git.cworth.org Git - fips/blob - metrics.c

   1 /* Copyright © 2013, Intel Corporation
   2  *
   3  * Permission is hereby granted, free of charge, to any person obtaining a copy
   4  * of this software and associated documentation files (the "Software"), to deal
   5  * in the Software without restriction, including without limitation the rights
   6  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
   7  * copies of the Software, and to permit persons to whom the Software is
   8  * furnished to do so, subject to the following conditions:
   9  *
  10  * The above copyright notice and this permission notice shall be included in
  11  * all copies or substantial portions of the Software.
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  16  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  17  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  18  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  19  * THE SOFTWARE.
  20  */
  21
  22 #define _GNU_SOURCE
  23
  24 #include <stdio.h>
  25 #include <stdlib.h>
  26 #include <assert.h>
  27 #include <sys/time.h>
  28
  29 #include "fips-dispatch-gl.h"
  30
  31 #include "metrics.h"
  32 #include "context.h"
  33 #include "metrics-info.h"
  34 #include "xmalloc.h"
  35
  36 int frames;
  37 int verbose;
  38
  39 #define MAX_MONITORS_IN_FLIGHT 1000
  40
  41 /* Timer query */
  42 typedef struct timer_query
  43 {
  44         unsigned id;
  45
  46         metrics_op_t op;
  47         struct timer_query *next;
  48 } timer_query_t;
  49
  50 /* Performance-monitor query */
  51 typedef struct monitor
  52 {
  53         unsigned id;
  54
  55         metrics_op_t op;
  56         struct monitor *next;
  57 } monitor_t;
  58
  59 typedef struct op_metrics
  60 {
  61         /* This happens to also be the index into the
  62          * ctx->op_metrics array currently
  63          */
  64         metrics_op_t op;
  65         double time_ns;
  66
  67         double **counters;
  68 } op_metrics_t;
  69
  70 struct metrics
  71 {
  72         metrics_op_t op;
  73
  74         /* GL_TIME_ELAPSED query for which glEndQuery has not yet
  75          * been called. */
  76         unsigned timer_begun_id;
  77
  78         /* GL_TIME_ELAPSED queries for which glEndQuery has been
  79          * called, (but results have not yet been queried). */
  80         timer_query_t *timer_head;
  81         timer_query_t *timer_tail;
  82
  83         /* Performance monitor for which glEndPerfMonitorAMD has not
  84          * yet been called. */
  85         unsigned monitor_begun_id;
  86
  87         /* Performance monitors for which glEndPerfMonitorAMD has
  88          * been called, (but results have not yet been queried). */
  89         monitor_t *monitor_head;
  90         monitor_t *monitor_tail;
  91
  92         int monitors_in_flight;
  93
  94         unsigned num_op_metrics;
  95         op_metrics_t *op_metrics;
  96 };
  97
  98 static const char *
  99 metrics_op_string (metrics_op_t op)
 100 {
 101         if (op >= METRICS_OP_SHADER)
 102                 return "Shader program";
 103
 104         switch (op)
 105         {
 106         case METRICS_OP_ACCUM:
 107                 return "glAccum*(+)";
 108         case METRICS_OP_BUFFER_DATA:
 109                 return "glBufferData(+)";
 110         case METRICS_OP_BUFFER_SUB_DATA:
 111                 return "glCopyBufferSubData*";
 112         case METRICS_OP_BITMAP:
 113                 return "glBitmap*";
 114         case METRICS_OP_BLIT_FRAMEBUFFER:
 115                 return "glBlitFramebuffer*";
 116         case METRICS_OP_CLEAR:
 117                 return "glClear(+)";
 118         case METRICS_OP_CLEAR_BUFFER_DATA:
 119                 return "glCearBufferData(+)";
 120         case METRICS_OP_CLEAR_TEX_IMAGE:
 121                 return "glClearTexImage(+)";
 122         case METRICS_OP_COPY_PIXELS:
 123                 return "glCopyPixels";
 124         case METRICS_OP_COPY_TEX_IMAGE:
 125                 return "glCopyTexImage(+)";
