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"
16 static struct pmu_event pme_test[] = {
18 .metric_expr = "inst_retired.any / cpu_clk_unhalted.thread",
20 .metric_group = "group1",
23 .metric_expr = "idq_uops_not_delivered.core / (4 * (( ( cpu_clk_unhalted.thread / 2 ) * "
24 "( 1 + cpu_clk_unhalted.one_thread_active / cpu_clk_unhalted.ref_xclk ) )))",
25 .metric_name = "Frontend_Bound_SMT",
28 .metric_expr = "l1d\\-loads\\-misses / inst_retired.any",
29 .metric_name = "dcache_miss_cpi",
32 .metric_expr = "l1i\\-loads\\-misses / inst_retired.any",
33 .metric_name = "icache_miss_cycles",
36 .metric_expr = "(dcache_miss_cpi + icache_miss_cycles)",
37 .metric_name = "cache_miss_cycles",
38 .metric_group = "group1",
41 .metric_expr = "l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit",
42 .metric_name = "DCache_L2_All_Hits",
45 .metric_expr = "max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + "
46 "l2_rqsts.pf_miss + l2_rqsts.rfo_miss",
47 .metric_name = "DCache_L2_All_Miss",
50 .metric_expr = "dcache_l2_all_hits + dcache_l2_all_miss",
51 .metric_name = "DCache_L2_All",
54 .metric_expr = "d_ratio(dcache_l2_all_hits, dcache_l2_all)",
55 .metric_name = "DCache_L2_Hits",
58 .metric_expr = "d_ratio(dcache_l2_all_miss, dcache_l2_all)",
59 .metric_name = "DCache_L2_Misses",
62 .metric_expr = "ipc + m2",
66 .metric_expr = "ipc + m1",
70 .metric_expr = "1/m3",
74 .metric_expr = "64 * l1d.replacement / 1000000000 / duration_time",
75 .metric_name = "L1D_Cache_Fill_BW",
82 static struct pmu_events_map map = {
94 static u64 find_value(const char *name, struct value *values)
96 struct value *v = values;
99 if (!strcmp(name, v->event))
106 static void load_runtime_stat(struct runtime_stat *st, struct evlist *evlist,
112 evlist__for_each_entry(evlist, evsel) {
113 count = find_value(evsel->name, vals);
114 perf_stat__update_shadow_stats(evsel, count, 0, st);
115 if (!strcmp(evsel->name, "duration_time"))
116 update_stats(&walltime_nsecs_stats, count);
120 static double compute_single(struct rblist *metric_events, struct evlist *evlist,
121 struct runtime_stat *st, const char *name)
123 struct metric_expr *mexp;
124 struct metric_event *me;
127 evlist__for_each_entry(evlist, evsel) {
128 me = metricgroup__lookup(metric_events, evsel, false);
130 list_for_each_entry (mexp, &me->head, nd) {
131 if (strcmp(mexp->metric_name, name))
133 return test_generic_metric(mexp, 0, st);
140 static int __compute_metric(const char *name, struct value *vals,
141 const char *name1, double *ratio1,
142 const char *name2, double *ratio2)
144 struct rblist metric_events = {
147 struct perf_cpu_map *cpus;
148 struct runtime_stat st;
149 struct evlist *evlist;
153 * We need to prepare evlist for stat mode running on CPU 0
154 * because that's where all the stats are going to be created.
156 evlist = evlist__new();
160 cpus = perf_cpu_map__new("0");
162 evlist__delete(evlist);
166 perf_evlist__set_maps(&evlist->core, cpus, NULL);
167 runtime_stat__init(&st);
169 /* Parse the metric into metric_events list. */
170 err = metricgroup__parse_groups_test(evlist, &map, name,
176 err = evlist__alloc_stats(evlist, false);
180 /* Load the runtime stats with given numbers for events. */
181 load_runtime_stat(&st, evlist, vals);
183 /* And execute the metric */
185 *ratio1 = compute_single(&metric_events, evlist, &st, name1);
187 *ratio2 = compute_single(&metric_events, evlist, &st, name2);
191 metricgroup__rblist_exit(&metric_events);
192 runtime_stat__exit(&st);
193 evlist__free_stats(evlist);
194 perf_cpu_map__put(cpus);
195 evlist__delete(evlist);
199 static int compute_metric(const char *name, struct value *vals, double *ratio)
201 return __compute_metric(name, vals, name, ratio, NULL, NULL);
204 static int compute_metric_group(const char *name, struct value *vals,
205 const char *name1, double *ratio1,
206 const char *name2, double *ratio2)
208 return __compute_metric(name, vals, name1, ratio1, name2, ratio2);
211 static int test_ipc(void)
214 struct value vals[] = {
215 { .event = "inst_retired.any", .val = 300 },
216 { .event = "cpu_clk_unhalted.thread", .val = 200 },
220 TEST_ASSERT_VAL("failed to compute metric",
