/// <summary>
/// Get the speech activity level in Q8
/// </summary>
/// <param name="psEncC">I/O Encoder state</param>
/// <param name="pIn">I PCM input</param>
/// <param name="pIn_ptr"></param>
/// <returns>0 if success</returns>
internal static int silk_VAD_GetSA_Q8(
SilkChannelEncoder psEncC,
short[] pIn,
int pIn_ptr)
{
int SA_Q15, pSNR_dB_Q7, input_tilt;
int decimated_framelength1, decimated_framelength2;
int decimated_framelength;
int dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s;
int sumSquared = 0, smooth_coef_Q16;
short HPstateTmp;
short[] X;
int[] Xnrg = new int[SilkConstants.VAD_N_BANDS];
int[] NrgToNoiseRatio_Q8 = new int[SilkConstants.VAD_N_BANDS];
int speech_nrg, x_tmp;
int[] X_offset = new int[SilkConstants.VAD_N_BANDS];
int ret = 0;
SilkVADState psSilk_VAD = psEncC.sVAD;
/* Safety checks */
Inlines.OpusAssert(SilkConstants.VAD_N_BANDS == 4);
Inlines.OpusAssert(SilkConstants.MAX_FRAME_LENGTH >= psEncC.frame_length);
Inlines.OpusAssert(psEncC.frame_length <= 512);
Inlines.OpusAssert(psEncC.frame_length == 8 * Inlines.silk_RSHIFT(psEncC.frame_length, 3));
/***********************/
/* Filter and Decimate */
/***********************/
decimated_framelength1 = Inlines.silk_RSHIFT(psEncC.frame_length, 1);
decimated_framelength2 = Inlines.silk_RSHIFT(psEncC.frame_length, 2);
decimated_framelength = Inlines.silk_RSHIFT(psEncC.frame_length, 3);
/* Decimate into 4 bands:
* 0 L 3L L 3L 5L
* - -- - -- --
* 8 8 2 4 4
*
* [0-1 kHz| temp. |1-2 kHz| 2-4 kHz | 4-8 kHz |
*
* They're arranged to allow the minimal ( frame_length / 4 ) extra
* scratch space during the downsampling process */
X_offset[0] = 0;
X_offset[1] = decimated_framelength + decimated_framelength2;
X_offset[2] = X_offset[1] + decimated_framelength;
X_offset[3] = X_offset[2] + decimated_framelength2;
X = new short[X_offset[3] + decimated_framelength1];
/* 0-8 kHz to 0-4 kHz and 4-8 kHz */
Filters.silk_ana_filt_bank_1(pIn, pIn_ptr, psSilk_VAD.AnaState,
X, X, X_offset[3], psEncC.frame_length);
/* 0-4 kHz to 0-2 kHz and 2-4 kHz */
Filters.silk_ana_filt_bank_1(X, 0, psSilk_VAD.AnaState1,
X, X, X_offset[2], decimated_framelength1);
/* 0-2 kHz to 0-1 kHz and 1-2 kHz */
Filters.silk_ana_filt_bank_1(X, 0, psSilk_VAD.AnaState2,
X, X, X_offset[1], decimated_framelength2);
/*********************************************/
/* HP filter on lowest band (differentiator) */
/*********************************************/
X[decimated_framelength - 1] = (short)(Inlines.silk_RSHIFT(X[decimated_framelength - 1], 1));
HPstateTmp = X[decimated_framelength - 1];
for (i = decimated_framelength - 1; i > 0; i--)
{
X[i - 1] = (short)(Inlines.silk_RSHIFT(X[i - 1], 1));
X[i] -= X[i - 1];
}
X[0] -= psSilk_VAD.HPstate;
psSilk_VAD.HPstate = HPstateTmp;
/*************************************/
/* Calculate the energy in each band */
/*************************************/
for (b = 0; b < SilkConstants.VAD_N_BANDS; b++)
{
/* Find the decimated framelength in the non-uniformly divided bands */
decimated_framelength = Inlines.silk_RSHIFT(psEncC.frame_length, Inlines.silk_min_int(SilkConstants.VAD_N_BANDS - b, SilkConstants.VAD_N_BANDS - 1));
/* Split length into subframe lengths */
dec_subframe_length = Inlines.silk_RSHIFT(decimated_framelength, SilkConstants.VAD_INTERNAL_SUBFRAMES_LOG2);
dec_subframe_offset = 0;
/* Compute energy per sub-frame */
/* initialize with summed energy of last subframe */
Xnrg[b] = psSilk_VAD.XnrgSubfr[b];
for (s = 0; s < SilkConstants.VAD_INTERNAL_SUBFRAMES; s++)
{
sumSquared = 0;
for (i = 0; i < dec_subframe_length; i++)
{
/* The energy will be less than dec_subframe_length * ( silk_int16_MIN / 8 ) ^ 2. */
/* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128) */
x_tmp = Inlines.silk_RSHIFT(
X[X_offset[b] + i + dec_subframe_offset], 3);
sumSquared = Inlines.silk_SMLABB(sumSquared, x_tmp, x_tmp);
/* Safety check */
Inlines.OpusAssert(sumSquared >= 0);
}
/* Add/saturate summed energy of current subframe */
if (s < SilkConstants.VAD_INTERNAL_SUBFRAMES - 1)
{
