opus_int silk_VAD_Init( /* O Return value, 0 if success */ silk_VAD_state *psSilk_VAD /* I/O Pointer to Silk VAD state */ ) { opus_int b, ret = 0; /* reset state memory */ silk_memset( psSilk_VAD, 0, sizeof( silk_VAD_state ) ); /* init noise levels */ /* Initialize array with approx pink noise levels (psd proportional to inverse of frequency) */ for( b = 0; b < VAD_N_BANDS; b++ ) { psSilk_VAD->NoiseLevelBias[ b ] = silk_max_32( silk_DIV32_16( VAD_NOISE_LEVELS_BIAS, b + 1 ), 1 ); } /* Initialize state */ for( b = 0; b < VAD_N_BANDS; b++ ) { psSilk_VAD->NL[ b ] = silk_MUL( 100, psSilk_VAD->NoiseLevelBias[ b ] ); psSilk_VAD->inv_NL[ b ] = silk_DIV32( silk_int32_MAX, psSilk_VAD->NL[ b ] ); } psSilk_VAD->counter = 15; /* init smoothed energy-to-noise ratio*/ for( b = 0; b < VAD_N_BANDS; b++ ) { psSilk_VAD->NrgRatioSmth_Q8[ b ] = 100 * 256; /* 100 * 256 --> 20 dB SNR */ } return( ret ); }
/* 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; } }
/* Control SNR of redidual quantizer */ opus_int silk_control_SNR( silk_encoder_state *psEncC, /* I/O Pointer to Silk encoder state */ opus_int32 TargetRate_bps /* I Target max bitrate (bps) */ ) { opus_int k, ret = SILK_NO_ERROR; opus_int32 frac_Q6; const opus_int32 *rateTable; /* Set bitrate/coding quality */ TargetRate_bps = silk_LIMIT(TargetRate_bps, MIN_TARGET_RATE_BPS, MAX_TARGET_RATE_BPS); if (TargetRate_bps != psEncC->TargetRate_bps) { psEncC->TargetRate_bps = TargetRate_bps; /* If new TargetRate_bps, translate to SNR_dB value */ if (psEncC->fs_kHz == 8) { rateTable = silk_TargetRate_table_NB; } else if (psEncC->fs_kHz == 12) { rateTable = silk_TargetRate_table_MB; } else { rateTable = silk_TargetRate_table_WB; } /* Reduce bitrate for 10 ms modes in these calculations */ if (psEncC->nb_subfr == 2) { TargetRate_bps -= REDUCE_BITRATE_10_MS_BPS; } /* Find bitrate interval in table and interpolate */ for (k = 1; k < TARGET_RATE_TAB_SZ; k++) { if (TargetRate_bps <= rateTable[k]) { frac_Q6 = silk_DIV32(silk_LSHIFT(TargetRate_bps - rateTable[k - 1], 6), rateTable[k] - rateTable[k - 1]); psEncC->SNR_dB_Q7 = silk_LSHIFT(silk_SNR_table_Q1[k - 1], 6) + silk_MUL(frac_Q6, silk_SNR_table_Q1[k] - silk_SNR_table_Q1[ k - 1]); break; } } /* Reduce coding quality whenever LBRR is enabled, to free up some bits */ if (psEncC->LBRR_enabled) { psEncC->SNR_dB_Q7 = silk_SMLABB(psEncC->SNR_dB_Q7, 12 - psEncC->LBRR_GainIncreases, SILK_FIX_CONST(-0.25, 7)); } } return ret; }
/* Convert int32 coefficients to int16 coefs and make sure there's no wrap-around */ void silk_LPC_fit( opus_int16 *a_QOUT, /* O Output signal */ opus_int32 *a_QIN, /* I/O Input signal */ const opus_int QOUT, /* I Input Q domain */ const opus_int QIN, /* I Input Q domain */ const opus_int d /* I Filter order */ ) { opus_int i, k, idx = 0; opus_int32 maxabs, absval, chirp_Q16; /* Limit the maximum absolute value of the prediction coefficients, so that they'll fit in int16 */ for( i = 0; i < 10; i++ ) { /* Find maximum absolute value and its index */ maxabs = 0; for( k = 0; k < d; k++ ) { absval = silk_abs( a_QIN[k] ); if( absval > maxabs ) { maxabs = absval; idx = k; } } maxabs = silk_RSHIFT_ROUND( maxabs, QIN - QOUT ); if( maxabs > silk_int16_MAX ) { /* Reduce magnitude of prediction coefficients */ maxabs = silk_min( maxabs, 163838 ); /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */ chirp_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ), silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) ); silk_bwexpander_32( a_QIN, d, chirp_Q16 ); } else { break; } } if( i == 10 ) { /* Reached the last iteration, clip the coefficients */ for( k = 0; k < d; k++ ) { a_QOUT[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT ) ); a_QIN[ k ] = silk_LSHIFT( (opus_int32)a_QOUT[ k ], QIN - QOUT ); } } else { for( k = 0; k < d; k++ ) { a_QOUT[ k ] = (opus_int16)silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT ); } } }
static OPUS_INLINE void silk_PLC_update( silk_decoder_state *psDec, /* I/O Decoder state */ silk_decoder_control *psDecCtrl /* I/O Decoder control */ ) { opus_int32 LTP_Gain_Q14, temp_LTP_Gain_Q14; opus_int i, j; silk_PLC_struct *psPLC; psPLC = &psDec->sPLC; /* Update parameters used in case of packet loss */ psDec->prevSignalType = psDec->indices.signalType; LTP_Gain_Q14 = 0; if( psDec->indices.signalType == TYPE_VOICED ) { /* Find the parameters for the last subframe which contains a pitch pulse */ for( j = 0; j * psDec->subfr_length < psDecCtrl->pitchL[ psDec->nb_subfr - 1 ]; j++ ) { if( j == psDec->nb_subfr ) { break; } temp_LTP_Gain_Q14 = 0; for( i = 0; i < LTP_ORDER; i++ ) { temp_LTP_Gain_Q14 += psDecCtrl->LTPCoef_Q14[ ( psDec->nb_subfr - 1 - j ) * LTP_ORDER + i ]; } if( temp_LTP_Gain_Q14 > LTP_Gain_Q14 ) { LTP_Gain_Q14 = temp_LTP_Gain_Q14; silk_memcpy( psPLC->LTPCoef_Q14, &psDecCtrl->LTPCoef_Q14[ silk_SMULBB( psDec->nb_subfr - 1 - j, LTP_ORDER ) ], LTP_ORDER * sizeof( opus_int16 ) ); psPLC->pitchL_Q8 = silk_LSHIFT( psDecCtrl->pitchL[ psDec->nb_subfr - 1 - j ], 8 ); } } silk_memset( psPLC->LTPCoef_Q14, 0, LTP_ORDER * sizeof( opus_int16 ) ); psPLC->LTPCoef_Q14[ LTP_ORDER / 2 ] = LTP_Gain_Q14; /* Limit LT coefs */ if( LTP_Gain_Q14 < V_PITCH_GAIN_START_MIN_Q14 ) { opus_int scale_Q10; opus_int32 tmp; tmp = silk_LSHIFT( V_PITCH_GAIN_START_MIN_Q14, 10 ); scale_Q10 = silk_DIV32( tmp, silk_max( LTP_Gain_Q14, 1 ) ); for( i = 0; i < LTP_ORDER; i++ ) { psPLC->LTPCoef_Q14[ i ] = silk_RSHIFT( silk_SMULBB( psPLC->LTPCoef_Q14[ i ], scale_Q10 ), 10 ); } } else if( LTP_Gain_Q14 > V_PITCH_GAIN_START_MAX_Q14 ) { opus_int scale_Q14; opus_int32 tmp; tmp = silk_LSHIFT( V_PITCH_GAIN_START_MAX_Q14, 14 ); scale_Q14 = silk_DIV32( tmp, silk_max( LTP_Gain_Q14, 1 ) ); for( i = 0; i < LTP_ORDER; i++ ) { psPLC->LTPCoef_Q14[ i ] = silk_RSHIFT( silk_SMULBB( psPLC->LTPCoef_Q14[ i ], scale_Q14 ), 14 ); } } } else { psPLC->pitchL_Q8 = silk_LSHIFT( silk_SMULBB( psDec->fs_kHz, 18 ), 8 ); silk_memset( psPLC->LTPCoef_Q14, 0, LTP_ORDER * sizeof( opus_int16 )); } /* Save LPC coeficients */ silk_memcpy( psPLC->prevLPC_Q12, psDecCtrl->PredCoef_Q12[ 1 ], psDec->LPC_order * sizeof( opus_int16 ) ); psPLC->prevLTP_scale_Q14 = psDecCtrl->LTP_scale_Q14; /* Save last two gains */ silk_memcpy( psPLC->prevGain_Q16, &psDecCtrl->Gains_Q16[ psDec->nb_subfr - 2 ], 2 * sizeof( opus_int32 ) ); psPLC->subfr_length = psDec->subfr_length; psPLC->nb_subfr = psDec->nb_subfr; }
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; } }
/* If not all roots are found, the a_Q16 coefficients are bandwidth expanded until convergence. */ void silk_A2NLSF( opus_int16 *NLSF, /* O Normalized Line Spectral Frequencies in Q15 (0..2^15-1) [d] */ opus_int32 *a_Q16, /* I/O Monic whitening filter coefficients in Q16 [d] */ const opus_int d /* I Filter order (must be even) */ ) { opus_int i, k, m, dd, root_ix, ffrac; opus_int32 xlo, xhi, xmid; opus_int32 ylo, yhi, ymid, thr; opus_int32 nom, den; opus_int32 P[ SILK_MAX_ORDER_LPC / 2 + 1 ]; opus_int32 Q[ SILK_MAX_ORDER_LPC / 2 + 1 ]; opus_int32 *PQ[ 2 ]; opus_int32 *p; /* Store pointers to array */ PQ[ 0 ] = P; PQ[ 1 ] = Q; dd = silk_RSHIFT( d, 1 ); silk_A2NLSF_init( a_Q16, P, Q, dd ); /* Find roots, alternating between P and Q */ p = P; /* Pointer to polynomial */ xlo = silk_LSFCosTab_FIX_Q12[ 0 ]; /* Q12*/ ylo = silk_A2NLSF_eval_poly( p, xlo, dd ); if( ylo < 0 ) { /* Set the first NLSF to zero and move on to the next */ NLSF[ 0 ] = 0; p = Q; /* Pointer to polynomial */ ylo = silk_A2NLSF_eval_poly( p, xlo, dd ); root_ix = 1; /* Index of current root */ } else { root_ix = 0; /* Index of current root */ } k = 1; /* Loop counter */ i = 0; /* Counter for bandwidth expansions applied */ thr = 0; while( 1 ) { /* Evaluate polynomial */ xhi = silk_LSFCosTab_FIX_Q12[ k ]; /* Q12 */ yhi = silk_A2NLSF_eval_poly( p, xhi, dd ); /* Detect zero crossing */ if( ( ylo <= 0 && yhi >= thr ) || ( ylo >= 0 && yhi <= -thr ) ) { if( yhi == 0 ) { /* If the root lies exactly at the end of the current */ /* interval, look for the next root in the next interval */ thr = 1; } else { thr = 0; } /* Binary division */ ffrac = -256; for( m = 0; m < BIN_DIV_STEPS_A2NLSF_FIX; m++ ) { /* Evaluate polynomial */ xmid = silk_RSHIFT_ROUND( xlo + xhi, 1 ); ymid = silk_A2NLSF_eval_poly( p, xmid, dd ); /* Detect zero crossing */ if( ( ylo <= 0 && ymid >= 0 ) || ( ylo >= 0 && ymid <= 0 ) ) { /* Reduce frequency */ xhi = xmid; yhi = ymid; } else { /* Increase frequency */ xlo = xmid; ylo = ymid; ffrac = silk_ADD_RSHIFT( ffrac, 128, m ); } } /* Interpolate */ if( silk_abs( ylo ) < 65536 ) { /* Avoid dividing