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                 glGetPerfMonitorCounterStringAMD (group->id,
 167                                                   group->counter_ids[i],
 168                                                   0, &length, NULL);
 169
 170                 group->counter_names[i] = xmalloc (length + 1);
 171
 172                 glGetPerfMonitorCounterStringAMD (group->id,
 173                                                   group->counter_ids[i],
 174                                                   length + 1, NULL,
 175                                                   group->counter_names[i]);
 176         }
 177 }
 178
 179 static void
 180 metrics_group_info_fini (metrics_group_info_t *group)
 181 {
 182         unsigned i;
 183
 184         for (i = 0; i < group->num_counters; i++)
 185                 free (group->counter_names[i]);
 186
 187         free (group->counter_types);
 188         free (group->counter_names);
 189         free (group->counter_ids);
 190
 191         free (group->name);
 192 }
 193
 194 static void
 195 metrics_info_fini (metrics_info_t *info);
 196
 197 /* A helper function, part of metrics_info_init below. */
 198
 199 typedef enum {
 200         SHADER_ACTIVE,
 201         SHADER_STALL
 202 } shader_phase_t;
 203
 204 static void
 205 _add_shader_stage (metrics_info_t *info, const char *name,
 206                    GLuint group_index, GLuint counter_index,
 207                    shader_phase_t phase)
 208 {
 209         shader_stage_info_t *stage;
 210         char *stage_name, *space;
 211         unsigned i;
 212
 213         stage_name = xstrdup (name);
 214
 215         /* Terminate the stage name at the first space.
 216          *
 217          * This is valid for counter names such as:
 218          *
 219          *      "Vertex Shader Active Time"
 220          * or
 221          *      "Vertex Shader Stall Time - Core Stall"
 222          */
 223         space = strchr (stage_name, ' ');
 224         if (space)
 225                 *space = '\0';
 226
 227         /* Look for an existing stage of the given name. */
 228         stage = NULL;
 229
 230         for (i = 0; i < info->num_shader_stages; i++) {
 231                 if (strcmp (info->stages[i].name, stage_name) == 0) {
 232                         stage = &info->stages[i];
 233                         break;
 234                 }
 235         }
 236
 237         if (stage == NULL) {
 238                 info->num_shader_stages++;
 239                 info->stages = xrealloc (info->stages,
 240                                          info->num_shader_stages *
 241                                          sizeof (shader_stage_info_t));
 242                 stage = &info->stages[info->num_shader_stages - 1];
 243                 stage->name = xstrdup (stage_name);
 244                 stage->active_group_index = 0;
 245                 stage->active_counter_index = 0;
 246                 stage->stall_group_index = 0;
 247                 stage->stall_counter_index = 0;
 248         }
 249
 250         if (phase == SHADER_ACTIVE) {
 251                 stage->active_group_index = group_index;
 252                 stage->active_counter_index = counter_index;
 253         } else {
 254                 stage->stall_group_index = group_index;
 255                 stage->stall_counter_index = counter_index;
 256         }
 257
 258         free (stage_name);
 259 }
 260
 261 void
 262 metrics_info_init (void)
 263 {
 264         unsigned i, j;
 265         GLuint *group_ids;
 266         metrics_info_t *info = &current_context.metrics_info;
 267
 268         if (info->initialized)
 269                 metrics_info_fini (info);
 270
 271         glGetPerfMonitorGroupsAMD ((int *) &info->num_groups, 0, NULL);
 272
 273         group_ids = xmalloc (info->num_groups * sizeof (GLuint));
 274
 275         glGetPerfMonitorGroupsAMD (NULL, info->num_groups, group_ids);
 276
 277         info->groups = xmalloc (info->num_groups * sizeof (metrics_group_info_t));
 278
 279         for (i = 0; i < info->num_groups; i++)
 280                 metrics_group_info_init (&info->groups[i], group_ids[i]);
 281
 282         free (group_ids);
 283
 284         /* Identify each shader stage (by looking at
 285          * performance-counter names for specific patterns) and
 286          * initialize structures referring to the corresponding
 287          * counter numbers for each stage. */
 288         info->num_shader_stages = 0;
 289         info->stages = NULL;
 290
 291         for (i = 0; i < info->num_groups; i++) {
 292                 metrics_group_info_t *group = &info->groups[i];
 293                 for (j = 0; j < group->num_counters; j++) {
 294                         char *name = group->counter_names[j];
 295                         if (strstr (name, "Shader Active Time")) {