Xnrg[b] = Inlines.silk_ADD_POS_SAT32(Xnrg[b], sumSquared);
}
else
{
/* Look-ahead subframe */
Xnrg[b] = Inlines.silk_ADD_POS_SAT32(Xnrg[b], Inlines.silk_RSHIFT(sumSquared, 1));
}
dec_subframe_offset += dec_subframe_length;
}
psSilk_VAD.XnrgSubfr[b] = sumSquared;
}
/********************/
/* Noise estimation */
/********************/
silk_VAD_GetNoiseLevels(Xnrg, psSilk_VAD);
/***********************************************/
/* Signal-plus-noise to noise ratio estimation */
/***********************************************/
sumSquared = 0;
input_tilt = 0;
for (b = 0; b < SilkConstants.VAD_N_BANDS; b++)
{
speech_nrg = Xnrg[b] - psSilk_VAD.NL[b];
if (speech_nrg > 0)
{
/* Divide, with sufficient resolution */
if ((Xnrg[b] & 0xFF800000) == 0)
{
NrgToNoiseRatio_Q8[b] = Inlines.silk_DIV32(Inlines.silk_LSHIFT(Xnrg[b], 8), psSilk_VAD.NL[b] + 1);
}
else
{
NrgToNoiseRatio_Q8[b] = Inlines.silk_DIV32(Xnrg[b], Inlines.silk_RSHIFT(psSilk_VAD.NL[b], 8) + 1);
}
/* Convert to log domain */
SNR_Q7 = Inlines.silk_lin2log(NrgToNoiseRatio_Q8[b]) - 8 * 128;
/* Sum-of-squares */
sumSquared = Inlines.silk_SMLABB(sumSquared, SNR_Q7, SNR_Q7); /* Q14 */
/* Tilt measure */
if (speech_nrg < ((int)1 << 20))
{
/* Scale down SNR value for small subband speech energies */
SNR_Q7 = Inlines.silk_SMULWB(Inlines.silk_LSHIFT(Inlines.silk_SQRT_APPROX(speech_nrg), 6), SNR_Q7);
}
input_tilt = Inlines.silk_SMLAWB(input_tilt, tiltWeights[b], SNR_Q7);
}
else
{
NrgToNoiseRatio_Q8[b] = 256;
}
}
/* Mean-of-squares */
sumSquared = Inlines.silk_DIV32_16(sumSquared, SilkConstants.VAD_N_BANDS); /* Q14 */
/* Root-mean-square approximation, scale to dBs, and write to output pointer */
pSNR_dB_Q7 = (short)(3 * Inlines.silk_SQRT_APPROX(sumSquared)); /* Q7 */
/*********************************/
/* Speech Probability Estimation */
/*********************************/
SA_Q15 = Sigmoid.silk_sigm_Q15(Inlines.silk_SMULWB(SilkConstants.VAD_SNR_FACTOR_Q16, pSNR_dB_Q7) - SilkConstants.VAD_NEGATIVE_OFFSET_Q5);
/**************************/
/* Frequency Tilt Measure */
/**************************/
psEncC.input_tilt_Q15 = Inlines.silk_LSHIFT(Sigmoid.silk_sigm_Q15(input_tilt) - 16384, 1);
/**************************************************/
/* Scale the sigmoid output based on power levels */
/**************************************************/
speech_nrg = 0;
for (b = 0; b < SilkConstants.VAD_N_BANDS; b++)
{
/* Accumulate signal-without-noise energies, higher frequency bands have more weight */
speech_nrg += (b + 1) * Inlines.silk_RSHIFT(Xnrg[b] - psSilk_VAD.NL[b], 4);
}
/* Power scaling */
if (speech_nrg <= 0)
{
SA_Q15 = Inlines.silk_RSHIFT(SA_Q15, 1);
}
else if (speech_nrg < 32768)
{
if (psEncC.frame_length == 10 * psEncC.fs_kHz)
{
speech_nrg = Inlines.silk_LSHIFT_SAT32(speech_nrg, 16);
}
else
{
speech_nrg = Inlines.silk_LSHIFT_SAT32(speech_nrg, 15);
}
/* square-root */
speech_nrg = Inlines.silk_SQRT_APPROX(speech_nrg);
SA_Q15 = Inlines.silk_SMULWB(32768 + speech_nrg, SA_Q15);
}
/* Copy the resulting speech activity in Q8 */
psEncC.speech_activity_Q8 = Inlines.silk_min_int(Inlines.silk_RSHIFT(SA_Q15, 7), byte.MaxValue);
/***********************************/
/* Energy Level and SNR estimation */
/***********************************/
/* Smoothing coefficient */
smooth_coef_Q16 = Inlines.silk_SMULWB(SilkConstants.VAD_SNR_SMOOTH_COEF_Q18, Inlines.silk_SMULWB((int)SA_Q15, SA_Q15));
if (psEncC.frame_length == 10 * psEncC.fs_kHz)
{
smooth_coef_Q16 >>= 1;
}
for (b = 0; b < SilkConstants.VAD_N_BANDS; b++)
{
/* compute smoothed energy-to-noise ratio per band */
psSilk_VAD.NrgRatioSmth_Q8[b] = Inlines.silk_SMLAWB(psSilk_VAD.NrgRatioSmth_Q8[b],
NrgToNoiseRatio_Q8[b] - psSilk_VAD.NrgRatioSmth_Q8[b], smooth_coef_Q16);
/* signal to noise ratio in dB per band */
SNR_Q7 = 3 * (Inlines.silk_lin2log(psSilk_VAD.NrgRatioSmth_Q8[b]) - 8 * 128);
/* quality = sigmoid( 0.25 * ( SNR_dB - 16 ) ); */
psEncC.input_quality_bands_Q15[b] = Sigmoid.silk_sigm_Q15(Inlines.silk_RSHIFT(SNR_Q7 - 16 * 128, 4));
}
return(ret);
}
/* Processing of gains */
