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                 ctx->timer_begun_id = 0;
 336         }
 337
 338         for (timer = ctx->timer_head;
 339              timer;
 340              timer = timer_next)
 341         {
 342                 timer_next = timer->next;
 343                 free (timer);
 344         }
 345         ctx->timer_head = NULL;
 346         ctx->timer_tail = NULL;
 347
 348         if (ctx->monitor_begun_id) {
 349                 ctx->monitor_begun_id = 0;
 350         }
 351
 352         for (monitor = ctx->monitor_head;
 353              monitor;
 354              monitor = monitor_next)
 355         {
 356                 monitor_next = monitor->next;
 357                 free (monitor);
 358         }
 359         ctx->monitor_head = NULL;
 360         ctx->monitor_tail = NULL;
 361
 362         current_context.monitors_in_flight = 0;
 363
 364         for (i = 0; i < info->num_groups; i++)
 365                 metrics_group_info_fini (&info->groups[i]);
 366
 367         free (info->groups);
 368         info->groups = NULL;
 369
 370         for (i = 0; i < info->num_shader_stages; i++)
 371                 free (info->stages[i].name);
 372
 373         free (info->stages);
 374         info->stages = NULL;
 375
 376         info->initialized = 0;
 377 }
 378
 379 static const char *
 380 metrics_op_string (metrics_op_t op)
 381 {
 382         if (op >= METRICS_OP_SHADER)
 383                 return "Shader program";
 384
 385         switch (op)
 386         {
 387         case METRICS_OP_ACCUM:
 388                 return "glAccum*(+)";
 389         case METRICS_OP_BUFFER_DATA:
 390                 return "glBufferData(+)";
 391         case METRICS_OP_BUFFER_SUB_DATA:
 392                 return "glCopyBufferSubData*";
 393         case METRICS_OP_BITMAP:
 394                 return "glBitmap*";
 395         case METRICS_OP_BLIT_FRAMEBUFFER:
 396                 return "glBlitFramebuffer*";
 397         case METRICS_OP_CLEAR:
 398                 return "glClear(+)";
 399         case METRICS_OP_CLEAR_BUFFER_DATA:
 400                 return "glCearBufferData(+)";
 401         case METRICS_OP_CLEAR_TEX_IMAGE:
 402                 return "glClearTexImage(+)";
 403         case METRICS_OP_COPY_PIXELS:
 404                 return "glCopyPixels";
 405         case METRICS_OP_COPY_TEX_IMAGE:
 406                 return "glCopyTexImage(+)";
 407         case METRICS_OP_DRAW_PIXELS:
 408                 return "glDrawPixels";
 409         case METRICS_OP_GET_TEX_IMAGE:
 410                 return "glGetTexImage(+)";
 411         case METRICS_OP_READ_PIXELS:
 412                 return "glReadPixels*";
 413         case METRICS_OP_TEX_IMAGE:
 414                 return "glTexImage*(+)";
 415         default:
 416                 fprintf (stderr, "fips: Internal error: "
 417                          "Unknown metrics op value: %d\n", op);
 418                 exit (1);
 419         }
 420
 421         return "";
 422 }
 423
 424 void
 425 metrics_counter_start (void)
 426 {
 427         context_t *ctx = &current_context;
 428         unsigned i;
 429
 430         /* Initialize the timer_query and monitor objects */
 431         glGenQueries (1, &ctx->timer_begun_id);
 432
 433         glGenPerfMonitorsAMD (1, &ctx->monitor_begun_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(ctx->monitor_begun_id,
 453                                                GL_TRUE, group->id,
 454                                                num_counters,
 455                                                group->counter_ids);
 456         }
 457
 458         /* Start the queries */
 459         glBeginQuery (GL_TIME_ELAPSED, ctx->timer_begun_id);
 460
 461         glBeginPerfMonitorAMD (ctx->monitor_begun_id);
 462 }
 463
 464 void
 465 metrics_counter_stop (void)
 466 {
 467         context_t *ctx = &current_context;
 468         timer_query_t *timer;
 469         monitor_t *monitor;
 470
 471         /* Stop the current timer and monitor. */
 472         glEndQuery (GL_TIME_ELAPSED);
 473         glEndPerfMonitorAMD (ctx->monitor_begun_id);
 474
 475         /* Add these IDs to our lists of outstanding queries and
 476          * monitors so the results can be collected later. */
 477         timer = xmalloc (sizeof (timer_query_t));
 478
 479         timer->op = ctx->op;
 480         timer->id = ctx->timer_begun_id;
 481         timer->next = NULL;
 482
 483         if (ctx->timer_tail) {
 484                 ctx->timer_tail->next = timer;
 485                 ctx->timer_tail = timer;
 486         } else {
 487                 ctx->timer_tail = timer;
