Commit eee7fdb4 authored by Sandesh Venkatesh's avatar Sandesh Venkatesh
Browse files

Merge branch 'complexity_optimizations_sba_path' into 'main' 

Complexity optimization for SBA path decoding [allow regression]

See merge request !1026
parents 7b212b76 2ec0301c

lib_com/cnst.h

+1 −0
Original line number Diff line number Diff line
@@ -137,6 +137,7 @@ 
#define FOUR_IN_Q28                     1073741824



#define MAX_WORD16                      32767

#define ONE_IN_Q45                      (Word64)0x200000000000

#define ONE_IN_Q62                      (Word64)0x4000000000000000

/*----------------------------------------------------------------------------------*

 * General constants

lib_com/ivas_cnst.h

+1 −0
Original line number Diff line number Diff line
@@ -150,6 +150,7 @@ typedef enum 
    RENDERER_OSBA_LS

} RENDERER_TYPE;



#define MAX_FREQUENCY_BANDS                    64



/*----------------------------------------------------------------------------------*

 * IVAS general constants

lib_com/ivas_dirac_com.c

+32 −17
Original line number Diff line number Diff line
@@ -966,38 +966,55 @@ void computeDiffuseness_fixed( 
        p_tmp_c = buffer_energy + i * num_freq_bands;



        q_tmp = add( q_factor_energy[i], min_q_shift1 );



        Word16 shift_q = sub( q_tmp, q_ene );

        IF( shift_q < 0 )

        {

            FOR( k = 0; k < num_freq_bands; k++ )

            {

                tmp = L_shl( p_tmp_c[k], min_q_shift1 );

            IF( LT_16( q_tmp, q_ene ) )

            {

                energy_slow[k] = L_add( L_shr( energy_slow[k], sub( q_ene, q_tmp ) ), tmp );

                energy_slow[k] = L_add( L_shl( energy_slow[k], shift_q ), tmp );

                move32();

            }

        }

        ELSE

        {

                energy_slow[k] = L_add( energy_slow[k], L_shr( tmp, sub( q_tmp, q_ene ) ) );

            FOR( k = 0; k < num_freq_bands; k++ )

            {

                tmp = L_shl( p_tmp_c[k], min_q_shift1 );

                energy_slow[k] = L_add( energy_slow[k], L_shr( tmp, shift_q ) );

                move32();

            }

        }



        q_ene = s_min( q_ene, q_tmp );



        /* Intensity slow */

        q_tmp = add( q_factor_intensity[i], min_q_shift2 );



        shift_q = sub( q_tmp, q_intensity );

        IF( shift_q > 0 )

        {

            FOR( j = 0; j < DIRAC_NUM_DIMS; ++j )

            {

                p_tmp = buffer_intensity[j][i];

                FOR( k = 0; k < num_freq_bands; k++ )

                {

                    tmp = L_shl( p_tmp[k], min_q_shift2 );

                IF( LT_16( q_intensity, q_tmp ) )

                {

                    intensity_slow[j * num_freq_bands + k] = L_add( intensity_slow[j * num_freq_bands + k], L_shr( tmp, sub( q_tmp, q_intensity ) ) );

                    intensity_slow[j * num_freq_bands + k] = L_add( intensity_slow[j * num_freq_bands + k], L_shr( tmp, shift_q ) );

                    move32();

                }

            }

        }

        ELSE

        {

                    intensity_slow[j * num_freq_bands + k] = L_add( L_shr( intensity_slow[j * num_freq_bands + k], sub( q_intensity, q_tmp ) ), tmp );

            FOR( j = 0; j < DIRAC_NUM_DIMS; ++j )

            {

                p_tmp = buffer_intensity[j][i];

                FOR( k = 0; k < num_freq_bands; k++ )

                {

                    tmp = L_shl( p_tmp[k], min_q_shift2 );

                    intensity_slow[j * num_freq_bands + k] = L_add( L_shl( intensity_slow[j * num_freq_bands + k], shift_q ), tmp );

                    move32();

