Merge branch 'MLD_improvements_7' into 'main' (ce1ed6e6) · Commits · SA4 / Audio / IVAS BASOP

lib_com/enhancer_fx.c

+240 −1

Original line number	Diff line number	Diff line
		@@ -520,7 +520,7 @@ void enhancer_ivas_fx(
		{
		/* tmp = 0.150 * (1.0 + voice_fac) */
		/* 0.30=voiced, 0=unvoiced */
		tmp = mac_r( 0x10000000L, voice_fac, 4915 ); /Q15 /
		tmp = mac_r( 0x13333333L, voice_fac, 4915 ); /Q15 /
		}
		ELSE
		{
		@@ -581,6 +581,245 @@ void enhancer_ivas_fx(
		}
		}

		void enhancer_ivas_fx2(
		const Word32 core_brate, /* i : decoder bitrate */
		const Word16 Opt_AMR_WB, /* i : flag indicating AMR-WB IO mode */
		const Word16 coder_type, /* i : coder type */
		const Word16 i_subfr, /* i : subframe number */
		const Word16 L_frame, /* i : frame size */
		const Word16 voice_fac, /* i : subframe voicing estimation Q15 */
		const Word16 stab_fac, /* i : LP filter stablility measure Q15 */
		Word32 norm_gain_code, /* i : normalised innovative cb. gain Q16 */
		const Word16 gain_inov, /* i : gain of the unscaled innovation Q12 */
		Word32 gc_threshold, / i/o: gain code threshold Q16 */
		Word16 code, / i/o: innovation Q12 */
		Word16 exc2, / i/o: adapt. excitation/total exc. Q_exc*/
		const Word16 gain_pit, /* i : quantized pitch gain Q14 */
		struct dispMem_fx dm_fx, / i/o: phase dispersion algorithm memory */
		const Word16 Q_exc /* i : Q of the excitation */
		)
		{
		Word16 tmp, fac, *pt_exc2;
		Word16 i;
		Word32 L_tmp;
		Word16 gain_code_hi;
		Word16 pit_sharp, tmp16;
		Word16 excp[L_SUBFR], sc;


		pit_sharp = gain_pit;
		move16(); /* to remove gcc warning */
		pt_exc2 = exc2 + i_subfr;

		/------------------------------------------------------------
		* Phase dispersion to enhance noise at low bit rate
		------------------------------------------------------------/

		i = 2;
		move16(); /* no dispersion */
		IF( Opt_AMR_WB )
		{
		IF( LE_32( core_brate, ACELP_6k60 ) )
		{
		i = 0;
		move16(); /* high dispersion */
		}
		ELSE if ( LE_32( core_brate, ACELP_8k85 ) )
		{
		i = 1;
		move16(); /* low dispersion */
		}
		}
		ELSE IF( NE_16( coder_type, UNVOICED ) )

		{
		test();
		test();
		test();
		test();
		IF( LE_32( core_brate, ACELP_7k20 ) )
		{
		i = 0;
		move16(); /* high dispersion */
		}
		ELSE if ( ( EQ_16( coder_type, GENERIC ) \|\| EQ_16( coder_type, TRANSITION ) \|\| EQ_16( coder_type, AUDIO ) \|\| coder_type == INACTIVE ) && LE_32( core_brate, ACELP_9k60 ) )
		{
		i = 1;
		move16(); /* low dispersion */
		}
		}

		phase_dispersion_fx( norm_gain_code, gain_pit, code, i, dm_fx );

		/*------------------------------------------------------------
		* noise enhancer
		*
		* - Enhance excitation on noise. (modify gain of code)
		* If signal is noisy and LPC filter is stable, move gain
		* of code 1.5 dB toward gain of code threshold.
		* This decreases by 3 dB noise energy variation.
		-----------------------------------------------------------/

