int16_t WebRtcSpl_GetScalingSquare(int16_t* in_vector,
int in_vector_length,
int times)
{
int16_t nbits = WebRtcSpl_GetSizeInBits(times);
int i;
int16_t smax = -1;
int16_t sabs;
int16_t *sptr = in_vector;
int16_t t;
int looptimes = in_vector_length;
for (i = looptimes; i > 0; i--)
{
sabs = (*sptr > 0 ? *sptr++ : -*sptr++);
smax = (sabs > smax ? sabs : smax);
}
t = WebRtcSpl_NormW32(WEBRTC_SPL_MUL(smax, smax));
if (smax == 0)
{
return 0; // Since norm(0) returns 0
} else
{
return (t > nbits) ? 0 : nbits - t;
}
}
int WebRtcSpl_AutoCorrelation(const int16_t* in_vector,
int in_vector_length,
int order,
int32_t* result,
int* scale) {
int32_t sum = 0;
int i = 0, j = 0;
int16_t smax = 0;
int scaling = 0;
if (order > in_vector_length) {
/* Undefined */
return -1;
} else if (order < 0) {
order = in_vector_length;
}
// Find the maximum absolute value of the samples.
smax = WebRtcSpl_MaxAbsValueW16(in_vector, in_vector_length);
// In order to avoid overflow when computing the sum we should scale the
// samples so that (in_vector_length * smax * smax) will not overflow.
if (smax == 0) {
scaling = 0;
} else {
// Number of bits in the sum loop.
int nbits = WebRtcSpl_GetSizeInBits((uint32_t)in_vector_length);
// Number of bits to normalize smax.
int t = WebRtcSpl_NormW32(WEBRTC_SPL_MUL(smax, smax));
if (t > nbits) {
scaling = 0;
} else {
scaling = nbits - t;
}
}
// Perform the actual correlation calculation.
for (i = 0; i < order + 1; i++) {
sum = 0;
/* Unroll the loop to improve performance. */
for (j = 0; i + j + 3 < in_vector_length; j += 4) {
sum += (in_vector[j + 0] * in_vector[i + j + 0]) >> scaling;
sum += (in_vector[j + 1] * in_vector[i + j + 1]) >> scaling;
sum += (in_vector[j + 2] * in_vector[i + j + 2]) >> scaling;
sum += (in_vector[j + 3] * in_vector[i + j + 3]) >> scaling;
}
for (; j < in_vector_length - i; j++) {
sum += (in_vector[j] * in_vector[i + j]) >> scaling;
}
*result++ = sum;
}
*scale = scaling;
return order + 1;
}
// 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
}
}
}
}
}
// Contains a function for the core loop in the normalized lattice MA
// filter routine for iSAC codec, optimized for ARM Neon platform.
// It does:
// for 0 <= n < HALF_SUBFRAMELEN - 1:
// *ptr2 = input2 * (*ptr2) + input0 * (*ptr0));
// *ptr1 = input1 * (*ptr0) + input0 * (*ptr2);
// Output is not bit-exact with the reference C code, due to the replacement
// of WEBRTC_SPL_MUL_16_32_RSFT15 and LATTICE_MUL_32_32_RSFT16 with Neon
// instructions. The difference should not be bigger than 1.
void WebRtcIsacfix_FilterMaLoopNeon(int16_t input0, // Filter coefficient
int16_t input1, // Filter coefficient
int32_t input2, // Inverse coefficient
int32_t* ptr0, // Sample buffer
int32_t* ptr1, // Sample buffer
int32_t* ptr2) // Sample buffer
{
int n = 0;
int loop = (HALF_SUBFRAMELEN - 1) >> 3;
int loop_tail = (HALF_SUBFRAMELEN - 1) & 0x7;
int32x4_t input0_v = vdupq_n_s32((int32_t)input0 << 16);
int32x4_t input1_v = vdupq_n_s32((int32_t)input1 << 16);
int32x4_t input2_v = vdupq_n_s32(input2);
int32x4_t tmp0a, tmp1a, tmp2a, tmp3a;
int32x4_t tmp0b, tmp1b, tmp2b, tmp3b;
int32x4_t ptr0va, ptr1va, ptr2va;
int32x4_t ptr0vb, ptr1vb, ptr2vb;
// Unroll to process 8 samples at once.
for (n = 0; n < loop; n++) {
ptr0va = vld1q_s32(ptr0);
ptr0vb = vld1q_s32(ptr0 + 4);
ptr0 += 8;
ptr2va = vld1q_s32(ptr2);
ptr2vb = vld1q_s32(ptr2 + 4);
// Calculate tmp0 = (*ptr0) * input0.
tmp0a = vqrdmulhq_s32(ptr0va, input0_v);
tmp0b = vqrdmulhq_s32(ptr0vb, input0_v);
// Calculate tmp1 = (*ptr0) * input1.
tmp1a = vqrdmulhq_s32(ptr0va, input1_v);
tmp1b = vqrdmulhq_s32(ptr0vb, input1_v);
// Calculate tmp2 = tmp0 + *(ptr2).
