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 <inttypes.h>
  25
  26 #include <stdio.h>
  27 #include <stdlib.h>
  28 #include <assert.h>
  29 #include <time.h>
  30 #include <sys/time.h>
  31 #include <string.h>
  32 #include <errno.h>
  33
  34 #include "fips-dispatch-gl.h"
  35
  36 #include "metrics.h"
  37 #include "context.h"
  38 #include "metrics-info.h"
  39 #include "xmalloc.h"
  40
  41 int frames;
  42 int verbose;
  43
  44 #define MAX_MONITORS_IN_FLIGHT 1000
  45
  46 /* Timer query */
  47 typedef struct timer_query
  48 {
  49         unsigned id;
  50
  51         metrics_op_t op;
  52         struct timer_query *next;
  53 } timer_query_t;
  54
  55 /* Performance-monitor query */
  56 typedef struct monitor
  57 {
  58         unsigned id;
  59
  60         metrics_op_t op;
  61         struct monitor *next;
  62 } monitor_t;
  63
  64 typedef struct op_metrics
  65 {
  66         /* This happens to also be the index into the
  67          * metrics->op_metrics array currently
  68          */
  69         metrics_op_t op;
  70         double time_ns;
  71
  72         double **counters;
  73 } op_metrics_t;
  74
  75 struct metrics
  76 {
  77         /* Flag to indicate first frame is being processed. */
  78         bool first_frame;
  79
  80         /* Description of all available peformance counters, counter
  81          * groups, their names and IDs, etc. */
  82         metrics_info_t *info;
  83
  84         /* The current operation being measured. */
  85         metrics_op_t op;
  86
  87         /* GL_TIME_ELAPSED query for which glEndQuery has not yet
  88          * been called. */
  89         unsigned timer_begun_id;
  90
  91         /* GL_TIME_ELAPSED queries for which glEndQuery has been
  92          * called, (but results have not yet been queried). */
  93         timer_query_t *timer_head;
  94         timer_query_t *timer_tail;
  95
  96         /* Performance monitor for which glEndPerfMonitorAMD has not
  97          * yet been called. */
  98         unsigned monitor_begun_id;
  99
 100         /* Performance monitors for which glEndPerfMonitorAMD has
 101          * been called, (but results have not yet been queried). */
 102         monitor_t *monitor_head;
 103         monitor_t *monitor_tail;
 104
 105         int monitors_in_flight;
 106
 107         unsigned num_op_metrics;
 108         op_metrics_t *op_metrics;
 109
 110         /* Per-frame time and latency measurement. */
 111         GLint64 swap_begin_timestamp;
 112         unsigned swap_end_timestamp_id;
 113         GLint64 previous_swap_end_timestamp;
 114
 115         struct timespec previous_cpu_time_ts;
 116         struct timespec cpu_time_ts;
 117 };
 118
 119 /* Return (a - b) while accounting for the fact that between a and b,
 120  * the counter may have wrapped around (with a maximum number of bits
 121  * in the counter of GL_QUERY_COUNTER_BITS). */
 122 static uint64_t
 123 subtract_timestamp (uint64_t a, uint64_t b)
 124 {
 125         GLint timestamp_bits;
 126
 127         glGetQueryiv (GL_TIMESTAMP, GL_QUERY_COUNTER_BITS, &timestamp_bits);
 128
 129         /* Note: "<<" isn't defined for the number of bits in the
 130          * type, and the compiler will punish you.
