Commit bdde93eb authored by thomas dettbarn's avatar thomas dettbarn
Browse files

carefully replacing shifts.

parent 0e413edb

lib_dec/ivas_svd_dec_fx.c

+11 −11
Original line number Diff line number Diff line
@@ -1005,7 +1005,7 @@ static void biDiagonalReductionLeft_64( 


#define HEADROOM_LEFT_1 1

#define HEADROOM_LEFT_2 ( HEADROOM_LEFT_1 + 1 )

#define HEADROOM_LEFT_3 2



    Word16 iCh, jCh;

    Word32 norm_x;

    Word16 norm_x_e;
@@ -1048,15 +1048,15 @@ 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( ( *g_e ), HEADROOM_LEFT_1 );

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

        tmpmul = W_shr( tmpmul, tmp_e );

        tmpmul = W_shl( 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_e = sub( add( 1, ( HEADROOM_LEFT_1 + HEADROOM_LEFT_1 ) ), r_e );

        invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, r, &invVal_e );
@@ -1065,11 +1065,11 @@ static void biDiagonalReductionLeft_64( 


        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 +1081,10 @@ 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 = add( add( norm_x_e, 23 ), sub( add( 1, ( HEADROOM_LEFT_1 + HEADROOM_LEFT_1 ) ), r_e ) );

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

            {

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

                factor1 = W_extract_l( W_shr( singularVectors_Left_64[jCh][currChannel], HEADROOM_LEFT_2 ) );

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

            }

        }
@@ -1113,7 +1113,6 @@ static void biDiagonalReductionRight_64( 


#define HEADROOM_RIGHT_1 2

#define HEADROOM_RIGHT_2 ( HEADROOM_RIGHT_1 + 1 )

#define HEADROOM_RIGHT_3 3





    ( *g ) = 0;
@@ -1133,6 +1132,7 @@ static void biDiagonalReductionRight_64( 
        }

        norm_x_e = W_norm( norm_64 );

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

        norm_x_e = add( sub( ( HEADROOM_RIGHT_1 + HEADROOM_RIGHT_1 ), norm_x_e ), 1 );

        move16();



        IF( norm_x )
@@ -1149,7 +1149,7 @@ static void biDiagonalReductionRight_64( 
            Word32 invVal;

            Word16 invVal_e;



            ( *g_e ) = add( sub( ( HEADROOM_RIGHT_1 + HEADROOM_RIGHT_1 ), norm_x_e ), 1 );

            ( *g_e ) = norm_x_e;

            move16();

            ( *g ) = Sqrt32( norm_x, g_e );

            IF( GE_64( singularVectors_Left_64[currChannel][idx], 0 ) )
@@ -1188,11 +1188,11 @@ 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 = 25 + norm_x_e - r_e; // FIXME: Why does this work?



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

                {

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

                    factor2 = W_extract_l( W_shr( singularVectors_Left_64[currChannel][jCh], HEADROOM_RIGHT_2 ) );

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

                }

            }