aec->mult;
// Near and error power sums
float sdSum = 0, seSum = 0;
// Power estimate smoothing coefficients
const float gCoh[2][2] = {{0.9f, 0.1f}, {0.93f, 0.07f}};
const float *ptrGCoh = gCoh[aec->mult - 1];
// Filter energey
float wfEnMax = 0, wfEn = 0;
const int delayEstInterval = 10 * aec->mult;
aec->delayEstCtr++;
if (aec->delayEstCtr == delayEstInterval) {
aec->delayEstCtr = 0;
}
// initialize comfort noise for H band
memset(comfortNoiseHband, 0, sizeof(comfortNoiseHband));
nlpGainHband = (float)0.0;
dtmp = (float)0.0;
// Measure energy in each filter partition to determine delay.
// TODO: Spread by computing one partition per block?
if (aec->delayEstCtr == 0) {
wfEnMax = 0;
aec->delayIdx = 0;
for (i = 0; i < NR_PART; i++) {
pos = i * PART_LEN1;
wfEn = 0;
for (j = 0; j < PART_LEN1; j++) {
wfEn += aec->wfBuf[0][pos + j] * aec->wfBuf[0][pos + j] +
aec->wfBuf[1][pos + j] * aec->wfBuf[1][pos + j];
}
if (wfEn > wfEnMax) {
wfEnMax = wfEn;
aec->delayIdx = i;
}
}
}
// NLP
// Windowed far fft
for (i = 0; i < PART_LEN; i++) {
fft[i] = aec->xBuf[i] * sqrtHanning[i];
fft[PART_LEN + i] = aec->xBuf[PART_LEN + i] * sqrtHanning[PART_LEN - i];
}
aec_rdft_128(1, fft, ip, wfft);
xfw[0][1] = 0;
xfw[PART_LEN][1] = 0;
xfw[0][0] = fft[0];
xfw[PART_LEN][0] = fft[1];
for (i = 1; i < PART_LEN; i++) {
xfw[i][0] = fft[2 * i];
xfw[i][1] = fft[2 * i + 1];
}
// Buffer far.
memcpy(aec->xfwBuf, xfw, sizeof(xfw));
// Use delayed far.
memcpy(xfw, aec->xfwBuf + aec->delayIdx * PART_LEN1, sizeof(xfw));
// Windowed near fft
for (i = 0; i < PART_LEN; i++) {
fft[i] = aec->dBuf[i] * sqrtHanning[i];
fft[PART_LEN + i] = aec->dBuf[PART_LEN + i] * sqrtHanning[PART_LEN - i];
}
aec_rdft_128(1, fft, ip, wfft);
dfw[1][0] = 0;
dfw[1][PART_LEN] = 0;
dfw[0][0] = fft[0];
dfw[0][PART_LEN] = fft[1];
for (i = 1; i < PART_LEN; i++) {
dfw[0][i] = fft[2 * i];
dfw[1][i] = fft[2 * i + 1];
}
// Windowed error fft
for (i = 0; i < PART_LEN; i++) {
fft[i] = aec->eBuf[i] * sqrtHanning[i];
fft[PART_LEN + i] = aec->eBuf[PART_LEN + i] * sqrtHanning[PART_LEN - i];
}
aec_rdft_128(1, fft, ip, wfft);
efw[1][0] = 0;
efw[1][PART_LEN] = 0;
efw[0][0] = fft[0];
efw[0][PART_LEN] = fft[1];
for (i = 1; i < PART_LEN; i++) {
efw[0][i] = fft[2 * i];
efw[1][i] = fft[2 * i + 1];
}
// Smoothed PSD
for (i = 0; i < PART_LEN1; i++) {
aec->sd[i] = ptrGCoh[0] * aec->sd[i] + ptrGCoh[1] *
(dfw[0][i] * dfw[0][i] + dfw[1][i] * dfw[1][i]);
aec->se[i] = ptrGCoh[0] * aec->se[i] + ptrGCoh[1] *
(efw[0][i] * efw[0][i] + efw[1][i] * efw[1][i]);
// We threshold here to protect against the ill-effects of a zero farend.
// The threshold is not arbitrarily chosen, but balances protection and
// adverse interaction with the algorithm's tuning.
// TODO: investigate further why this is so sensitive.
aec->sx[i] = ptrGCoh[0] * aec->sx[i] + ptrGCoh[1] *
WEBRTC_SPL_MAX(xfw[i][0] * xfw[i][0] + xfw[i][1] * xfw[i][1], 15);
aec->sde[i][0] = ptrGCoh[0] * aec->sde[i][0] + ptrGCoh[1] *
(dfw[0][i] * efw[0][i] + dfw[1][i] * efw[1][i]);
aec->sde[i][1] = ptrGCoh[0] * aec->sde[i][1] + ptrGCoh[1] *
(dfw[0][i] * efw[1][i] - dfw[1][i] * efw[0][i]);
aec->sxd[i][0] = ptrGCoh[0] * aec->sxd[i][0] + ptrGCoh[1] *
(dfw[0][i] * xfw[i][0] + dfw[1][i] * xfw[i][1]);
aec->sxd[i][1] = ptrGCoh[0] * aec->sxd[i][1] + ptrGCoh[1] *
(dfw[0][i] * xfw[i][1] - dfw[1][i] * xfw[i][0]);
sdSum += aec->sd[i];
seSum += aec->se[i];
}
// Divergent filter safeguard.
if (aec->divergeState == 0) {
if (seSum > sdSum) {
aec->divergeState = 1;
}
}
else {
if (seSum * 1.05f < sdSum) {
aec->divergeState = 0;
}
}
if (aec->divergeState == 1) {
memcpy(efw, dfw, sizeof(efw));
}
// Reset if error is significantly larger than nearend (13 dB).
if (seSum > (19.95f * sdSum)) {
memset(aec->wfBuf, 0, sizeof(aec->wfBuf));
}
// Subband coherence
for (i = 0; i < PART_LEN1; i++) {
cohde[i] = (aec->sde[i][0] * aec->sde[i][0] + aec->sde[i][1] * aec->sde[i][1]) /
(aec->sd[i] * aec->se[i] + 1e-10f);
cohxd[i] = (aec->sxd[i][0] * aec->sxd[i][0] + aec->sxd[i][1] * aec->sxd[i][1]) /
(aec->sx[i] * aec->sd[i] + 1e-10f);
}
hNlXdAvg = 0;
for (i = minPrefBand; i < prefBandSize + minPrefBand; i++) {
hNlXdAvg += cohxd[i];
}
hNlXdAvg /= prefBandSize;
hNlXdAvg = 1 - hNlXdAvg;
hNlDeAvg = 0;
for (i = minPrefBand; i < prefBandSize + minPrefBand; i++) {
hNlDeAvg += cohde[i];
}
hNlDeAvg /= prefBandSize;
if (hNlXdAvg < 0.75f && hNlXdAvg < aec->hNlXdAvgMin) {
aec->hNlXdAvgMin = hNlXdAvg;
}
if (hNlDeAvg > 0.98f && hNlXdAvg > 0.9f) {
aec->stNearState = 1;
}
else if (hNlDeAvg < 0.95f || hNlXdAvg < 0.8f) {
aec->stNearState = 0;
}
if (aec->hNlXdAvgMin == 1) {
aec->echoState = 0;
aec->overDrive = aec->minOverDrive;
if (aec->stNearState == 1) {
memcpy(hNl, cohde, sizeof(hNl));
hNlFb = hNlDeAvg;
hNlFbLow = hNlDeAvg;
}
else {
for (i = 0; i < PART_LEN1; i++) {
hNl[i] = 1 - cohxd[i];
}
hNlFb = hNlXdAvg;
hNlFbLow = hNlXdAvg;
}
}
else {
if (aec->stNearState == 1) {
aec->echoState = 0;
memcpy(hNl, cohde, sizeof(hNl));
hNlFb = hNlDeAvg;
hNlFbLow = hNlDeAvg;
}
else {
aec->echoState = 1;
for (i = 0; i < PART_LEN1; i++) {
hNl[i] = WEBRTC_SPL_MIN(cohde[i], 1 - cohxd[i]);
}
// Select an order statistic from the preferred bands.
// TODO: Using quicksort now, but a selection algorithm may be preferred.
memcpy(hNlPref, &hNl[minPrefBand], sizeof(float) * prefBandSize);
qsort(hNlPref, prefBandSize, sizeof(float), CmpFloat);
hNlFb = hNlPref[(int)floor(prefBandQuant * (prefBandSize - 1))];
hNlFbLow = hNlPref[(int)floor(prefBandQuantLow * (prefBandSize - 1))];
}
}
// Track the local filter minimum to determine suppression overdrive.
if (hNlFbLow < 0.6f && hNlFbLow < aec->hNlFbLocalMin) {
aec->hNlFbLocalMin = hNlFbLow;
aec->hNlFbMin = hNlFbLow;
aec->hNlNewMin = 1;
aec->hNlMinCtr = 0;
}
aec->hNlFbLocalMin = WEBRTC_SPL_MIN(aec->hNlFbLocalMin + 0.0008f / aec->mult, 1);
aec->hNlXdAvgMin = WEBRTC_SPL_MIN(aec->hNlXdAvgMin + 0.0006f / aec->mult, 1);
if (aec->hNlNewMin == 1) {
aec->hNlMinCtr++;
}
if (aec->hNlMinCtr == 2) {
aec->hNlNewMin = 0;
aec->hNlMinCtr = 0;
aec->overDrive = WEBRTC_SPL_MAX(aec->targetSupp /
((float)log(aec->hNlFbMin + 1e-10f) + 1e-10f), aec->minOverDrive);
}
// Smooth the overdrive.
if (aec->overDrive < aec->overDriveSm) {
aec->overDriveSm = 0.99f * aec->overDriveSm + 0.01f * aec->overDrive;
}
else {
aec->overDriveSm = 0.9f * aec->overDriveSm + 0.1f * aec->overDrive;
}
WebRtcAec_OverdriveAndSuppress(aec, hNl, hNlFb, efw);
#ifdef G167
if (aec->cnToggle) {
ComfortNoise(aec, efw, comfortNoiseHband, aec->noisePow, hNl);
}
#else
// Add comfort noise.
