IBM releases Cell SDK and simulation tools

SubD said:
The problem for SIMD support in compilers, in my experience, has been that the areas you get for free from the compiler are either non-hotspots that make no noticeable performance difference or hotspots that you are going to be focusing heavily on anyways and will be tweaking by hand.
Click to expand...
You know, just having compiler native SIMD types would go a Long way in helping things. Autovectorization from what I've seen in the past has always just worked more as a bandaid then anything more - compilers need higher level domain knowledge then they typically have if we're to hope for real improvements.
Though I don't know about XLC itself - just referring to other compilers that did similar stuff I've seen.

I'm also of belief that a properly featured AoS ISA can be much more compiler friendly in general then the... that other stuff that basically all desktop CPUs have. :p
 
Fafalada said:
You know, just having compiler native SIMD types would go a Long way in helping things. Autovectorization from what I've seen in the past has always just worked more as a bandaid then anything more - compilers need higher level domain knowledge then they typically have if we're to hope for real improvements.
Though I don't know about XLC itself - just referring to other compilers that did similar stuff I've seen.

I'm also of belief that a properly featured AoS ISA can be much more compiler friendly in general then the... that other stuff that basically all desktop CPUs have. :p
Click to expand...

Look, if you do not hand over the VFPU I'll get Intel to hire you and make you work 24/7 on SSE1 and x87 optimization for THE REST OF YOUR LIFE MUAHAHAHAHAHAHAHAHAHAHAHHAHAHAAH!




Ahem... j/k ;)............... :( sorry.... bad Pana... bad...
 
Last edited by a moderator:
:p Actually... who said anything about VFPU? Aren't GPU shaders a good practical example of how compilers can do a good job with a horizontal ISA?
 
Fafalada said:
:p Actually... who said anything about VFPU? Aren't GPU shaders a good practical example of how compilers can do a good job with a horizontal ISA?
Click to expand...

Yes, they are, but I needed an excuse to annoyingly beg for the VFPU on a public forum... oh, no keyboard-thinking again...
 
Sweet, I'll try working with this if I have time tonight or tomorrow. I'd really like to try implementing a raytracer, and then a photon mapping implementation on cell. Very good news! :D

Nite_Hawk
 
I have a maybe very dumb question for the programmers here:
Assuming you have written a certain app with the SDK and run it on the system emulator, is there a way to find out how the program would run on the real cell performance-wise ?
I mean are there some stats shown about cpu usage etc you could look up and then do the math ?
 
macabre said:
I have a maybe very dumb question for the programmers here:
Assuming you have written a certain app with the SDK and run it on the system emulator, is there a way to find out how the program would run on the real cell performance-wise ?
I mean are there some stats shown about cpu usage etc you could look up and then do the math ?
Click to expand...
Its claimed to be cycle-exact, so emulated-Cell-cycles/clockspeed-of-the-real-thing should be runtime in seconds.

Any idea why this is limited to Fedora Core only? Are there only binaries available for the emulator ATM?

\Cries for a Version compatible with Cygwin
 
On the IBM forum it was mentioned that the whole SDK was internally developed on Fedora Core 4, hence it was the most stable candidate. They did mention they had several other Linux ports, but apparently didn't want to release more than one so not to confuse people, and as well they wanted an affordable dev platform.
 
A friends was telling me of the EULA that comes with this SDK. You're only allowed to create unproductive software, anything you create belongs to IBM, and if IBM decide to sell it you have to delete your free SDK within 10 days.

From the sounds of it this is just a 'sweetener' to let devs get their hand in, but isn't intended (or might not be intended - IBM might change their mind) for producing actual products. Though the idea that everything produced belings to IBM (if true. Haven't seen the EULA myself) is nonsense. Copyright law is very clear on that.
 
Got a chance to set everything up, finally. Was pretty painless, asides from one or two small things. The simulator is definitely interesting - lots to read now o_0
 
Titanio said:
Got a chance to set everything up, finally. Was pretty painless, asides from one or two small things. The simulator is definitely interesting - lots to read now o_0
Click to expand...

