diff --git a/lib_com/options.h b/lib_com/options.h index ccaeca46a2cb3f0d49e9cea01bb545fc753cd27c..ab7efb8fd580ec347b275933a94e9303417b95cc 100644 --- a/lib_com/options.h +++ b/lib_com/options.h @@ -176,4 +176,7 @@ #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 NONBE_FIX_1386_STEREO_DMX_EVS_PHA /* Orange: Fix for stereo DMX / PHA mode : Change the filter taps resolution (Q31->Q30), improve precision for the IR window, for the ILD & IPD smoothing in sub-bands, for the ISD counters and for ICCr. */ +#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 diff --git a/lib_rend/ivas_dirac_dec_binaural_functions_fx.c b/lib_rend/ivas_dirac_dec_binaural_functions_fx.c index b5dd1f8b93cf291367dee3c9db457930b17fec04..be86281ef7428818775c81b6d922041656571e5c 100644 --- a/lib_rend/ivas_dirac_dec_binaural_functions_fx.c +++ b/lib_rend/ivas_dirac_dec_binaural_functions_fx.c @@ -2151,6 +2151,14 @@ 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_SUBSTITUTE_DIV_SQRT_IOSQRT + { + 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( exp2, exp1 ) ); + } +#else { Word16 exp_temp; tmp2 = BASOP_Util_Add_Mant32Exp( tmp2, sub( 31, q_tmp2 ), EPSILON_MANT, EPSILON_EXP, &exp_temp ); @@ -2159,6 +2167,8 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( } gain_fx = Sqrt32( tmp2, &exp2 ); q_gain = sub( 31, exp2 ); +#endif + // 1073741824 = 4 in Q28 IF( LT_16( q_gain, Q28 ) ) @@ -3249,6 +3259,20 @@ static void eig2x2_fx( /* Numeric case, when input is practically zeros */ // IF( D_fx[0] < EPSILON_FX ) + +#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; + move32(); + Ure_fx[1][1] = ONE_IN_Q31; + move32(); + *q_U = Q31; + move16(); + + return; + } +#else IF( EQ_16( BASOP_Util_Cmp_Mant32Exp( D_fx[0], *q_D, EPSILON_MANT, EPSILON_EXP ), -1 ) ) { Ure_fx[0][0] = ONE_IN_Q31; @@ -3260,6 +3284,7 @@ static void eig2x2_fx( return; } +#endif /* Numeric case, when input is near an identity matrix with a gain */ tmp1 = Mpy_32_32( 2147484, add_fx ); // 2147484 = 1e-3f in Q31 @@ -3468,7 +3493,19 @@ static void eig2x2_fx( move16(); } } +#ifdef FIX_1326_SPEEDUP_eig2x2_fx + if ( q_U_1 != 0 ) + { + *q_U = q_U_1; + move16(); + } + if ( q_U_1 == 0 ) + { + *q_U = q_U_2; + move16(); + } +#else IF( q_U_1 != 0 ) { *q_U = q_U_1; @@ -3478,6 +3515,7 @@ static void eig2x2_fx( *q_U = q_U_2; } move16(); +#endif return; } @@ -4532,7 +4570,9 @@ static void formulate2x2MixingMatrix_fx( exp = sub( exp, sub( Q30, 62 ) ); #else temp = ONE_DIV_EPSILON_MANT; /* Result of 1.0/eps with full precision */ + move32(); exp = ONE_DIV_EPSILON_EXP; + move16(); #endif } ELSE @@ -4543,6 +4583,18 @@ static void formulate2x2MixingMatrix_fx( div_fx[0] = Sqrt32( temp, &exp ); // Q = 31 - exp move32(); +#ifdef FIX_1326_SUBSTITUTE_DIV_SQRT_IOSQRT + // Sqrt(1) + div_fx[1] = L_add( 0, 2047986068 ); // Q = 31 - exp1 + exp1 = add( 0, 20 ); + + IF( D_fx[1] != 0 ) // This is the new code: replace div sqrt by isqrt + { + exp1 = sub( 31, q_D ); + div_fx[1] = ISqrt32( D_fx[1], &exp1 ); + move32(); + } +#else IF( D_fx[1] == 0 ) { #ifndef FIX_1009_REPLACE_DIV_SQRT_BY_ISQRT_LC @@ -4560,7 +4612,7 @@ static void formulate2x2MixingMatrix_fx( } div_fx[1] = Sqrt32( temp, &exp1 ); // Q = 31 - exp1 move32(); - +#endif q_div = sub( 31, s_max( exp, exp1 ) ); div_fx[0] = L_shr( div_fx[0], sub( sub( 31, exp ), q_div ) ); // q_div