 488                 ctx->timer_head = timer;
 489         }
 490
 491         /* Create a new performance-monitor query */
 492         monitor = xmalloc (sizeof (monitor_t));
 493
 494         monitor->op = ctx->op;
 495         monitor->id = ctx->monitor_begun_id;
 496         monitor->next = NULL;
 497
 498         if (ctx->monitor_tail) {
 499                 ctx->monitor_tail->next = monitor;
 500                 ctx->monitor_tail = monitor;
 501         } else {
 502                 ctx->monitor_tail = monitor;
 503                 ctx->monitor_head = monitor;
 504         }
 505
 506         ctx->monitors_in_flight++;
 507
 508         /* Avoid being a resource hog and collect outstanding results
 509          * once we have sent off a large number of
 510          * queries. (Presumably, many of the outstanding queries are
 511          * available by now.)
 512          */
 513         if (ctx->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
 514                 metrics_collect_available ();
 515 }
 516
 517 void
 518 metrics_set_current_op (metrics_op_t op)
 519 {
 520         current_context.op = op;
 521 }
 522
 523 metrics_op_t
 524 metrics_get_current_op (void)
 525 {
 526         return current_context.op;
 527 }
 528
 529 static void
 530 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 531 {
 532         metrics_info_t *info = &ctx->metrics_info;
 533         unsigned i, j;
 534
 535         metrics->op = op;
 536         metrics->time_ns = 0.0;
 537
 538         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 539
 540         for (i = 0; i < info->num_groups; i++) {
 541                 metrics->counters[i] = xmalloc (sizeof (double) *
 542                                                 info->groups[i].num_counters);
 543                 for (j = 0; j < info->groups[i].num_counters; j++)
 544                         metrics->counters[i][j] = 0.0;
 545         }
 546 }
 547
 548 static op_metrics_t *
 549 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 550 {
 551         unsigned i;
 552
 553         if (op >= ctx->num_op_metrics)
 554         {
 555                 ctx->op_metrics = realloc (ctx->op_metrics,
 556                                            (op + 1) * sizeof (op_metrics_t));
 557                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 558                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 559
 560                 ctx->num_op_metrics = op + 1;
 561         }
 562
 563         return &ctx->op_metrics[op];
 564 }
 565
 566 static void
 567 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 568 {
 569 #define CONSUME(var)                                                    \
 570         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 571         {                                                               \
 572                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 573                          "parsing results\n");                          \
 574                 break;                                                  \
 575         }                                                               \
 576         (var) = *((typeof(var) *) p);                                   \
 577         p += sizeof(var);
 578
 579         context_t *ctx = &current_context;
 580         metrics_info_t *info = &ctx->metrics_info;
 581         op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
 582         unsigned char *p = (unsigned char *) result;
 583
 584         while (p < ((unsigned char *) result) + size)
 585         {
 586                 GLuint group_id, group_index;
 587                 GLuint counter_id, counter_index;
 588                 metrics_group_info_t *group;
 589                 double value;
 590                 unsigned i;
 591
 592                 CONSUME (group_id);
 593                 CONSUME (counter_id);
 594
 595                 for (i = 0; i < info->num_groups; i++) {
 596                         if (info->groups[i].id == group_id)
 597                                 break;
 598                 }
 599                 group_index = i;
 600                 assert (group_index < info->num_groups);
 601                 group = &info->groups[group_index];
 602
 603                 for (i = 0; i < group->num_counters; i++) {
 604                         if (group->counter_ids[i] == counter_id)
 605                                 break;
 606                 }
 607                 counter_index = i;
 608                 assert (counter_index < group->num_counters);
 609
 610                 switch (group->counter_types[counter_index])
