* [COMMITTED 2/3] benchtests: Memory walking benchmark for memset
2017-10-05 16:55 [COMMITTED 1/3] benchtests: Memory walking benchmark for memcpy Siddhesh Poyarekar
@ 2017-10-05 16:55 ` Siddhesh Poyarekar
2017-10-05 16:55 ` [COMMITTED 3/3] benchtests: Memory walking benchmark for memmove Siddhesh Poyarekar
1 sibling, 0 replies; 3+ messages in thread
From: Siddhesh Poyarekar @ 2017-10-05 16:55 UTC (permalink / raw)
To: libc-alpha
This benchmark is an attempt to eliminate cache effects from string
benchmarks. The benchmark walks backward through a large memory area
and sets different sizes of memory and alignments one at a time
instead of looping around in the same memory area. This is a good
metric to have alongside the simple memset benchmark (which is only
really useful for smaller sizes) especially for larger sizes where the
likelihood of the call being done only once is pretty high.
* benchtests/bench-memset-walk.c: New file.
* benchtests/Makefile (string-benchset): Add it.
---
ChangeLog | 3 +
benchtests/Makefile | 2 +-
benchtests/bench-memset-walk.c | 138 +++++++++++++++++++++++++++++++++++++++++
3 files changed, 142 insertions(+), 1 deletion(-)
create mode 100644 benchtests/bench-memset-walk.c
diff --git a/ChangeLog b/ChangeLog
index a86faeb..5821c5d 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,8 @@
2017-10-05 Siddhesh Poyarekar <siddhesh@sourceware.org>
+ * benchtests/bench-memset-walk.c: New file.
+ * benchtests/Makefile (string-benchset): Add it.
+
* benchtests/bench-memcpy-walk.c: New file.
* benchtests/Makefile (string-benchset): Add it.
diff --git a/benchtests/Makefile b/benchtests/Makefile
index d086cc6..5a07639 100644
--- a/benchtests/Makefile
+++ b/benchtests/Makefile
@@ -38,7 +38,7 @@ string-benchset := bcopy bzero 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
+ memcpy-walk memset-walk
# Build and run locale-dependent benchmarks only if we're building natively.
ifeq (no,$(cross-compiling))
diff --git a/benchtests/bench-memset-walk.c b/benchtests/bench-memset-walk.c
new file mode 100644
index 0000000..59d2626
--- /dev/null
+++ b/benchtests/bench-memset-walk.c
@@ -0,0 +1,138 @@
+/* Measure memset function throughput with large data sizes.
+ Copyright (C) 2017 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
+ <http://www.gnu.org/licenses/>. */
+
+#define TEST_MAIN
+#ifndef WIDE
+# define TEST_NAME "memset"
+#else
+# define TEST_NAME "wmemset"
+#endif /* WIDE */
+#define START_SIZE (1)
+#define MIN_PAGE_SIZE (getpagesize () + 32 * 1024 * 1024)
+#define TIMEOUT (20 * 60)
+#include "bench-string.h"
+
+#ifndef WIDE
+# define MEMSET memset
+# define CHAR char
+# define SIMPLE_MEMSET simple_memset
+# define MEMCMP memcmp
+#else
+# include <wchar.h>
+# define MEMSET wmemset
+# define CHAR wchar_t
+# define SIMPLE_MEMSET simple_wmemset
+# define MEMCMP wmemcmp
+#endif /* WIDE */
+
+#include <assert.h>
+#include "json-lib.h"
+
+
+typedef CHAR *(*proto_t) (CHAR *, int, size_t);
+
+CHAR *
+inhibit_loop_to_libcall
+SIMPLE_MEMSET (CHAR *s, int c, size_t n)
+{
+ CHAR *r = s, *end = s + n;
+ while (r < end)
+ *r++ = c;
+ return s;
+}
+
+IMPL (SIMPLE_MEMSET, 1)
+
+static void
+do_one_test (json_ctx_t *json_ctx, impl_t *impl, CHAR *s, CHAR *s_end,
+ int c __attribute ((unused)), size_t n)
+{
+ size_t i, iters = MIN_PAGE_SIZE / n;
+ timing_t start, stop, cur;
+
+ TIMING_NOW (start);
+ for (i = 0; i < iters && s <= s_end; s++, i++)
+ CALL (impl, s, c, n);
+ TIMING_NOW (stop);
+
+ TIMING_DIFF (cur, start, stop);
+
