Trial division with CUDA (mmff) -- used, but runs like new!

[ATH]

No, this is the new GPUSievePrimes auto-select feature causing it, because if I set it manually, the bug disappears, and I didn't have it in version 0.23 on the same known fermat factors.

[ATH]

Here is the worktodo.txt to test the 21 known fermat factors that are within mmff's "search space" along with how the results.txt should look. I recommend setting PrintMode=1 in mmff.ini when you run this to avoid all the spam, and until version 0.25 is out you need to set GPUSievePrimes to something like 200000 to avoid the auto-select feature. This worktodo.txt takes 1min20sec on a GTX 460.

worktodo.txt:
FermatFactor=36,25709e6,25710e6
FermatFactor=33,5460e9,5470e9
FermatFactor=39,69,70
FermatFactor=45,11131e10,11132e10
FermatFactor=45,212e9,213e9
FermatFactor=50,2139e9,2140e9
FermatFactor=54,78,79
FermatFactor=54,81900e9,81911e9
FermatFactor=74,100,101
FermatFactor=79,5e9,6e9
FermatFactor=87,1595e9,1596e9
FermatFactor=88,20018e9,20019e9
FermatFactor=90,119e9,120e9
FermatFactor=92,198e9,199e9
FermatFactor=97,482e9,483e9
FermatFactor=101,3334e9,3335e9
FermatFactor=111,141,142
FermatFactor=120,3e9,4e9
FermatFactor=135,880e8,881e8
FermatFactor=148,173,174
FermatFactor=149,175,176

