From abd4aa4c51a795cf555d675ef976e4e369abef02 Mon Sep 17 00:00:00 2001 From: Sandesh Venkatesh Date: Thu, 13 Mar 2025 09:53:26 +0530 Subject: [PATCH] Fix for 3GPP issue 1371: Energy burst in MDCT-Stereo coded item at 48kbps Link #1371 Restricting the Q-factor of spectrum and mdst_spectrum to Q-factor 9 before enc_ste_pre_mdct to avoid truncation of the spectrums resolved the issue --- lib_enc/ivas_mdct_core_enc_fx.c | 34 ++++++++++++++++++--------------- 1 file changed, 19 insertions(+), 15 deletions(-) diff --git a/lib_enc/ivas_mdct_core_enc_fx.c b/lib_enc/ivas_mdct_core_enc_fx.c index 9027b2669..9f674473e 100644 --- a/lib_enc/ivas_mdct_core_enc_fx.c +++ b/lib_enc/ivas_mdct_core_enc_fx.c @@ -1556,27 +1556,30 @@ void ivas_mdct_core_whitening_enc_fx( } IF( EQ_16( hTcxEnc0->transform_type[n], TCX_5 ) ) { - Word16 length; + Word16 length, len_sbfr; const Word16 tcx5SizeFB = sts[1]->hTcxCfg->tcx5SizeFB; move16(); /* length = max(nSampCore / (2 * NB_DIV), L_subframeTCX / (2 * NB_DIV), NB_DIV = 2 */ length = shr( s_max( nSampCore, L_subframeTCX ), 2 ); + len_sbfr = shr( hTcxEnc0->L_frameTCX, shift ); + assert( hTcxEnc0->L_frameTCX == hTcxEnc1->L_frameTCX ); exp_max = s_max( hTcxEnc0->spectrum_e[n], hTcxEnc1->spectrum_e[n] ); exp_max = s_max( exp_max, mdst_spectrum_e[0][n] ); exp_max = s_max( exp_max, mdst_spectrum_e[1][n] ); // calculate the headroom available - exp_com = s_min( L_norm_arr( hTcxEnc1->spectrum_fx[n], length ), L_norm_arr( hTcxEnc0->spectrum_fx[n], length ) ); - exp_com = s_min( exp_com, L_norm_arr( mdst_spectrum_fx[0][n], length ) ); - exp_com = s_min( exp_com, L_norm_arr( mdst_spectrum_fx[1][n], length ) ); + exp_com = s_min( L_norm_arr( hTcxEnc1->spectrum_fx[n], len_sbfr ), L_norm_arr( hTcxEnc0->spectrum_fx[n], len_sbfr ) ); + exp_com = s_min( exp_com, L_norm_arr( mdst_spectrum_fx[0][n], len_sbfr ) ); + exp_com = s_min( exp_com, L_norm_arr( mdst_spectrum_fx[1][n], len_sbfr ) ); q_com = sub( s_min( Q31, add( sub( Q31, exp_max ), exp_com ) ), 6 ); // 6 guarded bits + q_com = s_max( q_com, 9 ); // Keep the Q-factor at least 9 to avoid precision loss inside enc_ste_pre_mdct. exp_com = sub( Q31, q_com ); - Scale_sig32( hTcxEnc0->spectrum_fx[n], shr( hTcxEnc0->L_frameTCX, shift ) /* hTcxEnc0->L_frameTCX/nSubframes */, sub( hTcxEnc0->spectrum_e[n], exp_com ) ); // hTcxEnc0->spectrum_e - Scale_sig32( hTcxEnc1->spectrum_fx[n], shr( hTcxEnc1->L_frameTCX, shift ) /* hTcxEnc1->L_frameTCX/nSubframes */, sub( hTcxEnc1->spectrum_e[n], exp_com ) ); // hTcxEnc1->spectrum_e - Scale_sig32( mdst_spectrum_fx[0][n], shr( hTcxEnc0->L_frameTCX, shift ) /* hTcxEnc0->L_frameTCX/nSubframes */, sub( mdst_spectrum_e[0][n], exp_com ) ); // mdst_spectrum_e - Scale_sig32( mdst_spectrum_fx[1][n], shr( hTcxEnc1->L_frameTCX, shift ) /* hTcxEnc1->L_frameTCX/nSubframes */, sub( mdst_spectrum_e[1][n], exp_com ) ); // mdst_spectrum_e + scale_sig32( hTcxEnc0->spectrum_fx[n], len_sbfr /* hTcxEnc0->L_frameTCX/nSubframes */, sub( hTcxEnc0->spectrum_e[n], exp_com ) ); // hTcxEnc0->spectrum_e + scale_sig32( hTcxEnc1->spectrum_fx[n], len_sbfr /* hTcxEnc1->L_frameTCX/nSubframes */, sub( hTcxEnc1->spectrum_e[n], exp_com ) ); // hTcxEnc1->spectrum_e + scale_sig32( mdst_spectrum_fx[0][n], len_sbfr /* hTcxEnc0->L_frameTCX/nSubframes */, sub( mdst_spectrum_e[0][n], exp_com ) ); // mdst_spectrum_e + scale_sig32( mdst_spectrum_fx[1][n], len_sbfr /* hTcxEnc1->L_frameTCX/nSubframes */, sub( mdst_spectrum_e[1][n], exp_com ) ); // mdst_spectrum_e Word16 q_com_orig = q_com; move16(); @@ -1600,17 +1603,18 @@ void ivas_mdct_core_whitening_enc_fx( exp_max = sub( Q31, q_com ); move16(); - exp_com = s_min( L_norm_arr( hTcxEnc1->spectrum_fx[n], shr( hTcxEnc1->L_frameTCX, shift ) /* hTcxEnc1->L_frameTCX/nSubframes */ ), L_norm_arr( hTcxEnc0->spectrum_fx[n], shr( hTcxEnc0->L_frameTCX, shift ) /* hTcxEnc0->L_frameTCX/nSubframes */ ) ); - exp_com = s_min( exp_com, L_norm_arr( mdst_spectrum_fx[0][n], shr( hTcxEnc0->L_frameTCX, shift ) /* hTcxEnc0->L_frameTCX/nSubframes */ ) ); - exp_com = s_min( exp_com, L_norm_arr( mdst_spectrum_fx[1][n], shr( hTcxEnc1->L_frameTCX, shift ) /* hTcxEnc1->L_frameTCX/nSubframes */ ) ); + exp_com = s_min( L_norm_arr( hTcxEnc1->spectrum_fx[n], len_sbfr /* hTcxEnc1->L_frameTCX/nSubframes */ ), L_norm_arr( hTcxEnc0->spectrum_fx[n], len_sbfr /* hTcxEnc0->L_frameTCX/nSubframes */ ) ); + exp_com = s_min( exp_com, L_norm_arr( mdst_spectrum_fx[0][n], len_sbfr /* hTcxEnc0->L_frameTCX/nSubframes */ ) ); + exp_com = s_min( exp_com, L_norm_arr( mdst_spectrum_fx[1][n], len_sbfr /* hTcxEnc1->L_frameTCX/nSubframes */ ) ); q_com = sub( s_min( Q31, add( q_com, exp_com ) ), 6 ); + q_com = s_max( q_com, 9 ); // Keep the Q-factor at least 9 to avoid precision loss inside enc_ste_pre_mdct. exp_com = sub( Q31, q_com ); - Scale_sig32( hTcxEnc0->spectrum_fx[n], shr( hTcxEnc0->L_frameTCX, shift ) /* hTcxEnc0->L_frameTCX/nSubframes */, sub( exp_max, exp_com ) ); // exp_max - Scale_sig32( hTcxEnc1->spectrum_fx[n], shr( hTcxEnc1->L_frameTCX, shift ) /* hTcxEnc1->L_frameTCX/nSubframes */, sub( exp_max, exp_com ) ); // exp_max - Scale_sig32( mdst_spectrum_fx[0][n], shr( hTcxEnc0->L_frameTCX, shift ) /* hTcxEnc0->L_frameTCX/nSubframes */, sub( exp_max, exp_com ) ); // exp_max - Scale_sig32( mdst_spectrum_fx[1][n], shr( hTcxEnc1->L_frameTCX, shift ) /* hTcxEnc1->L_frameTCX/nSubframes */, sub( exp_max, exp_com ) ); // exp_max + scale_sig32( hTcxEnc0->spectrum_fx[n], len_sbfr /* hTcxEnc0->L_frameTCX/nSubframes */, sub( exp_max, exp_com ) ); // exp_max + scale_sig32( hTcxEnc1->spectrum_fx[n], len_sbfr /* hTcxEnc1->L_frameTCX/nSubframes */, sub( exp_max, exp_com ) ); // exp_max + scale_sig32( mdst_spectrum_fx[0][n], len_sbfr /* hTcxEnc0->L_frameTCX/nSubframes */, sub( exp_max, exp_com ) ); // exp_max + scale_sig32( mdst_spectrum_fx[1][n], len_sbfr /* hTcxEnc1->L_frameTCX/nSubframes */, sub( exp_max, exp_com ) ); // exp_max q_com_orig = q_com; move16(); -- GitLab