 296                                 _add_shader_stage (info, name, i, j,
 297                                                    SHADER_ACTIVE);
 298                         }
 299                         if (strstr (name, "Shader Stall Time")) {
 300                                 _add_shader_stage (info, name, i, j,
 301                                                    SHADER_STALL);
 302                         }
 303                 }
 304         }
 305
 306         info->initialized = 1;
 307 }
 308
 309 static void
 310 metrics_info_fini (metrics_info_t *info)
 311 {
 312         unsigned i;
 313
 314         for (i = 0; i < info->num_groups; i++)
 315                 metrics_group_info_fini (&info->groups[i]);
 316
 317         free (info->groups);
 318
 319         for (i = 0; i < info->num_shader_stages; i++)
 320                 free (info->stages[i].name);
 321
 322         free (info->stages);
 323 }
 324
 325 static const char *
 326 metrics_op_string (metrics_op_t op)
 327 {
 328         if (op >= METRICS_OP_SHADER)
 329                 return "Shader program";
 330
 331         switch (op)
 332         {
 333         case METRICS_OP_ACCUM:
 334                 return "glAccum*(+)";
 335         case METRICS_OP_BUFFER_DATA:
 336                 return "glBufferData(+)";
 337         case METRICS_OP_BUFFER_SUB_DATA:
 338                 return "glCopyBufferSubData*";
 339         case METRICS_OP_BITMAP:
 340                 return "glBitmap*";
 341         case METRICS_OP_BLIT_FRAMEBUFFER:
 342                 return "glBlitFramebuffer*";
 343         case METRICS_OP_CLEAR:
 344                 return "glClear(+)";
 345         case METRICS_OP_CLEAR_BUFFER_DATA:
 346                 return "glCearBufferData(+)";
 347         case METRICS_OP_CLEAR_TEX_IMAGE:
 348                 return "glClearTexImage(+)";
 349         case METRICS_OP_COPY_PIXELS:
 350                 return "glCopyPixels";
 351         case METRICS_OP_COPY_TEX_IMAGE:
 352                 return "glCopyTexImage(+)";
 353         case METRICS_OP_DRAW_PIXELS:
 354                 return "glDrawPixels";
 355         case METRICS_OP_GET_TEX_IMAGE:
 356                 return "glGetTexImage(+)";
 357         case METRICS_OP_READ_PIXELS:
 358                 return "glReadPixels*";
 359         case METRICS_OP_TEX_IMAGE:
 360                 return "glTexImage*(+)";
 361         default:
 362                 fprintf (stderr, "fips: Internal error: "
 363                          "Unknown metrics op value: %d\n", op);
 364                 exit (1);
 365         }
 366
 367         return "";
 368 }
 369
 370 void
 371 metrics_counter_start (void)
 372 {
 373         context_t *ctx = &current_context;
 374         timer_query_t *timer;
 375         monitor_t *monitor;
 376         unsigned i;
 377
 378         /* Create new timer query, add to list */
 379         timer = xmalloc (sizeof (timer_query_t));
 380
 381         timer->op = ctx->op;
 382         timer->next = NULL;
 383
 384         if (ctx->timer_tail) {
 385                 ctx->timer_tail->next = timer;
 386                 ctx->timer_tail = timer;
 387         } else {
 388                 ctx->timer_tail = timer;
 389                 ctx->timer_head = timer;
 390         }
 391
 392         /* Create a new performance-monitor query */
 393         monitor = xmalloc (sizeof (monitor_t));
 394
 395         monitor->op = ctx->op;
 396         monitor->next = NULL;
 397
 398         if (ctx->monitor_tail) {
 399                 ctx->monitor_tail->next = monitor;
 400                 ctx->monitor_tail = monitor;
 401         } else {
 402                 ctx->monitor_tail = monitor;
 403                 ctx->monitor_head = monitor;
 404         }
 405
 406         /* Initialize the timer_query and monitor objects */
 407         glGenQueries (1, &timer->id);
 408
 409         glGenPerfMonitorsAMD (1, &monitor->id);
 410
 411         for (i = 0; i < ctx->metrics_info.num_groups; i++)
 412         {
 413                 metrics_group_info_t *group;
 414                 int num_counters;
 415
 416                 group = &ctx->metrics_info.groups[i];
 417
 418                 num_counters = group->num_counters;
 419                 if (group->max_active_counters < group->num_counters)
 420                 {
 421                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 422                                  group->max_active_counters,
 423                                  group->num_counters, i);
 424                         num_counters = group->max_active_counters;
 425
 426                 }
 427
 428                 glSelectPerfMonitorCountersAMD(monitor->id,
 429                                                GL_TRUE, group->id,
 430                                                num_counters,
 431                                                group->counter_ids);
 432         }
 433
