1 /* Copyright © 2013, Intel Corporation
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:
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
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
29 #include "fips-dispatch-gl.h"
35 typedef struct timer_query
40 struct timer_query *next;
43 /* Performance-monitor query */
44 typedef struct monitor
53 typedef struct op_metrics
55 /* This happens to also be the index into the
56 * ctx->op_metrics array currently
64 typedef struct metrics_group_info
70 GLuint max_active_counters;
74 GLuint *counter_types;
76 } metrics_group_info_t;
78 typedef struct shader_stage_info
82 GLuint active_group_index;
83 GLuint active_counter_index;
85 GLuint stall_group_index;
86 GLuint stall_counter_index;
88 } shader_stage_info_t;
90 typedef struct metrics_info
95 metrics_group_info_t *groups;
97 unsigned num_shader_stages;
98 shader_stage_info_t *stages;
102 typedef struct context
104 metrics_info_t metrics_info;
108 /* GL_TIME_ELAPSED query for which glEndQuery has not yet
110 unsigned timer_begun_id;
112 /* GL_TIME_ELAPSED queries for which glEndQuery has been
113 * called, (but results have not yet been queried). */
114 timer_query_t *timer_head;
115 timer_query_t *timer_tail;
117 /* Performance monitor for which glEndPerfMonitorAMD has not
118 * yet been called. */
119 unsigned monitor_begun_id;
121 /* Performance monitors for which glEndPerfMonitorAMD has
122 * been called, (but results have not yet been queried). */
123 monitor_t *monitor_head;
124 monitor_t *monitor_tail;
126 int monitors_in_flight;
128 unsigned num_op_metrics;
129 op_metrics_t *op_metrics;
132 /* FIXME: Need a map from integers to context objects and track the
133 * current context with glXMakeContextCurrent, eglMakeCurrent, etc. */
135 context_t current_context;
140 #define MAX_MONITORS_IN_FLIGHT 1000
143 metrics_collect_available (void);
146 metrics_group_info_init (metrics_group_info_t *group, GLuint id)
154 glGetPerfMonitorGroupStringAMD (id, 0, &length, NULL);
156 group->name = xmalloc (length + 1);
158 glGetPerfMonitorGroupStringAMD (id, length + 1, NULL, group->name);
160 /* Get number of counters */
161 group->num_counters = 0;
162 group->max_active_counters = 0;
163 glGetPerfMonitorCountersAMD (group->id,
164 (int *) &group->num_counters,
165 (int *) &group->max_active_counters,
168 /* Get counter numbers */
169 group->counter_ids = xmalloc (group->num_counters * sizeof (GLuint));
171 glGetPerfMonitorCountersAMD (group->id, NULL, NULL,
175 /* Get counter names */
176 group->counter_names = xmalloc (group->num_counters * sizeof (char *));
177 group->counter_types = xmalloc (group->num_counters * sizeof (GLuint));
179 for (i = 0; i < group->num_counters; i++) {
180 glGetPerfMonitorCounterInfoAMD (group->id,
181 group->counter_ids[i],
183 &group->counter_types[i]);
185 glGetPerfMonitorCounterStringAMD (group->id,
186 group->counter_ids[i],
189 group->counter_names[i] = xmalloc (length + 1);
191 glGetPerfMonitorCounterStringAMD (group->id,
192 group->counter_ids[i],
194 group->counter_names[i]);
199 metrics_group_info_fini (metrics_group_info_t *group)
203 for (i = 0; i < group->num_counters; i++)
204 free (group->counter_names[i]);
206 free (group->counter_types);
207 free (group->counter_names);
208 free (group->counter_ids);
213 /* A helper function, part of metrics_info_init below. */
221 _add_shader_stage (metrics_info_t *info, const char *name,
222 GLuint group_index, GLuint counter_index,
223 shader_phase_t phase)
225 shader_stage_info_t *stage;
226 char *stage_name, *space;
229 stage_name = xstrdup (name);
231 /* Terminate the stage name at the first space.
