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 <time.h>
  30 #include <sys/time.h>
  31 #include <string.h>
  32 #include <errno.h>
  33
  34 #include "fips-dispatch-gl.h"
  35
  36 #include "metrics.h"
  37 #include "context.h"
  38 #include "metrics-info.h"
  39 #include "xmalloc.h"
  40
  41 int frames;
  42 int verbose;
  43
  44 #define MAX_MONITORS_IN_FLIGHT 1000
  45
  46 /* Timer query */
  47 typedef struct timer_query
  48 {
  49         unsigned id;
  50
  51         metrics_op_t op;
  52         struct timer_query *next;
  53 } timer_query_t;
  54
  55 /* Performance-monitor query */
  56 typedef struct monitor
  57 {
  58         unsigned id;
  59
  60         metrics_op_t op;
  61         struct monitor *next;
  62 } monitor_t;
  63
  64 typedef struct op_metrics
  65 {
  66         /* This happens to also be the index into the
  67          * metrics->op_metrics array currently
  68          */
  69         metrics_op_t op;
  70         double time_ns;
  71
  72         double **counters;
  73 } op_metrics_t;
  74
  75 struct metrics
  76 {
  77         /* Description of all available peformance counters, counter
  78          * groups, their names and IDs, etc. */
  79         metrics_info_t *info;
  80
  81         /* The current operation being measured. */
  82         metrics_op_t op;
  83
  84         /* GL_TIME_ELAPSED query for which glEndQuery has not yet
  85          * been called. */
  86         unsigned timer_begun_id;
  87
  88         /* GL_TIME_ELAPSED queries for which glEndQuery has been
  89          * called, (but results have not yet been queried). */
  90         timer_query_t *timer_head;
  91         timer_query_t *timer_tail;
  92
  93         /* Performance monitor for which glEndPerfMonitorAMD has not
  94          * yet been called. */
  95         unsigned monitor_begun_id;
  96
  97         /* Performance monitors for which glEndPerfMonitorAMD has
  98          * been called, (but results have not yet been queried). */
  99         monitor_t *monitor_head;
 100         monitor_t *monitor_tail;
 101
 102         int monitors_in_flight;
 103
 104         unsigned num_op_metrics;
 105         op_metrics_t *op_metrics;
 106
 107         /* Per-frame time and latency measurement. */
 108         GLint64 swap_begin_timestamp;
 109         unsigned swap_end_timestamp_id;
 110         GLint64 previous_swap_end_timestamp;
 111
 112         struct timespec previous_cpu_time_ts;
 113         struct timespec cpu_time_ts;
 114 };
 115
 116 metrics_t *
 117 metrics_create (metrics_info_t *info)
 118 {
 119         metrics_t *metrics;
 120
 121         metrics = xmalloc (sizeof (metrics_t));
 122
 123         metrics->info = info;
 124
 125         metrics->op = 0;
 126
 127         metrics->timer_begun_id = 0;
 128
 129         metrics->timer_head = NULL;
 130         metrics->timer_tail = NULL;
 131
 132         metrics->monitor_begun_id = 0;
 133
 134         metrics->monitor_head = NULL;
 135         metrics->monitor_tail = NULL;
 136
 137         metrics->monitors_in_flight = 0;
 138
 139         metrics->num_op_metrics = 0;
 140         metrics->op_metrics = NULL;
 141
 142         glGenQueries (1, &metrics->swap_end_timestamp_id);
 143
 144         /* Get the first frame timestamp started immediately. */
 145         glQueryCounter (metrics->swap_end_timestamp_id, GL_TIMESTAMP);
 146
 147         return metrics;
 148 }
 149
 150 void
 151 metrics_fini (metrics_t *metrics)
 152 {
 153         timer_query_t *timer, *timer_next;
 154         monitor_t *monitor, *monitor_next;
 155
 156         /* Discard and cleanup any outstanding queries. */
 157         if (metrics->timer_begun_id) {
 158                 glEndQuery (GL_TIME_ELAPSED);
 159                 glDeleteQueries (1, &metrics->timer_begun_id);
 160                 metrics->timer_begun_id = 0;
 161         }
 162
 163         for (timer = metrics->timer_head;
 164              timer;
 165              timer = timer_next)
 166         {
 167                 glDeleteQueries (1, &timer->id);
 168                 timer_next = timer->next;
 169                 free (timer);
 170         }
 171         metrics->timer_head = NULL;
 172         metrics->timer_tail = NULL;
 173
 174         if (metrics->info->have_perfmon) {
 175
 176                 if (metrics->monitor_begun_id) {
 177                         glEndPerfMonitorAMD (metrics->monitor_begun_id);
 178                         glDeletePerfMonitorsAMD (1, &metrics->monitor_begun_id);
 179                         metrics->monitor_begun_id = 0;
 180                 }
 181
 182                 for (monitor = metrics->monitor_head;
 183                      monitor;
 184                      monitor = monitor_next)
 185                 {
 186                         glDeletePerfMonitorsAMD (1, &monitor->id);
 187                         monitor_next = monitor->next;
