Commit 1e58f159 authored by multrus's avatar multrus
Browse files

encapsulate changes

parent 6a806386

lib_com/ivas_prot_fx.h

+5 −2
Original line number Diff line number Diff line
@@ -4706,6 +4706,8 @@ Word32 dot_product_cholesky_fx( 
    const Word32 *A, /* i  : Cholesky  matrix A              */

    const Word16 N   /* i  : vector & matrix size            */

);



#ifndef DOT_PROD_CHOLESKY_64BIT

Word32 dot_product_cholesky_fixed(

    const Word32 *x, /* i  : vector x                        */

    const Word32 *A, /* i  : Cholesky  matrix A              */
@@ -4713,12 +4715,13 @@ Word32 dot_product_cholesky_fixed( 
    const Word16 exp_x,

    const Word16 exp_A,

    Word16 *exp_sum );



Word64 dot_product_cholesky_fixed64(

#else

Word64 dot_product_cholesky_fixed(

    const Word32 *x, /* i  : vector x                        */

    const Word32 *A, /* i  : Cholesky  matrix A              */

    const Word16 N   /* i  : vector & matrix size            */

);

#endif



void v_mult_mat_fx(

    Word32 *y_fx, /* o  : the product x*A                         */

lib_com/ivas_tools.c

+4 −10
Original line number Diff line number Diff line
@@ -606,6 +606,7 @@ void v_sub32_fx( 
 * Therefore, S=A*A' where A is upper triangular matrix of size (m*m+m)/2 (zeros ommitted, column-wise)

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



#ifndef DOT_PROD_CHOLESKY_64BIT

/*! r: the dot product x'*A*A'*x */

Word32 dot_product_cholesky_fixed(

    const Word32 *x, /* i  : vector x                        Q31 - exp_x*/
@@ -642,17 +643,9 @@ Word32 dot_product_cholesky_fixed( 


    return suma;

}



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

 * dot_product_cholesky()

 *

 * Calculates dot product of type x'*A*A'*x, where x is column vector of size m,

 * and A is a Cholesky decomposition of some Hermitian matrix S whose size is m*m.

 * Therefore, S=A*A' where A is upper triangular matrix of size (m*m+m)/2 (zeros ommitted, column-wise)

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



#else

/*! r: the dot product x'*A*A'*x */

Word64 dot_product_cholesky_fixed64(

Word64 dot_product_cholesky_fixed(

    const Word32 *x, /* i  : vector x                        Q31 - exp_x*/

    const Word32 *A, /* i  : Cholesky  matrix A              Q31 - exp_A*/

    const Word16 N   /* i  : vector & matrix size            Q0*/
@@ -686,6 +679,7 @@ Word64 dot_product_cholesky_fixed64( 


    return suma;

}

#endif



void v_mult_mat_fixed(

    Word32 *y,       /* o  : the product x*A               Qx - guardbits*/

lib_com/options.h

+1 −0
Original line number Diff line number Diff line
@@ -169,4 +169,5 @@ 
#define FIX_1298                                /* VA: fix possible assert in gaus_enc */

#define FIX_1300_ICA_SHIFT_QUANT_IMPROV         /* VA: Fix to 1300 to improve precision of the lag quantizer */

#define FIX_1301_CORRECT_TD_CNST                /* VA: Fix 1301, correct wrong constant in TD stereo */

#define DOT_PROD_CHOLESKY_64BIT                 /* FhG: Issue 1323, optimized 64 bit implementation of dot_product_cholesky() */

#endif

lib_enc/speech_music_classif_fx.c

+33 −3
Original line number Diff line number Diff line
@@ -1683,8 +1683,16 @@ Word16 ivas_smc_gmm_fx( 
    Word16 flag_odv;

    Word32 lps_fx, lpm_fx, lpn_fx;

    Word32 ps_fx[N_SMC_MIXTURES], pm_fx[N_SMC_MIXTURES], pn_fx[N_SMC_MIXTURES];

#ifndef DOT_PROD_CHOLESKY_64BIT

    Word32 lprob_fx;

    Word16 lprob_exp = 0;

#else

    Word64 wprob_fx;

#endif

    Word32 fvm_fx[N_PCA_COEF];

#ifndef DOT_PROD_CHOLESKY_64BIT

    Word16 fvm_exp = 0;

#endif

    Word32 sum_PS_fx, ps_diff_fx, ps_sta_fx;

