ApplyFdCng_fx integration in acelp_core_dec and stereo_mdct_core_dec (b0b8b867) · Commits · SA4 / Audio / IVAS BASOP

lib_com/fd_cng_com_fx.c

+1 −1

Original line number	Diff line number	Diff line
		@@ -1193,7 +1193,7 @@ void minimum_statistics_fx(
		{
		/* calculate scalar with normalized msPeriodogSum[cnt], exponent -2s1 /
		s1 = norm_l( msPeriodogSum[cnt] );
		msPeriodogSum16 = round_fx( L_shl( msPeriodogSum[cnt], s1 ) );
		msPeriodogSum16 = round_fx_sat( L_shl( msPeriodogSum[cnt], s1 ) );
		scalar = L_mult( msPeriodogSum16, msPeriodogSum16 );

		/* calculate difference, both elements in 16Q15 format, use absolute value

lib_dec/acelp_core_dec.c

+381 −1

File changed.

Preview size limit exceeded, changes collapsed.

lib_dec/fd_cng_dec_fx.c

+77 −13

Original line number	Diff line number	Diff line
		@@ -139,6 +139,7 @@ void initFdCngDec(
		hFdCngDec->msPeriodog_ST_exp = 0;
		hFdCngDec->hFdCngCom->fftBuffer_exp = 0;
		hFdCngDec->hFdCngCom->periodog_exp = 0;
		set32_fx(hFdCngDec->smoothed_psd_fx, 0, L_FRAME16k);
		#ifdef IVAS_CODE_CNG
		set_f(hFdCngDec->hFdCngCom->sidNoiseEstLp, 0.0f, NPART);

		@@ -1076,9 +1077,8 @@ Word16 ApplyFdCng_fx(
		Flag Overflow = 0;
		Flag Carry = 0;
		#endif

		Word64 W_tmp;
		Word16 L_frame, last_L_frame;
		Word16 factor;
		Word32 *sidNoiseEst;

		hFdCngDec = st->hFdCngDec;
		@@ -1291,7 +1291,7 @@ Word16 ApplyFdCng_fx(
		L_tmp = norm_llQ31( L_c, L_tmp, &L_tmp_exp );
		L_tmp_exp = sub( add( L_tmp_exp, *cngNoiseLevel_exp ), 1 );

		// L_tmp = Mpy_32_16_1(L_tmp, hFdCngCom->fftlen); /Q16/
		L_tmp = Mpy_32_16_1( L_tmp, 1 );

