X-Git-Url: https://git.cworth.org/git?a=blobdiff_plain;f=metrics.c;h=996e7795e74cfe1929623bd9bf75762a6665269a;hb=b158889b37077e5700b3323fe5f6dd57a38548f6;hp=fdd794a726863bba4a4b756d81be9d95e82b10d1;hpb=3d421696b9e6cc28f590204fb290f775330e78e8;p=fips diff --git a/metrics.c b/metrics.c index fdd794a..996e779 100644 --- a/metrics.c +++ b/metrics.c @@ -21,6 +21,8 @@ #define _GNU_SOURCE +#include + #include #include #include @@ -59,7 +61,7 @@ typedef struct monitor 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; @@ -69,6 +71,11 @@ typedef struct op_metrics 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 @@ -93,8 +100,103 @@ struct metrics unsigned num_op_metrics; op_metrics_t *op_metrics; + + /* Per-frame time and latency measurement. */ + GLint64 swap_begin_timestamp; + unsigned swap_end_timestamp_id; + GLint64 previous_swap_end_timestamp; }; +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; + + glGenQueries (1, &metrics->swap_end_timestamp_id); + + /* Get the first frame timestamp started immediately. */ + glQueryCounter (metrics->swap_end_timestamp_id, GL_TIMESTAMP); + + 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) { @@ -141,23 +243,30 @@ metrics_op_string (metrics_op_t op) } void -metrics_counter_start (void) +metrics_counter_start (metrics_t *metrics) { - context_t *ctx = context_get_current (); - metrics_t *metrics = ctx->metrics; unsigned i; - /* Initialize the timer_query and monitor objects */ + /* 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; + } + + /* 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) @@ -182,16 +291,16 @@ metrics_counter_start (void) } void -metrics_counter_stop (void) +metrics_counter_stop (metrics_t *metrics) { - context_t *ctx = context_get_current (); - metrics_t *metrics = ctx->metrics; timer_query_t *timer; monitor_t *monitor; /* Stop the current timer and monitor. */ glEndQuery (GL_TIME_ELAPSED); - glEndPerfMonitorAMD (metrics->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. */ @@ -209,19 +318,21 @@ metrics_counter_stop (void) 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 = metrics->op; - monitor->id = metrics->monitor_begun_id; - monitor->next = NULL; + 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; + 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++; @@ -232,31 +343,24 @@ metrics_counter_stop (void) * available by now.) */ if (metrics->monitors_in_flight > MAX_MONITORS_IN_FLIGHT) - metrics_collect_available (); + 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 (); - metrics_t *metrics = ctx->metrics; - metrics->op = op; } metrics_op_t -metrics_get_current_op (void) +metrics_get_current_op (metrics_t *metrics) { - context_t *ctx = context_get_current (); - metrics_t *metrics = ctx->metrics; - 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; @@ -273,9 +377,8 @@ op_metrics_init (context_t *ctx, op_metrics_t *metrics, metrics_op_t 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) { - metrics_t *metrics = ctx->metrics; unsigned i; if (op >= metrics->num_op_metrics) @@ -283,7 +386,7 @@ ctx_get_op_metrics (context_t *ctx, metrics_op_t op) 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 (ctx, &metrics->op_metrics[i], i); + op_metrics_init (metrics->info, &metrics->op_metrics[i], i); metrics->num_op_metrics = op + 1; } @@ -292,7 +395,8 @@ ctx_get_op_metrics (context_t *ctx, metrics_op_t 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) \ @@ -304,9 +408,8 @@ accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint 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) @@ -360,19 +463,18 @@ accumulate_program_metrics (metrics_op_t op, GLuint *result, GLuint 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 @@ -411,12 +513,12 @@ _is_shader_stage_counter (metrics_info_t *info, } 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; @@ -426,12 +528,12 @@ print_per_stage_metrics (context_t *ctx, 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 (" "); @@ -468,7 +570,7 @@ print_per_stage_metrics (context_t *ctx, 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], @@ -493,11 +595,9 @@ time_compare(const void *in_a, const void *in_b, void *arg unused) } static void -print_program_metrics (void) +print_program_metrics (metrics_t *metrics) { - context_t *ctx = context_get_current (); - metrics_t *metrics = ctx->metrics; - 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; @@ -505,10 +605,24 @@ print_program_metrics (void) unsigned group_index, counter_index; unsigned i, j, num_sorted; + if (! info->have_perfmon && + ! info->printed_missing_perfmon_warning) + { + fprintf (stderr, + "Warning: The GL_AMD_performance_monitor extensions is not available.\n" + " Some fips features, (such as per-shader-stage timings),\n" + " will not be available.