 131          */
 132         if (timestamp_bits == 64)
 133                 return a - b;
 134
 135         if (b > a)
 136                 return (1ull << timestamp_bits) + a - b;
 137         else
 138                 return a - b;
 139 }
 140
 141 metrics_t *
 142 metrics_create (metrics_info_t *info)
 143 {
 144         metrics_t *metrics;
 145
 146         metrics = xmalloc (sizeof (metrics_t));
 147
 148         metrics->info = info;
 149
 150         metrics->first_frame = true;
 151
 152         metrics->op = 0;
 153
 154         metrics->timer_begun_id = 0;
 155
 156         metrics->timer_head = NULL;
 157         metrics->timer_tail = NULL;
 158
 159         metrics->monitor_begun_id = 0;
 160
 161         metrics->monitor_head = NULL;
 162         metrics->monitor_tail = NULL;
 163
 164         metrics->monitors_in_flight = 0;
 165
 166         metrics->num_op_metrics = 0;
 167         metrics->op_metrics = NULL;
 168
 169         glGenQueries (1, &metrics->swap_end_timestamp_id);
 170
 171         /* Get the first frame timestamp started immediately. */
 172         glQueryCounter (metrics->swap_end_timestamp_id, GL_TIMESTAMP);
 173
 174         /* As well as the initial CPU time. */
 175         clock_gettime (CLOCK_PROCESS_CPUTIME_ID, &metrics->cpu_time_ts);
 176
 177         return metrics;
 178 }
 179
 180 void
 181 metrics_fini (metrics_t *metrics)
 182 {
 183         timer_query_t *timer, *timer_next;
 184         monitor_t *monitor, *monitor_next;
 185
 186         /* Discard and cleanup any outstanding queries. */
 187         if (metrics->timer_begun_id) {
 188                 glEndQuery (GL_TIME_ELAPSED);
 189                 glDeleteQueries (1, &metrics->timer_begun_id);
 190                 metrics->timer_begun_id = 0;
 191         }
 192
 193         for (timer = metrics->timer_head;
 194              timer;
 195              timer = timer_next)
 196         {
 197                 glDeleteQueries (1, &timer->id);
 198                 timer_next = timer->next;
 199                 free (timer);
 200         }
 201         metrics->timer_head = NULL;
 202         metrics->timer_tail = NULL;
 203
 204         if (metrics->info->have_perfmon) {
 205
 206                 if (metrics->monitor_begun_id) {
 207                         glEndPerfMonitorAMD (metrics->monitor_begun_id);
 208                         glDeletePerfMonitorsAMD (1, &metrics->monitor_begun_id);
 209                         metrics->monitor_begun_id = 0;
 210                 }
 211
 212                 for (monitor = metrics->monitor_head;
 213                      monitor;
 214                      monitor = monitor_next)
 215                 {
 216                         glDeletePerfMonitorsAMD (1, &monitor->id);
 217                         monitor_next = monitor->next;
 218                         free (monitor);
 219                 }
 220                 metrics->monitor_head = NULL;
 221                 metrics->monitor_tail = NULL;
 222
 223         }
 224
 225         metrics->monitors_in_flight = 0;
 226 }
 227
 228 void
 229 metrics_destroy (metrics_t *metrics)
 230 {
 231         metrics_fini (metrics);
 232
 233         free (metrics);
 234 }
 235
 236 static const char *
 237 metrics_op_string (metrics_op_t op)
 238 {
 239         if (op >= METRICS_OP_SHADER)
 240                 return "Shader program";
 241
 242         switch (op)
 243         {
 244         case METRICS_OP_ACCUM:
 245                 return "glAccum*(+)";
 246         case METRICS_OP_BUFFER_DATA:
 247                 return "glBufferData(+)";
 248         case METRICS_OP_BUFFER_SUB_DATA:
 249                 return "glCopyBufferSubData*";
 250         case METRICS_OP_BITMAP:
 251                 return "glBitmap*";
 252         case METRICS_OP_BLIT_FRAMEBUFFER:
 253                 return "glBlitFramebuffer*";
 254         case METRICS_OP_CLEAR:
 255                 return "glClear(+)";
 256         case METRICS_OP_CLEAR_BUFFER_DATA:
 257                 return "glCearBufferData(+)";
 258         case METRICS_OP_CLEAR_TEX_IMAGE:
 259                 return "glClearTexImage(+)";
 260         case METRICS_OP_COPY_PIXELS:
 261                 return "glCopyPixels";
 262         case METRICS_OP_COPY_TEX_IMAGE:
 263                 return "glCopyTexImage(+)";
 264         case METRICS_OP_DRAW_PIXELS:
 265                 return "glDrawPixels";
