Commit 5e45b8d5 authored by Sandesh Venkatesh's avatar Sandesh Venkatesh
Browse files

Merge branch 'high-complexity-param-mc-prm-est' into 'main' 

improve high complexity of param_mc_prm_est: MC/7-1-4/128kBit reduced by 166 WMOPS [allow regression]

See merge request !1112
parents 86f63a90 6892a6fc

lib_com/options.h

+3 −0
Original line number Diff line number Diff line
@@ -72,4 +72,7 @@ 
#define OPT_AVOID_STATE_BUF_RESCALE             /* Optimization made to avoid rescale of synth state buffer */

#define FIX_1310_SPEEDUP_ivas_dirac_dec_get_response_fx                 /*FhG: WMOPS tuning, nonbe*/

#define FIX_1310_SPEEDUP_ivas_dirac_dec_output_synthesis_process_slot   /*FhG: WMOPS tuning, nonbe*/

/* Both following 2 macros (IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST*) are independent from each other, they refer to different code blocks */

#define IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_BE    /* FhG: reduces WMOPS of param_mc_prm_est, bit-exact to previous version */

#define IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_NONBE /* FhG: reduces WMOPS of param_mc_prm_est, not bit-exact to previous version */

#endif

lib_enc/ivas_mc_param_enc_fx.c

+148 −2
Original line number Diff line number Diff line
@@ -720,9 +720,16 @@ static void ivas_param_mc_param_est_enc_fx( 
        }

    }



#if defined( IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_BE ) || defined( IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_NONBE )

    Word16 gb = find_guarded_bits_fx( l_ts );

    Word16 add20gb = add( 20, gb );

#endif



    FOR( ts = start_ts; ts < num_time_slots; ts++ )

    {

#if !defined( IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_BE ) && !defined( IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_NONBE )

        Word16 gb = find_guarded_bits_fx( l_ts );

#endif

        ivas_fb_mixer_get_windowed_fr_fx( hParamMC->hFbMixer, pcm_in_fx, p_slot_frame_f_real_fx, p_slot_frame_f_imag_fx, l_ts, l_ts, hParamMC->hFbMixer->fb_cfg->num_in_chans, gb );

        ivas_fb_mixer_update_prior_input_fx( hParamMC->hFbMixer, pcm_in_fx, l_ts, hParamMC->hFbMixer->fb_cfg->num_in_chans );
@@ -731,6 +738,7 @@ static void ivas_param_mc_param_est_enc_fx( 
        FOR( i = 0; i < nchan_input; i++ )

        {

            pcm_in_fx[i] += l_ts;

            move32();

        }

        /* Computing the downmix */

        FOR( cur_param_band = 0; cur_param_band < hParamMC->max_param_band_abs_cov; cur_param_band++ )
@@ -748,6 +756,7 @@ static void ivas_param_mc_param_est_enc_fx( 


                FOR( ch_idx1 = 0; ch_idx1 < nchan_transport; ++ch_idx1 )

                {

#ifndef IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_BE

                    dmx_real_fx[ch_idx1] = 0;

                    move32();

                    dmx_real_e[ch_idx1] = 0;
@@ -766,6 +775,28 @@ static void ivas_param_mc_param_est_enc_fx( 
                        move32();

                        p_dmx_fac_fx++;

                    }

#else

                    Word32 real_fx = L_add( 0, 0 );

                    Word16 real_e = add( 0, 0 );

                    Word32 imag_fx = L_add( 0, 0 );

                    Word16 imag_e = add( 0, 0 );

                    FOR( inp_ch = 0; inp_ch < nchan_input; inp_ch++ )

                    {

                        L_tmp = Mpy_32_32( slot_frame_f_real_fx[inp_ch][cur_cldfb_band], ( *p_dmx_fac_fx ) );

                        real_fx = BASOP_Util_Add_Mant32Exp( real_fx, real_e, L_tmp, add20gb, &real_e );

