EVP errors on native-arm builds on apple silicon

[kburns]

It looks like theres a variety of problems with both the dense and sparse eigenvalue solvers when using a native-arm build on apple silicon. For now I've changed the conda install scripts to build the stack for x86 on apple silicon by default. This comes at a 1.5-2x performance penalty, but seems much more robust for eigenvalue problems.

For dense problems, it seems like pinning libopenblas < 0.3.20 helps things, otherwise there are ggev errors. But for sparse problems, there are still ARPACK convergence errors for problems that work fine on x86.

[kburns]

Still getting ggev errors with libopenblas 0.3.21.

[kburns]

Opened an issue on the openblas repo: https://github.com/xianyi/OpenBLAS/issues/3735.

[martin-frbg]

Less expensive workaround should be to build OpenBLAS on M1 with TARGET=ARMV8

[kburns]

The dense eigenvalue (ggev) errors are fixed with https://github.com/conda-forge/openblas-feedstock/pull/144 and https://github.com/conda-forge/openblas-feedstock/pull/145, so builds 1+ for libopenblas 0.3.21 from conda-forge should work. Sparse problems still persist.

[kburns]

Sparse problems seem to be fixed with scipy 1.9.1, so native-arm conda builds now pass all tests and run the examples. They still don't have higher-precision support in numpy, though, so we should add tests that can explicitly tell when e.g. spherical harmonic polynomial construction is failing due to underflow issues with the envelopes.

[BenWibking]

Is this fixed with the conda-forge packages for Dedalus3? (The documentation still says to install the x86_64 packages.)

[kburns]

No, we still need to work around the lack of higher-precision support on apple silicon for spherical problems. But for Cartesian problems I expect everything to work fine. But that said, we haven't tested thoroughly on arm builds yet.