GMP-ECM GPU for WSL problems

[Rubiksmath]

yo pog i think it worked really smoothly with not a single error everything first try. Thanks man. I didnt realise i didnt need to compile CGBN, lol my bad.

[Rubiksmath]

Nope, nevermind, test.cgbnecm failed and I dont know why, I checked test.cgbnecm.log and I dont see any obvious failures in there, It did a bunch of curves and then at one of the test numbers the file just ends after stage 1, Im guessing stage 2 init failed or something but there is no form of error report I can find. Anyone know how to fix?
Edit: hmm weird running curves with cgbn works just fine it seems.

[chris2be8]

Post the last few lines from test.cgbnecm.log here. If you are not sure how to do that post output from tail test.cgbnecm.log (or post the whole file if it's small).

You will get better responses if you post enough information to give us a chance to tell what's wrong.

Also tell us what sort of GPU you have. And how much memory it has.

[wombatman]

Originally Posted by Rubiksmath

Nope, nevermind, test.cgbnecm failed and I dont know why, I checked test.cgbnecm.log and I dont see any obvious failures in there, It did a bunch of curves and then at one of the test numbers the file just ends after stage 1, Im guessing stage 2 init failed or something but there is no form of error report I can find. Anyone know how to fix?
Edit: hmm weird running curves with cgbn works just fine it seems.

Originally Posted by chris2be8

Post the last few lines from test.cgbnecm.log here. If you are not sure how to do that post output from tail test.cgbnecm.log (or post the whole file if it's small).

You will get better responses if you post enough information to give us a chance to tell what's wrong.

Also tell us what sort of GPU you have. And how much memory it has.

It would also be helpful to check test.gpuecm and make sure you pass all those as well. That will help us know if the issue with CGBN or the GPU generally.

[Rubiksmath]

Last few lines:

Input number is 2^1019-1 (307 digits)
Using B1=4, B2=0, sigma=3:92-3:123 (32 curves)
GPU: Using device code targeted for architecture compile_86
GPU: Ptx version is 86
GPU: maxThreadsPerBlock = 640
GPU: numRegsPerThread = 88 sharedMemPerBlock = 0 bytes
Computing 32 Step 1 took 0ms of CPU time / 0ms of GPU time

Same sort of thing that is at the end of the other log files which passed just fine, but as I said running curves with CGBN throws no errors and the stage 1 is insanely fast (quite literally hundreds of times faster than it was with my previous setup, can now run 1792 stage 1 curves at B1=1e9 on a 248 digit number in 7 hours. GPU is a RTX 3060 with 12 GB RAM). Last thing: where does it store the results?? I can't seem to find where it stores them.

[VBCurtis]

If you don't tell ecm to save residues somewhere, there is no output file. The program just runs the curves and reports the result to screen.

You can redirect the screen output to a file to achieve a record of curves run, or if you're doing stage 2 separately you can ask it to save stage1 residues to a file of your choosing and then invoke gmp-ecm with -resume and that filename to run stage 2.

If you post the exact command line you're using, you're likely to get better feedback; you keep not telling us enough for us to figure out exactly what you (or the setup) might have wrong.

[Rubiksmath]

Sorry about that.
Anyway, I've read the documentation on the save feature and have one question: the documentation says that the residues are saved only after step 2 is completed, if stage 2 has been run in that same run of GMP-ECM. Now this implies that stage 1 and 2 can be run separately, however I cannot see where it mentions how to do this. So my question is: how to run stage 1 only and save residues after stage 1?

[VBCurtis]

try ecm -h for options.

You want the -save, -savea, and -resume options. If I recall, you want to use the same B1 with -resume as you did on your GPU run so that gmp-ecm picks a reasonable B2; or you can choose B2 directly but still have to put a B1 on the command line (else it'll think your desired B2 is actually B1).

[chris2be8]

Originally Posted by Rubiksmath

Last few lines:

Input number is 2^1019-1 (307 digits)
Using B1=4, B2=0, sigma=3:92-3:123 (32 curves)
GPU: Using device code targeted for architecture compile_86
GPU: Ptx version is 86
GPU: maxThreadsPerBlock = 640
GPU: numRegsPerThread = 88 sharedMemPerBlock = 0 bytes
Computing 32 Step 1 took 0ms of CPU time / 0ms of GPU time

Same sort of thing that is at the end of the other log files which passed just fine, but as I said running curves with CGBN throws no errors and the stage 1 is insanely fast (quite literally hundreds of times faster than it was with my previous setup, can now run 1792 stage 1 curves at B1=1e9 on a 248 digit number in 7 hours. GPU is a RTX 3060 with 12 GB RAM). Last thing: where does it store the results?? I can't seem to find where it stores them.

