Will Msieve (SVN 1043) work with CUDA 11.4 and an Arch 3.0 GPU? No!

[EdH]

When I installed the toolkit and CUDA from scratch in Ubuntu 20.04, they claimed to be 10.1. Even "NVCC --version" and NVIDIA-smi both showed 10.1. However, somehow the 10.1 has upgraded to 11.4 at some point.

Msieve builds without errors and tries to run, but has the following trouble:

$ ./msieve -g 0 -np -v
Msieve v. 1.54 (SVN 1043M)
Sun Jul 18 19:42:28 2021
random seeds: f434fda4 55303c82
factoring 442481795557623688831895262327430539285948212813785834989707204563288336300056437057 (84 digits)
searching for 15-digit factors
commencing number field sieve (84-digit input)
commencing number field sieve polynomial selection
polynomial degree: 4
max stage 1 norm: 1.61e+14
max stage 2 norm: 3.32e+13
min E-value: 7.30e-08
poly select deadline: 59
time limit set to 0.02 CPU-hours
expecting poly E from 1.39e-07 to > 1.60e-07
searching leading coefficients from 1 to 3114368
using GPU 0 (NVS 510)
selected card has CUDA arch 3.0
cannot load library 'cub/sort_engine.so': cub/sort_engine.so: cannot open shared object file: No such file or directory
error: failed to load GPU sorting engine from "cub/sort_engine.so"

Is this a problem with Msieve or CUDA?

[RichD]

You may have upgraded your software beyond the hardware supported driver levels. I've seen this before but don't recall the consequences. Others, more knowledgable, will give a better explanation.

[frmky]

You need to copy over not only the msieve binary but also the built CUDA libraries in their directories. Create a cub directory and copy sort_engine.so from your build ./cub directory to that one.

[EdH]

Originally Posted by frmky

You need to copy over not only the msieve binary but also the built CUDA libraries in their directories. Create a cub directory and copy sort_engine.so from your build ./cub directory to that one.

Thanks. I'm running in the original build directory, but I see that there is no sort_engine.so anywhere in the sub/directories, so I guess it didn't get built.

[EdH]

I manually ran the Makefile in the cub directory and now have sort_engine.so. I don't remember having to do that in the distant past and didn't note that step in the docs.

However, now I have a different error message:

. . .
using GPU 0 (NVS 510)
selected card has CUDA arch 3.0
deadline: 8640000 CPU-seconds per coefficient
error (line 1111): CUDA_ERROR_FILE_NOT_FOUND

"line 1111" in which file?

[EdH]

This is probably my answer:

$ make
"/usr/local/cuda-11.4/bin/nvcc" -gencode=arch=compute_30,code=\"sm_30,compute_30\" -DSM300 -o sort_engine.so sort_engine.cu -Xptxas -v -Xcudafe -# -shared -Xcompiler -fPIC -Xcompiler -fvisibility=hidden -I"/usr/local/cuda-11.4/bin/..//include" -I. -O3  
nvcc fatal   : Unsupported gpu architecture 'compute_30'
make: *** [Makefile:67: sort_engine.so] Error 1

When I simply used "make" in the cub directory, it defaulted to sm_52, so it built the wrong .so. The "make" driven by the main Makefile probably called the cub Makefile as designed and that failed, but didn't provide an error I could catch in the displayed stream.

[RichD]

I have upgraded my system to CUDA 11.2. I made the necessary changes to sm_(xx) in the two Makefile(s). When making the package I got an error stating 'no' is an invalid value for -abi option. I couldn't find anything related with a Google search so I removed "-Xptxas -abi=no" from the Makefile. It built a binary successfully. Now my question, before I attempt to use it, did I break something because I don't know what is -abi ?

P.S. I am running Linux.

[henryzz]

Originally Posted by RichD

I have upgraded my system to CUDA 11.2. I made the necessary changes to sm_(xx) in the two Makefile(s). When making the package I got an error stating 'no' is an invalid value for -abi option. I couldn't find anything related with a Google search so I removed "-Xptxas -abi=no" from the Makefile. It built a binary successfully. Now my question, before I attempt to use it, did I break something because I don't know what is -abi ?

P.S. I am running Linux.

I think https://stackoverflow.com/questions/...osts-registers explains this.




I think --abi-compile no is the modern equivalent. I think the logic for turning it off is often to save registers. Newer generations of GPU/compiler might benefit from turning it back on. Alternatively it may not work because it could avoid a bug.

[RichD]

I still couldn't get the MAKE to work without errors. I looked at Greg's version of Msieve (with GPU-LA) and he has it removed. After thinking a bit more, I believe I can use Greg's branch of Msieve for all my needs: factMsieve, GPU poly search and GPU-LA. I found the following in the wiki write up about CUDA incase others are thinking of upgrading.

• CUDA SDK 9.0 – 9.2 support for compute capability 3.0 – 7.0 (Kepler, Maxwell, Pascal, Volta) (Pascal GTX 1070Ti Not Supported. CUDA SDK 9.0 and support CUDA SDK 9.2).
• CUDA SDK 10.0 – 10.2 support for compute capability 3.0 – 7.5 (Kepler, Maxwell, Pascal, Volta, Turing). Last version with support for compute capability 3.0 and 3.2 (Kepler in part). 10.2 is the last official release for macOS, as support will not be available for macOS in newer releases.
• CUDA SDK 11.0 support for compute capability 3.5 – 8.0 (Kepler (in part), Maxwell, Pascal, Volta, Turing, Ampere (in part)).[37]
• CUDA SDK 11.1 – 11.4 support for compute capability 3.5 – 8.6 (Kepler (in part), Maxwell, Pascal, Volta, Turing, Ampere (in part)).[38]
• CUDA SDK 11.5 – 11.7.1 support for compute capability 3.5 – 8.7 (Kepler (in part), Maxwell, Pascal, Volta, Turing, Ampere).[39]