internal static void silk_process_gains(
SilkChannelEncoder psEnc, /* I/O Encoder state */
SilkEncoderControl psEncCtrl, /* I/O Encoder control */
int condCoding /* I The type of conditional coding to use */
)
{
SilkShapeState psShapeSt = psEnc.sShape;
int k;
int s_Q16, InvMaxSqrVal_Q16, gain, gain_squared, ResNrg, ResNrgPart, quant_offset_Q10;
/* Gain reduction when LTP coding gain is high */
if (psEnc.indices.signalType == SilkConstants.TYPE_VOICED)
{
/*s = -0.5f * silk_sigmoid( 0.25f * ( psEncCtrl.LTPredCodGain - 12.0f ) ); */
s_Q16 = 0 - Sigmoid.silk_sigm_Q15(Inlines.silk_RSHIFT_ROUND(psEncCtrl.LTPredCodGain_Q7 - ((int)((12.0f) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(12.0f, 7)*/, 4));
for (k = 0; k < psEnc.nb_subfr; k++)
{
psEncCtrl.Gains_Q16[k] = Inlines.silk_SMLAWB(psEncCtrl.Gains_Q16[k], psEncCtrl.Gains_Q16[k], s_Q16);
}
}
/* Limit the quantized signal */
/* InvMaxSqrVal = pow( 2.0f, 0.33f * ( 21.0f - SNR_dB ) ) / subfr_length; */
InvMaxSqrVal_Q16 = Inlines.silk_DIV32_16(Inlines.silk_log2lin(
Inlines.silk_SMULWB(((int)((21 + 16 / 0.33f) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(21 + 16 / 0.33f, 7)*/ - psEnc.SNR_dB_Q7, ((int)((0.33f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(0.33f, 16)*/)), psEnc.subfr_length);
for (k = 0; k < psEnc.nb_subfr; k++)
{
/* Soft limit on ratio residual energy and squared gains */
ResNrg = psEncCtrl.ResNrg[k];
ResNrgPart = Inlines.silk_SMULWW(ResNrg, InvMaxSqrVal_Q16);
if (psEncCtrl.ResNrgQ[k] > 0)
{
ResNrgPart = Inlines.silk_RSHIFT_ROUND(ResNrgPart, psEncCtrl.ResNrgQ[k]);
}
else
{
if (ResNrgPart >= Inlines.silk_RSHIFT(int.MaxValue, -psEncCtrl.ResNrgQ[k]))
{
ResNrgPart = int.MaxValue;
}
else
{
ResNrgPart = Inlines.silk_LSHIFT(ResNrgPart, -psEncCtrl.ResNrgQ[k]);
}
}
gain = psEncCtrl.Gains_Q16[k];
gain_squared = Inlines.silk_ADD_SAT32(ResNrgPart, Inlines.silk_SMMUL(gain, gain));
if (gain_squared < short.MaxValue)
{
/* recalculate with higher precision */
gain_squared = Inlines.silk_SMLAWW(Inlines.silk_LSHIFT(ResNrgPart, 16), gain, gain);
Inlines.OpusAssert(gain_squared > 0);
gain = Inlines.silk_SQRT_APPROX(gain_squared); /* Q8 */
gain = Inlines.silk_min(gain, int.MaxValue >> 8);
psEncCtrl.Gains_Q16[k] = Inlines.silk_LSHIFT_SAT32(gain, 8); /* Q16 */
}
else
{
gain = Inlines.silk_SQRT_APPROX(gain_squared); /* Q0 */
gain = Inlines.silk_min(gain, int.MaxValue >> 16);
psEncCtrl.Gains_Q16[k] = Inlines.silk_LSHIFT_SAT32(gain, 16); /* Q16 */
}
}
/* Save unquantized gains and gain Index */
Array.Copy(psEncCtrl.Gains_Q16, psEncCtrl.GainsUnq_Q16, psEnc.nb_subfr);
psEncCtrl.lastGainIndexPrev = psShapeSt.LastGainIndex;
/* Quantize gains */
BoxedValueSbyte boxed_lastGainIndex = new BoxedValueSbyte(psShapeSt.LastGainIndex);
GainQuantization.silk_gains_quant(psEnc.indices.GainsIndices, psEncCtrl.Gains_Q16,
boxed_lastGainIndex, condCoding == SilkConstants.CODE_CONDITIONALLY ? 1 : 0, psEnc.nb_subfr);
psShapeSt.LastGainIndex = boxed_lastGainIndex.Val;
/* Set quantizer offset for voiced signals. Larger offset when LTP coding gain is low or tilt is high (ie low-pass) */
if (psEnc.indices.signalType == SilkConstants.TYPE_VOICED)
{
if (psEncCtrl.LTPredCodGain_Q7 + Inlines.silk_RSHIFT(psEnc.input_tilt_Q15, 8) > ((int)((1.0f) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 7)*/)
{
psEnc.indices.quantOffsetType = 0;
}
else
{
psEnc.indices.quantOffsetType = 1;
}
}
/* Quantizer boundary adjustment */
quant_offset_Q10 = Tables.silk_Quantization_Offsets_Q10[psEnc.indices.signalType >> 1][psEnc.indices.quantOffsetType];
psEncCtrl.Lambda_Q10 = ((int)((TuningParameters.LAMBDA_OFFSET) * ((long)1 << (10)) + 0.5))/*Inlines.SILK_CONST(TuningParameters.LAMBDA_OFFSET, 10)*/
+ Inlines.silk_SMULBB(((int)((TuningParameters.LAMBDA_DELAYED_DECISIONS) * ((long)1 << (10)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LAMBDA_DELAYED_DECISIONS, 10)*/, psEnc.nStatesDelayedDecision)
+ Inlines.silk_SMULWB(((int)((TuningParameters.LAMBDA_SPEECH_ACT) * ((long)1 << (18)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LAMBDA_SPEECH_ACT, 18)*/, psEnc.speech_activity_Q8)
