/* 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));
}
}
/* 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 */
}
}
opus_int silk_init_encoder(
silk_encoder_state_Fxx *psEnc /* I/O Pointer to Silk FIX encoder state */
)
{
opus_int ret = 0;
/* Clear the entire encoder state */
silk_memset( psEnc, 0, sizeof( silk_encoder_state_Fxx ) );
psEnc->sCmn.variable_HP_smth1_Q15 = silk_LSHIFT( silk_lin2log( SILK_FIX_CONST( VARIABLE_HP_MIN_CUTOFF_HZ, 16 ) ) - ( 16 << 7 ), 8 );
psEnc->sCmn.variable_HP_smth2_Q15 = psEnc->sCmn.variable_HP_smth1_Q15;
/* Used to deactivate LSF interpolation, pitch prediction */
psEnc->sCmn.first_frame_after_reset = 1;
/* Initialize Silk VAD */
ret += silk_VAD_Init( &psEnc->sCmn.sVAD );
return ret;
}
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;
}
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;
}
}
void silk_quant_LTP_gains(
opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* I/O (un)quantized LTP gains */
opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */
opus_int8 *periodicity_index, /* O Periodicity Index */
opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */
const opus_int32 W_Q18[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Error Weights in Q18 */
opus_int mu_Q9, /* I Mu value (R/D tradeoff) */
opus_int lowComplexity, /* I Flag for low complexity */
const opus_int nb_subfr, /* I number of subframes */
int arch /* I Run-time architecture */
)
{
opus_int j, k, cbk_size;
opus_int8 temp_idx[ MAX_NB_SUBFR ];
const opus_uint8 *cl_ptr_Q5;
const opus_int8 *cbk_ptr_Q7;
const opus_uint8 *cbk_gain_ptr_Q7;
const opus_int16 *b_Q14_ptr;
const opus_int32 *W_Q18_ptr;
opus_int32 rate_dist_Q14_subfr, rate_dist_Q14, min_rate_dist_Q14;
opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7, gain_Q7;
/***************************************************/
/* iterate over different codebooks with different */
/* rates/distortions, and choose best */
/***************************************************/
min_rate_dist_Q14 = silk_int32_MAX;
best_sum_log_gain_Q7 = 0;
for(k = 0; k < 3; k++) {
/* Safety margin for pitch gain control, to take into account factors
such as state rescaling/rewhitening. */
opus_int32 gain_safety = SILK_FIX_CONST(0.4, 7);
cl_ptr_Q5 = silk_LTP_gain_BITS_Q5_ptrs[ k ];
cbk_ptr_Q7 = silk_LTP_vq_ptrs_Q7[ k ];
cbk_gain_ptr_Q7 = silk_LTP_vq_gain_ptrs_Q7[ k ];
cbk_size = silk_LTP_vq_sizes[ k ];
/* Set up pointer to first subframe */
W_Q18_ptr = W_Q18;
b_Q14_ptr = B_Q14;
rate_dist_Q14 = 0;
sum_log_gain_tmp_Q7 = *sum_log_gain_Q7;
for(j = 0; j < nb_subfr; j++) {
max_gain_Q7 = silk_log2lin((SILK_FIX_CONST(MAX_SUM_LOG_GAIN_DB / 6.0, 7) - sum_log_gain_tmp_Q7)
+ SILK_FIX_CONST(7, 7)) - gain_safety;
silk_VQ_WMat_EC(
&temp_idx[ j ], /* O index of best codebook vector */
&rate_dist_Q14_subfr, /* O best weighted quantization error + mu * rate */
&gain_Q7, /* O sum of absolute LTP coefficients */
b_Q14_ptr, /* I input vector to be quantized */
W_Q18_ptr, /* I weighting matrix */
cbk_ptr_Q7, /* I codebook */
cbk_gain_ptr_Q7, /* I codebook effective gains */
cl_ptr_Q5, /* I code length for each codebook vector */
mu_Q9, /* I tradeoff between weighted error and rate */
max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
cbk_size, /* I number of vectors in codebook */
arch /* I Run-time architecture */
);
rate_dist_Q14 = silk_ADD_POS_SAT32(rate_dist_Q14, rate_dist_Q14_subfr);
sum_log_gain_tmp_Q7 = silk_max(0, sum_log_gain_tmp_Q7
+ silk_lin2log(gain_safety + gain_Q7) - SILK_FIX_CONST(7, 7));
b_Q14_ptr += LTP_ORDER;
W_Q18_ptr += LTP_ORDER * LTP_ORDER;
}
/* Avoid never finding a codebook */
rate_dist_Q14 = silk_min(silk_int32_MAX - 1, rate_dist_Q14);
if(rate_dist_Q14 < min_rate_dist_Q14) {
min_rate_dist_Q14 = rate_dist_Q14;
*periodicity_index = (opus_int8)k;
silk_memcpy(cbk_index, temp_idx, nb_subfr * sizeof(opus_int8));
best_sum_log_gain_Q7 = sum_log_gain_tmp_Q7;
}
/* Break early in low-complexity mode if rate distortion is below threshold */
if(lowComplexity && (rate_dist_Q14 < silk_LTP_gain_middle_avg_RD_Q14)) {
break;
