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 "xmalloc.h"
  33
  34 /* Timer query */
  35 typedef struct timer_query
  36 {
  37         unsigned id;
  38
  39         metrics_op_t op;
  40         struct timer_query *next;
  41 } timer_query_t;
  42
  43 /* Performance-monitor query */
  44 typedef struct monitor
  45 {
  46         unsigned id;
  47
  48         metrics_op_t op;
  49         struct monitor *next;
  50 } monitor_t;
  51
  52
  53 typedef struct op_metrics
  54 {
  55         /* This happens to also be the index into the
  56          * ctx->op_metrics array currently
  57          */
  58         metrics_op_t op;
  59         double time_ns;
  60
  61         double **counters;
  62 } op_metrics_t;
  63
  64 typedef struct metrics_group_info
  65 {
  66         GLuint id;
  67         char *name;
  68
  69         GLuint num_counters;
  70         GLuint max_active_counters;
  71
  72         GLuint *counter_ids;
  73         char **counter_names;
  74         GLuint *counter_types;
  75
  76 } metrics_group_info_t;
  77
  78 typedef struct shader_stage_info
  79 {
  80         char *name;
  81
  82         GLuint active_group_index;
  83         GLuint active_counter_index;
  84
  85         GLuint stall_group_index;
  86         GLuint stall_counter_index;
  87
  88 } shader_stage_info_t;
  89
  90 typedef struct metrics_info
  91 {
  92         int initialized;
  93
  94         unsigned num_groups;
  95         metrics_group_info_t *groups;
  96
  97         unsigned num_shader_stages;
  98         shader_stage_info_t *stages;
  99
 100 } metrics_info_t;
 101
 102 typedef struct context
 103 {
 104         metrics_info_t metrics_info;
 105
 106         metrics_op_t op;
 107
 108         timer_query_t *timer_head;
 109         timer_query_t *timer_tail;
 110
 111         monitor_t *monitor_head;
 112         monitor_t *monitor_tail;
 113
 114         unsigned num_op_metrics;
 115         op_metrics_t *op_metrics;
 116 } context_t;
 117
 118 /* FIXME: Need a map from integers to context objects and track the
 119  * current context with glXMakeContextCurrent, eglMakeCurrent, etc. */
 120
 121 context_t current_context;
 122
 123 int frames;
 124 int verbose;
 125
 126 static void
 127 metrics_group_info_init (metrics_group_info_t *group, GLuint id)
 128 {
 129         GLsizei length;
 130         unsigned i;
 131
 132         group->id = id;
 133
 134         /* Get name */
 135         glGetPerfMonitorGroupStringAMD (id, 0, &length, NULL);
 136
 137         group->name = xmalloc (length + 1);
 138
 139         glGetPerfMonitorGroupStringAMD (id, length + 1, NULL, group->name);
 140
 141         /* Get number of counters */
 142         group->num_counters = 0;
 143         group->max_active_counters = 0;
 144         glGetPerfMonitorCountersAMD (group->id,
 145                                      (int *) &group->num_counters,
 146                                      (int *) &group->max_active_counters,
 147                                      0, NULL);
 148
 149         /* Get counter numbers */
 150         group->counter_ids = xmalloc (group->num_counters * sizeof (GLuint));
 151
 152         glGetPerfMonitorCountersAMD (group->id, NULL, NULL,
 153                                      group->num_counters,
 154                                      group->counter_ids);
 155
 156         /* Get counter names */
 157         group->counter_names = xmalloc (group->num_counters * sizeof (char *));
 158         group->counter_types = xmalloc (group->num_counters * sizeof (GLuint));
 159
 160         for (i = 0; i < group->num_counters; i++) {
 161                 glGetPerfMonitorCounterInfoAMD (group->id,
 162                                                 group->counter_ids[i],
 163                                                 GL_COUNTER_TYPE_AMD,
 164                                                 &group->counter_types[i]);
 165
 166                 /* We assume that all peformance counters are made
 167                  * available as uint32 values. The code calling
 168                  * CONSUME in accumulate_program_metrics will need to
 169                  * be extended to accomodate other counter values. */
 170                 if (group->counter_types[i] != GL_UNSIGNED_INT) {
 171                         fprintf (stderr, "fips: Internal error: No support for non-uint counter values\n");
 172                         exit (1);
 173                 }
 174
 175                 glGetPerfMonitorCounterStringAMD (group->id,
 176                                                   group->counter_ids[i],