 266         case METRICS_OP_GET_TEX_IMAGE:
 267                 return "glGetTexImage(+)";
 268         case METRICS_OP_READ_PIXELS:
 269                 return "glReadPixels*";
 270         case METRICS_OP_TEX_IMAGE:
 271                 return "glTexImage*(+)";
 272         default:
 273                 fprintf (stderr, "fips: Internal error: "
 274                          "Unknown metrics op value: %d\n", op);
 275                 exit (1);
 276         }
 277
 278         return "";
 279 }
 280
 281 void
 282 metrics_counter_start (metrics_t *metrics)
 283 {
 284         unsigned i;
 285
 286         /* Initialize the timer_query object. */
 287         glGenQueries (1, &metrics->timer_begun_id);
 288
 289         /* Most everything else in this function is
 290          * performance-monitor related. If we don't have that
 291          * extension, just start the timer query and be done. */
 292         if (! metrics->info->have_perfmon) {
 293                 glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
 294                 return;
 295         }
 296
 297         /* Initialize the performance-monitor object */
 298         glGenPerfMonitorsAMD (1, &metrics->monitor_begun_id);
 299
 300         for (i = 0; i < metrics->info->num_groups; i++)
 301         {
 302                 metrics_group_info_t *group;
 303                 int num_counters;
 304
 305                 group = &metrics->info->groups[i];
 306
 307                 num_counters = group->num_counters;
 308                 if (group->max_active_counters < group->num_counters)
 309                 {
 310                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 311                                  group->max_active_counters,
 312                                  group->num_counters, i);
 313                         num_counters = group->max_active_counters;
 314
 315                 }
 316
 317                 glSelectPerfMonitorCountersAMD(metrics->monitor_begun_id,
 318                                                GL_TRUE, group->id,
 319                                                num_counters,
 320                                                group->counter_ids);
 321         }
 322
 323         /* Start the queries */
 324         glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
 325
 326         glBeginPerfMonitorAMD (metrics->monitor_begun_id);
 327 }
 328
 329 void
 330 metrics_counter_stop (metrics_t *metrics)
 331 {
 332         timer_query_t *timer;
 333         monitor_t *monitor;
 334
 335         /* Stop the current timer and monitor. */
 336         glEndQuery (GL_TIME_ELAPSED);
 337
 338         if (metrics->info->have_perfmon)
 339                 glEndPerfMonitorAMD (metrics->monitor_begun_id);
 340
 341         /* Add these IDs to our lists of outstanding queries and
 342          * monitors so the results can be collected later. */
 343         timer = xmalloc (sizeof (timer_query_t));
 344
 345         timer->op = metrics->op;
 346         timer->id = metrics->timer_begun_id;
 347         timer->next = NULL;
 348
 349         if (metrics->timer_tail) {
 350                 metrics->timer_tail->next = timer;
 351                 metrics->timer_tail = timer;
 352         } else {
 353                 metrics->timer_tail = timer;
 354                 metrics->timer_head = timer;
 355         }
 356
 357         if (metrics->info->have_perfmon) {
 358                 /* Create a new performance-monitor query */
 359                 monitor = xmalloc (sizeof (monitor_t));
 360
 361                 monitor->op = metrics->op;
 362                 monitor->id = metrics->monitor_begun_id;
 363                 monitor->next = NULL;
 364
 365                 if (metrics->monitor_tail) {
 366                         metrics->monitor_tail->next = monitor;
 367                         metrics->monitor_tail = monitor;
 368                 } else {
 369                         metrics->monitor_tail = monitor;
 370                         metrics->monitor_head = monitor;
 371                 }
 372         }
 373
 374         metrics->monitors_in_flight++;
 375
 376         /* Avoid being a resource hog and collect outstanding results
 377          * once we have sent off a large number of
 378          * queries. (Presumably, many of the outstanding queries are
 379          * available by now.)
