Commit c008f4d1 authored by ber's avatar ber
Browse files

cleaned a lot of push/pop wmops

parent 95f38b3d

lib_rend/ivas_dirac_dec_binaural_functions.c

+8 −61
Original line number Diff line number Diff line
@@ -48,7 +48,7 @@ 
#include "wmc_auto.h"



#if ( BINAURAL_CHANNELS == 2 )

#define FIX1072_SPEEDUP_combinedMatrixCalculations // 2 calls: approx 10.5 mhz

//#define FIX1072_SPEEDUP_combinedMatrixCalculations // 2 calls: approx 10.5 mhz - accept regression in failing pipelines

#endif
@@ -92,6 +92,7 @@ 
//                  0       : 282.704 --> 8 Mhz gain



#include <stdio.h>



//#define FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_0 // 8Mhz with 1072 issue stream

//#define FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_1 // 8Mhz with 1072 issue stream

//#define FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_2 //--> leave this inactive (pipeline 374321)
@@ -822,7 +823,6 @@ void ivas_dirac_dec_binaural_render_fx( 


    output_length = 0;

    move16();

    push_wmops( "IDR: ivas_dirac_dec_binaural_internal_fx (IDDBI)" );

    FOR( subframe_idx = first_sf; subframe_idx < last_sf; subframe_idx++ )

    {

        Word16 n_samples_sf = imult1616( slot_size, hSpatParamRendCom->subframe_nbslots[subframe_idx] );
@@ -838,7 +838,6 @@ void ivas_dirac_dec_binaural_render_fx( 
        /* update combined orientation access index */

        ivas_combined_orientation_update_index( st_ivas->hCombinedOrientationData, n_samples_sf );

    }

    pop_wmops();



    FOR( ch = 0; ch < nchan_out; ch++ )

    {
@@ -1031,7 +1030,6 @@ static void ivas_dirac_dec_binaural_internal_fx( 
        }

    }

    /* CLDFB Analysis of input */

    push_wmops( "IDDBI: CLDFB Analysis of input" );

    FOR( slot = 0; slot < hSpatParamRendCom->subframe_nbslots[subframe]; slot++ )

    {

        FOR( ch = 0; ch < numInChannels; ch++ )
@@ -1157,12 +1155,10 @@ static void ivas_dirac_dec_binaural_internal_fx( 
            }

        }

    }

    pop_wmops();



    test();

    IF( EQ_32( config_data.ivas_format, SBA_FORMAT ) || EQ_32( config_data.ivas_format, SBA_ISM_FORMAT ) )

    {

        push_wmops( "IDDBI: ivas_sba_prototype_renderer_fx" );

        hDiracDecBin->hDiffuseDist = &diffuseDistData;

        ivas_spar_param_to_masa_param_mapping_fx( st_ivas, Cldfb_RealBuffer_in_fx, Cldfb_ImagBuffer_in_fx, q_cldfb, subframe );
@@ -1178,7 +1174,6 @@ static void ivas_dirac_dec_binaural_internal_fx( 
        {

            Scale_sig32( st_ivas->hSpar->hFbMixer->pFb->fb_bin_to_band.pp_cldfb_weights_per_spar_band_fx[idx], IVAS_MAX_NUM_FB_BANDS, Q22 - Q31 ); /*Q31 to Q22*/

        }

        pop_wmops();

    }



    Word16 q_inp = Q6;
@@ -1220,10 +1215,8 @@ static void ivas_dirac_dec_binaural_internal_fx( 
    }



    test();

    push_wmops( "IDDBI: ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matrices_fx" );

    ivas_dirac_dec_binaural_formulate_input_and_target_covariance_matrices_fx( hDiracDecBin, hSpatParamRendCom, &config_data, Cldfb_RealBuffer_in_fx, Cldfb_ImagBuffer_in_fx, Rmat_fx, subframe,

                                                                               hCombinedOrientationData && hCombinedOrientationData->enableCombinedOrientation[hCombinedOrientationData->subframe_idx] > 0, st_ivas->hMasaIsmData, q_inp );

    pop_wmops();

    IF( EQ_32( config_data.ivas_format, ISM_FORMAT ) )

    {

        max_band_decorr = 0;
@@ -1260,10 +1253,8 @@ static void ivas_dirac_dec_binaural_internal_fx( 
        move16();

