cleanup useless speedup macros (e957073d) · Commits · SA4 / Audio / IVAS BASOP

lib_rend/ivas_dirac_dec_binaural_functions_fx.c

+14 −108

Original line number	Diff line number	Diff line
		@@ -47,22 +47,16 @@

		// MHZ NUMBERS:
		// NULL: 178.407
		// ALL: 169.499
		// ALL: 169.499 77 (170.650 wo 17)


		#define FIX_1326_SPEEDUP_01 // optimize matrixT1mul->eig2x2_fx // .4 WMOPS --> USE
		#define FIX_1326_SPEEDUP_02 // speedup eig2x2_fx // .3 WMOPS --> USE
		#define FIX_1326_SPEEDUP_03 // speedup eig2x2_fx // .1 WMOPS --> USE
		#define FIX_1326_SPEEDUP_04 // speedup eig2x2_fx // .2 WMOPS --> USE
		#define FIX_1326_SPEEDUP_08 // "-" // 3.0 WMOPS //small diffs --> USE
		#define FIX_1326_SPEEDUP_09 // tiny speedup // .1 WMOPS pipe 48851 fails --> DONTUSEYET
		#define FIX_1326_SPEEDUP_10 // tiny speedup // .1 WMOPS pipe 48851 fails --> DONTUSEYET
		#define FIX_1326_SPEEDUP_11 // tiny speedup // .1 WMOPS pipe 48851 fails --> DONTUSEYET
		#define FIX_1326_SPEEDUP_13 // replace div/sqrt->isqrt // 2.9 WMOPS --> USE

		#define FIX_1326_SPEEDUP_15 // replace Ladd(Mpy) -> Madd // .1 WMOPS --> USE
		#define FIX_1326_SPEEDUP_16 // tiny speedup like 04 // .2 WMOPS --> USE
		//#define FIX_1326_SPEEDUP_17 // use 1/x // 1.25WMOPS --> USE
		#define FIX_1326_SPEEDUP_18 // structural speedup // 1 WMOPS --> USE

		Word16 slot_fx[4] = { 32767, 16384, 10922, 8192 };
		@@ -1317,7 +1311,7 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		}
		pop_wmops(); /push_wmops( "IDRBCM apply EQ_low" );/

		push_wmops( "IDRBCM target matrix" );
		push_wmops( "IDRBCM target matrix (IDRBCMtm)" );
		/* Determine target covariance matrix containing target binaural properties */
		FOR( bin = 0; bin < nBins; bin++ )
		{
		@@ -1343,6 +1337,7 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		meanEnePerCh_fx = Mpy_32_32( hDiracDecBin->earlyPartEneCorrection_fx[bin], subFrameTotalEne_fx[bin] ); // Q( q_meanEnePerCh )
		q_meanEnePerCh = add( sub( q_earlyPartEneCorrection, subFrameTotalEne_e[bin] ), 1 ); // q_earlyPartEneCorrection + 31 - subFrameTotalEne_e[bin] - 31 + Q1(0.5f)
		/* Determine direct part target covariance matrix (for 1 or 2 directions) */
		push_wmops( "IDRBCMtm LOOP1" );
		FOR( dirIndex = 0; dirIndex < hSpatParamRendCom->numSimultaneousDirections; dirIndex++ )
		{
		Word16 aziDeg, eleDeg;
		@@ -1421,6 +1416,7 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		diffuseness_fx = 0;
		move32();
		}

		IF( isIsmDirection )
		{
		/* Objects cause lesser decorrelation reduction, to avoid removing all decorrelation when only objects are present */
		@@ -1430,7 +1426,6 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		{
		diffusenessValForDecorrelationReduction_fx = L_sub( diffusenessValForDecorrelationReduction_fx, ratio_fx ); /Q30/
		}

		IF( separateCenterChannelRendering )
		{
		/* In masa + mono rendering mode, the center directions originate from phantom sources, so the
		@@ -1477,14 +1472,12 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		Word32 hrtfEneCenter_fx, hrtfEneSides_fx, hrtfEneRealized_fx;
		Word16 eneCorrectionFactor_fx, eneCorrectionFactor_e;
		Word16 w1_fx, w2_fx, w3_fx, eq_fx;
		#ifdef FIX_1326_SPEEDUP_15
		hrtfEneCenter_fx = Madd_32_32( Madd_32_32( Madd_32_32( Mpy_32_32( rRealp_fx, rRealp_fx ), rImagp_fx, rImagp_fx ), lImagp_fx, lImagp_fx ), lRealp_fx, lRealp_fx ); // Q25
		#else

		hrtfEneCenter_fx = L_add( Mpy_32_32( lRealp_fx, lRealp_fx ), // Q25
		L_add( Mpy_32_32( lImagp_fx, lImagp_fx ), // Q25
		L_add( Mpy_32_32( rRealp_fx, rRealp_fx ), // Q25
		Mpy_32_32( rImagp_fx, rImagp_fx ) ) ) ); // Q25
		#endif

