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         /* Description of all available peformance counters, counter
  73          * groups, their names and IDs, etc. */
  74         metrics_info_t *info;
  75
  76         /* The current operation being measured. */
  77         metrics_op_t op;
  78
  79         /* GL_TIME_ELAPSED query for which glEndQuery has not yet
  80          * been called. */
  81         unsigned timer_begun_id;
  82
  83         /* GL_TIME_ELAPSED queries for which glEndQuery has been
  84          * called, (but results have not yet been queried). */
  85         timer_query_t *timer_head;
  86         timer_query_t *timer_tail;
  87
  88         /* Performance monitor for which glEndPerfMonitorAMD has not
  89          * yet been called. */
  90         unsigned monitor_begun_id;
  91
  92         /* Performance monitors for which glEndPerfMonitorAMD has
  93          * been called, (but results have not yet been queried). */
  94         monitor_t *monitor_head;
  95         monitor_t *monitor_tail;
  96
  97         int monitors_in_flight;
  98
  99         unsigned num_op_metrics;
 100         op_metrics_t *op_metrics;
 101 };
 102
 103 metrics_t *
 104 metrics_create (metrics_info_t *info)
 105 {
 106         metrics_t *metrics;
 107
 108         metrics = xmalloc (sizeof (metrics_t));
 109
 110         metrics->info = info;
 111
 112         metrics->op = 0;
 113
 114         metrics->timer_begun_id = 0;
 115
 116         metrics->timer_head = NULL;
 117         metrics->timer_tail = NULL;
 118
 119         metrics->monitor_begun_id = 0;
 120
 121         metrics->monitor_head = NULL;
 122         metrics->monitor_tail = NULL;
 123
 124         metrics->monitors_in_flight = 0;
 125
 126         metrics->num_op_metrics = 0;
 127         metrics->op_metrics = NULL;
 128
 129         return metrics;
 130 }
 131
 132 void
 133 metrics_fini (metrics_t *metrics)
 134 {
 135         timer_query_t *timer, *timer_next;
 136         monitor_t *monitor, *monitor_next;
 137
 138         /* Discard and cleanup any outstanding queries. */
 139         if (metrics->timer_begun_id) {
 140                 glEndQuery (GL_TIME_ELAPSED);
 141                 glDeleteQueries (1, &metrics->timer_begun_id);
 142                 metrics->timer_begun_id = 0;
 143         }
 144
 145         for (timer = metrics->timer_head;
 146              timer;
 147              timer = timer_next)
 148         {
 149                 glDeleteQueries (1, &timer->id);
 150                 timer_next = timer->next;
 151                 free (timer);
 152         }
 153         metrics->timer_head = NULL;
 154         metrics->timer_tail = NULL;
 155
 156         if (metrics->monitor_begun_id) {
 157                 glEndPerfMonitorAMD (metrics->monitor_begun_id);
 158                 glDeletePerfMonitorsAMD (1, &metrics->monitor_begun_id);
 159                 metrics->monitor_begun_id = 0;
 160         }
 161
 162         for (monitor = metrics->monitor_head;
 163              monitor;
 164              monitor = monitor_next)
 165         {
 166                 glDeletePerfMonitorsAMD (1, &monitor->id);
 167                 monitor_next = monitor->next;
 168                 free (monitor);
 169         }
 170         metrics->monitor_head = NULL;
 171         metrics->monitor_tail = NULL;
 172
 173         metrics->monitors_in_flight = 0;
 174 }
 175
 176 void
 177 metrics_destroy (metrics_t *metrics)
 178 {
 179         metrics_fini (metrics);
 180
 181         free (metrics);
 182 }
 183
 184 static const char *
 185 metrics_op_string (metrics_op_t op)
 186 {
 187         if (op >= METRICS_OP_SHADER)
 188                 return "Shader program";
 189
 190         switch (op)
 191         {
 192         case METRICS_OP_ACCUM:
 193                 return "glAccum*(+)";
 194         case METRICS_OP_BUFFER_DATA:
 195                 return "glBufferData(+)";
 196         case METRICS_OP_BUFFER_SUB_DATA:
 197                 return "glCopyBufferSubData*";
 198         case METRICS_OP_BITMAP:
 199                 return "glBitmap*";
 200         case METRICS_OP_BLIT_FRAMEBUFFER:
 201                 return "glBlitFramebuffer*";
 202         case METRICS_OP_CLEAR:
 203                 return "glClear(+)";
 204         case METRICS_OP_CLEAR_BUFFER_DATA:
 205                 return "glCearBufferData(+)";
 206         case METRICS_OP_CLEAR_TEX_IMAGE:
 207                 return "glClearTexImage(+)";
 208         case METRICS_OP_COPY_PIXELS:
 209                 return "glCopyPixels";
 210         case METRICS_OP_COPY_TEX_IMAGE:
 211                 return "glCopyTexImage(+)";
 212         case METRICS_OP_DRAW_PIXELS:
 213                 return "glDrawPixels";
 214         case METRICS_OP_GET_TEX_IMAGE:
 215                 return "glGetTexImage(+)";
 216         case METRICS_OP_READ_PIXELS:
 217                 return "glReadPixels*";
 218         case METRICS_OP_TEX_IMAGE:
 219                 return "glTexImage*(+)";
 220         default:
 221                 fprintf (stderr, "fips: Internal error: "
 222                          "Unknown metrics op value: %d\n", op);
 223                 exit (1);
 224         }
 225
 226         return "";
 227 }
 228
 229 void
 230 metrics_counter_start (void)
 231 {
 232         context_t *ctx = context_get_current ();
 233         metrics_t *metrics = ctx->metrics;
 234         unsigned i;
 235
 236         /* Initialize the timer_query and monitor objects */
 237         glGenQueries (1, &metrics->timer_begun_id);
 238
 239         glGenPerfMonitorsAMD (1, &metrics->monitor_begun_id);
 240
 241         for (i = 0; i < metrics->info->num_groups; i++)
 242         {
 243                 metrics_group_info_t *group;
 244                 int num_counters;
 245
 246                 group = &metrics->info->groups[i];
 247
 248                 num_counters = group->num_counters;
 249                 if (group->max_active_counters < group->num_counters)
 250                 {
 251                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 252                                  group->max_active_counters,
 253                                  group->num_counters, i);
 254                         num_counters = group->max_active_counters;
 255
 256                 }
 257
 258                 glSelectPerfMonitorCountersAMD(metrics->monitor_begun_id,
 259                                                GL_TRUE, group->id,
 260                                                num_counters,
 261                                                group->counter_ids);
 262         }
 263
 264         /* Start the queries */
 265         glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
 266
 267         glBeginPerfMonitorAMD (metrics->monitor_begun_id);
 268 }
 269
 270 void
 271 metrics_counter_stop (void)
 272 {
 273         context_t *ctx = context_get_current ();
 274         metrics_t *metrics = ctx->metrics;
 275         timer_query_t *timer;
 276         monitor_t *monitor;
 277
 278         /* Stop the current timer and monitor. */
 279         glEndQuery (GL_TIME_ELAPSED);
 280         glEndPerfMonitorAMD (metrics->monitor_begun_id);
 281
 282         /* Add these IDs to our lists of outstanding queries and
 283          * monitors so the results can be collected later. */
 284         timer = xmalloc (sizeof (timer_query_t));
 285
 286         timer->op = metrics->op;
 287         timer->id = metrics->timer_begun_id;
 288         timer->next = NULL;
 289
 290         if (metrics->timer_tail) {
 291                 metrics->timer_tail->next = timer;
 292                 metrics->timer_tail = timer;
 293         } else {
 294                 metrics->timer_tail = timer;
 295                 metrics->timer_head = timer;
 296         }
 297
 298         /* Create a new performance-monitor query */
 299         monitor = xmalloc (sizeof (monitor_t));
 300
 301         monitor->op = metrics->op;
 302         monitor->id = metrics->monitor_begun_id;
 303         monitor->next = NULL;
 304
 305         if (metrics->monitor_tail) {
 306                 metrics->monitor_tail->next = monitor;
 307                 metrics->monitor_tail = monitor;
 308         } else {
 309                 metrics->monitor_tail = monitor;
 310                 metrics->monitor_head = monitor;
 311         }
 312
 313         metrics->monitors_in_flight++;
 314
 315         /* Avoid being a resource hog and collect outstanding results
 316          * once we have sent off a large number of
 317          * queries. (Presumably, many of the outstanding queries are
 318          * available by now.)
