#include "fips-dispatch-gl.h"
#include "metrics.h"
+#include "context.h"
+#include "metrics-info.h"
#include "xmalloc.h"
+int frames;
+int verbose;
+
+#define MAX_MONITORS_IN_FLIGHT 1000
+
/* Timer query */
typedef struct timer_query
{
struct monitor *next;
} monitor_t;
-
typedef struct op_metrics
{
/* This happens to also be the index into the
- * ctx->op_metrics array currently
+ * metrics->op_metrics array currently
*/
metrics_op_t op;
double time_ns;
double **counters;
} op_metrics_t;
-typedef struct metrics_group_info
-{
- GLuint id;
- char *name;
-
- GLuint num_counters;
- GLuint max_active_counters;
-
- GLuint *counter_ids;
- char **counter_names;
- GLuint *counter_types;
-
-} metrics_group_info_t;
-
-typedef struct shader_stage_info
-{
- char *name;
-
- GLuint active_group_index;
- GLuint active_counter_index;
-
- GLuint stall_group_index;
- GLuint stall_counter_index;
-
-} shader_stage_info_t;
-
-typedef struct metrics_info
+struct metrics
{
- int initialized;
-
- unsigned num_groups;
- metrics_group_info_t *groups;
-
- unsigned num_shader_stages;
- shader_stage_info_t *stages;
-
-} metrics_info_t;
-
-typedef struct context
-{
- metrics_info_t metrics_info;
+ /* 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;
-} context_t;
-
-/* FIXME: Need a map from integers to context objects and track the
- * current context with glXMakeContextCurrent, eglMakeCurrent, etc. */
-
-context_t current_context;
-
-int frames;
-int verbose;
+};
-static void
-metrics_group_info_init (metrics_group_info_t *group, GLuint id)
+metrics_t *
+metrics_create (metrics_info_t *info)
{
- GLsizei length;
- unsigned i;
-
- group->id = id;
-
- /* Get name */
- glGetPerfMonitorGroupStringAMD (id, 0, &length, NULL);
-
- group->name = xmalloc (length + 1);
+ metrics_t *metrics;
- glGetPerfMonitorGroupStringAMD (id, length + 1, NULL, group->name);
+ metrics = xmalloc (sizeof (metrics_t));
- /* Get number of counters */
- group->num_counters = 0;
- group->max_active_counters = 0;
- glGetPerfMonitorCountersAMD (group->id,
- (int *) &group->num_counters,
- (int *) &group->max_active_counters,
- 0, NULL);
+ metrics->info = info;
- /* Get counter numbers */
- group->counter_ids = xmalloc (group->num_counters * sizeof (GLuint));
+ metrics->op = 0;
- glGetPerfMonitorCountersAMD (group->id, NULL, NULL,
- group->num_counters,
- group->counter_ids);
+ metrics->timer_begun_id = 0;
- /* Get counter names */
- group->counter_names = xmalloc (group->num_counters * sizeof (char *));
- group->counter_types = xmalloc (group->num_counters * sizeof (GLuint));
+ metrics->timer_head = NULL;
+ metrics->timer_tail = NULL;
- for (i = 0; i < group->num_counters; i++) {
- glGetPerfMonitorCounterInfoAMD (group->id,
- group->counter_ids[i],
- GL_COUNTER_TYPE_AMD,
- &group->counter_types[i]);
+ metrics->monitor_begun_id = 0;
- glGetPerfMonitorCounterStringAMD (group->id,
- group->counter_ids[i],
- 0, &length, NULL);
+ metrics->monitor_head = NULL;
+ metrics->monitor_tail = NULL;
- group->counter_names[i] = xmalloc (length + 1);
+ metrics->monitors_in_flight = 0;
- glGetPerfMonitorCounterStringAMD (group->id,
- group->counter_ids[i],
- length + 1, NULL,
- group->counter_names[i]);
- }
-}
-
-static void
-metrics_group_info_fini (metrics_group_info_t *group)
-{
- unsigned i;
+ metrics->num_op_metrics = 0;
+ metrics->op_metrics = NULL;
- for (i = 0; i < group->num_counters; i++)
- free (group->counter_names[i]);
-
- free (group->counter_types);
- free (group->counter_names);
- free (group->counter_ids);
-
- free (group->name);
-}
-
-/* A helper function, part of metrics_info_init below. */
-
-typedef enum {
- SHADER_ACTIVE,
- SHADER_STALL
-} shader_phase_t;
-
-static void
-_add_shader_stage (metrics_info_t *info, const char *name,
- GLuint group_index, GLuint counter_index,
- shader_phase_t phase)
-{
- shader_stage_info_t *stage;
- char *stage_name, *space;
- unsigned i;
-
- stage_name = xstrdup (name);
-
- /* Terminate the stage name at the first space.