233 * This is valid for counter names such as:
235 * "Vertex Shader Active Time"
237 * "Vertex Shader Stall Time - Core Stall"
239 space = strchr (stage_name, ' ');
243 /* Look for an existing stage of the given name. */
246 for (i = 0; i < info->num_shader_stages; i++) {
247 if (strcmp (info->stages[i].name, stage_name) == 0) {
248 stage = &info->stages[i];
254 info->num_shader_stages++;
255 info->stages = xrealloc (info->stages,
256 info->num_shader_stages *
257 sizeof (shader_stage_info_t));
258 stage = &info->stages[info->num_shader_stages - 1];
259 stage->name = xstrdup (stage_name);
260 stage->active_group_index = 0;
261 stage->active_counter_index = 0;
262 stage->stall_group_index = 0;
263 stage->stall_counter_index = 0;
266 if (phase == SHADER_ACTIVE) {
267 stage->active_group_index = group_index;
268 stage->active_counter_index = counter_index;
270 stage->stall_group_index = group_index;
271 stage->stall_counter_index = counter_index;
278 metrics_info_init (void)
282 metrics_info_t *info = ¤t_context.metrics_info;
284 glGetPerfMonitorGroupsAMD ((int *) &info->num_groups, 0, NULL);
286 group_ids = xmalloc (info->num_groups * sizeof (GLuint));
288 glGetPerfMonitorGroupsAMD (NULL, info->num_groups, group_ids);
290 info->groups = xmalloc (info->num_groups * sizeof (metrics_group_info_t));
292 for (i = 0; i < info->num_groups; i++)
293 metrics_group_info_init (&info->groups[i], group_ids[i]);
297 /* Identify each shader stage (by looking at
298 * performance-counter names for specific patterns) and
299 * initialize structures referring to the corresponding
300 * counter numbers for each stage. */
301 info->num_shader_stages = 0;
304 for (i = 0; i < info->num_groups; i++) {
305 metrics_group_info_t *group = &info->groups[i];
306 for (j = 0; j < group->num_counters; j++) {
307 char *name = group->counter_names[j];
308 if (strstr (name, "Shader Active Time")) {
309 _add_shader_stage (info, name, i, j,
312 if (strstr (name, "Shader Stall Time")) {
313 _add_shader_stage (info, name, i, j,
319 info->initialized = 1;
323 metrics_info_fini (void)
325 context_t *ctx = ¤t_context;
326 metrics_info_t *info = &ctx->metrics_info;
328 timer_query_t *timer, *timer_next;
329 monitor_t *monitor, *monitor_next;
331 if (! info->initialized)
334 metrics_collect_available ();
336 if (ctx->timer_begun_id) {
337 glEndQuery (GL_TIME_ELAPSED);
338 glDeleteQueries (1, &ctx->timer_begun_id);
339 ctx->timer_begun_id = 0;
342 for (timer = ctx->timer_head;
346 glDeleteQueries (1, &timer->id);
347 timer_next = timer->next;
350 ctx->timer_head = NULL;
351 ctx->timer_tail = NULL;
353 if (ctx->monitor_begun_id) {
354 glEndPerfMonitorAMD (ctx->monitor_begun_id);
355 glDeletePerfMonitorsAMD (1, &ctx->monitor_begun_id);
356 ctx->monitor_begun_id = 0;
359 for (monitor = ctx->monitor_head;
361 monitor = monitor_next)
363 glDeletePerfMonitorsAMD (1, &monitor->id);
364 monitor_next = monitor->next;
367 ctx->monitor_head = NULL;
368 ctx->monitor_tail = NULL;
370 current_context.monitors_in_flight = 0;
372 for (i = 0; i < info->num_groups; i++)
373 metrics_group_info_fini (&info->groups[i]);
378 for (i = 0; i < info->num_shader_stages; i++)
379 free (info->stages[i].name);
384 info->initialized = 0;
388 metrics_op_string (metrics_op_t op)
390 if (op >= METRICS_OP_SHADER)
391 return "Shader program";
395 case METRICS_OP_ACCUM:
396 return "glAccum*(+)";
397 case METRICS_OP_BUFFER_DATA:
398 return "glBufferData(+)";
399 case METRICS_OP_BUFFER_SUB_DATA:
400 return "glCopyBufferSubData*";
