void WebRtcVad_Allpass(WebRtc_Word16 *in_vector,
WebRtc_Word16 *out_vector,
WebRtc_Word16 filter_coefficients,
int vector_length,
WebRtc_Word16 *filter_state)
{
// The filter can only cause overflow (in the w16 output variable)
// if more than 4 consecutive input numbers are of maximum value and
// has the the same sign as the impulse responses first taps.
// First 6 taps of the impulse response: 0.6399 0.5905 -0.3779
// 0.2418 -0.1547 0.0990
int n;
WebRtc_Word16 tmp16;
WebRtc_Word32 tmp32, in32, state32;
state32 = WEBRTC_SPL_LSHIFT_W32(((WebRtc_Word32)(*filter_state)), 16); // Q31
for (n = 0; n < vector_length; n++)
{
tmp32 = state32 + WEBRTC_SPL_MUL_16_16(filter_coefficients, (*in_vector));
tmp16 = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(tmp32, 16);
*out_vector++ = tmp16;
in32 = WEBRTC_SPL_LSHIFT_W32(((WebRtc_Word32)(*in_vector)), 14);
state32 = in32 - WEBRTC_SPL_MUL_16_16(filter_coefficients, tmp16);
state32 = WEBRTC_SPL_LSHIFT_W32(state32, 1);
in_vector += 2;
}
*filter_state = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(state32, 16);
}
/* The conversion is implemented by the step-down algorithm */
void WebRtcSpl_AToK_JSK(
WebRtc_Word16 *a16, /* Q11 */
WebRtc_Word16 useOrder,
WebRtc_Word16 *k16 /* Q15 */
)
{
int m, k;
WebRtc_Word32 tmp32[MAX_AR_MODEL_ORDER];
WebRtc_Word32 tmp32b;
WebRtc_Word32 tmp_inv_denum32;
WebRtc_Word16 tmp_inv_denum16;
k16[useOrder-1]= WEBRTC_SPL_LSHIFT_W16(a16[useOrder], 4); //Q11<<4 => Q15
for (m=useOrder-1; m>0; m--) {
tmp_inv_denum32 = ((WebRtc_Word32) 1073741823) - WEBRTC_SPL_MUL_16_16(k16[m], k16[m]); // (1 - k^2) in Q30
tmp_inv_denum16 = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(tmp_inv_denum32, 15); // (1 - k^2) in Q15
for (k=1; k<=m; k++) {
tmp32b = WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)a16[k], 16) -
WEBRTC_SPL_LSHIFT_W32(WEBRTC_SPL_MUL_16_16(k16[m], a16[m-k+1]), 1);
tmp32[k] = WebRtcSpl_DivW32W16(tmp32b, tmp_inv_denum16); //Q27/Q15 = Q12
}
for (k=1; k<m; k++) {
a16[k] = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(tmp32[k], 1); //Q12>>1 => Q11
}
tmp32[m] = WEBRTC_SPL_SAT(4092, tmp32[m], -4092);
k16[m-1] = (WebRtc_Word16) WEBRTC_SPL_LSHIFT_W32(tmp32[m], 3); //Q12<<3 => Q15
}
return;
}
static __inline int32_t Log2Q8( uint32_t x ) {
int32_t zeros, lg2;
int16_t frac;
zeros=WebRtcSpl_NormU32(x);
frac=(int16_t)WEBRTC_SPL_RSHIFT_W32(((uint32_t)(WEBRTC_SPL_LSHIFT_W32(x, zeros))&0x7FFFFFFF), 23);
/* log2(magn(i)) */
lg2= (WEBRTC_SPL_LSHIFT_W32((31-zeros), 8)+frac);
return lg2;
}
void WebRtcIlbcfix_CbMemEnergy(int16_t range, int16_t * CB, /* (i) The CB memory (1:st section) */
int16_t * filteredCB, /* (i) The filtered CB memory (2:nd section) */
int16_t lMem, /* (i) Length of the CB memory */
int16_t lTarget, /* (i) Length of the target vector */
int16_t * energyW16, /* (o) Energy in the CB vectors */
int16_t * energyShifts, /* (o) Shift value of the energy */
int16_t scale, /* (i) The scaling of all energy values */
int16_t base_size /* (i) Index to where the energy values should be stored */
)
{
int16_t *ppi, *ppo, *pp;
int32_t energy, tmp32;
/* Compute the energy and store it in a vector. Also the
* corresponding shift values are stored. The energy values
* are reused in all three stages. */
/* Calculate the energy in the first block of 'lTarget' sampels. */
ppi = CB + lMem - lTarget - 1;
ppo = CB + lMem - 1;
pp = CB + lMem - lTarget;
energy = WebRtcSpl_DotProductWithScale(pp, pp, lTarget, scale);
/* Normalize the energy and store the number of shifts */
energyShifts[0] = (int16_t) WebRtcSpl_NormW32(energy);
tmp32 = WEBRTC_SPL_LSHIFT_W32(energy, energyShifts[0]);
energyW16[0] = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmp32, 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. */
WebRtcIlbcfix_CbMemEnergyCalc(energy, range, ppi, ppo, energyW16,
energyShifts, scale, 0);
/* Next, precompute the energy values for the filtered cb section */
pp = filteredCB + lMem - lTarget;
energy = WebRtcSpl_DotProductWithScale(pp, pp, lTarget, scale);
/* Normalize the energy and store the number of shifts */
energyShifts[base_size] = (int16_t) WebRtcSpl_NormW32(energy);
tmp32 = WEBRTC_SPL_LSHIFT_W32(energy, energyShifts[base_size]);
energyW16[base_size] = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmp32, 16);
ppi = filteredCB + lMem - 1 - lTarget;
ppo = filteredCB + lMem - 1;
WebRtcIlbcfix_CbMemEnergyCalc(energy, range, ppi, ppo, energyW16,
energyShifts, scale, base_size);
}
void WebRtcSpl_LpcToReflCoef(int16_t* a16, int use_order, int16_t* k16)
{
int m, k;
int32_t tmp32[SPL_LPC_TO_REFL_COEF_MAX_AR_MODEL_ORDER];
int32_t tmp_inv_denom32;
int16_t tmp_inv_denom16;
k16[use_order - 1] = a16[use_order] << 3; // Q12<<3 => Q15
for (m = use_order - 1; m > 0; m--)
{
// (1 - k^2) in Q30
tmp_inv_denom32 = 1073741823 - k16[m] * k16[m];
// (1 - k^2) in Q15
tmp_inv_denom16 = (int16_t)(tmp_inv_denom32 >> 15);
for (k = 1; k <= m; k++)
{
// tmp[k] = (a[k] - RC[m] * a[m-k+1]) / (1.0 - RC[m]*RC[m]);
// [Q12<<16 - (Q15*Q12)<<1] = [Q28 - Q28] = Q28
tmp32[k] = (a16[k] << 16) - (k16[m] * a16[m - k + 1] << 1);
tmp32[k] = WebRtcSpl_DivW32W16(tmp32[k], tmp_inv_denom16); //Q28/Q15 = Q13
}
for (k = 1; k < m; k++)
{
a16[k] = (int16_t)(tmp32[k] >> 1); // Q13>>1 => Q12
}
tmp32[m] = WEBRTC_SPL_SAT(8191, tmp32[m], -8191);
k16[m - 1] = (int16_t)WEBRTC_SPL_LSHIFT_W32(tmp32[m], 2); //Q13<<2 => Q15
}
return;
}
static __inline int32_t CalcLrIntQ(int32_t fixVal,
int16_t qDomain) {
int32_t intgr;
int32_t roundVal;
roundVal = WEBRTC_SPL_LSHIFT_W32((int32_t)1, qDomain - 1);
intgr = WEBRTC_SPL_RSHIFT_W32(fixVal + roundVal, qDomain);
return intgr;
}
WebRtc_UWord32 WebRtcNetEQ_ScaleTimestampExternalToInternal(const MCUInst_t *MCU_inst,
WebRtc_UWord32 externalTS)
{
WebRtc_Word32 timestampDiff;
WebRtc_UWord32 internalTS;
/* difference between this and last incoming timestamp */
timestampDiff = externalTS - MCU_inst->externalTS;
switch (MCU_inst->scalingFactor)
{
case kTSscalingTwo:
{
/* multiply with 2 */
timestampDiff = WEBRTC_SPL_LSHIFT_W32(timestampDiff, 1);
break;
}
case kTSscalingTwoThirds:
{
/* multiply with 2/3 */
timestampDiff = WEBRTC_SPL_LSHIFT_W32(timestampDiff, 1);
timestampDiff = WebRtcSpl_DivW32W16(timestampDiff, 3);
break;
}
case kTSscalingFourThirds:
{
/* multiply with 4/3 */
timestampDiff = WEBRTC_SPL_LSHIFT_W32(timestampDiff, 2);
timestampDiff = WebRtcSpl_DivW32W16(timestampDiff, 3);
break;
}
default:
{
/* no scaling */
}
}
/* add the scaled difference to last scaled timestamp and save ... */
internalTS = MCU_inst->internalTS + timestampDiff;
return internalTS;
}
void WebRtcSpl_SynthesisQMF(const int16_t* low_band, const int16_t* high_band,
size_t band_length, int16_t* out_data,
int32_t* filter_state1, int32_t* filter_state2)
{
int32_t tmp;
int32_t half_in1[kMaxBandFrameLength];
int32_t half_in2[kMaxBandFrameLength];
int32_t filter1[kMaxBandFrameLength];
int32_t filter2[kMaxBandFrameLength];
size_t i;
int16_t k;
assert(band_length <= kMaxBandFrameLength);
// Obtain the sum and difference channels out of upper and lower-band channels.
// Also shift to Q10 domain.
for (i = 0; i < band_length; i++)
{
tmp = (int32_t)low_band[i] + (int32_t)high_band[i];
half_in1[i] = WEBRTC_SPL_LSHIFT_W32(tmp, 10);
tmp = (int32_t)low_band[i] - (int32_t)high_band[i];
half_in2[i] = WEBRTC_SPL_LSHIFT_W32(tmp, 10);
}
// all-pass filter the sum and difference channels
WebRtcSpl_AllPassQMF(half_in1, band_length, filter1,
WebRtcSpl_kAllPassFilter2, filter_state1);
WebRtcSpl_AllPassQMF(half_in2, band_length, filter2,
WebRtcSpl_kAllPassFilter1, filter_state2);
// The filtered signals are even and odd samples of the output. Combine
// them. The signals are Q10 should shift them back to Q0 and take care of
// saturation.
for (i = 0, k = 0; i < band_length; i++)
{
tmp = (filter2[i] + 512) >> 10;
out_data[k++] = WebRtcSpl_SatW32ToW16(tmp);
tmp = (filter1[i] + 512) >> 10;
out_data[k++] = WebRtcSpl_SatW32ToW16(tmp);
}
}
void WebRtcSpl_AnalysisQMF(const int16_t* in_data, size_t in_data_length,
int16_t* low_band, int16_t* high_band,
int32_t* filter_state1, int32_t* filter_state2)
{
size_t i;
int16_t k;
int32_t tmp;
int32_t half_in1[kMaxBandFrameLength];
int32_t half_in2[kMaxBandFrameLength];
int32_t filter1[kMaxBandFrameLength];
int32_t filter2[kMaxBandFrameLength];
const size_t band_length = in_data_length / 2;
assert(in_data_length % 2 == 0);
assert(band_length <= kMaxBandFrameLength);
// Split even and odd samples. Also shift them to Q10.
for (i = 0, k = 0; i < band_length; i++, k += 2)
{
half_in2[i] = WEBRTC_SPL_LSHIFT_W32((int32_t)in_data[k], 10);
half_in1[i] = WEBRTC_SPL_LSHIFT_W32((int32_t)in_data[k + 1], 10);
}
// All pass filter even and odd samples, independently.
