On 3/5/18 10:22 PM, Jens Axboe wrote:
> On 1/18/18 1:36 PM, Rebecca Cran wrote:
>> 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.
>
> I'm worried that this will break some platforms. A few comments that
> might help rectify that:
>
> 1) Use fio_memalign() instead of relying on a c11 function
> 2) Add a configure check for the avx and sse, separately.
> This includes -msse for the compiler, includes for the
> xmm/imm intrinsics, etc.
>
> I don't care if it segfaults if avx/sse isn't there, but I don't
> want to risk failing a build on some rarely tested platform, just
> to have this as part of the memcpy test. The risk/reward just
> isn't there then.
Something like the below. It fixes up the two above points, and also
corrects some style and declaring 'int i' within the for-loop.
diff --git a/Makefile b/Makefile
index c25b4222e437..1f5cb429f2ef 100644
--- a/Makefile
+++ b/Makefile
@@ -28,6 +28,13 @@ LIBS += -lm $(EXTLIBS)
PROGS = fio
SCRIPTS = $(addprefix $(SRCDIR)/,tools/fio_generate_plots tools/plot/fio2gnuplot tools/genfio tools/fiologparser.py tools/hist/fiologparser_hist.py tools/fio_jsonplus_clat2csv)
+ifdef CONFIG_HAVE_SSE
+CFLAGS += -msse
+endif
+ifdef CONFIG_HAVE_AVX
+CFLAGS += -mavx
+endif
+
ifndef CONFIG_FIO_NO_OPT
CFLAGS += -O3 -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=2
endif
diff --git a/configure b/configure
index 2e8eb180ef50..61ed3d6d1f1b 100755
--- a/configure
+++ b/configure
@@ -2131,6 +2131,48 @@ if compile_prog "" "" "mkdir(a, b)"; then
fi
print_config "mkdir(a, b)" "$mkdir_two"
+##########################################
+# sse probe
+sse="no"
+cat > $TMPC << EOF
+#include <xmmintrin.h>
+#include <immintrin.h>
+int main(int argc, char **argv)
+{
+ __m128 val;
+ float const *src = NULL;
+ float *dst = NULL;
+ val = _mm_load_ps(src);
+ _mm_store_ps(dst, val);
+ return 0;
+}
+EOF
+if compile_prog "-msse" "" "sse"; then
+ sse="yes"
+fi
+print_config "SSE (compiler)" "$sse"
+
+##########################################
+# avx probe
+avx="no"
+cat > $TMPC << EOF
+#include <xmmintrin.h>
+#include <immintrin.h>
+int main(int argc, char **argv)
+{
+ __m256 val;
+ float const *src = NULL;
+ float *dst = NULL;
+ val = _mm256_load_ps(src);
+ _mm256_store_ps(dst, val);
+ return 0;
+}
+EOF
+if compile_prog "-mavx" "" "avx"; then
+ avx="yes"
+fi
+print_config "AVX (compiler)" "$avx"
+
#############################################################################
if test "$wordsize" = "64" ; then
@@ -2377,6 +2419,12 @@ fi
if test "$mkdir_two" = "yes" ; then
output_sym "CONFIG_HAVE_MKDIR_TWO"
fi
+if test "$sse" = "yes" ; then
+ output_sym "CONFIG_HAVE_SSE"
+fi
+if test "$avx" = "yes" ; then
+ output_sym "CONFIG_HAVE_AVX"
+fi
echo "LIBS+=$LIBS" >> $config_host_mak
echo "GFIO_LIBS+=$GFIO_LIBS" >> $config_host_mak
diff --git a/lib/memcpy.c b/lib/memcpy.c
index 00e65aa7d50a..b978a1197fd2 100644
--- a/lib/memcpy.c
+++ b/lib/memcpy.c
@@ -2,9 +2,15 @@
#include <stdlib.h>
#include <string.h>
+#if defined(CONFIG_HAVE_SSE) || defined(CONFIG_HAVE_AVX)
+#include <xmmintrin.h>
+#include <immintrin.h>
+#endif
+
#include "memcpy.h"
#include "rand.h"
#include "../fio_time.h"
+#include "../lib/memalign.h"
#include "../gettime.h"
#include "../fio.h"
@@ -80,6 +86,12 @@ enum {
T_MEMMOVE = 1U << 1,
T_SIMPLE = 1U << 2,
T_HYBRID = 1U << 3,
+#if defined(CONFIG_HAVE_SSE)
+ T_SSE = 1U << 4,
+#endif
+#if defined(CONFIG_HAVE_AVX)
+ T_AVX = 1U << 5,
+#endif
};
#define do_test(test, fn) do { \
@@ -122,6 +134,46 @@ static void simple_memcpy(void *dst, void const *src, size_t len)
*d++ = *s++;
}
+#if defined(CONFIG_HAVE_SSE)
+static void sse_memcpy(void *dst, void const *src, size_t len)
+{
+ __m128 val;
+ float *d = dst;
+ float const *s = src;
+ int i;
+
+ if (len < sizeof(__m128))
+ return;
+
+ for (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);
+ }
+}
+#endif
+
+#if defined(CONFIG_HAVE_AVX)
+static void avx_memcpy(void *dst, void const *src, size_t len)
+{
+ __m256 val;
+ float *d = dst;
+ float const *s = src;
+ int i;
+
+ if (len < sizeof(__m256))
+ return;
+
+ for (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);
+ }
+}
+#endif
+
static void t_simple(struct memcpy_test *test)
{
do_test(test, simple_memcpy);
@@ -135,6 +187,20 @@ static void t_hybrid(struct memcpy_test *test)
do_test(test, memcpy);
}
+#if defined(CONFIG_HAVE_SSE)
+static void t_sse(struct memcpy_test *test)
+{
+ do_test(test, sse_memcpy);
+}
+#endif
+
+#if defined(CONFIG_HAVE_AVX)
+static void t_avx(struct memcpy_test *test)
+{
+ do_test(test, avx_memcpy);
+}
+#endif
+
static struct memcpy_type t[] = {
{
.name = "memcpy",
@@ -156,6 +222,20 @@ static struct memcpy_type t[] = {
.mask = T_HYBRID,
.fn = t_hybrid,
},
+#if defined(CONFIG_HAVE_SSE)
+ {
+ .name = "sse",
+ .mask = T_SSE,
+ .fn = t_sse,
+ },
+#endif
+#if defined(CONFIG_HAVE_AVX)
+ {
+ .name = "avx",
+ .mask = T_AVX,
+ .fn = t_avx,
+ },
+#endif
{
.name = NULL,
},
@@ -200,8 +280,8 @@ static int setup_tests(void)
void *src, *dst;
int i;
- src = malloc(BUF_SIZE);
- dst = malloc(BUF_SIZE);
+ src = fio_memalign(64, BUF_SIZE);
+ dst = fio_memalign(64, BUF_SIZE);
if (!src || !dst) {
free(src);
free(dst);
@@ -222,8 +302,8 @@ static int setup_tests(void)
static void free_tests(void)
{
- free(tests[0].src);
- free(tests[0].dst);
+ fio_memfree(tests[0].src, BUF_SIZE);
+ fio_memfree(tests[0].dst, BUF_SIZE);
}
int fio_memcpy_test(const char *type)
--
Jens Axboe
--
To unsubscribe from this list: send the line "unsubscribe fio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
