Fletcher4 algorithm implemented in pure NEON for Aarch64 / ARMv8 64 bits

This is not useful on micro-architecture with a weak NEON
implementation (only 64 bits); the native version is slower &
the byteswap barely faster than scalar.  On A53 or A57, it's
a small improvement on scalar but OK for byteswap.

Results from an A53 system:
0 0 0x01 -1 0 1499068294333000 1499101101878000
implementation   native         byteswap       
scalar           1008227510     755880264      
aarch64_neon     1198098720     1044818671     
fastest          aarch64_neon   aarch64_neon 

Results from a A57 system:
0 0 0x01 -1 0 4407214734807033 4407233933777404
implementation   native         byteswap       
scalar           2302071241     1124873346     
aarch64_neon     2542214946     2245570352     
fastest          aarch64_neon   aarch64_neon 

Reviewed-by: Gvozden Neskovic <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Romain Dolbeau <[email protected]>
Closes #5248 

6 files changed, 232 insertions, 1 deletions

diff --git a/include/zfs_fletcher.h b/include/zfs_fletcher.h
index 85c2b5a7e..1f113ae2f 100644
--- a/include/zfs_fletcher.h
+++ b/include/zfs_fletcher.h
@@ -77,6 +77,10 @@ typedef struct zfs_fletcher_avx512 {
 	uint64_t v[8] __attribute__((aligned(64)));
 } zfs_fletcher_avx512_t;
 
+typedef struct zfs_fletcher_aarch64_neon {
+	uint64_t v[2] __attribute__((aligned(16)));
+} zfs_fletcher_aarch64_neon_t;
+
 
 typedef union fletcher_4_ctx {
 	zio_cksum_t scalar;
@@ -90,6 +94,9 @@ typedef union fletcher_4_ctx {
 #if defined(__x86_64) && defined(HAVE_AVX512F)
 	zfs_fletcher_avx512_t avx512[4];
 #endif
+#if defined(__aarch64__)
+	zfs_fletcher_aarch64_neon_t aarch64_neon[4];
+#endif
 } fletcher_4_ctx_t;
 
 /*
@@ -128,6 +135,10 @@ extern const fletcher_4_ops_t fletcher_4_avx2_ops;
 extern const fletcher_4_ops_t fletcher_4_avx512f_ops;
 #endif
 
+#if defined(__aarch64__)
+extern const fletcher_4_ops_t fletcher_4_aarch64_neon_ops;
+#endif
+
 #ifdef	__cplusplus
 }
 #endif
diff --git a/lib/libzpool/Makefile.am b/lib/libzpool/Makefile.am
index c2f5a50b1..13353a58a 100644
--- a/lib/libzpool/Makefile.am
+++ b/lib/libzpool/Makefile.am
@@ -25,6 +25,7 @@ KERNEL_C = \
 	zfs_fletcher_intel.c \
 	zfs_fletcher_sse.c \
 	zfs_fletcher_avx512.c \
+	zfs_fletcher_aarch64_neon.c \
 	zfs_namecheck.c \
 	zfs_prop.c \
 	zfs_uio.c \
diff --git a/man/man5/zfs-module-parameters.5 b/man/man5/zfs-module-parameters.5
index 932342cfd..95527906b 100644
--- a/man/man5/zfs-module-parameters.5
+++ b/man/man5/zfs-module-parameters.5
@@ -900,7 +900,7 @@ Default value: \fB67,108,864\fR.
 Select a fletcher 4 implementation.
 .sp
 Supported selectors are: \fBfastest\fR, \fBscalar\fR, \fBsse2\fR, \fBssse3\fR,
-\fBavx2\fR, and \fBavx512f\fR.
+\fBavx2\fR, \fBavx512f\fR, and \fBaarch64_neon\fR.
 All of the selectors except \fBfastest\fR and \fBscalar\fR require instruction
 set extensions to be available and will only appear if ZFS detects that they are
 present at runtime. If multiple implementations of fletcher 4 are available,
diff --git a/module/zcommon/Makefile.in b/module/zcommon/Makefile.in
index 7dffd5228..bcef47709 100644
--- a/module/zcommon/Makefile.in
+++ b/module/zcommon/Makefile.in
@@ -19,3 +19,4 @@ $(MODULE)-objs += zpool_prop.o
 $(MODULE)-$(CONFIG_X86) += zfs_fletcher_intel.o
 $(MODULE)-$(CONFIG_X86) += zfs_fletcher_sse.o
 $(MODULE)-$(CONFIG_X86) += zfs_fletcher_avx512.o
+$(MODULE)-$(CONFIG_ARM64) += zfs_fletcher_aarch64_neon.o
diff --git a/module/zcommon/zfs_fletcher.c b/module/zcommon/zfs_fletcher.c
index 355384f50..508ac4a1d 100644
--- a/module/zcommon/zfs_fletcher.c
+++ b/module/zcommon/zfs_fletcher.c
@@ -176,6 +176,9 @@ static const fletcher_4_ops_t *fletcher_4_impls[] = {
 #if defined(__x86_64) && defined(HAVE_AVX512F)
 	&fletcher_4_avx512f_ops,
 #endif
+#if defined(__aarch64__)
+	&fletcher_4_aarch64_neon_ops,
+#endif
 };
 
