tools/perf/tests/parse-metric.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/compiler.h>
   3 #include <string.h>
   4 #include <perf/cpumap.h>
   5 #include <perf/evlist.h>
   6 #include "metricgroup.h"
   7 #include "tests.h"
   8 #include "pmu-events/pmu-events.h"
   9 #include "evlist.h"
  10 #include "rblist.h"
  11 #include "debug.h"
  12 #include "expr.h"
  13 #include "stat.h"
  14 #include <perf/cpumap.h>
  15 #include <perf/evlist.h>
  16
  17 static struct pmu_event pme_test[] = {
  18 {
  19         .metric_expr    = "inst_retired.any / cpu_clk_unhalted.thread",
  20         .metric_name    = "IPC",
  21 },
  22 {
  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",
  26 },
  27 {
  28         .metric_expr    = "l1d\\-loads\\-misses / inst_retired.any",
  29         .metric_name    = "dcache_miss_cpi",
  30 },
  31 {
  32         .metric_expr    = "l1i\\-loads\\-misses / inst_retired.any",
  33         .metric_name    = "icache_miss_cycles",
  34 },
  35 {
  36         .metric_expr    = "(dcache_miss_cpi + icache_miss_cycles)",
  37         .metric_name    = "cache_miss_cycles",
  38 },
  39 {
  40         .metric_expr    = "l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit",
  41         .metric_name    = "DCache_L2_All_Hits",
  42 },
  43 {
  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",
  47 },
  48 {
  49         .metric_expr    = "dcache_l2_all_hits + dcache_l2_all_miss",
  50         .metric_name    = "DCache_L2_All",
  51 },
  52 {
  53         .metric_expr    = "d_ratio(dcache_l2_all_hits, dcache_l2_all)",
  54         .metric_name    = "DCache_L2_Hits",
  55 },
  56 {
  57         .metric_expr    = "d_ratio(dcache_l2_all_miss, dcache_l2_all)",
  58         .metric_name    = "DCache_L2_Misses",
  59 },
  60 {
  61         .metric_expr    = "ipc + m2",
  62         .metric_name    = "M1",
  63 },
  64 {
  65         .metric_expr    = "ipc + m1",
  66         .metric_name    = "M2",
  67 },
  68 {
  69         .metric_expr    = "1/m3",
  70         .metric_name    = "M3",
  71 }
  72 };
  73
  74 static struct pmu_events_map map = {
  75         .cpuid          = "test",
  76         .version        = "1",
  77         .type           = "core",
  78         .table          = pme_test,
  79 };
  80
  81 struct value {
  82         const char      *event;
  83         u64              val;
  84 };
  85
  86 static u64 find_value(const char *name, struct value *values)
  87 {
  88         struct value *v = values;
  89
  90         while (v->event) {
  91                 if (!strcmp(name, v->event))
  92                         return v->val;
  93                 v++;
  94         };
  95         return 0;
  96 }
  97
  98 static void load_runtime_stat(struct runtime_stat *st, struct evlist *evlist,
  99                               struct value *vals)
 100 {
 101         struct evsel *evsel;
 102         u64 count;
 103
 104         evlist__for_each_entry(evlist, evsel) {
 105                 count = find_value(evsel->name, vals);
 106                 perf_stat__update_shadow_stats(evsel, count, 0, st);
 107         }
 108 }
 109
 110 static double compute_single(struct rblist *metric_events, struct evlist *evlist,
 111                              struct runtime_stat *st)
 112 {
 113         struct evsel *evsel = evlist__first(evlist);
 114         struct metric_event *me;
 115
 116         me = metricgroup__lookup(metric_events, evsel, false);
 117         if (me != NULL) {
 118                 struct metric_expr *mexp;
 119
 120                 mexp = list_first_entry(&me->head, struct metric_expr, nd);
 121                 return test_generic_metric(mexp, 0, st);
 122         }
 123         return 0.;
 124 }
 125
 126 static int compute_metric(const char *name, struct value *vals, double *ratio)
 127 {
 128         struct rblist metric_events = {
 129                 .nr_entries = 0,
 130         };
 131         struct perf_cpu_map *cpus;
 132         struct runtime_stat st;
 133         struct evlist *evlist;
 134         int err;
 135
 136         /*
 137          * We need to prepare evlist for stat mode running on CPU 0
 138          * because that's where all the stats are going to be created.
