TensorFlow does not work uisng AVX-512

[hajgato]

On Skylake/KNL TensorFlow compilation does not work. Based on @akesandgren suggestion, it compiles not using AVX512 code-set. With Intel -xCORE-AVX2 or -xAVX2, with GCC -march=native -mno-avx512f solves the problem (tried with both TF 1.4.0 and 1.5.0)

(with GCC using -mavx2 does not use FMA3 instruction set)

[boegel]

@hajgato This? tensorflow/tensorflow#17043

[akesandgren]

I think that @hajgato has the same problem that we have on the KNLs. I.e., that it produces NaNs in the output. The above problem is TF 1.6 and may not apply to TF 1.5

[hajgato]

Its also the NaN problem on Skylake.

[boegel]

Possible fix in tensorflow/tensorflow#15606?