.module/ram do_DIF16;        {*** 1 16-way DIF ***}
.external tmp;
.entry DIF16;

DIF16:
        i0=^tmp;
        i1=^tmp+8;
        m1=1;

        ax1=dm(i3,m3);

        af=pass ax1, ax0=dm(i2,m2);
        ar=ax0+af, ax1=dm(i3,m3), my0=pm(i6,m6);
        ar=ax0-af, dm(i0,m1)=ar;
        mr=ar*my0(ss);

        cntr=6;
        do x16 until ce;
            af=pass ax1, ax0=dm(i2,m2);
            ar=ax0+af, ax1=dm(i3,m3), my0=pm(i6,m6);
            ar=ax0-af, dm(i0,m1)=ar;
x16:        mr=ar*my0(ss), dm(i1,m1)=mr1;

        af=pass ax1, ax0=dm(i2,m2);
        ar=ax0+af,                my0=pm(i6,m6);
        ar=ax0-af, dm(i0,m1)=ar;
        mr=ar*my0(ss), dm(i1,m1)=mr1;

        dm(i1,m1)=mr1;
        rts;                                       {*** end 1 16-way DIF ***}
.endmod;