Commit 1cae4d03 authored by Fabian Bauer's avatar Fabian Bauer
Browse files

lib_com/tools_fx.c 

lib_com/weight_a_fx.c
lib_com/wi_fx.c
lib_com/window_fx.c : completed overflow op replacement
parent 742a5b64

lib_com/tools_fx.c

+167 −14
Original line number Diff line number Diff line
@@ -308,28 +308,36 @@ Word16 usquant_fx( /* o: index of the winning codeword */ 
    Word16 idx;

    Word16 tmp, exp;

    Word32 L_tmp;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif



    /*    idx = (short)( (x - qlow)/delta + 0.5f); */

    exp = norm_s( delta );

    tmp = div_s( shl( 1, sub( 14, exp ) ), delta );     /*Q(29-exp-(Qx-1))->Q(30-exp-Qx) */

    L_tmp = L_mult( sub_o( x, qlow, &Overflow ), tmp ); /*Q(31-exp) */

#ifdef ISSUE_1796_replace_shl_o

    idx = extract_l( L_shr_r( L_add( L_tmp, shl_sat( 1, sub( 30, exp ) ) ), sub( 31, exp ) ) ); /*Q0 */

#ifdef ISSUE_1836_replace_overflow_libcom

    L_tmp = L_mult( sub_sat( x, qlow ), tmp ); /*Q(31-exp) */   //??sat

    idx = extract_l( L_shr_r( L_add( L_tmp, shl_sat( 1, sub( 30, exp ) ) ), sub( 31, exp ) ) ); /*Q0 */ //??sat

#else

    L_tmp = L_mult( sub_o( x, qlow, &Overflow ), tmp ); /*Q(31-exp) */

    idx = extract_l( L_shr_r( L_add( L_tmp, shl_o( 1, sub( 30, exp ), &Overflow ) ), sub( 31, exp ) ) ); /*Q0 */

#endif





    idx = s_min( idx, sub( cbsize, 1 ) );

    idx = s_max( idx, 0 );



    /*    *xq = idx*delta + qlow; */

    L_tmp = L_deposit_l( qlow );                                    /*Qx */

    L_tmp = L_mac( L_tmp, idx, delta );                             /*Qx */

#ifdef ISSUE_1836_replace_overflow_libcom

    *xq = round_fx_sat( L_shl_sat( L_tmp, 16 ) );                   /*Qx */

#else

    *xq = round_fx_o( L_shl_o( L_tmp, 16, &Overflow ), &Overflow ); /*Qx */

#endif

    return idx;

}

/*-------------------------------------------------------------------*
@@ -395,14 +403,20 @@ Word32 sum2_fx( /* o : sum of all squared vector elements 
{

    Word16 i;

    Word32 L_tmp;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif

    L_tmp = L_deposit_l( 0 );

    FOR( i = 0; i < lvec; i++ )

    {

#ifdef ISSUE_1836_replace_overflow_libcom

        L_tmp = L_mac_sat( L_tmp, vec[i], vec[i] ); /*Q(2x+1) */

#else

        L_tmp = L_mac_o( L_tmp, vec[i], vec[i], &Overflow ); /*Q(2x+1) */

#endif

    }



    return L_tmp;
@@ -456,15 +470,21 @@ Word32 sum2_fx_mod( /* o : sum of all squared vector element 
{

    Word16 i;

    Word32 L_tmp;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif



    L_tmp = L_deposit_l( 0 );

    FOR( i = 0; i < lvec; i++ )

    {

#ifdef ISSUE_1836_replace_overflow_libcom

        L_tmp = L_add_sat( L_tmp, L_shr( L_mult_sat( vec[i], vec[i] ), 9 ) );

#else

        L_tmp = L_add_o( L_tmp, L_shr( L_mult_o( vec[i], vec[i], &Overflow ), 9 ), &Overflow );

#endif

    }



    return L_tmp;
@@ -684,7 +704,7 @@ void Copy_Scale_sig( 
{

    Word16 i;

    Word16 tmp;

#ifndef ISSUE_1796_replace_shl_o

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();
@@ -711,8 +731,8 @@ void Copy_Scale_sig( 
    }

    FOR( i = 0; i < lg; i++ )

    {

#ifdef ISSUE_1796_replace_shl_o

        y[i] = shl_sat( x[i], exp0 );

#ifdef ISSUE_1836_replace_overflow_libcom

        y[i] = shl_sat( x[i], exp0 );   //??sat

#else

        y[i] = shl_o( x[i], exp0, &Overflow );

