static inline opus_int silk_setup_LBRR(
silk_encoder_state *psEncC, /* I/O */
const opus_int32 TargetRate_bps /* I */
)
{
opus_int ret = SILK_NO_ERROR;
opus_int32 LBRR_rate_thres_bps;
psEncC->LBRR_enabled = 0;
if( psEncC->useInBandFEC && psEncC->PacketLoss_perc > 0 ) {
if( psEncC->fs_kHz == 8 ) {
LBRR_rate_thres_bps = LBRR_NB_MIN_RATE_BPS;
} else if( psEncC->fs_kHz == 12 ) {
LBRR_rate_thres_bps = LBRR_MB_MIN_RATE_BPS;
} else {
LBRR_rate_thres_bps = LBRR_WB_MIN_RATE_BPS;
}
LBRR_rate_thres_bps = silk_SMULWB( silk_MUL( LBRR_rate_thres_bps, 125 - silk_min( psEncC->PacketLoss_perc, 25 ) ), SILK_FIX_CONST( 0.01, 16 ) );
if( TargetRate_bps > LBRR_rate_thres_bps ) {
/* Set gain increase for coding LBRR excitation */
psEncC->LBRR_enabled = 1;
psEncC->LBRR_GainIncreases = silk_max_int( 7 - silk_SMULWB( psEncC->PacketLoss_perc, SILK_FIX_CONST( 0.4, 16 ) ), 2 );
}
}
return ret;
}
static OPUS_INLINE opus_int silk_setup_LBRR(
silk_encoder_state *psEncC, /* I/O */
const opus_int32 TargetRate_bps /* I */
)
{
opus_int LBRR_in_previous_packet, ret = SILK_NO_ERROR;
opus_int32 LBRR_rate_thres_bps;
LBRR_in_previous_packet = psEncC->LBRR_enabled;
psEncC->LBRR_enabled = 0;
if( psEncC->useInBandFEC && psEncC->PacketLoss_perc > 0 ) {
if( psEncC->fs_kHz == 8 ) {
LBRR_rate_thres_bps = LBRR_NB_MIN_RATE_BPS;
} else if( psEncC->fs_kHz == 12 ) {
LBRR_rate_thres_bps = LBRR_MB_MIN_RATE_BPS;
} else {
LBRR_rate_thres_bps = LBRR_WB_MIN_RATE_BPS;
}
LBRR_rate_thres_bps = silk_SMULWB( silk_MUL( LBRR_rate_thres_bps, 125 - silk_min( psEncC->PacketLoss_perc, 25 ) ), SILK_FIX_CONST( 0.01, 16 ) );
if( TargetRate_bps > LBRR_rate_thres_bps ) {
/* Set gain increase for coding LBRR excitation */
if( LBRR_in_previous_packet == 0 ) {
/* Previous packet did not have LBRR, and was therefore coded at a higher bitrate */
psEncC->LBRR_GainIncreases = 7;
} else {
psEncC->LBRR_GainIncreases = silk_max_int( 7 - silk_SMULWB( (opus_int32)psEncC->PacketLoss_perc, SILK_FIX_CONST( 0.4, 16 ) ), 2 );
}
psEncC->LBRR_enabled = 1;
}
}
return ret;
}
/* Find least-squares prediction gain for one signal based on another and quantize it */
int32_t silk_stereo_find_predictor( /* O Returns predictor in Q13 */
int32_t * ratio_Q14, /* O Ratio of residual and mid energies */
const int16_t x[], /* I Basis signal */
const int16_t y[], /* I Target signal */
int32_t mid_res_amp_Q0[], /* I/O Smoothed mid, residual norms */
int length, /* I Number of samples */
int smooth_coef_Q16 /* I Smoothing coefficient */
)
{
int scale, scale1, scale2;
int32_t nrgx, nrgy, corr, pred_Q13, pred2_Q10;
/* Find predictor */
silk_sum_sqr_shift(&nrgx, &scale1, x, length);
silk_sum_sqr_shift(&nrgy, &scale2, y, length);
scale = silk_max_int(scale1, scale2);
scale = scale + (scale & 1); /* make even */
nrgy = silk_RSHIFT32(nrgy, scale - scale2);
nrgx = silk_RSHIFT32(nrgx, scale - scale1);
nrgx = silk_max_int(nrgx, 1);
corr = silk_inner_prod_aligned_scale(x, y, scale, length);
pred_Q13 = silk_DIV32_varQ(corr, nrgx, 13);
pred_Q13 = silk_LIMIT(pred_Q13, -(1 << 14), 1 << 14);
pred2_Q10 = silk_SMULWB(pred_Q13, pred_Q13);
/* Faster update for signals with large prediction parameters */
smooth_coef_Q16 =
(int) silk_max_int(smooth_coef_Q16, silk_abs(pred2_Q10));
/* Smoothed mid and residual norms */
assert(smooth_coef_Q16 < 32768);
scale = silk_RSHIFT(scale, 1);
mid_res_amp_Q0[0] =
silk_SMLAWB(mid_res_amp_Q0[0],
silk_LSHIFT(silk_SQRT_APPROX(nrgx),
scale) - mid_res_amp_Q0[0],
smooth_coef_Q16);
/* Residual energy = nrgy - 2 * pred * corr + pred^2 * nrgx */
nrgy = silk_SUB_LSHIFT32(nrgy, silk_SMULWB(corr, pred_Q13), 3 + 1);
nrgy = silk_ADD_LSHIFT32(nrgy, silk_SMULWB(nrgx, pred2_Q10), 6);
mid_res_amp_Q0[1] =
silk_SMLAWB(mid_res_amp_Q0[1],
silk_LSHIFT(silk_SQRT_APPROX(nrgy),
scale) - mid_res_amp_Q0[1],
smooth_coef_Q16);
/* Ratio of smoothed residual and mid norms */
*ratio_Q14 =
silk_DIV32_varQ(mid_res_amp_Q0[1], silk_max(mid_res_amp_Q0[0], 1),
14);
*ratio_Q14 = silk_LIMIT(*ratio_Q14, 0, 32767);
return pred_Q13;
}
/* Gain scalar quantization with hysteresis, uniform on log scale */
void silk_gains_quant(
opus_int8 ind[ MAX_NB_SUBFR ], /* O gain indices */
opus_int32 gain_Q16[ MAX_NB_SUBFR ], /* I/O gains (quantized out) */
opus_int8 *prev_ind, /* I/O last index in previous frame */
const opus_int conditional, /* I first gain is delta coded if 1 */
const opus_int nb_subfr /* I number of subframes */
)
{
opus_int k, double_step_size_threshold;
for( k = 0; k < nb_subfr; k++ ) {
/* Convert to log scale, scale, floor() */
ind[ k ] = silk_SMULWB( SCALE_Q16, silk_lin2log( gain_Q16[ k ] ) - OFFSET );
/* Round towards previous quantized gain (hysteresis) */
if( ind[ k ] < *prev_ind ) {
ind[ k ]++;
}
ind[ k ] = silk_LIMIT_int( ind[ k ], 0, N_LEVELS_QGAIN - 1 );
/* Compute delta indices and limit */
if( k == 0 && conditional == 0 ) {
/* Full index */
ind[ k ] = silk_LIMIT_int( ind[ k ], *prev_ind + MIN_DELTA_GAIN_QUANT, N_LEVELS_QGAIN - 1 );
*prev_ind = ind[ k ];
} else {
/* Delta index */
ind[ k ] = ind[ k ] - *prev_ind;
/* Double the quantization step size for large gain increases, so that the max gain level can be reached */
double_step_size_threshold = 2 * MAX_DELTA_GAIN_QUANT - N_LEVELS_QGAIN + *prev_ind;
if( ind[ k ] > double_step_size_threshold ) {
ind[ k ] = double_step_size_threshold + silk_RSHIFT( ind[ k ] - double_step_size_threshold + 1, 1 );
}
ind[ k ] = silk_LIMIT_int( ind[ k ], MIN_DELTA_GAIN_QUANT, MAX_DELTA_GAIN_QUANT );
/* Accumulate deltas */
if( ind[ k ] > double_step_size_threshold ) {
*prev_ind += silk_LSHIFT( ind[ k ], 1 ) - double_step_size_threshold;
*prev_ind = silk_min_int( *prev_ind, N_LEVELS_QGAIN - 1 );
} else {
*prev_ind += ind[ k ];
}
/* Shift to make non-negative */
ind[ k ] -= MIN_DELTA_GAIN_QUANT;
}
/* Scale and convert to linear scale */
gain_Q16[ k ] = silk_log2lin( silk_min_32( silk_SMULWB( INV_SCALE_Q16, *prev_ind ) + OFFSET, 3967 ) ); /* 3967 = 31 in Q7 */
}
}
/* Gains scalar dequantization, uniform on log scale */
void silk_gains_dequant(
opus_int32 gain_Q16[ MAX_NB_SUBFR ], /* O quantized gains */
const opus_int8 ind[ MAX_NB_SUBFR ], /* I gain indices */
opus_int8 *prev_ind, /* I/O last index in previous frame */
const opus_int conditional, /* I first gain is delta coded if 1 */
const opus_int nb_subfr /* I number of subframes */
)
{
opus_int k, ind_tmp, double_step_size_threshold;
for( k = 0; k < nb_subfr; k++ ) {
if( k == 0 && conditional == 0 ) {
/* Gain index is not allowed to go down more than 16 steps (~21.8 dB) */
*prev_ind = silk_max_int( ind[ k ], *prev_ind - 16 );
} else {
/* Delta index */
ind_tmp = ind[ k ] + MIN_DELTA_GAIN_QUANT;
/* Accumulate deltas */
double_step_size_threshold = 2 * MAX_DELTA_GAIN_QUANT - N_LEVELS_QGAIN + *prev_ind;
if( ind_tmp > double_step_size_threshold ) {
*prev_ind += silk_LSHIFT( ind_tmp, 1 ) - double_step_size_threshold;
} else {
*prev_ind += ind_tmp;
}
}
*prev_ind = silk_LIMIT_int( *prev_ind, 0, N_LEVELS_QGAIN - 1 );
/* Scale and convert to linear scale */
gain_Q16[ k ] = silk_log2lin( silk_min_32( silk_SMULWB( INV_SCALE_Q16, *prev_ind ) + OFFSET, 3967 ) ); /* 3967 = 31 in Q7 */
}
}
static OPUS_INLINE opus_int16 *silk_resampler_private_IIR_FIR_INTERPOL(
opus_int16 *out,
opus_int16 *buf,
opus_int32 max_index_Q16,
opus_int32 index_increment_Q16
)
{
opus_int32 index_Q16, res_Q15;
opus_int16 *buf_ptr;
opus_int32 table_index;
/* Interpolate upsampled signal and store in output array */
for( index_Q16 = 0; index_Q16 < max_index_Q16; index_Q16 += index_increment_Q16 ) {
table_index = silk_SMULWB( index_Q16 & 0xFFFF, 12 );
buf_ptr = &buf[ index_Q16 >> 16 ];
res_Q15 = silk_SMULBB( buf_ptr[ 0 ], silk_resampler_frac_FIR_12[ table_index ][ 0 ] );
res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 1 ], silk_resampler_frac_FIR_12[ table_index ][ 1 ] );
res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 2 ], silk_resampler_frac_FIR_12[ table_index ][ 2 ] );
res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 3 ], silk_resampler_frac_FIR_12[ table_index ][ 3 ] );
res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 4 ], silk_resampler_frac_FIR_12[ 11 - table_index ][ 3 ] );
res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 5 ], silk_resampler_frac_FIR_12[ 11 - table_index ][ 2 ] );
res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 6 ], silk_resampler_frac_FIR_12[ 11 - table_index ][ 1 ] );
res_Q15 = silk_SMLABB( res_Q15, buf_ptr[ 7 ], silk_resampler_frac_FIR_12[ 11 - table_index ][ 0 ] );
*out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q15, 15 ) );
}
return out;
}
/* Decode mid/side predictors */
void silk_stereo_decode_pred(
ec_dec *psRangeDec, /* I/O Compressor data structure */
opus_int32 pred_Q13[] /* O Predictors */
