4 * memcpy: Simple memory copy in various ways
6 * Written by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
10 #include "../util/util.h"
11 #include "../util/parse-options.h"
12 #include "../util/header.h"
13 #include "../util/cloexec.h"
15 #include "mem-memcpy-arch.h"
16 #include "mem-memset-arch.h"
26 static const char *length_str
= "1MB";
27 static const char *routine
= "default";
28 static int iterations
= 1;
29 static bool use_cycle
;
31 static bool only_prefault
;
32 static bool no_prefault
;
34 static const struct option options
[] = {
35 OPT_STRING('l', "length", &length_str
, "1MB",
36 "Specify length of memory to copy. "
37 "Available units: B, KB, MB, GB and TB (upper and lower)"),
38 OPT_STRING('r', "routine", &routine
, "default",
39 "Specify routine to copy"),
40 OPT_INTEGER('i', "iterations", &iterations
,
41 "repeat memcpy() invocation this number of times"),
42 OPT_BOOLEAN('c', "cycle", &use_cycle
,
43 "Use cycles event instead of gettimeofday() for measuring"),
44 OPT_BOOLEAN('o', "only-prefault", &only_prefault
,
45 "Show only the result with page faults before memcpy()"),
46 OPT_BOOLEAN('n', "no-prefault", &no_prefault
,
47 "Show only the result without page faults before memcpy()"),
51 typedef void *(*memcpy_t
)(void *, const void *, size_t);
52 typedef void *(*memset_t
)(void *, int, size_t);
63 struct routine memcpy_routines
[] = {
65 .desc
= "Default memcpy() provided by glibc",
66 .fn
.memcpy
= memcpy
},
67 #ifdef HAVE_ARCH_X86_64_SUPPORT
69 #define MEMCPY_FN(_fn, _name, _desc) {.name = _name, .desc = _desc, .fn.memcpy = _fn},
70 #include "mem-memcpy-x86-64-asm-def.h"
80 static const char * const bench_mem_memcpy_usage
[] = {
81 "perf bench mem memcpy <options>",
85 static struct perf_event_attr cycle_attr
= {
86 .type
= PERF_TYPE_HARDWARE
,
87 .config
= PERF_COUNT_HW_CPU_CYCLES
90 static void init_cycle(void)
92 cycle_fd
= sys_perf_event_open(&cycle_attr
, getpid(), -1, -1,
93 perf_event_open_cloexec_flag());
95 if (cycle_fd
< 0 && errno
== ENOSYS
)
96 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
101 static u64
get_cycle(void)
106 ret
= read(cycle_fd
, &clk
, sizeof(u64
));
107 BUG_ON(ret
!= sizeof(u64
));
112 static double timeval2double(struct timeval
*ts
)
114 return (double)ts
->tv_sec
+
115 (double)ts
->tv_usec
/ (double)1000000;
118 #define pf (no_prefault ? 0 : 1)
120 #define print_bps(x) do { \
122 printf(" %14lf B/Sec", x); \
123 else if (x < K * K) \
124 printf(" %14lfd KB/Sec", x / K); \
125 else if (x < K * K * K) \
126 printf(" %14lf MB/Sec", x / K / K); \
128 printf(" %14lf GB/Sec", x / K / K / K); \
131 struct bench_mem_info
{
132 const struct routine
*routines
;
133 u64 (*do_cycle
)(const struct routine
*r
, size_t len
, bool prefault
);
134 double (*do_gettimeofday
)(const struct routine
*r
, size_t len
, bool prefault
);
135 const char *const *usage
;
138 static int bench_mem_common(int argc
, const char **argv
,
139 const char *prefix __maybe_unused
,
140 struct bench_mem_info
*info
)
145 double result_bps
[2];
148 argc
= parse_options(argc
, argv
, options
,
151 if (no_prefault
&& only_prefault
) {
152 fprintf(stderr
, "Invalid options: -o and -n are mutually exclusive\n");
159 len
= (size_t)perf_atoll((char *)length_str
);
160 totallen
= (double)len
* iterations
;
162 result_cycle
[0] = result_cycle
[1] = 0ULL;
163 result_bps
[0] = result_bps
[1] = 0.0;
166 fprintf(stderr
, "Invalid length:%s\n", length_str
);
170 /* same to without specifying either of prefault and no-prefault */
171 if (only_prefault
&& no_prefault
)
172 only_prefault
= no_prefault
= false;
174 for (i
= 0; info
->routines
[i
].name
; i
++) {
175 if (!strcmp(info
->routines
[i
].name
, routine
))
178 if (!info
->routines
[i
].name
) {
179 printf("Unknown routine:%s\n", routine
);
180 printf("Available routines...\n");
181 for (i
= 0; info
->routines
[i
].name
; i
++) {
182 printf("\t%s ... %s\n",
183 info
->routines
[i
].name
, info
->routines
[i
].desc
);
188 if (bench_format
== BENCH_FORMAT_DEFAULT
)
189 printf("# Copying %s Bytes ...\n\n", length_str
);
191 if (!only_prefault
&& !no_prefault
) {
192 /* show both of results */
195 info
->do_cycle(&info
->routines
[i
], len
, false);
197 info
->do_cycle(&info
->routines
[i
], len
, true);
200 info
->do_gettimeofday(&info
->routines
[i
],
203 info