by zero */ den = ylo - yhi; nom = silk_LSHIFT( ylo, 8 - BIN_DIV_STEPS_A2NLSF_FIX ) + silk_RSHIFT( den, 1 ); if( den != 0 ) { ffrac += silk_DIV32( nom, den ); } } else { /* No risk of dividing by zero because abs(ylo - yhi) >= abs(ylo) >= 65536 */ ffrac += silk_DIV32( ylo, silk_RSHIFT( ylo - yhi, 8 - BIN_DIV_STEPS_A2NLSF_FIX ) ); } NLSF[ root_ix ] = (opus_int16)silk_min_32( silk_LSHIFT( (opus_int32)k, 8 ) + ffrac, silk_int16_MAX ); silk_assert( NLSF[ root_ix ] >= 0 ); root_ix++; /* Next root */ if( root_ix >= d ) { /* Found all roots */ break; } /* Alternate pointer to polynomial */ p = PQ[ root_ix & 1 ]; /* Evaluate polynomial */ xlo = silk_LSFCosTab_FIX_Q12[ k - 1 ]; /* Q12*/ ylo = silk_LSHIFT( 1 - ( root_ix & 2 ), 12 ); } else { /* Increment loop counter */ k++; xlo = xhi; ylo = yhi; thr = 0; if( k > LSF_COS_TAB_SZ_FIX ) { i++; if( i > MAX_ITERATIONS_A2NLSF_FIX ) { /* Set NLSFs to white spectrum and exit */ NLSF[ 0 ] = (opus_int16)silk_DIV32_16( 1 << 15, d + 1 ); for( k = 1; k < d; k++ ) { NLSF[ k ] = (opus_int16)silk_SMULBB( k + 1, NLSF[ 0 ] ); } return; } /* Error: Apply progressively more bandwidth expansion and run again */ silk_bwexpander_32( a_Q16, d, 65536 - silk_SMULBB( 10 + i, i ) ); /* 10_Q16 = 0.00015*/ silk_A2NLSF_init( a_Q16, P, Q, dd ); p = P; /* Pointer to polynomial */ xlo = silk_LSFCosTab_FIX_Q12[ 0 ]; /* Q12*/ ylo = silk_A2NLSF_eval_poly( p, xlo, dd ); if( ylo < 0 ) { /* Set the first NLSF to zero and move on to the next */ NLSF[ 0 ] = 0; p = Q; /* Pointer to polynomial */ ylo = silk_A2NLSF_eval_poly( p, xlo, dd ); root_ix = 1; /* Index of current root */ } else { root_ix = 0; /* Index of current root */ } k = 1; /* Reset loop counter */ } } } }
/* Initialize/reset the resampler state for a given pair of input/output sampling rates */ opus_int silk_resampler_init( silk_resampler_state_struct *S, /* I/O Resampler state */ opus_int32 Fs_Hz_in, /* I Input sampling rate (Hz) */ opus_int32 Fs_Hz_out, /* I Output sampling rate (Hz) */ opus_int forEnc /* I If 1: encoder; if 0: decoder */ ) { opus_int up2x; /* Clear state */ silk_memset( S, 0, sizeof( silk_resampler_state_struct ) ); /* Input checking */ if( forEnc ) { if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 && Fs_Hz_in != 24000 && Fs_Hz_in != 48000 ) || ( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 ) ) { silk_assert( 0 ); return -1; } S->inputDelay = delay_matrix_enc[ rateID( Fs_Hz_in ) ][ rateID( Fs_Hz_out ) ]; } else { if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 ) || ( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 && Fs_Hz_out != 24000 && Fs_Hz_out != 48000 ) ) { silk_assert( 0 ); return -1; } S->inputDelay = delay_matrix_dec[ rateID( Fs_Hz_in ) ][ rateID( Fs_Hz_out ) ]; } S->Fs_in_kHz = silk_DIV32_16( Fs_Hz_in, 1000 ); S->Fs_out_kHz = silk_DIV32_16( Fs_Hz_out, 1000 ); /* Number of samples processed per batch */ S->batchSize = S->Fs_in_kHz * RESAMPLER_MAX_BATCH_SIZE_MS; /* Find resampler with the right sampling ratio */ up2x = 0; if( Fs_Hz_out > Fs_Hz_in ) { /* Upsample */ if( Fs_Hz_out == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 1 */ /* Special case: directly use 2x upsampler */ S->resampler_function = USE_silk_resampler_private_up2_HQ_wrapper; } else { /* Default resampler */ S->resampler_function = USE_silk_resampler_private_IIR_FIR; up2x = 1; } } else if ( Fs_Hz_out < Fs_Hz_in ) { /* Downsample */ S->resampler_function = USE_silk_resampler_private_down_FIR; if( silk_MUL( Fs_Hz_out, 4 ) == silk_MUL( Fs_Hz_in, 3 ) ) { /* Fs_out : Fs_in = 3 : 4 */ S->FIR_Fracs = 3; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR0; S->Coefs = silk_Resampler_3_4_COEFS; } else if( silk_MUL( Fs_Hz_out, 3 ) == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 3 */ S->FIR_Fracs = 2; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR0; S->Coefs = silk_Resampler_2_3_COEFS; } else if( silk_MUL( Fs_Hz_out, 2 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 2 */ S->FIR_Fracs = 1; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR1; S->Coefs = silk_Resampler_1_2_COEFS; } else if( silk_MUL( Fs_Hz_out, 3 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 3 */ S->FIR_Fracs = 1; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2; S->Coefs = silk_Resampler_1_3_COEFS; } else if( silk_MUL( Fs_Hz_out, 4 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 4 */ S->FIR_Fracs = 1; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2; S->Coefs = silk_Resampler_1_4_COEFS; } else if( silk_MUL( Fs_Hz_out, 6 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 6 */ S->FIR_Fracs = 1; S->FIR_Order = RESAMPLER_DOWN_ORDER_FIR2; S->Coefs = silk_Resampler_1_6_COEFS; } else { /* None available */ silk_assert( 0 ); return -1; } } else { /* Input and output sampling rates are equal: copy */ S->resampler_function = USE_silk_resampler_copy; } /* Ratio of input/output samples */ S->invRatio_Q16 = silk_LSHIFT32( silk_DIV32( silk_LSHIFT32( Fs_Hz_in, 14 + up2x ), Fs_Hz_out ), 2 ); /* Make sure the ratio is rounded up */ while( silk_SMULWW( S->invRatio_Q16, Fs_Hz_out ) < silk_LSHIFT32( Fs_Hz_in, up2x ) ) { S->invRatio_Q16++; } return 0; }
/* compute whitening filter coefficients from normalized line spectral frequencies */ void silk_NLSF2A( opus_int16 *a_Q12, /* O monic whitening filter coefficients in Q12, [ d ] */ const opus_int16 *NLSF, /* I normalized line spectral frequencies in Q15, [ d ] */ const opus_int d /* I filter order (should be even) */ ) { /* This ordering was found to maximize quality. It improves numerical accuracy of silk_NLSF2A_find_poly() compared to "standard" ordering. */ static const unsigned char ordering16[16] = { 0, 15, 8, 7, 4, 11, 12, 3, 2, 13, 10, 5, 6, 9, 14, 1 }; static const unsigned char ordering10[10] = { 0, 9, 6, 3, 4, 5, 8, 1, 2, 7 }; const unsigned char *ordering; opus_int k, i, dd; opus_int32 cos_LSF_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 P[ SILK_MAX_ORDER_LPC / 2 + 1 ], Q[ SILK_MAX_ORDER_LPC / 2 + 1 ]; opus_int32 Ptmp, Qtmp, f_int, f_frac, cos_val, delta; opus_int32 a32_QA1[ SILK_MAX_ORDER_LPC ]; opus_int32 maxabs, absval, idx=0, sc_Q16; silk_assert( LSF_COS_TAB_SZ_FIX == 128 ); silk_assert( d==10||d==16 ); /* convert LSFs to 2*cos(LSF), using piecewise linear curve from table */ ordering = d == 16 ? ordering16 : ordering10; for( k = 0; k < d; k++ ) { silk_assert(NLSF[k] >= 0 ); /* f_int on a scale 0-127 (rounded down) */ f_int = silk_RSHIFT( NLSF[k], 15 - 7 ); /* f_frac, range: 0..255 */ f_frac = NLSF[k] - silk_LSHIFT( f_int, 15 - 7 ); silk_assert(f_int >= 0); silk_assert(f_int < LSF_COS_TAB_SZ_FIX ); /* Read start and end value from table */ cos_val = silk_LSFCosTab_FIX_Q12[ f_int ]; /* Q12 */ delta = silk_LSFCosTab_FIX_Q12[ f_int + 1 ] - cos_val; /* Q12, with a range of 0..200 */ /* Linear interpolation */ cos_LSF_QA[ordering[k]] = silk_RSHIFT_ROUND( silk_LSHIFT( cos_val, 8 ) + silk_MUL( delta, f_frac ), 20 - QA ); /* QA */ } dd = silk_RSHIFT( d, 1 ); /* generate even and odd polynomials using convolution */ silk_NLSF2A_find_poly( P, &cos_LSF_QA[ 0 ], dd ); silk_NLSF2A_find_poly( Q, &cos_LSF_QA[ 1 ], dd ); /* convert even and odd polynomials to opus_int32 Q12 filter coefs */ for( k = 0; k < dd; k++ ) { Ptmp = P[ k+1 ] + P[ k ]; Qtmp = Q[ k+1 ] - Q[ k ]; /* the Ptmp and Qtmp values at this stage need to fit in int32 */ a32_QA1[ k ] = -Qtmp - Ptmp; /* QA+1 */ a32_QA1[ d-k-1 ] = Qtmp - Ptmp; /* QA+1 */ } /* Limit the maximum absolute value of the prediction coefficients, so that they'll fit in int16 */ for( i = 0; i < 10; i++ ) { /* Find maximum absolute value and its index */ maxabs = 0; for( k = 0; k < d; k++ ) { absval = silk_abs( a32_QA1[k] ); if( absval > maxabs ) { maxabs = absval; idx = k; } } maxabs = silk_RSHIFT_ROUND( maxabs, QA + 1 - 12 ); /* QA+1 -> Q12 */ if( maxabs > silk_int16_MAX ) { /* Reduce magnitude of prediction coefficients */ maxabs = silk_min( maxabs, 163838 ); /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */ sc_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ), silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) ); silk_bwexpander_32( a32_QA1, d, sc_Q16 ); } else { break; } } if( i == 10 ) { /* Reached the last iteration, clip the coefficients */ for( k = 0; k < d; k++ ) { a_Q12[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ) ); /* QA+1 -> Q12 */ a32_QA1[ k ] = silk_LSHIFT( (opus_int32)a_Q12[ k ], QA + 1 - 12 ); } } else { for( k = 0; k < d; k++ ) { a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */ } } for( i = 0; i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) { if( silk_LPC_inverse_pred_gain( a_Q12, d ) < SILK_FIX_CONST( 1.0 / MAX_PREDICTION_POWER_GAIN, 30 ) ) { /* Prediction coefficients are (too close to) unstable; apply bandwidth expansion */ /* on the unscaled coefficients, convert to Q12 and measure again */ silk_bwexpander_32( a32_QA1, d, 65536 - silk_LSHIFT( 2, i ) ); for( k = 0; k < d; k++ ) { a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 ); /* QA+1 -> Q12 */ } } else { break; } } }