WebRtcSpl_AllPassQMF(half_in1, band_length, filter1,
WebRtcSpl_kAllPassFilter1, filter_state1);
WebRtcSpl_AllPassQMF(half_in2, band_length, filter2,
WebRtcSpl_kAllPassFilter2, filter_state2);
// Take the sum and difference of filtered version of odd and even
// branches to get upper & lower band.
for (i = 0; i < band_length; i++)
{
tmp = (filter1[i] + filter2[i] + 1024) >> 11;
low_band[i] = WebRtcSpl_SatW32ToW16(tmp);
tmp = (filter1[i] - filter2[i] + 1024) >> 11;
high_band[i] = WebRtcSpl_SatW32ToW16(tmp);
}
}
static __inline WebRtc_Word32 log2_Q8_LPC( WebRtc_UWord32 x ) {
WebRtc_Word32 zeros, lg2;
WebRtc_Word16 frac;
zeros=WebRtcSpl_NormU32(x);
frac=(WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(((WebRtc_UWord32)WEBRTC_SPL_LSHIFT_W32(x, zeros)&0x7FFFFFFF), 23);
/* log2(x) */
lg2= (WEBRTC_SPL_LSHIFT_W16((31-zeros), 8)+frac);
return lg2;
}
static void AllpassFilterForDec32(WebRtc_Word16 *InOut16, //Q0
const WebRtc_Word32 *APSectionFactors, //Q15
WebRtc_Word16 lengthInOut,
WebRtc_Word32 *FilterState) //Q16
{
int n, j;
WebRtc_Word32 a, b;
for (j=0; j<ALLPASSSECTIONS; j++) {
for (n=0;n<lengthInOut;n+=2){
a = WEBRTC_SPL_MUL_16_32_RSFT16(InOut16[n], APSectionFactors[j]); //Q0*Q31=Q31 shifted 16 gives Q15
a = WEBRTC_SPL_LSHIFT_W32(a, 1); // Q15 -> Q16
b = WEBRTC_SPL_ADD_SAT_W32(a, FilterState[j]); //Q16+Q16=Q16
a = WEBRTC_SPL_MUL_16_32_RSFT16(
(WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(b, 16),
-APSectionFactors[j]); //Q0*Q31=Q31 shifted 16 gives Q15
FilterState[j] = WEBRTC_SPL_ADD_SAT_W32(
WEBRTC_SPL_LSHIFT_W32(a,1),
WEBRTC_SPL_LSHIFT_W32((WebRtc_UWord32)InOut16[n], 16)); // Q15<<1 + Q0<<16 = Q16 + Q16 = Q16
InOut16[n] = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(b, 16); //Save as Q0
}
}
}
void WebRtcIlbcfix_CbMemEnergyAugmentation(
int16_t *interpSamples, /* (i) The interpolated samples */
int16_t *CBmem, /* (i) The CB memory */
int16_t scale, /* (i) The scaling of all energy values */
int16_t base_size, /* (i) Index to where the energy values should be stored */
int16_t *energyW16, /* (o) Energy in the CB vectors */
int16_t *energyShifts /* (o) Shift value of the energy */
){
int32_t energy, tmp32;
int16_t *ppe, *pp, *interpSamplesPtr;
int16_t *CBmemPtr, lagcount;
int16_t *enPtr=&energyW16[base_size-20];
int16_t *enShPtr=&energyShifts[base_size-20];
int32_t nrjRecursive;
CBmemPtr = CBmem+147;
interpSamplesPtr = interpSamples;
/* Compute the energy for the first (low-5) noninterpolated samples */
nrjRecursive = WebRtcSpl_DotProductWithScale( CBmemPtr-19, CBmemPtr-19, 15, scale);
ppe = CBmemPtr - 20;
for (lagcount=20; lagcount<=39; lagcount++) {
/* Update the energy recursively to save complexity */
nrjRecursive = nrjRecursive +
WEBRTC_SPL_MUL_16_16_RSFT(*ppe, *ppe, scale);
ppe--;
energy = nrjRecursive;
/* interpolation */
energy += WebRtcSpl_DotProductWithScale(interpSamplesPtr, interpSamplesPtr, 4, scale);
interpSamplesPtr += 4;
/* Compute energy for the remaining samples */
pp = CBmemPtr - lagcount;
energy += WebRtcSpl_DotProductWithScale(pp, pp, SUBL-lagcount, scale);
/* Normalize the energy and store the number of shifts */
(*enShPtr) = (int16_t)WebRtcSpl_NormW32(energy);
tmp32 = WEBRTC_SPL_LSHIFT_W32(energy, (*enShPtr));
(*enPtr) = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32, 16);
enShPtr++;
enPtr++;
}
}
/* 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 */
int16_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 */
int16_t scale, /* (i) The scaling of all energy values */
int16_t base_size /* (i) Index to where the energy values should be stored */
)
{
int16_t j,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<range-1;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 = WEBRTC_SPL_LSHIFT_W32(energy, shft);
*eW16_ptr++ = (int16_t)(tmp >> 16);
}
}
void WebRtcIsacfix_DecimateAllpass32(const WebRtc_Word16 *in,
WebRtc_Word32 *state_in, /* array of size: 2*ALLPASSSECTIONS+1 */
WebRtc_Word16 N, /* number of input samples */
WebRtc_Word16 *out) /* array of size N/2 */
{
int n;
WebRtc_Word16 data_vec[PITCH_FRAME_LEN];
/* copy input */
memcpy(data_vec+1, in, WEBRTC_SPL_MUL_16_16(sizeof(WebRtc_Word16), (N-1)));
data_vec[0] = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(state_in[WEBRTC_SPL_MUL_16_16(2, ALLPASSSECTIONS)],16); //the z^(-1) state
state_in[WEBRTC_SPL_MUL_16_16(2, ALLPASSSECTIONS)] = WEBRTC_SPL_LSHIFT_W32((WebRtc_UWord32)in[N-1],16);
AllpassFilterForDec32(data_vec+1, kApUpperQ15, N, state_in);
AllpassFilterForDec32(data_vec, kApLowerQ15, N, state_in+ALLPASSSECTIONS);
for (n=0;n<N/2;n++) {
out[n]=WEBRTC_SPL_ADD_SAT_W16(data_vec[WEBRTC_SPL_MUL_16_16(2, n)], data_vec[WEBRTC_SPL_MUL_16_16(2, n)+1]);
}
}
void WebRtcIsacfix_PitchFilter(int16_t* indatQQ, // Q10 if type is 1 or 4,
// Q0 if type is 2.
int16_t* outdatQQ,
PitchFiltstr* pfp,
int16_t* lagsQ7,
int16_t* gainsQ12,
int16_t type) {
int k, ind, cnt;
int16_t sign = 1;
int16_t inystateQQ[PITCH_DAMPORDER];
int16_t ubufQQ[PITCH_INTBUFFSIZE + QLOOKAHEAD];
const int16_t Gain = 21299; // 1.3 in Q14
int16_t oldLagQ7;
int16_t oldGainQ12, lagdeltaQ7, curLagQ7, gaindeltaQ12, curGainQ12;
int indW32 = 0, frcQQ = 0;
int32_t tmpW32;
const int16_t* fracoeffQQ = NULL;
// Assumptions in ARM assembly for WebRtcIsacfix_PitchFilterCoreARM().
COMPILE_ASSERT(PITCH_FRACORDER == 9);
COMPILE_ASSERT(PITCH_DAMPORDER == 5);
// Set up buffer and states.
memcpy(ubufQQ, pfp->ubufQQ, sizeof(pfp->ubufQQ));
memcpy(inystateQQ, pfp->ystateQQ, sizeof(inystateQQ));
// Get old lag and gain value from memory.
oldLagQ7 = pfp->oldlagQ7;
oldGainQ12 = pfp->oldgainQ12;
if (type == 4) {
sign = -1;
// Make output more periodic.
for (k = 0; k < PITCH_SUBFRAMES; k++) {
gainsQ12[k] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(
gainsQ12[k], Gain, 14);
}
}
// No interpolation if pitch lag step is big.
if ((WEBRTC_SPL_MUL_16_16_RSFT(lagsQ7[0], 3, 1) < oldLagQ7) ||
(lagsQ7[0] > WEBRTC_SPL_MUL_16_16_RSFT(oldLagQ7, 3, 1))) {
oldLagQ7 = lagsQ7[0];
oldGainQ12 = gainsQ12[0];
}
ind = 0;
for (k = 0; k < PITCH_SUBFRAMES; k++) {
// Calculate interpolation steps.
lagdeltaQ7 = lagsQ7[k] - oldLagQ7;
lagdeltaQ7 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(
lagdeltaQ7, kDivFactor, 15);
curLagQ7 = oldLagQ7;
gaindeltaQ12 = gainsQ12[k] - oldGainQ12;
gaindeltaQ12 = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(
gaindeltaQ12, kDivFactor, 15);
curGainQ12 = oldGainQ12;
oldLagQ7 = lagsQ7[k];
oldGainQ12 = gainsQ12[k];
// Each frame has 4 60-sample pitch subframes, and each subframe has 5
// 12-sample segments. Each segment need to be processed with
// newly-updated parameters, so we break the pitch filtering into
// two for-loops (5 x 12) below. It's also why kDivFactor = 0.2 (in Q15).
for (cnt = 0; cnt < kSegments; cnt++) {
// Update parameters for each segment.
curGainQ12 += gaindeltaQ12;
curLagQ7 += lagdeltaQ7;
indW32 = CalcLrIntQ(curLagQ7, 7);
tmpW32 = WEBRTC_SPL_LSHIFT_W32(indW32, 7);
tmpW32 -= curLagQ7;
frcQQ = WEBRTC_SPL_RSHIFT_W32(tmpW32, 4);
frcQQ += 4;
if (frcQQ == PITCH_FRACS) {
frcQQ = 0;
}
fracoeffQQ = kIntrpCoef[frcQQ];
// Pitch filtering.
WebRtcIsacfix_PitchFilterCore(PITCH_SUBFRAME_LEN / kSegments, curGainQ12,
indW32, sign, inystateQQ, ubufQQ, fracoeffQQ, indatQQ, outdatQQ, &ind);
}
}
// Export buffer and states.
memcpy(pfp->ubufQQ, ubufQQ + PITCH_FRAME_LEN, sizeof(pfp->ubufQQ));
memcpy(pfp->ystateQQ, inystateQQ, sizeof(pfp->ystateQQ));
pfp->oldlagQ7 = oldLagQ7;
pfp->oldgainQ12 = oldGainQ12;
if (type == 2) {
// Filter look-ahead segment.
