Commit 8cd5a1ce authored by thomas dettbarn's avatar thomas dettbarn
Browse files

renamed the macros for the headroom shifts.

parent 617c0edf

lib_dec/ivas_svd_dec_fx.c

+22 −26
Original line number Diff line number Diff line
@@ -1002,11 +1002,11 @@ static void biDiagonalReductionLeft_64( 
    Word32 *g,

    Word16 *g_e )

{



#define HEADROOM_LEFT_1 2

#define HEADROOM_LEFT_2 ( 16 - norm_x_e0 / 4 ) //  10

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

#define HEADROOM_LEFT_4 ( 16 - norm_x_e0 / 4 ) //  14

/* 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 )



    Word16 iCh, jCh;

    Word32 norm_x;
@@ -1027,7 +1027,7 @@ static void biDiagonalReductionLeft_64( 
        move64();

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

        {

            tmp = W_extract_l( W_shr( singularVectors_Left_64[jCh][currChannel], HEADROOM_LEFT_1 ) );	// q(sing)-H1			// exp(sing)+H1

            tmp = W_extract_l( W_shr( singularVectors_Left_64[jCh][currChannel], MAGIC_HEADROOM_1 ) );	// q(sing)-H1			// exp(sing)+H1

            norm_64 = W_add( norm_64, W_mult0_32_32( tmp, tmp ) );					// q(norm)=2*q(sing)-2*H1	// exp(norm)=2*exp(sing)+2*H1

        }

        norm_x_e = W_norm( norm_64 );
@@ -1044,16 +1044,16 @@ static void biDiagonalReductionLeft_64( 
        Word32 r, invVal;

        Word16 r_e, invVal_e;



        ( *g_e ) = add( sub( add( HEADROOM_LEFT_1, HEADROOM_LEFT_1 ), norm_x_e ), 1 );		// exp(g)=(2*H1-exp(norm_x)+1)

        ( *g_e ) = add( sub( add( MAGIC_HEADROOM_1, MAGIC_HEADROOM_1 ), norm_x_e ), 1 );		// exp(g)=(2*H1-exp(norm_x)+1)

        move16();

        ( *g ) = Sqrt32( norm_x, g_e );								// --> exp(g)=((2*H1-exp(norm_x)+1)/2)

        IF( GE_64( singularVectors_Left_64[currChannel][currChannel], 0 ) )

        {

            ( *g ) = L_negate( *g );

        }

        factor2 = W_extract_l( W_shr( singularVectors_Left_64[currChannel][currChannel], HEADROOM_LEFT_2 ) );	// q(factor2)=q(sing)-H2	exp(factor2)=exp(qsing)+H2

        tmp_e = sub( 2 * HEADROOM_LEFT_1 - HEADROOM_LEFT_2, ( *g_e ) );						

        tmpmul = W_mult0_32_32( ( *g ), factor2 );	// q(tmpmul)=q(g)+q(factor2)

        factor2 = W_extract_l( W_shr( singularVectors_Left_64[currChannel][currChannel], MAGIC_HEADROOM_2 ) );	// q(factor2)=q(sing)-H2	exp(factor2)=exp(qsing)+H2

        tmp_e = sub( 2 * MAGIC_HEADROOM_1 - MAGIC_HEADROOM_2, ( *g_e ) );						

        tmpmul = W_mult0_32_32( ( *g ), factor2 );	// q(tmpmul)=q(g)+q(factor2) --> q(tmpmul) ~= q(norm)

        tmpmul = W_shr( tmpmul, tmp_e );		// --> q(tmpmul)=q(g)+q(factor2)-(2*H1-H2-q(g))

        r_64 = W_sub( tmpmul, norm_64 );		// q(r_64)=max(q(tmpmul),q(norm))

        r_e = W_norm( r_64 );
@@ -1077,17 +1077,17 @@ 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 - HEADROOM_LEFT_3 ) );	// q(factor1) = q(sing)-H3

                factor2 = W_extract_h( W_shl( singularVectors_Left_64[jCh][iCh], 32 - HEADROOM_LEFT_3 ) );		// q(factor2) = q(sing)-H3

                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

                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 * HEADROOM_LEFT_3 ) - ( r_e - 2 * HEADROOM_LEFT_1 ) + ( 32 - HEADROOM_LEFT_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;

