GFN factoring with mmff-gfn | Reservations

[Batalov]

I'll start with GFN7 and 11 for n>=201 up to limit of mmff-gfn v.0.28 (252 bits).

You will find two new tables above (GFN7 between GFN6 and GFN10, and GFN11).
I will update current known limits for n<=200 for these ranges when I will find them exactly; for now, it is known from Keller's table thus:

         n         Limit for k  
       26-50          2*10^12 
       51-100         1*10^12  
      101-200       100*10^9 
      201-300        15*10^9

[Batalov]

The only change needed to build the mmff-gfn7.exe and mmff-gfn11.exe is:

append the following to mmffgfn-0.28/src/tf_gfn.h :

#if (BASE == 7)
#define BASE_n1 4
#define BASE_preinit a.d2 = 0; a.d1 = 0x1E39; a.d0 = 0xA5057D81
#define BASE_n2 5
#define BASE_preinit2 a.d4 = a.d3 = 0; a.d2 = 0x3918FA8; a.d1 = 0x303C3358; a.d0 = 0x6E913B01
#define BASE_n3 6
#define BASE_preinit3 a.d7 = a.d6 = 0; a.d5 = 0xCBC21; a.d4 = 0xFE4561C8; a.d3 = 0xD63B78E7; a.d2 = 0x80E1341E; a.d1 = 0x199417C8; a.d0 = 0xC0BB7601
#endif

#if (BASE == 11)
#define BASE_n1 4
#define BASE_preinit a.d2 = 0; a.d1 = 0xA33F09; a.d0 = 0x2E0B1AC1
#define BASE_n2 5
#define BASE_preinit2 a.d4 = 0; a.d3 = 0x6819; a.d2 = 0x55362721; a.d1 = 0xCE00DDAA; a.d0 = 0xAB61C581
#define BASE_n3 6
#define BASE_preinit3 a.d7 = 0; a.d6 = 0x2A5497BD; a.d5 = 0xC0BC1D6B; a.d4 = 0xEB50C360; a.d3 = 0xC8E92890; a.d2 = 0x618FB0CC; a.d1 = 0x7B96CC85; a.d0 = 0x9A21CB01
#endif

[flashjh]

I can build, please refresh my memory on how to build them...

[Batalov]

Originally Posted by flashjh

I can build, please refresh my memory on how to build them...

Just like mfaktc (or like mmff). But first take the code from http://www.mersenneforum.org/mmff-gfn/ and add the patch (two messages above) to mmffgfn-0.28/src/tf_gfn.h .

The only difference is that you have to make many .exe files by changing (at the top of the same file mmffgfn-0.28/src/tf_gfn.h ) "#define BASE 3" to 5, 6, 7, 10, 11, and 12 (and always clean up all object files in between builds).

[Batalov]

Because I had not found any GFN₇s just yet (while with much less work I'd found 3 GFN₁₁s), I had of course built an extensive unit test. All of these find the expected factors just fine:

FermatFactor=25,620719e9,620720e9
FermatFactor=32,951e9,952e9
FermatFactor=44,205953e6,205954e6
FermatFactor=45,50766e6,50767e6
FermatFactor=46,580e6,581e6
FermatFactor=49,3867e6,3868e6
FermatFactor=52,30e6,31e6
FermatFactor=61,458e6,459e6
FermatFactor=66,63975e6,63976e6
FermatFactor=68,19250e6,19251e6
#FermatFactor=82,97,98 #--finds the factor but then stops
FermatFactor=91,22066e6,22067e6
FermatFactor=93,933772e6,933773e6
FermatFactor=96,135404e6,135405e6
FermatFactor=102,250e6,251e6
FermatFactor=106,26949e6,26950e6
FermatFactor=110,3e5,4e5
FermatFactor=121,18e6,19e6
FermatFactor=141,51e6,52e6
FermatFactor=144,92500e6,92501e6
FermatFactor=148,94283e6,94284e6
FermatFactor=152,10139e6,10140e6
FermatFactor=167,14375e6,14376e6
FermatFactor=180,224e6,225e6
FermatFactor=181,1160e6,1161e6
FermatFactor=217,71600,71800

[henryzz]

It strikes me that it should be possible to automate that sort of patch so that only one binary is needed and new bases don't need to be added in the future.

[Batalov]

It can be automated largely, but only as a code generator (not run-time; you'd want the constants to be in kernels for init, not fetched from global variables). The fat binary would simply drag along seven times more kernels than needed all the time. And of course the "worktodo.txt" definition would need to be extended.

[flashjh]

Well, that took a long time, sorry, no excuse on building just got busy and forgot :(

All builds 32 and 64 bit CUDA 8.0

Tests went well except for the new 7 and 11. 7 had trouble with some errors and 11 didn't find any factors in the list, but I may have misread the testing post. If you have some test lines for 11, let me know. Otherwise, I think 7 needs some work or the testing is a bit off.

So, the 7 and 11 need testing before actual use.

I included the tests and cudart file and the files I used to compile with updates makefiles for new versions

Let me know about the 7 and 11, if I need to update anything and rebuild, let me know

[flashjh]

Anyone had a chance to test the new builds, do they work as expected?