Regarding the math libraries included in a Psi4 installation, I see that these include MKL. Can these be used optimally by AMD CPUs when executing Psi4 code? If not, would it be possible to use the AOCL math libraries that have been optimized for AMD CPUs in conjunction with a Psi4 installation?
The latest MKL (OneAPI versions) work well on AMD CPUs to my knowledge.
In principle you can already build psi4 with any BLAS library you want, but there are some hurdles in the sense that also numpy ideally should use the same BLAS library. Another issue is the use of matching threading libraries. It can get a bit tricky with different compilers and different math libraries so that for safely and smooth operations we currently optimized the build process for the MKL.
But we are looking into expanding math library support.
Thank you for your encouraging reply! In my case, I chose the option of installing a pre-built psi4 into a newly created conda python environment. In my preliminary tests, psi4 appears to be running well on a system with an AMD TR Pro 5975wx CPU and 2 x Nvidia A-6000 GPU cards.