Commit 069d14ba authored by Fabian Bauer's avatar Fabian Bauer
Browse files

some more _o replacements

parent 296c090f

lib_enc/ppp_enc_fx.c

+46 −0
Original line number Diff line number Diff line
@@ -622,9 +622,11 @@ static Word16 DTFS_alignment_fine_new_fx( DTFS_STRUCTURE X1_fx, DTFS_STRUCTURE X 
    Word16 n, fshift_fx, HalfLag, ab1[MAXLAG_WI], ab2[MAXLAG_WI];

    Word32 corr_fx;

    Word32 maxcorr_fx, wcorr_fx, diff_corr;

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif

    IF( LT_16( X1_fx.lag_fx, X2_fx.lag_fx ) )

    {
@@ -663,9 +665,15 @@ static Word16 DTFS_alignment_fine_new_fx( DTFS_STRUCTURE X1_fx, DTFS_STRUCTURE X 


        FOR( k = 0; k <= HalfLag; k++ )

        {

#ifdef ISSUE_1867_replace_overflow_libenc

            corr_fx = L_mac_sat( corr_fx, ab1[k], C_fx[temp % ( 4 * X2_fx.lag_fx )] );  //??sat

            corr_fx = L_mac_sat( corr_fx, ab2[k], S_fx[temp % ( 4 * X2_fx.lag_fx )] );  //??sat

            temp = add_sat( temp, temp1 );  //??sat

#else

            corr_fx = L_mac_o( corr_fx, ab1[k], C_fx[temp % ( 4 * X2_fx.lag_fx )], &Overflow );

            corr_fx = L_mac_o( corr_fx, ab2[k], S_fx[temp % ( 4 * X2_fx.lag_fx )], &Overflow );

            temp = add_o( temp, temp1, &Overflow );

#endif

        }

        temp = sub( 32767, extract_l( L_shr( L_mult( 82, abs_s( n ) ), 1 ) ) ); /*  Q15 */

        Qcorr = norm_l( corr_fx );
@@ -675,23 +683,40 @@ static Word16 DTFS_alignment_fine_new_fx( DTFS_STRUCTURE X1_fx, DTFS_STRUCTURE X 
            move16();

        }



#ifdef ISSUE_1867_replace_overflow_libenc

        temp1 = round_fx_sat( (Word32) L_shl_sat( corr_fx, Qcorr ) ); /*  Q(Qcorr-16) */   //??sat  //??sat

        wcorr_fx = L_mult_sat( temp1, temp );                                  /*  Q(Qcorr-16+15+1)=Q(Qcorr) */   //??sat

#else

        temp1 = round_fx_o( (Word32) L_shl_o( corr_fx, Qcorr, &Overflow ), &Overflow ); /*  Q(Qcorr-16) */

        wcorr_fx = L_mult_o( temp1, temp, &Overflow );                                  /*  Q(Qcorr-16+15+1)=Q(Qcorr) */

#endif



        IF( GE_16( Qmaxcorr, Qcorr ) )

        {

#ifdef ISSUE_1867_replace_overflow_libenc

            diff_corr = L_sub_sat( wcorr_fx, L_shl_sat( maxcorr_fx, sub( Qcorr, Qmaxcorr ) ) ); /*  Qcorr */   //??sat  //??sat

#else

            diff_corr = L_sub_o( wcorr_fx, L_shl_o( maxcorr_fx, sub( Qcorr, Qmaxcorr ), &Overflow ), &Overflow ); /*  Qcorr */

#endif

        }

        ELSE

        {

#ifdef ISSUE_1867_replace_overflow_libenc

            diff_corr = L_sub_sat( L_shl_sat( wcorr_fx, sub( Qmaxcorr, Qcorr ) ), maxcorr_fx ); /*  Qmaxcorr */  //??sat  //??sat

#else

            diff_corr = L_sub_o( L_shl_o( wcorr_fx, sub( Qmaxcorr, Qcorr ), &Overflow ), maxcorr_fx, &Overflow ); /*  Qmaxcorr */

#endif

        }



        if ( diff_corr > 0 )

        {

            fshift_fx = n;

            move16();

#ifdef ISSUE_1867_replace_overflow_libenc

            maxcorr_fx = (Word32) L_shl_sat( corr_fx, Qcorr ); /*  Qcorr */   //??sat

#else

            maxcorr_fx = (Word32) L_shl_o( corr_fx, Qcorr, &Overflow ); /*  Qcorr */

#endif

            Qmaxcorr = Qcorr;

            move16();

        }
@@ -739,9 +764,11 @@ static void LPCPowSpect_fx( 
    Word32 Ltemp, Lw;

