diff --git a/silk/PLC.c b/silk/PLC.c index 50e0ddae7231a9d643fc74a59fd7a9b55346054e..7624a9189cc6d10a578220d5cf1597e36bf60370 100644 --- a/silk/PLC.c +++ b/silk/PLC.c @@ -166,28 +166,23 @@ static inline void silk_PLC_conceal( ) { opus_int i, j, k; - opus_int16 *B_Q14, exc_buf[ MAX_FRAME_LENGTH ], *exc_buf_ptr; - opus_int16 rand_scale_Q14, A_Q12_tmp[ MAX_LPC_ORDER ]; - opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15; opus_int lag, idx, sLTP_buf_idx, shift1, shift2; + opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q16, inv_gain_Q30; opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr; - opus_int32 sig_Q10[ MAX_FRAME_LENGTH ], *sig_Q10_ptr, LPC_exc_Q10, LPC_pred_Q10, LTP_pred_Q14; - silk_PLC_struct *psPLC; - psPLC = &psDec->sPLC; - - /* Update LTP buffer */ - silk_memmove( psDec->sLTP_Q16, &psDec->sLTP_Q16[ psDec->frame_length ], psDec->ltp_mem_length * sizeof( opus_int32 ) ); - - /* LPC concealment. Apply BWE to previous LPC */ - silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) ); + opus_int32 LPC_exc_Q14, LPC_pred_Q10, LTP_pred_Q12; + opus_int16 rand_scale_Q14; + opus_int16 *B_Q14, *exc_buf_ptr; + opus_int32 *sLPC_Q14_ptr; + opus_int16 exc_buf[ 2 * MAX_SUB_FRAME_LENGTH ]; + opus_int16 A_Q12[ MAX_LPC_ORDER ]; + opus_int16 sLTP[ MAX_FRAME_LENGTH ]; + opus_int32 sLTP_Q14[ 2 * MAX_FRAME_LENGTH ]; + silk_PLC_struct *psPLC = &psDec->sPLC; /* Find random noise component */ /* Scale previous excitation signal */ exc_buf_ptr = exc_buf; - /* FIXME: JMV: Is this the right fix? */ - for (i=0;i<MAX_FRAME_LENGTH;i++) - exc_buf[i] = 0; - for( k = ( psDec->nb_subfr >> 1 ); k < psDec->nb_subfr; k++ ) { + for( k = psDec->nb_subfr - 2; k < psDec->nb_subfr; k++ ) { for( i = 0; i < psDec->subfr_length; i++ ) { exc_buf_ptr[ i ] = ( opus_int16 )silk_RSHIFT( silk_SMULWW( psDec->exc_Q10[ i + k * psDec->subfr_length ], psPLC->prevGain_Q16[ k ] ), 10 ); @@ -200,7 +195,7 @@ static inline void silk_PLC_conceal( if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) { /* First sub-frame has lowest energy */ - rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, 3 * psDec->subfr_length - RAND_BUF_SIZE ) ]; + rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, psDec->frame_length - psDec->subfr_length - RAND_BUF_SIZE ) ]; } else { /* Second sub-frame has lowest energy */ rand_ptr = &psDec->exc_Q10[ silk_max_int( 0, psDec->frame_length - RAND_BUF_SIZE ) ]; @@ -218,6 +213,12 @@ static inline void silk_PLC_conceal( rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ]; } + /* LPC concealment. Apply BWE to previous LPC */ + silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) ); + + /* Preload LPC coeficients to array on stack. Gives small performance gain */ + silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) ); + /* First Lost frame */ if( psDec->lossCnt == 0 ) { rand_scale_Q14 = 1 << 14; @@ -247,37 +248,42 @@ static inline void silk_PLC_conceal( lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 ); sLTP_buf_idx = psDec->ltp_mem_length; + /* Rewhiten LTP state */ + idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2; + silk_assert( idx > 0 ); + silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order ); + /* Scale LTP state */ + inv_gain_Q16 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ psDec->nb_subfr - 1 ], 32 ); + inv_gain_Q16 = silk_min( inv_gain_Q16, silk_int16_MAX ); + inv_gain_Q30 = silk_LSHIFT( inv_gain_Q16, 14 ); + for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) { + sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] ); + } + /***************************/ /* LTP synthesis filtering */ /***************************/ - sig_Q10_ptr = sig_Q10; for( k = 0; k < psDec->nb_subfr; k++ ) { /* Setup pointer */ - pred_lag_ptr = &psDec->sLTP_Q16[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; + pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; for( i = 0; i < psDec->subfr_length; i++ ) { - rand_seed = silk_RAND( rand_seed ); - idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK; - /* Unrolled loop */ - LTP_pred_Q14 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); - LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], B_Q14[ 1 ] ); - LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], B_Q14[ 2 ] ); - LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], B_Q14[ 3 ] ); - LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], B_Q14[ 4 ] ); + LTP_pred_Q12 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); + LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] ); + LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] ); + LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] ); + LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] ); pred_lag_ptr++; - /* Generate LPC residual */ - LPC_exc_Q10 = silk_LSHIFT( silk_SMULWB( rand_ptr[ idx ], rand_scale_Q14 ), 2 ); /* Random noise part */ - LPC_exc_Q10 = silk_ADD32( LPC_exc_Q10, silk_RSHIFT_ROUND( LTP_pred_Q14, 4 ) ); /* Harmonic part */ - - /* Update states */ - psDec->sLTP_Q16[ sLTP_buf_idx ] = silk_LSHIFT( LPC_exc_Q10, 6 ); + /* Generate LPC excitation */ + rand_seed = silk_RAND( rand_seed ); + idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK; + LPC_exc_Q14 = silk_LSHIFT32( silk_SMULWB( rand_ptr[ idx ], rand_scale_Q14 ), 6 ); /* Random noise part */ + LPC_exc_Q14 = silk_ADD32( LPC_exc_Q14, silk_LSHIFT32( LTP_pred_Q12, 2 ) ); /* Harmonic part */ + sLTP_Q14[ sLTP_buf_idx ] = LPC_exc_Q14; sLTP_buf_idx++; - - /* Save LPC residual */ - sig_Q10_ptr[ i ] = LPC_exc_Q10; } - sig_Q10_ptr += psDec->subfr_length; + /* Gradually reduce LTP gain */ for( j = 0; j < LTP_ORDER; j++ ) { B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 ); @@ -286,7 +292,7 @@ static inline void silk_PLC_conceal( rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 ); /* Slowly increase pitch lag */ - psPLC->pitchL_Q8 += silk_SMULWB( psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 ); + psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 ); psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) ); lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 ); } @@ -294,44 +300,38 @@ static inline void silk_PLC_conceal( /***************************/ /* LPC synthesis filtering */ /***************************/ - sig_Q10_ptr = sig_Q10; - /* Preload LPC coeficients to array on stack. Gives small performance gain */ - silk_memcpy( A_Q12_tmp, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) ); - silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */ - for( k = 0; k < psDec->nb_subfr; k++ ) { - for( i = 0; i < psDec->subfr_length; i++ ){ - /* partly unrolled */ - LPC_pred_Q10 = silk_SMULWB( psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 1 ], A_Q12_tmp[ 0 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 2 ], A_Q12_tmp[ 1 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 3 ], A_Q12_tmp[ 2 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 4 ], A_Q12_tmp[ 3 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 5 ], A_Q12_tmp[ 4 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 6 ], A_Q12_tmp[ 5 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 7 ], A_Q12_tmp[ 6 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 8 ], A_Q12_tmp[ 7 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 9 ], A_Q12_tmp[ 8 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 10 ], A_Q12_tmp[ 9 ] ); - - for( j = 10; j < psDec->LPC_order; j++ ) { - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - j - 1 ], A_Q12_tmp[ j ] ); - } + sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ]; - /* Add prediction to LPC residual */ - sig_Q10_ptr[ i ] = silk_ADD32( sig_Q10_ptr[ i ], LPC_pred_Q10 ); + /* Copy LPC state */ + silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) ); - /* Update states */ - psDec->sLPC_Q14[ MAX_LPC_ORDER + i ] = silk_LSHIFT( sig_Q10_ptr[ i ], 4 ); + silk_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */ + for( i = 0; i < psDec->frame_length; i++ ) { + /* partly unrolled */ + LPC_pred_Q10 = silk_SMULWB( sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 1 ], A_Q12[ 0 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 2 ], A_Q12[ 1 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 3 ], A_Q12[ 2 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 4 ], A_Q12[ 3 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 5 ], A_Q12[ 4 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 6 ], A_Q12[ 5 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 7 ], A_Q12[ 6 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 8 ], A_Q12[ 7 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 9 ], A_Q12[ 8 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] ); + for( j = 10; j < psDec->LPC_order; j++ ) { + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] ); } - sig_Q10_ptr += psDec->subfr_length; - /* Update LPC filter state */ - silk_memcpy( psDec->sLPC_Q14, &psDec->sLPC_Q14[ psDec->subfr_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); - } - /* Scale with Gain */ - for( i = 0; i < psDec->frame_length; i++ ) { - frame[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sig_Q10[ i ], psPLC->prevGain_Q16[ psDec->nb_subfr - 1 ] ), 10 ) ); + /* Add prediction to LPC excitation */ + sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_LSHIFT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], LPC_pred_Q10, 4 ); + + /* Scale with Gain */ + frame[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], psPLC->prevGain_Q16[ psDec->nb_subfr - 1 ] ), 14 ) ); } + /* Save LPC state */ + silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); + /**************************************/ /* Update states */ /**************************************/ diff --git a/silk/dec_API.c b/silk/dec_API.c index 5e676932e07a6e0beb76f3a2d2dc37a8fd85abd6..a66eeaf418eed5dbe174dc5dfd48f299736042b3 100644 --- a/silk/dec_API.c +++ b/silk/dec_API.c @@ -227,7 +227,7 @@ opus_int silk_Decode( ( lostFlag == FLAG_DECODE_LBRR && channel_state[ 0 ].LBRR_flags[ channel_state[ 0 ].nFramesDecoded ] == 1 ) ) { silk_stereo_decode_pred( psRangeDec, MS_pred_Q13 ); - /* For LBRR data, only decode mid-only flag if side-channel's LBRR flag is false */ + /* For LBRR data, decode mid-only flag only if side-channel's LBRR flag is false */ if( lostFlag == FLAG_DECODE_NORMAL || ( lostFlag == FLAG_DECODE_LBRR && channel_state[ 1 ].LBRR_flags[ channel_state[ 0 ].nFramesDecoded ] == 0 ) ) { diff --git a/silk/decode_core.c b/silk/decode_core.c index b80a3895798bfb8b16b87cf078223b18eee76bad..fd7283fe0083e2eb7fc7711377772597a09135c4 100644 --- a/silk/decode_core.c +++ b/silk/decode_core.c @@ -44,10 +44,11 @@ void silk_decode_core( opus_int i, j, k, lag = 0, start_idx, sLTP_buf_idx, NLSF_interpolation_flag, signalType; opus_int16 *A_Q12, *B_Q14, *pxq, A_Q12_tmp[ MAX_LPC_ORDER ]; opus_int16 sLTP[ MAX_FRAME_LENGTH ]; + opus_int32 sLTP_Q16[ 2 * MAX_FRAME_LENGTH ]; opus_int32 LTP_pred_Q14, LPC_pred_Q10, Gain_Q10, inv_gain_Q16, inv_gain_Q32, 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 vec_Q10[ MAX_SUB_FRAME_LENGTH ]; + opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + MAX_LPC_ORDER ]; silk_assert( psDec->prev_inv_gain_Q16 != 0 ); @@ -80,8 +81,11 @@ void silk_decode_core( DEBUG_STORE_DATA( dec_q.dat, pulses, psDec->frame_length * sizeof( opus_int ) ); #endif + /* Copy LPC state */ + silk_memcpy( sLPC_Q14, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) ); + pexc_Q10 = psDec->exc_Q10; - pxq = &psDec->outBuf[ psDec->ltp_mem_length ]; + pxq = xq; sLTP_buf_idx = psDec->ltp_mem_length; /* Loop over subframes */ for( k = 0; k < psDec->nb_subfr; k++ ) { @@ -104,7 +108,7 @@ void silk_decode_core( /* Scale short term state */ for( i = 0; i < MAX_LPC_ORDER; i++ ) { - psDec->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDec->sLPC_Q14[ i ] ); + sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, sLPC_Q14[ i ] ); } } @@ -128,28 +132,32 @@ void silk_decode_core( lag = psDecCtrl->pitchL[ k ]; /* Re-whitening */ - if( ( k & ( 3 - silk_LSHIFT( NLSF_interpolation_flag, 1 ) ) ) == 0 ) { + if( k == 0 || ( k == 2 && NLSF_interpolation_flag ) ) { /* Rewhiten with new A coefs */ start_idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2; silk_assert( start_idx > 0 ); + if( k == 2 ) { + silk_memcpy( &psDec->outBuf[ psDec->ltp_mem_length ], xq, 2 * psDec->subfr_length * sizeof( opus_int16 ) ); + } + silk_LPC_analysis_filter( &sLTP[ start_idx ], &psDec->outBuf[ start_idx + k * psDec->subfr_length ], A_Q12, psDec->ltp_mem_length - start_idx, psDec->LPC_order ); /* After rewhitening the LTP state is unscaled */ inv_gain_Q32 = silk_LSHIFT( inv_gain_Q16, 16 ); if( k == 0 ) { - /* Do LTP downscaling */ + /* Do LTP downscaling to reduce inter-packet dependency */ inv_gain_Q32 = silk_LSHIFT( silk_SMULWB( inv_gain_Q32, psDecCtrl->LTP_scale_Q14 ), 2 ); } for( i = 0; i < lag + LTP_ORDER/2; i++ ) { - psDec->sLTP_Q16[ sLTP_buf_idx - i - 1 ] = silk_SMULWB( inv_gain_Q32, sLTP[ psDec->ltp_mem_length - i - 1 ] ); + sLTP_Q16[ sLTP_buf_idx - i - 1 ] = silk_SMULWB( inv_gain_Q32, sLTP[ psDec->ltp_mem_length - i - 1 ] ); } } else { /* Update LTP state when Gain changes */ if( gain_adj_Q16 != 1 << 16 ) { for( i = 0; i < lag + LTP_ORDER/2; i++ ) { - psDec->sLTP_Q16[ sLTP_buf_idx - i - 1 ] = silk_SMULWW( gain_adj_Q16, psDec->sLTP_Q16[ sLTP_buf_idx - i - 1 ] ); + sLTP_Q16[ sLTP_buf_idx - i - 1 ] = silk_SMULWW( gain_adj_Q16, sLTP_Q16[ sLTP_buf_idx - i - 1 ] ); } } } @@ -158,7 +166,7 @@ void silk_decode_core( /* Long-term prediction */ if( signalType == TYPE_VOICED ) { /* Setup pointer */ - pred_lag_ptr = &psDec->sLTP_Q16[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; + pred_lag_ptr = &sLTP_Q16[ sLTP_buf_idx - lag + LTP_ORDER / 2 ]; for( i = 0; i < psDec->subfr_length; i++ ) { /* Unrolled loop */ LTP_pred_Q14 = silk_SMULWB( pred_lag_ptr[ 0 ], B_Q14[ 0 ] ); @@ -172,7 +180,7 @@ void silk_decode_core( pres_Q10[ i ] = silk_ADD32( pexc_Q10[ i ], silk_RSHIFT_ROUND( LTP_pred_Q14, 4 ) ); /* Update states */ - psDec->sLTP_Q16[ sLTP_buf_idx ] = silk_LSHIFT( pres_Q10[ i ], 6 ); + sLTP_Q16[ sLTP_buf_idx ] = silk_LSHIFT( pres_Q10[ i ], 6 ); sLTP_buf_idx++; } } else { @@ -186,43 +194,38 @@ void silk_decode_core( for( i = 0; i < psDec->subfr_length; i++ ) { /* Partially unrolled */ - LPC_pred_Q10 = silk_SMULWB( psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 1 ], A_Q12_tmp[ 0 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 2 ], A_Q12_tmp[ 1 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 3 ], A_Q12_tmp[ 2 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 4 ], A_Q12_tmp[ 3 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 5 ], A_Q12_tmp[ 4 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 6 ], A_Q12_tmp[ 5 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 7 ], A_Q12_tmp[ 6 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 8 ], A_Q12_tmp[ 7 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 9 ], A_Q12_tmp[ 8 ] ); - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - 10 ], A_Q12_tmp[ 9 ] ); + LPC_pred_Q10 = silk_SMULWB( 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 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 5 ], A_Q12_tmp[ 4 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 6 ], A_Q12_tmp[ 5 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 7 ], A_Q12_tmp[ 6 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 8 ], A_Q12_tmp[ 7 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 9 ], A_Q12_tmp[ 