Commit 63bb612b authored by Sandesh Venkatesh's avatar Sandesh Venkatesh
Browse files

Merge branch 'switch_div32_newton_part04' into 'main' 

lib_enc: replace BASOP_Util_Divide3232_Scale_cadence() by BASOP_Util_Divide3232_Scale_newton() [allow regression]

See merge request !1413
parents aa97e07e f34b6383

lib_enc/find_tilt_fx.c

+1 −1
Original line number Diff line number Diff line
@@ -329,7 +329,7 @@ void find_tilt_ivas_fx( 
    {

        hp_bckr = L_deposit_l( 1 );

    }

    Ltmp = BASOP_Util_Divide3232_Scale_cadence( lp_bckr, hp_bckr, &e_tmp );

    Ltmp = BASOP_Util_Divide3232_Scale_newton( lp_bckr, hp_bckr, &e_tmp );

    Ltmp = Mpy_32_16_r( Ltmp, 3277 /* 0.1f in Q15 */ );

    Ltmp = L_shr_sat( Ltmp, sub( 15, e_tmp ) );

    *bckr_tilt_lt = L_add_sat( Mpy_32_16_r( *bckr_tilt_lt, 29491 /* 0.9f in Q15 */ ), Ltmp ); // Q16

lib_enc/hq_classifier_enc_fx.c

+3 −3
Original line number Diff line number Diff line
@@ -143,7 +143,7 @@ static Word16 hf_spectrum_sparseness_fx( 
            Word16 l_shift = W_norm( inv_rms_fx );

            inv_rms32_fx = W_extract_h( W_shl( inv_rms_fx, l_shift ) );                                     // Q15+l_shift-32

            Word16 q_rms = sub( add( Q15, l_shift ), 32 );                                                  // q_rms

            inv_rms32_div_fx = BASOP_Util_Divide3232_Scale_cadence( inv_rms32_fx, L_SPEC_HB, &inv_rms32_e ); /* Q31-inv_rms32_e */

            inv_rms32_div_fx = BASOP_Util_Divide3232_Scale_newton( inv_rms32_fx, L_SPEC_HB, &inv_rms32_e ); /* Q31-inv_rms32_e */

            inv_rms32_e = sub( 31, add( sub( Q31, inv_rms32_e ), q_rms ) );

            // inv_rms = 1.0f / (float) sqrtf( inv_rms / L_SPEC_HB );

            inv_rms32_fx = ISqrt32( inv_rms32_div_fx, &inv_rms32_e ); /* Q31-inv_rms32_e */

lib_enc/ivas_enc_cov_handler_fx.c

+1 −1
Original line number Diff line number Diff line
@@ -271,7 +271,7 @@ static Word16 ivas_spar_get_activeW_flag_fx( 
    }

    ELSE

    {

        L_tmp = BASOP_Util_Divide3232_Scale_cadence( hCovEnc->bb_var_lt_fx[0], side_ch_var, &exp_diff );       // (Q31 - exp_diff)

        L_tmp = BASOP_Util_Divide3232_Scale_newton( hCovEnc->bb_var_lt_fx[0], side_ch_var, &exp_diff );        // (Q31 - exp_diff)

        en_ratio = L_shl_sat( L_tmp, exp_diff );                                                               // Q31

        IF( LT_32( en_ratio, Mpy_32_32( IVAS_SPAR_DYN_ACTIVEW_THRESH_FX, IVAS_SPAR_DYN_ACTIVEW_THRESH_FX ) ) ) // LHS Q31 :: RHS Q31

        {

lib_enc/ivas_masa_enc_fx.c

+12 −12
Original line number Diff line number Diff line
@@ -1477,7 +1477,7 @@ static void combine_freqbands_and_subframes_fx( 
                    move16();



                    hMeta->directional_meta[i].energy_ratio_fx[j][k] =

                        BASOP_Util_Divide3232_Scale_cadence( vecLen, L_add( energySum, EPSILON_FX ), &exp_diff ); /*exp_diff+vecLen_e-energySum_e*/

