Acelp core encoder cleanup, pre_proc_front cleanup (585e4bb4) · Commits · SA4 / Audio / IVAS BASOP

lib_com/ivas_fb_mixer.c

+2 −0

Original line number	Diff line number	Diff line
		@@ -383,6 +383,8 @@ ivas_error ivas_FB_mixer_open_fx(
		return IVAS_ERROR( IVAS_ERR_FAILED_ALLOC, "Can not allocate memory for FB mixer encoder" );
		}
		set32_fx( hFbMixer->ppFilterbank_prior_input_fx[i], 0, fb_cfg->prior_input_length );
		hFbMixer->q_ppFilterbank_prior_input_fx[i] = Q31;
		move16();
		}

		test();

lib_com/lerp.c

+18 −3

Original line number	Diff line number	Diff line
		@@ -617,6 +617,7 @@ static void L_lerp_proc_fx(
		Word32 pos_fx, shift_fx, diff_fx;
		Word32 buf_fx[2 * L_FRAME_MAX];
		Word16 tmp;
		Word32 ffxsub;

		IF( EQ_16( bufferNewSize, bufferOldSize ) )
		{
		@@ -644,8 +645,15 @@ static void L_lerp_proc_fx(
		ELSE
		{
		idx = extract_l( L_shr( pos_fx, 21 ) );
		diff_fx = L_sub( pos_fx, L_shl( idx, 21 ) ); /Q21/
		buf_fx[0] = L_add( f_fx[idx], L_shl( Mpy_32_32( diff_fx, L_sub( f_fx[( idx + 1 )], f_fx[idx] ) ), 31 - 21 ) ); /Qx/

		diff_fx = L_sub( pos_fx, L_shl( idx, 21 ) );
		ffxsub = L_sub( f_fx[idx + 1], f_fx[idx] );

		Word64 f_fx64 = W_shl( f_fx[idx], 21 );
		Word64 mult64 = W_mult0_32_32( diff_fx, ffxsub );
		Word64 add64 = W_add( f_fx64, mult64 );

		buf_fx[0] = W_extract_l( W_shr( add64, 21 ) );
		move32();
		}

		@@ -656,8 +664,15 @@ static void L_lerp_proc_fx(
		idx = extract_l( L_shr( pos_fx, 21 ) );
		diff_fx = L_sub( pos_fx, L_shl( idx, 21 ) ); /Q21/

		buf_fx[i] = L_add( f_fx[idx], L_shl( Mpy_32_32( diff_fx, L_sub( f_fx[idx + 1], f_fx[idx] ) ), 31 - 21 ) ); /Qx/
		ffxsub = L_sub( f_fx[idx + 1], f_fx[idx] );

		Word64 f_fx64 = W_shl( f_fx[idx], 21 );
		Word64 mult64 = W_mult0_32_32( diff_fx, ffxsub );
		Word64 add64 = W_add( f_fx64, mult64 );

		buf_fx[i] = W_extract_l( W_shr( add64, 21 ) );
		move32();

