tools/perf/bench/mem-functions.c

   1 /*
   2  * mem-memcpy.c
   3  *
   4  * Simple memcpy() and memset() benchmarks
   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 "../util/cloexec.h"
  14 #include "bench.h"
  15 #include "mem-memcpy-arch.h"
  16 #include "mem-memset-arch.h"
  17
  18 #include <stdio.h>
  19 #include <stdlib.h>
  20 #include <string.h>
  21 #include <sys/time.h>
  22 #include <errno.h>
  23
  24 #define K 1024
  25
  26 static const char       *size_str       = "1MB";
  27 static const char       *routine_str    = "all";
  28 static int              iterations      = 1;
  29 static bool             use_cycles;
  30 static int              cycles_fd;
  31
  32 static const struct option options[] = {
  33         OPT_STRING('l', "size", &size_str, "1MB",
  34                     "Specify the size of the memory buffers. "
  35                     "Available units: B, KB, MB, GB and TB (case insensitive)"),
  36
  37         OPT_STRING('r', "routine", &routine_str, "all",
  38                     "Specify the routine to run, \"all\" runs all available routines, \"help\" lists them"),
  39
  40         OPT_INTEGER('i', "iterations", &iterations,
  41                     "Repeat the function this number of times"),
  42
  43         OPT_BOOLEAN('c', "cycles", &use_cycles,
  44                     "Use a cycles event instead of gettimeofday() to measure performance"),
  45
  46         OPT_END()
  47 };
  48
  49 typedef void *(*memcpy_t)(void *, const void *, size_t);
  50 typedef void *(*memset_t)(void *, int, size_t);
  51
  52 struct routine {
  53         const char *name;
  54         const char *desc;
  55         union {
  56                 memcpy_t memcpy;
  57                 memset_t memset;
  58         } fn;
  59 };
  60
  61 struct routine memcpy_routines[] = {
  62         { .name         = "default",
  63           .desc         = "Default memcpy() provided by glibc",
  64           .fn.memcpy    = memcpy },
  65
  66 #ifdef HAVE_ARCH_X86_64_SUPPORT
  67 # define MEMCPY_FN(_fn, _name, _desc) {.name = _name, .desc = _desc, .fn.memcpy = _fn},
  68 # include "mem-memcpy-x86-64-asm-def.h"
  69 # undef MEMCPY_FN
  70 #endif
  71
  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 cycle_attr = {
  81         .type           = PERF_TYPE_HARDWARE,
  82         .config         = PERF_COUNT_HW_CPU_CYCLES
  83 };
  84
  85 static void init_cycles(void)
  86 {
  87         cycles_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1, perf_event_open_cloexec_flag());
  88
  89         if (cycles_fd < 0 && errno == ENOSYS)
  90                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  91         else
  92                 BUG_ON(cycles_fd < 0);
  93 }
  94
  95 static u64 get_cycles(void)
  96 {
  97         int ret;
  98         u64 clk;
  99
 100         ret = read(cycles_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 + (double)ts->tv_usec / (double)1000000;
 109 }
 110
 111 #define print_bps(x) do {                                               \
 112                 if (x < K)                                              \
 113                         printf(" %14lf bytes/sec\n", x);                \
 114                 else if (x < K * K)                                     \
 115                         printf(" %14lfd KB/sec\n", x / K);              \
 116                 else if (x < K * K * K)                                 \
 117                         printf(" %14lf MB/sec\n", x / K / K);           \
 118                 else                                                    \
 119                         printf(" %14lf GB/sec\n", x / K / K / K);       \
 120         } while (0)
 121
 122 struct bench_mem_info {
 123         const struct routine *routines;
 124         u64 (*do_cycles)(const struct routine *r, size_t size);
 125         double (*do_gettimeofday)(const struct routine *r, size_t size);
 126         const char *const *usage;
 127 };
 128
 129 static void __bench_mem_routine(struct bench_mem_info *info, int r_idx, size_t size, double size_total)
 130 {
 131         const struct routine *r = &info->routines[r_idx];
 132         double result_bps = 0.0;
