src/gallium/drivers/iris/iris_monitor.c

   1 /*
   2  * Copyright © 2019 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included
  12  * in all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  15  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  18  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  19  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  20  * DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #include "iris_monitor.h"
  24
  25 #include <xf86drm.h>
  26
  27 #include "iris_screen.h"
  28 #include "iris_context.h"
  29 #include "iris_perf.h"
  30
  31 struct iris_monitor_object {
  32    int num_active_counters;
  33    int *active_counters;
  34
  35    size_t result_size;
  36    unsigned char *result_buffer;
  37
  38    struct gen_perf_query_object *query;
  39 };
  40
  41 int
  42 iris_get_monitor_info(struct pipe_screen *pscreen, unsigned index,
  43                       struct pipe_driver_query_info *info)
  44 {
  45    const struct iris_screen *screen = (struct iris_screen *)pscreen;
  46    assert(screen->monitor_cfg);
  47    if (!screen->monitor_cfg)
  48       return 0;
  49
  50    const struct iris_monitor_config *monitor_cfg = screen->monitor_cfg;
  51
  52    if (!info) {
  53       /* return the number of metrics */
  54       return monitor_cfg->num_counters;
  55    }
  56
  57    const struct gen_perf_config *perf_cfg = monitor_cfg->perf_cfg;
  58    const int group = monitor_cfg->counters[index].group;
  59    const int counter_index = monitor_cfg->counters[index].counter;
  60    struct gen_perf_query_counter *counter =
  61       &perf_cfg->queries[group].counters[counter_index];
  62
  63    info->group_id = group;
  64    info->name = counter->name;
  65    info->query_type = PIPE_QUERY_DRIVER_SPECIFIC + index;
  66
  67    if (counter->type == GEN_PERF_COUNTER_TYPE_THROUGHPUT)
  68       info->result_type = PIPE_DRIVER_QUERY_RESULT_TYPE_AVERAGE;
  69    else
  70       info->result_type = PIPE_DRIVER_QUERY_RESULT_TYPE_CUMULATIVE;
  71    switch (counter->data_type) {
  72    case GEN_PERF_COUNTER_DATA_TYPE_BOOL32:
  73    case GEN_PERF_COUNTER_DATA_TYPE_UINT32:
  74       info->type = PIPE_DRIVER_QUERY_TYPE_UINT;
  75       info->max_value.u32 = 0;
  76       break;
  77    case GEN_PERF_COUNTER_DATA_TYPE_UINT64:
  78       info->type = PIPE_DRIVER_QUERY_TYPE_UINT64;
  79       info->max_value.u64 = 0;
  80       break;
  81    case GEN_PERF_COUNTER_DATA_TYPE_FLOAT:
  82    case GEN_PERF_COUNTER_DATA_TYPE_DOUBLE:
  83       info->type = PIPE_DRIVER_QUERY_TYPE_FLOAT;
  84       info->max_value.u64 = -1;
  85       break;
  86    default:
  87       assert(false);
  88       break;
  89    }
  90
  91    /* indicates that this is an OA query, not a pipeline statistics query */
  92    info->flags = PIPE_DRIVER_QUERY_FLAG_BATCH;
  93    return 1;
  94 }
  95
  96 static bool
  97 iris_monitor_init_metrics(struct iris_screen *screen)
  98 {
  99    struct iris_monitor_config *monitor_cfg =
 100       rzalloc(screen, struct iris_monitor_config);
 101    struct gen_perf_config *perf_cfg = NULL;
 102    if (unlikely(!monitor_cfg))
 103       goto allocation_error;
 104    perf_cfg = gen_perf_new(monitor_cfg);
 105    if (unlikely(!perf_cfg))
 106       goto allocation_error;
 107
 108    monitor_cfg->perf_cfg = perf_cfg;
 109
 110    iris_perf_init_vtbl(perf_cfg);
 111
 112    gen_perf_init_metrics(perf_cfg, &screen->devinfo, screen->fd);
 113    screen->monitor_cfg = monitor_cfg;
 114
 115    /* a gallium "group" is equivalent to a gen "query"
 116     * a gallium "query" is equivalent to a gen "query_counter"
 117     *
 118     * Each gen_query supports a specific number of query_counters.  To
 119     * allocate the array of iris_monitor_counter, we need an upper bound
 120     * (ignoring duplicate query_counters).
 121     */
 122    int gen_query_counters_count = 0;
 123    for (int gen_query_id = 0;
 124         gen_query_id < perf_cfg->n_queries;
 125         ++gen_query_id) {
 126       gen_query_counters_count += perf_cfg->queries[gen_query_id].n_counters;
 127    }
 128
 129    monitor_cfg->counters = rzalloc_size(monitor_cfg,
 130                                         sizeof(struct iris_monitor_counter) *
 131                                         gen_query_counters_count);
 132    if (unlikely(!monitor_cfg->counters))
 133       goto allocation_error;
 134
 135    int iris_monitor_id = 0;
 136    for (int group = 0; group < perf_cfg->n_queries; ++group) {
 137       for (int counter = 0;
 138            counter < perf_cfg->queries[group].n_counters;
 139            ++counter) {
 140          /* Check previously identified metrics to filter out duplicates. The
 141           * user is not helped by having the same metric available in several
 142           * groups. (n^2 algorithm).
