Merge branch 'stereo_mch_optimizations' into 'main' (ee32b82e) · Commits · SA4 / Audio / IVAS BASOP

lib_com/cldfb.c

+1 −1

Original line number	Diff line number	Diff line
		@@ -1213,7 +1213,7 @@ void cldfbSynthesis_ivas_fx(
		/cplxMult(&iBuffer[2i], &iBuffer[2i+1],-imagBuffer[k][2i], imagBuffer[k][M1-1-2i], rot_vctr_re[i], rot_vctr_im[i]);/
		iBuffer_fx[2 * i] = Msub_32_32( Mpy_32_32( ( L_negate( imagBuffer_fx[k][2 * i] ) ), rot_vctr_re_fx[i] ), imagBuffer_fx[k][( M1 - 1 ) - ( i * 2 )], rot_vctr_im_fx[i] ); // Qx
		move32();
		iBuffer_fx[2 * i + 1] = Madd_32_32( Mpy_32_32( ( L_negate( imagBuffer_fx[k][2 * i] ) ), rot_vctr_im_fx[i] ), imagBuffer_fx[k][( M1 - 1 ) - ( i * 2 )], rot_vctr_re_fx[i] ); // Qx
		iBuffer_fx[2 * i + 1] = Msub_32_32( Mpy_32_32( imagBuffer_fx[k][( M1 - 1 ) - ( i * 2 )], rot_vctr_re_fx[i] ), imagBuffer_fx[k][2 * i], rot_vctr_im_fx[i] ); // Qx
		move32();
		}

lib_com/ivas_stereo_ica_com_fx.c

+24 −38

Original line number	Diff line number	Diff line
		@@ -131,24 +131,23 @@ static void interpTargetChannel_fx(
		Word32 spread_factor2_fx;
		Word64 tempD1_fx, tempD2_fx;

		d = negate( sub( currShift, prevShift ) );
		IF( d >= 0 )
		d = sub( prevShift, currShift );

		IF( d == 0 )
		{
		/* this can happen in DFT->TD switching */
		return;
		}

		signShift = 1;
		move16();
		}
		ELSE

		if ( d < 0 )
		{
		signShift = -1;
		move16();
		}

		IF( d == 0 )
		{
		/* this can happen in DFT->TD switching */
		return;
		}

		N = L_shift_adapt;
		move16();
		Word32 *table_pointer = NULL;
		@@ -207,7 +206,7 @@ static void interpTargetChannel_fx(

		FOR( j = lim1; j <= lim2; j++ )
		{
		ptr2_fx[i] = L_add( Mpy_32_32( win_fx[j * INTERP_FACTOR1 - i], ptr1_fx[j] ), ptr2_fx[i] ); // qsynth
		ptr2_fx[i] = Madd_32_32( ptr2_fx[i], win_fx[j * INTERP_FACTOR1 - i], ptr1_fx[j] ); // qsynth
		move32();
		}
		}
		@@ -225,44 +224,31 @@ static void interpTargetChannel_fx(
		tempD1_fx = W_deposit32_l( table_D1_pointer[abs( d )] ); // Q35
		tempD2_fx = W_mult0_32_32( 3, table_D1_pointer[abs( d )] ); // Q35

		IF( EQ_16( signShift, 1 ) )
		{
		tempF1_fx = ONE_IN_Q12; // Q12
		move32();
		}
		ELSE
		{
		tempF1_fx = -ONE_IN_Q12; // Q12
		move32();
		}
		tempF1_fx = L_sub( imult3216( factor_fx, d ), tempF1_fx ); // Q12
		FOR( k = 0; k < sub( N, 1 ); k++ )
		{
		Word32 local = L_sub( W_extract_l( W_shr( W_mult0_32_32( tempF1_fx, spread_factor2_fx ), 31 ) ), ONE_IN_Q12 ); // Q12
		Word32 sign_local;
		IF( local > 0 )

		if ( EQ_16( signShift, 1 ) )
		{
		sign_local = 1;
		tempF1_fx = ONE_IN_Q12; // Q12
		move32();
		}
		ELSE