    Word32 Lacc;

    Word16 tmp, exp;

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif

    FOR( k = 0; k < Nf; k++ )

    {
@@ -773,9 +800,15 @@ static void LPCPowSpect_fx( 


            t1 = add( t1, (Word16) L_shr( Ltemp, 16 ) ); /*  t1 is interpolated cos(w) */

            Ltemp = L_shr( L_mult( LPC[i], t1 ), 1 );    /*  Ltemp in Q27 */

#ifdef ISSUE_1867_replace_overflow_libenc

            Re = L_add_sat( Re, Ltemp );        /*  Re=1-sum(LPC[i]*cos(Lw)); */  //??sat

            Ltemp = L_add_sat( Lw, 0x6000 );    /*  add 0.75, which is 3pi/2 to convert sin to cos */   //??sat

            Ltemp = L_shl_sat( Ltemp, 10 );     /*  Q25 */  //??sat

#else

            Re = L_add_o( Re, Ltemp, &Overflow );        /*  Re=1-sum(LPC[i]*cos(Lw)); */

            Ltemp = L_add_o( Lw, 0x6000, &Overflow );    /*  add 0.75, which is 3pi/2 to convert sin to cos */

            Ltemp = L_shl_o( Ltemp, 10, &Overflow );     /*  Q25 */

#endif

            w = extract_h( Ltemp );                      /*  w is equivalent cos index */

            dl = extract_l( Ltemp );                     /*  dl is 6 bit left-over for interpolation */

            w = s_and( w, 511 );
@@ -797,8 +830,13 @@ static void LPCPowSpect_fx( 


            t1 = add( t1, (Word16) L_shr( Ltemp, 16 ) ); /*  t1 is interpolated cos(w) */

            Ltemp = L_shr( L_mult( LPC[i], t1 ), 1 );    /*  Ltemp in Q27 */

#ifdef ISSUE_1867_replace_overflow_libenc

            Im = L_sub_sat( Im, Ltemp );        /*  Im=sum(LPC[i]*sin(Lw)) */   //??sat

            Lw = L_add_sat( Lw, freq[k] );      /*  Lw=(i+1)*freq[k] */   //??sat

#else

            Im = L_sub_o( Im, Ltemp, &Overflow );        /*  Im=sum(LPC[i]*sin(Lw)) */

            Lw = L_add_o( Lw, freq[k], &Overflow );      /*  Lw=(i+1)*freq[k] */

#endif

        }

        /* If necessary, we can block-normalize Re and Im to improve precision */

        dh = extract_h( Re );
@@ -813,7 +851,11 @@ static void LPCPowSpect_fx( 
        ELSE

            Lacc = L_mult0( dh, dl );



#ifdef ISSUE_1867_replace_overflow_libenc

        Lacc = L_add_sat( L_shr( Lacc, 15 ), L_shr( L_mult_sat( dh, dh ), 1 ) ); /*  Lacc=Re*Re */   //??sat  //??sat

#else

        Lacc = L_add_o( L_shr( Lacc, 15 ), L_shr( L_mult_o( dh, dh, &Overflow ), 1 ), &Overflow ); /*  Lacc=Re*Re */

#endif

        dh = extract_h( Im );

        dl = extract_l( Im );
@@ -844,7 +886,11 @@ static void LPCPowSpect_fx( 
            move16();

        }

        Ltemp = L_deposit_h( tmp );

#ifdef ISSUE_1867_replace_overflow_libenc

        out[k] = round_fx_sat( L_shl_sat( Ltemp, negate( add( exp, 8 ) ) ) );  //??sat  //??sat

#else

        out[k] = round_fx_o( L_shl_o( Ltemp, negate( add( exp, 8 ) ), &Overflow ), &Overflow );

#endif

        move16();



        /* out[k] = shl(tmp,-exp-8); in Q7 */

lib_enc/pre_proc_fx.c

+4 −2
Original line number Diff line number Diff line
@@ -127,10 +127,12 @@ void pre_proc_fx( 
    FD_BWE_ENC_HANDLE hBWE_FD = st->hBWE_FD;



#ifndef ISSUE_1796_replace_shl_o

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif

#endif



    /*------------------------------------------------------------------*
@@ -1048,8 +1050,8 @@ void pre_proc_fx( 
    test();

    IF( ( ( st->tcxonly == 0 ) || ( EQ_16( st->codec_mode, MODE1 ) ) ) && GT_32( st->input_Fs, 8000 ) )

    {

#ifdef ISSUE_1796_replace_shl_o

        st->mem_preemph_enc = shl_sat( new_inp_16k[sub( L_frame_tmp, 1 )], 1 );