#endif
@@ -733,7 +753,7 @@ void Copy_Scale_sig_16_32_DEPREC( 
{

    Word16 i;

    Word16 tmp;

#ifndef ISSUE_1796_replace_shl_o

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();
@@ -755,8 +775,8 @@ void Copy_Scale_sig_16_32_DEPREC( 
        /*Should not happen */

        FOR( i = 0; i < lg; i++ )

        {

#ifdef ISSUE_1796_replace_shl_o

            y[i] = L_deposit_l( shl_sat( x[i], exp0 ) );

#ifdef ISSUE_1836_replace_overflow_libcom

            y[i] = L_deposit_l( shl_sat( x[i], exp0 ) );  //??sat

#else

            y[i] = L_deposit_l( shl_o( x[i], exp0, &Overflow ) );

#endif
@@ -772,8 +792,8 @@ void Copy_Scale_sig_16_32_DEPREC( 
#else

    assert( exp0 < 16 );

#endif

#ifdef ISSUE_1796_replace_shl_o

    tmp = shl_sat( 1, exp0 );

#ifdef ISSUE_1836_replace_overflow_libcom

    tmp = shl_sat( 1, exp0 );   //??sat

#else

    tmp = shl_o( 1, exp0, &Overflow );

#endif
@@ -793,10 +813,12 @@ void Copy_Scale_sig_16_32_no_sat( 
{

    Word16 i;

    Word32 L_tmp;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif





    IF( exp0 == 0 )
@@ -813,8 +835,8 @@ void Copy_Scale_sig_16_32_no_sat( 
        /*Should not happen */

        FOR( i = 0; i < lg; i++ )

        {

#ifdef ISSUE_1796_replace_shl_o

            y[i] = L_deposit_l( shl_sat( x[i], exp0 ) );

#ifdef ISSUE_1836_replace_overflow_libcom

            y[i] = L_deposit_l( shl_sat( x[i], exp0 ) );  //??sat

#else

            y[i] = L_deposit_l( shl_o( x[i], exp0, &Overflow ) );

#endif
@@ -822,7 +844,11 @@ void Copy_Scale_sig_16_32_no_sat( 
        }

        return;

    }

#ifdef ISSUE_1836_replace_overflow_libcom

    L_tmp = L_shl_sat( 1, exp0 - 1 ); //??sat

#else

    L_tmp = L_shl_o( 1, exp0 - 1, &Overflow );

#endif



    IF( L_tmp >= 0x7FFF )

    {
@@ -854,9 +880,11 @@ void Copy_Scale_sig_32_16( 
{

    Word16 i;

    Word16 tmp;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif



    tmp = add( 16, exp0 );
@@ -864,7 +892,11 @@ void Copy_Scale_sig_32_16( 
    {

        FOR( i = 0; i < lg; i++ )

        {

#ifdef ISSUE_1836_replace_overflow_libcom

            y[i] = round_fx_sat( L_shl_sat( x[i], tmp ) );

#else

            y[i] = round_fx_o( L_shl_o( x[i], tmp, &Overflow ), &Overflow );

#endif

            move16();

        }

    }
@@ -872,7 +904,11 @@ void Copy_Scale_sig_32_16( 
    {

        FOR( i = 0; i < lg; i++ )

        {

#ifdef ISSUE_1836_replace_overflow_libcom

            y[i] = round_fx_sat( x[i] );  //??sat

#else

            y[i] = round_fx_o( x[i], &Overflow );

#endif

            move16();

        }

    }
@@ -890,9 +926,11 @@ void Scale_sig32( 
)

{

    Word16 i;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif

    IF( 0 == exp0 )

    {
@@ -901,7 +939,11 @@ void Scale_sig32( 


    FOR( i = 0; i < lg; i++ )

    {

#ifdef ISSUE_1836_replace_overflow_libcom

        x[i] = L_shl_sat( x[i], exp0 );

#else

        x[i] = L_shl_o( x[i], exp0, &Overflow );

#endif

        move32(); /* saturation can occur here */

    }

}
@@ -1896,10 +1938,12 @@ Word16 w_vquant_fx( 
    Word16 tmp;

    Word16 c, idx, j;

    Word32 dist, minDist;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif





    idx = 0;
@@ -1916,6 +1960,28 @@ Word16 w_vquant_fx( 
        {

            dist = L_deposit_l( 0 );