 177                                                   0, &length, NULL);
 178
 179                 group->counter_names[i] = xmalloc (length + 1);
 180
 181                 glGetPerfMonitorCounterStringAMD (group->id,
 182                                                   group->counter_ids[i],
 183                                                   length + 1, NULL,
 184                                                   group->counter_names[i]);
 185         }
 186 }
 187
 188 static void
 189 metrics_group_info_fini (metrics_group_info_t *group)
 190 {
 191         unsigned i;
 192
 193         for (i = 0; i < group->num_counters; i++)
 194                 free (group->counter_names[i]);
 195
 196         free (group->counter_types);
 197         free (group->counter_names);
 198         free (group->counter_ids);
 199
 200         free (group->name);
 201 }
 202
 203 static void
 204 metrics_info_fini (metrics_info_t *info);
 205
 206 /* A helper function, part of metrics_info_init below. */
 207
 208 typedef enum {
 209         SHADER_ACTIVE,
 210         SHADER_STALL
 211 } shader_phase_t;
 212
 213 static void
 214 _add_shader_stage (metrics_info_t *info, const char *name,
 215                    GLuint group_index, GLuint counter_index,
 216                    shader_phase_t phase)
 217 {
 218         shader_stage_info_t *stage;
 219         char *stage_name, *space;
 220         unsigned i;
 221
 222         stage_name = xstrdup (name);
 223
 224         /* Terminate the stage name at the first space.
 225          *
 226          * This is valid for counter names such as:
 227          *
 228          *      "Vertex Shader Active Time"
 229          * or
 230          *      "Vertex Shader Stall Time - Core Stall"
 231          */
 232         space = strchr (stage_name, ' ');
 233         if (space)
 234                 *space = '\0';
 235
 236         /* Look for an existing stage of the given name. */
 237         stage = NULL;
 238
 239         for (i = 0; i < info->num_shader_stages; i++) {
 240                 if (strcmp (info->stages[i].name, stage_name) == 0) {
 241                         stage = &info->stages[i];
 242                         break;
 243                 }
 244         }
 245
 246         if (stage == NULL) {
 247                 info->num_shader_stages++;
 248                 info->stages = xrealloc (info->stages,
 249                                          info->num_shader_stages *
 250                                          sizeof (shader_stage_info_t));
 251                 stage = &info->stages[info->num_shader_stages - 1];
 252                 stage->name = xstrdup (stage_name);
 253                 stage->active_group_index = 0;
 254                 stage->active_counter_index = 0;
 255                 stage->stall_group_index = 0;
 256                 stage->stall_counter_index = 0;
 257         }
 258
 259         if (phase == SHADER_ACTIVE) {
 260                 stage->active_group_index = group_index;
 261                 stage->active_counter_index = counter_index;
 262         } else {
 263                 stage->stall_group_index = group_index;
 264                 stage->stall_counter_index = counter_index;
 265         }
 266
 267         free (stage_name);
 268 }
 269
 270 void
 271 metrics_info_init (void)
 272 {
 273         unsigned i, j;
 274         GLuint *group_ids;
 275         metrics_info_t *info = &current_context.metrics_info;
 276
 277         if (info->initialized)
 278                 metrics_info_fini (info);
 279
 280         glGetPerfMonitorGroupsAMD ((int *) &info->num_groups, 0, NULL);
 281
 282         group_ids = xmalloc (info->num_groups * sizeof (GLuint));
 283
 284         glGetPerfMonitorGroupsAMD (NULL, info->num_groups, group_ids);
 285
 286         info->groups = xmalloc (info->num_groups * sizeof (metrics_group_info_t));
 287
 288         for (i = 0; i < info->num_groups; i++)
 289                 metrics_group_info_init (&info->groups[i], group_ids[i]);
 290
 291         free (group_ids);
 292
 293         /* Identify each shader stage (by looking at
 294          * performance-counter names for specific patterns) and
 295          * initialize structures referring to the corresponding
 296          * counter numbers for each stage. */
 297         info->num_shader_stages = 0;
 298         info->stages = NULL;
 299
 300         for (i = 0; i < info->num_groups; i++) {
 301                 metrics_group_info_t *group = &info->groups[i];
 302                 for (j = 0; j < group->num_counters; j++) {
