Commit 93c38c5d authored by thomas dettbarn's avatar thomas dettbarn
Browse files

Brought the code to the same level as the svd_optimization_branch.

parent 6ea437ed

lib_dec/ivas_svd_dec_fx.c

+34 −33
Original line number Diff line number Diff line
@@ -69,11 +69,11 @@ static void HouseholderReduction_fx( 
static void biDiagonalReductionLeft_64(

    Word64 singularVectors_Left_64[MAX_OUTPUT_CHANNELS][MAX_OUTPUT_CHANNELS],

    Word16 bitwindow,

    Word32 singularValues[MAX_OUTPUT_CHANNELS],    /* exp(singularValues_e) */

    Word16 singularValues_e[MAX_OUTPUT_CHANNELS],

    const Word16 nChannelsL,  /* Q0 */

    const Word16 nChannelsC,  /* Q0 */

    const Word16 currChannel /* Q0 */

    const Word16 currChannel, /* Q0 */

    Word32 *g, /* Q31 */

    Word16 *g_e

);



static void biDiagonalReductionRight_64(
@@ -877,8 +877,12 @@ static void HouseholderReduction_fx( 
#ifdef	MYCHANGES



    Word64 singularVectors_Left_64[MAX_OUTPUT_CHANNELS][MAX_OUTPUT_CHANNELS];

    Word32 g_fx = 0;

    Word16 g_e = 0;

    Word32 g_left_fx = 0;

    Word16 g_left_e = 0;

    move32();

    move16();

    Word32 g_right_fx = 0;

    Word16 g_right_e = 0;

    move32();

    move16();
@@ -895,7 +899,6 @@ static void HouseholderReduction_fx( 
    Word16 iCh, jCh;

    Word16 singularVectors_Left_fx_e[MAX_OUTPUT_CHANNELS][MAX_OUTPUT_CHANNELS];

#ifdef	 MYCHANGES

	push_wmops("HouseholderReduction_fx 64");

    FOR( jCh = 0; jCh < nChannelsL; jCh++ )

    {

        FOR( iCh = 0; iCh < nChannelsC; iCh++ )
@@ -909,23 +912,26 @@ static void HouseholderReduction_fx( 
	    bitwindow=1;

	    biDiagonalReductionLeft_64(

			    singularVectors_Left_64,bitwindow,

			    singularValues_fx,singularValues_fx_e,

			    nChannelsL,

			    nChannelsC,

			    nCh

			    nCh,

                            &g_left_fx,

                            &g_left_e

			    );

	    singularValues_fx_e[nCh]=add(singularVectors_Left_e,singularValues_fx_e[nCh]);

	    secDiag_fx[nCh]=g_fx;

	    singularValues_fx[nCh]=g_left_fx;

            move32();

	    secDiag_fx_e[nCh]=add(singularVectors_Left_e,g_e);

	    singularValues_fx_e[nCh]=add(singularVectors_Left_e,g_left_e);

	    secDiag_fx[nCh]=g_right_fx; /* from the previous channel */

	    move32();

	    secDiag_fx_e[nCh]=add(singularVectors_Left_e,g_right_e);

	    bitwindow=2;

	    biDiagonalReductionRight_64(

			    singularVectors_Left_64,bitwindow,

			    nChannelsL,

			    nChannelsC,

			    nCh,

			    &g_fx,

			    &g_e	

			    &g_right_fx,

			    &g_right_e	

			    );

		{

        		Word16 L_temp_e;
@@ -953,7 +959,6 @@ static void HouseholderReduction_fx( 
	    }

        }

    }

	pop_wmops();

#else



    FOR( jCh = 0; jCh < nChannelsL; jCh++ )
@@ -1002,22 +1007,22 @@ static void HouseholderReduction_fx( 
static void biDiagonalReductionLeft_64(

    Word64 singularVectors_Left_64[MAX_OUTPUT_CHANNELS][MAX_OUTPUT_CHANNELS],

    Word16 bitwindow,

    Word32 singularValues[MAX_OUTPUT_CHANNELS],    /* exp(singularValues_e) */

    Word16 singularValues_e[MAX_OUTPUT_CHANNELS],

    const Word16 nChannelsL,  /* Q0 */

    const Word16 nChannelsC,  /* Q0 */

    const Word16 currChannel /* Q0 */

    const Word16 currChannel, /* Q0 */

    Word32 *g,

    Word16 *g_e

)

{







    Word16 iCh, jCh;

    Word32 norm_x, g;

    Word16 norm_x_e, g_e;

    Word32 norm_x;

    Word16 norm_x_e;

    Word64 norm_64;

    g=0;

    g_e=0;

    ( *g )=0;

    ( *g_e) =0;

    move32();

    move16();

    norm_x=0;
@@ -1047,16 +1052,16 @@ static void biDiagonalReductionLeft_64( 
        Word32 r, invVal;

        Word16 r_e, invVal_e;



        g_e = norm_x_e;

        ( *g_e ) = norm_x_e;

        move16();

        g = Sqrt32( norm_x, &g_e);

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

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

        {

           g = L_negate( g );

           ( *g ) = L_negate( *g );

        }

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

        tmp_e=sub( g_e, bitwindow) ;

        tmpmul=W_mult0_32_32( g, factor2);

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

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

        tmpmul=W_shl(tmpmul, tmp_e);

        r_64=W_sub(tmpmul, norm_64);

        r_e=W_norm(r_64 );
@@ -1068,8 +1073,8 @@ static void biDiagonalReductionLeft_64( 
        invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, r, &invVal_e);





        tmp_e = add(31, sub(bitwindow, 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.

        tmp_e = add(31, sub(bitwindow, *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.



        bitwindow=add(bitwindow, 1); // so does the bit window

        FOR ( iCh = add( currChannel, 1 ); iCh < nChannelsC; iCh++)
@@ -1098,10 +1103,6 @@ static void biDiagonalReductionLeft_64( 
            }

        }

    }

    singularValues[currChannel] = g;

    singularValues_e[currChannel] = g_e;

    move32();

    move16();

}



/*-------------------------------------------------------------------------