		pos_fx = L_add( pos_fx, shift_fx ); /Q21/
		}

lib_com/prot.h

+12 −16

Original line number	Diff line number	Diff line
		@@ -2391,25 +2391,21 @@ ivas_error acelp_core_enc(
		ivas_error acelp_core_enc(
		Encoder_State st, / i/o: encoder state structure */
		/const/ float inp[], /* i : input signal of the current frame */
		// const float ener, /* i : residual energy from Levinson-Durbin */
		float A[NB_SUBFR16k * ( M + 1 )], /* i : A(z) unquantized for the 4 subframes */
		float Aw[NB_SUBFR16k * ( M + 1 )], /* i : weighted A(z) unquant. for subframes */
		const float epsP[M + 1], /* i : LP prediction errors */
		float lsp_new[M], /* i : LSPs at the end of the frame */
		float lsp_mid[M], /* i : LSPs in the middle of the frame */
		const int16_t vad_hover_flag, /* i : VAD hangover flag */
		const int16_t attack_flag, /* i : attack flag (GSC or TC) */
		Word16 A[NB_SUBFR16k * ( M + 1 )], /* i : A(z) unquantized for the 4 subframes */
		Word16 Aw[NB_SUBFR16k * ( M + 1 )], /* i : weighted A(z) unquant. for subframes */
		const Word32 epsP[M + 1], /* i : LP prediction errors */
		Word16 lsp_new[M], /* i : LSPs at the end of the frame */
		Word16 lsp_mid[M], /* i : LSPs in the middle of the frame */
		const Word16 vad_hover_flag, /* i : VAD hangover flag */
		const Word16 attack_flag, /* i : attack flag (GSC or TC) */
		Word32 bwe_exc_extended_fx[], /* i/o: bandwidth extended excitation 2Q_new /
		#ifndef IVAS_FLOAT_FIXED
		float voice_factors, / o : voicing factors */
		#else
		Word16 voice_factors_fx, / o : voicing factors Q15 */
		#endif
		float old_syn_12k8_16k[], /* o : ACELP core synthesis at 12.8kHz or 16kHz to be used by SWB BWE */
		float pitch_buf[NB_SUBFR16k], /* o : floating pitch for each subframe */
		int16_t unbits, / o : number of unused bits */
		Word16 old_syn_12k8_16k[], /* o : ACELP core synthesis at 12.8kHz or 16kHz to be used by SWB BWE */
		Word16 *q_old_syn_12k8_16,
		Word16 pitch_buf[NB_SUBFR16k], /* o : floating pitch for each subframe */
		Word16 unbits, / o : number of unused bits */
		STEREO_TD_ENC_DATA_HANDLE hStereoTD, /* i/o: TD stereo encoder handle */
		float tdm_lsfQ_PCh[M], /* i : Q LSFs for primary channel */
		Word16 tdm_lsfQ_PCh[M], /* i : Q LSFs for primary channel */
		Word16 Q_new );
		#endif

lib_enc/acelp_core_enc.c

+82 −310

File changed.

Preview size limit exceeded, changes collapsed.

lib_enc/analy_sp_fx.c

+198 −16

Original line number	Diff line number	Diff line
		@@ -291,15 +291,23 @@ static void find_enr_dft_fx(
		c_fx = L_shr( c_fx, guarded_bits );
		WHILE( LE_32( freq, crit_bands_fx[i] ) )
		{
		BinE_fx[bin_cnt] = W_extract_l( W_shr( W_add( W_mult0_32_32( ptR_fx, ptR_fx ), W_mult0_32_32( ptI_fx, ptI_fx ) ), 31 ) ); // /* energy */
		Word64 te = ( W_add( W_mult0_32_32( ptR_fx, ptR_fx ), W_mult0_32_32( ptI_fx, ptI_fx ) ) );
		Word16 te_exp = W_norm( te );
		te = W_shl( te, te_exp ); // 2 * Q_inp_dmx + te_exp
		BinE_fx[bin_cnt] = W_extract_h( te ); // 2 * Q_inp_dmx + te_exp - 32
		move32();
		g_fx = L_sub( L_shr( g_1_fx, guarded_bits ), Mpy_32_32( g_2_fx, c_fx ) );
		tmp_fx = L_sub( Mpy_32_32( c_fx, c_1_fx ), Mpy_32_32( s_fx, s_1_fx ) );
		s_fx = L_add( Mpy_32_32( s_fx, c_1_fx ), Mpy_32_32( c_fx, s_1_fx ) );
		c_fx = tmp_fx;
		move32();
		BinE_fx[bin_cnt] = Mpy_32_32( Mpy_32_32( BinE_fx[bin_cnt], norm_val_fx ), g_fx ); /* normalization - corresponds to FFT normalization by 2/L_FFT */

