Skip to content
Peter Scheibel edited this page Mar 16, 2022 · 14 revisions

Attendees

  • Peter Scheibel (host)
  • David Alexander
  • Mark Krentel
  • Tammy Dahlgren
  • Massimiliano Culpo
  • Wileam Phan
  • Greg Becker
  • Todd Gamblin

Agenda items

  • (Greg and Wileam) Proposal for nvhpc vendoring CUDA: can we delete the vendored CUDA?
    • The problem: bundling vendored packages into nvhpc causes reference errors
    • (Greg): the nvhpc package should install everything, then delete everything but the compilers
      • This wastes some bandwidth, but ensures consistency
    • Amending the package to remove these vendored dependencies seems possible:
      • References can be replaced in text files
      • And likely for binaries too
        • Any string references in the binary to the vendored CUDA install (inside Spack nvhpc) are guaranteed to be shorter than Spack CUDA installation prefix (so there isn't an issue with trying to replace a shorter string with a longer string in the binary)
  • Update contributing.md to tell users to read the packaging guide
  • (Massimiliano) How should we deal with package methods that rely on the presence of some variant and fail when accessing spec installed before the variant was added? See https://github.com/spack/spack/issues/28188
    • Proposal: cache values (like libs) in a file so that you don't have to recalculate them
    • (Greg) why not use the package from the upstream as it was installed? (right now part of the problem is that we use the latest builtin package definition)
      • (Todd) mentioned this being problematic
    • Can https://github.com/spack/spack/issues/28188 be resolved with defensive variant access?
      • i.e. prepare for them to be undefined
      • This could be as simple as using .satisfies for the check
      • (Greg) Most raw variant access (variants[key]) in the repo is for the CUDA arch
      • Possible complication: variants that are accessed for their string value (e.g. to be fed into a build system as an option)
        • Preparing for 'none' here could be difficult
    • For now: make this a style issue (the package reviewer should be aware of)
      • Could have a guide for package reviewers

For next week

Clone this wiki locally