#include <stdio.h>
#include <stdlib.h>
-
+#include <assert.h>
#include <sys/time.h>
#include "fips-dispatch-gl.h"
#include "metrics.h"
+#include "context.h"
+#include "metrics-info.h"
#include "xmalloc.h"
-typedef struct counter
-{
- unsigned id;
-
- metrics_op_t op;
- struct counter *next;
-} counter_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;
-} op_metrics_t;
-
-typedef struct context
-{
- metrics_op_t op;
-
- counter_t *counter_head;
- counter_t *counter_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;
+#define MAX_MONITORS_IN_FLIGHT 1000
+
static const char *
metrics_op_string (metrics_op_t op)
{
case METRICS_OP_TEX_IMAGE:
return "glTexImage*(+)";
default:
- fprintf (stderr, "Internal error: "
+ fprintf (stderr, "fips: Internal error: "
"Unknown metrics op value: %d\n", op);
exit (1);
}
void
metrics_counter_start (void)
{
- counter_t *counter;
+ context_t *ctx = context_get_current ();
+ unsigned i;
- counter = xmalloc (sizeof(counter_t));
+ /* Initialize the timer_query and monitor objects */
+ glGenQueries (1, &ctx->timer_begun_id);
- glGenQueries (1, &counter->id);
+ glGenPerfMonitorsAMD (1, &ctx->monitor_begun_id);
- counter->op = current_context.op;
- counter->next = NULL;
+ for (i = 0; i < ctx->metrics_info.num_groups; i++)
+ {
+ metrics_group_info_t *group;
+ int num_counters;
- if (current_context.counter_tail) {
- current_context.counter_tail->next = counter;
- current_context.counter_tail = counter;
- } else {
- current_context.counter_tail = counter;
- current_context.counter_head = counter;
+ group = &ctx->metrics_info.groups[i];
+
+ num_counters = group->num_counters;
+ if (group->max_active_counters < group->num_counters)
+ {
+ fprintf (stderr, "Warning: Only monitoring %d/%d counters from group %d\n",
+ group->max_active_counters,
+ group->num_counters, i);
+ num_counters = group->max_active_counters;
+
+ }
+
+ glSelectPerfMonitorCountersAMD(ctx->monitor_begun_id,
+ GL_TRUE, group->id,
+ num_counters,
+ group->counter_ids);
}
- glBeginQuery (GL_TIME_ELAPSED, counter->id);
+ /* Start the queries */
+ glBeginQuery (GL_TIME_ELAPSED, ctx->timer_begun_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 (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
-accumulate_program_time (metrics_op_t op, unsigned time_ns)
+op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t op)
+{
+ metrics_info_t *info = &ctx->metrics_info;
+ unsigned i, j;
+
+ metrics->op = op;
+ metrics->time_ns = 0.0;
+
+ metrics->counters = xmalloc (sizeof(double *) * info->num_groups);
+
+ for (i = 0; i < info->num_groups; i++) {
+ metrics->counters[i] = xmalloc (sizeof (double) *
+ info->groups[i].num_counters);
+ for (j = 0; j < info->groups[i].num_counters; j++)
+ metrics->counters[i][j] = 0.0;
+ }
+}
+
+static op_metrics_t *
+ctx_get_op_metrics (context_t *ctx, metrics_op_t op)
{
- context_t *ctx = ¤t_context;
unsigned i;
- if (op >= ctx->num_op_metrics) {
+ if (op >= ctx->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++) {
- ctx->op_metrics[i].op = i;
- ctx->op_metrics[i].time_ns = 0.0;
- }
+ for (i = ctx->num_op_metrics; i < op + 1; i++)
+ op_metrics_init (ctx, &ctx->op_metrics[i], i);
ctx->num_op_metrics = op + 1;
}
- ctx->op_metrics[op].time_ns += time_ns;
+ return &ctx->op_metrics[op];
+}
+
+static void
+accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint size)
+{
+#define CONSUME(var) \
+ if (p + sizeof(var) > ((unsigned char *) result) + size) \
+ { \
+ fprintf (stderr, "Unexpected end-of-buffer while " \
+ "parsing results\n"); \
+ break; \
+ } \
+ (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);
+ unsigned char *p = (unsigned char *) result;
+
+ while (p < ((unsigned char *) result) + size)
+ {
+ GLuint group_id, group_index;
+ GLuint counter_id, counter_index;
+ metrics_group_info_t *group;
+ double value;
+ unsigned i;
+
+ CONSUME (group_id);
+ CONSUME (counter_id);
+
+ for (i = 0; i < info->num_groups; i++) {
+ if (info->groups[i].id == group_id)
+ break;
+ }
+ group_index = i;
+ assert (group_index < info->num_groups);
+ group = &info->groups[group_index];
+
+ for (i = 0; i < group->num_counters; i++) {
+ if (group->counter_ids[i] == counter_id)
+ break;
+ }
+ counter_index = i;
+ assert (counter_index < group->num_counters);
+
+ switch (group->counter_types[counter_index])
+ {
+ uint uint_value;
+ uint64_t uint64_value;
+ float float_value;
+ case GL_UNSIGNED_INT:
+ CONSUME (uint_value);
+ value = uint_value;
+ break;
+ case GL_UNSIGNED_INT64_AMD:
+ CONSUME (uint64_value);
+ value = uint64_value;
+ break;
+ case GL_PERCENTAGE_AMD:
+ case GL_FLOAT:
+ CONSUME (float_value);
+ value = float_value;
+ break;
+ default:
+ fprintf (stderr, "fips: Warning: Unknown counter value type (%d)\n",
+ group->counter_types[counter_index]);
+ value = 0.0;
+ break;
+ }
+
+ metrics->counters[group_index][counter_index] += value;
+ }
+}
+
+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 (ctx, op);
+
+ metrics->time_ns += time_ns;
+}
+
+typedef struct per_stage_metrics
+{
+ op_metrics_t *metrics;
+ shader_stage_info_t *stage;
+ double time_ns;
+ double active;
+} per_stage_metrics_t;
+
+static int
+_is_shader_stage_counter (metrics_info_t *info,
+ unsigned group_index,
+ unsigned counter_index)
+{
+ shader_stage_info_t *stage;
+ unsigned i;
+
+ for (i = 0; i < info->num_shader_stages; i++) {
+ stage = &info->stages[i];
+
+ if (stage->active_group_index == group_index &&
+ stage->active_counter_index == counter_index)
+ {
+ return 1;
+ }
+
+ if (stage->stall_group_index == group_index &&
+ stage->stall_counter_index == counter_index)
+ {
+ return 1;
+ }
+ }
+
+ return 0;
+}
+
+static void
+print_per_stage_metrics (context_t *ctx,
+ per_stage_metrics_t *per_stage,
+ double total)
+{
+ metrics_info_t *info = &ctx->metrics_info;
+ op_metrics_t *metric = per_stage->metrics;
+ metrics_group_info_t *group;
+ const char *op_string;
+ unsigned group_index, counter;
+ double value;
+
+ /* Don't print anything for stages with no alloted time. */
+ if (per_stage->time_ns == 0.0)
+ return;
+
+ op_string = metrics_op_string (metric->op);
+
+ printf ("%21s", op_string);
+
+ if (metric->op >= METRICS_OP_SHADER) {
+ printf (" %3d", metric->op - METRICS_OP_SHADER);
+ } else {
+ printf (" ");
+
+ }
+
+ 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,
+ per_stage->time_ns / total * 100);
+
+ if (per_stage->active)
+ printf (", %4.1f%% active", per_stage->active * 100);
+
+ printf ("\n");
+
+ /* I'm not seeing a lot of value printing the rest of these
+ * performance counters by default yet. Use --verbose to get
+ * them for now. */
+ if (! verbose)
