Merge branch 'be_renderer_opt_1' into 'main' (1e645e74) · Commits · SA4 / Audio / IVAS BASOP

lib_com/options.h

+1 −0

Original line number	Diff line number	Diff line
		@@ -75,6 +75,7 @@
		#define FIX_1379_MASA_ANGLE_ROUND

		/* Note: each compile switch (FIX_1101_...) is independent from the other ones */
		#define OPT_SBA_REND_V1_BE
		#define OPT_SBA_ENC_V1_BE
		#define OPT_BIN_RENDERER_V1
		#define OPT_BIN_RENDERER_V2

lib_rend/ivas_dirac_dec_binaural_functions_fx.c

+165 −0

Original line number	Diff line number	Diff line
		@@ -3415,6 +3415,169 @@ static void eig2x2_fx(
		move16();
		move16();

		#ifdef OPT_SBA_REND_V1_BE
		/* Eigenvectors */
		FOR( ch = 0; ch < BINAURAL_CHANNELS; ch++ )
		{
		Word16 q_diff = sub( q_e, *q_D );
		IF( q_diff > 0 )
		{
		tmp1 = L_sub( D_fx[ch], L_shr( e1, q_diff ) );
		tmp2 = L_sub( D_fx[ch], L_shr( e2, q_diff ) );
		q_tmp1 = *q_D;
		move16();
		}
		ELSE
		{
		tmp1 = L_sub( L_shl( D_fx[ch], q_diff ), e1 );
		tmp2 = L_sub( L_shl( D_fx[ch], q_diff ), e2 );
		q_tmp1 = q_e;
		move16();
		}

		IF( GT_32( L_abs( tmp2 ), L_abs( tmp1 ) ) )
		{
		s_fx = tmp2;
		move32();
		exp = sub( norm_l( s_fx ), 1 );
		tmp2 = Mpy_32_32( s_fx, s_fx );
		q_tmp2 = sub( add( q_tmp1, q_tmp1 ), 31 );

		tmp2 = BASOP_Util_Add_Mant32Exp( crossSquare_fx, sub( 31, q_crossSquare ), tmp2, sub( 31, q_tmp2 ), &q_tmp2 );
		q_tmp2 = sub( 31, q_tmp2 );

		tmp3 = BASOP_Util_Add_Mant32Exp( tmp2, sub( 31, q_tmp2 ), epsilon_mant, epsilon_exp, &exp_tmp3 );

		tmp2 = BASOP_Util_Divide3232_Scale_newton( ONE_IN_Q30, tmp3, &exp );
		exp = sub( exp, sub( Q30, sub( 31, exp_tmp3 ) ) );
		normVal_fx = Sqrt32( tmp2, &exp ); // q_tmp2
		q_tmp2 = sub( 31, exp );

		q_diff = sub( q_c, q_tmp1 );
		IF( q_diff > 0 )
		{
		c_re = L_shr( c_re, q_diff );
		c_im = L_shr( c_im, q_diff );
		q_c = q_tmp1;
		move16();
		}
		ELSE
		{
		s_fx = L_shl( s_fx, q_diff );
		q_tmp1 = q_c;
		move16();
		}

		Ure_fx[0][ch] = Mpy_32_32( s_fx, normVal_fx );
		move32();
		Ure_fx[1][ch] = Mpy_32_32( c_re, normVal_fx );
		move32();
		Uim_fx[1][ch] = Mpy_32_32( c_im, normVal_fx );
		move32();
		q_U_1 = sub( add( q_tmp1, q_tmp2 ), 31 );

		IF( q_U_2 != 0 )
		{
		q_diff = sub( q_U_2, q_U_1 );
		IF( q_diff > 0 )
		{
		Ure_fx[1][ch - 1] = L_shr( Ure_fx[1][ch - 1], q_diff );
		Ure_fx[0][ch - 1] = L_shr( Ure_fx[0][ch - 1], q_diff );
		Uim_fx[0][ch - 1] = L_shr( Uim_fx[0][ch - 1], q_diff );
		q_U_2 = q_U_1;
		move32();
		move32();
		move32();
		move16();
		}
		ELSE IF( GT_16( q_U_1, q_U_2 ) )
		{
		Ure_fx[1][ch] = L_shl( Ure_fx[1][ch], q_diff );
		Ure_fx[0][ch] = L_shl( Ure_fx[0][ch], q_diff );
		Uim_fx[1][ch] = L_shl( Uim_fx[1][ch], q_diff );
		q_U_1 = q_U_2;
		move32();
		move32();
		move32();
		move16();
		}
		}
		q_U_2 = q_U_1;
		move16();
		}
		ELSE
		{
		s_fx = tmp1;
		move32();