 319          */
 320         if (metrics->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
 321                 metrics_collect_available ();
 322 }
 323
 324 void
 325 metrics_set_current_op (metrics_op_t op)
 326 {
 327         context_t *ctx = context_get_current ();
 328         metrics_t *metrics = ctx->metrics;
 329
 330         metrics->op = op;
 331 }
 332
 333 metrics_op_t
 334 metrics_get_current_op (void)
 335 {
 336         context_t *ctx = context_get_current ();
 337         metrics_t *metrics = ctx->metrics;
 338
 339         return metrics->op;
 340 }
 341
 342 static void
 343 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 344 {
 345         metrics_info_t *info = ctx->metrics->info;
 346         unsigned i, j;
 347
 348         metrics->op = op;
 349         metrics->time_ns = 0.0;
 350
 351         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 352
 353         for (i = 0; i < info->num_groups; i++) {
 354                 metrics->counters[i] = xmalloc (sizeof (double) *
 355                                                 info->groups[i].num_counters);
 356                 for (j = 0; j < info->groups[i].num_counters; j++)
 357                         metrics->counters[i][j] = 0.0;
 358         }
 359 }
 360
 361 static op_metrics_t *
 362 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 363 {
 364         metrics_t *metrics = ctx->metrics;
 365         unsigned i;
 366
 367         if (op >= metrics->num_op_metrics)
 368         {
 369                 metrics->op_metrics = realloc (metrics->op_metrics,
 370                                                (op + 1) * sizeof (op_metrics_t));
 371                 for (i = metrics->num_op_metrics; i < op + 1; i++)
 372                         op_metrics_init (ctx, &metrics->op_metrics[i], i);
 373
 374                 metrics->num_op_metrics = op + 1;
 375         }
 376
 377         return &metrics->op_metrics[op];
 378 }
 379
 380 static void
 381 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 382 {
 383 #define CONSUME(var)                                                    \
 384         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 385         {                                                               \
 386                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 387                          "parsing results\n");                          \
 388                 break;                                                  \
 389         }                                                               \
 390         (var) = *((typeof(var) *) p);                                   \
 391         p += sizeof(var);
 392
 393         context_t *ctx = context_get_current ();
 394         metrics_info_t *info = ctx->metrics->info;
 395         op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
 396         unsigned char *p = (unsigned char *) result;
 397
 398         while (p < ((unsigned char *) result) + size)
 399         {
 400                 GLuint group_id, group_index;
 401                 GLuint counter_id, counter_index;
 402                 metrics_group_info_t *group;
 403                 double value;
 404                 unsigned i;
 405
 406                 CONSUME (group_id);
 407                 CONSUME (counter_id);
 408
 409                 for (i = 0; i < info->num_groups; i++) {
 410                         if (info->groups[i].id == group_id)
 411                                 break;
 412                 }
 413                 group_index = i;
 414                 assert (group_index < info->num_groups);
 415                 group = &info->groups[group_index];
 416
 417                 for (i = 0; i < group->num_counters; i++) {
 418                         if (group->counter_ids[i] == counter_id)
 419                                 break;
 420                 }
 421                 counter_index = i;
 422                 assert (counter_index < group->num_counters);
 423
 424                 switch (group->counter_types[counter_index])
 425                 {
 426                         uint uint_value;
 427                         uint64_t uint64_value;
 428                         float float_value;
 429                 case GL_UNSIGNED_INT:
 430                         CONSUME (uint_value);
 431                         value = uint_value;
 432                         break;
 433                 case GL_UNSIGNED_INT64_AMD:
 434                         CONSUME (uint64_value);
 435                         value = uint64_value;
 436                         break;
 437                 case GL_PERCENTAGE_AMD:
 438                 case GL_FLOAT:
 439                         CONSUME (float_value);
 440                         value = float_value;
 441                         break;
 442                 default:
 443                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 444                                  group->counter_types[counter_index]);
 445                         value = 0.0;
 446                         break;
 447                 }
 448
