Re: gnuplot numerical errors on Intel MacBook

[Daniel J S. <dan...@ie...>]

Joe Koski wrote:

> My total knowledge of this subject consists of owning a copy of Abramowitz
> and Stegun, so please excuse me it the following is a red herring.
> 
> On his daily blog on the g95 website, g95.sourceforge.net, Andy Vaught wrote
> =====================
> September 6
> Bill McKie pointed out that x86/OSX wasn't trapping overflows during real
> exponentiation. It turns out that OSX uses the MMX unit to do exponentiation
> instead of the x87. So I added code to set the bits for MMX traps. Not only
> does the MMX exponentiation use about six times the code that the x87 uses,
> but it turns out that OSX doesn't handle traps from the MMX unit correctly
> either, reporting only a generic floating point exception. To add insult to
> injury, it takes about 3/4 of second (!) to process a floating point
> exception on OSX, be it x86 or powerpc.
> 
> September 13
> The origin of this problem was detecting overflow during exponentiation on
> OSX/386, which uses the SSE unit to do the math. I ran a speed test
> comparing exponentiation using the SSE unit with x87, and the x87 is 25%
> faster, and delivers a more accurate result because the intermediate
> quantities are carried to 64 bits in the mantissa. So the solution is not to
> bend over backwards for the SSE unit, but rather use the x87 for
> exponentiation where it is available.
> =====================

Wonder why they would choose that?  The extra instructions to set up the MMX registers only pays off, I think, when one actually uses the multiple registers.  For a single computation, the more general and common approach would be better...  They aren't expecting an MMX version of gnuplot, I hope.  ;-)

Dan