 126         case METRICS_OP_DRAW_PIXELS:
 127                 return "glDrawPixels";
 128         case METRICS_OP_GET_TEX_IMAGE:
 129                 return "glGetTexImage(+)";
 130         case METRICS_OP_READ_PIXELS:
 131                 return "glReadPixels*";
 132         case METRICS_OP_TEX_IMAGE:
 133                 return "glTexImage*(+)";
 134         default:
 135                 fprintf (stderr, "fips: Internal error: "
 136                          "Unknown metrics op value: %d\n", op);
 137                 exit (1);
 138         }
 139
 140         return "";
 141 }
 142
 143 void
 144 metrics_counter_start (void)
 145 {
 146         context_t *ctx = context_get_current ();
 147         metrics_t *metrics = ctx->metrics;
 148         unsigned i;
 149
 150         /* Initialize the timer_query and monitor objects */
 151         glGenQueries (1, &metrics->timer_begun_id);
 152
 153         glGenPerfMonitorsAMD (1, &metrics->monitor_begun_id);
 154
 155         for (i = 0; i < ctx->metrics_info.num_groups; i++)
 156         {
 157                 metrics_group_info_t *group;
 158                 int num_counters;
 159
 160                 group = &ctx->metrics_info.groups[i];
 161
 162                 num_counters = group->num_counters;
 163                 if (group->max_active_counters < group->num_counters)
 164                 {
 165                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 166                                  group->max_active_counters,
 167                                  group->num_counters, i);
 168                         num_counters = group->max_active_counters;
 169
 170                 }
 171
 172                 glSelectPerfMonitorCountersAMD(metrics->monitor_begun_id,
 173                                                GL_TRUE, group->id,
 174                                                num_counters,
 175                                                group->counter_ids);
 176         }
 177
 178         /* Start the queries */
 179         glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
 180
 181         glBeginPerfMonitorAMD (metrics->monitor_begun_id);
 182 }
 183
 184 void
 185 metrics_counter_stop (void)
 186 {
 187         context_t *ctx = context_get_current ();
 188         metrics_t *metrics = ctx->metrics;
 189         timer_query_t *timer;
 190         monitor_t *monitor;
 191
 192         /* Stop the current timer and monitor. */
 193         glEndQuery (GL_TIME_ELAPSED);
 194         glEndPerfMonitorAMD (metrics->monitor_begun_id);
 195
 196         /* Add these IDs to our lists of outstanding queries and
 197          * monitors so the results can be collected later. */
 198         timer = xmalloc (sizeof (timer_query_t));
 199
 200         timer->op = metrics->op;
 201         timer->id = metrics->timer_begun_id;
 202         timer->next = NULL;
 203
 204         if (metrics->timer_tail) {
 205                 metrics->timer_tail->next = timer;
 206                 metrics->timer_tail = timer;
 207         } else {
 208                 metrics->timer_tail = timer;
 209                 metrics->timer_head = timer;
 210         }
 211
 212         /* Create a new performance-monitor query */
 213         monitor = xmalloc (sizeof (monitor_t));
 214
 215         monitor->op = metrics->op;
 216         monitor->id = metrics->monitor_begun_id;
 217         monitor->next = NULL;
 218
 219         if (metrics->monitor_tail) {
 220                 metrics->monitor_tail->next = monitor;
 221                 metrics->monitor_tail = monitor;
 222         } else {
 223                 metrics->monitor_tail = monitor;
 224                 metrics->monitor_head = monitor;
 225         }
 226
 227         metrics->monitors_in_flight++;
 228
 229         /* Avoid being a resource hog and collect outstanding results
 230          * once we have sent off a large number of
 231          * queries. (Presumably, many of the outstanding queries are
 232          * available by now.)