 /* Hold all supported implementations */
diff --git a/module/zcommon/zfs_fletcher_aarch64_neon.c b/module/zcommon/zfs_fletcher_aarch64_neon.c
new file mode 100644
index 000000000..f8991b171
--- /dev/null
+++ b/module/zcommon/zfs_fletcher_aarch64_neon.c
@@ -0,0 +1,215 @@
+/*
+ * Implement fast Fletcher4 with NEON instructions. (aarch64)
+ *
+ * Use the 128-bit NEON SIMD instructions and registers to compute
+ * Fletcher4 in four incremental 64-bit parallel accumulator streams,
+ * and then combine the streams to form the final four checksum words.
+ * This implementation is a derivative of the AVX SIMD implementation by
+ * James Guilford and Jinshan Xiong from Intel (see zfs_fletcher_intel.c).
+ *
+ * Copyright (C) 2016 Romain Dolbeau.
+ *
+ * Authors:
+ *	Romain Dolbeau <[email protected]>
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *	- Redistributions of source code must retain the above
+ *	  copyright notice, this list of conditions and the following
+ *	  disclaimer.
+ *
+ *	- Redistributions in binary form must reproduce the above
+ *	  copyright notice, this list of conditions and the following
+ *	  disclaimer in the documentation and/or other materials
+ *	  provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#if defined(__aarch64__)
+
+#include <linux/simd_aarch64.h>
+#include <sys/spa_checksum.h>
+#include <zfs_fletcher.h>
+#include <strings.h>
+
+static void
+fletcher_4_aarch64_neon_init(fletcher_4_ctx_t *ctx)
+{
+	bzero(ctx->aarch64_neon, 4 * sizeof (zfs_fletcher_aarch64_neon_t));
+}
+
+static void
+fletcher_4_aarch64_neon_fini(fletcher_4_ctx_t *ctx, zio_cksum_t *zcp)
+{
+	uint64_t A, B, C, D;
+	A = ctx->aarch64_neon[0].v[0] + ctx->aarch64_neon[0].v[1];
+	B = 2 * ctx->aarch64_neon[1].v[0] + 2 * ctx->aarch64_neon[1].v[1] -
+	    ctx->aarch64_neon[0].v[1];
+	C = 4 * ctx->aarch64_neon[2].v[0] - ctx->aarch64_neon[1].v[0] +
+	    4 * ctx->aarch64_neon[2].v[1] - 3 * ctx->aarch64_neon[1].v[1];
+	D = 8 * ctx->aarch64_neon[3].v[0] - 4 * ctx->aarch64_neon[2].v[0] +
+	    8 * ctx->aarch64_neon[3].v[1] - 8 * ctx->aarch64_neon[2].v[1] +
+	    ctx->aarch64_neon[1].v[1];
+	ZIO_SET_CHECKSUM(zcp, A, B, C, D);
+}
+
+#define	NEON_INIT_LOOP()			\
+	asm("eor %[ZERO].16b,%[ZERO].16b,%[ZERO].16b\n"	\
+	"ld1 { %[ACC0].4s }, %[CTX0]\n"		\
+	"ld1 { %[ACC1].4s }, %[CTX1]\n"		\
+	"ld1 { %[ACC2].4s }, %[CTX2]\n"		\
+	"ld1 { %[ACC3].4s }, %[CTX3]\n"		\
+	: [ZERO] "=w" (ZERO),			\
