-
Notifications
You must be signed in to change notification settings - Fork 2.2k
Telcon: 2022 03 30
Peter Scheibel edited this page Mar 30, 2022
·
23 revisions
- Peter Scheibel (host)
- Mark Krentel
- Massimiliano Culpo
- Phil Regier
- Tammy Dahlgren
- Richarda Butler
- Greg Becker
- Wileam Phan
- Harmen Stoppels
- Carlos Bederian
-
Possibly continue from https://github.com/spack/spack/wiki/Telcon%3A-2022-03-09: Packages with multiple build systems
-
How to reorganize the docs?
-
(Wileam) Can't bootstrap
llvm@13
if gold is enabled (which is the default behavior) https://github.com/spack/spack/issues/29350 -
(Wileam) Progress on
nvhpc
,nvhpc-slim
, andcuda
/cuda-toolkit
. Relevant links:- https://github.com/spack/spack/issues/19365
- https://github.com/spack/spack/pull/29155
- https://github.com/spack/spack/pull/29550
- https://github.com/spack/spack/pull/29742
- https://github.com/spack/spack/pull/29782
- Last we spoke: goal is removing CUDA from nvhpc install
- We're checking with Nvidia
- May want to defer
- Harmen: there is a flag you can pass to nvcc to choose the CUDA toolkit, so nvhcp may be somewhat integrated
- Greg: it looks like this is based on examining filesystem (vs. hardcoded path)
- Massimiliano:
- If nvhpc normally bundles several CUDA toolkits and several cuFFT versions, can they be combined arbitrarily?
- Peter: if nvhpc bundles multiple instances of the CUDA toolkit, how do we model that
- e.g. as of now, we don't have any way to model a package which depends on multiple instances of the same package
- that being said, virtuals may address this point in particular, since a package can provide multiple instances of something
- Peter: why not add a "contains" directive to say that nvhpc is actually a DAG
- Harmen: the CUDA toolkit provided with nvhpc is structured differently than the separate CUDA toolkit (moreover, the parts of CUDA toolkit are distributed across multiple prefixes)
- E.g. CMake had to address that
- Massimiliano: you could likely emulate the CUDA toolkit directory structure from nvhpc with symlinks
- Harmen: the CUDA toolkit provided with nvhpc is structured differently than the separate CUDA toolkit (moreover, the parts of CUDA toolkit are distributed across multiple prefixes)
-
(Wileam) AOCC 3.2.0 can't compile Python https://github.com/spack/spack/issues/29115