[PATCH v2 4/6] aarch64: Added optimized memset for A64FX

[Naohiro Tamura <naohirot@fujitsu.com>]

From: Naohiro Tamura <naohirot@jp.fujitsu.com>

This patch optimizes the performance of memset 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 prefetch.

SVE assembler code for memset 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
---
 sysdeps/aarch64/multiarch/Makefile          |   1 +
 sysdeps/aarch64/multiarch/ifunc-impl-list.c |   5 +-
 sysdeps/aarch64/multiarch/memset.c          |  11 +-
 sysdeps/aarch64/multiarch/memset_a64fx.S    | 268 ++++++++++++++++++++
 4 files changed, 283 insertions(+), 2 deletions(-)
 create mode 100644 sysdeps/aarch64/multiarch/memset_a64fx.S

diff --git a/sysdeps/aarch64/multiarch/Makefile b/sysdeps/aarch64/multiarch/Makefile
index 04c3f17121..7500cf1e93 100644
--- a/sysdeps/aarch64/multiarch/Makefile
+++ b/sysdeps/aarch64/multiarch/Makefile
@@ -2,6 +2,7 @@ ifeq ($(subdir),string)
 sysdep_routines += memcpy_generic memcpy_advsimd memcpy_thunderx memcpy_thunderx2 \
 		   memcpy_falkor memcpy_a64fx \
 		   memset_generic memset_falkor memset_emag memset_kunpeng \
+		   memset_a64fx \
 		   memchr_generic memchr_nosimd \
 		   strlen_mte strlen_asimd
 endif
diff --git a/sysdeps/aarch64/multiarch/ifunc-impl-list.c b/sysdeps/aarch64/multiarch/ifunc-impl-list.c
index 911393565c..4e1a641d9f 100644
--- a/sysdeps/aarch64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/aarch64/multiarch/ifunc-impl-list.c
@@ -37,7 +37,7 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 
   INIT_ARCH ();
 
-  /* Support sysdeps/aarch64/multiarch/memcpy.c and memmove.c.  */
+  /* Support sysdeps/aarch64/multiarch/memcpy.c, memmove.c and memset.c.  */
   IFUNC_IMPL (i, name, memcpy,
 	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_thunderx)
 	      IFUNC_IMPL_ADD (array, i, memcpy, !bti, __memcpy_thunderx2)
@@ -62,6 +62,9 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 	      IFUNC_IMPL_ADD (array, i, memset, (zva_size == 64), __memset_falkor)
 	      IFUNC_IMPL_ADD (array, i, memset, (zva_size == 64), __memset_emag)
 	      IFUNC_IMPL_ADD (array, i, memset, 1, __memset_kunpeng)
+#if HAVE_AARCH64_SVE_ASM
+	      IFUNC_IMPL_ADD (array, i, memset, sve, __memset_a64fx)
+#endif
 	      IFUNC_IMPL_ADD (array, i, memset, 1, __memset_generic))
   IFUNC_IMPL (i, name, memchr,
 	      IFUNC_IMPL_ADD (array, i, memchr, !mte, __memchr_nosimd)
diff --git a/sysdeps/aarch64/multiarch/memset.c b/sysdeps/aarch64/multiarch/memset.c
index 28d3926bc2..48a59574dd 100644
--- a/sysdeps/aarch64/multiarch/memset.c
+++ b/sysdeps/aarch64/multiarch/memset.c
@@ -31,6 +31,9 @@ extern __typeof (__redirect_memset) __libc_memset;
 extern __typeof (__redirect_memset) __memset_falkor attribute_hidden;
 extern __typeof (__redirect_memset) __memset_emag attribute_hidden;
 extern __typeof (__redirect_memset) __memset_kunpeng attribute_hidden;
+#if HAVE_AARCH64_SVE_ASM
+extern __typeof (__redirect_memset) __memset_a64fx attribute_hidden;
+#endif
 extern __typeof (__redirect_memset) __memset_generic attribute_hidden;
 
 libc_ifunc (__libc_memset,
@@ -40,7 +43,13 @@ libc_ifunc (__libc_memset,
 	     ? __memset_falkor
 	     : (IS_EMAG (midr) && zva_size == 64
 	       ? __memset_emag
-	       : __memset_generic)));
+#if HAVE_AARCH64_SVE_ASM
+	       : (IS_A64FX (midr)
+		  ? __memset_a64fx
+	          : __memset_generic))));
+#else
+	          : __memset_generic)));
+#endif
 