+	[ACC0] "=w" (ACC0), [ACC1] "=w" (ACC1),	\
+	[ACC2] "=w" (ACC2), [ACC3] "=w" (ACC3)	\
+	: [CTX0] "Q" (ctx->aarch64_neon[0]),	\
+	[CTX1] "Q" (ctx->aarch64_neon[1]),	\
+	[CTX2] "Q" (ctx->aarch64_neon[2]),	\
+	[CTX3] "Q" (ctx->aarch64_neon[3]))
+
+#define	NEON_DO_REVERSE "rev32 %[SRC].16b, %[SRC].16b\n"
+
+#define	NEON_DONT_REVERSE ""
+
+#define	NEON_MAIN_LOOP(REVERSE)				\
+	asm("ld1 { %[SRC].4s }, %[IP]\n"		\
+	REVERSE						\
+	"zip1 %[TMP1].4s, %[SRC].4s, %[ZERO].4s\n"	\
+	"zip2 %[TMP2].4s, %[SRC].4s, %[ZERO].4s\n"	\
+	"add %[ACC0].2d, %[ACC0].2d, %[TMP1].2d\n"	\
+	"add %[ACC1].2d, %[ACC1].2d, %[ACC0].2d\n"	\
+	"add %[ACC2].2d, %[ACC2].2d, %[ACC1].2d\n"	\
+	"add %[ACC3].2d, %[ACC3].2d, %[ACC2].2d\n"	\
+	"add %[ACC0].2d, %[ACC0].2d, %[TMP2].2d\n"	\
+	"add %[ACC1].2d, %[ACC1].2d, %[ACC0].2d\n"	\
+	"add %[ACC2].2d, %[ACC2].2d, %[ACC1].2d\n"	\
+	"add %[ACC3].2d, %[ACC3].2d, %[ACC2].2d\n"	\
+	: [SRC] "=&w" (SRC),				\
+	[TMP1] "=&w" (TMP1), [TMP2] "=&w" (TMP2),	\
+	[ACC0] "+w" (ACC0), [ACC1] "+w" (ACC1),		\
+	[ACC2] "+w" (ACC2), [ACC3] "+w" (ACC3)		\
+	: [ZERO] "w" (ZERO), [IP] "Q" (*ip))
+
+#define	NEON_FINI_LOOP()			\
+	asm("st1 { %[ACC0].4s },%[DST0]\n"	\
+	"st1 { %[ACC1].4s },%[DST1]\n"		\
+	"st1 { %[ACC2].4s },%[DST2]\n"		\
+	"st1 { %[ACC3].4s },%[DST3]\n"		\
+	: [DST0] "=Q" (ctx->aarch64_neon[0]),	\
+	[DST1] "=Q" (ctx->aarch64_neon[1]),	\
+	[DST2] "=Q" (ctx->aarch64_neon[2]),	\
+	[DST3] "=Q" (ctx->aarch64_neon[3])	\
+	: [ACC0] "w" (ACC0), [ACC1] "w" (ACC1),	\
+	[ACC2] "w" (ACC2), [ACC3] "w" (ACC3))
+
+static void
+fletcher_4_aarch64_neon_native(fletcher_4_ctx_t *ctx,
+    const void *buf, uint64_t size)
+{
+	const uint64_t *ip = buf;
+	const uint64_t *ipend = (uint64_t *)((uint8_t *)ip + size);
+#if defined(_KERNEL)
+register unsigned char ZERO asm("v0") __attribute__((vector_size(16)));
+register unsigned char ACC0 asm("v1") __attribute__((vector_size(16)));
+register unsigned char ACC1 asm("v2") __attribute__((vector_size(16)));
+register unsigned char ACC2 asm("v3") __attribute__((vector_size(16)));
+register unsigned char ACC3 asm("v4") __attribute__((vector_size(16)));
+register unsigned char TMP1 asm("v5") __attribute__((vector_size(16)));
+register unsigned char TMP2 asm("v6") __attribute__((vector_size(16)));
+register unsigned char SRC asm("v7") __attribute__((vector_size(16)));
+#else
+unsigned char ZERO __attribute__((vector_size(16)));
+unsigned char ACC0 __attribute__((vector_size(16)));