+ /* Get time taken per function call. */
+ json_element_double (json_ctx, (double) cur * n / i);
+}
+
+static void
+do_test (json_ctx_t *json_ctx, int c, size_t len)
+{
+ json_element_object_begin (json_ctx);
+ json_attr_uint (json_ctx, "length", len);
+ json_attr_uint (json_ctx, "char", c);
+ json_array_begin (json_ctx, "timings");
+
+ FOR_EACH_IMPL (impl, 0)
+ {
+ do_one_test (json_ctx, impl, (CHAR *) buf1,
+ (CHAR *) buf1 + MIN_PAGE_SIZE - len, c, len);
+ realloc_bufs ();
+ }
+
+ json_array_end (json_ctx);
+ json_element_object_end (json_ctx);
+}
+
+int
+test_main (void)
+{
+ json_ctx_t json_ctx;
+ size_t i;
+
+ 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", "walk");
+
+ 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");
+ for (i = START_SIZE; i <= MIN_PAGE_SIZE; i <<= 1)
+ /* Test length alignments from 0-16 bytes. */
+ for (int j = 0; j < i && j < 16; j++)
+ do_test (&json_ctx, 65, i + j);
+
+ for (i = START_SIZE; i <= MIN_PAGE_SIZE; i <<= 1)
+ for (int j = 0; j < i && j < 16; j++)
+ do_test (&json_ctx, 0, i + j);
+
+ 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>
--
2.7.4
^ permalink raw reply [flat|nested] 3+ messages in thread
* [COMMITTED 3/3] benchtests: Memory walking benchmark for memmove
2017-10-05 16:55 [COMMITTED 1/3] benchtests: Memory walking benchmark for memcpy Siddhesh Poyarekar
2017-10-05 16:55 ` [COMMITTED 2/3] benchtests: Memory walking benchmark for memset Siddhesh Poyarekar
@ 2017-10-05 16:55 ` Siddhesh Poyarekar
1 sibling, 0 replies; 3+ messages in thread
From: Siddhesh Poyarekar @ 2017-10-05 16:55 UTC (permalink / raw)
To: libc-alpha
This benchmark is an attempt to eliminate cache effects from string
benchmarks. The benchmark walks both ways through a large memory area
and copies different sizes of memory and alignments one at a time
instead of looping around in the same memory area. This is a good
metric to have alongside the simple memmove benchmark (which is only
really useful for smaller sizes) especially for larger sizes where the
likelihood of the call being done only once is pretty high.
This benchmark is different from memcpy in that it also tests
overlapping copies.
* benchtests/bench-memmove-walk.c: New file.
* benchtests/Makefile (string-benchset): Add it.
---
ChangeLog | 3 +
benchtests/Makefile | 2 +-
benchtests/bench-memmove-walk.c | 143 ++++++++++++++++++++++++++++++++++++++++
3 files changed, 147 insertions(+), 1 deletion(-)
create mode 100644 benchtests/bench-memmove-walk.c
diff --git a/ChangeLog b/ChangeLog
index 5821c5d..6d4b5f2 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,8 @@
2017-10-05 Siddhesh Poyarekar <siddhesh@sourceware.org>
+ * benchtests/bench-memmove-walk.c: New file.
+ * benchtests/Makefile (string-benchset): Add it.
+
* benchtests/bench-memset-walk.c: New file.
* benchtests/Makefile (string-benchset): Add it.
diff --git a/benchtests/Makefile b/benchtests/Makefile
index 5a07639..d8681fc 100644
--- a/benchtests/Makefile
+++ b/benchtests/Makefile
@@ -38,7 +38,7 @@ string-benchset := bcopy bzero 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
+ memcpy-walk memset-walk memmove-walk
# Build and run locale-dependent benchmarks only if we're building natively.
ifeq (no,$(cross-compiling))
diff --git a/benchtests/bench-memmove-walk.c b/benchtests/bench-memmove-walk.c
new file mode 100644
index 0000000..54dcd64
--- /dev/null
+++ b/benchtests/bench-memmove-walk.c
@@ -0,0 +1,143 @@
+/* Measure memmove function combined throughput for different alignments.
