Divergent Posts from Benchmark Thread

[storm5510]

Originally Posted by charybdis

You've already been talking to the author. He is "frmky". (Edit for clarification: he's only the author of that github repo with the GPU linear algebra and multithreaded square root. The main author of msieve is "jasonp".)

It looks like you're trying to run msieve NFS postprocessing without a polynomial file. What number are you trying to run it on? If you're trying to run it on the benchmarking job, have you copied the msieve.fb, msieve.dat and worktodo.ini files across to the directory that has the msieve binary?

What I am trying to run is a number from James Heinrich's Factoring Project. It is a 128 digits in length. msieve never runs long enough to generate the msieve.??? files you mention. "Error line 1062." All I have is the worktodo.ini. All other files are in the same folder as the main program.

I know the GPU setup is correct. It ran gpuOwl overnight last night without problems.

[frmky]

Originally Posted by storm5510

"Error line 1062."

That would be the line

CUDA_TRY(cuModuleLoad(&t->gpu_module, "stage1_core.ptx"))

Make sure you have the files stage1_core.ptx and lanczos_kernel.ptx in the same directory in which you are running msieve as well as sort_engine.so and spmv_engine.so in the cub subdirectory.

If you do have stage1_core.ptx in the same directory as msieve, then what is the output of

head -n 15 stage1_core.ptx

[storm5510]

Originally Posted by frmky

...Make sure you have the files stage1_core.ptx and lanczos_kernel.ptx in the same directory in which you are running msieve as well as sort_engine.so and spmv_engine.so in the cub subdirectory.

If you do have stage1_core.ptx in the same directory as msieve, then what is the output of

head -n 15 stage1_core.ptx

All files mentioned above are present and in their proper place.

$ ./msieve head -n 15 stage1_core.ptx

deadline: 8640000 CPU-seconds per coefficient
error (line 1062): unexpected error

[kruoli]

In the case of the head ... command, please leave out the ./msieve part. Thanks!

[storm5510]

Originally Posted by kruoli

In the case of the head ... command, please leave out the ./msieve part. Thanks!

//
// Compiler Build ID: CL-26907403
// Cuda compilation tools, release 10.1, V10.1.243
// Based on LLVM 3.4svn
//

.version 6.4
.target sm_75
.addresss_size 64

// globl  sieve_kernel_trans_pp32_r32

.visible .entry sieve_kernel_trans_pp32_r32(

The hanging "(" in the last line is not a typo. It took some neck-stretching to get this right.

[kruoli]

It looks like the stx file is made for a card with CC of 7.5 instead of the 6.1 you want to have.

Unfortunately, I am not sure why it is like this if you have cleaned (make clean) everything, but surely someone else will chime in here.

[storm5510]

Originally Posted by kruoli

It looks like the stx file is made for a card with CC of 7.5 instead of the 6.1 you want to have.

Unfortunately, I am not sure why it is like this if you have cleaned (make clean) everything, but surely someone else will chime in here.

It is a GTX 1080. CC is 7.5.

Two days ago, this was a clean HD. Totally wiped and reloaded with Ubuntu 20.04.4 LTS. There were no errors in the msieve compile. gpuOwl and mfaktc both run as they should.

IMHO, there are too many options for GPU drivers. I had been using 510. This time, I went back to 470.

FWIW, mfaktc reports it is using CUDA 10.xx

[frmky]

Originally Posted by storm5510

It is a GTX 1080. CC is 7.5.

Not according to

Originally Posted by storm5510

It sees the GPU and says it has CUDA arch 6.1.

[EdH]

Originally Posted by storm5510

It is a GTX 1080. CC is 7.5.

A check for the specs for this card shows CUDA 6.1 (as nvidia-smi reported), so you should use CUDA=61 as frmky suggested a bit ago.

ETA: "arch" stands for architecture, which is synonymous with Compute Capability, and is also shown as CUDA in some references. You can can see the specs on this page, and look at Graphics Features for the "CUDA" value.

[storm5510]

Originally Posted by EdH

A check for the specs for this card shows CUDA 6.1 (as nvidia-smi reported), so you should use CUDA=61 as frmky suggested a bit ago.

ETA: "arch" stands for architecture, which is synonymous with Compute Capability, and is also shown as CUDA in some references. You can can see the specs on this page, and look at Graphics Features for the "CUDA" value.

Honestly, I cannot remember where I got the "6.1" from. nvidia-smi does not show it.
.....
I found it in another location. mfaktc shows it in the information upon startup, "compute capability 6.1." So, I did this:

make all CUDA=61 VBITS=256

I think I need to dump the entire folder because I still gets the line 1062 error. Next:

make clean
make all CUDA=61 VBITS=256

It runs!!!

I really appreciate everyone here trying to help while putting up with my Linux ignorance!

[LaurV]

Originally Posted by storm5510

my Linux ignorance!

at least here, you are not alone... (edit: to be clear, I have been a programmer for over 30 years, and I consider myself a VERY good one, but never went deeper under the surface related to Linux - members of the forum helped me every time, as well as a lot of web search and trials, when I was stuck with my nose and ears into it and could not move my head...)