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         /* GL_TIME_ELAPSED query for which glEndQuery has not yet
 109          * been called. */
 110         unsigned timer_begun_id;
 111
 112         /* GL_TIME_ELAPSED queries for which glEndQuery has been
 113          * called, (but results have not yet been queried). */
 114         timer_query_t *timer_head;
 115         timer_query_t *timer_tail;
 116
 117         /* Performance monitor for which glEndPerfMonitorAMD has not
 118          * yet been called. */
 119         unsigned monitor_begun_id;
 120
 121         /* Performance monitors for which glEndPerfMonitorAMD has
 122          * been called, (but results have not yet been queried). */
 123         monitor_t *monitor_head;
 124         monitor_t *monitor_tail;
 125
 126         int monitors_in_flight;
 127
 128         unsigned num_op_metrics;
 129         op_metrics_t *op_metrics;
 130 } context_t;
 131
 132 /* FIXME: Need a map from integers to context objects and track the
 133  * current context with glXMakeContextCurrent, eglMakeCurrent, etc. */
 134
 135 context_t current_context;
 136
 137 int frames;
 138 int verbose;
 139
 140 #define MAX_MONITORS_IN_FLIGHT 1000
 141
 142 void
 143 metrics_collect_available (void);
 144
 145 static void
 146 metrics_group_info_init (metrics_group_info_t *group, GLuint id)
 147 {
 148         GLsizei length;
 149         unsigned i;
 150
 151         group->id = id;
 152
 153         /* Get name */
 154         glGetPerfMonitorGroupStringAMD (id, 0, &length, NULL);
 155
 156         group->name = xmalloc (length + 1);
 157
 158         glGetPerfMonitorGroupStringAMD (id, length + 1, NULL, group->name);
 159
 160         /* Get number of counters */
 161         group->num_counters = 0;
 162         group->max_active_counters = 0;
 163         glGetPerfMonitorCountersAMD (group->id,
 164                                      (int *) &group->num_counters,
 165                                      (int *) &group->max_active_counters,
 166                                      0, NULL);
 167
 168         /* Get counter numbers */
 169         group->counter_ids = xmalloc (group->num_counters * sizeof (GLuint));
 170
 171         glGetPerfMonitorCountersAMD (group->id, NULL, NULL,
 172                                      group->num_counters,
 173                                      group->counter_ids);
 174
 175         /* Get counter names */
 176         group->counter_names = xmalloc (group->num_counters * sizeof (char *));
 177         group->counter_types = xmalloc (group->num_counters * sizeof (GLuint));
 178
 179         for (i = 0; i < group->num_counters; i++) {
 180                 glGetPerfMonitorCounterInfoAMD (group->id,
 181                                                 group->counter_ids[i],
 182                                                 GL_COUNTER_TYPE_AMD,
 183                                                 &group->counter_types[i]);
 184
 185                 glGetPerfMonitorCounterStringAMD (group->id,
 186                                                   group->counter_ids[i],
 187                                                   0, &length, NULL);
 188
 189                 group->counter_names[i] = xmalloc (length + 1);
 190
 191                 glGetPerfMonitorCounterStringAMD (group->id,
 192                                                   group->counter_ids[i],
 193                                                   length + 1, NULL,
 194                                                   group->counter_names[i]);
 195         }
 196 }
 197
 198 static void
 199 metrics_group_info_fini (metrics_group_info_t *group)
 200 {
 201         unsigned i;
 202
 203         for (i = 0; i < group->num_counters; i++)
 204                 free (group->counter_names[i]);
 205
 206         free (group->counter_types);
 207         free (group->counter_names);
 208         free (group->counter_ids);
 209
 210         free (group->name);
 211 }
 212
 213 /* A helper function, part of metrics_info_init below. */
 214
 215 typedef enum {
 216         SHADER_ACTIVE,
 217         SHADER_STALL
 218 } shader_phase_t;
 219
 220 static void
 221 _add_shader_stage (metrics_info_t *info, const char *name,
 222                    GLuint group_index, GLuint counter_index,
 223                    shader_phase_t phase)
 224 {
 225         shader_stage_info_t *stage;
 226         char *stage_name, *space;
 227         unsigned i;
 228
 229         stage_name = xstrdup (name);
 230
 231         /* Terminate the stage name at the first space.
