Duly noted. This has bothered me, too, because a lot of the 64-bit
floating-point functions would be both simpler and faster if computed
using 80 bits.
Any 80-bit or 128-bit ops that run on hardware would either have to be
Racket primitives or be exposed by an FFI. I couldn't do the former, but
possibly the latter.
For those precisions and larger, two years ago, I wrote an FFI wrapper
for MPFR, an LGPL'd library for arbitrary-precision floating-point math
with correctly rounded results. (The authors prove and publish the
proofs of their algorithms' correctness.) Back then, it would have made
more sense as a Planet package. With a `math' collection, it might be
better as `math/bigfloat' - and it would also give me a good standard to
test the 64-bit floating-point functions against.
Neil ⊥
On 07/04/2012 01:02 AM, Dmitry Pavlov wrote:
Neil,
I kept in mind to initiate a topic about it, but
since you asked first (thanks!), I will say it now:
*** I want 80 bit IEEE floating point numbers in Racket. ***
Double precision (64 bit) is not always enough for scientific
calculations. Extended precision is demanded.
Is it possible to make something like "flonum"-s, bul 80-bit?
More than that, in the future we will surely need 128-bit
(quadruple precision) floating point numbers. Intel
and GCC compilers seem to support quadruple precision
already. They do it via emulation, but still it would be
interesting to try that in Racket, too. I believe we
will see real processors with 128-bit math someday.
Best regards,
Dmitry
____________________
Racket Users list:
http://lists.racket-lang.org/users