#ifdef ISSUE_1836_replace_overflow_libcom

            tmp = sub_sat( x[3], shr( cb[j++], Qx ) );  //??sat

            if ( weights[3] != 0 )

            {

                dist = L_mac0_sat( dist, tmp, tmp );    //??sat

            }

            tmp = sub_sat( x[2], shr( cb[j++], Qx ) );  //??sat

            if ( weights[2] != 0 )

            {

                dist = L_mac0_sat( dist, tmp, tmp );    //??sat

            }

            tmp = sub_sat( x[1], shr( cb[j++], Qx ) );  //??sat

            if ( weights[1] != 0 )

            {

                dist = L_mac0_sat( dist, tmp, tmp );    //??sat

            }

            tmp = sub_sat( x[0], shr( cb[j++], Qx ) );  //??sat

            if ( weights[0] != 0 )

            {

                dist = L_mac0_sat( dist, tmp, tmp );    //??sat

            }

#else

            tmp = sub_o( x[3], shr( cb[j++], Qx ), &Overflow );

            if ( weights[3] != 0 )

            {
@@ -1936,6 +2002,7 @@ Word16 w_vquant_fx( 
            {

                dist = L_mac0_o( dist, tmp, tmp, &Overflow );

            }

#endif

            if ( LT_32( dist, minDist ) )

            {

                idx = c;
@@ -1965,6 +2032,28 @@ Word16 w_vquant_fx( 
        {

            dist = L_deposit_l( 0 );



#ifdef ISSUE_1836_replace_overflow_libcom

            tmp = sub_sat( x[0], shr( cb[j++], Qx ) );    //??sat

            if ( weights[0] != 0 )

            {

                dist = L_mac0_sat( dist, tmp, tmp );      //??sat

            }

            tmp = sub_sat( x[1], shr( cb[j++], Qx ) );    //??sat

            if ( weights[1] != 0 )

            {

                dist = L_mac0_sat( dist, tmp, tmp );      //??sat

            }

            tmp = sub_sat( x[2], shr( cb[j++], Qx ) );    //??sat

            if ( weights[2] != 0 )

            {

                dist = L_mac0_sat( dist, tmp, tmp );      //??sat

            }

            tmp = sub_sat( x[3], shr( cb[j++], Qx ) );    //??sat

            if ( weights[3] != 0 )

            {

                dist = L_mac0_sat( dist, tmp, tmp );      //??sat

            }

#else

            tmp = sub_o( x[0], shr( cb[j++], Qx ), &Overflow );

            if ( weights[0] != 0 )

            {
@@ -1985,6 +2074,7 @@ Word16 w_vquant_fx( 
            {

                dist = L_mac0_o( dist, tmp, tmp, &Overflow );

            }

#endif

            if ( LT_32( dist, minDist ) )

            {

                idx = c;
@@ -2451,9 +2541,11 @@ Word32 root_a_over_b_fx( 
    Word16 tmp, num, den, scale;

    Word16 exp, exp_num, exp_den;

    Word32 L_tmp;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif

    test();

    IF( ( a <= 0 ) || ( b <= 0 ) )
@@ -2464,15 +2556,23 @@ Word32 root_a_over_b_fx( 
    }



    exp_num = norm_l( b );

#ifdef ISSUE_1836_replace_overflow_libcom

    num = round_fx_sat( L_shl_sat( b, exp_num ) );  //??sat

#else

    num = round_fx_o( L_shl_o( b, exp_num, &Overflow ), &Overflow );

#endif

    exp_num = sub( sub( 30, exp_num ), Q_b );



    exp_den = norm_l( a );

#ifdef ISSUE_1836_replace_overflow_libcom

    den = round_fx_sat( L_shl_sat( a, exp_den ) );  //??sat

#else

    den = round_fx_o( L_shl_o( a, exp_den, &Overflow ), &Overflow );

#endif

    exp_den = sub( sub( 30, exp_den ), Q_a );



    scale = shr( sub( den, num ), 15 );

#ifdef ISSUE_1796_replace_shl_o

#ifdef ISSUE_1836_replace_overflow_libcom     //??sat

    num = shl_sat( num, scale );

#else

    num = shl_o( num, scale, &Overflow );
@@ -2632,9 +2732,11 @@ void fir_fx( const Word16 x[], /* i : input vector 
    Word16 buf_in[L_FRAME32k + L_FILT_MAX];

    Word16 i, j;

