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

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

Hi Wilco-san,

Let me focus on L1_prefetch in this mail.

> From: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
 
> > Memcpy/memmove uses 8, 4, 2 unrolls, and memset uses 32, 8, 4, 2 unrolls.
> > This unroll configuration recorded the highest performance.

When I tested "4 unrolls", I modified the source code [1][2] in the mail [0]
such as followings:
in case of memcpy, 
   I commented out L(unroll8), L(unroll2), and left L(unroll4), L(unroll1) and L(last),
In case of memmove,
   I commented out L(bwd_unroll8), L(bwd_unroll2), and left L(bwd_unroll4), L(bwd_unroll1) and L(bwd_last),
In case of memset, 
   I commented out L(unroll32), L(unroll8), L(unroll2), and left L(unroll4), L(unroll1) and L(last).

[0] https://sourceware.org/pipermail/libc-alpha/2021-April/125002.html
[1] https://github.com/NaohiroTamura/glibc/blob/ec0b55a855529f75bd6f280e59dc2b1c25640490/sysdeps/aarch64/multiarch/memcpy_a64fx.S
[2] https://github.com/NaohiroTamura/glibc/blob/ec0b55a855529f75bd6f280e59dc2b1c25640490/sysdeps/aarch64/multiarch/memset_a64fx.S

> > In case that Memcpy/memmove uses 4 unrolls, and memset uses 4 unrolls,
> > The performance degraded minus 5 to 15 Gbps/sec at the peak.
> 
> So this is the L(L1_vl_64) loop right? I guess the problem is the large number of

So this is NOT the L(L1_vl_64) loop, but L(vl_agnostic).

> prefetches and all the extra code that is not strictly required (you can remove 5
> redundant mov/cmp instructions from the loop). Also assuming prefetching helps
> here (the good memmove results suggest it's not needed), prefetching directly
> into L1 should be better than first into L2 and then into L1. So I don't see a good
> reason why 4x unrolling would have to be any slower.

I tried to remove L(L1_prefetch) from both memcpy and memset, and also
I tried to remove L2 prefetch instructions (prfm pstl2keep and pldl2keep) in
L(L1_prefetch) from both memcpy and memset.

In case of memcpy, both removing L(L1_prefetch)[3] and removing L2 prefetch
instruction from L(L1_prefetch) increased the performance of the size range 64KB-4MB
from 18-20 GB/sec [4] to 20-22 GB/sec [5].

[3] https://github.com/NaohiroTamura/glibc/commit/22612299247e64dbffd62aa186513bde7328d104
[4] https://drive.google.com/file/d/1hGWz4eAYWc1ktdw74rzDPxtQQ48P0-Hv/view
[5] https://drive.google.com/file/d/11Pt1mWSCN2LBPHxXUE-rs7Q6JhtBfpyQ/view

In case of memset, removing L(L1_prefetch)[6] decreased the performance of the size range
128KB-4MB from 22-24 GB/sec [7] to 20-22 GB/sec[8].
But removing L2 prefetch instruction (prfm pstl2keep) in L(L1_prefetch) [9] kept the same
performance of the size range 128KB-4MB as 22-24 GB/sec [10].

[6] https://github.com/NaohiroTamura/glibc/blob/22612299247e64dbffd62aa186513bde7328d104/sysdeps/aarch64/multiarch/memset_a64fx.S#L146-L163
   Commented out L146-L163, I didn't commit because of decreasing the performance.
[7] https://drive.google.com/file/d/1MT1d2aBxSoYrzQuRZtv4U9NCXV4ZwHsJ/view
[8] https://drive.google.com/file/d/1qUzYklLvgXTZbP1wm9n4VryF3bgUOplo/view
[9] https://github.com/NaohiroTamura/glibc/commit/cc478c96bac051c9b98b9d9a1ae6f38326f77645
[10] https://drive.google.com/file/d/1bPKHFWyhzNWXX7A_S6_UpZ2BwP2QAJK4/view

In conclusion, I adopt to remove L(L1_prefetch) from memcpy [3] and to remove L2 prefetch
instruction (prfm pstl2keep) from L(L1_prefetch) [9].

Thanks.
Naohiro