WebRtcIsacfix_PitchFilterCore(QLOOKAHEAD, curGainQ12, indW32, 1, inystateQQ,
ubufQQ, fracoeffQQ, indatQQ, outdatQQ, &ind);
}
}
void WebRtcIsacfix_Time2SpecC(int16_t *inre1Q9,
int16_t *inre2Q9,
int16_t *outreQ7,
int16_t *outimQ7)
{
int k;
int32_t tmpreQ16[FRAMESAMPLES/2], tmpimQ16[FRAMESAMPLES/2];
int16_t tmp1rQ14, tmp1iQ14;
int32_t xrQ16, xiQ16, yrQ16, yiQ16;
int32_t v1Q16, v2Q16;
int16_t factQ19, sh;
/* Multiply with complex exponentials and combine into one complex vector */
factQ19 = 16921; // 0.5/sqrt(240) in Q19 is round(.5/sqrt(240)*(2^19)) = 16921
for (k = 0; k < FRAMESAMPLES/2; k++) {
tmp1rQ14 = WebRtcIsacfix_kCosTab1[k];
tmp1iQ14 = WebRtcIsacfix_kSinTab1[k];
xrQ16 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(tmp1rQ14, inre1Q9[k]) + WEBRTC_SPL_MUL_16_16(tmp1iQ14, inre2Q9[k]), 7);
xiQ16 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(tmp1rQ14, inre2Q9[k]) - WEBRTC_SPL_MUL_16_16(tmp1iQ14, inre1Q9[k]), 7);
tmpreQ16[k] = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_32_RSFT16(factQ19, xrQ16)+4, 3); // (Q16*Q19>>16)>>3 = Q16
tmpimQ16[k] = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_32_RSFT16(factQ19, xiQ16)+4, 3); // (Q16*Q19>>16)>>3 = Q16
}
xrQ16 = WebRtcSpl_MaxAbsValueW32(tmpreQ16, FRAMESAMPLES/2);
yrQ16 = WebRtcSpl_MaxAbsValueW32(tmpimQ16, FRAMESAMPLES/2);
if (yrQ16>xrQ16) {
xrQ16 = yrQ16;
}
sh = WebRtcSpl_NormW32(xrQ16);
sh = sh-24; //if sh becomes >=0, then we should shift sh steps to the left, and the domain will become Q(16+sh)
//if sh becomes <0, then we should shift -sh steps to the right, and the domain will become Q(16+sh)
//"Fastest" vectors
if (sh>=0) {
for (k=0; k<FRAMESAMPLES/2; k++) {
inre1Q9[k] = (int16_t) WEBRTC_SPL_LSHIFT_W32(tmpreQ16[k], sh); //Q(16+sh)
inre2Q9[k] = (int16_t) WEBRTC_SPL_LSHIFT_W32(tmpimQ16[k], sh); //Q(16+sh)
}
} else {
int32_t round = WEBRTC_SPL_LSHIFT_W32((int32_t)1, -sh-1);
for (k=0; k<FRAMESAMPLES/2; k++) {
inre1Q9[k] = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmpreQ16[k]+round, -sh); //Q(16+sh)
inre2Q9[k] = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmpimQ16[k]+round, -sh); //Q(16+sh)
}
}
/* Get DFT */
WebRtcIsacfix_FftRadix16Fastest(inre1Q9, inre2Q9, -1); // real call
//"Fastest" vectors
if (sh>=0) {
for (k=0; k<FRAMESAMPLES/2; k++) {
tmpreQ16[k] = WEBRTC_SPL_RSHIFT_W32((int32_t)inre1Q9[k], sh); //Q(16+sh) -> Q16
tmpimQ16[k] = WEBRTC_SPL_RSHIFT_W32((int32_t)inre2Q9[k], sh); //Q(16+sh) -> Q16
}
} else {
for (k=0; k<FRAMESAMPLES/2; k++) {
tmpreQ16[k] = WEBRTC_SPL_LSHIFT_W32((int32_t)inre1Q9[k], -sh); //Q(16+sh) -> Q16
tmpimQ16[k] = WEBRTC_SPL_LSHIFT_W32((int32_t)inre2Q9[k], -sh); //Q(16+sh) -> Q16
}
}
/* Use symmetry to separate into two complex vectors and center frames in time around zero */
for (k = 0; k < FRAMESAMPLES/4; k++) {
xrQ16 = tmpreQ16[k] + tmpreQ16[FRAMESAMPLES/2 - 1 - k];
yiQ16 = -tmpreQ16[k] + tmpreQ16[FRAMESAMPLES/2 - 1 - k];
xiQ16 = tmpimQ16[k] - tmpimQ16[FRAMESAMPLES/2 - 1 - k];
yrQ16 = tmpimQ16[k] + tmpimQ16[FRAMESAMPLES/2 - 1 - k];
tmp1rQ14 = -WebRtcIsacfix_kSinTab2[FRAMESAMPLES/4 - 1 - k];
tmp1iQ14 = WebRtcIsacfix_kSinTab2[k];
v1Q16 = WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, xrQ16) - WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, xiQ16);
v2Q16 = WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, xrQ16) + WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, xiQ16);
outreQ7[k] = (int16_t) WEBRTC_SPL_RSHIFT_W32(v1Q16, 9);
outimQ7[k] = (int16_t) WEBRTC_SPL_RSHIFT_W32(v2Q16, 9);
v1Q16 = -WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, yrQ16) - WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, yiQ16);
v2Q16 = -WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, yrQ16) + WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, yiQ16);
outreQ7[FRAMESAMPLES/2 - 1 - k] = (int16_t)WEBRTC_SPL_RSHIFT_W32(v1Q16, 9); //CalcLrIntQ(v1Q16, 9);
outimQ7[FRAMESAMPLES/2 - 1 - k] = (int16_t)WEBRTC_SPL_RSHIFT_W32(v2Q16, 9); //CalcLrIntQ(v2Q16, 9);
}
}
void WebRtcIsacfix_Spec2TimeC(int16_t *inreQ7, int16_t *inimQ7, int32_t *outre1Q16, int32_t *outre2Q16)
{
int k;
int16_t tmp1rQ14, tmp1iQ14;
int32_t xrQ16, xiQ16, yrQ16, yiQ16;
int32_t tmpInRe, tmpInIm, tmpInRe2, tmpInIm2;
int16_t factQ11;
int16_t sh;
for (k = 0; k < FRAMESAMPLES/4; k++) {
/* Move zero in time to beginning of frames */
tmp1rQ14 = -WebRtcIsacfix_kSinTab2[FRAMESAMPLES/4 - 1 - k];
tmp1iQ14 = WebRtcIsacfix_kSinTab2[k];
tmpInRe = WEBRTC_SPL_LSHIFT_W32((int32_t) inreQ7[k], 9); // Q7 -> Q16
tmpInIm = WEBRTC_SPL_LSHIFT_W32((int32_t) inimQ7[k], 9); // Q7 -> Q16
tmpInRe2 = WEBRTC_SPL_LSHIFT_W32((int32_t) inreQ7[FRAMESAMPLES/2 - 1 - k], 9); // Q7 -> Q16
tmpInIm2 = WEBRTC_SPL_LSHIFT_W32((int32_t) inimQ7[FRAMESAMPLES/2 - 1 - k], 9); // Q7 -> Q16
xrQ16 = WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, tmpInRe) + WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, tmpInIm);
xiQ16 = WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, tmpInIm) - WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, tmpInRe);
yrQ16 = -WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, tmpInIm2) - WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, tmpInRe2);
yiQ16 = -WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, tmpInRe2) + WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, tmpInIm2);
/* Combine into one vector, z = x + j * y */
outre1Q16[k] = xrQ16 - yiQ16;
outre1Q16[FRAMESAMPLES/2 - 1 - k] = xrQ16 + yiQ16;
outre2Q16[k] = xiQ16 + yrQ16;
outre2Q16[FRAMESAMPLES/2 - 1 - k] = -xiQ16 + yrQ16;
}
/* Get IDFT */
tmpInRe = WebRtcSpl_MaxAbsValueW32(outre1Q16, 240);
tmpInIm = WebRtcSpl_MaxAbsValueW32(outre2Q16, 240);
if (tmpInIm>tmpInRe) {
tmpInRe = tmpInIm;
}
sh = WebRtcSpl_NormW32(tmpInRe);
sh = sh-24; //if sh becomes >=0, then we should shift sh steps to the left, and the domain will become Q(16+sh)
//if sh becomes <0, then we should shift -sh steps to the right, and the domain will become Q(16+sh)
//"Fastest" vectors
if (sh>=0) {
for (k=0; k<240; k++) {
inreQ7[k] = (int16_t) WEBRTC_SPL_LSHIFT_W32(outre1Q16[k], sh); //Q(16+sh)
inimQ7[k] = (int16_t) WEBRTC_SPL_LSHIFT_W32(outre2Q16[k], sh); //Q(16+sh)
}
} else {
int32_t round = WEBRTC_SPL_LSHIFT_W32((int32_t)1, -sh-1);
for (k=0; k<240; k++) {
inreQ7[k] = (int16_t) WEBRTC_SPL_RSHIFT_W32(outre1Q16[k]+round, -sh); //Q(16+sh)
inimQ7[k] = (int16_t) WEBRTC_SPL_RSHIFT_W32(outre2Q16[k]+round, -sh); //Q(16+sh)
}
}
WebRtcIsacfix_FftRadix16Fastest(inreQ7, inimQ7, 1); // real call
//"Fastest" vectors
if (sh>=0) {
for (k=0; k<240; k++) {
outre1Q16[k] = WEBRTC_SPL_RSHIFT_W32((int32_t)inreQ7[k], sh); //Q(16+sh) -> Q16
outre2Q16[k] = WEBRTC_SPL_RSHIFT_W32((int32_t)inimQ7[k], sh); //Q(16+sh) -> Q16
}
} else {
for (k=0; k<240; k++) {
outre1Q16[k] = WEBRTC_SPL_LSHIFT_W32((int32_t)inreQ7[k], -sh); //Q(16+sh) -> Q16
outre2Q16[k] = WEBRTC_SPL_LSHIFT_W32((int32_t)inimQ7[k], -sh); //Q(16+sh) -> Q16
}
}
/* Divide through by the normalizing constant: */
/* scale all values with 1/240, i.e. with 273 in Q16 */
/* 273/65536 ~= 0.0041656 */
/* 1/240 ~= 0.0041666 */
for (k=0; k<240; k++) {
outre1Q16[k] = WEBRTC_SPL_MUL_16_32_RSFT16(273, outre1Q16[k]);
outre2Q16[k] = WEBRTC_SPL_MUL_16_32_RSFT16(273, outre2Q16[k]);
}
/* Demodulate and separate */
factQ11 = 31727; // sqrt(240) in Q11 is round(15.49193338482967*2048) = 31727
for (k = 0; k < FRAMESAMPLES/2; k++) {
tmp1rQ14 = WebRtcIsacfix_kCosTab1[k];
tmp1iQ14 = WebRtcIsacfix_kSinTab1[k];
xrQ16 = WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, outre1Q16[k]) - WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, outre2Q16[k]);
xiQ16 = WEBRTC_SPL_MUL_16_32_RSFT14(tmp1rQ14, outre2Q16[k]) + WEBRTC_SPL_MUL_16_32_RSFT14(tmp1iQ14, outre1Q16[k]);
xrQ16 = WEBRTC_SPL_MUL_16_32_RSFT11(factQ11, xrQ16);
xiQ16 = WEBRTC_SPL_MUL_16_32_RSFT11(factQ11, xiQ16);
outre2Q16[k] = xiQ16;
outre1Q16[k] = xrQ16;
}
}
static void PCorr2Q32(const int16_t *in, int32_t *logcorQ8)
{
int16_t scaling,n,k;
int32_t ysum32,csum32, lys, lcs;
int32_t oneQ8;
const int16_t *x, *inptr;
oneQ8 = WEBRTC_SPL_LSHIFT_W32((int32_t)1, 8); // 1.00 in Q8
x = in + PITCH_MAX_LAG/2 + 2;
scaling = WebRtcSpl_GetScalingSquare ((int16_t *) in, PITCH_CORR_LEN2, PITCH_CORR_LEN2);
ysum32 = 1;
csum32 = 0;
x = in + PITCH_MAX_LAG/2 + 2;
for (n = 0; n < PITCH_CORR_LEN2; n++) {
ysum32 += WEBRTC_SPL_MUL_16_16_RSFT( (int16_t) in[n],(int16_t) in[n], scaling); // Q0
csum32 += WEBRTC_SPL_MUL_16_16_RSFT((int16_t) x[n],(int16_t) in[n], scaling); // Q0
}
logcorQ8 += PITCH_LAG_SPAN2 - 1;
lys=Log2Q8((uint32_t) ysum32); // Q8
lys=WEBRTC_SPL_RSHIFT_W32(lys, 1); //sqrt(ysum);
if (csum32>0) {
lcs=Log2Q8((uint32_t) csum32); // 2log(csum) in Q8
if (lcs>(lys + oneQ8) ){ // csum/sqrt(ysum) > 2 in Q8
*logcorQ8 = lcs - lys; // log2(csum/sqrt(ysum))
} else {
*logcorQ8 = oneQ8; // 1.00
}
} else {
*logcorQ8 = 0;
}
for (k = 1; k < PITCH_LAG_SPAN2; k++) {
inptr = &in[k];
ysum32 -= WEBRTC_SPL_MUL_16_16_RSFT( (int16_t) in[k-1],(int16_t) in[k-1], scaling);
ysum32 += WEBRTC_SPL_MUL_16_16_RSFT( (int16_t) in[PITCH_CORR_LEN2 + k - 1],(int16_t) in[PITCH_CORR_LEN2 + k - 1], scaling);
#ifdef WEBRTC_ARCH_ARM_NEON
{
int32_t vbuff[4];
int32x4_t int_32x4_sum = vmovq_n_s32(0);
// Can't shift a Neon register to right with a non-constant shift value.
int32x4_t int_32x4_scale = vdupq_n_s32(-scaling);
// Assert a codition used in loop unrolling at compile-time.
COMPILE_ASSERT(PITCH_CORR_LEN2 %4 == 0);
for (n = 0; n < PITCH_CORR_LEN2; n += 4) {
int16x4_t int_16x4_x = vld1_s16(&x[n]);
int16x4_t int_16x4_in = vld1_s16(&inptr[n]);
int32x4_t int_32x4 = vmull_s16(int_16x4_x, int_16x4_in);
int_32x4 = vshlq_s32(int_32x4, int_32x4_scale);
int_32x4_sum = vaddq_s32(int_32x4_sum, int_32x4);
}
// Use vector store to avoid long stall from data trasferring
// from vector to general register.
vst1q_s32(vbuff, int_32x4_sum);
csum32 = vbuff[0] + vbuff[1];
csum32 += vbuff[2];
csum32 += vbuff[3];
}
#else
csum32 = 0;
if(scaling == 0) {
for (n = 0; n < PITCH_CORR_LEN2; n++) {
csum32 += x[n] * inptr[n];
}
} else {
for (n = 0; n < PITCH_CORR_LEN2; n++) {
csum32 += (x[n] * inptr[n]) >> scaling;
}
}
#endif
logcorQ8--;
lys=Log2Q8((uint32_t)ysum32); // Q8
lys=WEBRTC_SPL_RSHIFT_W32(lys, 1); //sqrt(ysum);
if (csum32>0) {
lcs=Log2Q8((uint32_t) csum32); // 2log(csum) in Q8
if (lcs>(lys + oneQ8) ){ // csum/sqrt(ysum) > 2
*logcorQ8 = lcs - lys; // log2(csum/sqrt(ysum))
} else {
*logcorQ8 = oneQ8; // 1.00
}
} else {
*logcorQ8 = 0;
}
}
}
int WebRtcNetEQ_RTCPGetStats(WebRtcNetEQ_RTCP_t *RTCP_inst,
uint16_t *puw16_fraction_lost,
uint32_t *puw32_cum_lost, uint32_t *puw32_ext_max,
uint32_t *puw32_jitter, int16_t doNotReset)
{
uint32_t uw32_exp_nr, uw32_exp_interval, uw32_rec_interval;
int32_t w32_lost;
/* Extended highest sequence number received */
*puw32_ext_max
= (uint32_t) WEBRTC_SPL_LSHIFT_W32((uint32_t)RTCP_inst->cycles, 16)
+ RTCP_inst->max_seq;
/*
* Calculate expected number of packets and compare it to the number of packets that
* were actually received => the cumulative number of packets lost can be extracted.
