On Tuesday, 31 July 2012 at 01:00:38 UTC, Nick Sabalausky wrote:
My (limited) understanding is that it's almost practically
impossible to get consistency in x86 floating point unless
you're using SSE:
http://www.yosefk.com/blog/consistency-how-to-defeat-the-purpose-of-ieee-floating-point.html
Maybe the effect you're observing could be related to what he
describes?
I don't suppose there's an emulated floating type (like is
optionally available in the linux kernel) that could be used in
place of normal floating point? I wonder if the C++ version is
wrong; In walter's article regarding floating point he brought
out that few applications in C++ set/reset the register for
floating point between operations or between functions (when they
change)... Or was that between applications as a whole (And an
occasional OS issue)?
Is the problem be even worse if you mixed MMX and FPU (Since
they use the same registers, just a different mode)?
