[PATCH] benchtests: Add memset zero fill benchmark tests

[PATCH] benchtests: Add memset zero fill benchmark tests

[Naohiro Tamura via Libc-alpha]

Memset takes 0 as the second parameter in most cases.
More than 95% of memset takes 0 as the second parameter in case of
Linux Kernel source code.
However, we cannot measure the zero fill performance by
bench-memset-zerofill.c and bench-memset-large-zerofill.c.
This patch provides bench-memset-zerofill.c and
bench-memset-large-zerofill.c which are suitable to see the
performance of zero fill by fixing the second parameter to 0.
---
 benchtests/Makefile                      |   3 +-
 benchtests/bench-memset-large-zerofill.c | 125 ++++++++++++++++++
 benchtests/bench-memset-zerofill.c       | 156 +++++++++++++++++++++++
 3 files changed, 283 insertions(+), 1 deletion(-)
 create mode 100644 benchtests/bench-memset-large-zerofill.c
 create mode 100644 benchtests/bench-memset-zerofill.c

diff --git a/benchtests/Makefile b/benchtests/Makefile
index 1530939a8ce8..1261f7650fc7 100644
--- a/benchtests/Makefile
+++ b/benchtests/Makefile
@@ -53,7 +53,8 @@ string-benchset := memccpy memchr memcmp memcpy memmem memmove \
 		   strncasecmp strncat strncmp strncpy strnlen strpbrk strrchr \
 		   strspn strstr strcpy_chk stpcpy_chk memrchr strsep strtok \
 		   strcoll memcpy-large memcpy-random memmove-large memset-large \
-		   memcpy-walk memset-walk memmove-walk
+		   memcpy-walk memset-walk memmove-walk memset-zerofill \
+		   memset-large-zerofill
 
