Commit 40a0b479 authored by Arthur Tritthart's avatar Arthur Tritthart
Browse files

tuned function matrix_product_mant_exp_fx (regular case) in WMOPS (about... 

tuned function matrix_product_mant_exp_fx (regular case) in WMOPS (about -36.2), all bit-exact to previous version
parent e499d56a

lib_com/ivas_tools.c

+38 −0
Original line number Diff line number Diff line
@@ -1040,6 +1040,7 @@ Word16 matrix_product_mant_exp_fx( 
    }

    ELSE /* Regular case */

    {

#ifndef FIX_TUNE_IVAS_TOOLS_MATRIX_PRODUCT_WMOPS_MR1010X

        IF( NE_16( colsX, rowsY ) )

        {

            return EXIT_FAILURE;
@@ -1068,6 +1069,43 @@ Word16 matrix_product_mant_exp_fx( 
        move16();

        col = rowsX; /*Q0*/

        move16();

#else

        assert( colsX == rowsY );

        assert ( colsX >= 1 );



        Word16 XY_fx_e = add( X_fx_e, Y_fx_e );



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

        {

            Word16 j_rowsY = imult1616( j, rowsY );

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

            {

                Word32 Z_fx_m;

                Word16 Z_fx_e;



                x_idx = add( i, 0 );

                y_idx = add( 0, j_rowsY );

                Z_fx_m = Mpy_32_32( X_fx[x_idx], Y_fx[y_idx]);

                Z_fx_e = XY_fx_e;

                FOR( k = 1; k < colsX; ++k )

                {

                    x_idx = add( x_idx, rowsX );

                    y_idx = add( 1, y_idx );

                    Z_fx_m = BASOP_Util_Add_Mant32Exp( Z_fx_m, Z_fx_e, Mpy_32_32( X_fx[x_idx], Y_fx[y_idx] ), XY_fx_e, &Z_fx_e ); /*Q31 - Zp_fx_e*/

                }

                *Zp_fx++ = Z_fx_m;

                move32();

                *Zp_fx_e++ = Z_fx_e;

                move16();

            }

        }





        row = colsY; /*Q0*/

        move16();

        col = rowsX; /*Q0*/

        move16();

#endif

    }

    Zp_fx = Z_fx; /*Q31 - Zp_fx_e*/

    Zp_fx_e = out_e;

lib_com/options.h

+2 −0
Original line number Diff line number Diff line
@@ -153,3 +153,5 @@ 
#define FIX_881_HILBERT_FILTER                  /* VA: improve the precision of the Hilbert filter to remove 2kHz unwanted tone */

#endif

#define FIX_MINOR_SVD_WMOPS_MR1010X             /* FhG: Minor WMOPS tuning, bit-exact to previous version, saves about 8.2 WMOPS for MR1010 */

#define FIX_TUNE_IVAS_TOOLS_MATRIX_PRODUCT_WMOPS_MR1010X /* FhG: Major WMOPS tuning in HQ matrix multiplication, bit-exact to previous version, saves about 37 WMOPS for MR1010 */