*/
uw32_exp_nr = *puw32_ext_max - RTCP_inst->base_seq + 1;
if (RTCP_inst->received == 0)
{
/* no packets received, assume none lost */
*puw32_cum_lost = 0;
}
else if (uw32_exp_nr > RTCP_inst->received)
{
*puw32_cum_lost = uw32_exp_nr - RTCP_inst->received;
if (*puw32_cum_lost > (uint32_t) 0xFFFFFF)
{
*puw32_cum_lost = 0xFFFFFF;
}
}
else
{
*puw32_cum_lost = 0;
}
/* Fraction lost (Since last report) */
uw32_exp_interval = uw32_exp_nr - RTCP_inst->exp_prior;
if (!doNotReset)
{
RTCP_inst->exp_prior = uw32_exp_nr;
}
uw32_rec_interval = RTCP_inst->received - RTCP_inst->rec_prior;
if (!doNotReset)
{
RTCP_inst->rec_prior = RTCP_inst->received;
}
w32_lost = (int32_t) (uw32_exp_interval - uw32_rec_interval);
if (uw32_exp_interval == 0 || w32_lost <= 0 || RTCP_inst->received == 0)
{
*puw16_fraction_lost = 0;
}
else
{
*puw16_fraction_lost = (uint16_t) (WEBRTC_SPL_LSHIFT_W32(w32_lost, 8)
/ uw32_exp_interval);
}
if (*puw16_fraction_lost > 0xFF)
{
*puw16_fraction_lost = 0xFF;
}
/* Inter-arrival jitter */
*puw32_jitter = (RTCP_inst->jitter) >> 4; /* scaling from Q4 */
return 0;
}
void WebRtcIsacfix_InitialPitch(const int16_t *in, /* Q0 */
PitchAnalysisStruct *State,
int16_t *lagsQ7 /* Q7 */
)
{
int16_t buf_dec16[PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2+2];
int32_t *crrvecQ8_1,*crrvecQ8_2;
int32_t cv1q[PITCH_LAG_SPAN2+2],cv2q[PITCH_LAG_SPAN2+2], peakvq[PITCH_LAG_SPAN2+2];
int k;
int16_t peaks_indq;
int16_t peakiq[PITCH_LAG_SPAN2];
int32_t corr;
int32_t corr32, corr_max32, corr_max_o32;
int16_t npkq;
int16_t best4q[4]={0,0,0,0};
int32_t xq[3],yq[1],fyq[1];
int32_t *fxq;
int32_t best_lag1q, best_lag2q;
int32_t tmp32a,tmp32b,lag32,ratq;
int16_t start;
int16_t oldgQ12, tmp16a, tmp16b, gain_bias16,tmp16c, tmp16d, bias16;
int32_t tmp32c,tmp32d, tmp32e;
int16_t old_lagQ;
int32_t old_lagQ8;
int32_t lagsQ8[4];
old_lagQ = State->PFstr_wght.oldlagQ7; // Q7
old_lagQ8= WEBRTC_SPL_LSHIFT_W32((int32_t)old_lagQ,1); //Q8
oldgQ12= State->PFstr_wght.oldgainQ12;
crrvecQ8_1=&cv1q[1];
crrvecQ8_2=&cv2q[1];
/* copy old values from state buffer */
memcpy(buf_dec16, State->dec_buffer16, WEBRTC_SPL_MUL_16_16(sizeof(int16_t), (PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2-PITCH_FRAME_LEN/2+2)));
/* decimation; put result after the old values */
WebRtcIsacfix_DecimateAllpass32(in, State->decimator_state32, PITCH_FRAME_LEN,
&buf_dec16[PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2-PITCH_FRAME_LEN/2+2]);
/* low-pass filtering */
start= PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2-PITCH_FRAME_LEN/2+2;
WebRtcSpl_FilterARFastQ12(&buf_dec16[start],&buf_dec16[start],(int16_t*)kACoefQ12,3, PITCH_FRAME_LEN/2);
/* copy end part back into state buffer */
for (k = 0; k < (PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2-PITCH_FRAME_LEN/2+2); k++)
State->dec_buffer16[k] = buf_dec16[k+PITCH_FRAME_LEN/2];
/* compute correlation for first and second half of the frame */
PCorr2Q32(buf_dec16, crrvecQ8_1);
PCorr2Q32(buf_dec16 + PITCH_CORR_STEP2, crrvecQ8_2);
/* bias towards pitch lag of previous frame */
tmp32a = Log2Q8((uint32_t) old_lagQ8) - 2304; // log2(0.5*oldlag) in Q8
tmp32b = WEBRTC_SPL_MUL_16_16_RSFT(oldgQ12,oldgQ12, 10); //Q12 & * 4.0;
gain_bias16 = (int16_t) tmp32b; //Q12
if (gain_bias16 > 3276) gain_bias16 = 3276; // 0.8 in Q12
for (k = 0; k < PITCH_LAG_SPAN2; k++)
{
if (crrvecQ8_1[k]>0) {
tmp32b = Log2Q8((uint32_t) (k + (PITCH_MIN_LAG/2-2)));
tmp16a = (int16_t) (tmp32b - tmp32a); // Q8 & fabs(ratio)<4
tmp32c = WEBRTC_SPL_MUL_16_16_RSFT(tmp16a,tmp16a, 6); //Q10
tmp16b = (int16_t) tmp32c; // Q10 & <8
tmp32d = WEBRTC_SPL_MUL_16_16_RSFT(tmp16b, 177 , 8); // mult with ln2 in Q8
tmp16c = (int16_t) tmp32d; // Q10 & <4
tmp16d = Exp2Q10((int16_t) -tmp16c); //Q10
tmp32c = WEBRTC_SPL_MUL_16_16_RSFT(gain_bias16,tmp16d,13); // Q10 & * 0.5
bias16 = (int16_t) (1024 + tmp32c); // Q10
tmp32b = Log2Q8((uint32_t) bias16) - 2560; // Q10 in -> Q8 out with 10*2^8 offset
crrvecQ8_1[k] += tmp32b ; // -10*2^8 offset
}
}
/* taper correlation functions */
for (k = 0; k < 3; k++) {
crrvecQ8_1[k] += kLogLagWinQ8[k];
crrvecQ8_2[k] += kLogLagWinQ8[k];
crrvecQ8_1[PITCH_LAG_SPAN2-1-k] += kLogLagWinQ8[k];
crrvecQ8_2[PITCH_LAG_SPAN2-1-k] += kLogLagWinQ8[k];
}
/* Make zeropadded corr vectors */
cv1q[0]=0;
cv2q[0]=0;
cv1q[PITCH_LAG_SPAN2+1]=0;
cv2q[PITCH_LAG_SPAN2+1]=0;
corr_max32 = 0;
for (k = 1; k <= PITCH_LAG_SPAN2; k++)
{
corr32=crrvecQ8_1[k-1];
if (corr32 > corr_max32)
corr_max32 = corr32;
corr32=crrvecQ8_2[k-1];
corr32 += -4; // Compensate for later (log2(0.99))
if (corr32 > corr_max32)
corr_max32 = corr32;
}
/* threshold value to qualify as a peak */
// corr_max32 += -726; // log(0.14)/log(2.0) in Q8
corr_max32 += -1000; // log(0.14)/log(2.0) in Q8
corr_max_o32 = corr_max32;
/* find peaks in corr1 */
peaks_indq = 0;
for (k = 1; k <= PITCH_LAG_SPAN2; k++)
{
corr32=cv1q[k];
if (corr32>corr_max32) { // Disregard small peaks
if ((corr32>=cv1q[k-1]) && (corr32>cv1q[k+1])) { // Peak?
peakvq[peaks_indq] = corr32;
peakiq[peaks_indq++] = k;
}
}
}
/* find highest interpolated peak */
corr_max32=0;
best_lag1q =0;
if (peaks_indq > 0) {
FindFour32(peakvq, (int16_t) peaks_indq, best4q);
npkq = WEBRTC_SPL_MIN(peaks_indq, 4);
for (k=0;k<npkq;k++) {
lag32 = peakiq[best4q[k]];
fxq = &cv1q[peakiq[best4q[k]]-1];
xq[0]= lag32;
xq[0] = WEBRTC_SPL_LSHIFT_W32(xq[0], 8);
Intrp1DQ8(xq, fxq, yq, fyq);
tmp32a= Log2Q8((uint32_t) *yq) - 2048; // offset 8*2^8
/* Bias towards short lags */
/* log(pow(0.8, log(2.0 * *y )))/log(2.0) */
tmp32b= WEBRTC_SPL_MUL_16_16_RSFT((int16_t) tmp32a, -42, 8);
tmp32c= tmp32b + 256;
*fyq += tmp32c;
if (*fyq > corr_max32) {
corr_max32 = *fyq;
best_lag1q = *yq;
}
}
tmp32a = best_lag1q - OFFSET_Q8;
tmp32b = WEBRTC_SPL_LSHIFT_W32(tmp32a, 1);
lagsQ8[0] = tmp32b + PITCH_MIN_LAG_Q8;
lagsQ8[1] = lagsQ8[0];
} else {
lagsQ8[0] = old_lagQ8;
lagsQ8[1] = lagsQ8[0];
}
/* Bias towards constant pitch */
tmp32a = lagsQ8[0] - PITCH_MIN_LAG_Q8;
ratq = WEBRTC_SPL_RSHIFT_W32(tmp32a, 1) + OFFSET_Q8;
for (k = 1; k <= PITCH_LAG_SPAN2; k++)
{
tmp32a = WEBRTC_SPL_LSHIFT_W32(k, 7); // 0.5*k Q8
tmp32b = (int32_t) (WEBRTC_SPL_LSHIFT_W32(tmp32a, 1)) - ratq; // Q8
tmp32c = WEBRTC_SPL_MUL_16_16_RSFT((int16_t) tmp32b, (int16_t) tmp32b, 8); // Q8
tmp32b = (int32_t) tmp32c + (int32_t) WEBRTC_SPL_RSHIFT_W32(ratq, 1); // (k-r)^2 + 0.5 * r Q8
tmp32c = Log2Q8((uint32_t) tmp32a) - 2048; // offset 8*2^8 , log2(0.5*k) Q8
tmp32d = Log2Q8((uint32_t) tmp32b) - 2048; // offset 8*2^8 , log2(0.5*k) Q8
tmp32e = tmp32c -tmp32d;
cv2q[k] += WEBRTC_SPL_RSHIFT_W32(tmp32e, 1);
}
/* find peaks in corr2 */
corr_max32 = corr_max_o32;
peaks_indq = 0;
for (k = 1; k <= PITCH_LAG_SPAN2; k++)
{
corr=cv2q[k];
if (corr>corr_max32) { // Disregard small peaks
if ((corr>=cv2q[k-1]) && (corr>cv2q[k+1])) { // Peak?
peakvq[peaks_indq] = corr;
peakiq[peaks_indq++] = k;
}
}
}
/* find highest interpolated peak */
corr_max32 = 0;
best_lag2q =0;
if (peaks_indq > 0) {
FindFour32(peakvq, (int16_t) peaks_indq, best4q);
npkq = WEBRTC_SPL_MIN(peaks_indq, 4);
for (k=0;k<npkq;k++) {
lag32 = peakiq[best4q[k]];
fxq = &cv2q[peakiq[best4q[k]]-1];
xq[0]= lag32;
xq[0] = WEBRTC_SPL_LSHIFT_W32(xq[0], 8);
Intrp1DQ8(xq, fxq, yq, fyq);
/* Bias towards short lags */
/* log(pow(0.8, log(2.0f * *y )))/log(2.0f) */
tmp32a= Log2Q8((uint32_t) *yq) - 2048; // offset 8*2^8
tmp32b= WEBRTC_SPL_MUL_16_16_RSFT((int16_t) tmp32a, -82, 8);
tmp32c= tmp32b + 256;
*fyq += tmp32c;
if (*fyq > corr_max32) {
corr_max32 = *fyq;
best_lag2q = *yq;
}
}
tmp32a = best_lag2q - OFFSET_Q8;
tmp32b = WEBRTC_SPL_LSHIFT_W32(tmp32a, 1);
lagsQ8[2] = tmp32b + PITCH_MIN_LAG_Q8;
lagsQ8[3] = lagsQ8[2];
} else {
lagsQ8[2] = lagsQ8[0];
lagsQ8[3] = lagsQ8[0];
}
lagsQ7[0]=(int16_t) WEBRTC_SPL_RSHIFT_W32(lagsQ8[0], 1);
lagsQ7[1]=(int16_t) WEBRTC_SPL_RSHIFT_W32(lagsQ8[1], 1);
lagsQ7[2]=(int16_t) WEBRTC_SPL_RSHIFT_W32(lagsQ8[2], 1);
lagsQ7[3]=(int16_t) WEBRTC_SPL_RSHIFT_W32(lagsQ8[3], 1);
}
// Update the noise estimation information.