+ Inlines.silk_SMULWB(((int)((TuningParameters.LAMBDA_INPUT_QUALITY) * ((long)1 << (12)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LAMBDA_INPUT_QUALITY, 12)*/, psEncCtrl.input_quality_Q14)
+ Inlines.silk_SMULWB(((int)((TuningParameters.LAMBDA_CODING_QUALITY) * ((long)1 << (12)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LAMBDA_CODING_QUALITY, 12)*/, psEncCtrl.coding_quality_Q14)
+ Inlines.silk_SMULWB(((int)((TuningParameters.LAMBDA_QUANT_OFFSET) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LAMBDA_QUANT_OFFSET, 16)*/, quant_offset_Q10);
Inlines.OpusAssert(psEncCtrl.Lambda_Q10 > 0);
Inlines.OpusAssert(psEncCtrl.Lambda_Q10 < ((int)((2) * ((long)1 << (10)) + 0.5)) /*Inlines.SILK_CONST(2, 10)*/);
}
/**************************************************************/
/* Compute noise shaping coefficients and initial gain values */
/**************************************************************/
internal static void silk_noise_shape_analysis(
SilkChannelEncoder psEnc, /* I/O Encoder state FIX */
SilkEncoderControl psEncCtrl, /* I/O Encoder control FIX */
short[] pitch_res, /* I LPC residual from pitch analysis */
int pitch_res_ptr,
short[] x, /* I Input signal [ frame_length + la_shape ] */
int x_ptr
)
{
SilkShapeState psShapeSt = psEnc.sShape;
int k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0;
int SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32;
int nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7;
int delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8;
int[] auto_corr = new int[SilkConstants.MAX_SHAPE_LPC_ORDER + 1];
int[] refl_coef_Q16 = new int[SilkConstants.MAX_SHAPE_LPC_ORDER];
int[] AR1_Q24 = new int[SilkConstants.MAX_SHAPE_LPC_ORDER];
int[] AR2_Q24 = new int[SilkConstants.MAX_SHAPE_LPC_ORDER];
short[] x_windowed;
int pitch_res_ptr2;
int x_ptr2;
/* Point to start of first LPC analysis block */
x_ptr2 = x_ptr - psEnc.la_shape;
/****************/
/* GAIN CONTROL */
/****************/
SNR_adj_dB_Q7 = psEnc.SNR_dB_Q7;
/* Input quality is the average of the quality in the lowest two VAD bands */
psEncCtrl.input_quality_Q14 = (int)Inlines.silk_RSHIFT((int)psEnc.input_quality_bands_Q15[0]
+ psEnc.input_quality_bands_Q15[1], 2);
/* Coding quality level, between 0.0_Q0 and 1.0_Q0, but in Q14 */
psEncCtrl.coding_quality_Q14 = Inlines.silk_RSHIFT(Sigmoid.silk_sigm_Q15(Inlines.silk_RSHIFT_ROUND(SNR_adj_dB_Q7 -
((int)((20.0f) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(20.0f, 7)*/, 4)), 1);
/* Reduce coding SNR during low speech activity */
if (psEnc.useCBR == 0)
{
b_Q8 = ((int)((1.0f) * ((long)1 << (8)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 8)*/ - psEnc.speech_activity_Q8;
b_Q8 = Inlines.silk_SMULWB(Inlines.silk_LSHIFT(b_Q8, 8), b_Q8);
SNR_adj_dB_Q7 = Inlines.silk_SMLAWB(SNR_adj_dB_Q7,
Inlines.silk_SMULBB(((int)((0 - TuningParameters.BG_SNR_DECR_dB) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(0 - TuningParameters.BG_SNR_DECR_dB, 7)*/ >> (4 + 1), b_Q8), /* Q11*/
Inlines.silk_SMULWB(((int)((1.0f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 14)*/ + psEncCtrl.input_quality_Q14, psEncCtrl.coding_quality_Q14)); /* Q12*/
}
if (psEnc.indices.signalType == SilkConstants.TYPE_VOICED)
{
/* Reduce gains for periodic signals */
SNR_adj_dB_Q7 = Inlines.silk_SMLAWB(SNR_adj_dB_Q7, ((int)((TuningParameters.HARM_SNR_INCR_dB) * ((long)1 << (8)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HARM_SNR_INCR_dB, 8)*/, psEnc.LTPCorr_Q15);
}
else
{
/* For unvoiced signals and low-quality input, adjust the quality slower than SNR_dB setting */
SNR_adj_dB_Q7 = Inlines.silk_SMLAWB(SNR_adj_dB_Q7,
Inlines.silk_SMLAWB(((int)((6.0f) * ((long)1 << (9)) + 0.5)) /*Inlines.SILK_CONST(6.0f, 9)*/, -((int)((0.4f) * ((long)1 << (18)) + 0.5)) /*Inlines.SILK_CONST(0.4f, 18)*/, psEnc.SNR_dB_Q7),
((int)((1.0f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 14)*/ - psEncCtrl.input_quality_Q14);
}
/*************************/
/* SPARSENESS PROCESSING */
/*************************/
