glibc/sysdeps/aarch64/multiarch/memcpy_a64fx.S
Naohiro Tamura fa527f345c aarch64: Added optimized memcpy and memmove for A64FX
This patch optimizes the performance of memcpy/memmove for A64FX [1]
which implements ARMv8-A SVE and has L1 64KB cache per core and L2 8MB
cache per NUMA node.

The performance optimization makes use of Scalable Vector Register
with several techniques such as loop unrolling, memory access
alignment, cache zero fill, and software pipelining.

SVE assembler code for memcpy/memmove is implemented as Vector Length
Agnostic code so theoretically it can be run on any SOC which supports
ARMv8-A SVE standard.

We confirmed that all testcases have been passed by running 'make
check' and 'make xcheck' not only on A64FX but also on ThunderX2.

And also we confirmed that the SVE 512 bit vector register performance
is roughly 4 times better than Advanced SIMD 128 bit register and 8
times better than scalar 64 bit register by running 'make bench'.

[1] https://github.com/fujitsu/A64FX

Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
Reviewed-by: Szabolcs Nagy <Szabolcs.Nagy@arm.com>
2021-05-27 09:47:53 +01:00

407 lines
11 KiB
ArmAsm

/* Optimized memcpy for Fujitsu A64FX processor.
Copyright (C) 2021 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library. If not, see
<https://www.gnu.org/licenses/>. */
#include <sysdep.h>
/* Assumptions:
*
* ARMv8.2-a, AArch64, unaligned accesses, sve
*
*/
#define L2_SIZE (8*1024*1024)/2 // L2 8MB/2
#define CACHE_LINE_SIZE 256
#define ZF_DIST (CACHE_LINE_SIZE * 21) // Zerofill distance
#define dest x0
#define src x1
#define n x2 // size
#define tmp1 x3
#define tmp2 x4
#define tmp3 x5
#define rest x6
#define dest_ptr x7
#define src_ptr x8
#define vector_length x9
#define cl_remainder x10 // CACHE_LINE_SIZE remainder
#if HAVE_AARCH64_SVE_ASM
# if IS_IN (libc)
# define MEMCPY __memcpy_a64fx
# define MEMMOVE __memmove_a64fx
.arch armv8.2-a+sve
.macro dc_zva times
dc zva, tmp1
add tmp1, tmp1, CACHE_LINE_SIZE
.if \times-1
dc_zva "(\times-1)"
.endif
.endm
.macro ld1b_unroll8
ld1b z0.b, p0/z, [src_ptr, #0, mul vl]
ld1b z1.b, p0/z, [src_ptr, #1, mul vl]
ld1b z2.b, p0/z, [src_ptr, #2, mul vl]
ld1b z3.b, p0/z, [src_ptr, #3, mul vl]
ld1b z4.b, p0/z, [src_ptr, #4, mul vl]
ld1b z5.b, p0/z, [src_ptr, #5, mul vl]
ld1b z6.b, p0/z, [src_ptr, #6, mul vl]
ld1b z7.b, p0/z, [src_ptr, #7, mul vl]
.endm
.macro stld1b_unroll4a
st1b z0.b, p0, [dest_ptr, #0, mul vl]
st1b z1.b, p0, [dest_ptr, #1, mul vl]
ld1b z0.b, p0/z, [src_ptr, #0, mul vl]
ld1b z1.b, p0/z, [src_ptr, #1, mul vl]
st1b z2.b, p0, [dest_ptr, #2, mul vl]
st1b z3.b, p0, [dest_ptr, #3, mul vl]
ld1b z2.b, p0/z, [src_ptr, #2, mul vl]
ld1b z3.b, p0/z, [src_ptr, #3, mul vl]
.endm
.macro stld1b_unroll4b
st1b z4.b, p0, [dest_ptr, #4, mul vl]
st1b z5.b, p0, [dest_ptr, #5, mul vl]