		/* tmp = 0.5f * (1.0f - voice_fac) */
		#ifdef BASOP_NOGLOB
		tmp = msu_r_sat( 0x40000000, voice_fac, 16384 ); /Q15 / /* 1=unvoiced, 0=voiced */
		#else
		tmp = msu_r( 0x40000000, voice_fac, 16384 ); /Q15 / /* 1=unvoiced, 0=voiced */
		#endif
		/* fac = stab_fac * tmp */
		fac = mult( stab_fac, tmp ); /Q15/

		IF( LT_32( norm_gain_code, *gc_threshold ) )
		{
		L_tmp = Madd_32_16( norm_gain_code, norm_gain_code, 6226 ); /Q16 /
		L_tmp = L_min( L_tmp, gc_threshold ); /Q16 */
		}
		ELSE
		{
		L_tmp = Mult_32_16( norm_gain_code, 27536 ); /Q16 /
		L_tmp = L_max( L_tmp, gc_threshold ); /Q16 */
		}
		*gc_threshold = L_tmp;
		move32(); /Q16 /

		/* gain_code = (fac * tmp) + (1.0 - fac) * gain_code ==> fac * (tmp - gain_code) + gain_code */
		L_tmp = L_sub( L_tmp, norm_gain_code ); /Q16 /
		norm_gain_code = Madd_32_16( norm_gain_code, L_tmp, fac ); /Q16 /

		/* gain_code = gain_inov - Inverse the normalization /
		L_tmp = Mult_32_16( norm_gain_code, gain_inov ); /Q13/ /* gain_inov in Q12 */

		sc = 6;
		move16();

		gain_code_hi = round_fx( L_shl( L_tmp, add( Q_exc, 3 ) ) ); /* in Q_exc */

		/------------------------------------------------------------
		* pitch enhancer
		*
		* - Enhance excitation on voiced. (HP filtering of code)
		* On voiced signal, filtering of code by a smooth fir HP
		* filter to decrease energy of code at low frequency.
		------------------------------------------------------------/
		test();
		IF( !Opt_AMR_WB && EQ_16( coder_type, UNVOICED ) )
		{
		/* Copy(code, exc2, L_SUBFR) */
		FOR( i = 0; i < L_SUBFR; i++ )
		{
		pt_exc2[i] = round_fx( L_shl( L_mult( gain_code_hi, code[i] ), sc ) ); /Q0 / /* code in Q12 (Q9 for encoder) */
		move16();
		}
		}
		ELSE
		{
		test();
		test();
		IF( Opt_AMR_WB && ( EQ_32( core_brate, ACELP_8k85 ) \|\| EQ_32( core_brate, ACELP_6k60 ) ) )
		{
		#ifdef BASOP_NOGLOB
		pit_sharp = shl_sat( gain_pit, 1 ); /* saturation can occur here Q14 -> Q15 */
		#else
		pit_sharp = shl( gain_pit, 1 ); /* saturation can occur here Q14 -> Q15 */
		#endif
		/* saturation takes care of "if (pit_sharp > 1.0) { pit_sharp=1.0; }" */
		IF( GT_16( pit_sharp, 16384 ) )
		{
		tmp16 = mult( pit_sharp, 8192 );
		FOR( i = 0; i < L_SUBFR; i++ )
		{
		/* excp[i] = pt_exc2[i] * pit_sharp * 0.25 */
		excp[i] = mult_r( pt_exc2[i], tmp16 );
		move16();
		}
		}
		}

		IF( EQ_16( L_frame, L_FRAME16k ) )
		{
		/* tmp = 0.150 * (1.0 + voice_fac) */
		/* 0.30=voiced, 0=unvoiced */
		tmp = mac_r( 0x13333333L, voice_fac, 4915 ); /Q15 /
		}
		ELSE
		{
		/* tmp = 0.125 * (1.0 + voice_fac) */
		/* 0.25=voiced, 0=unvoiced */
		tmp = mac_r( 0x10000000L, voice_fac, 4096 ); /Q15 /
		}

