Description of problem:



Version-Release number of selected component (if applicable):


How reproducible: always


Steps to Reproduce:
1. Install python3-numpy on Fedora 33
2. Run python3 script https://gist.githubusercontent.com/markus-beuckelmann/8bc25531b11158431a5b09a45abd6276/raw/660904cb770197c3c841ab9b7084657b1aea5f32/numpy-benchmark.py
3. Note down the execution times

# Now change backend to openblas-threads (as this one is used directly in F32)
4. Install flexiblas-openblas-threads
5. Change backend in /etc/flexiblasrc to "openblas-threads"
6. Run the Python script 2 again
7. Note down execution times

8. Install python3-numpy on F32 or build on F33 directly against threaded openblas
9. Run script of step 2 again
10. Note down execution times

11. Compare results

Actual results: 
Some operations like eigendecompostition are much slower with flexiblas in between

F33 with flexiblas and openblas-threads backend:
Dotted two 4096x4096 matrices in 3.36 s.
Dotted two vectors of length 524288 in 0.75 ms.
SVD of a 2048x1024 matrix in 10.47 s.
Cholesky decomposition of a 2048x2048 matrix in 1.62 s.
Eigendecomposition of a 2048x2048 matrix in 51.59 s.

This was obtained using the following Numpy configuration:
blas_mkl_info:
  NOT AVAILABLE
blis_info:
  NOT AVAILABLE
openblas_info:
    libraries = ['flexiblas', 'flexiblas']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
    runtime_library_dirs = ['/usr/lib64']
blas_opt_info:
    libraries = ['flexiblas', 'flexiblas']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
    runtime_library_dirs = ['/usr/lib64']
lapack_mkl_info:
  NOT AVAILABLE
openblas_lapack_info:
    libraries = ['flexiblas', 'flexiblas']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
    runtime_library_dirs = ['/usr/lib64']
lapack_opt_info:
    libraries = ['flexiblas', 'flexiblas']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
    runtime_library_dirs = ['/usr/lib64']

F33 with python3-numpy linked against libopenblasp directly (or F32 default, very similar):
Dotted two 4096x4096 matrices in 2.95 s.
Dotted two vectors of length 524288 in 0.57 ms.
SVD of a 2048x1024 matrix in 1.70 s.
Cholesky decomposition of a 2048x2048 matrix in 0.22 s.
Eigendecomposition of a 2048x2048 matrix in 14.05 s.

This was obtained using the following Numpy configuration:
blas_mkl_info:
  NOT AVAILABLE
blis_info:
  NOT AVAILABLE
openblas_info:
    libraries = ['openblasp', 'openblasp']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
    runtime_library_dirs = ['/usr/lib64']
blas_opt_info:
    libraries = ['openblasp', 'openblasp']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
    runtime_library_dirs = ['/usr/lib64']
lapack_mkl_info:
  NOT AVAILABLE
openblas_lapack_info:
    libraries = ['openblasp', 'openblasp']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
    runtime_library_dirs = ['/usr/lib64']
lapack_opt_info:
    libraries = ['openblasp', 'openblasp']
    library_dirs = ['/usr/lib64']
    language = c
    define_macros = [('HAVE_CBLAS', None)]
    runtime_library_dirs = ['/usr/lib64']




Expected results: Performance is similar with and without flexiblas