/* Set quantizer offset */
if (psEnc.indices.signalType == SilkConstants.TYPE_VOICED)
{
/* Initially set to 0; may be overruled in process_gains(..) */
psEnc.indices.quantOffsetType = 0;
psEncCtrl.sparseness_Q8 = 0;
}
else
{
/* Sparseness measure, based on relative fluctuations of energy per 2 milliseconds */
nSamples = Inlines.silk_LSHIFT(psEnc.fs_kHz, 1);
energy_variation_Q7 = 0;
log_energy_prev_Q7 = 0;
pitch_res_ptr2 = pitch_res_ptr;
for (k = 0; k < Inlines.silk_SMULBB(SilkConstants.SUB_FRAME_LENGTH_MS, psEnc.nb_subfr) / 2; k++)
{
SumSqrShift.silk_sum_sqr_shift(out nrg, out scale, pitch_res, pitch_res_ptr2, nSamples);
nrg += Inlines.silk_RSHIFT(nSamples, scale); /* Q(-scale)*/
log_energy_Q7 = Inlines.silk_lin2log(nrg);
if (k > 0)
{
energy_variation_Q7 += Inlines.silk_abs(log_energy_Q7 - log_energy_prev_Q7);
}
log_energy_prev_Q7 = log_energy_Q7;
pitch_res_ptr2 += nSamples;
}
psEncCtrl.sparseness_Q8 = Inlines.silk_RSHIFT(Sigmoid.silk_sigm_Q15(Inlines.silk_SMULWB(energy_variation_Q7 -
((int)((5.0f) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(5.0f, 7)*/, ((int)((0.1f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(0.1f, 16)*/)), 7);
/* Set quantization offset depending on sparseness measure */
if (psEncCtrl.sparseness_Q8 > ((int)((TuningParameters.SPARSENESS_THRESHOLD_QNT_OFFSET) * ((long)1 << (8)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.SPARSENESS_THRESHOLD_QNT_OFFSET, 8)*/)
{
psEnc.indices.quantOffsetType = 0;
}
else
{
psEnc.indices.quantOffsetType = 1;
}
/* Increase coding SNR for sparse signals */
SNR_adj_dB_Q7 = Inlines.silk_SMLAWB(SNR_adj_dB_Q7, ((int)((TuningParameters.SPARSE_SNR_INCR_dB) * ((long)1 << (15)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.SPARSE_SNR_INCR_dB, 15)*/, psEncCtrl.sparseness_Q8 - ((int)((0.5f) * ((long)1 << (8)) + 0.5)) /*Inlines.SILK_CONST(0.5f, 8)*/);
}
/*******************************/
/* Control bandwidth expansion */
/*******************************/
/* More BWE for signals with high prediction gain */
strength_Q16 = Inlines.silk_SMULWB(psEncCtrl.predGain_Q16, ((int)((TuningParameters.FIND_PITCH_WHITE_NOISE_FRACTION) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.FIND_PITCH_WHITE_NOISE_FRACTION, 16)*/);
BWExp1_Q16 = BWExp2_Q16 = Inlines.silk_DIV32_varQ(((int)((TuningParameters.BANDWIDTH_EXPANSION) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.BANDWIDTH_EXPANSION, 16)*/,
Inlines.silk_SMLAWW(((int)((1.0f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 16)*/, strength_Q16, strength_Q16), 16);
delta_Q16 = Inlines.silk_SMULWB(((int)((1.0f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 16)*/ - Inlines.silk_SMULBB(3, psEncCtrl.coding_quality_Q14),
((int)((TuningParameters.LOW_RATE_BANDWIDTH_EXPANSION_DELTA) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LOW_RATE_BANDWIDTH_EXPANSION_DELTA, 16)*/);
BWExp1_Q16 = Inlines.silk_SUB32(BWExp1_Q16, delta_Q16);
BWExp2_Q16 = Inlines.silk_ADD32(BWExp2_Q16, delta_Q16);
/* BWExp1 will be applied after BWExp2, so make it relative */
BWExp1_Q16 = Inlines.silk_DIV32_16(Inlines.silk_LSHIFT(BWExp1_Q16, 14), Inlines.silk_RSHIFT(BWExp2_Q16, 2));
if (psEnc.warping_Q16 > 0)
{
/* Slightly more warping in analysis will move quantization noise up in frequency, where it's better masked */
warping_Q16 = Inlines.silk_SMLAWB(psEnc.warping_Q16, (int)psEncCtrl.coding_quality_Q14, ((int)((0.01f) * ((long)1 << (18)) + 0.5)) /*Inlines.SILK_CONST(0.01f, 18)*/);
}
else
{
warping_Q16 = 0;
}
/********************************************/
/* Compute noise shaping AR coefs and gains */
/********************************************/
x_windowed = new short[psEnc.shapeWinLength];
for (k = 0; k < psEnc.nb_subfr; k++)
{
/* Apply window: sine slope followed by flat part followed by cosine slope */
int shift, slope_part, flat_part;
flat_part = psEnc.fs_kHz * 3;
slope_part = Inlines.silk_RSHIFT(psEnc.shapeWinLength - flat_part, 1);
ApplySineWindow.silk_apply_sine_window(x_windowed, 0, x, x_ptr2, 1, slope_part);
shift = slope_part;
