A Thought on the Discussion of DEC Alphas
d.love at dl.ac.uk
Mon Mar 14 13:10:37 EST 1994
>>>>> "Phil" == Phil Bourne <system at CUHHCA.HHMI.COLUMBIA.EDU> writes:
Phil> In response to an article I wrote some time ago on items to consider
Phil> in optimizing code for use on Alpha's I received this note re the
Phil> use of IEEE arithmetic. Given the current discussion I thought I
Phil> would pass it on.. if you would like code to test this problem
Phil> please let me know..
Phil> /Phil Bourne
Phil> ============ Note from Layne ==================
Phil> A real service to the DEC user community would be to point out
Phil> that the AXP does NOT support fully the IEEE 754 Standard for
Phil> floating point arithmetic. In particular, programs using IEEE
Phil> arithmetic on a DECstation, when ported to an Alpha (either via
Phil> mx or via recompiling), do NOT produce the same results as on
Phil> the DECstation! If you want, I can send you a program which
Phil> illustrates the different results on the MIPS RISC and the AXP.
If this is a worry, it's worth pointing out that Un*x boxes sometimes
(often?) don't implement IEEE 754 properly by default. One reason is
that they may have to implement `gradual rounding' by traps
(i.e. slowly). Before worrying about it, it would be interesting to
know what documentation claims in each case.
More information about the Xtal-log