		/*-----------------------------------------------------------------
		* Do a simple noncasual "sharpening": effectively an FIR
		* filter with coefs [-tmp 1.0 -tmp] where tmp=0...0.25.
		* This is applied to code and add_fxed to exc2
		-----------------------------------------------------------------/
		/* pt_exc2[0] += code[0] - tmp * code[1] */
		L_tmp = L_deposit_h( code[0] ); /* if Enc :Q9 * Q15 -> Q25 */
		L_tmp = L_msu( L_tmp, code[1], tmp ); /* Q12 * Q15 -> Q28 */
		#ifdef BASOP_NOGLOB
		L_tmp = L_shl_sat( L_mult( gain_code_hi, extract_h( L_tmp ) ), sc );
		pt_exc2[0] = msu_r_sat( L_tmp, -32768, pt_exc2[0] );
		move16();
		#else
		L_tmp = L_shl( L_mult( gain_code_hi, extract_h( L_tmp ) ), sc );
		pt_exc2[0] = msu_r( L_tmp, -32768, pt_exc2[0] );
		move16();
		#endif
		move16(); /* in Q_exc */

		FOR( i = 1; i < L_SUBFR - 1; i++ )
		{
		/* pt_exc2[i] += code[i] - tmp * code[i-1] - tmp * code[i+1] */
		L_tmp = L_msu( -32768, code[i], -32768 );
		L_tmp = L_msu( L_tmp, code[i + 1], tmp );
		#ifdef BASOP_NOGLOB
		tmp16 = msu_r_sat( L_tmp, code[i - 1], tmp );
		L_tmp = L_shl_sat( L_mult( gain_code_hi, tmp16 ), sc );
		pt_exc2[i] = msu_r_sat( L_tmp, -32768, pt_exc2[i] );
		move16();
		#else
		tmp16 = msu_r( L_tmp, code[i - 1], tmp );
		L_tmp = L_shl( L_mult( gain_code_hi, tmp16 ), sc );
		pt_exc2[i] = msu_r( L_tmp, -32768, pt_exc2[i] );
		#endif
		move16(); /* in Q_exc */
		}

		/* pt_exc2[L_SUBFR-1] += code[L_SUBFR-1] - tmp * code[L_SUBFR-2] */
		L_tmp = L_deposit_h( code[L_SUBFR - 1] ); /Q28 /
		L_tmp = L_msu( L_tmp, code[L_SUBFR - 2], tmp ); /Q28 /
		L_tmp = L_shl( L_mult( gain_code_hi, extract_h( L_tmp ) ), sc );
		#ifdef BASOP_NOGLOB
		pt_exc2[L_SUBFR - 1] = msu_r_sat( L_tmp, -32768, pt_exc2[L_SUBFR - 1] );
		move16();
		#else
		pt_exc2[L_SUBFR - 1] = msu_r( L_tmp, -32768, pt_exc2[L_SUBFR - 1] );
		move16();
		#endif
		move16(); /* in Q_exc */
		test();
		test();
		IF( Opt_AMR_WB && ( EQ_32( core_brate, ACELP_8k85 ) \|\| EQ_32( core_brate, ACELP_6k60 ) ) )
		{
		IF( GT_16( pit_sharp, 16384 ) )
		{
		FOR( i = 0; i < L_SUBFR; i++ )
		{
		/* excp[i] += pt_exc2[i] */
		#ifdef BASOP_NOGLOB
		excp[i] = add_sat( excp[i], pt_exc2[i] );
		#else
		excp[i] = add( excp[i], pt_exc2[i] );
		#endif
		move16();
		}
		agc2_fx( pt_exc2, excp, L_SUBFR );
		Copy( excp, pt_exc2, L_SUBFR );
		}
		}
		}
		}

