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 metrics_info
  79 {
  80         int initialized;
  81
  82         unsigned num_groups;
  83         metrics_group_info_t *groups;
  84 } metrics_info_t;
  85
  86 typedef struct context
  87 {
  88         metrics_info_t metrics_info;
  89
  90         metrics_op_t op;
  91
  92         timer_query_t *timer_head;
  93         timer_query_t *timer_tail;
  94
  95         monitor_t *monitor_head;
  96         monitor_t *monitor_tail;
  97
  98         unsigned num_op_metrics;
  99         op_metrics_t *op_metrics;
 100 } context_t;
 101
 102 /* FIXME: Need a map from integers to context objects and track the
 103  * current context with glXMakeContextCurrent, eglMakeCurrent, etc. */
 104
 105 context_t current_context;
 106
 107 int frames;
 108 int verbose;
 109
 110 static void
 111 metrics_group_info_init (metrics_group_info_t *group, GLuint id)
 112 {
 113         GLsizei length;
 114         unsigned i;
 115
 116         group->id = id;
 117
 118         /* Get name */
 119         glGetPerfMonitorGroupStringAMD (id, 0, &length, NULL);
 120
 121         group->name = xmalloc (length + 1);
 122
 123         glGetPerfMonitorGroupStringAMD (id, length + 1, NULL, group->name);
 124
 125         /* Get number of counters */
 126         group->num_counters = 0;
 127         group->max_active_counters = 0;
 128         glGetPerfMonitorCountersAMD (group->id,
 129                                      (int *) &group->num_counters,
 130                                      (int *) &group->max_active_counters,
 131                                      0, NULL);
 132
 133         /* Get counter numbers */
 134         group->counter_ids = xmalloc (group->num_counters * sizeof (GLuint));
 135
 136         glGetPerfMonitorCountersAMD (group->id, NULL, NULL,
 137                                      group->num_counters,
 138                                      group->counter_ids);
 139
 140         /* Get counter names */
 141         group->counter_names = xmalloc (group->num_counters * sizeof (char *));
 142         group->counter_types = xmalloc (group->num_counters * sizeof (GLuint));
 143
 144         for (i = 0; i < group->num_counters; i++) {
 145                 glGetPerfMonitorCounterInfoAMD (group->id,
 146                                                 group->counter_ids[i],
 147                                                 GL_COUNTER_TYPE_AMD,
 148                                                 &group->counter_types[i]);
 149
 150                 /* We assume that all peformance counters are made
 151                  * available as uint32 values. The code calling
 152                  * CONSUME in accumulate_program_metrics will need to
 153                  * be extended to accomodate other counter values. */
 154                 if (group->counter_types[i] != GL_UNSIGNED_INT) {
 155                         fprintf (stderr, "fips: Internal error: No support for non-uint counter values\n");
 156                         exit (1);
 157                 }
 158
 159                 glGetPerfMonitorCounterStringAMD (group->id,
 160                                                   group->counter_ids[i],
 161                                                   0, &length, NULL);
 162
 163                 group->counter_names[i] = xmalloc (length + 1);
 164
 165                 glGetPerfMonitorCounterStringAMD (group->id,
 166                                                   group->counter_ids[i],
 167                                                   length + 1, NULL,
 168                                                   group->counter_names[i]);
 169         }
 170 }
 171
 172 static void
 173 metrics_group_info_fini (metrics_group_info_t *group)
 174 {
 175         unsigned i;
 176
 177         for (i = 0; i < group->num_counters; i++)
 178                 free (group->counter_names[i]);
 179
 180         free (group->counter_types);
 181         free (group->counter_names);
 182         free (group->counter_ids);
 183
 184         free (group->name);
 185 }
 186
 187 static void
 188 metrics_info_fini (metrics_info_t *info);
 189
 190 void
 191 metrics_info_init (void)
 192 {
 193         unsigned i;
 194         GLuint *group_ids;
 195         metrics_info_t *info = &current_context.metrics_info;
 196
 197         if (info->initialized)
 198                 metrics_info_fini (info);
 199
 200         glGetPerfMonitorGroupsAMD ((int *) &info->num_groups, 0, NULL);
 201
 202         group_ids = xmalloc (info->num_groups * sizeof (GLuint));
 203
 204         glGetPerfMonitorGroupsAMD (NULL, info->num_groups, group_ids);
 205
 206         info->groups = xmalloc (info->num_groups * sizeof (metrics_group_info_t));
 207
 208         for (i = 0; i < info->num_groups; i++)
 209                 metrics_group_info_init (&info->groups[i], group_ids[i]);
 210
 211         free (group_ids);
 212
 213         info->initialized = 1;
 214 }
 215
 216 static void
 217 metrics_info_fini (metrics_info_t *info)
 218 {
 219         unsigned i;
 220
 221         for (i = 0; i < info->num_groups; i++)
 222                 metrics_group_info_fini (&info->groups[i]);
 223
 224         free (info->groups);
 225 }
 226
 227 static const char *
 228 metrics_op_string (metrics_op_t op)
 229 {
 230         if (op >= METRICS_OP_SHADER)
 231                 return "Shader program";
 232
 233         switch (op)
 234         {
 235         case METRICS_OP_ACCUM:
 236                 return "glAccum*(+)";
 237         case METRICS_OP_BUFFER_DATA:
 238                 return "glBufferData(+)";
 239         case METRICS_OP_BUFFER_SUB_DATA:
 240                 return "glCopyBufferSubData*";
 241         case METRICS_OP_BITMAP:
 242                 return "glBitmap*";
 243         case METRICS_OP_BLIT_FRAMEBUFFER:
 244                 return "glBlitFramebuffer*";
 245         case METRICS_OP_CLEAR:
 246                 return "glClear(+)";
 247         case METRICS_OP_CLEAR_BUFFER_DATA:
 248                 return "glCearBufferData(+)";
 249         case METRICS_OP_CLEAR_TEX_IMAGE:
 250                 return "glClearTexImage(+)";
 251         case METRICS_OP_COPY_PIXELS:
 252                 return "glCopyPixels";
 253         case METRICS_OP_COPY_TEX_IMAGE:
 254                 return "glCopyTexImage(+)";
 255         case METRICS_OP_DRAW_PIXELS:
 256                 return "glDrawPixels";
 257         case METRICS_OP_GET_TEX_IMAGE:
 258                 return "glGetTexImage(+)";
 259         case METRICS_OP_READ_PIXELS:
 260                 return "glReadPixels*";
 261         case METRICS_OP_TEX_IMAGE:
 262                 return "glTexImage*(+)";
 263         default:
 264                 fprintf (stderr, "fips: Internal error: "
 265                          "Unknown metrics op value: %d\n", op);
 266                 exit (1);
 267         }
 268
 269         return "";
 270 }
 271
 272 void
 273 metrics_counter_start (void)
 274 {
 275         context_t *ctx = &current_context;
 276         timer_query_t *timer;
 277         monitor_t *monitor;
 278         unsigned i;
 279
 280         /* Create new timer query, add to list */
 281         timer = xmalloc (sizeof (timer_query_t));
 282
 283         timer->op = ctx->op;
 284         timer->next = NULL;
 285
 286         if (ctx->timer_tail) {
 287                 ctx->timer_tail->next = timer;
 288                 ctx->timer_tail = timer;
 289         } else {
 290                 ctx->timer_tail = timer;
 291                 ctx->timer_head = timer;
 292         }
 293
 294         /* Create a new performance-monitor query */
 295         monitor = xmalloc (sizeof (monitor_t));
 296
 297         monitor->op = ctx->op;
 298         monitor->next = NULL;
 299
 300         if (ctx->monitor_tail) {
 301                 ctx->monitor_tail->next = monitor;
 302                 ctx->monitor_tail = monitor;
 303         } else {
 304                 ctx->monitor_tail = monitor;
 305                 ctx->monitor_head = monitor;
 306         }
 307
 308         /* Initialize the timer_query and monitor objects */
 309         glGenQueries (1, &timer->id);
 310
 311         glGenPerfMonitorsAMD (1, &monitor->id);
 312
 313         for (i = 0; i < ctx->metrics_info.num_groups; i++)
 314         {
 315                 metrics_group_info_t *group;
 316                 int num_counters;
 317
 318                 group = &ctx->metrics_info.groups[i];
 319
 320                 num_counters = group->num_counters;
 321                 if (group->max_active_counters < group->num_counters)
 322                 {
 323                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 324                                  group->max_active_counters,
 325                                  group->num_counters, i);
 326                         num_counters = group->max_active_counters;
 327
 328                 }
 329
 330                 glSelectPerfMonitorCountersAMD(monitor->id,
 331                                                GL_TRUE, group->id,
 332                                                num_counters,
 333                                                group->counter_ids);
 334         }
 335
 336         /* Start the queries */
 337         glBeginQuery (GL_TIME_ELAPSED, timer->id);
 338
 339         glBeginPerfMonitorAMD (monitor->id);
 340 }
 341
 342 void
 343 metrics_counter_stop (void)
 344 {
 345         glEndQuery (GL_TIME_ELAPSED);
 346         glEndPerfMonitorAMD (current_context.monitor_tail->id);
 347 }
 348
 349 void
 350 metrics_set_current_op (metrics_op_t op)
 351 {
 352         current_context.op = op;
 353 }
 354
 355 metrics_op_t
 356 metrics_get_current_op (void)
 357 {
 358         return current_context.op;
 359 }
 360
 361 static void
 362 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 363 {
 364         metrics_info_t *info = &ctx->metrics_info;
 365         unsigned i, j;
 366
 367         metrics->op = op;
 368         metrics->time_ns = 0.0;
 369
 370         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 371
 372         for (i = 0; i < info->num_groups; i++) {
 373                 metrics->counters[i] = xmalloc (sizeof (double) *
 374                                                 info->groups[i].num_counters);
 375                 for (j = 0; j < info->groups[i].num_counters; j++)
 376                         metrics->counters[i][j] = 0.0;
 377         }
 378 }
 379
 380 static op_metrics_t *
 381 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 382 {
 383         unsigned i;
 384
 385         if (op >= ctx->num_op_metrics)
 386         {
 387                 ctx->op_metrics = realloc (ctx->op_metrics,
 388                                            (op + 1) * sizeof (op_metrics_t));
 389                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 390                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 391
 392                 ctx->num_op_metrics = op + 1;
 393         }
 394
 395         return &ctx->op_metrics[op];
 396 }
 397
 398 static void
 399 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 400 {
 401 #define CONSUME(var)                                                    \
 402         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 403         {                                                               \
 404                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 405                          "parsing results\n");                          \
 406                 break;                                                  \
 407         }                                                               \
 408         (var) = *((typeof(var) *) p);                                   \
 409         p += sizeof(var);
 410
 411         context_t *ctx = &current_context;
 412         metrics_info_t *info = &ctx->metrics_info;
 413         unsigned char *p = (unsigned char *) result;
 414
 415         while (p < ((unsigned char *) result) + size)
 416         {
 417                 GLuint group_id, group_index;
 418                 GLuint counter_id, counter_index;
 419                 metrics_group_info_t *group;
 420                 uint32_t value;
 421                 unsigned i;
 422
 423                 CONSUME (group_id);
 424                 CONSUME (counter_id);
 425                 CONSUME (value);
 426
 427                 for (i = 0; i < info->num_groups; i++) {
 428                         if (info->groups[i].id == i)
 429                                 break;
 430                 }
 431                 group_index = i;
 432                 assert (group_index < info->num_groups);
 433                 group = &info->groups[group_index];
 434
 435                 for (i = 0; i < group->num_counters; i++) {
 436                         if (group->counter_ids[i] == counter_id)
 437                                 break;
 438                 }
 439                 counter_index = i;
 440                 assert (counter_index < group->num_counters);
 441
 442                 ctx->op_metrics[op].counters[group_index][counter_index] += value;
 443         }
 444 }
 445
 446 static void
 447 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 448 {
 449         op_metrics_t *metrics;
 450
 451         metrics = ctx_get_op_metrics (&current_context, op);
 452
 453         metrics->time_ns += time_ns;
 454 }
 455
 456 static int
 457 time_compare(const void *in_a, const void *in_b, void *arg)
 458 {
 459         int a = *(const int *)in_a;
 460         int b = *(const int *)in_b;
 461         struct op_metrics *metrics = arg;
 462
 463         if (metrics[a].time_ns < metrics[b].time_ns)
 464                 return -1;
 465         if (metrics[a].time_ns > metrics[b].time_ns)
 466                 return 1;
 467         return 0;
 468 }
 469
 470 static void
 471 print_op_metrics (context_t *ctx, op_metrics_t *metric, double total)
 472 {
 473         metrics_info_t *info = &ctx->metrics_info;
 474         metrics_group_info_t *group;
 475         const char *op_string;
 476         unsigned i, group_index, counter;
 477         double value;
 478
 479         /* Since we sparsely fill the array based on program
 480          * id, many "programs" have no time.
 481          */
 482         if (metric->time_ns == 0.0)
 483                 return;
 484
 485         op_string = metrics_op_string (metric->op);
 486
 487         printf ("%s", op_string);
 488         if (metric->op >= METRICS_OP_SHADER) {
 489                 printf (" %d:", metric->op - METRICS_OP_SHADER);
 490         } else {
 491                 printf (":");
 492                 for (i = strlen (op_string); i < 20; i++)
 493                         printf (" ");
 494         }
 495
 496         printf ("\t%7.2f ms (% 2.1f%%)",
 497                 metric->time_ns / 1e6,
 498                 metric->time_ns / total * 100);
 499
 500         printf ("[");
 501         for (group_index = 0; group_index < info->num_groups; group_index++) {
 502                 group = &info->groups[group_index];
 503                 for (counter = 0; counter < group->num_counters; counter++) {
 504                         value = metric->counters[group_index][counter];
 505                         if (value == 0.0)
 506                                 continue;
 507                         printf ("%s: %.2f ", group->counter_names[counter],
 508                                 value / 1e6);
 509                 }
 510         }
 511         printf ("]\n");
 512 }
 513
 514 static void
 515 print_program_metrics (void)
 516 {
 517         context_t *ctx = &current_context;
 518         int *sorted; /* Sorted indices into the ctx->op_metrics */
 519         double total = 0;
 520         unsigned i;
 521
 522         /* Make a sorted list of the operations by time used, and figure
 523          * out the total so we can print percentages.
