#include "util/parse-events.h"
#include <sys/prctl.h>
+#include <math.h>
static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
};
+#define MAX_RUN 100
+
static int system_wide = 0;
-static int inherit = 1;
static int verbose = 0;
+static int nr_cpus = 0;
+static int run_idx = 0;
+
+static int run_count = 1;
+static int inherit = 1;
+static int scale = 1;
+static int target_pid = -1;
+static int null_run = 0;
static int fd[MAX_NR_CPUS][MAX_COUNTERS];
-static int target_pid = -1;
-static int nr_cpus = 0;
-static unsigned int page_size;
+static u64 runtime_nsecs[MAX_RUN];
+static u64 walltime_nsecs[MAX_RUN];
+static u64 runtime_cycles[MAX_RUN];
-static int scale = 1;
+static u64 event_res[MAX_RUN][MAX_COUNTERS][3];
+static u64 event_scaled[MAX_RUN][MAX_COUNTERS];
-static const unsigned int default_count[] = {
- 1000000,
- 1000000,
- 10000,
- 10000,
- 1000000,
- 10000,
-};
+static u64 event_res_avg[MAX_COUNTERS][3];
+static u64 event_res_noise[MAX_COUNTERS][3];
+
+static u64 event_scaled_avg[MAX_COUNTERS];
-static __u64 event_res[MAX_COUNTERS][3];
-static __u64 event_scaled[MAX_COUNTERS];
+static u64 runtime_nsecs_avg;
+static u64 runtime_nsecs_noise;
-static __u64 runtime_nsecs;
-static __u64 walltime_nsecs;
-static __u64 runtime_cycles;
+static u64 walltime_nsecs_avg;
+static u64 walltime_nsecs_noise;
+
+static u64 runtime_cycles_avg;
+static u64 runtime_cycles_noise;
+
+#define ERR_PERF_OPEN \
+"Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
static void create_perf_stat_counter(int counter)
{
if (system_wide) {
int cpu;
- for (cpu = 0; cpu < nr_cpus; cpu ++) {
+ for (cpu = 0; cpu < nr_cpus; cpu++) {
fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
- if (fd[cpu][counter] < 0 && verbose) {
- printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[cpu][counter], strerror(errno));
- }
+ if (fd[cpu][counter] < 0 && verbose)
+ fprintf(stderr, ERR_PERF_OPEN, counter,
+ fd[cpu][counter], strerror(errno));
}
} else {
attr->inherit = inherit;
attr->disabled = 1;
fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
- if (fd[0][counter] < 0 && verbose) {
- printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[0][counter], strerror(errno));
- }
+ if (fd[0][counter] < 0 && verbose)
+ fprintf(stderr, ERR_PERF_OPEN, counter,
+ fd[0][counter], strerror(errno));
}
}
*/
static void read_counter(int counter)
{
- __u64 *count, single_count[3];
+ u64 *count, single_count[3];
ssize_t res;
int cpu, nv;
int scaled;
- count = event_res[counter];
+ count = event_res[run_idx][counter];
count[0] = count[1] = count[2] = 0;
nv = scale ? 3 : 1;
- for (cpu = 0; cpu < nr_cpus; cpu ++) {
+ for (cpu = 0; cpu < nr_cpus; cpu++) {
if (fd[cpu][counter] < 0)
continue;
- res = read(fd[cpu][counter], single_count, nv * sizeof(__u64));
- assert(res == nv * sizeof(__u64));
+ res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
+ assert(res == nv * sizeof(u64));
+ close(fd[cpu][counter]);
+ fd[cpu][counter] = -1;
count[0] += single_count[0];
if (scale) {
scaled = 0;
if (scale) {
if (count[2] == 0) {
- event_scaled[counter] = -1;
+ event_scaled[run_idx][counter] = -1;
count[0] = 0;
return;
}
if (count[2] < count[1]) {
- event_scaled[counter] = 1;
+ event_scaled[run_idx][counter] = 1;
count[0] = (unsigned long long)
((double)count[0] * count[1] / count[2] + 0.5);
}
*/
if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
- runtime_nsecs = count[0];
+ runtime_nsecs[run_idx] = count[0];
if (attrs[counter].type == PERF_TYPE_HARDWARE &&
attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
- runtime_cycles = count[0];
+ runtime_cycles[run_idx] = count[0];
+}
+
+static int run_perf_stat(int argc, const char **argv)
+{
+ unsigned long long t0, t1;
+ int status = 0;
+ int counter;
+ int pid;
+
+ if (!system_wide)
+ nr_cpus = 1;
+
+ for (counter = 0; counter < nr_counters; counter++)
+ create_perf_stat_counter(counter);
+
+ /*
+ * Enable counters and exec the command:
+ */
+ t0 = rdclock();
+ prctl(PR_TASK_PERF_COUNTERS_ENABLE);
+
+ if ((pid = fork()) < 0)
+ perror("failed to fork");