#ifdef ISSUE_1867_replace_overflow_libenc

        st->mem_preemph_enc = shl_sat( new_inp_16k[sub( L_frame_tmp, 1 )], 1 );   //??sat

#else

        st->mem_preemph_enc = shl_o( new_inp_16k[sub( L_frame_tmp, 1 )], 1, &Overflow );

#endif

lib_enc/pvq_core_enc_fx.c

+12 −12
Original line number Diff line number Diff line
@@ -267,7 +267,7 @@ void pvq_encode_frame_ivas_fx( 
    Word32 xy_corr, yy_corr;

    PVQ_ENC_DATA pvq_enc;

    PVQ_ENC_HANDLE hPVQ = &pvq_enc;

#ifndef ISSUE_1796_replace_shl_o

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();
@@ -333,8 +333,8 @@ void pvq_encode_frame_ivas_fx( 
            }



            tmp = ratio( xy_corr, yy_corr, &exp );

#ifdef ISSUE_1796_replace_shl_o

            gopt[is] = shl_sat( tmp, sub( sub( sub( 14, Q_coefs ), shift ), exp ) );

#ifdef ISSUE_1867_replace_overflow_libenc

            gopt[is] = shl_sat( tmp, sub( sub( sub( 14, Q_coefs ), shift ), exp ) );  //??sat

#else

            gopt[is] = shl_o( tmp, sub( sub( sub( 14, Q_coefs ), shift ), exp ), &Overflow );

#endif
@@ -397,7 +397,7 @@ void pvq_encode_frame_fx( 
    Word32 xy_corr, yy_corr;

    PVQ_ENC_DATA pvq_enc;

    PVQ_ENC_HANDLE hPVQ = &pvq_enc;

#ifndef ISSUE_1796_replace_shl_o

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();
@@ -463,8 +463,8 @@ void pvq_encode_frame_fx( 
            }



            tmp = ratio( xy_corr, yy_corr, &exp );

#ifdef ISSUE_1796_replace_shl_o

            gopt[is] = shl_sat( tmp, sub( sub( sub( 14, Q_coefs ), shift ), exp ) );

#ifdef ISSUE_1867_replace_overflow_libenc

            gopt[is] = shl_sat( tmp, sub( sub( sub( 14, Q_coefs ), shift ), exp ) ); //??sat

#else

            gopt[is] = shl_o( tmp, sub( sub( sub( 14, Q_coefs ), shift ), exp ), &Overflow );

#endif
@@ -943,7 +943,7 @@ static void densityIndexSymbolEncode_ivas_fx( 
    Word32 sym_freq, cum_freq, tot;

    Word32 acc;

    UWord16 lsb;

#ifndef ISSUE_1796_replace_shl_o

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();
@@ -955,8 +955,8 @@ static void densityIndexSymbolEncode_ivas_fx( 
    IF( s_and( (Word16) 0xFFFE, density ) != 0 ) /* even */

    {

        angle = atan2_fx( SQRT_DIM_fx[r_dim], SQRT_DIM_fx[l_dim] );

#ifdef ISSUE_1796_replace_shl_o

        angle = shl_sat( angle, 1 );

#ifdef ISSUE_1867_replace_overflow_libenc

        angle = shl_sat( angle, 1 );  //??sat

#else

        angle = shl_o( angle, 1, &Overflow );

#endif
@@ -1026,7 +1026,7 @@ static void densityIndexSymbolEncode_fx( 
    Word32 sym_freq, cum_freq, tot;

    Word32 acc;

    UWord16 lsb;

#ifndef ISSUE_1796_replace_shl_o

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();
@@ -1038,8 +1038,8 @@ static void densityIndexSymbolEncode_fx( 
    IF( s_and( (Word16) 0xFFFE, density ) != 0 ) /* even */

    {

        angle = atan2_fx( SQRT_DIM_fx[r_dim], SQRT_DIM_fx[l_dim] );

#ifdef ISSUE_1796_replace_shl_o

        angle = shl_sat( angle, 1 );

#ifdef ISSUE_1867_replace_overflow_libenc

        angle = shl_sat( angle, 1 );  //??sat

#else

        angle = shl_o( angle, 1, &Overflow );

#endif

lib_enc/pvq_encode_fx.c

+19 −6
Original line number Diff line number Diff line
@@ -60,9 +60,11 @@ static Word16 one_pulse_search( 
    UWord32 UL_left_l, UL_right_l, UL_dummy;

    Word32 L_tmp;

    UWord16 u_sgn;