static void UpdateNoiseEstimateNeon(NsxInst_t* inst, int offset) {
// int i = 0;
const int16_t kExp2Const = 11819; // Q13
int16_t* ptr_noiseEstLogQuantile = NULL;
int16_t* ptr_noiseEstQuantile = NULL;
int16x4_t kExp2Const16x4 = vdup_n_s16(kExp2Const);
int32x4_t twentyOne32x4 = vdupq_n_s32(21);
int32x4_t constA32x4 = vdupq_n_s32(0x1fffff);
int32x4_t constB32x4 = vdupq_n_s32(0x200000);
int16_t tmp16 = WebRtcSpl_MaxValueW16(inst->noiseEstLogQuantile + offset,
inst->magnLen);
// Guarantee a Q-domain as high as possible and still fit in int16
inst->qNoise = 14 - (int) WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(kExp2Const,
tmp16,
21);
int32x4_t qNoise32x4 = vdupq_n_s32(inst->qNoise);
for (ptr_noiseEstLogQuantile = &inst->noiseEstLogQuantile[offset],
ptr_noiseEstQuantile = &inst->noiseEstQuantile[0];
ptr_noiseEstQuantile < &inst->noiseEstQuantile[inst->magnLen - 3];
ptr_noiseEstQuantile += 4, ptr_noiseEstLogQuantile += 4) {
// tmp32no2 = WEBRTC_SPL_MUL_16_16(kExp2Const,
// inst->noiseEstLogQuantile[offset + i]);
int16x4_t v16x4 = vld1_s16(ptr_noiseEstLogQuantile);
int32x4_t v32x4B = vmull_s16(v16x4, kExp2Const16x4);
// tmp32no1 = (0x00200000 | (tmp32no2 & 0x001FFFFF)); // 2^21 + frac
int32x4_t v32x4A = vandq_s32(v32x4B, constA32x4);
v32x4A = vorrq_s32(v32x4A, constB32x4);
// tmp16 = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmp32no2, 21);
v32x4B = vshrq_n_s32(v32x4B, 21);
// tmp16 -= 21;// shift 21 to get result in Q0
v32x4B = vsubq_s32(v32x4B, twentyOne32x4);
// tmp16 += (int16_t) inst->qNoise;
// shift to get result in Q(qNoise)
v32x4B = vaddq_s32(v32x4B, qNoise32x4);
// if (tmp16 < 0) {
// tmp32no1 = WEBRTC_SPL_RSHIFT_W32(tmp32no1, -tmp16);
// } else {
// tmp32no1 = WEBRTC_SPL_LSHIFT_W32(tmp32no1, tmp16);
// }
v32x4B = vshlq_s32(v32x4A, v32x4B);
// tmp16 = WebRtcSpl_SatW32ToW16(tmp32no1);
v16x4 = vqmovn_s32(v32x4B);
//inst->noiseEstQuantile[i] = tmp16;
vst1_s16(ptr_noiseEstQuantile, v16x4);
}
// Last iteration:
// inst->quantile[i]=exp(inst->lquantile[offset+i]);
// in Q21
int32_t tmp32no2 = WEBRTC_SPL_MUL_16_16(kExp2Const,
*ptr_noiseEstLogQuantile);
int32_t tmp32no1 = (0x00200000 | (tmp32no2 & 0x001FFFFF)); // 2^21 + frac
tmp16 = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmp32no2, 21);
tmp16 -= 21;// shift 21 to get result in Q0
tmp16 += (int16_t) inst->qNoise; //shift to get result in Q(qNoise)
if (tmp16 < 0) {
tmp32no1 = WEBRTC_SPL_RSHIFT_W32(tmp32no1, -tmp16);
} else {
tmp32no1 = WEBRTC_SPL_LSHIFT_W32(tmp32no1, tmp16);
}
*ptr_noiseEstQuantile = WebRtcSpl_SatW32ToW16(tmp32no1);
}
/****************************************************************************
* WebRtcIsacfix_EncTerminate(...)
*
* Final call to the arithmetic coder for an encoder call. This function
* terminates and return byte stream.
*
* Input:
* - streamData : in-/output struct containing bitstream
*
* Return value : number of bytes in the stream
*/
WebRtc_Word16 WebRtcIsacfix_EncTerminate(Bitstr_enc *streamData)
{
WebRtc_UWord16 *streamPtr;
WebRtc_UWord16 negCarry;
/* point to the right place in the stream buffer */
streamPtr = streamData->stream + streamData->stream_index;
/* find minimum length (determined by current interval width) */
if ( streamData->W_upper > 0x01FFFFFF )
{
streamData->streamval += 0x01000000;
/* if result is less than the added value we must take care of the carry */
if (streamData->streamval < 0x01000000)
{
/* propagate carry */
if (streamData->full == 0) {
/* Add value to current value */
negCarry = *streamPtr;
negCarry += 0x0100;
*streamPtr = negCarry;
/* if value is too big, propagate carry to next byte, and so on */
while (!(negCarry))
{
negCarry = *--streamPtr;
negCarry++;
*streamPtr = negCarry;
}
} else {
/* propagate carry by adding one to the previous byte in the
* stream if that byte is 0xFFFF we need to propagate the carry
* furhter back in the stream */
while ( !(++(*--streamPtr)) );
}
/* put pointer back to the old value */
streamPtr = streamData->stream + streamData->stream_index;
}
/* write remaining data to bitstream, if "full == 0" first byte has data */
if (streamData->full == 0) {
*streamPtr++ += (WebRtc_UWord16) WEBRTC_SPL_RSHIFT_W32(streamData->streamval, 24);
streamData->full = 1;
} else {
*streamPtr = (WebRtc_UWord16) WEBRTC_SPL_LSHIFT_W32(
WEBRTC_SPL_RSHIFT_W32(streamData->streamval, 24), 8);
streamData->full = 0;
}
}
else
{
streamData->streamval += 0x00010000;
/* if result is less than the added value we must take care of the carry */
if (streamData->streamval < 0x00010000)
{
/* propagate carry */
if (streamData->full == 0) {
/* Add value to current value */
negCarry = *streamPtr;
negCarry += 0x0100;
*streamPtr = negCarry;
/* if value to big, propagate carry to next byte, and so on */
while (!(negCarry))
{
negCarry = *--streamPtr;
negCarry++;
*streamPtr = negCarry;
}
} else {
/* Add carry to previous byte */
while ( !(++(*--streamPtr)) );
}
/* put pointer back to the old value */
streamPtr = streamData->stream + streamData->stream_index;
}
/* write remaining data (2 bytes) to bitstream */
if (streamData->full) {
*streamPtr++ = (WebRtc_UWord16) WEBRTC_SPL_RSHIFT_W32(streamData->streamval, 16);
} else {
*streamPtr++ |= (WebRtc_UWord16) WEBRTC_SPL_RSHIFT_W32(streamData->streamval, 24);
*streamPtr = (WebRtc_UWord16) WEBRTC_SPL_RSHIFT_W32(streamData->streamval, 8)
& 0xFF00;
}
}
/* calculate stream length in bytes */
return (((streamPtr - streamData->stream)<<1) + !(streamData->full));
}
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;
}
}
/* filter the signal using normalized lattice filter */
void WebRtcIsacfix_NormLatticeFilterAr(int16_t orderCoef,
int16_t *stateGQ0,
int32_t *lat_inQ25,
int16_t *filt_coefQ15,
int32_t *gain_lo_hiQ17,
int16_t lo_hi,
int16_t *lat_outQ0)
{
int ii,n,k,i,u;
int16_t sthQ15[MAX_AR_MODEL_ORDER];
int16_t cthQ15[MAX_AR_MODEL_ORDER];
int32_t tmp32;
int16_t tmpAR;
int16_t ARfQ0vec[HALF_SUBFRAMELEN];
int16_t ARgQ0vec[MAX_AR_MODEL_ORDER+1];
int32_t inv_gain32;
int16_t inv_gain16;
int16_t den16;
int16_t sh;
int16_t temp2,temp3;
int16_t ord_1 = orderCoef+1;
for (u=0;u<SUBFRAMES;u++)
{
int32_t temp1 = WEBRTC_SPL_MUL_16_16(u, HALF_SUBFRAMELEN);
//set the denominator and numerator of the Direct Form
temp2 = (int16_t)WEBRTC_SPL_MUL_16_16(u, orderCoef);
temp3 = (int16_t)WEBRTC_SPL_MUL_16_16(2, u) + lo_hi;
for (ii=0; ii<orderCoef; ii++) {
sthQ15[ii] = filt_coefQ15[temp2+ii];
}
WebRtcSpl_SqrtOfOneMinusXSquared(sthQ15, orderCoef, cthQ15);
/* Simulation of the 25 files shows that maximum value in
the vector gain_lo_hiQ17[] is 441344, which means that
it is log2((2^31)/441344) = 12.2 shifting bits from
saturation. Therefore, it should be safe to use Q27 instead
of Q17. */
tmp32 = WEBRTC_SPL_LSHIFT_W32(gain_lo_hiQ17[temp3], 10); // Q27
for (k=0;k<orderCoef;k++) {
tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[k], tmp32); // Q15*Q27>>15 = Q27
}
sh = WebRtcSpl_NormW32(tmp32); // tmp32 is the gain
den16 = (int16_t) WEBRTC_SPL_SHIFT_W32(tmp32, sh-16); //Q(27+sh-16) = Q(sh+11) (all 16 bits are value bits)
inv_gain32 = WebRtcSpl_DivW32W16((int32_t)2147483647, den16); // 1/gain in Q31/Q(sh+11) = Q(20-sh)
//initial conditions
inv_gain16 = (int16_t)(inv_gain32 >> 2); // 1/gain in Q(20-sh-2) = Q(18-sh)
for (i=0;i<HALF_SUBFRAMELEN;i++)
{
tmp32 = WEBRTC_SPL_LSHIFT_W32(lat_inQ25[i + temp1], 1); //Q25->Q26
tmp32 = WEBRTC_SPL_MUL_16_32_RSFT16(inv_gain16, tmp32); //lat_in[]*inv_gain in (Q(18-sh)*Q26)>>16 = Q(28-sh)
tmp32 = WEBRTC_SPL_SHIFT_W32(tmp32, -(28-sh)); // lat_in[]*inv_gain in Q0
ARfQ0vec[i] = (int16_t)WebRtcSpl_SatW32ToW16(tmp32); // Q0
}
for (i=orderCoef-1;i>=0;i--) //get the state of f&g for the first input, for all orders
{
tmp32 = (cthQ15[i] * ARfQ0vec[0] - sthQ15[i] * stateGQ0[i] + 16384) >> 15;
tmpAR = (int16_t)WebRtcSpl_SatW32ToW16(tmp32); // Q0
tmp32 = (sthQ15[i] * ARfQ0vec[0] + cthQ15[i] * stateGQ0[i] + 16384) >> 15;
ARgQ0vec[i+1] = (int16_t)WebRtcSpl_SatW32ToW16(tmp32); // Q0
ARfQ0vec[0] = tmpAR;
}
ARgQ0vec[0] = ARfQ0vec[0];
// Filter ARgQ0vec[] and ARfQ0vec[] through coefficients cthQ15[] and sthQ15[].
