Commit f0002fbd authored by Sandesh Venkatesh's avatar Sandesh Venkatesh
Browse files

Encoder bug fixes, quality improvements 

[x] Shift correction in prep_tbe_exc_ivas_fx
[x] Fixed Q-mismatch for mem_shb_speech_ref_fx
[x] Removed saturation in attack_det and fixed q mismatch
[x] Fix for saturation in tcx_ltp_encode
parent 41efce11

lib_com/swb_tbe_com_fx.c

+1 −6
Original line number Diff line number Diff line
@@ -7847,13 +7847,8 @@ void prep_tbe_exc_ivas_fx( 
    }

    ELSE

    {

        Word16 shift = 8;

        Word16 shift = 4;

        move16();

        if ( element_mode == EVS_MONO )

        {

            shift = 4;

            move16();

        }

        IF( gain_preQ_fx != 0 )

        {

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

lib_enc/ivas_stereo_icbwe_enc.c

+3 −1
Original line number Diff line number Diff line
@@ -664,7 +664,8 @@ void stereo_icBWE_enc_ivas_fx( 
        Copy_Scale_sig( hStereoICBWE->mem_shb_speech_ref_fx, hStereoICBWE->mem_shb_speech_ref_fx, L_LOOK_16k, negate( sub( max_e, hStereoICBWE->mem_shb_speech_ref_e ) ) ); // mem_shb_speech_ref_e



        set32_fx( shb_frame_ref_fx, 0, L_LOOK_16k + L_FRAME16k );

        Copy_Scale_sig_16_32( hStereoICBWE->mem_shb_speech_ref_fx, shb_frame_ref_fx, L_LOOK_16k, negate( sub( max_e, hStereoICBWE->mem_shb_speech_ref_e ) ) + Q16 ); // mem_shb_speech_ref_e

        Copy_Scale_sig_16_32( hStereoICBWE->mem_shb_speech_ref_fx, shb_frame_ref_fx, L_LOOK_16k, 0 );                        // Q = 15 - mem_shb_speech_ref_e

        scale_sig32( shb_frame_ref_fx, L_LOOK_16k, add( negate( sub( max_e, hStereoICBWE->mem_shb_speech_ref_e ) ), Q16 ) ); // Q = 31 - mem_shb_speech_ref_e



        hStereoICBWE->mem_shb_speech_ref_e = max_e;

        shb_frame_ref_e = max_e;
@@ -678,6 +679,7 @@ void stereo_icBWE_enc_ivas_fx( 
        Copy_Scale_sig32( shb_speech_ref_fx, shb_frame_ref_fx + L_LOOK_16k - refMemLen, L_FRAME16k, negate( sub( max_e, shb_speech_ref_e ) ) );                                   /* shb_speech_ref_e */

        Copy_Scale_sig_32_16( shb_frame_ref_fx + L_FRAME16k, hStereoICBWE->mem_shb_speech_ref_fx, L_LOOK_16k, negate( sub( max_e, hStereoICBWE->mem_shb_speech_ref_e ) ) - Q16 ); // mem_shb_speech_ref_e





        test();

        test();

        IF( ( EQ_16( st->extl, SWB_TBE ) || EQ_16( st->extl, FB_TBE ) ) && EQ_16( st->flag_ACELP16k, 1 ) )

lib_enc/speech_music_classif_fx.c

+8 −2
Original line number Diff line number Diff line
@@ -2846,13 +2846,18 @@ static Word16 attack_det_ivas_fx( /* o : attack flag 
            Copy32( finc, finc_prev, attack1 );



            /* compute mean energy before the attack */

            etmp = L_shr( sum32_fx( finc_prev, ATT_NSEG ), 5 ); /*ATT_NSEG == 32*/

            Word64 W_etmp = W_deposit32_l( finc_prev[0] );

            FOR( Word16 idx = 1; idx < ATT_NSEG; idx++ )

            {

                W_etmp = W_add( W_etmp, W_deposit32_l( finc_prev[idx] ) );