		/---------------------------------------------------------
		* Enhancement of the excitation signal before synthesis
		---------------------------------------------------------/

lib_com/prot_fx.h

+18 −0

Original line number	Diff line number	Diff line
		@@ -5193,6 +5193,24 @@ void enhancer_ivas_fx(
		const Word16 Q_exc /* i : Q of the excitation */
		);

		void enhancer_ivas_fx2(
		const Word32 core_brate, /* i : decoder bitrate */
		const Word16 Opt_AMR_WB, /* i : flag indicating AMR-WB IO mode */
		const Word16 coder_type, /* i : coder type */
		const Word16 i_subfr, /* i : subframe number */
		const Word16 L_frame, /* i : frame size */
		const Word16 voice_fac, /* i : subframe voicing estimation Q15 */
		const Word16 stab_fac, /* i : LP filter stablility measure Q15 */
		Word32 norm_gain_code, /* i : normalised innovative cb. gain Q16 */
		const Word16 gain_inov, /* i : gain of the unscaled innovation Q12 */
		Word32 gc_threshold, / i/o: gain code threshold Q16 */
		Word16 code, / i/o: innovation Q12 */
		Word16 exc2, / i/o: adapt. excitation/total exc. Q_exc*/
		const Word16 gain_pit, /* i : quantized pitch gain Q14 */
		struct dispMem_fx dm_fx, / i/o: phase dispersion algorithm memory */
		const Word16 Q_exc /* i : Q of the excitation */
		);

		Word16 E_UTIL_enhancer(
		Word16 voice_fac, /* i : subframe voicing estimation Q15 */
		Word16 stab_fac, /* i : LP filter stability measure Q15 */

lib_dec/dec_gen_voic_fx.c

+2 −2

Original line number	Diff line number	Diff line
		@@ -856,7 +856,7 @@ ivas_error decod_gen_voic_ivas_fx(
		}
		ELSE
		{
		enhancer_fx( st_fx->core_brate, 0, st_fx->coder_type, i_subfr_fx, L_frame, voice_fac_fx, st_fx->stab_fac_fx,
		enhancer_ivas_fx2( st_fx->core_brate, 0, st_fx->coder_type, i_subfr_fx, L_frame, voice_fac_fx, st_fx->stab_fac_fx,
		norm_gain_code_fx, gain_inov_fx, &st_fx->gc_threshold_fx, code_fx, exc2_fx, gain_pit_fx, &( st_fx->dm_fx ), st_fx->Q_exc );
		}

lib_dec/ivas_dirac_output_synthesis_cov.c

+8 −8

Original line number	Diff line number	Diff line
		@@ -747,8 +747,8 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_collect_slot_fx(
		const Word16 nchan_in /* i : number of input channels */
		)
		{
		Word16 cx_init_e[PARAM_MC_MAX_PARAMETER_BANDS];
		Word16 cx_init_imag_e[PARAM_MC_MAX_PARAMETER_BANDS];
		Word16 cx_init_e;
		Word16 cx_init_imag_e;
		Word16 band_idx, ch_idx;
		Word16 brange[2];
		Word32 real_in_buffer_fx[PARAM_MC_MAX_BANDS_IN_PARAMETER_BAND * MAX_TRANSPORT_CHANNELS];
		@@ -809,9 +809,9 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_collect_slot_fx(
		v_add_fixed_me( cx_fx, *cx_e, real_buffer_fx, output_e, cx_fx, &tmp1_e, imult1616( nchan_in, nchan_in ), 1 );

		v_add_fixed_me( cx_imag_fx, *cx_imag_e, imag_buffer_fx, output_e, cx_imag_fx, &tmp2_e, imult1616( nchan_in, nchan_in ), 1 );
		cx_init_e[0] = tmp1_e;
		cx_init_e = tmp1_e;
		move16();
		cx_init_imag_e[0] = tmp2_e;
		cx_init_imag_e = tmp2_e;
		move16();

		// normalizing both the matrices to a common exponent for a better precision
		@@ -822,18 +822,18 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_collect_slot_fx(