 188                         free (monitor);
 189                 }
 190                 metrics->monitor_head = NULL;
 191                 metrics->monitor_tail = NULL;
 192
 193         }
 194
 195         metrics->monitors_in_flight = 0;
 196 }
 197
 198 void
 199 metrics_destroy (metrics_t *metrics)
 200 {
 201         metrics_fini (metrics);
 202
 203         free (metrics);
 204 }
 205
 206 static const char *
 207 metrics_op_string (metrics_op_t op)
 208 {
 209         if (op >= METRICS_OP_SHADER)
 210                 return "Shader program";
 211
 212         switch (op)
 213         {
 214         case METRICS_OP_ACCUM:
 215                 return "glAccum*(+)";
 216         case METRICS_OP_BUFFER_DATA:
 217                 return "glBufferData(+)";
 218         case METRICS_OP_BUFFER_SUB_DATA:
 219                 return "glCopyBufferSubData*";
 220         case METRICS_OP_BITMAP:
 221                 return "glBitmap*";
 222         case METRICS_OP_BLIT_FRAMEBUFFER:
 223                 return "glBlitFramebuffer*";
 224         case METRICS_OP_CLEAR:
 225                 return "glClear(+)";
 226         case METRICS_OP_CLEAR_BUFFER_DATA:
 227                 return "glCearBufferData(+)";
 228         case METRICS_OP_CLEAR_TEX_IMAGE:
 229                 return "glClearTexImage(+)";
 230         case METRICS_OP_COPY_PIXELS:
 231                 return "glCopyPixels";
 232         case METRICS_OP_COPY_TEX_IMAGE:
 233                 return "glCopyTexImage(+)";
 234         case METRICS_OP_DRAW_PIXELS:
 235                 return "glDrawPixels";
 236         case METRICS_OP_GET_TEX_IMAGE:
 237                 return "glGetTexImage(+)";
 238         case METRICS_OP_READ_PIXELS:
 239                 return "glReadPixels*";
 240         case METRICS_OP_TEX_IMAGE:
 241                 return "glTexImage*(+)";
 242         default:
 243                 fprintf (stderr, "fips: Internal error: "
 244                          "Unknown metrics op value: %d\n", op);
 245                 exit (1);
 246         }
 247
 248         return "";
 249 }
 250
 251 void
 252 metrics_counter_start (metrics_t *metrics)
 253 {
 254         unsigned i;
 255
 256         /* Initialize the timer_query object. */
 257         glGenQueries (1, &metrics->timer_begun_id);
 258
 259         /* Most everything else in this function is
 260          * performance-monitor related. If we don't have that
 261          * extension, just start the timer query and be done. */
 262         if (! metrics->info->have_perfmon) {
 263                 glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
 264                 return;
 265         }
 266
 267         /* Initialize the performance-monitor object */
 268         glGenPerfMonitorsAMD (1, &metrics->monitor_begun_id);
 269
 270         for (i = 0; i < metrics->info->num_groups; i++)
 271         {
 272                 metrics_group_info_t *group;
 273                 int num_counters;
 274
 275                 group = &metrics->info->groups[i];
 276
 277                 num_counters = group->num_counters;
 278                 if (group->max_active_counters < group->num_counters)
 279                 {
 280                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 281                                  group->max_active_counters,
 282                                  group->num_counters, i);
 283                         num_counters = group->max_active_counters;
 284
 285                 }
 286
 287                 glSelectPerfMonitorCountersAMD(metrics->monitor_begun_id,
 288                                                GL_TRUE, group->id,
 289                                                num_counters,
 290                                                group->counter_ids);
 291         }
 292
 293         /* Start the queries */
 294         glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
 295
 296         glBeginPerfMonitorAMD (metrics->monitor_begun_id);
 297 }
 298
 299 void
 300 metrics_counter_stop (metrics_t *metrics)
 301 {
 302         timer_query_t *timer;
 303         monitor_t *monitor;
 304
 305         /* Stop the current timer and monitor. */
 306         glEndQuery (GL_TIME_ELAPSED);
 307
 308         if (metrics->info->have_perfmon)
 309                 glEndPerfMonitorAMD (metrics->monitor_begun_id);
 310
 311         /* Add these IDs to our lists of outstanding queries and
 312          * monitors so the results can be collected later. */
 313         timer = xmalloc (sizeof (timer_query_t));
 314
 315         timer->op = metrics->op;
 316         timer->id = metrics->timer_begun_id;
 317         timer->next = NULL;
 318
 319         if (metrics->timer_tail) {
 320                 metrics->timer_tail->next = timer;
 321                 metrics->timer_tail = timer;
 322         } else {
 323                 metrics->timer_tail = timer;
 324                 metrics->timer_head = timer;
