1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/compiler.h>
4 #include <perf/cpumap.h>
5 #include <perf/evlist.h>
6 #include "metricgroup.h"
8 #include "pmu-events/pmu-events.h"
15 static struct pmu_event pme_test[] = {
17 .metric_expr = "inst_retired.any / cpu_clk_unhalted.thread",
19 .metric_group = "group1",
22 .metric_expr = "idq_uops_not_delivered.core / (4 * (( ( cpu_clk_unhalted.thread / 2 ) * "
23 "( 1 + cpu_clk_unhalted.one_thread_active / cpu_clk_unhalted.ref_xclk ) )))",
24 .metric_name = "Frontend_Bound_SMT",
27 .metric_expr = "l1d\\-loads\\-misses / inst_retired.any",
28 .metric_name = "dcache_miss_cpi",
31 .metric_expr = "l1i\\-loads\\-misses / inst_retired.any",
32 .metric_name = "icache_miss_cycles",
35 .metric_expr = "(dcache_miss_cpi + icache_miss_cycles)",
36 .metric_name = "cache_miss_cycles",
37 .metric_group = "group1",
40 .metric_expr = "l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit",
41 .metric_name = "DCache_L2_All_Hits",
44 .metric_expr = "max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + "
45 "l2_rqsts.pf_miss + l2_rqsts.rfo_miss",
46 .metric_name = "DCache_L2_All_Miss",
49 .metric_expr = "dcache_l2_all_hits + dcache_l2_all_miss",
50 .metric_name = "DCache_L2_All",
53 .metric_expr = "d_ratio(dcache_l2_all_hits, dcache_l2_all)",
54 .metric_name = "DCache_L2_Hits",
57 .metric_expr = "d_ratio(dcache_l2_all_miss, dcache_l2_all)",
58 .metric_name = "DCache_L2_Misses",
61 .metric_expr = "ipc + m2",
65 .metric_expr = "ipc + m1",
69 .metric_expr = "1/m3",
73 .metric_expr = "64 * l1d.replacement / 1000000000 / duration_time",
74 .metric_name = "L1D_Cache_Fill_BW",
81 static struct pmu_events_map map = {
93 static u64 find_value(const char *name, struct value *values)
95 struct value *v = values;
98 if (!strcmp(name, v->event))
105 static void load_runtime_stat(struct runtime_stat *st, struct evlist *evlist,
111 evlist__for_each_entry(evlist, evsel) {
112 count = find_value(evsel->name, vals);
113 perf_stat__update_shadow_stats(evsel, count, 0, st);
114 if (!strcmp(evsel->name, "duration_time"))
115 update_stats(&walltime_nsecs_stats, count);
119 static double compute_single(struct rblist *metric_events, struct evlist *evlist,
120 struct runtime_stat *st, const char *name)
122 struct metric_expr *mexp;
123 struct metric_event *me;
126 evlist__for_each_entry(evlist, evsel) {
127 me = metricgroup__lookup(metric_events, evsel, false);
129 list_for_each_entry (mexp, &me->head, nd) {
130 if (strcmp(mexp->metric_name, name))
132 return test_generic_metric(mexp, 0, st);
139 static int __compute_metric(const char *name, struct value *vals,
140 const char *name1, double *ratio1,
141 const char *name2, double *ratio2)
143 struct rblist metric_events = {
146 struct perf_cpu_map *cpus;
147 struct runtime_stat st;
148 struct evlist *evlist;
152 * We need to prepare evlist for stat mode running on CPU 0
153 * because that's where all the stats are going to be created.
155 evlist = evlist__new();
159 cpus = perf_cpu_map__new("0");
161 evlist__delete(evlist);
165 perf_evlist__set_maps(&evlist->core, cpus, NULL);
166 runtime_stat__init(&st);
168 /* Parse the metric into metric_events list. */
169 err = metricgroup__parse_groups_test(evlist, &map, name,
175 err = evlist__alloc_stats(evlist, false);
179 /* Load the runtime stats with given numbers for events. */
180 load_runtime_stat(&st, evlist, vals);
182 /* And execute the metric */
184 *ratio1 = compute_single(&metric_events, evlist, &st, name1);
186 *ratio2 = compute_single(&metric_events, evlist, &st, name2);
190 metricgroup__rblist_exit(&metric_events);
191 runtime_stat__exit(&st);
192 evlist__free_stats(evlist);
193 perf_cpu_map__put(cpus);
194 evlist__delete(evlist);
198 static int compute_metric(const char *name, struct value *vals, double *ratio)
200 return __compute_metric(name, vals, name, ratio, NULL, NULL);
203 static int compute_metric_group(const char *name, struct value *vals,
204 const char *name1, double *ratio1,
205 const char *name2, double *ratio2)
207 return __compute_metric(name, vals, name1, ratio1, name2, ratio2);
210 static int test_ipc(void)
213 struct value vals[] = {
214 { .event = "inst_retired.any", .val = 300 },
215 { .event = "cpu_clk_unhalted.thread", .val = 200 },
219 TEST_ASSERT_VAL("failed to compute metric",