		/* Spread coherence is synthesized as coherent sources at 30 degree horizontal spacing.
		* The following formulas determine the gains for these sources.
		* spreadCoh = 0: Only panning
		@@ -1513,14 +1506,12 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric

		/* Apply the gain for the left source of the three coherent sources */
		getDirectPartGains_fx( bin, add( aziDeg, 30 ), eleDeg, &lRealpTmp_fx, &lImagpTmp_fx, &rRealpTmp_fx, &rImagpTmp_fx, hDiracDecBin->renderStereoOutputInsteadOfBinaural, Rmat_fx, &gainCache[gainCacheBaseIndex + 1], isHeadtracked );
		#ifdef FIX_1326_SPEEDUP_15
		hrtfEneSides_fx = Madd_32_32( Madd_32_32( Madd_32_32( Mpy_32_32( rRealpTmp_fx, rRealpTmp_fx ), rImagpTmp_fx, rImagpTmp_fx ), lImagpTmp_fx, lImagpTmp_fx ), lRealpTmp_fx, lRealpTmp_fx ); // Q25
		#else

		hrtfEneSides_fx = L_add( Mpy_32_32( lRealpTmp_fx, lRealpTmp_fx ), // Q25
		L_add( Mpy_32_32( lImagpTmp_fx, lImagpTmp_fx ), // Q25
		L_add( Mpy_32_32( rRealpTmp_fx, rRealpTmp_fx ), // Q25
		Mpy_32_32( rImagpTmp_fx, rImagpTmp_fx ) ) ) ); // Q25
		#endif

		lRealp_fx = L_add( lRealp_fx, Mpy_32_32( sidesMul_fx, lRealpTmp_fx ) ); // Q25
		lImagp_fx = L_add( lImagp_fx, Mpy_32_32( sidesMul_fx, lImagpTmp_fx ) ); // Q25
		rRealp_fx = L_add( rRealp_fx, Mpy_32_32( sidesMul_fx, rRealpTmp_fx ) ); // Q25
		@@ -1608,21 +1599,12 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		move16();
		}

		#ifdef FIX_1326_SPEEDUP_15
		hrtfEne_fx[0] = Madd_32_32( Mpy_32_32( lRealp_fx, lRealp_fx ), lImagp_fx, lImagp_fx ); // Q( 2*q_lr - 31 )
		hrtfEne_fx[1] = Madd_32_32( Mpy_32_32( rRealp_fx, rRealp_fx ), rImagp_fx, rImagp_fx ); // Q( 2*q_lr - 31 )
		move32();
		move32();
		hrtfCrossRe_fx = Madd_32_32( Mpy_32_32( lRealp_fx, rRealp_fx ), lImagp_fx, rImagp_fx ); // Q( 2*q_lr - 31 )
		hrtfCrossIm_fx = Madd_32_32( Mpy_32_32( -lImagp_fx, rRealp_fx ), lRealp_fx, rImagp_fx ); // Q( 2*q_lr - 31 )
		#else
		hrtfEne_fx[0] = L_add( Mpy_32_32( lRealp_fx, lRealp_fx ), Mpy_32_32( lImagp_fx, lImagp_fx ) ); // Q( 2*q_lr - 31 )
		hrtfEne_fx[1] = L_add( Mpy_32_32( rRealp_fx, rRealp_fx ), Mpy_32_32( rImagp_fx, rImagp_fx ) ); // Q( 2*q_lr - 31 )
		move32();
		move32();
		hrtfCrossRe_fx = L_add( Mpy_32_32( lRealp_fx, rRealp_fx ), Mpy_32_32( lImagp_fx, rImagp_fx ) ); // Q( 2*q_lr - 31 )
		hrtfCrossIm_fx = L_add( Mpy_32_32( -lImagp_fx, rRealp_fx ), Mpy_32_32( lRealp_fx, rImagp_fx ) ); // Q( 2*q_lr - 31 )
		#endif