 # Build and run locale-dependent benchmarks only if we're building natively.
 ifeq (no,$(cross-compiling))
diff --git a/benchtests/bench-memset-large-zerofill.c b/benchtests/bench-memset-large-zerofill.c
new file mode 100644
index 000000000000..d8eae9d9789f
--- /dev/null
+++ b/benchtests/bench-memset-large-zerofill.c
@@ -0,0 +1,125 @@
+/* Measure memset functions with large data sizes.
+   Copyright (C) 2016-2021 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#define TEST_MAIN
+#define TEST_NAME "memset"
+#define START_SIZE (128 * 1024)
+#define MIN_PAGE_SIZE (getpagesize () + 64 * 1024 * 1024)
+#define TIMEOUT (20 * 60)
+#include "bench-string.h"
+
+#include <assert.h>
+#include "json-lib.h"
+
+void *generic_memset (void *, int, size_t);
+typedef void *(*proto_t) (void *, int, size_t);
+
+IMPL (MEMSET, 1)
+IMPL (generic_memset, 0)
+
+static void
+do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s,
+	     int c __attribute ((unused)), size_t n)
+{
+  size_t i, iters = 16;
+  timing_t start, stop, cur;
+
+  TIMING_NOW (start);
+  for (i = 0; i < iters; ++i)
+    {
+      CALL (impl, s, c, n);
+    }
+  TIMING_NOW (stop);
+
+  TIMING_DIFF (cur, start, stop);
+
+  json_element_double (json_ctx, (double) cur / (double) iters);
+}
+
+static void
+do_test (json_ctx_t *json_ctx, size_t align, int c, size_t len)
+{
+  align &= 63;
+  if ((align + len) * sizeof (CHAR) > page_size)
+    return;
+
+  json_element_object_begin (json_ctx);
+  json_attr_uint (json_ctx, "length", len);
+  json_attr_uint (json_ctx, "alignment", align);
+  json_attr_int (json_ctx, "char", c);
+  json_array_begin (json_ctx, "timings");
+
+  FOR_EACH_IMPL (impl, 0)
+    {
+      do_one_test (json_ctx, impl, (CHAR *) (buf1) + align, c, len);
+      alloc_bufs ();
+    }
+
+  json_array_end (json_ctx);
+  json_element_object_end (json_ctx);
+}
+
+int
+test_main (void)
+{
+  json_ctx_t json_ctx;
+  size_t i;
+  int c;
+
+  test_init ();
+
+  json_init (&json_ctx, 0, stdout);
+
+  json_document_begin (&json_ctx);
+  json_attr_string (&json_ctx, "timing_type", TIMING_TYPE);
+
+  json_attr_object_begin (&json_ctx, "functions");
+  json_attr_object_begin (&json_ctx, TEST_NAME);
+  json_attr_string (&json_ctx, "bench-variant", "large-zerofill");
+
+  json_array_begin (&json_ctx, "ifuncs");
+  FOR_EACH_IMPL (impl, 0)
+    json_element_string (&json_ctx, impl->name);
+  json_array_end (&json_ctx);
+
+  json_array_begin (&json_ctx, "results");
+
+  c = 0;
+  for (i = START_SIZE; i <= MIN_PAGE_SIZE; i <<= 1)
+    {
+      do_test (&json_ctx, 0, c, i);
+      do_test (&json_ctx, 3, c, i);
+    }
+
+  json_array_end (&json_ctx);
+  json_attr_object_end (&json_ctx);
+  json_attr_object_end (&json_ctx);
+  json_document_end (&json_ctx);
+
+  return ret;
+}
+
+#include <support/test-driver.c>
+
+#define libc_hidden_builtin_def(X)
+#define libc_hidden_def(X)
+#define libc_hidden_weak(X)
+#define weak_alias(X,Y)
+#undef MEMSET
+#define MEMSET generic_memset
+#include <string/memset.c>
diff --git a/benchtests/bench-memset-zerofill.c b/benchtests/bench-memset-zerofill.c
new file mode 100644
index 000000000000..ac20ae4c6537
--- /dev/null
+++ b/benchtests/bench-memset-zerofill.c
@@ -0,0 +1,156 @@
+/* Measure memset functions.
+   Copyright (C) 2013-2021 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#define TEST_MAIN
