From f3a52497e1298d393cf5d34fb298caf9534183a2 Mon Sep 17 00:00:00 2001 From: Sandesh Venkatesh Date: Wed, 5 Mar 2025 11:51:18 +0530 Subject: [PATCH] Fix for 3GPP issue 1313: Multi-frame attenuation for MDCT-stereo 48kbps with input +10dB Link #1313 --- lib_enc/ivas_mc_param_enc_fx.c | 10 +++--- lib_enc/ivas_stereo_mdct_core_enc_fx.c | 8 ++--- lib_enc/stat_enc.h | 4 +-- lib_enc/transient_detection_fx.c | 48 ++++++++++---------------- 4 files changed, 30 insertions(+), 40 deletions(-) diff --git a/lib_enc/ivas_mc_param_enc_fx.c b/lib_enc/ivas_mc_param_enc_fx.c index 859ce3a9b..36324ae0f 100644 --- a/lib_enc/ivas_mc_param_enc_fx.c +++ b/lib_enc/ivas_mc_param_enc_fx.c @@ -1639,8 +1639,8 @@ static void ivas_param_mc_transient_detection_fx( push_wmops( "param_mc_trn_det" ); attackRatioThreshold_fx = hTranDet->transientDetector.attackRatioThreshold; - pSubblockNrg_fx = &hTranDet->subblockEnergies.subblockNrg[hParamMC->transient_detector_delay]; // Q0 - pAccSubblockNrg_fx = &hTranDet->subblockEnergies.accSubblockNrg[hParamMC->transient_detector_delay]; // Q0 + pSubblockNrg_fx = &hTranDet->subblockEnergies.subblockNrg[hParamMC->transient_detector_delay]; // Q(-1) + pAccSubblockNrg_fx = &hTranDet->subblockEnergies.accSubblockNrg[hParamMC->transient_detector_delay]; // Q(-1) bIsAttackPresent = FALSE; move16(); @@ -1651,8 +1651,8 @@ static void ivas_param_mc_transient_detection_fx( * if we had an attack very late in the last frame, * make the current frame also a transient one... */ test(); - IF( EQ_16( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg_fx[-1], 31, Mpy_32_16_1( pAccSubblockNrg_fx[-1], attackRatioThreshold_fx ), ( 31 + ATTACKTHRESHOLD_E ) ), 1 ) || - EQ_16( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg_fx[-2], 31, Mpy_32_16_1( pAccSubblockNrg_fx[-2], attackRatioThreshold_fx ), ( 31 + ATTACKTHRESHOLD_E ) ), 1 ) ) + IF( EQ_16( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg_fx[-1], 32, Mpy_32_16_1( pAccSubblockNrg_fx[-1], attackRatioThreshold_fx ), ( 32 + ATTACKTHRESHOLD_E ) ), 1 ) || + EQ_16( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg_fx[-2], 32, Mpy_32_16_1( pAccSubblockNrg_fx[-2], attackRatioThreshold_fx ), ( 32 + ATTACKTHRESHOLD_E ) ), 1 ) ) { bIsAttackPresent = TRUE; move16(); @@ -1661,7 +1661,7 @@ static void ivas_param_mc_transient_detection_fx( } FOR( i = 0; i < NSUBBLOCKS; i++ ){ - IF( EQ_16( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg_fx[i], 31, Mpy_32_16_1( pAccSubblockNrg_fx[i], attackRatioThreshold_fx ), ( 31 + ATTACKTHRESHOLD_E ) ), 1 ) ){ + IF( EQ_16( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg_fx[i], 32, Mpy_32_16_1( pAccSubblockNrg_fx[i], attackRatioThreshold_fx ), ( 32 + ATTACKTHRESHOLD_E ) ), 1 ) ){ bIsAttackPresent = TRUE; move16(); attackIndex = i; diff --git a/lib_enc/ivas_stereo_mdct_core_enc_fx.c b/lib_enc/ivas_stereo_mdct_core_enc_fx.c index b43c8dfdb..2a87ea4f9 100644 --- a/lib_enc/ivas_stereo_mdct_core_enc_fx.c +++ b/lib_enc/ivas_stereo_mdct_core_enc_fx.c @@ -53,13 +53,13 @@ static void sync_tcx_mode_fx( Encoder_State **st /* i/o: Encoder state */ ) { - const Word32 prevAccNrg0 = st[0]->hTranDet->subblockEnergies.accSubblockNrg[st[0]->hTranDet->subblockEnergies.nDelay]; /* Q0 */ + const Word32 prevAccNrg0 = st[0]->hTranDet->subblockEnergies.accSubblockNrg[st[0]->hTranDet->subblockEnergies.nDelay]; /* Q(-1) */ move32(); - const Word32 prevAccNrg1 = st[1]->hTranDet->subblockEnergies.accSubblockNrg[st[1]->hTranDet->subblockEnergies.nDelay]; /* Q0 */ + const Word32 prevAccNrg1 = st[1]->hTranDet->subblockEnergies.accSubblockNrg[st[1]->hTranDet->subblockEnergies.nDelay]; /* Q(-1) */ move32(); - const Word32 lastAccNrg0 = st[0]->hTranDet->subblockEnergies.accSubblockNrg[( st[0]->hTranDet->subblockEnergies.nDelay + st[0]->hTranDet->transientDetector.nSubblocksToCheck )]; /* Q0 */ + const Word32 lastAccNrg0 = st[0]->hTranDet->subblockEnergies.accSubblockNrg[( st[0]->hTranDet->subblockEnergies.nDelay + st[0]->hTranDet->transientDetector.nSubblocksToCheck )]; /* Q(-1) */ move32(); - const Word32 lastAccNrg1 = st[1]->hTranDet->subblockEnergies.accSubblockNrg[( st[1]->hTranDet->subblockEnergies.nDelay + st[1]->hTranDet->transientDetector.nSubblocksToCheck )]; /* Q0 */ + const Word32 lastAccNrg1 = st[1]->hTranDet->subblockEnergies.accSubblockNrg[( st[1]->hTranDet->subblockEnergies.nDelay + st[1]->hTranDet->transientDetector.nSubblocksToCheck )]; /* Q(-1) */ move32(); test(); diff --git a/lib_enc/stat_enc.h b/lib_enc/stat_enc.h index 5067cb410..6292c4b33 100644 --- a/lib_enc/stat_enc.h +++ b/lib_enc/stat_enc.h @@ -138,8 +138,8 @@ typedef struct typedef struct { DelayBuffer *pDelayBuffer; /* Delay buffer. */ - Word32 subblockNrg[NSUBBLOCKS + MAX_TD_DELAY]; // IVAS Q0 - Word32 accSubblockNrg[NSUBBLOCKS + MAX_TD_DELAY + 1]; // IVAS Q0 + Word32 subblockNrg[NSUBBLOCKS + MAX_TD_DELAY]; // IVAS Q(-1) + Word32 accSubblockNrg[NSUBBLOCKS + MAX_TD_DELAY + 1]; // IVAS Q(-1) Word16 subblockNrgChange[NSUBBLOCKS + MAX_TD_DELAY]; /* EVS: Q7(15 - SUBBLOCK_NRG_CHANGE_E) */ /* IVAS: Q3 */ int16_t nDelay; /* Size of the delay buffer in use, as number of subblocks. Maximum delay from all users of this buffer. */ int16_t nPartialDelay; /* Delay of the input (modulo pDelayBuffer->nSubblockSize), nPartialDelay <= pDelayBuffer->nDelay. */ diff --git a/lib_enc/transient_detection_fx.c b/lib_enc/transient_detection_fx.c index d9859a1a1..82d24ceba 100644 --- a/lib_enc/transient_detection_fx.c +++ b/lib_enc/transient_detection_fx.c @@ -726,7 +726,7 @@ void