		L_tmp = Mpy_32_16_1( L_tmp, shr( T_DIV_L_Frame[L_shl( L_mac( -28000, NORM_MDCT_FACTOR, 95 ), 1 - 15 )], 1 ) ); /Q16,exp -7/
		L_tmp_exp = add( L_tmp_exp, -7 ); /->Q16, L_tmp_exp /
		@@ -1302,6 +1302,7 @@ Word16 ApplyFdCng_fx(
		#else
		st->hTcxDec->CngLevelBackgroundTrace_bfi_fx = ( Sqrt32( L_tmp, &L_tmp_exp ) );
		#endif
		L_tmp_exp = add( L_tmp_exp, 1 );
		st->hTcxDec->CngLevelBackgroundTrace_bfi_exp = L_tmp_exp;
		move16();
		}
		@@ -1409,24 +1410,30 @@ Word16 ApplyFdCng_fx(
		}
		test();
		test();
		L_tmp = 0;
		W_tmp = 0;
		FOR( j = hFdCngCom->startBand; j < hFdCngCom->stopFFTbin; j++ )
		{
		L_tmp = L_add( L_tmp, L_shr( cngNoiseLevel[j], 16 ) );
		W_tmp = W_add( W_tmp, L_shr( cngNoiseLevel[j], sub( 31, *cngNoiseLevel_exp ) ) );
		}
		L_tmp_exp = add( *cngNoiseLevel_exp, 16 );
		L_tmp = W_extract_h( W_shl( W_tmp, 32 ) );
		L_tmp_exp = 31;
		#ifdef BASOP_NOGLOB
		IF( EQ_16( concealWholeFrame, 1 ) && EQ_16( st->nbLostCmpt, 1 ) && ( L_shl_o( L_tmp, L_tmp_exp, &Overflow ) > 21474836 /0.01f Q31/ ) )
		#else
		IF( EQ_16( concealWholeFrame, 1 ) && EQ_16( st->nbLostCmpt, 1 ) && ( L_shl( L_tmp, L_tmp_exp ) > 21474836 /0.01f Q31/ ) )
		#endif
		{
		/* always set psychParameters for MDCT-Stereo ... */
		IF( EQ_16( st->element_mode, IVAS_CPE_MDCT ) && st->hTonalMDCTConc != NULL )
		{
		st->hTonalMDCTConc->psychParams = EQ_16( st->core, TCX_20_CORE ) ? &st->hTonalMDCTConc->psychParamsTCX20 : &st->hTonalMDCTConc->psychParamsTCX10;
		}
		/* update isf cng estimate for concealment. Do that during concealment, in order to avoid addition clean channel complexity*/
		IF( EQ_16( st->element_mode, IVAS_CPE_MDCT ) && NE_16( st->core, ACELP_CORE ) )
		{
		TonalMdctConceal_whiten_noise_shape_ivas_fx( st, L_frame, ON_FIRST_LOST_FRAME );
		}
		ELSE IF( NE_16( st->element_mode, IVAS_CPE_MDCT ) )
		ELSE IF( NE_16( st->element_mode, IVAS_CPE_MDCT ) \|\| EQ_16(st->core, ACELP_CORE))
		{
		lpc_from_spectrum( hFdCngCom, hFdCngCom->startBand, hFdCngCom->stopFFTbin, 0 );
		E_LPC_a_lsp_conversion( hFdCngCom->A_cng, lsp_cng, st->lspold_cng, M );
		@@ -1643,17 +1650,73 @@ Word16 ApplyFdCng_fx(
		{
		IF( !( hFdCngCom->msFrCnt_init_counter < hFdCngCom->msFrCnt_init_thresh ) )
		{
		Word16 scale;
		sidNoiseEst = hFdCngCom->sidNoiseEstLp;

		s2 = negate( sub( WORD32_BITS, 1 ) );
		move16();
		/* Shape the SID noise levels in each FFT bin */
		j = 0;
		move16();
		FOR( k = 0; k < hFdCngCom->nFFTpart; k++ )
		{
		factor = BASOP_Util_Divide3232_Scale( ( sidNoiseEst[k] + 1 ), ( hFdCngDec->partNoiseShape[k] + 1 ), &scale );
		assert( hFdCngDec->partNoiseShape[k] >= 0 );

		/* add DELTA as it is done in FLC version, in order to avoid num > denom */
		facTab[k] = 0;
		move16();
		IF( hFdCngDec->partNoiseShape[k] != 0 )
		{
		s1 = norm_l( hFdCngCom->sidNoiseEst[k] );
		L_tmp = L_shl( hFdCngCom->sidNoiseEst[k], s1 );
		L_tmp_exp = sub( hFdCngCom->sidNoiseEstExp, s1 );
		L_tmp = BASOP_Util_Add_Mant32Exp( hFdCngCom->sidNoiseEst[k], hFdCngCom->sidNoiseEstExp, DELTA_MANTISSA_W32, DELTA_EXPONENT, &L_tmp_exp );
		L_tmp = L_shr( L_tmp, 1 );
		s = add( L_tmp_exp, 1 );
		num = extract_h( L_tmp );

		s1 = norm_l( hFdCngDec->partNoiseShape[k] );
		L_tmp = L_shl( hFdCngDec->partNoiseShape[k], s1 );
		L_tmp_exp = sub( hFdCngDec->partNoiseShape_exp, s1 );
		L_tmp = BASOP_Util_Add_Mant32Exp( L_tmp, L_tmp_exp, DELTA_MANTISSA_W32, DELTA_EXPONENT, &L_tmp_exp );
		s = sub( s, L_tmp_exp );
		denom = extract_h( L_tmp );

		facTab[k] = div_s( num, denom );
		move16();
		facTabExp[k] = s;
		move16();
		}
		/* Set unique exponent, IF mantissa is equal to zero */
		IF( EQ_16( facTab[k], 0 ) )
		{
		facTabExp[k] = negate( sub( WORD32_BITS, 1 ) );
		move16();
		}
		s2 = s_max( s2, facTabExp[k] );
		}

