Commit a878b04b authored by Sandesh Venkatesh's avatar Sandesh Venkatesh
Browse files

Merge branch 'mch_nbe_opt' into 'main' 

[allow regression] Multichannel path decoder optimizations: Removal of BASOP_Util_Add_Mant32Exp...

See merge request !1864
parents 10f418e3 53143c0b

lib_com/options.h

+1 −0
Original line number Diff line number Diff line
@@ -79,6 +79,7 @@ 




/* Note: each compile switch (FIX_1101_...) is independent from the other ones */

#define OPT_MCH_DEC_V1_NBE

#define OPT_MCH_DEC_V1_BE

#define OPT_MCT_ENC_V2_NBE

#define OPT_SBA_DEC_V2_NBE

lib_dec/ivas_dirac_output_synthesis_cov_fx.c

+44 −1
Original line number Diff line number Diff line
@@ -886,7 +886,11 @@ Word16 computeMixingMatrices_fx( 
    move16();

    FOR( i = 1; i < lengthCx; i++ )

    {

#ifdef OPT_MCH_DEC_V1_NBE

        IF( GT_32( svd_s_buffer_fx[i], L_shl_sat( limit_fx, sub( limit_e, svd_s_buffer_e[i] ) ) ) )

#else  /* OPT_MCH_DEC_V1_NBE */

        IF( BASOP_Util_Cmp_Mant32Exp( svd_s_buffer_fx[i], svd_s_buffer_e[i], limit_fx, limit_e ) > 0 )

#endif /* OPT_MCH_DEC_V1_NBE */

        {

            limit_fx = svd_s_buffer_fx[i];

            move32();
@@ -896,6 +900,7 @@ Word16 computeMixingMatrices_fx( 
    }



    limit_e = add( limit_e, reg_Sx_e );



#ifdef OPT_MCH_DEC_V1_BE

    limit_fx = Madd_32_32( EPSILON_FX, limit_fx, reg_Sx_fx );

#else  /* OPT_MCH_DEC_V1_BE */
@@ -905,7 +910,11 @@ Word16 computeMixingMatrices_fx( 


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

    {

#ifdef OPT_MCH_DEC_V1_NBE

        IF( LT_32( L_shl_sat( svd_s_buffer_fx[i], sub( svd_s_buffer_e[i], limit_e ) ), limit_fx ) )

#else  /* OPT_MCH_DEC_V1_NBE */

        IF( BASOP_Util_Cmp_Mant32Exp( svd_s_buffer_fx[i], svd_s_buffer_e[i], limit_fx, limit_e ) < 0 )

#endif /* OPT_MCH_DEC_V1_NBE */

        {

            svd_s_buffer_fx[i] = limit_fx;

            move32();
@@ -950,9 +959,16 @@ Word16 computeMixingMatrices_fx( 
    matrix_product_diag_fx( Q_Cx_fx, Q_Cx_e, lengthCy, lengthCx, 0, Q_fx, Q_e, lengthCy, lengthCx, 1, Cy_hat_diag_fx, &Cy_hat_diag_e );





#ifdef OPT_MCH_DEC_V1_NBE

    Word16 com_e = sub( limit_e, Cy_hat_diag_e );

#endif /* OPT_MCH_DEC_V1_NBE */

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

    {

#ifdef OPT_MCH_DEC_V1_NBE

        IF( GT_32( Cy_hat_diag_fx[i], L_shl_sat( limit_fx, com_e ) ) )

#else  /* OPT_MCH_DEC_V1_NBE */

        IF( BASOP_Util_Cmp_Mant32Exp( Cy_hat_diag_fx[i], Cy_hat_diag_e, limit_fx, limit_e ) > 0 )

#endif /* OPT_MCH_DEC_V1_NBE */

        {

            limit_fx = Cy_hat_diag_fx[i];

            move32();
@@ -968,11 +984,19 @@ Word16 computeMixingMatrices_fx( 
#endif                                                          /* OPT_MCH_DEC_V1_BE */

    limit_e = add( limit_e, reg_ghat_e );



#ifdef OPT_MCH_DEC_V1_NBE

    com_e = sub( Cy_hat_diag_e, limit_e );

#endif /* OPT_MCH_DEC_V1_NBE */

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

    {

        Cy_hat_diag_buff_e[i] = Cy_hat_diag_e;

        move16();



#ifdef OPT_MCH_DEC_V1_NBE

        IF( GT_32( limit_fx, L_shl_sat( Cy_hat_diag_fx[i], com_e ) ) )

