Commit 72f5f26f authored by ber's avatar ber
Browse files

Replace some divisions, everything is encapsulated in FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT macro

parent 7847d456

lib_rend/ivas_dirac_dec_binaural_functions.c

+183 −2
Original line number Diff line number Diff line
@@ -47,6 +47,14 @@ 


#include "wmc_auto.h"



#define FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT



#ifndef FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT

#define FIX_1072_SET_DIV3232_RETURN_VAL /*only meaningful if FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT is not defined*/

#else

#define FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_ADDEPSIILONOBSOLETE

#endif



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



/*-------------------------------------------------------------------------
@@ -486,6 +494,7 @@ void ivas_dirac_dec_binaural_render_fx( 


    output_length = 0;

    move16();

    push_wmops( "IDR: ivas_dirac_dec_binaural_internal_fx (IDDBI)" );

    FOR( subframe_idx = first_sf; subframe_idx < last_sf; subframe_idx++ )

    {

        Word16 n_samples_sf = imult1616( slot_size, hSpatParamRendCom->subframe_nbslots[subframe_idx] );
@@ -501,6 +510,7 @@ void ivas_dirac_dec_binaural_render_fx( 
        /* update combined orientation access index */

        ivas_combined_orientation_update_index( st_ivas->hCombinedOrientationData, n_samples_sf );

    }

    pop_wmops();



    FOR( ch = 0; ch < nchan_out; ch++ )

    {
@@ -693,6 +703,7 @@ static void ivas_dirac_dec_binaural_internal_fx( 
        }

    }

    /* CLDFB Analysis of input */

    push_wmops( "IDDBI: CLDFB Analysis of input" );

    FOR( slot = 0; slot < hSpatParamRendCom->subframe_nbslots[subframe]; slot++ )

    {

        FOR( ch = 0; ch < numInChannels; ch++ )
@@ -818,10 +829,12 @@ static void ivas_dirac_dec_binaural_internal_fx( 
            }

        }

    }

    pop_wmops();



    test();

    IF( EQ_32( config_data.ivas_format, SBA_FORMAT ) || EQ_32( config_data.ivas_format, SBA_ISM_FORMAT ) )

    {

        push_wmops( "IDDBI: ivas_sba_prototype_renderer_fx" );

        hDiracDecBin->hDiffuseDist = &diffuseDistData;

        ivas_spar_param_to_masa_param_mapping_fx( st_ivas, Cldfb_RealBuffer_in_fx, Cldfb_ImagBuffer_in_fx, q_cldfb, subframe );
@@ -837,6 +850,7 @@ static void ivas_dirac_dec_binaural_internal_fx( 
        {

            Scale_sig32( st_ivas->hSpar->hFbMixer->pFb->fb_bin_to_band.pp_cldfb_weights_per_spar_band_fx[idx], IVAS_MAX_NUM_FB_BANDS, Q22 - Q31 ); /*Q31 to Q22*/

        }

        pop_wmops();

    }



    Word16 q_inp = Q6;
@@ -878,9 +892,10 @@ static void ivas_dirac_dec_binaural_internal_fx( 
    }



    test();

    push_wmops( "IDDBI: ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matrices_fx" );

    ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matrices_fx( hDiracDecBin, hSpatParamRendCom, &config_data, Cldfb_RealBuffer_in_fx, Cldfb_ImagBuffer_in_fx, Rmat_fx, subframe,

                                                                               hCombinedOrientationData && hCombinedOrientationData->enableCombinedOrientation[hCombinedOrientationData->subframe_idx] > 0, st_ivas->hMasaIsmData, q_inp );



    pop_wmops();

    IF( EQ_32( config_data.ivas_format, ISM_FORMAT ) )

    {

        max_band_decorr = 0;
@@ -917,8 +932,10 @@ static void ivas_dirac_dec_binaural_internal_fx( 
        move16();

    }



    push_wmops( "IDDBI: IDDB_detProcMat_fx" );

    ivas_dirac_dec_binaural_determine_processing_matrices_fx( hDiracDecBin, hSpatParamRendCom, &config_data, max_band_decorr, Rmat_fx, subframe,