 143           */
 144          bool duplicate = false;
 145          for (int existing_group = 0;
 146               existing_group < group && !duplicate;
 147               ++existing_group) {
 148             for (int existing_counter = 0;
 149                  existing_counter < perf_cfg->queries[existing_group].n_counters && !duplicate;
 150                  ++existing_counter) {
 151                const char *current_name =
 152                   perf_cfg->queries[group].counters[counter].name;
 153                const char *existing_name =
 154                   perf_cfg->queries[existing_group].counters[existing_counter].name;
 155                if (strcmp(current_name, existing_name) == 0) {
 156                   duplicate = true;
 157                }
 158             }
 159          }
 160          if (duplicate)
 161             continue;
 162          monitor_cfg->counters[iris_monitor_id].group = group;
 163          monitor_cfg->counters[iris_monitor_id].counter = counter;
 164          ++iris_monitor_id;
 165       }
 166    }
 167    monitor_cfg->num_counters = iris_monitor_id;
 168    return monitor_cfg->num_counters;
 169
 170 allocation_error:
 171    if (monitor_cfg)
 172       free(monitor_cfg->counters);
 173    free(perf_cfg);
 174    free(monitor_cfg);
 175    return false;
 176 }
 177
 178 int
 179 iris_get_monitor_group_info(struct pipe_screen *pscreen,
 180                             unsigned group_index,
 181                             struct pipe_driver_query_group_info *info)
 182 {
 183    struct iris_screen *screen = (struct iris_screen *)pscreen;
 184    if (!screen->monitor_cfg) {
 185       if (!iris_monitor_init_metrics(screen))
 186          return 0;
 187    }
 188
 189    const struct iris_monitor_config *monitor_cfg = screen->monitor_cfg;
 190    const struct gen_perf_config *perf_cfg = monitor_cfg->perf_cfg;
 191
 192    if (!info) {
 193       /* return the count that can be queried */
 194       return perf_cfg->n_queries;
 195    }
 196
 197    if (group_index >= perf_cfg->n_queries) {
 198       /* out of range */
 199       return 0;
 200    }
 201
 202    struct gen_perf_query_info *query = &perf_cfg->queries[group_index];
 203
 204    info->name = query->name;
 205    info->max_active_queries = query->n_counters;
 206    info->num_queries = query->n_counters;
 207
 208    return 1;
 209 }
 210
 211 static void
 212 iris_init_monitor_ctx(struct iris_context *ice)
 213 {
 214    struct iris_screen *screen = (struct iris_screen *) ice->ctx.screen;
 215    struct iris_monitor_config *monitor_cfg = screen->monitor_cfg;
 216
 217    ice->perf_ctx = gen_perf_new_context(ice);
 218    if (unlikely(!ice->perf_ctx))
 219       return;
 220
 221    struct gen_perf_context *perf_ctx = ice->perf_ctx;
 222    struct gen_perf_config *perf_cfg = monitor_cfg->perf_cfg;
 223    gen_perf_init_context(perf_ctx,
 224                          perf_cfg,
 225                          ice,
 226                          screen->bufmgr,
 227                          &screen->devinfo,
 228                          ice->batches[IRIS_BATCH_RENDER].hw_ctx_id,
 229                          screen->fd);
 230 }
 231
 232 /* entry point for GenPerfMonitorsAMD */
 233 struct iris_monitor_object *
 234 iris_create_monitor_object(struct iris_context *ice,
 235                            unsigned num_queries,
 236                            unsigned *query_types)
 237 {
 238    struct iris_screen *screen = (struct iris_screen *) ice->ctx.screen;
 239    struct iris_monitor_config *monitor_cfg = screen->monitor_cfg;
 240    struct gen_perf_config *perf_cfg = monitor_cfg->perf_cfg;
 241    struct gen_perf_query_object *query_obj = NULL;
 242
 243    /* initialize perf context if this has not already been done.  This
 244     * function is the first entry point that carries the gl context.