                        L_tmp = Mpy_32_32( slot_frame_f_imag_fx[inp_ch][cur_cldfb_band], ( *p_dmx_fac_fx ) );

                        imag_fx = BASOP_Util_Add_Mant32Exp( imag_fx, imag_e, L_tmp, add20gb, &imag_e );

                        p_dmx_fac_fx++;

                    }

                    dmx_real_fx[ch_idx1] = real_fx;

                    dmx_real_e[ch_idx1] = real_e;

                    dmx_imag_fx[ch_idx1] = imag_fx;

                    dmx_imag_e[ch_idx1] = imag_e;

                    move32();

                    move16();

                    move32();

                    move16();

#endif

                }



                /* Cx for transport channels */
@@ -805,13 +836,49 @@ static void ivas_param_mc_param_est_enc_fx( 
                /* Cy for input channels */

                FOR( ch_idx1 = 0; ch_idx1 < nchan_input; ++ch_idx1 )

                {

#ifdef IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_NONBE

                    a_e = norm_l( slot_frame_f_real_fx[ch_idx1][cur_cldfb_band] );

                    a_fx = L_shl( slot_frame_f_real_fx[ch_idx1][cur_cldfb_band], a_e );

                    a_e = sub( add20gb, a_e );

                    if ( a_fx == 0 )

                    {

                        a_e = 0;

                        move16();

                    }

                    b_e = norm_l( slot_frame_f_imag_fx[ch_idx1][cur_cldfb_band] );

                    b_fx = L_shl( slot_frame_f_imag_fx[ch_idx1][cur_cldfb_band], b_e );

                    b_e = sub( add20gb, b_e );

                    if ( b_fx == 0 )

                    {

                        b_e = 0;

                        move16();

                    }

#endif

                    FOR( ch_idx2 = ch_idx1; ch_idx2 < nchan_input; ++ch_idx2 )

                    {

#ifndef IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_NONBE

                        a_fx = BASOP_Util_Add_Mant32Exp( slot_frame_f_real_fx[ch_idx1][cur_cldfb_band], add( 20, gb ), 0, 0, &a_e );

                        b_fx = BASOP_Util_Add_Mant32Exp( slot_frame_f_imag_fx[ch_idx1][cur_cldfb_band], add( 20, gb ), 0, 0, &b_e );

                        c_fx = BASOP_Util_Add_Mant32Exp( slot_frame_f_real_fx[ch_idx2][cur_cldfb_band], add( 20, gb ), 0, 0, &c_e );

                        d_fx = BASOP_Util_Add_Mant32Exp( slot_frame_f_imag_fx[ch_idx2][cur_cldfb_band], add( 20, gb ), 0, 0, &d_e );



#else

                        c_e = norm_l( slot_frame_f_real_fx[ch_idx2][cur_cldfb_band] );

                        c_fx = L_shl( slot_frame_f_real_fx[ch_idx2][cur_cldfb_band], c_e );

                        c_e = sub( add20gb, c_e );

                        if ( c_fx == 0 )

                        {

                            c_e = 0;

                            move16();

                        }

                        d_e = norm_l( slot_frame_f_imag_fx[ch_idx2][cur_cldfb_band] );

                        d_fx = L_shl( slot_frame_f_imag_fx[ch_idx2][cur_cldfb_band], d_e );

                        d_e = sub( add20gb, d_e );

                        if ( d_fx == 0 )

                        {

                            d_e = 0;

                            move16();

                        }

#endif

                        /* (a-ib)(c+id) = ac + bd + i(ad-bc) */

                        L_tmp = BASOP_Util_Add_Mant32Exp( Mpy_32_32( a_fx, c_fx ), add( a_e, c_e ), Mpy_32_32( b_fx, d_fx ), add( b_e, d_e ), &tmp_e );

