On 03/20/12 00:03, Mark Knecht wrote:
> On Mon, Mar 19, 2012 at 7:05 AM, Andrew Lowe <a...@wht.com.au> wrote:
>> On 03/19/12 20:34, Mark Knecht wrote:
>>> On Mon, Mar 19, 2012 at 5:32 AM, Mark Knecht <markkne...@gmail.com> wrote:
>>>> On Sun, Mar 18, 2012 at 10:26 PM, Andrew Lowe <a...@wht.com.au> wrote:
>>>>> Hi all,
>>>>> Has anyone played around with the various "better known" compilers on
>>>>> Gentoo? By "better known", I'm referring to gcc, Intel, llvm, pathscale.
>>>>> My
>>>>> situation is that I've just started my PhD which requires me to do Finite
>>>>> Element Analysis, FEA, and Computational Fluid Dynamics, CFD, and I want
>>>>> to
>>>>> find the "best" compiler for the job. Before anyone says "Why bother, XXX
>>>>> compiler is only 1 - 2% faster than gcc", in the context of the work I'm
>>>>> doing this 1 - 2% IS important.
>>>>>
>>>>> What I'm looking for is any feedback people may have on ability to compile
>>>>> the Gentoo environment, the ability to change compilers easily, gcc-config
>>>>> or flags in make.conf, as to whether the compiler/linker can use the
>>>>> libraries as compiled by gcc on a "standard" gentoo install and so on.
>>>>> Obviously there is much web trawling to be done to find what other people
>>>>> are saying as well.
>>>>>
>>>>> Any thoughts, greatly appreciated,
>>>>> Andrew Lowe
>>>>>
>>>>>
>>>>
>>>> Think CUDA
>>>>
>>>> Mark
>>>
>>> Sorry. Meant to include this reference: <$15 on Kindle. Reads great on
>>> Kindle for PC.
>>>
>>> http://www.amazon.com/CUDA-Example-Introduction-General-Purpose-ebook/dp/B003VYBOSE/ref=sr_1_4?ie=UTF8&qid=1332160431&sr=8-4
>>>
>>>
>>
>> I'm sorry but I'm doing a PhD, not creating a career in Academia. The
>> concept of writing an FEA or CFD from scratch, with CUDA is laughable, I
>> just don't have the time to learn CUDA, research the field, small
>> displacement, large displacement, dynamics, material nonlinearities,
>> write the code, and then most importantly benchmark it to make sure it's
>> actually correct. This is all bearing in mind that I have 20+ years
>> experience as a C/C++ technical software developer, including FEA and
>> CFD. I'll actually be using Code Aster, an open source FEA code that
>> runs under Linux.
>>
>> Sorry if I sound narky, but compilers is the subject at hand, not how
>> to write FEA code.
>>
>> Anyway, thanks for answering,
>>
>> Andrew
>>
>
> Nahh, be as snarky as you like as long as you don't really mean it personally.
>
> My experience with CUDA, and I'm not a programmer, is that there is a
> fairly steep learning curve. However changing C compilers will get you
> maybe 5%. Changing to CUDA will get you 30,000%, assuming a mid-high
> range CUDA card and that you can parallel-ize FEA. I did a little
> Googling and it seems that FEA is a pretty common CUDA topic so I
> don't think at the outset that you'd find yourself all alone.
>
> Good luck whatever you do and know that I didn't mind the response at all! :-)
>
> Cheers,
> Mark
>
>
The thing is that I agree that CUDA is the way to go for things like
FEA & CFD, in fact the mob that runs the super computer I'm using is
installing another one that is top heavy in CUDA cards. But the thing is
as I'm using the FEA as a tool, rather than playing around with the
innards of the code, I need an established code, one that has
verification behind it. My topic looks at the way that steel connections
behave so I need an established FEA code that is verified to provide the
correct answers, I don't get that if I write my own code.
Most likely I'll write a short paper covering a comparison of existing
C/C++ compilers and their relative speeds, spend the next 18 months - 2
years doing my research and then to close things off, I'll probably be
able to write another short paper concerning CUDA speedups as the FEA
code bases will have caught up and been verified.
Thanks for peoples replies,
Andrew
p.s. Writing this I just had a sudden horrific though and checked the
FEA code I'm using, Aster. It's written mostly in FORTRAN - fat chance
I'm going to be hacking that code........
