static unsigned
cairo_perf_calibrate (cairo_perf_t *perf,
cairo_perf_func_t perf_func)
{
cairo_time_t calibration, calibration_max;
unsigned loops, min_loops;
min_loops = 1;
calibration = perf_func (perf->cr, perf->size, perf->size, min_loops);
if (!perf->fast_and_sloppy) {
calibration_max = _cairo_time_from_s (perf->ms_per_iteration * 0.0001 / 4);
while (calibration < calibration_max) {
min_loops *= 2;
calibration = perf_func (perf->cr, perf->size, perf->size, min_loops);
}
}
/* XXX
* Compute the number of loops required for the timing
* interval to be perf->ms_per_iteration milliseconds. This
* helps to eliminate sampling variance due to timing and
* other systematic errors. However, it also hides
* synchronisation overhead as we attempt to process a large
* batch of identical operations in a single shot. This can be
* considered both good and bad... It would be good to perform
* a more rigorous analysis of the synchronisation overhead,
* that is to estimate the time for loop=0.
*/
loops = _cairo_time_from_s (perf->ms_per_iteration * 0.001 * min_loops / calibration);
min_loops = perf->fast_and_sloppy ? 1 : 10;
if (loops < min_loops)
loops = min_loops;
return loops;
}
void
cairo_perf_run (cairo_perf_t *perf,
const char *name,
cairo_perf_func_t perf_func,
cairo_count_func_t count_func)
{
static cairo_bool_t first_run = TRUE;
unsigned int i, similar, has_similar;
cairo_time_t *times;
cairo_stats_t stats = {0.0, 0.0};
int low_std_dev_count;
if (perf->list_only) {
printf ("%s\n", name);
return;
}
if (first_run) {
if (perf->raw) {
printf ("[ # ] %s.%-s %s %s %s ...\n",
"backend", "content", "test-size", "ticks-per-ms", "time(ticks)");
}
if (perf->summary) {
fprintf (perf->summary,
"[ # ] %8s.%-4s %28s %8s %8s %5s %5s %s %s\n",
"backend", "content", "test-size", "min(ticks)", "min(ms)", "median(ms)",
"stddev.", "iterations", "overhead");
}
first_run = FALSE;
}
times = perf->times;
if (getenv ("CAIRO_PERF_OUTPUT") != NULL) { /* check output */
char *filename;
cairo_status_t status;
xasprintf (&filename, "%s.%s.%s.%d.out.png",
name, perf->target->name,
_content_to_string (perf->target->content, 0),
perf->size);
cairo_save (perf->cr);
perf_func (perf->cr, perf->size, perf->size, 1);
cairo_restore (perf->cr);
status = cairo_surface_write_to_png (cairo_get_target (perf->cr), filename);
if (status) {
fprintf (stderr, "Failed to generate output check '%s': %s\n",
filename, cairo_status_to_string (status));
return;
}
free (filename);
}
has_similar = cairo_perf_has_similar (perf);
for (similar = 0; similar <= has_similar; similar++) {
unsigned loops;
if (perf->summary) {
fprintf (perf->summary,
"[%3d] %8s.%-5s %26s.%-3d ",
perf->test_number, perf->target->name,
_content_to_string (perf->target->content, similar),
name, perf->size);
fflush (perf->summary);
}
/* We run one iteration in advance to warm caches and calibrate. */
cairo_perf_yield ();
if (similar)
cairo_push_group_with_content (perf->cr,
cairo_boilerplate_content (perf->target->content));
else
cairo_save (perf->cr);
perf_func (perf->cr, perf->size, perf->size, 1);
loops = cairo_perf_calibrate (perf, perf_func);
if (similar)
cairo_pattern_destroy (cairo_pop_group (perf->cr));
else
cairo_restore (perf->cr);
low_std_dev_count = 0;
for (i =0; i < perf->iterations; i++) {
cairo_perf_yield ();
if (similar)
cairo_push_group_with_content (perf->cr,
cairo_boilerplate_content (perf->target->content));
else
cairo_save (perf->cr);
times[i] = perf_func (perf->cr, perf->size, perf->size, loops) ;
if (similar)
cairo_pattern_destroy (cairo_pop_group (perf->cr));
else
cairo_restore (perf->cr);
if (perf->raw) {
if (i == 0)
printf ("[*] %s.%s %s.%d %g",
perf->target->name,
_content_to_string (perf->target->content, similar),
name, perf->size,
_cairo_time_to_double (_cairo_time_from_s (1.)) / 1000.);
printf (" %lld", (long long) (times[i] / (double) loops));
} else if (! perf->exact_iterations) {
if (i > 0) {