		FOR( k = 0; k < hFdCngCom->nFFTpart; k++ )
		{
		s = sub( facTabExp[k], s2 );
		s = s_max( s_min( s, sub( WORD32_BITS, 1 ) ), negate( sub( WORD32_BITS, 1 ) ) );
		FOR( ; j <= hFdCngCom->part[k]; j++ )
		{
		cngNoiseLevel[j] = Mpy_32_16_1( hFdCngDec->bandNoiseShape[j], factor );
		cngNoiseLevel[j] = L_shl( Mpy_32_16_1( hFdCngDec->bandNoiseShape[j], facTab[k] ), s );
		move32();
		}
		}
		/* adapt scaling for rest of the buffer */
		s = sub( *cngNoiseLevel_exp, add( hFdCngDec->bandNoiseShape_exp, s2 ) );
		FOR( ; k < hFdCngCom->npart; k++ )
		{
		FOR( ; j <= hFdCngCom->part[k]; j++ )
		{
		cngNoiseLevel[j] = L_shl( cngNoiseLevel[j], s );
		move32();
		}
		}
		*cngNoiseLevel_exp = add( hFdCngDec->bandNoiseShape_exp, s2 );
		move16();
		}
		}
		IF( EQ_16( st->codec_mode, MODE2 ) )
		@@ -2157,7 +2220,7 @@ void perform_noise_estimation_dec_fx(
		Word16 part, psize_inv, *psize_norm;
		Word32 fftBuffer, periodog, ptr_per, ptr_r, *ptr_i;
		Word32 temp, ftemp, delta, L_tmp;
		Word16 e_temp, wght;
		Word16 e_temp, wght, periodog_exp;
		Word32 enr, enr_tot, enr_tot0;
		Word16 enr_ratio, alpha;
		Word32 *msPeriodog;
		@@ -2602,7 +2665,7 @@ void perform_noise_estimation_dec_fx(
		{
		/* use power spectrum calculated in the MDCT-domain instead of calculating new power spectrum */
		periodog = power_spectrum;
		hFdCngDec->hFdCngCom->periodog_exp = 31;
		periodog_exp = 31;
		}
		ELSE
		{
		@@ -2691,10 +2754,11 @@ void perform_noise_estimation_dec_fx(
		periodog[p] = L_shl( periodog[p], tmp_s );
		}
		}
		periodog_exp = hFdCngDec->hFdCngCom->periodog_exp;
		}

		/* Adjust to the desired frequency resolution by averaging over spectral partitions for SID transmission */
		bandcombinepow( periodog, hFdCngDec->hFdCngCom->periodog_exp, sub( stopFFTbin, startBand ), part, npart, psize_inv, hFdCngDec->msPeriodog, &hFdCngDec->msPeriodog_exp );
		bandcombinepow( periodog, periodog_exp, sub( stopFFTbin, startBand ), part, npart, psize_inv, hFdCngDec->msPeriodog, &hFdCngDec->msPeriodog_exp );


		hFdCngDec->msPeriodog_exp_fft = hFdCngDec->msPeriodog_exp;

lib_dec/ivas_stereo_mdct_core_dec_fx.c

+144 −0

Original line number	Diff line number	Diff line
		@@ -913,7 +913,151 @@ static void run_min_stats_fx(

		if ( will_estimate_noise_on_channel[0] \|\| will_estimate_noise_on_channel[1] \|\| st->bfi )
		{
		#ifdef IVAS_FLOAT_FIXED
		Word32 power_spec_fx[L_FRAME16k];

		if ( !st->bfi )
		{
		for ( int p = 0; p < L_FRAME16k; p++ )
		{
		if ( power_spec[p] <= (float) MAX_32 )
		{
		power_spec_fx[p] = (Word32) ( power_spec[p] /** ( 1u << 0 )*/ );
		}
		else
		{
		power_spec_fx[p] = MAX_32;
		}
		}
		}