+#ifndef WIDE
+# define TEST_NAME "memset"
+#else
+# define TEST_NAME "wmemset"
+# define generic_memset generic_wmemset
+#endif /* WIDE */
+#define MIN_PAGE_SIZE 131072
+#include "bench-string.h"
+
+#include "json-lib.h"
+
+#ifdef WIDE
+CHAR *generic_wmemset (CHAR *, CHAR, size_t);
+#else
+void *generic_memset (void *, int, size_t);
+#endif
+
+typedef void *(*proto_t) (void *, int, size_t);
+
+IMPL (MEMSET, 1)
+IMPL (generic_memset, 0)
+
+static void
+do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s,
+	     int c __attribute ((unused)), size_t n)
+{
+  size_t i, iters = INNER_LOOP_ITERS;
+  timing_t start, stop, cur;
+
+  TIMING_NOW (start);
+  for (i = 0; i < iters; ++i)
+    {
+      CALL (impl, s, c, n);
+    }
+  TIMING_NOW (stop);
+
+  TIMING_DIFF (cur, start, stop);
+
+  json_element_double (json_ctx, (double) cur / (double) iters);
+}
+
+static void
+do_test (json_ctx_t *json_ctx, size_t align, int c, size_t len)
+{
+  align &= 4095;
+  if ((align + len) * sizeof (CHAR) > page_size)
+    return;
+
+  json_element_object_begin (json_ctx);
+  json_attr_uint (json_ctx, "length", len);
+  json_attr_uint (json_ctx, "alignment", align);
+  json_attr_int (json_ctx, "char", c);
+  json_array_begin (json_ctx, "timings");
+
+  FOR_EACH_IMPL (impl, 0)
+    {
+      do_one_test (json_ctx, impl, (CHAR *) (buf1) + align, c, len);
+      alloc_bufs ();
+    }
+
+  json_array_end (json_ctx);
+  json_element_object_end (json_ctx);
+}
+
+int
+test_main (void)
+{
+  json_ctx_t json_ctx;
+  size_t i;
+  int c = 0;
+
+  test_init ();
+
+  json_init (&json_ctx, 0, stdout);
+
+  json_document_begin (&json_ctx);
+  json_attr_string (&json_ctx, "timing_type", TIMING_TYPE);
+
+  json_attr_object_begin (&json_ctx, "functions");
+  json_attr_object_begin (&json_ctx, TEST_NAME);
+  json_attr_string (&json_ctx, "bench-variant", "default-zerofill");
+
+  json_array_begin (&json_ctx, "ifuncs");
+  FOR_EACH_IMPL (impl, 0)
+    json_element_string (&json_ctx, impl->name);
+  json_array_end (&json_ctx);
+
+  json_array_begin (&json_ctx, "results");
+
+  c = 0;
+  for (i = 0; i < 18; ++i)
+    do_test (&json_ctx, 0, c, 1 << i);
+  for (i = 1; i < 64; ++i)
+    {
+      do_test (&json_ctx, i, c, i);
+      do_test (&json_ctx, 4096 - i, c, i);
+      do_test (&json_ctx, 4095, c, i);
+      if (i & (i - 1))
+	do_test (&json_ctx, 0, c, i);
+    }
+  for (i = 32; i < 512; i+=32)
+    {
+      do_test (&json_ctx, 0, c, i);
+      do_test (&json_ctx, i, c, i);
+    }
+  do_test (&json_ctx, 1, c, 14);
+  do_test (&json_ctx, 3, c, 1024);
+  do_test (&json_ctx, 4, c, 64);
+  do_test (&json_ctx, 2, c, 25);
+  for (i = 33; i <= 256; i += 4)
+    {
+      do_test (&json_ctx, 0, c, 32 * i);
+      do_test (&json_ctx, i, c, 32 * i);
+    }
+
+  json_array_end (&json_ctx);
+  json_attr_object_end (&json_ctx);
+  json_attr_object_end (&json_ctx);
+  json_document_end (&json_ctx);
+
+  return ret;
+}
+
+#include <support/test-driver.c>
+
+#define libc_hidden_builtin_def(X)
+#define libc_hidden_def(X)
+#define libc_hidden_weak(X)
+#define weak_alias(X,Y)
+#ifndef WIDE
+# undef MEMSET
+# define MEMSET generic_memset
+# include <string/memset.c>
+#else
+# define WMEMSET generic_wmemset
+# include <wcsmbs/wmemset.c>
+#endif
-- 
2.17.1