}
}
cbk_ptr_Q7 = silk_LTP_vq_ptrs_Q7[ *periodicity_index ];
for(j = 0; j < nb_subfr; j++) {
for(k = 0; k < LTP_ORDER; k++) {
B_Q14[ j * LTP_ORDER + k ] = silk_LSHIFT(cbk_ptr_Q7[ cbk_index[ j ] * LTP_ORDER + k ], 7);
}
}
*sum_log_gain_Q7 = best_sum_log_gain_Q7;
}
/* Entropy constrained matrix-weighted VQ, hard-coded to 5-element vectors, for a single input data vector */
void silk_VQ_WMat_EC_c(
opus_int8 *ind, /* O index of best codebook vector */
opus_int32 *res_nrg_Q15, /* O best residual energy */
opus_int32 *rate_dist_Q8, /* O best total bitrate */
opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
const opus_int32 *XX_Q17, /* I correlation matrix */
const opus_int32 *xX_Q17, /* I correlation vector */
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 subfr_len, /* I number of samples per subframe */
const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
const opus_int L /* I number of vectors in codebook */
)
{
opus_int k, gain_tmp_Q7;
const opus_int8 *cb_row_Q7;
opus_int32 neg_xX_Q24[ 5 ];
opus_int32 sum1_Q15, sum2_Q24;
opus_int32 bits_res_Q8, bits_tot_Q8;
/* Negate and convert to new Q domain */
neg_xX_Q24[ 0 ] = -silk_LSHIFT32( xX_Q17[ 0 ], 7 );
neg_xX_Q24[ 1 ] = -silk_LSHIFT32( xX_Q17[ 1 ], 7 );
neg_xX_Q24[ 2 ] = -silk_LSHIFT32( xX_Q17[ 2 ], 7 );
neg_xX_Q24[ 3 ] = -silk_LSHIFT32( xX_Q17[ 3 ], 7 );
neg_xX_Q24[ 4 ] = -silk_LSHIFT32( xX_Q17[ 4 ], 7 );
/* Loop over codebook */
*rate_dist_Q8 = silk_int32_MAX;
*res_nrg_Q15 = silk_int32_MAX;
cb_row_Q7 = cb_Q7;
/* In things go really bad, at least *ind is set to something safe. */
*ind = 0;
for( k = 0; k < L; k++ ) {
opus_int32 penalty;
gain_tmp_Q7 = cb_gain_Q7[k];
/* Weighted rate */
/* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */
sum1_Q15 = SILK_FIX_CONST( 1.001, 15 );
/* Penalty for too large gain */
penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 11 );
/* first row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 0 ], XX_Q17[ 1 ], cb_row_Q7[ 1 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 2 ], cb_row_Q7[ 2 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 3 ], cb_row_Q7[ 3 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 4 ], cb_row_Q7[ 4 ] );
sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 0 ], cb_row_Q7[ 0 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 0 ] );
/* second row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 1 ], XX_Q17[ 7 ], cb_row_Q7[ 2 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 8 ], cb_row_Q7[ 3 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 9 ], cb_row_Q7[ 4 ] );
sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 6 ], cb_row_Q7[ 1 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 1 ] );
/* third row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 2 ], XX_Q17[ 13 ], cb_row_Q7[ 3 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 14 ], cb_row_Q7[ 4 ] );
sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 12 ], cb_row_Q7[ 2 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 2 ] );
/* fourth row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 3 ], XX_Q17[ 19 ], cb_row_Q7[ 4 ] );
sum2_Q24 = silk_LSHIFT32( sum2_Q24, 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 18 ], cb_row_Q7[ 3 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 3 ] );
/* last row of XX_Q17 */
sum2_Q24 = silk_LSHIFT32( neg_xX_Q24[ 4 ], 1 );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 24 ], cb_row_Q7[ 4 ] );
sum1_Q15 = silk_SMLAWB( sum1_Q15, sum2_Q24, cb_row_Q7[ 4 ] );
/* find best */
if( sum1_Q15 >= 0 ) {
/* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */
bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penalty) - (15 << 7) );
/* In the following line we reduce the codelength component by half ("-1"); seems to slghtly improve quality */
bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 );
if( bits_tot_Q8 <= *rate_dist_Q8 ) {
*rate_dist_Q8 = bits_tot_Q8;
*res_nrg_Q15 = sum1_Q15 + penalty;
*ind = (opus_int8)k;
*gain_Q7 = gain_tmp_Q7;
}
}
/* Go to next cbk vector */
cb_row_Q7 += LTP_ORDER;
}
}
