Commit f4c40432 authored by thomas dettbarn's avatar thomas dettbarn Committed by Manuel Jander
Browse files

merge candiate.

parent 78e3ae9c

lib_enc/ivas_sns_enc_fx.c

+50 −209
Original line number Diff line number Diff line
@@ -59,7 +59,9 @@ static Word16 sns_1st_cod_fx( 
    Word32 *snsq_fx /* o  : quantized sns    Q16 */

)

{

    Word16 index, i;

    IF( exp_sns == Q15)

    {

        Word16 index;

        const Word16 split_len = M / 2;

        move16();

        const Word16 *means;
@@ -81,62 +83,9 @@ static Word16 sns_1st_cod_fx( 
            default:

                assert( !"illegal frame length in sns_1st_cod" );

        }

    Word16 exp_snsq_buffer[M] = { 0 }, exp_snsq = 0;

    move16();

    move16();

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

    {

        Word32 tmp = L_mult( means[i], means_fix ); // Q16

        exp_snsq_buffer[i] = 0;

        move16();

        snsq_fx[i] = BASOP_Util_Add_Mant32Exp( sns_fx[i], exp_sns, L_negate( tmp ), 15, &exp_snsq_buffer[i] );

        move32();

    }

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

    {

        exp_snsq = s_max( exp_snsq_buffer[i], exp_snsq );

    }

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

    {

        snsq_fx[i] = L_shr( snsq_fx[i], exp_snsq - exp_snsq_buffer[i] );

        move32();

    }



    index = 0;

    move16();

    FOR( Word16 split = 0; split < 2; ++split )

    {

        const Word16 *cdbk_ptr;

        Word16 j0, j1, index_split;

        Word32 dist_min_fx;

        const Word16 cdbk_fix = 8; // 1.f / powf( 2, SNS_CDBKS_BITS_4_FRAC ) in Q15

        move16();

        const Word16 *const cdbk = &sns_1st_cdbk[split][core - 1][0];



        j0 = imult1616( split, split_len );

        j1 = add( j0, split_len );



        cdbk_ptr = cdbk;

        dist_min_fx = MAXVAL_WORD32;

        Word16 exp_dist_min = 31;

        index_split = 0;

        FOR( i = 0; i < 32; ++i )

        {

            case L_FRAME16k:

                means = &sns_1st_means_16k[core - 1][0]; // Q14

                break;

            case L_FRAME25_6k:

                means = &sns_1st_means_25k6[core - 1][0]; // Q14

                break;

            case L_FRAME32k:

                means = &sns_1st_means_32k[core - 1][0]; // Q14

                break;

            default:

                assert( !"illegal frame length in sns_1st_cod" );

        }

        FOR( Word16 i = 0; i < M; ++i )

        {

            Word32 tmp = L_mult( means[i], means_fix ); // Q14->Q16

            Word32 tmp = L_mult( means[i], means_fix ); // Q16

            snsq_fx[i] = L_sub( sns_fx[i], tmp ); // Q16

            move32();

        }
@@ -147,48 +96,49 @@ static Word16 sns_1st_cod_fx( 
        {

            const Word16 *cdbk_ptr;

            Word16 j0, j1;

            Word16 index_split;

            Word16 dist_split;

            Word32 dist_min_fx;

            const Word16 cdbk_fix = 8; // 1.f / powf( 2, SNS_CDBKS_BITS_4_FRAC ) in Q15

            move16();

            const Word16 *const cdbk = &sns_1st_cdbk[split][core - 1][0]; // Q12

            const Word16 *const cdbk = &sns_1st_cdbk[split][core - 1][0];

    

            j0 = imult1616( split, split_len );

            j1 = add( j0, split_len );

    

            cdbk_ptr = cdbk;

            dist_min_fx = MAXVAL_WORD32;

            index_split = 0;

            dist_split = 0;

            move32();

            move16();

            FOR( Word16 i = 0; i < 32; ++i )

            {

                Word32 dist_fx = 0;

                move32();

                FOR( Word16 j = j0; j < j1; ++j ) // j1-j0=split_len. split_len=M/2. M=16

                FOR( Word16 j = j0; j < j1; ++j )

                {

                    Word32 tmp;

                    Word32 dist;

                    Word16 tmp2;

    

                    tmp = L_mult( *cdbk_ptr++, cdbk_fix ); // Q16

                    dist = L_sub( snsq_fx[j],  tmp );

                    dist = L_shr( dist, 4 ); // make sure that the next multiplication does not overflow

    

                    tmp = L_mult( *cdbk_ptr++, cdbk_fix ); // Q12->Q16

                    dist = L_sub( snsq_fx[j], tmp );       // Q16

                    dist = L_shl( dist, 11 );              // cdbk_ptr is a 16 bit LUT with 3.12 values, used as 3.16. assumption: snsq_fx has the same representation. thus, the subtraction results are in 4.16, which leaves 11 bit headroom.

                    dist = Mpy_32_32( dist, dist );

                    dist = L_shr( dist, 3 ); // make sure that the sum of 8 values does not overflow

                    tmp2 = extract_l( dist );

                    dist = L_mult( tmp2, tmp2 );

                    dist = L_shr( dist, 4 ); // make sure that the sum does not overflow

                    dist_fx = L_add( dist_fx, dist );

                }

    

                IF( LT_32( dist_fx, dist_min_fx ) )

                {

                    dist_min_fx = dist_fx;

                    move32();

                    index_split = i;

                    move16();

                    dist_split = i;

