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 (upper and lower)"),
  36         OPT_STRING('r', "routine", &routine_str, "all",
  37                     "Specify the routine to run, \"all\" runs all available routines"),
  38         OPT_INTEGER('i', "iterations", &iterations,
  39                     "repeat memcpy() invocation this number of times"),
  40         OPT_BOOLEAN('c', "cycles", &use_cycles,
  41                     "Use a cycles event instead of gettimeofday() to measure performance"),
  42         OPT_END()
  43 };
  44
  45 typedef void *(*memcpy_t)(void *, const void *, size_t);
  46 typedef void *(*memset_t)(void *, int, size_t);
  47
  48 struct routine {
  49         const char *name;
  50         const char *desc;
  51         union {
  52                 memcpy_t memcpy;
  53                 memset_t memset;
  54         } fn;
  55 };
  56
  57 struct routine memcpy_routines[] = {
  58         { .name         = "default",
  59           .desc         = "Default memcpy() provided by glibc",
  60           .fn.memcpy    = memcpy },
  61
  62 #ifdef HAVE_ARCH_X86_64_SUPPORT
  63 # define MEMCPY_FN(_fn, _name, _desc) {.name = _name, .desc = _desc, .fn.memcpy = _fn},
  64 # include "mem-memcpy-x86-64-asm-def.h"
  65 # undef MEMCPY_FN
  66 #endif
  67
  68         { NULL, }
  69 };
  70
  71 static const char * const bench_mem_memcpy_usage[] = {
  72         "perf bench mem memcpy <options>",
  73         NULL
  74 };
  75
  76 static struct perf_event_attr cycle_attr = {
  77         .type           = PERF_TYPE_HARDWARE,
  78         .config         = PERF_COUNT_HW_CPU_CYCLES
  79 };
  80
  81 static void init_cycles(void)
  82 {
  83         cycles_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1, perf_event_open_cloexec_flag());
  84
  85         if (cycles_fd < 0 && errno == ENOSYS)
  86                 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
  87         else
  88                 BUG_ON(cycles_fd < 0);
  89 }
  90
  91 static u64 get_cycles(void)
  92 {
  93         int ret;
  94         u64 clk;
  95
  96         ret = read(cycles_fd, &clk, sizeof(u64));
  97         BUG_ON(ret != sizeof(u64));
  98
  99         return clk;
 100 }
 101
 102 static double timeval2double(struct timeval *ts)
 103 {
 104         return (double)ts->tv_sec + (double)ts->tv_usec / (double)1000000;
 105 }
 106
 107 #define print_bps(x) do {                                               \
 108                 if (x < K)                                              \
 109                         printf(" %14lf B/Sec\n", x);                    \
 110                 else if (x < K * K)                                     \
 111                         printf(" %14lfd KB/Sec\n", x / K);              \
 112                 else if (x < K * K * K)                                 \
 113                         printf(" %14lf MB/Sec\n", x / K / K);           \
 114                 else                                                    \
 115                         printf(" %14lf GB/Sec\n", x / K / K / K);       \
 116         } while (0)
 117
 118 struct bench_mem_info {
 119         const struct routine *routines;
 120         u64 (*do_cycles)(const struct routine *r, size_t size);
 121         double (*do_gettimeofday)(const struct routine *r, size_t size);
 122         const char *const *usage;
 123 };
 124
 125 static void __bench_mem_routine(struct bench_mem_info *info, int r_idx, size_t size, double size_total)
 126 {
 127         const struct routine *r = &info->routines[r_idx];
 128         double result_bps = 0.0;
 129         u64 result_cycles = 0;
 130
 131         printf("Routine %s (%s)\n", r->name, r->desc);
 132
 133         if (bench_format == BENCH_FORMAT_DEFAULT)
 134                 printf("# Copying %s Bytes ...\n\n", size_str);
 135
 136         if (use_cycles) {
 137                 result_cycles = info->do_cycles(r, size);
 138         } else {