+ Copyright (C) 2017 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
+ <http://www.gnu.org/licenses/>. */
+
+/* This microbenchmark measures the throughput of memmove for various sizes from
+ 1 byte to 32MiB, doubling every iteration and then misaligning by 0-15
+ bytes. The copies are done from source to destination and then back and the
+ source walks forward across the array and the destination walks backward by
+ one byte each, thus measuring misaligned accesses as well. The idea is to
+ avoid caching effects by copying a different string and far enough from each
+ other, walking in different directions so that we can measure prefetcher
+ efficiency (software or hardware) more closely than with a loop copying the
+ same data over and over, which eventually only gives us L1 cache
+ performance. */
+
+#ifndef MEMMOVE_RESULT
+# define MEMMOVE_RESULT(dst, len) dst
+# define START_SIZE 1
+# define MIN_PAGE_SIZE (getpagesize () + 32 * 1024 * 1024)
+# define TEST_MAIN
+# define TEST_NAME "memmove"
+# define TIMEOUT (20 * 60)
+# include "bench-string.h"
+
+IMPL (memmove, 1)
+#endif
+
+#include "json-lib.h"
+
+typedef char *(*proto_t) (char *, const char *, size_t);
+
+static void
+do_one_test (json_ctx_t *json_ctx, impl_t *impl, char *dst, char *src,
+ size_t len)
+{
+ size_t i, iters = MIN_PAGE_SIZE / len;
+ timing_t start, stop, cur;
+
+ char *dst_end = dst + MIN_PAGE_SIZE - len;
+ char *src_end = src + MIN_PAGE_SIZE - len;
+
+ TIMING_NOW (start);
+ /* Copy the entire buffer back and forth, LEN at a time. */
+ for (i = 0; i < iters && dst_end >= dst && src <= src_end; src++, dst_end--)
+ {
+ CALL (impl, dst_end, src, len);
+ CALL (impl, src, dst_end, len);
+ i += 2;
+ }
+ TIMING_NOW (stop);
+
+ TIMING_DIFF (cur, start, stop);
+
+ /* Get time taken per function call. */
+ json_element_double (json_ctx, (double) cur * len / i);
+}
+
+static void
+do_test (json_ctx_t *json_ctx, size_t len, bool overlap)
+{
+ json_element_object_begin (json_ctx);
+ json_attr_uint (json_ctx, "length", (double) len);
+ json_array_begin (json_ctx, "timings");
+
+ if (overlap)
+ buf2 = buf1;
+
+ /* First the non-overlapping moves. */
+ FOR_EACH_IMPL (impl, 0)
+ do_one_test (json_ctx, impl, (char *) buf2, (char *) buf1, len);
+
+ json_array_end (json_ctx);
+ json_element_object_end (json_ctx);
+}
+
+int
+test_main (void)
+{
+ json_ctx_t json_ctx;
+ size_t i;
+
+ 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, "memmove");
+ json_attr_string (&json_ctx, "bench-variant", "walk");
+
+ 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");
+ /* Non-overlapping buffers. */
+ for (i = START_SIZE; i <= MIN_PAGE_SIZE; i <<= 1)
+ {
+ /* Test length alignments from 0-16 bytes. */
+ for (int j = 0; j < 8; j++)
+ {
+ do_test (&json_ctx, i + j, false);
+ do_test (&json_ctx, i + 16 - j, false);
+ }
+ }
+
+ /* Overlapping buffers. */
+ for (i = START_SIZE; i <= MIN_PAGE_SIZE; i <<= 1)
+ {
+ /* Test length alignments from 0-16 bytes. */
+ for (int j = 0; j < 8; j++)
+ {
+ do_test (&json_ctx, i + j, true);
+ do_test (&json_ctx, i + 16 - j, true);
+ }
+ }
+
+ 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>
--
2.7.4
^ permalink raw reply [flat|nested] 3+ messages in thread
* [COMMITTED 1/3] benchtests: Memory walking benchmark for memcpy
@ 2017-10-05 16:55 Siddhesh Poyarekar
2017-10-05 16:55 ` [COMMITTED 2/3] benchtests: Memory walking benchmark for memset Siddhesh Poyarekar
2017-10-05 16:55 ` [COMMITTED 3/3] benchtests: Memory walking benchmark for memmove Siddhesh Poyarekar
0 siblings, 2 replies; 3+ messages in thread
From: Siddhesh Poyarekar @ 2017-10-05 16:55 UTC (permalink / raw)
To: libc-alpha
This benchmark is an attempt to eliminate cache effects from string
benchmarks. The benchmark walks both ways through a large memory area
and copies different sizes of memory and alignments one at a time
instead of looping around in the same memory area. This is a good
metric to have alongside the other memcpy benchmarks, especially for
larger sizes where the likelihood of the call being done only once is
pretty high.
* benchtests/bench-memcpy-walk.c: New file.
* benchtests/Makefile (string-benchset): Add it.
---
ChangeLog | 5 ++
benchtests/Makefile | 3 +-
benchtests/bench-memcpy-walk.c | 127 +++++++++++++++++++++++++++++++++++++++++
3 files changed, 134 insertions(+), 1 deletion(-)
create mode 100644 benchtests/bench-memcpy-walk.c
diff --git a/ChangeLog b/ChangeLog
index 7ddff74..a86faeb 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,8 @@
+2017-10-05 Siddhesh Poyarekar <siddhesh@sourceware.org>
+
+ * benchtests/bench-memcpy-walk.c: New file.
+ * benchtests/Makefile (string-benchset): Add it.