ComfortNoise(aec, efw, comfortNoiseHband, aec->noisePow, hNl);
#endif
// Inverse error fft.
fft[0] = efw[0][0];
fft[1] = efw[0][PART_LEN];
for (i = 1; i < PART_LEN; i++) {
fft[2*i] = efw[0][i];
// Sign change required by Ooura fft.
fft[2*i + 1] = -efw[1][i];
}
aec_rdft_128(-1, fft, ip, wfft);
// Overlap and add to obtain output.
scale = 2.0f / PART_LEN2;
for (i = 0; i < PART_LEN; i++) {
fft[i] *= scale; // fft scaling
fft[i] = fft[i]*sqrtHanning[i] + aec->outBuf[i];
// Saturation protection
output[i] = (short)WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, fft[i],
WEBRTC_SPL_WORD16_MIN);
fft[PART_LEN + i] *= scale; // fft scaling
aec->outBuf[i] = fft[PART_LEN + i] * sqrtHanning[PART_LEN - i];
}
// For H band
if (aec->sampFreq == 32000) {
// H band gain
// average nlp over low band: average over second half of freq spectrum
// (4->8khz)
GetHighbandGain(hNl, &nlpGainHband);
// Inverse comfort_noise
if (flagHbandCn == 1) {
fft[0] = comfortNoiseHband[0][0];
fft[1] = comfortNoiseHband[PART_LEN][0];
for (i = 1; i < PART_LEN; i++) {
fft[2*i] = comfortNoiseHband[i][0];
fft[2*i + 1] = comfortNoiseHband[i][1];
}
aec_rdft_128(-1, fft, ip, wfft);
scale = 2.0f / PART_LEN2;
}
// compute gain factor
for (i = 0; i < PART_LEN; i++) {
dtmp = (float)aec->dBufH[i];
dtmp = (float)dtmp * nlpGainHband; // for variable gain
// add some comfort noise where Hband is attenuated
if (flagHbandCn == 1) {
fft[i] *= scale; // fft scaling
dtmp += cnScaleHband * fft[i];
}
// Saturation protection
outputH[i] = (short)WEBRTC_SPL_SAT(WEBRTC_SPL_WORD16_MAX, dtmp,
WEBRTC_SPL_WORD16_MIN);
}
}
// Copy the current block to the old position.
memcpy(aec->xBuf, aec->xBuf + PART_LEN, sizeof(float) * PART_LEN);
memcpy(aec->dBuf, aec->dBuf + PART_LEN, sizeof(float) * PART_LEN);
memcpy(aec->eBuf, aec->eBuf + PART_LEN, sizeof(float) * PART_LEN);
// Copy the current block to the old position for H band
if (aec->sampFreq == 32000) {
memcpy(aec->dBufH, aec->dBufH + PART_LEN, sizeof(float) * PART_LEN);
}
memmove(aec->xfwBuf + PART_LEN1, aec->xfwBuf, sizeof(aec->xfwBuf) -
sizeof(complex_t) * PART_LEN1);
}
static void ComfortNoise(aec_t *aec, float efw[2][PART_LEN1],
complex_t *comfortNoiseHband, const float *noisePow, const float *lambda)
{
int i, num;
float rand[PART_LEN];
float noise, noiseAvg, tmp, tmpAvg;
WebRtc_Word16 randW16[PART_LEN];
complex_t u[PART_LEN1];
const float pi2 = 6.28318530717959f;
// Generate a uniform random array on [0 1]
WebRtcSpl_RandUArray(randW16, PART_LEN, &aec->seed);
for (i = 0; i < PART_LEN; i++) {
rand[i] = ((float)randW16[i]) / 32768;
}
// Reject LF noise
u[0][0] = 0;
u[0][1] = 0;
for (i = 1; i < PART_LEN1; i++) {
tmp = pi2 * rand[i - 1];
noise = sqrtf(noisePow[i]);
u[i][0] = noise * (float)cos(tmp);
u[i][1] = -noise * (float)sin(tmp);
}
u[PART_LEN][1] = 0;
for (i = 0; i < PART_LEN1; i++) {
// This is the proper weighting to match the background noise power
tmp = sqrtf(WEBRTC_SPL_MAX(1 - lambda[i] * lambda[i], 0));
//tmp = 1 - lambda[i];
efw[0][i] += tmp * u[i][0];
efw[1][i] += tmp * u[i][1];
}
// For H band comfort noise
// TODO: don't compute noise and "tmp" twice. Use the previous results.
noiseAvg = 0.0;
tmpAvg = 0.0;
num = 0;
if (aec->sampFreq == 32000 && flagHbandCn == 1) {
// average noise scale
// average over second half of freq spectrum (i.e., 4->8khz)
// TODO: we shouldn't need num. We know how many elements we're summing.
for (i = PART_LEN1 >> 1; i < PART_LEN1; i++) {
num++;
noiseAvg += sqrtf(noisePow[i]);
}
noiseAvg /= (float)num;
// average nlp scale
// average over second half of freq spectrum (i.e., 4->8khz)
// TODO: we shouldn't need num. We know how many elements we're summing.
num = 0;
for (i = PART_LEN1 >> 1; i < PART_LEN1; i++) {
num++;
tmpAvg += sqrtf(WEBRTC_SPL_MAX(1 - lambda[i] * lambda[i], 0));
}
tmpAvg /= (float)num;
// Use average noise for H band
// TODO: we should probably have a new random vector here.
// Reject LF noise
u[0][0] = 0;
u[0][1] = 0;
for (i = 1; i < PART_LEN1; i++) {
tmp = pi2 * rand[i - 1];
// Use average noise for H band
u[i][0] = noiseAvg * (float)cos(tmp);
u[i][1] = -noiseAvg * (float)sin(tmp);
}
u[PART_LEN][1] = 0;
for (i = 0; i < PART_LEN1; i++) {
// Use average NLP weight for H band
comfortNoiseHband[i][0] = tmpAvg * u[i][0];
comfortNoiseHband[i][1] = tmpAvg * u[i][1];
}
}
}
int32_t WebRtcAec_Process(void *aecInst, const int16_t *nearend,
const int16_t *nearendH, int16_t *out, int16_t *outH,
int16_t nrOfSamples, int16_t msInSndCardBuf,
int32_t skew)
{
aecpc_t *aecpc = aecInst;
int32_t retVal = 0;
short i;
short nBlocks10ms;
short nFrames;
// Limit resampling to doubling/halving of signal
const float minSkewEst = -0.5f;
const float maxSkewEst = 1.0f;
if (aecpc == NULL) {
return -1;
}
if (nearend == NULL) {
aecpc->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (out == NULL) {
aecpc->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (aecpc->initFlag != initCheck) {
aecpc->lastError = AEC_UNINITIALIZED_ERROR;
return -1;
}
// number of samples == 160 for SWB input
if (nrOfSamples != 80 && nrOfSamples != 160) {
aecpc->lastError = AEC_BAD_PARAMETER_ERROR;
return -1;
}
// Check for valid pointers based on sampling rate
if (aecpc->sampFreq == 32000 && nearendH == NULL) {
aecpc->lastError = AEC_NULL_POINTER_ERROR;
return -1;
}
if (msInSndCardBuf < 0) {
msInSndCardBuf = 0;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
retVal = -1;
}
else if (msInSndCardBuf > 500) {
msInSndCardBuf = 500;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
retVal = -1;
}
// TODO(andrew): we need to investigate if this +10 is really wanted.
msInSndCardBuf += 10;
aecpc->msInSndCardBuf = msInSndCardBuf;
if (aecpc->skewMode == kAecTrue) {
if (aecpc->skewFrCtr < 25) {
aecpc->skewFrCtr++;
}
else {
retVal = WebRtcAec_GetSkew(aecpc->resampler, skew, &aecpc->skew);
if (retVal == -1) {
aecpc->skew = 0;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
}
aecpc->skew /= aecpc->sampFactor*nrOfSamples;
if (aecpc->skew < 1.0e-3 && aecpc->skew > -1.0e-3) {
aecpc->resample = kAecFalse;
}
else {
aecpc->resample = kAecTrue;
}
if (aecpc->skew < minSkewEst) {
aecpc->skew = minSkewEst;
}
else if (aecpc->skew > maxSkewEst) {
aecpc->skew = maxSkewEst;
}
#ifdef WEBRTC_AEC_DEBUG_DUMP
(void)fwrite(&aecpc->skew, sizeof(aecpc->skew), 1, aecpc->skewFile);
#endif
}
}
nFrames = nrOfSamples / FRAME_LEN;
nBlocks10ms = nFrames / aecpc->rate_factor;
if (aecpc->ECstartup) {
if (nearend != out) {
// Only needed if they don't already point to the same place.
memcpy(out, nearend, sizeof(short) * nrOfSamples);
}
// The AEC is in the start up mode
// AEC is disabled until the system delay is OK
// Mechanism to ensure that the system delay is reasonably stable.
if (aecpc->checkBuffSize) {
aecpc->checkBufSizeCtr++;
// Before we fill up the far-end buffer we require the system delay
// to be stable (+/-8 ms) compared to the first value. This
// comparison is made during the following 6 consecutive 10 ms
// blocks. If it seems to be stable then we start to fill up the
// far-end buffer.
if (aecpc->counter == 0) {
aecpc->firstVal = aecpc->msInSndCardBuf;
aecpc->sum = 0;
}
if (abs(aecpc->firstVal - aecpc->msInSndCardBuf) <
WEBRTC_SPL_MAX(0.2 * aecpc->msInSndCardBuf, sampMsNb)) {
aecpc->sum += aecpc->msInSndCardBuf;
aecpc->counter++;
}
else {
aecpc->counter = 0;
}
if (aecpc->counter * nBlocks10ms >= 6) {
// The far-end buffer size is determined in partitions of
// PART_LEN samples. Use 75% of the average value of the system
// delay as buffer size to start with.
aecpc->bufSizeStart = WEBRTC_SPL_MIN((3 * aecpc->sum *
aecpc->rate_factor * 8) / (4 * aecpc->counter * PART_LEN),
kMaxBufSizeStart);
// Buffer size has now been determined.
aecpc->checkBuffSize = 0;
}
if (aecpc->checkBufSizeCtr * nBlocks10ms > 50) {
// For really bad systems, don't disable the echo canceller for
// more than 0.5 sec.
aecpc->bufSizeStart = WEBRTC_SPL_MIN((aecpc->msInSndCardBuf *
aecpc->rate_factor * 3) / 40, kMaxBufSizeStart);
aecpc->checkBuffSize = 0;
}
}
// If |checkBuffSize| changed in the if-statement above.
if (!aecpc->checkBuffSize) {
// The system delay is now reasonably stable (or has been unstable
// for too long). When the far-end buffer is filled with
// approximately the same amount of data as reported by the system
// we end the startup phase.
int overhead_elements =
WebRtcAec_system_delay(aecpc->aec) / PART_LEN -
aecpc->bufSizeStart;
if (overhead_elements == 0) {
// Enable the AEC
aecpc->ECstartup = 0;
} else if (overhead_elements > 0) {
// TODO(bjornv): Do we need a check on how much we actually
// moved the read pointer? It should always be possible to move
// the pointer |overhead_elements| since we have only added data
// to the buffer and no delay compensation nor AEC processing
// has been done.
WebRtcAec_MoveFarReadPtr(aecpc->aec, overhead_elements);
// Enable the AEC
aecpc->ECstartup = 0;
}
}
} else {
// AEC is enabled.
EstBufDelay(aecpc);
// Note that 1 frame is supported for NB and 2 frames for WB.
for (i = 0; i < nFrames; i++) {
// Call the AEC.
WebRtcAec_ProcessFrame(aecpc->aec,
&nearend[FRAME_LEN * i],
&nearendH[FRAME_LEN * i],
aecpc->knownDelay,
&out[FRAME_LEN * i],
&outH[FRAME_LEN * i]);
// TODO(bjornv): Re-structure such that we don't have to pass
// |aecpc->knownDelay| as input. Change name to something like
// |system_buffer_diff|.