 139                 result_bps = info->do_gettimeofday(r, size);
 140         }
 141
 142         switch (bench_format) {
 143         case BENCH_FORMAT_DEFAULT:
 144                 if (use_cycles) {
 145                         printf(" %14lf cycles/Byte\n", (double)result_cycles/size_total);
 146                 } else {
 147                         print_bps(result_bps);
 148                 }
 149                 break;
 150
 151         case BENCH_FORMAT_SIMPLE:
 152                 if (use_cycles) {
 153                         printf("%lf\n", (double)result_cycles/size_total);
 154                 } else {
 155                         printf("%lf\n", result_bps);
 156                 }
 157                 break;
 158
 159         default:
 160                 BUG_ON(1);
 161                 break;
 162         }
 163 }
 164
 165 static int bench_mem_common(int argc, const char **argv, struct bench_mem_info *info)
 166 {
 167         int i;
 168         size_t size;
 169         double size_total;
 170
 171         argc = parse_options(argc, argv, options, info->usage, 0);
 172
 173         if (use_cycles)
 174                 init_cycles();
 175
 176         size = (size_t)perf_atoll((char *)size_str);
 177         size_total = (double)size * iterations;
 178
 179         if ((s64)size <= 0) {
 180                 fprintf(stderr, "Invalid size:%s\n", size_str);
 181                 return 1;
 182         }
 183
 184         if (!strncmp(routine_str, "all", 3)) {
 185                 for (i = 0; info->routines[i].name; i++)
 186                         __bench_mem_routine(info, i, size, size_total);
 187                 return 0;
 188         }
 189
 190         for (i = 0; info->routines[i].name; i++) {
 191                 if (!strcmp(info->routines[i].name, routine_str))
 192                         break;
 193         }
 194         if (!info->routines[i].name) {
 195                 printf("Unknown routine: %s\n", routine_str);
 196                 printf("Available routines...\n");
 197                 for (i = 0; info->routines[i].name; i++) {
 198                         printf("\t%s ... %s\n",
 199                                info->routines[i].name, info->routines[i].desc);
 200                 }
 201                 return 1;
 202         }
 203
 204         __bench_mem_routine(info, i, size, size_total);
 205
 206         return 0;
 207 }
 208
 209 static void memcpy_alloc_mem(void **dst, void **src, size_t size)
 210 {
 211         *dst = zalloc(size);
 212         if (!*dst)
 213                 die("memory allocation failed - maybe size is too large?\n");
 214
 215         *src = zalloc(size);
 216         if (!*src)
 217                 die("memory allocation failed - maybe size is too large?\n");
 218
 219         /* Make sure to always prefault zero pages even if MMAP_THRESH is crossed: */
 220         memset(*src, 0, size);
 221 }
 222
 223 static u64 do_memcpy_cycles(const struct routine *r, size_t size)
 224 {
 225         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 226         void *src = NULL, *dst = NULL;
 227         memcpy_t fn = r->fn.memcpy;
 228         int i;
 229
 230         memcpy_alloc_mem(&dst, &src, size);
 231
 232         /*
 233          * We prefault the freshly allocated memory range here,
 234          * to not measure page fault overhead:
 235          */
 236         fn(dst, src, size);
 237
 238         cycle_start = get_cycles();
 239         for (i = 0; i < iterations; ++i)
 240                 fn(dst, src, size);
 241         cycle_end = get_cycles();
 242
 243         free(src);
 244         free(dst);
 245         return cycle_end - cycle_start;
 246 }
 247
 248 static double do_memcpy_gettimeofday(const struct routine *r, size_t size)
 249 {
 250         struct timeval tv_start, tv_end, tv_diff;
 251         memcpy_t fn = r->fn.memcpy;
 252         void *src = NULL, *dst = NULL;
 253         int i;
 254
 255         memcpy_alloc_mem(&dst, &src, size);
 256