 133         u64 result_cycles = 0;
 134
 135         printf("# Routine '%s' (%s)\n", r->name, r->desc);
 136
 137         if (bench_format == BENCH_FORMAT_DEFAULT)
 138                 printf("# Copying %s bytes ...\n\n", size_str);
 139
 140         if (use_cycles) {
 141                 result_cycles = info->do_cycles(r, size);
 142         } else {
 143                 result_bps = info->do_gettimeofday(r, size);
 144         }
 145
 146         switch (bench_format) {
 147         case BENCH_FORMAT_DEFAULT:
 148                 if (use_cycles) {
 149                         printf(" %14lf cycles/byte\n", (double)result_cycles/size_total);
 150                 } else {
 151                         print_bps(result_bps);
 152                 }
 153                 break;
 154
 155         case BENCH_FORMAT_SIMPLE:
 156                 if (use_cycles) {
 157                         printf("%lf\n", (double)result_cycles/size_total);
 158                 } else {
 159                         printf("%lf\n", result_bps);
 160                 }
 161                 break;
 162
 163         default:
 164                 BUG_ON(1);
 165                 break;
 166         }
 167 }
 168
 169 static int bench_mem_common(int argc, const char **argv, struct bench_mem_info *info)
 170 {
 171         int i;
 172         size_t size;
 173         double size_total;
 174
 175         argc = parse_options(argc, argv, options, info->usage, 0);
 176
 177         if (use_cycles)
 178                 init_cycles();
 179
 180         size = (size_t)perf_atoll((char *)size_str);
 181         size_total = (double)size * iterations;
 182
 183         if ((s64)size <= 0) {
 184                 fprintf(stderr, "Invalid size:%s\n", size_str);
 185                 return 1;
 186         }
 187
 188         if (!strncmp(routine_str, "all", 3)) {
 189                 for (i = 0; info->routines[i].name; i++)
 190                         __bench_mem_routine(info, i, size, size_total);
 191                 return 0;
 192         }
 193
 194         for (i = 0; info->routines[i].name; i++) {
 195                 if (!strcmp(info->routines[i].name, routine_str))
 196                         break;
 197         }
 198         if (!info->routines[i].name) {
 199                 if (strcmp(routine_str, "help") && strcmp(routine_str, "h"))
 200                         printf("Unknown routine: %s\n", routine_str);
 201                 printf("Available routines:\n");
 202                 for (i = 0; info->routines[i].name; i++) {
 203                         printf("\t%s ... %s\n",
 204                                info->routines[i].name, info->routines[i].desc);
 205                 }
 206                 return 1;
 207         }
 208
 209         __bench_mem_routine(info, i, size, size_total);
 210
 211         return 0;
 212 }
 213
 214 static void memcpy_alloc_mem(void **dst, void **src, size_t size)
 215 {
 216         *dst = zalloc(size);
 217         if (!*dst)
 218                 die("memory allocation failed - maybe size is too large?\n");
 219
 220         *src = zalloc(size);
 221         if (!*src)
 222                 die("memory allocation failed - maybe size is too large?\n");
 223
 224         /* Make sure to always prefault zero pages even if MMAP_THRESH is crossed: */
 225         memset(*src, 0, size);
 226 }
 227
 228 static u64 do_memcpy_cycles(const struct routine *r, size_t size)
 229 {
 230         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 231         void *src = NULL, *dst = NULL;
 232         memcpy_t fn = r->fn.memcpy;
 233         int i;
 234
 235         memcpy_alloc_mem(&dst, &src, size);
 236
 237         /*
 238          * We prefault the freshly allocated memory range here,
 239          * to not measure page fault overhead:
 240          */
 241         fn(dst, src, size);
 242
 243         cycle_start = get_cycles();
 244         for (i = 0; i < iterations; ++i)
 245                 fn(dst, src, size);
 246         cycle_end = get_cycles();
 247
 248         free(src);
 249         free(dst);
 250         return cycle_end - cycle_start;
 251 }
 252
 253 static double do_memcpy_gettimeofday(const struct routine *r, size_t size)
 254 {
