Commit a18e470a authored by thomas dettbarn's avatar thomas dettbarn
Browse files

the magic shift has gotten more deterministic.

parent 0387efd8

lib_dec/ivas_svd_dec_fx.c

+17 −14
Original line number Diff line number Diff line
@@ -1004,13 +1004,14 @@ static void biDiagonalReductionLeft_64( 
{



#define HEADROOM_LEFT_1 2

#define HEADROOM_LEFT_2 10

#define HEADROOM_LEFT_3 15

#define HEADROOM_LEFT_4 15

#define HEADROOM_LEFT_2 ( 16 - norm_x_e0 / 4 ) // 12

#define HEADROOM_LEFT_3 ( 16 - norm_x_e0 / 4 ) // 14

#define HEADROOM_LEFT_4 ( 16 - norm_x_e0 / 4 ) // 15



    Word16 iCh, jCh;

    Word32 norm_x;

    Word16 norm_x_e;

    Word16 norm_x_e0;

    Word64 norm_64;

    ( *g ) = 0;

    ( *g_e ) = 0;
@@ -1030,6 +1031,7 @@ static void biDiagonalReductionLeft_64( 
            norm_64 = W_add( norm_64, W_mult0_32_32( tmp, tmp ) );

        }

        norm_x_e = W_norm( norm_64 );

        norm_x_e0 = W_norm( norm_64 );

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

    }

    IF( norm_x )
@@ -1076,15 +1078,14 @@ static void biDiagonalReductionLeft_64( 
            norm_64 = 0;

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

            {

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

                factor2 = W_extract_l( W_shr( singularVectors_Left_64[jCh][iCh], HEADROOM_LEFT_3 ) );

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

                factor2 = W_extract_h( W_shl( singularVectors_Left_64[jCh][iCh], 32 - HEADROOM_LEFT_3 ) );

                norm_64 = W_add( norm_64, W_mult0_32_32( factor1, factor2 ) );

            }

            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 = norm_x_e - r_e + ( 32 + HEADROOM_LEFT_1 - 2 * HEADROOM_LEFT_3 - HEADROOM_LEFT_4 );

            magic_shift = norm_x_e - r_e + ( 30 + 2 * HEADROOM_LEFT_1 - 2 * HEADROOM_LEFT_3 - HEADROOM_LEFT_4 );

            magic_shift = ( norm_x_e - 2 * HEADROOM_LEFT_3 ) - ( r_e - 2 * HEADROOM_LEFT_1 ) + ( 32 - HEADROOM_LEFT_4 ) - 2;

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

            {

                factor1 = W_extract_h( W_shl( singularVectors_Left_64[jCh][currChannel], 32 - HEADROOM_LEFT_4 ) );
@@ -1111,13 +1112,14 @@ static void biDiagonalReductionRight_64( 
    Word16 iCh, jCh;

    Word32 norm_x;

    Word16 norm_x_e;

    Word16 norm_x_e0;

    Word64 norm_64;

    Word16 idx;



#define HEADROOM_RIGHT_1 2

#define HEADROOM_RIGHT_2 10

#define HEADROOM_RIGHT_3 15

#define HEADROOM_RIGHT_4 15

#define HEADROOM_RIGHT_2 ( 16 - norm_x_e0 / 4 ) // 12

#define HEADROOM_RIGHT_3 ( 16 - norm_x_e0 / 4 ) // 14

#define HEADROOM_RIGHT_4 ( 16 - norm_x_e0 / 4 ) // 15





    ( *g ) = 0;
@@ -1136,6 +1138,7 @@ static void biDiagonalReductionRight_64( 
            norm_64 = W_add( norm_64, W_mult0_32_32( tmp, tmp ) );

        }

        norm_x_e = W_norm( norm_64 );

        norm_x_e0 = W_norm( norm_64 );

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

        move16();
@@ -1184,15 +1187,15 @@ static void biDiagonalReductionRight_64( 
                move64();

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

                {

                    factor1 = W_extract_l( W_shr( singularVectors_Left_64[iCh][jCh], HEADROOM_RIGHT_3 ) );

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

                    factor1 = W_extract_h( W_shl( singularVectors_Left_64[iCh][jCh], 32 - HEADROOM_RIGHT_3 ) );

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

                    norm_64 = W_add( norm_64, W_mult0_32_32( factor1, factor2 ) );

                }



                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 = norm_x_e - r_e + ( 30 + 2 * HEADROOM_RIGHT_1 - 2 * HEADROOM_RIGHT_3 - HEADROOM_RIGHT_4 );

                magic_shift = ( norm_x_e - 2 * HEADROOM_RIGHT_3 ) - ( r_e - 2 * HEADROOM_RIGHT_1 ) + ( 32 - HEADROOM_RIGHT_4 ) - 2;



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

                {