                                                              hCombinedOrientationData && hCombinedOrientationData->enableCombinedOrientation[hCombinedOrientationData->subframe_idx] > 0, nchanSeparateChannels, st_ivas->hMasaIsmData );

    pop_wmops();



    q_inp = Q6;

    move16();
@@ -964,7 +981,9 @@ static void ivas_dirac_dec_binaural_internal_fx( 
    hDiracDecBin->q_processMtxDecPrev = q_mat;

    move16();



    push_wmops( "IDDBI: ivas_dirac_dec_binaural_process_output_fx" );

    ivas_dirac_dec_binaural_process_output_fx( hDiracDecBin, hSpatParamRendCom, st_ivas->cldfbSynDec, output_fx, &q_out, Cldfb_RealBuffer_in_fx, Cldfb_ImagBuffer_in_fx, q_inp, max_band_decorr, numInChannels, config_data.processReverb, subframe, q_mat );

    pop_wmops();



    hDiracDecBin->hDiffuseDist = NULL;
@@ -1634,6 +1653,7 @@ static void ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matric 
        /* Formulate average diffuseness over frame */

        frameMeanDiffuseness = BASOP_Util_Divide3232_Scale_cadence( frameMeanDiffuseness, L_max( EPSILLON_FX, frameMeanDiffusenessEneWeight_fx[bin] ), &exp ); // exp = exp + 31 - q_meanEnePerCh - exp1

        exp = sub( exp, sub( sub( 31, q_meanEnePerCh ), exp1 ) );



        hDiracDecBin->frameMeanDiffuseness_fx[bin] = L_shl( frameMeanDiffuseness, sub( exp, 2 ) ); // Q29

        move32();

    }
@@ -1802,6 +1822,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
        move16();

    }



    push_wmops( "IDDB_detProcMat_bigLoop1" );

    FOR( bin = 0; bin < nBins; bin++ )

    {

        Word32 tmpMtxRe_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], tmpMtxIm_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], resultMtxRe_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], resultMtxIm_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], gain_fx;
@@ -1947,6 +1968,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
         * signal energy to synthesize a signal with the target covariance matrix from the non-decorrelated signals */

        IF( LT_16( bin, max_band_decorr ) )

        {

            push_wmops( "IDDB_detProcMat_bigLoop1_bigBranch1" );

            Word32 decorrelationReductionFactor_fx;

            Word16 q_decorrelationReductionFactor;
@@ -2049,6 +2071,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
                                         Q31, CrEneL_fx, CrEneR_fx, q_CrEne,

                                         CrCrossRe_fx, CrCrossIm_fx, q_CrCross,

                                         prototypeMtx_fx, MdecRe_fx, MdecIm_fx, &q_Mdec, 3277 ); // 3277 = 0.2 in Q14

            pop_wmops();

        }

        ELSE

        {
@@ -2199,6 +2222,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
            move16();

            UWord8 instantChange = 0;

            move16();

            push_wmops( "IDDB_detProcMat_bigLoop1_bigBranch2" );



            exp = sub( 31, hDiracDecBin->q_earlyPartEneCorrection );

            tmp1 = Sqrt32( hDiracDecBin->earlyPartEneCorrection_fx[bin], &exp );
@@ -2274,8 +2298,11 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
                    move16();

                }

            }

            pop_wmops();

        }

    }

    pop_wmops();



    /* Aligning Q-factors of all bins in the processing matrices to a common Q-factor */

    minimum_s( q_processMtx, nBins, &hDiracDecBin->q_processMtx );

    minimum_s( q_processMtxPrev, nBins, &hDiracDecBin->q_processMtxPrev );
@@ -2295,6 +2322,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
    minimum_s( q_processMtxDec, nBins, &hDiracDecBin->q_processMtxDec );

    minimum_s( q_processMtxDecPrev, nBins, &hDiracDecBin->q_processMtxDecPrev );



    push_wmops( "IDDB_detProcMat_bigLoop2" );

    FOR( bin = 0; bin < nBins; bin++ )

    {

        FOR( chA = 0; chA < BINAURAL_CHANNELS; chA++ )
@@ -2334,6 +2362,7 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
            }

        }

    }

    pop_wmops();



    return;

}
@@ -3965,6 +3994,8 @@ static void chol2x2_fx( 
        // 4611686 = Q62

        IF( outRe[1][1] == 0 )

        {

            push_wmops( "DIV3232_0" );

#if !defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT ) || 1

            outRe[0][1] = BASOP_Util_Divide3232_Scale_cadence( c_re, 4611686, &exp );

            move32();

            q_re2 = add( sub( 31, exp ), sub( q_c, 62 ) );
@@ -3972,9 +4003,27 @@ static void chol2x2_fx( 
            outIm[0][1] = BASOP_Util_Divide3232_Scale_cadence( -c_im, 4611686, &exp );

            move32();

            q_im = add( sub( 31, exp ), sub( q_c, 62 ) );



#else /*ALternative code has to be verified, first*/

            CDK_ASSERT(0);

            //outRe[0][1] = BASOP_Util_Divide3232_Scale_cadence( c_re, 4611686, &exp );