Array.Copy(x, x_ptr2 + shift, x_windowed, shift, flat_part);
shift += flat_part;
ApplySineWindow.silk_apply_sine_window(x_windowed, shift, x, x_ptr2 + shift, 2, slope_part);
/* Update pointer: next LPC analysis block */
x_ptr2 += psEnc.subfr_length;
BoxedValueInt scale_boxed = new BoxedValueInt(scale);
if (psEnc.warping_Q16 > 0)
{
/* Calculate warped auto correlation */
Autocorrelation.silk_warped_autocorrelation(auto_corr, scale_boxed, x_windowed, warping_Q16, psEnc.shapeWinLength, psEnc.shapingLPCOrder);
}
else
{
/* Calculate regular auto correlation */
Autocorrelation.silk_autocorr(auto_corr, scale_boxed, x_windowed, psEnc.shapeWinLength, psEnc.shapingLPCOrder + 1);
}
scale = scale_boxed.Val;
/* Add white noise, as a fraction of energy */
auto_corr[0] = Inlines.silk_ADD32(auto_corr[0], Inlines.silk_max_32(Inlines.silk_SMULWB(Inlines.silk_RSHIFT(auto_corr[0], 4),
((int)((TuningParameters.SHAPE_WHITE_NOISE_FRACTION) * ((long)1 << (20)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.SHAPE_WHITE_NOISE_FRACTION, 20)*/), 1));
/* Calculate the reflection coefficients using schur */
nrg = Schur.silk_schur64(refl_coef_Q16, auto_corr, psEnc.shapingLPCOrder);
Inlines.OpusAssert(nrg >= 0);
/* Convert reflection coefficients to prediction coefficients */
K2A.silk_k2a_Q16(AR2_Q24, refl_coef_Q16, psEnc.shapingLPCOrder);
Qnrg = -scale; /* range: -12...30*/
Inlines.OpusAssert(Qnrg >= -12);
Inlines.OpusAssert(Qnrg <= 30);
/* Make sure that Qnrg is an even number */
if ((Qnrg & 1) != 0)
{
Qnrg -= 1;
nrg >>= 1;
}
tmp32 = Inlines.silk_SQRT_APPROX(nrg);
Qnrg >>= 1; /* range: -6...15*/
psEncCtrl.Gains_Q16[k] = Inlines.silk_LSHIFT_SAT32(tmp32, 16 - Qnrg);
if (psEnc.warping_Q16 > 0)
{
/* Adjust gain for warping */
gain_mult_Q16 = warped_gain(AR2_Q24, warping_Q16, psEnc.shapingLPCOrder);
Inlines.OpusAssert(psEncCtrl.Gains_Q16[k] >= 0);
if (Inlines.silk_SMULWW(Inlines.silk_RSHIFT_ROUND(psEncCtrl.Gains_Q16[k], 1), gain_mult_Q16) >= (int.MaxValue >> 1))
{
psEncCtrl.Gains_Q16[k] = int.MaxValue;
}
else
{
psEncCtrl.Gains_Q16[k] = Inlines.silk_SMULWW(psEncCtrl.Gains_Q16[k], gain_mult_Q16);
}
}
/* Bandwidth expansion for synthesis filter shaping */
BWExpander.silk_bwexpander_32(AR2_Q24, psEnc.shapingLPCOrder, BWExp2_Q16);
/* Compute noise shaping filter coefficients */
Array.Copy(AR2_Q24, AR1_Q24, psEnc.shapingLPCOrder);
/* Bandwidth expansion for analysis filter shaping */
Inlines.OpusAssert(BWExp1_Q16 <= ((int)((1.0f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 16)*/);
BWExpander.silk_bwexpander_32(AR1_Q24, psEnc.shapingLPCOrder, BWExp1_Q16);
/* Ratio of prediction gains, in energy domain */
pre_nrg_Q30 = LPCInversePredGain.silk_LPC_inverse_pred_gain_Q24(AR2_Q24, psEnc.shapingLPCOrder);
nrg = LPCInversePredGain.silk_LPC_inverse_pred_gain_Q24(AR1_Q24, psEnc.shapingLPCOrder);
/*psEncCtrl.GainsPre[ k ] = 1.0f - 0.7f * ( 1.0f - pre_nrg / nrg ) = 0.3f + 0.7f * pre_nrg / nrg;*/
pre_nrg_Q30 = Inlines.silk_LSHIFT32(Inlines.silk_SMULWB(pre_nrg_Q30, ((int)((0.7f) * ((long)1 << (15)) + 0.5)) /*Inlines.SILK_CONST(0.7f, 15)*/), 1);
psEncCtrl.GainsPre_Q14[k] = (int)((int)((0.3f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(0.3f, 14)*/ + Inlines.silk_DIV32_varQ(pre_nrg_Q30, nrg, 14);
/* Convert to monic warped prediction coefficients and limit absolute values */
limit_warped_coefs(AR2_Q24, AR1_Q24, warping_Q16, ((int)((3.999f) * ((long)1 << (24)) + 0.5)) /*Inlines.SILK_CONST(3.999f, 24)*/, psEnc.shapingLPCOrder);
/* Convert from Q24 to Q13 and store in int16 */
for (i = 0; i < psEnc.shapingLPCOrder; i++)
{
psEncCtrl.AR1_Q13[k * SilkConstants.MAX_SHAPE_LPC_ORDER + i] = (short)Inlines.silk_SAT16(Inlines.silk_RSHIFT_ROUND(AR1_Q24[i], 11));
psEncCtrl.AR2_Q13[k * SilkConstants.MAX_SHAPE_LPC_ORDER + i] = (short)Inlines.silk_SAT16(Inlines.silk_RSHIFT_ROUND(AR2_Q24[i], 11));
}
}
/*****************/
/* Gain tweaking */
/*****************/
/* Increase gains during low speech activity and put lower limit on gains */