 524          */
 525         sorted = calloc(ctx->num_op_metrics, sizeof(*sorted));
 526         for (i = 0; i < ctx->num_op_metrics; i++) {
 527                 sorted[i] = i;
 528                 total += ctx->op_metrics[i].time_ns;
 529         }
 530         qsort_r(sorted, ctx->num_op_metrics, sizeof(*sorted),
 531                 time_compare, ctx->op_metrics);
 532
 533         for (i = 0; i < ctx->num_op_metrics; i++)
 534                 print_op_metrics (ctx, &ctx->op_metrics[sorted[i]], total);
 535
 536         free (sorted);
 537 }
 538
 539 /* Called at program exit */
 540 static void
 541 metrics_exit (void)
 542 {
 543         if (verbose)
 544                 printf ("fips: terminating\n");
 545
 546         metrics_info_fini (&current_context.metrics_info);
 547 }
 548
 549
 550 void
 551 metrics_end_frame (void)
 552 {
 553         static int initialized = 0;
 554         static struct timeval tv_start, tv_now;
 555
 556         if (! initialized) {
 557                 gettimeofday (&tv_start, NULL);
 558                 atexit (metrics_exit);
 559                 if (getenv ("FIPS_VERBOSE"))
 560                         verbose = 1;
 561                 initialized = 1;
 562         }
 563
 564         if (verbose)
 565                 printf ("fips: frame %d complete\n", frames);
 566
 567         frames++;
 568         gettimeofday (&tv_now, NULL);
 569
 570         /* Consume all timer queries that are ready. */
 571         timer_query_t *timer = current_context.timer_head;
 572
 573         while (timer) {
 574                 GLuint available, elapsed;
 575
 576                 glGetQueryObjectuiv (timer->id,
 577                                      GL_QUERY_RESULT_AVAILABLE, &available);
 578                 if (! available)
 579                         break;
 580
 581                 glGetQueryObjectuiv (timer->id,
 582                                      GL_QUERY_RESULT, &elapsed);
 583
 584                 accumulate_program_time (timer->op, elapsed);
 585
 586                 current_context.timer_head = timer->next;
 587                 if (current_context.timer_head == NULL)
 588                         current_context.timer_tail = NULL;
 589
 590                 glDeleteQueries (1, &timer->id);
 591
 592                 free (timer);
 593                 timer = current_context.timer_head;
 594         }
 595
 596         /* And similarly for all performance monitors that are ready. */
 597         monitor_t *monitor = current_context.monitor_head;
 598
 599         while (monitor) {
 600                 GLuint available, result_size, *result;
 601                 GLint bytes_written;
 602
 603                 glGetPerfMonitorCounterDataAMD (monitor->id,
 604                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 605                                                 sizeof (available), &available,
 606                                                 NULL);
 607                 if (! available)
 608                         break;
 609
 610                 glGetPerfMonitorCounterDataAMD (monitor->id,
 611                                                 GL_PERFMON_RESULT_SIZE_AMD,
 612                                                 sizeof (result_size),
 613                                                 &result_size, NULL);
 614
 615                 result = xmalloc (result_size);
 616
 617                 glGetPerfMonitorCounterDataAMD (monitor->id,
 618                                                 GL_PERFMON_RESULT_AMD,
 619                                                 result_size, result,
 620                                                 &bytes_written);
 621
 622                 accumulate_program_metrics (monitor->op, result, result_size);
 623
 624                 free (result);
 625
 626                 current_context.monitor_head = monitor->next;
 627                 if (current_context.monitor_head == NULL)
 628                         current_context.monitor_tail = NULL;
 629
 630                 glDeletePerfMonitorsAMD (1, &monitor->id);
 631
 632                 free (monitor);
 633                 monitor = current_context.monitor_head;
 634         }
 635
 636         if (frames % 60 == 0) {
 637                 double fps;
 638
 639                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 640                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 641
 642                 printf("FPS: %.3f\n", fps);
 643
 644                 print_program_metrics ();
 645         }
 646 }