    Word32 s;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif

    /* prepare the input buffer (copy and update memory) */

    Copy( mem, buf_in, K );
@@ -2647,14 +2749,27 @@ void fir_fx( const Word16 x[], /* i : input vector 
    /* do the filtering */

    FOR( i = 0; i < L; i++ )

    {

#ifdef ISSUE_1836_replace_overflow_libcom

        s = L_mult_sat( buf_in[K + i], h[0] );   //??sat

#else

        s = L_mult_o( buf_in[K + i], h[0], &Overflow );

#endif



        FOR( j = 1; j <= K; j++ )

        {

#ifdef ISSUE_1836_replace_overflow_libcom

            s = L_mac_sat( s, h[j], buf_in[K + i - j] );  //??sat

#else

            s = L_mac_o( s, h[j], buf_in[K + i - j], &Overflow );

#endif

        }

#ifdef ISSUE_1836_replace_overflow_libcom

        s = L_shl_sat( s, shift );          //??sat

        y[i] = round_fx_sat( s ); /*Qx */   //??sat

#else

        s = L_shl_o( s, shift, &Overflow );

        y[i] = round_fx_o( s, &Overflow ); /*Qx */

#endif

        move16();

    }

}
@@ -2790,8 +2905,10 @@ Word16 squant_fx( /* o: index of the winning codeword */ 
    Word16 tmp;

    Word16 c, idx;

    Word32 L_mindist, L_dist;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

#endif

#endif



    idx = 0;
@@ -2802,10 +2919,18 @@ Word16 squant_fx( /* o: index of the winning codeword */ 
    FOR( c = 0; c < cbsize; c++ )

    {

        L_dist = L_deposit_l( 0 );

#ifdef ISSUE_1836_replace_overflow_libcom

        tmp = sub_sat( x, cb[c] );

#else

        tmp = sub_o( x, cb[c], &Overflow );

#endif



        /*dist += tmp*tmp; */

#ifdef ISSUE_1836_replace_overflow_libcom

        L_dist = L_mac_sat( L_dist, tmp, tmp );   //??sat

#else

        L_dist = L_mac_o( L_dist, tmp, tmp, &Overflow );

#endif



        if ( LT_32( L_dist, L_mindist ) )

        {
@@ -2943,9 +3068,11 @@ void Copy_Scale_sig32( 
    Word16 i;

    Word32 L_tmp;

    Word16 tmp = exp0;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif

    IF( exp0 == 0 )

    {
@@ -2960,12 +3087,20 @@ void Copy_Scale_sig32( 
    {

        FOR( i = 0; i < lg; i++ )

        {

#ifdef ISSUE_1836_replace_overflow_libcom

            y[i] = L_shl_sat( x[i], tmp );  //??sat

#else

            y[i] = L_shl_o( x[i], tmp, &Overflow );

#endif

            move16();

        }

        return;

    }

#ifdef ISSUE_1836_replace_overflow_libcom

    L_tmp = L_shl_sat( 1, exp0 - 1 );   //??sat

#else

    L_tmp = L_shl_o( 1, exp0 - 1, &Overflow );

#endif

    FOR( i = 0; i < lg; i++ )

    {

        y[i] = W_extract_l( W_mult_32_32( L_tmp, x[i] ) );
@@ -3142,9 +3277,11 @@ void add_vec_fx( 
)

{

    Word16 i, Qyx1, Qyx2;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move32();

#endif

#endif

    Qyx1 = sub( Qx1, Qy );

    Qyx2 = sub( Qx2, Qy );
@@ -3152,7 +3289,11 @@ void add_vec_fx( 
    {

        FOR( i = 0; i < N; i++ )

        {

#ifdef ISSUE_1836_replace_overflow_libcom

            y[i] = add_sat( x1[i], shr_r_sat( x2[i], Qyx2 ) );  //??sat //??sat

#else

            y[i] = add_o( x1[i], shr_r_sat( x2[i], Qyx2 ), &Overflow );

#endif

            move16();

        }

    }
@@ -3160,7 +3301,11 @@ void add_vec_fx( 
    {

        FOR( i = 0; i < N; i++ )

        {

#ifdef ISSUE_1836_replace_overflow_libcom

            y[i] = add_sat( shr_r_sat( x1[i], Qyx1 ), shr_r_sat( x2[i], Qyx2 ) );   //??sat //!!sat //!!sat

