[PATCH 2/8] xor/arm: Replace vectorized implementation with arm64's intrinsics
Ard Biesheuvel
ardb+git at google.com
Wed Apr 22 10:16:58 PDT 2026
From: Ard Biesheuvel <ardb at kernel.org>
Drop the XOR implementation generated by the vectorizer: this has always
been a bit of a hack, and now that arm64 has an intrinsics version that
works on ARM too, let's use that instead.
So copy the part of the arm64 code that can be shared (so not the EOR3
version). The arm64 code will be updated in a subsequent patch to share
this implementation.
Performance (QEMU mach-virt VM running on Synquacer [Cortex-A53 @ 1 GHz]
Before:
[ 3.519687] xor: measuring software checksum speed
[ 3.521725] neon : 1660 MB/sec
[ 3.524733] 32regs : 1105 MB/sec
[ 3.527751] 8regs : 1098 MB/sec
[ 3.529911] arm4regs : 1540 MB/sec
After:
[ 3.517654] xor: measuring software checksum speed
[ 3.519454] neon : 1896 MB/sec
[ 3.522499] 32regs : 1090 MB/sec
[ 3.525560] 8regs : 1083 MB/sec
[ 3.527700] arm4regs : 1556 MB/sec
Signed-off-by: Ard Biesheuvel <ardb at kernel.org>
---
lib/raid/xor/Makefile | 6 +-
lib/raid/xor/arm/xor-neon.c | 26 ---
lib/raid/xor/arm/xor-neon.h | 7 +
lib/raid/xor/arm/xor_arch.h | 7 +-
lib/raid/xor/xor-8regs.c | 2 -
lib/raid/xor/xor-neon.c | 175 ++++++++++++++++++++
6 files changed, 186 insertions(+), 37 deletions(-)
diff --git a/lib/raid/xor/Makefile b/lib/raid/xor/Makefile
index 4d633dfd5b90..d78400f2427a 100644
--- a/lib/raid/xor/Makefile
+++ b/lib/raid/xor/Makefile
@@ -17,7 +17,7 @@ endif
xor-$(CONFIG_ALPHA) += alpha/xor.o
xor-$(CONFIG_ARM) += arm/xor.o
ifeq ($(CONFIG_ARM),y)
-xor-$(CONFIG_KERNEL_MODE_NEON) += arm/xor-neon.o arm/xor-neon-glue.o
+xor-$(CONFIG_KERNEL_MODE_NEON) += xor-neon.o arm/xor-neon-glue.o
endif
xor-$(CONFIG_ARM64) += arm64/xor-neon.o arm64/xor-neon-glue.o
xor-$(CONFIG_CPU_HAS_LSX) += loongarch/xor_simd.o
@@ -31,8 +31,8 @@ xor-$(CONFIG_X86_32) += x86/xor-avx.o x86/xor-sse.o x86/xor-mmx.o
xor-$(CONFIG_X86_64) += x86/xor-avx.o x86/xor-sse.o
obj-y += tests/
-CFLAGS_arm/xor-neon.o += $(CC_FLAGS_FPU)
-CFLAGS_REMOVE_arm/xor-neon.o += $(CC_FLAGS_NO_FPU)
+CFLAGS_xor-neon.o += $(CC_FLAGS_FPU) -I$(src)/$(SRCARCH)
+CFLAGS_REMOVE_xor-neon.o += $(CC_FLAGS_NO_FPU)
CFLAGS_arm64/xor-neon.o += $(CC_FLAGS_FPU)
CFLAGS_REMOVE_arm64/xor-neon.o += $(CC_FLAGS_NO_FPU)
diff --git a/lib/raid/xor/arm/xor-neon.c b/lib/raid/xor/arm/xor-neon.c
deleted file mode 100644
index 23147e3a7904..000000000000
--- a/lib/raid/xor/arm/xor-neon.c
+++ /dev/null
@@ -1,26 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0-only
-/*
- * Copyright (C) 2013 Linaro Ltd <ard.biesheuvel at linaro.org>
- */
-
-#include "xor_impl.h"
-#include "xor_arch.h"
-
-#ifndef __ARM_NEON__
-#error You should compile this file with '-march=armv7-a -mfloat-abi=softfp -mfpu=neon'
-#endif
-
-/*
- * Pull in the reference implementations while instructing GCC (through
- * -ftree-vectorize) to attempt to exploit implicit parallelism and emit
- * NEON instructions. Clang does this by default at O2 so no pragma is
- * needed.