}
}
#ifdef WEBRTC_AEC_DEBUG_DUMP
{
int16_t far_buf_size_ms = (int16_t)(WebRtcAec_system_delay(aecpc->aec) /
(sampMsNb * aecpc->rate_factor));
(void)fwrite(&far_buf_size_ms, 2, 1, aecpc->bufFile);
(void)fwrite(&aecpc->knownDelay, sizeof(aecpc->knownDelay), 1,
aecpc->delayFile);
}
#endif
return retVal;
}
void WebRtcIlbcfix_CbSearch(
IlbcEncoder *iLBCenc_inst,
/* (i) the encoder state structure */
int16_t *index, /* (o) Codebook indices */
int16_t *gain_index, /* (o) Gain quantization indices */
int16_t *intarget, /* (i) Target vector for encoding */
int16_t *decResidual,/* (i) Decoded residual for codebook construction */
int16_t lMem, /* (i) Length of buffer */
int16_t lTarget, /* (i) Length of vector */
int16_t *weightDenum,/* (i) weighting filter coefficients in Q12 */
int16_t block /* (i) the subblock number */
) {
int16_t i, j, stage, range;
int16_t *pp, scale, tmp;
int16_t bits, temp1, temp2;
int16_t base_size;
int32_t codedEner, targetEner;
int16_t gains[CB_NSTAGES+1];
int16_t *cb_vecPtr;
int16_t indexOffset, sInd, eInd;
int32_t CritMax=0;
int16_t shTotMax=WEBRTC_SPL_WORD16_MIN;
int16_t bestIndex=0;
int16_t bestGain=0;
int16_t indexNew, CritNewSh;
int32_t CritNew;
int32_t *cDotPtr;
int16_t noOfZeros;
int16_t *gainPtr;
int32_t t32, tmpW32;
int16_t *WebRtcIlbcfix_kGainSq5_ptr;
/* Stack based */
int16_t CBbuf[CB_MEML+LPC_FILTERORDER+CB_HALFFILTERLEN];
int32_t cDot[128];
int32_t Crit[128];
int16_t targetVec[SUBL+LPC_FILTERORDER];
int16_t cbvectors[CB_MEML + 1]; /* Adding one extra position for
Coverity warnings. */
int16_t codedVec[SUBL];
int16_t interpSamples[20*4];
int16_t interpSamplesFilt[20*4];
int16_t energyW16[CB_EXPAND*128];
int16_t energyShifts[CB_EXPAND*128];
int16_t *inverseEnergy=energyW16; /* Reuse memory */
int16_t *inverseEnergyShifts=energyShifts; /* Reuse memory */
int16_t *buf = &CBbuf[LPC_FILTERORDER];
int16_t *target = &targetVec[LPC_FILTERORDER];
int16_t *aug_vec = (int16_t*)cDot; /* length [SUBL], reuse memory */
/* Determine size of codebook sections */
base_size=lMem-lTarget+1;
if (lTarget==SUBL) {
base_size=lMem-19;
}
/* weighting of the CB memory */
noOfZeros=lMem-WebRtcIlbcfix_kFilterRange[block];
WebRtcSpl_MemSetW16(&buf[-LPC_FILTERORDER], 0, noOfZeros+LPC_FILTERORDER);
WebRtcSpl_FilterARFastQ12(
decResidual+noOfZeros, buf+noOfZeros,
weightDenum, LPC_FILTERORDER+1, WebRtcIlbcfix_kFilterRange[block]);
/* weighting of the target vector */
WEBRTC_SPL_MEMCPY_W16(&target[-LPC_FILTERORDER], buf+noOfZeros+WebRtcIlbcfix_kFilterRange[block]-LPC_FILTERORDER, LPC_FILTERORDER);
WebRtcSpl_FilterARFastQ12(
intarget, target,
weightDenum, LPC_FILTERORDER+1, lTarget);
/* Store target, towards the end codedVec is calculated as
the initial target minus the remaining target */
WEBRTC_SPL_MEMCPY_W16(codedVec, target, lTarget);
/* Find the highest absolute value to calculate proper
vector scale factor (so that it uses 12 bits) */
temp1 = WebRtcSpl_MaxAbsValueW16(buf, (int16_t)lMem);
temp2 = WebRtcSpl_MaxAbsValueW16(target, (int16_t)lTarget);
if ((temp1>0)&&(temp2>0)) {
temp1 = WEBRTC_SPL_MAX(temp1, temp2);
scale = WebRtcSpl_GetSizeInBits(WEBRTC_SPL_MUL_16_16(temp1, temp1));
} else {
/* temp1 or temp2 is negative (maximum was -32768) */
scale = 30;
}
/* Scale to so that a mul-add 40 times does not overflow */
scale = scale - 25;
scale = WEBRTC_SPL_MAX(0, scale);
/* Compute energy of the original target */
targetEner = WebRtcSpl_DotProductWithScale(target, target, lTarget, scale);
/* Prepare search over one more codebook section. This section
is created by filtering the original buffer with a filter. */
WebRtcIlbcfix_FilteredCbVecs(cbvectors, buf, lMem, WebRtcIlbcfix_kFilterRange[block]);
range = WebRtcIlbcfix_kSearchRange[block][0];
if(lTarget == SUBL) {
/* Create the interpolated samples and store them for use in all stages */
/* First section, non-filtered half of the cb */
WebRtcIlbcfix_InterpolateSamples(interpSamples, buf, lMem);
/* Second section, filtered half of the cb */
WebRtcIlbcfix_InterpolateSamples(interpSamplesFilt, cbvectors, lMem);
/* Compute the CB vectors' energies for the first cb section (non-filtered) */
WebRtcIlbcfix_CbMemEnergyAugmentation(interpSamples, buf,
scale, 20, energyW16, energyShifts);
/* Compute the CB vectors' energies for the second cb section (filtered cb) */
WebRtcIlbcfix_CbMemEnergyAugmentation(interpSamplesFilt, cbvectors,
scale, (int16_t)(base_size+20), energyW16, energyShifts);
/* Compute the CB vectors' energies and store them in the vector
* energyW16. Also the corresponding shift values are stored. The
* energy values are used in all three stages. */
WebRtcIlbcfix_CbMemEnergy(range, buf, cbvectors, lMem,
lTarget, energyW16+20, energyShifts+20, scale, base_size);
} else {
/* Compute the CB vectors' energies and store them in the vector
* energyW16. Also the corresponding shift values are stored. The
* energy values are used in all three stages. */
WebRtcIlbcfix_CbMemEnergy(range, buf, cbvectors, lMem,
lTarget, energyW16, energyShifts, scale, base_size);
/* Set the energy positions 58-63 and 122-127 to zero
(otherwise they are uninitialized) */
WebRtcSpl_MemSetW16(energyW16+range, 0, (base_size-range));
WebRtcSpl_MemSetW16(energyW16+range+base_size, 0, (base_size-range));
}
/* Calculate Inverse Energy (energyW16 is already normalized
and will contain the inverse energy in Q29 after this call */
WebRtcIlbcfix_EnergyInverse(energyW16, base_size*CB_EXPAND);
/* The gain value computed in the previous stage is used
* as an upper limit to what the next stage gain value
* is allowed to be. In stage 0, 16384 (1.0 in Q14) is used as
* the upper limit. */
gains[0] = 16384;
for (stage=0; stage<CB_NSTAGES; stage++) {
/* Set up memories */
range = WebRtcIlbcfix_kSearchRange[block][stage];
/* initialize search measures */
CritMax=0;
shTotMax=-100;
bestIndex=0;
bestGain=0;
/* loop over lags 40+ in the first codebook section, full search */
cb_vecPtr = buf+lMem-lTarget;
/* Calculate all the cross correlations (augmented part of CB) */
if (lTarget==SUBL) {
WebRtcIlbcfix_AugmentedCbCorr(target, buf+lMem,
interpSamples, cDot,
20, 39, scale);
cDotPtr=&cDot[20];
} else {
cDotPtr=cDot;
}
/* Calculate all the cross correlations (main part of CB) */
WebRtcSpl_CrossCorrelation(cDotPtr, target, cb_vecPtr, lTarget, range, scale, -1);
/* Adjust the search range for the augmented vectors */
if (lTarget==SUBL) {
range=WebRtcIlbcfix_kSearchRange[block][stage]+20;
} else {
range=WebRtcIlbcfix_kSearchRange[block][stage];
}
indexOffset=0;
/* Search for best index in this part of the vector */
WebRtcIlbcfix_CbSearchCore(
cDot, range, stage, inverseEnergy,
inverseEnergyShifts, Crit,
&indexNew, &CritNew, &CritNewSh);
/* Update the global best index and the corresponding gain */
WebRtcIlbcfix_CbUpdateBestIndex(
CritNew, CritNewSh, (int16_t)(indexNew+indexOffset), cDot[indexNew+indexOffset],
inverseEnergy[indexNew+indexOffset], inverseEnergyShifts[indexNew+indexOffset],
&CritMax, &shTotMax, &bestIndex, &bestGain);
sInd=bestIndex-(int16_t)(CB_RESRANGE>>1);
eInd=sInd+CB_RESRANGE;
if (sInd<0) {
eInd-=sInd;
sInd=0;
}
if (eInd>=range) {
eInd=range-1;
sInd=eInd-CB_RESRANGE;
}
range = WebRtcIlbcfix_kSearchRange[block][stage];
if (lTarget==SUBL) {
i=sInd;
if (sInd<20) {
WebRtcIlbcfix_AugmentedCbCorr(target, cbvectors+lMem,
interpSamplesFilt, cDot,
(int16_t)(sInd+20), (int16_t)(WEBRTC_SPL_MIN(39, (eInd+20))), scale);
i=20;
}
cDotPtr=&cDot[WEBRTC_SPL_MAX(0,(20-sInd))];
cb_vecPtr = cbvectors+lMem-20-i;
/* Calculate the cross correlations (main part of the filtered CB) */
WebRtcSpl_CrossCorrelation(cDotPtr, target, cb_vecPtr, lTarget, (int16_t)(eInd-i+1), scale, -1);
} else {
cDotPtr = cDot;
cb_vecPtr = cbvectors+lMem-lTarget-sInd;
/* Calculate the cross correlations (main part of the filtered CB) */
WebRtcSpl_CrossCorrelation(cDotPtr, target, cb_vecPtr, lTarget, (int16_t)(eInd-sInd+1), scale, -1);
}
/* Adjust the search range for the augmented vectors */
indexOffset=base_size+sInd;
/* Search for best index in this part of the vector */
WebRtcIlbcfix_CbSearchCore(
cDot, (int16_t)(eInd-sInd+1), stage, inverseEnergy+indexOffset,
inverseEnergyShifts+indexOffset, Crit,
&indexNew, &CritNew, &CritNewSh);
/* Update the global best index and the corresponding gain */
WebRtcIlbcfix_CbUpdateBestIndex(
CritNew, CritNewSh, (int16_t)(indexNew+indexOffset), cDot[indexNew],
inverseEnergy[indexNew+indexOffset], inverseEnergyShifts[indexNew+indexOffset],
&CritMax, &shTotMax, &bestIndex, &bestGain);
index[stage] = bestIndex;
bestGain = WebRtcIlbcfix_GainQuant(bestGain,
(int16_t)WEBRTC_SPL_ABS_W16(gains[stage]), stage, &gain_index[stage]);
/* Extract the best (according to measure) codebook vector
Also adjust the index, so that the augmented vectors are last.
Above these vectors were first...