		tempF1_fx = L_sub( imult3216( factor_fx, d ), tempF1_fx ); // Q12
		FOR( k = 0; k < N - 1; k++ )
		{
		sign_local = -1;
		move32();
		}
		Word32 local_int = W_extract_l( W_shr( W_abs( local ), 12 ) ); // Q0
		Word32 local = Madd_32_32( -ONE_IN_Q12, tempF1_fx, spread_factor2_fx ); // Q12
		Word32 local_int = L_shr( local, 12 ); // Q0
		Word32 res_a1, res_a2, res_a3;
		Word32 res_b1, res_b2, res_b3;
		Word32 res_c1, res_c2, res_c3;
		Word32 res_d1, res_d2, res_d3;
		Word64 local_int_scaled;
		Word32 local_int_scaled;
		Word64 res_a, res_b, res_c, res_d;
		Word64 tempa, tempb;
		Word64 mult_a_D1, mult_b_D2;
		local_int = W_extract_l( W_mult0_32_32( sign_local, local_int ) ); // Q0
		local_int_scaled = W_deposit32_l( L_shl( local_int, 12 ) ); // Q12
		local_int_scaled = L_shl( local_int, 12 ); // Q12
		lim1 = extract_l( local_int ); // Q0
		IF( W_sub( local_int_scaled, local ) > 0 ) // Q21
		if ( L_sub( local_int_scaled, local ) > 0 ) // Q12
		{
		lim1 = sub( lim1, 1 ); // Q0
		}
		@@ -387,7 +373,7 @@ static void targetCh_AlignStereoDFT_fx(
		}
		FOR( i = 0; i < L_shift_adapt; i++ )
		{
		target_fx[i] = L_add( Mpy_32_32( alpha_fx, fadeInBuff_fx[i] ), Mpy_32_32( L_sub( ONE_IN_Q31, alpha_fx ), fadeOutBuff_fx[i] ) ); // qsynth
		target_fx[i] = Madd_32_32( Mpy_32_32( alpha_fx, fadeInBuff_fx[i] ), L_sub( ONE_IN_Q31, alpha_fx ), fadeOutBuff_fx[i] ); // qsynth
		move32();

		alpha_fx = L_add_sat( alpha_fx, winSlope_fx ); // Q31

lib_com/ivas_tools.c

+81 −43

Original line number	Diff line number	Diff line
		@@ -942,6 +942,12 @@ Word16 matrix_product_mant_exp_fx(
		Word16 *Zp_fx_e = out_e;
		Word16 row, col;
		Word16 x_idx, y_idx;
		Word64 temp;
		Word16 temp_e;
		Word16 prod_e = add( X_fx_e, Y_fx_e );

		Word16 max_exp = -31;
		move16();

		/* Processing */
		test();
		@@ -957,17 +963,28 @@ Word16 matrix_product_mant_exp_fx(
		{
		FOR( i = 0; i < colsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		( *Zp_fx_e ) = 0;
		move16();
		temp = 0;
		move64();

		FOR( k = 0; k < rowsX; ++k )
		{
		x_idx = add( k, imult1616( i, rowsX ) );
		y_idx = add( k, imult1616( j, rowsY ) );
		( Zp_fx ) = BASOP_Util_Add_Mant32Exp( Zp_fx, Zp_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), add( X_fx_e, Y_fx_e ), Zp_fx_e ); /Q31 - Zp_fx_e*/
		move32();
		x_idx = k + i * rowsX;
		y_idx = k + j * rowsY;
		temp = W_mac_32_32( temp, X_fx[x_idx], Y_fx[y_idx] ); // X_fx_e + Y_fx_e
		}
		/* Maximize accumulated value to 32-bit */
		temp_e = W_norm( temp );
		temp = W_shl( temp, temp_e );
		if ( 0 == temp )
		{
		temp_e = prod_e;
		move16();
		}
		*Zp_fx_e = sub( prod_e, temp_e );
		move16();
		( *Zp_fx ) = W_extract_h( temp );
		move32();
		max_exp = s_max( max_exp, *Zp_fx_e ); // Find the max exp
		Zp_fx++;
		Zp_fx_e++;
		}
		@@ -987,17 +1004,27 @@ Word16 matrix_product_mant_exp_fx(
		{
		FOR( i = 0; i < rowsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		( *Zp_fx_e ) = 0;
		move16();
		temp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		x_idx = add( i, imult1616( k, rowsX ) );
		y_idx = add( j, imult1616( k, rowsY ) );
		( Zp_fx ) = BASOP_Util_Add_Mant32Exp( Zp_fx, Zp_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), add( X_fx_e, Y_fx_e ), Zp_fx_e ); /Q31 - Zp_fx_e*/
		move32();
		x_idx = i + k * rowsX;
		y_idx = j + k * rowsY;
		temp = W_mac_32_32( temp, X_fx[x_idx], Y_fx[y_idx] ); // X_fx_e + Y_fx_e
		}
		/* Maximize accumulated value to 32-bit */
		temp_e = W_norm( temp );
		temp = W_shl( temp, temp_e );
		if ( 0 == temp )
		{
		temp_e = prod_e;
		move16();
		}
		*Zp_fx_e = sub( prod_e, temp_e );
		move16();
		( *Zp_fx ) = W_extract_h( temp );
		move32();
		max_exp = s_max( max_exp, *Zp_fx_e ); // Find the max exp
		Zp_fx++;
		Zp_fx_e++;
		}
		@@ -1017,18 +1044,27 @@ Word16 matrix_product_mant_exp_fx(
		{
		FOR( i = 0; i < colsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		( *Zp_fx_e ) = 0;
		move16();
		temp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		x_idx = add( k, imult1616( i, rowsX ) );
		y_idx = add( j, imult1616( k, rowsY ) );
		( Zp_fx ) = BASOP_Util_Add_Mant32Exp( Zp_fx, Zp_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), add( X_fx_e, Y_fx_e ), Zp_fx_e ); /Q31 - Zp_fx_e*/
		move32();
		x_idx = k + i * rowsX;
		y_idx = j + k * rowsY;
		temp = W_mac_32_32( temp, X_fx[x_idx], Y_fx[y_idx] ); // X_fx_e + Y_fx_e
		}