 380          */
 381         if (metrics->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
 382                 metrics_collect_available (metrics);
 383 }
 384
 385 void
 386 metrics_set_current_op (metrics_t *metrics, metrics_op_t op)
 387 {
 388         metrics->op = op;
 389 }
 390
 391 metrics_op_t
 392 metrics_get_current_op (metrics_t *metrics)
 393 {
 394         return metrics->op;
 395 }
 396
 397 static void
 398 op_metrics_init (metrics_info_t *info, op_metrics_t *metrics, metrics_op_t op)
 399 {
 400         unsigned i, j;
 401
 402         metrics->op = op;
 403         metrics->time_ns = 0.0;
 404
 405         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 406
 407         for (i = 0; i < info->num_groups; i++) {
 408                 metrics->counters[i] = xmalloc (sizeof (double) *
 409                                                 info->groups[i].num_counters);
 410                 for (j = 0; j < info->groups[i].num_counters; j++)
 411                         metrics->counters[i][j] = 0.0;
 412         }
 413 }
 414
 415 static op_metrics_t *
 416 _get_op_metrics (metrics_t *metrics, metrics_op_t op)
 417 {
 418         unsigned i;
 419
 420         if (op >= metrics->num_op_metrics)
 421         {
 422                 metrics->op_metrics = realloc (metrics->op_metrics,
 423                                                (op + 1) * sizeof (op_metrics_t));
 424                 for (i = metrics->num_op_metrics; i < op + 1; i++)
 425                         op_metrics_init (metrics->info, &metrics->op_metrics[i], i);
 426
 427                 metrics->num_op_metrics = op + 1;
 428         }
 429
 430         return &metrics->op_metrics[op];
 431 }
 432
 433 static void
 434 accumulate_program_metrics (metrics_t *metrics, metrics_op_t op,
 435                             GLuint *result, GLuint size)
 436 {
 437 #define CONSUME(var)                                                    \
 438         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 439         {                                                               \
 440                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 441                          "parsing results\n");                          \
 442                 value = 0.0;                                            \
 443                 break;                                                  \
 444         }                                                               \
 445         (var) = *((typeof(var) *) p);                                   \
 446         p += sizeof(var);
 447
 448         metrics_info_t *info = metrics->info;
 449         op_metrics_t *op_metrics = _get_op_metrics (metrics, op);
 450         unsigned char *p = (unsigned char *) result;
 451
 452         while (p < ((unsigned char *) result) + size)
 453         {
 454                 GLuint group_id, group_index;
 455                 GLuint counter_id, counter_index;
 456                 metrics_group_info_t *group;
 457                 double value;
 458                 unsigned i;
 459
 460                 CONSUME (group_id);
 461                 CONSUME (counter_id);
 462
 463                 for (i = 0; i < info->num_groups; i++) {
 464                         if (info->groups[i].id == group_id)
 465                                 break;
 466                 }
 467                 group_index = i;
 468                 assert (group_index < info->num_groups);
 469                 group = &info->groups[group_index];
 470
 471                 for (i = 0; i < group->num_counters; i++) {
 472                         if (group->counter_ids[i] == counter_id)
 473                                 break;
 474                 }
 475                 counter_index = i;
 476                 assert (counter_index < group->num_counters);
 477
 478                 switch (group->counter_types[counter_index])
 479                 {
 480                         uint uint_value;
 481                         uint64_t uint64_value;
 482                         float float_value;
 483                 case GL_UNSIGNED_INT:
 484                         CONSUME (uint_value);
 485                         value = uint_value;
 486                         break;
 487                 case GL_UNSIGNED_INT64_AMD:
 488                         CONSUME (uint64_value);
 489                         value = uint64_value;
 490                         break;
 491                 case GL_PERCENTAGE_AMD:
 492                 case GL_FLOAT:
 493                         CONSUME (float_value);
 494                         value = float_value;
 495                         break;
 496                 default:
 497                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 498                                  group->counter_types[counter_index]);
 499                         value = 0.0;
 500                         break;
 501                 }
 502
 503                 op_metrics->counters[group_index][counter_index] += value;
 504         }
 505 }
 506
 507 static void
 508 accumulate_program_time (metrics_t *metrics, metrics_op_t op, unsigned time_ns)
 509 {
 510         op_metrics_t *op_metrics;
 511
 512         op_metrics = _get_op_metrics (metrics, op);
