Commit 76fc3d30 authored by Nicolas Roussin's avatar Nicolas Roussin
Browse files

Finalize optimization.

parent 890cd054

lib_basop/enh64.h

+0 −102
Original line number Diff line number Diff line
@@ -21,108 +21,6 @@ 
 *

 *****************************************************************************/

#ifdef ENH_64_BIT_OPERATOR





/*______________________________________________________________________________

|                                                                              |

|   Function Name : W_min                                                      |

|                                                                              |

|   Purpose :                                                                  |

|                                                                              |

|    Compares L64_var1 and L64_var2 and returns the minimum value.             |

|                                                                              |

|   Complexity weight : 1                                                      |

|                                                                              |

|   Inputs :                                                                   |

|                                                                              |

|    L64_var1   64 bit long signed integer (Word64) whose value falls in the   |

|       range : 0x80000000 00000000LL <= L64_var1 <= 0x7fffffff ffffffffLL.    |

|                                                                              |

|    L64_var2   64 bit long signed integer (Word64) whose value falls in the   |

|       range : 0x80000000 00000000LL <= L64_var2 <= 0x7fffffff ffffffffLL.    |

|                                                                              |

|   Outputs :                                                                  |

|                                                                              |

|    none                                                                      |

|                                                                              |

|   Return Value :                                                             |

|                                                                              |

|    L64_var_out                                                               |

|             64 bit long signed integer (Word64) whose value falls in the     |

|       range : 0x80000000 00000000LL <= L64_var_out <= 0x7fffffff ffffffffLL. |

|______________________________________________________________________________|

*/

static __inline Word64 W_min( Word64 L64_var1, Word64 L64_var2 )

{

    Word64 L64_var_out;



    if ( L64_var1 <= L64_var2 )

    {

        L64_var_out = L64_var1;

    }

    else

    {

        L64_var_out = L64_var2;

    }



#ifdef WMOPS

    multiCounter[currCounter].W_min++;

#endif /* ifdef WMOPS */



    return ( L64_var_out );

}





/*______________________________________________________________________________

|                                                                              |

|   Function Name : W_max                                                      |

|                                                                              |

|   Purpose :                                                                  |

|                                                                              |

|    Compares L64_var1 and L64_var2 and returns the maximum value.             |

|                                                                              |

|   Complexity weight : 1                                                      |

|                                                                              |

|   Inputs :                                                                   |

|                                                                              |

|    L64_var1   64 bit long signed integer (Word64) whose value falls in the   |

|       range : 0x80000000 00000000LL <= L64_var1 <= 0x7fffffff ffffffffLL.    |

|                                                                              |

|    L64_var2   64 bit long signed integer (Word64) whose value falls in the   |

|       range : 0x80000000 00000000LL <= L64_var2 <= 0x7fffffff ffffffffLL.    |

|                                                                              |

|   Outputs :                                                                  |

|                                                                              |

|    none                                                                      |

|                                                                              |

|   Return Value :                                                             |

|                                                                              |

|    L64_var_out                                                               |

|             64 bit long signed integer (Word64) whose value falls in the     |

|       range : 0x80000000 00000000LL <= L64_var_out <= 0x7fffffff ffffffffLL. |

|______________________________________________________________________________|

*/

static __inline Word64 W_max( Word64 L64_var1, Word64 L64_var2 )

{

    Word64 L64_var_out;



    if ( L64_var1 >= L64_var2 )

    {

        L64_var_out = L64_var1;

    }

    else

    {

        L64_var_out = L64_var2;

    }



#ifdef WMOPS

    multiCounter[currCounter].W_max++;

#endif /* ifdef WMOPS */



    return ( L64_var_out );

}





Word64 W_add_nosat( Word64 L64_var1, Word64 L64_var2 );

Word64 W_sub_nosat( Word64 L64_var1, Word64 L64_var2 );

Word64 W_shl( Word64 L64_var1, Word16 var2 );

lib_com/options.h

+6 −0
Original line number Diff line number Diff line
@@ -159,4 +159,10 @@ 


/* #################### End BASOP porting switches ############################ */



