Commit b94c2a1b authored by Nicolas Roussin's avatar Nicolas Roussin
Browse files

Optimize matrix mul output-format fix.

parent fdfcd002

lib_com/options.h

+1 −0
Original line number Diff line number Diff line
@@ -148,6 +148,7 @@ 


#define OPT_2181_MATRIX_TRANSP_1_MUL                    /* Dolby: Issue 2181, optimize matrixTransp1Mul_fx. */

#define OPT_2182_MATRIX_SCALE_OPS                       /* Dolby: Issue 2181, move matrix scale operations outside mul operations. */

#define OPT_XXXX_MATRIX_OUT_SCALING



/* #################### End BASOP optimization switches ############################ */

lib_rend/ivas_dirac_dec_binaural_functions_fx.c

+76 −30
Original line number Diff line number Diff line
@@ -3986,7 +3986,6 @@ static void matrixMul_fx( 
{

#ifdef OPT_2182_MATRIX_SCALE_OPS

    Word16 chA, chB;

    Word16 size = i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS );

#else

    Word16 chA, chB;

    Word16 min_q_shift1, min_q_shift2;
@@ -4005,6 +4004,9 @@ static void matrixMul_fx( 
    move16();

    move16();

#endif

#ifdef OPT_XXXX_MATRIX_OUT_SCALING

    Word32 not_zero = 0;

#endif



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

    {
@@ -4031,16 +4033,31 @@ static void matrixMul_fx( 
                                             Are_fx[chA][1], Bim_fx[1][chB] );

            move32();

#endif /* #ifdef IVAS_ENH64_CADENCE_CHANGES */

#ifdef OPT_XXXX_MATRIX_OUT_SCALING

            not_zero = L_or( not_zero, outRe_fx[chA][chB] );

            not_zero = L_or( not_zero, outIm_fx[chA][chB] );

#endif

        }

    }

    *q_out = sub( add( *q_A, *q_B ), 31 );



    *q_out = sub( add( *q_A, *q_B ), 31 );

    move16();

#ifdef OPT_XXXX_MATRIX_OUT_SCALING

    if ( !not_zero )

    {

        *q_out = Q31;

        move16();

    }

#else

#ifdef OPT_2182_MATRIX_SCALE_OPS

    Word16 size = i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS );

#endif

    if ( L_and( is_zero_arr( outRe_fx[0], size ), is_zero_arr( outIm_fx[0], size ) ) )

    {

        *q_out = Q31;

        move16();

    }

#endif



    return;

}
@@ -4057,7 +4074,6 @@ static void matrixTransp1Mul_fx( 
    Word16 *q_out )

{

    Word16 chA, chB;

    Word16 size = i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS );



    Word64 tmp_outRe_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

    Word64 tmp_outIm_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];
@@ -4067,11 +4083,19 @@ static void matrixTransp1Mul_fx( 
    q = add( add( q_A, q_B ), 1 );

    common_lsh = sub( 63, q );

    move16();

#else

    Word16 q_tmp_outRe_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

    Word16 q_tmp_outIm_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

    Word64 tmp64_1, tmp64_2;

    Word16 tmp16, q_common = 63;

    move16();

#endif



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

    {

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

        {

#ifdef OPT_2181_MATRIX_TRANSP_1_MUL

            tmp64 = W_mult_32_32( Are_fx[0][chA], Bre_fx[0][chB] );

            tmp64 = W_mac_32_32( tmp64, Are_fx[1][chA], Bre_fx[1][chB] );

            tmp64 = W_mac_32_32( tmp64, Aim_fx[0][chA], Bim_fx[0][chB] );
@@ -4085,31 +4109,7 @@ static void matrixTransp1Mul_fx( 
            tmp_outIm_fx[chA][chB] = tmp64 = W_mac_32_32( tmp64, Are_fx[1][chA], Bim_fx[1][chB] );

            move64();

            common_lsh = s_min( common_lsh, W_norm( tmp64 ) );

        }

    }

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

    {

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

        {

            outRe_fx[chA][chB] = W_extract_h( W_shl( tmp_outRe_fx[chA][chB], common_lsh ) );

            move32();

            outIm_fx[chA][chB] = W_extract_h( W_shl( tmp_outIm_fx[chA][chB], common_lsh ) );

            move32();

        }

    }

    *q_out = sub( add( q, common_lsh ), 32 );

    move16();

#else

    Word16 q_tmp_outRe_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

    Word16 q_tmp_outIm_fx[BINAURAL_CHANNELS][BINAURAL_CHANNELS];

    Word64 tmp64_1, tmp64_2;

    Word16 tmp16, q_common = 63;

    move16();



