Commit 40b6da87 authored by thomas dettbarn's avatar thomas dettbarn Committed by Sandesh Venkatesh
Browse files

inling the matrix multiplication in... 

inling the matrix multiplication in ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx().
parent f48701fc

lib_com/options.h

+1 −0
Original line number Diff line number Diff line
@@ -86,6 +86,7 @@ 
//#define HARM_SCE_INIT

#define DIV32_OPT_NEWTON                               /* FhG: faster 32 by 32 bit division */ 

#define	MERGE_REQUEST_1378_SPEEDUP_ivas_mc_param_enc_fx_NONBE /* FhG: reduce WMOPS of Cy calculation in ivas_param_mc_param_est_enc_fx() by using 64 Bit addition. Obsoletes IMPROVE_HIGH_COMPLEXITY_PARAM_MC_PRM_EST_NONBE. */

#define	MERGE_REQUEST_1564_SPEEDUP_ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx_NONBE /* FhG: reduce WMOPS by inlining the matrix multiplications for the smoothing operation. */

#define FIX_1439_SPEEDUP_Copy_Scale_sig_16_32_no_sat            /*FhG: reduces WMOPS - bit-exact*/

#define FIX_1439_SPEEDUP_stereo_icBWE_dec_fx                    /*FhG: reduces WMOPS - bit-exact*/

#define FIX_1439_SPEEDUP_ivas_swb_tbe_dec_fx                    /*FhG: reduces WMOPS - bit-exact*/

lib_dec/ivas_dirac_output_synthesis_cov_fx.c

+70 −0
Original line number Diff line number Diff line
@@ -30,6 +30,8 @@ 


*******************************************************************************************************/



#define	MERGE_REQUEST_1564_SPEEDUP_ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx_NONBE



#include <stdint.h>

#include <string.h>

#include <stdio.h>
@@ -516,10 +518,14 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx( 
    Word16 mixing_matrix_buffer_e;

    Word32 input_f_real_fx[PARAM_MC_MAX_TRANSPORT_CHANS];

    Word32 input_f_imag_fx[PARAM_MC_MAX_TRANSPORT_CHANS];

#ifdef	MERGE_REQUEST_1564_SPEEDUP_ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx_NONBE



#else

    Word32 output_f_real_fx[MAX_CICP_CHANNELS];

    Word32 output_f_imag_fx[MAX_CICP_CHANNELS];

    Word16 output_f_real_e;

    Word16 output_f_imag_e;

#endif

    Word32 diff_f_real_fx[MAX_CICP_CHANNELS];

    Word32 diff_f_imag_fx[MAX_CICP_CHANNELS];
@@ -527,8 +533,12 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx( 


    set_zero_fx( input_f_real_fx, PARAM_MC_MAX_TRANSPORT_CHANS );

    set_zero_fx( input_f_imag_fx, PARAM_MC_MAX_TRANSPORT_CHANS );

#ifdef	MERGE_REQUEST_1564_SPEEDUP_ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx_NONBE



#else

    set_zero_fx( output_f_real_fx, MAX_CICP_CHANNELS );

    set_zero_fx( output_f_imag_fx, MAX_CICP_CHANNELS );

#endif

    set_zero_fx( diff_f_real_fx, MAX_CICP_CHANNELS );

    set_zero_fx( diff_f_imag_fx, MAX_CICP_CHANNELS );
@@ -592,6 +602,36 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx( 
                }



                /* apply residual mixing */

#ifdef MERGE_REQUEST_1564_SPEEDUP_ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx_NONBE

		{	

			Word16 shifter;



			shifter=31-mixing_matrix_res_smooth_e;

			FOR( ch_idx = 0; ch_idx < nY; ch_idx++ )

			{

				int i;

				Word16 idx;

				Word64 temp_real, temp_imag;





				idx = ch_idx;

				temp_real = 0;

				temp_imag = 0;

				move64();

				move64();

				for (i=0;i<nY;i++)

				{

					temp_real = W_add( temp_real, W_mult0_32_32( mixing_matrix_res_smooth_fx[idx], diff_f_real_fx[i] ) ); 

					temp_imag = W_add( temp_imag, W_mult0_32_32( mixing_matrix_res_smooth_fx[idx], diff_f_imag_fx[i] ) ); 

					idx = add(idx, nY );

				}

				Cldfb_RealBuffer_fx[ch_idx][slot_idx_sfr][band] = W_extract_l( W_shr( temp_real, shifter ) );

				Cldfb_ImagBuffer_fx[ch_idx][slot_idx_sfr][band] = W_extract_l( W_shr( temp_imag, shifter ) );

			}

		}





#else



                matrix_product_mant_exp_fx( mixing_matrix_res_smooth_fx, mixing_matrix_res_smooth_e, nY, nY, 0, diff_f_real_fx, 25, nY, 1, 0, output_f_real_fx, &output_f_real_e );

                scale_sig32( output_f_real_fx, nY, sub( Q6, sub( Q31, output_f_real_e ) ) ); // Q6
@@ -607,6 +647,7 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx( 
                    Cldfb_ImagBuffer_fx[ch_idx][slot_idx_sfr][band] = output_f_imag_fx[ch_idx]; // Q6

                    move32();

                }

#endif

            }

            ELSE

            {
@@ -630,6 +671,34 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx( 
            }



            /* apply mixing matrix */

#ifdef MERGE_REQUEST_1564_SPEEDUP_ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx_NONBE

	{

            Word16 shifter;

            shifter=31-mixing_matrix_smooth_e;



	    FOR( ch_idx = 0; ch_idx < nY; ch_idx++ )

	    {

                int i;

                Word16 idx;

                Word64 temp_real, temp_imag;





                idx = ch_idx;

                temp_real = 0;

                temp_imag = 0;

                move64();

                move64();

                for (i=0;i<nX;i++)

                {

                    temp_real = W_add( temp_real, W_mult0_32_32( mixing_matrix_smooth_fx[idx], input_f_real_fx[i] ) ); 

                    temp_imag = W_add( temp_imag, W_mult0_32_32( mixing_matrix_smooth_fx[idx], input_f_imag_fx[i] ) ); 

                    idx = add(idx, nY );

                }

                Cldfb_RealBuffer_fx[ch_idx][slot_idx_sfr][band] = L_add( Cldfb_RealBuffer_fx[ch_idx][slot_idx_sfr][band], W_extract_l( W_shr( temp_real, shifter ) ) );

                Cldfb_ImagBuffer_fx[ch_idx][slot_idx_sfr][band] = L_add( Cldfb_ImagBuffer_fx[ch_idx][slot_idx_sfr][band], W_extract_l( W_shr( temp_imag, shifter ) ) );

	    }

	}

#else



            matrix_product_mant_exp_fx( mixing_matrix_smooth_fx, mixing_matrix_smooth_e, nY, nX, 0, input_f_real_fx, 25, nX, 1, 0, output_f_real_fx, &output_f_real_e );

            scale_sig32( output_f_real_fx, MAX_CICP_CHANNELS, sub( 6, sub( 31, output_f_real_e ) ) ); // Q6
@@ -646,6 +715,7 @@ void ivas_dirac_dec_output_synthesis_cov_param_mc_synthesise_slot_fx( 
                Cldfb_ImagBuffer_fx[ch_idx][slot_idx_sfr][band] = L_add( Cldfb_ImagBuffer_fx[ch_idx][slot_idx_sfr][band], output_f_imag_fx[ch_idx] );

                move32();

            }

#endif

        }

    }