 233          */
 234         if (metrics->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
 235                 metrics_collect_available ();
 236 }
 237
 238 void
 239 metrics_set_current_op (metrics_op_t op)
 240 {
 241         context_t *ctx = context_get_current ();
 242         metrics_t *metrics = ctx->metrics;
 243
 244         metrics->op = op;
 245 }
 246
 247 metrics_op_t
 248 metrics_get_current_op (void)
 249 {
 250         context_t *ctx = context_get_current ();
 251         metrics_t *metrics = ctx->metrics;
 252
 253         return metrics->op;
 254 }
 255
 256 static void
 257 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 258 {
 259         metrics_info_t *info = &ctx->metrics_info;
 260         unsigned i, j;
 261
 262         metrics->op = op;
 263         metrics->time_ns = 0.0;
 264
 265         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 266
 267         for (i = 0; i < info->num_groups; i++) {
 268                 metrics->counters[i] = xmalloc (sizeof (double) *
 269                                                 info->groups[i].num_counters);
 270                 for (j = 0; j < info->groups[i].num_counters; j++)
 271                         metrics->counters[i][j] = 0.0;
 272         }
 273 }
 274
 275 static op_metrics_t *
 276 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 277 {
 278         metrics_t *metrics = ctx->metrics;
 279         unsigned i;
 280
 281         if (op >= metrics->num_op_metrics)
 282         {
 283                 metrics->op_metrics = realloc (metrics->op_metrics,
 284                                                (op + 1) * sizeof (op_metrics_t));
 285                 for (i = metrics->num_op_metrics; i < op + 1; i++)
 286                         op_metrics_init (ctx, &metrics->op_metrics[i], i);
 287
 288                 metrics->num_op_metrics = op + 1;
 289         }
 290
 291         return &metrics->op_metrics[op];
 292 }
 293
 294 static void
 295 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 296 {
 297 #define CONSUME(var)                                                    \
 298         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 299         {                                                               \
 300                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 301                          "parsing results\n");                          \
 302                 break;                                                  \
 303         }                                                               \
 304         (var) = *((typeof(var) *) p);                                   \
 305         p += sizeof(var);
 306
 307         context_t *ctx = context_get_current ();
 308         metrics_info_t *info = &ctx->metrics_info;
 309         op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
 310         unsigned char *p = (unsigned char *) result;
 311
 312         while (p < ((unsigned char *) result) + size)
 313         {
 314                 GLuint group_id, group_index;
 315                 GLuint counter_id, counter_index;
 316                 metrics_group_info_t *group;
 317                 double value;
 318                 unsigned i;
 319
 320                 CONSUME (group_id);
 321                 CONSUME (counter_id);
 322
 323                 for (i = 0; i < info->num_groups; i++) {
 324                         if (info->groups[i].id == group_id)
 325                                 break;
 326                 }
 327                 group_index = i;
 328                 assert (group_index < info->num_groups);
 329                 group = &info->groups[group_index];
 330
 331                 for (i = 0; i < group->num_counters; i++) {
 332                         if (group->counter_ids[i] == counter_id)
 333                                 break;
 334                 }
 335                 counter_index = i;
 336                 assert (counter_index < group->num_counters);
 337
 338                 switch (group->counter_types[counter_index])
 339                 {
 340                         uint uint_value;
 341                         uint64_t uint64_value;
 342                         float float_value;
 343                 case GL_UNSIGNED_INT:
 344                         CONSUME (uint_value);
 345                         value = uint_value;
 346                         break;
 347                 case GL_UNSIGNED_INT64_AMD:
 348                         CONSUME (uint64_value);
 349                         value = uint64_value;
 350                         break;
 351                 case GL_PERCENTAGE_AMD:
 352                 case GL_FLOAT:
 353                         CONSUME (float_value);
 354                         value = float_value;
 355                         break;
 356                 default:
 357                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 358                                  group->counter_types[counter_index]);
 359                         value = 0.0;
 360                         break;
 361                 }
 362
 363                 metrics->counters[group_index][counter_index] += value;
 364         }
 365 }
 366
 367 static void
 368 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 369 {
 370         context_t *ctx = context_get_current ();
 371         op_metrics_t *metrics;
 372
 373         metrics = ctx_get_op_metrics (ctx, op);
 374
 375         metrics->time_ns += time_ns;
 376 }
 377
 378 typedef struct per_stage_metrics
 379 {