 232          *
 233          * This is valid for counter names such as:
 234          *
 235          *      "Vertex Shader Active Time"
 236          * or
 237          *      "Vertex Shader Stall Time - Core Stall"
 238          */
 239         space = strchr (stage_name, ' ');
 240         if (space)
 241                 *space = '\0';
 242
 243         /* Look for an existing stage of the given name. */
 244         stage = NULL;
 245
 246         for (i = 0; i < info->num_shader_stages; i++) {
 247                 if (strcmp (info->stages[i].name, stage_name) == 0) {
 248                         stage = &info->stages[i];
 249                         break;
 250                 }
 251         }
 252
 253         if (stage == NULL) {
 254                 info->num_shader_stages++;
 255                 info->stages = xrealloc (info->stages,
 256                                          info->num_shader_stages *
 257                                          sizeof (shader_stage_info_t));
 258                 stage = &info->stages[info->num_shader_stages - 1];
 259                 stage->name = xstrdup (stage_name);
 260                 stage->active_group_index = 0;
 261                 stage->active_counter_index = 0;
 262                 stage->stall_group_index = 0;
 263                 stage->stall_counter_index = 0;
 264         }
 265
 266         if (phase == SHADER_ACTIVE) {
 267                 stage->active_group_index = group_index;
 268                 stage->active_counter_index = counter_index;
 269         } else {
 270                 stage->stall_group_index = group_index;
 271                 stage->stall_counter_index = counter_index;
 272         }
 273
 274         free (stage_name);
 275 }
 276
 277 void
 278 metrics_info_init (void)
 279 {
 280         unsigned i, j;
 281         GLuint *group_ids;
 282         metrics_info_t *info = &current_context.metrics_info;
 283
 284         glGetPerfMonitorGroupsAMD ((int *) &info->num_groups, 0, NULL);
 285
 286         group_ids = xmalloc (info->num_groups * sizeof (GLuint));
 287
 288         glGetPerfMonitorGroupsAMD (NULL, info->num_groups, group_ids);
 289
 290         info->groups = xmalloc (info->num_groups * sizeof (metrics_group_info_t));
 291
 292         for (i = 0; i < info->num_groups; i++)
 293                 metrics_group_info_init (&info->groups[i], group_ids[i]);
 294
 295         free (group_ids);
 296
 297         /* Identify each shader stage (by looking at
 298          * performance-counter names for specific patterns) and
 299          * initialize structures referring to the corresponding
 300          * counter numbers for each stage. */
 301         info->num_shader_stages = 0;
 302         info->stages = NULL;
 303
 304         for (i = 0; i < info->num_groups; i++) {
 305                 metrics_group_info_t *group = &info->groups[i];
 306                 for (j = 0; j < group->num_counters; j++) {
 307                         char *name = group->counter_names[j];
 308                         if (strstr (name, "Shader Active Time")) {
 309                                 _add_shader_stage (info, name, i, j,
 310                                                    SHADER_ACTIVE);
 311                         }
 312                         if (strstr (name, "Shader Stall Time")) {
 313                                 _add_shader_stage (info, name, i, j,
 314                                                    SHADER_STALL);
 315                         }
 316                 }
 317         }
 318
 319         info->initialized = 1;
 320 }
 321
 322 void
 323 metrics_info_fini (void)
 324 {
 325         context_t *ctx = &current_context;
 326         metrics_info_t *info = &ctx->metrics_info;
 327         unsigned i;
 328         timer_query_t *timer, *timer_next;
 329         monitor_t *monitor, *monitor_next;
 330
 331         if (! info->initialized)
 332                 return;
 333
 334         if (ctx->timer_begun_id) {
 335                 glEndQuery (GL_TIME_ELAPSED);
 336                 glDeleteQueries (1, &ctx->timer_begun_id);
 337                 ctx->timer_begun_id = 0;
 338         }
 339
 340         for (timer = ctx->timer_head;
 341              timer;
 342              timer = timer_next)
 343         {
 344                 glDeleteQueries (1, &timer->id);
 345                 timer_next = timer->next;
 346                 free (timer);
 347         }
 348         ctx->timer_head = NULL;
 349         ctx->timer_tail = NULL;
 350
 351         if (ctx->monitor_begun_id) {
 352                 glEndPerfMonitorAMD (ctx->monitor_begun_id);
 353                 glDeletePerfMonitorsAMD (1, &ctx->monitor_begun_id);
 354                 ctx->monitor_begun_id = 0;
 355         }
 356
 357         for (monitor = ctx->monitor_head;
 358              monitor;
 359              monitor = monitor_next)
 360         {
 361                 glDeletePerfMonitorsAMD (1, &monitor->id);
 362                 monitor_next = monitor->next;
 363                 free (monitor);
 364         }
 365         ctx->monitor_head = NULL;
