Divergent Posts from Benchmark Thread

[charybdis]

Originally Posted by storm5510

Then where can they be found?

The msieve-lacuda-nfsathome branch of https://github.com/gchilders/msieve_nfsathome.git.
As frmky said a few posts back, there is no precompiled Windows version. You should follow the instructions he gave in the post above.

[storm5510]

I did a web search and found one on Sourceforge called msieve152_gpu.zip.

It seems to start then gives the message below.

Cannot load library 'b40c/sort_engine_sm20.dll', error 126
Error: failed to load GPU sorting engine

This is becoming a major waste of time. Incomplete packages seem to be the norm on GitHub and other sites, like Sourceforge. I searched and searched and found nothing.

No more Linux references please!

[frmky]

As I said earlier...

Originally Posted by frmky

I'm not aware of anyone that's compiled a Windows version with my changes.

[storm5510]

Originally Posted by frmky

Once you have installed the compiler and CUDA toolkit...
https://docs.nvidia.com/cuda/cuda-in...u-installation

sudo apt install git libgmp-dev
git clone https://github.com/gchilders/msieve_nfsathome.git -b msieve-lacuda-nfsathome
cd msieve_nfsathome
make all CUDA=75 VBITS=256

If you don't want to run from that directory, then copy msieve, *.ptx, and cub/*.so to the directory you are running in.

I had an extra blank hard drive so I installed Ubuntu 21.10 on a Xeon system I have. As I mentioned, I have some notes from before:

sudo apt install nvidia-driver-510. (specific for GTX 1080).
sudo apt-get install build-essential
sudo apt-get install nvidia-cuda-toolkit

Installed the compiler? This seems familiar, but I do not know if is is part of the above or not...

Edit: It tells me there is no longer support for 21.10.

[storm5510]

Originally Posted by frmky

Once you have installed the compiler and CUDA toolkit...
https://docs.nvidia.com/cuda/cuda-in...u-installation

sudo apt install git libgmp-dev
git clone https://github.com/gchilders/msieve_nfsathome.git -b msieve-lacuda-nfsathome
cd msieve_nfsathome
make all CUDA=75 VBITS=256

If you don't want to run from that directory, then copy msieve, *.ptx, and cub/*.so to the directory you are running in.

Just so you know, the "make" does not work with Ubuntu 22.04.1 LTS. Ubuntu's text editor claims there is a "missing bracket" on the last line of the makefile. I didn't see one. I installed the Nvidia driver before attempting the compile.

Error limit reached.
100 errors detected in the compilation of "sort_engine.cu".
Compilation terminated.
make[1]: *** [Makefile:106: sort_engine.so] Error 1
make[1]: Leaving directory '/home/storm5510/msieve_nfsathome/cub'
make: *** [Makefile:360: cub/built] Error 2
storm5510@1620XE:~/msieve_nfsathome$

I tried...

[storm5510]

I would sincerely like to apologize to anyone here I offended. I was totally out of line with some of my comments.

[Uncwilly]

We all can have a bad day now and then.

[EdH]

I'll attempt to remember and cover some of the roadblocks I experienced and had to overcome in order to get my GPUs working. Perhaps something will illuminate a shortfall:

- Although the system insisted on installing CUDA 11, I had to install CUDA 10.2. If you can't find it in the repositories, you may need to use a .run file. We should be able to find that in one of the prior threads.

- The NVidia driver has to be installed and the machine rebooted after installation.

At this point, nvidia-smi should run and show the card and driver info among other things. It will probably show CUDA 11 even though you have 10.2 installed.

- You need the msieve branch mentioned earlier.

- This is VERY important: You need to add some CUDA 10.2 directories to two PATH environment variables:

- - Use nautilus (files) to look at the /usr/local directory. Get there by choosing Other Locations and then Computer, followed by usr and finally local. In that directory look for cuda-10.2 to make sure it's there.

- - If cuda-10.2 is found type the following exactly (or copy/paste):

export PATH=$PATH:/usr/local/cuda/bin:/usr/local/cuda/targets/x86_64-linux/include

and press enter. You can verify it worked with echo $PATH, which should show the entire variable.

- - Again, if cuda-10.2 is found type the following exactly (or copy/paste):

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-10.2/lib64

and press enter. You can verify it worked with echo $LD_LIBRARY_PATH, which should show the entire variable.


- - Note that these PATH additions will disappear after any restart and would have to be added again before running make again.


- Move into the nfsathome_msieve directory and use the make all CUDA=75 VBITS=256 command from earlier.

I'll step back for a time and see if any of this helps and await others' additions/corrections.

[storm5510]

Originally Posted by Uncwilly

We all can have a bad day now and then.

I have too many. I need to address that seriously and not bring it here.

Originally Posted by EdH

I'll attempt to remember and cover some of the roadblocks I experienced and had to overcome in order to get my GPUs working. Perhaps something will illuminate a shortfall...

I decided to start from scratch with Ubuntu 20.04.4 LTS. By now, it's been patched pretty good. There was a few updates that needed to be added, just a few.

I am familiar with nvidia-smi, and use it on my Win10 system. In this case, CUDA 11.6 is what it shows. I don't know if this voids a lot of what you wrote above or not. I went to "usr/local" as you suggest. No CUDA folder to be seen.

A different direction for a moment. Below is what the complier reported this morning:

/usr/bin/ld: cannot find -lz
collect2: error: ld returned 1 exit status
make: *** [makefile:316: all] error 1

I looked at line 316 in the makefile. It didn't seem relevant.

I have a list of notes here about installing various things. One is the nvidia-cuda-toolkit. I am thinking maybe I forgot to run that one based on what I found while browsing around. I remember sitting and watching the actual driver install. It reported it didn't need "build-essential." Toolkit is the next item on my list. I did libgmp-dev. and the clone in frmky's list above.

I think I will just sit here and cool my heels and think for a while.

[frmky]

sudo apt install zlib1g zlib1g-dev

will resolve the error with -lz.

With CUDA 11.6, you might need to use the update_cub branch instead of msieve-lacuda-nfsathome branch. After the git clone, go into the msieve_nfsathome directory and

git checkout update_cub

Then use the make command from before.

This branch is still a work-in-progress and it will be renamed at some point. nVidia made a number of changes to cub in CUDA 11.6, and this branch is modified to support those changes.

[storm5510]

Thank you for the reply. Most kind.

Below is what I get back from zliblg:

Unable to locate package zliblg
Unable to locate package zliblg-dev

Cub appears to have been successful:

Branch 'update_cub' set up to track remote branch 'update_cub' from 'origin'
Switched to new branch 'update_cub'

Perhaps there is a path issue with zliblg. I don't know how environment variables work with Linux. I spent some time browsing around with the GUI file app and couldn't find anything similar. Maybe it is not here and needs to be pulled in from outside?

Update: I found the difficulty. I did a web search and found my spelling was incorrect. It is zlib1g. It appeared to me as an alphabetic 'l' above. So now, I have an msieve binary. Now I need to test drive it.