 513
 514         op_metrics->time_ns += time_ns;
 515 }
 516
 517 typedef struct per_stage_metrics
 518 {
 519         op_metrics_t *metrics;
 520         shader_stage_info_t *stage;
 521         double time_ns;
 522         double active;
 523 } per_stage_metrics_t;
 524
 525 static int
 526 _is_shader_stage_counter (metrics_info_t *info,
 527                           unsigned group_index,
 528                           unsigned counter_index)
 529 {
 530         shader_stage_info_t *stage;
 531         unsigned i;
 532
 533         for (i = 0; i < info->num_shader_stages; i++) {
 534                 stage = &info->stages[i];
 535
 536                 if (stage->active_group_index == group_index &&
 537                     stage->active_counter_index == counter_index)
 538                 {
 539                         return 1;
 540                 }
 541
 542                 if (stage->stall_group_index == group_index &&
 543                     stage->stall_counter_index == counter_index)
 544                 {
 545                         return 1;
 546                 }
 547         }
 548
 549         return 0;
 550 }
 551
 552 static void
 553 print_per_stage_metrics (metrics_t *metrics,
 554                          per_stage_metrics_t *per_stage,
 555                          double total)
 556 {
 557         metrics_info_t *info = metrics->info;
 558         op_metrics_t *op_metrics = per_stage->metrics;
 559         metrics_group_info_t *group;
 560         const char *op_string;
 561         unsigned group_index, counter;
 562         double value;
 563
 564         /* Don't print anything for stages with no alloted time. */
 565         if (per_stage->time_ns == 0.0)
 566                 return;
 567
 568         op_string = metrics_op_string (op_metrics->op);
 569
 570         printf ("%21s", op_string);
 571
 572         if (op_metrics->op >= METRICS_OP_SHADER) {
 573                 printf (" %3d", op_metrics->op - METRICS_OP_SHADER);
 574         } else {
 575                 printf ("    ");
 576
 577         }
 578
 579         if (per_stage->stage)
 580                 printf (" %cS:", per_stage->stage->name[0]);
 581         else
 582                 printf ("   :");
 583
 584         printf ("\t%7.2f ms (%4.1f%%)",
 585                 per_stage->time_ns / 1e6,
 586                 per_stage->time_ns / total * 100);
 587
 588         if (per_stage->active)
 589                 printf (", %4.1f%% active", per_stage->active * 100);
 590
 591         printf ("\n");
 592
 593         /* I'm not seeing a lot of value printing the rest of these
 594          * performance counters by default yet. Use --verbose to get
 595          * them for now. */
 596         if (! verbose)
 597                 return;
 598
 599         printf ("[");
 600         for (group_index = 0; group_index < info->num_groups; group_index++) {
 601                 group = &info->groups[group_index];
 602                 for (counter = 0; counter < group->num_counters; counter++) {
 603
 604                         /* Don't print this counter value if it's a
 605                          * per-stage cycle counter, (which we have
 606                          * already accounted for). */
 607                         if (_is_shader_stage_counter (info, group_index, counter))
 608                                 continue;
 609
 610                         value = op_metrics->counters[group_index][counter];
 611                         if (value == 0.0)
 612                                 continue;
 613                         printf ("%s: %.2f ", group->counter_names[counter],
 614                                 value / 1e6);
 615                 }
 616         }
 617         printf ("]\n");
 618 }
 619
 620 static int
 621 time_compare(const void *in_a, const void *in_b, void *arg unused)
 622 {
 623         const per_stage_metrics_t *a = in_a;
 624         const per_stage_metrics_t *b = in_b;
 625
 626
 627         if (a->time_ns < b->time_ns)
 628                 return -1;
 629         if (a->time_ns > b->time_ns)
 630                 return 1;
 631         return 0;
 632 }
 633
 634 static void
 635 print_program_metrics (metrics_t *metrics)
 636 {
 637         metrics_info_t *info = metrics->info;
 638         unsigned num_shader_stages = info->num_shader_stages;
 639         per_stage_metrics_t *sorted, *per_stage;
 640         double total_time, op_cycles;
 641         op_metrics_t *op;
 642         unsigned group_index, counter_index;
 643         unsigned i, j, num_sorted;
 644
 645         if (! info->have_perfmon &&
 646             ! info->printed_missing_perfmon_warning)
 647         {
 648                 fprintf (stderr,
 649                          "Warning: The GL_AMD_performance_monitor extensions is not available.\n"
 650                          "         Some fips features, (such as per-shader-stage timings),\n"
 651                          "         will not be available.\n");
 652
 653                 info->printed_missing_perfmon_warning = true;
 654         }
 655
 656         /* Make a sorted list of the per-stage operations by time
 657          * used, and figure out the total so we can print percentages.
