On 2010-Aug-21 05:26:36 +0100, "Dr. David Kirkby" <david.kir...@onetel.net> wrote: >it is clear that the quad double algorithm assumes that the floating point >processor rounds to 64-bits, which things like PowerPC and SPARC do. > >But Intel and AMD CPUs round to 80 bits by default. As such, on Intel/AMD >CPUs, >the quad double package does some low level stuff with the control register on >the FPU to make it round to 64 bits rather than the default 80. ... >If you build Sage on bsd.math, it does report > >"You do not appear to have an x86 based system --- not using fpu.c"
On FreeBSD, the x87 FPU is configured to round to 53/64-bits by default. This prevents FP results from varying depending on how temporaries are spilled to memory. Traditionally, Apple has been very conscious of FP math quality (the Macintosh SANE library was one of the first good microprocessor maths libraries). Since OS-X draws heavily on the FreeBSD kernel, it's possible it inherited the the "round to double" FPU initialisation. I'd suggest you try running paranoia (or similar package) on bsd.math and see what it reports. -- Peter Jeremy
pgpWuaybUZLMY.pgp
Description: PGP signature