		for ( int p = 0; p < st->hFdCngDec->hFdCngCom->fftlen; p++ )
		{
		st->hFdCngDec->hFdCngCom->fftBuffer[p] = (Word32) ( st->hFdCngDec->hFdCngCom->fftBuffer_flt[p] * ( 1u << ( 31 - st->hFdCngDec->hFdCngCom->fftBuffer_exp ) ) );
		}
		st->hFdCngDec->msNoiseEst_exp = 31 - Q4;
		for ( int p = 0; p < NPART_SHAPING; p++ )
		{
		st->hFdCngDec->msNoiseEst[p] = (Word32) ( st->hFdCngDec->msNoiseEst_float[p] * ( 1u << ( 31 - st->hFdCngDec->msNoiseEst_exp ) ) );
		st->hFdCngDec->msPeriodog_ST_fx[p] = (Word32) ( st->hFdCngDec->msPeriodog_ST[p] * ( 1u << ( 31 - st->hFdCngDec->msPeriodog_ST_exp ) ) );
		}
		st->hFdCngDec->hFdCngCom->cngNoiseLevelExp = 31 - Q3; // Q3
		for ( int p = 0; p < FFTCLDFBLEN; p++ )
		{
		st->hFdCngDec->hFdCngCom->cngNoiseLevel[p] = (Word32) ( st->hFdCngDec->hFdCngCom->cngNoiseLevel_flt[p] * ( 1u << ( 31 - st->hFdCngDec->hFdCngCom->cngNoiseLevelExp ) ) );
		}
		for ( int p = 0; p < MSBUFLEN * NPART_SHAPING; p++ )
		{
		if ( ( st->hFdCngDec->msMinBuf_float[p] * ( 1u << Q15 ) ) >= (float) MAX_32 )
		{
		st->hFdCngDec->msMinBuf[p] = MAX_32;
		}
		else
		{
		st->hFdCngDec->msMinBuf[p] = (Word32) ( st->hFdCngDec->msMinBuf_float[p] * ( 1u << ( 31 - 6 ) ) );
		}
		}
		/=================================================/
		ApplyFdCng_fx( NULL, 0, st->bfi ? NULL : power_spec_fx, NULL, NULL, NULL, st, st->bfi, 0 );
		/=================================================/
		if ( !st->bfi )
		{
		for ( int p = 0; p < st->hFdCngDec->hFdCngCom->fftlen; p++ )
		{
		st->hFdCngDec->hFdCngCom->fftBuffer_flt[p] = ( (float) st->hFdCngDec->hFdCngCom->fftBuffer[p] / ( 1u << ( 31 - st->hFdCngDec->hFdCngCom->fftBuffer_exp ) ) );
		}
		for ( int p = 0; p < st->hFdCngDec->npart_shaping; p++ )
		{
		st->hFdCngDec->msNoiseEst_float[p] = ( (float) st->hFdCngDec->msNoiseEst[p] / ( 1u << ( 31 - st->hFdCngDec->msNoiseEst_exp ) ) );
		}

		st->hTcxDec->CngLevelBackgroundTrace_bfi = (float) ( st->hTcxDec->CngLevelBackgroundTrace_bfi_fx / powf( 2, ( 31 - st->hTcxDec->CngLevelBackgroundTrace_bfi_exp ) ) );
		st->cngTDLevel_float = fix16_to_float( st->cngTDLevel, ( 15 - st->cngTDLevel_e ) );