+ return;
+
+ printf ("[");
+ for (group_index = 0; group_index < info->num_groups; group_index++) {
+ group = &info->groups[group_index];
+ for (counter = 0; counter < group->num_counters; counter++) {
+
+ /* Don't print this counter value if it's a
+ * per-stage cycle counter, (which we have
+ * already accounted for). */
+ if (_is_shader_stage_counter (info, group_index, counter))
+ continue;
+
+ value = metric->counters[group_index][counter];
+ if (value == 0.0)
+ continue;
+ printf ("%s: %.2f ", group->counter_names[counter],
+ value / 1e6);
+ }
+ }
+ printf ("]\n");
}
static int
-time_compare(const void *in_a, const void *in_b, void *arg)
+time_compare(const void *in_a, const void *in_b, void *arg unused)
{
- int a = *(const int *)in_a;
- int b = *(const int *)in_b;
- struct op_metrics *metrics = arg;
+ const per_stage_metrics_t *a = in_a;
+ const per_stage_metrics_t *b = in_b;
+
- if (metrics[a].time_ns < metrics[b].time_ns)
+ if (a->time_ns < b->time_ns)
return -1;
- if (metrics[a].time_ns > metrics[b].time_ns)
+ if (a->time_ns > b->time_ns)
return 1;
return 0;
}
static void
print_program_metrics (void)
{
- context_t *ctx = ¤t_context;
- unsigned i, j;
- int *sorted; /* Sorted indices into the ctx->op_metrics */
- double total = 0;
-
- /* Make a sorted list of the operations by time used, and figure
- * out the total so we can print percentages.
+ 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;
+ double total_time, op_cycles;
+ op_metrics_t *op;
+ unsigned group_index, counter_index;
+ unsigned i, j, num_sorted;
+
+ /* Make a sorted list of the per-stage operations by time
+ * used, and figure out the total so we can print percentages.
*/
- sorted = calloc(ctx->num_op_metrics, sizeof(*sorted));
- for (i = 0; i < ctx->num_op_metrics; i++) {
- sorted[i] = i;
- total += ctx->op_metrics[i].time_ns;
- }
- qsort_r(sorted, ctx->num_op_metrics, sizeof(*sorted),
- time_compare, ctx->op_metrics);
+ num_sorted = ctx->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++) {
- const char *op_string;
- op_metrics_t *metric =&ctx->op_metrics[sorted[i]];
-
- /* Since we sparsely fill the array based on program
- * id, many "programs" have no time.
- */
- if (metric->time_ns == 0.0)
- continue;
-
- op_string = metrics_op_string (metric->op);
-
- printf ("%s", op_string);
- if (metric->op >= METRICS_OP_SHADER) {
- printf (" %d:", metric->op - METRICS_OP_SHADER);
- } else {
- printf (":");
- for (j = strlen (op_string); j < 20; j++)
- printf (" ");
+
+ op = &ctx->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;
+
+ for (j = 0; j < num_shader_stages; j++) {
+ /* Active cycles */
+ group_index = info->stages[j].active_group_index;
+ counter_index = info->stages[j].active_counter_index;
+ op_cycles += op->counters[group_index][counter_index];
+
+ /* Stall cycles */
+ group_index = info->stages[j].stall_group_index;
+ counter_index = info->stages[j].stall_counter_index;
+ op_cycles += op->counters[group_index][counter_index];
+ }
+
+ for (j = 0; j < num_shader_stages; j++) {
+ double active_cycles, stall_cycles, stage_cycles;
+
+ /* Active cycles */
+ group_index = info->stages[j].active_group_index;
+ counter_index = info->stages[j].active_counter_index;
+ active_cycles = op->counters[group_index][counter_index];
+
+ /* Stall cycles */