*/
if(lTarget==(STATE_LEN-iLBCenc_inst->state_short_len)) {
if(index[stage]<base_size) {
pp=buf+lMem-lTarget-index[stage];
} else {
pp=cbvectors+lMem-lTarget-
index[stage]+base_size;
}
} else {
if (index[stage]<base_size) {
if (index[stage]>=20) {
/* Adjust index and extract vector */
index[stage]-=20;
pp=buf+lMem-lTarget-index[stage];
} else {
/* Adjust index and extract vector */
index[stage]+=(base_size-20);
WebRtcIlbcfix_CreateAugmentedVec((int16_t)(index[stage]-base_size+40),
buf+lMem, aug_vec);
pp = aug_vec;
}
} else {
if ((index[stage] - base_size) >= 20) {
/* Adjust index and extract vector */
index[stage]-=20;
pp=cbvectors+lMem-lTarget-
index[stage]+base_size;
} else {
/* Adjust index and extract vector */
index[stage]+=(base_size-20);
WebRtcIlbcfix_CreateAugmentedVec((int16_t)(index[stage]-2*base_size+40),
cbvectors+lMem, aug_vec);
pp = aug_vec;
}
}
}
/* Subtract the best codebook vector, according
to measure, from the target vector */
WebRtcSpl_AddAffineVectorToVector(target, pp, (int16_t)(-bestGain), (int32_t)8192, (int16_t)14, (int)lTarget);
/* record quantized gain */
gains[stage+1] = bestGain;
} /* end of Main Loop. for (stage=0;... */
/* Calculte the coded vector (original target - what's left) */
for (i=0;i<lTarget;i++) {
codedVec[i]-=target[i];
}
/* Gain adjustment for energy matching */
codedEner = WebRtcSpl_DotProductWithScale(codedVec, codedVec, lTarget, scale);
j=gain_index[0];
temp1 = (int16_t)WebRtcSpl_NormW32(codedEner);
temp2 = (int16_t)WebRtcSpl_NormW32(targetEner);
if(temp1 < temp2) {
bits = 16 - temp1;
} else {
bits = 16 - temp2;
}
tmp = (int16_t) WEBRTC_SPL_MUL_16_16_RSFT(gains[1],gains[1], 14);
targetEner = WEBRTC_SPL_MUL_16_16(
WEBRTC_SPL_SHIFT_W32(targetEner, -bits), tmp);
tmpW32 = ((int32_t)(gains[1]-1))<<1;
/* Pointer to the table that contains
gain_sq5TblFIX * gain_sq5TblFIX in Q14 */
gainPtr=(int16_t*)WebRtcIlbcfix_kGainSq5Sq+gain_index[0];
temp1 = (int16_t)WEBRTC_SPL_SHIFT_W32(codedEner, -bits);
WebRtcIlbcfix_kGainSq5_ptr = (int16_t*)&WebRtcIlbcfix_kGainSq5[j];
/* targetEner and codedEner are in Q(-2*scale) */
for (i=gain_index[0];i<32;i++) {
/* Change the index if
(codedEnergy*gainTbl[i]*gainTbl[i])<(targetEn*gain[0]*gain[0]) AND
gainTbl[i] < 2*gain[0]
*/
t32 = WEBRTC_SPL_MUL_16_16(temp1, (*gainPtr));
t32 = t32 - targetEner;
if (t32 < 0) {
if ((*WebRtcIlbcfix_kGainSq5_ptr) < tmpW32) {
j=i;
WebRtcIlbcfix_kGainSq5_ptr = (int16_t*)&WebRtcIlbcfix_kGainSq5[i];
}
}
gainPtr++;
}
gain_index[0]=j;
return;
}
static void ProcessExtended(Aec* self,
const float* const* near,
size_t num_bands,
float* const* out,
size_t num_samples,
int16_t reported_delay_ms,
int32_t skew) {
size_t i;
const int delay_diff_offset = kDelayDiffOffsetSamples;
#if defined(WEBRTC_UNTRUSTED_DELAY)
reported_delay_ms = kFixedDelayMs;
#else
// This is the usual mode where we trust the reported system delay values.
// Due to the longer filter, we no longer add 10 ms to the reported delay
// to reduce chance of non-causality. Instead we apply a minimum here to avoid
// issues with the read pointer jumping around needlessly.
reported_delay_ms = reported_delay_ms < kMinTrustedDelayMs
? kMinTrustedDelayMs
: reported_delay_ms;
// If the reported delay appears to be bogus, we attempt to recover by using
// the measured fixed delay values. We use >= here because higher layers
// may already clamp to this maximum value, and we would otherwise not
// detect it here.
reported_delay_ms = reported_delay_ms >= kMaxTrustedDelayMs
? kFixedDelayMs
: reported_delay_ms;
#endif
self->msInSndCardBuf = reported_delay_ms;
if (!self->farend_started) {
for (i = 0; i < num_bands; ++i) {
// Only needed if they don't already point to the same place.
if (near[i] != out[i]) {
memcpy(out[i], near[i], sizeof(near[i][0]) * num_samples);
}
}
return;
}
if (self->startup_phase) {
// In the extended mode, there isn't a startup "phase", just a special
// action on the first frame. In the trusted delay case, we'll take the
// current reported delay, unless it's less then our conservative
// measurement.
int startup_size_ms =
reported_delay_ms < kFixedDelayMs ? kFixedDelayMs : reported_delay_ms;
#if defined(WEBRTC_ANDROID) || defined(WEBRTC_GONK)
int target_delay = startup_size_ms * self->rate_factor * 8;
#else
// To avoid putting the AEC in a non-causal state we're being slightly
// conservative and scale by 2. On Android we use a fixed delay and
// therefore there is no need to scale the target_delay.
int target_delay = startup_size_ms * self->rate_factor * 8 / 2;
#endif
int overhead_elements =
(WebRtcAec_system_delay(self->aec) - target_delay) / PART_LEN;
WebRtcAec_MoveFarReadPtr(self->aec, overhead_elements);
self->startup_phase = 0;
}
EstBufDelayExtended(self);
{
// |delay_diff_offset| gives us the option to manually rewind the delay on
// very low delay platforms which can't be expressed purely through
// |reported_delay_ms|.
const int adjusted_known_delay =
WEBRTC_SPL_MAX(0, self->knownDelay + delay_diff_offset);
WebRtcAec_ProcessFrames(self->aec,
near,
num_bands,
num_samples,
adjusted_known_delay,
out);
}
}
// Updates the following smoothed Power Spectral Densities (PSD):
// - sd : near-end
// - se : residual echo
// - sx : far-end
// - sde : cross-PSD of near-end and residual echo
// - sxd : cross-PSD of near-end and far-end
//
// In addition to updating the PSDs, also the filter diverge state is determined
// upon actions are taken.
static void SmoothedPSD(AecCore* aec,
float efw[2][PART_LEN1],
float dfw[2][PART_LEN1],
float xfw[2][PART_LEN1],
int* extreme_filter_divergence) {
// Power estimate smoothing coefficients.
const float* ptrGCoh = aec->extended_filter_enabled
? WebRtcAec_kExtendedSmoothingCoefficients[aec->mult - 1]
: WebRtcAec_kNormalSmoothingCoefficients[aec->mult - 1];
int i;
float sdSum = 0, seSum = 0;
const float32x4_t vec_15 = vdupq_n_f32(WebRtcAec_kMinFarendPSD);
float32x4_t vec_sdSum = vdupq_n_f32(0.0f);
float32x4_t vec_seSum = vdupq_n_f32(0.0f);
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const float32x4_t vec_dfw0 = vld1q_f32(&dfw[0][i]);
const float32x4_t vec_dfw1 = vld1q_f32(&dfw[1][i]);
const float32x4_t vec_efw0 = vld1q_f32(&efw[0][i]);
const float32x4_t vec_efw1 = vld1q_f32(&efw[1][i]);
const float32x4_t vec_xfw0 = vld1q_f32(&xfw[0][i]);
const float32x4_t vec_xfw1 = vld1q_f32(&xfw[1][i]);
float32x4_t vec_sd = vmulq_n_f32(vld1q_f32(&aec->sd[i]), ptrGCoh[0]);
float32x4_t vec_se = vmulq_n_f32(vld1q_f32(&aec->se[i]), ptrGCoh[0]);
float32x4_t vec_sx = vmulq_n_f32(vld1q_f32(&aec->sx[i]), ptrGCoh[0]);
float32x4_t vec_dfw_sumsq = vmulq_f32(vec_dfw0, vec_dfw0);
float32x4_t vec_efw_sumsq = vmulq_f32(vec_efw0, vec_efw0);
float32x4_t vec_xfw_sumsq = vmulq_f32(vec_xfw0, vec_xfw0);
vec_dfw_sumsq = vmlaq_f32(vec_dfw_sumsq, vec_dfw1, vec_dfw1);
vec_efw_sumsq = vmlaq_f32(vec_efw_sumsq, vec_efw1, vec_efw1);
vec_xfw_sumsq = vmlaq_f32(vec_xfw_sumsq, vec_xfw1, vec_xfw1);
vec_xfw_sumsq = vmaxq_f32(vec_xfw_sumsq, vec_15);
vec_sd = vmlaq_n_f32(vec_sd, vec_dfw_sumsq, ptrGCoh[1]);
vec_se = vmlaq_n_f32(vec_se, vec_efw_sumsq, ptrGCoh[1]);
vec_sx = vmlaq_n_f32(vec_sx, vec_xfw_sumsq, ptrGCoh[1]);
vst1q_f32(&aec->sd[i], vec_sd);
vst1q_f32(&aec->se[i], vec_se);
vst1q_f32(&aec->sx[i], vec_sx);
{
float32x4x2_t vec_sde = vld2q_f32(&aec->sde[i][0]);
float32x4_t vec_dfwefw0011 = vmulq_f32(vec_dfw0, vec_efw0);
float32x4_t vec_dfwefw0110 = vmulq_f32(vec_dfw0, vec_efw1);
vec_sde.val[0] = vmulq_n_f32(vec_sde.val[0], ptrGCoh[0]);
vec_sde.val[1] = vmulq_n_f32(vec_sde.val[1], ptrGCoh[0]);
vec_dfwefw0011 = vmlaq_f32(vec_dfwefw0011, vec_dfw1, vec_efw1);
vec_dfwefw0110 = vmlsq_f32(vec_dfwefw0110, vec_dfw1, vec_efw0);
vec_sde.val[0] = vmlaq_n_f32(vec_sde.val[0], vec_dfwefw0011, ptrGCoh[1]);
vec_sde.val[1] = vmlaq_n_f32(vec_sde.val[1], vec_dfwefw0110, ptrGCoh[1]);
vst2q_f32(&aec->sde[i][0], vec_sde);
}
{
float32x4x2_t vec_sxd = vld2q_f32(&aec->sxd[i][0]);
float32x4_t vec_dfwxfw0011 = vmulq_f32(vec_dfw0, vec_xfw0);
float32x4_t vec_dfwxfw0110 = vmulq_f32(vec_dfw0, vec_xfw1);
vec_sxd.val[0] = vmulq_n_f32(vec_sxd.val[0], ptrGCoh[0]);
vec_sxd.val[1] = vmulq_n_f32(vec_sxd.val[1], ptrGCoh[0]);
vec_dfwxfw0011 = vmlaq_f32(vec_dfwxfw0011, vec_dfw1, vec_xfw1);
vec_dfwxfw0110 = vmlsq_f32(vec_dfwxfw0110, vec_dfw1, vec_xfw0);
vec_sxd.val[0] = vmlaq_n_f32(vec_sxd.val[0], vec_dfwxfw0011, ptrGCoh[1]);
vec_sxd.val[1] = vmlaq_n_f32(vec_sxd.val[1], vec_dfwxfw0110, ptrGCoh[1]);
vst2q_f32(&aec->sxd[i][0], vec_sxd);
}
vec_sdSum = vaddq_f32(vec_sdSum, vec_sd);
vec_seSum = vaddq_f32(vec_seSum, vec_se);
}
{
float32x2_t vec_sdSum_total;
float32x2_t vec_seSum_total;
// A B C D
vec_sdSum_total = vpadd_f32(vget_low_f32(vec_sdSum),
vget_high_f32(vec_sdSum));
vec_seSum_total = vpadd_f32(vget_low_f32(vec_seSum),
vget_high_f32(vec_seSum));
// A+B C+D
vec_sdSum_total = vpadd_f32(vec_sdSum_total, vec_sdSum_total);
vec_seSum_total = vpadd_f32(vec_seSum_total, vec_seSum_total);
// A+B+C+D A+B+C+D
sdSum = vget_lane_f32(vec_sdSum_total, 0);
seSum = vget_lane_f32(vec_seSum_total, 0);
}
// scalar code for the remaining items.
for (; i < PART_LEN1; i++) {
aec->sd[i] = ptrGCoh[0] * aec->sd[i] +
ptrGCoh[1] * (dfw[0][i] * dfw[0][i] + dfw[1][i] * dfw[1][i]);
aec->se[i] = ptrGCoh[0] * aec->se[i] +
ptrGCoh[1] * (efw[0][i] * efw[0][i] + efw[1][i] * efw[1][i]);
// We threshold here to protect against the ill-effects of a zero farend.