 366         ctx->monitor_tail = NULL;
 367
 368         current_context.monitors_in_flight = 0;
 369
 370         for (i = 0; i < info->num_groups; i++)
 371                 metrics_group_info_fini (&info->groups[i]);
 372
 373         free (info->groups);
 374         info->groups = NULL;
 375
 376         for (i = 0; i < info->num_shader_stages; i++)
 377                 free (info->stages[i].name);
 378
 379         free (info->stages);
 380         info->stages = NULL;
 381
 382         info->initialized = 0;
 383 }
 384
 385 static const char *
 386 metrics_op_string (metrics_op_t op)
 387 {
 388         if (op >= METRICS_OP_SHADER)
 389                 return "Shader program";
 390
 391         switch (op)
 392         {
 393         case METRICS_OP_ACCUM:
 394                 return "glAccum*(+)";
 395         case METRICS_OP_BUFFER_DATA:
 396                 return "glBufferData(+)";
 397         case METRICS_OP_BUFFER_SUB_DATA:
 398                 return "glCopyBufferSubData*";
 399         case METRICS_OP_BITMAP:
 400                 return "glBitmap*";
 401         case METRICS_OP_BLIT_FRAMEBUFFER:
 402                 return "glBlitFramebuffer*";
 403         case METRICS_OP_CLEAR:
 404                 return "glClear(+)";
 405         case METRICS_OP_CLEAR_BUFFER_DATA:
 406                 return "glCearBufferData(+)";
 407         case METRICS_OP_CLEAR_TEX_IMAGE:
 408                 return "glClearTexImage(+)";
 409         case METRICS_OP_COPY_PIXELS:
 410                 return "glCopyPixels";
 411         case METRICS_OP_COPY_TEX_IMAGE:
 412                 return "glCopyTexImage(+)";
 413         case METRICS_OP_DRAW_PIXELS:
 414                 return "glDrawPixels";
 415         case METRICS_OP_GET_TEX_IMAGE:
 416                 return "glGetTexImage(+)";
 417         case METRICS_OP_READ_PIXELS:
 418                 return "glReadPixels*";
 419         case METRICS_OP_TEX_IMAGE:
 420                 return "glTexImage*(+)";
 421         default:
 422                 fprintf (stderr, "fips: Internal error: "
 423                          "Unknown metrics op value: %d\n", op);
 424                 exit (1);
 425         }
 426
 427         return "";
 428 }
 429
 430 void
 431 metrics_counter_start (void)
 432 {
 433         context_t *ctx = &current_context;
 434         unsigned i;
 435
 436         /* Initialize the timer_query and monitor objects */
 437         glGenQueries (1, &ctx->timer_begun_id);
 438
 439         glGenPerfMonitorsAMD (1, &ctx->monitor_begun_id);
 440
 441         for (i = 0; i < ctx->metrics_info.num_groups; i++)
 442         {
 443                 metrics_group_info_t *group;
 444                 int num_counters;
 445
 446                 group = &ctx->metrics_info.groups[i];
 447
 448                 num_counters = group->num_counters;
 449                 if (group->max_active_counters < group->num_counters)
 450                 {
 451                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 452                                  group->max_active_counters,
 453                                  group->num_counters, i);
 454                         num_counters = group->max_active_counters;
 455
 456                 }
 457
 458                 glSelectPerfMonitorCountersAMD(ctx->monitor_begun_id,
 459                                                GL_TRUE, group->id,
 460                                                num_counters,
 461                                                group->counter_ids);
 462         }
 463
 464         /* Start the queries */
 465         glBeginQuery (GL_TIME_ELAPSED, ctx->timer_begun_id);
 466
 467         glBeginPerfMonitorAMD (ctx->monitor_begun_id);
 468 }
 469
 470 void
 471 metrics_counter_stop (void)
 472 {
 473         context_t *ctx = &current_context;
 474         timer_query_t *timer;
 475         monitor_t *monitor;
 476
 477         /* Stop the current timer and monitor. */
 478         glEndQuery (GL_TIME_ELAPSED);
 479         glEndPerfMonitorAMD (ctx->monitor_begun_id);
 480
 481         /* Add these IDs to our lists of outstanding queries and
 482          * monitors so the results can be collected later. */
 483         timer = xmalloc (sizeof (timer_query_t));
 484
 485         timer->op = ctx->op;
 486         timer->id = ctx->timer_begun_id;
 487         timer->next = NULL;
 488
 489         if (ctx->timer_tail) {
 490                 ctx->timer_tail->next = timer;
 491                 ctx->timer_tail = timer;
 492         } else {
 493                 ctx->timer_tail = timer;
 494                 ctx->timer_head = timer;
 495         }
 496
 497         /* Create a new performance-monitor query */
 498         monitor = xmalloc (sizeof (monitor_t));
 499
 500         monitor->op = ctx->op;
 501         monitor->id = ctx->monitor_begun_id;
 502         monitor->next = NULL;
 503
 504         if (ctx->monitor_tail) {
 505                 ctx->monitor_tail->next = monitor;
 506                 ctx->monitor_tail = monitor;
 507         } else {
 508                 ctx->monitor_tail = monitor;