 # undef memset
 strong_alias (__libc_memset, memset);
diff --git a/sysdeps/aarch64/multiarch/memset_a64fx.S b/sysdeps/aarch64/multiarch/memset_a64fx.S
new file mode 100644
index 0000000000..9bd58cab6d
--- /dev/null
+++ b/sysdeps/aarch64/multiarch/memset_a64fx.S
@@ -0,0 +1,268 @@
+/* Optimized memset for Fujitsu A64FX processor.
+   Copyright (C) 2012-2021 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>
+#include <sysdeps/aarch64/memset-reg.h>
+
+#if HAVE_AARCH64_SVE_ASM
+#if IS_IN (libc)
+# define MEMSET __memset_a64fx
+
+/* Assumptions:
+ *
+ * ARMv8.2-a, AArch64, unaligned accesses, sve
+ *
+ */
+
+#define L1_SIZE         (64*1024)       // L1 64KB
+#define L2_SIZE         (8*1024*1024)   // L2 8MB - 1MB
+#define CACHE_LINE_SIZE 256
+#define PF_DIST_L1      (CACHE_LINE_SIZE * 16)  // Prefetch distance L1
+#define ZF_DIST         (CACHE_LINE_SIZE * 21)  // Zerofill distance
+#define rest            x8
+#define vector_length   x9
+#define vl_remainder    x10     // vector_length remainder
+#define cl_remainder    x11     // CACHE_LINE_SIZE remainder
+
+    .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 st1b_unroll first=0, last=7
+    st1b        z0.b, p0, [dst, #\first, mul vl]
+    .if \last-\first
+    st1b_unroll "(\first+1)", \last
+    .endif
+    .endm
+
+    .macro shortcut_for_small_size exit
+    // if rest <= vector_length * 2
+    whilelo     p0.b, xzr, count
+    whilelo     p1.b, vector_length, count
+    b.last      1f
+    st1b        z0.b, p0, [dstin, #0, mul vl]
+    st1b        z0.b, p1, [dstin, #1, mul vl]
+    ret
+1:  // if rest > vector_length * 8
+    cmp         count, 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, count
+    incb        tmp1
+    whilelo     p3.b, tmp1, count
+    b.last      1f
+    st1b        z0.b, p0, [dstin, #0, mul vl]
+    st1b        z0.b, p1, [dstin, #1, mul vl]
+    st1b        z0.b, p2, [dstin, #2, mul vl]
+    st1b        z0.b, p3, [dstin, #3, mul vl]
+    ret
+1:  // if rest <= vector_length * 8
+    lsl         tmp1, vector_length, 2  // vector_length * 4
+    whilelo     p4.b, tmp1, count
+    incb        tmp1
+    whilelo     p5.b, tmp1, count
+    b.last      1f
+    st1b        z0.b, p0, [dstin, #0, mul vl]
+    st1b        z0.b, p1, [dstin, #1, mul vl]
+    st1b        z0.b, p2, [dstin, #2, mul vl]
+    st1b        z0.b, p3, [dstin, #3, mul vl]
+    st1b        z0.b, p4, [dstin, #4, mul vl]
+    st1b        z0.b, p5, [dstin, #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, count
+    incb        tmp1
+    whilelo     p7.b, tmp1, count
+    st1b        z0.b, p0, [dstin, #0, mul vl]
+    st1b        z0.b, p1, [dstin, #1, mul vl]
+    st1b        z0.b, p2, [dstin, #2, mul vl]
+    st1b        z0.b, p3, [dstin, #3, mul vl]
+    st1b        z0.b, p4, [dstin, #4, mul vl]
+    st1b        z0.b, p5, [dstin, #5, mul vl]
+    st1b        z0.b, p6, [dstin, #6, mul vl]
+    st1b        z0.b, p7, [dstin, #7, mul vl]
+    ret
+    .endm
+
+ENTRY (MEMSET)
+
+    PTR_ARG (0)
+    SIZE_ARG (2)
+
+    cbnz        count, 1f
+    ret
+1:  dup         z0.b, valw
+    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, count
+    mov         dst, dstin
+    add         dstend, dstin, count
+    // 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)
+    // if rest >= L1_SIZE && vector_length == 64 then L(L1_prefetch)
+    cmp         rest, L1_SIZE
+    ccmp        vector_length, tmp1, 0, cs
+    b.eq        L(L1_prefetch)
+
+L(unroll32):
+    lsl         tmp1, vector_length, 3  // vector_length * 8