		for ( int p = 0; p < ( st->hFdCngDec->hFdCngCom->stopFFTbin - st->hFdCngDec->hFdCngCom->startBand ); p++ )
		{
		st->hFdCngDec->bandNoiseShape_float[p] = ( (float) st->hFdCngDec->bandNoiseShape[p] / ( 1u << ( 31 - st->hFdCngDec->bandNoiseShape_exp ) ) );
		}
		for ( int p = 0; p < MSBUFLEN * NPART_SHAPING; p++ )
		{
		if ( ( st->hFdCngDec->msMinBuf[p] == MAX_32 ) )
		{
		st->hFdCngDec->msMinBuf_float[p] = FLT_MAX;
		}
		else
		{
		st->hFdCngDec->msMinBuf_float[p] = (float) st->hFdCngDec->msMinBuf[p] / ( 1u << ( 31 - 6 ) ); // CNG_S = 6
		}
		}
		for ( int p = 0; p < st->hFdCngDec->npart_shaping; p++ )
		{
		st->hFdCngDec->msPsd_float[p] = ( (float) st->hFdCngDec->msPsd[p] / ( 1u << Q9 ) );
		st->hFdCngDec->msPeriodog_float[p] = ( (float) st->hFdCngDec->msPeriodog[p] / ( 1u << ( 31 - st->hFdCngDec->msPeriodog_exp ) ) );
		st->hFdCngDec->msPeriodogBuf_float[p] = ( (float) st->hFdCngDec->msPeriodogBuf[p] / ( 1u << Q9 ) );
		st->hFdCngDec->msPeriodog_ST[p] = ( (float) st->hFdCngDec->msPeriodog_ST_fx[p] / ( 1u << ( 31 - st->hFdCngDec->msPeriodog_ST_exp ) ) );
		st->hFdCngDec->msPsdFirstMoment_float[p] = ( (float) st->hFdCngDec->msPsdFirstMoment[p] / ( 1u << Q9 ) );
		st->hFdCngDec->msPsdSecondMoment_float[p] = ( (float) st->hFdCngDec->msPsdSecondMoment[p] / ( 1u << Q19 ) );
		st->hFdCngDec->msNoiseFloor_float[p] = ( (float) st->hFdCngDec->msNoiseFloor[p] / ( 1u << Q9 ) );
		st->hFdCngDec->msAlpha_float[p] = ( (float) st->hFdCngDec->msAlpha[p] / ( 1u << Q31 ) );
		st->hFdCngDec->msBminWin_float[p] = ( (float) st->hFdCngDec->msBminWin[p] / ( 1u << Q27 ) );
		st->hFdCngDec->msBminSubWin_float[p] = ( (float) st->hFdCngDec->msBminSubWin[p] / ( 1u << Q27 ) );
		if ( st->hFdCngDec->msCurrentMin[p] == MAX_32 )
		{
		st->hFdCngDec->msCurrentMin_float[p] = FLT_MAX;
		}
		else
		{
		st->hFdCngDec->msCurrentMin_float[p] = ( (float) st->hFdCngDec->msCurrentMin[p] / ( 1u << ( 31 - ( 6 + 1 + 4 ) ) ) ); // exp : CNG_S + 1 + 4
		}
		if ( st->hFdCngDec->msCurrentMinOut[p] == MAX_32 )
		{
		st->hFdCngDec->msCurrentMinOut_float[p] = FLT_MAX;
		}
		else
		{
		st->hFdCngDec->msCurrentMinOut_float[p] = ( (float) st->hFdCngDec->msCurrentMinOut[p] / ( 1u << ( 31 - 6 ) ) );
		}
		if ( st->hFdCngDec->msCurrentMinSubWindow[p] == MAX_32 )
		{
		st->hFdCngDec->msCurrentMinSubWindow_float[p] = FLT_MAX;
		}
		else
		{
		st->hFdCngDec->msCurrentMinSubWindow_float[p] = ( (float) st->hFdCngDec->msCurrentMinSubWindow[p] / ( 1u << ( 31 - 6 ) ) );
		}
		}
		for ( int p = 0; p < ( st->hFdCngDec->hFdCngCom->stopFFTbin - st->hFdCngDec->hFdCngCom->startBand ); p++ )
		{
		st->hFdCngDec->hFdCngCom->cngNoiseLevel_flt[p] = ( (float) st->hFdCngDec->hFdCngCom->cngNoiseLevel[p] / ( 1u << ( 31 - st->hFdCngDec->hFdCngCom->cngNoiseLevelExp ) ) );
		}
		}
		else if ( ( ( st->bfi == 1 ) && ( st->nbLostCmpt == 1 ) ) && ( sum_f( st->hFdCngDec->hFdCngCom->cngNoiseLevel_flt + st->hFdCngDec->hFdCngCom->startBand, st->hFdCngDec->hFdCngCom->stopFFTbin - st->hFdCngDec->hFdCngCom->startBand ) > 0.01f ) )
		{
		if ( st->element_mode == IVAS_CPE_MDCT && st->core != ACELP_CORE )
		{
		for ( int p = 0; p < ( st->hFdCngDec->hFdCngCom->stopFFTbin - st->hFdCngDec->hFdCngCom->startBand ); p++ )
		{
		st->hFdCngDec->hFdCngCom->cngNoiseLevel_flt[p] = ( (float) st->hFdCngDec->hFdCngCom->cngNoiseLevel[p] / ( 1u << ( 31 - st->hFdCngDec->hFdCngCom->cngNoiseLevelExp ) ) );
		}
		for ( int p = 0; p < FDNS_NPTS; p++ )
		{
		st->hTonalMDCTConc->scaleFactorsBackground_flt[p] = (float) st->hTonalMDCTConc->scaleFactorsBackground_fx[p] / ONE_IN_Q16;
		}
		}
		if ( st->element_mode != IVAS_CPE_MDCT \|\| st->core == ACELP_CORE )
		{
		int A_cng_q = 15 - norm_s( st->hFdCngDec->hFdCngCom->A_cng[0] );
		for ( int p = 0; p < M; p++ )
		{
		st->hFdCngDec->hFdCngCom->A_cng_flt[p] = ( (float) st->hFdCngDec->hFdCngCom->A_cng[p] / ( 1u << A_cng_q ) );
		st->lspold_cng_float[p] = ( (float) st->lspold_cng[p] / ( 1u << Q15 ) );
		st->lsf_cng_float[p] = ( (float) st->lsf_cng[p] / 2.56f ); // Q2.56
		}
		}
		}
		#else
		ApplyFdCng_flt( NULL, st->bfi ? NULL : power_spec, NULL, NULL, st, st->bfi, 0 );
		#endif // IVAS_FLOAT_FIXED
		}