    }



    push_wmops( "IDDBI: IDDB_detProcMat_fx" );

    ivas_dirac_dec_binaural_determine_processing_matrices_fx( hDiracDecBin, hSpatParamRendCom, &config_data, max_band_decorr, Rmat_fx, subframe,

                                                              hCombinedOrientationData && hCombinedOrientationData->enableCombinedOrientation[hCombinedOrientationData->subframe_idx] > 0, nchanSeparateChannels, st_ivas->hMasaIsmData );

    pop_wmops();



    q_inp = Q6;

    move16();
@@ -1309,9 +1300,7 @@ static void ivas_dirac_dec_binaural_internal_fx( 
    hDiracDecBin->q_processMtxDecPrev = q_mat;

    move16();



    push_wmops( "IDDBI: ivas_dirac_dec_binaural_process_output_fx" );

    ivas_dirac_dec_binaural_process_output_fx( hDiracDecBin, hSpatParamRendCom, st_ivas->cldfbSynDec, output_fx, &q_out, Cldfb_RealBuffer_in_fx, Cldfb_ImagBuffer_in_fx, q_inp, max_band_decorr, numInChannels, config_data.processReverb, subframe, q_mat );

    pop_wmops();



    hDiracDecBin->hDiffuseDist = NULL;
@@ -2150,7 +2139,6 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
        move16();

    }



    push_wmops( "IDDB_detProcMat_bigLoop1" );

    FOR( bin = 0; bin < nBins; bin++ )

    {

        Word32 resultMtxRe_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], resultMtxIm_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS], gain_fx;
@@ -2311,7 +2299,6 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
         * signal energy to synthesize a signal with the target covariance matrix from the non-decorrelated signals */

        IF( LT_16( bin, max_band_decorr ) )

        {

            push_wmops( "IDDB_detProcMat_bigLoop1_bigBranch1" );

            Word32 decorrelationReductionFactor_fx;

            Word16 q_decorrelationReductionFactor;
@@ -2414,7 +2401,6 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
                                         Q31, CrEneL_fx, CrEneR_fx, q_CrEne,

                                         CrCrossRe_fx, CrCrossIm_fx, q_CrCross,

                                         prototypeMtx_fx, MdecRe_fx, MdecIm_fx, &q_Mdec, 3277 ); // 3277 = 0.2 in Q14

            pop_wmops();                                                                         /*IDDB_detProcMat_bigLoop1_bigBranch1*/

        }

        ELSE

        {
@@ -2565,7 +2551,6 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
            move16();

            UWord8 instantChange = 0;

            move16();

            push_wmops( "IDDB_detProcMat_bigLoop1_bigBranch2" );



            exp = sub( 31, hDiracDecBin->q_earlyPartEneCorrection );

            tmp1 = Sqrt32( hDiracDecBin->earlyPartEneCorrection_fx[bin], &exp );
@@ -2641,10 +2626,8 @@ static void ivas_dirac_dec_binaural_determine_processing_matrices_fx( 
                    move16();

                }

            }

            pop_wmops(); /*IDDB_detProcMat_bigLoop1_bigBranch2*/

       }

    }

    pop_wmops(); /*IDDB_detProcMat_bigLoop1*/



    /* Aligning Q-factors of all bins in the processing matrices to a common Q-factor */

    minimum_s( q_processMtx, nBins, &hDiracDecBin->q_processMtx );
@@ -3645,13 +3628,14 @@ static void eig2x2_fx( 
            tmp2 = BASOP_Util_Add_Mant32Exp( crossSquare_fx, sub( 31, q_crossSquare ), tmp2, sub( 31, q_tmp2 ), &q_tmp2 );

            q_tmp2 = sub( 31, q_tmp2 );



#if defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT ) && defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_0 )

            tmp3 = BASOP_Util_Add_Mant32Exp( tmp2, sub( 31, q_tmp2 ), epsilon_mant, epsilon_exp, &exp_tmp3 );



#if defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT ) && defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_0 )

            exp = exp_tmp3;

            normVal_fx = ISqrt32( tmp3, &exp );

            q_tmp2 = sub( 31, exp );