_cairo_stats_compute (&stats, times, i+1);
if (stats.std_dev <= CAIRO_PERF_LOW_STD_DEV) {
low_std_dev_count++;
if (low_std_dev_count >= CAIRO_PERF_STABLE_STD_DEV_COUNT)
break;
} else {
low_std_dev_count = 0;
}
}
}
}
if (perf->raw)
printf ("\n");
if (perf->summary) {
_cairo_stats_compute (&stats, times, i);
if (count_func != NULL) {
double count = count_func (perf->cr, perf->size, perf->size);
fprintf (perf->summary,
"%.3f [%10lld/%d] %#8.3f %#8.3f %#5.2f%% %3d: %.2f\n",
stats.min_ticks /(double) loops,
(long long) stats.min_ticks, loops,
_cairo_time_to_s (stats.min_ticks) * 1000.0 / loops,
_cairo_time_to_s (stats.median_ticks) * 1000.0 / loops,
stats.std_dev * 100.0, stats.iterations,
count / _cairo_time_to_s (stats.min_ticks));
} else {
fprintf (perf->summary,
"%.3f [%10lld/%d] %#8.3f %#8.3f %#5.2f%% %3d\n",
stats.min_ticks /(double) loops,
(long long) stats.min_ticks, loops,
_cairo_time_to_s (stats.min_ticks) * 1000.0 / loops,
_cairo_time_to_s (stats.median_ticks) * 1000.0 / loops,
stats.std_dev * 100.0, stats.iterations);
}
fflush (perf->summary);
}
perf->test_number++;
}
}
static void
cairo_perf_trace (cairo_perf_t *perf,
const cairo_boilerplate_target_t *target,
const char *trace)
{
static cairo_bool_t first_run = TRUE;
unsigned int i;
cairo_time_t *times, *paint, *mask, *fill, *stroke, *glyphs;
cairo_stats_t stats = {0.0, 0.0};
struct trace args = { target };
int low_std_dev_count;
char *trace_cpy, *name;
const cairo_script_interpreter_hooks_t hooks = {
&args,
perf->tile_size ? _tiling_surface_create : _similar_surface_create,
NULL, /* surface_destroy */
_context_create,
NULL, /* context_destroy */
NULL, /* show_page */
NULL, /* copy_page */
_source_image_create,
};
args.tile_size = perf->tile_size;
args.observe = perf->observe;
trace_cpy = xstrdup (trace);
name = basename_no_ext (trace_cpy);
if (perf->list_only) {
printf ("%s\n", name);
free (trace_cpy);
return;
}
if (first_run) {
if (perf->raw) {
printf ("[ # ] %s.%-s %s %s %s ...\n",
"backend", "content", "test-size", "ticks-per-ms", "time(ticks)");
}
if (perf->summary) {
if (perf->observe) {
fprintf (perf->summary,
"[ # ] %8s %28s %9s %9s %9s %9s %9s %9s %5s\n",
"backend", "test",
"total(s)", "paint(s)", "mask(s)", "fill(s)", "stroke(s)", "glyphs(s)",
"count");
} else {
fprintf (perf->summary,
"[ # ] %8s %28s %8s %5s %5s %s\n",
"backend", "test", "min(s)", "median(s)",
"stddev.", "count");
}
}
first_run = FALSE;
}
times = perf->times;
paint = times + perf->iterations;
mask = paint + perf->iterations;
stroke = mask + perf->iterations;
fill = stroke + perf->iterations;
glyphs = fill + perf->iterations;
low_std_dev_count = 0;
for (i = 0; i < perf->iterations && ! user_interrupt; i++) {
cairo_script_interpreter_t *csi;
cairo_status_t status;
unsigned int line_no;
args.surface = target->create_surface (NULL,
CAIRO_CONTENT_COLOR_ALPHA,
1, 1,
1, 1,
CAIRO_BOILERPLATE_MODE_PERF,
&args.closure);
fill_surface(args.surface); /* remove any clear flags */
if (perf->observe) {
cairo_surface_t *obs;
obs = cairo_surface_create_observer (args.surface,
CAIRO_SURFACE_OBSERVER_NORMAL);
cairo_surface_destroy (args.surface);
args.surface = obs;
}
if (cairo_surface_status (args.surface)) {
fprintf (stderr,
"Error: Failed to create target surface: %s\n",
target->name);
return;
}
cairo_perf_timer_set_synchronize (target->synchronize, args.closure);
if (i == 0) {
describe (perf, args.closure);
if (perf->summary) {
fprintf (perf->summary,
"[%3d] %8s %28s ",
perf->test_number,
perf->target->name,
name);
fflush (perf->summary);
}
}
csi = cairo_script_interpreter_create ();
cairo_script_interpreter_install_hooks (csi, &hooks);
if (! perf->observe) {
cairo_perf_yield ();
cairo_perf_timer_start ();
}
cairo_script_interpreter_run (csi, trace);
line_no = cairo_script_interpreter_get_line_number (csi);
/* Finish before querying timings in case we are using an intermediate
* target and so need to destroy all surfaces before rendering
* commences.
