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 "context.h"
  33 #include "metrics-info.h"
  34 #include "xmalloc.h"
  35
  36 int frames;
  37 int verbose;
  38
  39 #define MAX_MONITORS_IN_FLIGHT 1000
  40
  41 static const char *
  42 metrics_op_string (metrics_op_t op)
  43 {
  44         if (op >= METRICS_OP_SHADER)
  45                 return "Shader program";
  46
  47         switch (op)
  48         {
  49         case METRICS_OP_ACCUM:
  50                 return "glAccum*(+)";
  51         case METRICS_OP_BUFFER_DATA:
  52                 return "glBufferData(+)";
  53         case METRICS_OP_BUFFER_SUB_DATA:
  54                 return "glCopyBufferSubData*";
  55         case METRICS_OP_BITMAP:
  56                 return "glBitmap*";
  57         case METRICS_OP_BLIT_FRAMEBUFFER:
  58                 return "glBlitFramebuffer*";
  59         case METRICS_OP_CLEAR:
  60                 return "glClear(+)";
  61         case METRICS_OP_CLEAR_BUFFER_DATA:
  62                 return "glCearBufferData(+)";
  63         case METRICS_OP_CLEAR_TEX_IMAGE:
  64                 return "glClearTexImage(+)";
  65         case METRICS_OP_COPY_PIXELS:
  66                 return "glCopyPixels";
  67         case METRICS_OP_COPY_TEX_IMAGE:
  68                 return "glCopyTexImage(+)";
  69         case METRICS_OP_DRAW_PIXELS:
  70                 return "glDrawPixels";
  71         case METRICS_OP_GET_TEX_IMAGE:
  72                 return "glGetTexImage(+)";
  73         case METRICS_OP_READ_PIXELS:
  74                 return "glReadPixels*";
  75         case METRICS_OP_TEX_IMAGE:
  76                 return "glTexImage*(+)";
  77         default:
  78                 fprintf (stderr, "fips: Internal error: "
  79                          "Unknown metrics op value: %d\n", op);
  80                 exit (1);
  81         }
  82
  83         return "";
  84 }
  85
  86 void
  87 metrics_counter_start (void)
  88 {
  89         context_t *ctx = context_get_current ();
  90         unsigned i;
  91
  92         /* Initialize the timer_query and monitor objects */
  93         glGenQueries (1, &ctx->timer_begun_id);
  94
  95         glGenPerfMonitorsAMD (1, &ctx->monitor_begun_id);
  96
  97         for (i = 0; i < ctx->metrics_info.num_groups; i++)
  98         {
  99                 metrics_group_info_t *group;
 100                 int num_counters;
 101
 102                 group = &ctx->metrics_info.groups[i];
 103
 104                 num_counters = group->num_counters;
 105                 if (group->max_active_counters < group->num_counters)
 106                 {
 107                         fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
 108                                  group->max_active_counters,
 109                                  group->num_counters, i);
 110                         num_counters = group->max_active_counters;
 111
 112                 }
 113
 114                 glSelectPerfMonitorCountersAMD(ctx->monitor_begun_id,
 115                                                GL_TRUE, group->id,
 116                                                num_counters,
 117                                                group->counter_ids);
 118         }
 119
 120         /* Start the queries */
 121         glBeginQuery (GL_TIME_ELAPSED, ctx->timer_begun_id);
 122
 123         glBeginPerfMonitorAMD (ctx->monitor_begun_id);
 124 }
 125
 126 void
 127 metrics_counter_stop (void)
 128 {
 129         context_t *ctx = context_get_current ();
 130         timer_query_t *timer;
 131         monitor_t *monitor;
 132
 133         /* Stop the current timer and monitor. */
 134         glEndQuery (GL_TIME_ELAPSED);
 135         glEndPerfMonitorAMD (ctx->monitor_begun_id);
 136
 137         /* Add these IDs to our lists of outstanding queries and
 138          * monitors so the results can be collected later. */
 139         timer = xmalloc (sizeof (timer_query_t));
 140
 141         timer->op = ctx->op;
 142         timer->id = ctx->timer_begun_id;
 143         timer->next = NULL;
 144
 145         if (ctx->timer_tail) {
 146                 ctx->timer_tail->next = timer;
 147                 ctx->timer_tail = timer;
 148         } else {
 149                 ctx->timer_tail = timer;
 150                 ctx->timer_head = timer;
 151         }
 152
 153         /* Create a new performance-monitor query */
 154         monitor = xmalloc (sizeof (monitor_t));
 155
 156         monitor->op = ctx->op;
 157         monitor->id = ctx->monitor_begun_id;
 158         monitor->next = NULL;
 159
 160         if (ctx->monitor_tail) {
 161                 ctx->monitor_tail->next = monitor;
 162                 ctx->monitor_tail = monitor;
 163         } else {
 164                 ctx->monitor_tail = monitor;
 165                 ctx->monitor_head = monitor;
 166         }
 167
 168         ctx->monitors_in_flight++;
 169
 170         /* Avoid being a resource hog and collect outstanding results
 171          * once we have sent off a large number of
 172          * queries. (Presumably, many of the outstanding queries are
 173          * available by now.)