/* #################### Start BASOP optimization switches ############################ */



#define NONBE_OPT_2193_EIG2X2                           /* Dolby: Issue 2193, optimize eig2x2_fx. */



/* #################### End BASOP optimization switches ############################ */



#endif

lib_debug/wmc_auto.c

+1 −1
Original line number Diff line number Diff line
@@ -133,7 +133,7 @@ static BASIC_OP op_weight = { 
#ifdef ENH_64_BIT_OPERATOR

    /* Weights of new 64 bit basops */

    ,

    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1

    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1

#endif /* #ifdef ENH_64_BIT_OPERATOR */



#ifdef ENH_32_BIT_OPERATOR

lib_debug/wmc_auto.h

+0 −2
Original line number Diff line number Diff line
@@ -877,8 +877,6 @@ typedef struct 
/* New 64 bit basops */

#ifdef ENH_64_BIT_OPERATOR

    unsigned int move64;        /* Complexity Weight of 1 */

    unsigned int W_min;         /* Complexity Weight of 1 */

    unsigned int W_max;         /* Complexity Weight of 1 */

    unsigned int W_add_nosat;   /* Complexity Weight of 1 */

    unsigned int W_sub_nosat;   /* Complexity Weight of 1 */

    unsigned int W_shl;         /* Complexity Weight of 1 */

lib_rend/ivas_dirac_dec_binaural_functions_fx.c

+35 −202
Original line number Diff line number Diff line
@@ -3518,63 +3518,20 @@ static void ivas_dirac_dec_binaural_check_and_switch_transports_headtracked_fx( 
    return;

}



#if 1

static void check(

    Word32 computed_fx,

    Word16 computed_q,

    Word32 expected_fx,

    Word16 expected_q,

    Word32 max_abs_err );



static void check(

    Word32 computed_fx,

    Word16 computed_q,

    Word32 expected_fx,

    Word16 expected_q,

    Word32 max_abs_err )

{

    Word16 qd = computed_q - expected_q;

    Word32 cf = computed_fx >> +max( qd, 0 );

    Word32 ef = expected_fx >> -min( qd, 0 );

    Word32 abs_error = abs( cf - ef );

    if ( abs_error > max_abs_err )

    {

        assert( false );

    }

}

#endif



Word32 __pm_fx;

Word16 __pm_q;

Word32 __add_fx;

Word16 __add_q;



Word32 __as[BINAURAL_CHANNELS];



Word32 __s_fx[BINAURAL_CHANNELS];

Word16 __s_q[BINAURAL_CHANNELS];



Word32 __nval_fx[BINAURAL_CHANNELS];

Word16 __nval_q[BINAURAL_CHANNELS];



Word32 __Ure_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

Word32 __Uim_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

Word16 __U_q;



Word32 __D_fx[BINAURAL_CHANNELS];

Word16 __D_q;



Word32 __diff_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

Word16 __diff_q;



static void eig2x2_opt(

static void eig2x2_fx(

    const Word32 E1_fx, /*q_E*/

    const Word32 E2_fx, /*q_E*/

    Word16 q_E,

    const Word32 Cre_fx, /*q_C*/

    const Word32 Cim_fx, /*q_C*/

    Word16 q_C )

    Word16 q_C,

    Word32 Ure_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], /*q_U*/

    Word32 Uim_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], /*q_U*/

    Word16 *q_U,

    Word32 D_fx[BINAURAL_CHANNELS], /*q_D*/

    Word16 *q_D )

{

#ifdef NONBE_OPT_2193_EIG2X2

    Word32 s_fx[BINAURAL_CHANNELS];

    Word32 pm_fx, add_fx;

    Word32 tmp1, tmp2, e1, e2, c_re, c_im, c0_im, c1_im;
@@ -3586,13 +3543,13 @@ static void eig2x2_opt( 
    move32();

    move16();



    set32_fx( (Word32 *) __Ure_fx, 0, BINAURAL_CHANNELS * BINAURAL_CHANNELS );

    set32_fx( (Word32 *) __Uim_fx, 0, BINAURAL_CHANNELS * BINAURAL_CHANNELS );