)
{
opus_int n, ix[ 2 ][ 3 ];
opus_int32 low_Q13, step_Q13;
/* Entropy decoding */
n = ec_dec_icdf( psRangeDec, silk_stereo_pred_joint_iCDF, 8 );
ix[ 0 ][ 2 ] = silk_DIV32_16( n, 5 );
ix[ 1 ][ 2 ] = n - 5 * ix[ 0 ][ 2 ];
for( n = 0; n < 2; n++ ) {
ix[ n ][ 0 ] = ec_dec_icdf( psRangeDec, silk_uniform3_iCDF, 8 );
ix[ n ][ 1 ] = ec_dec_icdf( psRangeDec, silk_uniform5_iCDF, 8 );
}
/* Dequantize */
for( n = 0; n < 2; n++ ) {
ix[ n ][ 0 ] += 3 * ix[ n ][ 2 ];
low_Q13 = silk_stereo_pred_quant_Q13[ ix[ n ][ 0 ] ];
step_Q13 = silk_SMULWB( silk_stereo_pred_quant_Q13[ ix[ n ][ 0 ] + 1 ] - low_Q13,
SILK_FIX_CONST( 0.5 / STEREO_QUANT_SUB_STEPS, 16 ) );
pred_Q13[ n ] = silk_SMLABB( low_Q13, step_Q13, 2 * ix[ n ][ 1 ] + 1 );
}
/* Subtract second from first predictor (helps when actually applying these) */
pred_Q13[ 0 ] -= pred_Q13[ 1 ];
}
/* High-pass filter with cutoff frequency adaptation based on pitch lag statistics */
void silk_HP_variable_cutoff(silk_encoder_state_Fxx state_Fxx[] /* I/O Encoder states */
)
{
int quality_Q15;
int32_t pitch_freq_Hz_Q16, pitch_freq_log_Q7, delta_freq_Q7;
silk_encoder_state *psEncC1 = &state_Fxx[0].sCmn;
/* Adaptive cutoff frequency: estimate low end of pitch frequency range */
if (psEncC1->prevSignalType == TYPE_VOICED) {
/* difference, in log domain */
pitch_freq_Hz_Q16 =
silk_DIV32_16(silk_LSHIFT
(silk_MUL(psEncC1->fs_kHz, 1000), 16),
psEncC1->prevLag);
pitch_freq_log_Q7 = silk_lin2log(pitch_freq_Hz_Q16) - (16 << 7);
/* adjustment based on quality */
quality_Q15 = psEncC1->input_quality_bands_Q15[0];
pitch_freq_log_Q7 =
silk_SMLAWB(pitch_freq_log_Q7,
silk_SMULWB(silk_LSHIFT(-quality_Q15, 2),
quality_Q15),
pitch_freq_log_Q7 -
(silk_lin2log
(SILK_FIX_CONST(VARIABLE_HP_MIN_CUTOFF_HZ, 16))
- (16 << 7)));
/* delta_freq = pitch_freq_log - psEnc->variable_HP_smth1; */
delta_freq_Q7 =
pitch_freq_log_Q7 -
silk_RSHIFT(psEncC1->variable_HP_smth1_Q15, 8);
if (delta_freq_Q7 < 0) {
/* less smoothing for decreasing pitch frequency, to track something close to the minimum */
delta_freq_Q7 = silk_MUL(delta_freq_Q7, 3);
}
/* limit delta, to reduce impact of outliers in pitch estimation */
delta_freq_Q7 =
silk_LIMIT_32(delta_freq_Q7,
-SILK_FIX_CONST(VARIABLE_HP_MAX_DELTA_FREQ,
7),
SILK_FIX_CONST(VARIABLE_HP_MAX_DELTA_FREQ,
7));
/* update smoother */
psEncC1->variable_HP_smth1_Q15 =
silk_SMLAWB(psEncC1->variable_HP_smth1_Q15,
silk_SMULBB(psEncC1->speech_activity_Q8,
delta_freq_Q7),
SILK_FIX_CONST(VARIABLE_HP_SMTH_COEF1, 16));
/* limit frequency range */
psEncC1->variable_HP_smth1_Q15 =
silk_LIMIT_32(psEncC1->variable_HP_smth1_Q15,
silk_LSHIFT(silk_lin2log
(VARIABLE_HP_MIN_CUTOFF_HZ), 8),
silk_LSHIFT(silk_lin2log
(VARIABLE_HP_MAX_CUTOFF_HZ), 8));
}
}
/* Glues concealed frames with new good received frames */
void silk_PLC_glue_frames(
silk_decoder_state *psDec, /* I/O decoder state */
opus_int16 frame[], /* I/O signal */
opus_int length /* I length of signal */
)
{
opus_int i, energy_shift;
opus_int32 energy;
silk_PLC_struct *psPLC;
psPLC = &psDec->sPLC;
if( psDec->lossCnt ) {
/* Calculate energy in concealed residual */
silk_sum_sqr_shift( &psPLC->conc_energy, &psPLC->conc_energy_shift, frame, length );
psPLC->last_frame_lost = 1;
} else {
if( psDec->sPLC.last_frame_lost ) {
/* Calculate residual in decoded signal if last frame was lost */
silk_sum_sqr_shift( &energy, &energy_shift, frame, length );
/* Normalize energies */
if( energy_shift > psPLC->conc_energy_shift ) {
psPLC->conc_energy = silk_RSHIFT( psPLC->conc_energy, energy_shift - psPLC->conc_energy_shift );
} else if( energy_shift < psPLC->conc_energy_shift ) {
energy = silk_RSHIFT( energy, psPLC->conc_energy_shift - energy_shift );
}
/* Fade in the energy difference */
if( energy > psPLC->conc_energy ) {
opus_int32 frac_Q24, LZ;
opus_int32 gain_Q16, slope_Q16;
LZ = silk_CLZ32( psPLC->conc_energy );
LZ = LZ - 1;
psPLC->conc_energy = silk_LSHIFT( psPLC->conc_energy, LZ );
energy = silk_RSHIFT( energy, silk_max_32( 24 - LZ, 0 ) );
frac_Q24 = silk_DIV32( psPLC->conc_energy, silk_max( energy, 1 ) );
gain_Q16 = silk_LSHIFT( silk_SQRT_APPROX( frac_Q24 ), 4 );
slope_Q16 = silk_DIV32_16( ( (opus_int32)1 << 16 ) - gain_Q16, length );
/* Make slope 4x steeper to avoid missing onsets after DTX */
slope_Q16 = silk_LSHIFT( slope_Q16, 2 );
for( i = 0; i < length; i++ ) {
frame[ i ] = silk_SMULWB( gain_Q16, frame[ i ] );
gain_Q16 += slope_Q16;
if( gain_Q16 > (opus_int32)1 << 16 ) {
break;
}
}
}
}
psPLC->last_frame_lost = 0;
}
}
/* Prefilter for finding Quantizer input signal */
static OPUS_INLINE void silk_prefilt_FIX(
silk_prefilter_state_FIX *P, /* I/O state */
opus_int32 st_res_Q12[], /* I short term residual signal */
opus_int32 xw_Q3[], /* O prefiltered signal */
opus_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */
opus_int Tilt_Q14, /* I Tilt shaping coeficient */
opus_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients */
opus_int lag, /* I Lag for harmonic shaping */
opus_int length /* I Length of signals */
) {
opus_int i, idx, LTP_shp_buf_idx;
opus_int32 n_LTP_Q12, n_Tilt_Q10, n_LF_Q10;
opus_int32 sLF_MA_shp_Q12, sLF_AR_shp_Q12;
opus_int16 *LTP_shp_buf;
/* To speed up use temp variables instead of using the struct */
LTP_shp_buf = P->sLTP_shp;
LTP_shp_buf_idx = P->sLTP_shp_buf_idx;
sLF_AR_shp_Q12 = P->sLF_AR_shp_Q12;
sLF_MA_shp_Q12 = P->sLF_MA_shp_Q12;
for (i = 0; i < length; i++) {
if (lag > 0) {
/* unrolled loop */
silk_assert(HARM_SHAPE_FIR_TAPS == 3);
idx = lag + LTP_shp_buf_idx;
n_LTP_Q12 = silk_SMULBB(LTP_shp_buf[(idx - HARM_SHAPE_FIR_TAPS / 2 - 1) & LTP_MASK],
HarmShapeFIRPacked_Q12);
n_LTP_Q12 = silk_SMLABT(n_LTP_Q12,
LTP_shp_buf[(idx - HARM_SHAPE_FIR_TAPS / 2) & LTP_MASK],
HarmShapeFIRPacked_Q12);
n_LTP_Q12 = silk_SMLABB(n_LTP_Q12,
LTP_shp_buf[(idx - HARM_SHAPE_FIR_TAPS / 2 + 1) & LTP_MASK],
HarmShapeFIRPacked_Q12);
} else {
n_LTP_Q12 = 0;
}
n_Tilt_Q10 = silk_SMULWB(sLF_AR_shp_Q12, Tilt_Q14);
n_LF_Q10 = silk_SMLAWB(silk_SMULWT(sLF_AR_shp_Q12, LF_shp_Q14), sLF_MA_shp_Q12, LF_shp_Q14);
sLF_AR_shp_Q12 = silk_SUB32(st_res_Q12[i], silk_LSHIFT(n_Tilt_Q10, 2));
sLF_MA_shp_Q12 = silk_SUB32(sLF_AR_shp_Q12, silk_LSHIFT(n_LF_Q10, 2));
LTP_shp_buf_idx = (LTP_shp_buf_idx - 1) & LTP_MASK;
LTP_shp_buf[LTP_shp_buf_idx] = (opus_int16) silk_SAT16(
silk_RSHIFT_ROUND(sLF_MA_shp_Q12, 12));
xw_Q3[i] = silk_RSHIFT_ROUND(silk_SUB32(sLF_MA_shp_Q12, n_LTP_Q12), 9);
}
/* Copy temp variable back to state */
P->sLF_AR_shp_Q12 = sLF_AR_shp_Q12;
P->sLF_MA_shp_Q12 = sLF_MA_shp_Q12;
P->sLTP_shp_buf_idx = LTP_shp_buf_idx;
}
/* Second order ARMA filter, alternative implementation */
void silk_biquad_alt(const int16_t * in, /* I input signal */
const int32_t * B_Q28, /* I MA coefficients [3] */
const int32_t * A_Q28, /* I AR coefficients [2] */
int32_t * S, /* I/O State vector [2] */
int16_t * out, /* O output signal */
const int32_t len, /* I signal length (must be even) */
int stride /* I Operate on interleaved signal if > 1 */
)
{
/* DIRECT FORM II TRANSPOSED (uses 2 element state vector) */
int k;
int32_t inval, A0_U_Q28, A0_L_Q28, A1_U_Q28, A1_L_Q28, out32_Q14;
/* Negate A_Q28 values and split in two parts */
A0_L_Q28 = (-A_Q28[0]) & 0x00003FFF; /* lower part */
A0_U_Q28 = silk_RSHIFT(-A_Q28[0], 14); /* upper part */
A1_L_Q28 = (-A_Q28[1]) & 0x00003FFF; /* lower part */
A1_U_Q28 = silk_RSHIFT(-A_Q28[1], 14); /* upper part */
for (k = 0; k < len; k++) {
/* S[ 0 ], S[ 1 ]: Q12 */
inval = in[k * stride];
out32_Q14 = silk_LSHIFT(silk_SMLAWB(S[0], B_Q28[0], inval), 2);
S[0] =
S[1] + silk_RSHIFT_ROUND(silk_SMULWB(out32_Q14, A0_L_Q28),
14);
S[0] = silk_SMLAWB(S[0], out32_Q14, A0_U_Q28);
S[0] = silk_SMLAWB(S[0], B_Q28[1], inval);
S[1] = silk_RSHIFT_ROUND(silk_SMULWB(out32_Q14, A1_L_Q28), 14);
S[1] = silk_SMLAWB(S[1], out32_Q14, A1_U_Q28);
S[1] = silk_SMLAWB(S[1], B_Q28[2], inval);
/* Scale back to Q0 and saturate */
out[k * stride] =
(int16_t)
silk_SAT16(silk_RSHIFT(out32_Q14 + (1 << 14) - 1, 14));
}
}
/* Copy and multiply a vector by a constant */
void silk_scale_copy_vector16(
opus_int16 *data_out,
const opus_int16 *data_in,
opus_int32 gain_Q16, /* I Gain in Q16 */
const opus_int dataSize /* I Length */
)
{
opus_int i;
opus_int32 tmp32;
for(i = 0; i < dataSize; i++) {
tmp32 = silk_SMULWB(gain_Q16, data_in[ i ]);
data_out[ i ] = (opus_int16)silk_CHECK_FIT16(tmp32);
}
}
/* Processing of gains */
void silk_process_gains_FIX(
silk_encoder_state_FIX *psEnc, /* I/O Encoder state */
silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control */
opus_int condCoding /* I The type of conditional coding to use */
)
{
silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
opus_int k;