                        BASOP_Util_Divide3232_Scale_newton( vecLen, L_add( energySum, EPSILON_FX ), &exp_diff ); /*exp_diff+vecLen_e-energySum_e*/

                    move32();

                    exp_diff = add( exp_diff, sub( vecLen_e, energySum_e ) );

                    hMeta->directional_meta[i].energy_ratio_fx[j][k] =
@@ -1585,7 +1585,7 @@ static void combine_freqbands_and_subframes_fx( 
                {

                    energyRatioSum = BASOP_Util_Add_Mant32Exp( energyRatioSum, energyRatioSum_e, Mpy_32_32( energy[j][k], hMeta->directional_meta[i].energy_ratio_fx[j][k] ), add( energy_e[j][k], 1 ), &energyRatioSum_e ); // 31-energyRatioSum_e

                }

                energyRatioTemp = BASOP_Util_Divide3232_Scale_cadence( energyRatioSum, L_add( energySum, EPSILON_FX ), &exp_diff );

                energyRatioTemp = BASOP_Util_Divide3232_Scale_newton( energyRatioSum, L_add( energySum, EPSILON_FX ), &exp_diff );

                exp_diff = add( exp_diff, sub( energyRatioSum_e, energySum_e ) );

                energyRatioTemp = L_shl( energyRatioTemp, sub( exp_diff, 1 ) ); // Q30
@@ -2264,13 +2264,13 @@ static void compensate_energy_ratios_fx( 
                FOR( dir = 0; dir < numDirs; dir++ )

                {

                    hMeta->directional_meta[dir].energy_ratio_fx[sf][band] =

                        BASOP_Util_Divide3232_Scale_cadence( hMeta->directional_meta[dir].energy_ratio_fx[sf][band], ratioSum, &exp_diff );

                        BASOP_Util_Divide3232_Scale_newton( hMeta->directional_meta[dir].energy_ratio_fx[sf][band], ratioSum, &exp_diff );

                    move32();

                    hMeta->directional_meta[dir].energy_ratio_fx[sf][band] = L_shl( hMeta->directional_meta[dir].energy_ratio_fx[sf][band], sub( exp_diff, Q1 ) ); // Q30

                    move32();

                }

                hMeta->common_meta.diffuse_to_total_ratio_fx[sf][band] =

                    BASOP_Util_Divide3232_Scale_cadence( hMeta->common_meta.diffuse_to_total_ratio_fx[sf][band], ratioSum, &exp_diff );

                    BASOP_Util_Divide3232_Scale_newton( hMeta->common_meta.diffuse_to_total_ratio_fx[sf][band], ratioSum, &exp_diff );

                move32();

                hMeta->common_meta.diffuse_to_total_ratio_fx[sf][band] = L_shl( hMeta->common_meta.diffuse_to_total_ratio_fx[sf][band], sub( exp_diff, Q1 ) ); // Q30

                move32();
@@ -2406,7 +2406,7 @@ static void reduce_metadata_further_fx( 


    IF( hMasa->data.onset_detector_1_fx != 0 )

    {

        onset_filter = L_max( BASOP_Util_Divide3232_Scale_cadence( hMasa->data.onset_detector_2_fx, hMasa->data.onset_detector_1_fx, &onset_filter_e ), 0 );

        onset_filter = L_max( BASOP_Util_Divide3232_Scale_newton( hMasa->data.onset_detector_2_fx, hMasa->data.onset_detector_1_fx, &onset_filter_e ), 0 );



        IF( BASOP_Util_Cmp_Mant32Exp( onset_filter, onset_filter_e, ONE_IN_Q31, 0 ) > 0 )

        {
@@ -2433,7 +2433,7 @@ static void reduce_metadata_further_fx( 
            Word32 bandRatio;



            // threshold = totalEnergySum / ( MAX_PARAM_SPATIAL_SUBFRAMES * LOWBITRATE_NUM_BANDS ) * 0.5f; /* Average energy multiplied with energy ratio of 0.5f */

            threshold = BASOP_Util_Divide3232_Scale_cadence( totalEnergySum, ( MAX_PARAM_SPATIAL_SUBFRAMES * LOWBITRATE_NUM_BANDS ) * 2, &exp ); /* Average energy multiplied with energy ratio of 0.5f */

            threshold = BASOP_Util_Divide3232_Scale_newton( totalEnergySum, ( MAX_PARAM_SPATIAL_SUBFRAMES * LOWBITRATE_NUM_BANDS ) * 2, &exp ); /* Average energy multiplied with energy ratio of 0.5f */

            exp = add( exp, sub( sub( 31, add( hMasa->data.q_energy, tmp2 ) ), 31 ) );

            bandRatio = hqmetadata->q_direction[0].band_data[band].energy_ratio_fx[0]; // Q30

            move32();
@@ -2468,7 +2468,7 @@ static void reduce_metadata_further_fx( 
        }

