#include "fips-dispatch-gl.h"
#include "metrics.h"
+#include "context.h"
+#include "metrics-info.h"
#include "xmalloc.h"
-/* 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
- * ctx->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
-{
- 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;
-
- metrics_op_t op;
-
- timer_query_t *timer_head;
- timer_query_t *timer_tail;
-
- monitor_t *monitor_head;
- monitor_t *monitor_tail;
-
- 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)
-{
- GLsizei length;
- unsigned i;
-
- group->id = id;
-
- /* Get name */
- glGetPerfMonitorGroupStringAMD (id, 0, &length, NULL);
-
- group->name = xmalloc (length + 1);
-
- glGetPerfMonitorGroupStringAMD (id, length + 1, NULL, group->name);
-
- /* 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);
-
- /* Get counter numbers */
- group->counter_ids = xmalloc (group->num_counters * sizeof (GLuint));
-
- glGetPerfMonitorCountersAMD (group->id, NULL, NULL,
- group->num_counters,
- group->counter_ids);
-
- /* Get counter names */
- group->counter_names = xmalloc (group->num_counters * sizeof (char *));
- group->counter_types = xmalloc (group->num_counters * sizeof (GLuint));
-
- for (i = 0; i < group->num_counters; i++) {
- glGetPerfMonitorCounterInfoAMD (group->id,
- group->counter_ids[i],
- GL_COUNTER_TYPE_AMD,
- &group->counter_types[i]);
-
- glGetPerfMonitorCounterStringAMD (group->id,
- group->counter_ids[i],
- 0, &length, NULL);
-
- group->counter_names[i] = xmalloc (length + 1);
-
- 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;
-
- 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);
-}
-
-void
-metrics_info_init (void)
-{
- 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;
-
- for (timer = ctx->timer_head;
- timer;
- timer = timer_next)
- {
- timer_next = timer->next;
- free (timer);
- }
- ctx->timer_head = NULL;
- ctx->timer_tail = NULL;
-
- for (monitor = ctx->monitor_head;
- monitor;
- monitor = monitor_next)
- {
- monitor_next = monitor->next;
- free (monitor);
- }
- ctx->monitor_head = NULL;
- ctx->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);
-
- free (info->stages);
- info->stages = NULL;
-
- info->initialized = 0;
-}
+#define MAX_MONITORS_IN_FLIGHT 1000
static const char *
metrics_op_string (metrics_op_t op)
void
metrics_counter_start (void)
{
- context_t *ctx = ¤t_context;
- timer_query_t *timer;
- monitor_t *monitor;
+ context_t *ctx = context_get_current ();
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, &ctx->timer_begun_id);
- glGenPerfMonitorsAMD (1, &monitor->id);
+ glGenPerfMonitorsAMD (1, &ctx->monitor_begun_id);
for (i = 0; i < ctx->metrics_info.num_groups; i++)
{
}
- glSelectPerfMonitorCountersAMD(monitor->id,
+ glSelectPerfMonitorCountersAMD(ctx->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, ctx->timer_begun_id);
- glBeginPerfMonitorAMD (monitor->id);
+ glBeginPerfMonitorAMD (ctx->monitor_begun_id);
}
void
metrics_counter_stop (void)
{
+ context_t *ctx = context_get_current ();
+ timer_query_t *timer;
+ monitor_t *monitor;
+
+ /* Stop the current timer and monitor. */
glEndQuery (GL_TIME_ELAPSED);
- glEndPerfMonitorAMD (current_context.monitor_tail->id);
+ glEndPerfMonitorAMD (ctx->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->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->id = ctx->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;
+ }
+
+ ctx->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 ();
}
void
metrics_set_current_op (metrics_op_t op)
{
- current_context.op = op;
+ context_t *ctx = context_get_current ();
+
+ ctx->op = op;
}
metrics_op_t
metrics_get_current_op (void)
{
- return current_context.op;
+ context_t *ctx = context_get_current ();
+
+ return ctx->op;
}
static void
(var) = *((typeof(var) *) p); \
p += sizeof(var);
- context_t *ctx = ¤t_context;
+ context_t *ctx = context_get_current ();
metrics_info_t *info = &ctx->metrics_info;
op_metrics_t *metrics = ctx_get_op_metrics (ctx, op);
unsigned char *p = (unsigned char *) result;
static void
accumulate_program_time (metrics_op_t op, unsigned time_ns)
{
+ context_t *ctx = context_get_current ();
op_metrics_t *metrics;
- metrics = ctx_get_op_metrics (¤t_context, op);
+ metrics = ctx_get_op_metrics (ctx, op);
metrics->time_ns += time_ns;
}
printf (" ");
}
- printf (" %cS:", per_stage->stage->name[0]);
+
+ if (per_stage->stage)
+ printf (" %cS:", per_stage->stage->name[0]);
+ else
+ printf (" :");
printf ("\t%7.2f ms (%4.1f%%)",
per_stage->time_ns / 1e6,
static void
print_program_metrics (void)
{
- context_t *ctx = ¤t_context;
+ context_t *ctx = context_get_current ();
metrics_info_t *info = &ctx->metrics_info;
unsigned num_shader_stages = info->num_shader_stages;
per_stage_metrics_t *sorted, *per_stage;
per_stage = &sorted[i * num_shader_stages + j];
per_stage->metrics = op;
- per_stage->stage = &info->stages[j];
- if (op_cycles)
+
+ if (op_cycles) {
+ per_stage->stage = &info->stages[j];
per_stage->time_ns = op->time_ns * (stage_cycles / op_cycles);
- else
- per_stage->time_ns = 0.0;
- if (stage_cycles)
+ } else {
+ /* If we don't have any per-stage cycle counts
+ * for this operation, then use the first
+ * stage as a placeholder for all the time,
+ * but NULL-ify the stage info so that the
+ * report doesn't lie about this time being
+ * from any particular stage. */
+ per_stage->stage = NULL;
+ if (j == 0) {
+ per_stage->time_ns = op->time_ns;
+ } else {
+ per_stage->time_ns = 0.0;
+ }
+ }
+
+ if (stage_cycles) {
per_stage->active = active_cycles / stage_cycles;
- else
+ } else {
per_stage->active = 0.0;
+ }
}
}
free (sorted);
}
-/* Called at program exit */
+/* 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");
- metrics_info_fini ();
-}
+ if (! info->initialized)
+ return;
+ for (timer = ctx->timer_head;
+ timer;
+ timer = timer_next)
+ {
+ timer_next = timer->next;
+ free (timer);
+ }
-void
-metrics_end_frame (void)
-{
- context_t *ctx = ¤t_context;
- static int initialized = 0;
- static struct timeval tv_start, tv_now;
+ for (monitor = ctx->monitor_head;
+ monitor;
+ monitor = monitor_next)
+ {
+ monitor_next = monitor->next;
+ free (monitor);
+ }
- if (! initialized) {
- gettimeofday (&tv_start, NULL);
- atexit (metrics_exit);
- if (getenv ("FIPS_VERBOSE"))
- verbose = 1;
- initialized = 1;
+ 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);
}
- frames++;
- gettimeofday (&tv_now, NULL);
+ 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)
+{
+ context_t *ctx = context_get_current ();
/* Consume all timer queries that are ready. */
timer_query_t *timer = ctx->timer_head;
glDeletePerfMonitorsAMD (1, &monitor->id);
free (monitor);
+
+ ctx->monitors_in_flight--;
+
monitor = ctx->monitor_head;
}
+}
- if (frames % 60 == 0) {
+
+void
+metrics_end_frame (void)
+{
+ 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++;
+
+ metrics_collect_available ();
+
+ 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);