/* 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 {
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;
}
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;
}
}
opus_int silk_encode_frame_FLP(
silk_encoder_state_FLP *psEnc, /* I/O Encoder state FLP */
opus_int32 *pnBytesOut, /* O Number of payload bytes; */
ec_enc *psRangeEnc, /* I/O compressor data structure */
opus_int condCoding, /* I The type of conditional coding to use */
opus_int maxBits, /* I If > 0: maximum number of output bits */
opus_int useCBR /* I Flag to force constant-bitrate operation */
)
{
silk_encoder_control_FLP sEncCtrl;
opus_int i, iter, maxIter, found_upper, found_lower, ret = 0;
silk_float *x_frame, *res_pitch_frame;
silk_float xfw[ MAX_FRAME_LENGTH ];
silk_float res_pitch[ 2 * MAX_FRAME_LENGTH + LA_PITCH_MAX ];
ec_enc sRangeEnc_copy, sRangeEnc_copy2;
silk_nsq_state sNSQ_copy, sNSQ_copy2;
opus_int32 seed_copy, nBits, nBits_lower, nBits_upper, gainMult_lower, gainMult_upper;
opus_int32 gainsID, gainsID_lower, gainsID_upper;
opus_int16 gainMult_Q8;
opus_int16 ec_prevLagIndex_copy;
opus_int ec_prevSignalType_copy;
opus_int8 LastGainIndex_copy2;
opus_int32 pGains_Q16[ MAX_NB_SUBFR ];
opus_uint8 ec_buf_copy[ 1275 ];
/* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */
LastGainIndex_copy2 = nBits_lower = nBits_upper = gainMult_lower = gainMult_upper = 0;
psEnc->sCmn.indices.Seed = psEnc->sCmn.frameCounter++ & 3;
/**************************************************************/
/* Set up Input Pointers, and insert frame in input buffer */
/**************************************************************/
/* pointers aligned with start of frame to encode */
x_frame = psEnc->x_buf + psEnc->sCmn.ltp_mem_length; /* start of frame to encode */
res_pitch_frame = res_pitch + psEnc->sCmn.ltp_mem_length; /* start of pitch LPC residual frame */
/***************************************/
/* Ensure smooth bandwidth transitions */
/***************************************/
silk_LP_variable_cutoff( &psEnc->sCmn.sLP, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.frame_length );
/*******************************************/
/* Copy new frame to front of input buffer */
/*******************************************/
silk_short2float_array( x_frame + LA_SHAPE_MS * psEnc->sCmn.fs_kHz, psEnc->sCmn.inputBuf + 1, psEnc->sCmn.frame_length );
/* Add tiny signal to avoid high CPU load from denormalized floating point numbers */
for( i = 0; i < 8; i++ ) {
x_frame[ LA_SHAPE_MS * psEnc->sCmn.fs_kHz + i * ( psEnc->sCmn.frame_length >> 3 ) ] += ( 1 - ( i & 2 ) ) * 1e-6f;
}
if( !psEnc->sCmn.prefillFlag ) {
/*****************************************/
/* Find pitch lags, initial LPC analysis */
/*****************************************/
silk_find_pitch_lags_FLP( psEnc, &sEncCtrl, res_pitch, x_frame );
/************************/
/* Noise shape analysis */
/************************/
silk_noise_shape_analysis_FLP( psEnc, &sEncCtrl, res_pitch_frame, x_frame );
/***************************************************/
/* Find linear prediction coefficients (LPC + LTP) */
/***************************************************/
silk_find_pred_coefs_FLP( psEnc, &sEncCtrl, res_pitch, x_frame, condCoding );
/****************************************/
/* Process gains */
/****************************************/
silk_process_gains_FLP( psEnc, &sEncCtrl, condCoding );