		FOR( j = 0; j < PARAM_MC_MAX_TRANSPORT_CHANS * PARAM_MC_MAX_TRANSPORT_CHANS; j++ )
		{
		L_tmp = BASOP_Util_Add_Mant32Exp( cx_fx[j], cx_init_e[0], 0, 0, &tmp1_e );
		L_tmp = BASOP_Util_Add_Mant32Exp( cx_imag_fx[j], cx_init_imag_e[0], 0, 0, &tmp2_e );
		L_tmp = BASOP_Util_Add_Mant32Exp( cx_fx[j], cx_init_e, 0, 0, &tmp1_e );
		L_tmp = BASOP_Util_Add_Mant32Exp( cx_imag_fx[j], cx_init_imag_e, 0, 0, &tmp2_e );
		tmp1 = s_max( tmp1, tmp1_e );
		tmp2 = s_max( tmp2, tmp2_e );
		}

		FOR( j = 0; j < PARAM_MC_MAX_TRANSPORT_CHANS * PARAM_MC_MAX_TRANSPORT_CHANS; j++ )
		{
		L_tmp = BASOP_Util_Add_Mant32Exp( cx_fx[j], cx_init_e[0], 0, 0, &tmp1_e );
		L_tmp = BASOP_Util_Add_Mant32Exp( cx_fx[j], cx_init_e, 0, 0, &tmp1_e );
		cx_fx[j] = L_shr( L_tmp, sub( tmp1, tmp1_e ) );
		move32();
		L_tmp = BASOP_Util_Add_Mant32Exp( cx_imag_fx[j], cx_init_imag_e[0], 0, 0, &tmp2_e );
		L_tmp = BASOP_Util_Add_Mant32Exp( cx_imag_fx[j], cx_init_imag_e, 0, 0, &tmp2_e );
		cx_imag_fx[j] = L_shr( L_tmp, sub( tmp2, tmp2_e ) );
		move32();
		}

lib_dec/ivas_mc_param_dec.c

+37 −21

Original line number	Diff line number	Diff line
		@@ -2968,9 +2968,10 @@ void ivas_param_mc_dec_digest_tc_fx(
		Word32 cx_next_band_fx[PARAM_MC_MAX_TRANSPORT_CHANS * PARAM_MC_MAX_TRANSPORT_CHANS];
		Word32 cx_imag_next_band_fx[PARAM_MC_MAX_TRANSPORT_CHANS * PARAM_MC_MAX_TRANSPORT_CHANS];

		Word16 cx_buff_e[PARAM_MC_MAX_TRANSPORT_CHANS * PARAM_MC_MAX_TRANSPORT_CHANS];
		Word16 cx_buff_e[2][PARAM_MC_MAX_TRANSPORT_CHANS * PARAM_MC_MAX_TRANSPORT_CHANS];
		Word16 cx_e;
		Word16 cx_imag_e, tmp_e;
		Word16 cx_imag_next_band_e, cx_next_band_e;
		Word16 qout = 0;
		move16();

		@@ -3055,9 +3056,10 @@ void ivas_param_mc_dec_digest_tc_fx(
		move16();
		cx_imag_e = 0;
		move16();
		// cldfb_slots = DEFAULT_JBM_CLDFB_TIMESLOTS;
		cx_next_band_e = 0;
		move16();
		cx_imag_next_band_e = 0;
		move16();

		/* slot loop for gathering the input data */
		FOR( slot_idx = 0; slot_idx < nCldfbSlots; slot_idx++ )
		{
		@@ -3099,9 +3101,9 @@ void ivas_param_mc_dec_digest_tc_fx(
		&hParamMC->Cldfb_ImagBuffer_tc_fx[L_mult0( imult1616( slot_idx, hParamMC->num_freq_bands ), nchan_transport )],
		/hParamMC->Cldfb_ImagBuffer_tc_e/ Q31 - Q6,
		cx_next_band_fx,
		&cx_e,
		&cx_next_band_e,
		cx_imag_next_band_fx,
		&cx_imag_e,
		&cx_imag_next_band_e,
		hParamMC,
		add( param_band_idx, is_next_band ),
		nchan_transport );
		@@ -3124,6 +3126,7 @@ void ivas_param_mc_dec_digest_tc_fx(
		}
		}

