Implementation of SSE optimized Fletcher-4

Builds off of 1eeb4562 (Implementation of AVX2 optimized Fletcher-4)
This commit adds another implementation of the Fletcher-4 algorithm.
It is automatically selected at module load if it benchmarks higher
than all other available implementations.

The module benchmark was also amended to analyze the performance of
the byteswap-ed version of Fletcher-4, as well as the non-byteswaped
version. The average performance of the two is used to select the
the fastest implementation available on the host system.

Adds a pair of fields to an existing zcommon module parameter:
-  zfs_fletcher_4_impl (str)
    "sse2"    - new SSE2 implementation if available
    "ssse3"   - new SSSE3 implementation if available

Signed-off-by: Tyler J. Stachecki <[email protected]>
Signed-off-by: Gvozden Neskovic <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #4789

6 files changed, 243 insertions, 5 deletions

diff --git a/include/zfs_fletcher.h b/include/zfs_fletcher.h
index 4ccb4a2f4..afc3936c0 100644
--- a/include/zfs_fletcher.h
+++ b/include/zfs_fletcher.h
@@ -61,6 +61,14 @@ typedef struct fletcher_4_func {
 	const char *name;
 } fletcher_4_ops_t;
 
+#if defined(HAVE_SSE2)
+extern const fletcher_4_ops_t fletcher_4_sse2_ops;
+#endif
+
+#if defined(HAVE_SSE2) && defined(HAVE_SSSE3)
+extern const fletcher_4_ops_t fletcher_4_ssse3_ops;
+#endif
+
 #if defined(HAVE_AVX) && defined(HAVE_AVX2)
 extern const fletcher_4_ops_t fletcher_4_avx2_ops;
 #endif
diff --git a/lib/libzpool/Makefile.am b/lib/libzpool/Makefile.am
index 9a06eae5b..b40e9787c 100644
--- a/lib/libzpool/Makefile.am
+++ b/lib/libzpool/Makefile.am
@@ -23,6 +23,7 @@ KERNEL_C = \
 	zfs_deleg.c \
 	zfs_fletcher.c \
 	zfs_fletcher_intel.c \
+	zfs_fletcher_sse.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 7b0199de2..79855e72d 100644
--- a/man/man5/zfs-module-parameters.5
+++ b/man/man5/zfs-module-parameters.5
@@ -838,11 +838,15 @@ Default value: \fB67,108,864\fR.
 .RS 12n
 Select a fletcher 4 implementation.
 .sp
-Supported selectors are: \fBfastest\fR, \fBscalar\fR, and \fBavx2\fR when
-AVX2 is supported by the processor.  If multiple implementations of fletcher 4
-are available the \fBfastest\fR will be chosen using a micro benchmark.
-Selecting \fBscalar\fR results in the original CPU based calculation being
-used, \fBavx2\fR uses the AVX2 vector instructions to compute a fletcher 4.
+Supported selectors are: \fBfastest\fR, \fBscalar\fR, \fBsse2\fR, \fBssse3\fR,
+and \fBavx2\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, the \fBfastest\fR will be chosen using a micro
+benchmark. Selecting \fBscalar\fR results in the original CPU based calculation
+being used. Selecting any option other than \fBfastest\fR and \fBscalar\fR
+results in vector instructions from the respective CPU instruction set being
+used.
 .sp
 Default value: \fBfastest\fR.
 .RE
diff --git a/module/zcommon/Makefile.in b/module/zcommon/Makefile.in
index 0882fac2d..958835edf 100644
--- a/module/zcommon/Makefile.in
+++ b/module/zcommon/Makefile.in
@@ -17,3 +17,4 @@ $(MODULE)-objs += zfs_uio.o
 $(MODULE)-objs += zpool_prop.o
 
 $(MODULE)-$(CONFIG_X86) += zfs_fletcher_intel.o
+$(MODULE)-$(CONFIG_X86) += zfs_fletcher_sse.o
diff --git a/module/zcommon/zfs_fletcher.c b/module/zcommon/zfs_fletcher.c
index e76c5b8a5..a3888a32b 100644
--- a/module/zcommon/zfs_fletcher.c
+++ b/module/zcommon/zfs_fletcher.c
@@ -149,6 +149,12 @@ static const fletcher_4_ops_t fletcher_4_scalar_ops = {
 