Re: [PATCH] benchtests: Add memset zero fill benchmark tests

[Lucas A. M. Magalhaes via Libc-alpha]

Hi Naohiro,
Thanks for working on this.

I like the idea of a benchmark specific for 0 on memset. However having two
implementations seems too much. I would rather see just one
bench-memset-zerofill.c. What I guess would be even better is to have this
performance test inside bench-memset.c and bench-memset-large.c.

Quoting Naohiro Tamura via Libc-alpha (2021-07-13 05:22:14)
> Memset takes 0 as the second parameter in most cases.
> More than 95% of memset takes 0 as the second parameter in case of
> Linux Kernel source code.
The Linux Kernel does not use glibc, it has his own memset implementation.
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/lib/string.c#n784
Therefore IMO this argument is not suited for this commit.

> However, we cannot measure the zero fill performance by
> bench-memset-zerofill.c and bench-memset-large-zerofill.c.
> This patch provides bench-memset-zerofill.c and
> bench-memset-large-zerofill.c which are suitable to see the
> performance of zero fill by fixing the second parameter to 0.
In this section I guess you mistake bench-memset.c and bench-memset-large.c for
bench-memset-zerofill.c and bench-memset-large-zerofill.c.

> ---
>  benchtests/Makefile                      |   3 +-
>  benchtests/bench-memset-large-zerofill.c | 125 ++++++++++++++++++
>  benchtests/bench-memset-zerofill.c       | 156 +++++++++++++++++++++++
>  3 files changed, 283 insertions(+), 1 deletion(-)
>  create mode 100644 benchtests/bench-memset-large-zerofill.c
>  create mode 100644 benchtests/bench-memset-zerofill.c
> 
> diff --git a/benchtests/Makefile b/benchtests/Makefile
> index 1530939a8ce8..1261f7650fc7 100644
> --- a/benchtests/Makefile
> +++ b/benchtests/Makefile
> @@ -53,7 +53,8 @@ string-benchset := memccpy memchr memcmp memcpy memmem memmove \
>                    strncasecmp strncat strncmp strncpy strnlen strpbrk strrchr \
>                    strspn strstr strcpy_chk stpcpy_chk memrchr strsep strtok \
>                    strcoll memcpy-large memcpy-random memmove-large memset-large \
> -                  memcpy-walk memset-walk memmove-walk
> +                  memcpy-walk memset-walk memmove-walk memset-zerofill \
> +                  memset-large-zerofill
>  
>  # Build and run locale-dependent benchmarks only if we're building natively.
>  ifeq (no,$(cross-compiling))
> diff --git a/benchtests/bench-memset-large-zerofill.c b/benchtests/bench-memset-large-zerofill.c
> new file mode 100644
> index 000000000000..d8eae9d9789f
> --- /dev/null
> +++ b/benchtests/bench-memset-large-zerofill.c
> @@ -0,0 +1,125 @@
> +/* Measure memset functions with large data sizes.
Please fix this description.

> +   Copyright (C) 2016-2021 Free Software Foundation, Inc.
Just 2021 here.

> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +#define TEST_MAIN
> +#define TEST_NAME "memset"
> +#define START_SIZE (128 * 1024)
> +#define MIN_PAGE_SIZE (getpagesize () + 64 * 1024 * 1024)
> +#define TIMEOUT (20 * 60)
> +#include "bench-string.h"
> +
> +#include <assert.h>
> +#include "json-lib.h"
> +
This code don't need the assert.h.

> +void *generic_memset (void *, int, size_t);
> +typedef void *(*proto_t) (void *, int, size_t);
> +
> +IMPL (MEMSET, 1)
> +IMPL (generic_memset, 0)
> +
> +static void
> +do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s,
> +            int c __attribute ((unused)), size_t n)
> +{
> +  size_t i, iters = 16;
> +  timing_t start, stop, cur;
> +
> +  TIMING_NOW (start);
> +  for (i = 0; i < iters; ++i)
> +    {
> +      CALL (impl, s, c, n);
> +    }
> +  TIMING_NOW (stop);
> +
> +  TIMING_DIFF (cur, start, stop);
> +
> +  json_element_double (json_ctx, (double) cur / (double) iters);
> +}
> +
> +static void
> +do_test (json_ctx_t *json_ctx, size_t align, int c, size_t len)
> +{
> +  align &= 63;
> +  if ((align + len) * sizeof (CHAR) > page_size)
> +    return;
> +
> +  json_element_object_begin (json_ctx);
> +  json_attr_uint (json_ctx, "length", len);
> +  json_attr_uint (json_ctx, "alignment", align);
> +  json_attr_int (json_ctx, "char", c);
> +  json_array_begin (json_ctx, "timings");
> +
> +  FOR_EACH_IMPL (impl, 0)
> +    {
> +      do_one_test (json_ctx, impl, (CHAR *) (buf1) + align, c, len);
> +      alloc_bufs ();
> +    }
> +
> +  json_array_end (json_ctx);
> +  json_element_object_end (json_ctx);
> +}
> +
> +int
> +test_main (void)
> +{
> +  json_ctx_t json_ctx;
> +  size_t i;
> +  int c;
> +
> +  test_init ();
> +
> +  json_init (&json_ctx, 0, stdout);
> +
> +  json_document_begin (&json_ctx);
> +  json_attr_string (&json_ctx, "timing_type", TIMING_TYPE);
> +
> +  json_attr_object_begin (&json_ctx, "functions");
> +  json_attr_object_begin (&json_ctx, TEST_NAME);
> +  json_attr_string (&json_ctx, "bench-variant", "large-zerofill");
> +
> +  json_array_begin (&json_ctx, "ifuncs");
> +  FOR_EACH_IMPL (impl, 0)
> +    json_element_string (&json_ctx, impl->name);
> +  json_array_end (&json_ctx);
> +
> +  json_array_begin (&json_ctx, "results");
> +
> +  c = 0;
> +  for (i = START_SIZE; i <= MIN_PAGE_SIZE; i <<= 1)
> +    {
> +      do_test (&json_ctx, 0, c, i);
> +      do_test (&json_ctx, 3, c, i);
> +    }
> +
> +  json_array_end (&json_ctx);
> +  json_attr_object_end (&json_ctx);
> +  json_attr_object_end (&json_ctx);
> +  json_document_end (&json_ctx);
> +
> +  return ret;
> +}
> +
> +#include <support/test-driver.c>
> +
> +#define libc_hidden_builtin_def(X)
> +#define libc_hidden_def(X)
> +#define libc_hidden_weak(X)
> +#define weak_alias(X,Y)
> +#undef MEMSET
> +#define MEMSET generic_memset
> +#include <string/memset.c>
> diff --git a/benchtests/bench-memset-zerofill.c b/benchtests/bench-memset-zerofill.c
> new file mode 100644
> index 000000000000..ac20ae4c6537
> --- /dev/null
> +++ b/benchtests/bench-memset-zerofill.c
> @@ -0,0 +1,156 @@
> +/* Measure memset functions.
Fix the description.