/*****************************************/
/* Prefiltering for noise shaper */
/*****************************************/
silk_prefilter_FLP( psEnc, &sEncCtrl, xfw, x_frame );
/****************************************/
/* Low Bitrate Redundant Encoding */
/****************************************/
silk_LBRR_encode_FLP( psEnc, &sEncCtrl, xfw, condCoding );
/* Loop over quantizer and entroy coding to control bitrate */
maxIter = 6;
gainMult_Q8 = SILK_FIX_CONST( 1, 8 );
found_lower = 0;
found_upper = 0;
gainsID = silk_gains_ID( psEnc->sCmn.indices.GainsIndices, psEnc->sCmn.nb_subfr );
gainsID_lower = -1;
gainsID_upper = -1;
/* Copy part of the input state */
silk_memcpy( &sRangeEnc_copy, psRangeEnc, sizeof( ec_enc ) );
silk_memcpy( &sNSQ_copy, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) );
seed_copy = psEnc->sCmn.indices.Seed;
ec_prevLagIndex_copy = psEnc->sCmn.ec_prevLagIndex;
ec_prevSignalType_copy = psEnc->sCmn.ec_prevSignalType;
for( iter = 0; ; iter++ ) {
if( gainsID == gainsID_lower ) {
nBits = nBits_lower;
} else if( gainsID == gainsID_upper ) {
nBits = nBits_upper;
} else {
/* Restore part of the input state */
if( iter > 0 ) {
silk_memcpy( psRangeEnc, &sRangeEnc_copy, sizeof( ec_enc ) );
silk_memcpy( &psEnc->sCmn.sNSQ, &sNSQ_copy, sizeof( silk_nsq_state ) );
psEnc->sCmn.indices.Seed = seed_copy;
psEnc->sCmn.ec_prevLagIndex = ec_prevLagIndex_copy;
psEnc->sCmn.ec_prevSignalType = ec_prevSignalType_copy;
}
/*****************************************/
/* Noise shaping quantization */
/*****************************************/
silk_NSQ_wrapper_FLP( psEnc, &sEncCtrl, &psEnc->sCmn.indices, &psEnc->sCmn.sNSQ, psEnc->sCmn.pulses, xfw );
/****************************************/
/* Encode Parameters */
/****************************************/
silk_encode_indices( &psEnc->sCmn, psRangeEnc, psEnc->sCmn.nFramesEncoded, 0, condCoding );
/****************************************/
/* Encode Excitation Signal */
/****************************************/
silk_encode_pulses( psRangeEnc, psEnc->sCmn.indices.signalType, psEnc->sCmn.indices.quantOffsetType,
psEnc->sCmn.pulses, psEnc->sCmn.frame_length );
nBits = ec_tell( psRangeEnc );
if( useCBR == 0 && iter == 0 && nBits <= maxBits ) {
break;
}
}
if( iter == maxIter ) {
if( found_lower && ( gainsID == gainsID_lower || nBits > maxBits ) ) {
/* Restore output state from earlier iteration that did meet the bitrate budget */
silk_memcpy( psRangeEnc, &sRangeEnc_copy2, sizeof( ec_enc ) );
silk_assert( sRangeEnc_copy2.offs <= 1275 );
silk_memcpy( psRangeEnc->buf, ec_buf_copy, sRangeEnc_copy2.offs );
silk_memcpy( &psEnc->sCmn.sNSQ, &sNSQ_copy2, sizeof( silk_nsq_state ) );
psEnc->sShape.LastGainIndex = LastGainIndex_copy2;
}
break;
}
if( nBits > maxBits ) {
if( found_lower == 0 && iter >= 2 ) {
/* Adjust the quantizer's rate/distortion tradeoff and discard previous "upper" results */
sEncCtrl.Lambda *= 1.5f;
found_upper = 0;
gainsID_upper = -1;
} else {
found_upper = 1;
nBits_upper = nBits;
gainMult_upper = gainMult_Q8;
gainsID_upper = gainsID;
}
} else if( nBits < maxBits - 5 ) {
found_lower = 1;
nBits_lower = nBits;
gainMult_lower = gainMult_Q8;
if( gainsID != gainsID_lower ) {
gainsID_lower = gainsID;
/* Copy part of the output state */
silk_memcpy( &sRangeEnc_copy2, psRangeEnc, sizeof( ec_enc ) );
silk_assert( psRangeEnc->offs <= 1275 );
silk_memcpy( ec_buf_copy, psRangeEnc->buf, psRangeEnc->offs );