 174          */
 175         if (ctx->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
 176                 metrics_collect_available ();
 177 }
 178
 179 void
 180 metrics_set_current_op (metrics_op_t op)
 181 {
 182         context_t *ctx = context_get_current ();
 183
 184         ctx->op = op;
 185 }
 186
 187 metrics_op_t
 188 metrics_get_current_op (void)
 189 {
 190         context_t *ctx = context_get_current ();
 191
 192         return ctx->op;
 193 }
 194
 195 static void
 196 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
 197 {
 198         metrics_info_t *info = &ctx->metrics_info;
 199         unsigned i, j;
 200
 201         metrics->op = op;
 202         metrics->time_ns = 0.0;
 203
 204         metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
 205
 206         for (i = 0; i < info->num_groups; i++) {
 207                 metrics->counters[i] = xmalloc (sizeof (double) *
 208                                                 info->groups[i].num_counters);
 209                 for (j = 0; j < info->groups[i].num_counters; j++)
 210                         metrics->counters[i][j] = 0.0;
 211         }
 212 }
 213
 214 static op_metrics_t *
 215 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
 216 {
 217         unsigned i;
 218
 219         if (op >= ctx->num_op_metrics)
 220         {
 221                 ctx->op_metrics = realloc (ctx->op_metrics,
 222                                            (op + 1) * sizeof (op_metrics_t));
 223                 for (i = ctx->num_op_metrics; i < op + 1; i++)
 224                         op_metrics_init (ctx, &ctx->op_metrics[i], i);
 225
 226                 ctx->num_op_metrics = op + 1;
 227         }
 228
 229         return &ctx->op_metrics[op];
 230 }
 231
 232 static void
 233 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
 234 {
 235 #define CONSUME(var)                                                    \
 236         if (p + sizeof(var) > ((unsigned char *) result) + size)        \
 237         {                                                               \
 238                 fprintf (stderr, "Unexpected end-of-buffer while "      \
 239                          "parsing results\n");                          \
 240                 break;                                                  \
 241         }                                                               \
 242         (var) = *((typeof(var) *) p);                                   \
 243         p += sizeof(var);
 244
 245         context_t *ctx = context_get_current ();
 246         metrics_info_t *info = &ctx->metrics_info;
 247         op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
 248         unsigned char *p = (unsigned char *) result;
 249
 250         while (p < ((unsigned char *) result) + size)
 251         {
 252                 GLuint group_id, group_index;
 253                 GLuint counter_id, counter_index;
 254                 metrics_group_info_t *group;
 255                 double value;
 256                 unsigned i;
 257
 258                 CONSUME (group_id);
 259                 CONSUME (counter_id);
 260
 261                 for (i = 0; i < info->num_groups; i++) {
 262                         if (info->groups[i].id == group_id)
 263                                 break;
 264                 }
 265                 group_index = i;
 266                 assert (group_index < info->num_groups);
 267                 group = &info->groups[group_index];
 268
 269                 for (i = 0; i < group->num_counters; i++) {
 270                         if (group->counter_ids[i] == counter_id)
 271                                 break;
 272                 }
 273                 counter_index = i;
 274                 assert (counter_index < group->num_counters);
 275
 276                 switch (group->counter_types[counter_index])
 277                 {
 278                         uint uint_value;
 279                         uint64_t uint64_value;
 280                         float float_value;
 281                 case GL_UNSIGNED_INT:
 282                         CONSUME (uint_value);
 283                         value = uint_value;
 284                         break;
 285                 case GL_UNSIGNED_INT64_AMD:
 286                         CONSUME (uint64_value);
 287                         value = uint64_value;
 288                         break;
 289                 case GL_PERCENTAGE_AMD:
 290                 case GL_FLOAT:
 291                         CONSUME (float_value);
 292                         value = float_value;
 293                         break;
 294                 default:
 295                         fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
 296                                  group->counter_types[counter_index]);
 297                         value = 0.0;
 298                         break;
 299                 }
 300
 301                 metrics->counters[group_index][counter_index] += value;