 static const fletcher_4_ops_t *fletcher_4_algos[] = {
 	&fletcher_4_scalar_ops,
+#if defined(HAVE_SSE2)
+	&fletcher_4_sse2_ops,
+#endif
+#if defined(HAVE_SSE2) && defined(HAVE_SSSE3)
+	&fletcher_4_ssse3_ops,
+#endif
 #if defined(HAVE_AVX) && defined(HAVE_AVX2)
 	&fletcher_4_avx2_ops,
 #endif
@@ -157,6 +163,12 @@ static const fletcher_4_ops_t *fletcher_4_algos[] = {
 static enum fletcher_selector {
 	FLETCHER_FASTEST = 0,
 	FLETCHER_SCALAR,
+#if defined(HAVE_SSE2)
+	FLETCHER_SSE2,
+#endif
+#if defined(HAVE_SSE2) && defined(HAVE_SSSE3)
+	FLETCHER_SSSE3,
+#endif
 #if defined(HAVE_AVX) && defined(HAVE_AVX2)
 	FLETCHER_AVX2,
 #endif
@@ -169,6 +181,12 @@ static struct fletcher_4_impl_selector {
 } fletcher_4_impl_selectors[] = {
 	[ FLETCHER_FASTEST ]	= { "fastest", NULL },
 	[ FLETCHER_SCALAR ]	= { "scalar", &fletcher_4_scalar_ops },
+#if defined(HAVE_SSE2)
+	[ FLETCHER_SSE2 ]	= { "sse2", &fletcher_4_sse2_ops },
+#endif
+#if defined(HAVE_SSE2) && defined(HAVE_SSSE3)
+	[ FLETCHER_SSSE3 ]	= { "ssse3", &fletcher_4_ssse3_ops },
+#endif
 #if defined(HAVE_AVX) && defined(HAVE_AVX2)
 	[ FLETCHER_AVX2 ]	= { "avx2", &fletcher_4_avx2_ops },
 #endif
@@ -407,6 +425,7 @@ fletcher_4_init(void)
 		ops->init(&zc);
 		do {
 			ops->compute(databuf, data_size, &zc);
+			ops->compute_byteswap(databuf, data_size, &zc);
 			run_count++;
 		} while (gethrtime() < start + bench_ns);
 		if (ops->fini != NULL)
diff --git a/module/zcommon/zfs_fletcher_sse.c b/module/zcommon/zfs_fletcher_sse.c
new file mode 100644
index 000000000..734ae0853
--- /dev/null
+++ b/module/zcommon/zfs_fletcher_sse.c
@@ -0,0 +1,205 @@
+/*
+ * Implement fast Fletcher4 with SSE2,SSSE3 instructions. (x86)
+ *
+ * Use the 128-bit SSE2/SSSE3 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 Tyler J. Stachecki.
+ *
+ * Authors:
+ *	Tyler J. Stachecki <[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(HAVE_SSE2)
+
+#include <linux/simd_x86.h>
+#include <sys/spa_checksum.h>
+#include <zfs_fletcher.h>
+
+struct zfs_fletcher_sse_array {
+	uint64_t v[2] __attribute__((aligned(16)));
+};
+
+static void
+fletcher_4_sse2_init(zio_cksum_t *zcp)
+{
+	kfpu_begin();
+
+	/* clear sse registers */
+	asm volatile("pxor %xmm0, %xmm0");
+	asm volatile("pxor %xmm1, %xmm1");
+	asm volatile("pxor %xmm2, %xmm2");
+	asm volatile("pxor %xmm3, %xmm3");
+}
+
+static void
+fletcher_4_sse2_fini(zio_cksum_t *zcp)
+{
+	struct zfs_fletcher_sse_array a, b, c, d;
+	uint64_t A, B, C, D;
+
+	asm volatile("movdqa %%xmm0, %0":"=m" (a.v));
+	asm volatile("movdqa %%xmm1, %0":"=m" (b.v));
+	asm volatile("psllq $0x2, %xmm2");
+	asm volatile("movdqa %%xmm2, %0":"=m" (c.v));
+	asm volatile("psllq $0x3, %xmm3");
+	asm volatile("movdqa %%xmm3, %0":"=m" (d.v));
+
+	kfpu_end();
+
+	/*
+	 * The mixing matrix for checksum calculation is:
+	 * a = a0 + a1
+	 * b = 2b0 + 2b1 - a1
+	 * c = 4c0 - b0 + 4c1 -3b1
+	 * d = 8d0 - 4c0 + 8d1 - 8c1 + b1;
+	 *
+	 * c and d are multiplied by 4 and 8, respectively,
+	 * before spilling the vectors out to memory.
+	 */
+	A = a.v[0] + a.v[1];
+	B = 2*b.v[0] + 2*b.v[1] - a.v[1];
+	C = c.v[0] - b.v[0] + c.v[1] - 3*b.v[1];
+	D = d.v[0] - c.v[0] + d.v[1] - 2*c.v[1] + b.v[1];
+
+	ZIO_SET_CHECKSUM(zcp, A, B, C, D);
+}
+