 303                         char *name = group->counter_names[j];
 304                         if (strstr (name, "Shader Active Time")) {
 305                                 _add_shader_stage (info, name, i, j,
 306                                                    SHADER_ACTIVE);
 307                         }
 308                         if (strstr (name, "Shader Stall Time")) {
 309                                 _add_shader_stage (info, name, i, j,
 310                                                    SHADER_STALL);
 311                         }
 312                 }
 313         }
 314
 315         info->initialized = 1;
 316 }
 317
 318 static void
 319 metrics_info_fini (metrics_info_t *info)
 320 {
 321         unsigned i;
 322
 323         for (i = 0; i < info->num_groups; i++)
 324                 metrics_group_info_fini (&info->groups[i]);
 325
 326         free (info->groups);
 327
 328         for (i = 0; i < info->num_shader_stages; i++)
 329                 free (info->stages[i].name);
 330
 331         free (info->stages);
 332 }
 333
 334 static const char *
 335 metrics_op_string (metrics_op_t op)
 336 {
 337         if (op >= METRICS_OP_SHADER)
 338                 return "Shader program";
 339
 340         switch (op)
 341         {
 342         case METRICS_OP_ACCUM:
 343                 return "glAccum*(+)";
 344         case METRICS_OP_BUFFER_DATA:
 345                 return "glBufferData(+)";
 346         case METRICS_OP_BUFFER_SUB_DATA:
 347                 return "glCopyBufferSubData*";
 348         case METRICS_OP_BITMAP:
 349                 return "glBitmap*";
 350         case METRICS_OP_BLIT_FRAMEBUFFER:
 351                 return "glBlitFramebuffer*";
 352         case METRICS_OP_CLEAR:
 353                 return "glClear(+)";
 354         case METRICS_OP_CLEAR_BUFFER_DATA:
 355                 return "glCearBufferData(+)";
 356         case METRICS_OP_CLEAR_TEX_IMAGE:
 357                 return "glClearTexImage(+)";
 358         case METRICS_OP_COPY_PIXELS:
 359                 return "glCopyPixels";
 360         case METRICS_OP_COPY_TEX_IMAGE:
 361                 return "glCopyTexImage(+)";
 362         case METRICS_OP_DRAW_PIXELS:
 363                 return "glDrawPixels";
 364         case METRICS_OP_GET_TEX_IMAGE:
 365                 return "glGetTexImage(+)";
 366         case METRICS_OP_READ_PIXELS:
 367                 return "glReadPixels*";
 368         case METRICS_OP_TEX_IMAGE:
 369                 return "glTexImage*(+)";
 370         default:
 371                 fprintf (stderr, "fips: Internal error: "
 372                          "Unknown metrics op value: %d\n", op);
 373                 exit (1);
 374         }
 375
 376         return "";
 377 }
 378
 379 void
 380 metrics_counter_start (void)
 381 {
 382         context_t *ctx = &current_context;
 383         timer_query_t *timer;
 384         monitor_t *monitor;
 385         unsigned i;
 386
 387         /* Create new timer query, add to list */
 388         timer = xmalloc (sizeof (timer_query_t));
 389
 390         timer->op = ctx->op;
 391         timer->next = NULL;
 392
 393         if (ctx->timer_tail) {
 394                 ctx->timer_tail->next = timer;
 395                 ctx->timer_tail = timer;
 396         } else {
 397                 ctx->timer_tail = timer;
 398                 ctx->timer_head = timer;
 399         }
 400
 401         /* Create a new performance-monitor query */
 402         monitor = xmalloc (sizeof (monitor_t));
 403
 404         monitor->op = ctx->op;
 405         monitor->next = NULL;
 406
 407         if (ctx->monitor_tail) {
 408                 ctx->monitor_tail->next = monitor;
 409                 ctx->monitor_tail = monitor;
 410         } else {
 411                 ctx->monitor_tail = monitor;
 412                 ctx->monitor_head = monitor;
 413         }
 414
 415         /* Initialize the timer_query and monitor objects */
 416         glGenQueries (1, &timer->id);
 417
 418         glGenPerfMonitorsAMD (1, &monitor->id);
 419
 420         for (i = 0; i < ctx->metrics_info.num_groups; i++)
 421         {
 422                 metrics_group_info_t *group;
 423                 int num_counters;
 424
 425                 group = &ctx->metrics_info.groups[i];
 426
 427                 num_counters = group->num_counters;
 428                 if (group->max_active_counters < group->num_counters)
 429                 {
 430                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 431                                  group->max_active_counters,
