Merge remote-tracking branch 'refs/remotes/origin/main' into 2003_fix_saturation_at_con_tcx (07fb862e) · Commits · SA4 / Audio / IVAS BASOP

lib_com/options.h

+2 −0

Original line number	Diff line number	Diff line
		@@ -91,6 +91,7 @@
		#define FIX_1990_SANITIZER_IN_REVERB_LOAD /* Nokia: Fix issue part of issue 1990 by introducing missing free of structure. */
		#define FIX_1995_REVERB_INIT /* VA/Nokia: issue 1995: Fix use-of-uninitialized-value in ivas_binaural_reverb_init() */
		#define FIX_1996_MASKING_NOISE /* Dlb: Heavy precision loss in ola buffers causing discontinuity*/
		#define FIX1998_APA_EXEC_SCALING /* FhG: fix scaling of apa_exec_ivas_fx(); avoid continuously worse scaling with previous data */

		#define FIX_1962_FORMAT_CONV_SPECTRAL_DIFF /* FhG: Improved precision of targetEnergy in ivas_ls_setup_conversion_process_mdct_fx() */
		/* #################### Start BASOP porting switches ############################ */
		@@ -105,6 +106,7 @@
		#define NONBE_FIX_1261_MASA_EXT_META_JBM /* Nokia: issue #1261: MASA metadata EXT output delay buffer init in JBM */
		#define NONBE_FIX_1143_MASA_BRSW /* Nok: Fix for issue 1143: MSAN use of uninitialized value in masa decoding to binaural with dtx bitrate switching and 5 % FER */
		#define NONBE_FIX_1220_OMASA_JBM_EXT_USAN /* Nokia: fix issue 1220 OMASA EXT JBM USAN, also fix similar cases of free to avoid future problems */
		#define NONBE_FIX_1376_MDCT_CONCEALMENT /* FhG: fix concealment artifact in MDCT Stereo with DTX, in case transition frame gets lost */

		/* #################### End BASOP porting switches ############################ */

lib_dec/ivas_mdct_core_dec_fx.c

+10 −0

Original line number	Diff line number	Diff line
		@@ -1691,7 +1691,17 @@ void ivas_mdct_core_tns_ns_fx(
		q_x = add( q_x, 1 );
		}

		#ifdef NONBE_FIX_1376_MDCT_CONCEALMENT
		/*
		2025-09-07, mul:
		in case of PLC, applying SNS up to L_spec might not be enough: In case the transition frame from DTX after an inactive period is lost, L_spec is assumed to represent a regular TCX frame,
		however, this frame is nevertheless acting as an transition frame as also visible in L_frameTCX; thus, the safer approach to prevent high frequency artifacts is to apply the SNS up to L_frameTCX;
		in case this is not necessary, x[] is filled with zeros, and the multiplication is not causing any additional harm
		*/
		v_multc_fixed( x_fx[ch][k] + st->hTcxCfg->psychParamsCurrent->nBins, sns_int_scf_fx[FDNS_NPTS - 1], x_fx[ch][k] + st->hTcxCfg->psychParamsCurrent->nBins, sub( s_max( L_spec[ch], L_frameTCX[ch] ), st->hTcxCfg->psychParamsCurrent->nBins ) );
		#else
		v_multc_fixed( x_fx[ch][k] + st->hTcxCfg->psychParamsCurrent->nBins, sns_int_scf_fx[FDNS_NPTS - 1], x_fx[ch][k] + st->hTcxCfg->psychParamsCurrent->nBins, sub( L_spec[ch], st->hTcxCfg->psychParamsCurrent->nBins ) );
		#endif

		q_2 = sub( add( q_x, q_sns_int_scf ), 31 );
		Scale_sig32( &x_fx[ch][k][0], st->hTcxCfg->psychParamsCurrent->nBins, sub( q_2, q_x ) );

lib_dec/jbm_pcmdsp_apa_fx.c

+32 −1

Original line number	Diff line number	Diff line
		@@ -76,7 +76,10 @@ struct apa_state_t
		bool evs_compat_mode;