 611                 {
 612                         uint uint_value;
 613                         uint64_t uint64_value;
 614                         float float_value;
 615                 case GL_UNSIGNED_INT:
 616                         CONSUME (uint_value);
 617                         value = uint_value;
 618                         break;
 619                 case GL_UNSIGNED_INT64_AMD:
 620                         CONSUME (uint64_value);
 621                         value = uint64_value;
 622                         break;
 623                 case GL_PERCENTAGE_AMD:
 624                 case GL_FLOAT:
 625                         CONSUME (float_value);
 626                         value = float_value;
 627                         break;
 628                 default:
 629                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 630                                  group->counter_types[counter_index]);
 631                         value = 0.0;
 632                         break;
 633                 }
 634
 635                 metrics->counters[group_index][counter_index] += value;
 636         }
 637 }
 638
 639 static void
 640 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 641 {
 642         op_metrics_t *metrics;
 643
 644         metrics = ctx_get_op_metrics (&current_context, op);
 645
 646         metrics->time_ns += time_ns;
 647 }
 648
 649 typedef struct per_stage_metrics
 650 {
 651         op_metrics_t *metrics;
 652         shader_stage_info_t *stage;
 653         double time_ns;
 654         double active;
 655 } per_stage_metrics_t;
 656
 657 static int
 658 _is_shader_stage_counter (metrics_info_t *info,
 659                           unsigned group_index,
 660                           unsigned counter_index)
 661 {
 662         shader_stage_info_t *stage;
 663         unsigned i;
 664
 665         for (i = 0; i < info->num_shader_stages; i++) {
 666                 stage = &info->stages[i];
 667
 668                 if (stage->active_group_index == group_index &&
 669                     stage->active_counter_index == counter_index)
 670                 {
 671                         return 1;
 672                 }
 673
 674                 if (stage->stall_group_index == group_index &&
 675                     stage->stall_counter_index == counter_index)
 676                 {
 677                         return 1;
 678                 }
 679         }
 680
 681         return 0;
 682 }
 683
 684 static void
 685 print_per_stage_metrics (context_t *ctx,
 686                          per_stage_metrics_t *per_stage,
 687                          double total)
 688 {
 689         metrics_info_t *info = &ctx->metrics_info;
 690         op_metrics_t *metric = per_stage->metrics;
 691         metrics_group_info_t *group;
 692         const char *op_string;
 693         unsigned group_index, counter;
 694         double value;
 695
 696         /* Don't print anything for stages with no alloted time. */
 697         if (per_stage->time_ns == 0.0)
 698                 return;
 699
 700         op_string = metrics_op_string (metric->op);
 701
 702         printf ("%21s", op_string);
 703
 704         if (metric->op >= METRICS_OP_SHADER) {
 705                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 706         } else {
 707                 printf ("    ");
 708
 709         }
 710
 711         if (per_stage->stage)
 712                 printf (" %cS:", per_stage->stage->name[0]);
 713         else
 714                 printf ("   :");
 715
 716         printf ("\t%7.2f ms (%4.1f%%)",
 717                 per_stage->time_ns / 1e6,
 718                 per_stage->time_ns / total * 100);
 719
 720         if (per_stage->active)
 721                 printf (", %4.1f%% active", per_stage->active * 100);
 722
 723         printf ("\n");
 724
 725         /* I'm not seeing a lot of value printing the rest of these
 726          * performance counters by default yet. Use --verbose to get
 727          * them for now. */
 728         if (! verbose)
 729                 return;
 730
 731         printf ("[");
 732         for (group_index = 0; group_index < info->num_groups; group_index++) {
 733                 group = &info->groups[group_index];
 734                 for (counter = 0; counter < group->num_counters; counter++) {
 735
 736                         /* Don't print this counter value if it's a
 737                          * per-stage cycle counter, (which we have
 738                          * already accounted for). */
 739                         if (_is_shader_stage_counter (info, group_index, counter))