 325         }
 326
 327         if (metrics->info->have_perfmon) {
 328                 /* Create a new performance-monitor query */
 329                 monitor = xmalloc (sizeof (monitor_t));
 330
 331                 monitor->op = metrics->op;
 332                 monitor->id = metrics->monitor_begun_id;
 333                 monitor->next = NULL;
 334
 335                 if (metrics->monitor_tail) {
 336                         metrics->monitor_tail->next = monitor;
 337                         metrics->monitor_tail = monitor;
 338                 } else {
 339                         metrics->monitor_tail = monitor;
 340                         metrics->monitor_head = monitor;
 341                 }
 342         }
 343
 344         metrics->monitors_in_flight++;
 345
 346         /* Avoid being a resource hog and collect outstanding results
 347          * once we have sent off a large number of
 348          * queries. (Presumably, many of the outstanding queries are
 349          * available by now.)
 350          */
 351         if (metrics->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
 352                 metrics_collect_available (metrics);
 353 }
 354
 355 void
 356 metrics_set_current_op (metrics_t *metrics, metrics_op_t op)
 357 {
 358         metrics->op = op;
 359 }
 360
 361 metrics_op_t
 362 metrics_get_current_op (metrics_t *metrics)
 363 {
 364         return metrics->op;
 365 }
 366
 367 static void
 368 op_metrics_init (metrics_info_t *info, op_metrics_t *metrics, metrics_op_t op)
 369 {
 370         unsigned i, j;
 371
 372         metrics->op = op;
 373         metrics->time_ns = 0.0;
 374
 375         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 376
 377         for (i = 0; i < info->num_groups; i++) {
 378                 metrics->counters[i] = xmalloc (sizeof (double) *
 379                                                 info->groups[i].num_counters);
 380                 for (j = 0; j < info->groups[i].num_counters; j++)
 381                         metrics->counters[i][j] = 0.0;
 382         }
 383 }
 384
 385 static op_metrics_t *
 386 _get_op_metrics (metrics_t *metrics, metrics_op_t op)
 387 {
 388         unsigned i;
 389
 390         if (op >= metrics->num_op_metrics)
 391         {
 392                 metrics->op_metrics = realloc (metrics->op_metrics,
 393                                                (op + 1) * sizeof (op_metrics_t));
 394                 for (i = metrics->num_op_metrics; i < op + 1; i++)
 395                         op_metrics_init (metrics->info, &metrics->op_metrics[i], i);
 396
 397                 metrics->num_op_metrics = op + 1;
 398         }
 399
 400         return &metrics->op_metrics[op];
 401 }
 402
 403 static void
 404 accumulate_program_metrics (metrics_t *metrics, metrics_op_t op,
 405                             GLuint *result, GLuint size)
 406 {
 407 #define CONSUME(var)                                                    \
 408         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 409         {                                                               \
 410                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 411                          "parsing results\n");                          \
 412                 break;                                                  \
 413         }                                                               \
 414         (var) = *((typeof(var) *) p);                                   \
 415         p += sizeof(var);
 416
 417         metrics_info_t *info = metrics->info;
 418         op_metrics_t *op_metrics = _get_op_metrics (metrics, op);
 419         unsigned char *p = (unsigned char *) result;
 420
 421         while (p < ((unsigned char *) result) + size)
 422         {
 423                 GLuint group_id, group_index;
 424                 GLuint counter_id, counter_index;
 425                 metrics_group_info_t *group;
 426                 double value;
 427                 unsigned i;
 428
 429                 CONSUME (group_id);
 430                 CONSUME (counter_id);
 431
 432                 for (i = 0; i < info->num_groups; i++) {
 433                         if (info->groups[i].id == group_id)
 434                                 break;
 435                 }
 436                 group_index = i;
 437                 assert (group_index < info->num_groups);
 438                 group = &info->groups[group_index];
 439
 440                 for (i = 0; i < group->num_counters; i++) {
 441                         if (group->counter_ids[i] == counter_id)
 442                                 break;
 443                 }
 444                 counter_index = i;
 445                 assert (counter_index < group->num_counters);
 446
 447                 switch (group->counter_types[counter_index])
 448                 {