		/* Add direct part (1 or 2) covariance matrix */
		dirEne_fx = Mpy_32_32( ratio_fx, meanEnePerCh_fx ); // Q(q_meanEnePerCh - 1)
		@@ -1639,6 +1621,7 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		move32();
		move32();
		}
		pop_wmops(); //push_wmops( "IDRBCMtm LOOP1" );

		/* Add diffuse / ambient part covariance matrix */
		diffuseness_fx = L_max( 0, diffuseness_fx ); // Q30
		@@ -1697,11 +1680,7 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		}
		ELSE
		{
		#ifdef FIX_1326_SPEEDUP_15
		hDiracDecBin->ChCrossReOut_fx[bin] = BASOP_Util_Add_Mant32Exp( hDiracDecBin->ChCrossReOut_fx[bin], hDiracDecBin->ChCrossReOut_e[bin], Mpy_32_32( Madd_32_16( L_shl( surCoh_fx, 16 ), hDiracDecBin->diffuseFieldCoherence_fx[bin], sub( 32767, surCoh_fx ) ), diffEne_fx ), sub( 31, q_diffEne ), &hDiracDecBin->ChCrossReOut_e[bin] );
		#else
		hDiracDecBin->ChCrossReOut_fx[bin] = BASOP_Util_Add_Mant32Exp( hDiracDecBin->ChCrossReOut_fx[bin], hDiracDecBin->ChCrossReOut_e[bin], Mpy_32_32( L_add( Mpy_32_16_1( hDiracDecBin->diffuseFieldCoherence_fx[bin], sub( 32767, surCoh_fx ) ), L_shl( surCoh_fx, 16 ) ), diffEne_fx ), sub( 31, q_diffEne ), &hDiracDecBin->ChCrossReOut_e[bin] );
		#endif
		}
		move32();
		}
		@@ -1717,7 +1696,7 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric
		hDiracDecBin->frameMeanDiffuseness_fx[bin] = L_shl( frameMeanDiffuseness, sub( exp, 2 ) ); // Q29
		move32();
		}
		pop_wmops(); /push_wmops( "IDRBCM target matrix" );/
		pop_wmops(); /push_wmops( "IDRBCM target matrix (IDRBCMtm)" );;/

		test();
		/* Temporal IIR-type smoothing of covariance matrices. Also apply encoding quality based smoothing factor. */
		@@ -2165,14 +2144,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx(
		exp = sub( get_min_scalefactor( resultMtxRe_fx[0][0], resultMtxRe_fx[1][1] ), 2 );
		tmp2 = L_add( L_shl( resultMtxRe_fx[0][0], exp ), L_shl( resultMtxRe_fx[1][1], exp ) );
		q_tmp2 = add( q_res, exp );
		#ifdef FIX_1326_SPEEDUP_11
		{
		Word16 shift1 = s_max( 0, sub( q_tmp2, q_CrEne ) );
		Word16 shift2 = s_max( 0, sub( q_CrEne, q_tmp2 ) );
		realizedOutputEne_fx = L_add( L_shr( tmp1, shift2 ), L_shr( tmp2, shift1 ) );
		q_realizedOutputEne = s_min( q_CrEne, q_tmp2 );
		}
		#else

		IF( LT_16( q_CrEne, q_tmp2 ) )
		{
		realizedOutputEne_fx = L_add( tmp1, L_shr( tmp2, sub( q_tmp2, q_CrEne ) ) );
		@@ -2185,7 +2157,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx(
		q_realizedOutputEne = q_tmp2;
		move16();
		}
		#endif

		exp = sub( get_min_scalefactor( hDiracDecBin->ChEneOut_fx[0][bin], hDiracDecBin->ChEneOut_fx[1][bin] ), 1 );
		targetOutputEne_fx = L_add( L_shl( hDiracDecBin->ChEneOut_fx[0][bin], exp ), L_shl( hDiracDecBin->ChEneOut_fx[1][bin], exp ) );
		q_targetOutputEne = add( hDiracDecBin->q_ChEneOut, exp );
		@@ -3351,21 +3323,7 @@ static void eig2x2_fx(
		#endif

		/* Numeric case, when input is near an identity matrix with a gain */
		#ifdef FIX_1326_SPEEDUP_03
		tmp1 = Mpy_32_32( 2147484, add_fx ); // 2147484 = 1e-3f in Q31