		exp = sub( norm_l( s_fx ), 1 );
		tmp2 = Mpy_32_32( s_fx, s_fx );
		q_tmp2 = sub( add( q_tmp1, q_tmp1 ), 31 );

		tmp2 = BASOP_Util_Add_Mant32Exp( crossSquare_fx, sub( 31, q_crossSquare ), tmp2, sub( 31, q_tmp2 ), &q_tmp2 );
		q_tmp2 = sub( 31, q_tmp2 );

		tmp3 = BASOP_Util_Add_Mant32Exp( tmp2, sub( 31, q_tmp2 ), epsilon_mant, epsilon_exp, &exp_tmp3 );

		tmp2 = BASOP_Util_Divide3232_Scale_newton( ONE_IN_Q30, tmp3, &exp );
		exp = sub( exp, sub( Q30, sub( 31, exp_tmp3 ) ) );
		normVal_fx = Sqrt32( tmp2, &exp ); // q_tmp2
		q_tmp2 = sub( 31, exp );

		q_diff = sub( q_c, q_tmp1 );
		IF( q_diff > 0 )
		{
		c_re = L_shr( c_re, q_diff );
		c_im = L_shr( c_im, q_diff );
		q_c = q_tmp1;
		move16();
		}
		ELSE
		{
		s_fx = L_shl( s_fx, q_diff );
		q_tmp1 = q_c;
		move16();
		}

		Ure_fx[1][ch] = Mpy_32_32( s_fx, normVal_fx );
		move32();
		Ure_fx[0][ch] = Mpy_32_32( c_re, normVal_fx );
		move32();
		Uim_fx[0][ch] = Mpy_32_32( L_negate( c_im ), normVal_fx );
		move32();
		q_U_2 = sub( add( q_tmp1, q_tmp2 ), 31 );

		IF( q_U_1 != 0 )
		{
		q_diff = sub( q_U_2, q_U_1 );
		IF( q_diff > 0 )
		{
		Ure_fx[1][ch] = L_shr( Ure_fx[1][ch], q_diff );
		Ure_fx[0][ch] = L_shr( Ure_fx[0][ch], q_diff );
		Uim_fx[0][ch] = L_shr( Uim_fx[0][ch], q_diff );
		q_U_2 = q_U_1;
		move32();
		move32();
		move32();
		move16();
		}
		ELSE IF( GT_16( q_U_1, q_U_2 ) )
		{
		Ure_fx[1][ch - 1] = L_shl( Ure_fx[1][ch - 1], q_diff );
		Ure_fx[0][ch - 1] = L_shl( Ure_fx[0][ch - 1], q_diff );
		Uim_fx[1][ch - 1] = L_shl( Uim_fx[1][ch - 1], q_diff );
		q_U_1 = q_U_2;
		move32();
		move32();
		move32();
		move16();
		}
		}
		q_U_1 = q_U_2;
		move16();
		}
		}
		#else /* OPT_SBA_REND_V1_BE */
		/* Eigenvectors */
		FOR( ch = 0; ch < BINAURAL_CHANNELS; ch++ )
		{
		@@ -3571,6 +3734,8 @@ static void eig2x2_fx(
		move16();
		}
		}
		#endif /* OPT_SBA_REND_V1_BE */

		if ( q_U_1 != 0 )
		{
		*q_U = q_U_1;

lib_rend/ivas_dirac_output_synthesis_dec_fx.c

+61 −2

Original line number	Diff line number	Diff line
		@@ -2465,6 +2465,11 @@ void ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx(
		move16();
		move16();

		#ifdef OPT_SBA_REND_V1_BE
		Word32 cmp = W_shl_sat_l( DIRAC_GAIN_LIMIT_Q26, sub( h_dirac_output_synthesis_state->gains_dir_prev_q, 26 ) );
		Word32 cmp2 = W_extract_h( W_shl( W_mult_32_32( DIRAC_GAIN_LIMIT_Q26, L_shl( 1, h_dirac_output_synthesis_state->gains_diff_prev_q ) ), Q5 ) );
		#endif /* OPT_SBA_REND_V1_BE */

