Commit a791a152 authored by Nicolas Roussin's avatar Nicolas Roussin
Browse files

Optimize eig2x2_fx part 1.

parent 4048efa5

lib_dec/ivas_ism_metadata_dec_fx.c

+1 −1
Original line number Diff line number Diff line
@@ -30,7 +30,6 @@ 


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



#include "move.h"

#include <stdint.h>

#include "options.h"

#include "ivas_cnst.h"
@@ -40,6 +39,7 @@ 
#include "ivas_stat_enc.h"

#include <math.h>

#include "wmc_auto.h"

#include "move.h"

#include "ivas_prot_fx.h"

lib_rend/ivas_dirac_dec_binaural_functions_fx.c

+131 −0
Original line number Diff line number Diff line
@@ -3518,6 +3518,31 @@ static void ivas_dirac_dec_binaural_check_and_switch_transports_headtracked_fx( 
    return;

}



#if 0

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



static void eig2x2_fx(

    const Word32 E1_fx, /*q_E*/
@@ -3532,6 +3557,7 @@ static void eig2x2_fx( 
    Word32 D_fx[BINAURAL_CHANNELS], /*q_D*/

    Word16 *q_D )

{

#if 0

    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;
@@ -3672,7 +3698,112 @@ static void eig2x2_fx( 
        *q_D = sub( q_tmp2, 1 );

        move16();

    }

#else

    Word16 chA, chB, ch;

    FOR( chA = 0; chA < BINAURAL_CHANNELS; chA++ )

    {

        FOR( chB = 0; chB < BINAURAL_CHANNELS; chB++ )

        {

            Ure_fx[chA][chB] = 0;

            move32();

            Uim_fx[chA][chB] = 0;

            move32();

        }

    }



    // ===================================================================================================

    /*crossSquare_fx = (c_re * c_re) + (c_im * c_im)

    a_fx = (e1 + e2) * (e1 + e2) - 4.0f * ((e1 * e2) - crossSquare_fx) = (e1 - e2)^2 + 4 * crossSquare_fx

    pm_fx = 0.5f * sqrtf(max(0.0f, a_fx))

    add_fx = 0.5f * (e1 + e2)*/



    Word16 q1, q2, qm, qd, lshift;



    // (e1 - e2)^2 -> Q: 2 * q_E

    q1 = shl( q_E, 1 );

    // 4 * ((c_re * c_re) + (c_im * c_im)) -> Q: 2 * q_C - 2

    q2 = sub( shl( q_C, 1 ), 2 );



    // (e1 - e2)^2

    Word32 es = L_sub( E1_fx, E2_fx );

    Word64 es2 = W_mult0_32_32( es, es );

    lshift = sub( W_norm( es2 ), 1 );

    es2 = W_shl( es2, lshift );

    q1 = add( q1, lshift );

    if ( !es2 )

    {

        q1 = 63;

        move16();

    }



    // 4 * ((c_re * c_re) + (c_im * c_im))

    Word64 cs = W_add( W_mult0_32_32( Cre_fx, Cre_fx ), W_mult0_32_32( Cim_fx, Cim_fx ) ); // 2*q_C-2

    lshift = sub( W_norm( cs ), 1 );

    cs = W_shl( cs, lshift );

    q2 = add( q2, lshift );

    if ( !cs )

    {

        q2 = 63;

        move16();

    }



    Word32 crossSquare_fx = (Word32) ( cs >> 32 ); // FIXME

    Word16 q_crossSquare = 2 * q_C + lshift - 32;  // FIXME



    // a = max(0, (e1 - e2)^2 + 4 * crossSquare_fx)

    qm = s_min( q1, q2 );

    qd = sub( q1, q2 );

    Word64 a = W_max( W_add( W_shr( es2, s_max( qd, 0 ) ), W_shr( cs, negate( s_min( qd, 0 ) ) ) ), 0 );



    // pm = 0.5f * sqrtf(a)

    // a = 0.5f * ( E1 + E2 );

    lshift = W_norm( a );

    Word32 pm = W_extract_h( W_shl( a, lshift ) );

    Word16 e = sub( sub( 63, lshift ), qm );

    pm = L_shr( Sqrt32( pm, &e ), 1 );

    q2 = sub( 31, e );

    // check( pm, q2, pm_fx, q_tmp2, 1 << 16 );

    a = L_add( E1_fx, E2_fx );

    lshift = sub( norm_l( a ), 1 );

    a = W_shl( a, lshift );

    q1 = add( add( q_E, 1 ), lshift );

    // check( a, q1, add_fx, q_tmp1, 1 << 16 );



    Word32 add_fx = a;  // FIXME

    Word16 q_tmp1 = q1; // FIXME

    Word32 pm_fx = pm;  // FIXME

    Word16 q_tmp2 = q2; // FIXME



    // D[0] = add + pm;

    // D[1] = max( 0.0f, add - pm );

    qm = s_min( q1, q2 );

    qd = sub( q1, q2 );

    a = W_shr( a, s_max( qd, 0 ) );

    pm = W_shr( pm, negate( s_min( qd, 0 ) ) );

    Word32 d0 = L_add( a, pm );

    Word32 d1 = L_max( L_sub( a, pm ), 0 );

    // check( d0, qm, D_fx[0], *q_D, 1 << 16 );

    // check( d1, qm, D_fx[1], *q_D, 1 << 16 );



    D_fx[0] = d0; // FIXME

    D_fx[1] = d1; // FIXME

    *q_D = qm;    // FIXME



    Word32 tmp1, tmp2, tmp3, e1, e2, s_fx, normVal_fx, c_re, c_im; // FIXME

    Word16 q_U_1, q_U_2, q_c, q_e, exp, exp_tmp3;                  // FIXME

    Word32 epsilon_mant = 1180591621;                              // FIXME

    Word16 epsilon_exp = -39;                                      // FIXME



    exp = sub( get_min_scalefactor( Cre_fx, Cim_fx ), 2 ); // FIXME

    c_re = L_shl( Cre_fx, exp );                           // FIXME

    c_im = L_shl( Cim_fx, exp );                           // FIXME

    q_c = add( q_C, exp );                                 // FIXME



    exp = sub( get_min_scalefactor( E1_fx, E2_fx ), 2 ); // FIXME

    e1 = L_shl( E1_fx, exp );                            // FIXME

    e2 = L_shl( E2_fx, exp );                            // FIXME

    q_e = add( q_E, exp );                               // FIXME

#endif

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

    // IF( D_fx[0] < EPSILON_FX )