From:	jazz AT opennt DOT com (Jason Zions)
Subject:	Re: Floating Point Usage
5 May 1998 00:39:38 -0700 :
Message-ID:	<31AA903A2A1FD111A06300805F4B6D640174A631.cygnus.gnu-win32@ssi2.opennt.com>
Mime-Version:	1.0
To:	"'paslawsk AT cadvision DOT com'" <paslawsk AT cadvision DOT com>,
	"'gnu-win32 AT cygnus DOT com'" <gnu-win32 AT cygnus DOT com>

> I am porting an application from Linux to Windows NT which makes intensive
> usage of floating point operations.  I am finding slight differences in
> calculations

It's been our experience that NT on Intel sets the FP unit's internal
precision for 53-bit mantissas (64-bit register size) while linux (and other
Intel unices) sets it to 80-bit internal precision. Thus, all intermediate
computations under Linux carry another dozen or so bits of precision that
only get lost when an FP register is copied out of the FP stack into a
64-bit register.

We decided to remain compatible with the rest of the NT world and not enable
full 80-bit mode. (There's one or two conversion routines where the easiest
solution was to enable 80-bit mode, do the FP operation to convert the
number, then reset 64-bit mode.)

Jason Zions
Softway Systems, makers of OpenNT
Enterprise Middleware providing Unix Services for Windows NT

-
For help on using this list (especially unsubscribing), send a message to
"gnu-win32-request AT cygnus DOT com" with one line of text: "help".

- Raw text -