ld1b z4.b, p0/z, [src_ptr, #4, mul vl]
ld1b z5.b, p0/z, [src_ptr, #5, mul vl]
st1b z6.b, p0, [dest_ptr, #6, mul vl]
st1b z7.b, p0, [dest_ptr, #7, mul vl]
ld1b z6.b, p0/z, [src_ptr, #6, mul vl]
ld1b z7.b, p0/z, [src_ptr, #7, mul vl]
.endm
.macro stld1b_unroll8
stld1b_unroll4a
stld1b_unroll4b
.endm
.macro st1b_unroll8
st1b z0.b, p0, [dest_ptr, #0, mul vl]
st1b z1.b, p0, [dest_ptr, #1, mul vl]
st1b z2.b, p0, [dest_ptr, #2, mul vl]
st1b z3.b, p0, [dest_ptr, #3, mul vl]
st1b z4.b, p0, [dest_ptr, #4, mul vl]
st1b z5.b, p0, [dest_ptr, #5, mul vl]
st1b z6.b, p0, [dest_ptr, #6, mul vl]
st1b z7.b, p0, [dest_ptr, #7, mul vl]
.endm
.macro shortcut_for_small_size exit
// if rest <= vector_length * 2
whilelo p0.b, xzr, n
whilelo p1.b, vector_length, n
b.last 1f
ld1b z0.b, p0/z, [src, #0, mul vl]
ld1b z1.b, p1/z, [src, #1, mul vl]
st1b z0.b, p0, [dest, #0, mul vl]
st1b z1.b, p1, [dest, #1, mul vl]
ret
1: // if rest > vector_length * 8
cmp n, vector_length, lsl 3 // vector_length * 8
b.hi \exit
// if rest <= vector_length * 4
lsl tmp1, vector_length, 1 // vector_length * 2
whilelo p2.b, tmp1, n
incb tmp1
whilelo p3.b, tmp1, n
b.last 1f
ld1b z0.b, p0/z, [src, #0, mul vl]
ld1b z1.b, p1/z, [src, #1, mul vl]
ld1b z2.b, p2/z, [src, #2, mul vl]
ld1b z3.b, p3/z, [src, #3, mul vl]
st1b z0.b, p0, [dest, #0, mul vl]
st1b z1.b, p1, [dest, #1, mul vl]
st1b z2.b, p2, [dest, #2, mul vl]
st1b z3.b, p3, [dest, #3, mul vl]
ret
1: // if rest <= vector_length * 8
lsl tmp1, vector_length, 2 // vector_length * 4
whilelo p4.b, tmp1, n
incb tmp1
whilelo p5.b, tmp1, n
b.last 1f
ld1b z0.b, p0/z, [src, #0, mul vl]
ld1b z1.b, p1/z, [src, #1, mul vl]
ld1b z2.b, p2/z, [src, #2, mul vl]
ld1b z3.b, p3/z, [src, #3, mul vl]
ld1b z4.b, p4/z, [src, #4, mul vl]
ld1b z5.b, p5/z, [src, #5, mul vl]
st1b z0.b, p0, [dest, #0, mul vl]
st1b z1.b, p1, [dest, #1, mul vl]
st1b z2.b, p2, [dest, #2, mul vl]
st1b z3.b, p3, [dest, #3, mul vl]
st1b z4.b, p4, [dest, #4, mul vl]
st1b z5.b, p5, [dest, #5, mul vl]
ret
1: lsl tmp1, vector_length, 2 // vector_length * 4
incb tmp1 // vector_length * 5
incb tmp1 // vector_length * 6
whilelo p6.b, tmp1, n
incb tmp1
whilelo p7.b, tmp1, n
ld1b z0.b, p0/z, [src, #0, mul vl]
ld1b z1.b, p1/z, [src, #1, mul vl]
ld1b z2.b, p2/z, [src, #2, mul vl]
ld1b z3.b, p3/z, [src, #3, mul vl]
ld1b z4.b, p4/z, [src, #4, mul vl]
ld1b z5.b, p5/z, [src, #5, mul vl]
ld1b z6.b, p6/z, [src, #6, mul vl]
ld1b z7.b, p7/z, [src, #7, mul vl]
st1b z0.b, p0, [dest, #0, mul vl]
st1b z1.b, p1, [dest, #1, mul vl]
st1b z2.b, p2, [dest, #2, mul vl]
st1b z3.b, p3, [dest, #3, mul vl]
st1b z4.b, p4, [dest, #4, mul vl]
st1b z5.b, p5, [dest, #5, mul vl]
st1b z6.b, p6, [dest, #6, mul vl]
st1b z7.b, p7, [dest, #7, mul vl]
ret
.endm
ENTRY (MEMCPY)
PTR_ARG (0)
PTR_ARG (1)
SIZE_ARG (2)
L(memcpy):
cntb vector_length
// shortcut for less than vector_length * 8
