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 *counters;
  73         char **counter_names;
  74         GLuint *counter_types;
  75
  76 } metrics_group_info_t;
  77
  78 typedef struct metrics_info
  79 {
  80         unsigned num_groups;
  81         metrics_group_info_t *groups;
  82 } metrics_info_t;
  83
  84 typedef struct context
  85 {
  86         metrics_info_t metrics_info;
  87
  88         metrics_op_t op;
  89
  90         timer_query_t *timer_head;
  91         timer_query_t *timer_tail;
  92
  93         monitor_t *monitor_head;
  94         monitor_t *monitor_tail;
  95
  96         unsigned num_op_metrics;
  97         op_metrics_t *op_metrics;
  98 } context_t;
  99
 100 /* FIXME: Need a map from integers to context objects and track the
 101  * current context with glXMakeContextCurrent, eglMakeCurrent, etc. */
 102
 103 context_t current_context;
 104
 105 int frames;
 106 int verbose;
 107
 108 static void
 109 metrics_group_info_init (metrics_group_info_t *group, GLuint id)
 110 {
 111         GLsizei length;
 112         unsigned i;
 113
 114         group->id = id;
 115
 116         /* Get name */
 117         glGetPerfMonitorGroupStringAMD (id, 0, &length, NULL);
 118
 119         group->name = xmalloc (length + 1);
 120
 121         glGetPerfMonitorGroupStringAMD (id, length + 1, NULL, group->name);
 122
 123         /* Get number of counters */
 124         group->num_counters = 0;
 125         group->max_active_counters = 0;
 126         glGetPerfMonitorCountersAMD (group->id,
 127                                      (int *) &group->num_counters,
 128                                      (int *) &group->max_active_counters,
 129                                      0, NULL);
 130
 131         /* Get counter numbers */
 132         group->counters = xmalloc (group->num_counters * sizeof (GLuint));
 133
 134         glGetPerfMonitorCountersAMD (group->id, NULL, NULL,
 135                                      group->num_counters,
 136                                      group->counters);
 137
 138         /* Get counter names */
 139         group->counter_names = xmalloc (group->num_counters * sizeof (char *));
 140         group->counter_types = xmalloc (group->num_counters * sizeof (GLuint));
 141
 142         for (i = 0; i < group->num_counters; i++) {
 143                 glGetPerfMonitorCounterInfoAMD (group->id,
 144                                                 group->counters[i],
 145                                                 GL_COUNTER_TYPE_AMD,
 146                                                 &group->counter_types[i]);
 147
 148                 /* We assume that all peformance counters are made
 149                  * available as uint32 values. The code calling
 150                  * CONSUME in accumulate_program_metrics will need to
 151                  * be extended to accomodate other counter values. */
 152                 if (group->counter_types[i] != GL_UNSIGNED_INT) {
 153                         fprintf (stderr, "fips: Internal error: No support for non-uint counter values\n");
 154                         exit (1);
 155                 }
 156
 157                 glGetPerfMonitorCounterStringAMD (group->id,
 158                                                   group->counters[i],
 159                                                   0, &length, NULL);
 160
 161                 group->counter_names[i] = xmalloc (length + 1);
 162
 163                 glGetPerfMonitorCounterStringAMD (group->id,
 164                                                   group->counters[i],
 165                                                   length + 1, NULL,
 166                                                   group->counter_names[i]);
 167         }
 168 }
 169
 170 void
 171 metrics_info_init (void)
 172 {
 173         unsigned i;
 174         GLuint *group_ids;
 175         metrics_info_t *metrics_info = &current_context.metrics_info;
 176
 177         glGetPerfMonitorGroupsAMD ((int *) &metrics_info->num_groups, 0, NULL);
 178
 179         group_ids = xmalloc (metrics_info->num_groups * sizeof (GLuint));
 180
 181         glGetPerfMonitorGroupsAMD (NULL, metrics_info->num_groups, group_ids);
 182
 183         metrics_info->groups = xmalloc (metrics_info->num_groups * sizeof (metrics_group_info_t));
 184
 185         for (i = 0; i < metrics_info->num_groups; i++)
 186                 metrics_group_info_init (&metrics_info->groups[i], i);
 187
 188         free (group_ids);
 189 }
 190
 191 static const char *
 192 metrics_op_string (metrics_op_t op)
 193 {
 194         if (op >= METRICS_OP_SHADER)
 195                 return "Shader program";
 196
 197         switch (op)
 198         {
 199         case METRICS_OP_ACCUM:
 200                 return "glAccum*(+)";
 201         case METRICS_OP_BUFFER_DATA:
 202                 return "glBufferData(+)";
 203         case METRICS_OP_BUFFER_SUB_DATA:
 204                 return "glCopyBufferSubData*";
 205         case METRICS_OP_BITMAP:
 206                 return "glBitmap*";
 207         case METRICS_OP_BLIT_FRAMEBUFFER:
 208                 return "glBlitFramebuffer*";
 209         case METRICS_OP_CLEAR:
 210                 return "glClear(+)";
 211         case METRICS_OP_CLEAR_BUFFER_DATA:
 212                 return "glCearBufferData(+)";
 213         case METRICS_OP_CLEAR_TEX_IMAGE:
 214                 return "glClearTexImage(+)";
 215         case METRICS_OP_COPY_PIXELS:
 216                 return "glCopyPixels";
 217         case METRICS_OP_COPY_TEX_IMAGE:
 218                 return "glCopyTexImage(+)";
 219         case METRICS_OP_DRAW_PIXELS:
 220                 return "glDrawPixels";
 221         case METRICS_OP_GET_TEX_IMAGE:
 222                 return "glGetTexImage(+)";
 223         case METRICS_OP_READ_PIXELS:
 224                 return "glReadPixels*";
 225         case METRICS_OP_TEX_IMAGE:
 226                 return "glTexImage*(+)";
 227         default:
 228                 fprintf (stderr, "fips: Internal error: "
 229                          "Unknown metrics op value: %d\n", op);
 230                 exit (1);
 231         }
 232
 233         return "";
 234 }
 235
 236 void
 237 metrics_counter_start (void)
 238 {
 239         context_t *ctx = &current_context;
 240         timer_query_t *timer;
 241         monitor_t *monitor;
 242         unsigned i;
 243
 244         /* Create new timer query, add to list */
 245         timer = xmalloc (sizeof (timer_query_t));
 246
 247         timer->op = ctx->op;
 248         timer->next = NULL;
 249
 250         if (ctx->timer_tail) {
 251                 ctx->timer_tail->next = timer;
 252                 ctx->timer_tail = timer;
 253         } else {
 254                 ctx->timer_tail = timer;
 255                 ctx->timer_head = timer;
 256         }
 257
 258         /* Create a new performance-monitor query */
 259         monitor = xmalloc (sizeof (monitor_t));
 260
 261         monitor->op = ctx->op;
 262         monitor->next = NULL;
 263
 264         if (ctx->monitor_tail) {
 265                 ctx->monitor_tail->next = monitor;
 266                 ctx->monitor_tail = monitor;
 267         } else {
 268                 ctx->monitor_tail = monitor;
 269                 ctx->monitor_head = monitor;
 270         }
 271
 272         /* Initialize the timer_query and monitor objects */
 273         glGenQueries (1, &timer->id);
 274
 275         glGenPerfMonitorsAMD (1, &monitor->id);
 276
 277         for (i = 0; i < ctx->metrics_info.num_groups; i++)
 278         {
 279                 metrics_group_info_t *group;
 280                 int num_counters;
 281
 282                 group = &ctx->metrics_info.groups[i];
 283
 284                 num_counters = group->num_counters;
 285                 if (group->max_active_counters < group->num_counters)
 286                 {
 287                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 288                                  group->max_active_counters,
 289                                  group->num_counters, i);
 290                         num_counters = group->max_active_counters;
 291
 292                 }
 293
 294                 glSelectPerfMonitorCountersAMD(monitor->id,
 295                                                GL_TRUE, group->id,
 296                                                num_counters,
 297                                                group->counters);
 298         }
 299
 300         /* Start the queries */
 301         glBeginQuery (GL_TIME_ELAPSED, timer->id);
 302
 303         glBeginPerfMonitorAMD (monitor->id);
 304 }
 305
 306 void
 307 metrics_counter_stop (void)
 308 {
 309         glEndQuery (GL_TIME_ELAPSED);
 310         glEndPerfMonitorAMD (current_context.monitor_tail->id);
 311 }
 312
 313 void
 314 metrics_set_current_op (metrics_op_t op)
 315 {
 316         current_context.op = op;
 317 }
 318
 319 metrics_op_t
 320 metrics_get_current_op (void)
 321 {
 322         return current_context.op;
 323 }
 324
 325 static void
 326 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 327 {
 328         metrics_info_t *info = &ctx->metrics_info;
 329         unsigned i, j;
 330
 331         metrics->op = op;
 332         metrics->time_ns = 0.0;
 333
 334         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 335
 336         for (i = 0; i < info->num_groups; i++) {
 337                 metrics->counters[i] = xmalloc (sizeof (double) *
 338                                                 info->groups[i].num_counters);
 339                 for (j = 0; j < info->groups[i].num_counters; j++)
 340                         metrics->counters[i][j] = 0.0;
 341         }
 342 }
 343
 344 static op_metrics_t *
 345 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 346 {
 347         unsigned i;
 348
 349         if (op >= ctx->num_op_metrics)
 350         {