    __Ure_fx[0][0] = ONE_IN_Q30;

    set32_fx( (Word32 *) Ure_fx, 0, BINAURAL_CHANNELS * BINAURAL_CHANNELS );

    set32_fx( (Word32 *) Uim_fx, 0, BINAURAL_CHANNELS * BINAURAL_CHANNELS );

    Ure_fx[0][0] = ONE_IN_Q30;

    move32();

    __Ure_fx[1][1] = ONE_IN_Q30;

    Ure_fx[1][1] = ONE_IN_Q30;

    move32();

    __U_q = Q30;

    *q_U = Q30;

    move16();



    exp = sub( get_min_scalefactor( Cre_fx, Cim_fx ), 2 );
@@ -3662,11 +3619,6 @@ static void eig2x2_opt( 
    q_tmp1 = q_e;

    move16();



    __pm_fx = pm_fx;   // FIXME

    __pm_q = q_tmp2;   // FIXME

    __add_fx = add_fx; // FIXME

    __add_q = q_tmp1;  // FIXME



    // D[0] = add + pm;

    // D[1] = max( 0.0f, add - pm );
@@ -3686,17 +3638,17 @@ static void eig2x2_opt( 
        tmp2 = L_shl( tmp2, q_diff );

    }



    __D_fx[0] = L_add( tmp1, tmp2 );

    D_fx[0] = L_add( tmp1, tmp2 );

    move32();

    __D_fx[1] = L_max( L_sub( tmp1, tmp2 ), 0 );

    D_fx[1] = L_max( L_sub( tmp1, tmp2 ), 0 );

    move32();

    __D_q = s_min( q_tmp1, q_tmp2 );

    *q_D = s_min( q_tmp1, q_tmp2 );

    move32();



    // Numeric case, when input is practically zeros

    // if ( __D_fx[0] < EPSILON_FX )

    // if ( D_fx[0] < EPSILON_FX )



    if ( LT_32( L_shl_sat( __D_fx[0], sub( 31 - EPSILON_EXP, __D_q ) ), EPSILON_MANT ) )

    if ( LT_32( L_shl_sat( D_fx[0], sub( 31 - EPSILON_EXP, *q_D ) ), EPSILON_MANT ) )

    {

        return;

    }
@@ -3715,17 +3667,17 @@ static void eig2x2_opt( 


    // Eigenvectors



    q_diff = sub( q_e, __D_q );

    q_tmp1 = s_min( q_e, __D_q );

    q_diff = sub( q_e, *q_D );

    q_tmp1 = s_min( q_e, *q_D );



    tmp1 = __D_fx[0];

    tmp1 = D_fx[0];

    move32();

    if ( q_diff > 0 )

    {

        tmp1 = L_shr( tmp1, q_diff );

    }



    tmp2 = __D_fx[1];

    tmp2 = D_fx[1];

    move32();

    if ( q_diff > 0 )

    {
@@ -3742,23 +3694,13 @@ static void eig2x2_opt( 
        e2 = L_shl( e2, q_diff );

    }



    s_fx[0] = L_sub( tmp1, e1 ); // __D_fx[0] - e1

    tmp1 = L_sub( tmp1, e2 );    // __D_fx[0] - e2

    s_fx[1] = L_sub( tmp2, e1 ); // __D_fx[1] - e1

    tmp2 = L_sub( tmp2, e2 );    // __D_fx[1] - e2



    __diff_fx[0][0] = s_fx[0];

    __diff_fx[0][1] = tmp1;

    __diff_fx[1][0] = s_fx[1];

    __diff_fx[1][1] = tmp2;

    s_fx[0] = L_sub( tmp1, e1 ); // D_fx[0] - e1

    tmp1 = L_sub( tmp1, e2 );    // D_fx[0] - e2

    s_fx[1] = L_sub( tmp2, e1 ); // D_fx[1] - e1

    tmp2 = L_sub( tmp2, e2 );    // D_fx[1] - e2



    __diff_q = q_tmp1;



    i01 = GT_32( L_abs( tmp1 ), L_abs( s_fx[0] ) ); // fabsf( __D_fx[0] - e2 ) > fabsf( __D_fx[0] - e1 )

    i11 = GT_32( L_abs( tmp2 ), L_abs( s_fx[1] ) ); // fabsf( __D_fx[1] - e2 ) > fabsf( __D_fx[1] - e1 )