// gives a free ptrue to p0.b for n >= vector_length
shortcut_for_small_size L(vl_agnostic)
// end of shortcut
L(vl_agnostic): // VL Agnostic
mov rest, n
mov dest_ptr, dest
mov src_ptr, src
// if rest >= L2_SIZE && vector_length == 64 then L(L2)
mov tmp1, 64
cmp rest, L2_SIZE
ccmp vector_length, tmp1, 0, cs
b.eq L(L2)
L(unroll8): // unrolling and software pipeline
lsl tmp1, vector_length, 3 // vector_length * 8
.p2align 3
cmp rest, tmp1
b.cc L(last)
ld1b_unroll8
add src_ptr, src_ptr, tmp1
sub rest, rest, tmp1
cmp rest, tmp1
b.cc 2f
.p2align 3
1: stld1b_unroll8
add dest_ptr, dest_ptr, tmp1
add src_ptr, src_ptr, tmp1
sub rest, rest, tmp1
cmp rest, tmp1
b.ge 1b
2: st1b_unroll8
add dest_ptr, dest_ptr, tmp1
.p2align 3
L(last):
whilelo p0.b, xzr, rest
whilelo p1.b, vector_length, rest
b.last 1f
ld1b z0.b, p0/z, [src_ptr, #0, mul vl]
ld1b z1.b, p1/z, [src_ptr, #1, mul vl]
st1b z0.b, p0, [dest_ptr, #0, mul vl]
st1b z1.b, p1, [dest_ptr, #1, mul vl]
ret
1: lsl tmp1, vector_length, 1 // vector_length * 2
whilelo p2.b, tmp1, rest
incb tmp1
whilelo p3.b, tmp1, rest
b.last 1f
ld1b z0.b, p0/z, [src_ptr, #0, mul vl]
ld1b z1.b, p1/z, [src_ptr, #1, mul vl]
ld1b z2.b, p2/z, [src_ptr, #2, mul vl]
ld1b z3.b, p3/z, [src_ptr, #3, mul vl]
st1b z0.b, p0, [dest_ptr, #0, mul vl]
st1b z1.b, p1, [dest_ptr, #1, mul vl]
st1b z2.b, p2, [dest_ptr, #2, mul vl]
st1b z3.b, p3, [dest_ptr, #3, mul vl]
ret
1: lsl tmp1, vector_length, 2 // vector_length * 4
whilelo p4.b, tmp1, rest
incb tmp1
whilelo p5.b, tmp1, rest
incb tmp1
whilelo p6.b, tmp1, rest
incb tmp1
whilelo p7.b, tmp1, rest
ld1b z0.b, p0/z, [src_ptr, #0, mul vl]
ld1b z1.b, p1/z, [src_ptr, #1, mul vl]
ld1b z2.b, p2/z, [src_ptr, #2, mul vl]
ld1b z3.b, p3/z, [src_ptr, #3, mul vl]
ld1b z4.b, p4/z, [src_ptr, #4, mul vl]
ld1b z5.b, p5/z, [src_ptr, #5, mul vl]
ld1b z6.b, p6/z, [src_ptr, #6, mul vl]
ld1b z7.b, p7/z, [src_ptr, #7, mul vl]
st1b z0.b, p0, [dest_ptr, #0, mul vl]
st1b z1.b, p1, [dest_ptr, #1, mul vl]
st1b z2.b, p2, [dest_ptr, #2, mul vl]
st1b z3.b, p3, [dest_ptr, #3, mul vl]
st1b z4.b, p4, [dest_ptr, #4, mul vl]
st1b z5.b, p5, [dest_ptr, #5, mul vl]
st1b z6.b, p6, [dest_ptr, #6, mul vl]
st1b z7.b, p7, [dest_ptr, #7, mul vl]
ret
L(L2):
// align dest address at CACHE_LINE_SIZE byte boundary
mov tmp1, CACHE_LINE_SIZE
ands tmp2, dest_ptr, CACHE_LINE_SIZE - 1
// if cl_remainder == 0
b.eq L(L2_dc_zva)
sub cl_remainder, tmp1, tmp2
// process remainder until the first CACHE_LINE_SIZE boundary
whilelo p1.b, xzr, cl_remainder // keep p0.b all true
whilelo p2.b, vector_length, cl_remainder
b.last 1f
ld1b z1.b, p1/z, [src_ptr, #0, mul vl]
ld1b z2.b, p2/z, [src_ptr, #1, mul vl]
st1b z1.b, p1, [dest_ptr, #0, mul vl]
st1b z2.b, p2, [dest_ptr, #1, mul vl]
b 2f
1: lsl tmp1, vector_length, 1 // vector_length * 2
whilelo p3.b, tmp1, cl_remainder