- *
- * This is valid for counter names such as:
- *
- * "Vertex Shader Active Time"
- * or
- * "Vertex Shader Stall Time - Core Stall"
- */
- space = strchr (stage_name, ' ');
- if (space)
- *space = '\0';
-
- /* Look for an existing stage of the given name. */
- stage = NULL;
-
- for (i = 0; i < info->num_shader_stages; i++) {
- if (strcmp (info->stages[i].name, stage_name) == 0) {
- stage = &info->stages[i];
- break;
- }
- }
-
- if (stage == NULL) {
- info->num_shader_stages++;
- info->stages = xrealloc (info->stages,
- info->num_shader_stages *
- sizeof (shader_stage_info_t));
- stage = &info->stages[info->num_shader_stages - 1];
- stage->name = xstrdup (stage_name);
- stage->active_group_index = 0;
- stage->active_counter_index = 0;
- stage->stall_group_index = 0;
- stage->stall_counter_index = 0;
- }
-
- if (phase == SHADER_ACTIVE) {
- stage->active_group_index = group_index;
- stage->active_counter_index = counter_index;
- } else {
- stage->stall_group_index = group_index;
- stage->stall_counter_index = counter_index;
- }
-
- free (stage_name);
+ return metrics;
}
void
-metrics_info_init (void)
+metrics_fini (metrics_t *metrics)
{
- unsigned i, j;
- GLuint *group_ids;
- metrics_info_t *info = ¤t_context.metrics_info;
-
- glGetPerfMonitorGroupsAMD ((int *) &info->num_groups, 0, NULL);
-
- group_ids = xmalloc (info->num_groups * sizeof (GLuint));
-
- glGetPerfMonitorGroupsAMD (NULL, info->num_groups, group_ids);
-
- info->groups = xmalloc (info->num_groups * sizeof (metrics_group_info_t));
-
- for (i = 0; i < info->num_groups; i++)
- metrics_group_info_init (&info->groups[i], group_ids[i]);
-
- free (group_ids);
-
- /* Identify each shader stage (by looking at
- * performance-counter names for specific patterns) and
- * initialize structures referring to the corresponding
- * counter numbers for each stage. */
- info->num_shader_stages = 0;
- info->stages = NULL;
-
- for (i = 0; i < info->num_groups; i++) {
- metrics_group_info_t *group = &info->groups[i];
- for (j = 0; j < group->num_counters; j++) {
- char *name = group->counter_names[j];
- if (strstr (name, "Shader Active Time")) {
- _add_shader_stage (info, name, i, j,
- SHADER_ACTIVE);
- }
- if (strstr (name, "Shader Stall Time")) {
- _add_shader_stage (info, name, i, j,
- SHADER_STALL);
- }
- }
- }
-
- info->initialized = 1;
-}
-
-void
-metrics_info_fini (void)
-{
- context_t *ctx = ¤t_context;
- metrics_info_t *info = &ctx->metrics_info;
- unsigned i;
timer_query_t *timer, *timer_next;
monitor_t *monitor, *monitor_next;
- if (! info->initialized)
- return;
+ /* 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 = ctx->timer_head;
+ for (timer = metrics->timer_head;
timer;
timer = timer_next)
{
+ glDeleteQueries (1, &timer->id);
timer_next = timer->next;
free (timer);
}
- ctx->timer_head = NULL;
- ctx->timer_tail = NULL;
+ metrics->timer_head = NULL;
+ metrics->timer_tail = NULL;
- for (monitor = ctx->monitor_head;
+ 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);
}
- ctx->monitor_head = NULL;
- ctx->monitor_tail = NULL;
+ metrics->monitor_head = NULL;
+ metrics->monitor_tail = NULL;
- for (i = 0; i < info->num_groups; i++)
- metrics_group_info_fini (&info->groups[i]);
-
- free (info->groups);
- info->groups = NULL;
-
- for (i = 0; i < info->num_shader_stages; i++)
- free (info->stages[i].name);
+ metrics->monitors_in_flight = 0;
+}
- free (info->stages);
- info->stages = NULL;
+void
+metrics_destroy (metrics_t *metrics)
+{
+ metrics_fini (metrics);
- info->initialized = 0;
+ free (metrics);
}
static const char *
}
void
-metrics_counter_start (void)
+metrics_counter_start (metrics_t *metrics)
{
- context_t *ctx = ¤t_context;
- timer_query_t *timer;
- monitor_t *monitor;
unsigned i;
- /* Create new timer query, add to list */
- timer = xmalloc (sizeof (timer_query_t));
-
- timer->op = ctx->op;
- timer->next = NULL;
-
- if (ctx->timer_tail) {
- ctx->timer_tail->next = timer;
- ctx->timer_tail = timer;