 449                         uint uint_value;
 450                         uint64_t uint64_value;
 451                         float float_value;
 452                 case GL_UNSIGNED_INT:
 453                         CONSUME (uint_value);
 454                         value = uint_value;
 455                         break;
 456                 case GL_UNSIGNED_INT64_AMD:
 457                         CONSUME (uint64_value);
 458                         value = uint64_value;
 459                         break;
 460                 case GL_PERCENTAGE_AMD:
 461                 case GL_FLOAT:
 462                         CONSUME (float_value);
 463                         value = float_value;
 464                         break;
 465                 default:
 466                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 467                                  group->counter_types[counter_index]);
 468                         value = 0.0;
 469                         break;
 470                 }
 471
 472                 op_metrics->counters[group_index][counter_index] += value;
 473         }
 474 }
 475
 476 static void
 477 accumulate_program_time (metrics_t *metrics, metrics_op_t op, unsigned time_ns)
 478 {
 479         op_metrics_t *op_metrics;
 480
 481         op_metrics = _get_op_metrics (metrics, op);
 482
 483         op_metrics->time_ns += time_ns;
 484 }
 485
 486 typedef struct per_stage_metrics
 487 {
 488         op_metrics_t *metrics;
 489         shader_stage_info_t *stage;
 490         double time_ns;
 491         double active;
 492 } per_stage_metrics_t;
 493
 494 static int
 495 _is_shader_stage_counter (metrics_info_t *info,
 496                           unsigned group_index,
 497                           unsigned counter_index)
 498 {
 499         shader_stage_info_t *stage;
 500         unsigned i;
 501
 502         for (i = 0; i < info->num_shader_stages; i++) {
 503                 stage = &info->stages[i];
 504
 505                 if (stage->active_group_index == group_index &&
 506                     stage->active_counter_index == counter_index)
 507                 {
 508                         return 1;
 509                 }
 510
 511                 if (stage->stall_group_index == group_index &&
 512                     stage->stall_counter_index == counter_index)
 513                 {
 514                         return 1;
 515                 }
 516         }
 517
 518         return 0;
 519 }
 520
 521 static void
 522 print_per_stage_metrics (metrics_t *metrics,
 523                          per_stage_metrics_t *per_stage,
 524                          double total)
 525 {
 526         metrics_info_t *info = metrics->info;
 527         op_metrics_t *op_metrics = per_stage->metrics;
 528         metrics_group_info_t *group;
 529         const char *op_string;
 530         unsigned group_index, counter;
 531         double value;
 532
 533         /* Don't print anything for stages with no alloted time. */
 534         if (per_stage->time_ns == 0.0)
 535                 return;
 536
 537         op_string = metrics_op_string (op_metrics->op);
 538
 539         printf ("%21s", op_string);
 540
 541         if (op_metrics->op >= METRICS_OP_SHADER) {
 542                 printf (" %3d", op_metrics->op - METRICS_OP_SHADER);
 543         } else {
 544                 printf ("    ");
 545
 546         }
 547
 548         if (per_stage->stage)
 549                 printf (" %cS:", per_stage->stage->name[0]);
 550         else
 551                 printf ("   :");
 552
 553         printf ("\t%7.2f ms (%4.1f%%)",
 554                 per_stage->time_ns / 1e6,
 555                 per_stage->time_ns / total * 100);
 556
 557         if (per_stage->active)
 558                 printf (", %4.1f%% active", per_stage->active * 100);
 559
 560         printf ("\n");
 561
 562         /* I'm not seeing a lot of value printing the rest of these
 563          * performance counters by default yet. Use --verbose to get
 564          * them for now. */
 565         if (! verbose)
 566                 return;
 567
 568         printf ("[");
 569         for (group_index = 0; group_index < info->num_groups; group_index++) {
 570                 group = &info->groups[group_index];
 571                 for (counter = 0; counter < group->num_counters; counter++) {
 572
 573                         /* Don't print this counter value if it's a
 574                          * per-stage cycle counter, (which we have
 575                          * already accounted for). */
 576                         if (_is_shader_stage_counter (info, group_index, counter))
 577                                 continue;
 578
 579                         value = op_metrics->counters[group_index][counter];
 580                         if (value == 0.0)
 581                                 continue;
 582                         printf ("%s: %.2f ", group->counter_names[counter],