 658          */
 659         if (num_shader_stages)
 660                 num_sorted = metrics->num_op_metrics * num_shader_stages;
 661         else
 662                 num_sorted = metrics->num_op_metrics;
 663
 664         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 665
 666         total_time = 0.0;
 667
 668         for (i = 0; i < metrics->num_op_metrics; i++) {
 669
 670                 op = &metrics->op_metrics[i];
 671
 672                 /* Accumulate total time across all ops. */
 673                 total_time += op->time_ns;
 674
 675                 /* Also, find total cycles in all stages of this op. */
 676                 op_cycles = 0.0;
 677
 678                 if (num_shader_stages == 0) {
 679                         per_stage = &sorted[i];
 680                         per_stage->metrics = op;
 681                         per_stage->stage = NULL;
 682                         per_stage->time_ns = op->time_ns;
 683                         per_stage->active = 0.0;
 684                 }
 685
 686                 for (j = 0; j < num_shader_stages; j++) {
 687                         /* Active cycles */
 688                         group_index = info->stages[j].active_group_index;
 689                         counter_index = info->stages[j].active_counter_index;
 690                         op_cycles += op->counters[group_index][counter_index];
 691
 692                         /* Stall cycles */
 693                         group_index = info->stages[j].stall_group_index;
 694                         counter_index = info->stages[j].stall_counter_index;
 695                         op_cycles += op->counters[group_index][counter_index];
 696                 }
 697
 698                 for (j = 0; j < num_shader_stages; j++) {
 699                         double active_cycles, stall_cycles, stage_cycles;
 700
 701                         /* Active cycles */
 702                         group_index = info->stages[j].active_group_index;
 703                         counter_index = info->stages[j].active_counter_index;
 704                         active_cycles = op->counters[group_index][counter_index];
 705
 706                         /* Stall cycles */
 707                         group_index = info->stages[j].stall_group_index;
 708                         counter_index = info->stages[j].stall_counter_index;
 709                         stall_cycles = op->counters[group_index][counter_index];
 710
 711                         stage_cycles = active_cycles + stall_cycles;
 712
 713                         per_stage = &sorted[i * num_shader_stages + j];
 714                         per_stage->metrics = op;
 715
 716                         if (op_cycles) {
 717                                 per_stage->stage = &info->stages[j];
 718                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 719                         } else {
 720                                 /* If we don't have any per-stage cycle counts
 721                                  * for this operation, then use the first
 722                                  * stage as a placeholder for all the time,
 723                                  * but NULL-ify the stage info so that the
 724                                  * report doesn't lie about this time being
 725                                  * from any particular stage. */
 726                                 per_stage->stage = NULL;
 727                                 if (j == 0) {
 728                                         per_stage->time_ns = op->time_ns;
 729                                 } else {
 730                                         per_stage->time_ns = 0.0;
 731                                 }
 732                         }
 733
 734                         if (stage_cycles) {
 735                                 per_stage->active = active_cycles / stage_cycles;
 736                         } else {
 737                                 per_stage->active = 0.0;
 738                         }
 739                 }
 740         }
 741