#else  /* OPT_MCH_DEC_V1_NBE */

        IF( BASOP_Util_Cmp_Mant32Exp( limit_fx, limit_e, Cy_hat_diag_fx[i], Cy_hat_diag_buff_e[i] ) > 0 ) /* Computing Cy_hat_diag = max(Cy_hat_diag,limit) */

#endif /* OPT_MCH_DEC_V1_NBE */

        {

            Cy_hat_diag_fx[i] = limit_fx;

            move32();
@@ -1392,7 +1416,11 @@ Word16 computeMixingMatricesResidual_fx( 


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

    {

#ifdef OPT_MCH_DEC_V1_NBE

        IF( GT_32( Kx_fx[i], L_shl_sat( limit_fx, sub( limit_e, Kx_fx_e[i] ) ) ) )

#else  /* OPT_MCH_DEC_V1_NBE */

        IF( BASOP_Util_Cmp_Mant32Exp( Kx_fx[i], Kx_fx_e[i], limit_fx, limit_e ) > 0 )

#endif /* OPT_MCH_DEC_V1_NBE */

        {

            div_tmp = Kx_fx[i];

            move32();
@@ -1433,9 +1461,16 @@ Word16 computeMixingMatricesResidual_fx( 
    Cy_hat_diag_e = Cx_e;

    move16();



#ifdef OPT_MCH_DEC_V1_NBE

    Word16 com_e = sub( limit_e, Cy_hat_diag_e );

#endif /* OPT_MCH_DEC_V1_NBE */

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

    {

#ifdef OPT_MCH_DEC_V1_NBE

        IF( GT_32( Cy_hat_diag_fx[i], L_shl_sat( limit_fx, com_e ) ) )

#else  /* OPT_MCH_DEC_V1_NBE */

        IF( BASOP_Util_Cmp_Mant32Exp( Cy_hat_diag_fx[i], Cy_hat_diag_e, limit_fx, limit_e ) > 0 )

#endif /* OPT_MCH_DEC_V1_NBE */

        {

            limit_fx = Cy_hat_diag_fx[i];

            move32();
@@ -1453,11 +1488,19 @@ Word16 computeMixingMatricesResidual_fx( 
    limit_e = add( limit_e, reg_ghat_e );



    /* Computing G_hat */



#ifdef OPT_MCH_DEC_V1_NBE

    com_e = sub( Cy_hat_diag_e, limit_e );

#endif /* OPT_MCH_DEC_V1_NBE */

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

    {

        Cy_hat_diag_fx_e[i] = Cy_hat_diag_e;

        move16();

#ifdef OPT_MCH_DEC_V1_NBE

        IF( GT_32( limit_fx, L_shl_sat( Cy_hat_diag_fx[i], com_e ) ) ) /* Computing Cy_hat_diag = max(Cy_hat_diag,limit) */

#else                                                                  /* OPT_MCH_DEC_V1_NBE */

        IF( BASOP_Util_Cmp_Mant32Exp( limit_fx, limit_e, Cy_hat_diag_fx[i], Cy_hat_diag_e ) > 0 ) /* Computing Cy_hat_diag = max(Cy_hat_diag,limit) */

#endif                                                                 /* OPT_MCH_DEC_V1_NBE */

        {

            Cy_hat_diag_fx[i] = limit_fx;

            move32();

lib_dec/ivas_svd_dec_fx.c

+121 −9
Original line number Diff line number Diff line
@@ -322,7 +322,11 @@ Word16 svd_fx( 
        move16();

        FOR( iCh = 0; iCh < lengthSingularValues - 1; iCh++ )

        {

#ifdef OPT_MCH_DEC_V1_NBE

            IF( LT_32( L_shl_sat( singularValues_fx[iCh], sub( singularValues_fx_e[iCh], singularValues_fx_e[iCh + 1] ) ), singularValues_fx[iCh + 1] ) )

#else  /* OPT_MCH_DEC_V1_NBE */

            IF( BASOP_Util_Cmp_Mant32Exp( singularValues_fx[iCh], singularValues_fx_e[iCh], singularValues_fx[iCh + 1], singularValues_fx_e[iCh + 1] ) < 0 )

#endif /* OPT_MCH_DEC_V1_NBE */

            {

                condition = 1;

                move16();
@@ -427,14 +431,24 @@ static Word16 BidagonalDiagonalisation_fx( 


            FOR( jCh = iCh; jCh >= 0; jCh-- )

            {

                split = sub( jCh, 1 );                                                                                                                         /* Q0 */

