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         context_t *ctx = &current_context;
 307         metrics_info_t *info = &ctx->metrics_info;
 308         unsigned i;
 309         timer_query_t *timer, *timer_next;
 310         monitor_t *monitor, *monitor_next;
 311
 312         if (! info->initialized)
 313                 return;
 314
 315         for (timer = ctx->timer_head;
 316              timer;
 317              timer = timer_next)
 318         {
 319                 timer_next = timer->next;
 320                 free (timer);
 321         }
 322         ctx->timer_head = NULL;
 323         ctx->timer_tail = NULL;
 324
 325         for (monitor = ctx->monitor_head;
 326              monitor;
 327              monitor = monitor_next)
 328         {
 329                 monitor_next = monitor->next;
 330                 free (monitor);
 331         }
 332         ctx->monitor_head = NULL;
 333         ctx->monitor_tail = NULL;
 334
 335         for (i = 0; i < info->num_groups; i++)
 336                 metrics_group_info_fini (&info->groups[i]);
 337
 338         free (info->groups);
 339         info->groups = NULL;
 340
 341         for (i = 0; i < info->num_shader_stages; i++)
 342                 free (info->stages[i].name);
 343
 344         free (info->stages);
 345         info->stages = NULL;
 346
 347         info->initialized = 0;
 348 }
 349
 350 static const char *
 351 metrics_op_string (metrics_op_t op)
 352 {
 353         if (op >= METRICS_OP_SHADER)
 354                 return "Shader program";
 355
 356         switch (op)
 357         {
 358         case METRICS_OP_ACCUM:
 359                 return "glAccum*(+)";
 360         case METRICS_OP_BUFFER_DATA:
 361                 return "glBufferData(+)";
 362         case METRICS_OP_BUFFER_SUB_DATA:
 363                 return "glCopyBufferSubData*";
 364         case METRICS_OP_BITMAP:
 365                 return "glBitmap*";
 366         case METRICS_OP_BLIT_FRAMEBUFFER:
 367                 return "glBlitFramebuffer*";
 368         case METRICS_OP_CLEAR:
 369                 return "glClear(+)";
 370         case METRICS_OP_CLEAR_BUFFER_DATA:
 371                 return "glCearBufferData(+)";
 372         case METRICS_OP_CLEAR_TEX_IMAGE:
 373                 return "glClearTexImage(+)";
 374         case METRICS_OP_COPY_PIXELS:
 375                 return "glCopyPixels";
 376         case METRICS_OP_COPY_TEX_IMAGE:
 377                 return "glCopyTexImage(+)";
 378         case METRICS_OP_DRAW_PIXELS:
 379                 return "glDrawPixels";
 380         case METRICS_OP_GET_TEX_IMAGE:
 381                 return "glGetTexImage(+)";
 382         case METRICS_OP_READ_PIXELS:
 383                 return "glReadPixels*";
 384         case METRICS_OP_TEX_IMAGE:
 385                 return "glTexImage*(+)";
 386         default:
 387                 fprintf (stderr, "fips: Internal error: "
 388                          "Unknown metrics op value: %d\n", op);
 389                 exit (1);
 390         }
 391
 392         return "";
 393 }
 394
 395 void
 396 metrics_counter_start (void)
 397 {
 398         context_t *ctx = &current_context;
 399         timer_query_t *timer;
 400         monitor_t *monitor;
 401         unsigned i;
 402
 403         /* Create new timer query, add to list */
 404         timer = xmalloc (sizeof (timer_query_t));
 405
 406         timer->op = ctx->op;
 407         timer->next = NULL;
 408
 409         if (ctx->timer_tail) {
 410                 ctx->timer_tail->next = timer;
 411                 ctx->timer_tail = timer;
 412         } else {
 413                 ctx->timer_tail = timer;
 414                 ctx->timer_head = timer;
 415         }
 416
 417         /* Create a new performance-monitor query */
 418         monitor = xmalloc (sizeof (monitor_t));
 419
 420         monitor->op = ctx->op;
 421         monitor->next = NULL;
 422
 423         if (ctx->monitor_tail) {
 424                 ctx->monitor_tail->next = monitor;
 425                 ctx->monitor_tail = monitor;
 426         } else {
 427                 ctx->monitor_tail = monitor;
 428                 ctx->monitor_head = monitor;
 429         }
 430
 431         /* Initialize the timer_query and monitor objects */
 432         glGenQueries (1, &timer->id);
 433
 434         glGenPerfMonitorsAMD (1, &monitor->id);
 435
 436         for (i = 0; i < ctx->metrics_info.num_groups; i++)
 437         {
 438                 metrics_group_info_t *group;
 439                 int num_counters;
 440
 441                 group = &ctx->metrics_info.groups[i];
 442
 443                 num_counters = group->num_counters;
 444                 if (group->max_active_counters < group->num_counters)
 445                 {
 446                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 447                                  group->max_active_counters,
 448                                  group->num_counters, i);
 449                         num_counters = group->max_active_counters;
 450
 451                 }
 452
 453                 glSelectPerfMonitorCountersAMD(monitor->id,
 454                                                GL_TRUE, group->id,
 455                                                num_counters,
 456                                                group->counter_ids);
 457         }
 458
 459         /* Start the queries */
 460         glBeginQuery (GL_TIME_ELAPSED, timer->id);
 461
 462         glBeginPerfMonitorAMD (monitor->id);
 463 }
 464
 465 void
 466 metrics_counter_stop (void)
 467 {
 468         glEndQuery (GL_TIME_ELAPSED);
 469         glEndPerfMonitorAMD (current_context.monitor_tail->id);
 470 }
 471
 472 void
 473 metrics_set_current_op (metrics_op_t op)
 474 {
 475         current_context.op = op;
 476 }
 477
 478 metrics_op_t
 479 metrics_get_current_op (void)
 480 {
 481         return current_context.op;
 482 }
 483
 484 static void
 485 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 486 {
 487         metrics_info_t *info = &ctx->metrics_info;
 488         unsigned i, j;
 489
 490         metrics->op = op;
 491         metrics->time_ns = 0.0;
 492
 493         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 494
 495         for (i = 0; i < info->num_groups; i++) {
 496                 metrics->counters[i] = xmalloc (sizeof (double) *
 497                                                 info->groups[i].num_counters);
 498                 for (j = 0; j < info->groups[i].num_counters; j++)
 499                         metrics->counters[i][j] = 0.0;
 500         }
 501 }
 502
 503 static op_metrics_t *
 504 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 505 {
 506         unsigned i;
 507
 508         if (op >= ctx->num_op_metrics)
 509         {
 510                 ctx->op_metrics = realloc (ctx->op_metrics,
 511                                            (op + 1) * sizeof (op_metrics_t));
 512                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 513                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 514
 515                 ctx->num_op_metrics = op + 1;
 516         }
 517
 518         return &ctx->op_metrics[op];
 519 }
 520
 521 static void
 522 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 523 {
 524 #define CONSUME(var)                                                    \
 525         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 526         {                                                               \
 527                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 528                          "parsing results\n");                          \
 529                 break;                                                  \
 530         }                                                               \
 531         (var) = *((typeof(var) *) p);                                   \
 532         p += sizeof(var);
 533
 534         context_t *ctx = &current_context;
 535         metrics_info_t *info = &ctx->metrics_info;
 536         op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
 537         unsigned char *p = (unsigned char *) result;
 538
 539         while (p < ((unsigned char *) result) + size)
 540         {
 541                 GLuint group_id, group_index;
 542                 GLuint counter_id, counter_index;
 543                 metrics_group_info_t *group;
 544                 double value;
 545                 unsigned i;
 546
 547                 CONSUME (group_id);
 548                 CONSUME (counter_id);
 549
 550                 for (i = 0; i < info->num_groups; i++) {
 551                         if (info->groups[i].id == group_id)
 552                                 break;
 553                 }
 554                 group_index = i;
 555                 assert (group_index < info->num_groups);
 556                 group = &info->groups[group_index];
 557
 558                 for (i = 0; i < group->num_counters; i++) {
 559                         if (group->counter_ids[i] == counter_id)
 560                                 break;
 561                 }
 562                 counter_index = i;
 563                 assert (counter_index < group->num_counters);
 564
 565                 switch (group->counter_types[counter_index])
 566                 {
 567                         uint uint_value;
 568                         uint64_t uint64_value;
 569                         float float_value;
 570                 case GL_UNSIGNED_INT:
 571                         CONSUME (uint_value);
 572                         value = uint_value;
 573                         break;
 574                 case GL_UNSIGNED_INT64_AMD:
 575                         CONSUME (uint64_value);
 576                         value = uint64_value;
 577                         break;
 578                 case GL_PERCENTAGE_AMD:
 579                 case GL_FLOAT:
 580                         CONSUME (float_value);
 581                         value = float_value;
 582                         break;
 583                 default:
 584                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 585                                  group->counter_types[counter_index]);
 586                         value = 0.0;
 587                         break;
 588                 }
 589
 590                 metrics->counters[group_index][counter_index] += value;
 591         }
 592 }
 593
 594 static void
 595 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 596 {
 597         op_metrics_t *metrics;
 598
 599         metrics = ctx_get_op_metrics (&current_context, op);
 600
 601         metrics->time_ns += time_ns;
 602 }
 603
 604 typedef struct per_stage_metrics
 605 {
 606         op_metrics_t *metrics;
 607         shader_stage_info_t *stage;
 608         double time_ns;
 609         double active;
 610 } per_stage_metrics_t;
 611
 612 static int
 613 _is_shader_stage_counter (metrics_info_t *info,
 614                           unsigned group_index,
 615                           unsigned counter_index)
 616 {
 617         shader_stage_info_t *stage;
 618         unsigned i;
 619
 620         for (i = 0; i < info->num_shader_stages; i++) {
 621                 stage = &info->stages[i];
 622
 623                 if (stage->active_group_index == group_index &&
 624                     stage->active_counter_index == counter_index)
 625                 {
 626                         return 1;
 627                 }
 628
 629                 if (stage->stall_group_index == group_index &&
 630                     stage->stall_counter_index == counter_index)
 631                 {
 632                         return 1;
 633                 }
 634         }
 635
 636         return 0;
 637 }
 638
 639 static void
 640 print_per_stage_metrics (context_t *ctx,
 641                          per_stage_metrics_t *per_stage,
 642                          double total)
 643 {
 644         metrics_info_t *info = &ctx->metrics_info;
 645         op_metrics_t *metric = per_stage->metrics;
 646         metrics_group_info_t *group;
 647         const char *op_string;
 648         unsigned group_index, counter;
 649         double value;
 650
 651         /* Don't print anything for stages with no alloted time. */
 652         if (per_stage->time_ns == 0.0)
 653                 return;
 654
 655         op_string = metrics_op_string (metric->op);
 656
 657         printf ("%21s", op_string);
 658
 659         if (metric->op >= METRICS_OP_SHADER) {
 660                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 661         } else {
 662                 printf ("    ");
 663
 664         }
 665
 666         if (per_stage->stage)
 667                 printf (" %cS:", per_stage->stage->name[0]);
 668         else
 669                 printf ("   :");
 670
 671         printf ("\t%7.2f ms (%4.1f%%)",
 672                 per_stage->time_ns / 1e6,
 673                 per_stage->time_ns / total * 100);
 674
 675         if (per_stage->active)
 676                 printf (", %4.1f%% active", per_stage->active * 100);
 677
 678         printf ("\n");
 679
 680         /* I'm not seeing a lot of value printing the rest of these
 681          * performance counters by default yet. Use --verbose to get
 682          * them for now. */
 683         if (! verbose)
 684                 return;
 685
 686         printf ("[");
 687         for (group_index = 0; group_index < info->num_groups; group_index++) {
 688                 group = &info->groups[group_index];
 689                 for (counter = 0; counter < group->num_counters; counter++) {
 690
 691                         /* Don't print this counter value if it's a
 692                          * per-stage cycle counter, (which we have
 693                          * already accounted for). */
 694                         if (_is_shader_stage_counter (info, group_index, counter))
 695                                 continue;
 696
 697                         value = metric->counters[group_index][counter];
 698                         if (value == 0.0)
 699                                 continue;
 700                         printf ("%s: %.2f ", group->counter_names[counter],
 701                                 value / 1e6);
 702                 }
 703         }
 704         printf ("]\n");
 705 }
 706
 707 static int
 708 time_compare(const void *in_a, const void *in_b, void *arg unused)
 709 {
 710         const per_stage_metrics_t *a = in_a;
 711         const per_stage_metrics_t *b = in_b;
 712
 713
 714         if (a->time_ns < b->time_ns)
 715                 return -1;
 716         if (a->time_ns > b->time_ns)
 717                 return 1;
 718         return 0;
 719 }
 720
 721 static void
 722 print_program_metrics (void)
 723 {
 724         context_t *ctx = &current_context;
 725         metrics_info_t *info = &ctx->metrics_info;
 726         unsigned num_shader_stages = info->num_shader_stages;
 727         per_stage_metrics_t *sorted, *per_stage;
 728         double total_time, op_cycles;
 729         op_metrics_t *op;
 730         unsigned group_index, counter_index;
 731         unsigned i, j, num_sorted;
 732
 733         /* Make a sorted list of the per-stage operations by time
 734          * used, and figure out the total so we can print percentages.
 735          */
 736         num_sorted = ctx->num_op_metrics * num_shader_stages;
 737
 738         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 739
 740         total_time = 0.0;
 741
 742         for (i = 0; i < ctx->num_op_metrics; i++) {
 743
 744                 op = &ctx->op_metrics[i];
 745
 746                 /* Accumulate total time across all ops. */
 747                 total_time += op->time_ns;
 748
 749                 /* Also, find total cycles in all stages of this op. */
 750                 op_cycles = 0.0;
 751
 752                 for (j = 0; j < num_shader_stages; j++) {
 753                         /* Active cycles */
 754                         group_index = info->stages[j].active_group_index;
 755                         counter_index = info->stages[j].active_counter_index;
 756                         op_cycles += op->counters[group_index][counter_index];
 757
 758                         /* Stall cycles */
 759                         group_index = info->stages[j].stall_group_index;
 760                         counter_index = info->stages[j].stall_counter_index;
 761                         op_cycles += op->counters[group_index][counter_index];
 762                 }
 763
 764                 for (j = 0; j < num_shader_stages; j++) {
 765                         double active_cycles, stall_cycles, stage_cycles;
 766
 767                         /* Active cycles */
 768                         group_index = info->stages[j].active_group_index;
 769                         counter_index = info->stages[j].active_counter_index;
 770                         active_cycles = op->counters[group_index][counter_index];
 771
 772                         /* Stall cycles */
 773                         group_index = info->stages[j].stall_group_index;
 774                         counter_index = info->stages[j].stall_counter_index;
 775                         stall_cycles = op->counters[group_index][counter_index];
 776
 777                         stage_cycles = active_cycles + stall_cycles;
 778
 779                         per_stage = &sorted[i * num_shader_stages + j];
 780                         per_stage->metrics = op;
 781
 782                         if (op_cycles) {
 783                                 per_stage->stage = &info->stages[j];
 784                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 785                         } else {
 786                                 /* If we don't have any per-stage cycle counts
 787                                  * for this operation, then use the first
 788                                  * stage as a placeholder for all the time,
 789                                  * but NULL-ify the stage info so that the
 790                                  * report doesn't lie about this time being
 791                                  * from any particular stage. */
 792                                 per_stage->stage = NULL;
 793                                 if (j == 0) {
 794                                         per_stage->time_ns = op->time_ns;
 795                                 } else {
 796                                         per_stage->time_ns = 0.0;
 797                                 }
 798                         }
 799
 800                         if (stage_cycles) {
 801                                 per_stage->active = active_cycles / stage_cycles;
 802                         } else {
 803                                 per_stage->active = 0.0;
 804                         }
 805                 }
 806         }
 807
 808         qsort_r (sorted, num_sorted, sizeof (*sorted),
 809                  time_compare, ctx->op_metrics);
 810
 811         for (i = 0; i < num_sorted; i++)
 812                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 813
 814         free (sorted);
 815 }
 816
 817 /* Called at program exit */
 818 static void
 819 metrics_exit (void)
 820 {
 821         if (verbose)
 822                 printf ("fips: terminating\n");
 823
 824         metrics_info_fini ();
 825 }
 826
 827
 828 void
 829 metrics_end_frame (void)
 830 {
 831         context_t *ctx = &current_context;
 832         static int initialized = 0;
 833         static struct timeval tv_start, tv_now;
 834
 835         if (! initialized) {
 836                 gettimeofday (&tv_start, NULL);
 837                 atexit (metrics_exit);
 838                 if (getenv ("FIPS_VERBOSE"))
 839                         verbose = 1;
 840                 initialized = 1;
 841         }
 842
 843         frames++;
 844         gettimeofday (&tv_now, NULL);
 845
 846         /* Consume all timer queries that are ready. */
 847         timer_query_t *timer = ctx->timer_head;
 848
 849         while (timer) {
 850                 GLuint available, elapsed;
 851
 852                 glGetQueryObjectuiv (timer->id,
 853                                      GL_QUERY_RESULT_AVAILABLE, &available);
 854                 if (! available)
 855                         break;
 856
 857                 glGetQueryObjectuiv (timer->id,
 858                                      GL_QUERY_RESULT, &elapsed);
 859
 860                 accumulate_program_time (timer->op, elapsed);
 861
 862                 ctx->timer_head = timer->next;
 863                 if (ctx->timer_head == NULL)
 864                         ctx->timer_tail = NULL;
 865
 866                 glDeleteQueries (1, &timer->id);
 867
 868                 free (timer);
 869                 timer = ctx->timer_head;
 870         }
 871
 872         /* And similarly for all performance monitors that are ready. */
 873         monitor_t *monitor = ctx->monitor_head;
 874
 875         while (monitor) {
 876                 GLuint available, result_size, *result;
 877                 GLint bytes_written;
 878
 879                 glGetPerfMonitorCounterDataAMD (monitor->id,
 880                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 881                                                 sizeof (available), &available,
 882                                                 NULL);
 883                 if (! available)
 884                         break;
 885
 886                 glGetPerfMonitorCounterDataAMD (monitor->id,
 887                                                 GL_PERFMON_RESULT_SIZE_AMD,
 888                                                 sizeof (result_size),
 889                                                 &result_size, NULL);
 890
 891                 result = xmalloc (result_size);
 892
 893                 glGetPerfMonitorCounterDataAMD (monitor->id,
 894                                                 GL_PERFMON_RESULT_AMD,
 895                                                 result_size, result,
 896                                                 &bytes_written);
 897
 898                 accumulate_program_metrics (monitor->op, result, result_size);
 899
 900                 free (result);
 901
 902                 ctx->monitor_head = monitor->next;
 903                 if (ctx->monitor_head == NULL)
 904                         ctx->monitor_tail = NULL;
 905
 906                 glDeletePerfMonitorsAMD (1, &monitor->id);
 907
 908                 free (monitor);
 909                 monitor = ctx->monitor_head;
 910         }
 911
 912         if (frames % 15 == 0) {
 913                 double fps;
 914
 915                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 916                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 917
 918                 printf("FPS: %.3f\n", fps);
 919
 920                 print_program_metrics ();
 921         }
 922 }