// The threshold is not arbitrarily chosen, but balances protection and
// adverse interaction with the algorithm's tuning.
// TODO(bjornv): investigate further why this is so sensitive.
aec->sx[i] =
ptrGCoh[0] * aec->sx[i] +
ptrGCoh[1] * WEBRTC_SPL_MAX(
xfw[0][i] * xfw[0][i] + xfw[1][i] * xfw[1][i],
WebRtcAec_kMinFarendPSD);
aec->sde[i][0] =
ptrGCoh[0] * aec->sde[i][0] +
ptrGCoh[1] * (dfw[0][i] * efw[0][i] + dfw[1][i] * efw[1][i]);
aec->sde[i][1] =
ptrGCoh[0] * aec->sde[i][1] +
ptrGCoh[1] * (dfw[0][i] * efw[1][i] - dfw[1][i] * efw[0][i]);
aec->sxd[i][0] =
ptrGCoh[0] * aec->sxd[i][0] +
ptrGCoh[1] * (dfw[0][i] * xfw[0][i] + dfw[1][i] * xfw[1][i]);
aec->sxd[i][1] =
ptrGCoh[0] * aec->sxd[i][1] +
ptrGCoh[1] * (dfw[0][i] * xfw[1][i] - dfw[1][i] * xfw[0][i]);
sdSum += aec->sd[i];
seSum += aec->se[i];
}
// Divergent filter safeguard update.
aec->divergeState = (aec->divergeState ? 1.05f : 1.0f) * seSum > sdSum;
// Signal extreme filter divergence if the error is significantly larger
// than the nearend (13 dB).
*extreme_filter_divergence = (seSum > (19.95f * sdSum));
}
static int ProcessNormal(Aec* aecpc,
const float* const* nearend,
int num_bands,
float* const* out,
int16_t nrOfSamples,
int16_t msInSndCardBuf,
int32_t skew) {
int retVal = 0;
short i;
short nBlocks10ms;
// Limit resampling to doubling/halving of signal
const float minSkewEst = -0.5f;
const float maxSkewEst = 1.0f;
msInSndCardBuf =
msInSndCardBuf > kMaxTrustedDelayMs ? kMaxTrustedDelayMs : msInSndCardBuf;
// TODO(andrew): we need to investigate if this +10 is really wanted.
msInSndCardBuf += 10;
aecpc->msInSndCardBuf = msInSndCardBuf;
if (aecpc->skewMode == kAecTrue) {
if (aecpc->skewFrCtr < 25) {
aecpc->skewFrCtr++;
} else {
retVal = WebRtcAec_GetSkew(aecpc->resampler, skew, &aecpc->skew);
if (retVal == -1) {
aecpc->skew = 0;
aecpc->lastError = AEC_BAD_PARAMETER_WARNING;
}
aecpc->skew /= aecpc->sampFactor * nrOfSamples;
if (aecpc->skew < 1.0e-3 && aecpc->skew > -1.0e-3) {
aecpc->resample = kAecFalse;
} else {
aecpc->resample = kAecTrue;
}
if (aecpc->skew < minSkewEst) {
aecpc->skew = minSkewEst;
} else if (aecpc->skew > maxSkewEst) {
aecpc->skew = maxSkewEst;
}
#ifdef WEBRTC_AEC_DEBUG_DUMP
(void)fwrite(&aecpc->skew, sizeof(aecpc->skew), 1, aecpc->skewFile);
#endif
}
}
nBlocks10ms = nrOfSamples / (FRAME_LEN * aecpc->rate_factor);
if (aecpc->startup_phase) {
for (i = 0; i < num_bands; ++i) {
// Only needed if they don't already point to the same place.
if (nearend[i] != out[i]) {
memcpy(out[i], nearend[i], sizeof(nearend[i][0]) * nrOfSamples);
}
}
// The AEC is in the start up mode
// AEC is disabled until the system delay is OK
// Mechanism to ensure that the system delay is reasonably stable.
if (aecpc->checkBuffSize) {
aecpc->checkBufSizeCtr++;
// Before we fill up the far-end buffer we require the system delay
// to be stable (+/-8 ms) compared to the first value. This
// comparison is made during the following 6 consecutive 10 ms
// blocks. If it seems to be stable then we start to fill up the
// far-end buffer.
if (aecpc->counter == 0) {
aecpc->firstVal = aecpc->msInSndCardBuf;
aecpc->sum = 0;
}
if (abs(aecpc->firstVal - aecpc->msInSndCardBuf) <
WEBRTC_SPL_MAX(0.2 * aecpc->msInSndCardBuf, sampMsNb)) {
aecpc->sum += aecpc->msInSndCardBuf;
aecpc->counter++;
} else {
aecpc->counter = 0;
}
if (aecpc->counter * nBlocks10ms >= 6) {
// The far-end buffer size is determined in partitions of
// PART_LEN samples. Use 75% of the average value of the system
// delay as buffer size to start with.
aecpc->bufSizeStart =
WEBRTC_SPL_MIN((3 * aecpc->sum * aecpc->rate_factor * 8) /
(4 * aecpc->counter * PART_LEN),
kMaxBufSizeStart);
// Buffer size has now been determined.
aecpc->checkBuffSize = 0;
}
if (aecpc->checkBufSizeCtr * nBlocks10ms > 50) {
// For really bad systems, don't disable the echo canceller for
// more than 0.5 sec.
aecpc->bufSizeStart = WEBRTC_SPL_MIN(
(aecpc->msInSndCardBuf * aecpc->rate_factor * 3) / 40,
kMaxBufSizeStart);
aecpc->checkBuffSize = 0;
}
}
// If |checkBuffSize| changed in the if-statement above.
if (!aecpc->checkBuffSize) {
// The system delay is now reasonably stable (or has been unstable
// for too long). When the far-end buffer is filled with
// approximately the same amount of data as reported by the system
// we end the startup phase.
int overhead_elements =
WebRtcAec_system_delay(aecpc->aec) / PART_LEN - aecpc->bufSizeStart;
if (overhead_elements == 0) {
// Enable the AEC
aecpc->startup_phase = 0;
} else if (overhead_elements > 0) {
// TODO(bjornv): Do we need a check on how much we actually
// moved the read pointer? It should always be possible to move
// the pointer |overhead_elements| since we have only added data
// to the buffer and no delay compensation nor AEC processing
// has been done.
WebRtcAec_MoveFarReadPtr(aecpc->aec, overhead_elements);
// Enable the AEC
aecpc->startup_phase = 0;
}
}
} else {
// AEC is enabled.
if (WebRtcAec_reported_delay_enabled(aecpc->aec)) {
EstBufDelayNormal(aecpc);
}
// Call the AEC.
// TODO(bjornv): Re-structure such that we don't have to pass
// |aecpc->knownDelay| as input. Change name to something like
// |system_buffer_diff|.
WebRtcAec_ProcessFrames(aecpc->aec,
nearend,
num_bands,
nrOfSamples,
aecpc->knownDelay,
out);
}
return retVal;
}
static void ProcessExtended(aecpc_t* self,
const int16_t* near,
const int16_t* near_high,
int16_t* out,
int16_t* out_high,
int16_t num_samples,
int16_t reported_delay_ms,
int32_t skew) {
int i;
const int num_frames = num_samples / FRAME_LEN;
#if defined(WEBRTC_UNTRUSTED_DELAY)
const int delay_diff_offset = kDelayDiffOffsetSamples;
reported_delay_ms = kFixedDelayMs;
#else
// This is the usual mode where we trust the reported system delay values.
const int delay_diff_offset = 0;
// Due to the longer filter, we no longer add 10 ms to the reported delay
// to reduce chance of non-causality. Instead we apply a minimum here to avoid
// issues with the read pointer jumping around needlessly.
reported_delay_ms = reported_delay_ms < kMinTrustedDelayMs
? kMinTrustedDelayMs
: reported_delay_ms;
// If the reported delay appears to be bogus, we attempt to recover by using
// the measured fixed delay values. We use >= here because higher layers
// may already clamp to this maximum value, and we would otherwise not
// detect it here.
reported_delay_ms = reported_delay_ms >= kMaxTrustedDelayMs
? kFixedDelayMs
: reported_delay_ms;
#endif
self->msInSndCardBuf = reported_delay_ms;
if (!self->farend_started) {
// Only needed if they don't already point to the same place.
if (near != out) {
memcpy(out, near, sizeof(short) * num_samples);
}
if (near_high != out_high) {
memcpy(out_high, near_high, sizeof(short) * num_samples);
}
return;
}
if (self->startup_phase) {
// In the extended mode, there isn't a startup "phase", just a special
// action on the first frame. In the trusted delay case, we'll take the
// current reported delay, unless it's less then our conservative
// measurement.
int startup_size_ms =
reported_delay_ms < kFixedDelayMs ? kFixedDelayMs : reported_delay_ms;
int overhead_elements = (WebRtcAec_system_delay(self->aec) -
startup_size_ms / 2 * self->rate_factor * 8) /
PART_LEN;
WebRtcAec_MoveFarReadPtr(self->aec, overhead_elements);
self->startup_phase = 0;
}
EstBufDelayExtended(self);
{
// |delay_diff_offset| gives us the option to manually rewind the delay on
// very low delay platforms which can't be expressed purely through
// |reported_delay_ms|.