 583                                 value / 1e6);
 584                 }
 585         }
 586         printf ("]\n");
 587 }
 588
 589 static int
 590 time_compare(const void *in_a, const void *in_b, void *arg unused)
 591 {
 592         const per_stage_metrics_t *a = in_a;
 593         const per_stage_metrics_t *b = in_b;
 594
 595
 596         if (a->time_ns < b->time_ns)
 597                 return -1;
 598         if (a->time_ns > b->time_ns)
 599                 return 1;
 600         return 0;
 601 }
 602
 603 static void
 604 print_program_metrics (metrics_t *metrics)
 605 {
 606         metrics_info_t *info = metrics->info;
 607         unsigned num_shader_stages = info->num_shader_stages;
 608         per_stage_metrics_t *sorted, *per_stage;
 609         double total_time, op_cycles;
 610         op_metrics_t *op;
 611         unsigned group_index, counter_index;
 612         unsigned i, j, num_sorted;
 613
 614         if (! info->have_perfmon &&
 615             ! info->printed_missing_perfmon_warning)
 616         {
 617                 fprintf (stderr,
 618                          "Warning: The GL_AMD_performance_monitor extensions is not available.\n"
 619                          "         Some fips features, (such as per-shader-stage timings),\n"
 620                          "         will not be available.\n");
 621
 622                 info->printed_missing_perfmon_warning = true;
 623         }
 624
 625         /* Make a sorted list of the per-stage operations by time
 626          * used, and figure out the total so we can print percentages.
 627          */
 628         if (num_shader_stages)
 629                 num_sorted = metrics->num_op_metrics * num_shader_stages;
 630         else
 631                 num_sorted = metrics->num_op_metrics;
 632
 633         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 634
 635         total_time = 0.0;
 636
 637         for (i = 0; i < metrics->num_op_metrics; i++) {
 638
 639                 op = &metrics->op_metrics[i];
 640
 641                 /* Accumulate total time across all ops. */
 642                 total_time += op->time_ns;
 643
 644                 /* Also, find total cycles in all stages of this op. */
 645                 op_cycles = 0.0;
 646
 647                 if (num_shader_stages == 0) {
 648                         per_stage = &sorted[i];
 649                         per_stage->metrics = op;
 650                         per_stage->stage = NULL;
 651                         per_stage->time_ns = op->time_ns;
 652                         per_stage->active = 0.0;
 653                 }
 654
 655                 for (j = 0; j < num_shader_stages; j++) {
 656                         /* Active cycles */
 657                         group_index = info->stages[j].active_group_index;
 658                         counter_index = info->stages[j].active_counter_index;
 659                         op_cycles += op->counters[group_index][counter_index];
 660
 661                         /* Stall cycles */
 662                         group_index = info->stages[j].stall_group_index;
 663                         counter_index = info->stages[j].stall_counter_index;
 664                         op_cycles += op->counters[group_index][counter_index];
 665                 }
 666
 667                 for (j = 0; j < num_shader_stages; j++) {
 668                         double active_cycles, stall_cycles, stage_cycles;
 669
 670                         /* Active cycles */
 671                         group_index = info->stages[j].active_group_index;
 672                         counter_index = info->stages[j].active_counter_index;
 673                         active_cycles = op->counters[group_index][counter_index];
 674
 675                         /* Stall cycles */
 676                         group_index = info->stages[j].stall_group_index;
 677                         counter_index = info->stages[j].stall_counter_index;
 678                         stall_cycles = op->counters[group_index][counter_index];
 679
 680                         stage_cycles = active_cycles + stall_cycles;
 681
 682                         per_stage = &sorted[i * num_shader_stages + j];
 683                         per_stage->metrics = op;
 684
 685                         if (op_cycles) {
 686                                 per_stage->stage = &info->stages[j];
 687                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 688                         } else {
 689                                 /* If we don't have any per-stage cycle counts
 690                                  * for this operation, then use the first
 691                                  * stage as a placeholder for all the time,
 692                                  * but NULL-ify the stage info so that the
 693                                  * report doesn't lie about this time being
