Commit f47fd3d8 authored by thomas dettbarn's avatar thomas dettbarn
Browse files

More deterministic "magic_shifts" in biDiagonalReductionLeft_64() and... 

More deterministic "magic_shifts" in biDiagonalReductionLeft_64() and biDiagonalReductionRight_64().
parent f5737f6c

lib_dec/ivas_svd_dec_fx.c

+19 −18
Original line number Diff line number Diff line
@@ -1004,8 +1004,9 @@ static void biDiagonalReductionLeft_64( 
{



#define HEADROOM_LEFT_1 1 

#define HEADROOM_LEFT_2 ( HEADROOM_LEFT_1 + 1 )

#define HEADROOM_LEFT_2 2

#define HEADROOM_LEFT_3 2



    Word16 iCh, jCh;

    Word32 norm_x;

    Word16 norm_x_e;
@@ -1048,28 +1049,28 @@ static void biDiagonalReductionLeft_64( 
            ( *g ) = L_negate( *g );

        }

        factor2 = W_extract_l( W_shr( singularVectors_Left_64[currChannel][currChannel], HEADROOM_LEFT_1 ) );

        tmp_e = shr( sub( norm_x_e, 1 ), 1 );

        tmp_e = sub( HEADROOM_LEFT_1, ( *g_e ) );

        tmpmul = W_mult0_32_32( ( *g ), factor2 );

        tmpmul = W_shr( tmpmul, tmp_e );

        r_64 = W_sub( tmpmul, norm_64 );

        r_e = W_norm( r_64 );

        r = W_extract_h( W_shl( r_64, r_e ) );





        invVal_e = sub( add( 1, ( HEADROOM_LEFT_1 + HEADROOM_LEFT_1 ) ), r_e );

        invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, r, &invVal_e );

        invVal_e = 0;

        move16();

        invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, maxWithSign_fx( r ), &invVal_e );





        tmp_e = add( 31, sub( HEADROOM_LEFT_1, *g_e ) );

        tmp_e = sub( 32, *g_e );

        singularVectors_Left_64[currChannel][currChannel] = W_sub( singularVectors_Left_64[currChannel][currChannel], W_shr( W_deposit32_h( *g ), tmp_e ) ); // here, the exponent goes up.



        FOR( iCh = add( currChannel, 1 ); iCh < nChannelsC; iCh++ )

        {

            Word16 magic_shift;

            Word32 factor1;

            Word32 factor2;

            Word32 f; // = norm / r

                      //            Word16 f_e;      // not really needed

            Word16 magic_shift;



            norm_64 = 0;

            for ( jCh = currChannel; jCh < nChannelsL; jCh++ )
@@ -1081,10 +1082,10 @@ static void biDiagonalReductionLeft_64( 
            norm_x_e = W_norm( norm_64 );

            norm_x = W_extract_h( W_shl( norm_64, norm_x_e ) );

            f = Mpy_32_32( norm_x, invVal );

            magic_shift = 31 - HEADROOM_LEFT_3 + norm_x_e - r_e;

            magic_shift = norm_x_e-r_e+(28-HEADROOM_LEFT_3);	// works with headroom_left: 1,2,2

            FOR( jCh = currChannel; jCh < nChannelsL; jCh++ )

            {

                factor1 = W_extract_l( W_shr( singularVectors_Left_64[jCh][currChannel], HEADROOM_LEFT_3 ) );

                factor1 = W_extract_h( W_shl( singularVectors_Left_64[jCh][currChannel], 32-HEADROOM_LEFT_3 ) );

                singularVectors_Left_64[jCh][iCh] = W_add( singularVectors_Left_64[jCh][iCh], W_shr( W_mult0_32_32( f, factor1 ), magic_shift ) );

            }

        }
@@ -1112,8 +1113,8 @@ static void biDiagonalReductionRight_64( 
    Word16 idx;



#define HEADROOM_RIGHT_1 2

#define HEADROOM_RIGHT_2 ( HEADROOM_RIGHT_1 + 1 )

#define HEADROOM_RIGHT_3 3

#define HEADROOM_RIGHT_2 2

#define HEADROOM_RIGHT_3 2





    ( *g ) = 0;
@@ -1159,9 +1160,9 @@ static void biDiagonalReductionRight_64( 
            move32();

            move16();

            factor2 = W_extract_l( W_shr( singularVectors_Left_64[currChannel][idx], HEADROOM_RIGHT_1 ) );

            tmp_e = sub( *g_e, HEADROOM_RIGHT_1 );

            tmp_e = sub( HEADROOM_RIGHT_1, *g_e );

            tmpmul = W_mult0_32_32( *g, factor2 );

            tmpmul = W_shl( tmpmul, tmp_e );

            tmpmul = W_shr( tmpmul, tmp_e );

            r_64 = W_sub( tmpmul, norm_64 );

            r_e = W_norm( r_64 );

            r = W_extract_h( W_shl( r_64, r_e ) );
@@ -1170,8 +1171,8 @@ static void biDiagonalReductionRight_64( 
            move16();

            invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, maxWithSign_fx( r ), &invVal_e );



            magic_shift = 32 - *g_e;

            singularVectors_Left_64[currChannel][idx] = W_sub( singularVectors_Left_64[currChannel][idx], W_shr( W_deposit32_h( *g ), magic_shift ) ); // here, the exponent goes up

            tmp_e = sub( 32, *g_e );

            singularVectors_Left_64[currChannel][idx] = W_sub( singularVectors_Left_64[currChannel][idx], W_shr( W_deposit32_h( *g ), tmp_e ) ); // here, the exponent goes up



            FOR( iCh = idx; iCh < nChannelsL; iCh++ )

            {
@@ -1188,11 +1189,11 @@ static void biDiagonalReductionRight_64( 
                norm_x_e = W_norm( norm_64 );

                norm_x = W_extract_h( W_shl( norm_64, norm_x_e ) );

                f = Mpy_32_32( norm_x, invVal );

                magic_shift = 25 + norm_x_e - r_e; // headroom 3 FIXME: Why does this work?

                magic_shift = norm_x_e-r_e   +29-(HEADROOM_RIGHT_3); // works with headroom_right: 2,2,2



                FOR( jCh = idx; jCh < nChannelsC; jCh++ )

                {

                    factor2 = W_extract_l( W_shr( singularVectors_Left_64[currChannel][jCh], HEADROOM_RIGHT_3 ) );

                    factor2 = W_extract_h( W_shl( singularVectors_Left_64[currChannel][jCh], 32-HEADROOM_RIGHT_3 ) );

                    singularVectors_Left_64[iCh][jCh] = W_add( singularVectors_Left_64[iCh][jCh], W_shr( W_mult0_32_32( f, factor2 ), magic_shift ) );

                }

            }