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif



    en_tmp = en_dn_shift; /* dummy assignment to avoid compiler warning for unused parameter  */
@@ -85,8 +87,13 @@ static Word16 one_pulse_search( 


        FOR( i = 0; i < dim; i++ ) /* FOR 3 ops  */

        {

#ifdef ISSUE_1867_replace_overflow_libenc

            L_tmp_corr = L_shl_sat( L_mac_sat( *L_xy_ptr, 1, x_abs[i] ), corr_up_shift ); /*  actual in-loop target    value, 2 ops  */  //??sat  //??sat

            corr_tmp = round_fx_sat( L_tmp_corr );                                                 /*     1 op   */   //??sat

#else

            L_tmp_corr = L_shl_o( L_mac_o( *L_xy_ptr, 1, x_abs[i], &Overflow ), corr_up_shift, &Overflow ); /*  actual in-loop target    value, 2 ops  */

            corr_tmp = round_fx_o( L_tmp_corr, &Overflow );                                                 /*     1 op   */

#endif

            corr_sq_tmp = mult( corr_tmp, corr_tmp );                                                       /* CorrSq, is a 16bit for low compelxity cross multiplication    1 op   */



            L_tmp_en_lc = L_mac( *L_yy_ptr, 1, y[i] ); /*Q1 result ,  energy may span up to ~14+1(Q1)+1(sign)=16 bits,  1 op */
@@ -202,9 +209,11 @@ void pvq_encode_ivas_fx( 
    Word16 neg_gain_norm, shift_tot;

    Word16 high_pulse_density_flag;

    PvqEntry entry;

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif



    L_proj_fac = 4096;
@@ -359,12 +368,13 @@ void pvq_encode_ivas_fx( 
        }

        Mpy_32_16_ss( L_isqrt, tmp, &L_tmp, &u16_tmp );         /* Q31*Q(0+x)  +1         */

        Mpy_32_16_ss( L_tmp, neg_gain_norm, &L_tmp, &u16_tmp ); /* Q31*Q(0+x) *Q15 +1     */

#ifdef ISSUE_1799_replace_L_shr_o

        L_tmp = L_shr_sat( L_tmp, shift_tot ); /* Q31+x                  */

#ifdef ISSUE_1867_replace_overflow_libenc

        L_tmp = L_shr_sat( L_tmp, shift_tot ); /* Q31+x                  */   //??sat

        xq[i] = round_fx_sat( L_tmp ); /* Q15, array move        */   //??sat

#else

        L_tmp = L_shr_o( L_tmp, shift_tot, &Overflow ); /* Q31+x                  */

#endif

        xq[i] = round_fx_o( L_tmp, &Overflow ); /* Q15, array move        */

#endif

        move16();

        L_xq[i] = L_tmp; /* Q31 currently  unused  */

        move32();
@@ -414,9 +424,11 @@ void pvq_encode_fx( 
    Word16 neg_gain_norm, shift_tot;

    Word16 high_pulse_density_flag;

    PvqEntry entry;

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif



    L_proj_fac = 4096;
@@ -571,12 +583,13 @@ void pvq_encode_fx( 
        }

        Mpy_32_16_ss( L_isqrt, tmp, &L_tmp, &u16_tmp );         /* Q31*Q(0+x)  +1         */

        Mpy_32_16_ss( L_tmp, neg_gain_norm, &L_tmp, &u16_tmp ); /* Q31*Q(0+x) *Q15 +1     */

#ifdef ISSUE_1799_replace_L_shr_o

        L_tmp = L_shr_sat( L_tmp, shift_tot ); /* Q31+x                  */

#ifdef ISSUE_1867_replace_overflow_libenc

        L_tmp = L_shr_sat( L_tmp, shift_tot ); /* Q31+x                  */   //??sat

        xq[i] = round_fx_sat( L_tmp ); /* Q15, array move        */   //??sat

#else

        L_tmp = L_shr_o( L_tmp, shift_tot, &Overflow ); /* Q31+x                  */

#endif

        xq[i] = round_fx_o( L_tmp, &Overflow ); /* Q15, array move        */

#endif

        move16();

        L_xq[i] = L_tmp; /* Q31 currently  unused  */

        move32();

lib_enc/q_gain2p_fx.c

+28 −0
Original line number Diff line number Diff line
@@ -145,10 +145,12 @@ static Word16 gain_enc( /* o : quantization pitch index 
    const Word16 *p;

    const Word16 *t_qua_gain;

    Word32 L_tmp, dist_min, L_tmp1;

