Commit 23009853 authored by Arthur Tritthart's avatar Arthur Tritthart
Browse files

added minor WMOPS tuning for SVD module, extends MR1010

parent e6334462

lib_com/options.h

+2 −0
Original line number Diff line number Diff line
@@ -150,3 +150,5 @@ 
#define FIX_ISSUE_1237                          /* VA: replacement of Copy_Scale_sig_16_32_DEPREC() that are doing 16 bits left shift by Copy_Scale_sig_16_32_no_sat() */

#define FIX_ISSUE_1237_KEEP_EVS_BE              /* VA: Fix to keep EVS bitexactness to 26.444 */

#endif

#define FIX_MINOR_SVD_WMOPS_MR1010X             /* FhG: Minor WMOPS tuning, bit-exact to previous version, saves about 8.2 WMOPS for MR1010 */

lib_dec/ivas_svd_dec.c

+41 −0
Original line number Diff line number Diff line
@@ -919,6 +919,7 @@ static void ApplyRotation_fx( 
    *g = BASOP_Util_Add_Mant32Exp( Mpy_32_32( c, x12 ), add( c_e, x12_e ), Mpy_32_32( L_negate( s ), x11 ), add( s_e, x11_e ), g_e ); /* exp(g_e) */

    move32();



#ifndef FIX_MINOR_SVD_WMOPS_MR1010X

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

    {

        x11 = singularVector[ch][currentIndex2];
@@ -934,6 +935,24 @@ static void ApplyRotation_fx( 
        singularVector[ch][currentIndex1] = L_shl_sat( singularVector[ch][currentIndex1], temp_exp ); /* exp(temp_exp) */

        move32();

    }

#else

    Word32 s_neg = L_negate(s);

    Word32 temp;

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

    {

        x11 = singularVector[ch][currentIndex2];

        move32();

        x12 = singularVector[ch][currentIndex1];

        move32();

        temp = BASOP_Util_Add_Mant32Exp( Mpy_32_32( c, x11 ), c_e, Mpy_32_32( s, x12 ), s_e, &temp_exp ); /* exp(temp_exp) */

        singularVector[ch][currentIndex2] = L_shl_sat( temp, temp_exp ); /* exp(temp_exp) */

        move32();

        temp = BASOP_Util_Add_Mant32Exp( Mpy_32_32( c, x12 ), c_e, Mpy_32_32( s_neg, x11 ), s_e, &temp_exp ); /* exp(temp_exp) */

        singularVector[ch][currentIndex1] = L_shl_sat( temp, temp_exp ); /* exp(temp_exp) */

        move32();

    }



#endif



    return;

}
@@ -1160,6 +1179,7 @@ IF( LT_16( currChannel, nChannelsL ) ) /* i <= m */ 
        L_temp = Sqrt32( norm_x, &L_temp_e );

        L_temp = L_shl_r( L_temp, L_temp_e ); // Q31

        //( *g ) = L_negate( GE_32( singularVectors[currChannel][idx], 0 ) ? L_temp : L_negate( L_temp ) );

#ifndef FIX_MINOR_SVD_WMOPS_MR1010X

        IF( singularVectors[currChannel][idx] >= 0 )

        {

            ( *g ) = L_negate( L_temp );
@@ -1170,6 +1190,14 @@ IF( LT_16( currChannel, nChannelsL ) ) /* i <= m */ 
            ( *g ) = L_negate( L_negate( L_temp ) );

            move32();

        }

#else

        if ( singularVectors[currChannel][idx] >= 0 )

        {

            L_temp = L_negate( L_temp );

        }

        ( *g ) = L_temp ;

        move32();

#endif



#ifndef FIX_1010_OPT_SINGLE_RESCALE

        r = BASOP_Util_Add_Mant32Exp( Mpy_32_32( ( *g ), singularVectors[currChannel][idx] ), sing_exp[currChannel], -norm_x, norm_x_e, &r_e );                       /* exp(r_e) */
@@ -1868,6 +1896,7 @@ static Word32 maxWithSign_fx( 
    const Word32 a /* Qx */

)

{

#ifndef FIX_MINOR_SVD_WMOPS_MR1010X

    IF( GT_32( L_abs( a ), SVD_MINIMUM_VALUE_FX ) )

    {

        return a;
@@ -1880,6 +1909,18 @@ static Word32 maxWithSign_fx( 
    {

        return SVD_MINIMUM_VALUE_FX;

    }

#else

    Word32 result;

    if (a >= 0)

    {

        result = L_max( a, SVD_MINIMUM_VALUE_FX );

    }

    if (a < 0)

    {

        result = L_min( a, -SVD_MINIMUM_VALUE_FX );

    }

     return result;

#endif

}



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