> +   Copyright (C) 2013-2021 Free Software Foundation, Inc.
Only 2021 here.

> +   This file is part of the GNU C Library.
> +
> +   The GNU C Library is free software; you can redistribute it and/or
> +   modify it under the terms of the GNU Lesser General Public
> +   License as published by the Free Software Foundation; either
> +   version 2.1 of the License, or (at your option) any later version.
> +
> +   The GNU C Library is distributed in the hope that it will be useful,
> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> +   Lesser General Public License for more details.
> +
> +   You should have received a copy of the GNU Lesser General Public
> +   License along with the GNU C Library; if not, see
> +   <https://www.gnu.org/licenses/>.  */
> +
> +#define TEST_MAIN
> +#ifndef WIDE
> +# define TEST_NAME "memset"
> +#else
> +# define TEST_NAME "wmemset"
> +# define generic_memset generic_wmemset
> +#endif /* WIDE */
> +#define MIN_PAGE_SIZE 131072
> +#include "bench-string.h"
> +
> +#include "json-lib.h"
> +
> +#ifdef WIDE
> +CHAR *generic_wmemset (CHAR *, CHAR, size_t);
> +#else
> +void *generic_memset (void *, int, size_t);
> +#endif
> +
> +typedef void *(*proto_t) (void *, int, size_t);
> +
> +IMPL (MEMSET, 1)
> +IMPL (generic_memset, 0)
> +
> +static void
> +do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s,
> +            int c __attribute ((unused)), size_t n)
> +{
> +  size_t i, iters = INNER_LOOP_ITERS;
> +  timing_t start, stop, cur;
> +
> +  TIMING_NOW (start);
> +  for (i = 0; i < iters; ++i)
> +    {
> +      CALL (impl, s, c, n);
> +    }
> +  TIMING_NOW (stop);
> +
> +  TIMING_DIFF (cur, start, stop);
> +
> +  json_element_double (json_ctx, (double) cur / (double) iters);
> +}
> +
> +static void
> +do_test (json_ctx_t *json_ctx, size_t align, int c, size_t len)
> +{
> +  align &= 4095;
> +  if ((align + len) * sizeof (CHAR) > page_size)
> +    return;
> +
> +  json_element_object_begin (json_ctx);
> +  json_attr_uint (json_ctx, "length", len);
> +  json_attr_uint (json_ctx, "alignment", align);
> +  json_attr_int (json_ctx, "char", c);
> +  json_array_begin (json_ctx, "timings");
> +
> +  FOR_EACH_IMPL (impl, 0)
> +    {
> +      do_one_test (json_ctx, impl, (CHAR *) (buf1) + align, c, len);
> +      alloc_bufs ();
> +    }
> +
> +  json_array_end (json_ctx);
> +  json_element_object_end (json_ctx);
> +}
> +
> +int
> +test_main (void)
> +{
> +  json_ctx_t json_ctx;
> +  size_t i;
> +  int c = 0;
> +
> +  test_init ();
> +
> +  json_init (&json_ctx, 0, stdout);
> +
> +  json_document_begin (&json_ctx);
> +  json_attr_string (&json_ctx, "timing_type", TIMING_TYPE);
> +
> +  json_attr_object_begin (&json_ctx, "functions");
> +  json_attr_object_begin (&json_ctx, TEST_NAME);
> +  json_attr_string (&json_ctx, "bench-variant", "default-zerofill");
> +
> +  json_array_begin (&json_ctx, "ifuncs");
> +  FOR_EACH_IMPL (impl, 0)
> +    json_element_string (&json_ctx, impl->name);
> +  json_array_end (&json_ctx);
> +
> +  json_array_begin (&json_ctx, "results");
> +
> +  c = 0;
> +  for (i = 0; i < 18; ++i)
> +    do_test (&json_ctx, 0, c, 1 << i);
> +  for (i = 1; i < 64; ++i)
> +    {
> +      do_test (&json_ctx, i, c, i);
> +      do_test (&json_ctx, 4096 - i, c, i);
> +      do_test (&json_ctx, 4095, c, i);
> +      if (i & (i - 1))
> +       do_test (&json_ctx, 0, c, i);
> +    }
> +  for (i = 32; i < 512; i+=32)
> +    {
> +      do_test (&json_ctx, 0, c, i);
> +      do_test (&json_ctx, i, c, i);
> +    }
> +  do_test (&json_ctx, 1, c, 14);
> +  do_test (&json_ctx, 3, c, 1024);
> +  do_test (&json_ctx, 4, c, 64);
> +  do_test (&json_ctx, 2, c, 25);
> +  for (i = 33; i <= 256; i += 4)
> +    {
> +      do_test (&json_ctx, 0, c, 32 * i);
> +      do_test (&json_ctx, i, c, 32 * i);
> +    }
> +
> +  json_array_end (&json_ctx);
> +  json_attr_object_end (&json_ctx);
> +  json_attr_object_end (&json_ctx);
> +  json_document_end (&json_ctx);
> +
> +  return ret;
> +}
> +
> +#include <support/test-driver.c>
> +
> +#define libc_hidden_builtin_def(X)
> +#define libc_hidden_def(X)
> +#define libc_hidden_weak(X)
> +#define weak_alias(X,Y)
> +#ifndef WIDE
> +# undef MEMSET
> +# define MEMSET generic_memset
> +# include <string/memset.c>
> +#else
> +# define WMEMSET generic_wmemset
> +# include <wcsmbs/wmemset.c>
> +#endif
> -- 
> 2.17.1
> 