*/
cairo_script_interpreter_finish (csi);
if (perf->observe) {
cairo_device_t *observer = cairo_surface_get_device (args.surface);
times[i] = _cairo_time_from_s (1.e-9 * cairo_device_observer_elapsed (observer));
paint[i] = _cairo_time_from_s (1.e-9 * cairo_device_observer_paint_elapsed (observer));
mask[i] = _cairo_time_from_s (1.e-9 * cairo_device_observer_mask_elapsed (observer));
stroke[i] = _cairo_time_from_s (1.e-9 * cairo_device_observer_stroke_elapsed (observer));
fill[i] = _cairo_time_from_s (1.e-9 * cairo_device_observer_fill_elapsed (observer));
glyphs[i] = _cairo_time_from_s (1.e-9 * cairo_device_observer_glyphs_elapsed (observer));
} else {
fill_surface (args.surface); /* queue a write to the sync'ed surface */
cairo_perf_timer_stop ();
times[i] = cairo_perf_timer_elapsed ();
}
scache_clear ();
cairo_surface_destroy (args.surface);
if (target->cleanup)
target->cleanup (args.closure);
status = cairo_script_interpreter_destroy (csi);
if (status) {
if (perf->summary) {
fprintf (perf->summary, "Error during replay, line %d: %s\n",
line_no,
cairo_status_to_string (status));
}
goto out;
}
if (perf->raw) {
if (i == 0)
printf ("[*] %s.%s %s.%d %g",
perf->target->name,
"rgba",
name,
0,
_cairo_time_to_double (_cairo_time_from_s (1)) / 1000.);
printf (" %lld", (long long) times[i]);
fflush (stdout);
} else if (! perf->exact_iterations) {
if (i > CAIRO_PERF_MIN_STD_DEV_COUNT) {
_cairo_stats_compute (&stats, times, i+1);
if (stats.std_dev <= CAIRO_PERF_LOW_STD_DEV) {
if (++low_std_dev_count >= CAIRO_PERF_STABLE_STD_DEV_COUNT)
break;
} else {
low_std_dev_count = 0;
}
}
}
if (perf->summary && perf->summary_continuous) {
_cairo_stats_compute (&stats, times, i+1);
fprintf (perf->summary,
"\r[%3d] %8s %28s ",
perf->test_number,
perf->target->name,
name);
if (perf->observe) {
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, paint, i+1);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, mask, i+1);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, fill, i+1);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, stroke, i+1);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, glyphs, i+1);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
fprintf (perf->summary,
" %5d", i+1);
} else {
fprintf (perf->summary,
"%#8.3f %#8.3f %#6.2f%% %4d/%d",
_cairo_time_to_s (stats.min_ticks),
_cairo_time_to_s (stats.median_ticks),
stats.std_dev * 100.0,
stats.iterations, i+1);
}
fflush (perf->summary);
}
}
user_interrupt = 0;
if (perf->summary) {
_cairo_stats_compute (&stats, times, i);
if (perf->summary_continuous) {
fprintf (perf->summary,
"\r[%3d] %8s %28s ",
perf->test_number,
perf->target->name,
name);
}
if (perf->observe) {
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, paint, i);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, mask, i);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, fill, i);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, stroke, i);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
_cairo_stats_compute (&stats, glyphs, i);
fprintf (perf->summary,
" %#9.3f", _cairo_time_to_s (stats.median_ticks));
fprintf (perf->summary,
" %5d\n", i);
} else {
fprintf (perf->summary,
"%#8.3f %#8.3f %#6.2f%% %4d/%d\n",
_cairo_time_to_s (stats.min_ticks),
_cairo_time_to_s (stats.median_ticks),
stats.std_dev * 100.0,
stats.iterations, i);
}
fflush (perf->summary);
}
out:
if (perf->raw) {
printf ("\n");
fflush (stdout);
}
perf->test_number++;
free (trace_cpy);
}