opus_int32 s_Q16, InvMaxSqrVal_Q16, gain, gain_squared, ResNrg, ResNrgPart, quant_offset_Q10;
/* Gain reduction when LTP coding gain is high */
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
/*s = -0.5f * silk_sigmoid( 0.25f * ( psEncCtrl->LTPredCodGain - 12.0f ) ); */
s_Q16 = -silk_sigm_Q15( silk_RSHIFT_ROUND( psEncCtrl->LTPredCodGain_Q7 - SILK_FIX_CONST( 12.0, 7 ), 4 ) );
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
psEncCtrl->Gains_Q16[ k ] = 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 = silk_DIV32_16( silk_log2lin(
silk_SMULWB( SILK_FIX_CONST( 21 + 16 / 0.33, 7 ) - psEnc->sCmn.SNR_dB_Q7, SILK_FIX_CONST( 0.33, 16 ) ) ), psEnc->sCmn.subfr_length );
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
/* Soft limit on ratio residual energy and squared gains */
ResNrg = psEncCtrl->ResNrg[ k ];
ResNrgPart = silk_SMULWW( ResNrg, InvMaxSqrVal_Q16 );
if( psEncCtrl->ResNrgQ[ k ] > 0 ) {
ResNrgPart = silk_RSHIFT_ROUND( ResNrgPart, psEncCtrl->ResNrgQ[ k ] );
} else {
if( ResNrgPart >= silk_RSHIFT( silk_int32_MAX, -psEncCtrl->ResNrgQ[ k ] ) ) {
ResNrgPart = silk_int32_MAX;
} else {
ResNrgPart = silk_LSHIFT( ResNrgPart, -psEncCtrl->ResNrgQ[ k ] );
}
}
gain = psEncCtrl->Gains_Q16[ k ];
gain_squared = silk_ADD_SAT32( ResNrgPart, silk_SMMUL( gain, gain ) );
if( gain_squared < silk_int16_MAX ) {
/* recalculate with higher precision */
gain_squared = silk_SMLAWW( silk_LSHIFT( ResNrgPart, 16 ), gain, gain );
silk_assert( gain_squared > 0 );
gain = silk_SQRT_APPROX( gain_squared ); /* Q8 */
gain = silk_min( gain, silk_int32_MAX >> 8 );
psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT_SAT32( gain, 8 ); /* Q16 */
} else {
void silk_noise_shape_analysis_FIX(
silk_encoder_state_FIX *psEnc, /* I/O Encoder state FIX */
silk_encoder_control_FIX *psEncCtrl, /* I/O Encoder control FIX */
const opus_int16 *pitch_res, /* I LPC residual from pitch analysis */
const opus_int16 *x, /* I Input signal [ frame_length + la_shape ] */
int arch /* I Run-time architecture */
)
{
silk_shape_state_FIX *psShapeSt = &psEnc->sShape;
opus_int k, i, nSamples, Qnrg, b_Q14, warping_Q16, scale = 0;
opus_int32 SNR_adj_dB_Q7, HarmBoost_Q16, HarmShapeGain_Q16, Tilt_Q16, tmp32;
opus_int32 nrg, pre_nrg_Q30, log_energy_Q7, log_energy_prev_Q7, energy_variation_Q7;
opus_int32 delta_Q16, BWExp1_Q16, BWExp2_Q16, gain_mult_Q16, gain_add_Q16, strength_Q16, b_Q8;
opus_int32 auto_corr[ MAX_SHAPE_LPC_ORDER + 1 ];
opus_int32 refl_coef_Q16[ MAX_SHAPE_LPC_ORDER ];
opus_int32 AR1_Q24[ MAX_SHAPE_LPC_ORDER ];
opus_int32 AR2_Q24[ MAX_SHAPE_LPC_ORDER ];
VARDECL( opus_int16, x_windowed );
const opus_int16 *x_ptr, *pitch_res_ptr;
SAVE_STACK;
/* Point to start of first LPC analysis block */
x_ptr = x - psEnc->sCmn.la_shape;
/****************/
/* GAIN CONTROL */
/****************/
SNR_adj_dB_Q7 = psEnc->sCmn.SNR_dB_Q7;
/* Input quality is the average of the quality in the lowest two VAD bands */
psEncCtrl->input_quality_Q14 = ( opus_int )silk_RSHIFT( (opus_int32)psEnc->sCmn.input_quality_bands_Q15[ 0 ]
+ psEnc->sCmn.input_quality_bands_Q15[ 1 ], 2 );
/* Coding quality level, between 0.0_Q0 and 1.0_Q0, but in Q14 */
psEncCtrl->coding_quality_Q14 = silk_RSHIFT( silk_sigm_Q15( silk_RSHIFT_ROUND( SNR_adj_dB_Q7 -
SILK_FIX_CONST( 20.0, 7 ), 4 ) ), 1 );
/* Reduce coding SNR during low speech activity */
if( psEnc->sCmn.useCBR == 0 ) {
b_Q8 = SILK_FIX_CONST( 1.0, 8 ) - psEnc->sCmn.speech_activity_Q8;
b_Q8 = silk_SMULWB( silk_LSHIFT( b_Q8, 8 ), b_Q8 );
SNR_adj_dB_Q7 = silk_SMLAWB( SNR_adj_dB_Q7,
silk_SMULBB( SILK_FIX_CONST( -BG_SNR_DECR_dB, 7 ) >> ( 4 + 1 ), b_Q8 ), /* Q11*/
silk_SMULWB( SILK_FIX_CONST( 1.0, 14 ) + psEncCtrl->input_quality_Q14, psEncCtrl->coding_quality_Q14 ) ); /* Q12*/
}
/* Second order AR filter with single delay elements */
void silk_resampler_private_AR2(int32_t S[], /* I/O State vector [ 2 ] */
int32_t out_Q8[], /* O Output signal */
const int16_t in[], /* I Input signal */
const int16_t A_Q14[], /* I AR coefficients, Q14 */
int32_t len /* I Signal length */
)
{
int32_t k;
int32_t out32;
for (k = 0; k < len; k++) {
out32 = silk_ADD_LSHIFT32(S[0], (int32_t) in[k], 8);
out_Q8[k] = out32;
out32 = silk_LSHIFT(out32, 2);
S[0] = silk_SMLAWB(S[1], out32, A_Q14[0]);
S[1] = silk_SMULWB(out32, A_Q14[1]);
}
}
/* Second order AR filter with single delay elements */
void silk_resampler_private_AR2(
opus_int32 S[], /* I/O State vector [ 2 ] */
opus_int32 out_Q8[], /* O Output signal */
const opus_int16 in[], /* I Input signal */
const opus_int16 A_Q14[], /* I AR coefficients, Q14 */
opus_int32 len /* I Signal length */
)
{
opus_int32 k;
opus_int32 out32;
for(k = 0; k < len; k++) {
out32 = silk_ADD_LSHIFT32(S[ 0 ], (opus_int32)in[ k ], 8);
out_Q8[ k ] = out32;
out32 = silk_LSHIFT(out32, 2);
S[ 0 ] = silk_SMLAWB(S[ 1 ], out32, A_Q14[ 0 ]);
S[ 1 ] = silk_SMULWB(out32, A_Q14[ 1 ]);
}
}
/* Calculation of LTP state scaling */
void silk_LTP_scale_ctrl_FIX(
silk_encoder_state_FIX *psEnc, /* I/O encoder state */
silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */
opus_int condCoding /* I The type of conditional coding to use */
)
{
opus_int round_loss;
if( condCoding == CODE_INDEPENDENTLY ) {
/* Only scale if first frame in packet */
round_loss = psEnc->sCmn.PacketLoss_perc + psEnc->sCmn.nFramesPerPacket;
psEnc->sCmn.indices.LTP_scaleIndex = (opus_int8)silk_LIMIT(
silk_SMULWB( silk_SMULBB( round_loss, psEncCtrl->LTPredCodGain_Q7 ), SILK_FIX_CONST( 0.1, 9 ) ), 0, 2 );
} else {
/* Default is minimum scaling */
psEnc->sCmn.indices.LTP_scaleIndex = 0;
}
psEncCtrl->LTP_scale_Q14 = silk_LTPScales_table_Q14[ psEnc->sCmn.indices.LTP_scaleIndex ];
}
void silk_warped_LPC_analysis_filter_FIX(
opus_int32 state[], /* I/O State [order + 1] */
opus_int16 res[], /* O Residual signal [length] */
const opus_int16 coef_Q13[], /* I Coefficients [order] */
const opus_int16 input[], /* I Input signal [length] */
const opus_int16 lambda_Q16, /* I Warping factor */
const opus_int length, /* I Length of input signal */
const opus_int order /* I Filter order (even) */
)
{
opus_int n, i;
opus_int32 acc_Q11, tmp1, tmp2;
/* Order must be even */
silk_assert( ( order & 1 ) == 0 );
for( n = 0; n < length; n++ ) {
/* Output of lowpass section */
tmp2 = silk_SMLAWB( state[ 0 ], state[ 1 ], lambda_Q16 );
state[ 0 ] = silk_LSHIFT( input[ n ], 14 );
/* Output of allpass section */
tmp1 = silk_SMLAWB( state[ 1 ], state[ 2 ] - tmp2, lambda_Q16 );
state[ 1 ] = tmp2;
acc_Q11 = silk_SMULWB( tmp2, coef_Q13[ 0 ] );
/* Loop over allpass sections */
for( i = 2; i < order; i += 2 ) {
/* Output of allpass section */
tmp2 = silk_SMLAWB( state[ i ], state[ i + 1 ] - tmp1, lambda_Q16 );
state[ i ] = tmp1;
acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ i - 1 ] );
/* Output of allpass section */
tmp1 = silk_SMLAWB( state[ i + 1 ], state[ i + 2 ] - tmp2, lambda_Q16 );
state[ i + 1 ] = tmp2;
acc_Q11 = silk_SMLAWB( acc_Q11, tmp2, coef_Q13[ i ] );
}
state[ order ] = tmp1;
acc_Q11 = silk_SMLAWB( acc_Q11, tmp1, coef_Q13[ order - 1 ] );
res[ n ] = ( opus_int16 )silk_SAT16( ( opus_int32 )input[ n ] - silk_RSHIFT_ROUND( acc_Q11, 11 ) );
}
}
/* Quantize mid/side predictors */
void silk_stereo_quant_pred(
opus_int32 pred_Q13[], /* I/O Predictors (out: quantized) */
opus_int8 ix[ 2 ][ 3 ] /* O Quantization indices */
)
{
opus_int i, j, n;
opus_int32 low_Q13, step_Q13, lvl_Q13, err_min_Q13, err_Q13, quant_pred_Q13 = 0;
/* Quantize */
for( n = 0; n < 2; n++ ) {
/* Brute-force search over quantization levels */
err_min_Q13 = silk_int32_MAX;
for( i = 0; i < STEREO_QUANT_TAB_SIZE - 1; i++ ) {
low_Q13 = silk_stereo_pred_quant_Q13[ i ];
step_Q13 = silk_SMULWB( silk_stereo_pred_quant_Q13[ i + 1 ] - low_Q13,
SILK_FIX_CONST( 0.5 / STEREO_QUANT_SUB_STEPS, 16 ) );
for( j = 0; j < STEREO_QUANT_SUB_STEPS; j++ ) {
lvl_Q13 = silk_SMLABB( low_Q13, step_Q13, 2 * j + 1 );
err_Q13 = silk_abs( pred_Q13[ n ] - lvl_Q13 );
if( err_Q13 < err_min_Q13 ) {
err_min_Q13 = err_Q13;
quant_pred_Q13 = lvl_Q13;
ix[ n ][ 0 ] = i;
ix[ n ][ 1 ] = j;
} else {
/* Error increasing, so we're past the optimum */
goto done;
}
}
}
done:
ix[ n ][ 2 ] = silk_DIV32_16( ix[ n ][ 0 ], 3 );
ix[ n ][ 0 ] -= ix[ n ][ 2 ] * 3;
pred_Q13[ n ] = quant_pred_Q13;
}
/* Subtract second from first predictor (helps when actually applying these) */
pred_Q13[ 0 ] -= pred_Q13[ 1 ];
}
/* Split signal into two decimated bands using first-order allpass filters */
void silk_ana_filt_bank_1(
const opus_int16 *in, /* I Input signal [N] */
opus_int32 *S, /* I/O State vector [2] */
opus_int16 *outL, /* O Low band [N/2] */
opus_int16 *outH, /* O High band [N/2] */
const opus_int32 N /* I Number of input samples */
)
{
opus_int k, N2 = silk_RSHIFT( N, 1 );
opus_int32 in32, X, Y, out_1, out_2;
