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