compile: update gcc flags

[jasperzhong]

I tried Ofast instead of O2. I do observe somewhat performance gain in my experiments, roughly from 4% to 20%, without accuracy loss. Ofast enables all O3 optimizations and some non-standard optimizations according to gcc docs.

Also, I tried march=native to replace avx flags. According to gcc docs, it enables all instruction subsets supported by the local machine, including avx if possible. So I suppose the flag is more general.

[ymjiang]

Thank you! We will look into this and see if it also boosts our performance.

[bobzhuyb]

@ymjiang Have you tested this?

[ymjiang]

@bobzhuyb Yes, but did not see performance gain.

@vycezhong Can you please share more context, i.e., In what kind of scenario did you see 4-20% improvement?

[jasperzhong]

That's exact the same MNIST exp I mentioned in the #225.

I just tried to reproduce the results but failed. Maybe 4%-20% is due to some turbulence.