gain_mult_Q16 = Inlines.silk_log2lin(-Inlines.silk_SMLAWB(-((int)((16.0f) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(16.0f, 7)*/, SNR_adj_dB_Q7, ((int)((0.16f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(0.16f, 16)*/));
gain_add_Q16 = Inlines.silk_log2lin(Inlines.silk_SMLAWB(((int)((16.0f) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(16.0f, 7)*/, ((int)((SilkConstants.MIN_QGAIN_DB) * ((long)1 << (7)) + 0.5)) /*Inlines.SILK_CONST(SilkConstants.MIN_QGAIN_DB, 7)*/, ((int)((0.16f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(0.16f, 16)*/));
Inlines.OpusAssert(gain_mult_Q16 > 0);
for (k = 0; k < psEnc.nb_subfr; k++)
{
psEncCtrl.Gains_Q16[k] = Inlines.silk_SMULWW(psEncCtrl.Gains_Q16[k], gain_mult_Q16);
Inlines.OpusAssert(psEncCtrl.Gains_Q16[k] >= 0);
psEncCtrl.Gains_Q16[k] = Inlines.silk_ADD_POS_SAT32(psEncCtrl.Gains_Q16[k], gain_add_Q16);
}
gain_mult_Q16 = ((int)((1.0f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 16)*/ + Inlines.silk_RSHIFT_ROUND(Inlines.silk_MLA(((int)((TuningParameters.INPUT_TILT) * ((long)1 << (26)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.INPUT_TILT, 26)*/,
psEncCtrl.coding_quality_Q14, ((int)((TuningParameters.HIGH_RATE_INPUT_TILT) * ((long)1 << (12)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HIGH_RATE_INPUT_TILT, 12)*/), 10);
for (k = 0; k < psEnc.nb_subfr; k++)
{
psEncCtrl.GainsPre_Q14[k] = Inlines.silk_SMULWB(gain_mult_Q16, psEncCtrl.GainsPre_Q14[k]);
}
/************************************************/
/* Control low-frequency shaping and noise tilt */
/************************************************/
/* Less low frequency shaping for noisy inputs */
strength_Q16 = Inlines.silk_MUL(((int)((TuningParameters.LOW_FREQ_SHAPING) * ((long)1 << (4)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LOW_FREQ_SHAPING, 4)*/, Inlines.silk_SMLAWB(((int)((1.0f) * ((long)1 << (12)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 12)*/,
((int)((TuningParameters.LOW_QUALITY_LOW_FREQ_SHAPING_DECR) * ((long)1 << (13)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LOW_QUALITY_LOW_FREQ_SHAPING_DECR, 13)*/, psEnc.input_quality_bands_Q15[0] - ((int)((1.0f) * ((long)1 << (15)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 15)*/));
strength_Q16 = Inlines.silk_RSHIFT(Inlines.silk_MUL(strength_Q16, psEnc.speech_activity_Q8), 8);
if (psEnc.indices.signalType == SilkConstants.TYPE_VOICED)
{
/* Reduce low frequencies quantization noise for periodic signals, depending on pitch lag */
/*f = 400; freqz([1, -0.98 + 2e-4 * f], [1, -0.97 + 7e-4 * f], 2^12, Fs); axis([0, 1000, -10, 1])*/
int fs_kHz_inv = Inlines.silk_DIV32_16(((int)((0.2f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(0.2f, 14)*/, psEnc.fs_kHz);
for (k = 0; k < psEnc.nb_subfr; k++)
{
b_Q14 = fs_kHz_inv + Inlines.silk_DIV32_16(((int)((3.0f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(3.0f, 14)*/, psEncCtrl.pitchL[k]);
/* Pack two coefficients in one int32 */
psEncCtrl.LF_shp_Q14[k] = Inlines.silk_LSHIFT(((int)((1.0f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 14)*/ - b_Q14 - Inlines.silk_SMULWB(strength_Q16, b_Q14), 16);
psEncCtrl.LF_shp_Q14[k] |= (b_Q14 - ((int)((1.0f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 14)*/) & 0xFFFF; // opus bug: again, cast to ushort was done here where bitwise masking was intended
}
Inlines.OpusAssert(((int)((TuningParameters.HARM_HP_NOISE_COEF) * ((long)1 << (24)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HARM_HP_NOISE_COEF, 24)*/ < ((int)((0.5f) * ((long)1 << (24)) + 0.5)) /*Inlines.SILK_CONST(0.5f, 24)*/); /* Guarantees that second argument to SMULWB() is within range of an short*/
Tilt_Q16 = -((int)((TuningParameters.HP_NOISE_COEF) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HP_NOISE_COEF, 16)*/ -
Inlines.silk_SMULWB(((int)((1.0f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 16)*/ - ((int)((TuningParameters.HP_NOISE_COEF) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HP_NOISE_COEF, 16)*/,
Inlines.silk_SMULWB(((int)((TuningParameters.HARM_HP_NOISE_COEF) * ((long)1 << (24)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HARM_HP_NOISE_COEF, 24)*/, psEnc.speech_activity_Q8));