401 case METRICS_OP_BITMAP:
403 case METRICS_OP_BLIT_FRAMEBUFFER:
404 return "glBlitFramebuffer*";
405 case METRICS_OP_CLEAR:
407 case METRICS_OP_CLEAR_BUFFER_DATA:
408 return "glCearBufferData(+)";
409 case METRICS_OP_CLEAR_TEX_IMAGE:
410 return "glClearTexImage(+)";
411 case METRICS_OP_COPY_PIXELS:
412 return "glCopyPixels";
413 case METRICS_OP_COPY_TEX_IMAGE:
414 return "glCopyTexImage(+)";
415 case METRICS_OP_DRAW_PIXELS:
416 return "glDrawPixels";
417 case METRICS_OP_GET_TEX_IMAGE:
418 return "glGetTexImage(+)";
419 case METRICS_OP_READ_PIXELS:
420 return "glReadPixels*";
421 case METRICS_OP_TEX_IMAGE:
422 return "glTexImage*(+)";
424 fprintf (stderr, "fips: Internal error: "
425 "Unknown metrics op value: %d\n", op);
433 metrics_counter_start (void)
435 context_t *ctx = ¤t_context;
438 /* Initialize the timer_query and monitor objects */
439 glGenQueries (1, &ctx->timer_begun_id);
441 glGenPerfMonitorsAMD (1, &ctx->monitor_begun_id);
443 for (i = 0; i < ctx->metrics_info.num_groups; i++)
445 metrics_group_info_t *group;
448 group = &ctx->metrics_info.groups[i];
450 num_counters = group->num_counters;
451 if (group->max_active_counters < group->num_counters)
453 fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
454 group->max_active_counters,
455 group->num_counters, i);
456 num_counters = group->max_active_counters;
460 glSelectPerfMonitorCountersAMD(ctx->monitor_begun_id,
466 /* Start the queries */
467 glBeginQuery (GL_TIME_ELAPSED, ctx->timer_begun_id);
469 glBeginPerfMonitorAMD (ctx->monitor_begun_id);
473 metrics_counter_stop (void)
475 context_t *ctx = ¤t_context;
476 timer_query_t *timer;
479 /* Stop the current timer and monitor. */
480 glEndQuery (GL_TIME_ELAPSED);
481 glEndPerfMonitorAMD (ctx->monitor_begun_id);
483 /* Add these IDs to our lists of outstanding queries and
484 * monitors so the results can be collected later. */
485 timer = xmalloc (sizeof (timer_query_t));
488 timer->id = ctx->timer_begun_id;
491 if (ctx->timer_tail) {
492 ctx->timer_tail->next = timer;
493 ctx->timer_tail = timer;
495 ctx->timer_tail = timer;
496 ctx->timer_head = timer;
499 /* Create a new performance-monitor query */
500 monitor = xmalloc (sizeof (monitor_t));
502 monitor->op = ctx->op;
503 monitor->id = ctx->monitor_begun_id;
504 monitor->next = NULL;
506 if (ctx->monitor_tail) {
507 ctx->monitor_tail->next = monitor;
508 ctx->monitor_tail = monitor;
510 ctx->monitor_tail = monitor;
511 ctx->monitor_head = monitor;
514 ctx->monitors_in_flight++;
516 /* Avoid being a resource hog and collect outstanding results
517 * once we have sent off a large number of
518 * queries. (Presumably, many of the outstanding queries are
521 if (ctx->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
522 metrics_collect_available ();
526 metrics_set_current_op (metrics_op_t op)
528 current_context.op = op;
532 metrics_get_current_op (void)
534 return current_context.op;
538 op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
540 metrics_info_t *info = &ctx->metrics_info;
544 metrics->time_ns = 0.0;
546 metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
548 for (i = 0; i < info->num_groups; i++) {
549 metrics->counters[i] = xmalloc (sizeof (double) *
550 info->groups[i].num_counters);
551 for (j = 0; j < info->groups[i].num_counters; j++)
552 metrics->counters[i][j] = 0.0;
556 static op_metrics_t *
557 ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
561 if (op >= ctx->num_op_metrics)
563 ctx->op_metrics = realloc (ctx->op_metrics,
