Re: [AArch64] Add precision choices for the reciprocal square root approximation

[Evandro Menezes <e dot menezes at samsung dot com>]

On 03/18/16 10:21, Wilco Dijkstra wrote:
> Hi Evandro,
>
> > For example, though this approximation is improves the performance
> > noticeably for DF on A57, for SF, not so much, if at all.
> I'm still skeptical that you ever can get any gain on scalars. I bet the only gain is on
> 4x vectorized floats.

I created a simple test that loops around an inline asm version of the 
Newton series using scalar insns and got these results on A57:

   1/sqrt(x):    18290898/s
   Fast:         45896823/s

   1/sqrtf(x):   69618490/s
   Fast:         61865874/s


> So what I would like to see is this implemented in a more general way. We should
> be able choose whether to expand depending on the mode - including whether it is
> vectorized. For example enable on V4SFmode and maybe V2DFmode, but not
> on any scalars.
>
> Then we'd add new CPU tuning settings for division, sqrt and rsqrt (rather than adding lots
> of extra tune flags).

If I understood you correctly, would something like coarse tuning flags 
along with target-specific cost or parameters tables be what you have in 
mind?

> Note the md file should call a function in aarch64.c to decide whether to
> expand or not (your division approximation patch makes the decision in the md file which
> does not seem a good idea).

I agree.  Will modify it.

Thank you,

--
Evandro Menezes