		FOR( k = 0; k < nchan_out_woLFE; k++ )
		{
		Word32 power_smooth_temp;
		@@ -2512,11 +2517,19 @@ void ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx(
		*( p_gains_dir ) = 0;
		move32();
		}
		#ifdef OPT_SBA_REND_V1_BE
		ELSE IF( GT_32( *( p_gains_dir ), cmp ) )
		{
		( p_gains_dir ) = cmp; /26 + h_dirac_output_synthesis_state->gains_dir_prev_q + 1 + 5 - 32 -> h_dirac_output_synthesis_state->gains_dir_prev_q*/
		move32();
		}
		#else /* OPT_SBA_REND_V1_BE */
		ELSE IF( GT_32( *( p_gains_dir ), W_extract_h( W_shl( W_mult_32_32( DIRAC_GAIN_LIMIT_Q26, L_shl( 1, h_dirac_output_synthesis_state->gains_dir_prev_q ) ), Q5 ) ) ) )
		{
		( p_gains_dir ) = W_extract_h( W_shl( W_mult_32_32( DIRAC_GAIN_LIMIT_Q26, L_shl( 1, h_dirac_output_synthesis_state->gains_dir_prev_q ) ), Q5 ) ); /26 + h_dirac_output_synthesis_state->gains_dir_prev_q + 1 + 5 - 32 -> h_dirac_output_synthesis_state->gains_dir_prev_q*/
		move32();
		}
		#endif /* OPT_SBA_REND_V1_BE */

		IF( *( p_cy_cross_dir_smooth_prev++ ) < 0 )
		{
		@@ -2549,11 +2562,19 @@ void ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx(
		*( p_gains_diff ) = 0;
		move32();
		}
		#ifdef OPT_SBA_REND_V1_BE
		ELSE IF( GT_32( ( p_gains_diff ), cmp2 ) ) /h_dirac_output_synthesis_state->gains_diff_prev_q*/
		{
		( p_gains_diff ) = cmp2; /h_dirac_output_synthesis_state->gains_diff_prev_q*/
		move32();
		}
		#else /* OPT_SBA_REND_V1_BE */
		ELSE IF( GT_32( ( p_gains_diff ), W_extract_h( W_shl( W_mult_32_32( DIRAC_GAIN_LIMIT_Q26, L_shl( 1, h_dirac_output_synthesis_state->gains_diff_prev_q ) ), Q5 ) ) ) ) /h_dirac_output_synthesis_state->gains_diff_prev_q*/
		{
		( p_gains_diff ) = W_extract_h( W_shl( W_mult_32_32( DIRAC_GAIN_LIMIT_Q26, L_shl( 1, h_dirac_output_synthesis_state->gains_diff_prev_q ) ), Q5 ) ); /h_dirac_output_synthesis_state->gains_diff_prev_q*/
		move32();
		}
		#endif /* OPT_SBA_REND_V1_BE */
		p_gains_diff++;
		}

		@@ -2564,15 +2585,25 @@ void ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx(
		g1 = alpha[l]; // Q31
		move32();
		g2 = L_sub( ONE_IN_Q31, g1 ); // Q31
		#ifdef OPT_SBA_REND_V1_BE
		W_temp = W_mac_32_32( W_mult_32_32( g1, ( *( p_cy_auto_dir_smooth++ ) ) ),
		g2, ( ( p_cy_auto_dir_smooth_prev ) ) ); /32+q_cy_auto_dir_smooth_prev_local*/
		#else /* OPT_SBA_REND_V1_BE */
		W_temp = W_add( W_mult_32_32( g1, ( *( p_cy_auto_dir_smooth++ ) ) ),
		W_mult_32_32( g2, ( ( p_cy_auto_dir_smooth_prev ) ) ) ); /32+q_cy_auto_dir_smooth_prev_local*/
		#endif /* OPT_SBA_REND_V1_BE */
		q_tmp = W_norm( W_temp );
		L_tmp = W_extract_h( W_shl( W_temp, q_tmp ) ); // q_cy_auto_dir_smooth_prev_local + q_tmp
		*( p_cy_auto_dir_smooth_prev++ ) = L_shr_r( L_tmp, q_tmp ); // q_cy_auto_dir_smooth_prev_local