#else

            tmp3 = BASOP_Util_Add_Mant32Exp( tmp2, sub( 31, q_tmp2 ), epsilon_mant, epsilon_exp, &exp_tmp3 );

            tmp2 = BASOP_Util_Divide3232_Scale_cadence( ONE_IN_Q30, tmp3, &exp );

            exp = sub( exp, sub( Q30, sub( 31, exp_tmp3 ) ) );

            normVal_fx = Sqrt32( tmp2, &exp ); // q_tmp2
@@ -5026,9 +5010,6 @@ static void formulate2x2MixingMatrix_fx( 
    Word16 q_Pre[BINAURAL_CHANNELS][BINAURAL_CHANNELS], q_Pim[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

    Word16 hdrm_re[BINAURAL_CHANNELS][BINAURAL_CHANNELS], hdrm_im[BINAURAL_CHANNELS][BINAURAL_CHANNELS];



    push_wmops( "IDDB_detProcMat_bigLoop1_2x2Matrix" );

    push_wmops( "IDDB_2x2Matrix_Part1" );



#ifdef FIX1072_SPEEDUP_formulate2x2MixingMatrix_fx_simple

#else

    set16_fx( hdrm_re[0], 63, i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS ) );
@@ -5139,18 +5120,12 @@ static void formulate2x2MixingMatrix_fx( 
    Cout_im = Mpy_32_32( Cout_im, maxEneDiv_fx );

    q_cout = sub( add( q_cout, q_maxEneDiv ), 31 );



    pop_wmops(); //( "IDDB_2x2Matrix_Part1" );

    push_wmops( "IDDB_2x2Matrix_Part2" );



    push_wmops( "IDDB_2x2Matrix_Part2.1" );

    push_wmops( "IDDB_2x2Matrix_Part2.1.1" );

    /* Cholesky decomposition of target / output covariance matrix */

    chol2x2_fx( E_out1, E_out2, q_eout, Cout_re, Cout_im, q_cout, KyRe_fx, KyIm_fx, &q_ky );



    /* Eigendecomposition of input covariance matrix */

    eig2x2_fx( E_in1, E_in2, q_ein, Cin_re, Cin_im, q_cin, Uxre_fx, Uxim_fx, &q_Ux, Sx_fx, &q_Sx );

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.1.1" );

    push_wmops( "IDDB_2x2Matrix_Part2.1.2" );



    /* Eigendecomposition to Kx -- Ux Sx Ux' -> Kx Kx'*/

    exp = sub( 31, q_Sx );

    exp1 = sub( 31, q_Sx );
@@ -5164,8 +5139,6 @@ static void formulate2x2MixingMatrix_fx( 
    Sx_fx[1] = L_shr( Sx_fx[1], sub( sub( 31, exp1 ), q_Sx ) ); // q_Sx

    move32();



    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.1.2" );

    push_wmops( "IDDB_2x2Matrix_Part2.1.3" );

    matrixDiagMul_fx( Uxre_fx, Uxim_fx, q_Ux, Sx_fx, q_Sx, Kxre_fx, Kxim_fx, &q_Kx );



    /* Regularize the diagonal Sx for matrix inversion */
@@ -5177,9 +5150,6 @@ static void formulate2x2MixingMatrix_fx( 


    temp = Mpy_32_32( E_in2, 2147484 ); // 2147484 = 0.001f in Q31

    temp = L_max( temp, E_in1 );

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.1.3" );

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.1" );

    push_wmops( "IDDB_2x2Matrix_Part2.2" );

#if defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT ) && defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_2 )

    /*IF (E_out1 == 0)*/

    {
@@ -5303,8 +5273,6 @@ static void formulate2x2MixingMatrix_fx( 
    move32();

    Ghat_fx[1] = L_shr( Ghat_fx[1], sub( sub( 31, exp1 ), q_Ghat ) ); // q_Ghat

    move32();

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.2" );

    push_wmops( "IDDB_2x2Matrix_Part2.3" );



    /* Matrix multiplication, tmp = Ky' * G_hat * Q */

    FOR( chA = 0; chA < BINAURAL_CHANNELS; chA++ )
@@ -5335,9 +5303,6 @@ static void formulate2x2MixingMatrix_fx( 


    q_temp = sub( add( q_ky, q_GhatQ ), 31 );



    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.3" );

    push_wmops( "IDDB_2x2Matrix_Part2.4" );

    push_wmops( "IDDB_2x2Matrix_Part2.4.1" );