RunTransientDetection_ivas_fx( e0_fx = dotp_fx( filteredInput_fx + /* length / 2 */ shr( length, 1 ), filteredInput_fx + /* length / 2 */ shr( length, 1 ), shr( pSubblockEnergies->pDelayBuffer->nSubblockSize, 1 ) /* pSubblockEnergies->pDelayBuffer->nSubblockSize / 2 */, &exp ); exp = sub( exp, 2 * 0 - 1 ); // Q = 2*0 + (30-exp), E = 31 - (2*0 + 30 - exp) = 1 + exp - 2*0 = exp - (2*0-1) e0_fx = BASOP_Util_Add_Mant32Exp( MIN_BLOCK_ENERGY_IVAS_FX_Q7 / 2, 31 - Q7, e0_fx, exp, &exp ); - e1_fx = BASOP_Util_Add_Mant32Exp( pSubblockEnergies->subblockNrg[pSubblockEnergies->nDelay + 4], 31 - 0, L_negate( e0_fx ), exp, &exp1 ); + e1_fx = BASOP_Util_Add_Mant32Exp( pSubblockEnergies->subblockNrg[pSubblockEnergies->nDelay + 4], 32, L_negate( e0_fx ), exp, &exp1 ); IF( BASOP_Util_Cmp_Mant32Exp( e1_fx, exp1, e0_fx, exp ) > 0 ) { pSubblockEnergies->ramp_up_flag = (UWord16) L_or( pSubblockEnergies->ramp_up_flag, 0x0001 ); @@ -1122,8 +1122,8 @@ static void InitSubblockEnergies_ivas_fx( Word16 nFrameLength, Word16 nDelay, De assert( ( pDelayBuffer != NULL ) && ( pSubblockEnergies != NULL ) && ( pDelayBuffer->nSubblockSize * NSUBBLOCKS == nFrameLength ) && ( pDelayBuffer->nSubblockSize > 0 ) ); - set32_fx( pSubblockEnergies->subblockNrg, 107 /* 107.37 in Q0 */, nMaxBuffSize ); - set32_fx( pSubblockEnergies->accSubblockNrg, 107 /* 107.37 in Q0 */, nMaxBuffSize + 1 ); + set32_fx( pSubblockEnergies->subblockNrg, 54 /* 107.37 in Q(-1) */, nMaxBuffSize ); + set32_fx( pSubblockEnergies->accSubblockNrg, 54 /* 107.37 in Q(-1) */, nMaxBuffSize + 1 ); set16_fx( pSubblockEnergies->subblockNrgChange, ONE_IN_Q3, nMaxBuffSize ); IF( nDelay != 0 ) { @@ -1495,9 +1495,9 @@ static void CalculateSubblockEnergies_ivas_fx( Word16 const *input, Word16 nSamp Word16 *pSubblockNrgChange; Word32 *pAccSubblockTmp; Word16 nWindows; - Word16 w, k, k2, tmp; + Word16 w, k, k2; Word32 w0, w1; - Word32 accu; + Word64 accu; move16(); pDelayBuffer = pSubblockEnergies->pDelayBuffer; @@ -1522,45 +1522,35 @@ static void CalculateSubblockEnergies_ivas_fx( Word16 const *input, Word16 nSamp IF( nWindows > 0 ) { /* Process left over samples from the previous frame. */ - accu = 107; // 107.37f in Q0 - move32(); - assert( ( SUBBLOCK_NRG_E & 1 ) == 0 ); + accu = 215; // 107.37f in Q1 + move64(); FOR( k = 0; k < nPartialDelay; k++ ) { - tmp = delayBuffer[k]; // Q0 - move16(); - accu = L_mac0( accu, tmp, tmp ); // Q0 + accu = W_mac_16_16( accu, delayBuffer[k], delayBuffer[k] ); // Q1 } /* Process new samples in the 0. subblock. */ - w = sub( nSubblockSize, nPartialDelay ); - assert( ( SUBBLOCK_NRG_E & 1 ) == 0 ); - FOR( k = 0; k < w; k++ ) + FOR( k = 0; k < ( nSubblockSize - nPartialDelay ); k++ ) { - tmp = input[k]; // Q0 - move16(); - accu = L_mac0_sat( accu, tmp, tmp ); // Q0 + accu = W_mac_16_16( accu, input[k], input[k] ); // Q1 } - pSubblockNrg[0] = accu; // Q0 + pSubblockNrg[0] = W_shl_sat_l( accu, -2 ); // Q(-1) move32(); /* Set accumulated subblock energy at this point. */ UpdatedAndStoreAccWindowNrg( pSubblockNrg[0], pAccSubblockTmp, facAccSubblockNrg, &pAccSubblockNrg[0] ); FOR( w = 1; w < nWindows; w++ ) { - accu = 107; // 107.37f in Q0 - move32(); + accu = 215; // 107.37f in Q1 + move64(); /* Process new samples in the w. subblock. */ k2 = add( k, nSubblockSize ); - assert( ( SUBBLOCK_NRG_E & 1 ) == 0 ); FOR( ; k < k2; k++ ) { - tmp = input[k]; // Q0 - move16(); - accu = L_mac0_sat( accu, tmp, tmp ); // Q0 + accu = W_mac_16_16( accu, input[k], input[k] ); // Q1 } - pSubblockNrg[w] = accu; // Q0 + pSubblockNrg[w] = W_shl_sat_l( accu, -2 ); // Q(-1) move32(); /* Set accumulated subblock energy at this point. */ UpdatedAndStoreAccWindowNrg( pSubblockNrg[w], pAccSubblockTmp, facAccSubblockNrg, &pAccSubblockNrg[w] ); @@ -1743,7 +1733,7 @@ Word16 transient_analysis_ivas_fx( /* Forward attack analysis */ FOR( i = -2; i < 7; i++ ) { - IF( BASOP_Util_Cmp_Mant32Exp( hTranDet->subblockEnergies.subblockNrg[nRelativeDelay + i], 31, Mpy_32_16_1( hTranDet->subblockEnergies.accSubblockNrg[nRelativeDelay + i], thr_fwd_fx ), ( 31 + 4 ) ) > 0 ) + IF( BASOP_Util_Cmp_Mant32Exp( hTranDet->subblockEnergies.subblockNrg[nRelativeDelay + i], 32, Mpy_32_16_1( hTranDet->subblockEnergies.accSubblockNrg[nRelativeDelay + i], thr_fwd_fx ), ( 32 + 4 ) ) > 0 ) { prel_force_td = s_or( prel_force_td, 0x0001 ); } @@ -1780,7 +1770,7 @@ Word16 transient_analysis_ivas_fx( /* -3 check */ test(); - IF( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg[1 + offset], 31, Mpy_32_16_1( accSubblockNrgRev_fx[1], thr_rev_fx ), ( 31 + 4 ) ) > 0 ) + IF( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg[1 + offset], 32, Mpy_32_16_1( accSubblockNrgRev_fx[1], thr_rev_fx ), ( 32 + 4 ) ) > 0 ) { prel_force_td = s_or( prel_force_td, 0x0002 ); move16(); @@ -1788,10 +1778,10 @@ Word16 transient_analysis_ivas_fx( /* -4 check */ test(); - IF( prel_force_td == 0 && BASOP_Util_Cmp_Mant32Exp( pSubblockNrg[offset], 31, Mpy_32_16_1( accSubblockNrgRev_fx[0], thr_rev_fx ), ( 31 + 4 ) ) > 0 ) + IF( prel_force_td == 0 && BASOP_Util_Cmp_Mant32Exp( pSubblockNrg[offset], 32, Mpy_32_16_1( accSubblockNrgRev_fx[0], thr_rev_fx ), ( 32 + 4 ) ) > 0 ) { - IF( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg[offset], 31, Mpy_32_16_1( accSubblockNrgRev_fx[0], add( thr_rev_fx, THR_LOW_STEP_FX ) ), ( 31 + 4 ) ) > 0 ) + IF( BASOP_Util_Cmp_Mant32Exp( pSubblockNrg[offset], 32, Mpy_32_16_1( accSubblockNrgRev_fx[0], add( thr_rev_fx, THR_LOW_STEP_FX ) ), ( 32 + 4 ) ) > 0 ) { prel_force_td = s_or( prel_force_td, 0x0004 ); } -- GitLab