---
Lucas A. M. Magalhães

[PATCH] benchtests: Add memset zero fill benchmark tests

[Wilco Dijkstra via Libc-alpha]

Hi,

> I like the idea of a benchmark specific for 0 on memset. However having two
> implementations seems too much. I would rather see just one
> bench-memset-zerofill.c. What I guess would be even better is to have this
> performance test inside bench-memset.c and bench-memset-large.c.

I agree just copying the files is not a good idea. Currently bench-memset and
bench-memset-walk already test zero memsets. Bench-memset-large could just
test zero since that is the most common, especially for large sizes. Reducing the
number of non-zero tests in bench-memset would make it more representative -
you could do the main set of tests with zero only and then have a small selection
where it alternates between zero and non-zero.

> Quoting Naohiro Tamura via Libc-alpha (2021-07-13 05:22:14)
>> Memset takes 0 as the second parameter in most cases.
>> More than 95% of memset takes 0 as the second parameter in case of
>> Linux Kernel source code.
> The Linux Kernel does not use glibc, it has his own memset implementation.
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/lib/string.c#n784
> Therefore IMO this argument is not suited for this commit.

The argument is true in general - you could simply state that almost all memset
calls are zeroing without mentioning the Linux kernel. In some old stats from
SPEC I saw about 1.8% non-zero memsets.