#ifdef OPT_MCH_DEC_V1_NBE

                Word16 com_e = s_max( secDiag_new_e[jCh], eps_x_e );

                IF( LE_32( L_shr( L_abs( secDiag_fx[jCh] ), sub( com_e, secDiag_new_e[jCh] ) ), L_shr( Mpy_32_32( CONVERGENCE_FACTOR_FX, eps_x ), sub( com_e, eps_x_e ) ) ) ) /* is secDiag[ch] vanishing compared to eps_x */

#else

                split = sub( jCh, 1 ); /* Q0 */                                                                                                                                  /* OPT_MCH_DEC_V1_NBE */

                IF( LE_16( BASOP_Util_Cmp_Mant32Exp( L_abs( secDiag_fx[jCh] ), secDiag_new_e[jCh], Mpy_32_32( CONVERGENCE_FACTOR_FX, eps_x ), eps_x_e ), 0 ) )                   /* is secDiag[ch] vanishing compared to eps_x */

#endif /* OPT_MCH_DEC_V1_NBE */

                {

                    found_split = 0;

                    move16();

                    BREAK;

                }

#ifdef OPT_MCH_DEC_V1_NBE

                com_e = s_max( singularValues_new_e[jCh - 1], eps_x_e );

                IF( LE_32( L_shr( L_abs( singularValues_fx[jCh - 1] ), sub( com_e, singularValues_new_e[jCh - 1] ) ), L_shr( Mpy_32_32( CONVERGENCE_FACTOR_FX, eps_x ), sub( com_e, eps_x_e ) ) ) ) /* is singularValues[jCh - 1] vanishing compared to eps_x */

#else                                                                                                                                                                                               /* OPT_MCH_DEC_V1_NBE */

                IF( LE_16( BASOP_Util_Cmp_Mant32Exp( L_abs( singularValues_fx[split] ), singularValues_new_e[split], Mpy_32_32( CONVERGENCE_FACTOR_FX, eps_x ), eps_x_e ), 0 ) ) /* is singularValues[split] vanishing compared to eps_x */

#endif                                                                                                                                                                                              /* OPT_MCH_DEC_V1_NBE */

                {

                    BREAK;

                }
@@ -462,14 +476,21 @@ static Word16 BidagonalDiagonalisation_fx( 
                move32();

                c_e = 0;

                move16();



#ifdef OPT_MCH_DEC_V1_NBE

                split = sub( jCh, 1 ); /* Q0 */

#endif                                 /* OPT_MCH_DEC_V1_NBE */

                FOR( kCh = jCh; kCh <= iCh; kCh++ )

                {

                    g = Mpy_32_32( s, secDiag_fx[kCh] ); /* exp(s_e + secDiag_new_e) */

                    g_e = add( s_e, secDiag_new_e[kCh] );

                    secDiag_fx[kCh] = Mpy_32_32( c, secDiag_fx[kCh] ); /* exp(c_e + secDiag_new_e) */

                    secDiag_new_e[kCh] = add( c_e, secDiag_new_e[kCh] );

#ifdef OPT_MCH_DEC_V1_NBE

                    Word16 com_e = s_max( g_e, eps_x_e );

                    IF( LE_32( L_shr( L_abs( g ), sub( com_e, g_e ) ), L_shr( Mpy_32_32( CONVERGENCE_FACTOR_FX, eps_x ), sub( com_e, eps_x_e ) ) ) )

#else  /* OPT_MCH_DEC_V1_NBE */

                    IF( LE_16( BASOP_Util_Cmp_Mant32Exp( L_abs( g ), g_e, Mpy_32_32( CONVERGENCE_FACTOR_FX, eps_x ), eps_x_e ), 0 ) )                                            /* is singularValues[split] vanishing compared to eps_x */

#endif /* OPT_MCH_DEC_V1_NBE */

                    {

                        BREAK;

                    }
@@ -929,9 +950,15 @@ static void biDiagonalReductionLeft_fx( 
            Word16 invVal_e;

            Word32 invVal;

            invVal = BASOP_Util_Divide3232_Scale_newton( MAXVAL_WORD32, maxWithSign_fx( *sig_x ), &invVal_e );

#ifdef OPT_MCH_DEC_V1_NBE

            Word64 temp = 0;

            move64();

            Word16 max_e = MIN_16;

#else  /* OPT_MCH_DEC_V1_NBE */

            norm_x = 0;

            move32();

            norm_x_e = 0;

#endif /* OPT_MCH_DEC_V1_NBE */

            move16();

            FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */

            {
@@ -940,8 +967,25 @@ static void biDiagonalReductionLeft_fx( 
                move32();

                singularVectors2_e[jCh][currChannel] = sub( add( invVal_e, sub( singularVectors2_e[jCh][currChannel], *sig_x_e ) ), temp_e );

                move16();

#ifdef OPT_MCH_DEC_V1_NBE

                max_e = s_max( max_e, singularVectors2_e[jCh][currChannel] );

#else  /* OPT_MCH_DEC_V1_NBE */

                norm_x = BASOP_Util_Add_Mant32Exp( norm_x, norm_x_e, Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][currChannel] ), shl( singularVectors2_e[jCh][currChannel], 1 ), &norm_x_e ); /* exp(norm_x_e) */

#endif /* OPT_MCH_DEC_V1_NBE */

            }



#ifdef OPT_MCH_DEC_V1_NBE

            FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */

            {

                temp = W_add( temp, L_shr( Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][currChannel] ), shl( sub( max_e, singularVectors2_e[jCh][currChannel] ), 1 ) ) );

            }



            Word16 nrm = W_norm( temp );

            nrm = sub( nrm, 32 );

            norm_x = W_shl_sat_l( temp, nrm );

            norm_x_e = sub( add( max_e, max_e ), nrm );

#endif /* OPT_MCH_DEC_V1_NBE */



            IF( GT_16( norm_x_e, 0 ) )

            {

                norm_x = MAX_32;
@@ -969,6 +1013,30 @@ static void biDiagonalReductionLeft_fx( 


            FOR( iCh = currChannel + 1; iCh < nChannelsC; iCh++ ) /* nChannelsC */

            {

#ifdef OPT_MCH_DEC_V1_NBE

                Word16 max2_e = MIN_16;

                max_e = MIN_16;

                move16();

                move16();

                temp = 0;

                move64();



                FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */

                {

                    max_e = s_max( max_e, singularVectors2_e[jCh][currChannel] ); /* exp(norm_x_e) */

                    max2_e = s_max( max2_e, singularVectors2_e[jCh][iCh] );       /* exp(norm_x_e) */

                }

                max_e = add( max_e, max2_e );



                FOR( jCh = idx; jCh < nChannelsL; jCh++ ) /* nChannelsL */

                {

                    temp = W_add( temp, L_shr( Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][iCh] ), sub( max_e, add( singularVectors2_e[jCh][currChannel], singularVectors2_e[jCh][iCh] ) ) ) );

                }

                nrm = W_norm( temp );

                nrm = sub( nrm, 32 );

                norm_x = W_shl_sat_l( temp, nrm );

                norm_x_e = sub( max_e, nrm );

#else  /* OPT_MCH_DEC_V1_NBE */

                norm_x = 0;

                move32();

                norm_x_e = 0;
@@ -977,6 +1045,7 @@ static void biDiagonalReductionLeft_fx( 
                {

                    norm_x = BASOP_Util_Add_Mant32Exp( norm_x, norm_x_e, Mpy_32_32( singularVectors[jCh][currChannel], singularVectors[jCh][iCh] ), add( singularVectors2_e[jCh][currChannel], singularVectors2_e[jCh][iCh] ), &norm_x_e ); /* exp(norm_x_e) */

                }

#endif /* OPT_MCH_DEC_V1_NBE */



                f = Mpy_32_32( norm_x, invVal ); /* invVal_e + (norm_x_e - r_e) */

                f_e = add( invVal_e, sub( norm_x_e, r_e ) );
@@ -1228,8 +1297,16 @@ static void singularVectorsAccumulationLeft_fx( 
                move32();

            }

        }

#ifdef OPT_MCH_DEC_V1_NBE

        Word16 exp = s_max( singularVectors_Left_e[nCh][nCh], 1 );

        singularVectors_Left[nCh][nCh] = L_sub( L_shr( singularVectors_Left[nCh][nCh], sub( exp, singularVectors_Left_e[nCh][nCh] ) ), L_shr( MINUS_ONE_IN_Q31, exp ) ); /* exp(sing_exp2) */

        move32();

        singularVectors_Left_e[nCh][nCh] = exp;

        move16();

#else  /* OPT_MCH_DEC_V1_NBE */

        singularVectors_Left[nCh][nCh] = BASOP_Util_Add_Mant32Exp( singularVectors_Left[nCh][nCh], singularVectors_Left_e[nCh][nCh], ONE_IN_Q30, 1, &singularVectors_Left_e[nCh][nCh] ); /* exp(sing_exp2) */

        move32();