+
+ if (!pid) {
+ if (execvp(argv[0], (char **)argv)) {
+ perror(argv[0]);
+ exit(-1);
+ }
+ }
+
+ wait(&status);
+
+ prctl(PR_TASK_PERF_COUNTERS_DISABLE);
+ t1 = rdclock();
+
+ walltime_nsecs[run_idx] = t1 - t0;
+
+ for (counter = 0; counter < nr_counters; counter++)
+ read_counter(counter);
+
+ return WEXITSTATUS(status);
+}
+
+static void print_noise(u64 *count, u64 *noise)
+{
+ if (run_count > 1)
+ fprintf(stderr, " ( +- %7.3f%% )",
+ (double)noise[0]/(count[0]+1)*100.0);
+}
+
+static void nsec_printout(int counter, u64 *count, u64 *noise)
+{
+ double msecs = (double)count[0] / 1000000;
+
+ fprintf(stderr, " %14.6f %-20s", msecs, event_name(counter));
+
+ if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
+ attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
+
+ if (walltime_nsecs_avg)
+ fprintf(stderr, " # %10.3f CPUs ",
+ (double)count[0] / (double)walltime_nsecs_avg);
+ }
+ print_noise(count, noise);
+}
+
+static void abs_printout(int counter, u64 *count, u64 *noise)
+{
+ fprintf(stderr, " %14Ld %-20s", count[0], event_name(counter));
+
+ if (runtime_cycles_avg &&
+ attrs[counter].type == PERF_TYPE_HARDWARE &&
+ attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
+
+ fprintf(stderr, " # %10.3f IPC ",
+ (double)count[0] / (double)runtime_cycles_avg);
+ } else {
+ if (runtime_nsecs_avg) {
+ fprintf(stderr, " # %10.3f M/sec",
+ (double)count[0]/runtime_nsecs_avg*1000.0);
+ }
+ }
+ print_noise(count, noise);
}
/*
*/
static void print_counter(int counter)
{
- __u64 *count;
+ u64 *count, *noise;
int scaled;
- count = event_res[counter];
- scaled = event_scaled[counter];
+ count = event_res_avg[counter];
+ noise = event_res_noise[counter];
+ scaled = event_scaled_avg[counter];
if (scaled == -1) {
fprintf(stderr, " %14s %-20s\n",
return;
}
- if (nsec_counter(counter)) {
- double msecs = (double)count[0] / 1000000;
+ if (nsec_counter(counter))
+ nsec_printout(counter, count, noise);
+ else
+ abs_printout(counter, count, noise);
- fprintf(stderr, " %14.6f %-20s",
- msecs, event_name(counter));
- if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
- attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
-
- if (walltime_nsecs)
- fprintf(stderr, " # %11.3f CPU utilization factor",
- (double)count[0] / (double)walltime_nsecs);
- }
- } else {
- fprintf(stderr, " %14Ld %-20s",
- count[0], event_name(counter));
- if (runtime_nsecs)
- fprintf(stderr, " # %11.3f M/sec",
- (double)count[0]/runtime_nsecs*1000.0);
- if (runtime_cycles &&
- attrs[counter].type == PERF_TYPE_HARDWARE &&
- attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
-
- fprintf(stderr, " # %1.3f per cycle",
- (double)count[0] / (double)runtime_cycles);
- }
- }
if (scaled)
fprintf(stderr, " (scaled from %.2f%%)",
(double) count[2] / count[1] * 100);
+
fprintf(stderr, "\n");
}
-static int do_perf_stat(int argc, const char **argv)
+/*
+ * normalize_noise noise values down to stddev:
+ */
+static void normalize_noise(u64 *val)
{
- unsigned long long t0, t1;
- int counter;
- int status;
- int pid;
- int i;
+ double res;
- if (!system_wide)
- nr_cpus = 1;
+ res = (double)*val / (run_count * sqrt((double)run_count));
- for (counter = 0; counter < nr_counters; counter++)
- create_perf_stat_counter(counter);
+ *val = (u64)res;
+}
- /*
- * Enable counters and exec the command:
- */
- t0 = rdclock();
- prctl(PR_TASK_PERF_COUNTERS_ENABLE);
+static void update_avg(const char *name, int idx, u64 *avg, u64 *val)
+{
+ *avg += *val;
- if ((pid = fork()) < 0)
- perror("failed to fork");
+ if (verbose > 1)
+ fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
+}
+/*
+ * Calculate the averages and noises:
+ */
+static void calc_avg(void)
+{
+ int i, j;
+
+ if (verbose > 1)
+ fprintf(stderr, "\n");
+
+ for (i = 0; i < run_count; i++) {
+ update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
+ update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
+ update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
+
+ for (j = 0; j < nr_counters; j++) {
+ update_avg("counter/0", j,
