Skip to content
Prev 27619 / 398500 Next

floating point question

On Fri, 31 Jan 2003, Thomas Lumley wrote:

            
(Sparc FPUs have some guard digits, as far as I recall, and the fine 
details do vary by hardware, or at least they used to and we still have 
some of the machines that varied running.)
I used to think that!  But when it became possible to build R for Windows
not chopping in-FPU calculations to 53 bits, the results were more
consistent, and we changed over.  At that point most of the
inconsistencies were in printing, hence my remark earlier.  The print
routines in Windows `libc' (really msvcrt.dll) does lose a digit or two
quite often, quite possibly to avoid giving meaningless ones (which glibc
is happy to if asked).