564 (op + 1) * sizeof (op_metrics_t));
565 for (i = ctx->num_op_metrics; i < op + 1; i++)
566 op_metrics_init (ctx, &ctx->op_metrics[i], i);
568 ctx->num_op_metrics = op + 1;
571 return &ctx->op_metrics[op];
575 accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
577 #define CONSUME(var) \
578 if (p + sizeof(var) > ((unsigned char *) result) + size) \
580 fprintf (stderr, "Unexpected end-of-buffer while " \
581 "parsing results\n"); \
584 (var) = *((typeof(var) *) p); \
587 context_t *ctx = ¤t_context;
588 metrics_info_t *info = &ctx->metrics_info;
589 op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
590 unsigned char *p = (unsigned char *) result;
592 while (p < ((unsigned char *) result) + size)
594 GLuint group_id, group_index;
595 GLuint counter_id, counter_index;
596 metrics_group_info_t *group;
601 CONSUME (counter_id);
603 for (i = 0; i < info->num_groups; i++) {
604 if (info->groups[i].id == group_id)
608 assert (group_index < info->num_groups);
609 group = &info->groups[group_index];
611 for (i = 0; i < group->num_counters; i++) {
612 if (group->counter_ids[i] == counter_id)
616 assert (counter_index < group->num_counters);
618 switch (group->counter_types[counter_index])
621 uint64_t uint64_value;
623 case GL_UNSIGNED_INT:
624 CONSUME (uint_value);
627 case GL_UNSIGNED_INT64_AMD:
628 CONSUME (uint64_value);
629 value = uint64_value;
631 case GL_PERCENTAGE_AMD:
633 CONSUME (float_value);
637 fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
638 group->counter_types[counter_index]);
643 metrics->counters[group_index][counter_index] += value;
648 accumulate_program_time (metrics_op_t op, unsigned time_ns)
650 op_metrics_t *metrics;
652 metrics = ctx_get_op_metrics (¤t_context, op);
654 metrics->time_ns += time_ns;
657 typedef struct per_stage_metrics
659 op_metrics_t *metrics;
660 shader_stage_info_t *stage;
663 } per_stage_metrics_t;
666 _is_shader_stage_counter (metrics_info_t *info,
667 unsigned group_index,
668 unsigned counter_index)
670 shader_stage_info_t *stage;
673 for (i = 0; i < info->num_shader_stages; i++) {
674 stage = &info->stages[i];
676 if (stage->active_group_index == group_index &&
677 stage->active_counter_index == counter_index)
682 if (stage->stall_group_index == group_index &&
683 stage->stall_counter_index == counter_index)
693 print_per_stage_metrics (context_t *ctx,
694 per_stage_metrics_t *per_stage,
697 metrics_info_t *info = &ctx->metrics_info;
698 op_metrics_t *metric = per_stage->metrics;
699 metrics_group_info_t *group;
700 const char *op_string;
701 unsigned group_index, counter;
704 /* Don't print anything for stages with no alloted time. */
705 if (per_stage->time_ns == 0.0)
708 op_string = metrics_op_string (metric->op);
710 printf ("%21s", op_string);
712 if (metric->op >= METRICS_OP_SHADER) {
713 printf (" %3d", metric->op - METRICS_OP_SHADER);
719 if (per_stage->stage)
720 printf (" %cS:", per_stage->stage->name[0]);
724 printf ("\t%7.2f ms (%4.1f%%)",
725 per_stage->time_ns / 1e6,
726 per_stage->time_ns / total * 100);
728 if (per_stage->active)
729 printf (", %4.1f%% active", per_stage->active * 100);
733 /* I'm not seeing a lot of value printing the rest of these
734 * performance counters by default yet. Use --verbose to get
740 for (group_index = 0; group_index < info->num_groups; group_index++) {
741 group = &info->groups[group_index];
742 for (counter = 0; counter < group->num_counters; counter++) {
744 /* Don't print this counter value if it's a