+ group_index = info->stages[j].stall_group_index;
+ counter_index = info->stages[j].stall_counter_index;
+ stall_cycles = op->counters[group_index][counter_index];
+
+ stage_cycles = active_cycles + stall_cycles;
+
+ per_stage = &sorted[i * num_shader_stages + j];
+ per_stage->metrics = op;
+
+ if (op_cycles) {
+ per_stage->stage = &info->stages[j];
+ per_stage->time_ns = op->time_ns * (stage_cycles / op_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 {
+ per_stage->active = 0.0;
+ }
}
- printf ("\t%7.2f ms (% 2.1f%%)\n",
- metric->time_ns / 1e6,
- metric->time_ns / total * 100);
}
+
+ qsort_r (sorted, num_sorted, sizeof (*sorted),
+ time_compare, ctx->op_metrics);
+
+ for (i = 0; i < num_sorted; i++)
+ print_per_stage_metrics (ctx, &sorted[i], total_time);
+
+ 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");
+
+ 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)
+{
+ context_t *ctx = context_get_current ();
+
+ /* Consume all timer queries that are ready. */
+ timer_query_t *timer = ctx->timer_head;
+
+ while (timer) {
+ GLuint available, elapsed;
+
+ glGetQueryObjectuiv (timer->id,
+ GL_QUERY_RESULT_AVAILABLE, &available);
+ if (! available)
+ break;
+
+ glGetQueryObjectuiv (timer->id,
+ GL_QUERY_RESULT, &elapsed);
+
+ accumulate_program_time (timer->op, elapsed);
+
+ ctx->timer_head = timer->next;
+ if (ctx->timer_head == NULL)
+ ctx->timer_tail = NULL;
+
+ glDeleteQueries (1, &timer->id);
+
+ free (timer);
+ timer = ctx->timer_head;
+ }
+
+ /* And similarly for all performance monitors that are ready. */
+ monitor_t *monitor = ctx->monitor_head;
+
+ while (monitor) {
+ GLuint available, result_size, *result;
+ GLint bytes_written;
+
+ glGetPerfMonitorCounterDataAMD (monitor->id,
+ GL_PERFMON_RESULT_AVAILABLE_AMD,
+ sizeof (available), &available,
+ NULL);
+ if (! available)
+ break;
+
+ glGetPerfMonitorCounterDataAMD (monitor->id,
+ GL_PERFMON_RESULT_SIZE_AMD,
+ sizeof (result_size),
+ &result_size, NULL);
+
+ result = xmalloc (result_size);
+
+ glGetPerfMonitorCounterDataAMD (monitor->id,
+ GL_PERFMON_RESULT_AMD,
+ result_size, result,
+ &bytes_written);
+
+ accumulate_program_metrics (monitor->op, result, result_size);
+
+ free (result);
+
+ ctx->monitor_head = monitor->next;
+ if (ctx->monitor_head == NULL)
+ ctx->monitor_tail = NULL;
+
+ glDeletePerfMonitorsAMD (1, &monitor->id);
+
+ free (monitor);
+
+ ctx->monitors_in_flight--;
+
+ monitor = ctx->monitor_head;
+ }
}
initialized = 1;
}
- if (verbose)
- printf ("fips: frame %d complete\n", frames);
-
frames++;
- gettimeofday (&tv_now, NULL);
- /* Consume all counters that are ready. */
- counter_t *counter = current_context.counter_head;
+ metrics_collect_available ();
- while (counter) {
- GLuint available, elapsed;
-
- glGetQueryObjectuiv (counter->id, GL_QUERY_RESULT_AVAILABLE,
- &available);
- if (! available)
- break;
-
- glGetQueryObjectuiv (counter->id, GL_QUERY_RESULT, &elapsed);
-
- accumulate_program_time (counter->op, elapsed);
-
- current_context.counter_head = counter->next;
- if (current_context.counter_head == NULL)
- current_context.counter_tail = NULL;
-
- glDeleteQueries (1, &counter->id);
-
- free (counter);
- counter = current_context.counter_head;
- }
-
- if (frames % 60 == 0) {
+ 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);