 434         /* Start the queries */
 435         glBeginQuery (GL_TIME_ELAPSED, timer->id);
 436
 437         glBeginPerfMonitorAMD (monitor->id);
 438 }
 439
 440 void
 441 metrics_counter_stop (void)
 442 {
 443         glEndQuery (GL_TIME_ELAPSED);
 444         glEndPerfMonitorAMD (current_context.monitor_tail->id);
 445 }
 446
 447 void
 448 metrics_set_current_op (metrics_op_t op)
 449 {
 450         current_context.op = op;
 451 }
 452
 453 metrics_op_t
 454 metrics_get_current_op (void)
 455 {
 456         return current_context.op;
 457 }
 458
 459 static void
 460 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 461 {
 462         metrics_info_t *info = &ctx->metrics_info;
 463         unsigned i, j;
 464
 465         metrics->op = op;
 466         metrics->time_ns = 0.0;
 467
 468         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 469
 470         for (i = 0; i < info->num_groups; i++) {
 471                 metrics->counters[i] = xmalloc (sizeof (double) *
 472                                                 info->groups[i].num_counters);
 473                 for (j = 0; j < info->groups[i].num_counters; j++)
 474                         metrics->counters[i][j] = 0.0;
 475         }
 476 }
 477
 478 static op_metrics_t *
 479 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 480 {
 481         unsigned i;
 482
 483         if (op >= ctx->num_op_metrics)
 484         {
 485                 ctx->op_metrics = realloc (ctx->op_metrics,
 486                                            (op + 1) * sizeof (op_metrics_t));
 487                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 488                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 489
 490                 ctx->num_op_metrics = op + 1;
 491         }
 492
 493         return &ctx->op_metrics[op];
 494 }
 495
 496 static void
 497 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 498 {
 499 #define CONSUME(var)                                                    \
 500         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 501         {                                                               \
 502                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 503                          "parsing results\n");                          \
 504                 break;                                                  \
 505         }                                                               \
 506         (var) = *((typeof(var) *) p);                                   \
 507         p += sizeof(var);
 508
 509         context_t *ctx = &current_context;
 510         metrics_info_t *info = &ctx->metrics_info;
 511         unsigned char *p = (unsigned char *) result;
 512
 513         while (p < ((unsigned char *) result) + size)
 514         {
 515                 GLuint group_id, group_index;
 516                 GLuint counter_id, counter_index;
 517                 metrics_group_info_t *group;
 518                 double value;
 519                 unsigned i;
 520
 521                 CONSUME (group_id);
 522                 CONSUME (counter_id);
 523
 524                 for (i = 0; i < info->num_groups; i++) {
 525                         if (info->groups[i].id == group_id)
 526                                 break;
 527                 }
 528                 group_index = i;
 529                 assert (group_index < info->num_groups);
 530                 group = &info->groups[group_index];
 531
 532                 for (i = 0; i < group->num_counters; i++) {
 533                         if (group->counter_ids[i] == counter_id)
 534                                 break;
 535                 }
 536                 counter_index = i;
 537                 assert (counter_index < group->num_counters);
 538
 539                 switch (group->counter_types[counter_index])
 540                 {
 541                         uint uint_value;
 542                         uint64_t uint64_value;
 543                         float float_value;
 544                 case GL_UNSIGNED_INT:
 545                         CONSUME (uint_value);
 546                         value = uint_value;
 547                         break;
 548                 case GL_UNSIGNED_INT64_AMD:
 549                         CONSUME (uint64_value);
 550                         value = uint64_value;
 551                         break;
 552                 case GL_PERCENTAGE_AMD:
 553                 case GL_FLOAT:
 554                         CONSUME (float_value);
 555                         value = float_value;
 556                         break;
 557                 default:
 558                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 559                                  group->counter_types[counter_index]);
 560                         value = 0.0;
 561                         break;
 562                 }
 563