            Word32 tmp1 = 1953125005;

            exp = 9;

            outRe[0][1] = Mpy_32_32(tmp1, c_re);

            

            move32();

            q_re2 = add( sub( 31, exp ), sub( q_c, 62 ) );



            //outIm[0][1] = BASOP_Util_Divide3232_Scale_cadence( -c_im, 4611686, &exp );

            outIm[0][1] = Mpy_32_32(tmp1, -c_im);

            move32();

            q_im = add( sub( 31, exp ), sub( q_c, 62 ) );

#endif

            pop_wmops();

        }

        ELSE

        {

#if !defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT) || 1

            outRe[0][1] = BASOP_Util_Divide3232_Scale_cadence( c_re, outRe[1][1], &exp );

            move32();

            q_re2 = add( sub( 31, exp ), sub( q_c, q_re1 ) );
@@ -3982,6 +4031,21 @@ static void chol2x2_fx( 
            outIm[0][1] = BASOP_Util_Divide3232_Scale_cadence( -c_im, outRe[1][1], &exp );

            move32();

            q_im = add( sub( 31, exp ), sub( q_c, q_re1 ) );

#else /*alternative code has to be verified, first*/

            {

                CDK_ASSERT( 0 );

                //outRe[0][1] = BASOP_Util_Divide3232_Scale_cadence( c_re, outRe[1][1], &exp );

                Word32 tmp1 = BASOP_Util_Divide3232_Scale_cadence( 0x7FFFFFFF, outRe[1][1], &exp );

                outRe[0][1] = Mpy_32_32(tmp1, c_re);

                move32();

                q_re2 = add( sub( 31, exp ), sub( q_c, q_re1 ) );



                //outIm[0][1] = BASOP_Util_Divide3232_Scale_cadence( -c_im, outRe[1][1], &exp );

                outIm[0][1] = Mpy_32_32(tmp1, -c_im);

                move32();

                q_im = add( sub( 31, exp ), sub( q_c, q_re1 ) );

            }

#endif

        }

        if ( outRe[0][1] == 0 )

        {
@@ -4092,6 +4156,9 @@ static void formulate2x2MixingMatrix_fx( 
    Word32 temp;

    Word16 q_Pre[BINAURAL_CHANNELS][BINAURAL_CHANNELS], q_Pim[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

    Word16 hdrm_re[BINAURAL_CHANNELS][BINAURAL_CHANNELS], hdrm_im[BINAURAL_CHANNELS][BINAURAL_CHANNELS];



    push_wmops( "IDDB_detProcMat_bigLoop1_2x2Matrix" );



    set16_fx( hdrm_re[0], 63, i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS ) );

    set16_fx( hdrm_im[0], 63, i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS ) );

    set16_fx( q_Pre[0], Q31, i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS ) );
@@ -4205,9 +4272,48 @@ static void formulate2x2MixingMatrix_fx( 
    temp = Mpy_32_32( E_in2, 2147484 ); // 2147484 = 0.001f in Q31

    temp = L_max( temp, E_in1 );



#if defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT )

    /*IF (E_out1 == 0)*/

    {

        Ghat_fx[0] = 0;

        move32();

        exp = -19;

    }



    IF( E_out1 != 0 )

    {

        IF( temp == 0 )

        {

            // ASSERT for testing

            temp = BASOP_Util_Divide3232_Scale_cadence( E_out1, 4611686, &exp ); // 4611686 = Q62

            exp = sub( exp, sub( q_eout, 62 ) );

            Ghat_fx[0] = Sqrt32( temp, &exp ); // Q = 31 - exp

            move32();

        }

        ELSE

        {

#ifndef FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_ADDEPSIILONOBSOLETE

            temp = BASOP_Util_Add_Mant32Exp( temp, sub( 31, q_ein ), EPSILON_MANT, EPSILON_EXP, &exp_temp ); //obsolete?

            exp_temp = add( exp_temp, sub( 31, q_eout ) );

#else

            exp_temp = add( sub( 31, q_ein ), sub( 31, q_eout ) );

#endif

            temp = Mpy_32_32( temp, E_out1 );

            temp = ISqrt32( temp, &exp_temp );

            Ghat_fx[0] = Mpy_32_32( temp, E_out1 );

            move32();

            exp = add( exp_temp, sub( 31, q_eout ) );

        }

    }



#else

    IF( temp == 0 )

    {

#ifdef FIX_1072_SET_DIV3232_RETURN_VAL

        temp = BASOP_Util_Divide3232_Scale_cadence( E_out1, 4611686, &exp ); // 4611686 = Q62

