The P-1 factoring CUDA program

[James Heinrich]

Originally Posted by frmky

Here's a Windows version to try
Included is a worktodo.txt that should find a factor.
please try that first, make sure the factor is found, and please let me know if it worked!

Starting a run on my GTX 570. One thing that might be a concern (especially in light of Stef42's comment about stage2 memory) is the references to "zu" as a quantity of graphics memory:

Selected B1=605000, B2=16637500, 4.1% chance of finding a factor
CUDA reports zuM of zuM GPU memory free.
Using e=6, d=2310, nrp=16
Using approximately zuM GPU memory.
Starting stage 1 P-1, M61262347, B1 = 605000, B2 = 16637500, e = 6, fft length =
3360K
Doing 873133 iterations

Running at 7.0ms/it, stage1 should be done in 1h40m and I'll report back with what happens when stage2 starts.

[owftheevil]

The zu is not a big deal, simply a size specifier specific to gcc. For windows we need Iu instead.

[Stef42]

Right after stage 1 finished and stage 2 was initiated, I got a popup saying that CUDAPm1 crashed.
The Windows error log showed an APPCRASH, which is not very useful I think.

When followed was that after the gpu load dropped from 99% to 0%,
the memory remained at 1134MB usage until, I guess because of a time-out, was flushed.

Maybe I'll try a smaller P-1 exponent with a factor found to check.

[Karl M Johnson]

Feedback so far:
1. Does not create checkpoints.
2. Beats CUDALucas in memory stability stress testing(60M exponents were free from errors on CL, found errors at 50K iterations on CP+1)
3. Fails at the beginning of stage 2 with out of memory error, should not be the case (6GB of vRAM, 16GB of RAM).

[James Heinrich]

Originally Posted by owftheevil

The zu is not a big deal, simply a size specifier specific to gcc. For windows we need Iu instead.

Would it be a big deal that we're seeing "zu" instead of "Iu" on Windows?

[owftheevil]

I think I found one problem. chalsall has his SPEs. Well, this was an ISPE, I for Ineffably. Its an easy fix, but will have to wait until I get home from work. In the meantime, running with b2 = even multiple of 2310 should bypass the error.

@Karl M Johnson:

1. Checkpoints are coming soon, maybe this weekend.
2. CPm1 during stage 1 does do more global memory reads than CuLu, so maybe thats why.
3. Unexpected. What is the error message?

Thank you all for your input.

[owftheevil]

Originally Posted by James Heinrich

Would it be a big deal that we're seeing "zu" instead of "Iu" on Windows?

%zu in printf prints size_t variable values, you need %Iu in windows to do the same thing.

[James Heinrich]

Ah, now I understand what you mean.

[Stef42]

I restarted CUDAPm1 using a b2 value, of a known P-1 with a factor found by me earlier.

This was the command-line:

cudapm1.exe -b2 550000

Output:

Iteration 164000 M9090017, 0xd7661b0c859fa9e5, n = 512K, CUDAPm1 v0.00 err = 0.0
2734 (0:01 real, 0.7921 ms/iter, ETA 0:01)
Iteration 165000 M9090017, 0x7d3f99a08f445b8b, n = 512K, CUDAPm1 v0.00 err = 0.0
2734 (0:01 real, 0.7878 ms/iter, ETA 0:00)
M9090017, 0x1d50507696eeef9f, offset = 0, n = 512K, CUDAPm1 v0.00
Stage 1 complete, estimated total time = 2:14
Starting stage 1 gcd.
M9090017 Stage 1 found no factor (P-1, B1=115000, B2=1495000, e=6, n=512K CUDAPm
1 v0.00)
Starting stage 2.
Zeros: 59077, Ones: 84923, Pairs: 18379
itime: 14.921770, transforms: 1, average: 14921.770000
ptime: 35.394836, transforms: 88612, average: 0.399436
ETA: 0:50
itime: 17.911887, transforms: 1, average: 17911.887000
ptime: 35.547328, transforms: 88434, average: 0.401964
ETA: 0:00
Stage 2 complete, estimated total time = 1:43
Accumulated Product: M9090017, 0x1a6840caa5d05db3, n = 512K, CUDAPm1 v0.00
Starting stage 2 gcd.
M9090017 has a factor: 516770062491225473521 (P-1, B1=115000, B2=1495000, e=6, n
=512K CUDAPm1 v0.00)

As you can see, there is a different B2 value. Still, it finished well.
Earlier on, the program would crash when starting stage 2. Any thoughts? I must have done something wrong

Bit more surprising: according to mersenne.ca,
in the past the factor was found in stage 1 using prime95, but CudaPm1 reports stage 2 in the output... ?
Exponent 9090017

[James Heinrich]

Originally Posted by Stef42

M9090017 Stage 1 found no factor (P-1, B1=115000, B2=1495000, e=6, n=512K CUDAPm
1 v0.00)
Bit more surprising: according to mersenne.ca,
in the past the factor was found in stage 1 using prime95, but CudaPm1 reports stage 2 in the output... ?

That is a bit disturbing.

M9090017 has factor 516770062491225473521, with a k of 28425142796280
k-factored = 2³ × 3 × 5 × 61 × 97 × 389 × 102913
minimal bounds to find this factor in stage2 would be B1=389,B2=102913
minimal bounds to find this factor in stage1 would be B1=102913

You ran this with B1=115000 so it should have found the factor, at least according to my understand of P-1

[Stef42]

It can still find it, although I wonder, as you mentioned, in stage 2 rather than stage 1....
I will do some further testing on a different exponent.

Did the same test again on prime95 to verify:

[May 2 16:52] Worker starting
[May 2 16:52] Setting affinity to run worker on any logical CPU.
[May 2 16:52] P-1 on M9090017 with B1=110000
[May 2 16:54] M9090017 stage 1 complete. 317502 transforms. Time: 128.915 sec.
[May 2 16:54] Stage 1 GCD complete. Time: 6.593 sec.
[May 2 16:54] P-1 found a factor in stage #1, B1=110000.
[May 2 16:54] M9090017 has a factor: 516770062491225473521
[May 2 16:54] No work to do at the present time.  Waiting.