search

adc-connect / adcc

adcc: Seamlessly connect your program to ADC
https://adc-connect.org
GNU General Public License v3.0
32 stars 19 forks source link

mkl not properly built with conda #151

Open t-fransson opened 2 years ago

t-fransson commented 2 years ago

It seems like conda is no longer able to properly install mkl-libraries during the build, resulting in error codes such as this:

INTEL MKL ERROR: /home/thomas/Notebook/anaconda/envs/adcc_test/lib/python3.10/site-packages/../.././libmkl_avx2.so.1: undefined symbol: mkl_sparse_optimize_bsr_trsm_i8. Intel MKL FATAL ERROR: Cannot load libmkl_avx2.so.1 or libmkl_def.so.1. INTEL MKL ERROR: /home/thomas/Notebook/anaconda/envs/adcc_test/lib/python3.10/site-packages/../.././libmkl_avx2.so.1: undefined symbol: mkl_sparse_optimize_bsr_trsm_i8. Intel MKL FATAL ERROR: Cannot load libmkl_avx2.so.1 or libmkl_def.so.1.

Explicitly adding the flollowing in a .yml-file:

- libblas==_mkl

...makes the calculations run, but they sometimes have issues with properly converging the correct eigenstates, yielding resulting eigenenergies of very high negative values... This seems to primarily be the case when using small basis sets.

For VeloxChem I think they had similar issues with conda install suddenly not working, and solved this by some additional explicit requirement of mkl-libraries. I guess something like that and new conda builds might be needed? Can others verify that they have similar issues?