#else

            y[i] = add_o( shr_r_sat( x1[i], Qyx1 ), shr_r_sat( x2[i], Qyx2 ), &Overflow );

#endif

            move16();

        }

    }
@@ -3287,7 +3432,11 @@ Word32 Calc_Energy_Autoscaled( /* o: Result (Energy) 
    FOR( i = 0; i < j; i++ )

    {

        /* divide by 2 so energy will be divided by 4 */

#ifdef ISSUE_1836_replace_overflow_libcom

        temp = mult_r( *signal++, 16384 );

#else

        temp = mult_ro( *signal++, 16384, &Overflow );

#endif

        L_Energy = L_mac0_o( L_Energy, temp, temp, &Overflow );

    }

    FOR( i = j; i < len; i += 8 ) /* Process 8 Samples at a time */
@@ -3297,7 +3446,11 @@ Word32 Calc_Energy_Autoscaled( /* o: Result (Energy) 
        L_temp = L_mult0( temp, temp );

        FOR( j = 1; j < 8; j++ )

        {

#ifdef ISSUE_1836_replace_overflow_libcom

            temp = mult_r( *signal++, 16384 );

#else

            temp = mult_ro( *signal++, 16384, &Overflow );

#endif

            L_temp = L_mac0_o( L_temp, temp, temp, &Overflow );

        }

#ifdef ISSUE_1799_replace_L_shr_o

lib_com/weight_a_fx.c

+6 −0
Original line number Diff line number Diff line
@@ -100,10 +100,12 @@ void weight_a_fx( 
    Word16 i, fac;

    Word32 Amax;

    Word16 shift;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow;

    Overflow = 0;

    move32();

#endif

#endif



    fac = gamma; /* Q15 */
@@ -122,7 +124,11 @@ void weight_a_fx( 
    move16();

    FOR( i = 1; i < m; i++ )

    {

#ifdef ISSUE_1836_replace_overflow_libcom

        ap[i] = round_fx_sat( L_shl( L_mult0( a[i], fac ), shift ) ); /* Q11 + shift */   //??sat

#else

        ap[i] = round_fx_o( L_shl( L_mult0( a[i], fac ), shift ), &Overflow ); /* Q11 + shift */

#endif

        move16();

        fac = mult_r( fac, gamma ); /* Q15 */

    }

lib_com/wi_fx.c

+361 −23

File changed.

Preview size limit exceeded, changes collapsed.

lib_com/window_fx.c

+8 −2
Original line number Diff line number Diff line
@@ -29,10 +29,12 @@ void ham_cos_window( 
{

    Word16 i;

    Word32 cte, cc;

#ifndef ISSUE_1836_replace_overflow_libcom

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif





    assert( n1 >= 102 ); /* if n1 is too low -> overflow in div_l */
@@ -46,7 +48,11 @@ void ham_cos_window( 
    {

        /* fh_f[i] = 0.54f - 0.46f * (Float32)cos(cc);	*/

        BASOP_SATURATE_WARNING_OFF_EVS

#ifdef ISSUE_1836_replace_overflow_libcom

        fh[i] = sub_sat( P54_0Q15, mult_r( getCosWord16( round_fx_sat( L_shl_sat( cc, 9 ) ) ), P92_0Q15 ) ); /*0Q15*/   //??sat //??sat //??sat

#else

        fh[i] = sub_o( P54_0Q15, mult_r( getCosWord16( round_fx_o( L_shl_o( cc, 9, &Overflow ), &Overflow ) ), P92_0Q15 ), &Overflow ); /*0Q15*/

#endif

        move16();

        BASOP_SATURATE_WARNING_ON_EVS

        cc = L_add( cc, cte ); /*0Q15*/
@@ -63,8 +69,8 @@ void ham_cos_window( 
    FOR( i = n1; i < n1 + n2; i++ )

    {

        /* fh_f[i] = (Float32)cos(cc);		*/

#ifdef ISSUE_1796_replace_shl_o

        fh[i] = shl_sat( getCosWord16( round_fx( L_shl( cc, 10 ) ) ), 1 ); /*0Q15*/

#ifdef ISSUE_1836_replace_overflow_libcom

        fh[i] = shl_sat( getCosWord16( round_fx( L_shl( cc, 10 ) ) ), 1 ); /*0Q15*/     //??sat

#else

        fh[i] = shl_o( getCosWord16( round_fx( L_shl( cc, 10 ) ) ), 1, &Overflow ); /*0Q15*/

#endif