 740                                 continue;
 741
 742                         value = metric->counters[group_index][counter];
 743                         if (value == 0.0)
 744                                 continue;
 745                         printf ("%s: %.2f ", group->counter_names[counter],
 746                                 value / 1e6);
 747                 }
 748         }
 749         printf ("]\n");
 750 }
 751
 752 static int
 753 time_compare(const void *in_a, const void *in_b, void *arg unused)
 754 {
 755         const per_stage_metrics_t *a = in_a;
 756         const per_stage_metrics_t *b = in_b;
 757
 758
 759         if (a->time_ns < b->time_ns)
 760                 return -1;
 761         if (a->time_ns > b->time_ns)
 762                 return 1;
 763         return 0;
 764 }
 765
 766 static void
 767 print_program_metrics (void)
 768 {
 769         context_t *ctx = &current_context;
 770         metrics_info_t *info = &ctx->metrics_info;
 771         unsigned num_shader_stages = info->num_shader_stages;
 772         per_stage_metrics_t *sorted, *per_stage;
 773         double total_time, op_cycles;
 774         op_metrics_t *op;
 775         unsigned group_index, counter_index;
 776         unsigned i, j, num_sorted;
 777
 778         /* Make a sorted list of the per-stage operations by time
 779          * used, and figure out the total so we can print percentages.
 780          */
 781         num_sorted = ctx->num_op_metrics * num_shader_stages;
 782
 783         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 784
 785         total_time = 0.0;
 786
 787         for (i = 0; i < ctx->num_op_metrics; i++) {
 788
 789                 op = &ctx->op_metrics[i];
 790
 791                 /* Accumulate total time across all ops. */
 792                 total_time += op->time_ns;
 793
 794                 /* Also, find total cycles in all stages of this op. */
 795                 op_cycles = 0.0;
 796
 797                 for (j = 0; j < num_shader_stages; j++) {
 798                         /* Active cycles */
 799                         group_index = info->stages[j].active_group_index;
 800                         counter_index = info->stages[j].active_counter_index;
 801                         op_cycles += op->counters[group_index][counter_index];
 802
 803                         /* Stall cycles */
 804                         group_index = info->stages[j].stall_group_index;
 805                         counter_index = info->stages[j].stall_counter_index;
 806                         op_cycles += op->counters[group_index][counter_index];
 807                 }
 808
 809                 for (j = 0; j < num_shader_stages; j++) {
 810                         double active_cycles, stall_cycles, stage_cycles;
 811
 812                         /* Active cycles */
 813                         group_index = info->stages[j].active_group_index;
 814                         counter_index = info->stages[j].active_counter_index;
 815                         active_cycles = op->counters[group_index][counter_index];
 816
 817                         /* Stall cycles */
 818                         group_index = info->stages[j].stall_group_index;
 819                         counter_index = info->stages[j].stall_counter_index;
 820                         stall_cycles = op->counters[group_index][counter_index];
 821
 822                         stage_cycles = active_cycles + stall_cycles;
 823
 824                         per_stage = &sorted[i * num_shader_stages + j];
 825                         per_stage->metrics = op;
 826
 827                         if (op_cycles) {
 828                                 per_stage->stage = &info->stages[j];
 829                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 830                         } else {
 831                                 /* If we don't have any per-stage cycle counts
 832                                  * for this operation, then use the first
 833                                  * stage as a placeholder for all the time,
 834                                  * but NULL-ify the stage info so that the
 835                                  * report doesn't lie about this time being
 836                                  * from any particular stage. */
 837                                 per_stage->stage = NULL;
 838                                 if (j == 0) {
 839                                         per_stage->time_ns = op->time_ns;
 840                                 } else {
 841                                         per_stage->time_ns = 0.0;
