4 * memcpy: Simple memory copy in various ways
6 * Written by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
11 #include "../util/util.h"
12 #include "../util/parse-options.h"
13 #include "../util/header.h"
15 #include "mem-memcpy-arch.h"
25 static const char *length_str = "1MB";
26 static const char *routine = "default";
27 static bool use_clock;
29 static bool only_prefault;
30 static bool no_prefault;
32 static const struct option options[] = {
33 OPT_STRING('l', "length", &length_str, "1MB",
34 "Specify length of memory to copy. "
35 "available unit: B, MB, GB (upper and lower)"),
36 OPT_STRING('r', "routine", &routine, "default",
37 "Specify routine to copy"),
38 OPT_BOOLEAN('c', "clock", &use_clock,
39 "Use CPU clock for measuring"),
40 OPT_BOOLEAN('o', "only-prefault", &only_prefault,
41 "Show only the result with page faults before memcpy()"),
42 OPT_BOOLEAN('n', "no-prefault", &no_prefault,
43 "Show only the result without page faults before memcpy()"),
47 typedef void *(*memcpy_t)(void *, const void *, size_t);
55 struct routine routines[] = {
57 "Default memcpy() provided by glibc",
61 #define MEMCPY_FN(fn, name, desc) { name, desc, fn },
62 #include "mem-memcpy-x86-64-asm-def.h"
72 static const char * const bench_mem_memcpy_usage[] = {
73 "perf bench mem memcpy <options>",
77 static struct perf_event_attr clock_attr = {
78 .type = PERF_TYPE_HARDWARE,
79 .config = PERF_COUNT_HW_CPU_CYCLES
82 static void init_clock(void)
84 clock_fd = sys_perf_event_open(&clock_attr, getpid(), -1, -1, 0);
86 if (clock_fd < 0 && errno == ENOSYS)
87 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
92 static u64 get_clock(void)
97 ret = read(clock_fd, &clk, sizeof(u64));
98 BUG_ON(ret != sizeof(u64));
103 static double timeval2double(struct timeval *ts)
105 return (double)ts->tv_sec +
106 (double)ts->tv_usec / (double)1000000;
109 static void alloc_mem(void **dst, void **src, size_t length)
111 *dst = zalloc(length);
113 die("memory allocation failed - maybe length is too large?\n");
115 *src = zalloc(length);
117 die("memory allocation failed - maybe length is too large?\n");
120 static u64 do_memcpy_clock(memcpy_t fn, size_t len, bool prefault)
122 u64 clock_start = 0ULL, clock_end = 0ULL;
123 void *src = NULL, *dst = NULL;
125 alloc_mem(&src, &dst, len);
130 clock_start = get_clock();
132 clock_end = get_clock();
136 return clock_end - clock_start;
139 static double do_memcpy_gettimeofday(memcpy_t fn, size_t len, bool prefault)
141 struct timeval tv_start, tv_end, tv_diff;
142 void *src = NULL, *dst = NULL;
144 alloc_mem(&src, &dst, len);
149 BUG_ON(gettimeofday(&tv_start, NULL));
151 BUG_ON(gettimeofday(&tv_end, NULL));
153 timersub(&tv_end, &tv_start, &tv_diff);
157 return (double)((double)len / timeval2double(&tv_diff));
160 #define pf (no_prefault ? 0 : 1)
162 #define print_bps(x) do { \
164 printf(" %14lf B/Sec", x); \
165 else if (x < K * K) \
166 printf(" %14lfd KB/Sec", x / K); \
167 else if (x < K * K * K) \
168 printf(" %14lf MB/Sec", x / K / K); \
170 printf(" %14lf GB/Sec", x / K / K / K); \
173 int bench_mem_memcpy(int argc, const char **argv,
174 const char *prefix __used)
178 double result_bps[2];
181 argc = parse_options(argc, argv, options,
182 bench_mem_memcpy_usage, 0);
187 len = (size_t)perf_atoll((char *)length_str);
189 result_clock[0] = result_clock[1] = 0ULL;
190 result_bps[0] = result_bps[1] = 0.0;
193 fprintf(stderr, "Invalid length:%s\n", length_str);
197 /* same to without specifying either of prefault and no-prefault */
198 if (only_prefault && no_prefault)
199 only_prefault = no_prefault = false;
201 for (i = 0; routines[i].name; i++) {
202 if (!strcmp(routines[i].name, routine))
205 if (!routines[i].name) {
206 printf("Unknown routine:%s\n", routine);
207 printf("Available routines...\n");
208 for (i = 0; routines[i].name; i++) {
209 printf("\t%s ... %s\n",
210 routines[i].name, routines[i].desc);
215 if (bench_format == BENCH_FORMAT_DEFAULT)
216 printf("# Copying %s Bytes ...\n\n", length_str);
218 if (!only_prefault && !no_prefault) {
219 /* show both of results */
222 do_memcpy_clock(routines[i].fn, len, false);
224 do_memcpy_clock(routines[i].fn, len, true);
227 do_memcpy_gettimeofday(routines[i].fn,
230 do_memcpy_gettimeofday(routines[i].fn,
236 do_memcpy_clock(routines[i].fn,
240 do_memcpy_gettimeofday(routines[i].fn,
245 switch (bench_format) {
246 case BENCH_FORMAT_DEFAULT:
247 if (!only_prefault && !no_prefault) {
249 printf(" %14lf Clock/Byte\n",
250 (double)result_clock[0]
252 printf(" %14lf Clock/Byte (with prefault)\n",
253 (double)result_clock[1]
256 print_bps(result_bps[0]);
258 print_bps(result_bps[1]);
259 printf(" (with prefault)\n");
263 printf(" %14lf Clock/Byte",
264 (double)result_clock[pf]
267 print_bps(result_bps[pf]);
269 printf("%s\n", only_prefault ? " (with prefault)" : "");
272 case BENCH_FORMAT_SIMPLE:
273 if (!only_prefault && !no_prefault) {
276 (double)result_clock[0] / (double)len,
277 (double)result_clock[1] / (double)len);
280 result_bps[0], result_bps[1]);
284 printf("%lf\n", (double)result_clock[pf]
287 printf("%lf\n", result_bps[pf]);
291 /* reaching this means there's some disaster: */
292 die("unknown format: %d\n", bench_format);