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif





    assert( ( func_type != FUNC_GAIN_ENC_UV ) && ( func_type != FUNC_GAIN_ENC_GACELP_UV ) );
@@ -165,7 +167,11 @@ static Word16 gain_enc( /* o : quantization pitch index 
    /* gain_inov = 1.0f / sqrt((dot_product(code, code, L_SUBFR) + 0.01) / L_SUBFR) */

    L_tmp = calc_gain_inov( code, lcode, &L_tmp1, &exp_L_tmp1 );

    move16();

#ifdef ISSUE_1867_replace_overflow_libenc

    *gain_inov = round_fx_sat( L_shl_sat( L_tmp, 15 - 3 ) ); /* gain_inov in Q12 */  //??sat  //??sat

#else

    *gain_inov = round_fx_o( L_shl_o( L_tmp, 15 - 3, &Overflow ), &Overflow ); /* gain_inov in Q12 */

#endif

    move16();

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

     * calculate the predicted gain code
@@ -377,7 +383,11 @@ static Word16 gain_enc( /* o : quantization pitch index 
        /* Here, we use L_mult0 to compensate the factor 0.5 applied to coeff[1..4] before */

        L_tmp = L_add( L_tmp, L_shr( L_mult0( p[2 * i + 0], p[2 * i + 0] ), shr_coeff0 ) );

        L_tmp = L_sub( L_tmp, L_shr( L_mult( p[2 * i + 0], coeff1 ), shr_coeff1 ) );

#ifdef ISSUE_1867_replace_overflow_libenc

        L_tmp1 = L_sub_sat( L_tmp, dist_min );  //??sat

#else

        L_tmp1 = L_sub_o( L_tmp, dist_min, &Overflow );

#endif

        BASOP_SATURATE_WARNING_ON_EVS

        if ( L_tmp1 < 0 )

        {
@@ -397,13 +407,21 @@ static Word16 gain_enc( /* o : quantization pitch index 


    L_tmp = L_mult( g_code, gcode0 ); /* Q11*Q15 -> Q27 */

    exp_gcode0 = add( exp_gcode0, -11 );

#ifdef ISSUE_1867_replace_overflow_libenc

    L_tmp = L_shl_sat( L_tmp, exp_gcode0 ); /*   Q27 -> Q16 */  //??sat

#else

    L_tmp = L_shl_o( L_tmp, exp_gcode0, &Overflow ); /*   Q27 -> Q16 */

#endif



    *gain_code = L_tmp;

    move32();

    /* Q16/Q12 => Q5 */

    L_tmp = L_deposit_h( BASOP_Util_Divide3216_Scale( L_tmp, *gain_inov, &i ) );

#ifdef ISSUE_1867_replace_overflow_libenc

    *past_gcode = L_shl_sat( L_tmp, sub( i, 15 - 12 ) );  //??sat

#else

    *past_gcode = L_shl_o( L_tmp, sub( i, 15 - 12 ), &Overflow );

#endif

    move16();



    return index;
@@ -436,9 +454,11 @@ Word16 gain_enc_uv_fx( /* o : quantization pitch ind 
    Word16 index2;

    const Word16 log2_scale = 16;

    move16();

#ifndef ISSUE_1867_replace_overflow_libenc

#ifdef BASOP_NOGLOB_DECLARE_LOCAL

    Flag Overflow = 0;

    move16();

#endif

#endif

    pred_nrg_frame = 0; /* to suppress compilation warnings */

    g_code2 = 0;        /* to suppress compilation warnings */
@@ -595,7 +615,11 @@ Word16 gain_enc_uv_fx( /* o : quantization pitch ind 
        }



        s1 = norm_l( *gain_code );

#ifdef ISSUE_1867_replace_overflow_libenc

        tmp1 = round_fx_sat( L_shl( *gain_code, s1 ) );  //??sat

#else

        tmp1 = round_fx_o( L_shl_o( *gain_code, s1, &Overflow ), &Overflow );

#endif

        s1 = sub( 15, s1 );

        tmp1 = mult_r( mult_r( tmp1, tmp1 ), g_coeff->y2y2 );
@@ -608,7 +632,11 @@ Word16 gain_enc_uv_fx( /* o : quantization pitch ind 
        L_tmp1 = L_add( L_tmp, 0 );



        s1 = norm_l( *gain_code );

#ifdef ISSUE_1867_replace_overflow_libenc

        tmp1 = round_fx_sat( L_shl( *gain_code, s1 ) );  //??sat

#else

        tmp1 = round_fx_o( L_shl_o( *gain_code, s1, &Overflow ), &Overflow );

#endif

        s1 = sub( 15, s1 );



        c_index2 = 0x7FFF;