+unsigned char ACC1 __attribute__((vector_size(16)));
+unsigned char ACC2 __attribute__((vector_size(16)));
+unsigned char ACC3 __attribute__((vector_size(16)));
+unsigned char TMP1 __attribute__((vector_size(16)));
+unsigned char TMP2 __attribute__((vector_size(16)));
+unsigned char SRC __attribute__((vector_size(16)));
+#endif
+
+	kfpu_begin();
+
+	NEON_INIT_LOOP();
+
+	for (; ip < ipend; ip += 2) {
+		NEON_MAIN_LOOP(NEON_DONT_REVERSE);
+	}
+
+	NEON_FINI_LOOP();
+
+	kfpu_end();
+}
+
+static void
+fletcher_4_aarch64_neon_byteswap(fletcher_4_ctx_t *ctx,
+    const void *buf, uint64_t size)
+{
+	const uint64_t *ip = buf;
+	const uint64_t *ipend = (uint64_t *)((uint8_t *)ip + size);
+#if defined(_KERNEL)
+register unsigned char ZERO asm("v0") __attribute__((vector_size(16)));
+register unsigned char ACC0 asm("v1") __attribute__((vector_size(16)));
+register unsigned char ACC1 asm("v2") __attribute__((vector_size(16)));
+register unsigned char ACC2 asm("v3") __attribute__((vector_size(16)));
+register unsigned char ACC3 asm("v4") __attribute__((vector_size(16)));
+register unsigned char TMP1 asm("v5") __attribute__((vector_size(16)));
+register unsigned char TMP2 asm("v6") __attribute__((vector_size(16)));
+register unsigned char SRC asm("v7") __attribute__((vector_size(16)));
+#else
+unsigned char ZERO __attribute__((vector_size(16)));
+unsigned char ACC0 __attribute__((vector_size(16)));
+unsigned char ACC1 __attribute__((vector_size(16)));
+unsigned char ACC2 __attribute__((vector_size(16)));
+unsigned char ACC3 __attribute__((vector_size(16)));
+unsigned char TMP1 __attribute__((vector_size(16)));
+unsigned char TMP2 __attribute__((vector_size(16)));
+unsigned char SRC __attribute__((vector_size(16)));
+#endif
+
+	kfpu_begin();
+
+	NEON_INIT_LOOP();
+
+	for (; ip < ipend; ip += 2) {
+		NEON_MAIN_LOOP(NEON_DO_REVERSE);
+	}
+
+	NEON_FINI_LOOP();
+
+	kfpu_end();
+}
+
+static boolean_t fletcher_4_aarch64_neon_valid(void)
+{
+	return (B_TRUE);
+}
+
+const fletcher_4_ops_t fletcher_4_aarch64_neon_ops = {
+	.init_native = fletcher_4_aarch64_neon_init,
+	.compute_native = fletcher_4_aarch64_neon_native,
+	.fini_native = fletcher_4_aarch64_neon_fini,
+	.init_byteswap = fletcher_4_aarch64_neon_init,
+	.compute_byteswap = fletcher_4_aarch64_neon_byteswap,
+	.fini_byteswap = fletcher_4_aarch64_neon_fini,
+	.valid = fletcher_4_aarch64_neon_valid,
+	.name = "aarch64_neon"
+};
+
+#endif /* defined(__aarch64__) */