        IF( totalEnergySum != 0 )

        {

            meanRatio = BASOP_Util_Divide3232_Scale_cadence( meanRatio, totalEnergySum, &exp );

            meanRatio = BASOP_Util_Divide3232_Scale_newton( meanRatio, totalEnergySum, &exp );

            exp = add( add( exp, 1 ), tmp2 ); // 31 - (hMasa->data.q_energy - 1) - 31 - hMasa->data.q_energy - tmp2 => 1 + tmp2

        }

        ELSE
@@ -4730,7 +4730,7 @@ static void ivas_encode_masaism_metadata_fx( 


                FOR( obj = 0; obj < nchan_ism; obj++ )

                {

                    hOmasaData->energy_ratio_ism_fx[sf][0][obj] = BASOP_Util_Divide3232_Scale_cadence( energy_ism_ind[obj], energy_ism, &L_tmp_e );

                    hOmasaData->energy_ratio_ism_fx[sf][0][obj] = BASOP_Util_Divide3232_Scale_newton( energy_ism_ind[obj], energy_ism, &L_tmp_e );

                    move32();

                    L_tmp_e = add( L_tmp_e, sub( energy_ism_ind_e[obj], energy_ism_e ) );

                    /* Scaling to Q30 */
@@ -4740,7 +4740,7 @@ static void ivas_encode_masaism_metadata_fx( 
                L_tmp = BASOP_Util_Add_Mant32Exp( eneBand32, eneBand_exp, energy_ism, energy_ism_e, &L_tmp_e );

                IF( L_tmp != 0 )

                {

                    hOmasaData->masa_to_total_energy_ratio_fx[sf][0] = BASOP_Util_Divide3232_Scale_cadence( eneBand32, L_tmp, &tmp );

                    hOmasaData->masa_to_total_energy_ratio_fx[sf][0] = BASOP_Util_Divide3232_Scale_newton( eneBand32, L_tmp, &tmp );

                    move32();

                    tmp = add( tmp, sub( eneBand_exp, L_tmp_e ) );

                    /* Scaling to Q30 */
@@ -4791,7 +4791,7 @@ static void ivas_encode_masaism_metadata_fx( 
            {

                FOR( obj = 0; obj < nchan_ism; obj++ )

                {

                    hOmasaData->energy_ratio_ism_fx[0][band][obj] = BASOP_Util_Divide3232_Scale_cadence( energy_ism_ind[obj], energy_ism, &L_tmp_e );

                    hOmasaData->energy_ratio_ism_fx[0][band][obj] = BASOP_Util_Divide3232_Scale_newton( energy_ism_ind[obj], energy_ism, &L_tmp_e );

                    move32();

                    L_tmp_e = add( L_tmp_e, sub( energy_ism_ind_e[obj], energy_ism_e ) );

                    /* Scaling to Q30 */
@@ -4819,7 +4819,7 @@ static void ivas_encode_masaism_metadata_fx( 
                L_tmp = BASOP_Util_Add_Mant32Exp( eneBand32, eneBand_exp, energy_ism, energy_ism_e, &L_tmp_e );

                IF( L_tmp != 0 )

                {

                    hOmasaData->masa_to_total_energy_ratio_fx[0][band] = BASOP_Util_Divide3232_Scale_cadence( eneBand32, L_tmp, &tmp );

                    hOmasaData->masa_to_total_energy_ratio_fx[0][band] = BASOP_Util_Divide3232_Scale_newton( eneBand32, L_tmp, &tmp );

                    move32();

                    tmp = add( tmp, sub( eneBand_exp, L_tmp_e ) );

                    /* Scaling to Q30 */
@@ -4876,7 +4876,7 @@ static void ivas_encode_masaism_metadata_fx( 
                    L_tmp = BASOP_Util_Add_Mant32Exp( eneBand32, eneBand_exp, hOmasaData->energy_ism_fx[sf][band], hOmasaData->energy_ism_fx_e[sf][band], &L_tmp_e );

                    IF( L_tmp != 0 )

                    {

                        hOmasaData->masa_to_total_energy_ratio_fx[sf][band] = BASOP_Util_Divide3232_Scale_cadence( eneBand32, L_tmp, &tmp );

