tools/perf/bench/mem-memcpy.c

   1 /*
   2  * mem-memcpy.c
   3  *
   4  * memcpy: Simple memory copy in various ways
   5  *
   6  * Written by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
   7  */
   8
   9 #include "../perf.h"
  10 #include "../util/util.h"
  11 #include "../util/parse-options.h"
  12 #include "../util/header.h"
  13 #include "bench.h"
  14 #include "mem-memcpy-arch.h"
  15
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <string.h>
  19 #include <sys/time.h>
  20 #include <errno.h>
  21
  22 #define K 1024
  23
  24 static const char       *length_str     = "1MB";
  25 static const char       *routine        = "default";
  26 static int              iterations      = 1;
  27 static bool             use_clock;
  28 static int              clock_fd;
  29 static bool             only_prefault;
  30 static bool             no_prefault;
  31
  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_INTEGER('i', "iterations", &iterations,
  39                     "repeat memcpy() invocation this number of times"),
  40         OPT_BOOLEAN('c', "clock", &use_clock,
  41                     "Use CPU clock for measuring"),
  42         OPT_BOOLEAN('o', "only-prefault", &only_prefault,
  43                     "Show only the result with page faults before memcpy()"),
  44         OPT_BOOLEAN('n', "no-prefault", &no_prefault,
  45                     "Show only the result without page faults before memcpy()"),
  46         OPT_END()
  47 };
  48
  49 typedef void *(*memcpy_t)(void *, const void *, size_t);
  50
  51 struct routine {
  52         const char *name;
  53         const char *desc;
  54         memcpy_t fn;
  55 };
  56
  57 struct routine routines[] = {
  58         { "default",
  59           "Default memcpy() provided by glibc",
  60           memcpy },
  61 #ifdef ARCH_X86_64
  62
  63 #define MEMCPY_FN(fn, name, desc) { name, desc, fn },
  64 #include "mem-memcpy-x86-64-asm-def.h"
  65 #undef MEMCPY_FN
  66
  67 #endif
  68
  69         { NULL,
  70           NULL,
  71           NULL   }
  72 };
  73
  74 static const char * const bench_mem_memcpy_usage[] = {
  75         "perf bench mem memcpy <options>",
  76         NULL
  77 };
  78
  79 static struct perf_event_attr clock_attr = {
  80         .type           = PERF_TYPE_HARDWARE,
  81         .config         = PERF_COUNT_HW_CPU_CYCLES
  82 };
  83
  84 static void init_clock(void)
  85 {
  86         clock_fd = sys_perf_event_open(&clock_attr, getpid(), -1, -1, 0);
  87
  88         if (clock_fd < 0 && errno == ENOSYS)
  89                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  90         else
  91                 BUG_ON(clock_fd < 0);
  92 }
  93
  94 static u64 get_clock(void)
  95 {
  96         int ret;
  97         u64 clk;
  98
  99         ret = read(clock_fd, &clk, sizeof(u64));
 100         BUG_ON(ret != sizeof(u64));
 101
 102         return clk;
 103 }
 104
 105 static double timeval2double(struct timeval *ts)
 106 {
 107         return (double)ts->tv_sec +
 108                 (double)ts->tv_usec / (double)1000000;
 109 }
 110
 111 static void alloc_mem(void **dst, void **src, size_t length)
 112 {
 113         *dst = zalloc(length);
 114         if (!dst)
 115                 die("memory allocation failed - maybe length is too large?\n");
 116
 117         *src = zalloc(length);
 118         if (!src)
 119                 die("memory allocation failed - maybe length is too large?\n");
 120 }
 121
 122 static u64 do_memcpy_clock(memcpy_t fn, size_t len, bool prefault)
 123 {
 124         u64 clock_start = 0ULL, clock_end = 0ULL;
 125         void *src = NULL, *dst = NULL;
 126         int i;
 127
 128         alloc_mem(&src, &dst, len);
 129
 130         if (prefault)
 131                 fn(dst, src, len);
 132
 133         clock_start = get_clock();
 134         for (i = 0; i < iterations; ++i)
 135                 fn(dst, src, len);
 136         clock_end = get_clock();
 137
 138         free(src);
 139         free(dst);
 140         return clock_end - clock_start;
 141 }
 142
 143 static double do_memcpy_gettimeofday(memcpy_t fn, size_t len, bool prefault)
 144 {
 145         struct timeval tv_start, tv_end, tv_diff;
 146         void *src = NULL, *dst = NULL;
 147         int i;
 148
 149         alloc_mem(&src, &dst, len);
 150
 151         if (prefault)
 152                 fn(dst, src, len);
 153
 154         BUG_ON(gettimeofday(&tv_start, NULL));
 155         for (i = 0; i < iterations; ++i)
 156                 fn(dst, src, len);
 157         BUG_ON(gettimeofday(&tv_end, NULL));
 158
 159         timersub(&tv_end, &tv_start, &tv_diff);
 160
 161         free(src);
 162         free(dst);
 163         return (double)((double)len / timeval2double(&tv_diff));
 164 }
 165
 166 #define pf (no_prefault ? 0 : 1)
 167
 168 #define print_bps(x) do {                                       \
 169                 if (x < K)                                      \
 170                         printf(" %14lf B/Sec", x);              \
 171                 else if (x < K * K)                             \
 172                         printf(" %14lfd KB/Sec", x / K);        \
 173                 else if (x < K * K * K)                         \
 174                         printf(" %14lf MB/Sec", x / K / K);     \
 175                 else                                            \
 176                         printf(" %14lf GB/Sec", x / K / K / K); \
 177         } while (0)
 178
 179 int bench_mem_memcpy(int argc, const char **argv,
