Telcon: 2024 03 13
Peter Scheibel edited this page Mar 13, 2024
·
12 revisions
Wednesday March 13th, 9am PT (UTC -7:00)
- Peter Scheibel (host)
- Mark Krentel
- Tammy Dahlgren
- Prentice Bisbal
- Jakov Petrina
- Davide DelVento
(This week the meeting is for Q&A, there are no pre-planned topics)
- (resolved) Prentice:
- Trying to build aoclsparse
- aocl-sparse package says it needs amdblis for @4.1:, but it accesses amdblis for @4.0:
- Either the
depends_on
needs to move "back" or the phase logic needs to do only access for @4.1:
- Either the
- It would be nice to show details about failing package (e.g. compiler)
-
spack find -c
should be able to show details of concretized specs within environments
-
- Davide: building Python, numpy, scipy w/ Intel compiler and MKL
- Works, performance (e.g. for SVD) is similar to pre-compiled
- But slower than GCC-compiled version
- In some cases faster, but generally slower
- Installing numpy w/pip on top of gcc-based spack-installed python is fast
- Intel-based spack installed python is slow
- Intel had instructions for compiling
- But as of now tells users not to use it
- See: https://www.intel.com/content/www/us/en/developer/articles/technical/numpyscipy-with-intel-mkl.html
- Intel website has a 1GB binary which provides numpy etc. that seems to work well
- But: This doesn't specify associated version (only built for Python 3.9)
- Prentice: Some things seem slow, can I get more info?
- Tammy:
spack -d
e.g.spack -d concretize
,spack -d install
, etc.
- Tammy: