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 /* A helper function, part of metrics_info_init below. */
 195
 196 typedef enum {
 197         SHADER_ACTIVE,
 198         SHADER_STALL
 199 } shader_phase_t;
 200
 201 static void
 202 _add_shader_stage (metrics_info_t *info, const char *name,
 203                    GLuint group_index, GLuint counter_index,
 204                    shader_phase_t phase)
 205 {
 206         shader_stage_info_t *stage;
 207         char *stage_name, *space;
 208         unsigned i;
 209
 210         stage_name = xstrdup (name);
 211
 212         /* Terminate the stage name at the first space.
 213          *
 214          * This is valid for counter names such as:
 215          *
 216          *      "Vertex Shader Active Time"
 217          * or
 218          *      "Vertex Shader Stall Time - Core Stall"
 219          */
 220         space = strchr (stage_name, ' ');
 221         if (space)
 222                 *space = '\0';
 223
 224         /* Look for an existing stage of the given name. */
 225         stage = NULL;
 226
 227         for (i = 0; i < info->num_shader_stages; i++) {
 228                 if (strcmp (info->stages[i].name, stage_name) == 0) {
 229                         stage = &info->stages[i];
 230                         break;
 231                 }
 232         }
 233
 234         if (stage == NULL) {
 235                 info->num_shader_stages++;
 236                 info->stages = xrealloc (info->stages,
 237                                          info->num_shader_stages *
 238                                          sizeof (shader_stage_info_t));
 239                 stage = &info->stages[info->num_shader_stages - 1];
 240                 stage->name = xstrdup (stage_name);
 241                 stage->active_group_index = 0;
 242                 stage->active_counter_index = 0;
 243                 stage->stall_group_index = 0;
 244                 stage->stall_counter_index = 0;
 245         }
 246
 247         if (phase == SHADER_ACTIVE) {
 248                 stage->active_group_index = group_index;
 249                 stage->active_counter_index = counter_index;
 250         } else {
 251                 stage->stall_group_index = group_index;
 252                 stage->stall_counter_index = counter_index;
 253         }
 254
 255         free (stage_name);
 256 }
 257
 258 void
 259 metrics_info_init (void)
 260 {
 261         unsigned i, j;
 262         GLuint *group_ids;
 263         metrics_info_t *info = &current_context.metrics_info;
 264
 265         glGetPerfMonitorGroupsAMD ((int *) &info->num_groups, 0, NULL);
 266
 267         group_ids = xmalloc (info->num_groups * sizeof (GLuint));
 268
 269         glGetPerfMonitorGroupsAMD (NULL, info->num_groups, group_ids);
 270
 271         info->groups = xmalloc (info->num_groups * sizeof (metrics_group_info_t));
 272
 273         for (i = 0; i < info->num_groups; i++)
 274                 metrics_group_info_init (&info->groups[i], group_ids[i]);
 275
 276         free (group_ids);
 277
 278         /* Identify each shader stage (by looking at
 279          * performance-counter names for specific patterns) and
 280          * initialize structures referring to the corresponding
 281          * counter numbers for each stage. */
 282         info->num_shader_stages = 0;
 283         info->stages = NULL;
 284
 285         for (i = 0; i < info->num_groups; i++) {
 286                 metrics_group_info_t *group = &info->groups[i];
 287                 for (j = 0; j < group->num_counters; j++) {
 288                         char *name = group->counter_names[j];
 289                         if (strstr (name, "Shader Active Time")) {
 290                                 _add_shader_stage (info, name, i, j,
 291                                                    SHADER_ACTIVE);
 292                         }
 293                         if (strstr (name, "Shader Stall Time")) {
 294                                 _add_shader_stage (info, name, i, j,
 295                                                    SHADER_STALL);
 296                         }
 297                 }
 298         }
 299
 300         info->initialized = 1;
 301 }
 302
 303 void
 304 metrics_info_fini (void)
 305 {
 306         metrics_info_t *info = &current_context.metrics_info;
 307         unsigned i;
 308         timer_query_t *timer, *timer_next;
 309         monitor_t *monitor, *monitor_next;
 310
 311         if (! info->initialized)
 312                 return;
 313
 314         for (timer = current_context.timer_head;
 315              timer;
 316              timer = timer_next)
 317         {
 318                 timer_next = timer->next;
 319                 free (timer);
 320         }
 321         current_context.timer_head = NULL;
 322         current_context.timer_tail = NULL;
 323
 324         for (monitor = current_context.monitor_head;
 325              monitor;
 326              monitor = monitor_next)
 327         {
 328                 monitor_next = monitor->next;
 329                 free (monitor);
 330         }
 331         current_context.monitor_head = NULL;
 332         current_context.monitor_tail = NULL;
 333
 334         for (i = 0; i < info->num_groups; i++)
 335                 metrics_group_info_fini (&info->groups[i]);
 336
 337         free (info->groups);
 338         info->groups = NULL;
 339
 340         for (i = 0; i < info->num_shader_stages; i++)
 341                 free (info->stages[i].name);
 342
 343         free (info->stages);
 344         info->stages = NULL;
 345
 346         info->initialized = 0;
 347 }
 348
 349 static const char *
 350 metrics_op_string (metrics_op_t op)
 351 {
 352         if (op >= METRICS_OP_SHADER)
 353                 return "Shader program";
 354
 355         switch (op)
 356         {
 357         case METRICS_OP_ACCUM:
 358                 return "glAccum*(+)";
 359         case METRICS_OP_BUFFER_DATA:
 360                 return "glBufferData(+)";
 361         case METRICS_OP_BUFFER_SUB_DATA:
 362                 return "glCopyBufferSubData*";
 363         case METRICS_OP_BITMAP:
 364                 return "glBitmap*";
 365         case METRICS_OP_BLIT_FRAMEBUFFER:
 366                 return "glBlitFramebuffer*";
 367         case METRICS_OP_CLEAR:
 368                 return "glClear(+)";
 369         case METRICS_OP_CLEAR_BUFFER_DATA:
 370                 return "glCearBufferData(+)";
 371         case METRICS_OP_CLEAR_TEX_IMAGE:
 372                 return "glClearTexImage(+)";
 373         case METRICS_OP_COPY_PIXELS:
 374                 return "glCopyPixels";
 375         case METRICS_OP_COPY_TEX_IMAGE:
 376                 return "glCopyTexImage(+)";
 377         case METRICS_OP_DRAW_PIXELS:
 378                 return "glDrawPixels";
 379         case METRICS_OP_GET_TEX_IMAGE:
 380                 return "glGetTexImage(+)";
 381         case METRICS_OP_READ_PIXELS:
 382                 return "glReadPixels*";
 383         case METRICS_OP_TEX_IMAGE:
 384                 return "glTexImage*(+)";
 385         default:
 386                 fprintf (stderr, "fips: Internal error: "
 387                          "Unknown metrics op value: %d\n", op);
 388                 exit (1);
 389         }
 390
 391         return "";
 392 }
 393
 394 void
 395 metrics_counter_start (void)
 396 {
 397         context_t *ctx = &current_context;
 398         timer_query_t *timer;
 399         monitor_t *monitor;
 400         unsigned i;
 401
 402         /* Create new timer query, add to list */
 403         timer = xmalloc (sizeof (timer_query_t));
 404
 405         timer->op = ctx->op;
 406         timer->next = NULL;
 407
 408         if (ctx->timer_tail) {
 409                 ctx->timer_tail->next = timer;
 410                 ctx->timer_tail = timer;
 411         } else {
 412                 ctx->timer_tail = timer;
 413                 ctx->timer_head = timer;
 414         }
 415
 416         /* Create a new performance-monitor query */
 417         monitor = xmalloc (sizeof (monitor_t));
 418
 419         monitor->op = ctx->op;
 420         monitor->next = NULL;
 421
 422         if (ctx->monitor_tail) {
 423                 ctx->monitor_tail->next = monitor;
 424                 ctx->monitor_tail = monitor;
 425         } else {
 426                 ctx->monitor_tail = monitor;
 427                 ctx->monitor_head = monitor;
 428         }
 429
 430         /* Initialize the timer_query and monitor objects */
 431         glGenQueries (1, &timer->id);
 432
 433         glGenPerfMonitorsAMD (1, &monitor->id);
 434
 435         for (i = 0; i < ctx->metrics_info.num_groups; i++)
 436         {
 437                 metrics_group_info_t *group;
 438                 int num_counters;
 439
 440                 group = &ctx->metrics_info.groups[i];
 441
 442                 num_counters = group->num_counters;
 443                 if (group->max_active_counters < group->num_counters)
 444                 {
 445                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 446                                  group->max_active_counters,
 447                                  group->num_counters, i);
 448                         num_counters = group->max_active_counters;
 449
 450                 }
 451
 452                 glSelectPerfMonitorCountersAMD(monitor->id,
 453                                                GL_TRUE, group->id,
 454                                                num_counters,
 455                                                group->counter_ids);
 456         }
 457
 458         /* Start the queries */
 459         glBeginQuery (GL_TIME_ELAPSED, timer->id);
 460
 461         glBeginPerfMonitorAMD (monitor->id);
 462 }
 463
 464 void
 465 metrics_counter_stop (void)
 466 {
 467         glEndQuery (GL_TIME_ELAPSED);
 468         glEndPerfMonitorAMD (current_context.monitor_tail->id);
 469 }
 470
 471 void
 472 metrics_set_current_op (metrics_op_t op)
 473 {
 474         current_context.op = op;
 475 }
 476
 477 metrics_op_t
 478 metrics_get_current_op (void)
 479 {
 480         return current_context.op;
 481 }
 482
 483 static void
 484 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 485 {
 486         metrics_info_t *info = &ctx->metrics_info;
 487         unsigned i, j;
 488
 489         metrics->op = op;
 490         metrics->time_ns = 0.0;
 491
 492         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 493
 494         for (i = 0; i < info->num_groups; i++) {
 495                 metrics->counters[i] = xmalloc (sizeof (double) *
 496                                                 info->groups[i].num_counters);
 497                 for (j = 0; j < info->groups[i].num_counters; j++)
 498                         metrics->counters[i][j] = 0.0;
 499         }
 500 }
 501
 502 static op_metrics_t *
 503 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 504 {
 505         unsigned i;
 506
 507         if (op >= ctx->num_op_metrics)
 508         {
 509                 ctx->op_metrics = realloc (ctx->op_metrics,
 510                                            (op + 1) * sizeof (op_metrics_t));
 511                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 512                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 513
 514                 ctx->num_op_metrics = op + 1;
 515         }
 516
 517         return &ctx->op_metrics[op];
 518 }
 519
 520 static void
 521 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 522 {
 523 #define CONSUME(var)                                                    \
 524         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 525         {                                                               \
 526                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 527                          "parsing results\n");                          \
 528                 break;                                                  \
 529         }                                                               \
 530         (var) = *((typeof(var) *) p);                                   \
 531         p += sizeof(var);
 532
 533         context_t *ctx = &current_context;
 534         metrics_info_t *info = &ctx->metrics_info;
 535         unsigned char *p = (unsigned char *) result;
 536
 537         while (p < ((unsigned char *) result) + size)
 538         {
 539                 GLuint group_id, group_index;
 540                 GLuint counter_id, counter_index;
 541                 metrics_group_info_t *group;
 542                 double value;
 543                 unsigned i;
 544
 545                 CONSUME (group_id);
 546                 CONSUME (counter_id);
 547
 548                 for (i = 0; i < info->num_groups; i++) {
 549                         if (info->groups[i].id == group_id)
 550                                 break;
 551                 }
 552                 group_index = i;
 553                 assert (group_index < info->num_groups);
 554                 group = &info->groups[group_index];
 555
 556                 for (i = 0; i < group->num_counters; i++) {
 557                         if (group->counter_ids[i] == counter_id)
 558                                 break;
 559                 }
 560                 counter_index = i;
 561                 assert (counter_index < group->num_counters);
 562
 563                 switch (group->counter_types[counter_index])
 564                 {
 565                         uint uint_value;
 566                         uint64_t uint64_value;
 567                         float float_value;
 568                 case GL_UNSIGNED_INT:
 569                         CONSUME (uint_value);
 570                         value = uint_value;
 571                         break;
 572                 case GL_UNSIGNED_INT64_AMD:
 573                         CONSUME (uint64_value);
 574                         value = uint64_value;
 575                         break;
 576                 case GL_PERCENTAGE_AMD:
 577                 case GL_FLOAT:
 578                         CONSUME (float_value);
 579                         value = float_value;
 580                         break;
 581                 default:
 582                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 583                                  group->counter_types[counter_index]);
 584                         value = 0.0;
 585                         break;
 586                 }
 587
 588                 ctx->op_metrics[op].counters[group_index][counter_index] += value;
 589         }
 590 }
 591
 592 static void
 593 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 594 {
 595         op_metrics_t *metrics;
 596
 597         metrics = ctx_get_op_metrics (&current_context, op);
 598
 599         metrics->time_ns += time_ns;
 600 }
 601
 602 typedef struct per_stage_metrics
 603 {
 604         op_metrics_t *metrics;
 605         shader_stage_info_t *stage;
 606         double time_ns;
 607         double active;
 608 } per_stage_metrics_t;
 609
 610 static int
 611 _is_shader_stage_counter (metrics_info_t *info,
 612                           unsigned group_index,
 613                           unsigned counter_index)
 614 {
 615         shader_stage_info_t *stage;
 616         unsigned i;
 617
 618         for (i = 0; i < info->num_shader_stages; i++) {
 619                 stage = &info->stages[i];
 620
 621                 if (stage->active_group_index == group_index &&
 622                     stage->active_counter_index == counter_index)
 623                 {
 624                         return 1;
 625                 }
 626
 627                 if (stage->stall_group_index == group_index &&
 628                     stage->stall_counter_index == counter_index)
 629                 {
 630                         return 1;
 631                 }
 632         }
 633
 634         return 0;
 635 }
 636
 637 static void
 638 print_per_stage_metrics (context_t *ctx,
 639                          per_stage_metrics_t *per_stage,
 640                          double total)
 641 {
 642         metrics_info_t *info = &ctx->metrics_info;
 643         op_metrics_t *metric = per_stage->metrics;
 644         metrics_group_info_t *group;
 645         const char *op_string;
 646         unsigned group_index, counter;
 647         double value;
 648
 649         /* Don't print anything for stages with no alloted time. */
 650         if (per_stage->time_ns == 0.0)
 651                 return;
 652
 653         op_string = metrics_op_string (metric->op);
 654
 655         printf ("%21s", op_string);
 656
 657         if (metric->op >= METRICS_OP_SHADER) {
 658                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 659         } else {
 660                 printf ("    ");
 661
 662         }
 663         printf (" %cS:", per_stage->stage->name[0]);
 664
 665         printf ("\t%7.2f ms (%4.1f%%)",
 666                 per_stage->time_ns / 1e6,
 667                 per_stage->time_ns / total * 100);
 668
 669         if (per_stage->active)
 670                 printf (", %4.1f%% active", per_stage->active * 100);
 671
 672         printf ("\n");
 673
 674         /* I'm not seeing a lot of value printing the rest of these
 675          * performance counters by default yet. Use --verbose to get
 676          * them for now. */
 677         if (! verbose)
 678                 return;
 679
 680         printf ("[");
 681         for (group_index = 0; group_index < info->num_groups; group_index++) {
 682                 group = &info->groups[group_index];
 683                 for (counter = 0; counter < group->num_counters; counter++) {
 684
 685                         /* Don't print this counter value if it's a
 686                          * per-stage cycle counter, (which we have
 687                          * already accounted for). */
 688                         if (_is_shader_stage_counter (info, group_index, counter))
 689                                 continue;
 690
 691                         value = metric->counters[group_index][counter];
 692                         if (value == 0.0)
 693                                 continue;
 694                         printf ("%s: %.2f ", group->counter_names[counter],
 695                                 value / 1e6);
 696                 }
 697         }
 698         printf ("]\n");
 699 }
 700
 701 static int
 702 time_compare(const void *in_a, const void *in_b, void *arg unused)
 703 {
 704         const per_stage_metrics_t *a = in_a;
 705         const per_stage_metrics_t *b = in_b;
 706
 707
 708         if (a->time_ns < b->time_ns)
 709                 return -1;
 710         if (a->time_ns > b->time_ns)
 711                 return 1;
 712         return 0;
 713 }
 714
 715 static void
 716 print_program_metrics (void)
 717 {
 718         context_t *ctx = &current_context;
 719         metrics_info_t *info = &ctx->metrics_info;
 720         unsigned num_shader_stages = info->num_shader_stages;
 721         per_stage_metrics_t *sorted, *per_stage;
 722         double total_time, op_cycles;
 723         op_metrics_t *op;
 724         unsigned group_index, counter_index;
 725         unsigned i, j, num_sorted;
 726
 727         /* Make a sorted list of the per-stage operations by time
 728          * used, and figure out the total so we can print percentages.
 729          */
 730         num_sorted = ctx->num_op_metrics * num_shader_stages;
 731
 732         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 733
 734         total_time = 0.0;
 735
 736         for (i = 0; i < ctx->num_op_metrics; i++) {
 737
 738                 op = &ctx->op_metrics[i];
 739
 740                 /* Accumulate total time across all ops. */
 741                 total_time += op->time_ns;
 742
 743                 /* Also, find total cycles in all stages of this op. */
 744                 op_cycles = 0.0;
 745
 746                 for (j = 0; j < num_shader_stages; j++) {
 747                         /* Active cycles */
 748                         group_index = info->stages[j].active_group_index;
 749                         counter_index = info->stages[j].active_counter_index;
 750                         op_cycles += op->counters[group_index][counter_index];
 751
 752                         /* Stall cycles */
 753                         group_index = info->stages[j].stall_group_index;
 754                         counter_index = info->stages[j].stall_counter_index;
 755                         op_cycles += op->counters[group_index][counter_index];
 756                 }
 757
 758                 for (j = 0; j < num_shader_stages; j++) {
 759                         double active_cycles, stall_cycles, stage_cycles;
 760
 761                         /* Active cycles */
 762                         group_index = info->stages[j].active_group_index;
 763                         counter_index = info->stages[j].active_counter_index;
 764                         active_cycles = op->counters[group_index][counter_index];
 765
 766                         /* Stall cycles */
 767                         group_index = info->stages[j].stall_group_index;
 768                         counter_index = info->stages[j].stall_counter_index;
 769                         stall_cycles = op->counters[group_index][counter_index];
 770
 771                         stage_cycles = active_cycles + stall_cycles;
 772
 773                         per_stage = &sorted[i * num_shader_stages + j];
 774                         per_stage->metrics = op;
 775                         per_stage->stage = &info->stages[j];
 776                         if (op_cycles)
 777                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 778                         else
 779                                 per_stage->time_ns = 0.0;
 780                         if (stage_cycles)
 781                                 per_stage->active = active_cycles / stage_cycles;
 782                         else
 783                                 per_stage->active = 0.0;
 784                 }
 785         }
 786
 787         qsort_r (sorted, num_sorted, sizeof (*sorted),
 788                  time_compare, ctx->op_metrics);
 789
 790         for (i = 0; i < num_sorted; i++)
 791                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 792
 793         free (sorted);
 794 }
 795
 796 /* Called at program exit */
 797 static void
 798 metrics_exit (void)
 799 {
 800         if (verbose)
 801                 printf ("fips: terminating\n");
 802
 803         metrics_info_fini ();
 804 }
 805
 806
 807 void
 808 metrics_end_frame (void)
 809 {
 810         static int initialized = 0;
 811         static struct timeval tv_start, tv_now;
 812
 813         if (! initialized) {
 814                 gettimeofday (&tv_start, NULL);
 815                 atexit (metrics_exit);
 816                 if (getenv ("FIPS_VERBOSE"))
 817                         verbose = 1;
 818                 initialized = 1;
 819         }
 820
 821         frames++;
 822         gettimeofday (&tv_now, NULL);
 823
 824         /* Consume all timer queries that are ready. */
 825         timer_query_t *timer = current_context.timer_head;
 826
 827         while (timer) {
 828                 GLuint available, elapsed;
 829
 830                 glGetQueryObjectuiv (timer->id,
 831                                      GL_QUERY_RESULT_AVAILABLE, &available);
 832                 if (! available)
 833                         break;
 834
 835                 glGetQueryObjectuiv (timer->id,
 836                                      GL_QUERY_RESULT, &elapsed);
 837
 838                 accumulate_program_time (timer->op, elapsed);
 839
 840                 current_context.timer_head = timer->next;
 841                 if (current_context.timer_head == NULL)
 842                         current_context.timer_tail = NULL;
 843
 844                 glDeleteQueries (1, &timer->id);
 845
 846                 free (timer);
 847                 timer = current_context.timer_head;
 848         }
 849
 850         /* And similarly for all performance monitors that are ready. */
 851         monitor_t *monitor = current_context.monitor_head;
 852
 853         while (monitor) {
 854                 GLuint available, result_size, *result;
 855                 GLint bytes_written;
 856
 857                 glGetPerfMonitorCounterDataAMD (monitor->id,
 858                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 859                                                 sizeof (available), &available,
 860                                                 NULL);
 861                 if (! available)
 862                         break;
 863
 864                 glGetPerfMonitorCounterDataAMD (monitor->id,
 865                                                 GL_PERFMON_RESULT_SIZE_AMD,
 866                                                 sizeof (result_size),
 867                                                 &result_size, NULL);
 868
 869                 result = xmalloc (result_size);
 870
 871                 glGetPerfMonitorCounterDataAMD (monitor->id,
 872                                                 GL_PERFMON_RESULT_AMD,
 873                                                 result_size, result,
 874                                                 &bytes_written);
 875
 876                 accumulate_program_metrics (monitor->op, result, result_size);
 877
 878                 free (result);
 879
 880                 current_context.monitor_head = monitor->next;
 881                 if (current_context.monitor_head == NULL)
 882                         current_context.monitor_tail = NULL;
 883
 884                 glDeletePerfMonitorsAMD (1, &monitor->id);
 885
 886                 free (monitor);
 887                 monitor = current_context.monitor_head;
 888         }
 889
 890         if (frames % 60 == 0) {
 891                 double fps;
 892
 893                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 894                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 895
 896                 printf("FPS: %.3f\n", fps);
 897
 898                 print_program_metrics ();
 899         }
 900 }