 302         }
 303 }
 304
 305 static void
 306 accumulate_program_time (metrics_op_t op, unsigned time_ns)
 307 {
 308         context_t *ctx = context_get_current ();
 309         op_metrics_t *metrics;
 310
 311         metrics = ctx_get_op_metrics (ctx, op);
 312
 313         metrics->time_ns += time_ns;
 314 }
 315
 316 typedef struct per_stage_metrics
 317 {
 318         op_metrics_t *metrics;
 319         shader_stage_info_t *stage;
 320         double time_ns;
 321         double active;
 322 } per_stage_metrics_t;
 323
 324 static int
 325 _is_shader_stage_counter (metrics_info_t *info,
 326                           unsigned group_index,
 327                           unsigned counter_index)
 328 {
 329         shader_stage_info_t *stage;
 330         unsigned i;
 331
 332         for (i = 0; i < info->num_shader_stages; i++) {
 333                 stage = &info->stages[i];
 334
 335                 if (stage->active_group_index == group_index &&
 336                     stage->active_counter_index == counter_index)
 337                 {
 338                         return 1;
 339                 }
 340
 341                 if (stage->stall_group_index == group_index &&
 342                     stage->stall_counter_index == counter_index)
 343                 {
 344                         return 1;
 345                 }
 346         }
 347
 348         return 0;
 349 }
 350
 351 static void
 352 print_per_stage_metrics (context_t *ctx,
 353                          per_stage_metrics_t *per_stage,
 354                          double total)
 355 {
 356         metrics_info_t *info = &ctx->metrics_info;
 357         op_metrics_t *metric = per_stage->metrics;
 358         metrics_group_info_t *group;
 359         const char *op_string;
 360         unsigned group_index, counter;
 361         double value;
 362
 363         /* Don't print anything for stages with no alloted time. */
 364         if (per_stage->time_ns == 0.0)
 365                 return;
 366
 367         op_string = metrics_op_string (metric->op);
 368
 369         printf ("%21s", op_string);
 370
 371         if (metric->op >= METRICS_OP_SHADER) {
 372                 printf (" %3d", metric->op - METRICS_OP_SHADER);
 373         } else {
 374                 printf ("    ");
 375
 376         }
 377
 378         if (per_stage->stage)
 379                 printf (" %cS:", per_stage->stage->name[0]);
 380         else
 381                 printf ("   :");
 382
 383         printf ("\t%7.2f ms (%4.1f%%)",
 384                 per_stage->time_ns / 1e6,
 385                 per_stage->time_ns / total * 100);
 386
 387         if (per_stage->active)
 388                 printf (", %4.1f%% active", per_stage->active * 100);
 389
 390         printf ("\n");
 391
 392         /* I'm not seeing a lot of value printing the rest of these
 393          * performance counters by default yet. Use --verbose to get
 394          * them for now. */
 395         if (! verbose)
 396                 return;
 397
 398         printf ("[");
 399         for (group_index = 0; group_index < info->num_groups; group_index++) {
 400                 group = &info->groups[group_index];
 401                 for (counter = 0; counter < group->num_counters; counter++) {
 402
 403                         /* Don't print this counter value if it's a
 404                          * per-stage cycle counter, (which we have
 405                          * already accounted for). */
 406                         if (_is_shader_stage_counter (info, group_index, counter))
 407                                 continue;
 408
 409                         value = metric->counters[group_index][counter];
 410                         if (value == 0.0)
 411                                 continue;
 412                         printf ("%s: %.2f ", group->counter_names[counter],
 413                                 value / 1e6);
 414                 }
 415         }
 416         printf ("]\n");
 417 }
 418
 419 static int
 420 time_compare(const void *in_a, const void *in_b, void *arg unused)
 421 {
 422         const per_stage_metrics_t *a = in_a;
 423         const per_stage_metrics_t *b = in_b;
 424
 425
 426         if (a->time_ns < b->time_ns)
 427                 return -1;
 428         if (a->time_ns > b->time_ns)
 429                 return 1;
 430         return 0;
 431 }
 432
 433 static void
 434 print_program_metrics (void)
 435 {
 436         context_t *ctx = context_get_current ();
 437         metrics_info_t *info = &ctx->metrics_info;
 438         unsigned num_shader_stages = info->num_shader_stages;
 439         per_stage_metrics_t *sorted, *per_stage;
 440         double total_time, op_cycles;
 441         op_metrics_t *op;
 442         unsigned group_index, counter_index;
 443         unsigned i, j, num_sorted;
 444
 445         /* Make a sorted list of the per-stage operations by time
 446          * used, and figure out the total so we can print percentages.
 447          */
 448         num_sorted = ctx->num_op_metrics * num_shader_stages;
 449
 450         sorted = xmalloc (sizeof (*sorted) * num_sorted);
 451
 452         total_time = 0.0;
 453
 454         for (i = 0; i < ctx->num_op_metrics; i++) {
 455
 456                 op = &ctx->op_metrics[i];
 457
 458                 /* Accumulate total time across all ops. */
 459                 total_time += op->time_ns;
 460
 461                 /* Also, find total cycles in all stages of this op. */
 462                 op_cycles = 0.0;
 463
 464                 for (j = 0; j < num_shader_stages; j++) {
 465                         /* Active cycles */
 466                         group_index = info->stages[j].active_group_index;
 467                         counter_index = info->stages[j].active_counter_index;
 468                         op_cycles += op->counters[group_index][counter_index];
 469
 470                         /* Stall cycles */
 471                         group_index = info->stages[j].stall_group_index;
 472                         counter_index = info->stages[j].stall_counter_index;
 473                         op_cycles += op->counters[group_index][counter_index];
 474                 }
 475
 476                 for (j = 0; j < num_shader_stages; j++) {
 477                         double active_cycles, stall_cycles, stage_cycles;
 478
 479                         /* Active cycles */
 480                         group_index = info->stages[j].active_group_index;
 481                         counter_index = info->stages[j].active_counter_index;
 482                         active_cycles = op->counters[group_index][counter_index];
 483
 484                         /* Stall cycles */
 485                         group_index = info->stages[j].stall_group_index;
 486                         counter_index = info->stages[j].stall_counter_index;
 487                         stall_cycles = op->counters[group_index][counter_index];
 488
 489                         stage_cycles = active_cycles + stall_cycles;
 490
 491                         per_stage = &sorted[i * num_shader_stages + j];
 492                         per_stage->metrics = op;
 493
 494                         if (op_cycles) {
 495                                 per_stage->stage = &info->stages[j];
 496                                 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
 497                         } else {
 498                                 /* If we don't have any per-stage cycle counts
 499                                  * for this operation, then use the first
 500                                  * stage as a placeholder for all the time,
 501                                  * but NULL-ify the stage info so that the
 502                                  * report doesn't lie about this time being
 503                                  * from any particular stage. */
 504                                 per_stage->stage = NULL;
 505                                 if (j == 0) {
 506                                         per_stage->time_ns = op->time_ns;
 507                                 } else {
 508                                         per_stage->time_ns = 0.0;
 509                                 }
 510                         }
 511
 512                         if (stage_cycles) {
 513                                 per_stage->active = active_cycles / stage_cycles;
 514                         } else {
 515                                 per_stage->active = 0.0;
 516                         }
 517                 }
 518         }
 519
 520         qsort_r (sorted, num_sorted, sizeof (*sorted),
 521                  time_compare, ctx->op_metrics);
 522
 523         for (i = 0; i < num_sorted; i++)
 524                 print_per_stage_metrics (ctx, &sorted[i], total_time);
 525
 526         free (sorted);
 527 }
 528
 529 /* Called at program exit.
 530  *
 531  * This is similar to metrics_info_fini, but only frees any used
 532  * memory. Notably, it does not call any OpenGL functions, (since the
 533  * OpenGL context no longer exists at program exit).