const int adjusted_known_delay =
WEBRTC_SPL_MAX(0, self->knownDelay + delay_diff_offset);
for (i = 0; i < num_frames; ++i) {
WebRtcAec_ProcessFrame(self->aec,
&near[FRAME_LEN * i],
&near_high[FRAME_LEN * i],
adjusted_known_delay,
&out[FRAME_LEN * i],
&out_high[FRAME_LEN * i]);
}
}
}
void WebRtcIlbcfix_CbSearchCore(
int32_t *cDot, /* (i) Cross Correlation */
int16_t range, /* (i) Search range */
int16_t stage, /* (i) Stage of this search */
int16_t *inverseEnergy, /* (i) Inversed energy */
int16_t *inverseEnergyShift, /* (i) Shifts of inversed energy
with the offset 2*16-29 */
int32_t *Crit, /* (o) The criteria */
int16_t *bestIndex, /* (o) Index that corresponds to
maximum criteria (in this
vector) */
int32_t *bestCrit, /* (o) Value of critera for the
chosen index */
int16_t *bestCritSh) /* (o) The domain of the chosen
criteria */
{
int32_t maxW32, tmp32;
int16_t max, sh, tmp16;
int i;
int32_t *cDotPtr;
int16_t cDotSqW16;
int16_t *inverseEnergyPtr;
int32_t *critPtr;
int16_t *inverseEnergyShiftPtr;
/* Don't allow negative values for stage 0 */
if (stage==0) {
cDotPtr=cDot;
for (i=0;i<range;i++) {
*cDotPtr=WEBRTC_SPL_MAX(0, (*cDotPtr));
cDotPtr++;
}
}
/* Normalize cDot to int16_t, calculate the square of cDot and store the upper int16_t */
maxW32 = WebRtcSpl_MaxAbsValueW32(cDot, range);
sh = (int16_t)WebRtcSpl_NormW32(maxW32);
cDotPtr = cDot;
inverseEnergyPtr = inverseEnergy;
critPtr = Crit;
inverseEnergyShiftPtr=inverseEnergyShift;
max=WEBRTC_SPL_WORD16_MIN;
for (i=0;i<range;i++) {
/* Calculate cDot*cDot and put the result in a int16_t */
tmp32 = WEBRTC_SPL_LSHIFT_W32(*cDotPtr,sh);
tmp16 = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32,16);
cDotSqW16 = (int16_t)(((int32_t)(tmp16)*(tmp16))>>16);
/* Calculate the criteria (cDot*cDot/energy) */
*critPtr=WEBRTC_SPL_MUL_16_16(cDotSqW16, (*inverseEnergyPtr));
/* Extract the maximum shift value under the constraint
that the criteria is not zero */
if ((*critPtr)!=0) {
max = WEBRTC_SPL_MAX((*inverseEnergyShiftPtr), max);
}
inverseEnergyPtr++;
inverseEnergyShiftPtr++;
critPtr++;
cDotPtr++;
}
/* If no max shifts still at initialization value, set shift to zero */
if (max==WEBRTC_SPL_WORD16_MIN) {
max = 0;
}
/* Modify the criterias, so that all of them use the same Q domain */
critPtr=Crit;
inverseEnergyShiftPtr=inverseEnergyShift;
for (i=0;i<range;i++) {
/* Guarantee that the shift value is less than 16
in order to simplify for DSP's (and guard against >31) */
tmp16 = WEBRTC_SPL_MIN(16, max-(*inverseEnergyShiftPtr));
(*critPtr)=WEBRTC_SPL_SHIFT_W32((*critPtr),-tmp16);
critPtr++;
inverseEnergyShiftPtr++;
}
/* Find the index of the best value */
*bestIndex = WebRtcSpl_MaxIndexW32(Crit, range);
*bestCrit = Crit[*bestIndex];
/* Calculate total shifts of this criteria */
*bestCritSh = 32 - 2*sh + max;
return;
}
// Updates the following smoothed Power Spectral Densities (PSD):
// - sd : near-end
// - se : residual echo
// - sx : far-end
// - sde : cross-PSD of near-end and residual echo
// - sxd : cross-PSD of near-end and far-end
//
// In addition to updating the PSDs, also the filter diverge state is determined
// upon actions are taken.
static void SmoothedPSD(AecCore* aec,
float efw[2][PART_LEN1],
float dfw[2][PART_LEN1],
float xfw[2][PART_LEN1],
int* extreme_filter_divergence) {
// Power estimate smoothing coefficients.
const float* ptrGCoh = aec->extended_filter_enabled
? WebRtcAec_kExtendedSmoothingCoefficients[aec->mult - 1]
: WebRtcAec_kNormalSmoothingCoefficients[aec->mult - 1];
int i;
float sdSum = 0, seSum = 0;
const __m128 vec_15 = _mm_set1_ps(WebRtcAec_kMinFarendPSD);
const __m128 vec_GCoh0 = _mm_set1_ps(ptrGCoh[0]);
const __m128 vec_GCoh1 = _mm_set1_ps(ptrGCoh[1]);
__m128 vec_sdSum = _mm_set1_ps(0.0f);
__m128 vec_seSum = _mm_set1_ps(0.0f);
for (i = 0; i + 3 < PART_LEN1; i += 4) {
const __m128 vec_dfw0 = _mm_loadu_ps(&dfw[0][i]);
const __m128 vec_dfw1 = _mm_loadu_ps(&dfw[1][i]);
const __m128 vec_efw0 = _mm_loadu_ps(&efw[0][i]);
const __m128 vec_efw1 = _mm_loadu_ps(&efw[1][i]);
const __m128 vec_xfw0 = _mm_loadu_ps(&xfw[0][i]);
const __m128 vec_xfw1 = _mm_loadu_ps(&xfw[1][i]);
__m128 vec_sd = _mm_mul_ps(_mm_loadu_ps(&aec->sd[i]), vec_GCoh0);
__m128 vec_se = _mm_mul_ps(_mm_loadu_ps(&aec->se[i]), vec_GCoh0);
__m128 vec_sx = _mm_mul_ps(_mm_loadu_ps(&aec->sx[i]), vec_GCoh0);
__m128 vec_dfw_sumsq = _mm_mul_ps(vec_dfw0, vec_dfw0);
__m128 vec_efw_sumsq = _mm_mul_ps(vec_efw0, vec_efw0);
__m128 vec_xfw_sumsq = _mm_mul_ps(vec_xfw0, vec_xfw0);
vec_dfw_sumsq = _mm_add_ps(vec_dfw_sumsq, _mm_mul_ps(vec_dfw1, vec_dfw1));
vec_efw_sumsq = _mm_add_ps(vec_efw_sumsq, _mm_mul_ps(vec_efw1, vec_efw1));
vec_xfw_sumsq = _mm_add_ps(vec_xfw_sumsq, _mm_mul_ps(vec_xfw1, vec_xfw1));
vec_xfw_sumsq = _mm_max_ps(vec_xfw_sumsq, vec_15);
vec_sd = _mm_add_ps(vec_sd, _mm_mul_ps(vec_dfw_sumsq, vec_GCoh1));
vec_se = _mm_add_ps(vec_se, _mm_mul_ps(vec_efw_sumsq, vec_GCoh1));
vec_sx = _mm_add_ps(vec_sx, _mm_mul_ps(vec_xfw_sumsq, vec_GCoh1));
_mm_storeu_ps(&aec->sd[i], vec_sd);
_mm_storeu_ps(&aec->se[i], vec_se);
_mm_storeu_ps(&aec->sx[i], vec_sx);
{
const __m128 vec_3210 = _mm_loadu_ps(&aec->sde[i][0]);
const __m128 vec_7654 = _mm_loadu_ps(&aec->sde[i + 2][0]);
__m128 vec_a = _mm_shuffle_ps(vec_3210, vec_7654,
_MM_SHUFFLE(2, 0, 2, 0));
__m128 vec_b = _mm_shuffle_ps(vec_3210, vec_7654,
_MM_SHUFFLE(3, 1, 3, 1));
__m128 vec_dfwefw0011 = _mm_mul_ps(vec_dfw0, vec_efw0);
__m128 vec_dfwefw0110 = _mm_mul_ps(vec_dfw0, vec_efw1);
vec_a = _mm_mul_ps(vec_a, vec_GCoh0);
vec_b = _mm_mul_ps(vec_b, vec_GCoh0);
vec_dfwefw0011 = _mm_add_ps(vec_dfwefw0011,
_mm_mul_ps(vec_dfw1, vec_efw1));
vec_dfwefw0110 = _mm_sub_ps(vec_dfwefw0110,
_mm_mul_ps(vec_dfw1, vec_efw0));
vec_a = _mm_add_ps(vec_a, _mm_mul_ps(vec_dfwefw0011, vec_GCoh1));
vec_b = _mm_add_ps(vec_b, _mm_mul_ps(vec_dfwefw0110, vec_GCoh1));
_mm_storeu_ps(&aec->sde[i][0], _mm_unpacklo_ps(vec_a, vec_b));
_mm_storeu_ps(&aec->sde[i + 2][0], _mm_unpackhi_ps(vec_a, vec_b));
}
{
const __m128 vec_3210 = _mm_loadu_ps(&aec->sxd[i][0]);
const __m128 vec_7654 = _mm_loadu_ps(&aec->sxd[i + 2][0]);
__m128 vec_a = _mm_shuffle_ps(vec_3210, vec_7654,
_MM_SHUFFLE(2, 0, 2, 0));
__m128 vec_b = _mm_shuffle_ps(vec_3210, vec_7654,
_MM_SHUFFLE(3, 1, 3, 1));
__m128 vec_dfwxfw0011 = _mm_mul_ps(vec_dfw0, vec_xfw0);
__m128 vec_dfwxfw0110 = _mm_mul_ps(vec_dfw0, vec_xfw1);
vec_a = _mm_mul_ps(vec_a, vec_GCoh0);
vec_b = _mm_mul_ps(vec_b, vec_GCoh0);
vec_dfwxfw0011 = _mm_add_ps(vec_dfwxfw0011,
_mm_mul_ps(vec_dfw1, vec_xfw1));
vec_dfwxfw0110 = _mm_sub_ps(vec_dfwxfw0110,
_mm_mul_ps(vec_dfw1, vec_xfw0));
vec_a = _mm_add_ps(vec_a, _mm_mul_ps(vec_dfwxfw0011, vec_GCoh1));
vec_b = _mm_add_ps(vec_b, _mm_mul_ps(vec_dfwxfw0110, vec_GCoh1));
_mm_storeu_ps(&aec->sxd[i][0], _mm_unpacklo_ps(vec_a, vec_b));
_mm_storeu_ps(&aec->sxd[i + 2][0], _mm_unpackhi_ps(vec_a, vec_b));
}
vec_sdSum = _mm_add_ps(vec_sdSum, vec_sd);
vec_seSum = _mm_add_ps(vec_seSum, vec_se);
}
_mm_add_ps_4x1(vec_sdSum, &sdSum);
_mm_add_ps_4x1(vec_seSum, &seSum);
for (; i < PART_LEN1; i++) {
aec->sd[i] = ptrGCoh[0] * aec->sd[i] +
ptrGCoh[1] * (dfw[0][i] * dfw[0][i] + dfw[1][i] * dfw[1][i]);
aec->se[i] = ptrGCoh[0] * aec->se[i] +
ptrGCoh[1] * (efw[0][i] * efw[0][i] + efw[1][i] * efw[1][i]);
// We threshold here to protect against the ill-effects of a zero farend.
// The threshold is not arbitrarily chosen, but balances protection and
// adverse interaction with the algorithm's tuning.
// TODO(bjornv): investigate further why this is so sensitive.
aec->sx[i] =
ptrGCoh[0] * aec->sx[i] +
ptrGCoh[1] * WEBRTC_SPL_MAX(
xfw[0][i] * xfw[0][i] + xfw[1][i] * xfw[1][i],
WebRtcAec_kMinFarendPSD);
aec->sde[i][0] =
ptrGCoh[0] * aec->sde[i][0] +
ptrGCoh[1] * (dfw[0][i] * efw[0][i] + dfw[1][i] * efw[1][i]);
aec->sde[i][1] =
ptrGCoh[0] * aec->sde[i][1] +
ptrGCoh[1] * (dfw[0][i] * efw[1][i] - dfw[1][i] * efw[0][i]);
aec->sxd[i][0] =
ptrGCoh[0] * aec->sxd[i][0] +
ptrGCoh[1] * (dfw[0][i] * xfw[0][i] + dfw[1][i] * xfw[1][i]);
aec->sxd[i][1] =
ptrGCoh[0] * aec->sxd[i][1] +
ptrGCoh[1] * (dfw[0][i] * xfw[1][i] - dfw[1][i] * xfw[0][i]);
sdSum += aec->sd[i];
seSum += aec->se[i];
}
// Divergent filter safeguard update.
aec->divergeState = (aec->divergeState ? 1.05f : 1.0f) * seSum > sdSum;
// Signal extreme filter divergence if the error is significantly larger
// than the nearend (13 dB).