 139          */
 140         evlist = evlist__new();
 141         if (!evlist)
 142                 return -ENOMEM;
 143
 144         cpus = perf_cpu_map__new("0");
 145         if (!cpus)
 146                 return -ENOMEM;
 147
 148         perf_evlist__set_maps(&evlist->core, cpus, NULL);
 149
 150         /* Parse the metric into metric_events list. */
 151         err = metricgroup__parse_groups_test(evlist, &map, name,
 152                                              false, false,
 153                                              &metric_events);
 154         if (err)
 155                 return err;
 156
 157         if (perf_evlist__alloc_stats(evlist, false))
 158                 return -1;
 159
 160         /* Load the runtime stats with given numbers for events. */
 161         runtime_stat__init(&st);
 162         load_runtime_stat(&st, evlist, vals);
 163
 164         /* And execute the metric */
 165         *ratio = compute_single(&metric_events, evlist, &st);
 166
 167         /* ... clenup. */
 168         metricgroup__rblist_exit(&metric_events);
 169         runtime_stat__exit(&st);
 170         perf_evlist__free_stats(evlist);
 171         perf_cpu_map__put(cpus);
 172         evlist__delete(evlist);
 173         return 0;
 174 }
 175
 176 static int test_ipc(void)
 177 {
 178         double ratio;
 179         struct value vals[] = {
 180                 { .event = "inst_retired.any",        .val = 300 },
 181                 { .event = "cpu_clk_unhalted.thread", .val = 200 },
 182                 { .event = NULL, },
 183         };
 184
 185         TEST_ASSERT_VAL("failed to compute metric",
 186                         compute_metric("IPC", vals, &ratio) == 0);
 187
 188         TEST_ASSERT_VAL("IPC failed, wrong ratio",
 189                         ratio == 1.5);
 190         return 0;
 191 }
 192
 193 static int test_frontend(void)
 194 {
 195         double ratio;
 196         struct value vals[] = {
 197                 { .event = "idq_uops_not_delivered.core",        .val = 300 },
 198                 { .event = "cpu_clk_unhalted.thread",            .val = 200 },
 199                 { .event = "cpu_clk_unhalted.one_thread_active", .val = 400 },
 200                 { .event = "cpu_clk_unhalted.ref_xclk",          .val = 600 },
 201                 { .event = NULL, },
 202         };
 203
 204         TEST_ASSERT_VAL("failed to compute metric",
 205                         compute_metric("Frontend_Bound_SMT", vals, &ratio) == 0);
 206
 207         TEST_ASSERT_VAL("Frontend_Bound_SMT failed, wrong ratio",
 208                         ratio == 0.45);
 209         return 0;
 210 }
 211
 212 static int test_cache_miss_cycles(void)
 213 {
 214         double ratio;
 215         struct value vals[] = {
 216                 { .event = "l1d-loads-misses",  .val = 300 },
 217                 { .event = "l1i-loads-misses",  .val = 200 },
 218                 { .event = "inst_retired.any",  .val = 400 },
 219                 { .event = NULL, },
 220         };
 221
 222         TEST_ASSERT_VAL("failed to compute metric",
 223                         compute_metric("cache_miss_cycles", vals, &ratio) == 0);
 224
 225         TEST_ASSERT_VAL("cache_miss_cycles failed, wrong ratio",
 226                         ratio == 1.25);
 227         return 0;
