rename optimiztion macros, move macros to options.h (8cc9439e) · Commits · SA4 / Audio / IVAS BASOP

lib_com/options.h

+3 −0

Original line number	Diff line number	Diff line
		@@ -175,4 +175,7 @@
		#define NONBE_FIX_1402_WAVEADJUST /* VA: BASOP iisue 1402: fix waveform adjustment decoder PLC */
		#define FIX_ISSUE_1376 /* VA: Fix for issue 1376 (issue with GSC excitation) */
		#define OPT_SBA_AVOID_SPAR_RESCALE /* Optimization made to spar decoder and IGF */
		#define FIX_1326_SUBSTITUTE_CMPMANT32EXP /* FhG: Minor WMOPS tuning*/
		#define FIX_1326_SUBSTITUTE_DIV_SQRT_IOSQRT /* FhG: WMOPS tuning */
		#define FIX_1326_SPEEDUP_eig2x2_fx /* FhG: Minor WMOPS tuning*/
		#endif

lib_rend/ivas_dirac_dec_binaural_functions_fx.c

+8 −132

Original line number	Diff line number	Diff line
		@@ -46,18 +46,6 @@
		#include "wmc_auto.h"

		// MHZ NUMBERS:
		// NULL: 178.407
		// 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_04 // speedup eig2x2_fx // .2 WMOPS --> USE
		#define FIX_1326_SPEEDUP_08 // "-" // 3.0 WMOPS //small diffs --> USE
		#define FIX_1326_SPEEDUP_13 // replace div/sqrt->isqrt // 2.9 WMOPS --> USE

		#define FIX_1326_SPEEDUP_16 // tiny speedup like 04 // .2 WMOPS --> USE
		#define FIX_1326_SPEEDUP_18 // structural speedup // 1 WMOPS --> USE

		Word16 slot_fx[4] = { 32767, 16384, 10922, 8192 };

		@@ -1892,11 +1880,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx(
		move16();
		Word32 tmp1, tmp2, res1, res2;
		Word16 q_tmp1, q_tmp2, q_realizedOutputEne, q_targetOutputEne, q_missingOutputEne, q_gain;
		#ifdef FIX_1326_SPEEDUP_13
		Word16 exp1, q_processMtx_bin, q_processMtxDec_bin;
		#else
		Word16 exp1, exp2, q_processMtx_bin, q_processMtxDec_bin;
		#endif

		CrEneL_fx = 0;
		move32();
		@@ -2199,13 +2183,12 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx(

		tmp1 = BASOP_Util_Add_Mant32Exp( tmp2, sub( 31, q_tmp2 ), missingOutputEne_fx, sub( 31, q_missingOutputEne ), &exp1 );

		#ifdef FIX_1326_SPEEDUP_13
		#ifdef FIX_1326_SUBSTITUTE_DIV_SQRT_IOSQRT
		{
		Word16 exp_temp;
		tmp2 = BASOP_Util_Add_Mant32Exp( tmp2, sub( 31, q_tmp2 ), EPSILON_MANT, EPSILON_EXP, &exp_temp );
		tmp2 = ISqrt32( tmp2, &exp_temp );
		tmp2 = BASOP_Util_Add_Mant32Exp( tmp2, sub( 31, q_tmp2 ), EPSILON_MANT, EPSILON_EXP, &exp2 );
		tmp2 = ISqrt32( tmp2, &exp2 );
		gain_fx = Mpy_32_32( tmp2, Sqrt32( tmp1, &exp1 ) );
		q_gain = sub( 31, add( exp_temp, exp1 ) );
		q_gain = sub( 31, add( exp2, exp1 ) );
		}
		#else
		{
		@@ -3317,7 +3300,7 @@ static void eig2x2_fx(
		/* Numeric case, when input is practically zeros */
		// IF( D_fx[0] < EPSILON_FX )