 351                 ctx->op_metrics = realloc (ctx->op_metrics,
 352                                            (op + 1) * sizeof (op_metrics_t));
 353                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 354                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 355
 356                 ctx->num_op_metrics = op + 1;
 357         }
 358
 359         return &ctx->op_metrics[op];
 360 }
 361
 362 static void
 363 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 364 {
 365 #define CONSUME(var)                                                    \
 366         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 367         {                                                               \
 368                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 369                          "parsing results\n");                          \
 370                 break;                                                  \
 371         }                                                               \
 372         (var) = *((typeof(var) *) p);                                   \
 373         p += sizeof(var);
 374
 375         context_t *ctx = &current_context;
 376         unsigned char *p = (unsigned char *) result;
 377
 378         while (p < ((unsigned char *) result) + size)
 379         {
 380                 GLuint group_id, counter_id, counter_index;
 381                 metrics_group_info_t *group;
 382                 uint32_t value;
 383                 unsigned i;
 384
 385                 CONSUME (group_id);
 386                 CONSUME (counter_id);
 387                 CONSUME (value);
 388
 389                 assert (group_id < ctx->metrics_info.num_groups);
 390                 group = &ctx->metrics_info.groups[group_id];
 391
 392                 for (i = 0; i < group->num_counters; i++) {
 393                         if (group->counters[i] == counter_id)
 394                                 break;
 395                 }
 396                 counter_index = i;
 397                 assert (counter_index < group->num_counters);
 398
 399                 ctx->op_metrics[op].counters[group_id][counter_index] += value;
 400         }
 401 }
 402
 403 static void
 404 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 405 {
 406         op_metrics_t *metrics;
 407
 408         metrics = ctx_get_op_metrics (&current_context, op);
 409
 410         metrics->time_ns += time_ns;
 411 }
 412
 413 static int
 414 time_compare(const void *in_a, const void *in_b, void *arg)
 415 {
 416         int a = *(const int *)in_a;
 417         int b = *(const int *)in_b;
 418         struct op_metrics *metrics = arg;
 419
 420         if (metrics[a].time_ns < metrics[b].time_ns)
 421                 return -1;
 422         if (metrics[a].time_ns > metrics[b].time_ns)
 423                 return 1;
 424         return 0;
 425 }
 426
 427 static void
 428 print_op_metrics (context_t *ctx, op_metrics_t *metric, double total)
 429 {
 430         metrics_info_t *info = &ctx->metrics_info;
 431         metrics_group_info_t *group;
 432         const char *op_string;
 433         unsigned i, group_id, counter;
 434         double value;
 435
 436         /* Since we sparsely fill the array based on program
 437          * id, many "programs" have no time.
 438          */
 439         if (metric->time_ns == 0.0)
 440                 return;
 441
 442         op_string = metrics_op_string (metric->op);
 443
 444         printf ("%s", op_string);
 445         if (metric->op >= METRICS_OP_SHADER) {
 446                 printf (" %d:", metric->op - METRICS_OP_SHADER);
 447         } else {
 448                 printf (":");
 449                 for (i = strlen (op_string); i < 20; i++)
 450                         printf (" ");
 451         }
 452
 453         printf ("\t%7.2f ms (% 2.1f%%)",
 454                 metric->time_ns / 1e6,
 455                 metric->time_ns / total * 100);
 456
 457         printf ("[");
 458         for (group_id = 0; group_id < info->num_groups; group_id++) {
 459                 group = &info->groups[group_id];
 460                 for (counter = 0; counter < group->num_counters; counter++) {
 461                         value = metric->counters[group_id][counter];
 462                         if (value == 0.0)
 463                                 continue;
 464                         printf ("%s: %.2f ", group->counter_names[counter],
 465                                 value / 1e6);
 466                 }
 467         }
 468         printf ("]\n");
 469 }
 470
 471 static void
 472 print_program_metrics (void)
 473 {
 474         context_t *ctx = &current_context;
 475         int *sorted; /* Sorted indices into the ctx->op_metrics */
 476         double total = 0;
 477         unsigned i;
 478
 479         /* Make a sorted list of the operations by time used, and figure
 480          * out the total so we can print percentages.