                }

            }
@@ -1017,9 +1034,7 @@ void computeDiffuseness_fixed( 


        FOR( k = 0; k < num_freq_bands; k++ )

        {

            p_tmp[k] = Mpy_32_32( p_tmp[k], p_tmp[k] );

            move32();

            intensity_slow_abs[k] = L_add( intensity_slow_abs[k], p_tmp[k] );

            intensity_slow_abs[k] = Madd_32_32( intensity_slow_abs[k], p_tmp[k], p_tmp[k] );

            move32();

        }

    }

lib_com/ivas_spar_com.c

+33 −43
Original line number Diff line number Diff line
@@ -7044,7 +7044,7 @@ void ivas_dirac_dec_get_response_fx( 
    Word16 Q_out )

{

    Word16 index_azimuth, index_elevation;

    Word16 el, e, az;

    Word16 el, e, az, q_diff;

    Word32 cos_1_fx, cos_2_fx, sin_1_fx, cos_az_fx[3];

    Word32 sin_az_fx[3];

    Word32 f_fx;
@@ -7055,49 +7055,39 @@ void ivas_dirac_dec_get_response_fx( 
    index_azimuth = add( azimuth, 180 ) % 360;

    move16();

    index_elevation = add( elevation, 90 );

    IF( GT_16( index_elevation, 90 ) )

    {

        e = -1;

        move16();

    }

    ELSE

    {



    e = 1;

    move16();

    }



    IF( GT_16( index_elevation, 90 ) )

    if ( GT_16( index_elevation, 90 ) )

    {

        el = sub( 180, index_elevation );

        move16();

        e = -1;

    }

    ELSE

    {



    el = index_elevation;

    move16();

    }



    IF( GT_16( index_azimuth, 180 ) )

    if ( GT_16( index_elevation, 90 ) )

    {

        az = sub( 360, index_azimuth );

        move16();

        el = sub( 180, index_elevation );

    }

    ELSE

    {



    az = index_azimuth;

    move16();

    }



    IF( GT_16( index_azimuth, 180 ) )

    if ( GT_16( index_azimuth, 180 ) )

    {

        f_fx = -1;

        move16();

        az = sub( 360, index_azimuth );

    }

    ELSE

    {



    f_fx = 1;

    move16();



    if ( GT_16( index_azimuth, 180 ) )

    {

        f_fx = -1;

    }



    cos_1_fx = L_shr( dirac_gains_trg_term_fx[az][0], 1 );  // q30

    cos_2_fx = L_shl( Mpy_32_32( cos_1_fx, cos_1_fx ), 1 ); // q30

    sin_1_fx = L_shr( dirac_gains_trg_term_fx[az][1], 1 );  // q30
@@ -7122,54 +7112,54 @@ void ivas_dirac_dec_get_response_fx( 
    response_fx[0] = L_shl_sat( 1, Q_out ); // Q_out

    move32();



    q_diff = sub( Q_out, 29 );



    FOR( l = 1; l <= ambisonics_order; l++ )

    {

        b_2 = imult1616( l, l );

        b1_2 = add( imult1616( l, l ), shl( l, 1 ) );

        b1_2 = add( b_2, shl( l, 1 ) );

        FOR( m = 0; m < l; m += 2 )

        {

            b = add( b_2, m );

            b = b_2 + m;

            a = dirac_gains_P_idx[b];

            move16();



            c_fx_better = local_result_table[el][a]; // q30

            move32();

            response_fx[b] = L_shl( Mpy_32_32( c_fx_better, sin_az_fx[l - m - 1] ), sub( Q_out, 29 ) ); // Q_out

            response_fx[b] = L_shl( Mpy_32_32( c_fx_better, sin_az_fx[l - m - 1] ), q_diff ); // Q_out

            move32();



            b1 = sub( b1_2, m );

            response_fx[b1] = L_shl( Mpy_32_32( c_fx_better, cos_az_fx[l - m - 1] ), sub( Q_out, 29 ) ); // Q_out

            b1 = b1_2 - m;

            response_fx[b1] = L_shl( Mpy_32_32( c_fx_better, cos_az_fx[l - m - 1] ), q_diff ); // Q_out

            move32();

