4 * Simple memcpy() and memset() benchmarks
6 * Written by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
11 #include "../util/util.h"
12 #include <subcmd/parse-options.h>
13 #include "../util/header.h"
14 #include "../util/cloexec.h"
15 #include "../util/string2.h"
17 #include "mem-memcpy-arch.h"
18 #include "mem-memset-arch.h"
25 #include <linux/time64.h>
29 static const char *size_str = "1MB";
30 static const char *function_str = "all";
31 static int nr_loops = 1;
32 static bool use_cycles;
35 static const struct option options[] = {
36 OPT_STRING('s', "size", &size_str, "1MB",
37 "Specify the size of the memory buffers. "
38 "Available units: B, KB, MB, GB and TB (case insensitive)"),
40 OPT_STRING('f', "function", &function_str, "all",
41 "Specify the function to run, \"all\" runs all available functions, \"help\" lists them"),
43 OPT_INTEGER('l', "nr_loops", &nr_loops,
44 "Specify the number of loops to run. (default: 1)"),
46 OPT_BOOLEAN('c', "cycles", &use_cycles,
47 "Use a cycles event instead of gettimeofday() to measure performance"),
52 typedef void *(*memcpy_t)(void *, const void *, size_t);
53 typedef void *(*memset_t)(void *, int, size_t);
64 static struct perf_event_attr cycle_attr = {
65 .type = PERF_TYPE_HARDWARE,
66 .config = PERF_COUNT_HW_CPU_CYCLES
69 static int init_cycles(void)
71 cycles_fd = sys_perf_event_open(&cycle_attr, getpid(), -1, -1, perf_event_open_cloexec_flag());
73 if (cycles_fd < 0 && errno == ENOSYS) {
74 pr_debug("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
81 static u64 get_cycles(void)
86 ret = read(cycles_fd, &clk, sizeof(u64));
87 BUG_ON(ret != sizeof(u64));
92 static double timeval2double(struct timeval *ts)
94 return (double)ts->tv_sec + (double)ts->tv_usec / (double)USEC_PER_SEC;
97 #define print_bps(x) do { \
99 printf(" %14lf bytes/sec\n", x); \
100 else if (x < K * K) \
101 printf(" %14lfd KB/sec\n", x / K); \
102 else if (x < K * K * K) \
103 printf(" %14lf MB/sec\n", x / K / K); \
105 printf(" %14lf GB/sec\n", x / K / K / K); \
108 struct bench_mem_info {
109 const struct function *functions;
110 u64 (*do_cycles)(const struct function *r, size_t size, void *src, void *dst);
111 double (*do_gettimeofday)(const struct function *r, size_t size, void *src, void *dst);
112 const char *const *usage;
116 static void __bench_mem_function(struct bench_mem_info *info, int r_idx, size_t size, double size_total)
118 const struct function *r = &info->functions[r_idx];
119 double result_bps = 0.0;
120 u64 result_cycles = 0;
121 void *src = NULL, *dst = zalloc(size);
123 printf("# function '%s' (%s)\n", r->name, r->desc);
126 goto out_alloc_failed;
128 if (info->alloc_src) {
131 goto out_alloc_failed;
134 if (bench_format == BENCH_FORMAT_DEFAULT)
135 printf("# Copying %s bytes ...\n\n", size_str);
138 result_cycles = info->do_cycles(r, size, src, dst);
140 result_bps = info->do_gettimeofday(r, size, src, dst);
143 switch (bench_format) {
144 case BENCH_FORMAT_DEFAULT:
146 printf(" %14lf cycles/byte\n", (double)result_cycles/size_total);
148 print_bps(result_bps);
152 case BENCH_FORMAT_SIMPLE:
154 printf("%lf\n", (double)result_cycles/size_total);
156 printf("%lf\n", result_bps);
170 printf("# Memory allocation failed - maybe size (%s) is too large?\n", size_str);
174 static int bench_mem_common(int argc, const char **argv, struct bench_mem_info *info)
180 argc = parse_options(argc, argv, options, info->usage, 0);
185 fprintf(stderr, "Failed to open cycles counter\n");
190 size = (size_t)perf_atoll((char *)size_str);
191 size_total = (double)size * nr_loops;
193 if ((s64)size <= 0) {
194 fprintf(stderr, "Invalid size:%s\n", size_str);
198 if (!strncmp(function_str, "all", 3)) {
199 for (i = 0; info->functions[i].name; i++)
200 __bench_mem_function(info, i, size, size_total);