+ event_res_avg[j]+0, event_res[i][j]+0);
+ update_avg("counter/1", j,
+ event_res_avg[j]+1, event_res[i][j]+1);
+ update_avg("counter/2", j,
+ event_res_avg[j]+2, event_res[i][j]+2);
+ if (event_scaled[i][j] != -1)
+ update_avg("scaled", j,
+ event_scaled_avg + j, event_scaled[i]+j);
+ else
+ event_scaled_avg[j] = -1;
+ }
+ }
+ runtime_nsecs_avg /= run_count;
+ walltime_nsecs_avg /= run_count;
+ runtime_cycles_avg /= run_count;
+
+ for (j = 0; j < nr_counters; j++) {
+ event_res_avg[j][0] /= run_count;
+ event_res_avg[j][1] /= run_count;
+ event_res_avg[j][2] /= run_count;
+ }
- if (!pid) {
- if (execvp(argv[0], (char **)argv)) {
- perror(argv[0]);
- exit(-1);
+ for (i = 0; i < run_count; i++) {
+ runtime_nsecs_noise +=
+ abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
+ walltime_nsecs_noise +=
+ abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
+ runtime_cycles_noise +=
+ abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
+
+ for (j = 0; j < nr_counters; j++) {
+ event_res_noise[j][0] +=
+ abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
+ event_res_noise[j][1] +=
+ abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
+ event_res_noise[j][2] +=
+ abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
}
}
- while (wait(&status) >= 0)
- ;
+ normalize_noise(&runtime_nsecs_noise);
+ normalize_noise(&walltime_nsecs_noise);
+ normalize_noise(&runtime_cycles_noise);
- prctl(PR_TASK_PERF_COUNTERS_DISABLE);
- t1 = rdclock();
+ for (j = 0; j < nr_counters; j++) {
+ normalize_noise(&event_res_noise[j][0]);
+ normalize_noise(&event_res_noise[j][1]);
+ normalize_noise(&event_res_noise[j][2]);
+ }
+}
- walltime_nsecs = t1 - t0;
+static void print_stat(int argc, const char **argv)
+{
+ int i, counter;
+
+ calc_avg();
fflush(stdout);
for (i = 1; i < argc; i++)
fprintf(stderr, " %s", argv[i]);
- fprintf(stderr, "\':\n");
- fprintf(stderr, "\n");
-
- for (counter = 0; counter < nr_counters; counter++)
- read_counter(counter);
+ fprintf(stderr, "\'");
+ if (run_count > 1)
+ fprintf(stderr, " (%d runs)", run_count);
+ fprintf(stderr, ":\n\n");
for (counter = 0; counter < nr_counters; counter++)
print_counter(counter);
fprintf(stderr, "\n");
- fprintf(stderr, " Wall-clock time elapsed: %12.6f msecs\n",
- (double)(t1-t0)/1e6);
- fprintf(stderr, "\n");
-
- return 0;
+ fprintf(stderr, " %14.9f seconds time elapsed",
+ (double)walltime_nsecs_avg/1e9);
+ if (run_count > 1) {
+ fprintf(stderr, " ( +- %7.3f%% )",
+ 100.0*(double)walltime_nsecs_noise/(double)walltime_nsecs_avg);
+ }
+ fprintf(stderr, "\n\n");
}
static volatile int signr = -1;
OPT_INTEGER('p', "pid", &target_pid,
"stat events on existing pid"),
OPT_BOOLEAN('a', "all-cpus", &system_wide,
- "system-wide collection from all CPUs"),
+ "system-wide collection from all CPUs"),
OPT_BOOLEAN('S', "scale", &scale,
- "scale/normalize counters"),
+ "scale/normalize counters"),
OPT_BOOLEAN('v', "verbose", &verbose,
"be more verbose (show counter open errors, etc)"),
+ OPT_INTEGER('r', "repeat", &run_count,
+ "repeat command and print average + stddev (max: 100)"),
+ OPT_BOOLEAN('n', "null", &null_run,
+ "null run - dont start any counters"),
OPT_END()
};
int cmd_stat(int argc, const char **argv, const char *prefix)
{
- page_size = sysconf(_SC_PAGE_SIZE);
+ int status;
memcpy(attrs, default_attrs, sizeof(attrs));
argc = parse_options(argc, argv, options, stat_usage, 0);
if (!argc)
usage_with_options(stat_usage, options);
+ if (run_count <= 0 || run_count > MAX_RUN)
+ usage_with_options(stat_usage, options);
- if (!nr_counters)
+ if (!null_run && !nr_counters)
nr_counters = 8;
nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
signal(SIGALRM, skip_signal);
signal(SIGABRT, skip_signal);
- return do_perf_stat(argc, argv);
+ status = 0;
+ for (run_idx = 0; run_idx < run_count; run_idx++) {
+ if (run_count != 1 && verbose)
+ fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
+ status = run_perf_stat(argc, argv);
+ }
+
+ print_stat(argc, argv);
+
+ return status;
}
projects / firefly-linux-kernel-4.4.55.git / blobdiff