SGI: Development

didn't find a fitting thread but if there's one please merge


as you've probably noticed i've been playing around with all kinds of irix versions starting from 4 lately and naturally at some point i wondered about the benefits of all the upgrades over time. namely coff, (o)32 and n32 so i ran a little test with gzip:

• irix4.0.5, coff: 2.12
• irix5.3, coff: 2.12
• irix5.3, o32: 2.12
• irix6.2, n32: 2.04
• irix6.2, o32: 2.12

these are average values to keep it fair. i ran each test a couple of times to determine the average.
no surprises so far you might say but only at first sight. all gzip builds i used i compiled using the original mips(pro) compilers for each irix version. but only the compiler i used for the n32 version does support -O3 for object files i.e. all others i built using -O2. not fair you might say and that's what i thought as well so i built the n32 version with -O2, too and the result was 2.19! and that although it's mips3 and all others are just mips1.
so i went ahead and compiled a mips1 -O2 version on 6.2 and to my surprise that one beat all others except for the n32 getting a 2.09. might be the result of the newer compiler i have for 6.2 compared to the ones for 5.3 and 4.
intrigued by that i built one more gzip but this time taking the detour and doing a 2 phase build; ucode first and then the final link. the only way to use -O3 for mips1 o32 if there's more than one source file. did that and ran the test once more: 2.05!

so the good old mips1 build with just using -O3 could compete with the fancy n32 mips3 build built with not just -O3 but the whole spiffy -OPT:Olimit=0:roundoff=3:div...

sure, the test is far from comprehensive but still interesting i think.

foetz wrote: so the good old mips1 build with just using -O3 could compete with the fancy n32 mips3 build built with not just -O3 but the whole spiffy -OPT:Olimit=0:roundoff=3:div...

sure, the test is far from comprehensive but still interesting i think.

What happens if you do something more floatingpointy ? for integer, a p-iii will kick Irix butt every which way but loose ...

hamei wrote:

foetz wrote: so the good old mips1 build with just using -O3 could compete with the fancy n32 mips3 build built with not just -O3 but the whole spiffy -OPT:Olimit=0:roundoff=3:div...

sure, the test is far from comprehensive but still interesting i think.

What happens if you do something more floatingpointy ? for integer, a p-iii will kick Irix butt every which way but loose ...

sure, as i said it's a simple test but i have no idea how a p3 matters here. this was meant to see the differences between the irix binary types, nothing else. so, obviously, anything else but what runs irix is irrelevant.

foetz wrote: i have no idea how a p3 matters here.

'cuz even a p-iii is so much faster than any MIPS cpu at integer that it doesn't seem like they'd have spent much effort optimizing for integer ? If one were going to write integer-heavy apps you'd have to be a lunatic to choose Irix ...

hamei wrote:

foetz wrote: i have no idea how a p3 matters here.

'cuz even a p-iii is so much faster than any MIPS cpu at integer that it doesn't seem like they'd have spent much effort optimizing for integer ? If one were going to write integer-heavy apps you'd have to be a lunatic to choose Irix ...

hehe yes true but again missing the point. this test was solely meant for getting a quick impression of how much the different irix binary types (and only them) differ and not how fast or slow irix is compared to whatever else under whichever conditions.
and as i said 2 times already now there could of course be many more tests like that to give a better picture

hamei wrote: If one were going to write integer-heavy apps you'd have to be a lunatic to choose Irix ...

If one were going to write integer-only apps for scientific computing, you'd have to be a lunatic to use them...

nongrato wrote:

hamei wrote: If one were going to write integer-heavy apps you'd have to be a lunatic to choose Irix ...

If one were going to write integer-only apps for scientific computing, you'd have to be a lunatic to use them...

Uh, a lunatic to use what ?

Among other things, we ran Groundwars and the NATO Reference Mobility Model on IRIX for many years and it worked great...