 432                                  group->num_counters, i);
 433                         num_counters = group->max_active_counters;
 434
 435                 }
 436
 437                 glSelectPerfMonitorCountersAMD(monitor->id,
 438                                                GL_TRUE, group->id,
 439                                                num_counters,
 440                                                group->counter_ids);
 441         }
 442
 443         /* Start the queries */
 444         glBeginQuery (GL_TIME_ELAPSED, timer->id);
 445
 446         glBeginPerfMonitorAMD (monitor->id);
 447 }
 448
 449 void
 450 metrics_counter_stop (void)
 451 {
 452         glEndQuery (GL_TIME_ELAPSED);
 453         glEndPerfMonitorAMD (current_context.monitor_tail->id);
 454 }
 455
 456 void
 457 metrics_set_current_op (metrics_op_t op)
 458 {
 459         current_context.op = op;
 460 }
 461
 462 metrics_op_t
 463 metrics_get_current_op (void)
 464 {
 465         return current_context.op;
 466 }
 467
 468 static void
 469 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 470 {
 471         metrics_info_t *info = &ctx->metrics_info;
 472         unsigned i, j;
 473
 474         metrics->op = op;
 475         metrics->time_ns = 0.0;
 476
 477         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 478
 479         for (i = 0; i < info->num_groups; i++) {
 480                 metrics->counters[i] = xmalloc (sizeof (double) *
 481                                                 info->groups[i].num_counters);
 482                 for (j = 0; j < info->groups[i].num_counters; j++)
 483                         metrics->counters[i][j] = 0.0;
 484         }
 485 }
 486
 487 static op_metrics_t *
 488 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 489 {
 490         unsigned i;
 491
 492         if (op >= ctx->num_op_metrics)
 493         {
 494                 ctx->op_metrics = realloc (ctx->op_metrics,
 495                                            (op + 1) * sizeof (op_metrics_t));
 496                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 497                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 498
 499                 ctx->num_op_metrics = op + 1;
 500         }
 501
 502         return &ctx->op_metrics[op];
 503 }
 504
 505 static void
 506 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 507 {
 508 #define CONSUME(var)                                                    \
 509         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 510         {                                                               \
 511                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 512                          "parsing results\n");                          \
 513                 break;                                                  \
 514         }                                                               \
 515         (var) = *((typeof(var) *) p);                                   \
 516         p += sizeof(var);
 517
 518         context_t *ctx = &current_context;
 519         metrics_info_t *info = &ctx->metrics_info;
 520         unsigned char *p = (unsigned char *) result;
 521
 522         while (p < ((unsigned char *) result) + size)
 523         {
 524                 GLuint group_id, group_index;
 525                 GLuint counter_id, counter_index;
 526                 metrics_group_info_t *group;
 527                 uint32_t value;
 528                 unsigned i;
 529
 530                 CONSUME (group_id);
 531                 CONSUME (counter_id);
 532                 CONSUME (value);
 533
 534                 for (i = 0; i < info->num_groups; i++) {
 535                         if (info->groups[i].id == i)
 536                                 break;
 537                 }
 538                 group_index = i;
 539                 assert (group_index < info->num_groups);
 540                 group = &info->groups[group_index];
 541
 542                 for (i = 0; i < group->num_counters; i++) {
 543                         if (group->counter_ids[i] == counter_id)
 544                                 break;
 545                 }
 546                 counter_index = i;
 547                 assert (counter_index < group->num_counters);
 548
 549                 ctx->op_metrics[op].counters[group_index][counter_index] += value;
 550         }
 551 }
 552
 553 static void
 554 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 555 {
 556         op_metrics_t *metrics;
 557
 558         metrics = ctx_get_op_metrics (&current_context, op);
 559