void silk_find_pred_coefs_FIX( silk_encoder_state_FIX *psEnc, /* I/O encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ const opus_int16 res_pitch[], /* I Residual from pitch analysis */ const opus_int16 x[], /* I Speech signal */ opus_int condCoding /* I The type of conditional coding to use */ ) { opus_int i; opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const opus_int16 *x_ptr; opus_int16 *x_pre_ptr; VARDECL( opus_int16, LPC_in_pre ); opus_int32 tmp, min_gain_Q16, minInvGain_Q30; opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ]; SAVE_STACK; /* weighting for weighted least squares */ min_gain_Q16 = silk_int32_MAX >> 6; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { min_gain_Q16 = silk_min( min_gain_Q16, psEncCtrl->Gains_Q16[ i ] ); } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { /* Divide to Q16 */ silk_assert( psEncCtrl->Gains_Q16[ i ] > 0 ); /* Invert and normalize gains, and ensure that maximum invGains_Q16 is within range of a 16 bit int */ invGains_Q16[ i ] = silk_DIV32_varQ( min_gain_Q16, psEncCtrl->Gains_Q16[ i ], 16 - 2 ); /* Ensure Wght_Q15 a minimum value 1 */ invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 363 ); /* Square the inverted gains */ silk_assert( invGains_Q16[ i ] == silk_SAT16( invGains_Q16[ i ] ) ); tmp = silk_SMULWB( invGains_Q16[ i ], invGains_Q16[ i ] ); Wght_Q15[ i ] = silk_RSHIFT( tmp, 1 ); /* Invert the inverted and normalized gains */ local_gains[ i ] = silk_DIV32( ( (opus_int32)1 << 16 ), invGains_Q16[ i ] ); } ALLOC( LPC_in_pre, psEnc->sCmn.nb_subfr * psEnc->sCmn.predictLPCOrder + psEnc->sCmn.frame_length, opus_int16 ); if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { VARDECL( opus_int32, WLTP ); /**********/ /* VOICED */ /**********/ silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); ALLOC( WLTP, psEnc->sCmn.nb_subfr * LTP_ORDER * LTP_ORDER, opus_int32 ); /* LTP analysis */ silk_find_LTP_FIX( psEncCtrl->LTPCoef_Q14, WLTP, &psEncCtrl->LTPredCodGain_Q7, res_pitch, psEncCtrl->pitchL, Wght_Q15, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length, LTP_corrs_rshift ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains( psEncCtrl->LTPCoef_Q14, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, &psEnc->sCmn.sum_log_gain_Q7, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr); /* Control LTP scaling */ silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding ); /* Create LTP residual */ silk_LTP_analysis_filter_FIX( LPC_in_pre, x - psEnc->sCmn.predictLPCOrder, psEncCtrl->LTPCoef_Q14, psEncCtrl->pitchL, invGains_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); } else { /************/ /* UNVOICED */ /************/ /* Create signal with prepended subframes, scaled by inverse gains */ x_ptr = x - psEnc->sCmn.predictLPCOrder; x_pre_ptr = LPC_in_pre; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { silk_scale_copy_vector16( x_pre_ptr, x_ptr, invGains_Q16[ i ], psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder ); x_pre_ptr += psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder; x_ptr += psEnc->sCmn.subfr_length; } silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) ); psEncCtrl->LTPredCodGain_Q7 = 0; psEnc->sCmn.sum_log_gain_Q7 = 0; } /* Limit on total predictive coding gain */ if( psEnc->sCmn.first_frame_after_reset ) { minInvGain_Q30 = SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET, 30 ); } else { minInvGain_Q30 = silk_log2lin( silk_SMLAWB( 16 << 7, (opus_int32)psEncCtrl->LTPredCodGain_Q7, SILK_FIX_CONST( 1.0 / 3, 16 ) ) ); /* Q16 */ minInvGain_Q30 = silk_DIV32_varQ( minInvGain_Q30, silk_SMULWW( SILK_FIX_CONST( MAX_PREDICTION_POWER_GAIN, 0 ), silk_SMLAWB( SILK_FIX_CONST( 0.25, 18 ), SILK_FIX_CONST( 0.75, 18 ), psEncCtrl->coding_quality_Q14 ) ), 14 ); } /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ silk_find_LPC_FIX( &psEnc->sCmn, NLSF_Q15, LPC_in_pre, minInvGain_Q30 ); /* Quantize LSFs */ silk_process_NLSFs( &psEnc->sCmn, psEncCtrl->PredCoef_Q12, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 ); /* Calculate residual energy using quantized LPC coefficients */ silk_residual_energy_FIX( psEncCtrl->ResNrg, psEncCtrl->ResNrgQ, LPC_in_pre, psEncCtrl->PredCoef_Q12, local_gains, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); /* Copy to prediction struct for use in next frame for interpolation */ silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); RESTORE_STACK; }
/* Compute reflection coefficients from input signal */ void silk_burg_modified_sse4_1( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ const opus_int16 x[], /* I Input signal, length: nb_subfr * (D + subfr_length) */ const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ const opus_int nb_subfr, /* I Number of subframes stacked in x */ const opus_int D, /* I Order */ int arch /* I Run-time architecture */ ) { opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain; opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const opus_int16 *x_ptr; opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 xcorr[ SILK_MAX_ORDER_LPC ]; __m128i FIRST_3210, LAST_3210, ATMP_3210, TMP1_3210, TMP2_3210, T1_3210, T2_3210, PTR_3210, SUBFR_3210, X1_3210, X2_3210; __m128i CONST1 = _mm_set1_epi32(1); silk_assert(subfr_length * nb_subfr <= MAX_FRAME_SIZE); /* Compute autocorrelations, added over subframes */ silk_sum_sqr_shift(&C0, &rshifts, x, nb_subfr * subfr_length); if(rshifts > MAX_RSHIFTS) { C0 = silk_LSHIFT32(C0, rshifts - MAX_RSHIFTS); silk_assert(C0 > 0); rshifts = MAX_RSHIFTS; } else { lz = silk_CLZ32(C0) - 1; rshifts_extra = N_BITS_HEAD_ROOM - lz; if(rshifts_extra > 0) { rshifts_extra = silk_min(rshifts_extra, MAX_RSHIFTS - rshifts); C0 = silk_RSHIFT32(C0, rshifts_extra); } else { rshifts_extra = silk_max(rshifts_extra, MIN_RSHIFTS - rshifts); C0 = silk_LSHIFT32(C0, -rshifts_extra); } rshifts += rshifts_extra; } CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */ silk_memset(C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof(opus_int32)); if(rshifts > 0) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; for(n = 1; n < D + 1; n++) { C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64(x_ptr, x_ptr + n, subfr_length - n, arch), rshifts); } } } else { for(s = 0; s < nb_subfr; s++) { int i; opus_int32 d; x_ptr = x + s * subfr_length; celt_pitch_xcorr(x_ptr, x_ptr + 1, xcorr, subfr_length - D, D, arch); for(n = 1; n < D + 1; n++) { for (i = n + subfr_length - D, d = 0; i < subfr_length; i++) d = MAC16_16(d, x_ptr[ i ], x_ptr[ i - n ]); xcorr[ n - 1 ] += d; } for(n = 1; n < D + 1; n++) { C_first_row[ n - 1 ] += silk_LSHIFT32(xcorr[ n - 1 ], -rshifts); } } } silk_memcpy(C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof(opus_int32)); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0) + 1; /* Q(-rshifts) */ invGain_Q30 = (opus_int32)1 << 30; reached_max_gain = 0; for(n = 0; n < D; n++) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ if(rshifts > -2) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], 16 - rshifts); /* Q(16-rshifts) */ x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts); /* Q(16-rshifts) */ tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], QA - 16); /* Q(QA-16) */ tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16); /* Q(QA-16) */ for(k = 0; k < n; k++) { C_first_row[ k ] = silk_SMLAWB(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */ C_last_row[ k ] = silk_SMLAWB(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */ Atmp_QA = Af_QA[ k ]; tmp1 = silk_SMLAWB(tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ tmp2 = silk_SMLAWB(tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ]); /* Q(QA-16) */ } tmp1 = silk_LSHIFT32(-tmp1, 32 - QA - rshifts); /* Q(16-rshifts) */ tmp2 = silk_LSHIFT32(-tmp2, 32 - QA - rshifts); /* Q(16-rshifts) */ for(k = 0; k <= n; k++) { CAf[ k ] = silk_SMLAWB(CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q(-rshift) */ CAb[ k ] = silk_SMLAWB(CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ]); /* Q(-rshift) */ } } } else { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32((opus_int32)x_ptr[ n ], -rshifts); /* Q(-rshifts) */ x2 = -silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts); /* Q(-rshifts) */ tmp1 = silk_LSHIFT32((opus_int32)x_ptr[ n ], 17); /* Q17 */ tmp2 = silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n - 1 ], 17); /* Q17 */ X1_3210 = _mm_set1_epi32(x1); X2_3210 = _mm_set1_epi32(x2); TMP1_3210 = _mm_setzero_si128(); TMP2_3210 = _mm_setzero_si128(); for(k = 0; k < n - 3; k += 4) { PTR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 1 - 3 ]); SUBFR_3210 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k ]); FIRST_3210 = _mm_loadu_si128((__m128i *)&C_first_row[ k ]); PTR_3210 = _mm_shuffle_epi32(PTR_3210, _MM_SHUFFLE(0, 1, 2, 3)); LAST_3210 = _mm_loadu_si128((__m128i *)&C_last_row[ k ]); ATMP_3210 = _mm_loadu_si128((__m128i *)&Af_QA[ k ]); T1_3210 = _mm_mullo_epi32(PTR_3210, X1_3210); T2_3210 = _mm_mullo_epi32(SUBFR_3210, X2_3210); ATMP_3210 = _mm_srai_epi32(ATMP_3210, 7); ATMP_3210 = _mm_add_epi32(ATMP_3210, CONST1); ATMP_3210 = _mm_srai_epi32(ATMP_3210, 1); FIRST_3210 = _mm_add_epi32(FIRST_3210, T1_3210); LAST_3210 = _mm_add_epi32(LAST_3210, T2_3210); PTR_3210 = _mm_mullo_epi32(ATMP_3210, PTR_3210); SUBFR_3210 = _mm_mullo_epi32(ATMP_3210, SUBFR_3210); _mm_storeu_si128((__m128i *)&C_first_row[ k ], FIRST_3210); _mm_storeu_si128((__m128i *)&C_last_row[ k ], LAST_3210); TMP1_3210 = _mm_add_epi32(TMP1_3210, PTR_3210); TMP2_3210 = _mm_add_epi32(TMP2_3210, SUBFR_3210); } TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_unpackhi_epi64(TMP1_3210, TMP1_3210)); TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_unpackhi_epi64(TMP2_3210, TMP2_3210)); TMP1_3210 = _mm_add_epi32(TMP1_3210, _mm_shufflelo_epi16(TMP1_3210, 0x0E)); TMP2_3210 = _mm_add_epi32(TMP2_3210, _mm_shufflelo_epi16(TMP2_3210, 0x0E)); tmp1 += _mm_cvtsi128_si32(TMP1_3210); tmp2 += _mm_cvtsi128_si32(TMP2_3210); for(; k < n; k++) { C_first_row[ k ] = silk_MLA(C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q(-rshifts) */ C_last_row[ k ] = silk_MLA(C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ]); /* Q(-rshifts) */ Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 17); /* Q17 */ tmp1 = silk_MLA(tmp1, x_ptr[ n - k - 1 ], Atmp1); /* Q17 */ tmp2 = silk_MLA(tmp2, x_ptr[ subfr_length - n + k ], Atmp1); /* Q17 */ } tmp1 = -tmp1; /* Q17 */ tmp2 = -tmp2; /* Q17 */ { __m128i xmm_tmp1, xmm_tmp2; __m128i xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1; __m128i xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1; xmm_tmp1 = _mm_set1_epi32(tmp1); xmm_tmp2 = _mm_set1_epi32(tmp2); for(k = 0; k <= n - 3; k += 4) { xmm_x_ptr_n_k_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ n - k - 3 ]); xmm_x_ptr_sub_x2x0 = OP_CVTEPI16_EPI32_M64(&x_ptr[ subfr_length - n + k - 1 ]); xmm_x_ptr_n_k_x2x0 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 1, 2, 3)); xmm_x_ptr_n_k_x2x0 = _mm_slli_epi32(xmm_x_ptr_n_k_x2x0, -rshifts - 1); xmm_x_ptr_sub_x2x0 = _mm_slli_epi32(xmm_x_ptr_sub_x2x0, -rshifts - 1); /* equal shift right 4 bytes, xmm_x_ptr_n_k_x3x1 = _mm_srli_si128(xmm_x_ptr_n_k_x2x0, 4)*/ xmm_x_ptr_n_k_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_n_k_x2x0, _MM_SHUFFLE(0, 3, 2, 1)); xmm_x_ptr_sub_x3x1 = _mm_shuffle_epi32(xmm_x_ptr_sub_x2x0, _MM_SHUFFLE(0, 3, 2, 1)); xmm_x_ptr_n_k_x2x0 = _mm_mul_epi32(xmm_x_ptr_n_k_x2x0, xmm_tmp1); xmm_x_ptr_n_k_x3x1 = _mm_mul_epi32(xmm_x_ptr_n_k_x3x1, xmm_tmp1); xmm_x_ptr_sub_x2x0 = _mm_mul_epi32(xmm_x_ptr_sub_x2x0, xmm_tmp2); xmm_x_ptr_sub_x3x1 = _mm_mul_epi32(xmm_x_ptr_sub_x3x1, xmm_tmp2); xmm_x_ptr_n_k_x2x0 = _mm_srli_epi64(xmm_x_ptr_n_k_x2x0, 16); xmm_x_ptr_n_k_x3x1 = _mm_slli_epi64(xmm_x_ptr_n_k_x3x1, 16); xmm_x_ptr_sub_x2x0 = _mm_srli_epi64(xmm_x_ptr_sub_x2x0, 16); xmm_x_ptr_sub_x3x1 = _mm_slli_epi64(xmm_x_ptr_sub_x3x1, 16); xmm_x_ptr_n_k_x2x0 = _mm_blend_epi16(xmm_x_ptr_n_k_x2x0, xmm_x_ptr_n_k_x3x1, 0xCC); xmm_x_ptr_sub_x2x0 = _mm_blend_epi16(xmm_x_ptr_sub_x2x0, xmm_x_ptr_sub_x3x1, 0xCC); X1_3210 = _mm_loadu_si128((__m128i *)&CAf[ k ]); PTR_3210 = _mm_loadu_si128((__m128i *)&CAb[ k ]); X1_3210 = _mm_add_epi32(X1_3210, xmm_x_ptr_n_k_x2x0); PTR_3210 = _mm_add_epi32(PTR_3210, xmm_x_ptr_sub_x2x0); _mm_storeu_si128((__m128i *)&CAf[ k ], X1_3210); _mm_storeu_si128((__m128i *)&CAb[ k ], PTR_3210); } for(; k <= n; k++) { CAf[ k ] = silk_SMLAWW(CAf[ k ], tmp1, silk_LSHIFT32((opus_int32)x_ptr[ n - k ], -rshifts - 1)); /* Q(-rshift) */ CAb[ k ] = silk_SMLAWW(CAb[ k ], tmp2, silk_LSHIFT32((opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1)); /* Q(-rshift) */ } } } } /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ tmp1 = C_first_row[ n ]; /* Q(-rshifts) */ tmp2 = C_last_row[ n ]; /* Q(-rshifts) */ num = 0; /* Q(-rshifts) */ nrg = silk_ADD32(CAb[ 0 ], CAf[ 0 ]); /* Q(1-rshifts) */ for(k = 0; k < n; k++) { Atmp_QA = Af_QA[ k ]; lz = silk_CLZ32(silk_abs(Atmp_QA)) - 1; lz = silk_min(32 - QA, lz); Atmp1 = silk_LSHIFT32(Atmp_QA, lz); /* Q(QA + lz) */ tmp1 = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(C_last_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ tmp2 = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(C_first_row[ n - k - 1 ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ num = silk_ADD_LSHIFT32(num, silk_SMMUL(CAb[ n - k ], Atmp1), 32 - QA - lz); /* Q(-rshifts) */ nrg = silk_ADD_LSHIFT32(nrg, silk_SMMUL(silk_ADD32(CAb[ k + 1 ], CAf[ k + 1 ]), Atmp1), 32 - QA - lz); /* Q(1-rshifts) */ } CAf[ n + 1 ] = tmp1; /* Q(-rshifts) */ CAb[ n + 1 ] = tmp2; /* Q(-rshifts) */ num = silk_ADD32(num, tmp2); /* Q(-rshifts) */ num = silk_LSHIFT32(-num, 1); /* Q(1-rshifts) */ /* Calculate the next order reflection (parcor) coefficient */ if(silk_abs(num) < nrg) { rc_Q31 = silk_DIV32_varQ(num, nrg, 31); } else { rc_Q31 = (num > 0) ? silk_int32_MAX : silk_int32_MIN; } /* Update inverse prediction gain */ tmp1 = ((opus_int32)1 << 30) - silk_SMMUL(rc_Q31, rc_Q31); tmp1 = silk_LSHIFT(silk_SMMUL(invGain_Q30, tmp1), 2); if(tmp1 <= minInvGain_Q30) { /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ tmp2 = ((opus_int32)1 << 30) - silk_DIV32_varQ(minInvGain_Q30, invGain_Q30, 30); /* Q30 */ rc_Q31 = silk_SQRT_APPROX(tmp2); /* Q15 */ /* Newton-Raphson iteration */ rc_Q31 = silk_RSHIFT32(rc_Q31 + silk_DIV32(tmp2, rc_Q31), 1); /* Q15 */ rc_Q31 = silk_LSHIFT32(rc_Q31, 16); /* Q31 */ if(num < 0) { /* Ensure adjusted reflection coefficients has the original sign */ rc_Q31 = -rc_Q31; } invGain_Q30 = minInvGain_Q30; reached_max_gain = 1; } else { invGain_Q30 = tmp1; } /* Update the AR coefficients */ for(k = 0; k < (n + 1) >> 1; k++) { tmp1 = Af_QA[ k ]; /* QA */ tmp2 = Af_QA[ n - k - 1 ]; /* QA */ Af_QA[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* QA */ Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* QA */ } Af_QA[ n ] = silk_RSHIFT32(rc_Q31, 31 - QA); /* QA */ if(reached_max_gain) { /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ for(k = n + 1; k < D; k++) { Af_QA[ k ] = 0; } break; } /* Update C * Af and C * Ab */ for(k = 0; k <= n + 1; k++) { tmp1 = CAf[ k ]; /* Q(-rshifts) */ tmp2 = CAb[ n - k + 1 ]; /* Q(-rshifts) */ CAf[ k ] = silk_ADD_LSHIFT32(tmp1, silk_SMMUL(tmp2, rc_Q31), 1); /* Q(-rshifts) */ CAb[ n - k + 1 ] = silk_ADD_LSHIFT32(tmp2, silk_SMMUL(tmp1, rc_Q31), 1); /* Q(-rshifts) */ } } if(reached_max_gain) { for(k = 0; k < D; k++) { /* Scale coefficients */ A_Q16[ k ] = -silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); } /* Subtract energy of preceding samples from C0 */ if(rshifts > 0) { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; C0 -= (opus_int32)silk_RSHIFT64(silk_inner_prod16_aligned_64(x_ptr, x_ptr, D, arch), rshifts); } } else { for(s = 0; s < nb_subfr; s++) { x_ptr = x + s * subfr_length; C0 -= silk_LSHIFT32(silk_inner_prod_aligned(x_ptr, x_ptr, D, arch), -rshifts); } } /* Approximate residual energy */ *res_nrg = silk_LSHIFT(silk_SMMUL(invGain_Q30, C0), 2); *res_nrg_Q = -rshifts; } else { /* Return residual energy */ nrg = CAf[ 0 ]; /* Q(-rshifts) */ tmp1 = (opus_int32)1 << 16; /* Q16 */ for(k = 0; k < D; k++) { Atmp1 = silk_RSHIFT_ROUND(Af_QA[ k ], QA - 16); /* Q16 */ nrg = silk_SMLAWW(nrg, CAf[ k + 1 ], Atmp1); /* Q(-rshifts) */ tmp1 = silk_SMLAWW(tmp1, Atmp1, Atmp1); /* Q16 */ A_Q16[ k ] = -Atmp1; } *res_nrg = silk_SMLAWW(nrg, silk_SMMUL(SILK_FIX_CONST(FIND_LPC_COND_FAC, 32), C0), -tmp1);/* Q(-rshifts) */ *res_nrg_Q = -rshifts; } }
