Seti@Home optimized science apps and information
Optimized Seti@Home apps => Windows => Topic started by: Devaster on 14 Mar 2007, 03:40:04 pm
-
wanna try AMD FFT optimized SETI ?
wait to tomorrow ... ;)
-
Sounds like fun!
-
based on ACML (aka Intel KML)....
-
I thought that's what you had used; how much did you have to change in the sources to make it work, and is only FFT done via ACML?
Regards,
Simon.
-
fft and sincos in data chirping ....
-
Hey that sounds cool.. I wonder if good o'l ACML speeds up for real..
I saw a paper comparing FFTW analysis and ACML and IPP on the same page doing FFT analysis but ACML didn't really come all the way up to IPPs performance though, but it is sure as h*ll interesting to see..
Nowadays i've turned to the dark side (In*el) but it could be really fun to see the benefits and implementations of ACML.
Wait until AMD releases their Quaddie later this year with a highly re-optimized core structure.. That can be VERY interesting but i really doubt that they will get around of beating Intel with their 2nd gen Quaddie also around the corner though..
Keep up the good work ..
Kind Regards Vyper
-
sincos is for 64-bit only
-
Now for a fun question :)
Does ACML work on Intel 64-bit CPUs, too?
Regards,
Simon.
-
I will try/test....I have 2 AMD systems Athlon XP 3000+ and X2 4800+ (Socket 939).
I will check back tomorrow to see if it is available.
-citroja
-
try....
http://www.megaupload.com/?d=EOPVNO5M
its for all AMD with SSE minimum ....
not optimized, only fft replaced with ACML
TEST VERSION !!!!
give me feedback if is good in validation
-
Tried to run the test app in Knabench on my FX60. Got some kind of DLL error, and would not run.
How about you, Simon?
-
shit what dll ???
maybe rapidmind ....
http://www.megaupload.com/?d=WFUTXIAE
copy dlls to exe file ....
version with APL is coming soon ....
-
I'll run it again and see if I can tell you more in 20-25 minutes.
-
"Knabench" - where i found it ???
ií cannont found any download ...
-
First, it looks llike the app is trying to find WBEM in the DLL error.
Second, Knabench is the test and benchmark program we use in pre-release testing. There is a link to it under downloads. Test and Benchmark, but it appears that Simon has not updated it in a while. Simon, whaddya think?
EDIT....My bad....I think the Knabench in the download section IS current, was looking at the pre-release test package number, not the Knabench revision number.
-
I think I haven't had time to test it yet ;)
Sometime tonight, I'll be able to tell you more.
HTH,
Simon.
-
I just tried again. Copied the 3 DLL files into the science apps file along with the test app, still no go. Got the same WBEM dll error. That's alll I have time to try before I go to work today, I'll check when I get home late tonight, and if y'all have found a fix, I will do a test run on my FX60 against the patched -PM app I am currently running.
-
I've tested it ... CPU PD@3.6 GHz
AMD client: <wu_cpu_time>243.687500</wu_cpu_time>
2.2B : <wu_cpu_time>179.421875</wu_cpu_time>
-
I've tested it ... CPU PD@3.6 GHz
AMD client: <wu_cpu_time>243.687500</wu_cpu_time>
2.2B : <wu_cpu_time>179.421875</wu_cpu_time>
thanx cruncher.and validation was ok ?
next version has : malloc,vectorized powerspectrum,and vectorized memcopy (APL)....
would be uploaded tomorrow about 1500(GMT+1)
any results from AMD based machine ?
-
I would test it on my FX60 rig if you or Simon can advise me how to get it to run without the DLL error I posted this morning.
-
I would test it on my FX60 rig if you or Simon can advise me how to get it to run without the DLL error I posted this morning.
1. download ---> http://www.dependencywalker.com/depends22_x86.zip
2. extract the program
3. run "depends.exe" and open "seti_5.17_devaster_ACML_GFX.exe"
4. after that it'll show you what dlll's are missing to run thhe app.
-------------------------------------------------------------------------------------------
P.S. Personally this is my opinion ......
ACML is to way slow .... the whole thing was tried back while we had non_enhanced seti ... IPP is way faster even using AMD CPU's it's not worth the hassle.
The thing to have a look at is to use IPP or MKL for powerspectrum an Transpose.
Anything else is not even worth a try if you're going for speed.
-
I have already downloaded and unzipped the program. What or where is 'depends.exe'? I simply cut and pasted the app into Knabench and get the DLL error. I get a DLL error when I just try to run the app by double clicking it in the temp folder. MSVCR80.dll could not be found.
-
my fault ... I always forgot add a MS VC runtime .... :-[
in last APL (comparable with IPP) is a fft too so i try this version ...
a question : how many threads wanna use in APL ? one or more ?
today morning i have a 4 hours power outage, so sorry i have do only a vectorized powerspectrum ....
-
now AMD version is veeeery slow in search pulses and triplets ... say CodeAnalyst.
iam trying maximal vectorize it....
and transpose functions too.
maybe it would be better when a MS c++ compiler can unroll the loops ???
-
If only you can make it faster than generic AMD (Barton core) SSE version client... I`d be very happy!
-
for now i have stopped any work on this version beacuse i am working on gpu version ....