                +---------------------------------------------------+
                |       What does Convergent rounding mean?         |
                |       Is it without bias?                         |
                +---------------------------------------------------+

                The DSP56000 family provides "round to nearest even", 
                or RN, rounding, which is performed during MPYR, MACR
                and RND instructions to round the double precision 
                48-bit accumulator results into a single precision 
                24-bit result.  This type of rounding is also called 
                "convergent" rounding.

                The DSP56000 rounding algorithm:
                Where:
                A is the double precision accumulator (could be A or B).
                A1 is the most significant portion.  
                A0 is the least significant portion.

                First:
                        If A0 is less than 1/2 the ls bit of A1
                        (A0 < $800000) then:
                                Add nothing to A1.
                                (i.e. round down)

                        If A0 is greater than 1/2 the ls bit of A1
                        (A0 > $800000) then:
                                Add one at the ls bit position of A1.
                                (i.e. round up)

                        If A0 is exactly equal to 1/2 the ls bit of A1
                        (A0 = $800000) then:
                                If the ls bit of A1 is a "1" then: 
                                        Add one at the ls bit position of A1. 
                                        (round up)
                                ELSE (ls bit of A1 = "0")
                                        Add nothing to A1.
                                        (round down)
                Then:
                        Clear A0.


                DSP56000 rounding, RN rounding, is without bias.  
                
                Most other rounding implementations have a slight 
                positive bias because they always round up the case
                of A0 = 1/2 ($800000).

                One last thing.  The scaling mode bits, S1 and S0 
                affect the rounding position as described in
                Chapter 4 of the DSP56000 User's Manual (sections
                4.8.9 and 4.15).
