On 15 May, 15:48, "RB" <NoMail@NoSpam> wrote:
> I read somewhere awhile back (unless I am confused) that the
> modern day math coprocessors did operations on 80bit
> floating point values. If the input value was a smaller bit length
> it was converted up and then converted back down after.
> Is this true ?
> And if so I would surmise that it is a trivial performance factor of
> conversion so that using a float as oppose to a double ( if the
> precision offered by a float was sufficient ) is advantageous only
> in the amount of storage it requires ?
For x86, yes, you are right. IIRC the maths coprocessor *can* be
changed to use floating point numbers less precise than 80-bit and
that's sometimes worth doing - such as for fast divides.
Check out instrucion timing tables documented by Agner Fog for lots of
details.
James