 694                                  * from any particular stage. */
 695                                 per_stage->stage = NULL;
 696                                 if (j == 0) {
 697                                         per_stage->time_ns = op->time_ns;
 698                                 } else {
 699                                         per_stage->time_ns = 0.0;
 700                                 }
 701                         }
 702
 703                         if (stage_cycles) {
 704                                 per_stage->active = active_cycles / stage_cycles;
 705                         } else {
 706                                 per_stage->active = 0.0;
 707                         }
 708                 }
 709         }
 710
 711         qsort_r (sorted, num_sorted, sizeof (*sorted),
 712                  time_compare, metrics->op_metrics);
 713
 714         for (i = 0; i < num_sorted; i++)
 715                 print_per_stage_metrics (metrics, &sorted[i], total_time);
 716
 717         free (sorted);
 718 }
 719
 720 void
 721 metrics_collect_available (metrics_t *metrics)
 722 {
 723         /* Consume all timer queries that are ready. */
 724         timer_query_t *timer = metrics->timer_head;
 725
 726         while (timer) {
 727                 GLuint available, elapsed;
 728
 729                 glGetQueryObjectuiv (timer->id,
 730                                      GL_QUERY_RESULT_AVAILABLE, &available);
 731                 if (! available)
 732                         break;
 733
 734                 glGetQueryObjectuiv (timer->id,
 735                                      GL_QUERY_RESULT, &elapsed);
 736
 737                 accumulate_program_time (metrics, timer->op, elapsed);
 738
 739                 metrics->timer_head = timer->next;
 740                 if (metrics->timer_head == NULL)
 741                         metrics->timer_tail = NULL;
 742
 743                 glDeleteQueries (1, &timer->id);
 744
 745                 free (timer);
 746                 timer = metrics->timer_head;
 747         }
 748
 749         if (! metrics->info->have_perfmon)
 750                 return;
 751
 752         /* And similarly for all performance monitors that are ready. */
 753         monitor_t *monitor = metrics->monitor_head;
 754
 755         while (monitor) {
 756                 GLuint available, result_size, *result;
 757                 GLint bytes_written;
 758
 759                 glGetPerfMonitorCounterDataAMD (monitor->id,
 760                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 761                                                 sizeof (available), &available,
 762                                                 NULL);
 763                 if (! available)
 764                         break;
 765
 766                 glGetPerfMonitorCounterDataAMD (monitor->id,
 767                                                 GL_PERFMON_RESULT_SIZE_AMD,
 768                                                 sizeof (result_size),
 769                                                 &result_size, NULL);
 770
 771                 result = xmalloc (result_size);
 772
 773                 glGetPerfMonitorCounterDataAMD (monitor->id,
 774                                                 GL_PERFMON_RESULT_AMD,
 775                                                 result_size, result,
 776                                                 &bytes_written);
 777
 778                 accumulate_program_metrics (metrics, monitor->op, result, result_size);
 779
 780                 free (result);
 781
 782                 metrics->monitor_head = monitor->next;
 783                 if (metrics->monitor_head == NULL)
 784                         metrics->monitor_tail = NULL;
 785
 786                 glDeletePerfMonitorsAMD (1, &monitor->id);
 787
 788                 free (monitor);
 789
 790                 metrics->monitors_in_flight--;
 791
 792                 monitor = metrics->monitor_head;
 793         }
 794 }
 795
 796 /* Return the difference from ts0 to ts1 as floating-point seconds. */
 797 #define TIMESPEC_DIFF(ts0, ts1) (double) (ts1.tv_sec - ts0.tv_sec + \
 798                                           (ts1.tv_nsec - ts0.tv_nsec) / 1e9)
 799
 800 void
 801 metrics_end_frame_pre_swap (metrics_t *metrics)
 802 {
 803         GLuint64 swap_end_timestamp, frame_time_ns, latency_ns;
 804         double cpu_time;
 805
 806         /* Don't leave any counters running over the end_frame work we
 807          * do here. The counters will be started again at the end of
 808          * metrics_end_frame_post_swap.  */
 809
 810         metrics_counter_stop (metrics);
 811
 812         /* Now that an entire frame's worth of content has gone by, we
 813          * can be sure that the timer quiery for the previous frame's
 814          * swap is available.