 449                 metrics->counters[group_index][counter_index] += value;
 450         }
 451 }
 452
 453 static void
 454 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 455 {
 456         context_t *ctx = context_get_current ();
 457         op_metrics_t *metrics;
 458
 459         metrics = ctx_get_op_metrics (ctx, op);
 460
 461         metrics->time_ns += time_ns;
 462 }
 463
 464 typedef struct per_stage_metrics
 465 {
 466         op_metrics_t *metrics;
 467         shader_stage_info_t *stage;
 468         double time_ns;
 469         double active;
 470 } per_stage_metrics_t;
 471
 472 static int
 473 _is_shader_stage_counter (metrics_info_t *info,
 474                           unsigned group_index,
 475                           unsigned counter_index)
 476 {
 477         shader_stage_info_t *stage;
 478         unsigned i;
 479
 480         for (i = 0; i < info->num_shader_stages; i++) {
 481                 stage = &info->stages[i];
 482
 483                 if (stage->active_group_index == group_index &&
 484                     stage->active_counter_index == counter_index)
 485                 {
 486                         return 1;
 487                 }
 488
 489                 if (stage->stall_group_index == group_index &&
 490                     stage->stall_counter_index == counter_index)
 491                 {
 492                         return 1;
 493                 }
 494         }
 495
 496         return 0;
 497 }
 498
 499 static void
 500 print_per_stage_metrics (context_t *ctx,
 501                          per_stage_metrics_t *per_stage,
 502                          double total)
 503 {
 504         metrics_info_t *info = ctx->metrics->info;
 505         op_metrics_t *metric = per_stage->metrics;
 506         metrics_group_info_t *group;
 507         const char *op_string;
 508         unsigned group_index, counter;
 509         double value;
 510
 511         /* Don't print anything for stages with no alloted time. */
 512         if (per_stage->time_ns == 0.0)
 513                 return;
 514
 515         op_string = metrics_op_string (metric->op);
 516
 517         printf ("%21s", op_string);
 518
 519         if (metric->op >= METRICS_OP_SHADER) {
 520                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 521         } else {
 522                 printf ("    ");
 523
 524         }
 525
 526         if (per_stage->stage)
 527                 printf (" %cS:", per_stage->stage->name[0]);
 528         else
 529                 printf ("   :");
 530
 531         printf ("\t%7.2f ms (%4.1f%%)",
 532                 per_stage->time_ns / 1e6,
 533                 per_stage->time_ns / total * 100);
 534
 535         if (per_stage->active)
 536                 printf (", %4.1f%% active", per_stage->active * 100);
 537
 538         printf ("\n");
 539
 540         /* I'm not seeing a lot of value printing the rest of these
 541          * performance counters by default yet. Use --verbose to get
 542          * them for now. */
 543         if (! verbose)
 544                 return;
 545
 546         printf ("[");
 547         for (group_index = 0; group_index < info->num_groups; group_index++) {
 548                 group = &info->groups[group_index];
 549                 for (counter = 0; counter < group->num_counters; counter++) {
 550
 551                         /* Don't print this counter value if it's a
 552                          * per-stage cycle counter, (which we have
 553                          * already accounted for). */
 554                         if (_is_shader_stage_counter (info, group_index, counter))
 555                                 continue;
 556
 557                         value = metric->counters[group_index][counter];
 558                         if (value == 0.0)
 559                                 continue;
 560                         printf ("%s: %.2f ", group->counter_names[counter],
 561                                 value / 1e6);
 562                 }
 563         }
 564         printf ("]\n");
 565 }
 566
 567 static int
 568 time_compare(const void *in_a, const void *in_b, void *arg unused)
 569 {
 570         const per_stage_metrics_t *a = in_a;
 571         const per_stage_metrics_t *b = in_b;
 572
 573
 574         if (a->time_ns < b->time_ns)
 575                 return -1;
 576         if (a->time_ns > b->time_ns)
 577                 return 1;
 578         return 0;
 579 }
 580
 581 static void
 582 print_program_metrics (void)
 583 {
 584         context_t *ctx = context_get_current ();
 585         metrics_t *metrics = ctx->metrics;
 586         metrics_info_t *info = metrics->info;
 587         unsigned num_shader_stages = info->num_shader_stages;
 588         per_stage_metrics_t *sorted, *per_stage;
 589         double total_time, op_cycles;
 590         op_metrics_t *op;
 591         unsigned group_index, counter_index;
 592         unsigned i, j, num_sorted;
 593
 594         /* Make a sorted list of the per-stage operations by time
 595          * used, and figure out the total so we can print percentages.