incb tmp1
whilelo p4.b, tmp1, cl_remainder
ld1b z1.b, p1/z, [src_ptr, #0, mul vl]
ld1b z2.b, p2/z, [src_ptr, #1, mul vl]
ld1b z3.b, p3/z, [src_ptr, #2, mul vl]
ld1b z4.b, p4/z, [src_ptr, #3, mul vl]
st1b z1.b, p1, [dest_ptr, #0, mul vl]
st1b z2.b, p2, [dest_ptr, #1, mul vl]
st1b z3.b, p3, [dest_ptr, #2, mul vl]
st1b z4.b, p4, [dest_ptr, #3, mul vl]
2: add dest_ptr, dest_ptr, cl_remainder
add src_ptr, src_ptr, cl_remainder
sub rest, rest, cl_remainder
L(L2_dc_zva):
// zero fill
and tmp1, dest, 0xffffffffffffff
and tmp2, src, 0xffffffffffffff
subs tmp1, tmp1, tmp2 // diff
b.ge 1f
neg tmp1, tmp1
1: mov tmp3, ZF_DIST + CACHE_LINE_SIZE * 2
cmp tmp1, tmp3
b.lo L(unroll8)
mov tmp1, dest_ptr
dc_zva (ZF_DIST / CACHE_LINE_SIZE) - 1
// unroll
ld1b_unroll8 // this line has to be after "b.lo L(unroll8)"
add src_ptr, src_ptr, CACHE_LINE_SIZE * 2
sub rest, rest, CACHE_LINE_SIZE * 2
mov tmp1, ZF_DIST
.p2align 3
1: stld1b_unroll4a
add tmp2, dest_ptr, tmp1 // dest_ptr + ZF_DIST
dc zva, tmp2
stld1b_unroll4b
add tmp2, tmp2, CACHE_LINE_SIZE
dc zva, tmp2
add dest_ptr, dest_ptr, CACHE_LINE_SIZE * 2
add src_ptr, src_ptr, CACHE_LINE_SIZE * 2
sub rest, rest, CACHE_LINE_SIZE * 2
cmp rest, tmp3 // ZF_DIST + CACHE_LINE_SIZE * 2
b.ge 1b
st1b_unroll8
add dest_ptr, dest_ptr, CACHE_LINE_SIZE * 2
b L(unroll8)
END (MEMCPY)
libc_hidden_builtin_def (MEMCPY)
ENTRY (MEMMOVE)
PTR_ARG (0)
PTR_ARG (1)
SIZE_ARG (2)
// remove tag address
// dest has to be immutable because it is the return value
// src has to be immutable because it is used in L(bwd_last)
and tmp2, dest, 0xffffffffffffff // save dest_notag into tmp2
and tmp3, src, 0xffffffffffffff // save src_notag intp tmp3
cmp n, 0
ccmp tmp2, tmp3, 4, ne
b.ne 1f
ret
1: cntb vector_length
// shortcut for less than vector_length * 8
// gives a free ptrue to p0.b for n >= vector_length
// tmp2 and tmp3 should not be used in this macro to keep
// notag addresses
shortcut_for_small_size L(dispatch)
// end of shortcut
L(dispatch):
// tmp2 = dest_notag, tmp3 = src_notag
// diff = dest_notag - src_notag
sub tmp1, tmp2, tmp3
// if diff <= 0 || diff >= n then memcpy
cmp tmp1, 0
ccmp tmp1, n, 2, gt
b.cs L(vl_agnostic)
L(bwd_start):
mov rest, n
add dest_ptr, dest, n // dest_end
add src_ptr, src, n // src_end
L(bwd_unroll8): // unrolling and software pipeline
lsl tmp1, vector_length, 3 // vector_length * 8
.p2align 3
cmp rest, tmp1
b.cc L(bwd_last)
sub src_ptr, src_ptr, tmp1
ld1b_unroll8
sub rest, rest, tmp1
cmp rest, tmp1
b.cc 2f
.p2align 3
1: sub src_ptr, src_ptr, tmp1
sub dest_ptr, dest_ptr, tmp1
stld1b_unroll8
sub rest, rest, tmp1
cmp rest, tmp1
b.ge 1b
2: sub dest_ptr, dest_ptr, tmp1
st1b_unroll8
L(bwd_last):
mov dest_ptr, dest
mov src_ptr, src
b L(last)
END (MEMMOVE)
libc_hidden_builtin_def (MEMMOVE)
# endif /* IS_IN (libc) */
#endif /* HAVE_AARCH64_SVE_ASM */