Slashdot Mirror
IBM Releases Compiler for Power4 and G5
davids-world.com writes "IBM offers its optimized XLC compiler not just for Intel CPUs, but also for its own G5 processor (article in German at Heise). Unlike gcc, it is optimized for the G5 and achieves a major boost in speed, as first results show. I guess we will have to compare the new benchmark data (once available) with the data we get with the optimized Intel compiler for Xeon. The compiler is available for download now."
Lest anyone forget, Apple beat Intel in real world benchmarks... so the PC fanboys cried that SPEC benchmarks are the real measurement to gague speed... (probably because the comparisons were much closer when conducted this way). When SPEC benchmarks were displayed, these same fanboys cried that Intel's compiler wasn't used (instead the same compiler between platforms). Apple replied that its fairer to normalize the compiler between platforms and that while Intel could have achieved higher results when their compiler was used, Apple could do the same. So, here is that compiler. When/if the G5 outperforms Intel's best, what will the fanboys rally cry be next?
And is there a particular reason why IBM couldn't apply their work towards gcc? So much for the whole open-source, contribute-to-the-community philosophy.
Maybe they just didn't feel like it. Maybe they just preferred working with their own code. IBM's contributed, a lot, it doesn't mean that they're now servants of the open source community.
Believe it or not, people DO sometimes run singular tasks on hardware which they want highly optimized...and believe it or not, you can actually install more than one compiler on your system at a time(yes, I know, amazing!)
Please help metamoderate.
IBM has contributed more to the open source community than I would venture to guess any other corporation on earth. (think of the code, projects, money, support, etc). I would not be so quick to blast them for not doing everything under an open source license. Encourage yes, but let's be civil about this. IBM is not just blowing smoke about OSS, they have put their money, time, and products where their mouth is.
Finkployd
Why would they use GCC as a base? It's not intended to compile over a wide range of platforms, just the pSeries processors (which includes the G5), so it's not like they need to be able to port it to x86. When you code a compiler for one fixed architecture you can (and almost always do) end up with HUGE speed increases, since you get to exploit all the fun capabilities of a chip without having to worry about compatibilty.
Put simply, IBM wasn't that interested in improving the cross-platform GCC this time. They were interested in improving a number of their OWN compilers (not just C/C++, Fortran too!) for their OWN architecture.
Because they aren't in love with RMS's conditional license? Because they have IP they dont want to release? Because they make money from their own compiler?
Is there any reason that gcc maintainers dont just improve gcc? Isn't that what they're supposed to do?
Gcc may be a complete compiler suite. Whee. It's hardly optimized for any platforms besides x86 and Alpha, though.
I don't need no instructions to know how to rock!!!!
they probably wrote this a long time ago. Writing custom compilers for one architecture is easier than writing generic backends that gcc needs. And IBM has always had a lot of people working on compilers (like Backus, who did Fortran). Perhaps they will be persuaded to contribute to gcc too.
And is there a particular reason why IBM couldn't apply their work towards gcc?
Yes.
GCC rejected alot of their patches. GCC is a cross paltform compiler, the maintainers can be very picky about adding code that only helps one platform.
IBM has contributed quite a bit to GCC, but GCC is not the proper place for a heavily optimised platform-specific compiler effort.
"The worst tyrannies were the ones where a governance required its own logic on every embedded node." - Vernor Vinge
And is there a particular reason why IBM couldn't apply their work towards gcc? So much for the whole open-source, contribute-to-the-community philosophy.
First off, most companies develop compilers for their hardware/software because they have the behind-the-scenes knowledge necessary to produce a superior product, and therefore, make money off of it. It's called having a competitive advantage, and I see nothing wrong with that.
Second, who says that they won't eventually work to get these performance enhancements included into a future release of GCC? Just because they haven't done so yet does not mean that they never will.
Rule #1 -- Politics always trumps technology.
I know everyone wants to know why they did not just contribute to GCC, but seriously people, I imagine they have their reasons.
Perhaps they wanted something done in a timly manner without waiting for the GCC people to accept their patches. Perhaps they felt a more elegant solution could be achieved by building their own compiler from scratch. Maybe they think GCC is a piece of crap (which for high performance computing, it is).
Just because IBM support open source does not mean they are obligated to do everything for the sole purpose of advancing the movement.