		/* restore VAD (see above) */

lib_dec/ivas_tcx_core_dec.c

+6 −3

Original line number	Diff line number	Diff line
		@@ -888,16 +888,19 @@ void stereo_tcx_core_dec(
		}
		}

		st->hFdCngDec->hFdCngCom->cngNoiseLevelExp = 24; // Q7
		st->hFdCngDec->hFdCngCom->cngNoiseLevelExp = 31 - Q4; // Q4
		for ( int p = 0; p < FFTCLDFBLEN; p++ )
		{
		st->hFdCngDec->hFdCngCom->cngNoiseLevel[p] = (Word32) ( st->hFdCngDec->hFdCngCom->cngNoiseLevel_flt[p] * ( 1u << ( 31 - st->hFdCngDec->hFdCngCom->cngNoiseLevelExp ) ) );
		}
		st->hFdCngDec->hFdCngCom->sidNoiseEstExp = 31 - Q4;
		st->hFdCngDec->partNoiseShape_exp = 31 - Q4;
		for ( int p = 0; p < 24; p++ )
		for ( int p = 0; p < NPART; p++ )
		{
		st->hFdCngDec->hFdCngCom->sidNoiseEst[p] = (Word32) ( st->hFdCngDec->hFdCngCom->sidNoiseEst_flt[p] * ( 1u << ( 31 - st->hFdCngDec->hFdCngCom->sidNoiseEstExp ) ) );
		}
		for ( int p = 0; p < 24; p++ )
		{
		st->hFdCngDec->partNoiseShape[p] = (Word32) ( st->hFdCngDec->partNoiseShape_float[p] * ( 1u << ( 31 - st->hFdCngDec->partNoiseShape_exp ) ) );
		}
		/=================================/
		@@ -1104,7 +1107,7 @@ void stereo_tcx_core_dec(
		}
		}

		st->hFdCngDec->hFdCngCom->cngNoiseLevelExp = 24; // Q7
		st->hFdCngDec->hFdCngCom->cngNoiseLevelExp = 31 - Q4; // Q4
		for ( int p = 0; p < FFTCLDFBLEN; p++ )
		{
		st->hFdCngDec->hFdCngCom->cngNoiseLevel[p] = (Word32) ( st->hFdCngDec->hFdCngCom->cngNoiseLevel_flt[p] * ( 1u << ( 31 - st->hFdCngDec->hFdCngCom->cngNoiseLevelExp ) ) );