So, what are you learning?
 
Shifty Geezer said:
A friends was telling me of the EULA that comes with this SDK. You're only allowed to create unproductive software, anything you create belongs to IBM, and if IBM decide to sell it you have to delete your free SDK within 10 days.
Click to expand...
This just relates to stuff like suggested improvements ... ie. so beta testers don't come back and claim they deserve a piece of the action. It's just a boiler plate license for pre-release software AFAICS.
 
Titanio said:
Got a chance to set everything up, finally. Was pretty painless, asides from one or two small things. The simulator is definitely interesting - lots to read now o_0
Click to expand...


Did you download the files from both the IBM site and the other one (the install script did not download for me the archives that were not at IBM's site :p) as well as freeglut (gave me an error in one of SCE's sample without freeglut and glut.h).

Yeah, besides those things it was relatively painless :).
 
Panajev2001a said:
Did you download the files from both the IBM site and the other one (the install script did not download for me the archives that were not at IBM's site :p) as well as freeglut (gave me an error in one of SCE's sample without freeglut and glut.h).

Yeah, besides those things it was relatively painless :).
Click to expand...

I managed to get the files off the other site OK, although I didn't initially realise you had to manually download those ones off the Barcelona site - I thought the install script was going to pull them down. After that, the actual build was a lot quicker for me than the IBM article suggested.

The only other problem I've had sofar was the missing "-m32" when compiling the hello world example. Thankfully someone else had that problem before me ;) I haven't looked at the other samples though - which sample relates to freeglut/glut.h?

mckmas8808 said:
So, what are you learning?
Click to expand...

Not much so far. Except maybe that I need more hours in the day ;)
 
Titanio said:
I managed to get the files off the other site OK, although I didn't initially realise you had to manually download those ones off the Barcelona site - I thought the install script was going to pull them down. After that, the actual build was a lot quicker for me than the IBM article suggested.
Click to expand...
Yeah, the "getting started" page gave me the impression that the install script would download the stuff off the Barcelona site too. I did a fresh install of FC4 and I got nailed by some SE Linux security setting that didn't let me build the ramdisk image of linux for cell too, but luckily that was covered in the getting started article.

The only other problem I've had sofar was the missing "-m32" when compiling the hello world example. Thankfully someone else had that problem before me ;) I haven't looked at the other samples though - which sample relates to freeglut/glut.h?
Click to expand...
That would be "realizer.c", part of the "vse_subdiv" (variable-sharpness edge subdivision) sample program. IINM that runs off the host system and not within the simulator itself.
 
Last edited by a moderator:
Silkworm said:
That would be "realizer.c", part of the "vse_subdiv" (variable-sharpness edge subdivision) sample program. IINM that runs off the host system and not within the simulator itself.
Click to expand...

Cheers! I'll keep an eye out for that.
 
Titanio said:
I managed to get the files off the other site OK, although I didn't initially realise you had to manually download those ones off the Barcelona site - I thought the install script was going to pull them down. After that, the actual build was a lot quicker for me than the IBM article suggested.

The only other problem I've had sofar was the missing "-m32" when compiling the hello world example. Thankfully someone else had that problem before me ;) I haven't looked at the other samples though - which sample relates to freeglut/glut.h?
Click to expand...

It relates to GL/glut.h, when I installed Fedora Core 4 it was the only file it was missing.

You discove this "issue" when you build the whole package with the script, the thing is that it does not stop the creation of the simulator and it won't end with BIG MISTAKE I CANNOT FINISH, it gives an error for that "sample" and it continues with the rest.

Log the output next time or just scroll up after the install.sh script has finished its work and you should see it... unless you already have glut.h in your GL folder.

Boh...
 
This may be tangentental to the subject of the SDK on the whole, but the word is that Mercury Systems will be featuring/demoing their 'Turismo' quad-Cell blades today. Anyone have any word on this?
 
You must log in or register to reply here.
Back
Top