745 * per-stage cycle counter, (which we have
746 * already accounted for). */
747 if (_is_shader_stage_counter (info, group_index, counter))
750 value = metric->counters[group_index][counter];
753 printf ("%s: %.2f ", group->counter_names[counter],
761 time_compare(const void *in_a, const void *in_b, void *arg unused)
763 const per_stage_metrics_t *a = in_a;
764 const per_stage_metrics_t *b = in_b;
767 if (a->time_ns < b->time_ns)
769 if (a->time_ns > b->time_ns)
775 op_metrics_reset (op_metrics_t *op)
777 context_t *ctx = ¤t_context;
778 metrics_info_t *info = &ctx->metrics_info;
779 unsigned group_index, counter;
780 metrics_group_info_t *group;
784 for (group_index = 0; group_index < info->num_groups; group_index++) {
785 group = &info->groups[group_index];
786 for (counter = 0; counter < group->num_counters; counter++)
788 op->counters[group_index][counter] = 0.0;
796 context_t *ctx = ¤t_context;
799 for (i = 0; i < ctx->num_op_metrics; i++) {
801 op_metrics_reset (&ctx->op_metrics[i]);
807 print_program_metrics (void)
809 context_t *ctx = ¤t_context;
810 metrics_info_t *info = &ctx->metrics_info;
811 unsigned num_shader_stages = info->num_shader_stages;
812 per_stage_metrics_t *sorted, *per_stage;
813 double total_time, op_cycles;
815 unsigned group_index, counter_index;
816 unsigned i, j, num_sorted;
818 /* Make a sorted list of the per-stage operations by time
819 * used, and figure out the total so we can print percentages.
821 num_sorted = ctx->num_op_metrics * num_shader_stages;
823 sorted = xmalloc (sizeof (*sorted) * num_sorted);
827 for (i = 0; i < ctx->num_op_metrics; i++) {
829 op = &ctx->op_metrics[i];
831 /* Accumulate total time across all ops. */
832 total_time += op->time_ns;
834 /* Also, find total cycles in all stages of this op. */
837 for (j = 0; j < num_shader_stages; j++) {
839 group_index = info->stages[j].active_group_index;
840 counter_index = info->stages[j].active_counter_index;
841 op_cycles += op->counters[group_index][counter_index];
844 group_index = info->stages[j].stall_group_index;
845 counter_index = info->stages[j].stall_counter_index;
846 op_cycles += op->counters[group_index][counter_index];
849 for (j = 0; j < num_shader_stages; j++) {
850 double active_cycles, stall_cycles, stage_cycles;
853 group_index = info->stages[j].active_group_index;
854 counter_index = info->stages[j].active_counter_index;
855 active_cycles = op->counters[group_index][counter_index];
858 group_index = info->stages[j].stall_group_index;
859 counter_index = info->stages[j].stall_counter_index;
860 stall_cycles = op->counters[group_index][counter_index];
862 stage_cycles = active_cycles + stall_cycles;
864 per_stage = &sorted[i * num_shader_stages + j];
865 per_stage->metrics = op;
868 per_stage->stage = &info->stages[j];
869 per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
871 /* If we don't have any per-stage cycle counts
872 * for this operation, then use the first
873 * stage as a placeholder for all the time,
874 * but NULL-ify the stage info so that the
875 * report doesn't lie about this time being
876 * from any particular stage. */
877 per_stage->stage = NULL;
879 per_stage->time_ns = op->time_ns;
881 per_stage->time_ns = 0.0;
886 per_stage->active = active_cycles / stage_cycles;
888 per_stage->active = 0.0;
893 qsort_r (sorted, num_sorted, sizeof (*sorted),
894 time_compare, ctx->op_metrics);
896 for (i = 0; i < num_sorted; i++)
897 print_per_stage_metrics (ctx, &sorted[i], total_time);
902 /* Called at program exit.