silk_memcpy( &sNSQ_copy2, &psEnc->sCmn.sNSQ, sizeof( silk_nsq_state ) );
LastGainIndex_copy2 = psEnc->sShape.LastGainIndex;
}
} else {
/* Within 5 bits of budget: close enough */
break;
}
if( ( found_lower & found_upper ) == 0 ) {
/* Adjust gain according to high-rate rate/distortion curve */
opus_int32 gain_factor_Q16;
gain_factor_Q16 = silk_log2lin( silk_LSHIFT( nBits - maxBits, 7 ) / psEnc->sCmn.frame_length + SILK_FIX_CONST( 16, 7 ) );
gain_factor_Q16 = silk_min_32( gain_factor_Q16, SILK_FIX_CONST( 2, 16 ) );
if( nBits > maxBits ) {
gain_factor_Q16 = silk_max_32( gain_factor_Q16, SILK_FIX_CONST( 1.3, 16 ) );
}
gainMult_Q8 = silk_SMULWB( gain_factor_Q16, gainMult_Q8 );
} else {
/* Adjust gain by interpolating */
gainMult_Q8 = gainMult_lower + ( ( gainMult_upper - gainMult_lower ) * ( maxBits - nBits_lower ) ) / ( nBits_upper - nBits_lower );
/* New gain multplier must be between 25% and 75% of old range (note that gainMult_upper < gainMult_lower) */
if( gainMult_Q8 > silk_ADD_RSHIFT32( gainMult_lower, gainMult_upper - gainMult_lower, 2 ) ) {
gainMult_Q8 = silk_ADD_RSHIFT32( gainMult_lower, gainMult_upper - gainMult_lower, 2 );
} else
if( gainMult_Q8 < silk_SUB_RSHIFT32( gainMult_upper, gainMult_upper - gainMult_lower, 2 ) ) {
gainMult_Q8 = silk_SUB_RSHIFT32( gainMult_upper, gainMult_upper - gainMult_lower, 2 );
}
}
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
pGains_Q16[ i ] = silk_LSHIFT_SAT32( silk_SMULWB( sEncCtrl.GainsUnq_Q16[ i ], gainMult_Q8 ), 8 );
}
/* Quantize gains */
psEnc->sShape.LastGainIndex = sEncCtrl.lastGainIndexPrev;
silk_gains_quant( psEnc->sCmn.indices.GainsIndices, pGains_Q16,
&psEnc->sShape.LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr );
/* Unique identifier of gains vector */
gainsID = silk_gains_ID( psEnc->sCmn.indices.GainsIndices, psEnc->sCmn.nb_subfr );
/* Overwrite unquantized gains with quantized gains and convert back to Q0 from Q16 */
for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) {
sEncCtrl.Gains[ i ] = pGains_Q16[ i ] / 65536.0f;
}
}
}
/* Update input buffer */
silk_memmove( psEnc->x_buf, &psEnc->x_buf[ psEnc->sCmn.frame_length ],
( psEnc->sCmn.ltp_mem_length + LA_SHAPE_MS * psEnc->sCmn.fs_kHz ) * sizeof( silk_float ) );
/* Parameters needed for next frame */
psEnc->sCmn.prevLag = sEncCtrl.pitchL[ psEnc->sCmn.nb_subfr - 1 ];
psEnc->sCmn.prevSignalType = psEnc->sCmn.indices.signalType;
/* Exit without entropy coding */
if( psEnc->sCmn.prefillFlag ) {
/* No payload */
*pnBytesOut = 0;
return ret;
}
/****************************************/
/* Finalize payload */
/****************************************/
psEnc->sCmn.first_frame_after_reset = 0;
/* Payload size */
*pnBytesOut = silk_RSHIFT( ec_tell( psRangeEnc ) + 7, 3 );
return ret;
}
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 {
gain = silk_SQRT_APPROX( gain_squared ); /* Q0 */
gain = silk_min( gain, silk_int32_MAX >> 16 );
psEncCtrl->Gains_Q16[ k ] = silk_LSHIFT_SAT32( gain, 16 ); /* Q16 */
}
}
/* Save unquantized gains and gain Index */
silk_memcpy( psEncCtrl->GainsUnq_Q16, psEncCtrl->Gains_Q16, psEnc->sCmn.nb_subfr * sizeof( opus_int32 ) );
psEncCtrl->lastGainIndexPrev = psShapeSt->LastGainIndex;
/* Quantize gains */
silk_gains_quant( psEnc->sCmn.indices.GainsIndices, psEncCtrl->Gains_Q16,
&psShapeSt->LastGainIndex, condCoding == CODE_CONDITIONALLY, psEnc->sCmn.nb_subfr );