/* Compute reflection coefficients from input signal */ void silk_burg_modified( opus_int32 *res_nrg, /* O Residual energy */ opus_int *res_nrg_Q, /* O Residual energy Q value */ opus_int32 A_Q16[], /* O Prediction coefficients (length order) */ const opus_int16 x[], /* I Input signal, length: nb_subfr * ( D + subfr_length ) */ const opus_int32 minInvGain_Q30, /* I Inverse of max prediction gain */ const opus_int subfr_length, /* I Input signal subframe length (incl. D preceding samples) */ const opus_int nb_subfr, /* I Number of subframes stacked in x */ const opus_int D /* I Order */ ) { opus_int k, n, s, lz, rshifts, rshifts_extra, reached_max_gain; opus_int32 C0, num, nrg, rc_Q31, invGain_Q30, Atmp_QA, Atmp1, tmp1, tmp2, x1, x2; const opus_int16 *x_ptr; opus_int32 C_first_row[ SILK_MAX_ORDER_LPC ]; opus_int32 C_last_row[ SILK_MAX_ORDER_LPC ]; opus_int32 Af_QA[ SILK_MAX_ORDER_LPC ]; opus_int32 CAf[ SILK_MAX_ORDER_LPC + 1 ]; opus_int32 CAb[ SILK_MAX_ORDER_LPC + 1 ]; silk_assert( subfr_length * nb_subfr <= MAX_FRAME_SIZE ); /* Compute autocorrelations, added over subframes */ silk_sum_sqr_shift( &C0, &rshifts, x, nb_subfr * subfr_length ); if( rshifts > MAX_RSHIFTS ) { C0 = silk_LSHIFT32( C0, rshifts - MAX_RSHIFTS ); silk_assert( C0 > 0 ); rshifts = MAX_RSHIFTS; } else { lz = silk_CLZ32( C0 ) - 1; rshifts_extra = N_BITS_HEAD_ROOM - lz; if( rshifts_extra > 0 ) { rshifts_extra = silk_min( rshifts_extra, MAX_RSHIFTS - rshifts ); C0 = silk_RSHIFT32( C0, rshifts_extra ); } else { rshifts_extra = silk_max( rshifts_extra, MIN_RSHIFTS - rshifts ); C0 = silk_LSHIFT32( C0, -rshifts_extra ); } rshifts += rshifts_extra; } CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ silk_memset( C_first_row, 0, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); if( rshifts > 0 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr + n, subfr_length - n ), rshifts ); } } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; for( n = 1; n < D + 1; n++ ) { C_first_row[ n - 1 ] += silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr + n, subfr_length - n ), -rshifts ); } } } silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( opus_int32 ) ); /* Initialize */ CAb[ 0 ] = CAf[ 0 ] = C0 + silk_SMMUL( SILK_FIX_CONST( FIND_LPC_COND_FAC, 32 ), C0 ) + 1; /* Q(-rshifts) */ invGain_Q30 = (opus_int32)1 << 30; reached_max_gain = 0; for( n = 0; n < D; n++ ) { /* Update first row of correlation matrix (without first element) */ /* Update last row of correlation matrix (without last element, stored in reversed order) */ /* Update C * Af */ /* Update C * flipud(Af) (stored in reversed order) */ if( rshifts > -2 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], 16 - rshifts ); /* Q(16-rshifts) */ x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 16 - rshifts ); /* Q(16-rshifts) */ tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], QA - 16 ); /* Q(QA-16) */ tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], QA - 16 ); /* Q(QA-16) */ for( k = 0; k < n; k++ ) { C_first_row[ k ] = silk_SMLAWB( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ C_last_row[ k ] = silk_SMLAWB( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Atmp_QA = Af_QA[ k ]; tmp1 = silk_SMLAWB( tmp1, Atmp_QA, x_ptr[ n - k - 1 ] ); /* Q(QA-16) */ tmp2 = silk_SMLAWB( tmp2, Atmp_QA, x_ptr[ subfr_length - n + k ] ); /* Q(QA-16) */ } tmp1 = silk_LSHIFT32( -tmp1, 32 - QA - rshifts ); /* Q(16-rshifts) */ tmp2 = silk_LSHIFT32( -tmp2, 32 - QA - rshifts ); /* Q(16-rshifts) */ for( k = 0; k <= n; k++ ) { CAf[ k ] = silk_SMLAWB( CAf[ k ], tmp1, x_ptr[ n - k ] ); /* Q( -rshift ) */ CAb[ k ] = silk_SMLAWB( CAb[ k ], tmp2, x_ptr[ subfr_length - n + k - 1 ] ); /* Q( -rshift ) */ } } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; x1 = -silk_LSHIFT32( (opus_int32)x_ptr[ n ], -rshifts ); /* Q( -rshifts ) */ x2 = -silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], -rshifts ); /* Q( -rshifts ) */ tmp1 = silk_LSHIFT32( (opus_int32)x_ptr[ n ], 17 ); /* Q17 */ tmp2 = silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n - 1 ], 17 ); /* Q17 */ for( k = 0; k < n; k++ ) { C_first_row[ k ] = silk_MLA( C_first_row[ k ], x1, x_ptr[ n - k - 1 ] ); /* Q( -rshifts ) */ C_last_row[ k ] = silk_MLA( C_last_row[ k ], x2, x_ptr[ subfr_length - n + k ] ); /* Q( -rshifts ) */ Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 17 ); /* Q17 */ tmp1 = silk_MLA( tmp1, x_ptr[ n - k - 1 ], Atmp1 ); /* Q17 */ tmp2 = silk_MLA( tmp2, x_ptr[ subfr_length - n + k ], Atmp1 ); /* Q17 */ } tmp1 = -tmp1; /* Q17 */ tmp2 = -tmp2; /* Q17 */ for( k = 0; k <= n; k++ ) { CAf[ k ] = silk_SMLAWW( CAf[ k ], tmp1, silk_LSHIFT32( (opus_int32)x_ptr[ n - k ], -rshifts - 1 ) ); /* Q( -rshift ) */ CAb[ k ] = silk_SMLAWW( CAb[ k ], tmp2, silk_LSHIFT32( (opus_int32)x_ptr[ subfr_length - n + k - 1 ], -rshifts - 1 ) ); /* Q( -rshift ) */ } } } /* Calculate nominator and denominator for the next order reflection (parcor) coefficient */ tmp1 = C_first_row[ n ]; /* Q( -rshifts ) */ tmp2 = C_last_row[ n ]; /* Q( -rshifts ) */ num = 0; /* Q( -rshifts ) */ nrg = silk_ADD32( CAb[ 0 ], CAf[ 0 ] ); /* Q( 1-rshifts ) */ for( k = 0; k < n; k++ ) { Atmp_QA = Af_QA[ k ]; lz = silk_CLZ32( silk_abs( Atmp_QA ) ) - 1; lz = silk_min( 32 - QA, lz ); Atmp1 = silk_LSHIFT32( Atmp_QA, lz ); /* Q( QA + lz ) */ tmp1 = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( C_last_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ tmp2 = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( C_first_row[ n - k - 1 ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ num = silk_ADD_LSHIFT32( num, silk_SMMUL( CAb[ n - k ], Atmp1 ), 32 - QA - lz ); /* Q( -rshifts ) */ nrg = silk_ADD_LSHIFT32( nrg, silk_SMMUL( silk_ADD32( CAb[ k + 1 ], CAf[ k + 1 ] ), Atmp1 ), 32 - QA - lz ); /* Q( 1-rshifts ) */ } CAf[ n + 1 ] = tmp1; /* Q( -rshifts ) */ CAb[ n + 1 ] = tmp2; /* Q( -rshifts ) */ num = silk_ADD32( num, tmp2 ); /* Q( -rshifts ) */ num = silk_LSHIFT32( -num, 1 ); /* Q( 1-rshifts ) */ /* Calculate the next order reflection (parcor) coefficient */ if( silk_abs( num ) < nrg ) { rc_Q31 = silk_DIV32_varQ( num, nrg, 31 ); } else { rc_Q31 = ( num > 0 ) ? silk_int32_MAX : silk_int32_MIN; } /* Update inverse prediction gain */ tmp1 = ( (opus_int32)1 << 30 ) - silk_SMMUL( rc_Q31, rc_Q31 ); tmp1 = silk_LSHIFT( silk_SMMUL( invGain_Q30, tmp1 ), 2 ); if( tmp1 <= minInvGain_Q30 ) { /* Max prediction gain exceeded; set reflection coefficient such that max prediction gain is exactly hit */ tmp2 = ( (opus_int32)1 << 30 ) - silk_DIV32_varQ( minInvGain_Q30, invGain_Q30, 30 ); /* Q30 */ rc_Q31 = silk_SQRT_APPROX( tmp2 ); /* Q15 */ /* Newton-Raphson iteration */ rc_Q31 = silk_RSHIFT32( rc_Q31 + silk_DIV32( tmp2, rc_Q31 ), 1 ); /* Q15 */ rc_Q31 = silk_LSHIFT32( rc_Q31, 16 ); /* Q31 */ if( num < 0 ) { /* Ensure adjusted reflection coefficients has the original sign */ rc_Q31 = -rc_Q31; } invGain_Q30 = minInvGain_Q30; reached_max_gain = 1; } else { invGain_Q30 = tmp1; } /* Update the AR coefficients */ for( k = 0; k < (n + 1) >> 1; k++ ) { tmp1 = Af_QA[ k ]; /* QA */ tmp2 = Af_QA[ n - k - 1 ]; /* QA */ Af_QA[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* QA */ Af_QA[ n - k - 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* QA */ } Af_QA[ n ] = silk_RSHIFT32( rc_Q31, 31 - QA ); /* QA */ if( reached_max_gain ) { /* Reached max prediction gain; set remaining coefficients to zero and exit loop */ for( k = n + 1; k < D; k++ ) { Af_QA[ k ] = 0; } break; } /* Update C * Af and C * Ab */ for( k = 0; k <= n + 1; k++ ) { tmp1 = CAf[ k ]; /* Q( -rshifts ) */ tmp2 = CAb[ n - k + 1 ]; /* Q( -rshifts ) */ CAf[ k ] = silk_ADD_LSHIFT32( tmp1, silk_SMMUL( tmp2, rc_Q31 ), 1 ); /* Q( -rshifts ) */ CAb[ n - k + 1 ] = silk_ADD_LSHIFT32( tmp2, silk_SMMUL( tmp1, rc_Q31 ), 1 ); /* Q( -rshifts ) */ } } if( reached_max_gain ) { for( k = 0; k < D; k++ ) { /* Scale coefficients */ A_Q16[ k ] = -silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); } /* Subtract energy of preceding samples from C0 */ if( rshifts > 0 ) { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; C0 -= (opus_int32)silk_RSHIFT64( silk_inner_prod16_aligned_64( x_ptr, x_ptr, D ), rshifts ); } } else { for( s = 0; s < nb_subfr; s++ ) { x_ptr = x + s * subfr_length; C0 -= silk_LSHIFT32( silk_inner_prod_aligned( x_ptr, x_ptr, D ), -rshifts ); } } /* Approximate residual energy */ *res_nrg = silk_LSHIFT( silk_SMMUL( invGain_Q30, C0 ), 2 ); *res_nrg_Q = -rshifts; } else { /* Return residual energy */ nrg = CAf[ 0 ]; /* Q( -rshifts ) */ tmp1 = (opus_int32)1 << 16; /* Q16 */ for( k = 0; k < D; k++ ) { Atmp1 = silk_RSHIFT_ROUND( Af_QA[ k ], QA - 16 ); /* Q16 */ nrg = silk_SMLAWW( nrg, CAf[ k + 1 ], Atmp1 ); /* Q( -rshifts ) */ tmp1 = silk_SMLAWW( tmp1, Atmp1, Atmp1 ); /* Q16 */ A_Q16[ k ] = -Atmp1; } *res_nrg = silk_SMLAWW( nrg, silk_SMMUL( FIND_LPC_COND_FAC, C0 ), -tmp1 ); /* Q( -rshifts ) */ *res_nrg_Q = -rshifts; } }
