Power9 VSX BLAS 3 single precision complex

[edelsohn]

Optimize BLAS 3 single precision complex for Power9 VSX. Tune for Power9 pipeline, cache and memory hierarchy. Exploit new Power9 instructions. Demonstrate equivalent performance to IBM ESSL.

[yuriy-yarosh]

I'd like to work on this and #2078 if @quickwritereader don't mind.

[quickwritereader]

@yuriy-yarosh Hi. Actually, I am going to do both of these after a few weeks break as I am busy preparing for am edx exam. I asked IBM team to split it into four tasks and started from crucial ones and postponed easier ones for later after talks.
I gently ask you to leave them for me as I don't have any rights to prevent you from doing.
Sorry for the late answer I just saw. Hope you understand me

[yuriy-yarosh]

@quickwritereader yup, no problem.

I'll leave some comments regarding the existing Power9 code.

[quickwritereader]

@yuriy-yarosh Thanks for understanding me. I am sorry such happened and I answered late but I was working just had to have a break due to exam preparations.

[yuriy-yarosh]

@quickwritereader long story short but it worth using extended mnemonics for dcbt/dcbtst just to make it a bit more readable.

In general, dcbt cache hints with TH=0b00000 are meh on Power9.
It worth using dcbt's / dcbtst's with TH=0b01010 setting the appropriate units count and Transient flags.

[quickwritereader]

#2172 initial cgemm kernels for power9.
@martin-frbg, for now, we can close this issue as it is far better than power8 kernel on new architecture. in future, I might return to improve it as there is still room for that. And also suggestions by @yuriy-yarosh still had not been applied and tested yet. So though we will close the issue if there be any imrpovements I will pr.
For now, I worked mostly for the correctness of kernels. Besides dgemm others are almost re-written from scratch.

[martin-frbg]

Alright, thanks for your work on this.