On 01/18/2018 10:47 AM, Jens Axboe wrote:
Adding the memcpy for avx/sse to the test case might be interesting
though, just to be able to compare performances with builtin
memcpy/memmove on a given system.
I've attached the patch. I think it should work on all Intel systems
newer than 2012. It causes a segfault on my Atom E3826 (Baytrail)
because it doesn't support AVX.
I've tested it using Clang 3.8, GCC 4.8 and 7.x on openSUSE, and Clang
4.0 on FreeBSD 11.1.
--
Rebecca
>From f9fa5d6fc83d5d924ac2cf9532d51a3b1d81c9ff Mon Sep 17 00:00:00 2001
From: Rebecca Cran <rebecca@xxxxxxxxxxxx>
Date: Thu, 18 Jan 2018 12:07:11 -0700
Subject: [PATCH] Add SSE and AVX tests to lib/memcpy.c and add -msse and -mavx
to compiler flags
---
Makefile | 2 +-
lib/memcpy.c | 65 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++--
2 files changed, 64 insertions(+), 3 deletions(-)
diff --git a/Makefile b/Makefile
index 3ce60646..4e32f269 100644
--- a/Makefile
+++ b/Makefile
@@ -23,7 +23,7 @@ endif
DEBUGFLAGS = -DFIO_INC_DEBUG
CPPFLAGS= -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 -DFIO_INTERNAL $(DEBUGFLAGS)
OPTFLAGS= -g -ffast-math
-CFLAGS = -std=gnu99 -Wwrite-strings -Wall -Wdeclaration-after-statement $(OPTFLAGS) $(EXTFLAGS) $(BUILD_CFLAGS) -I. -I$(SRCDIR)
+CFLAGS = -msse -mavx -std=gnu99 -Wwrite-strings -Wall -Wdeclaration-after-statement $(OPTFLAGS) $(EXTFLAGS) $(BUILD_CFLAGS) -I. -I$(SRCDIR)
LIBS += -lm $(EXTLIBS)
PROGS = fio
SCRIPTS = $(addprefix $(SRCDIR)/,tools/fio_generate_plots tools/plot/fio2gnuplot tools/genfio tools/fiologparser.py tools/hist/fiologparser_hist.py tools/fio_jsonplus_clat2csv)
diff --git a/lib/memcpy.c b/lib/memcpy.c
index 00e65aa7..1fb22dad 100644
--- a/lib/memcpy.c
+++ b/lib/memcpy.c
@@ -1,6 +1,8 @@
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
+#include <xmmintrin.h>
+#include <immintrin.h>
#include "memcpy.h"
#include "rand.h"
@@ -80,6 +82,8 @@ enum {
T_MEMMOVE = 1U << 1,
T_SIMPLE = 1U << 2,
T_HYBRID = 1U << 3,
+ T_SSE = 1U << 4,
+ T_AVX = 1U << 5,
};
#define do_test(test, fn) do { \
@@ -122,6 +126,43 @@ static void simple_memcpy(void *dst, void const *src, size_t len)
*d++ = *s++;
}
+static void sse_memcpy(void *dst, void const *src, size_t len)
+{
+ __m128 val;
+ float *d = dst;
+ float const *s = src;
+
+ if (len < sizeof(__m128))
+ return;
+
+
+ for (int i = 0; i < len; i += sizeof(__m128))
+ {
+ val = _mm_load_ps(s);
+ _mm_store_ps(d, val);
+ d += sizeof(__m128) / sizeof(float);
+ s += sizeof(__m128) / sizeof(float);
+ }
+}
+
+static void avx_memcpy(void *dst, void const *src, size_t len)
+{
+ __m256 val;
+ float *d = dst;
+ float const *s = src;
+
+ if (len < sizeof(__m256))
+ return;
+
+ for (int i = 0; i < len; i += sizeof(__m256))
+ {
+ val = _mm256_load_ps(s);
+ _mm256_store_ps(d, val);
+ d += sizeof(__m256) / sizeof(float);
+ s += sizeof(__m256) / sizeof(float);
+ }
+}
+
static void t_simple(struct memcpy_test *test)
{
do_test(test, simple_memcpy);
@@ -135,6 +176,16 @@ static void t_hybrid(struct memcpy_test *test)
do_test(test, memcpy);
}
+static void t_sse(struct memcpy_test *test)
+{
+ do_test(test, sse_memcpy);
+}
+
+static void t_avx(struct memcpy_test *test)
+{
+ do_test(test, avx_memcpy);
+}
+
static struct memcpy_type t[] = {
{
.name = "memcpy",
@@ -157,6 +208,16 @@ static struct memcpy_type t[] = {
.fn = t_hybrid,
},
{
+ .name = "sse",
+ .mask = T_SSE,
+ .fn = t_sse,
+ },
+ {
+ .name = "avx",
+ .mask = T_AVX,
+ .fn = t_avx,
+ },
+ {
.name = NULL,
},
};
@@ -200,8 +261,8 @@ static int setup_tests(void)
void *src, *dst;
int i;
- src = malloc(BUF_SIZE);
- dst = malloc(BUF_SIZE);
+ src = aligned_alloc(64, BUF_SIZE);
+ dst = aligned_alloc(64, BUF_SIZE);
if (!src || !dst) {
free(src);
free(dst);
--
2.13.6
