tools/perf/bench/mem-memset.c

   1 /*
   2  * mem-memset.c
   3  *
   4  * memset: Simple memory set in various ways
   5  *
   6  * Trivial clone of mem-memcpy.c.
   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-memset-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_cycle;
  28 static int              cycle_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 set. "
  35                     "Available units: B, KB, MB, GB and TB (upper and lower)"),
  36         OPT_STRING('r', "routine", &routine, "default",
  37                     "Specify routine to set"),
  38         OPT_INTEGER('i', "iterations", &iterations,
  39                     "repeat memset() invocation this number of times"),
  40         OPT_BOOLEAN('c', "cycle", &use_cycle,
  41                     "Use cycles event instead of gettimeofday() for measuring"),
  42         OPT_BOOLEAN('o', "only-prefault", &only_prefault,
  43                     "Show only the result with page faults before memset()"),
  44         OPT_BOOLEAN('n', "no-prefault", &no_prefault,
  45                     "Show only the result without page faults before memset()"),
  46         OPT_END()
  47 };
  48
  49 typedef void *(*memset_t)(void *, int, size_t);
  50
  51 struct routine {
  52         const char *name;
  53         const char *desc;
  54         memset_t fn;
  55 };
  56
  57 static const struct routine routines[] = {
  58         { "default",
  59           "Default memset() provided by glibc",
  60           memset },
  61 #ifdef HAVE_ARCH_X86_64_SUPPORT
  62
  63 #define MEMSET_FN(fn, name, desc) { name, desc, fn },
  64 #include "mem-memset-x86-64-asm-def.h"
  65 #undef MEMSET_FN
  66
  67 #endif
  68
  69         { NULL,
  70           NULL,
  71           NULL   }
  72 };
  73
  74 static const char * const bench_mem_memset_usage[] = {
  75         "perf bench mem memset <options>",
  76         NULL
  77 };
  78
  79 static struct perf_event_attr cycle_attr = {
  80         .type           = PERF_TYPE_HARDWARE,
  81         .config         = PERF_COUNT_HW_CPU_CYCLES
  82 };
  83
  84 static void init_cycle(void)
  85 {
  86         cycle_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1, 0);
  87
  88         if (cycle_fd < 0 && errno == ENOSYS)
  89                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  90         else
  91                 BUG_ON(cycle_fd < 0);
  92 }
  93
  94 static u64 get_cycle(void)
  95 {
  96         int ret;
  97         u64 clk;
  98
  99         ret = read(cycle_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, size_t length)
 112 {
 113         *dst = zalloc(length);
 114         if (!*dst)
 115                 die("memory allocation failed - maybe length is too large?\n");
 116 }
 117
 118 static u64 do_memset_cycle(memset_t fn, size_t len, bool prefault)
 119 {
 120         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 121         void *dst = NULL;
 122         int i;
 123
 124         alloc_mem(&dst, len);
 125
 126         if (prefault)
 127                 fn(dst, -1, len);
 128
 129         cycle_start = get_cycle();
 130         for (i = 0; i < iterations; ++i)
 131                 fn(dst, i, len);
 132         cycle_end = get_cycle();
 133
 134         free(dst);
 135         return cycle_end - cycle_start;
 136 }
 137
 138 static double do_memset_gettimeofday(memset_t fn, size_t len, bool prefault)
 139 {
 140         struct timeval tv_start, tv_end, tv_diff;
 141         void *dst = NULL;
 142         int i;
 143
 144         alloc_mem(&dst, len);
 145
 146         if (prefault)
 147                 fn(dst, -1, len);
 148
 149         BUG_ON(gettimeofday(&tv_start, NULL));
 150         for (i = 0; i < iterations; ++i)
 151                 fn(dst, i, len);
 152         BUG_ON(gettimeofday(&tv_end, NULL));
 153
 154         timersub(&tv_end, &tv_start, &tv_diff);
 155
 156         free(dst);
 157         return (double)((double)len / timeval2double(&tv_diff));
 158 }
 159
 160 #define pf (no_prefault ? 0 : 1)
 161
 162 #define print_bps(x) do {                                       \
 163                 if (x < K)                                      \
 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);     \
 169                 else                                            \
 170                         printf(" %14lf GB/Sec", x / K / K / K); \
 171         } while (0)
 172
 173 int bench_mem_memset(int argc, const char **argv,
 174                      const char *prefix __maybe_unused)
 175 {
 176         int i;
 177         size_t len;
 178         double result_bps[2];
 179         u64 result_cycle[2];
 180
 181         argc = parse_options(argc, argv, options,
 182                              bench_mem_memset_usage, 0);
 183
 184         if (no_prefault && only_prefault) {
 185                 fprintf(stderr, "Invalid options: -o and -n are mutually exclusive\n");
 186                 return 1;
 187         }
 188
 189         if (use_cycle)
 190                 init_cycle();
 191
 192         len = (size_t)perf_atoll((char *)length_str);
 193