WebRtcIsacfix_FilterArLoop(ARgQ0vec, ARfQ0vec, cthQ15, sthQ15, orderCoef);
for(n=0;n<HALF_SUBFRAMELEN;n++)
{
lat_outQ0[n + temp1] = ARfQ0vec[n];
}
/* cannot use memcpy in the following */
for (i=0;i<ord_1;i++)
{
stateGQ0[i] = ARgQ0vec[i];
}
}
return;
}
WebRtc_Word16 WebRtcSpl_LevinsonW32_JSK(
WebRtc_Word32 *R, /* (i) Autocorrelation of length >= order+1 */
WebRtc_Word16 *A, /* (o) A[0..order] LPC coefficients (Q11) */
WebRtc_Word16 *K, /* (o) K[0...order-1] Reflection coefficients (Q15) */
WebRtc_Word16 order /* (i) filter order */
) {
WebRtc_Word16 i, j;
WebRtc_Word16 R_hi[LEVINSON_MAX_ORDER+1], R_low[LEVINSON_MAX_ORDER+1];
/* Aurocorr coefficients in high precision */
WebRtc_Word16 A_hi[LEVINSON_MAX_ORDER+1], A_low[LEVINSON_MAX_ORDER+1];
/* LPC coefficients in high precicion */
WebRtc_Word16 A_upd_hi[LEVINSON_MAX_ORDER+1], A_upd_low[LEVINSON_MAX_ORDER+1];
/* LPC coefficients for next iteration */
WebRtc_Word16 K_hi, K_low; /* reflection coefficient in high precision */
WebRtc_Word16 Alpha_hi, Alpha_low, Alpha_exp; /* Prediction gain Alpha in high precision
and with scale factor */
WebRtc_Word16 tmp_hi, tmp_low;
WebRtc_Word32 temp1W32, temp2W32, temp3W32;
WebRtc_Word16 norm;
/* Normalize the autocorrelation R[0]...R[order+1] */
norm = WebRtcSpl_NormW32(R[0]);
for (i=order;i>=0;i--) {
temp1W32 = WEBRTC_SPL_LSHIFT_W32(R[i], norm);
/* Put R in hi and low format */
R_hi[i] = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp1W32, 16);
R_low[i] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp1W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)R_hi[i], 16)), 1);
}
/* K = A[1] = -R[1] / R[0] */
temp2W32 = WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)R_hi[1],16) +
WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)R_low[1],1); /* R[1] in Q31 */
temp3W32 = WEBRTC_SPL_ABS_W32(temp2W32); /* abs R[1] */
temp1W32 = WebRtcSpl_DivW32HiLow(temp3W32, R_hi[0], R_low[0]); /* abs(R[1])/R[0] in Q31 */
/* Put back the sign on R[1] */
if (temp2W32 > 0) {
temp1W32 = -temp1W32;
}
/* Put K in hi and low format */
K_hi = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp1W32, 16);
K_low = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp1W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)K_hi, 16)), 1);
/* Store first reflection coefficient */
K[0] = K_hi;
temp1W32 = WEBRTC_SPL_RSHIFT_W32(temp1W32, 4); /* A[1] in Q27 */
/* Put A[1] in hi and low format */
A_hi[1] = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp1W32, 16);
A_low[1] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp1W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)A_hi[1], 16)), 1);
/* Alpha = R[0] * (1-K^2) */
temp1W32 = WEBRTC_SPL_LSHIFT_W32((WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(K_hi, K_low), 14) +
WEBRTC_SPL_MUL_16_16(K_hi, K_hi)), 1); /* temp1W32 = k^2 in Q31 */
temp1W32 = WEBRTC_SPL_ABS_W32(temp1W32); /* Guard against <0 */
temp1W32 = (WebRtc_Word32)0x7fffffffL - temp1W32; /* temp1W32 = (1 - K[0]*K[0]) in Q31 */
/* Store temp1W32 = 1 - K[0]*K[0] on hi and low format */
tmp_hi = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp1W32, 16);
tmp_low = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp1W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)tmp_hi, 16)), 1);
/* Calculate Alpha in Q31 */
temp1W32 = WEBRTC_SPL_LSHIFT_W32((WEBRTC_SPL_MUL_16_16(R_hi[0], tmp_hi) +
WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(R_hi[0], tmp_low), 15) +
WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(R_low[0], tmp_hi), 15) ), 1);
/* Normalize Alpha and put it in hi and low format */
Alpha_exp = WebRtcSpl_NormW32(temp1W32);
temp1W32 = WEBRTC_SPL_LSHIFT_W32(temp1W32, Alpha_exp);
Alpha_hi = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp1W32, 16);
Alpha_low = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp1W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)Alpha_hi, 16)), 1);
/* Perform the iterative calculations in the
Levinson Durbin algorithm */
for (i=2; i<=order; i++)
{
/* ----
\
temp1W32 = R[i] + > R[j]*A[i-j]
/
----
j=1..i-1
*/
temp1W32 = 0;
for(j=1; j<i; j++) {
/* temp1W32 is in Q31 */
temp1W32 += (WEBRTC_SPL_LSHIFT_W32(WEBRTC_SPL_MUL_16_16(R_hi[j], A_hi[i-j]), 1) +
WEBRTC_SPL_LSHIFT_W32(( WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(R_hi[j], A_low[i-j]), 15) +
WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(R_low[j], A_hi[i-j]), 15) ), 1));
}
temp1W32 = WEBRTC_SPL_LSHIFT_W32(temp1W32, 4);
temp1W32 += (WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)R_hi[i], 16) +
WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)R_low[i], 1));
/* K = -temp1W32 / Alpha */
temp2W32 = WEBRTC_SPL_ABS_W32(temp1W32); /* abs(temp1W32) */
temp3W32 = WebRtcSpl_DivW32HiLow(temp2W32, Alpha_hi, Alpha_low); /* abs(temp1W32)/Alpha */
/* Put the sign of temp1W32 back again */
if (temp1W32 > 0) {
temp3W32 = -temp3W32;
}
/* Use the Alpha shifts from earlier to denormalize */
norm = WebRtcSpl_NormW32(temp3W32);
if ((Alpha_exp <= norm)||(temp3W32==0)) {
temp3W32 = WEBRTC_SPL_LSHIFT_W32(temp3W32, Alpha_exp);
} else {
if (temp3W32 > 0)
{
temp3W32 = (WebRtc_Word32)0x7fffffffL;
} else
{
temp3W32 = (WebRtc_Word32)0x80000000L;
}
}
/* Put K on hi and low format */
K_hi = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp3W32, 16);
K_low = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp3W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)K_hi, 16)), 1);
/* Store Reflection coefficient in Q15 */
K[i-1] = K_hi;
/* Test for unstable filter. If unstable return 0 and let the
user decide what to do in that case
*/
if ((WebRtc_Word32)WEBRTC_SPL_ABS_W16(K_hi) > (WebRtc_Word32)32740) {
return(-i); /* Unstable filter */
}
/*
Compute updated LPC coefficient: Anew[i]
Anew[j]= A[j] + K*A[i-j] for j=1..i-1
Anew[i]= K
*/
for(j=1; j<i; j++)
{
temp1W32 = WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)A_hi[j],16) +
WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)A_low[j],1); /* temp1W32 = A[j] in Q27 */
temp1W32 += WEBRTC_SPL_LSHIFT_W32(( WEBRTC_SPL_MUL_16_16(K_hi, A_hi[i-j]) +
WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(K_hi, A_low[i-j]), 15) +
WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(K_low, A_hi[i-j]), 15) ), 1); /* temp1W32 += K*A[i-j] in Q27 */
/* Put Anew in hi and low format */
A_upd_hi[j] = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp1W32, 16);
A_upd_low[j] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp1W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)A_upd_hi[j], 16)), 1);
}
temp3W32 = WEBRTC_SPL_RSHIFT_W32(temp3W32, 4); /* temp3W32 = K in Q27 (Convert from Q31 to Q27) */
/* Store Anew in hi and low format */
A_upd_hi[i] = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp3W32, 16);
A_upd_low[i] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp3W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)A_upd_hi[i], 16)), 1);
/* Alpha = Alpha * (1-K^2) */
temp1W32 = WEBRTC_SPL_LSHIFT_W32((WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(K_hi, K_low), 14) +
WEBRTC_SPL_MUL_16_16(K_hi, K_hi)), 1); /* K*K in Q31 */
temp1W32 = WEBRTC_SPL_ABS_W32(temp1W32); /* Guard against <0 */
temp1W32 = (WebRtc_Word32)0x7fffffffL - temp1W32; /* 1 - K*K in Q31 */
/* Convert 1- K^2 in hi and low format */
tmp_hi = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp1W32, 16);
tmp_low = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp1W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)tmp_hi, 16)), 1);
/* Calculate Alpha = Alpha * (1-K^2) in Q31 */
temp1W32 = WEBRTC_SPL_LSHIFT_W32(( WEBRTC_SPL_MUL_16_16(Alpha_hi, tmp_hi) +
WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(Alpha_hi, tmp_low), 15) +
WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_16_16(Alpha_low, tmp_hi), 15)), 1);
/* Normalize Alpha and store it on hi and low format */
norm = WebRtcSpl_NormW32(temp1W32);
temp1W32 = WEBRTC_SPL_LSHIFT_W32(temp1W32, norm);
Alpha_hi = (WebRtc_Word16) WEBRTC_SPL_RSHIFT_W32(temp1W32, 16);
Alpha_low = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32((temp1W32 - WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)Alpha_hi, 16)), 1);
/* Update the total nomalization of Alpha */
Alpha_exp = Alpha_exp + norm;
/* Update A[] */
for(j=1; j<=i; j++)
{
A_hi[j] =A_upd_hi[j];
A_low[j] =A_upd_low[j];
}
}
/*
Set A[0] to 1.0 and store the A[i] i=1...order in Q12
(Convert from Q27 and use rounding)
*/
A[0] = 2048;
for(i=1; i<=order; i++) {
/* temp1W32 in Q27 */
temp1W32 = WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)A_hi[i], 16) +
WEBRTC_SPL_LSHIFT_W32((WebRtc_Word32)A_low[i], 1);
/* Round and store upper word */
A[i] = (WebRtc_Word16)WEBRTC_SPL_RSHIFT_W32(temp1W32+(WebRtc_Word32)32768, 16);
}
return(1); /* Stable filters */
}
/* MA filter */
void WebRtcIsacfix_NormLatticeFilterMa(int16_t orderCoef,
int32_t *stateGQ15,
int16_t *lat_inQ0,
int16_t *filt_coefQ15,
int32_t *gain_lo_hiQ17,
int16_t lo_hi,
int16_t *lat_outQ9)
{
int16_t sthQ15[MAX_AR_MODEL_ORDER];
int16_t cthQ15[MAX_AR_MODEL_ORDER];
int u, i, k, n;
int16_t temp2,temp3;
int16_t ord_1 = orderCoef+1;
int32_t inv_cthQ16[MAX_AR_MODEL_ORDER];
int32_t gain32, fQtmp;
int16_t gain16;
int16_t gain_sh;
int32_t tmp32, tmp32b;
int32_t fQ15vec[HALF_SUBFRAMELEN];
int32_t gQ15[MAX_AR_MODEL_ORDER+1][HALF_SUBFRAMELEN];
int16_t sh;
int16_t t16a;
int16_t t16b;
for (u=0;u<SUBFRAMES;u++)
{
int32_t temp1 = WEBRTC_SPL_MUL_16_16(u, HALF_SUBFRAMELEN);
/* set the Direct Form coefficients */
temp2 = (int16_t)WEBRTC_SPL_MUL_16_16(u, orderCoef);
temp3 = (int16_t)WEBRTC_SPL_MUL_16_16(2, u)+lo_hi;
/* compute lattice filter coefficients */
memcpy(sthQ15, &filt_coefQ15[temp2], orderCoef * sizeof(int16_t));
WebRtcSpl_SqrtOfOneMinusXSquared(sthQ15, orderCoef, cthQ15);
/* compute the gain */
gain32 = gain_lo_hiQ17[temp3];
gain_sh = WebRtcSpl_NormW32(gain32);
gain32 = WEBRTC_SPL_LSHIFT_W32(gain32, gain_sh); //Q(17+gain_sh)
for (k=0;k<orderCoef;k++)
{
gain32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[k], gain32); //Q15*Q(17+gain_sh)>>15 = Q(17+gain_sh)
inv_cthQ16[k] = WebRtcSpl_DivW32W16((int32_t)2147483647, cthQ15[k]); // 1/cth[k] in Q31/Q15 = Q16
}
gain16 = (int16_t)(gain32 >> 16); // Q(1+gain_sh).