		Word64 ngmult = W_mult0_32_32( norm_val_fx, g_fx ); // Q31 + Q31 - gaurded_bits
		Word16 ngmult_exp = W_norm( ngmult );
		ngmult = W_shl( ngmult, ngmult_exp ); // Q31 + Q31 - gaurded_bits + ngmult_exp
		BinE_fx[bin_cnt] = Mpy_32_32( W_extract_h( ngmult ), BinE_fx[bin_cnt] ); //(Q31 + Q31 - gaurded_bits + ngmult_exp) - Q32 + (2 * Q_inp_dmx + te_exp - 32) - 31
		move32();
		BinE_fx[bin_cnt] = L_shl( BinE_fx[bin_cnt], guarded_bits ); // 2 * Q_inp_dmx - 32 - 1
		BinE_fx[bin_cnt] = L_shl( BinE_fx[bin_cnt], sub( Qout, sub( sub( add( add( shl( Q_inp_dmx, 1 ), te_exp ), ngmult_exp ), 34 ), guarded_bits ) ) );
		move32();
		band_fx[i] = L_add( BinE_fx[bin_cnt], band_fx[i] );
		move32();
		@@ -313,8 +321,6 @@ static void find_enr_dft_fx(
		c_fx = L_shl( c_fx, guarded_bits );
		band_fx[i] = Mpy_32_16_r( band_fx[i], inv_tbl_fx[cnt] ); /* normalization per frequency bin */
		move32();
		band_fx[i] = L_shl( band_fx[i], sub( Qout, sub( shl( Q_inp_dmx, 1 ), 32 ) ) );
		move32();
		band_ener_fx[i] = band_fx[i]; /* per band energy without E_MIN */
		move32();
		if ( LT_32( band_fx[i], L_shl( E_MIN_Q11_FX, sub( Qout, Q11 ) ) ) )
		@@ -351,12 +357,19 @@ static void find_enr_dft_fx(
		c_fx = L_shr( c_fx, guarded_bits );
		WHILE( LT_32( freq, 6399 ) )
		{
		BinE_fx[bin_cnt] = W_extract_l( W_shr( W_add( W_mult0_32_32( ptR_fx, ptR_fx ), W_mult0_32_32( ptI_fx, ptI_fx ) ), 31 ) );
		Word64 te = ( W_add( W_mult0_32_32( ptR_fx, ptR_fx ), W_mult0_32_32( ptI_fx, ptI_fx ) ) );
		Word16 te_exp = W_norm( te );
		te = W_shl( te, te_exp ); // 2 * Q_inp_dmx + te_exp
		BinE_fx[bin_cnt] = W_extract_h( te ); // 2 * Q_inp_dmx + te_exp - 32
		move32();
		g_fx = L_sub( L_shr( g_1_fx, guarded_bits ), Mpy_32_32( g_2_fx, c_fx ) );
		BinE_fx[bin_cnt] = Mpy_32_32( Mpy_32_32( BinE_fx[bin_cnt], norm_val_fx ), g_fx );

		Word64 ngmult = W_mult0_32_32( norm_val_fx, g_fx ); // Q31 + Q31 - gaurded_bits
		Word16 ngmult_exp = W_norm( ngmult );
		ngmult = W_shl( ngmult, ngmult_exp ); // Q31 + Q31 - gaurded_bits + ngmult_exp
		BinE_fx[bin_cnt] = Mpy_32_32( W_extract_h( ngmult ), BinE_fx[bin_cnt] ); //(Q31 + Q31 - gaurded_bits + ngmult_exp) - Q32 + (2 * Q_inp_dmx + te_exp - 32) - 31
		move32();
		BinE_fx[bin_cnt] = L_shl( BinE_fx[bin_cnt], guarded_bits ); // Q11
		BinE_fx[bin_cnt] = L_shl( BinE_fx[bin_cnt], sub( Qout, sub( sub( add( add( shl( Q_inp_dmx, 1 ), te_exp ), ngmult_exp ), 34 ), guarded_bits ) ) );
		move32();
		band_fx[i] = L_add( BinE_fx[bin_cnt], band_fx[i] );
		move32();
		@@ -368,7 +381,7 @@ static void find_enr_dft_fx(
		}
		c_fx = L_shl( c_fx, guarded_bits );
		band_fx[i] = Mpy_32_16_r( band_fx[i], inv_tbl_fx[cnt] );
		band_fx[i] = L_shl( band_fx[i], sub( Qout, sub( shl( Q_inp_dmx, 1 ), 32 ) ) );
		move32();
		band_ener_fx[i] = band_fx[i];
		move32();
		if ( LT_32( band_fx[i], L_shl( E_MIN_Q11_FX, sub( Qout, Q11 ) ) ) )
		@@ -381,13 +394,11 @@ static void find_enr_dft_fx(
		/* put bin energies from BinE into Bin_E[L_FFT/2-1] (interpolate 40 Hz bin values to fit into 50 Hz bins) */
		/* Last value of Bin_E is handled outside this function*/
		assert( bin_cnt == ( STEREO_DFT_N_12k8_ENC / 2 - 1 ) );
		Scale_sig32( BinE_fx, 256, sub( Qout, sub( shl( Q_inp_dmx, 1 ), 32 ) ) );
		BinE_fx[STEREO_DFT_N_12k8_ENC / 2 - 1] = BinE_fx[STEREO_DFT_N_12k8_ENC / 2 - 2];
		move32();
		L_lerp_fx( BinE_fx, Bin_E_fx, L_FFT / 2, STEREO_DFT_N_12k8_ENC / 2, &Qout );
		MVR2R_WORD32( Bin_E_fx, ptE_fx, VOIC_BINS ); // Qout