\n"); + + info->printed_missing_perfmon_warning = true; + } + /* Make a sorted list of the per-stage operations by time * used, and figure out the total so we can print percentages. */ - num_sorted = metrics->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); @@ -524,6 +638,14 @@ print_program_metrics (void) /* 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; @@ -584,76 +706,14 @@ print_program_metrics (void) 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_t *metrics = ctx->metrics; - 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 = metrics->timer_head; - timer; - timer = timer_next) - { - timer_next = timer->next; - free (timer); - } - - for (monitor = metrics->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 (); - metrics_t *metrics = ctx->metrics; - /* Consume all timer queries that are ready. */ timer_query_t *timer = metrics->timer_head; @@ -668,7 +728,7 @@ metrics_collect_available (void) glGetQueryObjectuiv (timer->id, GL_QUERY_RESULT, &elapsed); - accumulate_program_time (timer->op, elapsed); + accumulate_program_time (metrics, timer->op, elapsed); metrics->timer_head = timer->next; if (metrics->timer_head == NULL) @@ -680,6 +740,9 @@ metrics_collect_available (void) timer = metrics->timer_head; } + if (! metrics->info->have_perfmon) + return; + /* And similarly for all performance monitors that are ready. */ monitor_t *monitor = metrics->monitor_head; @@ -706,7 +769,7 @@ metrics_collect_available (void) result_size, result, &bytes_written); - accumulate_program_metrics (monitor->op, result, result_size); + accumulate_program_metrics (metrics, monitor->op, result, result_size); free (result); @@ -724,24 +787,81 @@ metrics_collect_available (void) } } +void +metrics_end_frame_pre_swap (metrics_t *metrics) +{ + GLuint64 swap_end_timestamp, frame_time_ns, latency_ns; + + /* Don't leave any counters running over the end_frame work we + * do here. The counters will be started again at the end of + * metrics_end_frame_post_swap. */ + + metrics_counter_stop (metrics); + + /* Now that an entire frame's worth of content has gone by, we + * can be sure that the timer quiery for the previous frame's + * swap is available. + * + * Note: The only case in which this query isn't immediately + * available would be if there was effectively nothing in this + * frame. In that case, triggering a block on this query + * result is not a concern (since an empty frame is not + * interesting, and all of our counters are stopped anyway). + * + * Accepting this one block is much better than adding a + * linked list for these results that would reaslisticly never + * have more than one entry anyway. + */ + + glGetQueryObjectui64v (metrics->swap_end_timestamp_id, + GL_QUERY_RESULT, &swap_end_timestamp); + + if (frames > 0) { + /* Subtract previous frame's timestamp to get frame time. */ + frame_time_ns = swap_end_timestamp - metrics->previous_swap_end_timestamp; + + latency_ns = swap_end_timestamp - metrics->swap_begin_timestamp; + + /* We've waited one frame to ensure we have a timestamp + * result. So the time we've actually measured here is + * for the previous frame. */ + printf ("Frame %d: %.2f ms (latency: %.2f ms)\n", + frames - 1, frame_time_ns / 1e6, latency_ns / 1e6); + } + + metrics->previous_swap_end_timestamp = swap_end_timestamp; + + /* Before the actual swap call, we get the current timestamp + * value. This is a synchronous get so we can use this as the + * baseline for a frame latency measurment. */ + + glGetInteger64v (GL_TIMESTAMP, &metrics->swap_begin_timestamp); +} void -metrics_end_frame (void) +metrics_end_frame_post_swap (metrics_t *metrics) { static int initialized = 0; static struct timeval tv_start, tv_now; + /* Now that the swap command has been queued, we issue an + * asynchronous query of the timestamp value. Comparing this + * to the synchronous get we just sent in + * metrics_end_frame_pre_swap allows us to measure the + * per-frame swap latency. */ + + glQueryCounter (metrics->swap_end_timestamp_id, GL_TIMESTAMP); + if (! initialized) { gettimeofday (&tv_start, NULL); - atexit (metrics_exit); if (getenv ("FIPS_VERBOSE")) verbose = 1; initialized = 1; } - frames++; + metrics_collect_available (metrics); - metrics_collect_available (); + frames++; if (frames % 15 == 0) { double fps; @@ -753,85 +873,9 @@ metrics_end_frame (void) printf("FPS: %.3f\n", fps); - print_program_metrics (); - } -} - -metrics_t * -metrics_create (void) -{ - metrics_t *metrics; - - metrics = xmalloc (sizeof (metrics_t)); - - 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->monitor_begun_id) { - glEndPerfMonitorAMD (metrics->monitor_begun_id); - glDeletePerfMonitorsAMD (1, &metrics->monitor_begun_id); - metrics->monitor_begun_id = 0; + print_program_metrics (metrics); } - 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); + /* Start the counter up again now that we're done. */ + metrics_counter_start (metrics); }