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 #include <ctype.h>
   9
  10 #include "../perf.h"
  11 #include "../util/util.h"
  12 #include "../util/parse-options.h"
  13 #include "../util/header.h"
  14 #include "bench.h"
  15 #include "mem-memcpy-arch.h"
  16
  17 #include <stdio.h>
  18 #include <stdlib.h>
  19 #include <string.h>
  20 #include <sys/time.h>
  21 #include <errno.h>
  22
  23 #define K 1024
  24
  25 static const char       *length_str     = "1MB";
  26 static const char       *routine        = "default";
  27 static int              iterations      = 1;
  28 static bool             use_clock;
  29 static int              clock_fd;
  30 static bool             only_prefault;
  31 static bool             no_prefault;
  32
  33 static const struct option options[] = {
  34         OPT_STRING('l', "length", &length_str, "1MB",
  35                     "Specify length of memory to copy. "
  36                     "available unit: B, MB, GB (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', "clock", &use_clock,
  42                     "Use CPU clock 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()"),
  47         OPT_END()
  48 };
  49
  50 typedef void *(*memcpy_t)(void *, const void *, size_t);
  51
  52 struct routine {
  53         const char *name;
  54         const char *desc;
  55         memcpy_t fn;
  56 };
  57
  58 struct routine routines[] = {
  59         { "default",
  60           "Default memcpy() provided by glibc",
  61           memcpy },
  62 #ifdef ARCH_X86_64
  63
  64 #define MEMCPY_FN(fn, name, desc) { name, desc, fn },
  65 #include "mem-memcpy-x86-64-asm-def.h"
  66 #undef MEMCPY_FN
  67
  68 #endif
  69
  70         { NULL,
  71           NULL,
  72           NULL   }
  73 };
  74
  75 static const char * const bench_mem_memcpy_usage[] = {
  76         "perf bench mem memcpy <options>",
  77         NULL
  78 };
  79
  80 static struct perf_event_attr clock_attr = {
  81         .type           = PERF_TYPE_HARDWARE,
  82         .config         = PERF_COUNT_HW_CPU_CYCLES
  83 };
  84
  85 static void init_clock(void)
  86 {
  87         clock_fd = sys_perf_event_open(&clock_attr, getpid(), -1, -1, 0);
  88
  89         if (clock_fd < 0 && errno == ENOSYS)
  90                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  91         else
  92                 BUG_ON(clock_fd < 0);
  93 }
  94
  95 static u64 get_clock(void)
  96 {
  97         int ret;
  98         u64 clk;
  99
 100         ret = read(clock_fd, &clk, sizeof(u64));
 101         BUG_ON(ret != sizeof(u64));
 102
 103         return clk;
 104 }
 105
 106 static double timeval2double(struct timeval *ts)
 107 {
 108         return (double)ts->tv_sec +
 109                 (double)ts->tv_usec / (double)1000000;
 110 }
 111
 112 static void alloc_mem(void **dst, void **src, size_t length)
 113 {
 114         *dst = zalloc(length);
 115         if (!dst)
 116                 die("memory allocation failed - maybe length is too large?\n");
 117
 118         *src = zalloc(length);
 119         if (!src)
 120                 die("memory allocation failed - maybe length is too large?\n");
 121 }
 122
 123 static u64 do_memcpy_clock(memcpy_t fn, size_t len, bool prefault)
 124 {
 125         u64 clock_start = 0ULL, clock_end = 0ULL;
 126         void *src = NULL, *dst = NULL;
 127         int i;
 128
 129         alloc_mem(&src, &dst, len);
 130
 131         if (prefault)
 132                 fn(dst, src, len);
 133
 134         clock_start = get_clock();
 135         for (i = 0; i < iterations; ++i)
 136                 fn(dst, src, len);
 137         clock_end = get_clock();
 138
 139         free(src);
 140         free(dst);
 141         return clock_end - clock_start;
 142 }
 143
 144 static double do_memcpy_gettimeofday(memcpy_t fn, size_t len, bool prefault)
 145 {
 146         struct timeval tv_start, tv_end, tv_diff;
 147         void *src = NULL, *dst = NULL;
 148         int i;
 149
 150         alloc_mem(&src, &dst, len);
 151
 152         if (prefault)
 153                 fn(dst, src, len);
 154
 155         BUG_ON(gettimeofday(&tv_start, NULL));
 156         for (i = 0; i < iterations; ++i)
 157                 fn(dst, src, len);
 158         BUG_ON(gettimeofday(&tv_end, NULL));
 159
 160         timersub(&tv_end, &tv_start, &tv_diff);
 161
 162         free(src);
 163         free(dst);
 164         return (double)((double)len / timeval2double(&tv_diff));
 165 }
 166
 167 #define pf (no_prefault ? 0 : 1)
 168
 169 #define print_bps(x) do {                                       \
 170                 if (x < K)                                      \
 171                         printf(" %14lf B/Sec", x);              \
 172                 else if (x < K * K)                             \
 173                         printf(" %14lfd KB/Sec", x / K);        \
 174                 else if (x < K * K * K)                         \
 175                         printf(" %14lf MB/Sec", x / K / K);     \
 176                 else                                            \
 177                         printf(" %14lf GB/Sec", x / K / K / K); \
 178         } while (0)
 179
 180 int bench_mem_memcpy(int argc, const char **argv,
 181                      const char *prefix __used)
 182 {
 183         int i;
 184         size_t len;
 185         double result_bps[2];