/* Set quantizer offset for voiced signals. Larger offset when LTP coding gain is low or tilt is high (ie low-pass) */
if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) {
if( psEncCtrl->LTPredCodGain_Q7 + silk_RSHIFT( psEnc->sCmn.input_tilt_Q15, 8 ) > SILK_FIX_CONST( 1.0, 7 ) ) {
psEnc->sCmn.indices.quantOffsetType = 0;
int silk_encode_frame_FIX(silk_encoder_state_FIX * psEnc, /* I/O Pointer to Silk FIX encoder state */
int32_t * pnBytesOut, /* O Pointer to number of payload bytes; */
ec_enc * psRangeEnc, /* I/O compressor data structure */
int condCoding, /* I The type of conditional coding to use */
int maxBits, /* I If > 0: maximum number of output bits */
int useCBR /* I Flag to force constant-bitrate operation */
) {
silk_encoder_control_FIX sEncCtrl;
int i, iter, maxIter, found_upper, found_lower, ret = 0;
int16_t *x_frame;
ec_enc sRangeEnc_copy, sRangeEnc_copy2;
silk_nsq_state sNSQ_copy, sNSQ_copy2;
int32_t seed_copy, nBits, nBits_lower, nBits_upper, gainMult_lower,
gainMult_upper;
int32_t gainsID, gainsID_lower, gainsID_upper;
int16_t gainMult_Q8;
int16_t ec_prevLagIndex_copy;
int ec_prevSignalType_copy;
int8_t LastGainIndex_copy2;
/* This is totally unnecessary but many compilers (including gcc) are too dumb to realise it */
LastGainIndex_copy2 = nBits_lower = nBits_upper = gainMult_lower =
gainMult_upper = 0;
psEnc->sCmn.indices.Seed = psEnc->sCmn.frameCounter++ & 3;
/**************************************************************/
/* Set up Input Pointers, and insert frame in input buffer */
/*************************************************************/
/* start of frame to encode */
x_frame = psEnc->x_buf + psEnc->sCmn.ltp_mem_length;
/***************************************/
/* Ensure smooth bandwidth transitions */
/***************************************/
silk_LP_variable_cutoff(&psEnc->sCmn.sLP, psEnc->sCmn.inputBuf + 1,
psEnc->sCmn.frame_length);
/*******************************************/
/* Copy new frame to front of input buffer */
/*******************************************/
memcpy(x_frame + LA_SHAPE_MS * psEnc->sCmn.fs_kHz,
psEnc->sCmn.inputBuf + 1,
psEnc->sCmn.frame_length * sizeof(int16_t));
if (!psEnc->sCmn.prefillFlag) {
int16_t *res_pitch_frame;
int16_t res_pitch[psEnc->sCmn.la_pitch +
psEnc->sCmn.frame_length +
psEnc->sCmn.ltp_mem_length];
/* start of pitch LPC residual frame */
res_pitch_frame = res_pitch + psEnc->sCmn.ltp_mem_length;
/*****************************************/
/* Find pitch lags, initial LPC analysis */
/*****************************************/
silk_find_pitch_lags_FIX(psEnc, &sEncCtrl, res_pitch, x_frame,
psEnc->sCmn.arch);
/************************/
/* Noise shape analysis */
/************************/
silk_noise_shape_analysis_FIX(psEnc, &sEncCtrl, res_pitch_frame,
x_frame, psEnc->sCmn.arch);
/***************************************************/
/* Find linear prediction coefficients (LPC + LTP) */
/***************************************************/
silk_find_pred_coefs_FIX(psEnc, &sEncCtrl, res_pitch, x_frame,
condCoding);
/****************************************/
/* Process gains */
/****************************************/
silk_process_gains_FIX(psEnc, &sEncCtrl, condCoding);
/*****************************************/
/* Prefiltering for noise shaper */
/*****************************************/
int32_t xfw_Q3[psEnc->sCmn.frame_length];