8 ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - 10 ], A_Q12_tmp[ 9 ] ); for( j = 10; j < psDec->LPC_order; j++ ) { - LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, psDec->sLPC_Q14[ MAX_LPC_ORDER + i - j - 1 ], A_Q12_tmp[ j ] ); + LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14[ MAX_LPC_ORDER + i - j - 1 ], A_Q12_tmp[ j ] ); } /* Add prediction to LPC excitation */ - vec_Q10[ i ] = silk_ADD32( pres_Q10[ i ], LPC_pred_Q10 ); + sLPC_Q14[ MAX_LPC_ORDER + i ] = silk_LSHIFT( silk_ADD32( pres_Q10[ i ], LPC_pred_Q10 ), 4 ); - /* Update states */ - psDec->sLPC_Q14[ MAX_LPC_ORDER + i ] = silk_LSHIFT( vec_Q10[ i ], 4 ); - } - - /* Scale with Gain */ - for( i = 0; i < psDec->subfr_length; i++ ) { - pxq[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( vec_Q10[ i ], Gain_Q10 ), 4 ) ); + /* Scale with Gain */ + pxq[ i ] = ( opus_int16 )silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14[ MAX_LPC_ORDER + i ], Gain_Q10 ), 8 ) ); } /* Update LPC filter state */ - silk_memcpy( psDec->sLPC_Q14, &psDec->sLPC_Q14[ psDec->subfr_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); + silk_memcpy( sLPC_Q14, &sLPC_Q14[ psDec->subfr_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) ); pexc_Q10 += psDec->subfr_length; pxq += psDec->subfr_length; } - /* Copy to output */ - silk_memcpy( xq, &psDec->outBuf[ psDec->ltp_mem_length ], psDec->frame_length * sizeof( opus_int16 ) ); + /* Save LPC state */ + silk_memcpy( psDec->sLPC_Q14_buf, sLPC_Q14, MAX_LPC_ORDER * sizeof( opus_int32 ) ); #ifdef SAVE_ALL_INTERNAL_DATA - DEBUG_STORE_DATA( dec_sLTP_Q16.dat, &psDec->sLTP_Q16[ psDec->ltp_mem_length ], psDec->frame_length * sizeof( opus_int32 )); + DEBUG_STORE_DATA( dec_sLTP_Q16.dat, &sLTP_Q16[ psDec->ltp_mem_length ], psDec->frame_length * sizeof( opus_int32 )); DEBUG_STORE_DATA( dec_xq.dat, xq, psDec->frame_length * sizeof( opus_int16 ) ); #endif } diff --git a/silk/decoder_set_fs.c b/silk/decoder_set_fs.c index be29a97763fc538fce3c4b5595e60b881d2d33ff..c2430f8c3ad195c15d679b82208de05b946fa1a5 100644 --- a/silk/decoder_set_fs.c +++ b/silk/decoder_set_fs.c @@ -72,7 +72,7 @@ void silk_decoder_set_fs( if( psDec->fs_kHz != fs_kHz) { /* Reset part of the decoder state */ - silk_memset( psDec->sLPC_Q14, 0, sizeof( psDec->sLPC_Q14 ) ); + silk_memset( psDec->sLPC_Q14_buf, 0, sizeof( psDec->sLPC_Q14_buf ) ); silk_memset( psDec->outBuf, 0, MAX_FRAME_LENGTH * sizeof( opus_int16 ) ); silk_memset( psDec->prevNLSF_Q15, 0, sizeof( psDec->prevNLSF_Q15 ) ); } diff --git a/silk/structs.h b/silk/structs.h index 70c81baeebff1cacc1f7aee7169225fbb32253a3..c3de024144d236e2cb7b184ea962ab614af59ee6 100644 --- a/silk/structs.h +++ b/silk/structs.h @@ -255,10 +255,9 @@ typedef struct { /********************************/ typedef struct { opus_int32 prev_inv_gain_Q16; - opus_int32 sLTP_Q16[ 2 * MAX_FRAME_LENGTH ]; - opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + MAX_LPC_ORDER ]; opus_int32 exc_Q10[ MAX_FRAME_LENGTH ]; - opus_int16 outBuf[ 2 * MAX_FRAME_LENGTH ]; /* Buffer for output signal */ + opus_int32 sLPC_Q14_buf[ MAX_LPC_ORDER ]; + opus_int16 outBuf[ MAX_FRAME_LENGTH + 2 * MAX_SUB_FRAME_LENGTH ]; /* Buffer for output signal */ opus_int16 delayBuf[ MAX_DECODER_DELAY ]; /* Buffer for delaying the SILK output prior to resampling */ opus_int delay; /* How much decoder delay to add */ opus_int lagPrev; /* Previous Lag */ diff --git a/src/opus_decoder.c b/src/opus_decoder.c index 6b72efa2fa2f91eb18fd9207aaf1ff5995ee4be3..cd4079d5077a2779677a48a9ccee4f643fb6a408 100644 --- a/src/opus_decoder.c +++ b/src/opus_decoder.c @@ -418,7 +418,7 @@ static int opus_decode_frame(OpusDecoder *st, const unsigned char *data, pcm[i] = SAT16(pcm[i] + pcm_silk[i]); #else for (i=0;i<frame_size*st->channels;i++) - pcm[i] = pcm[i] + (1./32768.)*pcm_silk[i]; + pcm[i] = pcm[i] + (opus_val16)((1./32768.)*pcm_silk[i]); #endif }