Finkployd
What's a "real-world" benchmark? Comparing the function of photoshop on mac vs. pc, when it's developed natively for the mac? That's not really fair. It's simply not the same code. We could take any of the many programs made natively on PC (which are then ported to mac) and do the same trick.
so the PC fanboys cried that SPEC benchmarks are the real measurement to gague speed... (probably because the comparisons were much closer when conducted this way).
And I'm sure that the macheads don't favor the "photoshop test" for the same reason? We need some objectivity here. Nothing wrong with a fair benchmark, and I'll go on record as always preferring a benchmark, regardless of who it favors.
Apple replied that its fairer to normalize the compiler between platforms and that while Intel could have achieved higher results when their compiler was used
Why fairer? Shouldn't tests use whatever compiler will *actually be used*? If apple/IBM made a chip without a decent compiler to support it, that's their problem. I'd say it was fair to use Intel's compiler for intel's chip. Now it will be fair to use IBM's compiler for IBM's chip. Also interesting was how you suspect the "fanboys" of being biased when they wanted benchmarks instead of "real-world" tests, but Apple just "realized that it's fairer" to do something that benefited them. It works both ways.
Apple could do the same. So, here is that compiler. When/if the G5 outperforms Intel's best, what will the fanboys rally cry be next?
That's perfectly fair. Let's use the best compiler for each machine. I've no problem with it.
Now, to ensure a completely fair test, let's make sure Apple uses both machines as they ship, and everything will be fine. But the pc "fanboys" as you call them were right to criticize a test in which Apple modified the machines, both their own and Dell's, fom the way they shipped. The changes it made to the Apple were beneficial, the ones to the Dell were all detrimental. That's not science.
I honestly like Apples, this is just a matter of scientific method to me. I really could give a shit who wins.
-Looking for a job as a materials chemist or multivariat
Hmm... Maybe because this compiler is something they plan to sell (from what I can tell it's a closed source binary-only download) and return a profit on.
If you really think that all IBM has contributed is some legacy drivers to Linux, I'm guessing you're either out of touch or I've severely misunderstood IBM's involvement. Look at the SCO lawsuit for a good example of many of the IBM developed/owned technologies that are now available as GPL in the kernel.
I think that it comes down to is they didn't use gcc because they weren't looking for some massive slow cross compiler to modify to fit their needs, they (as the makers of the hardware) wanted a package that was built from the ground up to build the best code it can for their architecture. It would seem almost impossible to make gcc as good on any one architecture as an architecture specific compiler written by the architecture's creator. If I want a high performance sportscar I don't start with a station wagon as a base (well, unless I have my own show on TNN) and modify it. I build it from the ground up.
Maxim: People cannot follow directions.
Increases in truth directly with the length of time spent explaining them
Eh, I wouldn't say that GCC is "worthless", it's just that its worth lies in an area that has nothing to do with high-performance computing. Or even mid-performance computing. :)
GCC's primary virtue is that it exists. For practically every computing platform on the planet, there's a famliar, stable toolchain that produces repeatable results. You can't optimize a program that doesn't exist in the first place, and GCC is the tool that's allowed many if not most of the programs we use on a daily basis to exist. There's value in that.
News for Nerds. Stuff that Matters? Like hell.
No, I don't mean scientific computing. I mean high performance computing. SP clusters, Linux clusters, weather modeling, the stuff that large universities do. No offense but where I am we run significantly larger clusters than SARA has ever seen.
GCC is a great compiler, I use it daily. Better than Intel and IBM for performance (expecially FORTRAN) it is NOT. It was designed as a cross compiler and in that it shines. But it is not all things to all people.
I certainly did not flame GCC and all the people who work on it, come back down off your high horse. All I did was list some reasons why IBM may not have used GCC, one of which is that most of the HPC people I know and work with have tested it and found it to be unacceptable for what they do.
That said, I certainly can see a day when GCC IS the best performing compiler, hands down. I believe open source will eventually overtake most commercial applications due to the momentium it currently has. However it would be silly of me to say that MySQL is better than Oracle, when it clearly is not (yet).
Finkployd
And is there a particular reason why IBM couldn't apply their work towards gcc? So much for the whole open-source, contribute-to-the-community philosophy.
Maybe because, oh, they didn't want to? Just because a company participates in Open Source projects doesn't mean that they have to participate ONLY in OS projects. In this case, they built a chip, then they built a compiler for that chip, and have freely distributed it. I, for one, think that's pretty nice of them. After all, there's NOTHING stopping you from using GCC instead of the new compiler. Your choice. Just like THEY made a choice. See how that works?
