...
One established solution for relaxing composability problems is to provide all the alternatives as threading layers. Therefore, user can choose at load or install time what threading runtime to use. MKL, IPP, Numba, Python multiprocessing pools (via tbb4py module) do this already. There is a desire to implement similar approach for: Numpy/Scipy, Cython, DAAL, TensorFlow, PyTorch, Caffe, to name a few just in Python ecosystem.
More on this topic and what was done for Python: http://conference.scipy.org/proceedings/scipy2018/pdfs/anton_malakhov.pdf
Parallel settings orchestration
...