From bbfc9c9ee5de680320aeae2f8bd981ebe7af6861 Mon Sep 17 00:00:00 2001 From: Koen Vos <koen.vos@skype.net> Date: Tue, 13 Dec 2011 14:50:12 -0500 Subject: [PATCH] Improves the accuracy such that it matches a float decoder much better --- silk/CNG.c | 6 ++-- silk/NSQ.c | 59 ++++++++++++++++++-------------- silk/NSQ_del_dec.c | 57 +++++++++++++++++------------- silk/PLC.c | 18 ++++++---- silk/control_codec.c | 2 +- silk/decode_core.c | 31 ++++++++++------- silk/define.h | 2 +- silk/enc_API.c | 2 +- silk/fixed/encode_frame_FIX.c | 18 +++++----- silk/fixed/find_pred_coefs_FIX.c | 2 +- silk/fixed/prefilter_FIX.c | 17 +++++---- silk/float/find_pred_coefs_FLP.c | 2 +- silk/float/wrappers_FLP.c | 8 ++--- silk/init_decoder.c | 2 +- silk/main.h | 4 +-- silk/structs.h | 4 +-- silk/tables_other.c | 2 +- 17 files changed, 132 insertions(+), 104 deletions(-) diff --git a/silk/CNG.c b/silk/CNG.c index 6885d83d8..c85353b05 100644 --- a/silk/CNG.c +++ b/silk/CNG.c @@ -51,7 +51,7 @@ static inline void silk_CNG_exc( seed = *rand_seed; for( i = 0; i < length; i++ ) { seed = silk_RAND( seed ); - idx = ( opus_int )( silk_RSHIFT( seed, 24 ) & exc_mask ); + idx = (opus_int)( silk_RSHIFT( seed, 24 ) & exc_mask ); silk_assert( idx >= 0 ); silk_assert( idx <= CNG_BUF_MASK_MAX ); residual_Q10[ i ] = (opus_int16)silk_SAT16( silk_SMULWW( exc_buf_Q10[ idx ], Gain_Q16 ) ); @@ -134,7 +134,9 @@ void silk_CNG( silk_memcpy( CNG_sig_Q10, psCNG->CNG_synth_state, MAX_LPC_ORDER * sizeof( opus_int32 ) ); for( i = 0; i < length; i++ ) { silk_assert( psDec->LPC_order == 10 || psDec->LPC_order == 16 ); - sum_Q6 = silk_SMULWB( CNG_sig_Q10[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] ); + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + sum_Q6 = silk_RSHIFT( psDec->LPC_order, 1 ); + sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] ); sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] ); sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] ); sum_Q6 = silk_SMLAWB( sum_Q6, CNG_sig_Q10[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] ); diff --git a/silk/NSQ.c b/silk/NSQ.c index 760042622..96fec5401 100644 --- a/silk/NSQ.c +++ b/silk/NSQ.c @@ -34,7 +34,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. static inline void silk_nsq_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ - const opus_int32 x_Q10[], /* I input in Q0 */ + const opus_int32 x_Q3[], /* I input in Q3 */ opus_int32 x_sc_Q10[], /* O input scaled with 1/Gain */ const opus_int16 sLTP[], /* I re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -71,7 +71,7 @@ void silk_NSQ( const silk_encoder_state *psEncC, /* I/O Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q10[], /* I Prefiltered input signal */ + const opus_int32 x_Q3[], /* I Prefiltered input signal */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ @@ -99,7 +99,7 @@ void silk_NSQ( /* Set unvoiced lag to the previous one, overwrite later for voiced */ lag = NSQ->lagPrev; - silk_assert( NSQ->prev_inv_gain_Q16 != 0 ); + silk_assert( NSQ->prev_inv_gain_Q31 != 0 ); offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ]; @@ -142,13 +142,13 @@ void silk_NSQ( } } - silk_nsq_scale_states( psEncC, NSQ, x_Q10, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType ); + silk_nsq_scale_states( psEncC, NSQ, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType ); silk_noise_shape_quantizer( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder ); - x_Q10 += psEncC->subfr_length; + x_Q3 += psEncC->subfr_length; pulses += psEncC->subfr_length; pxq += psEncC->subfr_length; } @@ -157,6 +157,7 @@ void silk_NSQ( NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; /* Save quantized speech and noise shaping signals */ + /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[ psEncC->ltp_mem_length ], psEncC->frame_length * sizeof( opus_int16 ) ) */ silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); silk_memmove( NSQ->sLTP_shp_Q10, &NSQ->sLTP_shp_Q10[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); } @@ -208,7 +209,9 @@ static inline void silk_noise_shape_quantizer( /* Short-term prediction */ silk_assert( predictLPCOrder == 10 || predictLPCOrder == 16 ); - LPC_pred_Q10 = silk_SMULWB( psLPC_Q14[ 0 ], a_Q12[ 0 ] ); + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LPC_pred_Q10 = silk_RSHIFT( predictLPCOrder, 1 ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ 0 ], a_Q12[ 0 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -1 ], a_Q12[ 1 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -2 ], a_Q12[ 2 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -3 ], a_Q12[ 3 ] ); @@ -230,7 +233,9 @@ static inline void silk_noise_shape_quantizer( /* Long-term prediction */ if( signalType == TYPE_VOICED ) { /* Unrolled loop */ - LTP_pred_Q13 = silk_SMULWB( pred_lag_ptr[ 0 ], b_Q14[ 0 ] ); + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LTP_pred_Q13 = 2; + LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ 0 ], b_Q14[ 0 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -1 ], b_Q14[ 1 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -2 ], b_Q14[ 2 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -3 ], b_Q14[ 3 ] ); @@ -273,7 +278,7 @@ static inline void silk_noise_shape_quantizer( n_LTP_Q14 = silk_LSHIFT( n_LTP_Q14, 6 ); shp_lag_ptr++; - tmp1 = silk_SUB32( LTP_pred_Q13 << 1, n_LTP_Q14 ); /* Add Q14 stuff */ + tmp1 = silk_SUB32( silk_LSHIFT32( LTP_pred_Q13, 1 ), n_LTP_Q14 ); /* Add Q14 stuff */ tmp1 = silk_RSHIFT( tmp1, 4 ); /* convert to Q10 */ tmp1 = silk_ADD32( tmp1, LPC_pred_Q10 ); /* add Q10 stuff */ tmp1 = silk_SUB32( tmp1, n_AR_Q10 ); /* subtract Q10 stuff */ @@ -337,7 +342,6 @@ static inline void silk_noise_shape_quantizer( /* Scale XQ back to normal level before saving */ xq[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( xq_Q10, Gain_Q16 ), 10 ) ); - /* DEBUG_STORE_DATA( enc.pcm, &xq[i], sizeof( opus_int16 ) ) */ /* Update states */ psLPC_Q14++; *psLPC_Q14 = silk_LSHIFT( xq_Q10, 4 ); @@ -360,7 +364,7 @@ static inline void silk_noise_shape_quantizer( static inline void silk_nsq_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ - const opus_int32 x_Q10[], /* I input in Q0 */ + const opus_int32 x_Q3[], /* I input in Q3 */ opus_int32 x_sc_Q10[], /* O input scaled with 1/Gain */ const opus_int16 sLTP[], /* I re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -372,14 +376,30 @@ static inline void silk_nsq_scale_states( ) { opus_int i, lag; - opus_int32 inv_gain_Q16, gain_adj_Q16, inv_gain_Q31; + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23; - inv_gain_Q16 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 32 ); + inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); lag = pitchL[ subfr ]; + /* Calculate gain adjustment factor */ + if( inv_gain_Q31 != NSQ->prev_inv_gain_Q31 ) { + gain_adj_Q16 = silk_DIV32_varQ( inv_gain_Q31, NSQ->prev_inv_gain_Q31, 16 ); + } else { + gain_adj_Q16 = 1 << 16; + } + + /* Scale input */ + inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 ); + for( i = 0; i < psEncC->subfr_length; i++ ) { + x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 ); + } + + /* Save inverse gain */ + silk_assert( inv_gain_Q31 != 0 ); + NSQ->prev_inv_gain_Q31 = inv_gain_Q31; + /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ if( NSQ->rewhite_flag ) { - inv_gain_Q31 = silk_LSHIFT( inv_gain_Q16, 15 ); if( subfr == 0 ) { /* Do LTP downscaling */ inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 ); @@ -391,9 +411,7 @@ static inline void silk_nsq_scale_states( } /* Adjust for changing gain */ - if( inv_gain_Q16 != NSQ->prev_inv_gain_Q16 ) { - gain_adj_Q16 = silk_DIV32_varQ( inv_gain_Q16, NSQ->prev_inv_gain_Q16, 16 ); - + if( gain_adj_Q16 != 1 << 16 ) { /* Scale long-term shaping state */ for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) { NSQ->sLTP_shp_Q10[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q10[ i ] ); @@ -416,13 +434,4 @@ static inline void silk_nsq_scale_states( NSQ->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sAR2_Q14[ i ] ); } } - - /* Scale input */ - for( i = 0; i < psEncC->subfr_length; i++ ) { - x_sc_Q10[ i ] = silk_SMULWW( x_Q10[ i ], inv_gain_Q16 ); - } - - /* Save inv_gain */ - silk_assert( inv_gain_Q16 != 0 ); - NSQ->prev_inv_gain_Q16 = inv_gain_Q16; } diff --git a/silk/NSQ_del_dec.c b/silk/NSQ_del_dec.c index 48d413b10..b13ecca86 100644 --- a/silk/NSQ_del_dec.c +++ b/silk/NSQ_del_dec.c @@ -58,7 +58,7 @@ static inline void silk_nsq_del_dec_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ - const opus_int32 x_Q10[], /* I Input in Q0 */ + const opus_int32 x_Q3[], /* I Input in Q3 */ opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -107,7 +107,7 @@ void silk_NSQ_del_dec( const silk_encoder_state *psEncC, /* I/O Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q10[], /* I Prefiltered input signal */ + const opus_int32 x_Q3[], /* I Prefiltered input signal */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ @@ -138,7 +138,7 @@ void silk_NSQ_del_dec( /* Set unvoiced lag to the previous one, overwrite later for voiced */ lag = NSQ->lagPrev; - silk_assert( NSQ->prev_inv_gain_Q16 != 0 ); + silk_assert( NSQ->prev_inv_gain_Q31 != 0 ); /* Initialize delayed decision states */ silk_memset( psDelDec, 0, psEncC->nStatesDelayedDecision * sizeof( NSQ_del_dec_struct ) ); @@ -241,7 +241,7 @@ void silk_NSQ_del_dec( } } - silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x_Q10, x_sc_Q10, sLTP, sLTP_Q15, k, + silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, psEncC->nStatesDelayedDecision, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay ); silk_noise_shape_quantizer_del_dec( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, @@ -249,7 +249,7 @@ void silk_NSQ_del_dec( Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay ); - x_Q10 += psEncC->subfr_length; + x_Q3 += psEncC->subfr_length; pulses += psEncC->subfr_length; pxq += psEncC->subfr_length; } @@ -342,7 +342,9 @@ static inline void silk_noise_shape_quantizer_del_dec( /* Long-term prediction */ if( signalType == TYPE_VOICED ) { /* Unrolled loop */ - LTP_pred_Q13 = silk_SMULWB( pred_lag_ptr[ 0 ], b_Q14[ 0 ] ); + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LTP_pred_Q13 = 2; + LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ 0 ], b_Q14[ 0 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -1 ], b_Q14[ 1 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -2 ], b_Q14[ 2 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -3 ], b_Q14[ 3 ] ); @@ -360,7 +362,7 @@ static inline void silk_noise_shape_quantizer_del_dec( n_LTP_Q14 = silk_LSHIFT( n_LTP_Q14, 6 ); shp_lag_ptr++; - LTP_Q10 = silk_RSHIFT( silk_SUB32( LTP_pred_Q13 << 1, n_LTP_Q14 ), 4 ); + LTP_Q10 = silk_RSHIFT( silk_SUB32( silk_LSHIFT( LTP_pred_Q13, 1 ), n_LTP_Q14 ), 4 ); } else { LTP_Q10 = 0; } @@ -382,7 +384,9 @@ static inline void silk_noise_shape_quantizer_del_dec( psLPC_Q14 = &psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 + i ]; /* Short-term prediction */ silk_assert( predictLPCOrder == 10 || predictLPCOrder == 16 ); - LPC_pred_Q10 = silk_SMULWB( psLPC_Q14[ 0 ], a_Q12[ 0 ] ); + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LPC_pred_Q10 = silk_RSHIFT( predictLPCOrder, 1 ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ 0 ], a_Q12[ 0 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -1 ], a_Q12[ 1 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -2 ], a_Q12[ 2 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psLPC_Q14[ -3 ], a_Q12[ 3 ] ); @@ -603,7 +607,7 @@ static inline void silk_nsq_del_dec_scale_states( const silk_encoder_state *psEncC, /* I Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ - const opus_int32 x_Q10[], /* I Input in Q0 */ + const opus_int32 x_Q3[], /* I Input in Q3 */ opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ @@ -617,15 +621,31 @@ static inline void silk_nsq_del_dec_scale_states( ) { opus_int i, k, lag; - opus_int32 inv_gain_Q16, gain_adj_Q16, inv_gain_Q31; + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23; NSQ_del_dec_struct *psDD; - inv_gain_Q16 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 32 ); + inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); lag = pitchL[ subfr ]; + /* Calculate gain adjustment factor */ + if( inv_gain_Q31 != NSQ->prev_inv_gain_Q31 ) { + gain_adj_Q16 = silk_DIV32_varQ( inv_gain_Q31, NSQ->prev_inv_gain_Q31, 16 ); + } else { + gain_adj_Q16 = 1 << 16; + } + + /* Scale input */ + inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 ); + for( i = 0; i < psEncC->subfr_length; i++ ) { + x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 ); + } + + /* Save inverse gain */ + silk_assert( inv_gain_Q31 != 0 ); + NSQ->prev_inv_gain_Q31 = inv_gain_Q31; + /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ if( NSQ->rewhite_flag ) { - inv_gain_Q31 = silk_LSHIFT( inv_gain_Q16, 15 ); if( subfr == 0 ) { /* Do LTP downscaling */ inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 ); @@ -637,9 +657,7 @@ static inline void silk_nsq_del_dec_scale_states( } /* Adjust for changing gain */ - if( inv_gain_Q16 != NSQ->prev_inv_gain_Q16 ) { - gain_adj_Q16 = silk_DIV32_varQ( inv_gain_Q16, NSQ->prev_inv_gain_Q16, 16 ); - + if( gain_adj_Q16 != 1 << 16 ) { /* Scale long-term shaping state */ for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) { NSQ->sLTP_shp_Q10[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q10[ i ] ); @@ -671,13 +689,4 @@ static inline void silk_nsq_del_dec_scale_states( } } } - - /* Scale input */ - for( i = 0; i < psEncC->subfr_length; i++ ) { - x_sc_Q10[ i ] = silk_SMULWW( x_Q10[ i ], inv_gain_Q16 ); - } - - /* save inv_gain */ - silk_assert( inv_gain_Q16 != 0 ); - NSQ->prev_inv_gain_Q16 = inv_gain_Q16; } diff --git a/silk/PLC.c b/silk/PLC.c index b76032ac9..fdf46bb78 100644 --- a/silk/PLC.c +++ b/silk/PLC.c @@ -172,7 +172,7 @@ static inline void silk_PLC_conceal( { opus_int i, j, k; opus_int lag, idx, sLTP_buf_idx, shift1, shift2; - opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q16, inv_gain_Q30; + opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q30; opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr; opus_int32 LPC_exc_Q14, LPC_pred_Q10, LTP_pred_Q12; opus_int16 rand_scale_Q14; @@ -184,8 +184,9 @@ static inline void silk_PLC_conceal( opus_int32 sLTP_Q14[ 2 * MAX_FRAME_LENGTH ]; silk_PLC_struct *psPLC = &psDec->sPLC; - if (psDec->first_frame_after_reset) + if( psDec->first_frame_after_reset ) { silk_memset(psPLC->prevLPC_Q12, 0, MAX_LPC_ORDER*sizeof(psPLC->prevLPC_Q12[ 0 ])); + } /* Find random noise component */ /* Scale previous excitation signal */ @@ -261,9 +262,8 @@ static inline void silk_PLC_conceal( silk_assert( idx > 0 ); silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order ); /* Scale LTP state */ - inv_gain_Q16 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 32 ); - inv_gain_Q16 = silk_min( inv_gain_Q16, silk_int16_MAX ); - inv_gain_Q30 = silk_LSHIFT( inv_gain_Q16, 14 ); + 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 ] ); } @@ -276,7 +276,9 @@ static inline void silk_PLC_conceal( pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; for( i = 0; i < psDec->subfr_length; i++ ) { /* Unrolled loop */ - LTP_pred_Q12 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); + /* 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 ] ); @@ -316,7 +318,9 @@ static inline void silk_PLC_conceal( silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */ for( i = 0; i < psDec->frame_length; i++ ) { /* partly unrolled */ - LPC_pred_Q10 = silk_SMULWB( sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] ); + /* 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 ] ); diff --git a/silk/control_codec.c b/silk/control_codec.c index 16666d68d..1615a16c9 100644 --- a/silk/control_codec.c +++ b/silk/control_codec.c @@ -247,7 +247,7 @@ opus_int silk_setup_fs( psEnc->sPrefilt.lagPrev = 100; psEnc->sShape.LastGainIndex = 10; psEnc->sCmn.sNSQ.lagPrev = 100; - psEnc->sCmn.sNSQ.prev_inv_gain_Q16 = 65536; + psEnc->sCmn.sNSQ.prev_inv_gain_Q31 = silk_int32_MAX; psEnc->sCmn.prevSignalType = TYPE_NO_VOICE_ACTIVITY; psEnc->sCmn.fs_kHz = fs_kHz; diff --git a/silk/decode_core.c b/silk/decode_core.c index 6633143d1..4b36aa07d 100644 --- a/silk/decode_core.c +++ b/silk/decode_core.c @@ -45,12 +45,12 @@ void silk_decode_core( opus_int16 *A_Q12, *B_Q14, *pxq, A_Q12_tmp[ MAX_LPC_ORDER ]; opus_int16 sLTP[ MAX_FRAME_LENGTH ]; opus_int32 sLTP_Q15[ 2 * MAX_FRAME_LENGTH ]; - opus_int32 LTP_pred_Q13, LPC_pred_Q10, Gain_Q10, inv_gain_Q16, inv_gain_Q31, gain_adj_Q16, rand_seed, offset_Q10; + opus_int32 LTP_pred_Q13, LPC_pred_Q10, Gain_Q10, inv_gain_Q31, gain_adj_Q16, rand_seed, offset_Q10; opus_int32 *pred_lag_ptr, *pexc_Q10, *pres_Q10; opus_int32 res_Q10[ MAX_SUB_FRAME_LENGTH ]; opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + MAX_LPC_ORDER ]; - silk_assert( psDec->prev_inv_gain_Q16 != 0 ); + silk_assert( psDec->prev_inv_gain_Q31 != 0 ); offset_Q10 = silk_Quantization_Offsets_Q10[ psDec->indices.signalType >> 1 ][ psDec->indices.quantOffsetType ]; @@ -94,22 +94,23 @@ void silk_decode_core( signalType = psDec->indices.signalType; Gain_Q10 = silk_RSHIFT( psDecCtrl->Gains_Q16[ k ], 6 ); - inv_gain_Q16 = silk_INVERSE32_varQ( psDecCtrl->Gains_Q16[ k ], 32 ); + inv_gain_Q31 = silk_INVERSE32_varQ( psDecCtrl->Gains_Q16[ k ], 47 ); - /* Calculate Gain adjustment factor */ - gain_adj_Q16 = 1 << 16; - if( inv_gain_Q16 != psDec->prev_inv_gain_Q16 ) { - gain_adj_Q16 = silk_DIV32_varQ( inv_gain_Q16, psDec->prev_inv_gain_Q16, 16 ); + /* Calculate gain adjustment factor */ + if( inv_gain_Q31 != psDec->prev_inv_gain_Q31 ) { + gain_adj_Q16 = silk_DIV32_varQ( inv_gain_Q31, psDec->prev_inv_gain_Q31, 16 ); /* Scale short term state */ for( i = 0; i < MAX_LPC_ORDER; i++ ) { sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, sLPC_Q14[ i ] ); } + } else { + gain_adj_Q16 = 1 << 16; } /* Save inv_gain */ - silk_assert( inv_gain_Q16 != 0 ); - psDec->prev_inv_gain_Q16 = inv_gain_Q16; + silk_assert( inv_gain_Q31 != 0 ); + psDec->prev_inv_gain_Q31 = inv_gain_Q31; /* Avoid abrupt transition from voiced PLC to unvoiced normal decoding */ if( psDec->lossCnt && psDec->prevSignalType == TYPE_VOICED && @@ -140,7 +141,6 @@ void silk_decode_core( A_Q12, psDec->ltp_mem_length - start_idx, psDec->LPC_order ); /* After rewhitening the LTP state is unscaled */ - inv_gain_Q31 = silk_LSHIFT( inv_gain_Q16, 15 ); if( k == 0 ) { /* Do LTP downscaling to reduce inter-packet dependency */ inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, psDecCtrl->LTP_scale_Q14 ), 2 ); @@ -164,7 +164,9 @@ void silk_decode_core( pred_lag_ptr = &sLTP_Q15[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; for( i = 0; i < psDec->subfr_length; i++ ) { /* Unrolled loop */ - LTP_pred_Q13 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ + LTP_pred_Q13 = 2; + LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -1 ], B_Q14[ 1 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -2 ], B_Q14[ 2 ] ); LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -3 ], B_Q14[ 3 ] ); @@ -185,7 +187,9 @@ void silk_decode_core( for( i = 0; i < psDec->subfr_length; i++ ) { /* Short-term prediction */ silk_assert( psDec->LPC_order == 10 || psDec->LPC_order == 16 ); - LPC_pred_Q10 = silk_SMULWB( sLPC_Q14[ MAX_LPC_ORDER + i - 1 ], A_Q12_tmp[ 0 ] ); + /* 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[ MAX_LPC_ORDER + i - 1 ], A_Q12_tmp[ 0 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 2 ], A_Q12_tmp[ 1 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 3 ], A_Q12_tmp[ 2 ] ); LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 4 ], A_Q12_tmp[ 3 ] ); @@ -207,11 +211,12 @@ void silk_decode_core( /* Add prediction to LPC excitation */ sLPC_Q14[ MAX_LPC_ORDER + i ] = silk_LSHIFT( silk_ADD32( pres_Q10[ i ], LPC_pred_Q10 ), 4 ); - /* Scale with Gain */ + /* Scale with gain */ pxq[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14[ MAX_LPC_ORDER + i ], Gain_Q10 ), 8 ) ); } /* DEBUG_STORE_DATA( dec.pcm, pxq, psDec->subfr_length * sizeof( opus_int16 ) ) */ + /* Update LPC filter state */ silk_memcpy( sLPC_Q14, &sLPC_Q14[ psDec->subfr_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); pexc_Q10 += psDec->subfr_length; diff --git a/silk/define.h b/silk/define.h index fe2a659a2..6730f1427 100644 --- a/silk/define.h +++ b/silk/define.h @@ -129,7 +129,7 @@ extern "C" #define QUANT_LEVEL_ADJUST_Q10 80 -/* Maximum numbers of iterations used to stabilize a LPC vector */ +/* Maximum numbers of iterations used to stabilize an LPC vector */ #define MAX_LPC_STABILIZE_ITERATIONS 30 #define MAX_PREDICTION_POWER_GAIN 1e4f #define MAX_PREDICTION_POWER_GAIN_AFTER_RESET 1e2f diff --git a/silk/enc_API.c b/silk/enc_API.c index b993ed595..d662bf252 100644 --- a/silk/enc_API.c +++ b/silk/enc_API.c @@ -392,7 +392,7 @@ opus_int silk_Encode( /* O Returns error co 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_inv_gain_Q16 = 65536; + psEnc->state_Fxx[ 1 ].sCmn.sNSQ.prev_inv_gain_Q31 = silk_int32_MAX; psEnc->state_Fxx[ 1 ].sCmn.first_frame_after_reset = 1; } silk_encode_do_VAD_Fxx( &psEnc->state_Fxx[ 1 ] ); diff --git a/silk/fixed/encode_frame_FIX.c b/silk/fixed/encode_frame_FIX.c index 11b9c69dd..601a07822 100644 --- a/silk/fixed/encode_frame_FIX.c +++ b/silk/fixed/encode_frame_FIX.c @@ -36,7 +36,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. static inline void silk_LBRR_encode_FIX( silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O Pointer to Silk FIX encoder control struct */ - const opus_int32 xfw_Q10[], /* I Input signal */ + const opus_int32 xfw_Q3[], /* I Input signal */ opus_int condCoding /* I The type of conditional coding used so far for this frame */ ); @@ -85,7 +85,7 @@ opus_int silk_encode_frame_FIX( silk_encoder_control_FIX sEncCtrl; opus_int i, iter, maxIter, found_upper, found_lower, ret = 0; opus_int16 *x_frame, *res_pitch_frame; - opus_int32 xfw_Q10[ MAX_FRAME_LENGTH ]; + opus_int32 xfw_Q3[ MAX_FRAME_LENGTH ]; opus_int16 res_pitch[ 2 * MAX_FRAME_LENGTH + LA_PITCH_MAX ]; ec_enc sRangeEnc_copy, sRangeEnc_copy2; silk_nsq_state sNSQ_copy, sNSQ_copy2; @@ -143,12 +143,12 @@ opus_int silk_encode_frame_FIX( /*****************************************/ /* Prefiltering for noise shaper */ /*****************************************/ - silk_prefilter_FIX( psEnc, &sEncCtrl, xfw_Q10, x_frame ); + silk_prefilter_FIX( psEnc, &sEncCtrl, xfw_Q3, x_frame ); /****************************************/ /* Low Bitrate Redundant Encoding */ /****************************************/ - silk_LBRR_encode_FIX( psEnc, &sEncCtrl, xfw_Q10, condCoding ); + silk_LBRR_encode_FIX( psEnc, &sEncCtrl, xfw_Q3, condCoding ); /* Loop over quantizer and entropy coding to control bitrate */ maxIter = 6; @@ -183,11 +183,11 @@ opus_int silk_encode_frame_FIX( /* 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_Q10, psEnc->sCmn.pulses, + 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_Q10, psEnc->sCmn.pulses, + 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 ); } @@ -316,7 +316,7 @@ opus_int silk_encode_frame_FIX( static inline void silk_LBRR_encode_FIX( silk_encoder_state_FIX *psEnc, /* I/O Pointer to Silk FIX encoder state */ silk_encoder_control_FIX *psEncCtrl, /* I/O Pointer to Silk FIX encoder control struct */ - const opus_int32 xfw_Q10[], /* I Input signal */ + const opus_int32 xfw_Q3[], /* I Input signal */ opus_int condCoding /* I The type of conditional coding used so far for this frame */ ) { @@ -355,12 +355,12 @@ static inline void silk_LBRR_encode_FIX( /* Noise shaping quantization */ /*****************************************/ if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q10, + silk_NSQ_del_dec( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q3, psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, psEncCtrl->AR2_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14 ); } else { - silk_NSQ( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q10, + silk_NSQ( &psEnc->sCmn, &sNSQ_LBRR, psIndices_LBRR, xfw_Q3, psEnc->sCmn.pulses_LBRR[ psEnc->sCmn.nFramesEncoded ], psEncCtrl->PredCoef_Q12[ 0 ], psEncCtrl->LTPCoef_Q14, psEncCtrl->AR2_Q13, psEncCtrl->HarmShapeGain_Q14, psEncCtrl->Tilt_Q14, psEncCtrl->LF_shp_Q14, psEncCtrl->Gains_Q16, psEncCtrl->pitchL, psEncCtrl->Lambda_Q10, psEncCtrl->LTP_scale_Q14 ); diff --git a/silk/fixed/find_pred_coefs_FIX.c b/silk/fixed/find_pred_coefs_FIX.c index 5f67e6387..2053905c3 100644 --- a/silk/fixed/find_pred_coefs_FIX.c +++ b/silk/fixed/find_pred_coefs_FIX.c @@ -118,7 +118,7 @@ void silk_find_pred_coefs_FIX( minInvGain_Q30 = silk_log2lin( silk_SMLAWB( 16 << 7, 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( 0.1, 18 ), SILK_FIX_CONST( 0.9, 18 ), psEncCtrl->coding_quality_Q14 ) ), 14 ); + 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 */ diff --git a/silk/fixed/prefilter_FIX.c b/silk/fixed/prefilter_FIX.c index fe663fa85..89727ed44 100644 --- a/silk/fixed/prefilter_FIX.c +++ b/silk/fixed/prefilter_FIX.c @@ -36,7 +36,7 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. static inline void silk_prefilt_FIX( silk_prefilter_state_FIX *P, /* I/O state */ opus_int32 st_res_Q12[], /* I short term residual signal */ - opus_int32 pxw_Q10[], /* O prefiltered signal */ + opus_int32 xw_Q3[], /* O prefiltered signal */ opus_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */ opus_int Tilt_Q14, /* I Tilt shaping coeficient */ opus_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients */ @@ -88,7 +88,7 @@ void silk_warped_LPC_analysis_filter_FIX( void silk_prefilter_FIX( silk_encoder_state_FIX *psEnc, /* I/O Encoder state */ const silk_encoder_control_FIX *psEncCtrl, /* I Encoder control */ - opus_int32 xw_Q10[], /* O Weighted signal */ + opus_int32 xw_Q3[], /* O Weighted signal */ const opus_int16 x[] /* I Speech signal */ ) { @@ -97,7 +97,7 @@ void silk_prefilter_FIX( opus_int32 tmp_32; const opus_int16 *AR1_shp_Q13; const opus_int16 *px; - opus_int32 *pxw_Q10; + opus_int32 *pxw_Q3; opus_int HarmShapeGain_Q12, Tilt_Q14; opus_int32 HarmShapeFIRPacked_Q12, LF_shp_Q14; opus_int32 x_filt_Q12[ MAX_SUB_FRAME_LENGTH ]; @@ -106,7 +106,7 @@ void silk_prefilter_FIX( /* Set up pointers */ px = x; - pxw_Q10 = xw_Q10; + pxw_Q3 = xw_Q3; lag = P->lagPrev; for( k = 0; k < psEnc->sCmn.nb_subfr; k++ ) { /* Update Variables that change per sub frame */ @@ -140,11 +140,10 @@ void silk_prefilter_FIX( } P->sHarmHP_Q2 = st_res_Q2[ psEnc->sCmn.subfr_length - 1 ]; - silk_prefilt_FIX( P, x_filt_Q12, pxw_Q10, HarmShapeFIRPacked_Q12, Tilt_Q14, - LF_shp_Q14, lag, psEnc->sCmn.subfr_length ); + silk_prefilt_FIX( P, x_filt_Q12, pxw_Q3, HarmShapeFIRPacked_Q12, Tilt_Q14, LF_shp_Q14, lag, psEnc->sCmn.subfr_length ); px += psEnc->sCmn.subfr_length; - pxw_Q10 += psEnc->sCmn.subfr_length; + pxw_Q3 += psEnc->sCmn.subfr_length; } P->lagPrev = psEncCtrl->pitchL[ psEnc->sCmn.nb_subfr - 1 ]; @@ -154,7 +153,7 @@ void silk_prefilter_FIX( static inline void silk_prefilt_FIX( silk_prefilter_state_FIX *P, /* I/O state */ opus_int32 st_res_Q12[], /* I short term residual signal */ - opus_int32 xw_Q10[], /* O prefiltered signal */ + opus_int32 xw_Q3[], /* O prefiltered signal */ opus_int32 HarmShapeFIRPacked_Q12, /* I Harmonic shaping coeficients */ opus_int Tilt_Q14, /* I Tilt shaping coeficient */ opus_int32 LF_shp_Q14, /* I Low-frequancy shaping coeficients */ @@ -194,7 +193,7 @@ static inline void silk_prefilt_FIX( LTP_shp_buf_idx = ( LTP_shp_buf_idx - 1 ) & LTP_MASK; LTP_shp_buf[ LTP_shp_buf_idx ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( sLF_MA_shp_Q12, 12 ) ); - xw_Q10[i] = silk_RSHIFT_ROUND( silk_SUB32( sLF_MA_shp_Q12, n_LTP_Q12 ), 2 ); + xw_Q3[i] = silk_RSHIFT_ROUND( silk_SUB32( sLF_MA_shp_Q12, n_LTP_Q12 ), 9 ); } /* Copy temp variable back to state */ diff --git a/silk/float/find_pred_coefs_FLP.c b/silk/float/find_pred_coefs_FLP.c index 718667788..9c0726a24 100644 --- a/silk/float/find_pred_coefs_FLP.c +++ b/silk/float/find_pred_coefs_FLP.c @@ -97,7 +97,7 @@ void silk_find_pred_coefs_FLP( minInvGain = 1.0f / MAX_PREDICTION_POWER_GAIN_AFTER_RESET; } else { minInvGain = (silk_float)powf( 