 481          */
 482         sorted = calloc(ctx->num_op_metrics, sizeof(*sorted));
 483         for (i = 0; i < ctx->num_op_metrics; i++) {
 484                 sorted[i] = i;
 485                 total += ctx->op_metrics[i].time_ns;
 486         }
 487         qsort_r(sorted, ctx->num_op_metrics, sizeof(*sorted),
 488                 time_compare, ctx->op_metrics);
 489
 490         for (i = 0; i < ctx->num_op_metrics; i++)
 491                 print_op_metrics (ctx, &ctx->op_metrics[sorted[i]], total);
 492
 493         free (sorted);
 494 }
 495
 496 /* Called at program exit */
 497 static void
 498 metrics_exit (void)
 499 {
 500         if (verbose)
 501                 printf ("fips: terminating\n");
 502 }
 503
 504
 505 void
 506 metrics_end_frame (void)
 507 {
 508         static int initialized = 0;
 509         static struct timeval tv_start, tv_now;
 510
 511         if (! initialized) {
 512                 gettimeofday (&tv_start, NULL);
 513                 atexit (metrics_exit);
 514                 if (getenv ("FIPS_VERBOSE"))
 515                         verbose = 1;
 516                 initialized = 1;
 517         }
 518
 519         if (verbose)
 520                 printf ("fips: frame %d complete\n", frames);
 521
 522         frames++;
 523         gettimeofday (&tv_now, NULL);
 524
 525         /* Consume all timer queries that are ready. */
 526         timer_query_t *timer = current_context.timer_head;
 527
 528         while (timer) {
 529                 GLuint available, elapsed;
 530
 531                 glGetQueryObjectuiv (timer->id,
 532                                      GL_QUERY_RESULT_AVAILABLE, &available);
 533                 if (! available)
 534                         break;
 535
 536                 glGetQueryObjectuiv (timer->id,
 537                                      GL_QUERY_RESULT, &elapsed);
 538
 539                 accumulate_program_time (timer->op, elapsed);
 540
 541                 current_context.timer_head = timer->next;
 542                 if (current_context.timer_head == NULL)
 543                         current_context.timer_tail = NULL;
 544
 545                 glDeleteQueries (1, &timer->id);
 546
 547                 free (timer);
 548                 timer = current_context.timer_head;
 549         }
 550
 551         /* And similarly for all performance monitors that are ready. */
 552         monitor_t *monitor = current_context.monitor_head;
 553
 554         while (monitor) {
 555                 GLuint available, result_size, *result;
 556                 GLint bytes_written;
 557
 558                 glGetPerfMonitorCounterDataAMD (monitor->id,
 559                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 560                                                 sizeof (available), &available,
 561                                                 NULL);
 562                 if (! available)
 563                         break;
 564
 565                 glGetPerfMonitorCounterDataAMD (monitor->id,
 566                                                 GL_PERFMON_RESULT_SIZE_AMD,
 567                                                 sizeof (result_size),
 568                                                 &result_size, NULL);
 569
 570                 result = xmalloc (result_size);
 571
 572                 glGetPerfMonitorCounterDataAMD (monitor->id,
 573                                                 GL_PERFMON_RESULT_AMD,
 574                                                 result_size, result,
 575                                                 &bytes_written);
 576
 577                 accumulate_program_metrics (monitor->op, result, result_size);
 578
 579                 free (result);
 580
 581                 current_context.monitor_head = monitor->next;
 582                 if (current_context.monitor_head == NULL)
 583                         current_context.monitor_tail = NULL;
 584
 585                 glDeletePerfMonitorsAMD (1, &monitor->id);
 586
 587                 free (monitor);
 588                 monitor = current_context.monitor_head;
 589         }
 590
 591         if (frames % 60 == 0) {
 592                 double fps;
 593
 594                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 595                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 596
 597                 printf("FPS: %.3f\n", fps);
 598
 599                 print_program_metrics ();
 600         }
 601 }