#define MAX_MONITORS_IN_FLIGHT 1000
+/* Timer query */
+typedef struct timer_query
+{
+ unsigned id;
+
+ metrics_op_t op;
+ struct timer_query *next;
+} timer_query_t;
+
+/* Performance-monitor query */
+typedef struct monitor
+{
+ unsigned id;
+
+ metrics_op_t op;
+ struct monitor *next;
+} monitor_t;
+
+typedef struct op_metrics
+{
+ /* This happens to also be the index into the
+ * metrics->op_metrics array currently
+ */
+ metrics_op_t op;
+ double time_ns;
+
+ double **counters;
+} op_metrics_t;
+
+struct metrics
+{
+ /* Description of all available peformance counters, counter
+ * groups, their names and IDs, etc. */
+ metrics_info_t *info;
+
+ /* The current operation being measured. */
+ metrics_op_t op;
+
+ /* GL_TIME_ELAPSED query for which glEndQuery has not yet
+ * been called. */
+ unsigned timer_begun_id;
+
+ /* GL_TIME_ELAPSED queries for which glEndQuery has been
+ * called, (but results have not yet been queried). */
+ timer_query_t *timer_head;
+ timer_query_t *timer_tail;
+
+ /* Performance monitor for which glEndPerfMonitorAMD has not
+ * yet been called. */
+ unsigned monitor_begun_id;
+
+ /* Performance monitors for which glEndPerfMonitorAMD has
+ * been called, (but results have not yet been queried). */
+ monitor_t *monitor_head;
+ monitor_t *monitor_tail;
+
+ int monitors_in_flight;
+
+ unsigned num_op_metrics;
+ op_metrics_t *op_metrics;
+};
+
+metrics_t *
+metrics_create (metrics_info_t *info)
+{
+ metrics_t *metrics;
+
+ metrics = xmalloc (sizeof (metrics_t));
+
+ metrics->info = info;
+
+ metrics->op = 0;
+
+ metrics->timer_begun_id = 0;
+
+ metrics->timer_head = NULL;
+ metrics->timer_tail = NULL;
+
+ metrics->monitor_begun_id = 0;
+
+ metrics->monitor_head = NULL;
+ metrics->monitor_tail = NULL;
+
+ metrics->monitors_in_flight = 0;
+
+ metrics->num_op_metrics = 0;
+ metrics->op_metrics = NULL;
+
+ return metrics;
+}
+
+void
+metrics_fini (metrics_t *metrics)
+{
+ timer_query_t *timer, *timer_next;
+ monitor_t *monitor, *monitor_next;
+
+ /* Discard and cleanup any outstanding queries. */
+ if (metrics->timer_begun_id) {
+ glEndQuery (GL_TIME_ELAPSED);
+ glDeleteQueries (1, &metrics->timer_begun_id);
+ metrics->timer_begun_id = 0;
+ }
+
+ for (timer = metrics->timer_head;
+ timer;
+ timer = timer_next)
+ {
+ glDeleteQueries (1, &timer->id);
+ timer_next = timer->next;
+ free (timer);
+ }
+ metrics->timer_head = NULL;
+ metrics->timer_tail = NULL;
+
+ if (metrics->info->have_perfmon) {
+
+ if (metrics->monitor_begun_id) {
+ glEndPerfMonitorAMD (metrics->monitor_begun_id);
+ glDeletePerfMonitorsAMD (1, &metrics->monitor_begun_id);
+ metrics->monitor_begun_id = 0;
+ }
+
+ for (monitor = metrics->monitor_head;
+ monitor;
+ monitor = monitor_next)
+ {
+ glDeletePerfMonitorsAMD (1, &monitor->id);
+ monitor_next = monitor->next;
+ free (monitor);
+ }
+ metrics->monitor_head = NULL;
+ metrics->monitor_tail = NULL;
+
+ }
+
+ metrics->monitors_in_flight = 0;
+}
+
+void
+metrics_destroy (metrics_t *metrics)
+{
+ metrics_fini (metrics);
+
+ free (metrics);
+}
+
static const char *
metrics_op_string (metrics_op_t op)
{
}
void
-metrics_counter_start (void)
+metrics_counter_start (metrics_t *metrics)
{
- context_t *ctx = context_get_current ();
unsigned i;
- /* Initialize the timer_query and monitor objects */