                        Cy_sum_fx[cur_param_band][ch_idx1][ch_idx2] = BASOP_Util_Add_Mant32Exp( Cy_sum_fx[cur_param_band][ch_idx1][ch_idx2], Cy_sum_e[cur_param_band][ch_idx1][ch_idx2],
@@ -841,6 +908,7 @@ static void ivas_param_mc_param_est_enc_fx( 


                FOR( ch_idx1 = 0; ch_idx1 < nchan_transport; ++ch_idx1 )

                {

#ifndef IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_BE

                    dmx_real_fx[ch_idx1] = 0;

                    move32();

                    dmx_real_e[ch_idx1] = 0;
@@ -860,13 +928,49 @@ static void ivas_param_mc_param_est_enc_fx( 
                        move32();

                        p_dmx_fac_fx++;

                    }

#else

                    Word32 real_fx = L_add( 0, 0 );

                    Word16 real_e = 0;

                    move16();

                    Word32 imag_fx = L_add( 0, 0 );

                    Word16 imag_e = 0;

                    move16();



                    FOR( inp_ch = 0; inp_ch < nchan_input; inp_ch++ )

                    {

                        L_tmp = Mpy_32_32( slot_frame_f_real_fx[inp_ch][cur_cldfb_band], ( *p_dmx_fac_fx ) );

                        real_fx = BASOP_Util_Add_Mant32Exp( real_fx, real_e, L_tmp, add20gb, &real_e );

                        L_tmp = Mpy_32_32( slot_frame_f_imag_fx[inp_ch][cur_cldfb_band], ( *p_dmx_fac_fx ) );

                        imag_fx = BASOP_Util_Add_Mant32Exp( imag_fx, imag_e, L_tmp, add20gb, &imag_e );

                        p_dmx_fac_fx++;

                    }

                    dmx_real_fx[ch_idx1] = real_fx;

                    move32();

                    dmx_real_e[ch_idx1] = real_e;

                    move16();

                    dmx_imag_fx[ch_idx1] = imag_fx;

                    move32();

                    dmx_imag_e[ch_idx1] = imag_e;

                    move16();

#endif

                }



                /* Cx for transport channels */

                FOR( ch_idx1 = 0; ch_idx1 < nchan_transport; ++ch_idx1 )

                {

#ifdef IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_BE

                    a_fx = dmx_real_fx[ch_idx1];

                    move32();

                    a_e = dmx_real_e[ch_idx1];

                    move16();

                    b_fx = dmx_imag_fx[ch_idx1];

                    move32();

                    b_e = dmx_imag_e[ch_idx1];

                    move16();

#endif

                    FOR( ch_idx2 = 0; ch_idx2 < nchan_transport; ++ch_idx2 )

                    {

#ifndef IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_BE

                        a_fx = dmx_real_fx[ch_idx1];

                        move32();

                        a_e = dmx_real_e[ch_idx1];
@@ -888,6 +992,12 @@ static void ivas_param_mc_param_est_enc_fx( 
                        L_tmp = BASOP_Util_Add_Mant32Exp( Mpy_32_32( a_fx, c_fx ), add( a_e, c_e ), Mpy_32_32( b_fx, d_fx ), add( b_e, d_e ), &tmp_e );

                        Cx_sum_fx[cur_param_band][ch_idx1][ch_idx2] = BASOP_Util_Add_Mant32Exp( Cx_sum_fx[cur_param_band][ch_idx1][ch_idx2], Cx_sum_e[cur_param_band][ch_idx1][ch_idx2], L_tmp, tmp_e,

                                                                                                &Cx_sum_e[cur_param_band][ch_idx1][ch_idx2] );

#else

                        /* (a-ib)(c+id) = ac + bd + i(ad-bc) */

                        L_tmp = BASOP_Util_Add_Mant32Exp( Mpy_32_32( a_fx, dmx_real_fx[ch_idx2] ), add( a_e, dmx_real_e[ch_idx2] ), Mpy_32_32( b_fx, dmx_imag_fx[ch_idx2] ), add( b_e, dmx_imag_e[ch_idx2] ), &tmp_e );