 186         u64 result_clock[2];
 187
 188         argc = parse_options(argc, argv, options,
 189                              bench_mem_memcpy_usage, 0);
 190
 191         if (use_clock)
 192                 init_clock();
 193
 194         len = (size_t)perf_atoll((char *)length_str);
 195
 196         result_clock[0] = result_clock[1] = 0ULL;
 197         result_bps[0] = result_bps[1] = 0.0;
 198
 199         if ((s64)len <= 0) {
 200                 fprintf(stderr, "Invalid length:%s\n", length_str);
 201                 return 1;
 202         }
 203
 204         /* same to without specifying either of prefault and no-prefault */
 205         if (only_prefault && no_prefault)
 206                 only_prefault = no_prefault = false;
 207
 208         for (i = 0; routines[i].name; i++) {
 209                 if (!strcmp(routines[i].name, routine))
 210                         break;
 211         }
 212         if (!routines[i].name) {
 213                 printf("Unknown routine:%s\n", routine);
 214                 printf("Available routines...\n");
 215                 for (i = 0; routines[i].name; i++) {
 216                         printf("\t%s ... %s\n",
 217                                routines[i].name, routines[i].desc);
 218                 }
 219                 return 1;
 220         }
 221
 222         if (bench_format == BENCH_FORMAT_DEFAULT)
 223                 printf("# Copying %s Bytes ...\n\n", length_str);
 224
 225         if (!only_prefault && !no_prefault) {
 226                 /* show both of results */
 227                 if (use_clock) {
 228                         result_clock[0] =
 229                                 do_memcpy_clock(routines[i].fn, len, false);
 230                         result_clock[1] =
 231                                 do_memcpy_clock(routines[i].fn, len, true);
 232                 } else {
 233                         result_bps[0] =
 234                                 do_memcpy_gettimeofday(routines[i].fn,
 235                                                 len, false);
 236                         result_bps[1] =
 237                                 do_memcpy_gettimeofday(routines[i].fn,
 238                                                 len, true);
 239                 }
 240         } else {
 241                 if (use_clock) {
 242                         result_clock[pf] =
 243                                 do_memcpy_clock(routines[i].fn,
 244                                                 len, only_prefault);
 245                 } else {
 246                         result_bps[pf] =
 247                                 do_memcpy_gettimeofday(routines[i].fn,
 248                                                 len, only_prefault);
 249                 }
 250         }
 251
 252         switch (bench_format) {
 253         case BENCH_FORMAT_DEFAULT:
 254                 if (!only_prefault && !no_prefault) {
 255                         if (use_clock) {
 256                                 printf(" %14lf Clock/Byte\n",
 257                                         (double)result_clock[0]
 258                                         / (double)len);
 259                                 printf(" %14lf Clock/Byte (with prefault)\n",
 260                                         (double)result_clock[1]
 261                                         / (double)len);
 262                         } else {
 263                                 print_bps(result_bps[0]);
 264                                 printf("\n");
 265                                 print_bps(result_bps[1]);
 266                                 printf(" (with prefault)\n");
 267                         }
 268                 } else {
 269                         if (use_clock) {
 270                                 printf(" %14lf Clock/Byte",
 271                                         (double)result_clock[pf]
 272                                         / (double)len);
 273                         } else
 274                                 print_bps(result_bps[pf]);
 275
 276                         printf("%s\n", only_prefault ? " (with prefault)" : "");
 277                 }
 278                 break;
 279         case BENCH_FORMAT_SIMPLE:
 280                 if (!only_prefault && !no_prefault) {
 281                         if (use_clock) {
 282                                 printf("%lf %lf\n",
 283                                         (double)result_clock[0] / (double)len,
 284                                         (double)result_clock[1] / (double)len);
 285                         } else {
 286                                 printf("%lf %lf\n",
 287                                         result_bps[0], result_bps[1]);
 288                         }
 289                 } else {
 290                         if (use_clock) {
 291                                 printf("%lf\n", (double)result_clock[pf]
 292                                         / (double)len);
 293                         } else
 294                                 printf("%lf\n", result_bps[pf]);
 295                 }
 296                 break;
 297         default:
 298                 /* reaching this means there's some disaster: */
 299                 die("unknown format: %d\n", bench_format);
 300                 break;
 301         }
 302
 303         return 0;
 304 }