		move32();
		#ifdef OPT_SBA_REND_V1_BE
		( p_cy_cross_dir_smooth_prev ) = Madd_32_32( Mpy_32_32( g1, ( ( p_cy_cross_dir_smooth ) ) ),
		g2, ( *( p_cy_cross_dir_smooth_prev ) ) ); // (Q31, q_cy_cross_dir_smooth_prev) -> q_cy_cross_dir_smooth_prev
		#else /* OPT_SBA_REND_V1_BE */
		( p_cy_cross_dir_smooth_prev ) = L_add( Mpy_32_32( g1, ( ( p_cy_cross_dir_smooth ) ) ),
		Mpy_32_32( g2, ( *( p_cy_cross_dir_smooth_prev ) ) ) ); // (Q31, q_cy_cross_dir_smooth_prev) -> q_cy_cross_dir_smooth_prev
		#endif /* OPT_SBA_REND_V1_BE */
		move32();
		test();
		if ( ( p_cy_cross_dir_smooth_prev ) == 0 && ( ( p_cy_cross_dir_smooth ) != 0 ) )
		@@ -2604,11 +2635,19 @@ void ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx(
		*( p_gains_dir ) = 0;
		move32();
		}
		#ifdef OPT_SBA_REND_V1_BE
		ELSE IF( GT_32( ( p_gains_dir ), cmp ) ) /gains_dir_prev_q*/
		{
		( p_gains_dir ) = cmp; /gains_dir_prev_q*/
		move32();
		}
		#else /* OPT_SBA_REND_V1_BE */
		ELSE IF( GT_32( ( p_gains_dir ), W_extract_h( W_shl( W_mult_32_32( DIRAC_GAIN_LIMIT_Q26, L_shl( 1, h_dirac_output_synthesis_state->gains_dir_prev_q ) ), Q5 ) ) ) ) /gains_dir_prev_q*/
		{
		( p_gains_dir ) = W_extract_h( W_shl( W_mult_32_32( DIRAC_GAIN_LIMIT_Q26, L_shl( 1, h_dirac_output_synthesis_state->gains_dir_prev_q ) ), Q5 ) ); /gains_dir_prev_q*/
		move32();
		}
		#endif /* OPT_SBA_REND_V1_BE */

		IF( *( p_cy_cross_dir_smooth_prev++ ) < 0 )
		{
		@@ -2695,7 +2734,11 @@ void ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx(
		shl( i_mult( proto_direct_index[k], num_freq_bands ), Q1 );
		FOR( l = 0; l < num_freq_bands; l++ )
		{
		#ifdef OPT_SBA_REND_V1_BE
		g = Madd_32_32( Mpy_32_32( g1, ( p_gain_1++ ) ), g2, ( p_gain_2++ ) ); // (Q31, gains_dir_prev_q) -> gains_dir_prev_q
		#else /* OPT_SBA_REND_V1_BE */
		g = L_add( Mpy_32_32( g1, ( p_gain_1++ ) ), Mpy_32_32( g2, ( p_gain_2++ ) ) ); // (Q31, gains_dir_prev_q) -> gains_dir_prev_q
		#endif /* OPT_SBA_REND_V1_BE */

		Cldfb_RealBuffer64_fx[k][buf_idx][l] = W_mult0_32_32( g, ( *( p_power_smooth++ ) ) ); // (gains_dir_prev_q, q_proto_direct_buffer) -> gains_dir_prev_q + q_proto_direct_buffer
		move64();
		@@ -2717,7 +2760,12 @@ void ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx(
		{
		FOR( l = 0; l < h_dirac_output_synthesis_params->max_band_decorr; l++ )
		{
		#ifdef OPT_SBA_REND_V1_BE
		g = Madd_32_32( Mpy_32_32( g1, ( p_gain_1++ ) ), g2, ( p_gain_2++ ) ); // (Q31, gains_diff_prev_q) -> gains_diff_prev_q
		#else /* OPT_SBA_REND_V1_BE */
		g = L_add( Mpy_32_32( g1, ( p_gain_1++ ) ), Mpy_32_32( g2, ( p_gain_2++ ) ) ); // (Q31, gains_diff_prev_q) -> gains_diff_prev_q

		#endif /* OPT_SBA_REND_V1_BE */
		Cldfb_RealBuffer64_fx[k][buf_idx][l] = W_add( Cldfb_RealBuffer64_fx[k][buf_idx][l],
		W_shr( W_mult0_32_32( g, ( *( p_power_smooth_diff++ ) ) ), negate( q_align ) ) ); // (gains_diff_prev_q, q_proto_direct_buffer) -> gains_diff_prev_q + q_proto_direct_buffer
		move64();
		@@ -2766,16 +2814,27 @@ void ivas_dirac_dec_output_synthesis_process_subframe_psd_ls_fx(
		}
		}
		q_align = W_norm( W_temp );
		#ifdef OPT_SBA_REND_V1_BE
		Word16 shift = sub( q_align, 32 );
		#endif /* OPT_SBA_REND_V1_BE */