/* encControl->payloadSize_ms is set to */ opus_int silk_Encode( /* O Returns error code */ void *encState, /* I/O State */ silk_EncControlStruct *encControl, /* I Control status */ const opus_int16 *samplesIn, /* I Speech sample input vector */ opus_int nSamplesIn, /* I Number of samples in input vector */ ec_enc *psRangeEnc, /* I/O Compressor data structure */ opus_int32 *nBytesOut, /* I/O Number of bytes in payload (input: Max bytes) */ const opus_int prefillFlag /* I Flag to indicate prefilling buffers no coding */ ) { opus_int n, i, nBits, flags, tmp_payloadSize_ms = 0, tmp_complexity = 0, ret = 0; opus_int nSamplesToBuffer, nSamplesToBufferMax, nBlocksOf10ms; opus_int nSamplesFromInput = 0, nSamplesFromInputMax; opus_int speech_act_thr_for_switch_Q8; opus_int32 TargetRate_bps, MStargetRates_bps[ 2 ], channelRate_bps, LBRR_symbol, sum; silk_encoder *psEnc = ( silk_encoder * )encState; VARDECL( opus_int16, buf ); opus_int transition, curr_block, tot_blocks; SAVE_STACK; if (encControl->reducedDependency) { psEnc->state_Fxx[0].sCmn.first_frame_after_reset = 1; psEnc->state_Fxx[1].sCmn.first_frame_after_reset = 1; } psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded = psEnc->state_Fxx[ 1 ].sCmn.nFramesEncoded = 0; /* Check values in encoder control structure */ if( ( ret = check_control_input( encControl ) ) != 0 ) { silk_assert( 0 ); RESTORE_STACK; return ret; } encControl->switchReady = 0; if( encControl->nChannelsInternal > psEnc->nChannelsInternal ) { /* Mono -> Stereo transition: init state of second channel and stereo state */ ret += silk_init_encoder( &psEnc->state_Fxx[ 1 ], psEnc->state_Fxx[ 0 ].sCmn.arch ); silk_memset( psEnc->sStereo.pred_prev_Q13, 0, sizeof( psEnc->sStereo.pred_prev_Q13 ) ); silk_memset( psEnc->sStereo.sSide, 0, sizeof( psEnc->sStereo.sSide ) ); psEnc->sStereo.mid_side_amp_Q0[ 0 ] = 0; psEnc->sStereo.mid_side_amp_Q0[ 1 ] = 1; psEnc->sStereo.mid_side_amp_Q0[ 2 ] = 0; psEnc->sStereo.mid_side_amp_Q0[ 3 ] = 1; psEnc->sStereo.width_prev_Q14 = 0; psEnc->sStereo.smth_width_Q14 = SILK_FIX_CONST( 1, 14 ); if( psEnc->nChannelsAPI == 2 ) { silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, sizeof( silk_resampler_state_struct ) ); silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.In_HP_State, &psEnc->state_Fxx[ 0 ].sCmn.In_HP_State, sizeof( psEnc->state_Fxx[ 1 ].sCmn.In_HP_State ) ); } } transition = (encControl->payloadSize_ms != psEnc->state_Fxx[ 0 ].sCmn.PacketSize_ms) || (psEnc->nChannelsInternal != encControl->nChannelsInternal); psEnc->nChannelsAPI = encControl->nChannelsAPI; psEnc->nChannelsInternal = encControl->nChannelsInternal; nBlocksOf10ms = silk_DIV32( 100 * nSamplesIn, encControl->API_sampleRate ); tot_blocks = ( nBlocksOf10ms > 1 ) ? nBlocksOf10ms >> 1 : 1; curr_block = 0; if( prefillFlag ) { /* Only accept input length of 10 ms */ if( nBlocksOf10ms != 1 ) { silk_assert( 0 ); RESTORE_STACK; return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; } /* Reset Encoder */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { ret = silk_init_encoder( &psEnc->state_Fxx[ n ], psEnc->state_Fxx[ n ].sCmn.arch ); silk_assert( !ret ); } tmp_payloadSize_ms = encControl->payloadSize_ms; encControl->payloadSize_ms = 10; tmp_complexity = encControl->complexity; encControl->complexity = 0; for( n = 0; n < encControl->nChannelsInternal; n++ ) { psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0; psEnc->state_Fxx[ n ].sCmn.prefillFlag = 1; } } else { /* Only accept input lengths that are a multiple of 10 ms */ if( nBlocksOf10ms * encControl->API_sampleRate != 100 * nSamplesIn || nSamplesIn < 0 ) { silk_assert( 0 ); RESTORE_STACK; return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; } /* Make sure no more than one packet can be produced */ if( 1000 * (opus_int32)nSamplesIn > encControl->payloadSize_ms * encControl->API_sampleRate ) { silk_assert( 0 ); RESTORE_STACK; return SILK_ENC_INPUT_INVALID_NO_OF_SAMPLES; } } TargetRate_bps = silk_RSHIFT32( encControl->bitRate, encControl->nChannelsInternal - 1 ); for( n = 0; n < encControl->nChannelsInternal; n++ ) { /* Force the side channel to the same rate as the mid */ opus_int force_fs_kHz = (n==1) ? psEnc->state_Fxx[0].sCmn.fs_kHz : 0; if( ( ret = silk_control_encoder( &psEnc->state_Fxx[ n ], encControl, TargetRate_bps, psEnc->allowBandwidthSwitch, n, force_fs_kHz ) ) != 0 ) { silk_assert( 0 ); RESTORE_STACK; return ret; } if( psEnc->state_Fxx[n].sCmn.first_frame_after_reset || transition ) { for( i = 0; i < psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket; i++ ) { psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ] = 0; } } psEnc->state_Fxx[ n ].sCmn.inDTX = psEnc->state_Fxx[ n ].sCmn.useDTX; } silk_assert( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 0 ].sCmn.fs_kHz == psEnc->state_Fxx[ 1 ].sCmn.fs_kHz ); /* Input buffering/resampling and encoding */ nSamplesToBufferMax = 10 * nBlocksOf10ms * psEnc->state_Fxx[ 0 ].sCmn.fs_kHz; nSamplesFromInputMax = silk_DIV32_16( nSamplesToBufferMax * psEnc->state_Fxx[ 0 ].sCmn.API_fs_Hz, psEnc->state_Fxx[ 0 ].sCmn.fs_kHz * 1000 ); ALLOC( buf, nSamplesFromInputMax, opus_int16 ); while( 1 ) { nSamplesToBuffer = psEnc->state_Fxx[ 0 ].sCmn.frame_length - psEnc->state_Fxx[ 0 ].sCmn.inputBufIx; nSamplesToBuffer = silk_min( nSamplesToBuffer, nSamplesToBufferMax ); nSamplesFromInput = silk_DIV32_16( nSamplesToBuffer * psEnc->state_Fxx[ 0 ].sCmn.API_fs_Hz, psEnc->state_Fxx[ 0 ].sCmn.fs_kHz * 1000 ); /* Resample and write to buffer */ if( encControl->nChannelsAPI == 2 && encControl->nChannelsInternal == 2 ) { opus_int id = psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded; for( n = 0; n < nSamplesFromInput; n++ ) { buf[ n ] = samplesIn[ 2 * n ]; } /* Making sure to start both resamplers from the same state when switching from mono to stereo */ if( psEnc->nPrevChannelsInternal == 1 && id==0 ) { silk_memcpy( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, sizeof(psEnc->state_Fxx[ 1 ].sCmn.resampler_state)); } ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer; nSamplesToBuffer = psEnc->state_Fxx[ 1 ].sCmn.frame_length - psEnc->state_Fxx[ 1 ].sCmn.inputBufIx; nSamplesToBuffer = silk_min( nSamplesToBuffer, 10 * nBlocksOf10ms * psEnc->state_Fxx[ 1 ].sCmn.fs_kHz ); for( n = 0; n < nSamplesFromInput; n++ ) { buf[ n ] = samplesIn[ 2 * n + 1 ]; } ret += silk_resampler( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); psEnc->state_Fxx[ 1 ].sCmn.inputBufIx += nSamplesToBuffer; } else if( encControl->nChannelsAPI == 2 && encControl->nChannelsInternal == 1 ) { /* Combine left and right channels before resampling */ for( n = 0; n < nSamplesFromInput; n++ ) { sum = samplesIn[ 2 * n ] + samplesIn[ 2 * n + 1 ]; buf[ n ] = (opus_int16)silk_RSHIFT_ROUND( sum, 1 ); } ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); /* On the first mono frame, average the results for the two resampler states */ if( psEnc->nPrevChannelsInternal == 2 && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == 0 ) { ret += silk_resampler( &psEnc->state_Fxx[ 1 ].sCmn.resampler_state, &psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); for( n = 0; n < psEnc->state_Fxx[ 0 ].sCmn.frame_length; n++ ) { psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx+n+2 ] = silk_RSHIFT(psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx+n+2 ] + psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ psEnc->state_Fxx[ 1 ].sCmn.inputBufIx+n+2 ], 1); } } psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer; } else { silk_assert( encControl->nChannelsAPI == 1 && encControl->nChannelsInternal == 1 ); silk_memcpy(buf, samplesIn, nSamplesFromInput*sizeof(opus_int16)); ret += silk_resampler( &psEnc->state_Fxx[ 0 ].sCmn.resampler_state, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.inputBufIx + 2 ], buf, nSamplesFromInput ); psEnc->state_Fxx[ 0 ].sCmn.inputBufIx += nSamplesToBuffer; } samplesIn += nSamplesFromInput * encControl->nChannelsAPI; nSamplesIn -= nSamplesFromInput; /* Default */ psEnc->allowBandwidthSwitch = 0; /* Silk encoder */ if( psEnc->state_Fxx[ 0 ].sCmn.inputBufIx >= psEnc->state_Fxx[ 0 ].sCmn.frame_length ) { /* Enough data in input buffer, so encode */ silk_assert( psEnc->state_Fxx[ 0 ].sCmn.inputBufIx == psEnc->state_Fxx[ 0 ].sCmn.frame_length ); silk_assert( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 1 ].sCmn.inputBufIx == psEnc->state_Fxx[ 1 ].sCmn.frame_length ); /* Deal with LBRR data */ if( psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == 0 && !prefillFlag ) { /* Create space at start of payload for VAD and FEC flags */ opus_uint8 iCDF[ 2 ] = { 0, 0 }; iCDF[ 0 ] = 256 - silk_RSHIFT( 256, ( psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket + 1 ) * encControl->nChannelsInternal ); ec_enc_icdf( psRangeEnc, 0, iCDF, 8 ); /* Encode any LBRR data from previous packet */ /* Encode LBRR flags */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { LBRR_symbol = 0; for( i = 0; i < psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket; i++ ) { LBRR_symbol |= silk_LSHIFT( psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ], i ); } psEnc->state_Fxx[ n ].sCmn.LBRR_flag = LBRR_symbol > 0 ? 