		#ifdef FIX_1326_SPEEDUP_02
		#ifdef FIX_1326_SUBSTITUTE_CMPMANT32EXP
		IF( LT_32( L_shl_sat( D_fx[0], sub( sub( 31, *q_D ), EPSILON_EXP ) ), EPSILON_MANT ) )
		{
		Ure_fx[0][0] = ONE_IN_Q31;
		@@ -3406,27 +3389,10 @@ static void eig2x2_fx(
		tmp2 = Mpy_32_32( s_fx, s_fx );
		q_tmp2 = sub( add( q_tmp1, q_tmp1 ), 31 );


		#ifdef FIX_1326_SPEEDUP_16

		{
		Word16 tmp2_exp;
		Word32 eps_tmp;
		tmp2 = BASOP_Util_Add_Mant32Exp( crossSquare_fx, sub( 31, q_crossSquare ), tmp2, sub( 31, q_tmp2 ), &tmp2_exp );

		// Add epsilon if relevant
		eps_tmp = L_shl_sat( epsilon_mant, sub( epsilon_exp, tmp2_exp ) );
		tmp3 = L_add( L_shr( tmp2, 1 ), L_shr( eps_tmp, 1 ) );

		exp_tmp3 = add( tmp2_exp, 1 );
		}
		#else
		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 );
		#endif

		#if 1
		tmp2 = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, tmp3, &exp );
		exp = sub( exp, sub( Q30, sub( 31, exp_tmp3 ) ) );
		@@ -3499,22 +3465,10 @@ static void eig2x2_fx(
		q_tmp2 = sub( add( q_tmp1, q_tmp1 ), 31 );


		#ifdef FIX_1326_SPEEDUP_04
		Word16 exp_tmp2;
		Word32 eps_tmp;

		tmp2 = BASOP_Util_Add_Mant32Exp( crossSquare_fx, sub( 31, q_crossSquare ), tmp2, sub( 31, q_tmp2 ), &exp_tmp2 );
		eps_tmp = L_shl_sat( epsilon_mant, sub( epsilon_exp, exp_tmp2 ) );

		tmp3 = L_add( L_shr( tmp2, 1 ), L_shr( eps_tmp, 1 ) ); // Add Epsilon if relevant

		exp_tmp3 = add( exp_tmp2, 1 );
		#else
		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 );
		#endif

		#if 1
		tmp2 = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, tmp3, &exp );
		@@ -3815,7 +3769,6 @@ static void matrixMul_fx(
		return;
		}

		#ifndef FIX_1326_SPEEDUP_01
		static void matrixTransp1Mul_fx(
		Word32 Are_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], /q_A/
		Word32 Aim_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], /q_A/
		@@ -3929,7 +3882,6 @@ static void matrixTransp1Mul_fx(

		return;
		}
		#endif /FIX_1326_SPEEDUP_01/

		static void matrixTransp2Mul_fx(
		Word32 Are_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], /q_A/
		@@ -4655,80 +4607,10 @@ static void formulate2x2MixingMatrix_fx(
		/* Find nearest orthonormal matrix P to A = Ky' * G_hat * Q * Kx
		For matrix A that is P = A(A'A)^0.5 */
		push_wmops( "oPtoA MT1M" );
		#ifdef FIX_1326_SPEEDUP_01
		// matrixTransp1Mul_fx( Are_fx, Aim_fx, q_A, Are_fx, Aim_fx, q_A, tmpRe_fx, tmpIm_fx, &q_temp );