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

            {

                factor1 = W_extract_h( W_shl( singularVectors_Left_64[jCh][currChannel], 32 - HEADROOM_LEFT_4 ) );

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

            }

        }
@@ -1115,10 +1115,6 @@ static void biDiagonalReductionRight_64( 
    Word64 norm_64;

    Word16 idx;



#define HEADROOM_RIGHT_1 2

#define HEADROOM_RIGHT_2 ( 16 - norm_x_e0 / 4 ) // 10

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

#define HEADROOM_RIGHT_4 ( 16 - norm_x_e0 / 4 ) // 14





    ( *g ) = 0;
@@ -1133,7 +1129,7 @@ static void biDiagonalReductionRight_64( 
        FOR( jCh = idx; jCh < nChannelsC; jCh++ )

        {

            Word32 tmp;

            tmp = W_extract_l( W_shr( singularVectors_Left_64[currChannel][jCh], HEADROOM_RIGHT_1 ) );	// q(sing)-H1			// exp(sing)+H1	

            tmp = W_extract_l( W_shr( singularVectors_Left_64[currChannel][jCh], MAGIC_HEADROOM_1 ) );	// q(sing)-H1			// exp(sing)+H1	

            norm_64 = W_add( norm_64, W_mult0_32_32( tmp, tmp ) );					// q(norm)=2*q(sing)-2*H1	// exp(norm)=2*exp(sing)+2*H1

        }

        norm_x_e = W_norm( norm_64 );
@@ -1155,7 +1151,7 @@ static void biDiagonalReductionRight_64( 
            Word32 invVal;

            Word16 invVal_e;



            ( *g_e ) = add( sub( ( HEADROOM_RIGHT_1 + HEADROOM_RIGHT_1 ), norm_x_e ), 1 );	// exp(g)=(2*H1-exp(norm_x)+1)

            ( *g_e ) = add( sub( ( MAGIC_HEADROOM_1 + MAGIC_HEADROOM_1 ), norm_x_e ), 1 );	// exp(g)=(2*H1-exp(norm_x)+1)

            move16();

            ( *g ) = Sqrt32( norm_x, g_e );							// --> exp(g)=((2*H1-exp(norm_x)+1)/2)

            IF( GE_64( singularVectors_Left_64[currChannel][idx], 0 ) )
@@ -1164,8 +1160,8 @@ static void biDiagonalReductionRight_64( 
            }

            move32();

            move16();

            factor2 = W_extract_l( W_shr( singularVectors_Left_64[currChannel][idx], HEADROOM_RIGHT_2 ) );	// q(factor2)=q(sing)-H2	exp(factor2)=exp(qsing)+H2

            tmp_e = sub( 2 * HEADROOM_RIGHT_1 - HEADROOM_RIGHT_2, *g_e );

            factor2 = W_extract_l( W_shr( singularVectors_Left_64[currChannel][idx], MAGIC_HEADROOM_2 ) );	// q(factor2)=q(sing)-H2	exp(factor2)=exp(qsing)+H2

            tmp_e = sub( 2 * MAGIC_HEADROOM_1 - MAGIC_HEADROOM_2, *g_e );

            tmpmul = W_mult0_32_32( *g, factor2 );	// q(tmpmul)=q(g)+q(factor2)

            tmpmul = W_shr( tmpmul, tmp_e );		// --> q(tmpmul)=q(g)+q(factor2)-(2*H1-H2-q(g))

            r_64 = W_sub( tmpmul, norm_64 );		// q(r_64)=max(q(tmpmul),q(norm))
@@ -1185,19 +1181,19 @@ static void biDiagonalReductionRight_64( 
                move64();

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

                {

                    factor1 = W_extract_h( W_shl( singularVectors_Left_64[iCh][jCh], 32 - HEADROOM_RIGHT_3 ) );		// q(factor1) = q(sing)-H3

                    factor2 = W_extract_h( W_shl( singularVectors_Left_64[currChannel][jCh], 32 - HEADROOM_RIGHT_3 ) );	// q(factor2) = q(sing)-H3

                    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

                    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 * HEADROOM_RIGHT_3 ) - ( r_e - 2 * HEADROOM_RIGHT_1 ) + ( 32 - HEADROOM_RIGHT_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;



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

                {

                    factor2 = W_extract_h( W_shl( singularVectors_Left_64[currChannel][jCh], 32 - HEADROOM_RIGHT_4 ) );

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

                }

            }