CUDALucas (a.k.a. MaclucasFFTW/CUDA 2.3/CUFFTW) - Page 189

LaurV · 2013-12-02, 17:24

Did 37500769 few times, it is already "multiple" checked

(I changed the version of cL to 2.05 on the way to a triple test, first was mismatch, then I realized that it didn't resume the work, but started from scratch. As the other test was almost finish, I decided to go back to 2.04 and finish both of them. So, now I got the same residue 3 times, so mine is good, for sure! Bookmarked, waiting for some P95 checkers to confirm...

The release 50 seems to work ok. I like the interactive options, and the thread tuning. The new drivers are a bit slower for older cards (gtx 580), but this is compensated by the tuning features and other small nice things...

Good job!

flashjh · 2013-12-02, 17:34

When I get some time, I'm going to try and compile from other CUDA versions to see if that helps with speed at all. I've gotten away from tracking the exact iteration times because I want to get the program stable. Maybe I'll worry about the speed when I'm racing to beat you all for DCing the next MP

So far when testing 2.05 the mismatches already had two LLs done, so my DC matched one. I completed 30612941 yesterday and it was a mismatch. petrw1 picked up the assignment for DC, so I'll have to wait until it's done. I could run it for a TC, maybe tomorrow? For 2.05s sake, hopefully the DC matches mine.

owftheevil · 2013-12-02, 17:56

Quote:

Originally Posted by Prime95

We have 3 options that I see:

1) Disallow CUDALucas from double-checking CUDALucas results (the status quo).
2) Allow double-checks as long as the shift counts are different. The downside: it is real easy to forge a double-check.
3) Add a security code to the CUDALucas final result output (a simple hash of the exponent and shift count). This code could be secret, which isn't allowed if CUDALucas is GPL. Those building executables would be entrusted with the secret code. Or, the code can be public. At least a forger of results has to go to the trouble of reading C code.

There is an optional add-on to options 2 and 3. The Primenet server and/or GPU72 server can be upgraded to only give double-check exponents that were first tested by prime95.

The code that generates the security code could be a separate application.

kracker · 2013-12-02, 18:26

Quote:

Originally Posted by LaurV

Did 37500769 few times, it is already "multiple" checked

Yay... I somehow got it. I'll probably toss it on my 7770.

LaurV · 2013-12-02, 18:44

Quote:

Originally Posted by kracker

Yay... I somehow got it. I'll probably toss it on my 7770.

If you do that, you will get no credit. PrimeNet will not accept a result for this expo, if it does not come from P95. That is why all the discussions about shifts and secret CRCs. In the past we had a thread for this type of exponents (which were DC-ed and TC-ed with CudaLucas, and still gave different residues from the original test), to warn the potential crunchers that they must use P95 for them, otherwise they waste the time. I even was moderator and did the maintenance for that thread, but since the rights were restricted, the thread is forgotten and I can not find it.

kracker · 2013-12-02, 18:52

Quote:

Originally Posted by LaurV

If you do that, you will get no credit. PrimeNet will not accept a result for this expo, if it does not come from P95. That is why all the discussions about shifts and secret CRCs. In the past we had a thread for this type of exponents (which were DC-ed and TC-ed with CudaLucas, and still gave different residues from the original test), to warn the potential crunchers that they must use P95 for them, otherwise they waste the time. I even was moderator and did the maintenance for that thread, but since the rights were restricted, the thread is forgotten and I can not find it.

I use clLucas. I guess it is same as CudaLucas or considered different?

EDIT: "if it does not come from P95" Sorry didn't see that. Tossing the exponent on a CPU.
EDIT2: This?

LaurV · 2013-12-02, 19:00

Quote:

Originally Posted by kracker

EDIT: "if it does not come from P95" Sorry didn't see that. Tossing the exponent on a CPU.

Yes, one of the checks, either the first LL or the DC, must come from P95, there is no other way, because up to now (i.e. before cudalucas 2.05, done last week) only P95 implemented the random shift. All other programs used shift zero (clLucas included) and in case of a FFT error, they will all get the same (wrong) residue.

Quote:

EDIT2: This?