    __as[0] = i01;

    __as[1] = i11;

    i01 = GT_32( L_abs( tmp1 ), L_abs( s_fx[0] ) ); // fabsf( D_fx[0] - e2 ) > fabsf( D_fx[0] - e1 )

    i11 = GT_32( L_abs( tmp2 ), L_abs( s_fx[1] ) ); // fabsf( D_fx[1] - e2 ) > fabsf( D_fx[1] - e1 )



    if ( i01 )

    {
@@ -3772,11 +3714,6 @@ static void eig2x2_opt( 
        move32();

    }



    __s_fx[0] = s_fx[0];

    __s_fx[1] = s_fx[1];

    __s_q[0] = q_tmp1;

    __s_q[1] = q_tmp1;



    // normVal = sqrtf( 1.0f / ( 1e-12f + crossSquare + s * s ) );



    Word32 nvalm[BINAURAL_CHANNELS];
@@ -3820,11 +3757,6 @@ static void eig2x2_opt( 
    nvalm[1] = Sqrt32( nvalm[1], &exp );

    nvalq[1] = sub( 31, exp );



    __nval_fx[0] = nvalm[0];

    __nval_q[0] = nvalq[0];

    __nval_fx[1] = nvalm[1];

    __nval_q[1] = nvalq[1];



    q_diff = sub( q_c, q_tmp1 );

    q_tmp1 = s_min( q_tmp1, q_c );
@@ -3861,7 +3793,7 @@ static void eig2x2_opt( 
        nvalm[1] = L_shl( nvalm[1], q_diff );

    }



    __U_q = sub( add( q_tmp1, q_tmp2 ), 31 );

    *q_U = sub( add( q_tmp1, q_tmp2 ), 31 );



    i00 = L_sub( 1, i01 );

    i10 = L_sub( 1, i11 );
@@ -3880,53 +3812,20 @@ static void eig2x2_opt( 
        c1_im = L_negate( c1_im );

    }



    __Ure_fx[i00][0] = Mpy_32_32( s_fx[0], nvalm[0] );

    Ure_fx[i00][0] = Mpy_32_32( s_fx[0], nvalm[0] );

    move32();

    __Ure_fx[i01][0] = Mpy_32_32( c_re, nvalm[0] );

    Ure_fx[i01][0] = Mpy_32_32( c_re, nvalm[0] );

    move32();

    __Uim_fx[i01][0] = Mpy_32_32( c0_im, nvalm[0] );

    Uim_fx[i01][0] = Mpy_32_32( c0_im, nvalm[0] );

    move32();



    __Ure_fx[i10][1] = Mpy_32_32( s_fx[1], nvalm[1] );

    Ure_fx[i10][1] = Mpy_32_32( s_fx[1], nvalm[1] );

    move32();

    __Ure_fx[i11][1] = Mpy_32_32( c_re, nvalm[1] );

    Ure_fx[i11][1] = Mpy_32_32( c_re, nvalm[1] );

    move32();

    __Uim_fx[i11][1] = Mpy_32_32( c1_im, nvalm[1] );

    Uim_fx[i11][1] = Mpy_32_32( c1_im, nvalm[1] );

    move32();



    return;

}



static void eig2x2_fx(

    const Word32 E1_fx, /*q_E*/

    const Word32 E2_fx, /*q_E*/

    Word16 q_E,

    const Word32 Cre_fx, /*q_C*/

    const Word32 Cim_fx, /*q_C*/

    Word16 q_C,

    Word32 Ure_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], /*q_U*/

    Word32 Uim_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], /*q_U*/

    Word16 *q_U,

    Word32 D_fx[BINAURAL_CHANNELS], /*q_D*/

    Word16 *q_D )