		/* find the total log energy */
		tmp_fx = *Etot_fx;
		move32();
		@@ -396,6 +407,9 @@ static void find_enr_dft_fx(
		tmp_fx = L_add( tmp_fx, band_fx[i] ); // Qout
		}
		*Etot_fx = tmp_fx;

		scale_sig32( band_fx, NB_BANDS, 2 );
		scale_sig32( band_ener_fx, NB_BANDS, 2 );
		move32();
		return;
		}
		@@ -550,6 +564,168 @@ void ivas_analy_sp_fx(
		}


		return;
		}

		void ivas_analy_sp_fx_front(
		const Word16 element_mode, /* i : element mode */
		CPE_ENC_HANDLE hCPE, /* i/o: CPE encoder structure */
		const Word32 input_Fs, /* i : input sampling rate */
		Word16 speech, / i : speech buffer Q_new - preemph_bits */
		const Word16 Q_new, /* i : current scaling exp Q0 */
		Word32 fr_bands, / o : energy in critical frequency bands Q_new + QSCALE */
		Word32 lf_E, / o : per bin E for first... Q_new + QSCALE - 2*/
		Word16 Etot, / o : total input energy Q8 */
		const Word16 min_band, /* i : minimum critical band Q0 */
		const Word16 max_band, /* i : maximum critical band Q0 */
		const Word32 e_min_scaled, /* i : minimum energy scaled Q_new + QSCALE */
		Word16 Scale_fac[2], /* o : FFT scales factors (2 values by frame) Q0 */
		Word32 Bin_E, / o : per-bin energy spectrum Q7 */
		Word32 Bin_E_old, / o : per-bin energy spectrum of the previous frame Q7 */
		Word32 PS, / o : per-bin energy spectrum Q_new + QSCALE */
		Word16 EspecdB, / o : per-bin log energy spectrum (with f=0) Q7 */
		Word32 band_energies, / o : energy in critical frequency bands without minimum noise floor MODE2_E_MIN (Q_new + QSCALE + 2)*/
		Word16 fft_buff / o : FFT coefficients (Q_new + QSCALE + 2) */
		)
		{
		Word16 *pt;
		Word16 i_subfr, i;
		Word32 *pt_bands;
		Word32 Ltmp, LEtot;
		Word16 *pt_fft;
		Word16 Min_val, Max_val;
		Word16 Scale_fac2;
		Word16 fft_temp[L_FFT];

		/-----------------------------------------------------------------
		* Compute spectrum
		* find energy per critical frequency band and total energy in dB
		-----------------------------------------------------------------/

		pt_bands = fr_bands;
		pt_fft = fft_buff;
		LEtot = L_deposit_l( 0 );
		IF( NE_16( element_mode, IVAS_CPE_DFT ) )
		{
		FOR( i_subfr = 0; i_subfr <= 1; i_subfr++ )
		{
		pt = speech + 3 * ( L_SUBFR / 2 ) - L_FFT / 2;
		IF( i_subfr != 0 )
		{
		pt = speech + 7 * ( L_SUBFR / 2 ) - L_FFT / 2;
		}