results.txt
F28 has a factor: 1766730974551267606529 [TF:70:71:mmff 0.24 mfaktc_barrett89_F32_63gs]
found 1 factor for k*2^36+1 in k range: 25709M to 25710M (71-bit factors) [mmff 0.24 mfaktc_barrett89_F32_63gs]
F31 has a factor: 46931635677864055013377 [TF:75:76:mmff 0.24 mfaktc_barrett89_F32_63gs]
found 1 factor for k*2^33+1 in k range: 5460G to 5470G (76-bit factors) [mmff 0.24 mfaktc_barrett89_F32_63gs]
F37 has a factor: 701179711390136401921 [TF:69:70:mmff 0.24 mfaktc_barrett89_F32_63gs]
found 1 factor for k*2^39+1 in k range: 1073741824 to 2147483648 (70-bit factors) [mmff 0.24 mfaktc_barrett89_F32_63gs]
F42 has a factor: 3916660235220715932328394753 [TF:91:92:mmff 0.24 mfaktc_barrett96_F32_63gs]
found 1 factor for k*2^45+1 in k range: 111310G to 111320G (92-bit factors) [mmff 0.24 mfaktc_barrett96_F32_63gs]
F43 has a factor: 7482850493766970889994241 [TF:82:83:mmff 0.24 mfaktc_barrett89_F32_63gs]
found 1 factor for k*2^45+1 in k range: 212G to 213G (83-bit factors) [mmff 0.24 mfaktc_barrett89_F32_63gs]
F48 has a factor: 2408911986953445595315961857 [TF:90:91:mmff 0.24 mfaktc_barrett96_F32_63gs]
found 1 factor for k*2^50+1 in k range: 2139G to 2140G (91-bit factors) [mmff 0.24 mfaktc_barrett96_F32_63gs]
F52 has a factor: 389591181597081096683521 [TF:78:79:mmff 0.24 mfaktc_barrett89_F32_63gs]
found 1 factor for k*2^54+1 in k range: 16777216 to 33554432 (79-bit factors) [mmff 0.24 mfaktc_barrett89_F32_63gs]
F52 has a factor: 1475547810493913550438096961537 [TF:100:101:mmff 0.24 mfaktc_barrett108_F32_63gs]
found 1 factor for k*2^54+1 in k range: 81900G to 81911G (101-bit factors) [mmff 0.24 mfaktc_barrett108_F32_63gs]
F72 has a factor: 1443765874709062348345951911937 [TF:100:101:mmff 0.24 mfaktc_barrett108_F64_95gs]
found 1 factor for k*2^74+1 in k range: 67108864 to 134217728 (101-bit factors) [mmff 0.24 mfaktc_barrett108_F64_95gs]
F77 has a factor: 3590715923977960355577974656860161 [TF:111:112:mmff 0.24 mfaktc_barrett120_F64_95gs]
found 1 factor for k*2^79+1 in k range: 5G to 6G (112-bit factors) [mmff 0.24 mfaktc_barrett120_F64_95gs]
F83 has a factor: 246947940268608417020015902258307792897 [TF:127:128:mmff 0.24 mfaktc_barrett128_F64_95gs]
found 1 factor for k*2^87+1 in k range: 1595G to 1596G (128-bit factors) [mmff 0.24 mfaktc_barrett128_F64_95gs]
F86 has a factor: 6195449970597928748332522715641578258433 [TF:132:133:mmff 0.24 mfaktc_barrett140_F64_95gs]
found 1 factor for k*2^88+1 in k range: 20018G to 20019G (133-bit factors) [mmff 0.24 mfaktc_barrett140_F64_95gs]
F88 has a factor: 148481934042154969241780501829489000449 [TF:126:127:mmff 0.24 mfaktc_barrett128_F64_95gs]
found 1 factor for k*2^90+1 in k range: 119G to 120G (127-bit factors) [mmff 0.24 mfaktc_barrett128_F64_95gs]
F90 has a factor: 985016348367230226078056532654006730753 [TF:129:130:mmff 0.24 mfaktc_barrett140_F64_95gs]
found 1 factor for k*2^92+1 in k range: 198G to 199G (130-bit factors) [mmff 0.24 mfaktc_barrett140_F64_95gs]
F94 has a factor: 76459067246115642538831634131564386844673 [TF:135:136:mmff 0.24 mfaktc_barrett140_F96_127gs]
found 1 factor for k*2^97+1 in k range: 482G to 483G (136-bit factors) [mmff 0.24 mfaktc_barrett140_F96_127gs]
F96 has a factor: 8453027931784477309850388309101819121893377 [TF:142:143:mmff 0.24 mfaktc_barrett152_F96_127gs]
found 1 factor for k*2^101+1 in k range: 3334G to 3335G (143-bit factors) [mmff 0.24 mfaktc_barrett152_F96_127gs]
F107 has a factor: 3346902437331832346018436558958369334886401 [TF:141:142:mmff 0.24 mfaktc_barrett152_F96_127gs]
found 1 factor for k*2^111+1 in k range: 1073741824 to 2147483648 (142-bit factors) [mmff 0.24 mfaktc_barrett152_F96_127gs]
F116 has a factor: 4563438810603420826872624280490561141381005313 [TF:151:152:mmff 0.24 mfaktc_barrett152_F96_127gs]
found 1 factor for k*2^120+1 in k range: 3G to 4G (152-bit factors) [mmff 0.24 mfaktc_barrett152_F96_127gs]
F133 has a factor: 3836232386548105510567872577199319351015739156856833 [TF:171:172:mmff 0.24 mfaktc_barrett172_F128_159gs]
found 1 factor for k*2^135+1 in k range: 88000M to 88100M (172-bit factors) [mmff 0.24 mfaktc_barrett172_F128_159gs]
F146 has a factor: 13235038053749721162769301995307025251972223086886913 [TF:173:174:mmff 0.24 mfaktc_barrett183_F128_159gs]
found 1 factor for k*2^148+1 in k range: 33554432 to 67108864 (174-bit factors) [mmff 0.24 mfaktc_barrett183_F128_159gs]
F147 has a factor: 88894220732640180500173831441107513117330143465963521 [TF:175:176:mmff 0.24 mfaktc_barrett183_F128_159gs]
found 1 factor for k*2^149+1 in k range: 67108864 to 134217728 (176-bit factors) [mmff 0.24 mfaktc_barrett183_F128_159gs]

[kladner]

Originally Posted by bcp19

This may have nothing to do with it, but I noticed new nVidia drivers were available recently (I have not upgraded mine yet). Could they be part of the cause?

306.23 is working fine for me so far. Performance is good.

[mognuts]

Originally Posted by Prime95

I need a bit more information to investigate this.

I'm running Windows 7 64 bit with a GTX570 card.

I have this in my mmff.ini file:

 
#GPUProgressHeader=    class | candidates |    time |    ETA | raw  rate | SievePrimes | CPU wait
#ProgressFormat=%C/4620 |    %n | %ts | %e | %rM/s |     %s |  %W%%
 
# print everything
GPUProgressHeader=[date    time]  exponent [TF bits]: percent  class #, seq    |     GHz |    time |    ETA |    #FCs |      rate | SieveP. | CPU wait | V5UserID@ComputerID
ProgressFormat=[%d %T] M%M [%l-%u]: %p%% %C/4620,%c/960 | %g | %ts | %e | %n | %rM/s | %s |  %W%% | %U@%H

.. and this in my worktodo.ini file:

MMFactor=127,4.0e15,4.3e15

The %g variable generates values such as 19071091533368053000000000000000000000.00

It doesn't matter if the %g variable is on its own i.e.

ProgressFormat=%g

or in a string with other variables such as the default string above.