Cheers,
Wilco

Re: [PATCH] benchtests: Add memset zero fill benchmark tests

[Noah Goldstein via Libc-alpha]

On Tue, Jul 13, 2021 at 12:15 PM Wilco Dijkstra via Libc-alpha <
libc-alpha@sourceware.org> wrote:

> Hi,
>
> > I like the idea of a benchmark specific for 0 on memset. However having
> two
> > implementations seems too much. I would rather see just one
> > bench-memset-zerofill.c. What I guess would be even better is to have
> this
> > performance test inside bench-memset.c and bench-memset-large.c.
>
> I agree just copying the files is not a good idea. Currently bench-memset
> and
> bench-memset-walk already test zero memsets. Bench-memset-large could just
> test zero since that is the most common, especially for large sizes.
> Reducing the
> number of non-zero tests in bench-memset would make it more representative
> -
> you could do the main set of tests with zero only and then have a small
> selection
> where it alternates between zero and non-zero.
>

I'm in favor of a seperate file. On some x86_64 systems writing zeros to a
cacheline
that has not been modified can leave the cacheline in an unmodified
state[1] which
affects memory bandwidth on the writeback to DRAM for larger regions. I can
imagine
we might want to test memset zero on unmodified vs modified region which
will require
unique setup that I think justifies a separate file (at least for
memset-large-zero).

[1] https://travisdowns.github.io/blog/2020/05/13/intel-zero-opt.html


>
> > Quoting Naohiro Tamura via Libc-alpha (2021-07-13 05:22:14)
> >> Memset takes 0 as the second parameter in most cases.
> >> More than 95% of memset takes 0 as the second parameter in case of
> >> Linux Kernel source code.
> > The Linux Kernel does not use glibc, it has his own memset
> implementation.
> >
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/lib/string.c#n784
> > Therefore IMO this argument is not suited for this commit.
>
> The argument is true in general - you could simply state that almost all
> memset
> calls are zeroing without mentioning the Linux kernel. In some old stats
> from
> SPEC I saw about 1.8% non-zero memsets.
>
> Cheers,
> Wilco

Re: [PATCH] benchtests: Add memset zero fill benchmark tests

[naohirot--- via Libc-alpha]

Hi Lucas, Wilco, Noah,

Thanks for the comments.

> From: Noah Goldstein <goldstein.w.n@gmail.com>
> Sent: Wednesday, 14 July 2021 03:47
> 
> On Tue, Jul 13, 2021 at 12:15 PM Wilco Dijkstra via Libc-alpha <libc-alpha@sourceware.org<mailto:libc-alpha@sourceware.org>> wrote:
> > Hi,
> > 
> > > I like the idea of a benchmark specific for 0 on memset. However having two
> > > implementations seems too much. I would rather see just one
> > > bench-memset-zerofill.c. What I guess would be even better is to have this
> > > performance test inside bench-memset.c and bench-memset-large.c.

Yes, that's one way. In this case, benchtests/scripts/plot_strings.py
cannot be used as is to visualize zero-fill performance.
We need to extract the zero fill data from the JSON bench results.

> > I agree just copying the files is not a good idea. Currently bench-memset and
> > bench-memset-walk already test zero memsets. Bench-memset-large could just
> > test zero since that is the most common, especially for large sizes. Reducing the
> > number of non-zero tests in bench-memset would make it more representative -
> > you could do the main set of tests with zero only and then have a small selection
> > where it alternates between zero and non-zero.

Yes, this is one way too.
However, we also need to extract the zero fill data from JSON bench result.

> 
> I'm in favor of a seperate file. On some x86_64 systems writing zeros to a cacheline
> that has not been modified can leave the cacheline in an unmodified state[1] which
> affects memory bandwidth on the writeback to DRAM for larger regions. I can imagine
> we might want to test memset zero on unmodified vs modified region which will require
> unique setup that I think justifies a separate file (at least for memset-large-zero).
> 
> [1] https://travisdowns.github.io/blog/2020/05/13/intel-zero-opt.html

I throughly read the blog, it's very interesting.
It seems that Skylake-S microarchitcture has some zero-over-zero
optimization in the size range of L3 and RAM.