		/* Clear 1st value of 1st part, copy 1st value of 2nd part */
		fft_temp[0] = 0;
		move16();
		fft_temp[L_FFT / 2] = pt[L_FFT / 2];
		move16();
		Max_val = s_max( fft_temp[0], fft_temp[L_FFT / 2] );
		Min_val = s_min( fft_temp[0], fft_temp[L_FFT / 2] );

		FOR( i = 1; i < L_FFT / 2; i++ )
		{
		/* 1st windowed part */
		fft_temp[i] = mult_r( pt[i], sqrt_han_window_fx[i] );
		move16();
		if ( fft_temp[i] > 0 )
		Max_val = s_max( Max_val, fft_temp[i] );
		if ( fft_temp[i] < 0 )
		Min_val = s_min( Min_val, fft_temp[i] );

		/* 2nd windowed part */
		fft_temp[L_FFT - i] = mult_r( pt[L_FFT - i], sqrt_han_window_fx[i] );
		move16();
		if ( fft_temp[L_FFT - i] > 0 )
		Max_val = s_max( Max_val, fft_temp[L_FFT - i] );
		if ( fft_temp[L_FFT - i] < 0 )
		Min_val = s_min( Min_val, fft_temp[L_FFT - i] );
		}

		/* Combine -Min_val and Max_val into one */
		Max_val = s_max( negate( Min_val ), Max_val );

		Scale_fac[i_subfr] = s_min( sub( norm_s( Max_val ), 1 ), 6 );
		move16();
		Scale_fac2 = shl( Scale_fac[i_subfr], 1 );
		Scale_sig( fft_temp, L_FRAME_12k8, Scale_fac[i_subfr] );

		r_fft_fx_lc( FFT_W128, SIZE_256, SIZE2_256, NUM_STAGE_256, fft_temp, pt_fft, 1 );
		Scale_sig( pt_fft, L_FFT, -1 ); // Q(-1)
		/e_min_scaled = Q_new + QSCALE + 2/
		find_enr( pt_fft, pt_bands, lf_E + i_subfr * VOIC_BINS, &LEtot, min_band, max_band,
		add( Q_new, Scale_fac2 ), e_min_scaled, &Bin_E[i_subfr * L_FFT / 2], BIN, band_energies + i_subfr * NB_BANDS );
		pt_bands += NB_BANDS;
		pt_fft += L_FFT;
		}
		LEtot = L_shl_sat( LEtot, 2 ); // Q_new + Q_SCALE - 2
		}
		ELSE
		{
		Word16 Q_inp_dmx = Q_factor_arrL( hCPE->hStereoDft->DFT[0], STEREO_DFT_N_MAX_ENC );
		floatToFixed_arrL( hCPE->hStereoDft->DFT[0], hCPE->hStereoDft->DFT_fx[0], Q_inp_dmx, STEREO_DFT_N_MAX_ENC );
		Word16 Qout = add( Q_new, QSCALE - 2 );
		find_enr_dft_fx( hCPE, input_Fs, hCPE->hStereoDft->DFT_fx[0], pt_bands, lf_E, &LEtot, min_band, max_band, Bin_E, band_energies, Q_inp_dmx, Qout );
		MVR2R_WORD32( lf_E, lf_E + VOIC_BINS, VOIC_BINS );
		MVR2R_WORD32( Bin_E, Bin_E + ( L_FFT / 2 ), L_FFT / 2 );
		MVR2R_WORD32( band_energies, band_energies + NB_BANDS, NB_BANDS );
		MVR2R_WORD32( pt_bands, pt_bands + NB_BANDS, NB_BANDS );
		LEtot = L_shl( LEtot, 1 );
		}
		Word32 temp32_log;
		temp32_log = L_add( BASOP_Util_Log2( L_shr( LEtot, 1 ) ), L_shl( sub( Q31, add( Q_new, QSCALE - Q2 - 1 ) ), Q25 ) );
		temp32_log = Mpy_32_32( temp32_log, 1616142483 ); // log10(x) = log2(x)/log2(10)
		/* 10.0 * log10( (float) tmp )*/
		/* 10.0/log2(10) in Q29 = 1616142483*/
		*Etot = extract_l( L_shr( temp32_log, 23 - 8 ) ); // Q8

