From 613673b0883bf0e0c18f0ba671f8905e0f73419a Mon Sep 17 00:00:00 2001 From: Sandesh Venkatesh Date: Wed, 12 Feb 2025 21:20:48 +0530 Subject: [PATCH] Bug fixes in orig-specturm q, in functions ITF_Detect_fx and Q_lsf_tcxlpc_fx --- lib_com/prot_fx.h | 10 +++ lib_com/tns_base.c | 139 ++++++++++++++++++++++++++++++ lib_enc/igf_enc.c | 4 +- lib_enc/ivas_core_enc.c | 8 +- lib_enc/ivas_tcx_core_enc.c | 2 +- lib_enc/lsf_msvq_ma_enc_fx.c | 159 +++++++++++++++++++++++++++++++++++ lib_enc/prot_fx_enc.h | 13 +++ lib_enc/tcx_utils_enc_fx.c | 2 +- 8 files changed, 329 insertions(+), 8 deletions(-) diff --git a/lib_com/prot_fx.h b/lib_com/prot_fx.h index 9fae6a1c1..0e093b77f 100644 --- a/lib_com/prot_fx.h +++ b/lib_com/prot_fx.h @@ -3642,6 +3642,16 @@ Word16 ITF_Detect_fx( Word32 const pSpectrum[], Word16 *curr_order, Word16 Q ); +Word16 ITF_Detect_ivas_fx( Word32 const pSpectrum[], + const Word16 startLine, + const Word16 stopLine, + const Word16 maxOrder, + Word16 *A, + Word16 *Q_A, + Word16 *predictionGain, + Word16 *curr_order, + Word16 Q ); + void ITF_Apply_fx( Word32 spectrum[], Word16 startLine, Word16 stopLine, diff --git a/lib_com/tns_base.c b/lib_com/tns_base.c index d4de526d6..af4a87d34 100644 --- a/lib_com/tns_base.c +++ b/lib_com/tns_base.c @@ -602,6 +602,145 @@ Word16 ITF_Detect_fx( return 1; } + +Word16 ITF_Detect_ivas_fx( + const Word32 pSpectrum[], /*Q*/ + const Word16 startLine, /*Q0*/ + const Word16 stopLine, /*Q0*/ + const Word16 maxOrder, /*Q0*/ + Word16 *A, /*Q_A*/ + Word16 *Q_A, + Word16 *predictionGain, /*Q7*/ + Word16 *curr_order, /*Q0*/ + Word16 Q ) +{ + Word32 norms[MAX_SUBDIVISIONS]; + Word16 num_subdivisions, i, length; + Word16 iStartLine, iEndLine, spectrumLength; + Word32 rxx[ITF_MAX_FILTER_ORDER + 1]; + Word16 q_rxx[ITF_MAX_FILTER_ORDER + 1]; + Word32 temp_spectrum[640]; + const Word16 *pWindow; + const Word32 *ptr_spectrum1, *ptr_spectrum2; + Word16 iSubdivisions, lag; + Word16 headroom, guard_bits, shift, q_min; + Word64 sum; + Word16 fac, q_fac, exp, q_temp; + Word32 temp; + + IF( maxOrder <= 0 ) + { + return 0; + } + pWindow = tnsAcfWindow_fx; + set_zero_fx( norms, MAX_SUBDIVISIONS ); + set_zero_fx( rxx, ITF_MAX_FILTER_ORDER + 1 ); /* This initialization is required */ + set16_fx( q_rxx, Q31, ITF_MAX_FILTER_ORDER + 1 ); /* This initialization is required */ + + spectrumLength = sub( stopLine, startLine ); + num_subdivisions = 0; + move16(); + + /* Calculate norms for each spectrum part */ + FOR( iSubdivisions = 0; iSubdivisions < MAX_SUBDIVISIONS; iSubdivisions++ ) + { + /* iStartLine = startLine + ( stopLine - startLine ) * iSubdivisions / nSubdivisions; */ + iStartLine = add( startLine, mult( imult1616( spectrumLength, iSubdivisions ), 10923 /* 1/MAX_SUBDIVISIONS in Q15 */ ) ); /*Q0*/ + /* iEndLine = startLine + ( stopLine - startLine ) * ( iSubdivisions + 1 ) / nSubdivisions; */ + iEndLine = add( startLine, mult( imult1616( spectrumLength, add( iSubdivisions, 1 ) ), 10923 /* 1/MAX_SUBDIVISIONS in Q15 */ ) ); /*Q0*/ + + + /* Variable initialization */ + /* norms[iSubdivisions] = sum2_f(pSpectrum + iStartLine - IGF_START_MN, iEndLine - iStartLine); */ + + ptr_spectrum1 = pSpectrum + sub( iStartLine, IGF_START_MN ); + length = sub( iEndLine, iStartLine ); + headroom = L_norm_arr( ptr_spectrum1, length ); + guard_bits = find_guarded_bits_fx( length ); + shift = sub( headroom, guard_bits ); + + Copy_Scale_sig32( ptr_spectrum1, temp_spectrum, length, shift ); // Q -> Q+shift + + sum = 0; + move64(); + FOR( i = 0; i < length; i++ ) + { + sum = W_mac_32_32( sum, temp_spectrum[i], temp_spectrum[i] ); // 2(Q+shift)+1 + } + + IF( LE_64( sum, W_shl( 32768 * 2 /* HLM_MIN_NRG in Q1 */, shl( add( Q, shift ), 1 ) ) ) ) + { + BREAK; + } + + /* fac = 1.0f / norms[iSubdivisions]; */ + exp = W_norm( sum ); + sum = W_shl( sum, exp ); // 2(Q+shift)+1+exp + fac = div_s( ONE_IN_Q14, extract_h( W_extract_h( sum ) ) ); // 15+14-(2(Q+shift)+1+exp-48) = 76-(2(Q+shift)+exp) + q_fac = sub( 76, add( shl( add( Q, shift ), 1 ), exp ) ); + pWindow = tnsAcfWindow_fx; + + /* For additional loop condition */ + /* Variable initialization */ + /*for ( lag = 1; lag <= maxOrder; lag++ ) + { + rxx[lag] += fac * ( *pWindow ) * dotp( pSpectrum + iStartLine - IGF_START_MN, pSpectrum + iStartLine - IGF_START_MN + lag, iEndLine - iStartLine - lag ); + pWindow++; + }*/ + + ptr_spectrum1 = temp_spectrum; // pSpectrum + iStartLine - IGF_START_MN; + FOR( lag = 1; lag <= maxOrder; lag++ ) + { + /* dotp( pSpectrum + iStartLine - IGF_START_MN, pSpectrum + iStartLine - IGF_START_MN + lag, iEndLine - iStartLine - lag ) */ + ptr_spectrum2 = temp_spectrum + lag; // pSpectrum + iStartLine - IGF_START_MN + lag; + + sum = 0; + move64(); + FOR( i = 0; i < iEndLine - iStartLine - lag; i++ ) + { + sum = W_mac_32_32( sum, ptr_spectrum1[i], ptr_spectrum2[i] ); // 2(Q+shift)+1 + } + exp = W_norm( sum ); + sum = W_shl( sum, exp ); // 2(Q+shift)+1+exp + temp = Mpy_32_32( L_mult0( fac, *pWindow ), W_extract_h( sum ) ); // (q_fac+15)+(2(Q+shift)+1+exp-32)-31 = q_fac+2(Q+shift)+exp-47 + q_temp = sub( add( q_fac, add( shl( add( Q, shift ), 1 ), exp ) ), 47 ); + + /* rxx[lag] += fac * (*pWindow) * dotp(pSpectrum + iStartLine - IGF_START_MN, pSpectrum + iStartLine - IGF_START_MN + lag, iEndLine - iStartLine - lag); */ + q_min = sub( s_min( q_temp, q_rxx[lag] ), 1 ); + rxx[lag] = L_add( L_shl( rxx[lag], sub( q_min, q_rxx[lag] ) ), L_shl( temp, sub( q_min, q_temp ) ) ); + q_rxx[lag] = q_min; + move32(); + move16(); + + pWindow++; + } + + num_subdivisions = add( num_subdivisions, 1 ); + } + + minimum_s( q_rxx + 1, ITF_MAX_FILTER_ORDER, &q_min ); + q_min = s_min( Q29, q_min ); + + FOR( i = 1; i < ITF_MAX_FILTER_ORDER; i++ ) + { + rxx[i] = L_shl( rxx[i], sub( q_min, q_rxx[i] ) ); + move32(); + } + + IF( EQ_16( iSubdivisions, MAX_SUBDIVISIONS ) ) /* meaning there is no subdivision with low energy */ + { + rxx[0] = L_shl( MAX_SUBDIVISIONS, q_min ); + move32(); + + /* Limit the maximum order to spectrum length/4 */ + ITF_GetFilterParameters_fx( rxx, s_min( maxOrder, shr( spectrumLength, 2 ) ), A, Q_A, predictionGain ); + + *curr_order = maxOrder; /*Q0*/ + move16(); + } + + return 1; +} /* Helper functions for Hufmann table coding */ diff --git a/lib_enc/igf_enc.c b/lib_enc/igf_enc.c index 6cccea43b..a85843be2 100644 --- a/lib_enc/igf_enc.c +++ b/lib_enc/igf_enc.c @@ -575,8 +575,8 @@ static void IGF_CalculateEnvelope_ivas_fx( strt_cpy = add( strt_cpy, 1 ); } - sfbEnergyTileR = L_deposit_l( BASOP_Util_Divide3232_Scale( sfbEnergyTileR, width, &tmp_e ) ); - sfbEnergyTileR_e = add( sub( sfbEnergyTileR_e, Q15 ), tmp_e ); + sfbEnergyTileR = L_deposit_h( BASOP_Util_Divide3232_Scale( sfbEnergyTileR, width, &tmp_e ) ); + sfbEnergyTileR_e = add( sub( sfbEnergyTileR_e, Q31 ), tmp_e ); IF( sfbEnergyTileR == 0 ) { diff --git a/lib_enc/ivas_core_enc.c b/lib_enc/ivas_core_enc.c index 3bc869675..0c7fe9094 100644 --- a/lib_enc/ivas_core_enc.c +++ b/lib_enc/ivas_core_enc.c @@ -498,14 +498,14 @@ ivas_error ivas_core_enc_fx( IF( EQ_16( nSubframes, NB_DIV ) ) { Word16 max_e = s_max( orig_spectrum_e[i][0], orig_spectrum_e[i][1] ); - Scale_sig32( hMCT->p_orig_spectrum_long_fx[cpe_id][i], N_TCX10_MAX, sub( orig_spectrum_e[i][0], max_e ) ); // exp(max_e) - Scale_sig32( hMCT->p_orig_spectrum_long_fx[cpe_id][i] + N_TCX10_MAX, N_TCX10_MAX, sub( orig_spectrum_e[i][1], max_e ) ); // exp(max_e) - hMCT->q_orig_spectrum_long_fx[cpe_id][i] = max_e; + scale_sig32( hMCT->p_orig_spectrum_long_fx[cpe_id][i], N_TCX10_MAX, sub( orig_spectrum_e[i][0], max_e ) ); // exp(max_e) + scale_sig32( hMCT->p_orig_spectrum_long_fx[cpe_id][i] + N_TCX10_MAX, N_TCX10_MAX, sub( orig_spectrum_e[i][1], max_e ) ); // exp(max_e) + hMCT->q_orig_spectrum_long_fx[cpe_id][i] = sub( Q31, max_e ); move16(); } ELSE { - hMCT->q_orig_spectrum_long_fx[cpe_id][i] = orig_spectrum_e[i][0]; + hMCT->q_orig_spectrum_long_fx[cpe_id][i] = sub( Q31, orig_spectrum_e[i][0] ); move16(); } diff --git a/lib_enc/ivas_tcx_core_enc.c b/lib_enc/ivas_tcx_core_enc.c index 0680fc795..f072e5436 100644 --- a/lib_enc/ivas_tcx_core_enc.c +++ b/lib_enc/ivas_tcx_core_enc.c @@ -466,7 +466,7 @@ void stereo_tcx_core_enc( E_LPC_lsp_lsf_conversion( lsp_fx, lsf_fx, M ); /* Quantize */ - Q_lsf_tcxlpc_fx( lsf_fx, lsf_tcx_q_fx, lspq_ind, param_lpc, M, st->narrowBand, tcx_lpc_cdk, st->mem_MA_fx, st->hTcxCfg->coder_type, st->Bin_E_fx, Q_ener ); + Q_lsf_tcxlpc_ivas_fx( lsf_fx, lsf_tcx_q_fx, lspq_ind, param_lpc, M, st->narrowBand, tcx_lpc_cdk, st->mem_MA_fx, st->hTcxCfg->coder_type, st->Bin_E_fx, st->q_Bin_E ); /* Account for consumed bits */ nbits_lpc[0] = TCXLPC_NUMBITS; diff --git a/lib_enc/lsf_msvq_ma_enc_fx.c b/lib_enc/lsf_msvq_ma_enc_fx.c index d30f2c249..b81cc18e8 100644 --- a/lib_enc/lsf_msvq_ma_enc_fx.c +++ b/lib_enc/lsf_msvq_ma_enc_fx.c @@ -652,6 +652,165 @@ Word16 Q_lsf_tcxlpc_fx( return NumIndices; } + +Word16 Q_lsf_tcxlpc_ivas_fx( + /* const */ Word16 lsf[], /* i : original lsf */ + Word16 lsf_q[], /* o : quantized lsf */ + Word16 lsp_q_ind[], /* o : quantized lsp (w/o MA prediction) */ + Word16 indices[], /* o : VQ indices */ + const Word16 lpcorder, /* i : LPC order */ + const Word16 narrowband, /* i : narrowband flag */ + const Word16 cdk, /* i : codebook selector */ + const Word16 mem_MA[], /* i : MA memory */ + const Word16 coder_type, + const Word32 *Bin_Ener, + const Word16 Q_ener ) +{ + Word16 weights[M + 1]; + Word16 pred[M16k]; + Word16 i; + Word16 NumIndices; + Word16 lsf_q_ind[M16k]; + const Word16 *means; + Word16 lsf_rem[M]; + Word16 lsf_rem_q_ind[M]; + + Unified_weighting_fx( &Bin_Ener[L_FFT / 2], Q_ener, lsf, weights, narrowband, (Word16) EQ_16( coder_type, UNVOICED ), 12800, M ); + + move16(); + NumIndices = 0; + + /* Put disabled flag */ + indices[NumIndices] = 0; + move16(); + NumIndices = add( NumIndices, 1 ); + + /* Inter-frame prediction */ + + means = lsf_means[narrowband]; /* 14Q1 * 1.28 */ + + FOR( i = 0; i < lpcorder; ++i ) + { + pred[i] = add( means[i], mult_r( MU_MA_FX, mem_MA[i] ) ); /* 14Q1 * 1.28 + ( 14Q1 * 1.28 * Q15 ) = 14Q1 * 1.28*/ + move16(); + } + + /* Subtract prediction */ + + FOR( i = 0; i < lpcorder; ++i ) + { + lsf[i] = sub( lsf[i], pred[i] ); /* 14Q1 * 1.28 */ + move16(); + } + + + msvq_enc_fx( + lsf_codebook[narrowband][cdk], + lsf_dims, + lsf_offs, + lsf, + lsf_numlevels, + kMaxC, + TCXLPC_NUMSTAGES, + weights, + lpcorder, + lpcorder, + indices + NumIndices ); + msvq_dec( + lsf_codebook[narrowband][cdk], + lsf_dims, + lsf_offs, + TCXLPC_NUMSTAGES, + lpcorder, + lpcorder, + indices + NumIndices, +#ifdef IVAS_MSVQ + 0, NULL, +#endif + lsf_q ); + NumIndices = add( NumIndices, TCXLPC_NUMSTAGES ); + + FOR( i = 0; i < lpcorder; ++i ) + { + lsf_q_ind[i] = lsf_q[i]; + move16(); + } + + /* Update flag */ + indices[0] = lsf_ind_is_active( lsf_q_ind, lsf_means[narrowband], narrowband, cdk ); + move16(); + + /* Get residual vector */ + FOR( i = 0; i < lpcorder; ++i ) + { + lsf_rem[i] = add( sub( pred[i], lsf_means[narrowband][i] ), sub( lsf[i], lsf_q_ind[i] ) ); + move16(); + } + + /* Quantize using extra stage(s) */ + msvq_enc_fx( + lsf_ind_codebook[narrowband][cdk], + lsf_ind_dims, + lsf_ind_offs, + lsf_rem, + lsf_ind_numlevels, + kMaxC, + TCXLPC_IND_NUMSTAGES, + weights, + lpcorder, + lpcorder, + indices + NumIndices ); + /* Only add contribution if flag is enabled */ + IF( indices[0] ) + { + /* Decode */ + msvq_dec( + lsf_ind_codebook[narrowband][cdk], + lsf_ind_dims, + lsf_ind_offs, + TCXLPC_IND_NUMSTAGES, + lpcorder, + lpcorder, + indices + NumIndices, +#ifdef IVAS_MSVQ + 0, NULL, +#endif + lsf_rem_q_ind ); + NumIndices = add( NumIndices, TCXLPC_IND_NUMSTAGES ); + + /* Add to MA-removed vector */ + FOR( i = 0; i < lpcorder; ++i ) + { + lsf_q_ind[i] = add( lsf_q_ind[i], lsf_rem_q_ind[i] ); + move16(); + } + } + + /* Add inter-frame prediction */ + FOR( i = 0; i < lpcorder; ++i ) + { + lsf_q[i] = add( lsf_q[i], pred[i] ); + lsf[i] = add( lsf[i], pred[i] ); + move16(); + move16(); + } + + reorder_lsf_fx( lsf_q, TCXLPC_LSF_GAP, lpcorder, INT_FS_FX ); + + FOR( i = 0; i < lpcorder; ++i ) + { + lsf_q_ind[i] = add( lsf_q_ind[i], lsf_means[narrowband][i] ); + move16(); + } + reorder_lsf_fx( lsf_q_ind, TCXLPC_LSF_GAP, lpcorder, INT_FS_FX ); + + IF( lsp_q_ind ) + { + E_LPC_lsf_lsp_conversion /*lsf2lsp*/ ( lsf_q_ind, lsp_q_ind, lpcorder ); + } + + return NumIndices; +} /*--------------------------------------------------------------------------* * enc_lsf_tcxlpc_fx() * diff --git a/lib_enc/prot_fx_enc.h b/lib_enc/prot_fx_enc.h index bc08aae24..280a76c94 100644 --- a/lib_enc/prot_fx_enc.h +++ b/lib_enc/prot_fx_enc.h @@ -2652,6 +2652,19 @@ Word16 Q_lsf_tcxlpc_fx( const Word32 *Bin_Ener, const Word16 Q_ener ); +Word16 Q_lsf_tcxlpc_ivas_fx( + /* const */ Word16 lsf[], /* i : original lsf */ + Word16 lsf_q[], /* o : quantized lsf */ + Word16 lsp_q_ind[], /* o : quantized lsp (w/o MA prediction) */ + Word16 indices[], /* o : VQ indices */ + const Word16 lpcorder, /* i : LPC order */ + const Word16 narrowband, /* i : narrowband flag */ + const Word16 cdk, /* i : codebook selector */ + const Word16 mem_MA[], /* i : MA memory */ + const Word16 coder_type, + const Word32 *Bin_Ener, + const Word16 Q_ener ); + Word16 signalling_mode1_tcx20_enc_fx( Encoder_State *st, /* i : encoder state structure */ Word16 push /*Q0*/ ); diff --git a/lib_enc/tcx_utils_enc_fx.c b/lib_enc/tcx_utils_enc_fx.c index 4d8f26031..35a72b07f 100644 --- a/lib_enc/tcx_utils_enc_fx.c +++ b/lib_enc/tcx_utils_enc_fx.c @@ -3825,7 +3825,7 @@ void ProcessIGF_ivas_fx( q_A = 0; move16(); - ITF_Detect_fx( hIGFEnc->spec_be_igf, hIGFEnc->infoStartLine, hIGFEnc->infoStopLine, 8 /*maxOrder*/, A, &q_A, &predictionGain, &curr_order, sub( 31, hIGFEnc->spec_be_igf_e ) ); + ITF_Detect_ivas_fx( hIGFEnc->spec_be_igf, hIGFEnc->infoStartLine, hIGFEnc->infoStopLine, 8 /*maxOrder*/, A, &q_A, &predictionGain, &curr_order, sub( 31, hIGFEnc->spec_be_igf_e ) ); test(); IF( LT_32( L_deposit_l( hIGFEnc->tns_predictionGain ), 9646899 /* 1.15 in Q23 */ ) && LT_32( L_deposit_l( predictionGain ), 9646899 /* 1.15 in Q23 */ ) ) -- GitLab