		/* Maximize accumulated value to 32-bit */
		temp_e = W_norm( temp );
		temp = W_shl( temp, temp_e );
		if ( 0 == temp )
		{
		temp_e = prod_e;
		move16();
		}
		*Zp_fx_e = sub( prod_e, temp_e );
		move16();
		( *Zp_fx ) = W_extract_h( temp );
		move32();
		max_exp = s_max( max_exp, *Zp_fx_e ); // Find the max exp
		Zp_fx++;
		Zp_fx_e++;
		}
		@@ -1049,17 +1085,26 @@ Word16 matrix_product_mant_exp_fx(
		{
		FOR( i = 0; i < rowsX; ++i )
		{
		( *Zp_fx ) = 0;
		move32();
		( *Zp_fx_e ) = 0;
		move16();
		temp = 0;
		move64();
		FOR( k = 0; k < colsX; ++k )
		{
		x_idx = add( i, imult1616( k, rowsX ) );
		y_idx = add( k, imult1616( j, rowsY ) );
		( Zp_fx ) = BASOP_Util_Add_Mant32Exp( Zp_fx, Zp_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), add( X_fx_e, Y_fx_e ), Zp_fx_e ); /Q31 - Zp_fx_e*/
		move32();
		x_idx = i + k * rowsX;
		y_idx = k + j * rowsY;
		temp = W_mac_32_32( temp, X_fx[x_idx], Y_fx[y_idx] ); // X_fx_e + Y_fx_e
		}
		/* Maximize accumulated value to 32-bit */
		temp_e = W_norm( temp );
		temp = W_shl( temp, temp_e );
		if ( 0 == temp )
		{
		temp_e = prod_e;
		}
		*Zp_fx_e = sub( prod_e, temp_e );
		move16();
		( *Zp_fx ) = W_extract_h( temp );
		move32();
		max_exp = s_max( max_exp, *Zp_fx_e ); // Find the max exp
		Zp_fx++;
		Zp_fx_e++;
		}
		@@ -1070,18 +1115,11 @@ Word16 matrix_product_mant_exp_fx(
		move16();
		}
		Zp_fx = Z_fx; /Q31 - Zp_fx_e/

		Zp_fx_e = out_e;
		Word16 max_exp = -31;
		move16();
		FOR( j = 0; j < row; ++j )
		{
		FOR( i = 0; i < col; ++i )
		{
		max_exp = s_max( max_exp, *Zp_fx_e );
		Zp_fx_e++;
		}
		}
		Zp_fx_e = out_e;


		*Z_fx_e = max_exp;
		move16();
		FOR( j = 0; j < row; ++j )

lib_com/options.h

+2 −1

Original line number	Diff line number	Diff line
		@@ -152,5 +152,6 @@
		#define FIX_ISSUE_1214 /* Ittiam: Fix for issue 1214: Energy leakage in IGF tiles for MDCT-stereo @64kbps SWB*/
		#define FIX_881_HILBERT_FILTER /* VA: improve the precision of the Hilbert filter to remove 2kHz unwanted tone */
		#define FIX_ISSUE_1245 /* Ittiam: Fix for issue 1245: Basop Encoder: Audible noise for silent Stereo input DTX on @24.4 kbps, @32 kbps*/
		#endif
		#define FIX_MINOR_SVD_WMOPS_MR1010X /* FhG: Minor WMOPS tuning, bit-exact to previous version, saves about 8.2 WMOPS for MR1010 */
		#define SVD_WMOPS_OPT /* Ittiam : SVD related optimizations */
		#endif

lib_dec/dec_tcx.c

+23 −12

Original line number	Diff line number	Diff line
		@@ -237,6 +237,7 @@ void decoder_tcx_imdct_fx(
		Word16 q_a_itf = 15;
		Word16 x_e = sub( 31, q_x );
		move16();
		Word16 shift_q = sub( q_x, q_win );