->do_gettimeofday(&info
->routines
[i
],
209 info
->do_cycle(&info
->routines
[i
],
213 info
->do_gettimeofday(&info
->routines
[i
],
218 switch (bench_format
) {
219 case BENCH_FORMAT_DEFAULT
:
220 if (!only_prefault
&& !no_prefault
) {
222 printf(" %14lf Cycle/Byte\n",
223 (double)result_cycle
[0]
225 printf(" %14lf Cycle/Byte (with prefault)\n",
226 (double)result_cycle
[1]
229 print_bps(result_bps
[0]);
231 print_bps(result_bps
[1]);
232 printf(" (with prefault)\n");
236 printf(" %14lf Cycle/Byte",
237 (double)result_cycle
[pf
]
240 print_bps(result_bps
[pf
]);
242 printf("%s\n", only_prefault
? " (with prefault)" : "");
245 case BENCH_FORMAT_SIMPLE
:
246 if (!only_prefault
&& !no_prefault
) {
249 (double)result_cycle
[0] / totallen
,
250 (double)result_cycle
[1] / totallen
);
253 result_bps
[0], result_bps
[1]);
257 printf("%lf\n", (double)result_cycle
[pf
]
260 printf("%lf\n", result_bps
[pf
]);
264 /* reaching this means there's some disaster: */
265 die("unknown format: %d\n", bench_format
);
272 static void memcpy_alloc_mem(void **dst
, void **src
, size_t length
)
274 *dst
= zalloc(length
);
276 die("memory allocation failed - maybe length is too large?\n");
278 *src
= zalloc(length
);
280 die("memory allocation failed - maybe length is too large?\n");
281 /* Make sure to always replace the zero pages even if MMAP_THRESH is crossed */
282 memset(*src
, 0, length
);
285 static u64
do_memcpy_cycle(const struct routine
*r
, size_t len
, bool prefault
)
287 u64 cycle_start
= 0ULL, cycle_end
= 0ULL;
288 void *src
= NULL
, *dst
= NULL
;
289 memcpy_t fn
= r
->fn
.memcpy
;
292 memcpy_alloc_mem(&dst
, &src
, len
);
297 cycle_start
= get_cycle();
298 for (i
= 0; i
< iterations
; ++i
)
300 cycle_end
= get_cycle();
304 return cycle_end
- cycle_start
;
307 static double do_memcpy_gettimeofday(const struct routine
*r
, size_t len
,
310 struct timeval tv_start
, tv_end
, tv_diff
;
311 memcpy_t fn
= r
->fn
.memcpy
;
312 void *src
= NULL
, *dst
= NULL
;
315 memcpy_alloc_mem(&dst
, &src
, len
);
320 BUG_ON(gettimeofday(&tv_start
, NULL
));
321 for (i
= 0; i
< iterations
; ++i
)
323 BUG_ON(gettimeofday(&tv_end
, NULL
));
325 timersub(&tv_end
, &tv_start
, &tv_diff
);
329 return (double)(((double)len
* iterations
) / timeval2double(&tv_diff
));
332 int bench_mem_memcpy(int argc
, const char **argv
,
333 const char *prefix __maybe_unused
)
335 struct bench_mem_info info
= {
336 .routines
= memcpy_routines
,
337 .do_cycle
= do_memcpy_cycle
,
338 .do_gettimeofday
= do_memcpy_gettimeofday
,
339 .usage
= bench_mem_memcpy_usage
,
342 return bench_mem_common(argc
, argv
, prefix
, &info
);
345 static void memset_alloc_mem(void **dst
, size_t length
)
347 *dst
= zalloc(length
);
349 die("memory allocation failed - maybe length is too large?\n");
352 static u64
do_memset_cycle(const struct routine
*r
, size_t len
, bool prefault
)
354 u64 cycle_start
= 0ULL, cycle_end
= 0ULL;
355 memset_t fn
= r
->fn
.memset
;
359 memset_alloc_mem(&dst
, len
);
364 cycle_start
= get_cycle();
365 for (i
= 0; i
< iterations
; ++i
)
367 cycle_end
= get_cycle();
370 return cycle_end
- cycle_start
;
373 static double do_memset_gettimeofday(const struct routine
*r
, size_t len
,
376 struct timeval tv_start
, tv_end
, tv_diff
;
377 memset_t fn
= r
->fn
.memset
;
381 memset_alloc_mem(&dst
, len
);
386 BUG_ON(gettimeofday(&tv_start
, NULL
));
387 for (i
= 0; i
< iterations
; ++i
)
389 BUG_ON(gettimeofday(&tv_end
, NULL
));
391 timersub(&tv_end
, &tv_start
, &tv_diff
);
394 return (double)(((double)len
* iterations
) / timeval2double(&tv_diff
));
397 static const char * const bench_mem_memset_usage
[] = {
398 "perf bench mem memset <options>",
402 static const struct routine memset_routines
[] = {
404 .desc
= "Default memset() provided by glibc",
405 .fn
.memset
= memset
},
406 #ifdef HAVE_ARCH_X86_64_SUPPORT
408 #define MEMSET_FN(_fn, _name, _desc) { .name = _name, .desc = _desc, .fn.memset = _fn },
409 #include "mem-memset-x86-64-asm-def.h"
419 int bench_mem_memset(int argc
, const char **argv
,
420 const char *prefix __maybe_unused
)
422 struct bench_mem_info info
= {
423 .routines
= memset_routines
,
424 .do_cycle
= do_memset_cycle
,
425 .do_gettimeofday
= do_memset_gettimeofday
,
426 .usage
= bench_mem_memset_usage
,
429 return bench_mem_common(argc
, argv
, prefix
, &info
);