So why don't we proceed like this?
- create a new benchtest file, ex. memset-zerofill.c.
  it measures zero-over-zero and zero-over-one from 32KB(L1), through
  L2 and L3, to 64MB(RAM)
- create a script to extract some part of data such as zero-over-zero
  or zero-over-one from JSON bench result to CSV in the first place in
  order to create graph in spreadsheet
- update benchtests/scripts/plot_strings.py later so that it can draw
  zero-over-zero and zero-over-one separately

> 
> > 
> > > Quoting Naohiro Tamura via Libc-alpha (2021-07-13 05:22:14)
> > >> Memset takes 0 as the second parameter in most cases.
> > >> More than 95% of memset takes 0 as the second parameter in case of
> > >> Linux Kernel source code.
> > > The Linux Kernel does not use glibc, it has his own memset implementation.
> > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/lib/string.c#n784
> > > Therefore IMO this argument is not suited for this commit.
> > 
> > The argument is true in general - you could simply state that almost all memset
> > calls are zeroing without mentioning the Linux kernel. In some old stats from
> > SPEC I saw about 1.8% non-zero memsets.

I'll fix that.

Thanks.
Naohiro

RE: [PATCH] benchtests: Add memset zero fill benchmark tests

[naohirot--- via Libc-alpha]

Hi Lucas, Wilco, Noah,

> From: Tamura, Naohiro/田村 直広 <naohirot@fujitsu.com>
> Sent: Thursday, July 15, 2021 5:16 PM

> > [1] https://travisdowns.github.io/blog/2020/05/13/intel-zero-opt.html
> 
> I throughly read the blog, it's very interesting.
> It seems that Skylake-S microarchitcture has some zero-over-zero
> optimization in the size range of L3 and RAM.
> 
> So why don't we proceed like this?
> - create a new benchtest file, ex. memset-zerofill.c.
>   it measures zero-over-zero and zero-over-one from 32KB(L1), through
>   L2 and L3, to 64MB(RAM)
> - create a script to extract some part of data such as zero-over-zero
>   or zero-over-one from JSON bench result to CSV in the first place in
>   order to create graph in spreadsheet
> - update benchtests/scripts/plot_strings.py later so that it can draw
>   zero-over-zero and zero-over-one separately

Here is a result of the memset zero fill patch v2 [1] using the source code A64FX memset patch v2 [2].

The first graph [3] is the zero-over-zero result by typing the command:

$ cat bench-memset-zerofill.out | jq -r '.functions.memset."bench-variant"="zerofill-0o0" | del(.functions.memset.results[] | select(.char1 != 0))' | plot_strings.py -l -p thru -v -

The second graph [4] is the zero-over-one result by typing the command:

$ cat bench-memset-zerofill.out | jq -r '.functions.memset."bench-variant"="zerofill-0o1" | del(.functions.memset.results[] | select(.char1 != 1))' | plot_strings.py -l -p thru -v -

In case of zero-over-zero, it's very interesting that memset_generic is faster than memset_a64fx.
On the other hand, however, in case of zero-over-one, memset_a64fx is faster is faster than memset_generic.
It may be due to SVE code.
Comparing the memset_generic performance, that is non SVE code, between zero-over-zero and zero-over-one, 
it seems that a64fx micro-architecture has some zero-over-zero optimization?.

[1] https://sourceware.org/pipermail/libc-alpha/2021-July/129290.html
[2] https://sourceware.org/pipermail/libc-alpha/2021-July/128857.html
[3] https://drive.google.com/file/d/1iNlp-srAq-p9fr7PuRh62TRj-V2eJi0U/view?usp=sharing
[4] https://drive.google.com/file/d/1UipGRHUWU8WJ1-EbV6YzoTcOtxtOhAy0/view?usp=sharing

Thanks.
Naohiro

Re: [PATCH] benchtests: Add memset zero fill benchmark tests

[Wilco Dijkstra via Libc-alpha]

Hi Naohiro,

> In case of zero-over-zero, it's very interesting that memset_generic is faster than memset_a64fx.

That's because it uses DC ZVA for the full range rather than only for sizes > 8MB like in memset v2.

> On the other hand, however, in case of zero-over-one, memset_a64fx is faster is faster than memset_generic.
> It may be due to SVE code.

Yes, you get more bandwidth with 512-bit SVE stores than 128-bit Neon stores.

> Comparing the memset_generic performance, that is non SVE code, between zero-over-zero and
> zero-over-one, it seems that a64fx micro-architecture has some zero-over-zero optimization?.

No, that does not prove that the previous value matters, just that storing zeroes is faster.

Cheers,
Wilco