+
2017-10-05 Florian Weimer <fweimer@redhat.com>
nscd: Eliminate compilation time dependency in the build output.
diff --git a/benchtests/Makefile b/benchtests/Makefile
index 3acc39c..d086cc6 100644
--- a/benchtests/Makefile
+++ b/benchtests/Makefile
@@ -37,7 +37,8 @@ string-benchset := bcopy bzero memccpy memchr memcmp memcpy memmem memmove \
strcat strchr strchrnul strcmp strcpy strcspn strlen \
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
+ strcoll memcpy-large memcpy-random memmove-large memset-large \
+ memcpy-walk
# Build and run locale-dependent benchmarks only if we're building natively.
ifeq (no,$(cross-compiling))
diff --git a/benchtests/bench-memcpy-walk.c b/benchtests/bench-memcpy-walk.c
new file mode 100644
index 0000000..69d467d
--- /dev/null
+++ b/benchtests/bench-memcpy-walk.c
@@ -0,0 +1,127 @@
+/* Measure memcpy function combined throughput for different alignments.
+ Copyright (C) 2017 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
+ <http://www.gnu.org/licenses/>. */
+
+/* This microbenchmark measures the throughput of memcpy for various sizes from
+ 1 byte to 32MiB, doubling every iteration and then misaligning by 0-15
+ bytes. The copies are done from source to destination and then back and the
+ source walks forward across the array and the destination walks backward by
+ one byte each, thus measuring misaligned accesses as well. The idea is to
+ avoid caching effects by copying a different string and far enough from each
+ other, walking in different directions so that we can measure prefetcher
+ efficiency (software or hardware) more closely than with a loop copying the
+ same data over and over, which eventually only gives us L1 cache
+ performance. */
+
+#ifndef MEMCPY_RESULT
+# define MEMCPY_RESULT(dst, len) dst
+# define START_SIZE 1
+# define MIN_PAGE_SIZE (getpagesize () + 32 * 1024 * 1024)
+# define TEST_MAIN
+# define TEST_NAME "memcpy"
+# define TIMEOUT (20 * 60)
+# include "bench-string.h"
+
+IMPL (memcpy, 1)
+#endif
+
+#include "json-lib.h"
+
+typedef char *(*proto_t) (char *, const char *, size_t);
+
+static void
+do_one_test (json_ctx_t *json_ctx, impl_t *impl, char *dst, char *src,
+ size_t len)
+{
+ size_t i, iters = MIN_PAGE_SIZE / len;
+ timing_t start, stop, cur;
+
+ char *dst_end = dst + MIN_PAGE_SIZE - len;
+ char *src_end = src + MIN_PAGE_SIZE - len;
+
+ TIMING_NOW (start);
+ /* Copy the entire buffer back and forth, LEN at a time. */
+ for (i = 0; i < iters && dst_end >= dst && src <= src_end; src++, dst_end--)
+ {
+ CALL (impl, dst_end, src, len);
+ CALL (impl, src, dst_end, len);
+ i += 2;
+ }
+ TIMING_NOW (stop);
+
+ TIMING_DIFF (cur, start, stop);
+
+ /* Get time taken per function call. */
+ json_element_double (json_ctx, (double) cur * len / i);
+}
+
+static void
+do_test (json_ctx_t *json_ctx, size_t len)
+{
+ json_element_object_begin (json_ctx);
+ json_attr_uint (json_ctx, "length", (double) len);
+ json_array_begin (json_ctx, "timings");
+
+ FOR_EACH_IMPL (impl, 0)
+ do_one_test (json_ctx, impl, (char *) buf2, (char *) buf1, len);
+
+ json_array_end (json_ctx);
+ json_element_object_end (json_ctx);
+}
+
+int
+test_main (void)
+{
+ json_ctx_t json_ctx;
+ size_t i;
+
+ 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, "memcpy");
+ json_attr_string (&json_ctx, "bench-variant", "walk");
+
+ 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");
+ for (i = START_SIZE; i <= MIN_PAGE_SIZE; i <<= 1)
+ {
+ /* Test length alignments from 0-16 bytes. */
+ for (int j = 0; j < 8; j++)
+ {
+ do_test (&json_ctx, i + j);
+ do_test (&json_ctx, i + 16 - j);
+ }
+ }
+
+ 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>
--
2.7.4
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2017-10-05 16:55 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-10-05 16:55 [COMMITTED 1/3] benchtests: Memory walking benchmark for memcpy Siddhesh Poyarekar
2017-10-05 16:55 ` [COMMITTED 2/3] benchtests: Memory walking benchmark for memset Siddhesh Poyarekar
2017-10-05 16:55 ` [COMMITTED 3/3] benchtests: Memory walking benchmark for memmove Siddhesh Poyarekar
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).