2, psEncCtrl->LTPredCodGain / 3 ) / MAX_PREDICTION_POWER_GAIN; - minInvGain /= 0.1f + 0.9f * psEncCtrl->coding_quality; + minInvGain /= 0.25f + 0.75f * psEncCtrl->coding_quality; } /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ diff --git a/silk/float/wrappers_FLP.c b/silk/float/wrappers_FLP.c index ad9c2a59f..5e705b1df 100644 --- a/silk/float/wrappers_FLP.c +++ b/silk/float/wrappers_FLP.c @@ -102,7 +102,7 @@ void silk_NSQ_wrapper_FLP( ) { opus_int i, j; - opus_int32 x_Q10[ MAX_FRAME_LENGTH ]; + opus_int32 x_Q3[ MAX_FRAME_LENGTH ]; opus_int32 Gains_Q16[ MAX_NB_SUBFR ]; silk_DWORD_ALIGN opus_int16 PredCoef_Q12[ 2 ][ MAX_LPC_ORDER ]; opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ]; @@ -155,15 +155,15 @@ void silk_NSQ_wrapper_FLP( /* Convert input to fix */ for( i = 0; i < psEnc->sCmn.frame_length; i++ ) { - x_Q10[ i ] = silk_float2int( 1024.0 * x[ i ] ); + x_Q3[ i ] = silk_float2int( 8.0 * x[ i ] ); } /* Call NSQ */ if( psEnc->sCmn.nStatesDelayedDecision > 1 || psEnc->sCmn.warping_Q16 > 0 ) { - silk_NSQ_del_dec( &psEnc->sCmn, psNSQ, psIndices, x_Q10, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, + silk_NSQ_del_dec( &psEnc->sCmn, psNSQ, psIndices, x_Q3, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, AR2_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14 ); } else { - silk_NSQ( &psEnc->sCmn, psNSQ, psIndices, x_Q10, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, + silk_NSQ( &psEnc->sCmn, psNSQ, psIndices, x_Q3, pulses, PredCoef_Q12[ 0 ], LTPCoef_Q14, AR2_Q13, HarmShapeGain_Q14, Tilt_Q14, LF_shp_Q14, Gains_Q16, psEncCtrl->pitchL, Lambda_Q10, LTP_scale_Q14 ); } } diff --git a/silk/init_decoder.c b/silk/init_decoder.c index f744bc6ed..c1c21fdb8 100644 --- a/silk/init_decoder.c +++ b/silk/init_decoder.c @@ -43,7 +43,7 @@ opus_int silk_init_decoder( /* Used to deactivate LSF interpolation */ psDec->first_frame_after_reset = 1; - psDec->prev_inv_gain_Q16 = 65536; + psDec->prev_inv_gain_Q31 = silk_int32_MAX; /* Reset CNG state */ silk_CNG_Reset( psDec ); diff --git a/silk/main.h b/silk/main.h index 1d9e7b830..8397384b6 100644 --- a/silk/main.h +++ b/silk/main.h @@ -229,7 +229,7 @@ void silk_NSQ( const silk_encoder_state *psEncC, /* I/O Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q10[], /* I Prefiltered input signal */ + const opus_int32 x_Q3[], /* I Prefiltered input signal */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ @@ -248,7 +248,7 @@ void silk_NSQ_del_dec( const silk_encoder_state *psEncC, /* I/O Encoder State */ silk_nsq_state *NSQ, /* I/O NSQ state */ SideInfoIndices *psIndices, /* I/O Quantization Indices */ - const opus_int32 x_Q10[], /* I Prefiltered input signal */ + const opus_int32 x_Q3[], /* I Prefiltered input signal */ opus_int8 pulses[], /* O Quantized pulse signal */ const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ diff --git a/silk/structs.h b/silk/structs.h index 338281b82..aee53ce8a 100644 --- a/silk/structs.h +++ b/silk/structs.h @@ -52,7 +52,7 @@ typedef struct { opus_int sLTP_buf_idx; opus_int sLTP_shp_buf_idx; opus_int32 rand_seed; - opus_int32 prev_inv_gain_Q16; + opus_int32 prev_inv_gain_Q31; opus_int rewhite_flag; } silk_nsq_state; @@ -255,7 +255,7 @@ typedef struct { /* Decoder state */ /********************************/ typedef struct { - opus_int32 prev_inv_gain_Q16; + opus_int32 prev_inv_gain_Q31; opus_int32 exc_Q10[ MAX_FRAME_LENGTH ]; opus_int32 sLPC_Q14_buf[ MAX_LPC_ORDER ]; opus_int16 outBuf[ MAX_FRAME_LENGTH + 2 * MAX_SUB_FRAME_LENGTH ]; /* Buffer for output signal */ diff --git a/silk/tables_other.c b/silk/tables_other.c index 7e1dc99c7..9a3313924 100644 --- a/silk/tables_other.c +++ b/silk/tables_other.c @@ -49,7 +49,7 @@ const opus_int32 silk_TargetRate_table_WB[ TARGET_RATE_TAB_SZ ] = { 0, 10500, 14000, 17000, 21500, 28500, 42000, MAX_TARGET_RATE_BPS }; const opus_int16 silk_SNR_table_Q1[ TARGET_RATE_TAB_SZ ] = { - 19, 29, 35, 39, 44, 50, 60, 80 + 18, 29, 38, 40, 46, 52, 62, 84 }; /* Tables for stereo predictor coding */ -- GitLab