When I ran the same test on my GPU I got:

Input number is 2^1019-1 (307 digits)
Using B1=4, B2=0, sigma=3:92-3:123 (32 curves)
No available CGBN Kernel large enough to process N(1019 bits)
Computing 32 Step 1 took 0ms of CPU time / 0ms of GPU time
Please report internal errors at <ecm-discuss@inria.fr>.

But that's supposed to happen. test.cgbnecm is basically ./test.gpuecm "${1:-./ecm -cgbn}"
And in test.gpuecm:

# try 2^1019-1: should give an error
echo "2^1019-1" | $ECM -sigma 3:92 4 0
checkcode $? 1

So it looks as if your ecm and GPU combination is giving a rc indicating success. B1=4 would take very little time so saying it took 0ms is probably OK.

Try commenting out that test in a copy of test.gpuecm and see if it passes the other tests. If it does you are probably OK.

[SethTro]

I sorry this has been so hard to setup under WSL. If there's anything that can be improved in documentation / autoconfig I'll endeavour to make it better for the next WSL user.

Originally Posted by Rubiksmath

Sorry about that.
Anyway, I've read the documentation on the save feature and have one question: the documentation says that the residues are saved only after step 2 is completed, if stage 2 has been run in that same run of GMP-ECM. Now this implies that stage 1 and 2 can be run separately, however I cannot see where it mentions how to do this. So my question is: how to run stage 1 only and save residues after stage 1?

like such

# the 0 here says don't run stage 2
$ echo "(2^257-1)/535006138814359" | ./ecm -cgbn -save 2_257_1e3.txt 1e3 0
# Resume the save file
$ ./ecm -q -resume 2_257_3e4.txt 3e4 1e6

Originally Posted by Rubiksmath

Last few lines:

Input number is 2^1019-1 (307 digits)
Using B1=4, B2=0, sigma=3:92-3:123 (32 curves)
GPU: Using device code targeted for architecture compile_86
GPU: Ptx version is 86
GPU: maxThreadsPerBlock = 640
GPU: numRegsPerThread = 88 sharedMemPerBlock = 0 bytes
Computing 32 Step 1 took 0ms of CPU time / 0ms of GPU time

Same sort of thing that is at the end of the other log files which passed just fine, but as I said running curves with CGBN throws no errors and the stage 1 is insanely fast (quite literally hundreds of times faster than it was with my previous setup, can now run 1792 stage 1 curves at B1=1e9 on a 248 digit number in 7 hours. GPU is a RTX 3060 with 12 GB RAM). Last thing: where does it store the results?? I can't seem to find where it stores them.

This test checks that the code fails, which it seems to be doing, but maybe WSL isn't returning the same error. On my system

$ echo "2^1019-1" | ./ecm -cgbn -sigma 3:92 4 0
...
$ echo $?
1

[SethTro]

Originally Posted by Rubiksmath

The stage 1 is insanely fast (quite literally hundreds of times faster than it was with my previous setup, can now run 1792 stage 1 curves at B1=1e9 on a 248 digit number in 7 hours. GPU is a RTX 3060 with 12 GB RAM). Last thing: where does it store the results?? I can't seem to find where it stores them.

I'm glad it's much faster! I'm curious how my 1080ti stacks up against the 30 series (if someone in this thread has a 3080/3090 I'd love to see some timing results too), would you mind sharing the result of running?

$ echo "10^248+123" | ./ecm -v -cgbn 2e5 0
...
CGBN<1024, 8> running kernel<56 block x 256 threads> input number is 824 bits
...
Computing 1694 bits/call, 134631/288578 (46.7%), ETA 8 + 7 = 14 seconds (~8 ms/curves)
Copying results back to CPU ...
Computing 1792 Step 1 took 37ms of CPU time / 14013ms of GPU time
Throughput: 127.885 curves per second (on average 7.82ms per Step 1)