    /* A = Ky' * G_hat * Q * Kx (see publication) */

    matrixMul_fx( tmpRe_fx, tmpIm_fx, &q_temp, Kxre_fx, Kxim_fx, &q_Kx, Are_fx, Aim_fx, &q_A );
@@ -5351,8 +5316,6 @@ static void formulate2x2MixingMatrix_fx( 


    eig2x2_fx( tmpRe_fx[0][0], tmpRe_fx[1][1], q_temp, tmpRe_fx[1][0], tmpIm_fx[1][0], q_temp, Ure_fx, Uim_fx, &q_U, D_fx, &q_D );



    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.4.1" );

    push_wmops( "IDDB_2x2Matrix_Part2.4.2" );

#if defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT ) && defined( FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT_4 )

    IF( D_fx[0] == 0 )

    {
@@ -5426,9 +5389,7 @@ static void formulate2x2MixingMatrix_fx( 
    div_fx[1] = Sqrt32( temp, &exp1 ); // Q = 31 - exp1

    move32();

#endif           /*FIX_1072_REPLACE_DIV_SQRT_BY_ISQRT*/

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.4.2" );

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.4" );

    push_wmops( "IDDB_2x2Matrix_Part2.5" );



    q_div = sub( 31, s_max( exp, exp1 ) );



    div_fx[0] = L_shr( div_fx[0], sub( sub( 31, exp ), q_div ) ); // q_div
@@ -5460,8 +5421,6 @@ static void formulate2x2MixingMatrix_fx( 
    scale_sig32( div_fx, BINAURAL_CHANNELS, exp );

    q_div = add( q_div, exp );



    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.5" );

    push_wmops( "IDDB_2x2Matrix_Part2.6" );

    FOR( chA = 0; chA < BINAURAL_CHANNELS; chA++ )

    {

        FOR( chB = 0; chB < BINAURAL_CHANNELS; chB++ )
@@ -5531,8 +5490,6 @@ static void formulate2x2MixingMatrix_fx( 
            }

        }

    }

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part2.6" );

    push_wmops( "IDDB_2x2Matrix_Part2.7" );

    minimum_s( hdrm_re[0], i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS ), &exp );

    q_temp = exp;

    move16();
@@ -5556,10 +5513,7 @@ static void formulate2x2MixingMatrix_fx( 
                         0 /*int Bscale*/,

#endif

                         Pre_fx, Pim_fx, &q_P ); /* Nearest orthonormal matrix P to matrix A formulated */

    pop_wmops();                                 // push_wmops( "IDDB_2x2Matrix_Part2.7" );

    pop_wmops();                                 //( "IDDB_2x2Matrix_Part2" );

    push_wmops( "IDDB_2x2Matrix_Part3" );

    push_wmops( "IDDB_2x2Matrix_Part3.1" );



    /* These are the final formulas of the JAES publication M = Ky P Kx^(-1) */

    FOR( chA = 0; chA < BINAURAL_CHANNELS; chA++ )

    {
@@ -5617,8 +5571,6 @@ static void formulate2x2MixingMatrix_fx( 
#endif

        }

    }

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part3.1" );

    push_wmops( "IDDB_2x2Matrix_Part3.2" );

    minimum_s( q_Pre[0], i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS ), &exp );

    q_P = s_min( q_P, exp );

    minimum_s( q_Pim[0], i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS ), &exp );
@@ -5634,8 +5586,6 @@ static void formulate2x2MixingMatrix_fx( 
            move32();

        }

    }

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part3.2" );

    push_wmops( "IDDB_2x2Matrix_Part3.3" );

#if defined( FIX1072_SPEEDUP_combinedMatrixCalculations )

    combinedMatrixCalculations( KyRe_fx, KyIm_fx, &q_ky, 0 /*scale*/, Pre_fx, Pim_fx, &q_P, 0 /*scale*/, Uxre_fx, Uxim_fx, &q_Ux, 0 /*scale*/, Mre_fx, Mim_fx, q_M );

#else
@@ -5648,9 +5598,6 @@ static void formulate2x2MixingMatrix_fx( 
#endif

                         Mre_fx, Mim_fx, q_M );

#endif           /*FIX1072_SPEEDUP_combinedMatrixCalculations*/

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part3.3" );

    pop_wmops(); // push_wmops( "IDDB_2x2Matrix_Part3" );

    pop_wmops(); /*IDDB_detProcMat_bigLoop1_2x2Matrix*/

    return;

}