 509                 ctx->monitor_head = monitor;
 510         }
 511
 512         ctx->monitors_in_flight++;
 513
 514         /* Avoid being a resource hog and collect outstanding results
 515          * once we have sent off a large number of
 516          * queries. (Presumably, many of the outstanding queries are
 517          * available by now.)
 518          */
 519         if (ctx->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
 520                 metrics_collect_available ();
 521 }
 522
 523 void
 524 metrics_set_current_op (metrics_op_t op)
 525 {
 526         current_context.op = op;
 527 }
 528
 529 metrics_op_t
 530 metrics_get_current_op (void)
 531 {
 532         return current_context.op;
 533 }
 534
 535 static void
 536 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 537 {
 538         metrics_info_t *info = &ctx->metrics_info;
 539         unsigned i, j;
 540
 541         metrics->op = op;
 542         metrics->time_ns = 0.0;
 543
 544         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 545
 546         for (i = 0; i < info->num_groups; i++) {
 547                 metrics->counters[i] = xmalloc (sizeof (double) *
 548                                                 info->groups[i].num_counters);
 549                 for (j = 0; j < info->groups[i].num_counters; j++)
 550                         metrics->counters[i][j] = 0.0;
 551         }
 552 }
 553
 554 static op_metrics_t *
 555 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 556 {
 557         unsigned i;
 558
 559         if (op >= ctx->num_op_metrics)
 560         {
 561                 ctx->op_metrics = realloc (ctx->op_metrics,
 562                                            (op + 1) * sizeof (op_metrics_t));
 563                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 564                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 565
 566                 ctx->num_op_metrics = op + 1;
 567         }
 568
 569         return &ctx->op_metrics[op];
 570 }
 571
 572 static void
 573 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 574 {
 575 #define CONSUME(var)                                                    \
 576         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 577         {                                                               \
 578                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 579                          "parsing results\n");                          \
 580                 break;                                                  \
 581         }                                                               \
 582         (var) = *((typeof(var) *) p);                                   \
 583         p += sizeof(var);
 584
 585         context_t *ctx = &current_context;
 586         metrics_info_t *info = &ctx->metrics_info;
 587         op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
 588         unsigned char *p = (unsigned char *) result;
 589
 590         while (p < ((unsigned char *) result) + size)
 591         {
 592                 GLuint group_id, group_index;
 593                 GLuint counter_id, counter_index;
 594                 metrics_group_info_t *group;
 595                 double value;
 596                 unsigned i;
 597
 598                 CONSUME (group_id);
 599                 CONSUME (counter_id);
 600
 601                 for (i = 0; i < info->num_groups; i++) {
 602                         if (info->groups[i].id == group_id)
 603                                 break;
 604                 }
 605                 group_index = i;
 606                 assert (group_index < info->num_groups);
 607                 group = &info->groups[group_index];
 608
 609                 for (i = 0; i < group->num_counters; i++) {
 610                         if (group->counter_ids[i] == counter_id)
 611                                 break;
 612                 }
 613                 counter_index = i;
 614                 assert (counter_index < group->num_counters);
 615
 616                 switch (group->counter_types[counter_index])
 617                 {
 618                         uint uint_value;
 619                         uint64_t uint64_value;
 620                         float float_value;
 621                 case GL_UNSIGNED_INT:
 622                         CONSUME (uint_value);
 623                         value = uint_value;
 624                         break;
 625                 case GL_UNSIGNED_INT64_AMD:
 626                         CONSUME (uint64_value);
 627                         value = uint64_value;
 628                         break;
 629                 case GL_PERCENTAGE_AMD:
 630                 case GL_FLOAT:
 631                         CONSUME (float_value);
 632                         value = float_value;
 633                         break;
 634                 default:
 635                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 636                                  group->counter_types[counter_index]);
 637                         value = 0.0;