 815          *
 816          * Note: The only case in which this query isn't immediately
 817          * available would be if there was effectively nothing in this
 818          * frame. In that case, triggering a block on this query
 819          * result is not a concern (since an empty frame is not
 820          * interesting, and all of our counters are stopped anyway).
 821          *
 822          * Accepting this one block is much better than adding a
 823          * linked list for these results that would reaslisticly never
 824          * have more than one entry anyway.
 825          */
 826
 827         glGetQueryObjectui64v (metrics->swap_end_timestamp_id,
 828                                GL_QUERY_RESULT, &swap_end_timestamp);
 829
 830         if (frames == 0) {
 831                 /* Print header */
 832                 printf ("# frame: Frame_Number Frame_Time_milliseconds Frame_latency_milliseconds CPU_load\n");
 833         } else {
 834                 /* Subtract previous frame's times to get frame times. */
 835                 frame_time_ns = swap_end_timestamp - metrics->previous_swap_end_timestamp;
 836
 837                 latency_ns = swap_end_timestamp - metrics->swap_begin_timestamp;
 838
 839                 cpu_time = TIMESPEC_DIFF (metrics->previous_cpu_time_ts, metrics->cpu_time_ts);
 840
 841                 /* We've waited one frame to ensure we have a timestamp
 842                  * result. So the time we've actually measured here is
 843                  * for the previous frame. */
 844                 printf ("frame: %d %g %g %g\n",
 845                         frames - 1,
 846                         (double) frame_time_ns / 1e6,
 847                         (double) latency_ns / 1e6,
 848                         cpu_time / (frame_time_ns / 1e9));
 849         }
 850
 851         metrics->previous_swap_end_timestamp = swap_end_timestamp;
 852         metrics->previous_cpu_time_ts = metrics->cpu_time_ts;
 853
 854         /* Before the actual swap call, we get the current timestamp
 855          * value. This is a synchronous get so we can use this as the
 856          * baseline for a frame latency measurment. */
 857
 858         glGetInteger64v (GL_TIMESTAMP, &metrics->swap_begin_timestamp);
 859 }
 860
 861 void
 862 metrics_end_frame_post_swap (metrics_t *metrics)
 863 {
 864         static int initialized = 0;
 865         static struct timeval tv_start, tv_now;
 866         int err;
 867
 868         /* Now that the swap command has been queued, we issue an
 869          * asynchronous query of the timestamp value. Comparing this
 870          * to the synchronous get we just sent in
 871          * metrics_end_frame_pre_swap allows us to measure the
 872          * per-frame swap latency. */
 873
 874         glQueryCounter (metrics->swap_end_timestamp_id, GL_TIMESTAMP);
 875
 876         /* Query the total amount of CPU time spent by this process. */
 877         err = clock_gettime (CLOCK_PROCESS_CPUTIME_ID, &metrics->cpu_time_ts);
 878         if (err == -1 ) {
 879                 fprintf (stderr, "Failed to query CPU time with clock_gettime: %s\n",
 880                          strerror (errno));
 881                 exit (1);
 882         }
 883
 884         if (! initialized) {
 885                 gettimeofday (&tv_start, NULL);
 886                 if (getenv ("FIPS_VERBOSE"))
 887                         verbose = 1;
 888                 initialized = 1;
 889         }
 890
 891         metrics_collect_available (metrics);
 892
 893         frames++;
 894
 895         if (frames % 15 == 0) {
 896                 double fps;
 897
 898                 gettimeofday (&tv_now, NULL);
 899
 900                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 901                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 902
 903                 printf("FPS: %.3f\n", fps);
 904
 905                 print_program_metrics (metrics);
 906         }
 907
 908         /* Start the counter up again now that we're done. */
 909         metrics_counter_start (metrics);
 910 }