Enable ppc64le

[jayfurmanek]

This PR adds support for building recent cudatoolkit versions for ppc64le.

• Nvidia's RHEL7 based runfiles don't use embedded runfiles
• Once the toolkit is extracted, it ends up in a directory called "cuda-toolkit'
• the --extract runfile command is used because letting the runfile do an "install" results in attempted installs of .pc and doc files into standard Linux locations, which is not what we want.

#4

[jakirkham]

cc @jjhelmus

[jjhelmus]

AnacondaCON starts tomorrow and I have a week of vacation following the conference. I will not have time to look at this until I'm back on June 15th.

[jakirkham]

Sounds good. Thanks for the update. Hope you have a good break 🙂

[jakirkham]

Friendly nudge @jjhelmus 🙂

[jjhelmus]

I tried building this on our ppc64le machine inside a docker container we use for all our build, this failed with:

...
cuda_10.2.89_440.33.01_linux_ppc64le.run: 2.19GB [11:25, 3.20MB/s]                                                                                                                  
md5sum.txt: 3.07kB [00:00, 39.7kB/s]                                                                                                                                                
 Failed to verify gcc version. See log at /var/log/cuda-installer.log for details.

The container has no compilers installed in order to insure that those provided by conda are used during the build. On x86_64 this setup works to build the package.

I'm working on testing in a container with a compiler installed.

[jayfurmanek]

hmm..I'm not sure why the runfile is wanting a compiler on ppc64le.. It's likely I happened to have one installed when I tested.

[jayfurmanek]

Looking back at this I do see an error when no gcc is installed. The cuda-intaller.log shows:

[INFO]: Setting extractPath=/home/conda/cudatoolkit
[INFO]: Setting toolkit=true
[INFO]: Driver not installed.
[INFO]: Checking compiler version...
[INFO]: gcc location: which: no gcc in ....
[ERROR]: Missing gcc. gcc is required to continue.

Installing the driver requires gcc (for the kernel module shim) but its not needed elsewhere. Perhaps the gcc check is just upfront in any case. It looks like there is an override that disables that check...and looks like that will do the trick:

[INFO]: Setting extractPath=/home/conda/cudatoolkit
[INFO]: Setting toolkit=true
[INFO]: Overriding compiler check
[INFO]: Driver not installed.
[INFO]: Extracting to /home/conda/cudatoolkit

I'll update the RP with the override flag set.

[jayfurmanek]

OK, updated.
@jjhelmus Can you try again?

Also - avoid the conda-forge patchelf for this. I was getting ELF load command alignment not page-aligned errors in the load tests.

[jakirkham]

Also - avoid the conda-forge patchelf for this. I was getting ELF load command alignment not page-aligned errors in the load tests.

Would you be able to add a minimal reproducer here?

[jjhelmus]

Thanks for the update @jayfurmanek. With the most recent change this builds inside our docker container. Merging. Packages will be available later today.

[jjhelmus]

cudatoolkit-10.2.89 packages are now available in defaults for the linux-ppc64le platform.