versions greater than 1.3.1 downgrade libblas, libcblas and liblapack

[ysarri]

[oleksandr-pavlyk]

Please provide steps to create the initial conda environment, and the command used to modify it that prompts response recorded in the screenshot.

[ysarri]

Thank you for your reply, directly after installing Miniforge3 23.1.0-4 in Windows 11:

conda create -n py11test python=3.11.4

conda install -n py11test numpy

conda install -n py11test mkl_fft

This latest command downgrades the versions of these three libraries resulting in much lower performance for some numpy functions. (By installing mkl_fft directly, without the numpy intermediate, the libraries installed are also the ones from netlib and not the mkl ones). It also happens for mambaforge. I have not tested miniconda.

[oleksandr-pavlyk]

@ysarri The trouble is that latest mkl_fft was built using a later version of MKL than the NumPy was.

Using conda install mkl_fft mkl=2022.1.0 -c conda-forge --override-channels to keep the same version of MKL as used by NumPy, I am getting an older mkl_fft=1.3.1, but the blas/cblas/clapack are not downgraded.