/* normalized lattice filter */
/*****************************/
/* initial conditions */
for (i=0;i<HALF_SUBFRAMELEN;i++)
{
fQ15vec[i] = WEBRTC_SPL_LSHIFT_W32((int32_t)lat_inQ0[i + temp1], 15); //Q15
gQ15[0][i] = WEBRTC_SPL_LSHIFT_W32((int32_t)lat_inQ0[i + temp1], 15); //Q15
}
fQtmp = fQ15vec[0];
/* get the state of f&g for the first input, for all orders */
for (i=1;i<ord_1;i++)
{
// Calculate f[i][0] = inv_cth[i-1]*(f[i-1][0] + sth[i-1]*stateG[i-1]);
tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(sthQ15[i-1], stateGQ15[i-1]);//Q15*Q15>>15 = Q15
tmp32b= fQtmp + tmp32; //Q15+Q15=Q15
tmp32 = inv_cthQ16[i-1]; //Q16
t16a = (int16_t)(tmp32 >> 16);
t16b = (int16_t) (tmp32-WEBRTC_SPL_LSHIFT_W32(((int32_t)t16a), 16));
if (t16b<0) t16a++;
tmp32 = LATTICE_MUL_32_32_RSFT16(t16a, t16b, tmp32b);
fQtmp = tmp32; // Q15
// Calculate g[i][0] = cth[i-1]*stateG[i-1] + sth[i-1]* f[i][0];
tmp32 = WEBRTC_SPL_MUL_16_32_RSFT15(cthQ15[i-1], stateGQ15[i-1]); //Q15*Q15>>15 = Q15
tmp32b = WEBRTC_SPL_MUL_16_32_RSFT15(sthQ15[i-1], fQtmp); //Q15*Q15>>15 = Q15
tmp32 = tmp32 + tmp32b;//Q15+Q15 = Q15
gQ15[i][0] = tmp32; // Q15
}
/* filtering */
/* save the states */
for(k=0;k<orderCoef;k++)
{
// for 0 <= n < HALF_SUBFRAMELEN - 1:
// f[k+1][n+1] = inv_cth[k]*(f[k][n+1] + sth[k]*g[k][n]);
// g[k+1][n+1] = cth[k]*g[k][n] + sth[k]* f[k+1][n+1];
WebRtcIsacfix_FilterMaLoopFix(sthQ15[k], cthQ15[k], inv_cthQ16[k],
&gQ15[k][0], &gQ15[k+1][1], &fQ15vec[1]);
}
fQ15vec[0] = fQtmp;
for(n=0;n<HALF_SUBFRAMELEN;n++)
{
//gain32 >>= gain_sh; // Q(17+gain_sh) -> Q17
tmp32 = WEBRTC_SPL_MUL_16_32_RSFT16(gain16, fQ15vec[n]); //Q(1+gain_sh)*Q15>>16 = Q(gain_sh)
sh = 9-gain_sh; //number of needed shifts to reach Q9
t16a = (int16_t) WEBRTC_SPL_SHIFT_W32(tmp32, sh);
lat_outQ9[n + temp1] = t16a;
}
/* save the states */
for (i=0;i<ord_1;i++)
{
stateGQ15[i] = gQ15[i][HALF_SUBFRAMELEN-1];
}
//process next frame
}
return;
}
// Compute speech/noise probability
// speech/noise probability is returned in: probSpeechFinal
//snrLocPrior is the prior SNR for each frequency (in Q11)
//snrLocPost is the post SNR for each frequency (in Q11)
void WebRtcNsx_SpeechNoiseProb(NsxInst_t* inst,
uint16_t* nonSpeechProbFinal,
uint32_t* priorLocSnr,
uint32_t* postLocSnr) {
uint32_t zeros, num, den, tmpU32no1, tmpU32no2, tmpU32no3;
int32_t invLrtFX, indPriorFX, tmp32, tmp32no1, tmp32no2, besselTmpFX32;
int32_t frac32, logTmp;
int32_t logLrtTimeAvgKsumFX;
int16_t indPriorFX16;
int16_t tmp16, tmp16no1, tmp16no2, tmpIndFX, tableIndex, frac, intPart;
int i, normTmp, normTmp2, nShifts;
// compute feature based on average LR factor
// this is the average over all frequencies of the smooth log LRT
logLrtTimeAvgKsumFX = 0;
for (i = 0; i < inst->magnLen; i++) {
besselTmpFX32 = (int32_t)postLocSnr[i]; // Q11
normTmp = WebRtcSpl_NormU32(postLocSnr[i]);
num = WEBRTC_SPL_LSHIFT_U32(postLocSnr[i], normTmp); // Q(11+normTmp)
if (normTmp > 10) {
den = WEBRTC_SPL_LSHIFT_U32(priorLocSnr[i], normTmp - 11); // Q(normTmp)
} else {
den = WEBRTC_SPL_RSHIFT_U32(priorLocSnr[i], 11 - normTmp); // Q(normTmp)
}
if (den > 0) {
besselTmpFX32 -= WEBRTC_SPL_UDIV(num, den); // Q11
} else {
besselTmpFX32 -= num; // Q11
}
// inst->logLrtTimeAvg[i] += LRT_TAVG * (besselTmp - log(snrLocPrior)
// - inst->logLrtTimeAvg[i]);
// Here, LRT_TAVG = 0.5
zeros = WebRtcSpl_NormU32(priorLocSnr[i]);
frac32 = (int32_t)(((priorLocSnr[i] << zeros) & 0x7FFFFFFF) >> 19);
tmp32 = WEBRTC_SPL_MUL(frac32, frac32);
tmp32 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL(tmp32, -43), 19);
tmp32 += WEBRTC_SPL_MUL_16_16_RSFT((int16_t)frac32, 5412, 12);
frac32 = tmp32 + 37;
// tmp32 = log2(priorLocSnr[i])
tmp32 = (int32_t)(((31 - zeros) << 12) + frac32) - (11 << 12); // Q12
logTmp = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL_32_16(tmp32, 178), 8);
// log2(priorLocSnr[i])*log(2)
tmp32no1 = WEBRTC_SPL_RSHIFT_W32(logTmp + inst->logLrtTimeAvgW32[i], 1);
// Q12
inst->logLrtTimeAvgW32[i] += (besselTmpFX32 - tmp32no1); // Q12
logLrtTimeAvgKsumFX += inst->logLrtTimeAvgW32[i]; // Q12
}
inst->featureLogLrt = WEBRTC_SPL_RSHIFT_W32(logLrtTimeAvgKsumFX * 5,
inst->stages + 10);
// 5 = BIN_SIZE_LRT / 2
// done with computation of LR factor
//
//compute the indicator functions
//
// average LRT feature
// FLOAT code
// indicator0 = 0.5 * (tanh(widthPrior *
// (logLrtTimeAvgKsum - threshPrior0)) + 1.0);
tmpIndFX = 16384; // Q14(1.0)
tmp32no1 = logLrtTimeAvgKsumFX - inst->thresholdLogLrt; // Q12
nShifts = 7 - inst->stages; // WIDTH_PR_MAP_SHIFT - inst->stages + 5;
//use larger width in tanh map for pause regions
if (tmp32no1 < 0) {
tmpIndFX = 0;
tmp32no1 = -tmp32no1;
//widthPrior = widthPrior * 2.0;
nShifts++;
}
tmp32no1 = WEBRTC_SPL_SHIFT_W32(tmp32no1, nShifts); // Q14
// compute indicator function: sigmoid map
tableIndex = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32no1, 14);
if ((tableIndex < 16) && (tableIndex >= 0)) {
tmp16no2 = kIndicatorTable[tableIndex];
tmp16no1 = kIndicatorTable[tableIndex + 1] - kIndicatorTable[tableIndex];
frac = (int16_t)(tmp32no1 & 0x00003fff); // Q14
tmp16no2 += (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(tmp16no1, frac, 14);
if (tmpIndFX == 0) {
tmpIndFX = 8192 - tmp16no2; // Q14
} else {
tmpIndFX = 8192 + tmp16no2; // Q14
}
}
indPriorFX = WEBRTC_SPL_MUL_16_16(inst->weightLogLrt, tmpIndFX); // 6*Q14
//spectral flatness feature
if (inst->weightSpecFlat) {
tmpU32no1 = WEBRTC_SPL_UMUL(inst->featureSpecFlat, 400); // Q10
tmpIndFX = 16384; // Q14(1.0)
//use larger width in tanh map for pause regions
tmpU32no2 = inst->thresholdSpecFlat - tmpU32no1; //Q10
nShifts = 4;
if (inst->thresholdSpecFlat < tmpU32no1) {
tmpIndFX = 0;
tmpU32no2 = tmpU32no1 - inst->thresholdSpecFlat;
//widthPrior = widthPrior * 2.0;
nShifts++;
}
tmp32no1 = (int32_t)WebRtcSpl_DivU32U16(WEBRTC_SPL_LSHIFT_U32(tmpU32no2,
nShifts), 25);
//Q14
tmpU32no1 = WebRtcSpl_DivU32U16(WEBRTC_SPL_LSHIFT_U32(tmpU32no2, nShifts),
25); //Q14
// compute indicator function: sigmoid map
// FLOAT code
// indicator1 = 0.5 * (tanh(sgnMap * widthPrior *
// (threshPrior1 - tmpFloat1)) + 1.0);
tableIndex = (int16_t)WEBRTC_SPL_RSHIFT_U32(tmpU32no1, 14);
if (tableIndex < 16) {
tmp16no2 = kIndicatorTable[tableIndex];
tmp16no1 = kIndicatorTable[tableIndex + 1] - kIndicatorTable[tableIndex];
frac = (int16_t)(tmpU32no1 & 0x00003fff); // Q14
tmp16no2 += (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(tmp16no1, frac, 14);
if (tmpIndFX) {
tmpIndFX = 8192 + tmp16no2; // Q14
} else {
tmpIndFX = 8192 - tmp16no2; // Q14
}
}
indPriorFX += WEBRTC_SPL_MUL_16_16(inst->weightSpecFlat, tmpIndFX); // 6*Q14
}
//for template spectral-difference
if (inst->weightSpecDiff) {
tmpU32no1 = 0;
if (inst->featureSpecDiff) {
normTmp = WEBRTC_SPL_MIN(20 - inst->stages,
WebRtcSpl_NormU32(inst->featureSpecDiff));
tmpU32no1 = WEBRTC_SPL_LSHIFT_U32(inst->featureSpecDiff, normTmp);
// Q(normTmp-2*stages)
tmpU32no2 = WEBRTC_SPL_RSHIFT_U32(inst->timeAvgMagnEnergy,
20 - inst->stages - normTmp);
if (tmpU32no2 > 0) {
// Q(20 - inst->stages)
tmpU32no1 = WEBRTC_SPL_UDIV(tmpU32no1, tmpU32no2);
} else {
tmpU32no1 = (uint32_t)(0x7fffffff);
}
}
tmpU32no3 = WEBRTC_SPL_UDIV(WEBRTC_SPL_LSHIFT_U32(inst->thresholdSpecDiff,
17),
25);
tmpU32no2 = tmpU32no1 - tmpU32no3;
nShifts = 1;
tmpIndFX = 16384; // Q14(1.0)
//use larger width in tanh map for pause regions
if (tmpU32no2 & 0x80000000) {
tmpIndFX = 0;
tmpU32no2 = tmpU32no3 - tmpU32no1;
//widthPrior = widthPrior * 2.0;
nShifts--;
}
tmpU32no1 = WEBRTC_SPL_RSHIFT_U32(tmpU32no2, nShifts);
// compute indicator function: sigmoid map
/* FLOAT code
indicator2 = 0.5 * (tanh(widthPrior * (tmpFloat1 - threshPrior2)) + 1.0);
*/
tableIndex = (int16_t)WEBRTC_SPL_RSHIFT_U32(tmpU32no1, 14);
if (tableIndex < 16) {
tmp16no2 = kIndicatorTable[tableIndex];
tmp16no1 = kIndicatorTable[tableIndex + 1] - kIndicatorTable[tableIndex];
frac = (int16_t)(tmpU32no1 & 0x00003fff); // Q14
tmp16no2 += (int16_t)WEBRTC_SPL_MUL_16_16_RSFT_WITH_ROUND(
tmp16no1, frac, 14);
if (tmpIndFX) {
tmpIndFX = 8192 + tmp16no2;
} else {
tmpIndFX = 8192 - tmp16no2;
}
}
indPriorFX += WEBRTC_SPL_MUL_16_16(inst->weightSpecDiff, tmpIndFX); // 6*Q14
}
//combine the indicator function with the feature weights
// FLOAT code
// indPrior = 1 - (weightIndPrior0 * indicator0 + weightIndPrior1 *
// indicator1 + weightIndPrior2 * indicator2);
indPriorFX16 = WebRtcSpl_DivW32W16ResW16(98307 - indPriorFX, 6); // Q14
// done with computing indicator function
//compute the prior probability
// FLOAT code
// inst->priorNonSpeechProb += PRIOR_UPDATE *
// (indPriorNonSpeech - inst->priorNonSpeechProb);
tmp16 = indPriorFX16 - inst->priorNonSpeechProb; // Q14