 596          */
 597         num_sorted = metrics->num_op_metrics * num_shader_stages;
 598
 599         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 600
 601         total_time = 0.0;
 602
 603         for (i = 0; i < metrics->num_op_metrics; i++) {
 604
 605                 op = &metrics->op_metrics[i];
 606
 607                 /* Accumulate total time across all ops. */
 608                 total_time += op->time_ns;
 609
 610                 /* Also, find total cycles in all stages of this op. */
 611                 op_cycles = 0.0;
 612
 613                 for (j = 0; j < num_shader_stages; j++) {
 614                         /* Active cycles */
 615                         group_index = info->stages[j].active_group_index;
 616                         counter_index = info->stages[j].active_counter_index;
 617                         op_cycles += op->counters[group_index][counter_index];
 618
 619                         /* Stall cycles */
 620                         group_index = info->stages[j].stall_group_index;
 621                         counter_index = info->stages[j].stall_counter_index;
 622                         op_cycles += op->counters[group_index][counter_index];
 623                 }
 624
 625                 for (j = 0; j < num_shader_stages; j++) {
 626                         double active_cycles, stall_cycles, stage_cycles;
 627
 628                         /* Active cycles */
 629                         group_index = info->stages[j].active_group_index;
 630                         counter_index = info->stages[j].active_counter_index;
 631                         active_cycles = op->counters[group_index][counter_index];
 632
 633                         /* Stall cycles */
 634                         group_index = info->stages[j].stall_group_index;
 635                         counter_index = info->stages[j].stall_counter_index;
 636                         stall_cycles = op->counters[group_index][counter_index];
 637
 638                         stage_cycles = active_cycles + stall_cycles;
 639
 640                         per_stage = &sorted[i * num_shader_stages + j];
 641                         per_stage->metrics = op;
 642
 643                         if (op_cycles) {
 644                                 per_stage->stage = &info->stages[j];
 645                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 646                         } else {
 647                                 /* If we don't have any per-stage cycle counts
 648                                  * for this operation, then use the first
 649                                  * stage as a placeholder for all the time,
 650                                  * but NULL-ify the stage info so that the
 651                                  * report doesn't lie about this time being
 652                                  * from any particular stage. */
 653                                 per_stage->stage = NULL;
 654                                 if (j == 0) {
 655                                         per_stage->time_ns = op->time_ns;
 656                                 } else {
 657                                         per_stage->time_ns = 0.0;
 658                                 }
 659                         }
 660
 661                         if (stage_cycles) {
 662                                 per_stage->active = active_cycles / stage_cycles;
 663                         } else {
 664                                 per_stage->active = 0.0;
 665                         }
 666                 }
 667         }
 668
 669         qsort_r (sorted, num_sorted, sizeof (*sorted),
 670                  time_compare, metrics->op_metrics);
 671
 672         for (i = 0; i < num_sorted; i++)
 673                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 674
 675         free (sorted);
 676 }
 677
 678 /* Called at program exit.
 679  *
 680  * This is similar to metrics_info_fini, but only frees any used
 681  * memory. Notably, it does not call any OpenGL functions, (since the
 682  * OpenGL context no longer exists at program exit).
 683  */
 684 static void
 685 metrics_exit (void)
 686 {
 687         context_t *ctx = context_get_current ();
 688         metrics_t *metrics = ctx->metrics;
 689         metrics_info_t *info = metrics->info;
 690         unsigned i, j;
 691         timer_query_t *timer, *timer_next;
 692         monitor_t *monitor, *monitor_next;
 693
 694         if (verbose)
 695                 printf ("fips: terminating\n");
 696
 697         if (! info->initialized)
 698                 return;
 699
 700         for (timer = metrics->timer_head;
 701              timer;
 702              timer = timer_next)
 703         {
 704                 timer_next = timer->next;
 705                 free (timer);
 706         }
 707
 708         for (monitor = metrics->monitor_head;
 709              monitor;
 710              monitor = monitor_next)
 711         {
 712                 monitor_next = monitor->next;
 713                 free (monitor);
 714         }
 715
 716         for (i = 0; i < info->num_groups; i++) {
 717                 metrics_group_info_t *group = &info->groups[i];
 718
 719                 for (j = 0; j < group->num_counters; i++)
 720                         free (group->counter_names[j]);
 721
 722                 free (group->counter_types);
 723                 free (group->counter_names);
 724                 free (group->counter_ids);
 725
 726                 free (group->name);
 727         }
 728
 729         free (info->groups);
 730
 731         for (i = 0; i < info->num_shader_stages; i++)
 732                 free (info->stages[i].name);
 733
 734         free (info->stages);
 735 }
 736
 737 void
 738 metrics_collect_available (void)
 739 {
 740         context_t *ctx = context_get_current ();
 741         metrics_t *metrics = ctx->metrics;
 742
 743         /* Consume all timer queries that are ready. */
 744         timer_query_t *timer = metrics->timer_head;
 745
 746         while (timer) {
 747                 GLuint available, elapsed;
 748
 749                 glGetQueryObjectuiv (timer->id,
 750                                      GL_QUERY_RESULT_AVAILABLE, &available);
 751                 if (! available)
 752                         break;
 753
 754                 glGetQueryObjectuiv (timer->id,
 755                                      GL_QUERY_RESULT, &elapsed);
 756
 757                 accumulate_program_time (timer->op, elapsed);
 758
 759                 metrics->timer_head = timer->next;
 760                 if (metrics->timer_head == NULL)
 761                         metrics->timer_tail = NULL;
 762
 763                 glDeleteQueries (1, &timer->id);
 764
 765                 free (timer);
 766                 timer = metrics->timer_head;
 767         }
 768
 769         /* And similarly for all performance monitors that are ready. */
 770         monitor_t *monitor = metrics->monitor_head;
 771
 772         while (monitor) {
 773                 GLuint available, result_size, *result;
 774                 GLint bytes_written;
 775
 776                 glGetPerfMonitorCounterDataAMD (monitor->id,
 777                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 778                                                 sizeof (available), &available,
 779                                                 NULL);
 780                 if (! available)
 781                         break;
 782
 783                 glGetPerfMonitorCounterDataAMD (monitor->id,
 784                                                 GL_PERFMON_RESULT_SIZE_AMD,
 785                                                 sizeof (result_size),
 786                                                 &result_size, NULL);
 787
 788                 result = xmalloc (result_size);
 789
 790                 glGetPerfMonitorCounterDataAMD (monitor->id,
 791                                                 GL_PERFMON_RESULT_AMD,
 792                                                 result_size, result,
 793                                                 &bytes_written);
 794
 795                 accumulate_program_metrics (monitor->op, result, result_size);
 796
 797                 free (result);
 798
 799                 metrics->monitor_head = monitor->next;
 800                 if (metrics->monitor_head == NULL)
 801                         metrics->monitor_tail = NULL;
 802
 803                 glDeletePerfMonitorsAMD (1, &monitor->id);
 804
 805                 free (monitor);
 806
 807                 metrics->monitors_in_flight--;
 808
 809                 monitor = metrics->monitor_head;
 810         }
 811 }
 812
 813
 814 void
 815 metrics_end_frame (void)
 816 {
 817         static int initialized = 0;
 818         static struct timeval tv_start, tv_now;
 819
 820         if (! initialized) {
 821                 gettimeofday (&tv_start, NULL);
 822                 atexit (metrics_exit);
 823                 if (getenv ("FIPS_VERBOSE"))
 824                         verbose = 1;
 825                 initialized = 1;
 826         }
 827
 828         frames++;
 829
 830         metrics_collect_available ();
 831
 832         if (frames % 15 == 0) {
 833                 double fps;
 834
 835                 gettimeofday (&tv_now, NULL);
 836
 837                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 838                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 839
 840                 printf("FPS: %.3f\n", fps);
 841
 842                 print_program_metrics ();
 843         }
 844 }