204 for (i = 0; info->functions[i].name; i++) {
205 if (!strcmp(info->functions[i].name, function_str))
208 if (!info->functions[i].name) {
209 if (strcmp(function_str, "help") && strcmp(function_str, "h"))
210 printf("Unknown function: %s\n", function_str);
211 printf("Available functions:\n");
212 for (i = 0; info->functions[i].name; i++) {
213 printf("\t%s ... %s\n",
214 info->functions[i].name, info->functions[i].desc);
219 __bench_mem_function(info, i, size, size_total);
224 static u64 do_memcpy_cycles(const struct function *r, size_t size, void *src, void *dst)
226 u64 cycle_start = 0ULL, cycle_end = 0ULL;
227 memcpy_t fn = r->fn.memcpy;
230 /* Make sure to always prefault zero pages even if MMAP_THRESH is crossed: */
231 memset(src, 0, size);
234 * We prefault the freshly allocated memory range here,
235 * to not measure page fault overhead:
239 cycle_start = get_cycles();
240 for (i = 0; i < nr_loops; ++i)
242 cycle_end = get_cycles();
244 return cycle_end - cycle_start;
247 static double do_memcpy_gettimeofday(const struct function *r, size_t size, void *src, void *dst)
249 struct timeval tv_start, tv_end, tv_diff;
250 memcpy_t fn = r->fn.memcpy;
254 * We prefault the freshly allocated memory range here,
255 * to not measure page fault overhead:
259 BUG_ON(gettimeofday(&tv_start, NULL));
260 for (i = 0; i < nr_loops; ++i)
262 BUG_ON(gettimeofday(&tv_end, NULL));
264 timersub(&tv_end, &tv_start, &tv_diff);
266 return (double)(((double)size * nr_loops) / timeval2double(&tv_diff));
269 struct function memcpy_functions[] = {
271 .desc = "Default memcpy() provided by glibc",
272 .fn.memcpy = memcpy },
274 #ifdef HAVE_ARCH_X86_64_SUPPORT
275 # define MEMCPY_FN(_fn, _name, _desc) {.name = _name, .desc = _desc, .fn.memcpy = _fn},
276 # include "mem-memcpy-x86-64-asm-def.h"
283 static const char * const bench_mem_memcpy_usage[] = {
284 "perf bench mem memcpy <options>",
288 int bench_mem_memcpy(int argc, const char **argv)
290 struct bench_mem_info info = {
291 .functions = memcpy_functions,
292 .do_cycles = do_memcpy_cycles,
293 .do_gettimeofday = do_memcpy_gettimeofday,
294 .usage = bench_mem_memcpy_usage,
298 return bench_mem_common(argc, argv, &info);
301 static u64 do_memset_cycles(const struct function *r, size_t size, void *src __maybe_unused, void *dst)
303 u64 cycle_start = 0ULL, cycle_end = 0ULL;
304 memset_t fn = r->fn.memset;
308 * We prefault the freshly allocated memory range here,
309 * to not measure page fault overhead:
313 cycle_start = get_cycles();
314 for (i = 0; i < nr_loops; ++i)
316 cycle_end = get_cycles();
318 return cycle_end - cycle_start;
321 static double do_memset_gettimeofday(const struct function *r, size_t size, void *src __maybe_unused, void *dst)
323 struct timeval tv_start, tv_end, tv_diff;
324 memset_t fn = r->fn.memset;
328 * We prefault the freshly allocated memory range here,
329 * to not measure page fault overhead:
333 BUG_ON(gettimeofday(&tv_start, NULL));
334 for (i = 0; i < nr_loops; ++i)
336 BUG_ON(gettimeofday(&tv_end, NULL));
338 timersub(&tv_end, &tv_start, &tv_diff);
340 return (double)(((double)size * nr_loops) / timeval2double(&tv_diff));
343 static const char * const bench_mem_memset_usage[] = {
344 "perf bench mem memset <options>",
348 static const struct function memset_functions[] = {
350 .desc = "Default memset() provided by glibc",
351 .fn.memset = memset },
353 #ifdef HAVE_ARCH_X86_64_SUPPORT
354 # define MEMSET_FN(_fn, _name, _desc) { .name = _name, .desc = _desc, .fn.memset = _fn },
355 # include "mem-memset-x86-64-asm-def.h"
362 int bench_mem_memset(int argc, const char **argv)
364 struct bench_mem_info info = {
365 .functions = memset_functions,
366 .do_cycles = do_memset_cycles,
367 .do_gettimeofday = do_memset_gettimeofday,
368 .usage = bench_mem_memset_usage,
371 return bench_mem_common(argc, argv, &info);