advantages of fused multiply-add, and a thesis project

[uunet!cwi.nl!dik]

 > If I recall correctly, the MC68881 has a Rounding-Precision state, which 
 > can cause all rounding to be done to 32-bit format, or 64-bit format, 
 > despite the internal 96-bit registers (80 mantissa, 16 exponent?).
Right, but internal size is 80 bits; 64 mantissa, 15 exponent, 1 sign.
 >                                                                     Some 
 > other hardware also offers this as a "mode" option.
Amongst others the 80x87.
The 68040 (which incorporates floating-point) allows this also as an encoding
in the instruction.
 >                                                      When utilized, these 
 > modes seem to cause a very noticeable performance degradation; so I would 
 > suppose they exist for precisely the same purpose you are invoking here?
On the 68881 there is indeed a performance degradation.  I do not know about
the others.  The 80x87 manuals state that it is required by IEEE; the
68881/68882 manual states that it is to emulate standard IEEE.

dik
--
dik t. winter, cwi, amsterdam, nederland
dikacwi.nl