 560         metrics->time_ns += time_ns;
 561 }
 562
 563 typedef struct per_stage_metrics
 564 {
 565         op_metrics_t *metrics;
 566         shader_stage_info_t *stage;
 567         double time_ns;
 568         double active;
 569 } per_stage_metrics_t;
 570
 571 static int
 572 _is_shader_stage_counter (metrics_info_t *info,
 573                           unsigned group_index,
 574                           unsigned counter_index)
 575 {
 576         shader_stage_info_t *stage;
 577         unsigned i;
 578
 579         for (i = 0; i < info->num_shader_stages; i++) {
 580                 stage = &info->stages[i];
 581
 582                 if (stage->active_group_index == group_index &&
 583                     stage->active_counter_index == counter_index)
 584                 {
 585                         return 1;
 586                 }
 587
 588                 if (stage->stall_group_index == group_index &&
 589                     stage->stall_counter_index == counter_index)
 590                 {
 591                         return 1;
 592                 }
 593         }
 594
 595         return 0;
 596 }
 597
 598 static void
 599 print_per_stage_metrics (context_t *ctx,
 600                          per_stage_metrics_t *per_stage,
 601                          double total)
 602 {
 603         metrics_info_t *info = &ctx->metrics_info;
 604         op_metrics_t *metric = per_stage->metrics;
 605         metrics_group_info_t *group;
 606         const char *op_string;
 607         unsigned group_index, counter;
 608         double value;
 609
 610         /* Don't print anything for stages with no alloted time. */
 611         if (per_stage->time_ns == 0.0)
 612                 return;
 613
 614         op_string = metrics_op_string (metric->op);
 615
 616         printf ("%21s", op_string);
 617
 618         if (metric->op >= METRICS_OP_SHADER) {
 619                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 620         } else {
 621                 printf ("    ");
 622
 623         }
 624         printf (" %cS:", per_stage->stage->name[0]);
 625
 626         printf ("\t%7.2f ms (%4.1f%%)",
 627                 per_stage->time_ns / 1e6,
 628                 per_stage->time_ns / total * 100);
 629
 630         if (per_stage->active)
 631                 printf (", %4.1f%% active", per_stage->active * 100);
 632
 633         printf ("\n");
 634
 635         /* I'm not seeing a lot of value printing the rest of these
 636          * performance counters by default yet. Use --verbose to get
 637          * them for now. */
 638         if (! verbose)
 639                 return;
 640
 641         printf ("[");
 642         for (group_index = 0; group_index < info->num_groups; group_index++) {
 643                 group = &info->groups[group_index];
 644                 for (counter = 0; counter < group->num_counters; counter++) {
 645
 646                         /* Don't print this counter value if it's a
 647                          * per-stage cycle counter, (which we have
 648                          * already accounted for). */
 649                         if (_is_shader_stage_counter (info, group_index, counter))
 650                                 continue;
 651
 652                         value = metric->counters[group_index][counter];
 653                         if (value == 0.0)
 654                                 continue;
 655                         printf ("%s: %.2f ", group->counter_names[counter],
 656                                 value / 1e6);
 657                 }
 658         }
 659         printf ("]\n");
 660 }
 661
 662 static int
 663 time_compare(const void *in_a, const void *in_b, void *arg unused)
 664 {
 665         const per_stage_metrics_t *a = in_a;
 666         const per_stage_metrics_t *b = in_b;
 667
 668
 669         if (a->time_ns < b->time_ns)
 670                 return -1;
 671         if (a->time_ns > b->time_ns)
 672                 return 1;
 673         return 0;
 674 }
 675
 676 static void
 677 print_program_metrics (void)
 678 {
 679         context_t *ctx = &current_context;
 680         metrics_info_t *info = &ctx->metrics_info;
 681         unsigned num_shader_stages = info->num_shader_stages;
 682         per_stage_metrics_t *sorted, *per_stage;
 683         double total_time, op_cycles;
 684         op_metrics_t *op;
 685         unsigned group_index, counter_index;
 686         unsigned i, j, num_sorted;
 687
 688         /* Make a sorted list of the per-stage operations by time
 689          * used, and figure out the total so we can print percentages.