 257         /*
 258          * We prefault the freshly allocated memory range here,
 259          * to not measure page fault overhead:
 260          */
 261         fn(dst, src, size);
 262
 263         BUG_ON(gettimeofday(&tv_start, NULL));
 264         for (i = 0; i < iterations; ++i)
 265                 fn(dst, src, size);
 266         BUG_ON(gettimeofday(&tv_end, NULL));
 267
 268         timersub(&tv_end, &tv_start, &tv_diff);
 269
 270         free(src);
 271         free(dst);
 272
 273         return (double)(((double)size * iterations) / timeval2double(&tv_diff));
 274 }
 275
 276 int bench_mem_memcpy(int argc, const char **argv, const char *prefix __maybe_unused)
 277 {
 278         struct bench_mem_info info = {
 279                 .routines               = memcpy_routines,
 280                 .do_cycles              = do_memcpy_cycles,
 281                 .do_gettimeofday        = do_memcpy_gettimeofday,
 282                 .usage                  = bench_mem_memcpy_usage,
 283         };
 284
 285         return bench_mem_common(argc, argv, &info);
 286 }
 287
 288 static void memset_alloc_mem(void **dst, size_t size)
 289 {
 290         *dst = zalloc(size);
 291         if (!*dst)
 292                 die("memory allocation failed - maybe size is too large?\n");
 293 }
 294
 295 static u64 do_memset_cycles(const struct routine *r, size_t size)
 296 {
 297         u64 cycle_start = 0ULL, cycle_end = 0ULL;
 298         memset_t fn = r->fn.memset;
 299         void *dst = NULL;
 300         int i;
 301
 302         memset_alloc_mem(&dst, size);
 303
 304         /*
 305          * We prefault the freshly allocated memory range here,
 306          * to not measure page fault overhead:
 307          */
 308         fn(dst, -1, size);
 309
 310         cycle_start = get_cycles();
 311         for (i = 0; i < iterations; ++i)
 312                 fn(dst, i, size);
 313         cycle_end = get_cycles();
 314
 315         free(dst);
 316         return cycle_end - cycle_start;
 317 }
 318
 319 static double do_memset_gettimeofday(const struct routine *r, size_t size)
 320 {
 321         struct timeval tv_start, tv_end, tv_diff;
 322         memset_t fn = r->fn.memset;
 323         void *dst = NULL;
 324         int i;
 325
 326         memset_alloc_mem(&dst, size);
 327
 328         /*
 329          * We prefault the freshly allocated memory range here,
 330          * to not measure page fault overhead:
 331          */
 332         fn(dst, -1, size);
 333
 334         BUG_ON(gettimeofday(&tv_start, NULL));
 335         for (i = 0; i < iterations; ++i)
 336                 fn(dst, i, size);
 337         BUG_ON(gettimeofday(&tv_end, NULL));
 338
 339         timersub(&tv_end, &tv_start, &tv_diff);
 340
 341         free(dst);
 342         return (double)(((double)size * iterations) / timeval2double(&tv_diff));
 343 }
 344
 345 static const char * const bench_mem_memset_usage[] = {
 346         "perf bench mem memset <options>",
 347         NULL
 348 };
 349
 350 static const struct routine memset_routines[] = {
 351         { .name         = "default",
 352           .desc         = "Default memset() provided by glibc",
 353           .fn.memset    = memset },
 354
 355 #ifdef HAVE_ARCH_X86_64_SUPPORT
 356 # define MEMSET_FN(_fn, _name, _desc) { .name = _name, .desc = _desc, .fn.memset = _fn },
 357 # include "mem-memset-x86-64-asm-def.h"
 358 # undef MEMSET_FN
 359 #endif
 360
 361         { NULL, }
 362 };
 363
 364 int bench_mem_memset(int argc, const char **argv, const char *prefix __maybe_unused)
 365 {
 366         struct bench_mem_info info = {
 367                 .routines               = memset_routines,
 368                 .do_cycles              = do_memset_cycles,
 369                 .do_gettimeofday        = do_memset_gettimeofday,
 370                 .usage                  = bench_mem_memset_usage,
 371         };
 372
 373         return bench_mem_common(argc, argv, &info);
 374 }