- */
-#ifdef CONFIG_CC_IS_GCC
-#pragma GCC optimize "tree-vectorize"
-#endif
-
-#define NO_TEMPLATE
-#include "../xor-8regs.c"
-
-__DO_XOR_BLOCKS(neon_inner, xor_8regs_2, xor_8regs_3, xor_8regs_4, xor_8regs_5);
diff --git a/lib/raid/xor/arm/xor-neon.h b/lib/raid/xor/arm/xor-neon.h
new file mode 100644
index 000000000000..406e0356f05b
--- /dev/null
+++ b/lib/raid/xor/arm/xor-neon.h
@@ -0,0 +1,7 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+
+extern struct xor_block_template xor_block_arm4regs;
+extern struct xor_block_template xor_block_neon;
+
+void xor_gen_neon_inner(void *dest, void **srcs, unsigned int src_cnt,
+ unsigned int bytes);
diff --git a/lib/raid/xor/arm/xor_arch.h b/lib/raid/xor/arm/xor_arch.h
index 775ff835df65..f1ddb64fe62a 100644
--- a/lib/raid/xor/arm/xor_arch.h
+++ b/lib/raid/xor/arm/xor_arch.h
@@ -3,12 +3,7 @@
* Copyright (C) 2001 Russell King
*/
#include <asm/neon.h>
-
-extern struct xor_block_template xor_block_arm4regs;
-extern struct xor_block_template xor_block_neon;
-
-void xor_gen_neon_inner(void *dest, void **srcs, unsigned int src_cnt,
- unsigned int bytes);
+#include "xor-neon.h"
static __always_inline void __init arch_xor_init(void)
{
diff --git a/lib/raid/xor/xor-8regs.c b/lib/raid/xor/xor-8regs.c
index 1edaed8acffe..46b3c8bdc27f 100644
--- a/lib/raid/xor/xor-8regs.c
+++ b/lib/raid/xor/xor-8regs.c
@@ -93,11 +93,9 @@ xor_8regs_5(unsigned long bytes, unsigned long * __restrict p1,
} while (--lines > 0);
}
-#ifndef NO_TEMPLATE
DO_XOR_BLOCKS(8regs, xor_8regs_2, xor_8regs_3, xor_8regs_4, xor_8regs_5);
struct xor_block_template xor_block_8regs = {
.name = "8regs",
.xor_gen = xor_gen_8regs,
};
-#endif /* NO_TEMPLATE */
diff --git a/lib/raid/xor/xor-neon.c b/lib/raid/xor/xor-neon.c
new file mode 100644
index 000000000000..a3e2b4af8d36
--- /dev/null
+++ b/lib/raid/xor/xor-neon.c
@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Authors: Jackie Liu <liuyun01 at kylinos.cn>
+ * Copyright (C) 2018,Tianjin KYLIN Information Technology Co., Ltd.
+ */
+
+#include "xor_impl.h"
+#include "xor-neon.h"
+
+#include <asm/neon-intrinsics.h>
+
+static void __xor_neon_2(unsigned long bytes, unsigned long * __restrict p1,
+ const unsigned long * __restrict p2)
+{
+ uint64_t *dp1 = (uint64_t *)p1;
+ uint64_t *dp2 = (uint64_t *)p2;
+
+ register uint64x2_t v0, v1, v2, v3;
+ long lines = bytes / (sizeof(uint64x2_t) * 4);
+
+ do {
+ /* p1 ^= p2 */
+ v0 = veorq_u64(vld1q_u64(dp1 + 0), vld1q_u64(dp2 + 0));
+ v1 = veorq_u64(vld1q_u64(dp1 + 2), vld1q_u64(dp2 + 2));
+ v2 = veorq_u64(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4));
+ v3 = veorq_u64(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6));
+
+ /* store */
+ vst1q_u64(dp1 + 0, v0);
+ vst1q_u64(dp1 + 2, v1);
+ vst1q_u64(dp1 + 4, v2);
+ vst1q_u64(dp1 + 6, v3);
+
+ dp1 += 8;
+ dp2 += 8;
+ } while (--lines > 0);
+}
+
+static void __xor_neon_3(unsigned long bytes, unsigned long * __restrict p1,
+ const unsigned long * __restrict p2,
+ const unsigned long * __restrict p3)
+{
+ uint64_t *dp1 = (uint64_t *)p1;
+ uint64_t *dp2 = (uint64_t *)p2;
+ uint64_t *dp3 = (uint64_t *)p3;
+
+ register uint64x2_t v0, v1, v2, v3;
+ long lines = bytes / (sizeof(uint64x2_t) * 4);
+
+ do {
+ /* p1 ^= p2 */
+ v0 = veorq_u64(vld1q_u64(dp1 + 0), vld1q_u64(dp2 + 0));
+ v1 = veorq_u64(vld1q_u64(dp1 + 2), vld1q_u64(dp2 + 2));
+ v2 = veorq_u64(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4));