 245     */
 246    if (ice->perf_ctx == NULL) {
 247       iris_init_monitor_ctx(ice);
 248    }
 249    struct gen_perf_context *perf_ctx = ice->perf_ctx;
 250
 251    assert(num_queries > 0);
 252    int query_index = query_types[0] - PIPE_QUERY_DRIVER_SPECIFIC;
 253    assert(query_index <= monitor_cfg->num_counters);
 254    const int group = monitor_cfg->counters[query_index].group;
 255
 256    struct iris_monitor_object *monitor =
 257       calloc(1, sizeof(struct iris_monitor_object));
 258    if (unlikely(!monitor))
 259       goto allocation_failure;
 260
 261    monitor->num_active_counters = num_queries;
 262    monitor->active_counters = calloc(num_queries, sizeof(int));
 263    if (unlikely(!monitor->active_counters))
 264       goto allocation_failure;
 265
 266    for (int i = 0; i < num_queries; ++i) {
 267       unsigned current_query = query_types[i];
 268       unsigned current_query_index = current_query - PIPE_QUERY_DRIVER_SPECIFIC;
 269
 270       /* all queries must be in the same group */
 271       assert(current_query_index <= monitor_cfg->num_counters);
 272       assert(monitor_cfg->counters[current_query_index].group == group);
 273       monitor->active_counters[i] =
 274          monitor_cfg->counters[current_query_index].counter;
 275    }
 276
 277    /* create the gen_perf_query */
 278    query_obj = gen_perf_new_query(perf_ctx, group);
 279    if (unlikely(!query_obj))
 280       goto allocation_failure;
 281
 282    monitor->query = query_obj;
 283    monitor->result_size = perf_cfg->queries[group].data_size;
 284    monitor->result_buffer = calloc(1, monitor->result_size);
 285    if (unlikely(!monitor->result_buffer))
 286       goto allocation_failure;
 287
 288    return monitor;
 289
 290 allocation_failure:
 291    if (monitor) {
 292       free(monitor->active_counters);
 293       free(monitor->result_buffer);
 294    }
 295    free(query_obj);
 296    free(monitor);
 297    return NULL;
 298 }
 299
 300 void
 301 iris_destroy_monitor_object(struct pipe_context *ctx,
 302                             struct iris_monitor_object *monitor)
 303 {
 304    struct iris_context *ice = (struct iris_context *)ctx;
 305
 306    gen_perf_delete_query(ice->perf_ctx, monitor->query);
 307    free(monitor->result_buffer);
 308    monitor->result_buffer = NULL;
 309    free(monitor->active_counters);
 310    monitor->active_counters = NULL;
 311    free(monitor);
 312 }
 313
 314 bool
 315 iris_begin_monitor(struct pipe_context *ctx,
 316                    struct iris_monitor_object *monitor)
 317 {
 318    struct iris_context *ice = (void *) ctx;
 319    struct gen_perf_context *perf_ctx = ice->perf_ctx;
 320
 321    return gen_perf_begin_query(perf_ctx, monitor->query);
 322 }
 323
 324 bool
 325 iris_end_monitor(struct pipe_context *ctx,
 326                  struct iris_monitor_object *monitor)
 327 {
 328    struct iris_context *ice = (void *) ctx;
 329    struct gen_perf_context *perf_ctx = ice->perf_ctx;
 330
 331    gen_perf_end_query(perf_ctx, monitor->query);
 332    return true;
 333 }
 334
 335 bool
 336 iris_get_monitor_result(struct pipe_context *ctx,
 337                         struct iris_monitor_object *monitor,
 338                         bool wait,
 339                         union pipe_numeric_type_union *result)
 340 {
 341    struct iris_context *ice = (void *) ctx;
 342    struct gen_perf_context *perf_ctx = ice->perf_ctx;
 343    struct iris_batch *batch = &ice->batches[IRIS_BATCH_RENDER];
 344
 345    bool monitor_ready =
 346       gen_perf_is_query_ready(perf_ctx, monitor->query, batch);
 347
 348    if (!monitor_ready) {
 349       if (!wait)
 350          return false;
 351       gen_perf_wait_query(perf_ctx, monitor->query, batch);
 352    }
 353
 354    assert(gen_perf_is_query_ready(perf_ctx, monitor->query, batch));
 355
 356    unsigned bytes_written;
 357    gen_perf_get_query_data(perf_ctx, monitor->query,
 358                            monitor->result_size,
 359                            (unsigned*) monitor->result_buffer,
 360                            &bytes_written);
 361    if (bytes_written != monitor->result_size)
 362       return false;
 363
 364    /* copy metrics into the batch result */
 365    for (int i = 0; i < monitor->num_active_counters; ++i) {
 366       int current_counter = monitor->active_counters[i];
 367       const struct gen_perf_query_info *info =
 368          gen_perf_query_info(monitor->query);
 369       const struct gen_perf_query_counter *counter =
 370          &info->counters[current_counter];
 371       assert(gen_perf_query_counter_get_size(counter));
 372       switch (counter->data_type) {
 373       case GEN_PERF_COUNTER_DATA_TYPE_UINT64:
 374          result[i].u64 = *(uint64_t*)(monitor->result_buffer + counter->offset);
 375          break;
 376       case GEN_PERF_COUNTER_DATA_TYPE_FLOAT:
 377          result[i].f = *(float*)(monitor->result_buffer + counter->offset);
 378          break;
 379       case GEN_PERF_COUNTER_DATA_TYPE_UINT32:
 380       case GEN_PERF_COUNTER_DATA_TYPE_BOOL32:
 381          result[i].u64 = *(uint32_t*)(monitor->result_buffer + counter->offset);
 382          break;
 383       case GEN_PERF_COUNTER_DATA_TYPE_DOUBLE: {
 384          double v = *(double*)(monitor->result_buffer + counter->offset);
 385          result[i].f = v;
 386          break;
 387       }
 388       default:
 389          unreachable("unexpected counter data type");
 390       }
 391    }
 392    return true;
 393 }