/* Internal variables and state are in Q10 format */
for( k = 0; k < N2; k++ ) {
/* Convert to Q10 */
in32 = silk_LSHIFT( (opus_int32)in[ 2 * k ], 10 );
/* All-pass section for even input sample */
Y = silk_SUB32( in32, S[ 0 ] );
X = silk_SMLAWB( Y, Y, A_fb1_21 );
out_1 = silk_ADD32( S[ 0 ], X );
S[ 0 ] = silk_ADD32( in32, X );
/* Convert to Q10 */
in32 = silk_LSHIFT( (opus_int32)in[ 2 * k + 1 ], 10 );
/* All-pass section for odd input sample, and add to output of previous section */
Y = silk_SUB32( in32, S[ 1 ] );
X = silk_SMULWB( Y, A_fb1_20 );
out_2 = silk_ADD32( S[ 1 ], X );
S[ 1 ] = silk_ADD32( in32, X );
/* Add/subtract, convert back to int16 and store to output */
outL[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_ADD32( out_2, out_1 ), 11 ) );
outH[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SUB32( out_2, out_1 ), 11 ) );
}
}
static OPUS_INLINE void silk_PLC_conceal(
silk_decoder_state *psDec, /* I/O Decoder state */
silk_decoder_control *psDecCtrl, /* I/O Decoder control */
opus_int16 frame[], /* O LPC residual signal */
int arch /* I Run-time architecture */
)
{
opus_int i, j, k;
opus_int lag, idx, sLTP_buf_idx, shift1, shift2;
opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q30;
opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr;
opus_int32 LPC_pred_Q10, LTP_pred_Q12;
opus_int16 rand_scale_Q14;
opus_int16 *B_Q14;
opus_int32 *sLPC_Q14_ptr;
opus_int16 A_Q12[ MAX_LPC_ORDER ];
#ifdef SMALL_FOOTPRINT
opus_int16 *sLTP;
#else
VARDECL( opus_int16, sLTP );
#endif
VARDECL( opus_int32, sLTP_Q14 );
silk_PLC_struct *psPLC = &psDec->sPLC;
opus_int32 prevGain_Q10[2];
SAVE_STACK;
ALLOC( sLTP_Q14, psDec->ltp_mem_length + psDec->frame_length, opus_int32 );
#ifdef SMALL_FOOTPRINT
/* Ugly hack that breaks aliasing rules to save stack: put sLTP at the very end of sLTP_Q14. */
sLTP = ((opus_int16*)&sLTP_Q14[psDec->ltp_mem_length + psDec->frame_length])-psDec->ltp_mem_length;
#else
ALLOC( sLTP, psDec->ltp_mem_length, opus_int16 );
#endif
prevGain_Q10[0] = silk_RSHIFT( psPLC->prevGain_Q16[ 0 ], 6);
prevGain_Q10[1] = silk_RSHIFT( psPLC->prevGain_Q16[ 1 ], 6);
if( psDec->first_frame_after_reset ) {
silk_memset( psPLC->prevLPC_Q12, 0, sizeof( psPLC->prevLPC_Q12 ) );
}
silk_PLC_energy(&energy1, &shift1, &energy2, &shift2, psDec->exc_Q14, prevGain_Q10, psDec->subfr_length, psDec->nb_subfr);
if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) {
/* First sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ];
} else {
/* Second sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ];
}
/* Set up Gain to random noise component */
B_Q14 = psPLC->LTPCoef_Q14;
rand_scale_Q14 = psPLC->randScale_Q14;
/* Set up attenuation gains */
harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
if( psDec->prevSignalType == TYPE_VOICED ) {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
} else {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
}
/* LPC concealment. Apply BWE to previous LPC */
silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) );
/* Preload LPC coeficients to array on stack. Gives small performance gain */
silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) );
/* First Lost frame */
if( psDec->lossCnt == 0 ) {
rand_scale_Q14 = 1 << 14;
/* Reduce random noise Gain for voiced frames */
if( psDec->prevSignalType == TYPE_VOICED ) {
for( i = 0; i < LTP_ORDER; i++ ) {
rand_scale_Q14 -= B_Q14[ i ];
}
rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */
rand_scale_Q14 = (opus_int16)silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 );
} else {
/* Reduce random noise for unvoiced frames with high LPC gain */
opus_int32 invGain_Q30, down_scale_Q30;
invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order, arch );
down_scale_Q30 = silk_min_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 );
down_scale_Q30 = silk_max_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 );
down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES );
rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 );
}
}
rand_seed = psPLC->rand_seed;
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
sLTP_buf_idx = psDec->ltp_mem_length;
/* Rewhiten LTP state */
idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2;
silk_assert( idx > 0 );
silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order, arch );
/* Scale LTP state */
inv_gain_Q30 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 46 );
inv_gain_Q30 = silk_min( inv_gain_Q30, silk_int32_MAX >> 1 );
for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) {
sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] );
}
/***************************/
/* LTP synthesis filtering */
/***************************/
for( k = 0; k < psDec->nb_subfr; k++ ) {
/* Set up pointer */
pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ];
for( i = 0; i < psDec->subfr_length; i++ ) {
/* Unrolled loop */
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LTP_pred_Q12 = 2;
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ 0 ], B_Q14[ 0 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] );
pred_lag_ptr++;
/* Generate LPC excitation */
rand_seed = silk_RAND( rand_seed );
idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK;
sLTP_Q14[ sLTP_buf_idx ] = silk_LSHIFT32( silk_SMLAWB( LTP_pred_Q12, rand_ptr[ idx ], rand_scale_Q14 ), 2 );
sLTP_buf_idx++;
}
/* Gradually reduce LTP gain */
for( j = 0; j < LTP_ORDER; j++ ) {
B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 );
}
if ( psDec->indices.signalType != TYPE_NO_VOICE_ACTIVITY ) {
/* Gradually reduce excitation gain */
rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 );
}
/* Slowly increase pitch lag */
psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 );
psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) );
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
}
/***************************/
/* LPC synthesis filtering */
/***************************/
sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ];
/* Copy LPC state */
silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) );
silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */
for( i = 0; i < psDec->frame_length; i++ ) {
/* partly unrolled */
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LPC_pred_Q10 = silk_RSHIFT( psDec->LPC_order, 1 );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
for( j = 10; j < psDec->LPC_order; j++ ) {
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] );
}
/* Add prediction to LPC excitation */
sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ],
silk_LSHIFT_SAT32( LPC_pred_Q10, 4 ));
/* Scale with Gain */
frame[ i ] = (opus_int16)silk_SAT16( silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], prevGain_Q10[ 1 ] ), 8 ) ) );
}
/* Save LPC state */
silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
/**************************************/
/* Update states */
/**************************************/
psPLC->rand_seed = rand_seed;
psPLC->randScale_Q14 = rand_scale_Q14;
for( i = 0; i < MAX_NB_SUBFR; i++ ) {
psDecCtrl->pitchL[ i ] = lag;
}
RESTORE_STACK;
}
opus_int silk_VAD_GetSA_Q8( /* O Return value, 0 if success */
silk_encoder_state *psEncC, /* I/O Encoder state */
const opus_int16 pIn[] /* I PCM input */
)
{
opus_int SA_Q15, pSNR_dB_Q7, input_tilt;
opus_int decimated_framelength1, decimated_framelength2;
opus_int decimated_framelength;
opus_int dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s;
opus_int32 sumSquared, smooth_coef_Q16;
opus_int16 HPstateTmp;
VARDECL( opus_int16, X );
opus_int32 Xnrg[ VAD_N_BANDS ];
opus_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ];
opus_int32 speech_nrg, x_tmp;
opus_int X_offset[ VAD_N_BANDS ];
opus_int ret = 0;
silk_VAD_state *psSilk_VAD = &psEncC->sVAD;
SAVE_STACK;
/* Safety checks */
silk_assert( VAD_N_BANDS == 4 );
silk_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
silk_assert( psEncC->frame_length <= 512 );
silk_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) );
/***********************/
/* Filter and Decimate */
/***********************/
decimated_framelength1 = silk_RSHIFT( psEncC->frame_length, 1 );
decimated_framelength2 = silk_RSHIFT( psEncC->frame_length, 2 );
decimated_framelength = 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;
ALLOC( X, X_offset[ 3 ] + decimated_framelength1, opus_int16 );
/* 0-8 kHz to 0-4 kHz and 4-8 kHz */
silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[ 0 ],
X, &X[ X_offset[ 3 ] ], psEncC->frame_length );
/* 0-4 kHz to 0-2 kHz and 2-4 kHz */
silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState1[ 0 ],
X, &X[ X_offset[ 2 ] ], decimated_framelength1 );
/* 0-2 kHz to 0-1 kHz and 1-2 kHz */
silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState2[ 0 ],
X, &X[ X_offset[ 1 ] ], decimated_framelength2 );
/*********************************************/
/* HP filter on lowest band (differentiator) */
/*********************************************/
X[ decimated_framelength - 1 ] = silk_RSHIFT( X[ decimated_framelength - 1 ], 1 );
HPstateTmp = X[ decimated_framelength - 1 ];
for( i = decimated_framelength - 1; i > 0; i-- ) {
X[ i - 1 ] = 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 < VAD_N_BANDS; b++ ) {
/* Find the decimated framelength in the non-uniformly divided bands */
decimated_framelength = silk_RSHIFT( psEncC->frame_length, silk_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) );
/* Split length into subframe lengths */
dec_subframe_length = silk_RSHIFT( decimated_framelength, 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 < 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 = silk_RSHIFT(
X[ X_offset[ b ] + i + dec_subframe_offset ], 3 );
sumSquared = silk_SMLABB( sumSquared, x_tmp, x_tmp );
/* Safety check */
silk_assert( sumSquared >= 0 );
}
/* Add/saturate summed energy of current subframe */
if( s < VAD_INTERNAL_SUBFRAMES - 1 ) {
Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], sumSquared );
} else {
/* Look-ahead subframe */
Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], silk_RSHIFT( sumSquared, 1 ) );
}
dec_subframe_offset += dec_subframe_length;
}
psSilk_VAD->XnrgSubfr[ b ] = sumSquared;
}
/********************/
/* Noise estimation */
/********************/
silk_VAD_GetNoiseLevels( &Xnrg[ 0 ], psSilk_VAD );
/***********************************************/
/* Signal-plus-noise to noise ratio estimation */
/***********************************************/
sumSquared = 0;
input_tilt = 0;
for( b = 0; b < 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 ] = silk_DIV32( silk_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 );
} else {
NrgToNoiseRatio_Q8[ b ] = silk_DIV32( Xnrg[ b ], silk_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 );
}
/* Convert to log domain */
SNR_Q7 = silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128;
/* Sum-of-squares */
sumSquared = silk_SMLABB( sumSquared, SNR_Q7, SNR_Q7 ); /* Q14 */
/* Tilt measure */
if( speech_nrg < ( (opus_int32)1 << 20 ) ) {
/* Scale down SNR value for small subband speech energies */
SNR_Q7 = silk_SMULWB( silk_LSHIFT( silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 );
}
input_tilt = silk_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 );
} else {
NrgToNoiseRatio_Q8[ b ] = 256;
}
}
/* Mean-of-squares */
sumSquared = silk_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */
/* Root-mean-square approximation, scale to dBs, and write to output pointer */
pSNR_dB_Q7 = (opus_int16)( 3 * silk_SQRT_APPROX( sumSquared ) ); /* Q7 */
/*********************************/
/* Speech Probability Estimation */
/*********************************/
SA_Q15 = silk_sigm_Q15( silk_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 );
/**************************/
/* Frequency Tilt Measure */
/**************************/
psEncC->input_tilt_Q15 = silk_LSHIFT( silk_sigm_Q15( input_tilt ) - 16384, 1 );
/**************************************************/
/* Scale the sigmoid output based on power levels */
/**************************************************/
speech_nrg = 0;
for( b = 0; b < VAD_N_BANDS; b++ ) {
/* Accumulate signal-without-noise energies, higher frequency bands have more weight */
speech_nrg += ( b + 1 ) * silk_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 );
}
/* Power scaling */
if( speech_nrg <= 0 ) {
SA_Q15 = silk_RSHIFT( SA_Q15, 1 );
} else if( speech_nrg < 32768 ) {
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 16 );
} else {
speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 15 );
}
/* square-root */
speech_nrg = silk_SQRT_APPROX( speech_nrg );
SA_Q15 = silk_SMULWB( 32768 + speech_nrg, SA_Q15 );
}
/* Copy the resulting speech activity in Q8 */
psEncC->speech_activity_Q8 = silk_min_int( silk_RSHIFT( SA_Q15, 7 ), silk_uint8_MAX );
/***********************************/
/* Energy Level and SNR estimation */
/***********************************/
/* Smoothing coefficient */
smooth_coef_Q16 = silk_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, silk_SMULWB( (opus_int32)SA_Q15, SA_Q15 ) );
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
smooth_coef_Q16 >>= 1;
}
/* amplitude of monic warped coefficients by using bandwidth expansion on the true coefficients */
static inline void limit_warped_coefs(
opus_int32 *coefs_syn_Q24,
opus_int32 *coefs_ana_Q24,
opus_int lambda_Q16,
opus_int32 limit_Q24,
opus_int order
) {
opus_int i, iter, ind = 0;
opus_int32 tmp, maxabs_Q24, chirp_Q16, gain_syn_Q16, gain_ana_Q16;
opus_int32 nom_Q16, den_Q24;
/* Convert to monic coefficients */
lambda_Q16 = -lambda_Q16;
for( i = order - 1; i > 0; i-- ) {
coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 );
coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 );
}
lambda_Q16 = -lambda_Q16;
nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -lambda_Q16, lambda_Q16 );
den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_syn_Q24[ 0 ], lambda_Q16 );
gain_syn_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 );
den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_ana_Q24[ 0 ], lambda_Q16 );
gain_ana_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 );
for( i = 0; i < order; i++ ) {
coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] );
coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] );
}
for( iter = 0; iter < 10; iter++ ) {
/* Find maximum absolute value */
maxabs_Q24 = -1;
for( i = 0; i < order; i++ ) {
tmp = silk_max( silk_abs_int32( coefs_syn_Q24[ i ] ), silk_abs_int32( coefs_ana_Q24[ i ] ) );
if( tmp > maxabs_Q24 ) {
maxabs_Q24 = tmp;
ind = i;
}
}
if( maxabs_Q24 <= limit_Q24 ) {
/* Coefficients are within range - done */
return;
}
/* Convert back to true warped coefficients */
for( i = 1; i < order; i++ ) {
coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 );
coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 );
}
gain_syn_Q16 = silk_INVERSE32_varQ( gain_syn_Q16, 32 );
gain_ana_Q16 = silk_INVERSE32_varQ( gain_ana_Q16, 32 );
for( i = 0; i < order; i++ ) {
coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] );
coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] );
}
/* Apply bandwidth expansion */
chirp_Q16 = SILK_FIX_CONST( 0.99, 16 ) - silk_DIV32_varQ(
silk_SMULWB( maxabs_Q24 - limit_Q24, silk_SMLABB( SILK_FIX_CONST( 0.8, 10 ), SILK_FIX_CONST( 0.1, 10 ), iter ) ),
silk_MUL( maxabs_Q24, ind + 1 ), 22 );
silk_bwexpander_32( coefs_syn_Q24, order, chirp_Q16 );
silk_bwexpander_32( coefs_ana_Q24, order, chirp_Q16 );
/* Convert to monic warped coefficients */
lambda_Q16 = -lambda_Q16;
for( i = order - 1; i > 0; i-- ) {
coefs_syn_Q24[ i - 1 ] = silk_SMLAWB( coefs_syn_Q24[ i - 1 ], coefs_syn_Q24[ i ], lambda_Q16 );
coefs_ana_Q24[ i - 1 ] = silk_SMLAWB( coefs_ana_Q24[ i - 1 ], coefs_ana_Q24[ i ], lambda_Q16 );
}
lambda_Q16 = -lambda_Q16;
nom_Q16 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 16 ), -lambda_Q16, lambda_Q16 );
den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_syn_Q24[ 0 ], lambda_Q16 );
gain_syn_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 );
den_Q24 = silk_SMLAWB( SILK_FIX_CONST( 1.0, 24 ), coefs_ana_Q24[ 0 ], lambda_Q16 );
gain_ana_Q16 = silk_DIV32_varQ( nom_Q16, den_Q24, 24 );
for( i = 0; i < order; i++ ) {
coefs_syn_Q24[ i ] = silk_SMULWW( gain_syn_Q16, coefs_syn_Q24[ i ] );
coefs_ana_Q24[ i ] = silk_SMULWW( gain_ana_Q16, coefs_ana_Q24[ i ] );
}
}
silk_assert( 0 );
}
static inline void silk_PLC_conceal(
silk_decoder_state *psDec, /* I/O Decoder state */
silk_decoder_control *psDecCtrl, /* I/O Decoder control */
opus_int16 frame[] /* O LPC residual signal */
)
{
opus_int i, j, k;
opus_int lag, idx, sLTP_buf_idx, shift1, shift2;
opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q16, inv_gain_Q30;
opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr;
opus_int32 LPC_exc_Q14, LPC_pred_Q10, LTP_pred_Q12;
opus_int16 rand_scale_Q14;
opus_int16 *B_Q14, *exc_buf_ptr;
opus_int32 *sLPC_Q14_ptr;
opus_int16 exc_buf[ 2 * MAX_SUB_FRAME_LENGTH ];
opus_int16 A_Q12[ MAX_LPC_ORDER ];
opus_int16 sLTP[ MAX_FRAME_LENGTH ];
opus_int32 sLTP_Q14[ 2 * MAX_FRAME_LENGTH ];
silk_PLC_struct *psPLC = &psDec->sPLC;
if (psDec->first_frame_after_reset)
silk_memset(psPLC->prevLPC_Q12, 0, MAX_LPC_ORDER*sizeof(psPLC->prevLPC_Q12[ 0 ]));
/* Find random noise component */
/* Scale previous excitation signal */
exc_buf_ptr = exc_buf;
for( k = 0; k < 2; k++ ) {