                        Cx_sum_fx[cur_param_band][ch_idx1][ch_idx2] = BASOP_Util_Add_Mant32Exp( Cx_sum_fx[cur_param_band][ch_idx1][ch_idx2], Cx_sum_e[cur_param_band][ch_idx1][ch_idx2], L_tmp, tmp_e,

                                                                                                &Cx_sum_e[cur_param_band][ch_idx1][ch_idx2] );

#endif

                        move32();

                    }

                }
@@ -895,13 +1005,49 @@ static void ivas_param_mc_param_est_enc_fx( 
                /* Cy for input channels */

                FOR( ch_idx1 = 0; ch_idx1 < nchan_input; ++ch_idx1 )

                {

#ifdef IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_NONBE

                    a_e = norm_l( slot_frame_f_real_fx[ch_idx1][cur_cldfb_band] );

                    a_fx = L_shl( slot_frame_f_real_fx[ch_idx1][cur_cldfb_band], a_e );

                    a_e = sub( add20gb, a_e );

                    if ( a_fx == 0 )

                    {

                        a_e = 0;

                        move16();

                    }

                    b_e = norm_l( slot_frame_f_imag_fx[ch_idx1][cur_cldfb_band] );

                    b_fx = L_shl( slot_frame_f_imag_fx[ch_idx1][cur_cldfb_band], b_e );

                    b_e = sub( add20gb, b_e );

                    if ( b_fx == 0 )

                    {

                        b_e = 0;

                        move16();

                    }

#endif

                    FOR( ch_idx2 = ch_idx1; ch_idx2 < nchan_input; ++ch_idx2 )

                    {

#ifndef IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_NONBE

                        a_fx = BASOP_Util_Add_Mant32Exp( slot_frame_f_real_fx[ch_idx1][cur_cldfb_band], add( 20, gb ), 0, 0, &a_e );

                        b_fx = BASOP_Util_Add_Mant32Exp( slot_frame_f_imag_fx[ch_idx1][cur_cldfb_band], add( 20, gb ), 0, 0, &b_e );

                        c_fx = BASOP_Util_Add_Mant32Exp( slot_frame_f_real_fx[ch_idx2][cur_cldfb_band], add( 20, gb ), 0, 0, &c_e );

                        d_fx = BASOP_Util_Add_Mant32Exp( slot_frame_f_imag_fx[ch_idx2][cur_cldfb_band], add( 20, gb ), 0, 0, &d_e );



#else

                        c_e = norm_l( slot_frame_f_real_fx[ch_idx2][cur_cldfb_band] );

                        c_fx = L_shl( slot_frame_f_real_fx[ch_idx2][cur_cldfb_band], c_e );

                        c_e = sub( add20gb, c_e );

                        if ( c_fx == 0 )

                        {

                            c_e = 0;

                            move16();

                        }

                        d_e = norm_l( slot_frame_f_imag_fx[ch_idx2][cur_cldfb_band] );

                        d_fx = L_shl( slot_frame_f_imag_fx[ch_idx2][cur_cldfb_band], d_e );

                        d_e = sub( add20gb, d_e );

                        if ( d_fx == 0 )

                        {

                            d_e = 0;

                            move16();

                        }

#endif

                        /* (a-ib)(c+id) = ac + bd + i(ad-bc) */

                        L_tmp = BASOP_Util_Add_Mant32Exp( Mpy_32_32( a_fx, c_fx ), add( a_e, c_e ), Mpy_32_32( b_fx, d_fx ), add( b_e, d_e ), &tmp_e );

                        Cy_sum_fx[cur_param_band][ch_idx1][ch_idx2] = BASOP_Util_Add_Mant32Exp( Cy_sum_fx[cur_param_band][ch_idx1][ch_idx2], Cy_sum_e[cur_param_band][ch_idx1][ch_idx2], L_tmp, tmp_e,