+ v3 = veorq_u64(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6));
+
+ /* p1 ^= p3 */
+ v0 = veorq_u64(v0, vld1q_u64(dp3 + 0));
+ v1 = veorq_u64(v1, vld1q_u64(dp3 + 2));
+ v2 = veorq_u64(v2, vld1q_u64(dp3 + 4));
+ v3 = veorq_u64(v3, vld1q_u64(dp3 + 6));
+
+ /* store */
+ vst1q_u64(dp1 + 0, v0);
+ vst1q_u64(dp1 + 2, v1);
+ vst1q_u64(dp1 + 4, v2);
+ vst1q_u64(dp1 + 6, v3);
+
+ dp1 += 8;
+ dp2 += 8;
+ dp3 += 8;
+ } while (--lines > 0);
+}
+
+static void __xor_neon_4(unsigned long bytes, unsigned long * __restrict p1,
+ const unsigned long * __restrict p2,
+ const unsigned long * __restrict p3,
+ const unsigned long * __restrict p4)
+{
+ uint64_t *dp1 = (uint64_t *)p1;
+ uint64_t *dp2 = (uint64_t *)p2;
+ uint64_t *dp3 = (uint64_t *)p3;
+ uint64_t *dp4 = (uint64_t *)p4;
+
+ register uint64x2_t v0, v1, v2, v3;
+ long lines = bytes / (sizeof(uint64x2_t) * 4);
+
+ do {
+ /* p1 ^= p2 */
+ v0 = veorq_u64(vld1q_u64(dp1 + 0), vld1q_u64(dp2 + 0));
+ v1 = veorq_u64(vld1q_u64(dp1 + 2), vld1q_u64(dp2 + 2));
+ v2 = veorq_u64(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4));
+ v3 = veorq_u64(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6));
+
+ /* p1 ^= p3 */
+ v0 = veorq_u64(v0, vld1q_u64(dp3 + 0));
+ v1 = veorq_u64(v1, vld1q_u64(dp3 + 2));
+ v2 = veorq_u64(v2, vld1q_u64(dp3 + 4));
+ v3 = veorq_u64(v3, vld1q_u64(dp3 + 6));
+
+ /* p1 ^= p4 */
+ v0 = veorq_u64(v0, vld1q_u64(dp4 + 0));
+ v1 = veorq_u64(v1, vld1q_u64(dp4 + 2));
+ v2 = veorq_u64(v2, vld1q_u64(dp4 + 4));
+ v3 = veorq_u64(v3, vld1q_u64(dp4 + 6));
+
+ /* store */
+ vst1q_u64(dp1 + 0, v0);
+ vst1q_u64(dp1 + 2, v1);
+ vst1q_u64(dp1 + 4, v2);
+ vst1q_u64(dp1 + 6, v3);
+
+ dp1 += 8;
+ dp2 += 8;
+ dp3 += 8;
+ dp4 += 8;
+ } while (--lines > 0);
+}
+
+static void __xor_neon_5(unsigned long bytes, unsigned long * __restrict p1,
+ const unsigned long * __restrict p2,
+ const unsigned long * __restrict p3,
+ const unsigned long * __restrict p4,
+ const unsigned long * __restrict p5)
+{
+ uint64_t *dp1 = (uint64_t *)p1;
+ uint64_t *dp2 = (uint64_t *)p2;
+ uint64_t *dp3 = (uint64_t *)p3;
+ uint64_t *dp4 = (uint64_t *)p4;
+ uint64_t *dp5 = (uint64_t *)p5;
+
+ register uint64x2_t v0, v1, v2, v3;
+ long lines = bytes / (sizeof(uint64x2_t) * 4);
+
+ do {
+ /* p1 ^= p2 */
+ v0 = veorq_u64(vld1q_u64(dp1 + 0), vld1q_u64(dp2 + 0));
+ v1 = veorq_u64(vld1q_u64(dp1 + 2), vld1q_u64(dp2 + 2));
+ v2 = veorq_u64(vld1q_u64(dp1 + 4), vld1q_u64(dp2 + 4));
+ v3 = veorq_u64(vld1q_u64(dp1 + 6), vld1q_u64(dp2 + 6));
+
+ /* p1 ^= p3 */
+ v0 = veorq_u64(v0, vld1q_u64(dp3 + 0));
+ v1 = veorq_u64(v1, vld1q_u64(dp3 + 2));
+ v2 = veorq_u64(v2, vld1q_u64(dp3 + 4));
+ v3 = veorq_u64(v3, vld1q_u64(dp3 + 6));
+
+ /* p1 ^= p4 */
+ v0 = veorq_u64(v0, vld1q_u64(dp4 + 0));
+ v1 = veorq_u64(v1, vld1q_u64(dp4 + 2));
+ v2 = veorq_u64(v2, vld1q_u64(dp4 + 4));
+ v3 = veorq_u64(v3, vld1q_u64(dp4 + 6));
+
+ /* p1 ^= p5 */
+ v0 = veorq_u64(v0, vld1q_u64(dp5 + 0));
+ v1 = veorq_u64(v1, vld1q_u64(dp5 + 2));
+ v2 = veorq_u64(v2, vld1q_u64(dp5 + 4));
+ v3 = veorq_u64(v3, vld1q_u64(dp5 + 6));
+
+ /* store */
+ vst1q_u64(dp1 + 0, v0);
+ vst1q_u64(dp1 + 2, v1);
+ vst1q_u64(dp1 + 4, v2);
+ vst1q_u64(dp1 + 6, v3);
+
+ dp1 += 8;
+ dp2 += 8;
+ dp3 += 8;
+ dp4 += 8;
+ dp5 += 8;
+ } while (--lines > 0);
+}
+
+__DO_XOR_BLOCKS(neon_inner, __xor_neon_2, __xor_neon_3, __xor_neon_4,
+ __xor_neon_5);
--
2.54.0.rc1.555.g9c883467ad-goog
More information about the linux-arm-kernel
mailing list