 380         op_metrics_t *metrics;
 381         shader_stage_info_t *stage;
 382         double time_ns;
 383         double active;
 384 } per_stage_metrics_t;
 385
 386 static int
 387 _is_shader_stage_counter (metrics_info_t *info,
 388                           unsigned group_index,
 389                           unsigned counter_index)
 390 {
 391         shader_stage_info_t *stage;
 392         unsigned i;
 393
 394         for (i = 0; i < info->num_shader_stages; i++) {
 395                 stage = &info->stages[i];
 396
 397                 if (stage->active_group_index == group_index &&
 398                     stage->active_counter_index == counter_index)
 399                 {
 400                         return 1;
 401                 }
 402
 403                 if (stage->stall_group_index == group_index &&
 404                     stage->stall_counter_index == counter_index)
 405                 {
 406                         return 1;
 407                 }
 408         }
 409
 410         return 0;
 411 }
 412
 413 static void
 414 print_per_stage_metrics (context_t *ctx,
 415                          per_stage_metrics_t *per_stage,
 416                          double total)
 417 {
 418         metrics_info_t *info = &ctx->metrics_info;
 419         op_metrics_t *metric = per_stage->metrics;
 420         metrics_group_info_t *group;
 421         const char *op_string;
 422         unsigned group_index, counter;
 423         double value;
 424
 425         /* Don't print anything for stages with no alloted time. */
 426         if (per_stage->time_ns == 0.0)
 427                 return;
 428
 429         op_string = metrics_op_string (metric->op);
 430
 431         printf ("%21s", op_string);
 432
 433         if (metric->op >= METRICS_OP_SHADER) {
 434                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 435         } else {
 436                 printf ("    ");
 437
 438         }
 439
 440         if (per_stage->stage)
 441                 printf (" %cS:", per_stage->stage->name[0]);
 442         else
 443                 printf ("   :");
 444
 445         printf ("\t%7.2f ms (%4.1f%%)",
 446                 per_stage->time_ns / 1e6,
 447                 per_stage->time_ns / total * 100);
 448
 449         if (per_stage->active)
 450                 printf (", %4.1f%% active", per_stage->active * 100);
 451
 452         printf ("\n");
 453
 454         /* I'm not seeing a lot of value printing the rest of these
 455          * performance counters by default yet. Use --verbose to get
 456          * them for now. */
 457         if (! verbose)
 458                 return;
 459
 460         printf ("[");
 461         for (group_index = 0; group_index < info->num_groups; group_index++) {
 462                 group = &info->groups[group_index];
 463                 for (counter = 0; counter < group->num_counters; counter++) {
 464
 465                         /* Don't print this counter value if it's a
 466                          * per-stage cycle counter, (which we have
 467                          * already accounted for). */
 468                         if (_is_shader_stage_counter (info, group_index, counter))
 469                                 continue;
 470
 471                         value = metric->counters[group_index][counter];
 472                         if (value == 0.0)
 473                                 continue;
 474                         printf ("%s: %.2f ", group->counter_names[counter],
 475                                 value / 1e6);
 476                 }
 477         }
 478         printf ("]\n");
 479 }
 480
 481 static int
 482 time_compare(const void *in_a, const void *in_b, void *arg unused)
 483 {
 484         const per_stage_metrics_t *a = in_a;
 485         const per_stage_metrics_t *b = in_b;
 486
 487
 488         if (a->time_ns < b->time_ns)
 489                 return -1;
 490         if (a->time_ns > b->time_ns)
 491                 return 1;
 492         return 0;
 493 }
 494
 495 static void
 496 print_program_metrics (void)
 497 {
 498         context_t *ctx = context_get_current ();
 499         metrics_t *metrics = ctx->metrics;
 500         metrics_info_t *info = &ctx->metrics_info;
 501         unsigned num_shader_stages = info->num_shader_stages;
 502         per_stage_metrics_t *sorted, *per_stage;
 503         double total_time, op_cycles;
 504         op_metrics_t *op;
 505         unsigned group_index, counter_index;
 506         unsigned i, j, num_sorted;
 507
 508         /* Make a sorted list of the per-stage operations by time
 509          * used, and figure out the total so we can print percentages.
 510          */
 511         num_sorted = metrics->num_op_metrics * num_shader_stages;
 512
 513         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 514
 515         total_time = 0.0;
 516
 517         for (i = 0; i < metrics->num_op_metrics; i++) {
 518
 519                 op = &metrics->op_metrics[i];
 520
 521                 /* Accumulate total time across all ops. */
 522                 total_time += op->time_ns;
 523
 524                 /* Also, find total cycles in all stages of this op. */
 525                 op_cycles = 0.0;
 526
 527                 for (j = 0; j < num_shader_stages; j++) {
 528                         /* Active cycles */
 529                         group_index = info->stages[j].active_group_index;
 530                         counter_index = info->stages[j].active_counter_index;
 531                         op_cycles += op->counters[group_index][counter_index];