		Bin_E[L_FFT / 2 - 1] = Bin_E[L_FFT / 2 - 2];
		move32();
		Bin_E[L_FFT - 1] = Bin_E[L_FFT - 2];
		move32();

		/* Per-bin log-energy spectrum */

		FOR( i = 0; i < L_FFT / 2; i++ )
		{
		Bin_E_old[i] = Bin_E[i];
		move32();
		/* tmp = (input[i] + input[i+Len]+0.001f)/2.0f */
		// Ltmp = L_max( L_shr( 21474, sub( Q31, add( Q_new, QSCALE ) ) ), L_add( L_shr( Bin_E[i], 1 ), L_shr( Bin_E[i + L_FFT / 2], 1 ) ) );
		Ltmp = L_shr( L_add( Bin_E[i], Bin_E[i + L_FFT / 2] ), 1 ); // average
		if ( Ltmp == 0 )
		{
		// Ltmp = L_min( L_shr( 21474, sub( Q31, add( Q_new, QSCALE - 2 ) ) ), 1 );
		Ltmp = L_max( L_shr( 21474, sub( Q31, add( Q_new, QSCALE - 2 ) ) ), 1 );

		if ( EspecdB != NULL )
		{
		// temp32_log = L_add( BASOP_Util_Log2( Ltmp ), L_shl( sub( Q31, Q_new + QSCALE ), Q25 ) );
		// EspecdB[i] = -1474; /* log(1e-5) in Q7 */
		EspecdB[i] = -14736; /* log(1e-5) in Q7 */
		move16();
		}
		}
		else
		{
		if ( EspecdB != NULL )
		{
		temp32_log = L_add( BASOP_Util_Log2( Ltmp ), L_shl( sub( Q31, Q_new + QSCALE - 2 - 1 ), Q25 ) );
		EspecdB[i] = extract_l( L_shr( Mpy_32_32( temp32_log, 1860652798 ), Q22 - Q7 ) ); /log(x) = log2(x)logf(2)*/
		/* 10.0log((float)tmp) /
		/* 10.0logf(2) in Q28 = 1860652798 /

		move16();
		}
		}
		if ( PS != NULL )
		{
		PS[i] = Ltmp;
		move32();
		}
		}


		return;
		}
		#endif
		@@ -629,9 +805,12 @@ static void find_enr(
		FOR( ; freq <= crit_bands[i]; freq += BIN_FREQ_FX )
		{
		/ptE = ptR * ptR + ptI * ptI / /* energy */
		Ltmp = L_mult( ptI, ptI );
		Ltmp = L_mac( Ltmp, ptR, ptR );
		Word64 te = ( W_add( W_mult0_32_32( ptR, ptR ), W_mult0_32_32( ptI, ptI ) ) ); // 2 * Qfft
		Word16 te_exp = W_norm( te );
		te = W_shl( te, te_exp ); // 2 * Qfft + te_exp
		Ltmp = W_extract_h( te ); // 2 * Qfft + te_exp - 32

		Ltmp = L_shr( Ltmp, sub( te_exp, 32 ) );
		/* ptE = 4.0 / (L_FFTL_FFT) /
		/* normalization - corresponds to FFT normalization by 2/L_FFT */
		BASOP_SATURATE_WARNING_OFF_EVS; /* saturation seems to have no effect (tested by simulation) */
		@@ -714,8 +893,11 @@ static void find_enr(
		FOR( ; freq <= crit_bands[i]; freq += BIN_FREQ_FX )
		{
		/* ptE = ptR * ptR + ptI * ptI /
		Ltmp = L_mult( ptI, ptI );
		Ltmp = L_mac( Ltmp, ptR, ptR );
		Word64 te = ( W_add( W_mult0_32_32( ptR, ptR ), W_mult0_32_32( ptI, ptI ) ) );
		Word16 te_exp = W_norm( te );
		te = W_shl( te, te_exp );
		Ltmp = W_extract_h( te );
		Ltmp = L_shr( Ltmp, sub( te_exp, 32 ) );

		/* ptE = 4.0 / (L_FFTL_FFT) /
		/* normalization - corresponds to FFT normalization by 2/L_FFT */