 638                         break;
 639                 }
 640
 641                 metrics->counters[group_index][counter_index] += value;
 642         }
 643 }
 644
 645 static void
 646 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 647 {
 648         op_metrics_t *metrics;
 649
 650         metrics = ctx_get_op_metrics (&current_context, op);
 651
 652         metrics->time_ns += time_ns;
 653 }
 654
 655 typedef struct per_stage_metrics
 656 {
 657         op_metrics_t *metrics;
 658         shader_stage_info_t *stage;
 659         double time_ns;
 660         double active;
 661 } per_stage_metrics_t;
 662
 663 static int
 664 _is_shader_stage_counter (metrics_info_t *info,
 665                           unsigned group_index,
 666                           unsigned counter_index)
 667 {
 668         shader_stage_info_t *stage;
 669         unsigned i;
 670
 671         for (i = 0; i < info->num_shader_stages; i++) {
 672                 stage = &info->stages[i];
 673
 674                 if (stage->active_group_index == group_index &&
 675                     stage->active_counter_index == counter_index)
 676                 {
 677                         return 1;
 678                 }
 679
 680                 if (stage->stall_group_index == group_index &&
 681                     stage->stall_counter_index == counter_index)
 682                 {
 683                         return 1;
 684                 }
 685         }
 686
 687         return 0;
 688 }
 689
 690 static void
 691 print_per_stage_metrics (context_t *ctx,
 692                          per_stage_metrics_t *per_stage,
 693                          double total)
 694 {
 695         metrics_info_t *info = &ctx->metrics_info;
 696         op_metrics_t *metric = per_stage->metrics;
 697         metrics_group_info_t *group;
 698         const char *op_string;
 699         unsigned group_index, counter;
 700         double value;
 701
 702         /* Don't print anything for stages with no alloted time. */
 703         if (per_stage->time_ns == 0.0)
 704                 return;
 705
 706         op_string = metrics_op_string (metric->op);
 707
 708         printf ("%21s", op_string);
 709
 710         if (metric->op >= METRICS_OP_SHADER) {
 711                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 712         } else {
 713                 printf ("    ");
 714
 715         }
 716
 717         if (per_stage->stage)
 718                 printf (" %cS:", per_stage->stage->name[0]);
 719         else
 720                 printf ("   :");
 721
 722         printf ("\t%7.2f ms (%4.1f%%)",
 723                 per_stage->time_ns / 1e6,
 724                 per_stage->time_ns / total * 100);
 725
 726         if (per_stage->active)
 727                 printf (", %4.1f%% active", per_stage->active * 100);
 728
 729         printf ("\n");
 730
 731         /* I'm not seeing a lot of value printing the rest of these
 732          * performance counters by default yet. Use --verbose to get
 733          * them for now. */
 734         if (! verbose)
 735                 return;
 736
 737         printf ("[");
 738         for (group_index = 0; group_index < info->num_groups; group_index++) {
 739                 group = &info->groups[group_index];
 740                 for (counter = 0; counter < group->num_counters; counter++) {
 741
 742                         /* Don't print this counter value if it's a
 743                          * per-stage cycle counter, (which we have
 744                          * already accounted for). */
 745                         if (_is_shader_stage_counter (info, group_index, counter))
 746                                 continue;
 747
 748                         value = metric->counters[group_index][counter];
 749                         if (value == 0.0)
 750                                 continue;
 751                         printf ("%s: %.2f ", group->counter_names[counter],
 752                                 value / 1e6);
 753                 }
 754         }
 755         printf ("]\n");
 756 }
 757
 758 static int
 759 time_compare(const void *in_a, const void *in_b, void *arg unused)
 760 {
 761         const per_stage_metrics_t *a = in_a;
 762         const per_stage_metrics_t *b = in_b;
 763
 764
 765         if (a->time_ns < b->time_ns)
 766                 return -1;
 767         if (a->time_ns > b->time_ns)
 768                 return 1;
 769         return 0;
 770 }
 771
 772 static void
 773 print_program_metrics (void)
 774 {
 775         context_t *ctx = &current_context;
 776         metrics_info_t *info = &ctx->metrics_info;
 777         unsigned num_shader_stages = info->num_shader_stages;
 778         per_stage_metrics_t *sorted, *per_stage;
 779         double total_time, op_cycles;
 780         op_metrics_t *op;
 781         unsigned group_index, counter_index;
 782         unsigned i, j, num_sorted;
 783
 784         /* Make a sorted list of the per-stage operations by time
 785          * used, and figure out the total so we can print percentages.