 532
 533                         /* Stall cycles */
 534                         group_index = info->stages[j].stall_group_index;
 535                         counter_index = info->stages[j].stall_counter_index;
 536                         op_cycles += op->counters[group_index][counter_index];
 537                 }
 538
 539                 for (j = 0; j < num_shader_stages; j++) {
 540                         double active_cycles, stall_cycles, stage_cycles;
 541
 542                         /* Active cycles */
 543                         group_index = info->stages[j].active_group_index;
 544                         counter_index = info->stages[j].active_counter_index;
 545                         active_cycles = op->counters[group_index][counter_index];
 546
 547                         /* Stall cycles */
 548                         group_index = info->stages[j].stall_group_index;
 549                         counter_index = info->stages[j].stall_counter_index;
 550                         stall_cycles = op->counters[group_index][counter_index];
 551
 552                         stage_cycles = active_cycles + stall_cycles;
 553
 554                         per_stage = &sorted[i * num_shader_stages + j];
 555                         per_stage->metrics = op;
 556
 557                         if (op_cycles) {
 558                                 per_stage->stage = &info->stages[j];
 559                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 560                         } else {
 561                                 /* If we don't have any per-stage cycle counts
 562                                  * for this operation, then use the first
 563                                  * stage as a placeholder for all the time,
 564                                  * but NULL-ify the stage info so that the
 565                                  * report doesn't lie about this time being
 566                                  * from any particular stage. */
 567                                 per_stage->stage = NULL;
 568                                 if (j == 0) {
 569                                         per_stage->time_ns = op->time_ns;
 570                                 } else {
 571                                         per_stage->time_ns = 0.0;
 572                                 }
 573                         }
 574
 575                         if (stage_cycles) {
 576                                 per_stage->active = active_cycles / stage_cycles;
 577                         } else {
 578                                 per_stage->active = 0.0;
 579                         }
 580                 }
 581         }
 582
 583         qsort_r (sorted, num_sorted, sizeof (*sorted),
 584                  time_compare, metrics->op_metrics);
 585
 586         for (i = 0; i < num_sorted; i++)
 587                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 588
 589         free (sorted);
 590 }
 591
 592 /* Called at program exit.
 593  *
 594  * This is similar to metrics_info_fini, but only frees any used
 595  * memory. Notably, it does not call any OpenGL functions, (since the
 596  * OpenGL context no longer exists at program exit).
 597  */
 598 static void
 599 metrics_exit (void)
 600 {
 601         context_t *ctx = context_get_current ();
 602         metrics_t *metrics = ctx->metrics;
 603         metrics_info_t *info = &ctx->metrics_info;
 604         unsigned i, j;
 605         timer_query_t *timer, *timer_next;
 606         monitor_t *monitor, *monitor_next;
 607
 608         if (verbose)
 609                 printf ("fips: terminating\n");
 610
 611         if (! info->initialized)
 612                 return;
 613
 614         for (timer = metrics->timer_head;
 615              timer;
 616              timer = timer_next)
 617         {
 618                 timer_next = timer->next;
 619                 free (timer);
 620         }
 621
 622         for (monitor = metrics->monitor_head;
 623              monitor;
 624              monitor = monitor_next)
 625         {
 626                 monitor_next = monitor->next;
 627                 free (monitor);
 628         }
 629
 630         for (i = 0; i < info->num_groups; i++) {
 631                 metrics_group_info_t *group = &info->groups[i];
 632
 633                 for (j = 0; j < group->num_counters; i++)
 634                         free (group->counter_names[j]);
 635
 636                 free (group->counter_types);
 637                 free (group->counter_names);
 638                 free (group->counter_ids);
 639
 640                 free (group->name);
 641         }
 642
 643         free (info->groups);
 644
 645         for (i = 0; i < info->num_shader_stages; i++)
 646                 free (info->stages[i].name);
 647
 648         free (info->stages);
 649 }
 650
 651 void
 652 metrics_collect_available (void)
 653 {
 654         context_t *ctx = context_get_current ();
 655         metrics_t *metrics = ctx->metrics;
 656
 657         /* Consume all timer queries that are ready. */
 658         timer_query_t *timer = metrics->timer_head;
 659
 660         while (timer) {
 661                 GLuint available, elapsed;
 662
 663                 glGetQueryObjectuiv (timer->id,
 664                                      GL_QUERY_RESULT_AVAILABLE, &available);
 665                 if (! available)
 666                         break;
 667
 668                 glGetQueryObjectuiv (timer->id,
