GKN%OAK.SAINET.MFENET@LLL-MFE.ARPA (02/24/86)
From: <GKN%OAK.SAINET.MFENET@LLL-MFE.Arpa> (Gerard K. Newman)
Date: Mon, 24-FEB-1986 08:42 EST
To: Info-VAX@SRI-KL.Arpa
Message-ID: <[OAK.SAINET.MFENET].6EEAB060.008EB05B.GKN>
US-Mail: Science Applications; P.O. Box 2501; Oak Ridge, TN 37831
Telephone: (615) 482-9031 x293
X-VMS-Mail-To: INFO_VAX
From: Arpa%"poneill@AMSAA.ARPA" 15-FEB-1986 20:11
Subj: VMS Fortran Problem
We have a VMS-based VAX 782 with Fortran V4.3 and floating point
accelerator. We compiled a model with and without the g_floating
option it takes on the order of 32 times longer to execute with the
option in. The model was developed on an IBM system and takes less
than 10 seconds to run on the IBM. On the 782 with the G floating
option it takes 16 minutes plus to run the same set of parameters.
Without the G option, it takes about 30 seconds(.5 minute) to run.
We need to use the G option to get correct numerical results
(precision is evidently a concern). Is this difference in run time
solely associated with going with more precision ? Any comments are
welcome!
Pat:
There are four floating point formats defined by the VAX architecture.
They are referred to as F, G, D, and H floating. Not all VAX processors
implement all four types in hardware or microcode. The 11/780 (and hence
the 11/782) do not implement the G and H floating point formats in hardware
without the installation of a special option (I'm not sure if this option
is hardware or software).
The difference you see in run-time is due to the fact that VMS is emulating
G floating instructions in software.
gkn