}
else
{
b_Q14 = Inlines.silk_DIV32_16(21299, psEnc.fs_kHz); /* 1.3_Q0 = 21299_Q14*/
/* Pack two coefficients in one int32 */
psEncCtrl.LF_shp_Q14[0] = Inlines.silk_LSHIFT(((int)((1.0f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 14)*/ - b_Q14 -
Inlines.silk_SMULWB(strength_Q16, Inlines.silk_SMULWB(((int)((0.6f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(0.6f, 16)*/, b_Q14)), 16);
psEncCtrl.LF_shp_Q14[0] |= (b_Q14 - ((int)((1.0f) * ((long)1 << (14)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 14)*/) & 0xFFFF; // opus bug: cast to ushort is better expressed as a bitwise operator, otherwise runtime analysis might flag it as an overflow error
for (k = 1; k < psEnc.nb_subfr; k++)
{
psEncCtrl.LF_shp_Q14[k] = psEncCtrl.LF_shp_Q14[0];
}
Tilt_Q16 = -((int)((TuningParameters.HP_NOISE_COEF) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HP_NOISE_COEF, 16)*/;
}
/****************************/
/* HARMONIC SHAPING CONTROL */
/****************************/
/* Control boosting of harmonic frequencies */
HarmBoost_Q16 = Inlines.silk_SMULWB(Inlines.silk_SMULWB(((int)((1.0f) * ((long)1 << (17)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 17)*/ - Inlines.silk_LSHIFT(psEncCtrl.coding_quality_Q14, 3),
psEnc.LTPCorr_Q15), ((int)((TuningParameters.LOW_RATE_HARMONIC_BOOST) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LOW_RATE_HARMONIC_BOOST, 16)*/);
/* More harmonic boost for noisy input signals */
HarmBoost_Q16 = Inlines.silk_SMLAWB(HarmBoost_Q16,
((int)((1.0f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 16)*/ - Inlines.silk_LSHIFT(psEncCtrl.input_quality_Q14, 2), ((int)((TuningParameters.LOW_INPUT_QUALITY_HARMONIC_BOOST) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.LOW_INPUT_QUALITY_HARMONIC_BOOST, 16)*/);
if (SilkConstants.USE_HARM_SHAPING != 0 && psEnc.indices.signalType == SilkConstants.TYPE_VOICED)
{
/* More harmonic noise shaping for high bitrates or noisy input */
HarmShapeGain_Q16 = Inlines.silk_SMLAWB(((int)((TuningParameters.HARMONIC_SHAPING) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HARMONIC_SHAPING, 16)*/,
((int)((1.0f) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 16)*/ - Inlines.silk_SMULWB(((int)((1.0f) * ((long)1 << (18)) + 0.5)) /*Inlines.SILK_CONST(1.0f, 18)*/ - Inlines.silk_LSHIFT(psEncCtrl.coding_quality_Q14, 4),
psEncCtrl.input_quality_Q14), ((int)((TuningParameters.HIGH_RATE_OR_LOW_QUALITY_HARMONIC_SHAPING) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.HIGH_RATE_OR_LOW_QUALITY_HARMONIC_SHAPING, 16)*/);
/* Less harmonic noise shaping for less periodic signals */
HarmShapeGain_Q16 = Inlines.silk_SMULWB(Inlines.silk_LSHIFT(HarmShapeGain_Q16, 1),
Inlines.silk_SQRT_APPROX(Inlines.silk_LSHIFT(psEnc.LTPCorr_Q15, 15)));
}
else
{
HarmShapeGain_Q16 = 0;
}
/*************************/
/* Smooth over subframes */
/*************************/
for (k = 0; k < SilkConstants.MAX_NB_SUBFR; k++)
{
psShapeSt.HarmBoost_smth_Q16 =
Inlines.silk_SMLAWB(psShapeSt.HarmBoost_smth_Q16, HarmBoost_Q16 - psShapeSt.HarmBoost_smth_Q16, ((int)((TuningParameters.SUBFR_SMTH_COEF) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.SUBFR_SMTH_COEF, 16)*/);
psShapeSt.HarmShapeGain_smth_Q16 =
Inlines.silk_SMLAWB(psShapeSt.HarmShapeGain_smth_Q16, HarmShapeGain_Q16 - psShapeSt.HarmShapeGain_smth_Q16, ((int)((TuningParameters.SUBFR_SMTH_COEF) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.SUBFR_SMTH_COEF, 16)*/);
psShapeSt.Tilt_smth_Q16 =
Inlines.silk_SMLAWB(psShapeSt.Tilt_smth_Q16, Tilt_Q16 - psShapeSt.Tilt_smth_Q16, ((int)((TuningParameters.SUBFR_SMTH_COEF) * ((long)1 << (16)) + 0.5)) /*Inlines.SILK_CONST(TuningParameters.SUBFR_SMTH_COEF, 16)*/);
psEncCtrl.HarmBoost_Q14[k] = (int)Inlines.silk_RSHIFT_ROUND(psShapeSt.HarmBoost_smth_Q16, 2);
psEncCtrl.HarmShapeGain_Q14[k] = (int)Inlines.silk_RSHIFT_ROUND(psShapeSt.HarmShapeGain_smth_Q16, 2);
psEncCtrl.Tilt_Q14[k] = (int)Inlines.silk_RSHIFT_ROUND(psShapeSt.Tilt_smth_Q16, 2);
}
}