for( i = 0; i < psPLC->subfr_length; i++ ) {
exc_buf_ptr[ i ] = ( opus_int16 )silk_RSHIFT(
silk_SMULWW( psDec->exc_Q10[ i + ( k + psPLC->nb_subfr - 2 ) * psPLC->subfr_length ], psPLC->prevGain_Q16[ k ] ), 10 );
}
exc_buf_ptr += psPLC->subfr_length;
}
/* Find the subframe with lowest energy of the last two and use that as random noise generator */
silk_sum_sqr_shift( &energy1, &shift1, exc_buf, psPLC->subfr_length );
silk_sum_sqr_shift( &energy2, &shift2, &exc_buf[ psPLC->subfr_length ], psPLC->subfr_length );
if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) {
/* First sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ];
} else {
/* Second sub-frame has lowest energy */
rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ];
}
/* Setup Gain to random noise component */
B_Q14 = psPLC->LTPCoef_Q14;
rand_scale_Q14 = psPLC->randScale_Q14;
/* Setup attenuation gains */
harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
if( psDec->prevSignalType == TYPE_VOICED ) {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
} else {
rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
}
/* LPC concealment. Apply BWE to previous LPC */
silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) );
/* Preload LPC coeficients to array on stack. Gives small performance gain */
silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) );
/* First Lost frame */
if( psDec->lossCnt == 0 ) {
rand_scale_Q14 = 1 << 14;
/* Reduce random noise Gain for voiced frames */
if( psDec->prevSignalType == TYPE_VOICED ) {
for( i = 0; i < LTP_ORDER; i++ ) {
rand_scale_Q14 -= B_Q14[ i ];
}
rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */
rand_scale_Q14 = ( opus_int16 )silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 );
} else {
/* Reduce random noise for unvoiced frames with high LPC gain */
opus_int32 invGain_Q30, down_scale_Q30;
silk_LPC_inverse_pred_gain( &invGain_Q30, psPLC->prevLPC_Q12, psDec->LPC_order );
down_scale_Q30 = silk_min_32( silk_RSHIFT( 1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 );
down_scale_Q30 = silk_max_32( silk_RSHIFT( 1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 );
down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES );
rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 );
}
}
rand_seed = psPLC->rand_seed;
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
sLTP_buf_idx = psDec->ltp_mem_length;
/* Rewhiten LTP state */
idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2;
silk_assert( idx > 0 );
silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order );
/* Scale LTP state */
inv_gain_Q16 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 32 );
inv_gain_Q16 = silk_min( inv_gain_Q16, silk_int16_MAX );
inv_gain_Q30 = silk_LSHIFT( inv_gain_Q16, 14 );
for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) {
sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] );
}
/***************************/
/* LTP synthesis filtering */
/***************************/
for( k = 0; k < psDec->nb_subfr; k++ ) {
/* Setup pointer */
pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ];
for( i = 0; i < psDec->subfr_length; i++ ) {
/* Unrolled loop */
LTP_pred_Q12 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] );
LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] );
pred_lag_ptr++;
/* Generate LPC excitation */
rand_seed = silk_RAND( rand_seed );
idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK;
LPC_exc_Q14 = silk_LSHIFT32( silk_SMULWB( rand_ptr[ idx ], rand_scale_Q14 ), 6 ); /* Random noise part */
LPC_exc_Q14 = silk_ADD32( LPC_exc_Q14, silk_LSHIFT32( LTP_pred_Q12, 2 ) ); /* Harmonic part */
sLTP_Q14[ sLTP_buf_idx ] = LPC_exc_Q14;
sLTP_buf_idx++;
}
/* Gradually reduce LTP gain */
for( j = 0; j < LTP_ORDER; j++ ) {
B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 );
}
/* Gradually reduce excitation gain */
rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 );
/* Slowly increase pitch lag */
psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 );
psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) );
lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
}
/***************************/
/* LPC synthesis filtering */
/***************************/
sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ];
/* Copy LPC state */
silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) );
silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */
for( i = 0; i < psDec->frame_length; i++ ) {
/* partly unrolled */
LPC_pred_Q10 = silk_SMULWB( sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] );
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
for( j = 10; j < psDec->LPC_order; j++ ) {
LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] );
}
/* Add prediction to LPC excitation */
sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], LPC_pred_Q10, 4 );
/* Scale with Gain */
frame[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], psPLC->prevGain_Q16[ 1 ] ), 14 ) );
}
/* Save LPC state */
silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
/**************************************/
/* Update states */
/**************************************/
psPLC->rand_seed = rand_seed;
psPLC->randScale_Q14 = rand_scale_Q14;
for( i = 0; i < MAX_NB_SUBFR; i++ ) {
psDecCtrl->pitchL[ i ] = lag;
}
}
void silk_prefilter_FIX(
silk_encoder_state_FIX *psEnc, /* I/O Encoder state */
const silk_encoder_control_FIX *psEncCtrl, /* I Encoder control */
opus_int32 xw_Q3[], /* O Weighted signal */
const opus_int16 x[] /* I Speech signal */
)
{
silk_prefilter_state_FIX *P = &psEnc->sPrefilt;
opus_int j, k, lag;
opus_int32 tmp_32;
const opus_int16 *AR1_shp_Q13;
const opus_int16 *px;
opus_int32 *pxw_Q3;
opus_int HarmShapeGain_Q12, Tilt_Q14;
opus_int32 HarmShapeFIRPacked_Q12, LF_shp_Q14;
VARDECL( opus_int32, x_filt_Q12 );
VARDECL( opus_int32, st_res_Q2 );
opus_int16 B_Q10[ 2 ];
SAVE_STACK;
/* Set up pointers */
px = x;
pxw_Q3 = xw_Q3;
lag = P->lagPrev;
ALLOC( x_filt_Q12, psEnc->sCmn.subfr_length, opus_int32 );
ALLOC( st_res_Q2, psEnc->sCmn.subfr_length, opus_int32 );
for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) {
/* Update Variables that change per sub frame */
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
lag = psEncCtrl->pitchL[ k ];
}
/* Noise shape parameters */
HarmShapeGain_Q12 = silk_SMULWB( (opus_int32)psEncCtrl->HarmShapeGain_Q14[ k ], 16384 - psEncCtrl->HarmBoost_Q14[ k ] );
silk_assert( HarmShapeGain_Q12 >= 0 );
HarmShapeFIRPacked_Q12 = silk_RSHIFT( HarmShapeGain_Q12, 2 );
HarmShapeFIRPacked_Q12 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q12, 1 ), 16 );
Tilt_Q14 = psEncCtrl->Tilt_Q14[ k ];
LF_shp_Q14 = psEncCtrl->LF_shp_Q14[ k ];
AR1_shp_Q13 = &psEncCtrl->AR1_Q13[ k * MAX_SHAPE_LPC_ORDER ];
/* Short term FIR filtering*/
silk_warped_LPC_analysis_filter_FIX( P->sAR_shp, st_res_Q2, AR1_shp_Q13, px,
psEnc->sCmn.warping_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.shapingLPCOrder );
/* Reduce (mainly) low frequencies during harmonic emphasis */
B_Q10[ 0 ] = silk_RSHIFT_ROUND( psEncCtrl->GainsPre_Q14[ k ], 4 );
tmp_32 = silk_SMLABB( SILK_FIX_CONST( INPUT_TILT, 26 ), psEncCtrl->HarmBoost_Q14[ k ], HarmShapeGain_Q12 ); /* Q26 */
tmp_32 = silk_SMLABB( tmp_32, psEncCtrl->coding_quality_Q14, SILK_FIX_CONST( HIGH_RATE_INPUT_TILT, 12 ) ); /* Q26 */
tmp_32 = silk_SMULWB( tmp_32, -psEncCtrl->GainsPre_Q14[ k ] ); /* Q24 */
tmp_32 = silk_RSHIFT_ROUND( tmp_32, 14 ); /* Q10 */
B_Q10[ 1 ]= silk_SAT16( tmp_32 );
x_filt_Q12[ 0 ] = silk_MLA( silk_MUL( st_res_Q2[ 0 ], B_Q10[ 0 ] ), P->sHarmHP_Q2, B_Q10[ 1 ] );
for( j = 1; j < psEnc->sCmn.subfr_length; j++ ) {
x_filt_Q12[ j ] = silk_MLA( silk_MUL( st_res_Q2[ j ], B_Q10[ 0 ] ), st_res_Q2[ j - 1 ], B_Q10[ 1 ] );
}
P->sHarmHP_Q2 = st_res_Q2[ psEnc->sCmn.subfr_length - 1 ];
silk_prefilt_FIX( P, x_filt_Q12, pxw_Q3, HarmShapeFIRPacked_Q12, Tilt_Q14, LF_shp_Q14, lag, psEnc->sCmn.subfr_length );
px += psEnc->sCmn.subfr_length;
pxw_Q3 += psEnc->sCmn.subfr_length;
}
P->lagPrev = psEncCtrl->pitchL[ psEnc->sCmn.nb_subfr - 1 ];
RESTORE_STACK;
}
void silk_find_LTP_FIX(
opus_int16 b_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O LTP coefs */
opus_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* O Weight for LTP quantization */
opus_int *LTPredCodGain_Q7, /* O LTP coding gain */
const opus_int16 r_lpc[], /* I residual signal after LPC signal + state for first 10 ms */
const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */
const opus_int32 Wght_Q15[ MAX_NB_SUBFR ], /* I weights */
const opus_int subfr_length, /* I subframe length */
const opus_int nb_subfr, /* I number of subframes */
const opus_int mem_offset, /* I number of samples in LTP memory */