		{
		// Word16 chA, chB;
		{
		chA = 0, chB = 0;
		tmpRe_fx[0][0] = Madd_32_32( Madd_32_32( Madd_32_32( Mpy_32_32( Are_fx[0][0], Are_fx[0][0] ),
		Are_fx[1][0], Are_fx[1][0] ),
		Aim_fx[0][0], Aim_fx[0][0] ),
		Aim_fx[1][0], Aim_fx[1][0] );
		move32();
		}
		{
		// chA = 0, chB = 1;
		tmpRe_fx[1][0] = Madd_32_32( Madd_32_32( Madd_32_32( Mpy_32_32( Are_fx[0][1], Are_fx[0][0] ),
		Are_fx[1][1], Are_fx[1][0] ),
		Aim_fx[0][1], Aim_fx[0][0] ),
		Aim_fx[1][1], Aim_fx[1][0] );
		move32();
		tmpIm_fx[1][0] = Msub_32_32( Msub_32_32( Madd_32_32( Mpy_32_32( Are_fx[0][1], Aim_fx[0][0] ),
		Are_fx[1][1], Aim_fx[1][0] ),
		Aim_fx[0][1], Are_fx[0][0] ),
		Aim_fx[1][1], Are_fx[1][0] );
		move32();
		}
		{
		// chA = 1, chB = 0;
		tmpRe_fx[1][1] = Madd_32_32( Madd_32_32( Madd_32_32( Mpy_32_32( Are_fx[0][1], Are_fx[0][1] ),
		Are_fx[1][1], Are_fx[1][1] ),
		Aim_fx[0][1], Aim_fx[0][1] ),
		Aim_fx[1][1], Aim_fx[1][1] );
		move32();
		}
		{
		// chA = 1, chB = 1;
		}

		q_temp = sub( add( q_A, q_A ), 31 );

		move16();
		Word16 ZeroState = add( 1, 0 );
		if ( tmpRe_fx[0][0] != 0 )
		{
		ZeroState = add( 0, 0 );
		}
		if ( tmpRe_fx[1][1] != 0 )
		{
		ZeroState = add( 0, 0 );
		}
		if ( tmpRe_fx[1][0] != 0 )
		{
		ZeroState = add( 0, 0 );
		}
		if ( tmpIm_fx[1][0] != 0 )
		{
		ZeroState = add( 0, 0 );
		}

		if ( sub( ZeroState, 1 ) == 0 )
		{
		q_temp = Q31;
		move16();
		}
		}


		eig2x2_fx( tmpRe_fx[0][0], tmpRe_fx[1][1], q_temp, tmpRe_fx[1][0], tmpIm_fx[1][0], q_temp, Ure_fx, Uim_fx, &q_U, D_fx, &q_D );
		#else
		matrixTransp1Mul_fx( Are_fx, Aim_fx, q_A, Are_fx, Aim_fx, q_A, tmpRe_fx, tmpIm_fx, &q_temp );

		eig2x2_fx( tmpRe_fx[0][0], tmpRe_fx[1][1], q_temp, tmpRe_fx[1][0], tmpIm_fx[1][0], q_temp, Ure_fx, Uim_fx, &q_U, D_fx, &q_D );

		#endif
		pop_wmops(); /push_wmops( "oPtoA MT1M" );/


		@@ -4753,11 +4635,9 @@ static void formulate2x2MixingMatrix_fx(
		move32();

		#ifdef FIX_1326_SPEEDUP_08
		// This is just a shortcut to already existing optimizations (FIX_1009_REPLACE_DIV_SQRT_BY_ISQRT_LC) - but makes everything even faster
		{
		//Sqrt(1)
		div_fx[1] = L_add( 0, 2047986068 ); // Q = 31 - exp1
		exp1 = add( 0, 20 ); // move32();
		}
		exp1 = add( 0, 20 );

		IF( D_fx[1] != 0 ) // This is the new code: replace div sqrt by isqrt
		{
		@@ -4841,11 +4721,7 @@ static void formulate2x2MixingMatrix_fx(
		W_tmp = W_mult0_32_32( tmpIm_fx[chA][chB], div_fx[chB] );
		IF( W_tmp != 0 )
		{
		#ifdef FIX_1326_SPEEDUP_10
		hdrm_im[chA][chB] = W_norm( W_tmp );
		#else
		hdrm_im[chA][chB] = sub( W_norm( W_tmp ), 0 );
		#endif
		move16();
		W_tmp = W_shl( W_tmp, hdrm_im[chA][chB] );
		tmpIm_fx[chA][chB] = W_extract_h( W_tmp );