220 compute_metric("IPC", vals, &ratio) == 0);
222 TEST_ASSERT_VAL("IPC failed, wrong ratio",
227 static int test_frontend(void)
230 struct value vals[] = {
231 { .event = "idq_uops_not_delivered.core", .val = 300 },
232 { .event = "cpu_clk_unhalted.thread", .val = 200 },
233 { .event = "cpu_clk_unhalted.one_thread_active", .val = 400 },
234 { .event = "cpu_clk_unhalted.ref_xclk", .val = 600 },
238 TEST_ASSERT_VAL("failed to compute metric",
239 compute_metric("Frontend_Bound_SMT", vals, &ratio) == 0);
241 TEST_ASSERT_VAL("Frontend_Bound_SMT failed, wrong ratio",
246 static int test_cache_miss_cycles(void)
249 struct value vals[] = {
250 { .event = "l1d-loads-misses", .val = 300 },
251 { .event = "l1i-loads-misses", .val = 200 },
252 { .event = "inst_retired.any", .val = 400 },
256 TEST_ASSERT_VAL("failed to compute metric",
257 compute_metric("cache_miss_cycles", vals, &ratio) == 0);
259 TEST_ASSERT_VAL("cache_miss_cycles failed, wrong ratio",
266 * DCache_L2_All_Hits = l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hi
267 * DCache_L2_All_Miss = max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) +
268 * l2_rqsts.pf_miss + l2_rqsts.rfo_miss
269 * DCache_L2_All = dcache_l2_all_hits + dcache_l2_all_miss
270 * DCache_L2_Hits = d_ratio(dcache_l2_all_hits, dcache_l2_all)
271 * DCache_L2_Misses = d_ratio(dcache_l2_all_miss, dcache_l2_all)
273 * l2_rqsts.demand_data_rd_hit = 100
274 * l2_rqsts.pf_hit = 200
275 * l2_rqsts.rfo_hi = 300
276 * l2_rqsts.all_demand_data_rd = 400
277 * l2_rqsts.pf_miss = 500
278 * l2_rqsts.rfo_miss = 600
280 * DCache_L2_All_Hits = 600
281 * DCache_L2_All_Miss = MAX(400 - 100, 0) + 500 + 600 = 1400
282 * DCache_L2_All = 600 + 1400 = 2000
283 * DCache_L2_Hits = 600 / 2000 = 0.3
284 * DCache_L2_Misses = 1400 / 2000 = 0.7
286 static int test_dcache_l2(void)
289 struct value vals[] = {
290 { .event = "l2_rqsts.demand_data_rd_hit", .val = 100 },
291 { .event = "l2_rqsts.pf_hit", .val = 200 },
292 { .event = "l2_rqsts.rfo_hit", .val = 300 },
293 { .event = "l2_rqsts.all_demand_data_rd", .val = 400 },
294 { .event = "l2_rqsts.pf_miss", .val = 500 },
295 { .event = "l2_rqsts.rfo_miss", .val = 600 },
299 TEST_ASSERT_VAL("failed to compute metric",
300 compute_metric("DCache_L2_Hits", vals, &ratio) == 0);
302 TEST_ASSERT_VAL("DCache_L2_Hits failed, wrong ratio",
305 TEST_ASSERT_VAL("failed to compute metric",
306 compute_metric("DCache_L2_Misses", vals, &ratio) == 0);
308 TEST_ASSERT_VAL("DCache_L2_Misses failed, wrong ratio",
313 static int test_recursion_fail(void)
316 struct value vals[] = {
317 { .event = "inst_retired.any", .val = 300 },
318 { .event = "cpu_clk_unhalted.thread", .val = 200 },
322 TEST_ASSERT_VAL("failed to find recursion",
323 compute_metric("M1", vals, &ratio) == -1);
325 TEST_ASSERT_VAL("failed to find recursion",
326 compute_metric("M3", vals, &ratio) == -1);
330 static int test_memory_bandwidth(void)
333 struct value vals[] = {
334 { .event = "l1d.replacement", .val = 4000000 },
335 { .event = "duration_time", .val = 200000000 },
339 TEST_ASSERT_VAL("failed to compute metric",
340 compute_metric("L1D_Cache_Fill_BW", vals, &ratio) == 0);
341 TEST_ASSERT_VAL("L1D_Cache_Fill_BW, wrong ratio",
347 static int test_metric_group(void)
349 double ratio1, ratio2;
350 struct value vals[] = {
351 { .event = "cpu_clk_unhalted.thread", .val = 200 },
352 { .event = "l1d-loads-misses", .val = 300 },
353 { .event = "l1i-loads-misses", .val = 200 },
354 { .event = "inst_retired.any", .val = 400 },
358 TEST_ASSERT_VAL("failed to find recursion",
359 compute_metric_group("group1", vals,
361 "cache_miss_cycles", &ratio2) == 0);
363 TEST_ASSERT_VAL("group IPC failed, wrong ratio",
366 TEST_ASSERT_VAL("group cache_miss_cycles failed, wrong ratio",
371 int test__parse_metric(struct test *test __maybe_unused, int subtest __maybe_unused)
373 TEST_ASSERT_VAL("IPC failed", test_ipc() == 0);
374 TEST_ASSERT_VAL("frontend failed", test_frontend() == 0);
375 TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0);
376 TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0);
377 TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0);
378 TEST_ASSERT_VAL("test metric group", test_metric_group() == 0);
379 TEST_ASSERT_VAL("Memory bandwidth", test_memory_bandwidth() == 0);