Yes, thanks, didn't really look (1.55 A.M. here!), had bookmark, which "expired", hehe. I will actualize the bookmark. You should read it, it has some basic theory which will help you understand the shifting process

flashjh · 2013-12-02, 19:17

Quote:

Originally Posted by owftheevil

The code that generates the security code could be a separate application.

Would it be a program called by CUDALucas or could it be in a .dll file? Do you have something in mind?

owftheevil · 2013-12-02, 19:26

Quote:

Originally Posted by flashjh

Would it be a program called by CUDALucas or could it be in a .dll file? Do you have something in mind?

Could be either, or something else, and yes I have something in mind.

kladner · 2013-12-03, 00:21

FWIW, I have run one DC on my GTX 580, and two on the 570. All have matched. All were single first-time LL's. These runs included both intentional restarts, and program crash restarts. Various interactive features were invoked, and I played with clock speeds for both core and VRAM in the course of the tests.

chappjc · 2013-12-03, 01:21

I haven't been around for a while, but I recently started TF'ing with mfaktc on a 580, 670 and C2075. Now, with the latest updates to CUDALucas 2.05, I am actually able to run it on these GPUs (although the 580 has random runtime API errors if FFT length is not tweaked).

Are results from CUDALucas 2.05 accepted at PrimeNet? I completed an exponent in a few days, and the system didn't find any CUDALucas lines in the results.txt, but there was a line there indicating version 2.05.

Also, are there any tests I should run and report here?

2013-12-02, 17:34	#2070
flashjh "Jerry" Nov 2011 Vancouver, WA 463₁₆ Posts	When I get some time, I'm going to try and compile from other CUDA versions to see if that helps with speed at all. I've gotten away from tracking the exact iteration times because I want to get the program stable. Maybe I'll worry about the speed when I'm racing to beat you all for DCing the next MP So far when testing 2.05 the mismatches already had two LLs done, so my DC matched one. I completed 30612941 yesterday and it was a mismatch. petrw1 picked up the assignment for DC, so I'll have to wait until it's done. I could run it for a TC, maybe tomorrow? For 2.05s sake, hopefully the DC matches mine. Last fiddled with by flashjh on 2013-12-02 at 17:55

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Don't DC/LL them with CudaLucas	LaurV	Data	131	2017-05-02 18:41
CUDALucas / cuFFT Performance on CUDA 7 / 7.5 / 8	Brain	GPU Computing	13	2016-02-19 15:53
CUDALucas: which binary to use?	Karl M Johnson	GPU Computing	15	2015-10-13 04:44
settings for cudaLucas	fairsky	GPU Computing	11	2013-11-03 02:08
Trying to run CUDALucas on Windows 8 CP	Rodrigo	GPU Computing	12	2012-03-07 23:20

2013-12-02, 17:24	#2069
LaurV Romulan Interpreter Jun 2011 Thailand 7²×197 Posts	Did 37500769 few times, it is already "multiple" checked (I changed the version of cL to 2.05 on the way to a triple test, first was mismatch, then I realized that it didn't resume the work, but started from scratch. As the other test was almost finish, I decided to go back to 2.04 and finish both of them. So, now I got the same residue 3 times, so mine is good, for sure! Bookmarked, waiting for some P95 checkers to confirm... The release 50 seems to work ok. I like the interactive options, and the thread tuning. The new drivers are a bit slower for older cards (gtx 580), but this is compensated by the tuning features and other small nice things... Good job!

2013-12-03, 00:21	#2078
kladner "Kieren" Jul 2011 In My Own Galaxy! 2·3·1,693 Posts	FWIW, I have run one DC on my GTX 580, and two on the 570. All have matched. All were single first-time LL's. These runs included both intentional restarts, and program crash restarts. Various interactive features were invoked, and I played with clock speeds for both core and VRAM in the course of the tests.

2013-12-03, 01:21	#2079
chappjc Jul 2007 2²·5 Posts	I haven't been around for a while, but I recently started TF'ing with mfaktc on a 580, 670 and C2075. Now, with the latest updates to CUDALucas 2.05, I am actually able to run it on these GPUs (although the 580 has random runtime API errors if FFT length is not tweaked). Are results from CUDALucas 2.05 accepted at PrimeNet? I completed an exponent in a few days, and the system didn't find any CUDALucas lines in the results.txt, but there was a line there indicating version 2.05. Also, are there any tests I should run and report here?