#else

        BASOP_Util_Divide3232_Scale_cadence( E_out1, 4611686, &exp );  // 4611686 = Q62

#endif

        exp = sub( exp, sub( q_eout, 62 ) );

    }

    ELSE
@@ -4219,12 +4325,53 @@ static void formulate2x2MixingMatrix_fx( 
    }

    Ghat_fx[0] = Sqrt32( temp, &exp ); // Q = 31 - exp

    move32();

#endif /*FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT*/



    temp = Mpy_32_32( E_in1, 2147484 ); // 2147484 = 0.001f in Q31

    temp = L_max( temp, E_in2 );        // q_ein





#if defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT )

    //IF( E_out2 == 0 )

    {

        Ghat_fx[1] = 0;

        exp1 = -19;

        move32();

    }



    IF( E_out2 != 0 )

    {

        IF( temp == 0 )

        {

            /*ASSERT for testing*/

            temp = BASOP_Util_Divide3232_Scale_cadence( E_out2, 4611686, &exp1 ); // 4611686 = Q62

            exp1 = sub( exp1, sub( q_eout, 62 ) );

            Ghat_fx[1] = Sqrt32( temp, &exp1 ); // Q = 31 - exp1

            move32();

        }

        ELSE

        {

#ifndef FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_ADDEPSIILONOBSOLETE

            temp = BASOP_Util_Add_Mant32Exp( temp, sub( 31, q_ein ), EPSILON_MANT, EPSILON_EXP, &exp_temp ); //obsolete?

            exp1 = add( sub( 31, q_eout ), exp_temp );

#else

            exp1 = add( sub( 31, q_eout ), sub( 31, q_ein ) );

#endif

            temp = Mpy_32_32( E_out2, temp );

            temp = ISqrt32( temp, &exp1 );

            Ghat_fx[1] = Mpy_32_32( temp, E_out2 );

            move32();

            exp1 = add( sub( 31, q_eout ), exp1 );

        }

    }

#else

    IF( temp == 0 )

    {

#ifdef FIX_1072_SET_DIV3232_RETURN_VAL

        temp = BASOP_Util_Divide3232_Scale_cadence( E_out2, 4611686, &exp1 ); // 4611686 = Q62

#else

        BASOP_Util_Divide3232_Scale_cadence( E_out2, 4611686, &exp1 ); // 4611686 = Q62

#endif

        exp1 = sub( exp1, sub( q_eout, 62 ) );

    }

    ELSE
@@ -4236,6 +4383,7 @@ static void formulate2x2MixingMatrix_fx( 
    }

    Ghat_fx[1] = Sqrt32( temp, &exp1 ); // Q = 31 - exp1

    move32();

#endif /*FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT*/



    q_Ghat = sub( 31, s_max( exp, exp1 ) );
@@ -4281,6 +4429,21 @@ static void formulate2x2MixingMatrix_fx( 


    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 );



#if defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT )

    IF( D_fx[0] == 0 )

    {

        temp = 2047986068;

        exp = 20;

        div_fx[0] = temp;

        move32();

    }

    ELSE

    {

        exp = sub( 31, q_D );

        div_fx[0] = ISqrt32( D_fx[0], &exp );

        move32();

    }

#else

    IF( D_fx[0] == 0 )

    {

        temp = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, 4611686, &exp ); // 4611686 = 1e-12 in Q62
@@ -4293,7 +4456,23 @@ static void formulate2x2MixingMatrix_fx( 
    }

    div_fx[0] = Sqrt32( temp, &exp ); // Q = 31 - exp

    move32();

#endif /*FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT*/



#if defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT )

    IF( D_fx[1] == 0 )

    {

        temp = 2047986068;

        exp1 = 20;

        div_fx[1] = temp;

        move32();

    }

    ELSE

    {

        exp1 = sub( 31, q_D );

        div_fx[1] = ISqrt32( D_fx[1], &exp1 );

        move32();

    }

#else

    IF( D_fx[1] == 0 )

    {

        temp = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, 4611686, &exp1 ); // 4611686 = 1e-12 in Q62
@@ -4306,6 +4485,7 @@ static void formulate2x2MixingMatrix_fx( 
    }

    div_fx[1] = Sqrt32( temp, &exp1 ); // Q = 31 - exp1

    move32();

#endif /*FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT*/



    q_div = sub( 31, s_max( exp, exp1 ) );
@@ -4446,6 +4626,7 @@ static void formulate2x2MixingMatrix_fx( 


    matrixTransp2Mul_fx( tmpRe_fx, tmpIm_fx, &q_temp, Uxre_fx, Uxim_fx, &q_Ux, Mre_fx, Mim_fx, q_M );



    pop_wmops();

    return;

}