AMD Optimization

[mmciastro]

This thread was started so I could include a link to it for someone at AMD to find.  This thread is so he can ask questions of developers and optimizers, and perhaps so developers and optimizers can ask question of him as well.

Please don't use this thread for general conversations.  I know it's in the "general forum", but with Simon being AWOL at the moment, we have little choice but to have the conversation here, as only Simon can get him into the other "prerelease testers fora".

Thanks for you cooperation

tony

I'm off to send him an email now.  Don't expect a quick response, atleast I don't.

[mmciastro]

Thank you all for keeping this thread clean for our AMD contact.  I was worried my email hit his spamfilter, so I sent a message a different route and this is how he responded

[snip]" I’m actually working very hard on this, drafting a plan of attack.  I’ll be in touch soon."

I thank you all for your patience.

tony

[mmciastro]

Not wanting to push anyone to do something, I gave him plenty of time.  Today, I sent an email to remind him I was still here and was interested in what he was working on.  He responded with the following:

I’m proposing that we start a BOINC optimization contest actually.  My last day at AMD is a week from today – would you mind if I had someone from my team get in touch with you on how to carry this forward?

So, I guess we wait for someone else to come along.  I ofcourse said it was fine to pass my name/email to the next person and asked for more info on the "contest" idea.

[mmciastro]

He has responded with this:

The contest would be conducted by Top Coder or an organization like it, and would include almost anyone that wants to participate.  A typical clause for a contest of this type would be that AMD employees and their family members would be exempt from receiving the prize (whatever that is).

http://www.topcoder.com/

I know open this thread for any discussions/comments you might have, as it appears he won't be asking developers here questions.

[Alex Kan]

Well, I'm a little disappointed that they're not planning on asking us questions here, and that seems like the exact opposite of the process that we've been using, but seeing as we've never tried it, it'll be interesting to see what comes of it. That said, there are some logistic issues involved:

• SETI@home is licensed under the GPL, and the currently-available optimized apps branched off from the main tree long ago. Although some optimizations have been merged back to the main line, thanks to Joe, all of the current active optimization work done here or elsewhere seems to be targeted towards those branched codebases. I'm assuming that AMD needs to level the playing field for the purpose of competition, so this creates an issue in that starting from the official source may cause certain wheels to be reinvented many times, and the availability of other optimized app source online may complicate things further. I'm not a TopCoder participant myself, but do most of their competitions usually start from scratch?
• I'm also assuming that AMD is particularly interested in targeting optimizations for K10, but without K10-based systems in wide circulation, the only thing people really have to go off of is the optimization guides that AMD has released. Past optimization work has been very much profile-driven, so how do they propose to do anything particularly hardware-specific?
• What happens after the contest ends and the extrinsic motivators disappear? No one's paying the current crop of optimizers, so I assume we're doing this because we enjoy it. Does AMD plan to support optimization efforts after the contest ends? I guess this is largely a function of what the contest produces, although I would be loath to ever say that we've found every optimization possible.

[Jason G]

Hmmm, only some musing about this,
    From what I understand [could be wrong], Top coder methods encompass the [important and worthy] tasks of algorithmic and software engineering refinement. Their primary platform, being Java and dotnet cross platform applications, would seem to me to necessitate such rigourous methodology due to the associated platform overhead.

 To me it would seem, that with seti@home, the high level algorithmic and developmental refinement was accomplished long ago [i.e. Fast fourier transform is already of NLogN complexity] so there are little gains to be had in that regard IMO.

Beyond algorithmic and design optimisations lie the more specialised hardware optimisations such as those currently filtering through the optimised apps to the stock app (thanks to alex, simon, joe, Ben and probably others I;ve missed).  AFAIK there are no mentions on top coder of vectorisation, cache management, or hardware specific optimisation. [ though clearly present in chicken soup code ]

So the part of me that spent years studying computer science, respects what methods used by top coder can do in terms of productivity.  The other part that has been studying  electronic engineering , recognises that you can only get so much performance from abstractions before you need some  implementations

[Jason G]

Posting For the sake of completion, and to correct myself:  I just finished reading a 2007 IEEE paper by MIT researchers Steven G. Johnson and Matteo Frigo, confirming/matching  "Recent results by Van Buskirk et al."

The paper describes a new derivation of the Yavne "Split Radix" Fast Fourier Transform that has held the record for lowest flop count since 1968.  The new algorithm is said to be ~6% fewer flops.

So I'm discarding my prior opinions, where I stated that little improvement could be found from an already NLogN algorthm, and also dismissing my prior opinion that there are little gains to be had in further high level algorthimic optimisation of the seti science app.  I was wrong, go for it Top Coders

Jason

[_heinz]

Hi,

New "BARCELONA" Software Developer Resources are available there.

Its worth to have a look at it.

regards heinz