		FOR( buf_idx = 0; buf_idx < nbslots; ++buf_idx )
		{
		FOR( k = 0; k < nchan_out_woLFE; k++ )
		{
		FOR( l = 0; l < num_freq_bands; l++ )
		{
		#ifdef OPT_SBA_REND_V1_BE
		RealBuffer[k][buf_idx][l] = W_shl_sat_l( Cldfb_RealBuffer64_fx[k][buf_idx][l], shift ); /( ( ( h_dirac_output_synthesis_state->proto_direct_buffer_f_q+h_dirac_output_synthesis_state->gains_dir_prev_q )+ q_align )- 32 )/
		move32();
		ImagBuffer[k][buf_idx][l] = W_shl_sat_l( Cldfb_ImagBuffer64_fx[k][buf_idx][l], shift ); /( ( ( h_dirac_output_synthesis_state->proto_direct_buffer_f_q+h_dirac_output_synthesis_state->gains_dir_prev_q )+ q_align )- 32 )/
		move32();
		#else /* OPT_SBA_REND_V1_BE */
		RealBuffer[k][buf_idx][l] = W_extract_h( W_shl( Cldfb_RealBuffer64_fx[k][buf_idx][l], q_align ) ); /( ( ( h_dirac_output_synthesis_state->proto_direct_buffer_f_q+h_dirac_output_synthesis_state->gains_dir_prev_q )+ q_align )- 32 )/
		move32();
		ImagBuffer[k][buf_idx][l] = W_extract_h( W_shl( Cldfb_ImagBuffer64_fx[k][buf_idx][l], q_align ) ); /( ( ( h_dirac_output_synthesis_state->proto_direct_buffer_f_q+h_dirac_output_synthesis_state->gains_dir_prev_q )+ q_align )- 32 )/
		move32();
		#endif /* OPT_SBA_REND_V1_BE */
		}
		}
		}

lib_rend/ivas_dirac_rend_fx.c

+6 −2

Original line number	Diff line number	Diff line
		@@ -3155,7 +3155,11 @@ void protoSignalComputation4_fx(

		sq_tmp_fx = Madd_32_32( Mpy_32_32( proto_frame_f_fx[idx], proto_frame_f_fx[idx] ), proto_frame_f_fx[idx + 1], proto_frame_f_fx[idx + 1] ); // 2*(proto_frame_f_q)-31
		sq_tmp_q = sub( add( proto_frame_f_q, proto_frame_f_q ), 31 );
		#ifdef OPT_SBA_REND_V1_BE
		proto_power_smooth_fx_q = s_min( *proto_power_smooth_q, sq_tmp_q );

		proto_power_smooth_fx[l + ( k * num_freq_bands )] = L_add( L_shr( proto_power_smooth_fx[l + ( k * num_freq_bands )], sub( *proto_power_smooth_q, proto_power_smooth_fx_q ) ), L_shr( sq_tmp_fx, sub( sq_tmp_q, proto_power_smooth_fx_q ) ) ); // proto_power_smooth_fx_q
		#else /* OPT_SBA_REND_V1_BE */
		IF( LT_16( *proto_power_smooth_q, sq_tmp_q ) )
		{
		proto_power_smooth_fx[l + ( k * num_freq_bands )] = L_add( proto_power_smooth_fx[l + ( k * num_freq_bands )], L_shr( sq_tmp_fx, sub( sq_tmp_q, *proto_power_smooth_q ) ) ); // proto_power_smooth_q
		@@ -3170,7 +3174,7 @@ void protoSignalComputation4_fx(
		proto_power_smooth_fx_q = sq_tmp_q;
		move16();
		}

		#endif /* OPT_SBA_REND_V1_BE */
		p_proto_buffer_fx[idx] = proto_frame_f_fx[idx]; // proto_frame_f_q
		move32();
		p_proto_buffer_fx[idx + 1] = proto_frame_f_fx[idx + 1]; // proto_frame_f_q

lib_rend/ivas_efap_fx.c

+10 −6

Original line number	Diff line number	Diff line
		@@ -2513,7 +2513,11 @@ static Word16 in_tri_fx(
		#endif /* VEC_ARITH_OPT_v1 */

		/* Verification of the non-colinearity : Q22 * Q22 = Q13 */
		#ifdef OPT_SBA_REND_V1_BE
		invFactor = Msub_32_32( Mpy_32_32( tmpDot1[0], tmpDot2[1] ), tmpDot1[1], tmpDot2[0] ); /q22+q22-q31->q13/
		#else /* OPT_SBA_REND_V1_BE */
		invFactor = L_sub( Mpy_32_32( tmpDot1[0], tmpDot2[1] ), Mpy_32_32( tmpDot1[1], tmpDot2[0] ) ); /q22+q22-q31->q13/
		#endif /* OPT_SBA_REND_V1_BE */

		IF( invFactor == 0 )
		{