904 * This is similar to metrics_info_fini, but only frees any used
905 * memory. Notably, it does not call any OpenGL functions, (since the
906 * OpenGL context no longer exists at program exit).
911 context_t *ctx = ¤t_context;
912 metrics_info_t *info = &ctx->metrics_info;
914 timer_query_t *timer, *timer_next;
915 monitor_t *monitor, *monitor_next;
918 printf ("fips: terminating\n");
920 if (! info->initialized)
923 for (timer = ctx->timer_head;
927 timer_next = timer->next;
931 for (monitor = ctx->monitor_head;
933 monitor = monitor_next)
935 monitor_next = monitor->next;
939 for (i = 0; i < info->num_groups; i++)
940 metrics_group_info_fini (&info->groups[i]);
944 for (i = 0; i < info->num_shader_stages; i++)
945 free (info->stages[i].name);
951 metrics_collect_available (void)
953 context_t *ctx = ¤t_context;
955 /* Consume all timer queries that are ready. */
956 timer_query_t *timer = ctx->timer_head;
959 GLuint available, elapsed;
961 glGetQueryObjectuiv (timer->id,
962 GL_QUERY_RESULT_AVAILABLE, &available);
966 glGetQueryObjectuiv (timer->id,
967 GL_QUERY_RESULT, &elapsed);
969 accumulate_program_time (timer->op, elapsed);
971 ctx->timer_head = timer->next;
972 if (ctx->timer_head == NULL)
973 ctx->timer_tail = NULL;
975 glDeleteQueries (1, &timer->id);
978 timer = ctx->timer_head;
981 /* And similarly for all performance monitors that are ready. */
982 monitor_t *monitor = ctx->monitor_head;
985 GLuint available, result_size, *result;
988 glGetPerfMonitorCounterDataAMD (monitor->id,
989 GL_PERFMON_RESULT_AVAILABLE_AMD,
990 sizeof (available), &available,
995 glGetPerfMonitorCounterDataAMD (monitor->id,
996 GL_PERFMON_RESULT_SIZE_AMD,
997 sizeof (result_size),
1000 result = xmalloc (result_size);
1002 glGetPerfMonitorCounterDataAMD (monitor->id,
1003 GL_PERFMON_RESULT_AMD,
1004 result_size, result,
1007 accumulate_program_metrics (monitor->op, result, result_size);
1011 ctx->monitor_head = monitor->next;
1012 if (ctx->monitor_head == NULL)
1013 ctx->monitor_tail = NULL;
1015 glDeletePerfMonitorsAMD (1, &monitor->id);
1019 ctx->monitors_in_flight--;
1021 monitor = ctx->monitor_head;
1027 metrics_end_frame (void)
1029 static int initialized = 0;
1030 static struct timeval tv_start, tv_now;
1032 if (! initialized) {
1033 gettimeofday (&tv_start, NULL);
1034 atexit (metrics_exit);
1035 if (getenv ("FIPS_VERBOSE"))
1042 metrics_collect_available ();
1044 if (frames % 15 == 0) {
1047 gettimeofday (&tv_now, NULL);
1049 fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
1050 (tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
1052 printf("FPS: %.3f\n", fps);
1054 print_program_metrics ();