1 : 0; if( LBRR_symbol && psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket > 1 ) { ec_enc_icdf( psRangeEnc, LBRR_symbol - 1, silk_LBRR_flags_iCDF_ptr[ psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket - 2 ], 8 ); } } /* Code LBRR indices and excitation signals */ for( i = 0; i < psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket; i++ ) { for( n = 0; n < encControl->nChannelsInternal; n++ ) { if( psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i ] ) { opus_int condCoding; if( encControl->nChannelsInternal == 2 && n == 0 ) { silk_stereo_encode_pred( psRangeEnc, psEnc->sStereo.predIx[ i ] ); /* For LBRR data there's no need to code the mid-only flag if the side-channel LBRR flag is set */ if( psEnc->state_Fxx[ 1 ].sCmn.LBRR_flags[ i ] == 0 ) { silk_stereo_encode_mid_only( psRangeEnc, psEnc->sStereo.mid_only_flags[ i ] ); } } /* Use conditional coding if previous frame available */ if( i > 0 && psEnc->state_Fxx[ n ].sCmn.LBRR_flags[ i - 1 ] ) { condCoding = CODE_CONDITIONALLY; } else { condCoding = CODE_INDEPENDENTLY; } silk_encode_indices( &psEnc->state_Fxx[ n ].sCmn, psRangeEnc, i, 1, condCoding ); silk_encode_pulses( psRangeEnc, psEnc->state_Fxx[ n ].sCmn.indices_LBRR[i].signalType, psEnc->state_Fxx[ n ].sCmn.indices_LBRR[i].quantOffsetType, psEnc->state_Fxx[ n ].sCmn.pulses_LBRR[ i ], psEnc->state_Fxx[ n ].sCmn.frame_length ); } } } /* Reset LBRR flags */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { silk_memset( psEnc->state_Fxx[ n ].sCmn.LBRR_flags, 0, sizeof( psEnc->state_Fxx[ n ].sCmn.LBRR_flags ) ); } psEnc->nBitsUsedLBRR = ec_tell( psRangeEnc ); } silk_HP_variable_cutoff( psEnc->state_Fxx ); /* Total target bits for packet */ nBits = silk_DIV32_16( silk_MUL( encControl->bitRate, encControl->payloadSize_ms ), 1000 ); /* Subtract bits used for LBRR */ if( !prefillFlag ) { nBits -= psEnc->nBitsUsedLBRR; } /* Divide by number of uncoded frames left in packet */ nBits = silk_DIV32_16( nBits, psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket ); /* Convert to bits/second */ if( encControl->payloadSize_ms == 10 ) { TargetRate_bps = silk_SMULBB( nBits, 100 ); } else { TargetRate_bps = silk_SMULBB( nBits, 50 ); } /* Subtract fraction of bits in excess of target in previous frames and packets */ TargetRate_bps -= silk_DIV32_16( silk_MUL( psEnc->nBitsExceeded, 1000 ), BITRESERVOIR_DECAY_TIME_MS ); if( !prefillFlag && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded > 0 ) { /* Compare actual vs target bits so far in this packet */ opus_int32 bitsBalance = ec_tell( psRangeEnc ) - psEnc->nBitsUsedLBRR - nBits * psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded; TargetRate_bps -= silk_DIV32_16( silk_MUL( bitsBalance, 1000 ), BITRESERVOIR_DECAY_TIME_MS ); } /* Never exceed input bitrate */ TargetRate_bps = silk_LIMIT( TargetRate_bps, encControl->bitRate, 5000 ); /* Convert Left/Right to Mid/Side */ if( encControl->nChannelsInternal == 2 ) { silk_stereo_LR_to_MS( &psEnc->sStereo, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ 2 ], &psEnc->state_Fxx[ 1 ].sCmn.inputBuf[ 2 ], psEnc->sStereo.predIx[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ], &psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ], MStargetRates_bps, TargetRate_bps, psEnc->state_Fxx[ 0 ].sCmn.speech_activity_Q8, encControl->toMono, psEnc->state_Fxx[ 0 ].sCmn.fs_kHz, psEnc->state_Fxx[ 0 ].sCmn.frame_length ); if( psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] == 0 ) { /* Reset side channel encoder memory for first frame with side coding */ if( psEnc->prev_decode_only_middle == 1 ) { silk_memset( &psEnc->state_Fxx[ 1 ].sShape, 0, sizeof( psEnc->state_Fxx[ 1 ].sShape ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sPrefilt, 0, sizeof( psEnc->state_Fxx[ 1 ].sPrefilt ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sNSQ, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sNSQ ) ); silk_memset( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.prev_NLSFq_Q15 ) ); silk_memset( &psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State, 0, sizeof( psEnc->state_Fxx[ 1 ].sCmn.sLP.In_LP_State ) ); psEnc->state_Fxx[ 1 ].sCmn.prevLag = 100; psEnc->state_Fxx[ 1 ].sCmn.sNSQ.lagPrev = 100; psEnc->state_Fxx[ 1 ].sShape.LastGainIndex = 10; psEnc->state_Fxx[ 1 ].sCmn.prevSignalType = TYPE_NO_VOICE_ACTIVITY; psEnc->state_Fxx[ 1 ].sCmn.sNSQ.prev_gain_Q16 = 65536; psEnc->state_Fxx[ 1 ].sCmn.first_frame_after_reset = 1; } silk_encode_do_VAD_Fxx( &psEnc->state_Fxx[ 1 ] ); } else { psEnc->state_Fxx[ 1 ].sCmn.VAD_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] = 0; } if( !prefillFlag ) { silk_stereo_encode_pred( psRangeEnc, psEnc->sStereo.predIx[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] ); if( psEnc->state_Fxx[ 1 ].sCmn.VAD_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] == 0 ) { silk_stereo_encode_mid_only( psRangeEnc, psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded ] ); } } } else { /* Buffering */ silk_memcpy( psEnc->state_Fxx[ 0 ].sCmn.inputBuf, psEnc->sStereo.sMid, 2 * sizeof( opus_int16 ) ); silk_memcpy( psEnc->sStereo.sMid, &psEnc->state_Fxx[ 0 ].sCmn.inputBuf[ psEnc->state_Fxx[ 0 ].sCmn.frame_length ], 2 * sizeof( opus_int16 ) ); } silk_encode_do_VAD_Fxx( &psEnc->state_Fxx[ 0 ] ); /* Encode */ for( n = 0; n < encControl->nChannelsInternal; n++ ) { opus_int maxBits, useCBR; /* Handling rate constraints */ maxBits = encControl->maxBits; if( tot_blocks == 2 && curr_block == 0 ) { maxBits = maxBits * 3 / 5; } else if( tot_blocks == 3 ) { if( curr_block == 0 ) { maxBits = maxBits * 2 / 5; } else if( curr_block == 1 ) { maxBits = maxBits * 3 / 4; } } useCBR = encControl->useCBR && curr_block == tot_blocks - 1; if( encControl->nChannelsInternal == 1 ) { channelRate_bps = TargetRate_bps; } else { channelRate_bps = MStargetRates_bps[ n ]; if( n == 0 && MStargetRates_bps[ 1 ] > 0 ) { useCBR = 0; /* Give mid up to 1/2 of the max bits for that frame */ maxBits -= encControl->maxBits / ( tot_blocks * 2 ); } } if( channelRate_bps > 0 ) { opus_int condCoding; silk_control_SNR( &psEnc->state_Fxx[ n ].sCmn, channelRate_bps ); /* Use independent coding if no previous frame available */ if( psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded - n <= 0 ) { condCoding = CODE_INDEPENDENTLY; } else if( n > 0 && psEnc->prev_decode_only_middle ) { /* If we skipped a side frame in this packet, we don't need LTP scaling; the LTP state is well-defined. */ condCoding = CODE_INDEPENDENTLY_NO_LTP_SCALING; } else { condCoding = CODE_CONDITIONALLY; } if( ( ret = silk_encode_frame_Fxx( &psEnc->state_Fxx[ n ], nBytesOut, psRangeEnc, condCoding, maxBits, useCBR ) ) != 0 ) { silk_assert( 0 ); } } psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0; psEnc->state_Fxx[ n ].sCmn.inputBufIx = 0; psEnc->state_Fxx[ n ].sCmn.nFramesEncoded++; } psEnc->prev_decode_only_middle = psEnc->sStereo.mid_only_flags[ psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded - 1 ]; /* Insert VAD and FEC flags at beginning of bitstream */ if( *nBytesOut > 0 && psEnc->state_Fxx[ 0 ].sCmn.nFramesEncoded == psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket) { flags = 0; for( n = 0; n < encControl->nChannelsInternal; n++ ) { for( i = 0; i < psEnc->state_Fxx[ n ].sCmn.nFramesPerPacket; i++ ) { flags = silk_LSHIFT( flags, 1 ); flags |= psEnc->state_Fxx[ n ].sCmn.VAD_flags[ i ]; } flags = silk_LSHIFT( flags, 1 ); flags |= psEnc->state_Fxx[ n ].sCmn.LBRR_flag; } if( !prefillFlag ) { ec_enc_patch_initial_bits( psRangeEnc, flags, ( psEnc->state_Fxx[ 0 ].sCmn.nFramesPerPacket + 1 ) * encControl->nChannelsInternal ); } /* Return zero bytes if all channels DTXed */ if( psEnc->state_Fxx[ 