 669                                      GL_QUERY_RESULT, &elapsed);
 670
 671                 accumulate_program_time (timer->op, elapsed);
 672
 673                 metrics->timer_head = timer->next;
 674                 if (metrics->timer_head == NULL)
 675                         metrics->timer_tail = NULL;
 676
 677                 glDeleteQueries (1, &timer->id);
 678
 679                 free (timer);
 680                 timer = metrics->timer_head;
 681         }
 682
 683         /* And similarly for all performance monitors that are ready. */
 684         monitor_t *monitor = metrics->monitor_head;
 685
 686         while (monitor) {
 687                 GLuint available, result_size, *result;
 688                 GLint bytes_written;
 689
 690                 glGetPerfMonitorCounterDataAMD (monitor->id,
 691                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 692                                                 sizeof (available), &available,
 693                                                 NULL);
 694                 if (! available)
 695                         break;
 696
 697                 glGetPerfMonitorCounterDataAMD (monitor->id,
 698                                                 GL_PERFMON_RESULT_SIZE_AMD,
 699                                                 sizeof (result_size),
 700                                                 &result_size, NULL);
 701
 702                 result = xmalloc (result_size);
 703
 704                 glGetPerfMonitorCounterDataAMD (monitor->id,
 705                                                 GL_PERFMON_RESULT_AMD,
 706                                                 result_size, result,
 707                                                 &bytes_written);
 708
 709                 accumulate_program_metrics (monitor->op, result, result_size);
 710
 711                 free (result);
 712
 713                 metrics->monitor_head = monitor->next;
 714                 if (metrics->monitor_head == NULL)
 715                         metrics->monitor_tail = NULL;
 716
 717                 glDeletePerfMonitorsAMD (1, &monitor->id);
 718
 719                 free (monitor);
 720
 721                 metrics->monitors_in_flight--;
 722
 723                 monitor = metrics->monitor_head;
 724         }
 725 }
 726
 727
 728 void
 729 metrics_end_frame (void)
 730 {
 731         static int initialized = 0;
 732         static struct timeval tv_start, tv_now;
 733
 734         if (! initialized) {
 735                 gettimeofday (&tv_start, NULL);
 736                 atexit (metrics_exit);
 737                 if (getenv ("FIPS_VERBOSE"))
 738                         verbose = 1;
 739                 initialized = 1;
 740         }
 741
 742         frames++;
 743
 744         metrics_collect_available ();
 745
 746         if (frames % 15 == 0) {
 747                 double fps;
 748
 749                 gettimeofday (&tv_now, NULL);
 750
 751                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 752                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 753
 754                 printf("FPS: %.3f\n", fps);
 755
 756                 print_program_metrics ();
 757         }
 758 }
 759
 760 metrics_t *
 761 metrics_create (void)
 762 {
 763         metrics_t *metrics;
 764
 765         metrics = xmalloc (sizeof (metrics_t));
 766
 767         metrics->op = 0;
 768
 769         metrics->timer_begun_id = 0;
 770
 771         metrics->timer_head = NULL;
 772         metrics->timer_tail = NULL;
 773
 774         metrics->monitor_begun_id = 0;
 775
 776         metrics->monitor_head = NULL;
 777         metrics->monitor_tail = NULL;
 778
 779         metrics->monitors_in_flight = 0;
 780
 781         metrics->num_op_metrics = 0;
 782         metrics->op_metrics = NULL;
 783
 784         return metrics;
 785 }
 786
 787 void
 788 metrics_fini (metrics_t *metrics)
 789 {
 790         timer_query_t *timer, *timer_next;
 791         monitor_t *monitor, *monitor_next;
 792
 793         /* Discard and cleanup any outstanding queries. */
 794         if (metrics->timer_begun_id) {
 795                 glEndQuery (GL_TIME_ELAPSED);
 796                 glDeleteQueries (1, &metrics->timer_begun_id);
 797                 metrics->timer_begun_id = 0;
 798         }
 799
 800         for (timer = metrics->timer_head;
 801              timer;
 802              timer = timer_next)
 803         {
 804                 glDeleteQueries (1, &timer->id);
 805                 timer_next = timer->next;
 806                 free (timer);
 807         }
 808         metrics->timer_head = NULL;
 809         metrics->timer_tail = NULL;
 810
 811         if (metrics->monitor_begun_id) {
 812                 glEndPerfMonitorAMD (metrics->monitor_begun_id);
 813                 glDeletePerfMonitorsAMD (1, &metrics->monitor_begun_id);
 814                 metrics->monitor_begun_id = 0;
 815         }
 816
 817         for (monitor = metrics->monitor_head;
 818              monitor;
 819              monitor = monitor_next)
 820         {
 821                 glDeletePerfMonitorsAMD (1, &monitor->id);
 822                 monitor_next = monitor->next;
 823                 free (monitor);
 824         }
 825         metrics->monitor_head = NULL;
 826         metrics->monitor_tail = NULL;
 827
 828         metrics->monitors_in_flight = 0;
 829 }
 830
 831 void
 832 metrics_destroy (metrics_t *metrics)
 833 {
 834         metrics_fini (metrics);
 835
 836         free (metrics);
 837 }