 564                 ctx->op_metrics[op].counters[group_index][counter_index] += value;
 565         }
 566 }
 567
 568 static void
 569 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 570 {
 571         op_metrics_t *metrics;
 572
 573         metrics = ctx_get_op_metrics (&current_context, op);
 574
 575         metrics->time_ns += time_ns;
 576 }
 577
 578 typedef struct per_stage_metrics
 579 {
 580         op_metrics_t *metrics;
 581         shader_stage_info_t *stage;
 582         double time_ns;
 583         double active;
 584 } per_stage_metrics_t;
 585
 586 static int
 587 _is_shader_stage_counter (metrics_info_t *info,
 588                           unsigned group_index,
 589                           unsigned counter_index)
 590 {
 591         shader_stage_info_t *stage;
 592         unsigned i;
 593
 594         for (i = 0; i < info->num_shader_stages; i++) {
 595                 stage = &info->stages[i];
 596
 597                 if (stage->active_group_index == group_index &&
 598                     stage->active_counter_index == counter_index)
 599                 {
 600                         return 1;
 601                 }
 602
 603                 if (stage->stall_group_index == group_index &&
 604                     stage->stall_counter_index == counter_index)
 605                 {
 606                         return 1;
 607                 }
 608         }
 609
 610         return 0;
 611 }
 612
 613 static void
 614 print_per_stage_metrics (context_t *ctx,
 615                          per_stage_metrics_t *per_stage,
 616                          double total)
 617 {
 618         metrics_info_t *info = &ctx->metrics_info;
 619         op_metrics_t *metric = per_stage->metrics;
 620         metrics_group_info_t *group;
 621         const char *op_string;
 622         unsigned group_index, counter;
 623         double value;
 624
 625         /* Don't print anything for stages with no alloted time. */
 626         if (per_stage->time_ns == 0.0)
 627                 return;
 628
 629         op_string = metrics_op_string (metric->op);
 630
 631         printf ("%21s", op_string);
 632
 633         if (metric->op >= METRICS_OP_SHADER) {
 634                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 635         } else {
 636                 printf ("    ");
 637
 638         }
 639         printf (" %cS:", per_stage->stage->name[0]);
 640
 641         printf ("\t%7.2f ms (%4.1f%%)",
 642                 per_stage->time_ns / 1e6,
 643                 per_stage->time_ns / total * 100);
 644
 645         if (per_stage->active)
 646                 printf (", %4.1f%% active", per_stage->active * 100);
 647
 648         printf ("\n");
 649
 650         /* I'm not seeing a lot of value printing the rest of these
 651          * performance counters by default yet. Use --verbose to get
 652          * them for now. */
 653         if (! verbose)
 654                 return;
 655
 656         printf ("[");
 657         for (group_index = 0; group_index < info->num_groups; group_index++) {
 658                 group = &info->groups[group_index];
 659                 for (counter = 0; counter < group->num_counters; counter++) {
 660
 661                         /* Don't print this counter value if it's a
 662                          * per-stage cycle counter, (which we have
 663                          * already accounted for). */
 664                         if (_is_shader_stage_counter (info, group_index, counter))
 665                                 continue;
 666
 667                         value = metric->counters[group_index][counter];
 668                         if (value == 0.0)
 669                                 continue;
 670                         printf ("%s: %.2f ", group->counter_names[counter],
 671                                 value / 1e6);
 672                 }
 673         }
 674         printf ("]\n");
 675 }
 676
 677 static int
 678 time_compare(const void *in_a, const void *in_b, void *arg unused)
 679 {
 680         const per_stage_metrics_t *a = in_a;
 681         const per_stage_metrics_t *b = in_b;
 682
 683
 684         if (a->time_ns < b->time_ns)
 685                 return -1;
 686         if (a->time_ns > b->time_ns)
 687                 return 1;
 688         return 0;
 689 }
 690
 691 static void
 692 print_program_metrics (void)
 693 {
 694         context_t *ctx = &current_context;
 695         metrics_info_t *info = &ctx->metrics_info;
 696         unsigned num_shader_stages = info->num_shader_stages;
 697         per_stage_metrics_t *sorted, *per_stage;
 698         double total_time, op_cycles;
 699         op_metrics_t *op;
 700         unsigned group_index, counter_index;
 701         unsigned i, j, num_sorted;
 702
 703         /* Make a sorted list of the per-stage operations by time
 704          * used, and figure out the total so we can print percentages.