 194         result_cycle[0] = result_cycle[1] = 0ULL;
 195         result_bps[0] = result_bps[1] = 0.0;
 196
 197         if ((s64)len <= 0) {
 198                 fprintf(stderr, "Invalid length:%s\n", length_str);
 199                 return 1;
 200         }
 201
 202         /* same to without specifying either of prefault and no-prefault */
 203         if (only_prefault && no_prefault)
 204                 only_prefault = no_prefault = false;
 205
 206         for (i = 0; routines[i].name; i++) {
 207                 if (!strcmp(routines[i].name, routine))
 208                         break;
 209         }
 210         if (!routines[i].name) {
 211                 printf("Unknown routine:%s\n", routine);
 212                 printf("Available routines...\n");
 213                 for (i = 0; routines[i].name; i++) {
 214                         printf("\t%s ... %s\n",
 215                                routines[i].name, routines[i].desc);
 216                 }
 217                 return 1;
 218         }
 219
 220         if (bench_format == BENCH_FORMAT_DEFAULT)
 221                 printf("# Copying %s Bytes ...\n\n", length_str);
 222
 223         if (!only_prefault && !no_prefault) {
 224                 /* show both of results */
 225                 if (use_cycle) {
 226                         result_cycle[0] =
 227                                 do_memset_cycle(routines[i].fn, len, false);
 228                         result_cycle[1] =
 229                                 do_memset_cycle(routines[i].fn, len, true);
 230                 } else {
 231                         result_bps[0] =
 232                                 do_memset_gettimeofday(routines[i].fn,
 233                                                 len, false);
 234                         result_bps[1] =
 235                                 do_memset_gettimeofday(routines[i].fn,
 236                                                 len, true);
 237                 }
 238         } else {
 239                 if (use_cycle) {
 240                         result_cycle[pf] =
 241                                 do_memset_cycle(routines[i].fn,
 242                                                 len, only_prefault);
 243                 } else {
 244                         result_bps[pf] =
 245                                 do_memset_gettimeofday(routines[i].fn,
 246                                                 len, only_prefault);
 247                 }
 248         }
 249
 250         switch (bench_format) {
 251         case BENCH_FORMAT_DEFAULT:
 252                 if (!only_prefault && !no_prefault) {
 253                         if (use_cycle) {
 254                                 printf(" %14lf Cycle/Byte\n",
 255                                         (double)result_cycle[0]
 256                                         / (double)len);
 257                                 printf(" %14lf Cycle/Byte (with prefault)\n ",
 258                                         (double)result_cycle[1]
 259                                         / (double)len);
 260                         } else {
 261                                 print_bps(result_bps[0]);
 262                                 printf("\n");
 263                                 print_bps(result_bps[1]);
 264                                 printf(" (with prefault)\n");
 265                         }
 266                 } else {
 267                         if (use_cycle) {
 268                                 printf(" %14lf Cycle/Byte",
 269                                         (double)result_cycle[pf]
 270                                         / (double)len);
 271                         } else
 272                                 print_bps(result_bps[pf]);
 273
 274                         printf("%s\n", only_prefault ? " (with prefault)" : "");
 275                 }
 276                 break;
 277         case BENCH_FORMAT_SIMPLE:
 278                 if (!only_prefault && !no_prefault) {
 279                         if (use_cycle) {
 280                                 printf("%lf %lf\n",
 281                                         (double)result_cycle[0] / (double)len,
 282                                         (double)result_cycle[1] / (double)len);
 283                         } else {
 284                                 printf("%lf %lf\n",
 285                                         result_bps[0], result_bps[1]);
 286                         }
 287                 } else {
 288                         if (use_cycle) {
 289                                 printf("%lf\n", (double)result_cycle[pf]
 290                                         / (double)len);
 291                         } else
 292                                 printf("%lf\n", result_bps[pf]);
 293                 }
 294                 break;
 295         default:
 296                 /* reaching this means there's some disaster: */
 297                 die("unknown format: %d\n", bench_format);
 298                 break;
 299         }
 300
 301         return 0;
 302 }