Commit df4ab5e6 authored by thomas dettbarn's avatar thomas dettbarn
Browse files

replaced the costly basop_util_mant2exp() function with a 64 bit addtion.... 

replaced the costly basop_util_mant2exp() function with a 64 bit addtion. there is a potential overflow which needs to be adressed first.
parent 4ac67249

lib_com/ivas_prot_fx.h

+6 −0
Original line number Diff line number Diff line
@@ -4714,6 +4714,12 @@ Word32 dot_product_cholesky_fixed( 
    const Word16 exp_A,

    Word16 *exp_sum );



Word64 dot_product_cholesky_fixed64(

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

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

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

);



void v_mult_mat_fx(

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

    Word16 *y_q_fx,

lib_com/ivas_tools.c

+39 −0
Original line number Diff line number Diff line
@@ -642,6 +642,45 @@ 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)

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



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

Word64 dot_product_cholesky_fixed64(

    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*/

)

{

    Word16 i, j;

    Word64 suma, tmp_sum;

    Word32 mul;

    const Word32 *pt_x, *pt_A;

    pt_A = A;

    suma = 0;

    move32();

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

    {

        tmp_sum = 0;

        move32();

        pt_x = x;



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

        {

            mul = Mpy_32_32( *pt_x++, *pt_A++ );

            tmp_sum = W_add( tmp_sum, W_deposit32_l( mul ) );

        }



        suma = W_mac_32_32( suma, tmp_sum, tmp_sum );	// TODO: make sure that this does not overflow. 

    }



    return suma;

}

void v_mult_mat_fixed(

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

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

lib_enc/speech_music_classif_fx.c

+7 −16
Original line number Diff line number Diff line
@@ -1683,10 +1683,8 @@ 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];

    Word32 lprob_fx;

    Word16 lprob_exp = 0;

    Word64 wprob_fx;

    Word32 fvm_fx[N_PCA_COEF];

    Word16 fvm_exp = 0;

    Word32 sum_PS_fx, ps_diff_fx, ps_sta_fx;

    Word32 dlp_fx, wrelE_fx, wdrop_fx, wght_fx;

    Word32 wrise_fx;
@@ -2273,23 +2271,16 @@ 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 );

        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

        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 );

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

        move32();

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

        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

        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 );

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

        move32();

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

        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

        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 );

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

        move32();

    }