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

    {

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

        {

            tmp64_1 = W_mac_32_32( W_mult_32_32( Are_fx[0][chA], Bre_fx[0][chB] ), Are_fx[1][chA], Bre_fx[1][chB] ); // Q: add( add( q_A, q_B ), 1 )

            tmp64_2 = W_mac_32_32( W_mult_32_32( Aim_fx[0][chA], Bim_fx[0][chB] ), Aim_fx[1][chA], Bim_fx[1][chB] ); // Q: add( add( q_A, q_B ), 1 )

            tmp_outRe_fx[chA][chB] = W_add( tmp64_1, tmp64_2 );                                                      // Q: add( add( q_A, q_B ), 1 )
@@ -4132,26 +4132,54 @@ static void matrixTransp1Mul_fx( 
            q_tmp_outIm_fx[chA][chB] = add( tmp16, add( add( q_A, q_B ), 1 ) );

            move16();

            q_common = s_min( q_tmp_outIm_fx[chA][chB], q_common );

#endif

        }

    }

#ifdef OPT_XXXX_MATRIX_OUT_SCALING

    Word32 not_zero = 0;

#endif

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

    {

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

        {

#ifdef OPT_2181_MATRIX_TRANSP_1_MUL

            outRe_fx[chA][chB] = W_extract_h( W_shl( tmp_outRe_fx[chA][chB], common_lsh ) );

            move32();

            outIm_fx[chA][chB] = W_extract_h( W_shl( tmp_outIm_fx[chA][chB], common_lsh ) );

            move32();

#else

            outRe_fx[chA][chB] = W_extract_h( W_shl( tmp_outRe_fx[chA][chB], s_max( -63, sub( q_common, q_tmp_outRe_fx[chA][chB] ) ) ) );

            move32();

            outIm_fx[chA][chB] = W_extract_h( W_shl( tmp_outIm_fx[chA][chB], s_max( -63, sub( q_common, q_tmp_outIm_fx[chA][chB] ) ) ) );

            move32();

#endif

#ifdef OPT_XXXX_MATRIX_OUT_SCALING

            not_zero = L_or( not_zero, outRe_fx[chA][chB] );

            not_zero = L_or( not_zero, outIm_fx[chA][chB] );

#endif

        }

    }

#ifdef OPT_2181_MATRIX_TRANSP_1_MUL

    *q_out = sub( add( q, common_lsh ), 32 );

    move16();

#else

    *q_out = sub( q_common, 32 );

    move16();

#endif

#ifdef OPT_XXXX_MATRIX_OUT_SCALING

    if ( !not_zero )

    {

        *q_out = Q31;

        move16();

    }

#else

    Word16 size = i_mult( BINAURAL_CHANNELS, BINAURAL_CHANNELS );

    if ( L_and( is_zero_arr( outRe_fx[0], size ), is_zero_arr( outIm_fx[0], size ) ) )

    {

        *q_out = Q31;

        move16();

    }

#endif



    return;

}
@@ -4184,7 +4212,7 @@ static void matrixTransp2Mul_fx( 
{

#ifdef OPT_2182_MATRIX_SCALE_OPS

    Word16 chA, chB;

    Word16 size = BINAURAL_CHANNELS * BINAURAL_CHANNELS;

    // Word16 size = BINAURAL_CHANNELS * BINAURAL_CHANNELS;

#else

    Word16 chA, chB;

    Word16 min_q_shift;
@@ -4208,7 +4236,9 @@ static void matrixTransp2Mul_fx( 
        move16();

    }

#endif



#ifdef OPT_XXXX_MATRIX_OUT_SCALING

    Word32 not_zero = 0;

#endif

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

    {

        FOR( chB = 0; chB < BINAURAL_CHANNELS; chB++ )
@@ -4234,16 +4264,32 @@ static void matrixTransp2Mul_fx( 
                                             Are_fx[chA][1], Bim_fx[chB][1] );

            move32();

#endif /* #ifdef IVAS_ENH64_CADENCE_CHANGES */

#ifdef OPT_XXXX_MATRIX_OUT_SCALING

            not_zero = L_or( not_zero, outRe_fx[chA][chB] );

            not_zero = L_or( not_zero, outIm_fx[chA][chB] );

#endif

        }

    }



    *q_out = sub( add( *q_A, *q_B ), 31 );

    move16();



#ifdef OPT_XXXX_MATRIX_OUT_SCALING

    if ( !not_zero )

    {

        *q_out = Q31;

        move16();

    }

#else

#ifdef OPT_2182_MATRIX_SCALE_OPS

    Word16 size = BINAURAL_CHANNELS * BINAURAL_CHANNELS;

#endif

    if ( L_and( is_zero_arr( outRe_fx[0], size ), is_zero_arr( outIm_fx[0], size ) ) )

    {

        *q_out = Q31;

        move16();

    }

#endif



    return;

}