mfaktc: a CUDA program for Mersenne prefactoring

[TheJudger]

Hello!

Find attached mfaktc 0.11.
Again, thank you Luigi, Dave, Kevin and Ethan for testing, comments, ...!

---

It is faster than the previous version:
- the 75 bit kernel is ~2% faster
- the sieve is 20% faster (measured on Intel Core i7)

The new debug code (you usually don't have to care about the debug code, it is used only during development) revealed on computational bug.

mfaktc compiled for compute capability 1.x:

factor size    | chance to miss a factor
---------------+------------------------
2^24 to 2^24.2 | < 0.1%
2^56 to 2^56.2 | < 0.1%
2^88 to 2^88.2 | < 0.1%
other ranges   | < 0.001%

mfaktc compiled for compute capability 2.x:

factor size    | chance to miss a factor
---------------+------------------------
2^24 to 2^24.2 | < 0.1%
2^56 to 2^56.2 | < 0.1%
2^88 to 2^88.2 | < 0.1%
other ranges   | very small if not 0%

The ranges around 2^24, 2^56 and 2^88 occur only in the 75 and 95 bit kernel when the most significant "digit" is in the range 2^24 to 2^24.2 (this can't happen in the 71 bit kernel). I think that these ranges don't affect typical mfaktc jobs so this issue isn't critical, right?

The "other ranges" issue has a very low chance so I would say it is not critical, too. This affects all kernels.

< 0.1% means that out of 1000 factors "less than" one factor isn't detected in average.

This issues are related to floating point accuracy. I was a bit too greedy on quotient estimate during long division.

---
Note on upgrading: mfaktc refuses to read checkpoint files from other versions than its own version. Continue your work with your current version until you've finished a bit level or exponent.

Oliver

[Karl M Johnson]

Now, who can compile that for Windows?

[ckdo]

Originally Posted by Karl M Johnson

Now, who can compile that for Windows?

Win64, compute capability 1.x, puh-leeze...

[TheJudger]

Hi Karl, ckdo,

for now I can't provide precompiled Windows binaries... but I'm working on this.

I'm still unsure which Windows version I should choose.
Windows 7 64bit... so far, so good... but which variant?
I don't want "desktop effects"! E.g. when I click on a minimized windows I don't want an animation the window just should open in virtually no time!
I want a small taskbar and small icons, too.
This is what I hate on modern GUIs...they are overloaded with stupid animations etc.

I think I might try Windows 7 "Professional". This is the "smallest" release with XP mode, doesn't install games by default and (according to wikipedia) allows the taskbar to act like in WinXP...

Any hints/suggestions?

Oliver

P.S. on Linux I'm using Xfce
P.P.S. my current Windows XP (32bit) installation runs the "classic" GUI with some modifications.. turned of all desktop effects, even mouse shadow...
I like it simple and fast!

[ET_]

Originally Posted by TheJudger

Hi Karl, ckdo,

for now I can't provide precompiled Windows binaries... but I'm working on this.

I'm still unsure which Windows version I should choose.
Windows 7 64bit... so far, so good... but which variant?
I don't want "desktop effects"! E.g. when I click on a minimized windows I don't want an animation the window just should open in virtually no time!
I want a small taskbar and small icons, too.
This is what I hate on modern GUIs...they are overloaded with stupid animations etc.

I think I might try Windows 7 "Professional". This is the "smallest" release with XP mode, doesn't install games by default and (according to wikipedia) allows the taskbar to act like in WinXP...

Any hints/suggestions?

Oliver

P.S. on Linux I'm using Xfce
P.P.S. my current Windows XP (32bit) installation runs the "classic" GUI with some modifications.. turned of all desktop effects, even mouse shadow...
I like it simple and fast!

If you plan a GUI, you may give an eye to Nokia Qt development system: based on GCC and Mingw, offers cross-compatibility with Linux, MAC and Windows to source level.

Luigi

[TheJudger]

Hi Luigi,

Originally Posted by ET_

If you plan a GUI, you may give an eye to Nokia Qt development system: based on GCC and Mingw, offers cross-compatibility with Linux, MAC and Windows to source level.

Luigi

false hope, I don't plan a GUI for mfaktc.
I was talking about the GUIs of Windows / Linux / whatever.

Oliver

[Karl M Johnson]

Oliver, If you're working on compiling mfaktc on windows that's good news!
IMO, first get it compiling for windows, then make a GUI for it.
Hm, actually, does it need a GUI ?
Easy to learn application

[ET_]

Originally Posted by TheJudger

Hi Luigi,



false hope, I don't plan a GUI for mfaktc.
I was talking about the GUIs of Windows / Linux / whatever.

Oliver

I see...

Luigi

[lavalamp]

Originally Posted by TheJudger

I'm still unsure which Windows version I should choose.
Windows 7 64bit... so far, so good... but which variant?
I don't want "desktop effects"! E.g. when I click on a minimized windows I don't want an animation the window just should open in virtually no time!
I want a small taskbar and small icons, too.
This is what I hate on modern GUIs...they are overloaded with stupid animations etc.

I think I might try Windows 7 "Professional". This is the "smallest" release with XP mode, doesn't install games by default and (according to wikipedia) allows the taskbar to act like in WinXP...

You'll find that there are a multitude of settings to alter the GUI you can pick and choose from, these can be as general as changing the whole theme to "classic", and as individual as disabling that minimise animation. I've attached a screen shot with these options highlighted.

I'm using Windows Vista, but the options will no doubt be similar for Windows 7.

[lavalamp]

Here's the shot of the same options in Windows 7. I should also point out that you can do all of the other stuff you said you wanted to do, like changing to the classic start menu, in any version of windows.

[TheJudger]

Thank you lavalamp!
The classic schema look familiar to me

---
In my 0.12 development version I've changed the statistic line a little bit:

old style:

class 0: tested 105633251328 candidates in 2243766ms (47078550/sec) (avg. wait: 11115usec)

new style:

class 0: tested 105.63G candidates in 2244s (47.08M/sec) (avg. wait: 11115usec)

The number of candidates changes between M and G, depending on the number.
The time changes between milliseconds (time is less than 100s) and seconds (time greater or equal 100s)

Which variant do you prefer? (I think I won't make this a runtime option!)


Oliver