[Theora-dev] Re: MMX/mmxext optimisations

[James Cloos]

>>>>> "Jim" == Jim Darby <jim at jimbocorp.uklinux.net> writes:

Jim> This is a little worrying. Surely the MMX code should just make
Jim> things happen faster rather than different?

Perhaps it is due to mmx/sse floating point doing true 64bit doubles
and 32bit floats whereas x87 does everything in 80bit and rounds the
results.  Depending on the functions the mmx could lose several bits
of accuracy.  

A good comparison would be a chip like alpha, sparc, ppc, mips et al
where (AIUI) all fp is true 32 or 64 bit.  In short, does the result
change when encoded on x86 (using the portable C) vs anything else?

-JimC
-- 
James H. Cloos, Jr. <cloos at jhcloos.com>