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"
25 static const char *length_str
= "1MB";
26 static const char *routine
= "default";
27 static int iterations
= 1;
28 static bool use_cycle
;
30 static bool only_prefault
;
31 static bool no_prefault
;
33 static const struct option options
[] = {
34 OPT_STRING('l', "length", &length_str
, "1MB",
35 "Specify length of memory to copy. "
36 "Available units: B, KB, MB, GB and TB (upper and lower)"),
37 OPT_STRING('r', "routine", &routine
, "default",
38 "Specify routine to copy"),
39 OPT_INTEGER('i', "iterations", &iterations
,
40 "repeat memcpy() invocation this number of times"),
41 OPT_BOOLEAN('c', "cycle", &use_cycle
,
42 "Use cycles event instead of gettimeofday() for measuring"),
43 OPT_BOOLEAN('o', "only-prefault", &only_prefault
,
44 "Show only the result with page faults before memcpy()"),
45 OPT_BOOLEAN('n', "no-prefault", &no_prefault
,
46 "Show only the result without page faults before memcpy()"),
50 typedef void *(*memcpy_t
)(void *, const void *, size_t);
58 struct routine routines
[] = {
60 "Default memcpy() provided by glibc",
62 #ifdef HAVE_ARCH_X86_64_SUPPORT
64 #define MEMCPY_FN(fn, name, desc) { name, desc, fn },
65 #include "mem-memcpy-x86-64-asm-def.h"
75 static const char * const bench_mem_memcpy_usage
[] = {
76 "perf bench mem memcpy <options>",
80 static struct perf_event_attr cycle_attr
= {
81 .type
= PERF_TYPE_HARDWARE
,
82 .config
= PERF_COUNT_HW_CPU_CYCLES
85 static void init_cycle(void)
87 cycle_fd
= sys_perf_event_open(&cycle_attr
, getpid(), -1, -1,
88 perf_event_open_cloexec_flag());
90 if (cycle_fd
< 0 && errno
== ENOSYS
)
91 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
96 static u64
get_cycle(void)
101 ret
= read(cycle_fd
, &clk
, sizeof(u64
));
102 BUG_ON(ret
!= sizeof(u64
));
107 static double timeval2double(struct timeval
*ts
)
109 return (double)ts
->tv_sec
+
110 (double)ts
->tv_usec
/ (double)1000000;
113 static void alloc_mem(void **dst
, void **src
, size_t length
)
115 *dst
= zalloc(length
);
117 die("memory allocation failed - maybe length is too large?\n");
119 *src
= zalloc(length
);
121 die("memory allocation failed - maybe length is too large?\n");
122 /* Make sure to always replace the zero pages even if MMAP_THRESH is crossed */
123 memset(*src
, 0, length
);
126 static u64
do_memcpy_cycle(memcpy_t fn
, size_t len
, bool prefault
)
128 u64 cycle_start
= 0ULL, cycle_end
= 0ULL;
129 void *src
= NULL
, *dst
= NULL
;
132 alloc_mem(&src
, &dst
, len
);
137 cycle_start
= get_cycle();
138 for (i
= 0; i
< iterations
; ++i
)
140 cycle_end
= get_cycle();
144 return cycle_end
- cycle_start
;
147 static double do_memcpy_gettimeofday(memcpy_t fn
, size_t len
, bool prefault
)
149 struct timeval tv_start
, tv_end
, tv_diff
;
150 void *src
= NULL
, *dst
= NULL
;
153 alloc_mem(&src
, &dst
, len
);
158 BUG_ON(gettimeofday(&tv_start
, NULL
));
159 for (i
= 0; i
< iterations
; ++i
)
161 BUG_ON(gettimeofday(&tv_end
, NULL
));
163 timersub(&tv_end
, &tv_start
, &tv_diff
);
167 return (double)((double)len
/ timeval2double(&tv_diff
));
170 #define pf (no_prefault ? 0 : 1)
172 #define print_bps(x) do { \
174 printf(" %14lf B/Sec", x); \
175 else if (x < K * K) \
176 printf(" %14lfd KB/Sec", x / K); \
177 else if (x < K * K * K) \
178 printf(" %14lf MB/Sec", x / K / K); \
180 printf(" %14lf GB/Sec", x / K / K / K); \
183 int bench_mem_memcpy(int argc
, const char **argv
,
184 const char *prefix __maybe_unused
)
188 double result_bps
[2];
191 argc
= parse_options(argc
, argv
, options
,
192 bench_mem_memcpy_usage
, 0);
194 if (no_prefault
&& only_prefault
) {
195 fprintf(stderr
, "Invalid options: -o and -n are mutually exclusive\n");
202 len
= (size_t)perf_atoll((char *)length_str
);
204 result_cycle
[0] = result_cycle
[1] = 0ULL;
205 result_bps
[0] = result_bps
[1] = 0.0;
208 fprintf(stderr
, "Invalid length:%s\n", length_str
);
212 /* same to without specifying either of prefault and no-prefault */
213 if (only_prefault
&& no_prefault
)
214 only_prefault
= no_prefault
= false;
216 for (i
= 0; routines
[i
].name
; i
++) {
217 if (!strcmp(routines
[i
].name
, routine
))
220 if (!routines
[i
].name
) {
221 printf("Unknown routine:%s\n", routine
);
222 printf("Available routines...\n");
223 for (i
= 0; routines
[i
].name
; i
++) {
224 printf("\t%s ... %s\n",
225 routines
[i
].name
, routines
[i
].desc
);
230 if (bench_format
== BENCH_FORMAT_DEFAULT
)
231 printf("# Copying %s Bytes ...\n\n", length_str
);
233 if (!only_prefault
&& !no_prefault
) {
234 /* show both of results */
237 do_memcpy_cycle(routines
[i
].fn
, len
, false);
239 do_memcpy_cycle(routines
[i
].fn
, len
, true);
242 do_memcpy_gettimeofday(routines
[i
].fn
,
245 do_memcpy_gettimeofday(routines
[i
].fn
,
251 do_memcpy_cycle(routines
[i
].fn
,
255 do_memcpy_gettimeofday(routines
[i
].fn
,
260 switch (bench_format
) {
261 case BENCH_FORMAT_DEFAULT
:
262 if (!only_prefault
&& !no_prefault
) {
264 printf(" %14lf Cycle/Byte\n",
265 (double)result_cycle
[0]
267 printf(" %14lf Cycle/Byte (with prefault)\n",
268 (double)result_cycle
[1]
271 print_bps(result_bps
[0]);
273 print_bps(result_bps
[1]);
274 printf(" (with prefault)\n");
278 printf(" %14lf Cycle/Byte",
279 (double)result_cycle
[pf
]
282 print_bps(result_bps
[pf
]);
284 printf("%s\n", only_prefault
? " (with prefault)" : "");
287 case BENCH_FORMAT_SIMPLE
:
288 if (!only_prefault
&& !no_prefault
) {
291 (double)result_cycle
[0] / (double)len
,
292 (double)result_cycle
[1] / (double)len
);
295 result_bps
[0], result_bps
[1]);
299 printf("%lf\n", (double)result_cycle
[pf
]
302 printf("%lf\n", result_bps
[pf
]);
306 /* reaching this means there's some disaster: */
307 die("unknown format: %d\n", bench_format
);