Build CuPy for PowerPC?

[leofang]

Continuing the discussion in cupy/cupy#2583 (comment).

@jakirkham @isuruf Do we have the compiler/docker/CI infrastructure for this task? A while ago I verified on Ascent (Summit's test node) that CuPy works on PowerPC, so it's just a matter of preparing and building.

cc: @berceanu

[leofang]

It seems we need conda-forge/nvcc-feedstock#21 to be closed first.

[leofang]

cf: AnacondaRecipes/cudatoolkit-feedstock#4

[leofang]

Pending AnacondaRecipes/cudatoolkit-feedstock#6 and conda-forge/docker-images#141.

[jakirkham]

The nvcc bit is easy once we have cudatoolkit.

[leofang]

AnacondaRecipes/cudatoolkit-feedstock#6 is merged!

[jakirkham]

Now we just need to figure out how to get the older versions in there 🙂

AnacondaRecipes/cudatoolkit-feedstock#8

[leofang]

Can we build on top of CF's Power8 cudatoolkit instead? conda-forge/cudatoolkit-feedstock#6

[jakirkham]

Right now we are still building out cudatoolkit versions. There's also some integration work needed. That said, yeah I think that should be possible.

[leofang]

@jakirkham How close we are to build for ppcle64? I lost track...

[leofang]

Looks like updating pinning is the last missing piece? conda-forge/conda-forge-pinning-feedstock#725

[jakirkham]

Maybe @jaimergp would know 😉

[jaimergp]

Hi! Yes, it's been on my list for a while, but it got deprioritized after finding out OpenMM has some issues on GCC9+PPC, so I focused on other things while those are solved.

I can work on it after I am done with conda-forge/nvcc-feedstock#59.

IIRC, we need:

• nvcc packages - see Arch Migrator nvcc-feedstock#62 (EDIT: Add PowerPC nvcc-feedstock#66 is merged)
• build cudatoolkit under x.y version (they didn't get built because of some issue with files being somewhere else wrt linux 64)
• update the pinnings

[leofang]

Note to self: cuSPARSELt currently does not support ppc64le.

[leofang]

Support for ppc64le + CUDA 10.2 is done for v9.2.0 (#134) and v10.0.0a2 (#135). Thank you @jaimergp again for making the infrastructure available!

[leofang]

Hi all, I am reopening this issue for discussion. I've been monitoring CuPy's download counts regularly, and I noticed the download counts have been consistently on the order of 15 for every version (Python x CUDA) on ppc64le, making me suspect we have almost no human user of the ppc64le packages. If it's the case, we should consider to not build them so as to save CI and anaconda.org resources.

[leofang]

Let me start by shrinking the build matrix for ppc64le.