- glGenQueries (1, &ctx->timer_begun_id);
+ /* Initialize the timer_query object. */
+ glGenQueries (1, &metrics->timer_begun_id);
+
+ /* Most everything else in this function is
+ * performance-monitor related. If we don't have that
+ * extension, just start the timer query and be done. */
+ if (! metrics->info->have_perfmon) {
+ glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
+ return;
+ }
- glGenPerfMonitorsAMD (1, &ctx->monitor_begun_id);
+ /* Initialize the performance-monitor object */
+ glGenPerfMonitorsAMD (1, &metrics->monitor_begun_id);
- for (i = 0; i < ctx->metrics_info.num_groups; i++)
+ for (i = 0; i < metrics->info->num_groups; i++)
{
metrics_group_info_t *group;
int num_counters;
- group = &ctx->metrics_info.groups[i];
+ group = &metrics->info->groups[i];
num_counters = group->num_counters;
if (group->max_active_counters < group->num_counters)
}
- glSelectPerfMonitorCountersAMD(ctx->monitor_begun_id,
+ glSelectPerfMonitorCountersAMD(metrics->monitor_begun_id,
GL_TRUE, group->id,
num_counters,
group->counter_ids);
}
/* Start the queries */
- glBeginQuery (GL_TIME_ELAPSED, ctx->timer_begun_id);
+ glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
- glBeginPerfMonitorAMD (ctx->monitor_begun_id);
+ glBeginPerfMonitorAMD (metrics->monitor_begun_id);
}
void
-metrics_counter_stop (void)
+metrics_counter_stop (metrics_t *metrics)
{
- context_t *ctx = context_get_current ();
timer_query_t *timer;
monitor_t *monitor;
/* Stop the current timer and monitor. */
glEndQuery (GL_TIME_ELAPSED);
- glEndPerfMonitorAMD (ctx->monitor_begun_id);
+
+ if (metrics->info->have_perfmon)
+ glEndPerfMonitorAMD (metrics->monitor_begun_id);
/* Add these IDs to our lists of outstanding queries and
* monitors so the results can be collected later. */
timer = xmalloc (sizeof (timer_query_t));
- timer->op = ctx->op;
- timer->id = ctx->timer_begun_id;
+ timer->op = metrics->op;
+ timer->id = metrics->timer_begun_id;
timer->next = NULL;
- if (ctx->timer_tail) {
- ctx->timer_tail->next = timer;
- ctx->timer_tail = timer;
+ if (metrics->timer_tail) {
+ metrics->timer_tail->next = timer;
+ metrics->timer_tail = timer;
} else {
- ctx->timer_tail = timer;
- ctx->timer_head = timer;
+ metrics->timer_tail = timer;
+ metrics->timer_head = timer;
}
- /* Create a new performance-monitor query */
- monitor = xmalloc (sizeof (monitor_t));
+ if (metrics->info->have_perfmon) {
+ /* Create a new performance-monitor query */
+ monitor = xmalloc (sizeof (monitor_t));
- monitor->op = ctx->op;
- monitor->id = ctx->monitor_begun_id;
- monitor->next = NULL;
+ monitor->op = metrics->op;
+ monitor->id = metrics->monitor_begun_id;
+ monitor->next = NULL;
- if (ctx->monitor_tail) {
- ctx->monitor_tail->next = monitor;
- ctx->monitor_tail = monitor;
- } else {
- ctx->monitor_tail = monitor;
- ctx->monitor_head = monitor;
+ if (metrics->monitor_tail) {
+ metrics->monitor_tail->next = monitor;
+ metrics->monitor_tail = monitor;
+ } else {
+ metrics->monitor_tail = monitor;
+ metrics->monitor_head = monitor;
+ }
}
- ctx->monitors_in_flight++;
+ metrics->monitors_in_flight++;
/* Avoid being a resource hog and collect outstanding results
* once we have sent off a large number of
* queries. (Presumably, many of the outstanding queries are
* available by now.)