                }

            }

    

            /* set quantized vector */

            cdbk_ptr = &cdbk[imult1616( index_split, split_len )];

            cdbk_ptr = &cdbk[imult1616( dist_split, split_len )];

            FOR( Word16 j = j0; j < j1; ++j )

            {

                Word32 tmp_3 = L_mult( means[j], means_fix );    // Q16
@@ -200,15 +150,13 @@ static Word16 sns_1st_cod_fx( 
            /* for second split shift by five bits to store both indices as one 10 bit value */

            if ( EQ_16( split, 1 ) )

            {

                index_split = shl( index_split, 5 );

                dist_split = shl( dist_split, 5 );

            }

    

            index = add( index, index_split );

            index = add( index, dist_split );

        }

        return index;

    }

    ELSE

    {

    } ELSE {

        Word16 index, i;

        const Word16 split_len = M / 2;

        move16();
@@ -321,112 +269,6 @@ static Word16 sns_1st_cod_fx( 
    }

}



static Word16 sns_1st_cod_fx_q15(

    const Word32 *sns_fx, /* i  : vector to quantize                */

    const Word16 L_frame,

    const Word16 core,

    Word32 *snsq_fx /* o  : quantized sns    Q16 */

)

{

    Word16 index;

    const Word16 split_len = M / 2;

    move16();

    const Word16 *means;

    const Word16 means_fix = 2; // Q15

    push_wmops( "sns_1st_cod_fx_q15" );

    move16();

    /* remove means */

    means = NULL;

    SWITCH( L_frame )

    {

        case L_FRAME16k:

            means = &sns_1st_means_16k[core - 1][0];

            break;

        case L_FRAME25_6k:

            means = &sns_1st_means_25k6[core - 1][0];

            break;

        case L_FRAME32k:

            means = &sns_1st_means_32k[core - 1][0];

            break;

        default:

            assert( !"illegal frame length in sns_1st_cod" );

    }

    FOR( Word16 i = 0; i < M; ++i )

    {

        Word32 tmp = L_mult( means[i], means_fix ); // Q16

        snsq_fx[i] = L_sub( sns_fx[i], tmp ); // Q16

        move32();

    }



    index = 0;

    move16();

    FOR( Word16 split = 0; split < 2; ++split )

    {

        const Word16 *cdbk_ptr;

        Word16 j0, j1;

        Word16 dist_split;

        Word32 dist_min_fx;

        const Word16 cdbk_fix = 8; // 1.f / powf( 2, SNS_CDBKS_BITS_4_FRAC ) in Q15

        move16();

        const Word16 *const cdbk = &sns_1st_cdbk[split][core - 1][0];



        j0 = imult1616( split, split_len );

        j1 = add( j0, split_len );



        cdbk_ptr = cdbk;

        dist_min_fx = MAXVAL_WORD32;

        dist_split = 0;

        move32();

        move16();

        FOR( Word16 i = 0; i < 32; ++i )

        {

            Word32 dist_fx = 0;

            move32();

            FOR( Word16 j = j0; j < j1; ++j )

            {

                Word32 tmp;

                Word32 dist;

                Word16 tmp2;



                tmp = L_mult( *cdbk_ptr++, cdbk_fix ); // Q16

                dist = L_sub( snsq_fx[j],  tmp );

                dist = L_shr( dist, 4 ); // make sure that the next multiplication does not overflow



                tmp2 = extract_l( dist );

                dist = L_mult( tmp2, tmp2 );

                dist = L_shr( dist, 4 ); // make sure that the sum does not overflow

                dist_fx = L_add( dist_fx, dist );

            }



            IF( LT_32( dist_fx, dist_min_fx ) )

            {

                dist_min_fx = dist_fx;

                dist_split = i;

            }

        }



        /* set quantized vector */

        cdbk_ptr = &cdbk[imult1616( dist_split, split_len )];

        FOR( Word16 j = j0; j < j1; ++j )

        {

            Word32 tmp_3 = L_mult( means[j], means_fix );    // Q16

            Word32 tmp_4 = L_mult( *cdbk_ptr++, cdbk_fix ); // Q16

            snsq_fx[j] = L_add( tmp_4, tmp_3 );                 // Q16

            move32();

        }



        /* for second split shift by five bits to store both indices as one 10 bit value */

        if ( EQ_16( split, 1 ) )

        {

            dist_split = shl( dist_split, 5 );

        }



        index = add( index, dist_split );

    }

    pop_wmops();

    return index;

}



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

 * sns_2st_cod()

 *
@@ -532,7 +374,7 @@ void sns_avq_cod_fx( 
    Word16 indxt[256], nbits, nbt, nit;

    Word32 snsmid_q0_fx[M];



    index[0] = sns_1st_cod_fx_q15( sns_fx, L_frame, core, sns_q_fx );

    index[0] = sns_1st_cod_fx( sns_fx, exp_sns, L_frame, core, sns_q_fx );

    move16();

    nit = 1 + 2;

    move16();
@@ -559,8 +401,7 @@ void sns_avq_cod_fx( 
    {

        index++;



        index[0] = sns_1st_cod_fx_q15( snsmid_fx, L_frame, core, snsmid_q_fx );

        //        index[0] = sns_1st_cod_fx( snsmid_fx, exp_snsmid, L_frame, core, snsmid_q_fx );

        index[0] = sns_1st_cod_fx( snsmid_fx, exp_snsmid, L_frame, core, snsmid_q_fx );

        move16();

        nit = 1 + 2;

        move16();