test_ard_accuracy_on_easy_problem fails on CPU with AVX-512

[tbenst]

Description

On sklearn 0.21.3, when I build numpy against MKL, the tests pass if I run on a cpu that does not have AVX-512, but running the same binaries on a CPU with AVX-512 results in a test failure on test_ard_accuracy_on_easy_problem

Versions

numpy 1.17.4
scipy 1.3.3
sklearn 0.21.3

Edit: for what it's worth, quite close:

        # Expect an accuracy of better than 1E-4 in most cases -                
        # Failure-case produces 0.16!                                           
>       assert abs_coef_error < 0.01                                            
E       assert 0.012673147014894082 < 0.01 

So the simple fix would be to increse the threshold to say 0.02. Continuation of #14055

[tbenst]

Problem persists on 0.22.1. I'd love to see #16102 merged.

[jnothman]

Can you please confirm that #16102 results in the test passing on architectures where is is currently failing?

[tbenst]

@jnothman: yes, #16102 fixes the build on my intel Skylake-X i7-9800X

[NicolasHug]

Closing in favor of #15186