Mfakto cl_barret15_73_gs_2 - Not A Factor? - Page 2

chalsall · 2015-12-17, 16:37

Quote:

Originally Posted by Mark Rose

...it might be worth spending time redoing certain classes of work.

Then again, it might not. As LaurV et al said, false positives are not really a big deal -- they're immediately checked by Primenet (that calculation is inexpensive).

False negatives /might/ be a bigger deal, for those anal about finding all low factors of candidates.

At worst case a low factor is missed so additional work is expended on LL'ing and DC'ing. Where do the economic lines cross between re-doing previous TF'ing vs. doing new TF'ing?

But, without question, this does strongly suggest a bug in mfakto.

TheJudger · 2015-12-17, 19:52

Might be a cornercase, I'm not sure. The number is interesting at least.

FC = factor candidate

What we NOT do is to check if 2^exp - 1 == 0 mod FC
What we do is to check if 2^exp == 1 mod FC

For exp = 47079733 and FC = 2742459209866315019473
we have the following numbers:
2^47079733 = 2742459209866315019472 = -1 mod 2742459209866315019473

So might be an issue with checking the final result of the calculation.

Oliver

airsquirrels · 2015-12-17, 20:54

I misspoke on seeing this before with a different number. Turns out that since that result was rejected by primenet, my checks against my GPU72 assignments for missing work kept pulling up that candidate as needing checked again. So I've actually run this on 4 different GCN cards with mfakto and only finally the Titan Z with mfaktc cleared it.

chalsall · 2015-12-17, 21:02

Quote:

Originally Posted by airsquirrels

So I've actually run this on 4 different GCN cards with mfakto and only finally the Titan Z with mfaktc cleared it.

Hmmmm... Interesting....

airsquirrels · 2015-12-17, 21:30

Quote:

Originally Posted by TheJudger

Might be a cornercase, I'm not sure. The number is interesting at least.

FC = factor candidate

What we NOT do is to check if 2^exp - 1 == 0 mod FC
What we do is to check if 2^exp == 1 mod FC

For exp = 47079733 and FC = 2742459209866315019473
we have the following numbers:
2^47079733 = 2742459209866315019472 = -1 mod 2742459209866315019473

So might be an issue with checking the final result of the calculation.

Oliver

Seems like in that case signed vs. unsigned would make this an easy bug that would only manifest as false positives, so the risk is negligible. We should of course verify that is in fact what happened.

chalsall · 2015-12-17, 21:42

Quote:

Originally Posted by airsquirrels

Seems like in that case signed vs. unsigned would make this an easy bug that would only manifest as false positives, so the risk is negligible

I would argue that "TheJudger" is a very good friend to have on your side. But, also, that "Bdot" needs to be brought into this; he is also a very good friend to have on your side.

Quote:

Originally Posted by airsquirrels

We should of course verify that is in fact what happened.

Of course. Trust, but verify.

TheJudger · 2015-12-17, 23:08

Quote:

Originally Posted by airsquirrels

Seems like in that case signed vs. unsigned would make this an easy bug that would only manifest as false positives, so the risk is negligible. We should of course verify that is in fact what happened.

No, I'm affraid it won't be that simple. If it would be really that simple mfakto should report lots of (false) mersenne factors. This is what mfaktc does when it searches for Wagstaff factors.

Best would be to trace the data while running mfakto. I've looked at the interim results of each step when doing with libgmp on CPU... nothing "special" execpt the final reminder.

Oliver

Bdot · 2016-01-04, 18:26

kracker made me aware of this thread. I will debug this when I'm back next week.

LaurV · 2016-01-05, 04:14

Quote:

Originally Posted by Bdot

kracker made me aware of this thread. I will debug this when I'm back next week.

If you do, you may also have a look to the "rip dctf" thread and see what they are talking there about not using global memory, or so, giving a boost of about 30% in speed to mfakto.

kracker · 2016-01-05, 04:25

Quote:

Originally Posted by LaurV

If you do, you may also have a look to the "rip dctf" thread and see what they are talking there about not using global memory, or so, giving a boost of about 30% in speed to mfakto.

It depends on the GPU... the faster/the more GPU's you have.. the bigger improvement(so I hear).

Also....

Bdot · 2016-01-06, 16:14

Quote:

Originally Posted by LaurV

If you do, you may also have a look to the "rip dctf" thread and see what they are talking there about not using global memory, or so, giving a boost of about 30% in speed to mfakto.

Not only did you make me look for the thread by not providing a link, you also wanted to make me read 300+ posts? I came up to the 500 vamps biting your neck ... it's already 1am over here and have to get some sleep.

Quote:

Originally Posted by kracker

It depends on the GPU... the faster/the more GPU's you have.. the bigger improvement(so I hear).

Also....

Of course

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto error while running	jacob1235	GPU Computing	1	2018-02-03 19:03
TF with mfakto win 0.14	Yoshi24517	GPU Computing	5	2017-04-01 18:01
mfakto 0.13 with Windows 8.1	bplenhart	GPU Computing	16	2013-12-05 01:24
trouble with submitting mfakto factor	dbaugh	PrimeNet	5	2013-04-30 22:18
How to tell mfakto to use IGP?	Rodrigo	GPU Computing	22	2013-03-23 02:25

2015-12-17, 19:52	#13
TheJudger "Oliver" Mar 2005 Germany 11·101 Posts	Might be a cornercase, I'm not sure. The number is interesting at least. FC = factor candidate What we NOT do is to check if 2^exp - 1 == 0 mod FC What we do is to check if 2^exp == 1 mod FC For exp = 47079733 and FC = 2742459209866315019473 we have the following numbers: 2^47079733 = 2742459209866315019472 = -1 mod 2742459209866315019473 So might be an issue with checking the final result of the calculation. Oliver

2015-12-17, 20:54	#14
airsquirrels "David" Jul 2015 Ohio 517₁₀ Posts	I misspoke on seeing this before with a different number. Turns out that since that result was rejected by primenet, my checks against my GPU72 assignments for missing work kept pulling up that candidate as needing checked again. So I've actually run this on 4 different GCN cards with mfakto and only finally the Titan Z with mfaktc cleared it.

2016-01-04, 18:26	#19
Bdot Nov 2010 Germany 1001010101₂ Posts	kracker made me aware of this thread. I will debug this when I'm back next week.