		IF( LT_32( pm_fx, L_shl_sat( tmp1, sub( q_tmp1, q_tmp2 ) ) ) )
		{
		Ure_fx[0][0] = ONE_IN_Q30;
		move32();
		Ure_fx[1][1] = ONE_IN_Q30;
		move32();
		*q_U = Q30;
		move16();

		return;
		}
		#else
		tmp1 = Mpy_32_32( 2147484, add_fx ); // 2147484 = 1e-3f in Q31

		IF( LT_16( q_tmp1, q_tmp2 ) )
		@@ -3396,7 +3354,6 @@ static void eig2x2_fx(
		return;
		}
		}
		#endif

		q_U_1 = 0;
		q_U_2 = 0;
		@@ -3450,11 +3407,7 @@ static void eig2x2_fx(
		#endif

		#if 1
		#ifdef FIX_1326_SPEEDUP_17
		tmp2 = BASOP_Util_Divide3232_Scale_cadence_1( ONE_IN_Q30, tmp3, &exp );
		#else
		tmp2 = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, tmp3, &exp );
		#endif
		exp = sub( exp, sub( Q30, sub( 31, exp_tmp3 ) ) );
		normVal_fx = Sqrt32( tmp2, &exp ); // q_tmp2
		q_tmp2 = sub( 31, exp );
		@@ -3542,11 +3495,7 @@ static void eig2x2_fx(
		#endif

		#if 1
		#ifdef FIX_1326_SPEEDUP_17
		tmp2 = BASOP_Util_Divide3232_Scale_cadence_1( ONE_IN_Q30, tmp3, &exp );
		#else
		tmp2 = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, tmp3, &exp );
		#endif
		exp = sub( exp, sub( Q30, sub( 31, exp_tmp3 ) ) );
		normVal_fx = Sqrt32( tmp2, &exp ); // q_tmp2
		q_tmp2 = sub( 31, exp );
		@@ -4214,13 +4163,7 @@ static void formulate2x2MixingMatrix_fx(
		}
		ELSE
		{
		push_wmops( "formulate2x2MixingMatrix Division" );
		#ifdef FIX_1326_SPEEDUP_17
		maxEneDiv_fx = BASOP_Util_Divide3232_Scale_cadence_1( ONE_IN_Q30, maxEne_fx, &exp );
		#else
		maxEneDiv_fx = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, maxEne_fx, &exp );
		#endif
		pop_wmops(); /push_wmops( "formulate2x2MixingMatrix Division" )/
		q_maxEneDiv = add( sub( 31, exp ), sub( Q30, q_maxEne ) );
		}
		exp = norm_l( maxEneDiv_fx );
		@@ -4298,9 +4241,7 @@ static void formulate2x2MixingMatrix_fx(
		ELSE
		{
		temp = BASOP_Util_Add_Mant32Exp( temp, sub( 31, q_ein ), EPSILON_MANT, EPSILON_EXP, &exp_temp );
		push_wmops( "formulate2x2MixingMatrix Division" );
		temp = BASOP_Util_Divide3232_Scale_cadence( E_out1, temp, &exp );
		pop_wmops(); /push_wmops( "formulate2x2MixingMatrix Division" )/
		exp = sub( exp, sub( q_eout, sub( 31, exp_temp ) ) );
		Ghat_fx[0] = Sqrt32( temp, &exp ); // Q = 31 - exp
		}
		@@ -4335,9 +4276,7 @@ static void formulate2x2MixingMatrix_fx(
		a++;
		}
		temp = BASOP_Util_Add_Mant32Exp( temp, sub( 31, q_ein ), EPSILON_MANT, EPSILON_EXP, &exp_temp );
		push_wmops( "formulate2x2MixingMatrix Division" );
		temp = BASOP_Util_Divide3232_Scale_cadence( E_out2, temp, &exp1 );
		pop_wmops(); /push_wmops( "formulate2x2MixingMatrix Division" )/
		exp1 = sub( exp1, sub( q_eout, sub( 31, exp_temp ) ) );
		Ghat_fx[1] = Sqrt32( temp, &exp1 ); // Q = 31 - exp1
		}
		@@ -4466,24 +4405,19 @@ static void formulate2x2MixingMatrix_fx(
		#endif
		pop_wmops(); /push_wmops( "oPtoA MT1M" );/