 786          */
 787         num_sorted = ctx->num_op_metrics * num_shader_stages;
 788
 789         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 790
 791         total_time = 0.0;
 792
 793         for (i = 0; i < ctx->num_op_metrics; i++) {
 794
 795                 op = &ctx->op_metrics[i];
 796
 797                 /* Accumulate total time across all ops. */
 798                 total_time += op->time_ns;
 799
 800                 /* Also, find total cycles in all stages of this op. */
 801                 op_cycles = 0.0;
 802
 803                 for (j = 0; j < num_shader_stages; j++) {
 804                         /* Active cycles */
 805                         group_index = info->stages[j].active_group_index;
 806                         counter_index = info->stages[j].active_counter_index;
 807                         op_cycles += op->counters[group_index][counter_index];
 808
 809                         /* Stall cycles */
 810                         group_index = info->stages[j].stall_group_index;
 811                         counter_index = info->stages[j].stall_counter_index;
 812                         op_cycles += op->counters[group_index][counter_index];
 813                 }
 814
 815                 for (j = 0; j < num_shader_stages; j++) {
 816                         double active_cycles, stall_cycles, stage_cycles;
 817
 818                         /* Active cycles */
 819                         group_index = info->stages[j].active_group_index;
 820                         counter_index = info->stages[j].active_counter_index;
 821                         active_cycles = op->counters[group_index][counter_index];
 822
 823                         /* Stall cycles */
 824                         group_index = info->stages[j].stall_group_index;
 825                         counter_index = info->stages[j].stall_counter_index;
 826                         stall_cycles = op->counters[group_index][counter_index];
 827
 828                         stage_cycles = active_cycles + stall_cycles;
 829
 830                         per_stage = &sorted[i * num_shader_stages + j];
 831                         per_stage->metrics = op;
 832
 833                         if (op_cycles) {
 834                                 per_stage->stage = &info->stages[j];
 835                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 836                         } else {
 837                                 /* If we don't have any per-stage cycle counts
 838                                  * for this operation, then use the first
 839                                  * stage as a placeholder for all the time,
 840                                  * but NULL-ify the stage info so that the
 841                                  * report doesn't lie about this time being
 842                                  * from any particular stage. */
 843                                 per_stage->stage = NULL;
 844                                 if (j == 0) {
 845                                         per_stage->time_ns = op->time_ns;
 846                                 } else {
 847                                         per_stage->time_ns = 0.0;
 848                                 }
 849                         }
 850
 851                         if (stage_cycles) {
 852                                 per_stage->active = active_cycles / stage_cycles;
 853                         } else {
 854                                 per_stage->active = 0.0;
 855                         }
 856                 }
 857         }
 858
 859         qsort_r (sorted, num_sorted, sizeof (*sorted),
 860                  time_compare, ctx->op_metrics);
 861
 862         for (i = 0; i < num_sorted; i++)
 863                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 864
 865         free (sorted);
 866 }
 867
 868 /* Called at program exit.
 869  *
 870  * This is similar to metrics_info_fini, but only frees any used
 871  * memory. Notably, it does not call any OpenGL functions, (since the
 872  * OpenGL context no longer exists at program exit).