opus_int corr_rshifts[ MAX_NB_SUBFR ] /* O right shifts applied to correlations */
)
{
opus_int i, k, lshift;
const opus_int16 *r_ptr, *lag_ptr;
opus_int16 *b_Q14_ptr;
opus_int32 regu;
opus_int32 *WLTP_ptr;
opus_int32 b_Q16[ LTP_ORDER ], delta_b_Q14[ LTP_ORDER ], d_Q14[ MAX_NB_SUBFR ], nrg[ MAX_NB_SUBFR ], g_Q26;
opus_int32 w[ MAX_NB_SUBFR ], WLTP_max, max_abs_d_Q14, max_w_bits;
opus_int32 temp32, denom32;
opus_int extra_shifts;
opus_int rr_shifts, maxRshifts, maxRshifts_wxtra, LZs;
opus_int32 LPC_res_nrg, LPC_LTP_res_nrg, div_Q16;
opus_int32 Rr[ LTP_ORDER ], rr[ MAX_NB_SUBFR ];
opus_int32 wd, m_Q12;
b_Q14_ptr = b_Q14;
WLTP_ptr = WLTP;
r_ptr = &r_lpc[ mem_offset ];
for( k = 0; k < nb_subfr; k++ ) {
lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 );
silk_sum_sqr_shift( &rr[ k ], &rr_shifts, r_ptr, subfr_length ); /* rr[ k ] in Q( -rr_shifts ) */
/* Assure headroom */
LZs = silk_CLZ32( rr[k] );
if( LZs < LTP_CORRS_HEAD_ROOM ) {
rr[ k ] = silk_RSHIFT_ROUND( rr[ k ], LTP_CORRS_HEAD_ROOM - LZs );
rr_shifts += ( LTP_CORRS_HEAD_ROOM - LZs );
}
corr_rshifts[ k ] = rr_shifts;
silk_corrMatrix_FIX( lag_ptr, subfr_length, LTP_ORDER, LTP_CORRS_HEAD_ROOM, WLTP_ptr, &corr_rshifts[ k ] ); /* WLTP_fix_ptr in Q( -corr_rshifts[ k ] ) */
/* The correlation vector always has lower max abs value than rr and/or RR so head room is assured */
silk_corrVector_FIX( lag_ptr, r_ptr, subfr_length, LTP_ORDER, Rr, corr_rshifts[ k ] ); /* Rr_fix_ptr in Q( -corr_rshifts[ k ] ) */
if( corr_rshifts[ k ] > rr_shifts ) {
rr[ k ] = silk_RSHIFT( rr[ k ], corr_rshifts[ k ] - rr_shifts ); /* rr[ k ] in Q( -corr_rshifts[ k ] ) */
}
silk_assert( rr[ k ] >= 0 );
regu = 1;
regu = silk_SMLAWB( regu, rr[ k ], SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, 0, 0, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
regu = silk_SMLAWB( regu, matrix_ptr( WLTP_ptr, LTP_ORDER-1, LTP_ORDER-1, LTP_ORDER ), SILK_FIX_CONST( LTP_DAMPING/3, 16 ) );
silk_regularize_correlations_FIX( WLTP_ptr, &rr[k], regu, LTP_ORDER );
silk_solve_LDL_FIX( WLTP_ptr, LTP_ORDER, Rr, b_Q16 ); /* WLTP_fix_ptr and Rr_fix_ptr both in Q(-corr_rshifts[k]) */
/* Limit and store in Q14 */
silk_fit_LTP( b_Q16, b_Q14_ptr );
/* Calculate residual energy */
nrg[ k ] = silk_residual_energy16_covar_FIX( b_Q14_ptr, WLTP_ptr, Rr, rr[ k ], LTP_ORDER, 14 ); /* nrg_fix in Q( -corr_rshifts[ k ] ) */
/* temp = Wght[ k ] / ( nrg[ k ] * Wght[ k ] + 0.01f * subfr_length ); */
extra_shifts = silk_min_int( corr_rshifts[ k ], LTP_CORRS_HEAD_ROOM );
denom32 = silk_LSHIFT_SAT32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 + extra_shifts ) + /* Q( -corr_rshifts[ k ] + extra_shifts ) */
silk_RSHIFT( silk_SMULWB( subfr_length, 655 ), corr_rshifts[ k ] - extra_shifts ); /* Q( -corr_rshifts[ k ] + extra_shifts ) */
denom32 = silk_max( denom32, 1 );
silk_assert( ((opus_int64)Wght_Q15[ k ] << 16 ) < silk_int32_MAX ); /* Wght always < 0.5 in Q0 */
temp32 = silk_DIV32( silk_LSHIFT( (opus_int32)Wght_Q15[ k ], 16 ), denom32 ); /* Q( 15 + 16 + corr_rshifts[k] - extra_shifts ) */
temp32 = silk_RSHIFT( temp32, 31 + corr_rshifts[ k ] - extra_shifts - 26 ); /* Q26 */
/* Limit temp such that the below scaling never wraps around */
WLTP_max = 0;
for( i = 0; i < LTP_ORDER * LTP_ORDER; i++ ) {
WLTP_max = silk_max( WLTP_ptr[ i ], WLTP_max );
}
lshift = silk_CLZ32( WLTP_max ) - 1 - 3; /* keep 3 bits free for vq_nearest_neighbor_fix */
silk_assert( 26 - 18 + lshift >= 0 );
if( 26 - 18 + lshift < 31 ) {
temp32 = silk_min_32( temp32, silk_LSHIFT( (opus_int32)1, 26 - 18 + lshift ) );
}
silk_scale_vector32_Q26_lshift_18( WLTP_ptr, temp32, LTP_ORDER * LTP_ORDER ); /* WLTP_ptr in Q( 18 - corr_rshifts[ k ] ) */
w[ k ] = matrix_ptr( WLTP_ptr, LTP_ORDER/2, LTP_ORDER/2, LTP_ORDER ); /* w in Q( 18 - corr_rshifts[ k ] ) */
silk_assert( w[k] >= 0 );
r_ptr += subfr_length;
b_Q14_ptr += LTP_ORDER;
WLTP_ptr += LTP_ORDER * LTP_ORDER;
}
maxRshifts = 0;
for( k = 0; k < nb_subfr; k++ ) {
maxRshifts = silk_max_int( corr_rshifts[ k ], maxRshifts );
}
/* Compute LTP coding gain */
if( LTPredCodGain_Q7 != NULL ) {
LPC_LTP_res_nrg = 0;
LPC_res_nrg = 0;
silk_assert( LTP_CORRS_HEAD_ROOM >= 2 ); /* Check that no overflow will happen when adding */
for( k = 0; k < nb_subfr; k++ ) {
LPC_res_nrg = silk_ADD32( LPC_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( rr[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */
LPC_LTP_res_nrg = silk_ADD32( LPC_LTP_res_nrg, silk_RSHIFT( silk_ADD32( silk_SMULWB( nrg[ k ], Wght_Q15[ k ] ), 1 ), 1 + ( maxRshifts - corr_rshifts[ k ] ) ) ); /* Q( -maxRshifts ) */
}
LPC_LTP_res_nrg = silk_max( LPC_LTP_res_nrg, 1 ); /* avoid division by zero */
div_Q16 = silk_DIV32_varQ( LPC_res_nrg, LPC_LTP_res_nrg, 16 );
*LTPredCodGain_Q7 = ( opus_int )silk_SMULBB( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) );
silk_assert( *LTPredCodGain_Q7 == ( opus_int )silk_SAT16( silk_MUL( 3, silk_lin2log( div_Q16 ) - ( 16 << 7 ) ) ) );
}
/* smoothing */
/* d = sum( B, 1 ); */
b_Q14_ptr = b_Q14;
for( k = 0; k < nb_subfr; k++ ) {
d_Q14[ k ] = 0;
for( i = 0; i < LTP_ORDER; i++ ) {
d_Q14[ k ] += b_Q14_ptr[ i ];
}
b_Q14_ptr += LTP_ORDER;
}
/* m = ( w * d' ) / ( sum( w ) + 1e-3 ); */
/* Find maximum absolute value of d_Q14 and the bits used by w in Q0 */
max_abs_d_Q14 = 0;
max_w_bits = 0;
for( k = 0; k < nb_subfr; k++ ) {
max_abs_d_Q14 = silk_max_32( max_abs_d_Q14, silk_abs( d_Q14[ k ] ) );
/* w[ k ] is in Q( 18 - corr_rshifts[ k ] ) */
/* Find bits needed in Q( 18 - maxRshifts ) */
max_w_bits = silk_max_32( max_w_bits, 32 - silk_CLZ32( w[ k ] ) + corr_rshifts[ k ] - maxRshifts );
}
/* max_abs_d_Q14 = (5 << 15); worst case, i.e. LTP_ORDER * -silk_int16_MIN */
silk_assert( max_abs_d_Q14 <= ( 5 << 15 ) );
/* How many bits is needed for w*d' in Q( 18 - maxRshifts ) in the worst case, of all d_Q14's being equal to max_abs_d_Q14 */
extra_shifts = max_w_bits + 32 - silk_CLZ32( max_abs_d_Q14 ) - 14;
/* Subtract what we got available; bits in output var plus maxRshifts */
extra_shifts -= ( 32 - 1 - 2 + maxRshifts ); /* Keep sign bit free as well as 2 bits for accumulation */
extra_shifts = silk_max_int( extra_shifts, 0 );
maxRshifts_wxtra = maxRshifts + extra_shifts;
temp32 = silk_RSHIFT( 262, maxRshifts + extra_shifts ) + 1; /* 1e-3f in Q( 18 - (maxRshifts + extra_shifts) ) */
wd = 0;
for( k = 0; k < nb_subfr; k++ ) {
/* w has at least 2 bits of headroom so no overflow should happen */
temp32 = silk_ADD32( temp32, silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ) ); /* Q( 18 - maxRshifts_wxtra ) */
wd = silk_ADD32( wd, silk_LSHIFT( silk_SMULWW( silk_RSHIFT( w[ k ], maxRshifts_wxtra - corr_rshifts[ k ] ), d_Q14[ k ] ), 2 ) ); /* Q( 18 - maxRshifts_wxtra ) */
}
m_Q12 = silk_DIV32_varQ( wd, temp32, 12 );
b_Q14_ptr = b_Q14;
for( k = 0; k < nb_subfr; k++ ) {
/* w_fix[ k ] from Q( 18 - corr_rshifts[ k ] ) to Q( 16 ) */
if( 2 - corr_rshifts[k] > 0 ) {
temp32 = silk_RSHIFT( w[ k ], 2 - corr_rshifts[ k ] );
} else {
temp32 = silk_LSHIFT_SAT32( w[ k ], corr_rshifts[ k ] - 2 );
}
g_Q26 = silk_MUL(
silk_DIV32(
SILK_FIX_CONST( LTP_SMOOTHING, 26 ),
silk_RSHIFT( SILK_FIX_CONST( LTP_SMOOTHING, 26 ), 10 ) + temp32 ), /* Q10 */
silk_LSHIFT_SAT32( silk_SUB_SAT32( (opus_int32)m_Q12, silk_RSHIFT( d_Q14[ k ], 2 ) ), 4 ) ); /* Q16 */
temp32 = 0;
for( i = 0; i < LTP_ORDER; i++ ) {
delta_b_Q14[ i ] = silk_max_16( b_Q14_ptr[ i ], 1638 ); /* 1638_Q14 = 0.1_Q0 */
temp32 += delta_b_Q14[ i ]; /* Q14 */
}
temp32 = silk_DIV32( g_Q26, temp32 ); /* Q14 -> Q12 */
for( i = 0; i < LTP_ORDER; i++ ) {
b_Q14_ptr[ i ] = silk_LIMIT_32( (opus_int32)b_Q14_ptr[ i ] + silk_SMULWB( silk_LSHIFT_SAT32( temp32, 4 ), delta_b_Q14[ i ] ), -16000, 28000 );
}
b_Q14_ptr += LTP_ORDER;
}
}
/* notch filter just above Nyquist. */
void silk_resampler_private_up2_HQ(
opus_int32 *S, /* I/O Resampler state [ 6 ] */
opus_int16 *out, /* O Output signal [ 2 * len ] */
const opus_int16 *in, /* I Input signal [ len ] */
opus_int32 len /* I Number of input samples */
)
{
opus_int32 k;
opus_int32 in32, out32_1, out32_2, Y, X;
silk_assert( silk_resampler_up2_hq_0[ 0 ] > 0 );
silk_assert( silk_resampler_up2_hq_0[ 1 ] > 0 );
silk_assert( silk_resampler_up2_hq_0[ 2 ] < 0 );
silk_assert( silk_resampler_up2_hq_1[ 0 ] > 0 );
silk_assert( silk_resampler_up2_hq_1[ 1 ] > 0 );
silk_assert( silk_resampler_up2_hq_1[ 2 ] < 0 );
/* Internal variables and state are in Q10 format */
for( k = 0; k < len; k++ ) {
/* Convert to Q10 */
in32 = silk_LSHIFT( (opus_int32)in[ k ], 10 );
/* First all-pass section for even output sample */
Y = silk_SUB32( in32, S[ 0 ] );
X = silk_SMULWB( Y, silk_resampler_up2_hq_0[ 0 ] );
out32_1 = silk_ADD32( S[ 0 ], X );
S[ 0 ] = silk_ADD32( in32, X );