 180                      const char *prefix __used)
 181 {
 182         int i;
 183         size_t len;
 184         double result_bps[2];
 185         u64 result_clock[2];
 186
 187         argc = parse_options(argc, argv, options,
 188                              bench_mem_memcpy_usage, 0);
 189
 190         if (use_clock)
 191                 init_clock();
 192
 193         len = (size_t)perf_atoll((char *)length_str);
 194
 195         result_clock[0] = result_clock[1] = 0ULL;
 196         result_bps[0] = result_bps[1] = 0.0;
 197
 198         if ((s64)len <= 0) {
 199                 fprintf(stderr, "Invalid length:%s\n", length_str);
 200                 return 1;
 201         }
 202
 203         /* same to without specifying either of prefault and no-prefault */
 204         if (only_prefault && no_prefault)
 205                 only_prefault = no_prefault = false;
 206
 207         for (i = 0; routines[i].name; i++) {
 208                 if (!strcmp(routines[i].name, routine))
 209                         break;
 210         }
 211         if (!routines[i].name) {
 212                 printf("Unknown routine:%s\n", routine);
 213                 printf("Available routines...\n");
 214                 for (i = 0; routines[i].name; i++) {
 215                         printf("\t%s ... %s\n",
 216                                routines[i].name, routines[i].desc);
 217                 }
 218                 return 1;
 219         }
 220
 221         if (bench_format == BENCH_FORMAT_DEFAULT)
 222                 printf("# Copying %s Bytes ...\n\n", length_str);
 223
 224         if (!only_prefault && !no_prefault) {
 225                 /* show both of results */
 226                 if (use_clock) {
 227                         result_clock[0] =
 228                                 do_memcpy_clock(routines[i].fn, len, false);
 229                         result_clock[1] =
 230                                 do_memcpy_clock(routines[i].fn, len, true);
 231                 } else {
 232                         result_bps[0] =
 233                                 do_memcpy_gettimeofday(routines[i].fn,
 234                                                 len, false);
 235                         result_bps[1] =
 236                                 do_memcpy_gettimeofday(routines[i].fn,
 237                                                 len, true);
 238                 }
 239         } else {
 240                 if (use_clock) {
 241                         result_clock[pf] =
 242                                 do_memcpy_clock(routines[i].fn,
 243                                                 len, only_prefault);
 244                 } else {
 245                         result_bps[pf] =
 246                                 do_memcpy_gettimeofday(routines[i].fn,
 247                                                 len, only_prefault);
 248                 }
 249         }
 250
 251         switch (bench_format) {
 252         case BENCH_FORMAT_DEFAULT:
 253                 if (!only_prefault && !no_prefault) {
 254                         if (use_clock) {
 255                                 printf(" %14lf Clock/Byte\n",
 256                                         (double)result_clock[0]
 257                                         / (double)len);
 258                                 printf(" %14lf Clock/Byte (with prefault)\n",
 259                                         (double)result_clock[1]
 260                                         / (double)len);
 261                         } else {
 262                                 print_bps(result_bps[0]);
 263                                 printf("\n");
 264                                 print_bps(result_bps[1]);
 265                                 printf(" (with prefault)\n");
 266                         }
 267                 } else {
 268                         if (use_clock) {
 269                                 printf(" %14lf Clock/Byte",
 270                                         (double)result_clock[pf]
 271                                         / (double)len);
 272                         } else
 273                                 print_bps(result_bps[pf]);
 274
 275                         printf("%s\n", only_prefault ? " (with prefault)" : "");
 276                 }
 277                 break;
 278         case BENCH_FORMAT_SIMPLE:
 279                 if (!only_prefault && !no_prefault) {
 280                         if (use_clock) {
 281                                 printf("%lf %lf\n",
 282                                         (double)result_clock[0] / (double)len,
 283                                         (double)result_clock[1] / (double)len);
 284                         } else {
 285                                 printf("%lf %lf\n",
 286                                         result_bps[0], result_bps[1]);
 287                         }
 288                 } else {
 289                         if (use_clock) {
 290                                 printf("%lf\n", (double)result_clock[pf]
 291                                         / (double)len);
 292                         } else
 293                                 printf("%lf\n", result_bps[pf]);
 294                 }
 295                 break;
 296         default:
 297                 /* reaching this means there's some disaster: */
 298                 die("unknown format: %d\n", bench_format);
 299                 break;
 300         }
 301
 302         return 0;
 303 }