 705          */
 706         num_sorted = ctx->num_op_metrics * num_shader_stages;
 707
 708         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 709
 710         total_time = 0.0;
 711
 712         for (i = 0; i < ctx->num_op_metrics; i++) {
 713
 714                 op = &ctx->op_metrics[i];
 715
 716                 /* Accumulate total time across all ops. */
 717                 total_time += op->time_ns;
 718
 719                 /* Also, find total cycles in all stages of this op. */
 720                 op_cycles = 0.0;
 721
 722                 for (j = 0; j < num_shader_stages; j++) {
 723                         /* Active cycles */
 724                         group_index = info->stages[j].active_group_index;
 725                         counter_index = info->stages[j].active_counter_index;
 726                         op_cycles += op->counters[group_index][counter_index];
 727
 728                         /* Stall cycles */
 729                         group_index = info->stages[j].stall_group_index;
 730                         counter_index = info->stages[j].stall_counter_index;
 731                         op_cycles += op->counters[group_index][counter_index];
 732                 }
 733
 734                 for (j = 0; j < num_shader_stages; j++) {
 735                         double active_cycles, stall_cycles, stage_cycles;
 736
 737                         /* Active cycles */
 738                         group_index = info->stages[j].active_group_index;
 739                         counter_index = info->stages[j].active_counter_index;
 740                         active_cycles = op->counters[group_index][counter_index];
 741
 742                         /* Stall cycles */
 743                         group_index = info->stages[j].stall_group_index;
 744                         counter_index = info->stages[j].stall_counter_index;
 745                         stall_cycles = op->counters[group_index][counter_index];
 746
 747                         stage_cycles = active_cycles + stall_cycles;
 748
 749                         per_stage = &sorted[i * num_shader_stages + j];
 750                         per_stage->metrics = op;
 751                         per_stage->stage = &info->stages[j];
 752                         if (op_cycles)
 753                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 754                         else
 755                                 per_stage->time_ns = 0.0;
 756                         if (stage_cycles)
 757                                 per_stage->active = active_cycles / stage_cycles;
 758                         else
 759                                 per_stage->active = 0.0;
 760                 }
 761         }
 762
 763         qsort_r (sorted, num_sorted, sizeof (*sorted),
 764                  time_compare, ctx->op_metrics);
 765
 766         for (i = 0; i < num_sorted; i++)
 767                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 768
 769         free (sorted);
 770 }
 771
 772 /* Called at program exit */
 773 static void
 774 metrics_exit (void)
 775 {
 776         if (verbose)
 777                 printf ("fips: terminating\n");
 778
 779         metrics_info_fini (&current_context.metrics_info);
 780 }
 781
 782
 783 void
 784 metrics_end_frame (void)
 785 {
 786         static int initialized = 0;
 787         static struct timeval tv_start, tv_now;
 788
 789         if (! initialized) {
 790                 gettimeofday (&tv_start, NULL);
 791                 atexit (metrics_exit);
 792                 if (getenv ("FIPS_VERBOSE"))
 793                         verbose = 1;
 794                 initialized = 1;
 795         }
 796
 797         frames++;
 798         gettimeofday (&tv_now, NULL);
 799
 800         /* Consume all timer queries that are ready. */
 801         timer_query_t *timer = current_context.timer_head;
 802
 803         while (timer) {
 804                 GLuint available, elapsed;
 805
 806                 glGetQueryObjectuiv (timer->id,
 807                                      GL_QUERY_RESULT_AVAILABLE, &available);
 808                 if (! available)
 809                         break;
 810
 811                 glGetQueryObjectuiv (timer->id,
 812                                      GL_QUERY_RESULT, &elapsed);
 813
 814                 accumulate_program_time (timer->op, elapsed);
 815
 816                 current_context.timer_head = timer->next;
 817                 if (current_context.timer_head == NULL)
 818                         current_context.timer_tail = NULL;
 819
 820                 glDeleteQueries (1, &timer->id);
 821
 822                 free (timer);
 823                 timer = current_context.timer_head;
 824         }
 825
 826         /* And similarly for all performance monitors that are ready. */
 827         monitor_t *monitor = current_context.monitor_head;
 828
 829         while (monitor) {
 830                 GLuint available, result_size, *result;
 831                 GLint bytes_written;
 832
 833                 glGetPerfMonitorCounterDataAMD (monitor->id,
 834                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 835                                                 sizeof (available), &available,
 836                                                 NULL);
 837                 if (! available)
 838                         break;
 839
 840                 glGetPerfMonitorCounterDataAMD (monitor->id,
 841                                                 GL_PERFMON_RESULT_SIZE_AMD,
 842                                                 sizeof (result_size),
 843                                                 &result_size, NULL);
 844
 845                 result = xmalloc (result_size);
 846
 847                 glGetPerfMonitorCounterDataAMD (monitor->id,
 848                                                 GL_PERFMON_RESULT_AMD,
 849                                                 result_size, result,
 850                                                 &bytes_written);
 851
 852                 accumulate_program_metrics (monitor->op, result, result_size);
 853
 854                 free (result);
 855
 856                 current_context.monitor_head = monitor->next;
 857                 if (current_context.monitor_head == NULL)
 858                         current_context.monitor_tail = NULL;
 859
 860                 glDeletePerfMonitorsAMD (1, &monitor->id);
 861
 862                 free (monitor);
 863                 monitor = current_context.monitor_head;
 864         }
 865
 866         if (frames % 60 == 0) {
 867                 double fps;
 868
 869                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 870                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 871
 872                 printf("FPS: %.3f\n", fps);
 873
 874                 print_program_metrics ();
 875         }
 876 }