 842                                 }
 843                         }
 844
 845                         if (stage_cycles) {
 846                                 per_stage->active = active_cycles / stage_cycles;
 847                         } else {
 848                                 per_stage->active = 0.0;
 849                         }
 850                 }
 851         }
 852
 853         qsort_r (sorted, num_sorted, sizeof (*sorted),
 854                  time_compare, ctx->op_metrics);
 855
 856         for (i = 0; i < num_sorted; i++)
 857                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 858
 859         free (sorted);
 860 }
 861
 862 /* Called at program exit */
 863 static void
 864 metrics_exit (void)
 865 {
 866         if (verbose)
 867                 printf ("fips: terminating\n");
 868
 869         metrics_info_fini ();
 870 }
 871
 872 void
 873 metrics_collect_available (void)
 874 {
 875         context_t *ctx = &current_context;
 876
 877         /* Consume all timer queries that are ready. */
 878         timer_query_t *timer = ctx->timer_head;
 879
 880         while (timer) {
 881                 GLuint available, elapsed;
 882
 883                 glGetQueryObjectuiv (timer->id,
 884                                      GL_QUERY_RESULT_AVAILABLE, &available);
 885                 if (! available)
 886                         break;
 887
 888                 glGetQueryObjectuiv (timer->id,
 889                                      GL_QUERY_RESULT, &elapsed);
 890
 891                 accumulate_program_time (timer->op, elapsed);
 892
 893                 ctx->timer_head = timer->next;
 894                 if (ctx->timer_head == NULL)
 895                         ctx->timer_tail = NULL;
 896
 897                 glDeleteQueries (1, &timer->id);
 898
 899                 free (timer);
 900                 timer = ctx->timer_head;
 901         }
 902
 903         /* And similarly for all performance monitors that are ready. */
 904         monitor_t *monitor = ctx->monitor_head;
 905
 906         while (monitor) {
 907                 GLuint available, result_size, *result;
 908                 GLint bytes_written;
 909
 910                 glGetPerfMonitorCounterDataAMD (monitor->id,
 911                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 912                                                 sizeof (available), &available,
 913                                                 NULL);
 914                 if (! available)
 915                         break;
 916
 917                 glGetPerfMonitorCounterDataAMD (monitor->id,
 918                                                 GL_PERFMON_RESULT_SIZE_AMD,
 919                                                 sizeof (result_size),
 920                                                 &result_size, NULL);
 921
 922                 result = xmalloc (result_size);
 923
 924                 glGetPerfMonitorCounterDataAMD (monitor->id,
 925                                                 GL_PERFMON_RESULT_AMD,
 926                                                 result_size, result,
 927                                                 &bytes_written);
 928
 929                 accumulate_program_metrics (monitor->op, result, result_size);
 930
 931                 free (result);
 932
 933                 ctx->monitor_head = monitor->next;
 934                 if (ctx->monitor_head == NULL)
 935                         ctx->monitor_tail = NULL;
 936
 937                 glDeletePerfMonitorsAMD (1, &monitor->id);
 938
 939                 free (monitor);
 940
 941                 ctx->monitors_in_flight--;
 942
 943                 monitor = ctx->monitor_head;
 944         }
 945 }
 946
 947
 948 void
 949 metrics_end_frame (void)
 950 {
 951         static int initialized = 0;
 952         static struct timeval tv_start, tv_now;
 953
 954         if (! initialized) {
 955                 gettimeofday (&tv_start, NULL);
 956                 atexit (metrics_exit);
 957                 if (getenv ("FIPS_VERBOSE"))
 958                         verbose = 1;
 959                 initialized = 1;
 960         }
 961
 962         frames++;
 963
 964         metrics_collect_available ();
 965
 966         if (frames % 15 == 0) {
 967                 double fps;
 968
 969                 gettimeofday (&tv_now, NULL);
 970
 971                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 972                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 973
 974                 printf("FPS: %.3f\n", fps);
 975
 976                 print_program_metrics ();
 977         }
 978 }