0 ].sCmn.inDTX && ( encControl->nChannelsInternal == 1 || psEnc->state_Fxx[ 1 ].sCmn.inDTX ) ) { *nBytesOut = 0; } psEnc->nBitsExceeded += *nBytesOut * 8; psEnc->nBitsExceeded -= silk_DIV32_16( silk_MUL( encControl->bitRate, encControl->payloadSize_ms ), 1000 ); psEnc->nBitsExceeded = silk_LIMIT( psEnc->nBitsExceeded, 0, 10000 ); /* Update flag indicating if bandwidth switching is allowed */ speech_act_thr_for_switch_Q8 = (opus_int) silk_SMLAWB( SILK_FIX_CONST( SPEECH_ACTIVITY_DTX_THRES, 8 ), SILK_FIX_CONST( ( 1 - SPEECH_ACTIVITY_DTX_THRES ) / MAX_BANDWIDTH_SWITCH_DELAY_MS, 16 + 8 ), psEnc->timeSinceSwitchAllowed_ms ); if( psEnc->state_Fxx[ 0 ].sCmn.speech_activity_Q8 < speech_act_thr_for_switch_Q8 ) { psEnc->allowBandwidthSwitch = 1; psEnc->timeSinceSwitchAllowed_ms = 0; } else { psEnc->allowBandwidthSwitch = 0; psEnc->timeSinceSwitchAllowed_ms += encControl->payloadSize_ms; } } if( nSamplesIn == 0 ) { break; } } else { break; } curr_block++; } psEnc->nPrevChannelsInternal = encControl->nChannelsInternal; encControl->allowBandwidthSwitch = psEnc->allowBandwidthSwitch; encControl->inWBmodeWithoutVariableLP = psEnc->state_Fxx[ 0 ].sCmn.fs_kHz == 16 && psEnc->state_Fxx[ 0 ].sCmn.sLP.mode == 0; encControl->internalSampleRate = silk_SMULBB( psEnc->state_Fxx[ 0 ].sCmn.fs_kHz, 1000 ); encControl->stereoWidth_Q14 = encControl->toMono ? 0 : psEnc->sStereo.smth_width_Q14; if( prefillFlag ) { encControl->payloadSize_ms = tmp_payloadSize_ms; encControl->complexity = tmp_complexity; for( n = 0; n < encControl->nChannelsInternal; n++ ) { psEnc->state_Fxx[ n ].sCmn.controlled_since_last_payload = 0; psEnc->state_Fxx[ n ].sCmn.prefillFlag = 0; } } RESTORE_STACK; return ret; }
/* Initialize/reset the resampler state for a given pair of input/output sampling rates */ opus_int silk_resampler_init( silk_resampler_state_struct *S, /* I/O Resampler state */ opus_int32 Fs_Hz_in, /* I Input sampling rate (Hz) */ opus_int32 Fs_Hz_out /* I Output sampling rate (Hz) */ ) { opus_int32 up2 = 0, down2 = 0; /* Clear state */ silk_memset( S, 0, sizeof( silk_resampler_state_struct ) ); /* Input checking */ if( ( Fs_Hz_in != 8000 && Fs_Hz_in != 12000 && Fs_Hz_in != 16000 && Fs_Hz_in != 24000 && Fs_Hz_in != 48000 ) || ( Fs_Hz_out != 8000 && Fs_Hz_out != 12000 && Fs_Hz_out != 16000 && Fs_Hz_out != 24000 && Fs_Hz_out != 48000 ) ) { silk_assert( 0 ); return -1; } /* Number of samples processed per batch */ S->batchSize = silk_DIV32_16( Fs_Hz_in, 100 ); /* Find resampler with the right sampling ratio */ if( Fs_Hz_out > Fs_Hz_in ) { /* Upsample */ if( Fs_Hz_out == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 1 */ /* Special case: directly use 2x upsampler */ S->resampler_function = USE_silk_resampler_private_up2_HQ_wrapper; } else { /* Default resampler */ S->resampler_function = USE_silk_resampler_private_IIR_FIR; up2 = 1; } } else if ( Fs_Hz_out < Fs_Hz_in ) { /* Downsample */ if( silk_MUL( Fs_Hz_out, 4 ) == silk_MUL( Fs_Hz_in, 3 ) ) { /* Fs_out : Fs_in = 3 : 4 */ S->FIR_Fracs = 3; S->Coefs = silk_Resampler_3_4_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 3 ) == silk_MUL( Fs_Hz_in, 2 ) ) { /* Fs_out : Fs_in = 2 : 3 */ S->FIR_Fracs = 2; S->Coefs = silk_Resampler_2_3_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 2 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 2 */ S->FIR_Fracs = 1; S->Coefs = silk_Resampler_1_2_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 3 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 3 */ S->FIR_Fracs = 1; S->Coefs = silk_Resampler_1_3_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 4 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 4 */ S->FIR_Fracs = 1; down2 = 1; S->Coefs = silk_Resampler_1_2_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else if( silk_MUL( Fs_Hz_out, 6 ) == Fs_Hz_in ) { /* Fs_out : Fs_in = 1 : 6 */ S->FIR_Fracs = 1; down2 = 1; S->Coefs = silk_Resampler_1_3_COEFS; S->resampler_function = USE_silk_resampler_private_down_FIR; } else { /* None available */ silk_assert( 0 ); return -1; } } else { /* Input and output sampling rates are equal: copy */ S->resampler_function = USE_silk_resampler_copy; } S->input2x = up2 | down2; /* Ratio of input/output samples */ S->invRatio_Q16 = silk_LSHIFT32( silk_DIV32( silk_LSHIFT32( Fs_Hz_in, 14 + up2 - down2 ), Fs_Hz_out ), 2 ); /* Make sure the ratio is rounded up */ while( silk_SMULWW( S->invRatio_Q16, silk_LSHIFT32( Fs_Hz_out, down2 ) ) < silk_LSHIFT32( Fs_Hz_in, up2 ) ) { S->invRatio_Q16++; } return 0; }
void silk_find_pred_coefs_FIX( silk_encoder_state_FIX *psEnc, /* I/O encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O encoder control */ const opus_int16 res_pitch[], /* I Residual from pitch analysis */ const opus_int16 x[], /* I Speech signal */ opus_int condCoding /* I The type of conditional coding to use */ ) { opus_int i; opus_int32 WLTP[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ]; opus_int32 invGains_Q16[ MAX_NB_SUBFR ], local_gains[ MAX_NB_SUBFR ], Wght_Q15[ MAX_NB_SUBFR ]; opus_int16 NLSF_Q15[ MAX_LPC_ORDER ]; const opus_int16 *x_ptr; opus_int16 *x_pre_ptr, LPC_in_pre[ MAX_NB_SUBFR * MAX_LPC_ORDER + MAX_FRAME_LENGTH ]; opus_int32 tmp, min_gain_Q16; opus_int LTP_corrs_rshift[ MAX_NB_SUBFR ]; /* weighting for weighted least squares */ min_gain_Q16 = silk_int32_MAX >> 6; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { min_gain_Q16 = silk_min( min_gain_Q16, psEncCtrl->Gains_Q16[ i ] ); } for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { /* Divide to Q16 */ silk_assert( psEncCtrl->Gains_Q16[ i ] > 0 ); /* Invert and normalize gains, and ensure that maximum invGains_Q16 is within range of a 16 bit int */ invGains_Q16[ i ] = silk_DIV32_varQ( min_gain_Q16, psEncCtrl->Gains_Q16[ i ], 16 - 2 ); /* Ensure Wght_Q15 a minimum value 1 */ invGains_Q16[ i ] = silk_max( invGains_Q16[ i ], 363 ); /* Square the inverted gains */ silk_assert( invGains_Q16[ i ] == silk_SAT16( invGains_Q16[ i ] ) ); tmp = silk_SMULWB( invGains_Q16[ i ], invGains_Q16[ i ] ); Wght_Q15[ i ] = silk_RSHIFT( tmp, 1 ); /* Invert the inverted and normalized gains */ local_gains[ i ] = silk_DIV32( ( 1 << 16 ), invGains_Q16[ i ] ); } if( psEnc->sCmn.indices.signalType == TYPE_VOICED ) { /**********/ /* VOICED */ /**********/ silk_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); /* LTP analysis */ silk_find_LTP_FIX( psEncCtrl->LTPCoef_Q14, WLTP, &psEncCtrl->LTPredCodGain_Q7, res_pitch, psEncCtrl->pitchL, Wght_Q15, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.ltp_mem_length, LTP_corrs_rshift ); /* Quantize LTP gain parameters */ silk_quant_LTP_gains( psEncCtrl->LTPCoef_Q14, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, WLTP, psEnc->sCmn.mu_LTP_Q9, psEnc->sCmn.LTPQuantLowComplexity, psEnc->sCmn.nb_subfr); /* Control LTP scaling */ silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding ); /* Create LTP residual */ silk_LTP_analysis_filter_FIX( LPC_in_pre, x - psEnc->sCmn.predictLPCOrder, psEncCtrl->LTPCoef_Q14, psEncCtrl->pitchL, invGains_Q16, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); } else { /************/ /* UNVOICED */ /************/ /* Create signal with prepended subframes, scaled by inverse gains */ x_ptr = x - psEnc->sCmn.predictLPCOrder; x_pre_ptr = LPC_in_pre; for( i = 0; i < psEnc->sCmn.nb_subfr; i++ ) { silk_scale_copy_vector16( x_pre_ptr, x_ptr, invGains_Q16[ i ], psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder ); x_pre_ptr += psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder; x_ptr += psEnc->sCmn.subfr_length; } silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) ); psEncCtrl->LTPredCodGain_Q7 = 0; } /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ silk_find_LPC_FIX( NLSF_Q15, &psEnc->sCmn.indices.NLSFInterpCoef_Q2, psEnc->sCmn.prev_NLSFq_Q15, psEnc->sCmn.useInterpolatedNLSFs, psEnc->sCmn.first_frame_after_reset, psEnc->sCmn.predictLPCOrder, LPC_in_pre, psEnc->sCmn.subfr_length + psEnc->sCmn.predictLPCOrder, psEnc->sCmn.nb_subfr ); /* Quantize LSFs */ silk_process_NLSFs( &psEnc->sCmn, psEncCtrl->PredCoef_Q12, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 ); /* Calculate residual energy using quantized LPC coefficients */ silk_residual_energy_FIX( psEncCtrl->ResNrg, psEncCtrl->ResNrgQ, LPC_in_pre, psEncCtrl->PredCoef_Q12, local_gains, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.predictLPCOrder ); /* Copy to prediction struct for use in next frame for interpolation */ silk_memcpy( psEnc->sCmn.prev_NLSFq_Q15, NLSF_Q15, sizeof( psEnc->sCmn.prev_NLSFq_Q15 ) ); }