I'm personally rather tired of companies which consider "contributing to the open-source community" to be "lets send in drivers for our proprietary hardware which only we will ever need." That's NOT contributing to the community, that's getting your foot into the kernel.
If I have an older (or newer) IBM box that I've picked up, and I want to run some Open Source operating system on it, like Linux, I'm DAMN HAPPY that IBM contributed the stuff to the kernel that lets it run on the piece of hardware that I want it to. That is a GOOD THING(TM) in my opinion, as they have provided me with ANOTHER CHOICE. It means that they more than likely didn't have to guess how to do the drivers for it, and it's more than likely in their vested interest to keep it up to date.
Sure, it may help them make money on hardware purchases, etc., but this is beside the point.
Where is it written that people are only allowed to make contributions if they can't or don't make money off of it? At the end of the day, the community benefits from the submission, and they benefit from some sales. This is another GOOD THING(TM) because maybe it will help to show other companies that there is a reason to contribute other than trying to build up the karma points.
But hey, what do I know... I only live in the real world.
$0.02 (CDN)
the Intel C compiler spanks every other available x86 compiler
... you point it at some real code, nothing happens. Why is this?
:)
It used to, but it's a close call these days.
SSE/SIMD support is, in the words of one of my assembly-programmer friends, "awe-inspiring."
No it's not, it's fucking awful. Sure, the hype is good, the docs (which are actually very good) show a loop being unrolled to some SIMD instructions, including a little cleanup at the end in scalar instructions. You code it up, it goes
It's not magic, that's why. You put an exit condition in the loop, you break vectorising. If everything isn't lined up on nice 16 (?) byte boundaries, you break vectorising. Once you've gone through making sure all the conditions are met you realise that it's easier just to use the intrinsics in the first place.
Intrinsics, BTW, are very cool and much easier to code than you would have thought. GCC has them too
Dave
I write a blog now, you should be afraid.
Interesting to note that Objective C is much closer to raw C than C++, and as such yes. You can do some stuff with it that's worthwhile.
Also note that David Stes released his POC, which is an Objective-C->C compiler with a runtime he uses (actually it's a bit more than Objective-C) to run ObjC code today.
Apple will probably use this new compiler to recompile their CoreFoundation stuff, which is all C stuff that Carbon and Cocoa tap into. So, a simple recompile of CoreFoundation should net good speed improvements on G4s and MUCH better results on a G5.
So even if IBM chooses not to directly support Objective-C, Apple can still benefit in the short term while they rustle up their very own ObjC->C system (which is NOT a very challenging feat, considering the ObjC runtime is a relatively simple-to-use C library).
Slashdot. It's Not For Common Sense
The Pentium scores you listed are from Pentium CPU's (not the xeons which did slightly higher) which are incapable of dual cpu operation. This was against a dual cpu system that was probably optimized well for dual use.
Whats interesting to note that the single pentium machine scored 836 in int performance and the xeons (dual machines) scored 840... Almost no performance increase from going dual here.
Hmmm... Pie...
Both those guys must be pissed!
I know, I know, but I couldn't resist. . .
If you are doing serious Photoshop or After Effects work, all the time. No matter how fast the newest machine is, there's an art director out there with a layout which will make the machine choke.I am a believer of momentum and curves.
Can't we talk about Macs on Slashdot without having to talk about whether they're overpriced or slow or whatever? I have a Mac, I like it, I use it for lots of things. I'd like to be able to discuss it, rather than just read x86 users' posts about how much better their platform is & why. This article is about a complier, not comparison shopping. Could we stick to that please??
It seems to me that this special super-duper compiler wouldn't benefit any other CPU's since it's designed to make code that the processor executes the most effeciently.
Open sourcing a compiler for a CPU wouldn't automatically make the Xeon compiler any better. It would make gcc compile for that processor better.
Either way.. where's my super-duper Athlon and Athlon64 compilers!!
- It's not the Macs I hate. It's Digg users. -
Looking at all these pretty graphs, and knowing that in most cases sane people will restrict their optimization to -O2 (-O3 and above are usually unsafe: not as well tested, may change the program behaviour, cut corners with IEEE math, etc), GCC is still looking pretty good.
In most cases in both FP and Integer, GCC matches XLC up to -O3, sometimes a bit slower, sometimes a bit faster.
I applaud the work of the GCC people. GCC is the most versatile and portable C compiler, and it's not half bad at optimizing either.
Thanks too to IBM. Their compiler will surely prove useful in a lot of cases, and a new compiler to try and benchmark is always good news!