All versions of mmff (from the first released one to the latest, or CUDA v4 or v5) behave simillarly.

mognuts

[Prime95]

Originally Posted by mognuts

It doesn't matter if the %g variable is on its own i.e.

ProgressFormat=%g

or in a string with other variables such as the default string above.

Ah, I must admit I'm not familiar with all of mfaktc's output options. You may not like my solution though. Since Primenet does not track these results, there are no GHz-days of credit to be had. Thus, the %g option will now output "n.a.".

[Prime95]

Here we go again -- v 0.25:

This hopefully fixes many previously reported problems:

1) The problem reported with auto-select GPUSievePrimes is fixed.
2) The "exponentiation failure" bug where a tested k value is too small for the bit-level being worked on is fixed. This was done by no longer rounding the minimum k value down to a multiple of the number of classes. As far as I can tell, this rounding down was only necessary for the self-test code. A side effect of this change is that you cannot use v0.24 save files. Finish your current work before upgrading to v0.25.
3) Exponentiation failures from testing k values that are too large should be fixed. I've also decreased the minimum acceptable k values for Fermat testing. Although inefficient, it will allow us to have more known Fermat factors retested.
4) The funny %g outputs are gone.
5) Lots of bloat from mfaktc has been removed.

The good news is none of these changes have been due to fundamental problems in the kernels doing the real work.

My recommendation is to not upgrade until ATH, flashjh, and others have had time to try this version for a little bit. They have been quite effective in verifying the quality of recent releases.

[MrRepunit]

Originally Posted by Prime95

Here we go again -- v 0.25:

I tried to compile it under Win 7 64 Bit with Visual Studio 2010. I only succeeded after changing the following line in mfaktc.c

extern int tf_class_barrett92_gs(unsigned long long int k_min, unsigned long long int k_max, mystuff_t *mystuff);

to

#ifdef _MSC_VER
extern "C" int tf_class_barrett92_gs(unsigned long long int k_min, unsigned long long int k_max, mystuff_t *mystuff);
#else
extern int tf_class_barrett92_gs(unsigned long long int k_min, unsigned long long int k_max, mystuff_t *mystuff);
#endif

[LaurV]

binaries? flashjh?

[Ralf Recker]

One tiny (and unimportant) issue:

Program output:

WARNING: Read GPUSieveSize=1 from mmff.ini, using min value (4)

From mmff.ini:

# GPUSieveSize defines how big a GPU sieve to use (in Mbits).  Bigger sieves are a little
# more efficient, but may produce laggy video response.
#
# Minimum: GPUSieveSize=1

Increasing GPUSieveSize to the max gives has no noticeable effect on the screen lag*. The raw rate increases (on a GTX 470 at 607 MHz):

got assignment: k*2^28+1, k range 1000000000000000 to 1100000000000000 (78-bit factors)
Starting trial factoring of k*2^28+1 in k range: 1000T to 1100T (78-bit factors)
 k_min = 1000000000000000
 k_max = 1100000000000000
Using GPU kernel "mfaktc_barrett89_F0_31gs"
    class | candidates |    time |    ETA | raw  rate | SievePrimes | CPU wait
  21/4620 |     21.65G | 13.875s |  3h40m | 1560.00M/s |       69941

got assignment: k*2^28+1, k range 1000000000000000 to 1100000000000000 (78-bit factors)
Starting trial factoring of k*2^28+1 in k range: 1000T to 1100T (78-bit factors)
 k_min = 1000000000000000
 k_max = 1100000000000000
Using GPU kernel "mfaktc_barrett89_F0_31gs"
    class | candidates |    time |    ETA | raw  rate | SievePrimes | CPU wait
  21/4620 |     21.65G | 13.009s |  3h27m | 1663.85M/s |       69941

On a GTX 460 (725 MHz factory overclocked) the raw rate increases from 1000 M/s to ca. 1045 M/s.

*Debian 6.0 squeeze GNOME 2 Desktop / CUDA SDK 4.1 for Ubuntu 11.04 / libcudart using the libstdc++ from a self compiled gcc 4.5.x
Compilation with the CUDA SDK 4.0 for Ubuntu 10.10 fails with an internal error in the nvopencc.

[ATH]

Originally Posted by Prime95

Here we go again -- v 0.25:

Thank you for all the time you spend on GIMPS/primenet and now also this program.

Originally Posted by Prime95

My recommendation is to not upgrade until ATH, flashjh, and others have had time to try this version for a little bit. They have been quite effective in verifying the quality of recent releases.

Sorry I keep giving you more work :)

[flashjh]

Originally Posted by LaurV

binaries? flashjh?

Windows 32-bit & 64-bit executables: