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 <inttypes.h>
  25
  26 #include <stdio.h>
  27 #include <stdlib.h>
  28 #include <assert.h>
  29 #include <sys/time.h>
  30
  31 #include "fips-dispatch-gl.h"
  32
  33 #include "metrics.h"
  34 #include "context.h"
  35 #include "metrics-info.h"
  36 #include "xmalloc.h"
  37
  38 int frames;
  39 int verbose;
  40
  41 #define MAX_MONITORS_IN_FLIGHT 1000
  42
  43 /* Timer query */
  44 typedef struct timer_query
  45 {
  46         unsigned id;
  47
  48         metrics_op_t op;
  49         struct timer_query *next;
  50 } timer_query_t;
  51
  52 /* Performance-monitor query */
  53 typedef struct monitor
  54 {
  55         unsigned id;
  56
  57         metrics_op_t op;
  58         struct monitor *next;
  59 } monitor_t;
  60
  61 typedef struct op_metrics
  62 {
  63         /* This happens to also be the index into the
  64          * metrics->op_metrics array currently
  65          */
  66         metrics_op_t op;
  67         double time_ns;
  68
  69         double **counters;
  70 } op_metrics_t;
  71
  72 struct metrics
  73 {
  74         /* Description of all available peformance counters, counter
  75          * groups, their names and IDs, etc. */
  76         metrics_info_t *info;
  77
  78         /* The current operation being measured. */
  79         metrics_op_t op;
  80
  81         /* GL_TIME_ELAPSED query for which glEndQuery has not yet
  82          * been called. */
  83         unsigned timer_begun_id;
  84
  85         /* GL_TIME_ELAPSED queries for which glEndQuery has been
  86          * called, (but results have not yet been queried). */
  87         timer_query_t *timer_head;
  88         timer_query_t *timer_tail;
  89
  90         /* Performance monitor for which glEndPerfMonitorAMD has not
  91          * yet been called. */
  92         unsigned monitor_begun_id;
  93
  94         /* Performance monitors for which glEndPerfMonitorAMD has
  95          * been called, (but results have not yet been queried). */
  96         monitor_t *monitor_head;
  97         monitor_t *monitor_tail;
  98
  99         int monitors_in_flight;
 100
 101         unsigned num_op_metrics;
 102         op_metrics_t *op_metrics;
 103
 104         /* Per-frame time and latency measurement. */
 105         GLint64 swap_begin_timestamp;
 106         unsigned swap_end_timestamp_id;
 107         GLint64 previous_swap_end_timestamp;
 108 };
 109
 110 metrics_t *
 111 metrics_create (metrics_info_t *info)
 112 {
 113         metrics_t *metrics;
 114
 115         metrics = xmalloc (sizeof (metrics_t));
 116
 117         metrics->info = info;
 118
 119         metrics->op = 0;
 120
 121         metrics->timer_begun_id = 0;
 122
 123         metrics->timer_head = NULL;
 124         metrics->timer_tail = NULL;
 125
 126         metrics->monitor_begun_id = 0;
 127
 128         metrics->monitor_head = NULL;
 129         metrics->monitor_tail = NULL;
 130
 131         metrics->monitors_in_flight = 0;
 132
 133         metrics->num_op_metrics = 0;
 134         metrics->op_metrics = NULL;
 135
 136         glGenQueries (1, &metrics->swap_end_timestamp_id);
 137
 138         /* Get the first frame timestamp started immediately. */
 139         glQueryCounter (metrics->swap_end_timestamp_id, GL_TIMESTAMP);
 140
 141         return metrics;
 142 }
 143
 144 void
 145 metrics_fini (metrics_t *metrics)
 146 {
 147         timer_query_t *timer, *timer_next;
 148         monitor_t *monitor, *monitor_next;
 149
 150         /* Discard and cleanup any outstanding queries. */
 151         if (metrics->timer_begun_id) {
 152                 glEndQuery (GL_TIME_ELAPSED);
 153                 glDeleteQueries (1, &metrics->timer_begun_id);
 154                 metrics->timer_begun_id = 0;
 155         }
 156
 157         for (timer = metrics->timer_head;
 158              timer;
 159              timer = timer_next)
 160         {
 161                 glDeleteQueries (1, &timer->id);
 162                 timer_next = timer->next;
 163                 free (timer);
 164         }
 165         metrics->timer_head = NULL;
 166         metrics->timer_tail = NULL;
 167
 168         if (metrics->info->have_perfmon) {
 169
 170                 if (metrics->monitor_begun_id) {
 171                         glEndPerfMonitorAMD (metrics->monitor_begun_id);
 172                         glDeletePerfMonitorsAMD (1, &metrics->monitor_begun_id);
 173                         metrics->monitor_begun_id = 0;
 174                 }
 175
 176                 for (monitor = metrics->monitor_head;
 177                      monitor;
 178                      monitor = monitor_next)
 179                 {
 180                         glDeletePerfMonitorsAMD (1, &monitor->id);
 181                         monitor_next = monitor->next;
 182                         free (monitor);
 183                 }
 184                 metrics->monitor_head = NULL;
 185                 metrics->monitor_tail = NULL;
 186
 187         }
 188
 189         metrics->monitors_in_flight = 0;
 190 }
 191
 192 void
 193 metrics_destroy (metrics_t *metrics)
 194 {
 195         metrics_fini (metrics);
 196
 197         free (metrics);
 198 }
 199
 200 static const char *
 201 metrics_op_string (metrics_op_t op)
 202 {
 203         if (op >= METRICS_OP_SHADER)
 204                 return "Shader program";
 205
 206         switch (op)
 207         {
 208         case METRICS_OP_ACCUM:
 209                 return "glAccum*(+)";
 210         case METRICS_OP_BUFFER_DATA:
 211                 return "glBufferData(+)";
 212         case METRICS_OP_BUFFER_SUB_DATA:
 213                 return "glCopyBufferSubData*";
 214         case METRICS_OP_BITMAP:
 215                 return "glBitmap*";
 216         case METRICS_OP_BLIT_FRAMEBUFFER:
 217                 return "glBlitFramebuffer*";
 218         case METRICS_OP_CLEAR:
 219                 return "glClear(+)";
 220         case METRICS_OP_CLEAR_BUFFER_DATA:
 221                 return "glCearBufferData(+)";
 222         case METRICS_OP_CLEAR_TEX_IMAGE:
 223                 return "glClearTexImage(+)";
 224         case METRICS_OP_COPY_PIXELS:
 225                 return "glCopyPixels";
 226         case METRICS_OP_COPY_TEX_IMAGE:
 227                 return "glCopyTexImage(+)";
 228         case METRICS_OP_DRAW_PIXELS:
 229                 return "glDrawPixels";
 230         case METRICS_OP_GET_TEX_IMAGE:
 231                 return "glGetTexImage(+)";
 232         case METRICS_OP_READ_PIXELS:
 233                 return "glReadPixels*";
 234         case METRICS_OP_TEX_IMAGE:
 235                 return "glTexImage*(+)";
 236         default:
 237                 fprintf (stderr, "fips: Internal error: "
 238                          "Unknown metrics op value: %d\n", op);
 239                 exit (1);
 240         }
 241
 242         return "";
 243 }
 244
 245 void
 246 metrics_counter_start (metrics_t *metrics)
 247 {
 248         unsigned i;
 249
 250         /* Initialize the timer_query object. */
 251         glGenQueries (1, &metrics->timer_begun_id);
 252
 253         /* Most everything else in this function is
 254          * performance-monitor related. If we don't have that
 255          * extension, just start the timer query and be done. */
 256         if (! metrics->info->have_perfmon) {
 257                 glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
 258                 return;
 259         }
 260
 261         /* Initialize the performance-monitor object */
 262         glGenPerfMonitorsAMD (1, &metrics->monitor_begun_id);
 263
 264         for (i = 0; i < metrics->info->num_groups; i++)
 265         {
 266                 metrics_group_info_t *group;
 267                 int num_counters;
 268
 269                 group = &metrics->info->groups[i];
 270
 271                 num_counters = group->num_counters;
 272                 if (group->max_active_counters < group->num_counters)
 273                 {
 274                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 275                                  group->max_active_counters,
 276                                  group->num_counters, i);
 277                         num_counters = group->max_active_counters;
 278
 279                 }
 280
 281                 glSelectPerfMonitorCountersAMD(metrics->monitor_begun_id,
 282                                                GL_TRUE, group->id,
 283                                                num_counters,
 284                                                group->counter_ids);
 285         }
 286
 287         /* Start the queries */
 288         glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
 289
 290         glBeginPerfMonitorAMD (metrics->monitor_begun_id);
 291 }
 292
 293 void
 294 metrics_counter_stop (metrics_t *metrics)
 295 {
 296         timer_query_t *timer;
 297         monitor_t *monitor;
 298
 299         /* Stop the current timer and monitor. */
 300         glEndQuery (GL_TIME_ELAPSED);
 301
 302         if (metrics->info->have_perfmon)
 303                 glEndPerfMonitorAMD (metrics->monitor_begun_id);
 304
 305         /* Add these IDs to our lists of outstanding queries and
 306          * monitors so the results can be collected later. */
 307         timer = xmalloc (sizeof (timer_query_t));
 308
 309         timer->op = metrics->op;
 310         timer->id = metrics->timer_begun_id;
 311         timer->next = NULL;
 312
 313         if (metrics->timer_tail) {
 314                 metrics->timer_tail->next = timer;
 315                 metrics->timer_tail = timer;
 316         } else {
 317                 metrics->timer_tail = timer;
 318                 metrics->timer_head = timer;
 319         }
 320
 321         if (metrics->info->have_perfmon) {
 322                 /* Create a new performance-monitor query */
 323                 monitor = xmalloc (sizeof (monitor_t));
 324
 325                 monitor->op = metrics->op;
 326                 monitor->id = metrics->monitor_begun_id;
 327                 monitor->next = NULL;
 328
 329                 if (metrics->monitor_tail) {
 330                         metrics->monitor_tail->next = monitor;
 331                         metrics->monitor_tail = monitor;
 332                 } else {
 333                         metrics->monitor_tail = monitor;
 334                         metrics->monitor_head = monitor;
 335                 }
 336         }
 337
 338         metrics->monitors_in_flight++;
 339
 340         /* Avoid being a resource hog and collect outstanding results
 341          * once we have sent off a large number of
 342          * queries. (Presumably, many of the outstanding queries are
 343          * available by now.)
 344          */
 345         if (metrics->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
 346                 metrics_collect_available (metrics);
 347 }
 348
 349 void
 350 metrics_set_current_op (metrics_t *metrics, metrics_op_t op)
 351 {
 352         metrics->op = op;
 353 }
 354
 355 metrics_op_t
 356 metrics_get_current_op (metrics_t *metrics)
 357 {
 358         return metrics->op;
 359 }
 360
 361 static void
 362 op_metrics_init (metrics_info_t *info, op_metrics_t *metrics, metrics_op_t op)
 363 {
 364         unsigned i, j;
 365
 366         metrics->op = op;
 367         metrics->time_ns = 0.0;
 368
 369         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 370
 371         for (i = 0; i < info->num_groups; i++) {
 372                 metrics->counters[i] = xmalloc (sizeof (double) *
 373                                                 info->groups[i].num_counters);
 374                 for (j = 0; j < info->groups[i].num_counters; j++)
 375                         metrics->counters[i][j] = 0.0;
 376         }
 377 }
 378
 379 static op_metrics_t *
 380 _get_op_metrics (metrics_t *metrics, metrics_op_t op)
 381 {
 382         unsigned i;
 383
 384         if (op >= metrics->num_op_metrics)
 385         {
 386                 metrics->op_metrics = realloc (metrics->op_metrics,
 387                                                (op + 1) * sizeof (op_metrics_t));
 388                 for (i = metrics->num_op_metrics; i < op + 1; i++)
 389                         op_metrics_init (metrics->info, &metrics->op_metrics[i], i);
 390
 391                 metrics->num_op_metrics = op + 1;
 392         }
 393
 394         return &metrics->op_metrics[op];
 395 }
 396
 397 static void
 398 accumulate_program_metrics (metrics_t *metrics, metrics_op_t op,
 399                             GLuint *result, GLuint size)
 400 {
 401 #define CONSUME(var)                                                    \
 402         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 403         {                                                               \
 404                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 405                          "parsing results\n");                          \
 406                 break;                                                  \
 407         }                                                               \
 408         (var) = *((typeof(var) *) p);                                   \
 409         p += sizeof(var);
 410
 411         metrics_info_t *info = metrics->info;
 412         op_metrics_t *op_metrics = _get_op_metrics (metrics, op);
 413         unsigned char *p = (unsigned char *) result;
 414
 415         while (p < ((unsigned char *) result) + size)
 416         {
 417                 GLuint group_id, group_index;
 418                 GLuint counter_id, counter_index;
 419                 metrics_group_info_t *group;
 420                 double value;
 421                 unsigned i;
 422
 423                 CONSUME (group_id);
 424                 CONSUME (counter_id);
 425
 426                 for (i = 0; i < info->num_groups; i++) {
 427                         if (info->groups[i].id == group_id)
 428                                 break;
 429                 }
 430                 group_index = i;
 431                 assert (group_index < info->num_groups);
 432                 group = &info->groups[group_index];
 433
 434                 for (i = 0; i < group->num_counters; i++) {
 435                         if (group->counter_ids[i] == counter_id)
 436                                 break;
 437                 }
 438                 counter_index = i;
 439                 assert (counter_index < group->num_counters);
 440
 441                 switch (group->counter_types[counter_index])
 442                 {
 443                         uint uint_value;
 444                         uint64_t uint64_value;
 445                         float float_value;
 446                 case GL_UNSIGNED_INT:
 447                         CONSUME (uint_value);
 448                         value = uint_value;
 449                         break;
 450                 case GL_UNSIGNED_INT64_AMD:
 451                         CONSUME (uint64_value);
 452                         value = uint64_value;
 453                         break;
 454                 case GL_PERCENTAGE_AMD:
 455                 case GL_FLOAT:
 456                         CONSUME (float_value);
 457                         value = float_value;
 458                         break;
 459                 default:
 460                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 461                                  group->counter_types[counter_index]);
 462                         value = 0.0;
 463                         break;
 464                 }
 465
 466                 op_metrics->counters[group_index][counter_index] += value;
 467         }
 468 }
 469
 470 static void
 471 accumulate_program_time (metrics_t *metrics, metrics_op_t op, unsigned time_ns)
 472 {
 473         op_metrics_t *op_metrics;
 474
 475         op_metrics = _get_op_metrics (metrics, op);
 476
 477         op_metrics->time_ns += time_ns;
 478 }
 479
 480 typedef struct per_stage_metrics
 481 {
 482         op_metrics_t *metrics;
 483         shader_stage_info_t *stage;
 484         double time_ns;
 485         double active;
 486 } per_stage_metrics_t;
 487
 488 static int
 489 _is_shader_stage_counter (metrics_info_t *info,
 490                           unsigned group_index,
 491                           unsigned counter_index)
 492 {
 493         shader_stage_info_t *stage;
 494         unsigned i;
 495
 496         for (i = 0; i < info->num_shader_stages; i++) {
 497                 stage = &info->stages[i];
 498
 499                 if (stage->active_group_index == group_index &&
 500                     stage->active_counter_index == counter_index)
 501                 {
 502                         return 1;
 503                 }
 504
 505                 if (stage->stall_group_index == group_index &&
 506                     stage->stall_counter_index == counter_index)
 507                 {
 508                         return 1;
 509                 }
 510         }
 511
 512         return 0;
 513 }
 514
 515 static void
 516 print_per_stage_metrics (metrics_t *metrics,
 517                          per_stage_metrics_t *per_stage,
 518                          double total)
 519 {
 520         metrics_info_t *info = metrics->info;
 521         op_metrics_t *op_metrics = per_stage->metrics;
 522         metrics_group_info_t *group;
 523         const char *op_string;
 524         unsigned group_index, counter;
 525         double value;
 526
 527         /* Don't print anything for stages with no alloted time. */
 528         if (per_stage->time_ns == 0.0)
 529                 return;
 530
 531         op_string = metrics_op_string (op_metrics->op);
 532
 533         printf ("%21s", op_string);
 534
 535         if (op_metrics->op >= METRICS_OP_SHADER) {
 536                 printf (" %3d", op_metrics->op - METRICS_OP_SHADER);
 537         } else {
 538                 printf ("    ");
 539
 540         }
 541
 542         if (per_stage->stage)
 543                 printf (" %cS:", per_stage->stage->name[0]);
 544         else
 545                 printf ("   :");
 546
 547         printf ("\t%7.2f ms (%4.1f%%)",
 548                 per_stage->time_ns / 1e6,
 549                 per_stage->time_ns / total * 100);
 550
 551         if (per_stage->active)
 552                 printf (", %4.1f%% active", per_stage->active * 100);
 553
 554         printf ("\n");
 555
 556         /* I'm not seeing a lot of value printing the rest of these
 557          * performance counters by default yet. Use --verbose to get
 558          * them for now. */
 559         if (! verbose)
 560                 return;
 561
 562         printf ("[");
 563         for (group_index = 0; group_index < info->num_groups; group_index++) {
 564                 group = &info->groups[group_index];
 565                 for (counter = 0; counter < group->num_counters; counter++) {
 566
 567                         /* Don't print this counter value if it's a
 568                          * per-stage cycle counter, (which we have
 569                          * already accounted for). */
 570                         if (_is_shader_stage_counter (info, group_index, counter))
 571                                 continue;
 572
 573                         value = op_metrics->counters[group_index][counter];
 574                         if (value == 0.0)
 575                                 continue;
 576                         printf ("%s: %.2f ", group->counter_names[counter],
 577                                 value / 1e6);
 578                 }
 579         }
 580         printf ("]\n");
 581 }
 582
 583 static int
 584 time_compare(const void *in_a, const void *in_b, void *arg unused)
 585 {
 586         const per_stage_metrics_t *a = in_a;
 587         const per_stage_metrics_t *b = in_b;
 588
 589
 590         if (a->time_ns < b->time_ns)
 591                 return -1;
 592         if (a->time_ns > b->time_ns)
 593                 return 1;
 594         return 0;
 595 }
 596
 597 static void
 598 print_program_metrics (metrics_t *metrics)
 599 {
 600         metrics_info_t *info = metrics->info;
 601         unsigned num_shader_stages = info->num_shader_stages;
 602         per_stage_metrics_t *sorted, *per_stage;
 603         double total_time, op_cycles;
 604         op_metrics_t *op;
 605         unsigned group_index, counter_index;
 606         unsigned i, j, num_sorted;
 607
 608         if (! info->have_perfmon &&
 609             ! info->printed_missing_perfmon_warning)
 610         {
 611                 fprintf (stderr,
 612                          "Warning: The GL_AMD_performance_monitor extensions is not available.\n"
 613                          "         Some fips features, (such as per-shader-stage timings),\n"
 614                          "         will not be available.\n");
 615
 616                 info->printed_missing_perfmon_warning = true;
 617         }
 618
 619         /* Make a sorted list of the per-stage operations by time
 620          * used, and figure out the total so we can print percentages.
 621          */
 622         if (num_shader_stages)
 623                 num_sorted = metrics->num_op_metrics * num_shader_stages;
 624         else
 625                 num_sorted = metrics->num_op_metrics;
 626
 627         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 628
 629         total_time = 0.0;
 630
 631         for (i = 0; i < metrics->num_op_metrics; i++) {
 632
 633                 op = &metrics->op_metrics[i];
 634
 635                 /* Accumulate total time across all ops. */
 636                 total_time += op->time_ns;
 637
 638                 /* Also, find total cycles in all stages of this op. */
 639                 op_cycles = 0.0;
 640
 641                 if (num_shader_stages == 0) {
 642                         per_stage = &sorted[i];
 643                         per_stage->metrics = op;
 644                         per_stage->stage = NULL;
 645                         per_stage->time_ns = op->time_ns;
 646                         per_stage->active = 0.0;
 647                 }
 648
 649                 for (j = 0; j < num_shader_stages; j++) {
 650                         /* Active cycles */
 651                         group_index = info->stages[j].active_group_index;
 652                         counter_index = info->stages[j].active_counter_index;
 653                         op_cycles += op->counters[group_index][counter_index];
 654
 655                         /* Stall cycles */
 656                         group_index = info->stages[j].stall_group_index;
 657                         counter_index = info->stages[j].stall_counter_index;
 658                         op_cycles += op->counters[group_index][counter_index];
 659                 }
 660
 661                 for (j = 0; j < num_shader_stages; j++) {
 662                         double active_cycles, stall_cycles, stage_cycles;
 663
 664                         /* Active cycles */
 665                         group_index = info->stages[j].active_group_index;
 666                         counter_index = info->stages[j].active_counter_index;
 667                         active_cycles = op->counters[group_index][counter_index];
 668
 669                         /* Stall cycles */
 670                         group_index = info->stages[j].stall_group_index;
 671                         counter_index = info->stages[j].stall_counter_index;
 672                         stall_cycles = op->counters[group_index][counter_index];
 673
 674                         stage_cycles = active_cycles + stall_cycles;
 675
 676                         per_stage = &sorted[i * num_shader_stages + j];
 677                         per_stage->metrics = op;
 678
 679                         if (op_cycles) {
 680                                 per_stage->stage = &info->stages[j];
 681                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 682                         } else {
 683                                 /* If we don't have any per-stage cycle counts
 684                                  * for this operation, then use the first
 685                                  * stage as a placeholder for all the time,
 686                                  * but NULL-ify the stage info so that the
 687                                  * report doesn't lie about this time being
 688                                  * from any particular stage. */
 689                                 per_stage->stage = NULL;
 690                                 if (j == 0) {
 691                                         per_stage->time_ns = op->time_ns;
 692                                 } else {
 693                                         per_stage->time_ns = 0.0;
 694                                 }
 695                         }
 696
 697                         if (stage_cycles) {
 698                                 per_stage->active = active_cycles / stage_cycles;
 699                         } else {
 700                                 per_stage->active = 0.0;
 701                         }
 702                 }
 703         }
 704
 705         qsort_r (sorted, num_sorted, sizeof (*sorted),
 706                  time_compare, metrics->op_metrics);
 707
 708         for (i = 0; i < num_sorted; i++)
 709                 print_per_stage_metrics (metrics, &sorted[i], total_time);
 710
 711         free (sorted);
 712 }
 713
 714 void
 715 metrics_collect_available (metrics_t *metrics)
 716 {
 717         /* Consume all timer queries that are ready. */
 718         timer_query_t *timer = metrics->timer_head;
 719
 720         while (timer) {
 721                 GLuint available, elapsed;
 722
 723                 glGetQueryObjectuiv (timer->id,
 724                                      GL_QUERY_RESULT_AVAILABLE, &available);
 725                 if (! available)
 726                         break;
 727
 728                 glGetQueryObjectuiv (timer->id,
 729                                      GL_QUERY_RESULT, &elapsed);
 730
 731                 accumulate_program_time (metrics, timer->op, elapsed);
 732
 733                 metrics->timer_head = timer->next;
 734                 if (metrics->timer_head == NULL)
 735                         metrics->timer_tail = NULL;
 736
 737                 glDeleteQueries (1, &timer->id);
 738
 739                 free (timer);
 740                 timer = metrics->timer_head;
 741         }
 742
 743         if (! metrics->info->have_perfmon)
 744                 return;
 745
 746         /* And similarly for all performance monitors that are ready. */
 747         monitor_t *monitor = metrics->monitor_head;
 748
 749         while (monitor) {
 750                 GLuint available, result_size, *result;
 751                 GLint bytes_written;
 752
 753                 glGetPerfMonitorCounterDataAMD (monitor->id,
 754                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 755                                                 sizeof (available), &available,
 756                                                 NULL);
 757                 if (! available)
 758                         break;
 759
 760                 glGetPerfMonitorCounterDataAMD (monitor->id,
 761                                                 GL_PERFMON_RESULT_SIZE_AMD,
 762                                                 sizeof (result_size),
 763                                                 &result_size, NULL);
 764
 765                 result = xmalloc (result_size);
 766
 767                 glGetPerfMonitorCounterDataAMD (monitor->id,
 768                                                 GL_PERFMON_RESULT_AMD,
 769                                                 result_size, result,
 770                                                 &bytes_written);
 771
 772                 accumulate_program_metrics (metrics, monitor->op, result, result_size);
 773
 774                 free (result);
 775
 776                 metrics->monitor_head = monitor->next;
 777                 if (metrics->monitor_head == NULL)
 778                         metrics->monitor_tail = NULL;
 779
 780                 glDeletePerfMonitorsAMD (1, &monitor->id);
 781
 782                 free (monitor);
 783
 784                 metrics->monitors_in_flight--;
 785
 786                 monitor = metrics->monitor_head;
 787         }
 788 }
 789
 790 void
 791 metrics_end_frame_pre_swap (metrics_t *metrics)
 792 {
 793         GLuint64 swap_end_timestamp, frame_time_ns, latency_ns;
 794
 795         /* Don't leave any counters running over the end_frame work we
 796          * do here. The counters will be started again at the end of
 797          * metrics_end_frame_post_swap.  */
 798
 799         metrics_counter_stop (metrics);
 800
 801         /* Now that an entire frame's worth of content has gone by, we
 802          * can be sure that the timer quiery for the previous frame's
 803          * swap is available.
 804          *
 805          * Note: The only case in which this query isn't immediately
 806          * available would be if there was effectively nothing in this
 807          * frame. In that case, triggering a block on this query
 808          * result is not a concern (since an empty frame is not
 809          * interesting, and all of our counters are stopped anyway).
 810          *
 811          * Accepting this one block is much better than adding a
 812          * linked list for these results that would reaslisticly never
 813          * have more than one entry anyway.
 814          */
 815
 816         glGetQueryObjectui64v (metrics->swap_end_timestamp_id,
 817                                GL_QUERY_RESULT, &swap_end_timestamp);
 818
 819         if (frames > 0) {
 820                 /* Subtract previous frame's timestamp to get frame time. */
 821                 frame_time_ns = swap_end_timestamp - metrics->previous_swap_end_timestamp;
 822
 823                 latency_ns = swap_end_timestamp - metrics->swap_begin_timestamp;
 824
 825                 /* We've waited one frame to ensure we have a timestamp
 826                  * result. So the time we've actually measured here is
 827                  * for the previous frame. */
 828                 printf ("Frame %d: %.2f ms (latency: %.2f ms)\n",
 829                         frames - 1, frame_time_ns / 1e6, latency_ns / 1e6);
 830         }
 831
 832         metrics->previous_swap_end_timestamp = swap_end_timestamp;
 833
 834         /* Before the actual swap call, we get the current timestamp
 835          * value. This is a synchronous get so we can use this as the
 836          * baseline for a frame latency measurment. */
 837
 838         glGetInteger64v (GL_TIMESTAMP, &metrics->swap_begin_timestamp);
 839 }
 840
 841 void
 842 metrics_end_frame_post_swap (metrics_t *metrics)
 843 {
 844         static int initialized = 0;
 845         static struct timeval tv_start, tv_now;
 846
 847         /* Now that the swap command has been queued, we issue an
 848          * asynchronous query of the timestamp value. Comparing this
 849          * to the synchronous get we just sent in
 850          * metrics_end_frame_pre_swap allows us to measure the
 851          * per-frame swap latency. */
 852
 853         glQueryCounter (metrics->swap_end_timestamp_id, GL_TIMESTAMP);
 854
 855         if (! initialized) {
 856                 gettimeofday (&tv_start, NULL);
 857                 if (getenv ("FIPS_VERBOSE"))
 858                         verbose = 1;
 859                 initialized = 1;
 860         }
 861
 862         metrics_collect_available (metrics);
 863
 864         frames++;
 865
 866         if (frames % 15 == 0) {
 867                 double fps;
 868
 869                 gettimeofday (&tv_now, NULL);
 870
 871                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 872                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 873
 874                 printf("FPS: %.3f\n", fps);
 875
 876                 print_program_metrics (metrics);
 877         }
 878
 879         /* Start the counter up again now that we're done. */
 880         metrics_counter_start (metrics);
 881 }