/* Second all-pass section for even output sample */
Y = silk_SUB32( out32_1, S[ 1 ] );
X = silk_SMULWB( Y, silk_resampler_up2_hq_0[ 1 ] );
out32_2 = silk_ADD32( S[ 1 ], X );
S[ 1 ] = silk_ADD32( out32_1, X );
/* Third all-pass section for even output sample */
Y = silk_SUB32( out32_2, S[ 2 ] );
X = silk_SMLAWB( Y, Y, silk_resampler_up2_hq_0[ 2 ] );
out32_1 = silk_ADD32( S[ 2 ], X );
S[ 2 ] = silk_ADD32( out32_2, X );
/* Apply gain in Q15, convert back to int16 and store to output */
out[ 2 * k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( out32_1, 10 ) );
/* First all-pass section for odd output sample */
Y = silk_SUB32( in32, S[ 3 ] );
X = silk_SMULWB( Y, silk_resampler_up2_hq_1[ 0 ] );
out32_1 = silk_ADD32( S[ 3 ], X );
S[ 3 ] = silk_ADD32( in32, X );
/* Second all-pass section for odd output sample */
Y = silk_SUB32( out32_1, S[ 4 ] );
X = silk_SMULWB( Y, silk_resampler_up2_hq_1[ 1 ] );
out32_2 = silk_ADD32( S[ 4 ], X );
S[ 4 ] = silk_ADD32( out32_1, X );
/* Third all-pass section for odd output sample */
Y = silk_SUB32( out32_2, S[ 5 ] );
X = silk_SMLAWB( Y, Y, silk_resampler_up2_hq_1[ 2 ] );
out32_1 = silk_ADD32( S[ 5 ], X );
S[ 5 ] = silk_ADD32( out32_2, X );
/* Apply gain in Q15, convert back to int16 and store to output */
out[ 2 * k + 1 ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( out32_1, 10 ) );
}
}
/* Downsample by a factor 2/3, low quality */
void silk_resampler_down2_3(
opus_int32 *S, /* I/O State vector [ 6 ] */
opus_int16 *out, /* O Output signal [ floor(2*inLen/3) ] */
const opus_int16 *in, /* I Input signal [ inLen ] */
opus_int32 inLen /* I Number of input samples */
)
{
opus_int32 nSamplesIn, counter, res_Q6;
VARDECL( opus_int32, buf );
opus_int32 *buf_ptr;
SAVE_STACK;
ALLOC( buf, RESAMPLER_MAX_BATCH_SIZE_IN + ORDER_FIR, opus_int32 );
/* Copy buffered samples to start of buffer */
silk_memcpy( buf, S, ORDER_FIR * sizeof( opus_int32 ) );
/* Iterate over blocks of frameSizeIn input samples */
while( 1 ) {
nSamplesIn = silk_min( inLen, RESAMPLER_MAX_BATCH_SIZE_IN );
/* Second-order AR filter (output in Q8) */
silk_resampler_private_AR2( &S[ ORDER_FIR ], &buf[ ORDER_FIR ], in,
silk_Resampler_2_3_COEFS_LQ, nSamplesIn );
/* Interpolate filtered signal */
buf_ptr = buf;
counter = nSamplesIn;
while( counter > 2 ) {
/* Inner product */
res_Q6 = silk_SMULWB( buf_ptr[ 0 ], silk_Resampler_2_3_COEFS_LQ[ 2 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 1 ], silk_Resampler_2_3_COEFS_LQ[ 3 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 2 ], silk_Resampler_2_3_COEFS_LQ[ 5 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 3 ], silk_Resampler_2_3_COEFS_LQ[ 4 ] );
/* Scale down, saturate and store in output array */
*out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q6, 6 ) );
res_Q6 = silk_SMULWB( buf_ptr[ 1 ], silk_Resampler_2_3_COEFS_LQ[ 4 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 2 ], silk_Resampler_2_3_COEFS_LQ[ 5 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 3 ], silk_Resampler_2_3_COEFS_LQ[ 3 ] );
res_Q6 = silk_SMLAWB( res_Q6, buf_ptr[ 4 ], silk_Resampler_2_3_COEFS_LQ[ 2 ] );
/* Scale down, saturate and store in output array */
*out++ = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( res_Q6, 6 ) );
buf_ptr += 3;
counter -= 3;
}
in += nSamplesIn;
inLen -= nSamplesIn;
if( inLen > 0 ) {
/* More iterations to do; copy last part of filtered signal to beginning of buffer */
silk_memcpy( buf, &buf[ nSamplesIn ], ORDER_FIR * sizeof( opus_int32 ) );
} else {
break;
}
}
/* Copy last part of filtered signal to the state for the next call */
silk_memcpy( S, &buf[ nSamplesIn ], ORDER_FIR * sizeof( opus_int32 ) );
RESTORE_STACK;
}
/* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */
void silk_VQ_WMat_EC(
opus_int8 *ind, /* O index of best codebook vector */
opus_int32 *rate_dist_Q14, /* O best weighted quant error + mu * rate */
opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
const opus_int16 *in_Q14, /* I input vector to be quantized */
const opus_int32 *W_Q18, /* I weighting matrix */
const opus_int8 *cb_Q7, /* I codebook */
const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
const opus_int mu_Q9, /* I tradeoff betw. weighted error and rate */
const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
opus_int L /* I number of vectors in codebook */
)
{
opus_int k, gain_tmp_Q7;
const opus_int8 *cb_row_Q7;
opus_int16 diff_Q14[ 5 ];
opus_int32 sum1_Q14, sum2_Q16;
/* Loop over codebook */
*rate_dist_Q14 = silk_int32_MAX;
cb_row_Q7 = cb_Q7;
for( k = 0; k < L; k++ ) {
gain_tmp_Q7 = cb_gain_Q7[k];
diff_Q14[ 0 ] = in_Q14[ 0 ] - silk_LSHIFT( cb_row_Q7[ 0 ], 7 );
diff_Q14[ 1 ] = in_Q14[ 1 ] - silk_LSHIFT( cb_row_Q7[ 1 ], 7 );
diff_Q14[ 2 ] = in_Q14[ 2 ] - silk_LSHIFT( cb_row_Q7[ 2 ], 7 );
diff_Q14[ 3 ] = in_Q14[ 3 ] - silk_LSHIFT( cb_row_Q7[ 3 ], 7 );
diff_Q14[ 4 ] = in_Q14[ 4 ] - silk_LSHIFT( cb_row_Q7[ 4 ], 7 );
/* Weighted rate */
sum1_Q14 = silk_SMULBB( mu_Q9, cl_Q5[ k ] );
/* Penalty for too large gain */
sum1_Q14 = silk_ADD_LSHIFT32( sum1_Q14, silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 10 );
silk_assert( sum1_Q14 >= 0 );
/* first row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 1 ], diff_Q14[ 1 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 2 ], diff_Q14[ 2 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 3 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 4 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 0 ], diff_Q14[ 0 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 0 ] );
/* second row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 7 ], diff_Q14[ 2 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 8 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 9 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 6 ], diff_Q14[ 1 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 1 ] );
/* third row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 13 ], diff_Q14[ 3 ] );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 14 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 12 ], diff_Q14[ 2 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 2 ] );
/* fourth row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 19 ], diff_Q14[ 4 ] );
sum2_Q16 = silk_LSHIFT( sum2_Q16, 1 );
sum2_Q16 = silk_SMLAWB( sum2_Q16, W_Q18[ 18 ], diff_Q14[ 3 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 3 ] );
/* last row of W_Q18 */
sum2_Q16 = silk_SMULWB( W_Q18[ 24 ], diff_Q14[ 4 ] );
sum1_Q14 = silk_SMLAWB( sum1_Q14, sum2_Q16, diff_Q14[ 4 ] );
silk_assert( sum1_Q14 >= 0 );
/* find best */
if( sum1_Q14 < *rate_dist_Q14 ) {
*rate_dist_Q14 = sum1_Q14;
*ind = (opus_int8)k;
*gain_Q7 = gain_tmp_Q7;
}
/* Go to next cbk vector */
cb_row_Q7 += LTP_ORDER;
}
}
/* Residual energy: nrg = wxx - 2 * wXx * c + c' * wXX * c */
opus_int32 silk_residual_energy16_covar_FIX(
const opus_int16 *c, /* I Prediction vector */
const opus_int32 *wXX, /* I Correlation matrix */
const opus_int32 *wXx, /* I Correlation vector */
opus_int32 wxx, /* I Signal energy */
opus_int D, /* I Dimension */
opus_int cQ /* I Q value for c vector 0 - 15 */
)
{
opus_int i, j, lshifts, Qxtra;
opus_int32 c_max, w_max, tmp, tmp2, nrg;
opus_int cn[ MAX_MATRIX_SIZE ];
const opus_int32 *pRow;
/* Safety checks */
silk_assert( D >= 0 );
silk_assert( D <= 16 );
silk_assert( cQ > 0 );
silk_assert( cQ < 16 );
lshifts = 16 - cQ;
Qxtra = lshifts;
c_max = 0;
for( i = 0; i < D; i++ ) {
c_max = silk_max_32( c_max, silk_abs( (opus_int32)c[ i ] ) );
}
Qxtra = silk_min_int( Qxtra, silk_CLZ32( c_max ) - 17 );
w_max = silk_max_32( wXX[ 0 ], wXX[ D * D - 1 ] );
Qxtra = silk_min_int( Qxtra, silk_CLZ32( silk_MUL( D, silk_RSHIFT( silk_SMULWB( w_max, c_max ), 4 ) ) ) - 5 );
Qxtra = silk_max_int( Qxtra, 0 );
for( i = 0; i < D; i++ ) {
cn[ i ] = silk_LSHIFT( ( opus_int )c[ i ], Qxtra );
silk_assert( silk_abs(cn[i]) <= ( silk_int16_MAX + 1 ) ); /* Check that silk_SMLAWB can be used */
}
lshifts -= Qxtra;
/* Compute wxx - 2 * wXx * c */
tmp = 0;
for( i = 0; i < D; i++ ) {
tmp = silk_SMLAWB( tmp, wXx[ i ], cn[ i ] );
}
nrg = silk_RSHIFT( wxx, 1 + lshifts ) - tmp; /* Q: -lshifts - 1 */
/* Add c' * wXX * c, assuming wXX is symmetric */
tmp2 = 0;
for( i = 0; i < D; i++ ) {
tmp = 0;
pRow = &wXX[ i * D ];
for( j = i + 1; j < D; j++ ) {
tmp = silk_SMLAWB( tmp, pRow[ j ], cn[ j ] );
}
tmp = silk_SMLAWB( tmp, silk_RSHIFT( pRow[ i ], 1 ), cn[ i ] );
tmp2 = silk_SMLAWB( tmp2, tmp, cn[ i ] );
}
nrg = silk_ADD_LSHIFT32( nrg, tmp2, lshifts ); /* Q: -lshifts - 1 */
/* Keep one bit free always, because we add them for LSF interpolation */
if( nrg < 1 ) {
nrg = 1;
} else if( nrg > silk_RSHIFT( silk_int32_MAX, lshifts + 2 ) ) {
nrg = silk_int32_MAX >> 1;
} else {