tmp2a = vaddq_s32(tmp0a, ptr2va);
tmp2b = vaddq_s32(tmp0b, ptr2vb);
tmp2a = vshlq_n_s32(tmp2a, 15);
tmp2b = vshlq_n_s32(tmp2b, 15);
// Calculate *ptr2 = input2 * tmp2.
ptr2va = vqrdmulhq_s32(tmp2a, input2_v);
ptr2vb = vqrdmulhq_s32(tmp2b, input2_v);
vst1q_s32(ptr2, ptr2va);
vst1q_s32(ptr2 + 4, ptr2vb);
ptr2 += 8;
// Calculate tmp3 = ptr2v * input0.
tmp3a = vqrdmulhq_s32(ptr2va, input0_v);
tmp3b = vqrdmulhq_s32(ptr2vb, input0_v);
// Calculate *ptr1 = tmp1 + tmp3.
ptr1va = vaddq_s32(tmp1a, tmp3a);
ptr1vb = vaddq_s32(tmp1b, tmp3b);
vst1q_s32(ptr1, ptr1va);
vst1q_s32(ptr1 + 4, ptr1vb);
ptr1 += 8;
}
// Process four more samples.
if (loop_tail & 0x4) {
ptr0va = vld1q_s32(ptr0);
ptr2va = vld1q_s32(ptr2);
ptr0 += 4;
// Calculate tmp0 = (*ptr0) * input0.
tmp0a = vqrdmulhq_s32(ptr0va, input0_v);
// Calculate tmp1 = (*ptr0) * input1.
tmp1a = vqrdmulhq_s32(ptr0va, input1_v);
// Calculate tmp2 = tmp0 + *(ptr2).
tmp2a = vaddq_s32(tmp0a, ptr2va);
tmp2a = vshlq_n_s32(tmp2a, 15);
// Calculate *ptr2 = input2 * tmp2.
ptr2va = vqrdmulhq_s32(tmp2a, input2_v);
vst1q_s32(ptr2, ptr2va);
ptr2 += 4;
// Calculate tmp3 = *(ptr2) * input0.
tmp3a = vqrdmulhq_s32(ptr2va, input0_v);
// Calculate *ptr1 = tmp1 + tmp3.
ptr1va = vaddq_s32(tmp1a, tmp3a);
vst1q_s32(ptr1, ptr1va);
ptr1 += 4;
}
// Process two more samples.
if (loop_tail & 0x2) {
int32x2_t ptr0v_tail, ptr2v_tail, ptr1v_tail;
int32x2_t tmp0_tail, tmp1_tail, tmp2_tail, tmp3_tail;
ptr0v_tail = vld1_s32(ptr0);
ptr2v_tail = vld1_s32(ptr2);
ptr0 += 2;
// Calculate tmp0 = (*ptr0) * input0.
tmp0_tail = vqrdmulh_s32(ptr0v_tail, vget_low_s32(input0_v));
// Calculate tmp1 = (*ptr0) * input1.
tmp1_tail = vqrdmulh_s32(ptr0v_tail, vget_low_s32(input1_v));
// Calculate tmp2 = tmp0 + *(ptr2).
tmp2_tail = vadd_s32(tmp0_tail, ptr2v_tail);
tmp2_tail = vshl_n_s32(tmp2_tail, 15);
// Calculate *ptr2 = input2 * tmp2.
ptr2v_tail = vqrdmulh_s32(tmp2_tail, vget_low_s32(input2_v));
vst1_s32(ptr2, ptr2v_tail);
ptr2 += 2;
// Calculate tmp3 = *(ptr2) * input0.
tmp3_tail = vqrdmulh_s32(ptr2v_tail, vget_low_s32(input0_v));
// Calculate *ptr1 = tmp1 + tmp3.
ptr1v_tail = vadd_s32(tmp1_tail, tmp3_tail);
vst1_s32(ptr1, ptr1v_tail);
ptr1 += 2;
}
// Process one more sample.
if (loop_tail & 0x1) {
int16_t t16a = (int16_t)(input2 >> 16);
int16_t t16b = (int16_t)input2;
if (t16b < 0) t16a++;
int32_t tmp32a;
int32_t tmp32b;
// Calculate *ptr2 = input2 * (*ptr2 + input0 * (*ptr0)).
tmp32a = WEBRTC_SPL_MUL_16_32_RSFT15(input0, *ptr0);
tmp32b = *ptr2 + tmp32a;
*ptr2 = (int32_t)(WEBRTC_SPL_MUL(t16a, tmp32b) +
(WEBRTC_SPL_MUL_16_32_RSFT16(t16b, tmp32b)));
// Calculate *ptr1 = input1 * (*ptr0) + input0 * (*ptr2).
tmp32a = WEBRTC_SPL_MUL_16_32_RSFT15(input1, *ptr0);
tmp32b = WEBRTC_SPL_MUL_16_32_RSFT15(input0, *ptr2);
*ptr1 = tmp32a + tmp32b;
}