silk_prefilter_FIX(psEnc, &sEncCtrl, xfw_Q3, x_frame);
/****************************************/
/* Low Bitrate Redundant Encoding */
/****************************************/
silk_LBRR_encode_FIX(psEnc, &sEncCtrl, xfw_Q3, condCoding);
/* Loop over quantizer and entropy coding to control bitrate */
maxIter = 6;
gainMult_Q8 = SILK_FIX_CONST(1, 8);
found_lower = 0;
found_upper = 0;
gainsID =
silk_gains_ID(psEnc->sCmn.indices.GainsIndices,
psEnc->sCmn.nb_subfr);
gainsID_lower = -1;
gainsID_upper = -1;
/* Copy part of the input state */
memcpy(&sRangeEnc_copy, psRangeEnc, sizeof(ec_enc));
memcpy(&sNSQ_copy, &psEnc->sCmn.sNSQ,
sizeof(silk_nsq_state));
seed_copy = psEnc->sCmn.indices.Seed;
ec_prevLagIndex_copy = psEnc->sCmn.ec_prevLagIndex;
ec_prevSignalType_copy = psEnc->sCmn.ec_prevSignalType;
uint8_t ec_buf_copy[1275];
for (iter = 0;; iter++) {
if (gainsID == gainsID_lower) {
nBits = nBits_lower;
} else if (gainsID == gainsID_upper) {
nBits = nBits_upper;
} else {
/* Restore part of the input state */
if (iter > 0) {
memcpy(psRangeEnc, &sRangeEnc_copy,
sizeof(ec_enc));
memcpy(&psEnc->sCmn.sNSQ,
&sNSQ_copy,
sizeof(silk_nsq_state));
psEnc->sCmn.indices.Seed = seed_copy;
psEnc->sCmn.ec_prevLagIndex =
ec_prevLagIndex_copy;
psEnc->sCmn.ec_prevSignalType =
ec_prevSignalType_copy;
}
/*****************************************/
/* Noise shaping quantization */
/*****************************************/
if (psEnc->sCmn.nStatesDelayedDecision > 1
|| psEnc->sCmn.warping_Q16 > 0) {
silk_NSQ_del_dec(&psEnc->sCmn,
&psEnc->sCmn.sNSQ,
&psEnc->sCmn.indices,
xfw_Q3,
psEnc->sCmn.pulses,
sEncCtrl.
PredCoef_Q12[0],
sEncCtrl.LTPCoef_Q14,
sEncCtrl.AR2_Q13,
sEncCtrl.
HarmShapeGain_Q14,
sEncCtrl.Tilt_Q14,
sEncCtrl.LF_shp_Q14,
sEncCtrl.Gains_Q16,
sEncCtrl.pitchL,
sEncCtrl.Lambda_Q10,
sEncCtrl.
LTP_scale_Q14);
} else {
silk_NSQ(&psEnc->sCmn,
&psEnc->sCmn.sNSQ,
&psEnc->sCmn.indices, xfw_Q3,
psEnc->sCmn.pulses,
sEncCtrl.PredCoef_Q12[0],
sEncCtrl.LTPCoef_Q14,
sEncCtrl.AR2_Q13,
sEncCtrl.HarmShapeGain_Q14,
sEncCtrl.Tilt_Q14,
sEncCtrl.LF_shp_Q14,
sEncCtrl.Gains_Q16,
sEncCtrl.pitchL,
sEncCtrl.Lambda_Q10,
sEncCtrl.LTP_scale_Q14);
}
/****************************************/
/* Encode Parameters */
/****************************************/
silk_encode_indices(&psEnc->sCmn, psRangeEnc,
psEnc->sCmn.nFramesEncoded,
0, condCoding);
/****************************************/
/* Encode Excitation Signal */
/****************************************/
silk_encode_pulses(psRangeEnc,
psEnc->sCmn.indices.
signalType,
psEnc->sCmn.indices.
quantOffsetType,
psEnc->sCmn.pulses,
psEnc->sCmn.frame_length);
nBits = ec_tell(psRangeEnc);
if (useCBR == 0 && iter == 0
&& nBits <= maxBits) {
break;
}
}
if (iter == maxIter) {
if (found_lower
&& (gainsID == gainsID_lower
|| nBits > maxBits)) {
/* Restore output state from earlier iteration that did meet the bitrate budget */
memcpy(psRangeEnc,
&sRangeEnc_copy2,
sizeof(ec_enc));
assert(sRangeEnc_copy2.offs <=
1275);
memcpy(psRangeEnc->buf,
ec_buf_copy,
sRangeEnc_copy2.offs);
memcpy(&psEnc->sCmn.sNSQ,
&sNSQ_copy2,
sizeof(silk_nsq_state));
psEnc->sShape.LastGainIndex =
LastGainIndex_copy2;
}
break;
}
if (nBits > maxBits) {
if (found_lower == 0 && iter >= 2) {
/* Adjust the quantizer's rate/distortion tradeoff and discard previous "upper" results */
sEncCtrl.Lambda_Q10 =
silk_ADD_RSHIFT32(sEncCtrl.
Lambda_Q10,
sEncCtrl.
