Matti Picus
Matti Picus
> safely remove ... everything 1.15.x and below. Done
We upgraded wheel builds to manylinux_2_28, so GLIBC is now 2.28. Routines that were blocklisted under GLIBC 2.17 now will use native glibc routines: https://github.com/numpy/numpy/blob/a811c05394a2a052a5b95b130f415bea4ccd542e/numpy/_core/src/common/npy_config.h#L149-L164
As the intel compiler is proprietary, we would need someone with a proper license to work through this. I do not have such a license.
Somehow the installer works differently via a remote desktop, it seems to work better with direct access.
The latest oneAPI compiler suite I installed came with `icx.exe` and not `icl.exe`. I needed to do this to build: ``` \setvars.bat intel64 set CC=icx.exe set CXX=icx.exe pip install scipy-openblas64...
Right, but we no longer use `site.cfg`
There are many warnings from CPython about a missing declaration of `struct timeval`. This was fixed for CPython 3.12, so I would recommend using that for compilation with the oneAPI...
There are also many warnings about "warning: explicit comparison with NaN in fast floating point mode". It seems meson should use at least `-fhonor-infinities -fhonor-nans` according to [this issue in...
OK. I will add that to #25044
Maybe related: numpy/numpy#29884 where using `nthreads == ncores` slows down np.matmul on 100x100 double matrices. This in turn uses OpenBLAS `syrk` or `gemm` routines. Limiting nthreads to `ncores - 2`...