*extreme_filter_divergence = (seSum > (19.95f * sdSum));
}
void WebRtcIlbcfix_EncodeImpl(
WebRtc_UWord16 *bytes, /* (o) encoded data bits iLBC */
const WebRtc_Word16 *block, /* (i) speech vector to encode */
iLBC_Enc_Inst_t *iLBCenc_inst /* (i/o) the general encoder
state */
){
int n, meml_gotten, Nfor, Nback;
WebRtc_Word16 diff, start_pos;
int index;
int subcount, subframe;
WebRtc_Word16 start_count, end_count;
WebRtc_Word16 *residual;
WebRtc_Word32 en1, en2;
WebRtc_Word16 scale, max;
WebRtc_Word16 *syntdenum;
WebRtc_Word16 *decresidual;
WebRtc_Word16 *reverseResidual;
WebRtc_Word16 *reverseDecresidual;
/* Stack based */
WebRtc_Word16 weightdenum[(LPC_FILTERORDER + 1)*NSUB_MAX];
WebRtc_Word16 dataVec[BLOCKL_MAX + LPC_FILTERORDER];
WebRtc_Word16 memVec[CB_MEML+CB_FILTERLEN];
WebRtc_Word16 bitsMemory[sizeof(iLBC_bits)/sizeof(WebRtc_Word16)];
iLBC_bits *iLBCbits_inst = (iLBC_bits*)bitsMemory;
#ifdef SPLIT_10MS
WebRtc_Word16 *weightdenumbuf = iLBCenc_inst->weightdenumbuf;
WebRtc_Word16 last_bit;
#endif
WebRtc_Word16 *data = &dataVec[LPC_FILTERORDER];
WebRtc_Word16 *mem = &memVec[CB_HALFFILTERLEN];
/* Reuse som buffers to save stack memory */
residual = &iLBCenc_inst->lpc_buffer[LPC_LOOKBACK+BLOCKL_MAX-iLBCenc_inst->blockl];
syntdenum = mem; /* syntdenum[(LPC_FILTERORDER + 1)*NSUB_MAX] and mem are used non overlapping in the code */
decresidual = residual; /* Already encoded residual is overwritten by the decoded version */
reverseResidual = data; /* data and reverseResidual are used non overlapping in the code */
reverseDecresidual = reverseResidual; /* Already encoded residual is overwritten by the decoded version */
#ifdef SPLIT_10MS
WebRtcSpl_MemSetW16 ( (WebRtc_Word16 *) iLBCbits_inst, 0,
(WebRtc_Word16) (sizeof(iLBC_bits) / sizeof(WebRtc_Word16)) );
start_pos = iLBCenc_inst->start_pos;
diff = iLBCenc_inst->diff;
if (iLBCenc_inst->section != 0){
WEBRTC_SPL_MEMCPY_W16 (weightdenum, weightdenumbuf,
SCRATCH_ENCODE_DATAVEC - SCRATCH_ENCODE_WEIGHTDENUM);
/* Un-Packetize the frame into parameters */
last_bit = WebRtcIlbcfix_UnpackBits (iLBCenc_inst->bytes, iLBCbits_inst, iLBCenc_inst->mode);
if (last_bit)
return;
/* adjust index */
WebRtcIlbcfix_IndexConvDec (iLBCbits_inst->cb_index);
if (iLBCenc_inst->section == 1){
/* Save first 80 samples of a 160/240 sample frame for 20/30msec */
WEBRTC_SPL_MEMCPY_W16 (iLBCenc_inst->past_samples, block, 80);
}
else{ // iLBCenc_inst->section == 2 AND mode = 30ms
/* Save second 80 samples of a 240 sample frame for 30msec */
WEBRTC_SPL_MEMCPY_W16 (iLBCenc_inst->past_samples + 80, block, 80);
}
}
else{ // iLBCenc_inst->section == 0
/* form a complete frame of 160/240 for 20msec/30msec mode */
WEBRTC_SPL_MEMCPY_W16 (data + (iLBCenc_inst->mode * 8) - 80, block, 80);
WEBRTC_SPL_MEMCPY_W16 (data, iLBCenc_inst->past_samples,
(iLBCenc_inst->mode * 8) - 80);
iLBCenc_inst->Nfor_flag = 0;
iLBCenc_inst->Nback_flag = 0;
#else
/* copy input block to data*/
WEBRTC_SPL_MEMCPY_W16(data,block,iLBCenc_inst->blockl);
#endif
/* high pass filtering of input signal and scale down the residual (*0.5) */
WebRtcIlbcfix_HpInput(data, (WebRtc_Word16*)WebRtcIlbcfix_kHpInCoefs,
iLBCenc_inst->hpimemy, iLBCenc_inst->hpimemx,
iLBCenc_inst->blockl);
/* LPC of hp filtered input data */
WebRtcIlbcfix_LpcEncode(syntdenum, weightdenum, iLBCbits_inst->lsf, data,
iLBCenc_inst);
/* Set up state */
WEBRTC_SPL_MEMCPY_W16(dataVec, iLBCenc_inst->anaMem, LPC_FILTERORDER);
/* inverse filter to get residual */
for (n=0; n<iLBCenc_inst->nsub; n++ ) {
WebRtcSpl_FilterMAFastQ12(
&data[n*SUBL], &residual[n*SUBL],
&syntdenum[n*(LPC_FILTERORDER+1)],
LPC_FILTERORDER+1, SUBL);
}
/* Copy the state for next frame */
WEBRTC_SPL_MEMCPY_W16(iLBCenc_inst->anaMem, &data[iLBCenc_inst->blockl-LPC_FILTERORDER], LPC_FILTERORDER);
/* find state location */
iLBCbits_inst->startIdx = WebRtcIlbcfix_FrameClassify(iLBCenc_inst,residual);
/* check if state should be in first or last part of the
two subframes */
index = (iLBCbits_inst->startIdx-1)*SUBL;
max=WebRtcSpl_MaxAbsValueW16(&residual[index], 2*SUBL);
scale=WebRtcSpl_GetSizeInBits(WEBRTC_SPL_MUL_16_16(max,max));
/* Scale to maximum 25 bits so that the MAC won't cause overflow */
scale = scale - 25;
if(scale < 0) {
scale = 0;
}
diff = STATE_LEN - iLBCenc_inst->state_short_len;
en1=WebRtcSpl_DotProductWithScale(&residual[index], &residual[index],
iLBCenc_inst->state_short_len, scale);
index += diff;
en2=WebRtcSpl_DotProductWithScale(&residual[index], &residual[index],
iLBCenc_inst->state_short_len, scale);
if (en1 > en2) {
iLBCbits_inst->state_first = 1;
start_pos = (iLBCbits_inst->startIdx-1)*SUBL;
} else {
iLBCbits_inst->state_first = 0;
start_pos = (iLBCbits_inst->startIdx-1)*SUBL + diff;
}
/* scalar quantization of state */
WebRtcIlbcfix_StateSearch(iLBCenc_inst, iLBCbits_inst, &residual[start_pos],
&syntdenum[(iLBCbits_inst->startIdx-1)*(LPC_FILTERORDER+1)],
&weightdenum[(iLBCbits_inst->startIdx-1)*(LPC_FILTERORDER+1)]);
WebRtcIlbcfix_StateConstruct(iLBCbits_inst->idxForMax, iLBCbits_inst->idxVec,
&syntdenum[(iLBCbits_inst->startIdx-1)*(LPC_FILTERORDER+1)],
&decresidual[start_pos], iLBCenc_inst->state_short_len
);
/* predictive quantization in state */
if (iLBCbits_inst->state_first) { /* put adaptive part in the end */
/* setup memory */
WebRtcSpl_MemSetW16(mem, 0, (WebRtc_Word16)(CB_MEML-iLBCenc_inst->state_short_len));
WEBRTC_SPL_MEMCPY_W16(mem+CB_MEML-iLBCenc_inst->state_short_len,
decresidual+start_pos, iLBCenc_inst->state_short_len);
/* encode subframes */
WebRtcIlbcfix_CbSearch(iLBCenc_inst, iLBCbits_inst->cb_index, iLBCbits_inst->gain_index,
&residual[start_pos+iLBCenc_inst->state_short_len],
mem+CB_MEML-ST_MEM_L_TBL, ST_MEM_L_TBL, diff,
&weightdenum[iLBCbits_inst->startIdx*(LPC_FILTERORDER+1)], 0);
/* construct decoded vector */
WebRtcIlbcfix_CbConstruct(&decresidual[start_pos+iLBCenc_inst->state_short_len],
iLBCbits_inst->cb_index, iLBCbits_inst->gain_index,
mem+CB_MEML-ST_MEM_L_TBL, ST_MEM_L_TBL,
diff
);
}
else { /* put adaptive part in the beginning */
/* create reversed vectors for prediction */
WebRtcSpl_MemCpyReversedOrder(&reverseResidual[diff-1],
&residual[(iLBCbits_inst->startIdx+1)*SUBL-STATE_LEN], diff);
/* setup memory */
meml_gotten = iLBCenc_inst->state_short_len;
WebRtcSpl_MemCpyReversedOrder(&mem[CB_MEML-1], &decresidual[start_pos], meml_gotten);
WebRtcSpl_MemSetW16(mem, 0, (WebRtc_Word16)(CB_MEML-iLBCenc_inst->state_short_len));
/* encode subframes */
WebRtcIlbcfix_CbSearch(iLBCenc_inst, iLBCbits_inst->cb_index, iLBCbits_inst->gain_index,
reverseResidual, mem+CB_MEML-ST_MEM_L_TBL, ST_MEM_L_TBL, diff,
&weightdenum[(iLBCbits_inst->startIdx-1)*(LPC_FILTERORDER+1)],
0);
/* construct decoded vector */
WebRtcIlbcfix_CbConstruct(reverseDecresidual,
iLBCbits_inst->cb_index, iLBCbits_inst->gain_index,
mem+CB_MEML-ST_MEM_L_TBL, ST_MEM_L_TBL,
diff
);
/* get decoded residual from reversed vector */
WebRtcSpl_MemCpyReversedOrder(&decresidual[start_pos-1], reverseDecresidual, diff);
}
#ifdef SPLIT_10MS
iLBCenc_inst->start_pos = start_pos;
iLBCenc_inst->diff = diff;
iLBCenc_inst->section++;
/* adjust index */
WebRtcIlbcfix_IndexConvEnc (iLBCbits_inst->cb_index);
/* Packetize the parameters into the frame */
WebRtcIlbcfix_PackBits (iLBCenc_inst->bytes, iLBCbits_inst, iLBCenc_inst->mode);
WEBRTC_SPL_MEMCPY_W16 (weightdenumbuf, weightdenum,
SCRATCH_ENCODE_DATAVEC - SCRATCH_ENCODE_WEIGHTDENUM);
return;
}
#endif
/* forward prediction of subframes */
Nfor = iLBCenc_inst->nsub-iLBCbits_inst->startIdx-1;
