Commit 8e5931aa authored by thomas dettbarn's avatar thomas dettbarn
Browse files

biDiagonalReductionLeft_64() and biDiagonalReductionRight_64() are ready for a... 

biDiagonalReductionLeft_64() and biDiagonalReductionRight_64() are ready for a non-draft merge request.
parent 31bfe505

lib_dec/ivas_svd_dec_fx.c

+19 −11
Original line number Diff line number Diff line
@@ -1004,9 +1004,9 @@ static void biDiagonalReductionLeft_64( 
{

/* TODO: For some reason, this is optimal. But why? why not ( 32 - 2 * MAGIC_HEADROOM_1 - norm_x_e0 + 1 ) , for example? */

#define MAGIC_HEADROOM_1 2

#define MAGIC_HEADROOM_2 ( 16 - norm_x_e0 / 4 )

#define MAGIC_HEADROOM_3 ( 16 - norm_x_e0 / 4 )

#define MAGIC_HEADROOM_4 ( 16 - norm_x_e0 / 4 )

#define MAGIC_HEADROOM_2 ( sub( 16, shr( norm_x_e0, 2 ) ) )

#define MAGIC_HEADROOM_3 ( sub( 16, shr( norm_x_e0, 2 ) ) )

#define MAGIC_HEADROOM_4 ( sub( 16, shr( norm_x_e0, 2 ) ) )



    Word16 iCh, jCh;

    Word32 norm_x;
@@ -1077,17 +1077,21 @@ static void biDiagonalReductionLeft_64( 
            norm_64 = 0;

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

            {

                factor1 = W_extract_h( W_shl( singularVectors_Left_64[jCh][currChannel], 32 - MAGIC_HEADROOM_3 ) ); // q(factor1) = q(sing)-H3

                factor2 = W_extract_h( W_shl( singularVectors_Left_64[jCh][iCh], 32 - MAGIC_HEADROOM_3 ) );         // q(factor2) = q(sing)-H3

                factor1 = W_extract_h( W_shl( singularVectors_Left_64[jCh][currChannel], sub( 32 , MAGIC_HEADROOM_3 ) ) ); // q(factor1) = q(sing)-H3

                factor2 = W_extract_h( W_shl( singularVectors_Left_64[jCh][iCh], sub( 32, MAGIC_HEADROOM_3 ) ) );         // q(factor2) = q(sing)-H3

                norm_64 = W_add( norm_64, W_mult0_32_32( factor1, factor2 ) );                                      // q(norm)=2*q(sing)-2*H3

            }

            norm_x_e = W_norm( norm_64 );

            norm_x = W_extract_h( W_shl( norm_64, norm_x_e ) ); // Note: different norm

            f = Mpy_32_32( norm_x, invVal );                    // q(f)=q(norm_x)-q(invVal)

            magic_shift = ( norm_x_e - 2 * MAGIC_HEADROOM_3 ) - ( r_e - 2 * MAGIC_HEADROOM_1 ) + ( 32 - MAGIC_HEADROOM_4 ) - 2 * invVal_e;

//            magic_shift = ( norm_x_e - 2 * MAGIC_HEADROOM_3 ) - ( r_e - 2 * MAGIC_HEADROOM_1 ) + ( 32 - MAGIC_HEADROOM_4 ) - 2 * invVal_e;

            magic_shift = sub( norm_x_e, shl( MAGIC_HEADROOM_3, 1) );

            magic_shift = sub( magic_shift, sub( r_e, ( shl( MAGIC_HEADROOM_1, 1 ) ) ) );

            magic_shift = add( magic_shift, sub( 32, MAGIC_HEADROOM_4 ) );

            magic_shift = sub( magic_shift, shl( invVal_e, 1 ) );

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

            {

                factor1 = W_extract_h( W_shl( singularVectors_Left_64[jCh][currChannel], 32 - MAGIC_HEADROOM_4 ) );

                factor1 = W_extract_h( W_shl( singularVectors_Left_64[jCh][currChannel], sub( 32, MAGIC_HEADROOM_4 ) ) );

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

            }

        }
@@ -1180,19 +1184,23 @@ static void biDiagonalReductionRight_64( 
                move64();

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

                {

                    factor1 = W_extract_h( W_shl( singularVectors_Left_64[iCh][jCh], 32 - MAGIC_HEADROOM_3 ) );         // q(factor1) = q(sing)-H3

                    factor2 = W_extract_h( W_shl( singularVectors_Left_64[currChannel][jCh], 32 - MAGIC_HEADROOM_3 ) ); // q(factor2) = q(sing)-H3

                    factor1 = W_extract_h( W_shl( singularVectors_Left_64[iCh][jCh], sub( 32, MAGIC_HEADROOM_3 ) ) );         // q(factor1) = q(sing)-H3

                    factor2 = W_extract_h( W_shl( singularVectors_Left_64[currChannel][jCh], sub( 32, MAGIC_HEADROOM_3 ) ) ); // q(factor2) = q(sing)-H3

                    norm_64 = W_add( norm_64, W_mult0_32_32( factor1, factor2 ) );                                      // q(norm)=2*q(sing)-2*H3

                }



                norm_x_e = W_norm( norm_64 );

                norm_x = W_extract_h( W_shl( norm_64, norm_x_e ) ); // Note: different norm

                f = Mpy_32_32( norm_x, invVal );                    // q(f)=q(norm_x)-q(invVal)

                magic_shift = ( norm_x_e - 2 * MAGIC_HEADROOM_3 ) - ( r_e - 2 * MAGIC_HEADROOM_1 ) + ( 32 - MAGIC_HEADROOM_4 ) - 2 * invVal_e;

                // magic_shift = ( norm_x_e - 2 * MAGIC_HEADROOM_3 ) - ( r_e - 2 * MAGIC_HEADROOM_1 ) + ( 32 - MAGIC_HEADROOM_4 ) - 2 * invVal_e;

                magic_shift = sub( norm_x_e, shl( MAGIC_HEADROOM_3, 1) );

                magic_shift = sub( magic_shift, sub( r_e, ( shl( MAGIC_HEADROOM_1, 1 ) ) ) );

                magic_shift = add( magic_shift, sub( 32, MAGIC_HEADROOM_4 ) );

                magic_shift = sub( magic_shift, shl( invVal_e, 1 ) );



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

                {

                    factor2 = W_extract_h( W_shl( singularVectors_Left_64[currChannel][jCh], 32 - MAGIC_HEADROOM_4 ) );

                    factor2 = W_extract_h( W_shl( singularVectors_Left_64[currChannel][jCh], sub( 32, MAGIC_HEADROOM_4 ) ) );

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

                }

            }