*/
- if (ctx->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
- metrics_collect_available ();
+ if (metrics->monitors_in_flight > MAX_MONITORS_IN_FLIGHT)
+ metrics_collect_available (metrics);
}
void
-metrics_set_current_op (metrics_op_t op)
+metrics_set_current_op (metrics_t *metrics, metrics_op_t op)
{
- context_t *ctx = context_get_current ();
-
- ctx->op = op;
+ metrics->op = op;
}
metrics_op_t
-metrics_get_current_op (void)
+metrics_get_current_op (metrics_t *metrics)
{
- context_t *ctx = context_get_current ();
-
- return ctx->op;
+ return metrics->op;
}
static void
-op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
+op_metrics_init (metrics_info_t *info, op_metrics_t *metrics, metrics_op_t op)
{
- metrics_info_t *info = &ctx->metrics_info;
unsigned i, j;
metrics->op = op;
}
static op_metrics_t *
-ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
+_get_op_metrics (metrics_t *metrics, metrics_op_t op)
{
unsigned i;
- if (op >= ctx->num_op_metrics)
+ if (op >= metrics->num_op_metrics)
{
- ctx->op_metrics = realloc (ctx->op_metrics,
- (op + 1) * sizeof (op_metrics_t));
- for (i = ctx->num_op_metrics; i < op + 1; i++)
- op_metrics_init (ctx, &ctx->op_metrics[i], i);
+ metrics->op_metrics = realloc (metrics->op_metrics,
+ (op + 1) * sizeof (op_metrics_t));
+ for (i = metrics->num_op_metrics; i < op + 1; i++)
+ op_metrics_init (metrics->info, &metrics->op_metrics[i], i);
- ctx->num_op_metrics = op + 1;
+ metrics->num_op_metrics = op + 1;
}
- return &ctx->op_metrics[op];
+ return &metrics->op_metrics[op];
}
static void
-accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
+accumulate_program_metrics (metrics_t *metrics, metrics_op_t op,
+ GLuint *result, GLuint size)
{
#define CONSUME(var) \
if (p + sizeof(var) > ((unsigned char *) result) + size) \
(var) = *((typeof(var) *) p); \
p += sizeof(var);
- context_t *ctx = context_get_current ();
- metrics_info_t *info = &ctx->metrics_info;
- op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
+ metrics_info_t *info = metrics->info;
+ op_metrics_t *op_metrics = _get_op_metrics (metrics, op);
unsigned char *p = (unsigned char *) result;
while (p < ((unsigned char *) result) + size)
break;
}
- metrics->counters[group_index][counter_index] += value;
+ op_metrics->counters[group_index][counter_index] += value;
}
}
static void
-accumulate_program_time (metrics_op_t op, unsigned time_ns)
+accumulate_program_time (metrics_t *metrics, metrics_op_t op, unsigned time_ns)
{
- context_t *ctx = context_get_current ();
- op_metrics_t *metrics;
+ op_metrics_t *op_metrics;
- metrics = ctx_get_op_metrics (ctx, op);
+ op_metrics = _get_op_metrics (metrics, op);
- metrics->time_ns += time_ns;
+ op_metrics->time_ns += time_ns;
}
typedef struct per_stage_metrics
}
static void
-print_per_stage_metrics (context_t *ctx,
+print_per_stage_metrics (metrics_t *metrics,
per_stage_metrics_t *per_stage,
double total)
{
- metrics_info_t *info = &ctx->metrics_info;
- op_metrics_t *metric = per_stage->metrics;
+ metrics_info_t *info = metrics->info;
+ op_metrics_t *op_metrics = per_stage->metrics;
metrics_group_info_t *group;
const char *op_string;
unsigned group_index, counter;
if (per_stage->time_ns == 0.0)
return;
- op_string = metrics_op_string (metric->op);
+ op_string = metrics_op_string (op_metrics->op);
printf ("%21s", op_string);
- if (metric->op >= METRICS_OP_SHADER) {
- printf (" %3d", metric->op - METRICS_OP_SHADER);
+ if (op_metrics->op >= METRICS_OP_SHADER) {
+ printf (" %3d", op_metrics->op - METRICS_OP_SHADER);
} else {
printf (" ");
if (_is_shader_stage_counter (info, group_index, counter))
continue;
- value = metric->counters[group_index][counter];
+ value = op_metrics->counters[group_index][counter];
if (value == 0.0)
continue;
printf ("%s: %.2f ", group->counter_names[counter],
}
static void
-print_program_metrics (void)
+print_program_metrics (metrics_t *metrics)
{
- context_t *ctx = context_get_current ();
- metrics_info_t *info = &ctx->metrics_info;
+ metrics_info_t *info = metrics->info;
unsigned num_shader_stages = info->num_shader_stages;
per_stage_metrics_t *sorted, *per_stage;
double total_time, op_cycles;
/* Make a sorted list of the per-stage operations by time
* used, and figure out the total so we can print percentages.