{

#if 0

    eig2x2_opt( E1_fx, E2_fx, q_E, Cre_fx, Cim_fx, q_C );

    Ure_fx[0][0] = __Ure_fx[0][0];

    Ure_fx[0][1] = __Ure_fx[0][1];

    Ure_fx[1][0] = __Ure_fx[1][0];

    Ure_fx[1][1] = __Ure_fx[1][1];

    Uim_fx[0][0] = __Uim_fx[0][0];

    Uim_fx[0][1] = __Uim_fx[0][1];

    Uim_fx[1][0] = __Uim_fx[1][0];

    Uim_fx[1][1] = __Uim_fx[1][1];

    *q_U = __U_q;

    D_fx[0] = __D_fx[0];

    D_fx[1] = __D_fx[1];

    *q_D = __D_q;

#else

    eig2x2_opt( E1_fx, E2_fx, q_E, Cre_fx, Cim_fx, q_C );



    Word16 chA, chB, ch;

    Word32 s_fx, normVal_fx, crossSquare_fx, a_fx, pm_fx, add_fx;

    Word32 tmp1, tmp2, tmp3, e1, e2, c_re, c_im;
@@ -4049,9 +3948,6 @@ static void eig2x2_fx( 
    q_tmp1 = q_e;

    move16();



    check( __pm_fx, __pm_q, pm_fx, q_tmp2, 1 << 26 );

    check( __add_fx, __add_q, add_fx, q_tmp1, 1 << 26 );



    IF( LT_16( q_tmp1, q_tmp2 ) )

    {

        D_fx[0] = L_add( L_shr( add_fx, 1 ), L_shr( pm_fx, add( sub( q_tmp2, q_tmp1 ), 1 ) ) );
@@ -4071,9 +3967,6 @@ static void eig2x2_fx( 
        move16();

    }



    check( __D_fx[0], __D_q, D_fx[0], *q_D, 1 << 26 );

    check( __D_fx[1], __D_q, D_fx[1], *q_D, 1 << 26 );



    /* Numeric case, when input is practically zeros */

    // IF( D_fx[0] < EPSILON_FX )
@@ -4086,15 +3979,6 @@ static void eig2x2_fx( 
        *q_U = Q31;

        move16();



        for ( int i = 0; i < BINAURAL_CHANNELS; ++i )

        {

            for ( int j = 0; j < BINAURAL_CHANNELS; ++j )

            {

                check( __Ure_fx[i][j], __U_q, Ure_fx[i][j], *q_U, 1 << 2 );

                check( __Uim_fx[i][j], __U_q, Uim_fx[i][j], *q_U, 1 << 2 );

            }

        }



        return;

    }
@@ -4112,15 +3996,6 @@ static void eig2x2_fx( 
            *q_U = Q30;

            move16();



            for ( int i = 0; i < BINAURAL_CHANNELS; ++i )

            {

                for ( int j = 0; j < BINAURAL_CHANNELS; ++j )

                {

                    check( __Ure_fx[i][j], __U_q, Ure_fx[i][j], *q_U, 1 << 2 );

                    check( __Uim_fx[i][j], __U_q, Uim_fx[i][j], *q_U, 1 << 2 );

                }

            }



            return;

        }

    }
@@ -4135,15 +4010,6 @@ static void eig2x2_fx( 
            *q_U = Q30;

            move16();



            for ( int i = 0; i < BINAURAL_CHANNELS; ++i )

            {

                for ( int j = 0; j < BINAURAL_CHANNELS; ++j )

                {

                    check( __Ure_fx[i][j], __U_q, Ure_fx[i][j], *q_U, 1 << 2 );

                    check( __Uim_fx[i][j], __U_q, Uim_fx[i][j], *q_U, 1 << 2 );

                }

            }



            return;