		/-----------------------------------------------------------------
		* Initializations
		@@ -364,9 +365,10 @@ void decoder_tcx_imdct_fx(

		IF( EQ_16( st->element_mode, IVAS_CPE_MDCT ) )
		{
		Word16 copy_len = s_min( L_FRAME48k, s_max( L_spec, s_max( L_frame, L_frameTCX ) ) );
		set32_fx( x_tmp_fx, 0, L_FRAME_PLUS );
		Copy32( x_fx, x_tmp_fx, s_min( L_FRAME48k, s_max( L_spec, s_max( L_frame, L_frameTCX ) ) ) ); // q_x
		Copy32( x_fx, xn_bufFB_fx, s_min( L_FRAME48k, s_max( L_spec, s_max( L_frame, L_frameTCX ) ) ) ); // q_x
		Copy32( x_fx, x_tmp_fx, copy_len ); // q_x
		Copy32( x_fx, xn_bufFB_fx, copy_len ); // q_x
		}
		ELSE IF( ( st->element_mode == EVS_MONO ) )
		{
		@@ -374,8 +376,9 @@ void decoder_tcx_imdct_fx(
		}
		ELSE
		{
		Copy32( x_fx, x_tmp_fx, s_max( L_spec, s_max( L_frame, L_frameTCX ) ) ); // q_x
		Copy32( x_fx, xn_bufFB_fx, s_max( L_spec, s_max( L_frame, L_frameTCX ) ) ); // q_x
		Word16 copy_len = s_max( L_spec, s_max( L_frame, L_frameTCX ) );
		Copy32( x_fx, x_tmp_fx, copy_len ); // q_x
		Copy32( x_fx, xn_bufFB_fx, copy_len ); // q_x
		}

		IF( ( st->igf != 0 ) )
		@@ -416,24 +419,29 @@ void decoder_tcx_imdct_fx(

		FOR( Word16 ind = 0; ind < L_MDCT_OVLP_MAX + L_FRAME_PLUS + L_MDCT_OVLP_MAX; ind++ )
		{
		xn_bufFB_fx_16[ind] = extract_l( L_shr( xn_bufFB_fx[ind], sub( q_x, q_win ) ) ); // q_x
		xn_bufFB_fx_16[ind] = extract_l( L_shr( xn_bufFB_fx[ind], shift_q ) ); // q_x
		move16();
		}

		Word16 ratio_e;
		Word16 ratio = BASOP_Util_Divide1616_Scale( L_frameTCX_glob, L_frame_glob, &ratio_e ); // Q = 15-ratio_e. * FSCALE_DENOM is (1 << 9)
		ratio = shr( ratio, sub( 6, ratio_e ) );

		IF( st->element_mode != EVS_MONO )
		{
		IMDCT_ivas_fx( x_tmp_fx, q_x, hTcxDec->syn_OverlFB, hTcxDec->syn_Overl_TDACFB, xn_bufFB_fx_16, hTcxCfg->tcx_aldo_window_1_FB, hTcxCfg->tcx_aldo_window_1_FB_trunc, hTcxCfg->tcx_aldo_window_2_FB,
		hTcxCfg->tcx_mdct_window_halfFB, hTcxCfg->tcx_mdct_window_minimumFB, hTcxCfg->tcx_mdct_window_transFB, hTcxCfg->tcx_mdct_window_half_lengthFB, hTcxCfg->tcx_mdct_window_min_lengthFB, index,
		kernelType, left_rect, tcx_offsetFB, overlapFB, L_frameTCX, L_frameTCX, max( L_frameTCX, L_spec ) >> 1, L_frameTCX_glob, frame_cnt, bfi, st->hHQ_core->old_out_fx, 1, st, FSCALE_DENOM * L_frameTCX_glob / L_frame_glob, acelp_zir_fx, q_win );
		kernelType, left_rect, tcx_offsetFB, overlapFB, L_frameTCX, L_frameTCX, shr( max( L_frameTCX, L_spec ), 1 ), L_frameTCX_glob, frame_cnt, bfi, st->hHQ_core->old_out_fx, 1, st, ratio, acelp_zir_fx, q_win );
		}
		ELSE
		{