            }

            W_etmp = W_shr( W_etmp, 5 ); /*ATT_NSEG == 32*/



            etmp2 = finc[attack1];

            move32();

            test();

            test();

            if ( ( LT_32( L_shl( etmp, 4 ), etmp2 ) ) || ( LT_32( L_add( L_shl( etmp, 3 ), L_shl( etmp, 2 ) ), etmp2 ) && EQ_16( last_clas, UNVOICED_CLAS ) ) )

            if ( ( LT_64( W_shl( W_etmp, 4 ), W_deposit32_l( etmp2 ) ) ) || ( LT_64( W_add( W_shl( W_etmp, 3 ), W_shl( W_etmp, 2 ) ), W_deposit32_l( etmp2 ) ) && EQ_16( last_clas, UNVOICED_CLAS ) ) )

            {

                attack = attack1;

                move16();
@@ -3045,6 +3050,7 @@ void ivas_smc_mode_selection_fx( 
    IF( NE_16( shl( Q_new, 1 ), hSpMusClas->Q_lt_finc ) )

    {

        Scale_sig32( &hSpMusClas->lt_finc_fx, 1, sub( shl( Q_new, 1 ), hSpMusClas->Q_lt_finc ) );

        Scale_sig32( hSpMusClas->finc_prev_fx, ATT_NSEG, sub( shl( Q_new, 1 ), hSpMusClas->Q_lt_finc ) );

        hSpMusClas->Q_lt_finc = shl( Q_new, 1 );

        move16();

    }

lib_enc/swb_pre_proc.c

+3 −1
Original line number Diff line number Diff line
@@ -393,7 +393,9 @@ void swb_pre_proc_ivas_fx( 
            /* delay corresponding to CLDFB delay */

            Copy( old_input_fx + STEREO_DFT_OVL_16k - 20, shb_speech, L_FRAME16k );

            Copy( old_input_fx, hBWE_TD->old_speech_shb_fx + L_LOOK_16k + L_SUBFR16k - ( STEREO_DFT_OVL_16k - 20 ), STEREO_DFT_OVL_16k - 20 );

            Copy( old_input_fx, hCPE->hStereoICBWE->mem_shb_speech_ref_fx, STEREO_DFT_OVL_16k - 20 );

            Copy( old_input_fx, hCPE->hStereoICBWE->mem_shb_speech_ref_fx, STEREO_DFT_OVL_16k - 20 ); // Q0

            hCPE->hStereoICBWE->mem_shb_speech_ref_e = 15;

            move16();



            IF( LE_32( CldfbHB_fx, 0 ) )

            {

lib_enc/tcx_ltp_enc_fx.c

+1 −2
Original line number Diff line number Diff line
@@ -1154,8 +1154,7 @@ void tcx_ltp_encode_ivas_fx( 
        move16();

        sqr = Sqrt16( hTcxEnc->tcxltp_on_mem, &exponent );

        move16();

        sqr = shl_sat( sqr, exponent );

        IF( ( EQ_16( tcxOnly, 1 ) && EQ_16( st->element_mode, IVAS_CPE_MDCT ) && ( GT_16( ( mult( sqr, mult( ( norm_corr_fx ), 29492 ) ) ), 14419 ) && ( LT_16( tempFlatness_fx, 768 ) || ( LT_16( tempFlatness_fx, 896 ) && LT_16( maxEnergyChange_fx, 176 ) ) ) ) ) ) // 0.9f Q15, 6.0f Q7, 7.0f Q7, 22.0f Q3

        IF( ( EQ_16( tcxOnly, 1 ) && EQ_16( st->element_mode, IVAS_CPE_MDCT ) && ( GT_16( ( mult( sqr, mult( ( norm_corr_fx ), 29492 ) ) ), shr( 14419, exponent ) ) && ( LT_16( tempFlatness_fx, 768 ) || ( LT_16( tempFlatness_fx, 896 ) && LT_16( maxEnergyChange_fx, 176 ) ) ) ) ) ) // 0.9f Q15, 6.0f Q7, 7.0f Q7, 22.0f Q3

        {

            ltp_on = 1;

            move16();