 742         qsort_r (sorted, num_sorted, sizeof (*sorted),
 743                  time_compare, metrics->op_metrics);
 744
 745         for (i = 0; i < num_sorted; i++)
 746                 print_per_stage_metrics (metrics, &sorted[i], total_time);
 747
 748         free (sorted);
 749 }
 750
 751 void
 752 metrics_collect_available (metrics_t *metrics)
 753 {
 754         /* Consume all timer queries that are ready. */
 755         timer_query_t *timer = metrics->timer_head;
 756
 757         while (timer) {
 758                 GLuint available, elapsed;
 759
 760                 glGetQueryObjectuiv (timer->id,
 761                                      GL_QUERY_RESULT_AVAILABLE, &available);
 762                 if (! available)
 763                         break;
 764
 765                 glGetQueryObjectuiv (timer->id,
 766                                      GL_QUERY_RESULT, &elapsed);
 767
 768                 accumulate_program_time (metrics, timer->op, elapsed);
 769
 770                 metrics->timer_head = timer->next;
 771                 if (metrics->timer_head == NULL)
 772                         metrics->timer_tail = NULL;
 773
 774                 glDeleteQueries (1, &timer->id);
 775
 776                 free (timer);
 777                 timer = metrics->timer_head;
 778         }
 779
 780         if (! metrics->info->have_perfmon)
 781                 return;
 782
 783         /* And similarly for all performance monitors that are ready. */
 784         monitor_t *monitor = metrics->monitor_head;
 785
 786         while (monitor) {
 787                 GLuint available, result_size, *result;
 788                 GLint bytes_written;
 789
 790                 glGetPerfMonitorCounterDataAMD (monitor->id,
 791                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 792                                                 sizeof (available), &available,
 793                                                 NULL);
 794                 if (! available)
 795                         break;
 796
 797                 glGetPerfMonitorCounterDataAMD (monitor->id,
 798                                                 GL_PERFMON_RESULT_SIZE_AMD,
 799                                                 sizeof (result_size),
 800                                                 &result_size, NULL);
 801
 802                 result = xmalloc (result_size);
 803
 804                 glGetPerfMonitorCounterDataAMD (monitor->id,
 805                                                 GL_PERFMON_RESULT_AMD,
 806                                                 result_size, result,
 807                                                 &bytes_written);
 808
 809                 accumulate_program_metrics (metrics, monitor->op, result, result_size);
 810
 811                 free (result);
 812
 813                 metrics->monitor_head = monitor->next;
 814                 if (metrics->monitor_head == NULL)
 815                         metrics->monitor_tail = NULL;
 816
 817                 glDeletePerfMonitorsAMD (1, &monitor->id);
 818
 819                 free (monitor);
 820
 821                 metrics->monitors_in_flight--;
 822
 823                 monitor = metrics->monitor_head;
 824         }
 825 }
 826
 827 /* Subtract timespec values: Return (a - b) in seconds as a double. */
 828 #define SUBTRACT_TIMESPEC(a, b) (double) (a.tv_sec - b.tv_sec + \
 829                                           (a.tv_nsec - b.tv_nsec) / 1e9)
 830
 831 void
 832 metrics_end_frame_pre_swap (metrics_t *metrics)
 833 {
 834         GLuint64 swap_end_timestamp, frame_time_ns, latency_ns;
 835         double cpu_time;
 836
 837         /* Don't leave any counters running over the end_frame work we
 838          * do here. The counters will be started again at the end of
 839          * metrics_end_frame_post_swap.  */
 840
 841         metrics_counter_stop (metrics);
 842
 843         /* Now that an entire frame's worth of content has gone by, we
 844          * can be sure that the timer quiery for the previous frame's
 845          * swap is available.