		Word16 *buf_out_fx;
		Word16 Q_buf_out;
		Word16 Q_buf_out; /* stores the scaling of buf_out_fx */
		#ifdef FIX1998_APA_EXEC_SCALING
		Word16 Q_a_out_init_old; /* stores initially determined max. scaling of data in buf_out_fx, before beeing adjusted to previous frame scaling */
		#endif
		UWord16 buf_out_capacity;
		UWord16 l_buf_out;

		@@ -190,6 +193,10 @@ ivas_error apa_init(
		}
		memset( ps->buf_out_fx, 0, ( sizeof( Word16 ) * ps->buf_out_capacity ) );
		ps->Q_buf_out = Q15;
		#ifdef FIX1998_APA_EXEC_SCALING
		ps->Q_a_out_init_old = Q15;
		move16();
		#endif
		move16();

		ps->evs_compat_mode = false;
		@@ -272,6 +279,10 @@ UWord8 apa_reconfigure(
		ps->buf_out_fx = (Word16 ) malloc( sizeof( Word16 ) ps->buf_out_capacity );
		memset( ps->buf_out_fx, 0, ( sizeof( Word16 ) * ps->buf_out_capacity ) );
		ps->Q_buf_out = Q15;
		#ifdef FIX1998_APA_EXEC_SCALING
		ps->Q_a_out_init_old = Q15;
		move16();
		#endif
		move16();
		IF( !ps->buf_out_fx )
		{
		@@ -897,8 +908,15 @@ UWord8 apa_exec_ivas_fx(
		Word32 expScaling, actScaling;
		UWord32 statsResetThreshold, statsResetShift;
		Word16 Q_a_out;
		#ifdef FIX1998_APA_EXEC_SCALING
		Word16 Q_a_out_init_old;
		#endif

		Q_a_out = add( getScaleFactor32_copy( a_in, L_mult0( ps->num_channels, APA_BUF_PER_CHANNEL ) ), Q11 - Q16 - Q1 );
		#ifdef FIX1998_APA_EXEC_SCALING
		Q_a_out_init_old = Q_a_out; /* store the possible scaling of a_in, to be re-used in the next frame */
		move16();
		#endif
		statsResetThreshold = 1637;
		move32();
		statsResetShift = 2;
		@@ -965,7 +983,15 @@ UWord8 apa_exec_ivas_fx(
		Word16 a_tmp[APA_BUF];
		Word16 *buf_out_ptr = &( ps->buf_out_fx[ps->l_buf_out - ps->l_frm] );

		#ifdef FIX1998_APA_EXEC_SCALING
		/*
		don't compare against actual scaling in ps->Q_buf_out, but possible scaling in ps->Q_a_out_init_old; otherwise we are constantly reducing the scaling over time, leading to precision issues
		alternative approach: determine scaling of ps->buf_out_fx, but this is costly due to the sheer amount of samples stored there...
		*/
		Q_a_out = s_min( Q_a_out, ps->Q_a_out_init_old );
		#else
		Q_a_out = s_min( Q_a_out, ps->Q_buf_out );
		#endif
		Scale_sig( ps->buf_out_fx, ps->buf_out_capacity, sub( Q_a_out, ps->Q_buf_out ) ); // Q_buf_out -> Q_a_out
		IF( EQ_32( ps->scale, 100 ) )
		{
		@@ -1079,6 +1105,10 @@ UWord8 apa_exec_ivas_fx(
		}
		ps->l_buf_out = (UWord16) L_add( ps->l_buf_out, l_frm_out );
		move16();
		#ifdef FIX1998_APA_EXEC_SCALING
		ps->Q_a_out_init_old = Q_a_out_init_old;
		move16();
		#endif

		*l_out = l_frm_out;
		move16();
		@@ -1086,6 +1116,7 @@ UWord8 apa_exec_ivas_fx(
		ps->l_in_total = UL_addNsD( ps->l_in_total, ps->l_frm );
		move32();


		test();
		IF( LT_32( L_abs( ps->diffSinceSetScale ), L_sub( 0x7FFFFF, L_sub( l_frm_out, ps->l_frm ) ) ) &&
		LT_64( ps->nFramesSinceSetScale, statsResetThreshold ) )