5 #include "thread_map.h"
7 void update_stats(struct stats *stats, u64 val)
12 delta = val - stats->mean;
13 stats->mean += delta / stats->n;
14 stats->M2 += delta*(val - stats->mean);
23 double avg_stats(struct stats *stats)
29 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
31 * (\Sum n_i^2) - ((\Sum n_i)^2)/n
32 * s^2 = -------------------------------
35 * http://en.wikipedia.org/wiki/Stddev
37 * The std dev of the mean is related to the std dev by:
44 double stddev_stats(struct stats *stats)
46 double variance, variance_mean;
51 variance = stats->M2 / (stats->n - 1);
52 variance_mean = variance / stats->n;
54 return sqrt(variance_mean);
57 double rel_stddev_stats(double stddev, double avg)
62 pct = 100.0 * stddev/avg;
67 bool __perf_evsel_stat__is(struct perf_evsel *evsel,
68 enum perf_stat_evsel_id id)
70 struct perf_stat *ps = evsel->priv;
75 #define ID(id, name) [PERF_STAT_EVSEL_ID__##id] = #name
76 static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
78 ID(CYCLES_IN_TX, cpu/cycles-t/),
79 ID(TRANSACTION_START, cpu/tx-start/),
80 ID(ELISION_START, cpu/el-start/),
81 ID(CYCLES_IN_TX_CP, cpu/cycles-ct/),
85 void perf_stat_evsel_id_init(struct perf_evsel *evsel)
87 struct perf_stat *ps = evsel->priv;
90 /* ps->id is 0 hence PERF_STAT_EVSEL_ID__NONE by default */
92 for (i = 0; i < PERF_STAT_EVSEL_ID__MAX; i++) {
93 if (!strcmp(perf_evsel__name(evsel), id_str[i])) {
100 void perf_evsel__reset_stat_priv(struct perf_evsel *evsel)
103 struct perf_stat *ps = evsel->priv;
105 for (i = 0; i < 3; i++)
106 init_stats(&ps->res_stats[i]);
108 perf_stat_evsel_id_init(evsel);
111 int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
113 evsel->priv = zalloc(sizeof(struct perf_stat));
114 if (evsel->priv == NULL)
116 perf_evsel__reset_stat_priv(evsel);
120 void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
125 int perf_evsel__alloc_prev_raw_counts(struct perf_evsel *evsel,
126 int ncpus, int nthreads)
128 struct perf_counts *counts;
130 counts = perf_counts__new(ncpus, nthreads);
132 evsel->prev_raw_counts = counts;
134 return counts ? 0 : -ENOMEM;
137 void perf_evsel__free_prev_raw_counts(struct perf_evsel *evsel)
139 perf_counts__delete(evsel->prev_raw_counts);
140 evsel->prev_raw_counts = NULL;
143 int perf_evsel__alloc_stats(struct perf_evsel *evsel, bool alloc_raw)
145 int ncpus = perf_evsel__nr_cpus(evsel);
146 int nthreads = thread_map__nr(evsel->threads);
148 if (perf_evsel__alloc_stat_priv(evsel) < 0 ||
149 perf_evsel__alloc_counts(evsel, ncpus, nthreads) < 0 ||
150 (alloc_raw && perf_evsel__alloc_prev_raw_counts(evsel, ncpus, nthreads) < 0))
156 int perf_evlist__alloc_stats(struct perf_evlist *evlist, bool alloc_raw)
158 struct perf_evsel *evsel;
160 evlist__for_each(evlist, evsel) {
161 if (perf_evsel__alloc_stats(evsel, alloc_raw))
168 perf_evlist__free_stats(evlist);
172 void perf_evlist__free_stats(struct perf_evlist *evlist)
174 struct perf_evsel *evsel;
176 evlist__for_each(evlist, evsel) {
177 perf_evsel__free_stat_priv(evsel);
178 perf_evsel__free_counts(evsel);
179 perf_evsel__free_prev_raw_counts(evsel);
183 void perf_evlist__reset_stats(struct perf_evlist *evlist)
185 struct perf_evsel *evsel;
187 evlist__for_each(evlist, evsel) {
188 perf_evsel__reset_stat_priv(evsel);
189 perf_evsel__reset_counts(evsel);
193 static void zero_per_pkg(struct perf_evsel *counter)
195 if (counter->per_pkg_mask)
196 memset(counter->per_pkg_mask, 0, MAX_NR_CPUS);
199 static int check_per_pkg(struct perf_evsel *counter,
200 struct perf_counts_values *vals, int cpu, bool *skip)
202 unsigned long *mask = counter->per_pkg_mask;
203 struct cpu_map *cpus = perf_evsel__cpus(counter);
208 if (!counter->per_pkg)
211 if (cpu_map__empty(cpus))
215 mask = zalloc(MAX_NR_CPUS);
219 counter->per_pkg_mask = mask;
223 * we do not consider an event that has not run as a good
224 * instance to mark a package as used (skip=1). Otherwise
225 * we may run into a situation where the first CPU in a package
226 * is not running anything, yet the second is, and this function
227 * would mark the package as used after the first CPU and would
228 * not read the values from the second CPU.
230 if (!(vals->run && vals->ena))
233 s = cpu_map__get_socket(cpus, cpu);
237 *skip = test_and_set_bit(s, mask) == 1;
242 process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel,
244 struct perf_counts_values *count)
246 struct perf_counts_values *aggr = &evsel->counts->aggr;
247 static struct perf_counts_values zero;
250 if (check_per_pkg(evsel, count, cpu, &skip)) {
251 pr_err("failed to read per-pkg counter\n");
258 switch (config->aggr_mode) {
263 if (!evsel->snapshot)
264 perf_evsel__compute_deltas(evsel, cpu, thread, count);
265 perf_counts_values__scale(count, config->scale, NULL);
266 if (config->aggr_mode == AGGR_NONE)
267 perf_stat__update_shadow_stats(evsel, count->values, cpu);
270 aggr->val += count->val;
272 aggr->ena += count->ena;
273 aggr->run += count->run;
282 static int process_counter_maps(struct perf_stat_config *config,
283 struct perf_evsel *counter)
285 int nthreads = thread_map__nr(counter->threads);
286 int ncpus = perf_evsel__nr_cpus(counter);
289 if (counter->system_wide)
292 for (thread = 0; thread < nthreads; thread++) {
293 for (cpu = 0; cpu < ncpus; cpu++) {
294 if (process_counter_values(config, counter, cpu, thread,
295 perf_counts(counter->counts, cpu, thread)))
303 int perf_stat_process_counter(struct perf_stat_config *config,
304 struct perf_evsel *counter)
306 struct perf_counts_values *aggr = &counter->counts->aggr;
307 struct perf_stat *ps = counter->priv;
308 u64 *count = counter->counts->aggr.values;
311 aggr->val = aggr->ena = aggr->run = 0;
312 init_stats(ps->res_stats);
314 if (counter->per_pkg)
315 zero_per_pkg(counter);
317 ret = process_counter_maps(config, counter);
321 if (config->aggr_mode != AGGR_GLOBAL)
324 if (!counter->snapshot)
325 perf_evsel__compute_deltas(counter, -1, -1, aggr);
326 perf_counts_values__scale(aggr, config->scale, &counter->counts->scaled);
328 for (i = 0; i < 3; i++)
329 update_stats(&ps->res_stats[i], count[i]);
332 fprintf(config->output, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
333 perf_evsel__name(counter), count[0], count[1], count[2]);
337 * Save the full runtime - to allow normalization during printout:
339 perf_stat__update_shadow_stats(counter, count, 0);