 846          *
 847          * Note: The only case in which this query isn't immediately
 848          * available would be if there was effectively nothing in this
 849          * frame. In that case, triggering a block on this query
 850          * result is not a concern (since an empty frame is not
 851          * interesting, and all of our counters are stopped anyway).
 852          *
 853          * Accepting this one block is much better than adding a
 854          * linked list for these results that would reaslisticly never
 855          * have more than one entry anyway.
 856          */
 857
 858         glGetQueryObjectui64v (metrics->swap_end_timestamp_id,
 859                                GL_QUERY_RESULT, &swap_end_timestamp);
 860
 861         if (metrics->first_frame) {
 862                 /* Print header */
 863                 printf ("# frame: Frame_Number Frame_Time_milliseconds Frame_latency_milliseconds CPU_load GPU_load\n");
 864
 865                 metrics->first_frame = false;
 866         } else {
 867                 /* Subtract previous frame's times to get frame times. */
 868                 frame_time_ns = subtract_timestamp (swap_end_timestamp,
 869                                                     metrics->previous_swap_end_timestamp);
 870
 871                 latency_ns = subtract_timestamp (swap_end_timestamp,
 872                                                  metrics->swap_begin_timestamp);
 873
 874                 cpu_time = SUBTRACT_TIMESPEC (metrics->cpu_time_ts,
 875                                               metrics->previous_cpu_time_ts);
 876
 877                 /* We've waited one frame to ensure we have a timestamp
 878                  * result. So the time we've actually measured here is
 879                  * for the previous frame. */
 880                 printf ("frame: %d %g %g %g %g\n",
 881                         frames - 1,
 882                         (double) frame_time_ns / 1e6,
 883                         (double) latency_ns / 1e6,
 884                         cpu_time / (frame_time_ns / 1e9),
 885                         0.0);
 886         }
 887
 888         metrics->previous_swap_end_timestamp = swap_end_timestamp;
 889         metrics->previous_cpu_time_ts = metrics->cpu_time_ts;
 890
 891         /* Before the actual swap call, we get the current timestamp
 892          * value. This is a synchronous get so we can use this as the
 893          * baseline for a frame latency measurment. */
 894
 895         glGetInteger64v (GL_TIMESTAMP, &metrics->swap_begin_timestamp);
 896 }
 897
 898 void
 899 metrics_end_frame_post_swap (metrics_t *metrics)
 900 {
 901         static int initialized = 0;
 902         static struct timeval tv_start, tv_now;
 903         int err;
 904
 905         /* Now that the swap command has been queued, we issue an
 906          * asynchronous query of the timestamp value. Comparing this
 907          * to the synchronous get we just sent in
 908          * metrics_end_frame_pre_swap allows us to measure the
 909          * per-frame swap latency. */
 910
 911         glQueryCounter (metrics->swap_end_timestamp_id, GL_TIMESTAMP);
 912
 913         /* Query the total amount of CPU time spent by this process. */
 914         err = clock_gettime (CLOCK_PROCESS_CPUTIME_ID, &metrics->cpu_time_ts);
 915         if (err == -1 ) {
 916                 fprintf (stderr, "Failed to query CPU time with clock_gettime: %s\n",
 917                          strerror (errno));
 918                 exit (1);
 919         }
 920
 921         if (! initialized) {
 922                 gettimeofday (&tv_start, NULL);
 923                 if (getenv ("FIPS_VERBOSE"))
 924                         verbose = 1;
 925                 initialized = 1;
 926         }
 927
 928         metrics_collect_available (metrics);
 929
 930         frames++;
 931
 932         if (frames % 15 == 0) {
 933                 double fps;
 934
 935                 gettimeofday (&tv_now, NULL);
 936
 937                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 938                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 939
 940                 printf("FPS: %.3f\n", fps);
 941
 942                 print_program_metrics (metrics);
 943         }
 944
 945         /* Start the counter up again now that we're done. */
 946         metrics_counter_start (metrics);
 947 }