Lambda_Q10, 1);
found_upper = 0;
gainsID_upper = -1;
} else {
found_upper = 1;
nBits_upper = nBits;
gainMult_upper = gainMult_Q8;
gainsID_upper = gainsID;
}
} else if (nBits < maxBits - 5) {
found_lower = 1;
nBits_lower = nBits;
gainMult_lower = gainMult_Q8;
if (gainsID != gainsID_lower) {
gainsID_lower = gainsID;
/* Copy part of the output state */
memcpy(&sRangeEnc_copy2,
psRangeEnc, sizeof(ec_enc));
assert(psRangeEnc->offs <= 1275);
memcpy(ec_buf_copy,
psRangeEnc->buf,
psRangeEnc->offs);
memcpy(&sNSQ_copy2,
&psEnc->sCmn.sNSQ,
sizeof(silk_nsq_state));
LastGainIndex_copy2 =
psEnc->sShape.LastGainIndex;
}
} else {
/* Within 5 bits of budget: close enough */
break;
}
if ((found_lower & found_upper) == 0) {
/* Adjust gain according to high-rate rate/distortion curve */
int32_t gain_factor_Q16;
gain_factor_Q16 =
silk_log2lin(silk_LSHIFT(nBits - maxBits, 7)
/ psEnc->sCmn.frame_length +
SILK_FIX_CONST(16, 7));
gain_factor_Q16 =
silk_min_32(gain_factor_Q16,
SILK_FIX_CONST(2, 16));
if (nBits > maxBits) {
gain_factor_Q16 =
silk_max_32(gain_factor_Q16,
SILK_FIX_CONST(1.3,
16));
}
gainMult_Q8 =
silk_SMULWB(gain_factor_Q16, gainMult_Q8);
} else {
/* Adjust gain by interpolating */
assert(nBits_upper != nBits_lower);
gainMult_Q8 =
gainMult_lower +
silk_DIV32_16(silk_MUL
(gainMult_upper -
gainMult_lower,
maxBits - nBits_lower),
nBits_upper - nBits_lower);
/* New gain multplier must be between 25% and 75% of old range (note that gainMult_upper < gainMult_lower) */
if (gainMult_Q8 >
silk_ADD_RSHIFT32(gainMult_lower,
gainMult_upper -
gainMult_lower, 2)) {
gainMult_Q8 =
silk_ADD_RSHIFT32(gainMult_lower,
gainMult_upper -
gainMult_lower,
2);
} else if (gainMult_Q8 <
silk_SUB_RSHIFT32(gainMult_upper,
gainMult_upper -
gainMult_lower,
2)) {
gainMult_Q8 =
silk_SUB_RSHIFT32(gainMult_upper,
gainMult_upper -
gainMult_lower,
2);
}
}
for (i = 0; i < psEnc->sCmn.nb_subfr; i++) {
sEncCtrl.Gains_Q16[i] =
silk_LSHIFT_SAT32(silk_SMULWB
(sEncCtrl.GainsUnq_Q16[i],
gainMult_Q8), 8);
}
/* Quantize gains */
psEnc->sShape.LastGainIndex =
sEncCtrl.lastGainIndexPrev;
silk_gains_quant(psEnc->sCmn.indices.GainsIndices,
sEncCtrl.Gains_Q16,
&psEnc->sShape.LastGainIndex,
condCoding == CODE_CONDITIONALLY,
psEnc->sCmn.nb_subfr);
/* Unique identifier of gains vector */
gainsID =
silk_gains_ID(psEnc->sCmn.indices.GainsIndices,
psEnc->sCmn.nb_subfr);
}
}
/* Update input buffer */
memmove(psEnc->x_buf, &psEnc->x_buf[psEnc->sCmn.frame_length],
(psEnc->sCmn.ltp_mem_length + LA_SHAPE_MS * psEnc->sCmn.fs_kHz) * sizeof(int16_t));
/* Exit without entropy coding */
if (psEnc->sCmn.prefillFlag) {
/* No payload */
*pnBytesOut = 0;
return ret;
}
/* Parameters needed for next frame */
psEnc->sCmn.prevLag = sEncCtrl.pitchL[psEnc->sCmn.nb_subfr - 1];
psEnc->sCmn.prevSignalType = psEnc->sCmn.indices.signalType;
/****************************************/
/* Finalize payload */
/****************************************/
psEnc->sCmn.first_frame_after_reset = 0;
/* Payload size */
*pnBytesOut = silk_RSHIFT(ec_tell(psRangeEnc) + 7, 3);
return ret;
}