        }

    }
@@ -4172,18 +4038,10 @@ static void eig2x2_fx( 
            move16();

        }



        check( __diff_fx[ch][0], __diff_q, tmp1, q_tmp1, 1 << 26 );

        check( __diff_fx[ch][1], __diff_q, tmp2, q_tmp1, 1 << 26 );



        IF( GT_32( L_abs( tmp2 ), L_abs( tmp1 ) ) )

        {

            check( __as[ch], 0, 1, 0, 0 );



            s_fx = tmp2;

            move32();



            check( __s_fx[ch], __s_q[ch], s_fx, q_tmp1, 1 << 26 );



            exp = sub( norm_l( s_fx ), 1 );

            tmp2 = Mpy_32_32( s_fx, s_fx );

            q_tmp2 = sub( add( q_tmp1, q_tmp1 ), 31 );
@@ -4198,8 +4056,6 @@ static void eig2x2_fx( 
            normVal_fx = Sqrt32( tmp2, &exp ); // q_tmp2

            q_tmp2 = sub( 31, exp );



            check( __nval_fx[ch], __nval_q[ch], normVal_fx, q_tmp2, 1 << 26 );



            q_diff = sub( q_c, q_tmp1 );

            IF( q_diff > 0 )

            {
@@ -4223,10 +4079,6 @@ static void eig2x2_fx( 
            move32();

            q_U_1 = sub( add( q_tmp1, q_tmp2 ), 31 );



            check( __Ure_fx[0][ch], __U_q, Ure_fx[0][ch], q_U_1, 1 << 26 );

            check( __Ure_fx[1][ch], __U_q, Ure_fx[1][ch], q_U_1, 1 << 26 );

            check( __Uim_fx[1][ch], __U_q, Uim_fx[1][ch], q_U_1, 1 << 26 );



            IF( q_U_2 != 0 )

            {

                q_diff = sub( q_U_2, q_U_1 );
@@ -4258,13 +4110,9 @@ static void eig2x2_fx( 
        }

        ELSE

        {

            check( __as[ch], 0, 0, 0, 0 );



            s_fx = tmp1;

            move32();



            check( __s_fx[ch], __s_q[ch], s_fx, q_tmp1, 1 << 26 );



            exp = sub( norm_l( s_fx ), 1 );

            tmp2 = Mpy_32_32( s_fx, s_fx );

            q_tmp2 = sub( add( q_tmp1, q_tmp1 ), 31 );
@@ -4279,8 +4127,6 @@ static void eig2x2_fx( 
            normVal_fx = Sqrt32( tmp2, &exp ); // q_tmp2

            q_tmp2 = sub( 31, exp );



            check( __nval_fx[ch], __nval_q[ch], normVal_fx, q_tmp2, 1 << 26 );



            q_diff = sub( q_c, q_tmp1 );

            IF( q_diff > 0 )

            {
@@ -4304,10 +4150,6 @@ static void eig2x2_fx( 
            move32();

            q_U_2 = sub( add( q_tmp1, q_tmp2 ), 31 );



            check( __Ure_fx[1][ch], __U_q, Ure_fx[1][ch], q_U_2, 1 << 26 );

            check( __Ure_fx[0][ch], __U_q, Ure_fx[0][ch], q_U_2, 1 << 26 );

            check( __Uim_fx[0][ch], __U_q, Uim_fx[0][ch], q_U_2, 1 << 26 );



            IF( q_U_1 != 0 )

            {

                q_diff = sub( q_U_2, q_U_1 );
@@ -4350,15 +4192,6 @@ static void eig2x2_fx( 
        *q_U = q_U_2;

        move16();

    }



    for ( int i = 0; i < BINAURAL_CHANNELS; ++i )

    {

        for ( int j = 0; j < BINAURAL_CHANNELS; ++j )

        {

            check( __Ure_fx[i][j], __U_q, Ure_fx[i][j], *q_U, 1 << 26 );

            check( __Uim_fx[i][j], __U_q, Uim_fx[i][j], *q_U, 1 << 26 );

        }

    }

#endif

    return;

}