 873  */
 874 static void
 875 metrics_exit (void)
 876 {
 877         context_t *ctx = &current_context;
 878         metrics_info_t *info = &ctx->metrics_info;
 879         unsigned i;
 880         timer_query_t *timer, *timer_next;
 881         monitor_t *monitor, *monitor_next;
 882
 883         if (verbose)
 884                 printf ("fips: terminating\n");
 885
 886         if (! info->initialized)
 887                 return;
 888
 889         for (timer = ctx->timer_head;
 890              timer;
 891              timer = timer_next)
 892         {
 893                 timer_next = timer->next;
 894                 free (timer);
 895         }
 896
 897         for (monitor = ctx->monitor_head;
 898              monitor;
 899              monitor = monitor_next)
 900         {
 901                 monitor_next = monitor->next;
 902                 free (monitor);
 903         }
 904
 905         for (i = 0; i < info->num_groups; i++)
 906                 metrics_group_info_fini (&info->groups[i]);
 907
 908         free (info->groups);
 909
 910         for (i = 0; i < info->num_shader_stages; i++)
 911                 free (info->stages[i].name);
 912
 913         free (info->stages);
 914 }
 915
 916 void
 917 metrics_collect_available (void)
 918 {
 919         context_t *ctx = &current_context;
 920
 921         /* Consume all timer queries that are ready. */
 922         timer_query_t *timer = ctx->timer_head;
 923
 924         while (timer) {
 925                 GLuint available, elapsed;
 926
 927                 glGetQueryObjectuiv (timer->id,
 928                                      GL_QUERY_RESULT_AVAILABLE, &available);
 929                 if (! available)
 930                         break;
 931
 932                 glGetQueryObjectuiv (timer->id,
 933                                      GL_QUERY_RESULT, &elapsed);
 934
 935                 accumulate_program_time (timer->op, elapsed);
 936
 937                 ctx->timer_head = timer->next;
 938                 if (ctx->timer_head == NULL)
 939                         ctx->timer_tail = NULL;
 940
 941                 glDeleteQueries (1, &timer->id);
 942
 943                 free (timer);
 944                 timer = ctx->timer_head;
 945         }
 946
 947         /* And similarly for all performance monitors that are ready. */
 948         monitor_t *monitor = ctx->monitor_head;
 949
 950         while (monitor) {
 951                 GLuint available, result_size, *result;
 952                 GLint bytes_written;
 953
 954                 glGetPerfMonitorCounterDataAMD (monitor->id,
 955                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 956                                                 sizeof (available), &available,
 957                                                 NULL);
 958                 if (! available)
 959                         break;
 960
 961                 glGetPerfMonitorCounterDataAMD (monitor->id,
 962                                                 GL_PERFMON_RESULT_SIZE_AMD,
 963                                                 sizeof (result_size),
 964                                                 &result_size, NULL);
 965
 966                 result = xmalloc (result_size);
 967
 968                 glGetPerfMonitorCounterDataAMD (monitor->id,
 969                                                 GL_PERFMON_RESULT_AMD,
 970                                                 result_size, result,
 971                                                 &bytes_written);
 972
 973                 accumulate_program_metrics (monitor->op, result, result_size);
 974
 975                 free (result);
 976
 977                 ctx->monitor_head = monitor->next;
 978                 if (ctx->monitor_head == NULL)
 979                         ctx->monitor_tail = NULL;
 980
 981                 glDeletePerfMonitorsAMD (1, &monitor->id);
 982
 983                 free (monitor);
 984
 985                 ctx->monitors_in_flight--;
 986
 987                 monitor = ctx->monitor_head;
 988         }
 989 }
 990
 991
 992 void
 993 metrics_end_frame (void)
 994 {
 995         static int initialized = 0;
 996         static struct timeval tv_start, tv_now;
 997
 998         if (! initialized) {
 999                 gettimeofday (&tv_start, NULL);
1000                 atexit (metrics_exit);
1001                 if (getenv ("FIPS_VERBOSE"))
1002                         verbose = 1;
1003                 initialized = 1;
1004         }
1005
1006         frames++;
1007
1008         metrics_collect_available ();
1009
1010         if (frames % 15 == 0) {
1011                 double fps;
1012
1013                 gettimeofday (&tv_now, NULL);
1014
1015                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
1016                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
1017
1018                 printf("FPS: %.3f\n", fps);
1019
1020                 print_program_metrics ();
1021         }
1022 }