+static void
+fletcher_4_sse2(const void *buf, uint64_t size, zio_cksum_t *unused)
+{
+	const uint64_t *ip = buf;
+	const uint64_t *ipend = (uint64_t *)((uint8_t *)ip + size);
+
+	asm volatile("pxor %xmm4, %xmm4");
+
+	for (; ip < ipend; ip += 2) {
+		asm volatile("movdqu %0, %%xmm5" :: "m"(*ip));
+		asm volatile("movdqa %xmm5, %xmm6");
+		asm volatile("punpckldq %xmm4, %xmm5");
+		asm volatile("punpckhdq %xmm4, %xmm6");
+		asm volatile("paddq %xmm5, %xmm0");
+		asm volatile("paddq %xmm0, %xmm1");
+		asm volatile("paddq %xmm1, %xmm2");
+		asm volatile("paddq %xmm2, %xmm3");
+		asm volatile("paddq %xmm6, %xmm0");
+		asm volatile("paddq %xmm0, %xmm1");
+		asm volatile("paddq %xmm1, %xmm2");
+		asm volatile("paddq %xmm2, %xmm3");
+	}
+}
+
+static void
+fletcher_4_sse2_byteswap(const void *buf, uint64_t size, zio_cksum_t *unused)
+{
+	const uint32_t *ip = buf;
+	const uint32_t *ipend = (uint32_t *)((uint8_t *)ip + size);
+
+	for (; ip < ipend; ip += 2) {
+		uint32_t scratch;
+
+		asm volatile("bswapl %0" : "=r"(scratch) : "0"(*ip));
+		asm volatile("movd %0, %%xmm5" :: "r"(scratch));
+		asm volatile("bswapl %0" : "=r"(scratch) : "0"(*(ip + 1)));
+		asm volatile("movd %0, %%xmm6" :: "r"(scratch));
+		asm volatile("punpcklqdq %xmm6, %xmm5");
+		asm volatile("paddq %xmm5, %xmm0");
+		asm volatile("paddq %xmm0, %xmm1");
+		asm volatile("paddq %xmm1, %xmm2");
+		asm volatile("paddq %xmm2, %xmm3");
+	}
+}
+
+static boolean_t fletcher_4_sse2_valid(void)
+{
+	return (zfs_sse2_available());
+}
+
+const fletcher_4_ops_t fletcher_4_sse2_ops = {
+	.init = fletcher_4_sse2_init,
+	.fini = fletcher_4_sse2_fini,
+	.compute = fletcher_4_sse2,
+	.compute_byteswap = fletcher_4_sse2_byteswap,
+	.valid = fletcher_4_sse2_valid,
+	.name = "sse2"
+};
+
+#endif /* defined(HAVE_SSE2) */
+
+#if defined(HAVE_SSE2) && defined(HAVE_SSSE3)
+static void
+fletcher_4_ssse3_byteswap(const void *buf, uint64_t size, zio_cksum_t *unused)
+{
+	static const struct zfs_fletcher_sse_array mask = {
+		.v = { 0x0405060700010203, 0x0C0D0E0F08090A0B }
+	};
+
+	const uint64_t *ip = buf;
+	const uint64_t *ipend = (uint64_t *)((uint8_t *)ip + size);
+
+	asm volatile("movdqa %0, %%xmm7"::"m" (mask));
+	asm volatile("pxor %xmm4, %xmm4");
+
+	for (; ip < ipend; ip += 2) {
+		asm volatile("movdqu %0, %%xmm5"::"m" (*ip));
+		asm volatile("pshufb %xmm7, %xmm5");
+		asm volatile("movdqa %xmm5, %xmm6");
+		asm volatile("punpckldq %xmm4, %xmm5");
+		asm volatile("punpckhdq %xmm4, %xmm6");
+		asm volatile("paddq %xmm5, %xmm0");
+		asm volatile("paddq %xmm0, %xmm1");
+		asm volatile("paddq %xmm1, %xmm2");
+		asm volatile("paddq %xmm2, %xmm3");
+		asm volatile("paddq %xmm6, %xmm0");
+		asm volatile("paddq %xmm0, %xmm1");
+		asm volatile("paddq %xmm1, %xmm2");
+		asm volatile("paddq %xmm2, %xmm3");
+	}
+}
+
+static boolean_t fletcher_4_ssse3_valid(void)
+{
+	return (zfs_sse2_available() && zfs_ssse3_available());
+}
+
+const fletcher_4_ops_t fletcher_4_ssse3_ops = {
+	.init = fletcher_4_sse2_init,
+	.fini = fletcher_4_sse2_fini,
+	.compute = fletcher_4_sse2,
+	.compute_byteswap = fletcher_4_ssse3_byteswap,
+	.valid = fletcher_4_ssse3_valid,
+	.name = "ssse3"
+};
+
+#endif /* defined(HAVE_SSE2) && defined(HAVE_SSSE3) */