 534  */
 535 static void
 536 metrics_exit (void)
 537 {
 538         context_t *ctx = context_get_current ();
 539         metrics_info_t *info = &ctx->metrics_info;
 540         unsigned i, j;
 541         timer_query_t *timer, *timer_next;
 542         monitor_t *monitor, *monitor_next;
 543
 544         if (verbose)
 545                 printf ("fips: terminating\n");
 546
 547         if (! info->initialized)
 548                 return;
 549
 550         for (timer = ctx->timer_head;
 551              timer;
 552              timer = timer_next)
 553         {
 554                 timer_next = timer->next;
 555                 free (timer);
 556         }
 557
 558         for (monitor = ctx->monitor_head;
 559              monitor;
 560              monitor = monitor_next)
 561         {
 562                 monitor_next = monitor->next;
 563                 free (monitor);
 564         }
 565
 566         for (i = 0; i < info->num_groups; i++) {
 567                 metrics_group_info_t *group = &info->groups[i];
 568
 569                 for (j = 0; j < group->num_counters; i++)
 570                         free (group->counter_names[j]);
 571
 572                 free (group->counter_types);
 573                 free (group->counter_names);
 574                 free (group->counter_ids);
 575
 576                 free (group->name);
 577         }
 578
 579         free (info->groups);
 580
 581         for (i = 0; i < info->num_shader_stages; i++)
 582                 free (info->stages[i].name);
 583
 584         free (info->stages);
 585 }
 586
 587 void
 588 metrics_collect_available (void)
 589 {
 590         context_t *ctx = context_get_current ();
 591
 592         /* Consume all timer queries that are ready. */
 593         timer_query_t *timer = ctx->timer_head;
 594
 595         while (timer) {
 596                 GLuint available, elapsed;
 597
 598                 glGetQueryObjectuiv (timer->id,
 599                                      GL_QUERY_RESULT_AVAILABLE, &available);
 600                 if (! available)
 601                         break;
 602
 603                 glGetQueryObjectuiv (timer->id,
 604                                      GL_QUERY_RESULT, &elapsed);
 605
 606                 accumulate_program_time (timer->op, elapsed);
 607
 608                 ctx->timer_head = timer->next;
 609                 if (ctx->timer_head == NULL)
 610                         ctx->timer_tail = NULL;
 611
 612                 glDeleteQueries (1, &timer->id);
 613
 614                 free (timer);
 615                 timer = ctx->timer_head;
 616         }
 617
 618         /* And similarly for all performance monitors that are ready. */
 619         monitor_t *monitor = ctx->monitor_head;
 620
 621         while (monitor) {
 622                 GLuint available, result_size, *result;
 623                 GLint bytes_written;
 624
 625                 glGetPerfMonitorCounterDataAMD (monitor->id,
 626                                                 GL_PERFMON_RESULT_AVAILABLE_AMD,
 627                                                 sizeof (available), &available,
 628                                                 NULL);
 629                 if (! available)
 630                         break;
 631
 632                 glGetPerfMonitorCounterDataAMD (monitor->id,
 633                                                 GL_PERFMON_RESULT_SIZE_AMD,
 634                                                 sizeof (result_size),
 635                                                 &result_size, NULL);
 636
 637                 result = xmalloc (result_size);
 638
 639                 glGetPerfMonitorCounterDataAMD (monitor->id,
 640                                                 GL_PERFMON_RESULT_AMD,
 641                                                 result_size, result,
 642                                                 &bytes_written);
 643
 644                 accumulate_program_metrics (monitor->op, result, result_size);
 645
 646                 free (result);
 647
 648                 ctx->monitor_head = monitor->next;
 649                 if (ctx->monitor_head == NULL)
 650                         ctx->monitor_tail = NULL;
 651
 652                 glDeletePerfMonitorsAMD (1, &monitor->id);
 653
 654                 free (monitor);
 655
 656                 ctx->monitors_in_flight--;
 657
 658                 monitor = ctx->monitor_head;
 659         }
 660 }
 661
 662
 663 void
 664 metrics_end_frame (void)
 665 {
 666         static int initialized = 0;
 667         static struct timeval tv_start, tv_now;
 668
 669         if (! initialized) {
 670                 gettimeofday (&tv_start, NULL);
 671                 atexit (metrics_exit);
 672                 if (getenv ("FIPS_VERBOSE"))
 673                         verbose = 1;
 674                 initialized = 1;
 675         }
 676
 677         frames++;
 678
 679         metrics_collect_available ();
 680
 681         if (frames % 15 == 0) {
 682                 double fps;
 683
 684                 gettimeofday (&tv_now, NULL);
 685
 686                 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
 687                                          (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
 688
 689                 printf("FPS: %.3f\n", fps);
 690
 691                 print_program_metrics ();
 692         }
 693 }