		Word16 tmp_cx_e, tmp_cx_imag_e;
		/* map from complex input covariance to real values */
		FOR( is_next_band = 0; is_next_band < 2; is_next_band++ )
		{
		@@ -3137,11 +3140,15 @@ void ivas_param_mc_dec_digest_tc_fx(
		{
		pCx = &cx_next_band_fx[0];
		pCx_imag = &cx_imag_next_band_fx[0];
		tmp_cx_e = cx_next_band_e;
		tmp_cx_imag_e = cx_imag_next_band_e;
		}
		ELSE
		{
		pCx = &cx_fx[0];
		pCx_imag = &cx_imag_fx[0];
		tmp_cx_e = cx_e;
		tmp_cx_imag_e = cx_imag_e;
		}

		FOR( i = 0; i < imult1616( nchan_transport, nchan_transport ); i++ )
		@@ -3150,51 +3157,60 @@ void ivas_param_mc_dec_digest_tc_fx(
		imag_part_fx = pCx_imag[i];
		move32();
		move32();
		cx_buff_e[i] = cx_e;
		cx_buff_e[is_next_band][i] = tmp_cx_e;
		move16();
		/* (a-ib)(c+id) = ac + bd + i(ad-bc) */
		IF( LT_16( param_band_idx, hParamMC->max_param_band_abs_cov ) )
		{
		L_tmp1 = Mpy_32_32( real_part_fx, real_part_fx );
		L_tmp2 = Mpy_32_32( imag_part_fx, imag_part_fx );
		L_tmp1 = BASOP_Util_Add_Mant32Exp( L_tmp1, add( cx_e, cx_e ), L_tmp2, add( cx_imag_e, cx_imag_e ), &tmp_e );
		L_tmp1 = BASOP_Util_Add_Mant32Exp( L_tmp1, add( tmp_cx_e, tmp_cx_e ), L_tmp2, add( tmp_cx_imag_e, tmp_cx_imag_e ), &tmp_e );
		pCx[i] = Sqrt32( L_tmp1, &tmp_e );
		cx_fx[i] = Sqrt32( L_tmp1, &tmp_e );
		move32();
		cx_buff_e[i] = tmp_e;
		cx_buff_e[is_next_band][i] = tmp_e;
		move16();
		}
		ELSE
		{
		pCx[i] = real_part_fx;
		move32();
		cx_buff_e[i] = cx_e;
		cx_buff_e[is_next_band][i] = tmp_cx_e;
		move16();
		}
		}
		}

		max_e = cx_buff_e[0];
		max_e = cx_buff_e[0][0];
		move16();

		/* Cx for transport channels */
		FOR( is_next_band = 0; is_next_band < 2; is_next_band++ )
		{
		FOR( i = 0; i < imult1616( nchan_transport, nchan_transport ); i++ )
		{

		IF( LT_16( max_e, cx_buff_e[i] ) )
		IF( LT_16( max_e, cx_buff_e[is_next_band][i] ) )
		{
		max_e = cx_buff_e[i];
		max_e = cx_buff_e[is_next_band][i];
		}
		}
		}

		/* Cx for transport channels */
		FOR( is_next_band = 0; is_next_band < 2; is_next_band++ )
		{
		FOR( i = 0; i < imult1616( nchan_transport, nchan_transport ); i++ )
		{

		cx_fx[i] = L_shr( cx_fx[i], sub( max_e, cx_buff_e[i] ) );
		if ( is_next_band )
		{
		cx_fx[i] = L_shr( cx_fx[i], sub( max_e, cx_buff_e[is_next_band][i] ) );
		}
		else
		{
		cx_next_band_fx[i] = L_shr( cx_next_band_fx[i], sub( max_e, cx_buff_e[is_next_band][i] ) );
		}
		move32();
		}

		}
		cx_e = max_e;
		move16();