 228 }
 229
 230
 231 /*
 232  * DCache_L2_All_Hits = l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hi
 233  * DCache_L2_All_Miss = max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) +
 234  *                      l2_rqsts.pf_miss + l2_rqsts.rfo_miss
 235  * DCache_L2_All      = dcache_l2_all_hits + dcache_l2_all_miss
 236  * DCache_L2_Hits     = d_ratio(dcache_l2_all_hits, dcache_l2_all)
 237  * DCache_L2_Misses   = d_ratio(dcache_l2_all_miss, dcache_l2_all)
 238  *
 239  * l2_rqsts.demand_data_rd_hit = 100
 240  * l2_rqsts.pf_hit             = 200
 241  * l2_rqsts.rfo_hi             = 300
 242  * l2_rqsts.all_demand_data_rd = 400
 243  * l2_rqsts.pf_miss            = 500
 244  * l2_rqsts.rfo_miss           = 600
 245  *
 246  * DCache_L2_All_Hits = 600
 247  * DCache_L2_All_Miss = MAX(400 - 100, 0) + 500 + 600 = 1400
 248  * DCache_L2_All      = 600 + 1400  = 2000
 249  * DCache_L2_Hits     = 600 / 2000  = 0.3
 250  * DCache_L2_Misses   = 1400 / 2000 = 0.7
 251  */
 252 static int test_dcache_l2(void)
 253 {
 254         double ratio;
 255         struct value vals[] = {
 256                 { .event = "l2_rqsts.demand_data_rd_hit", .val = 100 },
 257                 { .event = "l2_rqsts.pf_hit",             .val = 200 },
 258                 { .event = "l2_rqsts.rfo_hit",            .val = 300 },
 259                 { .event = "l2_rqsts.all_demand_data_rd", .val = 400 },
 260                 { .event = "l2_rqsts.pf_miss",            .val = 500 },
 261                 { .event = "l2_rqsts.rfo_miss",           .val = 600 },
 262                 { .event = NULL, },
 263         };
 264
 265         TEST_ASSERT_VAL("failed to compute metric",
 266                         compute_metric("DCache_L2_Hits", vals, &ratio) == 0);
 267
 268         TEST_ASSERT_VAL("DCache_L2_Hits failed, wrong ratio",
 269                         ratio == 0.3);
 270
 271         TEST_ASSERT_VAL("failed to compute metric",
 272                         compute_metric("DCache_L2_Misses", vals, &ratio) == 0);
 273
 274         TEST_ASSERT_VAL("DCache_L2_Misses failed, wrong ratio",
 275                         ratio == 0.7);
 276         return 0;
 277 }
 278
 279 static int test_recursion_fail(void)
 280 {
 281         double ratio;
 282         struct value vals[] = {
 283                 { .event = "inst_retired.any",        .val = 300 },
 284                 { .event = "cpu_clk_unhalted.thread", .val = 200 },
 285                 { .event = NULL, },
 286         };
 287
 288         TEST_ASSERT_VAL("failed to find recursion",
 289                         compute_metric("M1", vals, &ratio) == -1);
 290
 291         TEST_ASSERT_VAL("failed to find recursion",
 292                         compute_metric("M3", vals, &ratio) == -1);
 293         return 0;
 294 }
 295
 296 int test__parse_metric(struct test *test __maybe_unused, int subtest __maybe_unused)
 297 {
 298         TEST_ASSERT_VAL("IPC failed", test_ipc() == 0);
 299         TEST_ASSERT_VAL("frontend failed", test_frontend() == 0);
 300         TEST_ASSERT_VAL("cache_miss_cycles failed", test_cache_miss_cycles() == 0);
 301         TEST_ASSERT_VAL("DCache_L2 failed", test_dcache_l2() == 0);
 302         TEST_ASSERT_VAL("recursion fail failed", test_recursion_fail() == 0);
 303         return 0;
 304 }