Intel MKL

[ischoegl]

Supersedes #29

Some Cantera versions on the cantera channel are compiled with Intel MKL support, which requires correct versions. Due to recurring issues (example: #29), MKL is currently no longer used for compiled packages.

Originally posted by @speth in #29 (comment)

Recent builds of Numpy seem to have separate builds for different MKL versions. Partial output of mamba search --info -c defaults numpy=1.23.5 shows:

numpy 1.23.5 py310h5f9d8c6_1
----------------------------
file name   : numpy-1.23.5-py310h5f9d8c6_1.conda
name        : numpy
version     : 1.23.5
build       : py310h5f9d8c6_1
build number: 1
size        : 10 KB
license     : BSD-3-Clause
subdir      : linux-64
url         : https://repo.anaconda.com/pkgs/main/linux-64/numpy-1.23.5-py310h5f9d8c6_1.conda
md5         : 875377e83010e75870affa1797444ce7
timestamp   : 2023-05-01 15:03:28 UTC
dependencies: 
  - blas 1.0 mkl
  - libgcc-ng >=11.2.0
  - libstdcxx-ng >=11.2.0
  - mkl >=2023.1.0,<2024.0a0
  - mkl-service >=2.3.0,<3.0a0
  - mkl_fft
  - mkl_random
  - numpy-base 1.23.5 py310hb5e798b_1
  - python >=3.10,<3.11.0a0

numpy 1.23.5 py310hd5efca6_0
----------------------------
file name   : numpy-1.23.5-py310hd5efca6_0.conda
name        : numpy
version     : 1.23.5
build       : py310hd5efca6_0
build number: 0
size        : 10 KB
license     : BSD-3-Clause
subdir      : linux-64
url         : https://repo.anaconda.com/pkgs/main/linux-64/numpy-1.23.5-py310hd5efca6_0.conda
md5         : 14704ba32c38d4d8d78e5812f274a795
timestamp   : 2022-12-29 19:03:16 UTC
dependencies: 
  - blas 1.0 mkl
  - libgcc-ng >=11.2.0
  - libstdcxx-ng >=11.2.0
  - mkl >=2021.4.0,<2022.0a0
  - mkl-service >=2.3.0,<3.0a0
  - mkl_fft
  - mkl_random
  - numpy-base 1.23.5 py310h8e6c178_0
  - python >=3.10,<3.11.0a0

So, one build that specifically requires MKL 2021 and one that specifically requires MKL 2023, in addition to a build for OpenBLAS (and a conspicuous absence of a build for MKL 2022, which was the source of some of the original trouble reported here). I think for us, all the extra builds really aren't worth it, and we should just stick to using OpenBLAS for the packages on the cantera channel.

[speth]

As a bit of an update, for Cantera 3.0, we ended up going the following:

• For macOS, use the native Accelerate framework to provide BLAS/LAPACK functions
• For Linux, use OpenBLAS
• For Windows, use MKL, since OpenBLAS is not available from the defaults channel
  We're only building for one version of MKL for Windows. We'll see what happens next time there's a major update to the MKL packages.