inst->priorNonSpeechProb += (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(
PRIOR_UPDATE_Q14, tmp16, 14); // Q14
//final speech probability: combine prior model with LR factor:
memset(nonSpeechProbFinal, 0, sizeof(uint16_t) * inst->magnLen);
if (inst->priorNonSpeechProb > 0) {
for (i = 0; i < inst->magnLen; i++) {
// FLOAT code
// invLrt = exp(inst->logLrtTimeAvg[i]);
// invLrt = inst->priorSpeechProb * invLrt;
// nonSpeechProbFinal[i] = (1.0 - inst->priorSpeechProb) /
// (1.0 - inst->priorSpeechProb + invLrt);
// invLrt = (1.0 - inst->priorNonSpeechProb) * invLrt;
// nonSpeechProbFinal[i] = inst->priorNonSpeechProb /
// (inst->priorNonSpeechProb + invLrt);
if (inst->logLrtTimeAvgW32[i] < 65300) {
tmp32no1 = WEBRTC_SPL_RSHIFT_W32(WEBRTC_SPL_MUL(
inst->logLrtTimeAvgW32[i], 23637),
14); // Q12
intPart = (int16_t)WEBRTC_SPL_RSHIFT_W32(tmp32no1, 12);
if (intPart < -8) {
intPart = -8;
}
frac = (int16_t)(tmp32no1 & 0x00000fff); // Q12
// Quadratic approximation of 2^frac
tmp32no2 = WEBRTC_SPL_RSHIFT_W32(frac * frac * 44, 19); // Q12
tmp32no2 += WEBRTC_SPL_MUL_16_16_RSFT(frac, 84, 7); // Q12
invLrtFX = WEBRTC_SPL_LSHIFT_W32(1, 8 + intPart)
+ WEBRTC_SPL_SHIFT_W32(tmp32no2, intPart - 4); // Q8
normTmp = WebRtcSpl_NormW32(invLrtFX);
normTmp2 = WebRtcSpl_NormW16((16384 - inst->priorNonSpeechProb));
if (normTmp + normTmp2 >= 7) {
if (normTmp + normTmp2 < 15) {
invLrtFX = WEBRTC_SPL_RSHIFT_W32(invLrtFX, 15 - normTmp2 - normTmp);
// Q(normTmp+normTmp2-7)
tmp32no1 = WEBRTC_SPL_MUL_32_16(invLrtFX,
(16384 - inst->priorNonSpeechProb));
// Q(normTmp+normTmp2+7)
invLrtFX = WEBRTC_SPL_SHIFT_W32(tmp32no1, 7 - normTmp - normTmp2);
// Q14
} else {
tmp32no1 = WEBRTC_SPL_MUL_32_16(invLrtFX,
(16384 - inst->priorNonSpeechProb));
// Q22
invLrtFX = WEBRTC_SPL_RSHIFT_W32(tmp32no1, 8); // Q14
}
tmp32no1 = WEBRTC_SPL_LSHIFT_W32((int32_t)inst->priorNonSpeechProb,
8); // Q22
nonSpeechProbFinal[i] = (uint16_t)WEBRTC_SPL_DIV(tmp32no1,
(int32_t)inst->priorNonSpeechProb + invLrtFX); // Q8
}
}
}
}
}
int16_t WebRtcIsacfix_DecodeImpl(int16_t *signal_out16,
ISACFIX_DecInst_t *ISACdec_obj,
int16_t *current_framesamples)
{
int k;
int err;
int16_t BWno;
int16_t len = 0;
int16_t model;
int16_t Vector_Word16_1[FRAMESAMPLES/2];
int16_t Vector_Word16_2[FRAMESAMPLES/2];
int32_t Vector_Word32_1[FRAMESAMPLES/2];
int32_t Vector_Word32_2[FRAMESAMPLES/2];
int16_t lofilt_coefQ15[ORDERLO*SUBFRAMES]; //refl. coeffs
int16_t hifilt_coefQ15[ORDERHI*SUBFRAMES]; //refl. coeffs
int32_t gain_lo_hiQ17[2*SUBFRAMES];
int16_t PitchLags_Q7[PITCH_SUBFRAMES];
int16_t PitchGains_Q12[PITCH_SUBFRAMES];
int16_t AvgPitchGain_Q12;
int16_t tmp_1, tmp_2;
int32_t tmp32a, tmp32b;
int16_t gainQ13;
int16_t frame_nb; /* counter */
int16_t frame_mode; /* 0 for 20ms and 30ms, 1 for 60ms */
int16_t processed_samples;
/* PLC */
int16_t overlapWin[ 240 ];
(ISACdec_obj->bitstr_obj).W_upper = 0xFFFFFFFF;
(ISACdec_obj->bitstr_obj).streamval = 0;
(ISACdec_obj->bitstr_obj).stream_index = 0;
(ISACdec_obj->bitstr_obj).full = 1;
/* decode framelength and BW estimation - not used, only for stream pointer*/
err = WebRtcIsacfix_DecodeFrameLen(&ISACdec_obj->bitstr_obj, current_framesamples);
if (err<0) // error check
return err;
frame_mode = (int16_t)WEBRTC_SPL_DIV(*current_framesamples, MAX_FRAMESAMPLES); /* 0, or 1 */
processed_samples = (int16_t)WEBRTC_SPL_DIV(*current_framesamples, frame_mode+1); /* either 320 (20ms) or 480 (30, 60 ms) */
err = WebRtcIsacfix_DecodeSendBandwidth(&ISACdec_obj->bitstr_obj, &BWno);
if (err<0) // error check
return err;
/* one loop if it's one frame (20 or 30ms), 2 loops if 2 frames bundled together (60ms) */
for (frame_nb = 0; frame_nb <= frame_mode; frame_nb++) {
/* decode & dequantize pitch parameters */
err = WebRtcIsacfix_DecodePitchGain(&(ISACdec_obj->bitstr_obj), PitchGains_Q12);
if (err<0) // error check
return err;
err = WebRtcIsacfix_DecodePitchLag(&ISACdec_obj->bitstr_obj, PitchGains_Q12, PitchLags_Q7);
if (err<0) // error check
return err;
AvgPitchGain_Q12 = (int16_t)(((int32_t)PitchGains_Q12[0] + PitchGains_Q12[1] + PitchGains_Q12[2] + PitchGains_Q12[3])>>2);
/* decode & dequantize FiltCoef */
err = WebRtcIsacfix_DecodeLpc(gain_lo_hiQ17, lofilt_coefQ15, hifilt_coefQ15,
&ISACdec_obj->bitstr_obj, &model);
if (err<0) // error check
return err;
/* decode & dequantize spectrum */
len = WebRtcIsacfix_DecodeSpec(&ISACdec_obj->bitstr_obj, Vector_Word16_1, Vector_Word16_2, AvgPitchGain_Q12);
if (len < 0) // error check
return len;
// Why does this need Q16 in and out? /JS
WebRtcIsacfix_Spec2Time(Vector_Word16_1, Vector_Word16_2, Vector_Word32_1, Vector_Word32_2);
for (k=0; k<FRAMESAMPLES/2; k++) {
Vector_Word16_1[k] = (int16_t)WEBRTC_SPL_RSHIFT_W32(Vector_Word32_1[k]+64, 7); //Q16 -> Q9
}
/* ---- If this is recovery frame ---- */
if( (ISACdec_obj->plcstr_obj).used == PLC_WAS_USED )
{
(ISACdec_obj->plcstr_obj).used = PLC_NOT_USED;
if( (ISACdec_obj->plcstr_obj).B < 1000 )
{
(ISACdec_obj->plcstr_obj).decayCoeffPriodic = 4000;
}
ISACdec_obj->plcstr_obj.decayCoeffPriodic = WEBRTC_SPL_WORD16_MAX; /* DECAY_RATE is in Q15 */
ISACdec_obj->plcstr_obj.decayCoeffNoise = WEBRTC_SPL_WORD16_MAX; /* DECAY_RATE is in Q15 */
ISACdec_obj->plcstr_obj.pitchCycles = 0;
PitchGains_Q12[0] = (int16_t)WEBRTC_SPL_MUL_16_16_RSFT(PitchGains_Q12[0], 700, 10 );
/* ---- Add-overlap ---- */
WebRtcSpl_GetHanningWindow( overlapWin, RECOVERY_OVERLAP );
for( k = 0; k < RECOVERY_OVERLAP; k++ )
Vector_Word16_1[k] = WebRtcSpl_AddSatW16(
(int16_t)WEBRTC_SPL_MUL_16_16_RSFT( (ISACdec_obj->plcstr_obj).overlapLP[k], overlapWin[RECOVERY_OVERLAP - k - 1], 14),
(int16_t)WEBRTC_SPL_MUL_16_16_RSFT( Vector_Word16_1[k], overlapWin[k], 14) );
}
/* --- Store side info --- */
if( frame_nb == frame_mode )
{
/* --- LPC info */
WEBRTC_SPL_MEMCPY_W16( (ISACdec_obj->plcstr_obj).lofilt_coefQ15, &lofilt_coefQ15[(SUBFRAMES-1)*ORDERLO], ORDERLO );
WEBRTC_SPL_MEMCPY_W16( (ISACdec_obj->plcstr_obj).hifilt_coefQ15, &hifilt_coefQ15[(SUBFRAMES-1)*ORDERHI], ORDERHI );
(ISACdec_obj->plcstr_obj).gain_lo_hiQ17[0] = gain_lo_hiQ17[(SUBFRAMES-1) * 2];
(ISACdec_obj->plcstr_obj).gain_lo_hiQ17[1] = gain_lo_hiQ17[(SUBFRAMES-1) * 2 + 1];
/* --- LTP info */
(ISACdec_obj->plcstr_obj).AvgPitchGain_Q12 = PitchGains_Q12[3];
(ISACdec_obj->plcstr_obj).lastPitchGain_Q12 = PitchGains_Q12[3];
(ISACdec_obj->plcstr_obj).lastPitchLag_Q7 = PitchLags_Q7[3];
if( PitchLags_Q7[3] < 3000 )
(ISACdec_obj->plcstr_obj).lastPitchLag_Q7 += PitchLags_Q7[3];
WEBRTC_SPL_MEMCPY_W16( (ISACdec_obj->plcstr_obj).prevPitchInvIn, Vector_Word16_1, FRAMESAMPLES/2 );
}
/* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ */
/* inverse pitch filter */
WebRtcIsacfix_PitchFilter(Vector_Word16_1, Vector_Word16_2, &ISACdec_obj->pitchfiltstr_obj, PitchLags_Q7, PitchGains_Q12, 4);
if( frame_nb == frame_mode )
{
WEBRTC_SPL_MEMCPY_W16( (ISACdec_obj->plcstr_obj).prevPitchInvOut, &(Vector_Word16_2[FRAMESAMPLES/2 - (PITCH_MAX_LAG + 10)]), PITCH_MAX_LAG );
}
/* reduce gain to compensate for pitch enhancer */
/* gain = 1.0f - 0.45f * AvgPitchGain; */
tmp32a = WEBRTC_SPL_MUL_16_16_RSFT(AvgPitchGain_Q12, 29, 0); // Q18
tmp32b = 262144 - tmp32a; // Q18
gainQ13 = (int16_t) WEBRTC_SPL_RSHIFT_W32(tmp32b, 5); // Q13
for (k = 0; k < FRAMESAMPLES/2; k++)
{
Vector_Word32_1[k] = (int32_t) WEBRTC_SPL_LSHIFT_W32(WEBRTC_SPL_MUL_16_16(Vector_Word16_2[k], gainQ13), 3); // Q25
}
/* perceptual post-filtering (using normalized lattice filter) */
WebRtcIsacfix_NormLatticeFilterAr(ORDERLO, (ISACdec_obj->maskfiltstr_obj).PostStateLoGQ0,
Vector_Word32_1, lofilt_coefQ15, gain_lo_hiQ17, 0, Vector_Word16_1);
/* --- Store Highpass Residual --- */
for (k = 0; k < FRAMESAMPLES/2; k++)
Vector_Word32_1[k] = WEBRTC_SPL_LSHIFT_W32(Vector_Word32_2[k], 9); // Q16 -> Q25
for( k = 0; k < PITCH_MAX_LAG + 10; k++ )
(ISACdec_obj->plcstr_obj).prevHP[k] = Vector_Word32_1[FRAMESAMPLES/2 - (PITCH_MAX_LAG + 10) + k];
WebRtcIsacfix_NormLatticeFilterAr(ORDERHI, (ISACdec_obj->maskfiltstr_obj).PostStateHiGQ0,
Vector_Word32_1, hifilt_coefQ15, gain_lo_hiQ17, 1, Vector_Word16_2);
/* recombine the 2 bands */
/* Form the polyphase signals, and compensate for DC offset */
for (k=0;k<FRAMESAMPLES/2;k++) {
tmp_1 = (int16_t)WebRtcSpl_SatW32ToW16(((int32_t)Vector_Word16_1[k]+Vector_Word16_2[k] + 1)); /* Construct a new upper channel signal*/
tmp_2 = (int16_t)WebRtcSpl_SatW32ToW16(((int32_t)Vector_Word16_1[k]-Vector_Word16_2[k])); /* Construct a new lower channel signal*/
Vector_Word16_1[k] = tmp_1;
Vector_Word16_2[k] = tmp_2;
}
WebRtcIsacfix_FilterAndCombine1(Vector_Word16_1, Vector_Word16_2, signal_out16 + frame_nb * processed_samples, &ISACdec_obj->postfiltbankstr_obj);
}
return len;
}
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;
}