		IMDCT_ivas_fx( x_fx, q_x, hTcxDec->syn_OverlFB, hTcxDec->syn_Overl_TDACFB, xn_bufFB_fx_16, hTcxCfg->tcx_aldo_window_1_FB, hTcxCfg->tcx_aldo_window_1_FB_trunc, hTcxCfg->tcx_aldo_window_2_FB, hTcxCfg->tcx_mdct_window_halfFB, hTcxCfg->tcx_mdct_window_minimumFB, hTcxCfg->tcx_mdct_window_transFB, hTcxCfg->tcx_mdct_window_half_lengthFB, hTcxCfg->tcx_mdct_window_min_lengthFB, index,
		kernelType, left_rect, tcx_offsetFB, overlapFB, L_frameTCX, L_frameTCX, shr( s_max( L_frameTCX, L_spec ), 1 ), L_frameTCX_glob, frame_cnt, bfi, st->hHQ_core->old_out_fx, 1, st, FSCALE_DENOM * L_frameTCX_glob / L_frame_glob, acelp_zir_fx, q_win );
		kernelType, left_rect, tcx_offsetFB, overlapFB, L_frameTCX, L_frameTCX, shr( s_max( L_frameTCX, L_spec ), 1 ), L_frameTCX_glob, frame_cnt, bfi, st->hHQ_core->old_out_fx, 1, st, ratio, acelp_zir_fx, q_win );
		}
		FOR( Word16 ind = 0; ind < L_MDCT_OVLP_MAX + L_FRAME_PLUS + L_MDCT_OVLP_MAX; ind++ )
		{
		xn_bufFB_fx[ind] = L_shl( xn_bufFB_fx_16[ind], sub( q_x, q_win ) ); // Q_x
		xn_bufFB_fx[ind] = L_shl( L_deposit_l( xn_bufFB_fx_16[ind] ), shift_q ); // Q_x
		}

		IF( ( bfi == 0 ) )
		@@ -453,19 +461,22 @@ void decoder_tcx_imdct_fx(

		IF( EQ_16( st->element_mode, IVAS_CPE_MDCT ) )
		{
		res_m = BASOP_Util_Divide1616_Scale( L_frame_glob, L_FRAME, &res_e );
		st->old_fpitch = L_shl( Mpy_32_16_1( st->old_fpitch, res_m ), res_e );
		// Using sat as a single instruction shifts and extracts
		st->old_fpitch = W_shl_sat_l( W_mult0_32_32( st->old_fpitch, L_frame_glob ), -8 ); // Divide by 256 ==> SHR by 8
		move32();
		}

		IF( GT_16( st->element_mode, EVS_MONO ) )
		{
		res_m = BASOP_Util_Divide1616_Scale( L_frameTCX_glob, L_frame_glob, &res_e );
		st->old_fpitchFB = L_shl( Mpy_32_16_1( st->old_fpitch, res_m ), res_e );
		move32();
		}
		ELSE
		{
		res_m = BASOP_Util_Divide1616_Scale( L_frameTCX, L_frame, &res_e );
		st->old_fpitchFB = L_shl( Mpy_32_16_1( st->old_fpitch, res_m ), res_e );
		move32();
		}
		}

		@@ -475,7 +486,7 @@ void decoder_tcx_imdct_fx(
		Copy( xn_buf_fx + L_frame, hTcxDec->syn_Overl, overlap ); // Q(-2)
		FOR( Word16 ind = 0; ind < overlapFB; ind++ )
		{
		hTcxDec->syn_OverlFB[ind] = (Word16) L_shr( xn_bufFB_fx[( ind + L_frameTCX )], sub( q_x, q_win ) ); // q_x
		hTcxDec->syn_OverlFB[ind] = extract_l( L_shr( xn_bufFB_fx[( ind + L_frameTCX )], shift_q ) ); // q_x
		}
		}

		@@ -483,7 +494,7 @@ void decoder_tcx_imdct_fx(
		Copy( xn_buf_fx + sub( shr( overlap, 1 ), tcx_offset ), synth_fx, L_frame_glob ); // Q(-2)
		FOR( Word16 ind = 0; ind < L_frameTCX_glob; ind++ )
		{
		synthFB_fx[ind] = (Word16) L_shr( xn_bufFB_fx[( ind + ( ( overlapFB >> 1 ) - tcx_offsetFB ) )], sub( q_x, q_win ) ); // q_x
		synthFB_fx[ind] = extract_l( L_shr( xn_bufFB_fx[( ind + ( ( overlapFB >> 1 ) - tcx_offsetFB ) )], shift_q ) ); // q_x
		}

Admin message