Using simpler march flag in GCC for AVX512

[stefanatwork]

No description provided.

[svenwoop]

Looks good.

[christophriesinger]

Just for clarity, -march=skx and -march=skylake-avx512 are synonymous in Clang.

The -mprefer-vector-width=256 is redundant if -march=skylake-avx512 as can be seen here.

My question: Should we keep this limitation as spr, gnr, and other targets don't use TuningPrefer256Bit?
This limitation is only relevant for the auto-vectorizer. Is there significant Embree functionality not implemented by intrinsics and thus vectorized by the auto-vectorizer?

[svenwoop]

Good catch, @stefanatwork the prefer the -march=skx already includes the prefer256.

Embree does not rely on the auto vectorizer, thus there is no performance relevant part that should be affected by that perfer256 flag. If it gets removed you essentially make Embree regress (as just one 512bit instruction already causes issues on SKX) on SKX, with no gains on newer architectures.