221 compute_metric("IPC", vals, &ratio) == 0);
223 TEST_ASSERT_VAL("IPC failed, wrong ratio",
228 static int test_frontend(void)
231 struct value vals[] = {
232 { .event = "idq_uops_not_delivered.core", .val = 300 },
233 { .event = "cpu_clk_unhalted.thread", .val = 200 },
234 { .event = "cpu_clk_unhalted.one_thread_active", .val = 400 },
235 { .event = "cpu_clk_unhalted.ref_xclk", .val = 600 },
239 TEST_ASSERT_VAL("failed to compute metric",
240 compute_metric("Frontend_Bound_SMT", vals, &ratio) == 0);
242 TEST_ASSERT_VAL("Frontend_Bound_SMT failed, wrong ratio",
247 static int test_cache_miss_cycles(void)
250 struct value vals[] = {
251 { .event = "l1d-loads-misses", .val = 300 },
252 { .event = "l1i-loads-misses", .val = 200 },
253 { .event = "inst_retired.any", .val = 400 },
257 TEST_ASSERT_VAL("failed to compute metric",
258 compute_metric("cache_miss_cycles", vals, &ratio) == 0);
260 TEST_ASSERT_VAL("cache_miss_cycles failed, wrong ratio",
267 * DCache_L2_All_Hits = l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hi
268 * DCache_L2_All_Miss = max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) +
269 * l2_rqsts.pf_miss + l2_rqsts.rfo_miss
270 * DCache_L2_All = dcache_l2_all_hits + dcache_l2_all_miss
271 * DCache_L2_Hits = d_ratio(dcache_l2_all_hits, dcache_l2_all)
272 * DCache_L2_Misses = d_ratio(dcache_l2_all_miss, dcache_l2_all)
274 * l2_rqsts.demand_data_rd_hit = 100
275 * l2_rqsts.pf_hit = 200
276 * l2_rqsts.rfo_hi = 300
277 * l2_rqsts.all_demand_data_rd = 400
278 * l2_rqsts.pf_miss = 500
279 * l2_rqsts.rfo_miss = 600
281 * DCache_L2_All_Hits = 600
282 * DCache_L2_All_Miss = MAX(400 - 100, 0) + 500 + 600 = 1400
283 * DCache_L2_All = 600 + 1400 = 2000
284 * DCache_L2_Hits = 600 / 2000 = 0.3
285 * DCache_L2_Misses = 1400 / 2000 = 0.7
287 static int test_dcache_l2(void)
290 struct value vals[] = {
291 { .event = "l2_rqsts.demand_data_rd_hit", .val = 100 },
292 { .event = "l2_rqsts.pf_hit", .val = 200 },
293 { .event = "l2_rqsts.rfo_hit", .val = 300 },
294 { .event = "l2_rqsts.all_demand_data_rd", .val = 400 },
295 { .event = "l2_rqsts.pf_miss", .val = 500 },
296 { .event = "l2_rqsts.rfo_miss", .val = 600 },
300 TEST_ASSERT_VAL("failed to compute metric",
301 compute_metric("DCache_L2_Hits", vals, &ratio) == 0);
303 TEST_ASSERT_VAL("DCache_L2_Hits failed, wrong ratio",
306 TEST_ASSERT_VAL("failed to compute metric",
307 compute_metric("DCache_L2_Misses", vals, &ratio) == 0);
309 TEST_ASSERT_VAL("DCache_L2_Misses failed, wrong ratio",
314 static int test_recursion_fail(void)
317 struct value vals[] = {
318 { .event = "inst_retired.any", .val = 300 },
319 { .event = "cpu_clk_unhalted.thread", .val = 200 },
323 TEST_ASSERT_VAL("failed to find recursion",
324 compute_metric("M1", vals, &ratio) == -1);
326 TEST_ASSERT_VAL("failed to find recursion",
327 compute_metric("M3", vals, &ratio) == -1);
331 static int test_memory_bandwidth(void)
334 struct value vals[] = {
335 { .event = "l1d.replacement", .val = 4000000 },
336 { .event = "duration_time", .val = 200000000 },
340 TEST_ASSERT_VAL("failed to compute metric",
341 compute_metric("L1D_Cache_Fill_BW", vals, &ratio) == 0);
342 TEST_ASSERT_VAL("L1D_Cache_Fill_BW, wrong ratio",
348 static int test_metric_group(void)
350 double ratio1, ratio2;
351 struct value vals[] = {
352 { .event = "cpu_clk_unhalted.thread", .val = 200 },
353 { .event = "l1d-loads-misses", .val = 300 },
354 { .event = "l1i-loads-misses", .val = 200 },
355 { .event = "inst_retired.any", .val = 400 },
359 TEST_ASSERT_VAL("failed to find recursion",
360 compute_metric_group("group1", vals,
362 "cache_miss_cycles", &ratio2) == 0);
364 TEST_ASSERT_VAL("group IPC failed, wrong ratio",
367 TEST_ASSERT_VAL("group cache_miss_cycles failed, wrong ratio",
372 int test__parse_metric(struct test *test __maybe_unused, int subtest __maybe_unused)
374 TEST_ASSERT_VAL("IPC failed", test_ipc() == 0);
375 TEST_ASSERT_VAL("frontend failed", test_frontend() == 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("Memory bandwidth", test_memory_bandwidth() == 0);
380 if (!perf_pmu__has_hybrid()) {
381 TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0);
382 TEST_ASSERT_VAL("test metric group", test_metric_group() == 0);