#endif /* OPT_MCH_DEC_V1_NBE */

    }

    // fclose(fp);

    FOR( nCh = 0; nCh < nChannelsL; nCh++ )
@@ -1292,21 +1369,56 @@ static void singularVectorsAccumulationRight_fx( 


                FOR( iCh = nCh + 1; iCh < nChannelsC; iCh++ ) /* nChannelsC */

                {

#ifdef OPT_MCH_DEC_V1_NBE

                    Word64 norm_val = 0;

                    move64();

                    Word16 maxL_e = MIN_16;

                    Word16 maxR_e = MIN_16;

                    Word16 maxR2_e = MIN_16;

                    move16();

                    move16();

                    move16();

                    FOR( k = nCh + 1; k < nChannelsC; k++ ) /* nChannelsC */

                    {

                        maxL_e = s_max( maxL_e, singularVectors_Left_e[nCh][k] );

                        maxR_e = s_max( maxR_e, sing_right_exp[k][iCh] );

                        maxR2_e = s_max( maxR2_e, sing_right_exp[k][nCh] );

                    }

#else  /* OPT_MCH_DEC_V1_NBE */

                    norm_y = 0;

                    move32();

                    norm_y_e = 0;

                    move16();

#endif /* OPT_MCH_DEC_V1_NBE */



                    FOR( k = nCh + 1; k < nChannelsC; k++ ) /* nChannelsC */

                    {

#ifdef OPT_MCH_DEC_V1_NBE

                        norm_val = W_mac_32_32( norm_val, L_shr( singularVectors_Left[nCh][k], sub( maxL_e, singularVectors_Left_e[nCh][k] ) ), L_shr( singularVectors_Right[k][iCh], sub( maxR_e, sing_right_exp[k][iCh] ) ) );

#else  /* OPT_MCH_DEC_V1_NBE */

                        norm_y = BASOP_Util_Add_Mant32Exp( norm_y, norm_y_e, Mpy_32_32( singularVectors_Left[nCh][k], singularVectors_Right[k][iCh] ), add( singularVectors_Left_e[nCh][k], sing_right_exp[k][iCh] ), &norm_y_e );                               /* exp(norm_y_e) */

#endif /* OPT_MCH_DEC_V1_NBE */

                    }

#ifdef OPT_MCH_DEC_V1_NBE

                    norm_y_e = W_norm( norm_val );

                    norm_y = W_extract_h( W_shl( norm_val, norm_y_e ) );

                    norm_y_e = sub( add( maxL_e, maxR_e ), norm_y_e );



                    Word16 max_new = s_max( maxR_e, add( maxR2_e, norm_y_e ) );

#endif                                                      /* OPT_MCH_DEC_V1_NBE */

                    FOR( k = nCh + 1; k < nChannelsC; k++ ) /* nChannelsC */

                    {

#ifdef OPT_MCH_DEC_V1_NBE

                        Word32 temp = Mpy_32_32( norm_y, singularVectors_Right[k][nCh] );

                        Word32 op2 = L_shr( temp, sub( max_new, add( norm_y_e, sing_right_exp[k][nCh] ) ) );

                        singularVectors_Right[k][iCh] = L_add_sat( L_shr( singularVectors_Right[k][iCh], sub( max_new, sing_right_exp[k][iCh] ) ), op2 ); /* exp(sing_right_exp) */

                        move32();

                        singularVectors_Right[k][iCh] = L_shl_sat( singularVectors_Right[k][iCh], max_new ); /* Q31 */

#else                                                                                                        /* OPT_MCH_DEC_V1_NBE */

                        singularVectors_Right[k][iCh] = BASOP_Util_Add_Mant32Exp( singularVectors_Right[k][iCh], sing_right_exp[k][iCh], Mpy_32_32( norm_y, singularVectors_Right[k][nCh] ), add( norm_y_e, sing_right_exp[k][nCh] ), &sing_right_exp[k][iCh] ); /* exp(sing_right_exp) */

                        move32();

                        singularVectors_Right[k][iCh] = L_shl_sat( singularVectors_Right[k][iCh], sing_right_exp[k][iCh] ); /* Q31 */

#endif                                                                                                       /* OPT_MCH_DEC_V1_NBE */

                        move32();

                        sing_right_exp[k][iCh] = 0;

                        move16();

lib_rend/ivas_dirac_decorr_dec_fx.c

+97 −2

File changed.

Preview size limit exceeded, changes collapsed.