GPU GMP-ECM to higher input limit

[wombatman]

I've been working on removing the 2^1018 limit on GMP-ECM's GPU implementation, and I think I've gotten it done. I'm not going to completely double-post, but I've put more details about how I've been testing it in this thread: http://mersenneforum.org/showthread.php?t=16480&page=23 (specifically, posts #253, 254, and 255 starting at the bottom of page 23).

I would appreciate some of the more knowledgeable mathematicians and programmers on here trying out the file uploaded in post 253 and letting me know if you get similar results. At this point, it should be able to handle up to 2^4096-1 as an input. I've personally tried as high as 2^4000-1.

Thanks.

[wombatman]

Not sure if anybody is actually interested or not, but I've got the limit to 2^8186-1 now. Still working alright.

[RichD]

Is there a 64-bit Linux version available for trial?

[wombatman]

If you're able to compile it yourself on Linux, I can tell you the (small) changes I made. Unfortunately, the only Linux box I have has a CC 1.3 Nvidia card in it, so I don't know if it's able to compile at CC 2.0 or higher.

Edit: Also, a question for those who understand ECM fair better than I do: the sigma in the GPU version is only randomly chosen as a starting point. Subsequent sigmas are incremented from this initial random value. Is that a problem in terms of the probability of finding factors of a given size?

[LaurV]

It doesn't matter. Like when you play lottery and pick 10 numbers in a row, or 10 random numbers, your chances are the same, it is 10 over the total numbers possible. The random selection in the beginning helps to have different people run different curves therefore avoid duplication of work when more people are ECM-ing the same number, so it has to be kept.

[wombatman]

Just what I needed to know. Thanks!