                        hOmasaData->masa_to_total_energy_ratio_fx[sf][band] = BASOP_Util_Divide3232_Scale_newton( eneBand32, L_tmp, &tmp );

                        move32();

                        tmp = add( tmp, sub( eneBand_exp, L_tmp_e ) );

                        /* Scaling to Q30 */

lib_enc/ivas_mdct_core_enc_fx.c

+5 −5
Original line number Diff line number Diff line
@@ -233,7 +233,7 @@ static Word16 kernel_switch_detect_fx( 
    L_tmp1 = Sqrt32( L_tmp1, &exp_tmp1 );

    L_tmp = BASOP_Util_Add_Mant32Exp( L_tmp, exp_tmp, L_tmp1, exp_tmp1, &exp_tmp );

    L_tmp = BASOP_Util_Add_Mant32Exp( L_tmp, exp_tmp, ONE_IN_Q30, 1, &exp_tmp );

    cov00 = BASOP_Util_Divide3232_Scale_cadence( cov00, L_tmp, &exp_tmp1 );

    cov00 = BASOP_Util_Divide3232_Scale_newton( cov00, L_tmp, &exp_tmp1 );

    exp_tmp = add( exp_tmp1, sub( sub( Q31, q_com ), exp_tmp ) );

    /* Added saturation to handle overflows when value is very close to 1.f */

    cov00 = L_shl_sat( cov00, exp_tmp ); // Q31
@@ -247,7 +247,7 @@ static Word16 kernel_switch_detect_fx( 
    L_tmp1 = Sqrt32( L_tmp1, &exp_tmp1 );

    L_tmp = BASOP_Util_Add_Mant32Exp( L_tmp, exp_tmp, L_tmp1, exp_tmp1, &exp_tmp );

    L_tmp = BASOP_Util_Add_Mant32Exp( L_tmp, exp_tmp, ONE_IN_Q30, 1, &exp_tmp );

    cov90 = BASOP_Util_Divide3232_Scale_cadence( cov90, L_tmp, &exp_tmp1 );

    cov90 = BASOP_Util_Divide3232_Scale_newton( cov90, L_tmp, &exp_tmp1 );

    exp_tmp = add( exp_tmp1, sub( sub( Q31, q_com ), exp_tmp ) );

    /* Added saturation to handle overflows when value is very close to 1.f */

    cov90 = L_shl_sat( cov90, exp_tmp ); // Q31
@@ -735,7 +735,7 @@ static void applyStereoPreProcessingCplx( 
    d_q = sub( 31, d_q );



    deno = L_max( L_shl_sat( 1, d_q ), d_fx );

    sq_imp = BASOP_Util_Divide3232_Scale_cadence( n_fx, deno, &exp );

    sq_imp = BASOP_Util_Divide3232_Scale_newton( n_fx, deno, &exp );

    exp = add( exp, sub( d_q, n_q ) );

    d_fx = Sqrt32( sq_imp, &exp );

    *mdctSample1_fx = Mpy_32_32( dmxR1_fx, d_fx );
@@ -764,7 +764,7 @@ static void applyStereoPreProcessingCplx( 
    d_q = sub( 31, d_q );



    deno = L_max( L_shl_sat( 1, d_q ), d_fx );

    sq_imp = BASOP_Util_Divide3232_Scale_cadence( n_fx, deno, &exp );

    sq_imp = BASOP_Util_Divide3232_Scale_newton( n_fx, deno, &exp );

    exp = add( exp, sub( d_q, n_q ) );

    d_fx = Sqrt32( sq_imp, &exp );

    *mdctSample2_fx = Mpy_32_32( dmxR2_fx, d_fx );
@@ -966,7 +966,7 @@ static UWord16 enc_ste_pre_mdct( 
    ELSE

    {

        Word16 exp1;

        corr_fx = BASOP_Util_Divide3232_Scale_cadence( Mpy_32_32( corr_fx, corr_fx ), Mpy_32_32( sumL_fx, sumR_fx ), &exp1 );

        corr_fx = BASOP_Util_Divide3232_Scale_newton( Mpy_32_32( corr_fx, corr_fx ), Mpy_32_32( sumL_fx, sumR_fx ), &exp1 );

        corr_e = add( exp1, sub( shl( corr_e, 1 ), add( sumL_e, sumR_e ) ) );

    }

    test();