search

ENCCS / gpu-programming

Meta-GPU lesson covering general aspects of GPU programming as well as specific frameworks
https://enccs.github.io/gpu-programming/
Creative Commons Attribution 4.0 International
62 stars 20 forks source link

julia versions of stencil example #57

Closed wikfeldt closed 1 year ago

stepas-toliautas commented 1 year ago

Thanks! Will there be a module to load for Julia on LUMI, or should we fetch the distribution/ place it on project's scratch folder? I'll tweak the output lines/ default options to give identical results to the other variants. Also, what would be the idiomatic way to collect constants (initial temps etc.) in one place and use names in the performant code? (As it's done in C++ and Python versions.) And, finally: will we be able to test Python/numba on LUMI? (v.env with an older version of numba, probably?)

wikfeldt commented 1 year ago
stepas-toliautas commented 1 year ago

there will most likely be a module to load on LUMI, i'm working with a CSC person to have it up in the coming days.

I will be tweaking example episode through the next weekend, aren't I 😄

idiomatic way to define constants in one place is to have a file like const.jl where all constants are defined with const keyword (important for performance), e.g. const x = 1.0

OK, will do.

what do you think about python/numba, should we demo how to use an old version of numba which supports AMD GPUs? doesn't feel quite right considering that they stopped supporting AMD GPUs, but at least it would be something for the pythonistas. I'm not sure how the future looks for AMD GPU programming in python

This is one of the consequences of us choosing LUMI: as I understand, Python/numba works fine on Nvidia GPUs, and also as I understand, such code should be portable/ transparent to the architecture (because if even Python is not, then what are we doing here). So jumping through some loops this run to make sure that the same lesson will be applicable for Nvidia participants/ workshop hosts in the future.

Are you OK with keeping this as PR, or should I merge and then revise?

wikfeldt commented 1 year ago

we can keep the PR open until necessary revisions are done

btw, this episode will now be taught on friday, not monday next week!

regarding numba, how about spinning up a google colab notebook to demo the numba examples? maybe better than not having any hands-on at all

stepas-toliautas commented 1 year ago

regarding numba, how about spinning up a google colab notebook to demo the numba examples? maybe better than not having any hands-on at all

Yup, let's do that. I actually started developing Python variant on Colab, but it did not agree to ArgParse, so I moved to local setup later. (And of course the execution will be slow in comparison, but the gains should be visible in Colab-CPU vs. Colab-GPU at least.)

stepas-toliautas commented 1 year ago

What would be the steps to run this on LUMI -- or anywhere? Both julia example.jl and julia> include("example.jl") complain about missing HeatEquation package, which sounds reasonable, as none of the included files have the information about src/ directory.

wikfeldt commented 1 year ago

i'll look into this this evening and make sure we have a working example tomorrow!

stepas-toliautas commented 1 year ago

i'll look into this this evening and make sure we have a working example tomorrow!

Great! I got Python version working well which I'm finishing now, and I'm still clarifying the structure of that entire episode for tomorrow, so I have two small requests: a) please commit updated / tested Julia files (as well as instructions for LUMI if they happen not to be straightforward) to stencil/julia when you will, and b) have a read through the updated 13-examples.rst in existing draft pull request #76 and feel free to fix or comment on layout / callout box choices, to make them more consistent with the rest of the lesson.

wikfeldt commented 1 year ago

@stepas-toliautas i've now simplified from a Julia module to three scripts roughly matching the python scripts, and added LUMI step-by-step instructions. Currently only threaded CPU version, will do GPU now but i suggest to merge right away