EKOPath 4 Compiler - Performance

Discussion related to the organization of the source code, repository and code-level/compiler optimization.

Moderators: Dade, coordinators

EKOPath 4 Compiler - Performance

Postby lfrisken » Mon Jun 13, 2011 2:41 pm

I've been following phoronix news of late, and Michael Larabel has been raving about they performance improvements available when using EKOPath 4 Compiler suite, which has just been open sourced, and was previously $2000.
article:
http://www.phoronix.com/scan.php?page=article&item=pathscale_ekopath4_open&num=1

results:
http://www.phoronix.com/scan.php?page=article&item=pathscale_ekopath4_open&num=3
and more:
http://www.phoronix.com/scan.php?page=article&item=pathscale_ekopath4_open&num=4
And I think he may have good reason to be raving, look at the performance increase over gcc with c-ray! It may be a simple benchmark, and therefore not a particularly realistic one, but does anyone think it would be worth trying to compile luxrender using this compiler? Is anyone game enough? :roll: Apparently it's capable of building Qt, Kde4 and the linux kernel, so it should be capable of luxrender right? (maybe too big of an assumption there!). I'll try to give it a shot soon when I get linux on my desktop again.
Last edited by lfrisken on Mon Jun 13, 2011 2:43 pm, edited 1 time in total.
lfrisken
Developer
 
Posts: 129
Joined: Fri Apr 24, 2009 5:46 am

Re: EKOPath 4 Compiler - Performance

Postby Abel » Mon Jun 13, 2011 2:43 pm

Check this thread: viewtopic.php?f=13&t=6366 ;)
User avatar
Abel
Developer
 
Posts: 1847
Joined: Sat Oct 20, 2007 8:13 am
Location: Stuttgart, Germany

Re: EKOPath 4 Compiler - Performance

Postby daidai67 » Fri Jul 01, 2011 8:37 am

lfrisken wrote: increase over gcc with c-ray!
If you compile with -finline-limit=1000, GCC 4.6 take the lead over ekopath. ekopath win mostly because of very aggressive inlining, that gcc doesn't do by default as it increase too much the code size.

regards,
daidai67
 
Posts: 10
Joined: Wed Dec 31, 2008 2:25 am

Re: EKOPath 4 Compiler - Performance

Postby Juicyfruit » Fri Jul 01, 2011 8:58 am

It seems that for some cases locality is more important the binary size. I have been unable to compile with ekopath and dont have the time to really try to hard.
Juicyfruit
 
Posts: 23
Joined: Sun May 15, 2011 3:04 pm

Re: EKOPath 4 Compiler - Performance

Postby daidai67 » Fri Jul 01, 2011 6:56 pm

yes, in case of c-ray 1.1 at least. I tested both (gcc and ekopath) and with gcc -O3/-Ofast -march=native -flto -finline-limit=10000, gcc wins with a 1-2% margin. Compiling complex project with ekopath give me compiler crash and I give up without trying to hard myself. The same with intel compiler, most of the time it can give better performance, but is not as stable as gcc and sometimes give strange result (and I prefer accurate result every time over fast inaccurate..)
daidai67
 
Posts: 10
Joined: Wed Dec 31, 2008 2:25 am

Re: EKOPath 4 Compiler - Performance

Postby Dade » Sat Jul 02, 2011 2:23 am

daidai67 wrote:The same with intel compiler, most of the time it can give better performance, but is not as stable as gcc and sometimes give strange result (and I prefer accurate result every time over fast inaccurate..)


This is the same reason why we stop using Intel Compiler in the past: after of a couple of headaches with compiler bugs and not IEEE compliant floating point precision.
User avatar
Dade
Developer
 
Posts: 8404
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

Re: EKOPath 4 Compiler - Performance

Postby foobarbarian » Sat Jul 02, 2011 7:30 am

Dade wrote:
daidai67 wrote:The same with intel compiler, most of the time it can give better performance, but is not as stable as gcc and sometimes give strange result (and I prefer accurate result every time over fast inaccurate..)


This is the same reason why we stop using Intel Compiler in the past: after of a couple of headaches with compiler bugs and not IEEE compliant floating point precision.


I did a few test runs and wasn't able to find any errors in the generated results. Of course careful studying of the manual was required as some options do indeed change the floating point behavior / precision.
Considering that the Intel Compiler is used mostly for scientific calculations I think lux should run okay with it.

I'm only advocating it because the speedup was big while still maintaining correct math. The gap is much smaller compared to GCC than it is compared to MSVC.
MSVC with profiling around 4400 max - while the with Intel 5040 points. Funnily enough any try to use profiling runs to optimize later optimized builds with the Intel compiler resulted in slowdowns.

I've noted the settings used here:
http://www.luxrender.net/wiki/Foobarbarian_Compiles

cheers
amir
foobarbarian
Developer
 
Posts: 36
Joined: Sat Mar 19, 2011 10:21 am


Return to Organization & Optimization

Who is online

Users browsing this forum: No registered users and 1 guest