PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise
11) Message boards : Project Staging Area : AVX build of llr (20%-50% faster) (Message 46605)
Posted 3385 days ago by Profile Microcruncher*Project donor
hmm I'm feeling stupid here. I can't seem to get this working with PRPnet. -- Atleast, I have zero change/improvement.

I am downloading (i think) the correct llr.exe file, and i put it into my individual prpnet sub folder, and run the task, and trying to benchmark on PPSElow...

before and after, I'm getting roughly 52 seconds per WU. (0.170ms per bit)

i7-3930k @ 4.2Ghz hyperthreading on.


I haven't tried any experiments in BOINC with app_info.xml yet..

Is it this box?

ID: 235836

Microsoft Windows 7
Ultimate x64 Edition, (06.01.7600.00)

Windows 7 SP 1 is required for AVX support.
12) Message boards : Project Staging Area : AVX build of llr (20%-50% faster) (Message 46519)
Posted 3386 days ago by Profile Microcruncher*Project donor
One thing I have noticed over the last week-plus with the AVX version is increased heat output... the AVX version runs 2C-3C hotter than the stock app (at least on my box, under full load of 8 LLR threads on a 2600K). If you have been running o/c'd to your thermal limits before, you should definitely check your temperatures now under full load. That being said, my box doesn't have the greatest ventilation, so perhaps it is a localized phenomenon.

--Gary

Thank you very much for your work.

I see the 2-3°C temperature increase too (4 LLR threads on a 2500K).
13) Message boards : Number crunching : The Last Days of 2011 (Message 45929)
Posted 3397 days ago by Profile Microcruncher*Project donor
29 k=5/7 WUs in the queue including 10 resends from around n=3.3M and an additional 71 WUs from the k=100-500 range.

11 k=5/7 WUs in the queue including 9 resends from around n=3.3M and 89 k=100-500 WUs. 4 of the k=5/7 WUs will finish in the next hour.

Update:

While the resends have higher than expected runtimes (100 minutes instead of 80), the k=100-500 WUs are much faster than expected (Less than 14 minutes instead of 36).
14) Message boards : Number crunching : The Last Days of 2011 (Message 45874)
Posted 3398 days ago by Profile Microcruncher*Project donor
How does one determine which tasks have which k's?

I looked into the WU files on my box.

cat pps_llr_113908733 1000000000:P:1:2:257 7 4776716

Line two shows the k and the n values.

Entries for validated workunits also show the values. Here is an example:

http://www.primegrid.com/workunit.php?wuid=236811821

Is prime? 7*2^4778450+1 is not prime.
15) Message boards : Number crunching : The Last Days of 2011 (Message 45868)
Posted 3398 days ago by Profile Microcruncher*Project donor
Back to topic:

67 k=5/7 WUs done and validated (n> 4.3M), another 17 are pending.

29 k=5/7 WUs in the queue including 10 resends from around n=3.3M and an additional 71 WUs from the k=100-500 range.
16) Message boards : Number crunching : The Last Days of 2011 (Message 45859)
Posted 3398 days ago by Profile Microcruncher*Project donor
Gwnum is hand-written, expertly-tuned, assembly language code, and yes, it's gwnum that does all the heavy lifting. It detects your processor type and makes use of the proper SIMD instructions, depending on what's available on your CPU.

Re-compiling anything that uses the gwnum software won't have much of an affect on speed because you're not compiling the parts that affect the speed.

To make use of the AVX instructions, you would need to re-write gwnum.

Which I suspect is what George Woltman either is doing or is planning on doing.

So, the best way to get AVX support is to wait until Prime95 has it, after which you can incorporate the new gwnum in other software that uses it.

The source code for gwnum, should you decide to tackle this yourself, is part of the source code for Prime95; it's not a separate library. Again, however, it's not a compiler-switch thing; you have to write the assembly language for the code, using the new AVX instructions, yourself. Got your Sandy Bridge instruction timing manual handy? ;-)

Again: Prime95 / mprime / gwnum-lib has working AVX support (at least for 32 bit).

ftp://mersenne.org/gimps/source272.zip

That's why I initially asked :D
17) Message boards : Number crunching : The Last Days of 2011 (Message 45855)
Posted 3398 days ago by Profile Microcruncher*Project donor
It actually ran! But, the runtimes between the executable that I built and the one provided with the official installation package were basically identical. No benefit.

The gwnumlib version contained in llr386src.zip doesn't have AVX support ;)
18) Message boards : News : World Record Twin Primes (Message 45672)
Posted 3401 days ago by Profile Microcruncher*Project donor
Congratulations!
19) Message boards : Number crunching : The Last Days of 2011 (Message 45649)
Posted 3402 days ago by Profile Microcruncher*Project donor
There is an AVX c++ compiler available free, I think the sourcecode of all PG programming is in C++ or C so it may be possible to compile the code as is for the AVX with few if any modifications.

I am not going to try it, my C++ is too rusty.

Prime95 27.x / mprime 27.2 has working (32 bit) AVX support (speed gains of up to 25% on Sandy Bridges). I'm in the process of downloading the sources to build an accelerated gmp-ecm version (6.3.1SVN + mpir 2.5.0SVN + gwnum-lib 27.2) for k*2^n+-c numbers.*

I was just curious if anyone has already tried to build an AVX enabled LLR version.

*and probably a few more types with the new ecm-sources ;) I do not despise speed gains of up to 5x compared with a non gwnum-lib ecm build.
20) Message boards : Number crunching : The Last Days of 2011 (Message 45647)
Posted 3402 days ago by Profile Microcruncher*Project donor
Please have a look in the llr-sources from http://jpenne.free.fr/llr3/llr386src.zip.
Are you able to compile the linux64-version?

Do you really think, we need a 64 bit version? I think 32 bit AVX compilation should be done first...

Again: Has anyone tried build an AVX enabled version of LLR?


Next 10 posts
[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2021 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 2.90, 3.73, 4.04
Generated 19 Apr 2021 | 4:10:01 UTC