[Info-vax] Floating point emulation of OpenVMS / Alpha / ES

[Stephen Hoffman]

On 2022-10-12 14:26:04 +0000, Nick K said:

> Most of what I'm looking at is single-threaded software, think Fortran 
> scientific applications that have been around for a really long time.  
> The newer versions were compiled with the IEEE floating points (T), the 
> older versions were not.  If I narrow things down to a half dozen 
> emulators I want to look into, and all vendors want to give me a trial, 
> this would still take a week or two, for me to configure systems and 
> run benchmarks (unfortunately, my boss has other ideas for my time).  
> This is why I'm asking if something is already available, or 
> heuristically known.  SpecFP?  FLOPS?  Emulator vendor A's pet project 
> is IO bound, emulator vendor B's pet project is Int bound, emulator 
> vendor C's pet project is Fortran bound?  Given that EmuVM and AVT have 
> offices that are 5 km apart, I assume there must be some significant 
> differences between all the various emulators.

There's no way to know without running the apps under emulation. I've 
met performance and timing and stability surprises when updating apps 
to different native hardware of the same architecture, as well as with 
emulated hardware.

Emulation is not going to get you FLOPS. An emulator with JIT will 
normally do better than one without. But whether even a JIT is 
beneficial for your particular case, and your mix of code? Test.

I've also met weird floating point bugs. Old Fortran code with floating 
point data can be... interesting... to reproduce the same results, 
should even slight changes to floating point implementations and 
settings arise. Test.

If you have source code, it's likely the results of a native port to 
OpenVMS x86-64 will be faster than the path of emulation. Native 
Fortran is arriving later this year or sometime early next year, per 
the current VSI roadmap. And, of course. test.


-- 
Pure Personal Opinion | HoffmanLabs LLC