/* counter for predicted subframes */
#ifdef SPLIT_10MS
if (iLBCenc_inst->mode == 20)
{
subcount = 1;
}
if (iLBCenc_inst->mode == 30)
{
if (iLBCenc_inst->section == 1)
{
subcount = 1;
}
if (iLBCenc_inst->section == 2)
{
subcount = 3;
}
}
#else
subcount=1;
#endif
if( Nfor > 0 ){
/* setup memory */
WebRtcSpl_MemSetW16(mem, 0, CB_MEML-STATE_LEN);
WEBRTC_SPL_MEMCPY_W16(mem+CB_MEML-STATE_LEN,
decresidual+(iLBCbits_inst->startIdx-1)*SUBL, STATE_LEN);
#ifdef SPLIT_10MS
if (iLBCenc_inst->Nfor_flag > 0)
{
for (subframe = 0; subframe < WEBRTC_SPL_MIN (Nfor, 2); subframe++)
{
/* update memory */
WEBRTC_SPL_MEMCPY_W16 (mem, mem + SUBL, (CB_MEML - SUBL));
WEBRTC_SPL_MEMCPY_W16 (mem + CB_MEML - SUBL,
&decresidual[(iLBCbits_inst->startIdx + 1 +
subframe) * SUBL], SUBL);
}
}
iLBCenc_inst->Nfor_flag++;
if (iLBCenc_inst->mode == 20)
{
start_count = 0;
end_count = Nfor;
}
if (iLBCenc_inst->mode == 30)
{
if (iLBCenc_inst->section == 1)
{
start_count = 0;
end_count = WEBRTC_SPL_MIN (Nfor, 2);
}
if (iLBCenc_inst->section == 2)
{
start_count = WEBRTC_SPL_MIN (Nfor, 2);
end_count = Nfor;
}
}
#else
start_count = 0;
end_count = (WebRtc_Word16)Nfor;
#endif
/* loop over subframes to encode */
for (subframe = start_count; subframe < end_count; subframe++){
/* encode subframe */
WebRtcIlbcfix_CbSearch(iLBCenc_inst, iLBCbits_inst->cb_index+subcount*CB_NSTAGES,
iLBCbits_inst->gain_index+subcount*CB_NSTAGES,
&residual[(iLBCbits_inst->startIdx+1+subframe)*SUBL],
mem, MEM_LF_TBL, SUBL,
&weightdenum[(iLBCbits_inst->startIdx+1+subframe)*(LPC_FILTERORDER+1)],
(WebRtc_Word16)subcount);
/* construct decoded vector */
WebRtcIlbcfix_CbConstruct(&decresidual[(iLBCbits_inst->startIdx+1+subframe)*SUBL],
iLBCbits_inst->cb_index+subcount*CB_NSTAGES,
iLBCbits_inst->gain_index+subcount*CB_NSTAGES,
mem, MEM_LF_TBL,
SUBL
);
/* update memory */
WEBRTC_SPL_MEMMOVE_W16(mem, mem+SUBL, (CB_MEML-SUBL));
WEBRTC_SPL_MEMCPY_W16(mem+CB_MEML-SUBL,
&decresidual[(iLBCbits_inst->startIdx+1+subframe)*SUBL], SUBL);
subcount++;
}
}
#ifdef SPLIT_10MS
if ((iLBCenc_inst->section == 1) &&
(iLBCenc_inst->mode == 30) && (Nfor > 0) && (end_count == 2))
{
iLBCenc_inst->section++;
/* adjust index */
WebRtcIlbcfix_IndexConvEnc (iLBCbits_inst->cb_index);
/* Packetize the parameters into the frame */
WebRtcIlbcfix_PackBits (iLBCenc_inst->bytes, iLBCbits_inst, iLBCenc_inst->mode);
WEBRTC_SPL_MEMCPY_W16 (weightdenumbuf, weightdenum,
SCRATCH_ENCODE_DATAVEC - SCRATCH_ENCODE_WEIGHTDENUM);
return;
}
#endif
/* backward prediction of subframes */
Nback = iLBCbits_inst->startIdx-1;
if( Nback > 0 ){
/* create reverse order vectors
(The decresidual does not need to be copied since it is
contained in the same vector as the residual)
*/
WebRtcSpl_MemCpyReversedOrder(&reverseResidual[Nback*SUBL-1], residual, Nback*SUBL);
/* setup memory */
meml_gotten = SUBL*(iLBCenc_inst->nsub+1-iLBCbits_inst->startIdx);
if( meml_gotten > CB_MEML ) {
meml_gotten=CB_MEML;
}
WebRtcSpl_MemCpyReversedOrder(&mem[CB_MEML-1], &decresidual[Nback*SUBL], meml_gotten);
WebRtcSpl_MemSetW16(mem, 0, (WebRtc_Word16)(CB_MEML-meml_gotten));
#ifdef SPLIT_10MS
if (iLBCenc_inst->Nback_flag > 0)
{
for (subframe = 0; subframe < WEBRTC_SPL_MAX (2 - Nfor, 0); subframe++)
{
/* update memory */
WEBRTC_SPL_MEMCPY_W16 (mem, mem + SUBL, (CB_MEML - SUBL));
WEBRTC_SPL_MEMCPY_W16 (mem + CB_MEML - SUBL,
&reverseDecresidual[subframe * SUBL], SUBL);
}
}
iLBCenc_inst->Nback_flag++;
if (iLBCenc_inst->mode == 20)
{
start_count = 0;
end_count = Nback;
}
if (iLBCenc_inst->mode == 30)
{
if (iLBCenc_inst->section == 1)
{
start_count = 0;
end_count = WEBRTC_SPL_MAX (2 - Nfor, 0);
}
if (iLBCenc_inst->section == 2)
{
start_count = WEBRTC_SPL_MAX (2 - Nfor, 0);
end_count = Nback;
}
}
#else
start_count = 0;
end_count = (WebRtc_Word16)Nback;
#endif
/* loop over subframes to encode */
for (subframe = start_count; subframe < end_count; subframe++){
/* encode subframe */
WebRtcIlbcfix_CbSearch(iLBCenc_inst, iLBCbits_inst->cb_index+subcount*CB_NSTAGES,
iLBCbits_inst->gain_index+subcount*CB_NSTAGES, &reverseResidual[subframe*SUBL],
mem, MEM_LF_TBL, SUBL,
&weightdenum[(iLBCbits_inst->startIdx-2-subframe)*(LPC_FILTERORDER+1)],
(WebRtc_Word16)subcount);
/* construct decoded vector */
WebRtcIlbcfix_CbConstruct(&reverseDecresidual[subframe*SUBL],
iLBCbits_inst->cb_index+subcount*CB_NSTAGES,
iLBCbits_inst->gain_index+subcount*CB_NSTAGES,
mem, MEM_LF_TBL, SUBL
);
/* update memory */
WEBRTC_SPL_MEMMOVE_W16(mem, mem+SUBL, (CB_MEML-SUBL));
WEBRTC_SPL_MEMCPY_W16(mem+CB_MEML-SUBL,
&reverseDecresidual[subframe*SUBL], SUBL);
subcount++;
}
/* get decoded residual from reversed vector */
WebRtcSpl_MemCpyReversedOrder(&decresidual[SUBL*Nback-1], reverseDecresidual, SUBL*Nback);
}
/* end encoding part */
/* adjust index */
WebRtcIlbcfix_IndexConvEnc(iLBCbits_inst->cb_index);
/* Packetize the parameters into the frame */
#ifdef SPLIT_10MS
if( (iLBCenc_inst->mode==30) && (iLBCenc_inst->section==1) ){
WebRtcIlbcfix_PackBits(iLBCenc_inst->bytes, iLBCbits_inst, iLBCenc_inst->mode);
}
else{
WebRtcIlbcfix_PackBits(bytes, iLBCbits_inst, iLBCenc_inst->mode);
}
#else
WebRtcIlbcfix_PackBits(bytes, iLBCbits_inst, iLBCenc_inst->mode);
#endif
#ifndef WEBRTC_BIG_ENDIAN
/* Swap bytes for LITTLE ENDIAN since the packbits()
function assumes BIG_ENDIAN machine */
#ifdef SPLIT_10MS
if (( (iLBCenc_inst->section == 1) && (iLBCenc_inst->mode == 20) ) ||
( (iLBCenc_inst->section == 2) && (iLBCenc_inst->mode == 30) )){
WebRtcIlbcfix_SwapBytes(bytes, iLBCenc_inst->no_of_words, bytes);
}
#else
WebRtcIlbcfix_SwapBytes(bytes, iLBCenc_inst->no_of_words, bytes);
#endif
#endif
#ifdef SPLIT_10MS
if (subcount == (iLBCenc_inst->nsub - 1))
{
iLBCenc_inst->section = 0;
}
else
{
iLBCenc_inst->section++;
WEBRTC_SPL_MEMCPY_W16 (weightdenumbuf, weightdenum,
SCRATCH_ENCODE_DATAVEC - SCRATCH_ENCODE_WEIGHTDENUM);
}
#endif
}
WebRtc_Word16 WebRtcNetEQ_PeakDetection(WebRtc_Word16 *pw16_data, WebRtc_Word16 w16_dataLen,
WebRtc_Word16 w16_nmbPeaks, WebRtc_Word16 fs_mult,
WebRtc_Word16 *pw16_winIndex,
WebRtc_Word16 *pw16_winValue)
{
/* Local variables */
int i;
WebRtc_Word16 w16_tmp;
WebRtc_Word16 w16_tmp2;
WebRtc_Word16 indMin = 0;
WebRtc_Word16 indMax = 0;
/* Peak detection */
for (i = 0; i <= (w16_nmbPeaks - 1); i++)
{
if (w16_nmbPeaks == 1)
{
/*
* Single peak
* The parabola fit assumes that an extra point is available; worst case it gets
* a zero on the high end of the signal.
*/
w16_dataLen++;
}
pw16_winIndex[i] = WebRtcSpl_MaxIndexW16(pw16_data, (WebRtc_Word16) (w16_dataLen - 1));
if (i != w16_nmbPeaks - 1)
{
w16_tmp = pw16_winIndex[i] - 2; /* *fs_mult; */
indMin = WEBRTC_SPL_MAX(0, w16_tmp);
w16_tmp = pw16_winIndex[i] + 2; /* *fs_mult; */
w16_tmp2 = w16_dataLen - 1;
indMax = WEBRTC_SPL_MIN(w16_tmp2, w16_tmp);
}
if ((pw16_winIndex[i] != 0) && (pw16_winIndex[i] != (w16_dataLen - 2)))
{
/* Parabola fit*/
WebRtcNetEQ_PrblFit(&(pw16_data[pw16_winIndex[i] - 1]), &(pw16_winIndex[i]),
&(pw16_winValue[i]), fs_mult);
}
else
{
if (pw16_winIndex[i] == (w16_dataLen - 2))
{
if (pw16_data[pw16_winIndex[i]] > pw16_data[pw16_winIndex[i] + 1])
{
WebRtcNetEQ_PrblFit(&(pw16_data[pw16_winIndex[i] - 1]),
&(pw16_winIndex[i]), &(pw16_winValue[i]), fs_mult);
}
else if (pw16_data[pw16_winIndex[i]] <= pw16_data[pw16_winIndex[i] + 1])
{
pw16_winValue[i] = (pw16_data[pw16_winIndex[i]]
+ pw16_data[pw16_winIndex[i] + 1]) >> 1; /* lin approx */
pw16_winIndex[i] = (pw16_winIndex[i] * 2 + 1) * fs_mult;
}
}
else
{
pw16_winValue[i] = pw16_data[pw16_winIndex[i]];
pw16_winIndex[i] = pw16_winIndex[i] * 2 * fs_mult;
}
}