        }



        FOR( m = 1; m < l; m += 2 )

        {

            b = add( b_2, m );

            b = b_2 + m;

            a = dirac_gains_P_idx[b];

            move16();

            c_fx_better = local_result_table[el][a]; // q30

            move32();

            if ( EQ_16( e, -1 ) )

            {

                c_fx_better = L_negate( c_fx_better ); // q30

            }

            response_fx[b] = L_shl( Mpy_32_32( c_fx_better, sin_az_fx[l - m - 1] ), sub( Q_out, 29 ) ); // Q_out

            response_fx[b] = L_shl( Mpy_32_32( c_fx_better, sin_az_fx[l - m - 1] ), q_diff ); // Q_out

            move32();

            b1 = sub( b1_2, m );

            response_fx[b1] = L_shl( Mpy_32_32( c_fx_better, cos_az_fx[l - m - 1] ), sub( Q_out, 29 ) ); // Q_out



            b1 = b1_2 - m;

            response_fx[b1] = L_shl( Mpy_32_32( c_fx_better, cos_az_fx[l - m - 1] ), q_diff ); // Q_out

            move32();

        }



        b = add( b_2, l );

        a = dirac_gains_P_idx[b];

        move16();

        c_fx_better = local_result_table_2[el][a]; // q30

        move32();

        IF( EQ_16( ( l % 2 ), 1 ) )

        IF( EQ_16( s_and( l, 0x01 ), 1 ) )

        {

            if ( EQ_16( e, -1 ) )

            {

                c_fx_better = L_negate( c_fx_better ); // q30

                move32();

            }

        }

        response_fx[b] = L_shl( c_fx_better, sub( Q_out, 30 ) ); // Q_out

lib_com/ivas_transient_det.c

+9 −9
Original line number Diff line number Diff line
@@ -344,26 +344,26 @@ static Word32 ivas_calc_duck_gain_fx( 
    Word32 duck_gain_out, L_tmp;

    Word16 tmp_e;



    duck_gain_out = L_add( L_shl( Mpy_32_32( L_sub( duck_gain, ONE_IN_Q30 ), duck_coeff ), Q1 ), ONE_IN_Q30 ); /*Q30*/

    duck_gain_out = L_add( Mpy_32_32( L_sub( duck_gain, ONE_IN_Q30 ), duck_coeff ), ONE_IN_Q29 ); /*Q29*/



    IF( EQ_16( BASOP_Util_Cmp_Mant32Exp( Mpy_32_32( duck_mult_fac, env_1 ), add( 2, env1_e ), Mpy_32_32( duck_gain_out, env_2 ), add( 1, env2_e ) ), -1 ) )

    {

        test();

        IF( ( env_1 == 0 ) || ( env_2 == 0 ) )

    IF( LT_64( W_mult0_32_32( duck_mult_fac, env_1 ), W_shr( W_mult0_32_32( duck_gain_out, env_2 ), sub( env1_e, env2_e ) ) ) )

    {



        duck_gain_out = 0;

        move32();

        }

        ELSE



        test();

        IF( ( env_1 != 0 ) && ( env_2 != 0 ) )

        {

            L_tmp = BASOP_Util_Divide3232_Scale_cadence( env_1, env_2, &tmp_e );

            L_tmp = L_shl( L_tmp, add( sub( env1_e, env2_e ), tmp_e ) );



            duck_gain_out = Mpy_32_32( duck_mult_fac, L_tmp ); /*Q29*/

            duck_gain_out = L_shl( duck_gain_out, Q1 );        /*Q30*/

        }

    }



    duck_gain_out = L_shl( duck_gain_out, Q1 ); /*Q30*/



    return duck_gain_out; /*Q30*/

}

/*-----------------------------------------------------------------------------------------*