cyyever

I need this feature

@pytorchbot merge

> Test regressions seem relevant, aren't they? Most of them failed in test_torchinductor. I have to rebase and rerun them when the master become green.

@malfet Clang-7 failed with Openmp, but clang-7 is too old, I can't setup a local environment to reproduce it easily...

@malfet Can you take some time to review this PR? I have managed to pass all tests. The MKL detection is also enhanced and OMP prefers MKL if presented.

@malfet Thank you for acceptance, I am preparing more CMake refactoring

@malfet The failed case export USE_CUPTI_SO=0 in CI script. But CMake requires USE_CUPTI_SO=1, we should change the CI script. However, I can't grep 'USE_CUPTI_SO=0' in git repository. Maybe it comes...

@malfet USE_CUPTI_SO=0 and centos7 combination needs to link pthread to run the cmake test. I have added it.

@pytorchmergebot merge

cc @Skylion007 @zou3519 The tests all pass.