		IF( D_fx[0] == 0 )
		{
		temp = ONE_DIV_EPSILON_MANT; /* Result of 1.0/eps with full precision */
		move32();
		exp = ONE_DIV_EPSILON_EXP;
		move16();
		#endif
		}
		ELSE
		{
		push_wmops( "formulate2x2MixingMatrix Division" );
		#ifdef FIX_1326_SPEEDUP_17
		temp = BASOP_Util_Divide3232_Scale_cadence_1( ONE_IN_Q30, D_fx[0], &exp );
		#else

		temp = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, D_fx[0], &exp );
		#endif
		exp = sub( exp, sub( Q30, q_D ) );
		pop_wmops(); /push_wmops( "formulate2x2MixingMatrix Division" )/
		}
		div_fx[0] = Sqrt32( temp, &exp ); // Q = 31 - exp
		move32();
		@@ -4514,13 +4448,7 @@ static void formulate2x2MixingMatrix_fx(
		}
		ELSE
		{
		push_wmops( "formulate2x2MixingMatrix Division" );
		#ifdef FIX_1326_SPEEDUP_17
		temp = BASOP_Util_Divide3232_Scale_cadence_1( ONE_IN_Q30, D_fx[1], &exp1 );
		#else
		temp = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, D_fx[1], &exp1 );
		#endif
		pop_wmops(); /push_wmops( "formulate2x2MixingMatrix Division" )/
		exp1 = sub( exp1, sub( Q30, q_D ) );
		}
		div_fx[1] = Sqrt32( temp, &exp1 ); // Q = 31 - exp1
		@@ -4535,17 +4463,6 @@ static void formulate2x2MixingMatrix_fx(


		// 1310720000 = 10,000.0f in Q17
		#ifdef FIX_1326_SPEEDUP_09
		{
		Word16 shift1 = s_max( sub( Q17, q_div ), 0 );
		Word16 shift2 = s_max( sub( q_div, Q17 ), 0 );

		div_fx[0] = L_min( L_shr( 1310720000, shift1 ), L_shr( div_fx[0], shift2 ) ); // q_div
		move32();
		div_fx[1] = L_min( L_shr( 1310720000, shift1 ), L_shr( div_fx[1], shift2 ) ); // q_div
		move32();
		}
		#else
		IF( LT_16( q_div, Q17 ) )
		{
		div_fx[0] = L_min( L_shr( 1310720000, sub( Q17, q_div ) ), div_fx[0] ); // q_div
		@@ -4562,7 +4479,6 @@ static void formulate2x2MixingMatrix_fx(
		q_div = Q17;
		move16();
		}
		#endif

		matrixMul_fx( Are_fx, Aim_fx, &q_A, Ure_fx, Uim_fx, &q_U, tmpRe_fx, tmpIm_fx, &q_temp );

		@@ -4579,11 +4495,7 @@ static void formulate2x2MixingMatrix_fx(
		W_tmp = W_mult0_32_32( tmpRe_fx[chA][chB], div_fx[chB] );
		IF( W_tmp != 0 )
		{
		#ifdef FIX_1326_SPEEDUP_10
		hdrm_re[chA][chB] = W_norm( W_tmp );
		#else
		hdrm_re[chA][chB] = sub( W_norm( W_tmp ), 0 );
		#endif
		move16();
		W_tmp = W_shl( W_tmp, hdrm_re[chA][chB] );
		tmpRe_fx[chA][chB] = W_extract_h( W_tmp );
		@@ -4710,13 +4622,7 @@ static void formulate2x2MixingMatrix_fx(
		{
		Word16 Pre_shift, Pim_shift;
		temp = BASOP_Util_Add_Mant32Exp( Sx_fx[chB], sub( 31, q_Sx ), EPSILON_MANT, EPSILON_EXP, &exp_temp );
		push_wmops( "formulate2x2MixingMatrix Division" );
		#ifdef FIX_1326_SPEEDUP_17
		temp = BASOP_Util_Divide3232_Scale_cadence_1( ONE_IN_Q30, temp, &exp );
		#else
		temp = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, temp, &exp );
		#endif
		pop_wmops(); /push_wmops( "formulate2x2MixingMatrix Division" )/
		q_temp = add( sub( sub( q_P, exp ), sub( 31, Q30 ) ), exp_temp );

		Pre_shift = norm_l( Pre_fx[0][chB] );