RE: [PATCH 0/5] Added optimized memcpy/memmove/memset for A64FX

["naohirot@fujitsu.com" <naohirot@fujitsu.com>]

Hi Wilco-san,

Let me focus on DC_ZVA and L1/L2 prefetch in this mail.

> From: Wilco Dijkstra <Wilco.Dijkstra@arm.com>

> > Without DC_VZA and L2 prefetch, memcpy and memset performance degraded
> over 4MB.
> 
> > DC_VZA and L2 prefetch have to be pair, only DC_VZA or only L2 prefetch
> doesn't get any improvement.
> 
> That seems odd. Was that using the L1 prefetch with the L2 distance? It seems to
> me one of the L1 or L2 prefetches is unnecessary. 

I tested the following 4 cases.
The result was that Case 4 is the best.
Case 2 and 3 were almost same as Case 1.
Case 4 [1] improved the performance in the size range more than 4MB from Case 1
7.5-10 GB/sec [2] to 10-10.5 GB/sec [3].

Case 1: DC_ZVA + L1 prefetch + L2 + prefetch [2]
Case 2: DC_ZVA + L1 prefetch
Case 3: DC_ZVA + L2 prefetch
Case 4: DC_ZVA only [3]

[1] https://github.com/NaohiroTamura/glibc/commit/d57bed764a45383dfea8265d6a384646f4f07eed
[2] https://drive.google.com/file/d/1ws3lTLzMFK3lLrrwxVFvriERrs-IKdP9/view
[3] https://drive.google.com/file/d/1g7nuFOtkFw3b5INcAfuuv2lVODmASm-G/view


>                                                Also why would the DC_ZVA
> need to be done so early? It seems to me that cleaning the cacheline just before
> you write it works best since that avoids accidentally replacing it.
> 

Yes, I moved it closer, please look at the change [1].

> > Without DC_VZA and L2 prefetch, memmove didn't degraded over 4MB.
> >
> > The reason why I didn't implement DC_VZA and L2 prefetch is that
> > memmove calls memcpy in most cases, and memmove code only handles
> backward copy.
> > Maybe most of memmove-large benchtest cases are backward copy, I need to
> check.
> 
> Most of the memmove tests do indeed overlap (so DC_ZVA does not work).
> However it also shows that it performs well across the L2 cache size range
> without any prefetch or DC_ZVA.

That's right, I confirmed that only DC_ZVA was necessary [1].

Next, I'll remove redundant instructions.

Thanks.
Naohiro