 690          */
 691         num_sorted = ctx->num_op_metrics * num_shader_stages;
 692
 693         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 694
 695         total_time = 0.0;
 696
 697         for (i = 0; i < ctx->num_op_metrics; i++) {
 698
 699                 op = &ctx->op_metrics[i];
 700
 701                 /* Accumulate total time across all ops. */
 702                 total_time += op->time_ns;
 703
 704                 /* Also, find total cycles in all stages of this op. */
 705                 op_cycles = 0.0;
 706
 707                 for (j = 0; j < num_shader_stages; j++) {
 708                         /* Active cycles */
 709                         group_index = info->stages[j].active_group_index;
 710                         counter_index = info->stages[j].active_counter_index;
 711                         op_cycles += op->counters[group_index][counter_index];
 712
 713                         /* Stall cycles */
 714                         group_index = info->stages[j].stall_group_index;
 715                         counter_index = info->stages[j].stall_counter_index;
 716                         op_cycles += op->counters[group_index][counter_index];
 717                 }
 718
 719                 for (j = 0; j < num_shader_stages; j++) {
 720                         double active_cycles, stall_cycles, stage_cycles;
 721
 722                         /* Active cycles */
 723                         group_index = info->stages[j].active_group_index;
 724                         counter_index = info->stages[j].active_counter_index;
 725                         active_cycles = op->counters[group_index][counter_index];
 726
 727                         /* Stall cycles */
 728                         group_index = info->stages[j].stall_group_index;
 729                         counter_index = info->stages[j].stall_counter_index;
 730                         stall_cycles = op->counters[group_index][counter_index];
 731
 732                         stage_cycles = active_cycles + stall_cycles;
 733
 734                         per_stage = &sorted[i * num_shader_stages + j];
 735                         per_stage->metrics = op;
 736                         per_stage->stage = &info->stages[j];
 737                         if (op_cycles)
 738                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 739                         else
 740                                 per_stage->time_ns = 0.0;
 741                         if (stage_cycles)
 742                                 per_stage->active = active_cycles / stage_cycles;
 743                         else
 744                                 per_stage->active = 0.0;
 745                 }
 746         }
 747
 748         qsort_r (sorted, num_sorted, sizeof (*sorted),
 749                  time_compare, ctx->op_metrics);
 750
 751         for (i = 0; i < num_sorted; i++)
 752                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 753
 754         free (sorted);
 755 }
 756
 757 /* Called at program exit */
 758 static void
 759 metrics_exit (void)
 760 {
 761         if (verbose)
 762                 printf ("fips: terminating\n");
 763
 764         metrics_info_fini (&current_context.metrics_info);
 765 }
 766
 767
 768 void
 769 metrics_end_frame (void)
 770 {
 771         static int initialized = 0;
 772         static struct timeval tv_start, tv_now;
 773
 774         if (! initialized) {
 775                 gettimeofday (&tv_start, NULL);
 776                 atexit (metrics_exit);
 777                 if (getenv ("FIPS_VERBOSE"))
 778                         verbose = 1;
 779                 initialized = 1;
 780         }
 781
 782         frames++;
 783         gettimeofday (&tv_now, NULL);
 784
 785         /* Consume all timer queries that are ready. */
 786         timer_query_t *timer = current_context.timer_head;
 787
 788         while (timer) {
 789                 GLuint available, elapsed;
 790
 791                 glGetQueryObjectuiv (timer->id,
 792                                      GL_QUERY_RESULT_AVAILABLE, &available);
 793                 if (! available)
 794                         break;
 795
 796                 glGetQueryObjectuiv (timer->id,
 797                                      GL_QUERY_RESULT, &elapsed);
 798
 799                 accumulate_program_time (timer->op, elapsed);
 800
 801                 current_context.timer_head = timer->next;
 802                 if (current_context.timer_head == NULL)
 803                         current_context.timer_tail = NULL;
 804
 805                 glDeleteQueries (1, &timer->id);
 806
 807                 free (timer);
 808                 timer = current_context.timer_head;
 809         }
 810
 811         /* And similarly for all performance monitors that are ready. */
 812         monitor_t *monitor = current_context.monitor_head;
 813
 814         while (monitor) {
 815                 GLuint available, result_size, *result;
 816                 GLint bytes_written;
 817
 818                 glGetPerfMonitorCounterDataAMD (monitor->id,
 819                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 820                                                 sizeof (available), &available,
 821                                                 NULL);
 822                 if (! available)
 823                         break;
 824
 825                 glGetPerfMonitorCounterDataAMD (monitor->id,
 826                                                 GL_PERFMON_RESULT_SIZE_AMD,
 827                                                 sizeof (result_size),
 828                                                 &result_size, NULL);
 829
 830                 result = xmalloc (result_size);
 831
 832                 glGetPerfMonitorCounterDataAMD (monitor->id,
 833                                                 GL_PERFMON_RESULT_AMD,
 834                                                 result_size, result,
 835                                                 &bytes_written);
 836
 837                 accumulate_program_metrics (monitor->op, result, result_size);
 838
 839                 free (result);
 840
 841                 current_context.monitor_head = monitor->next;
 842                 if (current_context.monitor_head == NULL)
 843                         current_context.monitor_tail = NULL;
 844
 845                 glDeletePerfMonitorsAMD (1, &monitor->id);
 846
 847                 free (monitor);
 848                 monitor = current_context.monitor_head;
 849         }
 850
 851         if (frames % 60 == 0) {
 852                 double fps;
 853
 854                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 855                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 856
 857                 printf("FPS: %.3f\n", fps);
 858
 859                 print_program_metrics ();
 860         }
 861 }