- } else {
- ctx->timer_tail = timer;
- ctx->timer_head = timer;
- }
-
- /* Create a new performance-monitor query */
- monitor = xmalloc (sizeof (monitor_t));
-
- monitor->op = ctx->op;
- 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;
- }
-
/* Initialize the timer_query and monitor objects */
- glGenQueries (1, &timer->id);
+ glGenQueries (1, &metrics->timer_begun_id);
- glGenPerfMonitorsAMD (1, &monitor->id);
+ 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(monitor->id,
+ glSelectPerfMonitorCountersAMD(metrics->monitor_begun_id,
GL_TRUE, group->id,
num_counters,
group->counter_ids);
}
/* Start the queries */
- glBeginQuery (GL_TIME_ELAPSED, timer->id);
+ glBeginQuery (GL_TIME_ELAPSED, metrics->timer_begun_id);
- glBeginPerfMonitorAMD (monitor->id);
+ glBeginPerfMonitorAMD (metrics->monitor_begun_id);
}
void
-metrics_counter_stop (void)
+metrics_counter_stop (metrics_t *metrics)
{
+ timer_query_t *timer;
+ monitor_t *monitor;
+
+ /* Stop the current timer and monitor. */
glEndQuery (GL_TIME_ELAPSED);
- glEndPerfMonitorAMD (current_context.monitor_tail->id);
+ 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 = metrics->op;
+ timer->id = metrics->timer_begun_id;
+ timer->next = NULL;
+
+ if (metrics->timer_tail) {
+ metrics->timer_tail->next = timer;
+ metrics->timer_tail = timer;
+ } else {
+ metrics->timer_tail = timer;
+ metrics->timer_head = timer;
+ }
+
+ /* Create a new performance-monitor query */
+ monitor = xmalloc (sizeof (monitor_t));
+
+ monitor->op = metrics->op;
+ monitor->id = metrics->monitor_begun_id;
+ monitor->next = NULL;
+
+ if (metrics->monitor_tail) {
+ metrics->monitor_tail->next = monitor;
+ metrics->monitor_tail = monitor;
+ } else {
+ metrics->monitor_tail = monitor;
+ metrics->monitor_head = monitor;
+ }
+
+ 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 (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)
{
- current_context.op = op;
+ metrics->op = op;
}
metrics_op_t
-metrics_get_current_op (void)
+metrics_get_current_op (metrics_t *metrics)
{
- return current_context.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 = ¤t_context;
- 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)
{
- op_metrics_t *metrics;
+ op_metrics_t *op_metrics;
- metrics = ctx_get_op_metrics (¤t_context, 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 = ¤t_context;
- 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;
+ num_sorted = metrics->num_op_metrics * num_shader_stages;
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;
}
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 */
-static void
-metrics_exit (void)
-{
- if (verbose)
- printf ("fips: terminating\n");
-
- metrics_info_fini ();
-}
-
-
void
-metrics_end_frame (void)
+metrics_collect_available (metrics_t *metrics)
{
- context_t *ctx = ¤t_context;
- static int initialized = 0;
- static struct timeval tv_start, tv_now;
-
- if (! initialized) {
- gettimeofday (&tv_start, NULL);
- atexit (metrics_exit);
- if (getenv ("FIPS_VERBOSE"))
- verbose = 1;
- initialized = 1;
- }
-
- frames++;
- gettimeofday (&tv_now, NULL);
-
/* 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;
}
/* 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);
- monitor = ctx->monitor_head;
+
+ metrics->monitors_in_flight--;
+
+ monitor = metrics->monitor_head;
}
+}
+
- if (frames % 60 == 0) {
+void
+metrics_end_frame (metrics_t *metrics)
+{
+ static int initialized = 0;
+ static struct timeval tv_start, tv_now;
+
+ if (! initialized) {
+ gettimeofday (&tv_start, NULL);
+ if (getenv ("FIPS_VERBOSE"))
+ verbose = 1;
+ initialized = 1;
+ }
+
+ frames++;
+
+ metrics_collect_available (metrics);
+
+ if (frames % 15 == 0) {
double fps;
+ gettimeofday (&tv_now, NULL);
+
fps = (double) frames / (tv_now.tv_sec - tv_start.tv_sec +
(tv_now.tv_usec - tv_start.tv_usec) / 1.0e6);
printf("FPS: %.3f\n", fps);
- print_program_metrics ();
+ print_program_metrics (metrics);
}
}