*/
- num_sorted = ctx->num_op_metrics * num_shader_stages;
+ if (num_shader_stages)
+ num_sorted = metrics->num_op_metrics * num_shader_stages;
+ else
+ num_sorted = metrics->num_op_metrics;
sorted = xmalloc (sizeof (*sorted) * num_sorted);
total_time = 0.0;
- for (i = 0; i < ctx->num_op_metrics; i++) {
+ for (i = 0; i < metrics->num_op_metrics; i++) {
- op = &ctx->op_metrics[i];
+ op = &metrics->op_metrics[i];
/* Accumulate total time across all ops. */
total_time += op->time_ns;
/* Also, find total cycles in all stages of this op. */
op_cycles = 0.0;
+ if (num_shader_stages == 0) {
+ per_stage = &sorted[i];
+ per_stage->metrics = op;
+ per_stage->stage = NULL;
+ per_stage->time_ns = op->time_ns;
+ per_stage->active = 0.0;
+ }
+
for (j = 0; j < num_shader_stages; j++) {
/* Active cycles */
group_index = info->stages[j].active_group_index;
}
qsort_r (sorted, num_sorted, sizeof (*sorted),
- time_compare, ctx->op_metrics);
+ time_compare, metrics->op_metrics);
for (i = 0; i < num_sorted; i++)
- print_per_stage_metrics (ctx, &sorted[i], total_time);
+ print_per_stage_metrics (metrics, &sorted[i], total_time);
free (sorted);
}
-/* Called at program exit.
- *
- * This is similar to metrics_info_fini, but only frees any used
- * memory. Notably, it does not call any OpenGL functions, (since the
- * OpenGL context no longer exists at program exit).
- */
-static void
-metrics_exit (void)
-{
- context_t *ctx = context_get_current ();
- metrics_info_t *info = &ctx->metrics_info;
- unsigned i, j;
- timer_query_t *timer, *timer_next;
- monitor_t *monitor, *monitor_next;
-
- if (verbose)
- printf ("fips: terminating\n");
-
- if (! info->initialized)
- return;
-
- for (timer = ctx->timer_head;
- timer;
- timer = timer_next)
- {
- timer_next = timer->next;
- free (timer);
- }
-
- for (monitor = ctx->monitor_head;
- monitor;
- monitor = monitor_next)
- {
- monitor_next = monitor->next;
- free (monitor);
- }
-
- for (i = 0; i < info->num_groups; i++) {
- metrics_group_info_t *group = &info->groups[i];
-
- for (j = 0; j < group->num_counters; i++)
- free (group->counter_names[j]);
-
- free (group->counter_types);
- free (group->counter_names);
- free (group->counter_ids);
-
- free (group->name);
- }
-
- free (info->groups);
-
- for (i = 0; i < info->num_shader_stages; i++)
- free (info->stages[i].name);
-
- free (info->stages);
-}
-
void
-metrics_collect_available (void)
+metrics_collect_available (metrics_t *metrics)
{
- context_t *ctx = context_get_current ();
-
/* Consume all timer queries that are ready. */
- timer_query_t *timer = ctx->timer_head;
+ timer_query_t *timer = metrics->timer_head;
while (timer) {
GLuint available, elapsed;
glGetQueryObjectuiv (timer->id,
GL_QUERY_RESULT, &elapsed);
- accumulate_program_time (timer->op, elapsed);
+ accumulate_program_time (metrics, timer->op, elapsed);
- ctx->timer_head = timer->next;
- if (ctx->timer_head == NULL)
- ctx->timer_tail = NULL;
+ metrics->timer_head = timer->next;
+ if (metrics->timer_head == NULL)
+ metrics->timer_tail = NULL;
glDeleteQueries (1, &timer->id);
free (timer);
- timer = ctx->timer_head;
+ timer = metrics->timer_head;
}
+ if (! metrics->info->have_perfmon)
+ return;
+
/* And similarly for all performance monitors that are ready. */
- monitor_t *monitor = ctx->monitor_head;
+ monitor_t *monitor = metrics->monitor_head;
while (monitor) {
GLuint available, result_size, *result;
result_size, result,
&bytes_written);
- accumulate_program_metrics (monitor->op, result, result_size);
+ accumulate_program_metrics (metrics, monitor->op, result, result_size);
free (result);
- ctx->monitor_head = monitor->next;
- if (ctx->monitor_head == NULL)
- ctx->monitor_tail = NULL;
+ metrics->monitor_head = monitor->next;
+ if (metrics->monitor_head == NULL)
+ metrics->monitor_tail = NULL;
glDeletePerfMonitorsAMD (1, &monitor->id);
free (monitor);
- ctx->monitors_in_flight--;
+ metrics->monitors_in_flight--;
- monitor = ctx->monitor_head;
+ monitor = metrics->monitor_head;
}
}
+static void
+metrics_exit (void)
+{
+ if (verbose)
+ printf ("fips: terminating\n");
+}
void
-metrics_end_frame (void)
+metrics_end_frame (metrics_t *metrics)
{
static int initialized = 0;
static struct timeval tv_start, tv_now;
frames++;
- metrics_collect_available ();
+ metrics_collect_available (metrics);
if (frames % 15 == 0) {
double fps;
printf("FPS: %.3f\n", fps);
- print_program_metrics ();
+ print_program_metrics (metrics);
}
}