 255         struct timeval tv_start, tv_end, tv_diff;
 256         memcpy_t fn = r->fn.memcpy;
 257         void *src = NULL, *dst = NULL;
 258         int i;
 259
 260         memcpy_alloc_mem(&dst, &src, size);
 261
 262         /*
 263          * We prefault the freshly allocated memory range here,
 264          * to not measure page fault overhead:
 265          */
 266         fn(dst, src, size);
 267
 268         BUG_ON(gettimeofday(&tv_start, NULL));
 269         for (i = 0; i < iterations; ++i)
 270                 fn(dst, src, size);
 271         BUG_ON(gettimeofday(&tv_end, NULL));
 272
 273         timersub(&tv_end, &tv_start, &tv_diff);
 274
 275         free(src);
 276         free(dst);
 277
 278         return (double)(((double)size * iterations) / timeval2double(&tv_diff));
 279 }
 280
 281 int bench_mem_memcpy(int argc, const char **argv, const char *prefix __maybe_unused)
 282 {
 283         struct bench_mem_info info = {
 284                 .routines               = memcpy_routines,
 285                 .do_cycles              = do_memcpy_cycles,
 286                 .do_gettimeofday        = do_memcpy_gettimeofday,
 287                 .usage                  = bench_mem_memcpy_usage,
 288         };
 289
 290         return bench_mem_common(argc, argv, &info);
 291 }
 292
 293 static void memset_alloc_mem(void **dst, size_t size)
 294 {
 295         *dst = zalloc(size);
 296         if (!*dst)
 297                 die("memory allocation failed - maybe size is too large?\n");
 298 }
 299
 300 static u64 do_memset_cycles(const struct routine *r, size_t size)
 301 {
 302         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 303         memset_t fn = r->fn.memset;
 304         void *dst = NULL;
 305         int i;
 306
 307         memset_alloc_mem(&dst, size);
 308
 309         /*
 310          * We prefault the freshly allocated memory range here,
 311          * to not measure page fault overhead:
 312          */
 313         fn(dst, -1, size);
 314
 315         cycle_start = get_cycles();
 316         for (i = 0; i < iterations; ++i)
 317                 fn(dst, i, size);
 318         cycle_end = get_cycles();
 319
 320         free(dst);
 321         return cycle_end - cycle_start;
 322 }
 323
 324 static double do_memset_gettimeofday(const struct routine *r, size_t size)
 325 {
 326         struct timeval tv_start, tv_end, tv_diff;
 327         memset_t fn = r->fn.memset;
 328         void *dst = NULL;
 329         int i;
 330
 331         memset_alloc_mem(&dst, size);
 332
 333         /*
 334          * We prefault the freshly allocated memory range here,
 335          * to not measure page fault overhead:
 336          */
 337         fn(dst, -1, size);
 338
 339         BUG_ON(gettimeofday(&tv_start, NULL));
 340         for (i = 0; i < iterations; ++i)
 341                 fn(dst, i, size);
 342         BUG_ON(gettimeofday(&tv_end, NULL));
 343
 344         timersub(&tv_end, &tv_start, &tv_diff);
 345
 346         free(dst);
 347         return (double)(((double)size * iterations) / timeval2double(&tv_diff));
 348 }
 349
 350 static const char * const bench_mem_memset_usage[] = {
 351         "perf bench mem memset <options>",
 352         NULL
 353 };
 354
 355 static const struct routine memset_routines[] = {
 356         { .name         = "default",
 357           .desc         = "Default memset() provided by glibc",
 358           .fn.memset    = memset },
 359
 360 #ifdef HAVE_ARCH_X86_64_SUPPORT
 361 # define MEMSET_FN(_fn, _name, _desc) { .name = _name, .desc = _desc, .fn.memset = _fn },
 362 # include "mem-memset-x86-64-asm-def.h"
 363 # undef MEMSET_FN
 364 #endif
 365
 366         { NULL, }
 367 };
 368
 369 int bench_mem_memset(int argc, const char **argv, const char *prefix __maybe_unused)
 370 {
 371         struct bench_mem_info info = {
 372                 .routines               = memset_routines,
 373                 .do_cycles              = do_memset_cycles,
 374                 .do_gettimeofday        = do_memset_gettimeofday,
 375                 .usage                  = bench_mem_memset_usage,
 376         };
 377
 378         return bench_mem_common(argc, argv, &info);
 379 }