+    lsl         tmp2, vector_length, 5  // vector_length * 32
+    .p2align 3
+1:  cmp         rest, tmp2
+    b.cc        L(unroll8)
+    st1b_unroll
+    add         dst, dst, tmp1
+    st1b_unroll
+    add         dst, dst, tmp1
+    st1b_unroll
+    add         dst, dst, tmp1
+    st1b_unroll
+    add         dst, dst, tmp1
+    sub         rest, rest, tmp2
+    b           1b
+
+L(unroll8):
+    lsl         tmp1, vector_length, 3
+    .p2align 3
+1:  cmp         rest, tmp1
+    b.cc        L(last)
+    st1b_unroll
+    add         dst, dst, tmp1
+    sub         rest, rest, tmp1
+    b           1b
+
+L(last):
+    whilelo     p0.b, xzr, rest
+    whilelo     p1.b, vector_length, rest
+    b.last      1f
+    st1b        z0.b, p0, [dst, #0, mul vl]
+    st1b        z0.b, p1, [dst, #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
+    st1b        z0.b, p0, [dst, #0, mul vl]
+    st1b        z0.b, p1, [dst, #1, mul vl]
+    st1b        z0.b, p2, [dst, #2, mul vl]
+    st1b        z0.b, p3, [dst, #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
+    st1b        z0.b, p0, [dst, #0, mul vl]
+    st1b        z0.b, p1, [dst, #1, mul vl]
+    st1b        z0.b, p2, [dst, #2, mul vl]
+    st1b        z0.b, p3, [dst, #3, mul vl]
+    st1b        z0.b, p4, [dst, #4, mul vl]
+    st1b        z0.b, p5, [dst, #5, mul vl]
+    st1b        z0.b, p6, [dst, #6, mul vl]
+    st1b        z0.b, p7, [dst, #7, mul vl]
+    ret
+
+L(L1_prefetch): // if rest >= L1_SIZE
+    .p2align 3
+1:  st1b_unroll 0, 3
+    prfm        pstl1keep, [dst, PF_DIST_L1]
+    st1b_unroll 4, 7
+    prfm        pstl1keep, [dst, PF_DIST_L1 + CACHE_LINE_SIZE]
+    add         dst, dst, CACHE_LINE_SIZE * 2
+    sub         rest, rest, CACHE_LINE_SIZE * 2
+    cmp         rest, L1_SIZE
+    b.ge        1b
+    cbnz        rest, L(unroll32)
+    ret
+
+L(L2):
+    // align dst address at vector_length byte boundary
+    sub         tmp1, vector_length, 1
+    ands        tmp2, dst, tmp1
+    // if vl_remainder == 0
+    b.eq        1f
+    sub         vl_remainder, vector_length, tmp2
+    // process remainder until the first vector_length boundary
+    whilelt     p2.b, xzr, vl_remainder
+    st1b        z0.b, p2, [dst]
+    add         dst, dst, vl_remainder
+    sub         rest, rest, vl_remainder
+    // align dstin address at CACHE_LINE_SIZE byte boundary
+1:  mov         tmp1, CACHE_LINE_SIZE
+    ands        tmp2, dst, 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
+    mov         tmp1, xzr       // index
+2:  whilelt     p2.b, tmp1, cl_remainder
+    st1b        z0.b, p2, [dst, tmp1]
+    incb        tmp1
+    cmp         tmp1, cl_remainder
+    b.lo        2b
+    add         dst, dst, cl_remainder
+    sub         rest, rest, cl_remainder
+
+L(L2_dc_zva):
+    // zero fill
+    mov         tmp1, dst
+    dc_zva      (ZF_DIST / CACHE_LINE_SIZE) - 1
+    mov         zva_len, ZF_DIST
+    add         tmp1, zva_len, CACHE_LINE_SIZE * 2
+    // unroll
+    .p2align 3
+1:  st1b_unroll 0, 3
+    add         tmp2, dst, zva_len
+    dc          zva, tmp2
+    st1b_unroll 4, 7
+    add         tmp2, tmp2, CACHE_LINE_SIZE
+    dc          zva, tmp2
+    add         dst, dst, CACHE_LINE_SIZE * 2
+    sub         rest, rest, CACHE_LINE_SIZE * 2
+    cmp         rest, tmp1      // ZF_DIST + CACHE_LINE_SIZE * 2
+    b.ge        1b
+    cbnz        rest, L(unroll8)
+    ret
+
+END (MEMSET)
+libc_hidden_builtin_def (MEMSET)
+
+#endif /* IS_IN (libc) */
+#endif /* HAVE_AARCH64_SVE_ASM */
-- 
2.17.1