    Word32 dlp_fx, wrelE_fx, wdrop_fx, wght_fx;

    Word32 wrise_fx;
@@ -2271,16 +2279,38 @@ Word16 ivas_smc_gmm_fx( 
    FOR( m = 0; m < N_SMC_MIXTURES; m++ )

    {

        v_sub32_fx( FV_fx, &means_speech_fx[m * N_PCA_COEF], fvm_fx, N_PCA_COEF );

        wprob_fx = dot_product_cholesky_fixed64( fvm_fx, &prec_chol_speech_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF );               // Q10

#ifndef DOT_PROD_CHOLESKY_64BIT

        fvm_exp = sub( 31, Qfact_FV );

        lprob_exp = 0;

        move16();

        lprob_fx = dot_product_cholesky_fixed( fvm_fx, &prec_chol_speech_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF, fvm_exp, 31 - 28, &lprob_exp );

        ps_fx[m] = L_sub( L_sub( L_add( log_weights_speech_compute[m], log_det_chol_speech_fx[m] ), L_shl( lprob_fx, sub( Q18 - 1, sub( Q31, lprob_exp ) ) ) ), HALF_N_PCA_COEF_LOG_P12_Q18 ); // Q18

#else

        wprob_fx = dot_product_cholesky_fixed( fvm_fx, &prec_chol_speech_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF );               // Q10

        ps_fx[m] = L_sub( L_sub( L_add( log_weights_speech_compute[m], log_det_chol_speech_fx[m] ), W_shr( wprob_fx, Q10 ) ), HALF_N_PCA_COEF_LOG_P12_Q18 ); // Q18

#endif

        move32();

        v_sub32_fx( FV_fx, &means_music_fx[m * N_PCA_COEF], fvm_fx, N_PCA_COEF );

        wprob_fx = dot_product_cholesky_fixed64( fvm_fx, &prec_chol_music_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF );              // Q10

#ifndef DOT_PROD_CHOLESKY_64BIT

        lprob_exp = 0;

        move16();

        lprob_fx = dot_product_cholesky_fixed( fvm_fx, &prec_chol_music_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF, fvm_exp, 31 - 28, &lprob_exp );

        pm_fx[m] = L_sub( L_sub( L_add( log_weights_music_compute[m], log_det_chol_music_fx[m] ), L_shl( lprob_fx, sub( Q18 - 1, sub( Q31, lprob_exp ) ) ) ), HALF_N_PCA_COEF_LOG_P12_Q18 ); // Q18

#else

        wprob_fx = dot_product_cholesky_fixed( fvm_fx, &prec_chol_music_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF );              // Q10

        pm_fx[m] = L_sub( L_sub( L_add( log_weights_music_compute[m], log_det_chol_music_fx[m] ), W_shr( wprob_fx, Q10 ) ), HALF_N_PCA_COEF_LOG_P12_Q18 ); // Q18

#endif

        move32();

        v_sub32_fx( FV_fx, &means_noise_fx[m * N_PCA_COEF], fvm_fx, N_PCA_COEF );

        wprob_fx = dot_product_cholesky_fixed64( fvm_fx, &prec_chol_noise_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF );              // Q10

#ifndef DOT_PROD_CHOLESKY_64BIT

        lprob_exp = 0;

        move16();

        lprob_fx = dot_product_cholesky_fixed( fvm_fx, &prec_chol_noise_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF, fvm_exp, 31 - 28, &lprob_exp );

        pn_fx[m] = L_sub( L_sub( L_add( log_weights_noise_compute[m], log_det_chol_noise_fx[m] ), L_shl( lprob_fx, sub( Q18 - 1, sub( Q31, lprob_exp ) ) ) ), HALF_N_PCA_COEF_LOG_P12_Q18 ); // Q18

#else

        wprob_fx = dot_product_cholesky_fixed( fvm_fx, &prec_chol_noise_fx[m * ( N_PCA_COEF * N_PCA_COEF + N_PCA_COEF ) / 2], N_PCA_COEF );              // Q10

        pn_fx[m] = L_sub( L_sub( L_add( log_weights_noise_compute[m], log_det_chol_noise_fx[m] ), W_shr( wprob_fx, Q10 ) ), HALF_N_PCA_COEF_LOG_P12_Q18 ); // Q18

#endif

        move32();

    }