virtualx-engine/main/main_timer_sync.cpp
lawnjelly 3025b6d299 Delta smoothing - fix overflow for long frames
Extremely long frames caused by suspending and resuming the machine could result in an overflow in the delta smoothing because it uses 32 bit math on delta values measured in nanoseconds.

This PR puts a cap of a second as the maximum frame delta that will be processed by the smoothing, otherwise it returns the frame delta 64 bit value unaltered. It also converts internal math to explicitly use 64 bit integers.
2021-08-11 09:47:23 +01:00

481 lines
17 KiB
C++

/*************************************************************************/
/* main_timer_sync.cpp */
/*************************************************************************/
/* This file is part of: */
/* GODOT ENGINE */
/* https://godotengine.org */
/*************************************************************************/
/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur. */
/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md). */
/* */
/* Permission is hereby granted, free of charge, to any person obtaining */
/* a copy of this software and associated documentation files (the */
/* "Software"), to deal in the Software without restriction, including */
/* without limitation the rights to use, copy, modify, merge, publish, */
/* distribute, sublicense, and/or sell copies of the Software, and to */
/* permit persons to whom the Software is furnished to do so, subject to */
/* the following conditions: */
/* */
/* The above copyright notice and this permission notice shall be */
/* included in all copies or substantial portions of the Software. */
/* */
/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, */
/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF */
/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY */
/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, */
/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE */
/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. */
/*************************************************************************/
#include "main_timer_sync.h"
#include "core/math/math_funcs.h"
#include "core/os/os.h"
void MainFrameTime::clamp_idle(float min_idle_step, float max_idle_step) {
if (idle_step < min_idle_step) {
idle_step = min_idle_step;
} else if (idle_step > max_idle_step) {
idle_step = max_idle_step;
}
}
/////////////////////////////////
void MainTimerSync::DeltaSmoother::update_refresh_rate_estimator(int64_t p_delta) {
// the calling code should prevent 0 or negative values of delta
// (preventing divide by zero)
// note that if the estimate gets locked, and something external changes this
// (e.g. user changes to non-vsync in the OS), then the results may be less than ideal,
// but usually it will detect this via the FPS measurement and not attempt smoothing.
// This should be a rare occurrence anyway, and will be cured next time user restarts game.
if (_estimate_locked) {
return;
}
// First average the delta over NUM_READINGS
_estimator_total_delta += p_delta;
_estimator_delta_readings++;
const int NUM_READINGS = 60;
if (_estimator_delta_readings < NUM_READINGS) {
return;
}
// use average
p_delta = _estimator_total_delta / NUM_READINGS;
// reset the averager for next time
_estimator_delta_readings = 0;
_estimator_total_delta = 0;
///////////////////////////////
int fps = Math::round(1000000.0 / p_delta);
// initial estimation, to speed up converging, special case we will estimate the refresh rate
// from the first average FPS reading
if (_estimated_fps == 0) {
// below 50 might be chugging loading stuff, or else
// dropping loads of frames, so the estimate will be inaccurate
if (fps >= 50) {
_estimated_fps = fps;
#ifdef GODOT_DEBUG_DELTA_SMOOTHER
print_line("initial guess (average measured) refresh rate: " + itos(fps));
#endif
} else {
// can't get started until above 50
return;
}
}
// we hit our exact estimated refresh rate.
// increase our confidence in the estimate.
if (fps == _estimated_fps) {
// note that each hit is an average of NUM_READINGS frames
_hits_at_estimated++;
if (_estimate_complete && _hits_at_estimated == 20) {
_estimate_locked = true;
#ifdef GODOT_DEBUG_DELTA_SMOOTHER
print_line("estimate LOCKED at " + itos(_estimated_fps) + " fps");
#endif
return;
}
// if we are getting pretty confident in this estimate, decide it is complete
// (it can still be increased later, and possibly lowered but only for a short time)
if ((!_estimate_complete) && (_hits_at_estimated > 2)) {
// when the estimate is complete we turn on smoothing
if (_estimated_fps) {
_estimate_complete = true;
_vsync_delta = 1000000 / _estimated_fps;
#ifdef GODOT_DEBUG_DELTA_SMOOTHER
print_line("estimate complete. vsync_delta " + itos(_vsync_delta) + ", fps " + itos(_estimated_fps));
#endif
}
}
#ifdef GODOT_DEBUG_DELTA_SMOOTHER
if ((_hits_at_estimated % (400 / NUM_READINGS)) == 0) {
String sz = "hits at estimated : " + itos(_hits_at_estimated) + ", above : " + itos(_hits_above_estimated) + "( " + itos(_hits_one_above_estimated) + " ), below : " + itos(_hits_below_estimated) + " (" + itos(_hits_one_below_estimated) + " )";
print_line(sz);
}
#endif
return;
}
const int SIGNIFICANCE_UP = 1;
const int SIGNIFICANCE_DOWN = 2;
// we are not usually interested in slowing the estimate
// but we may have overshot, so make it possible to reduce
if (fps < _estimated_fps) {
// micro changes
if (fps == (_estimated_fps - 1)) {
_hits_one_below_estimated++;
if ((_hits_one_below_estimated > _hits_at_estimated) && (_hits_one_below_estimated > SIGNIFICANCE_DOWN)) {
_estimated_fps--;
made_new_estimate();
}
return;
} else {
_hits_below_estimated++;
// don't allow large lowering if we are established at a refresh rate, as it will probably be dropped frames
bool established = _estimate_complete && (_hits_at_estimated > 10);
// macro changes
// note there is a large barrier to macro lowering. That is because it is more likely to be dropped frames
// than mis-estimation of the refresh rate.
if (!established) {
if (((_hits_below_estimated / 8) > _hits_at_estimated) && (_hits_below_estimated > SIGNIFICANCE_DOWN)) {
// decrease the estimate
_estimated_fps--;
made_new_estimate();
}
}
return;
}
}
// Changes increasing the estimate.
// micro changes
if (fps == (_estimated_fps + 1)) {
_hits_one_above_estimated++;
if ((_hits_one_above_estimated > _hits_at_estimated) && (_hits_one_above_estimated > SIGNIFICANCE_UP)) {
_estimated_fps++;
made_new_estimate();
}
return;
} else {
_hits_above_estimated++;
// macro changes
if ((_hits_above_estimated > _hits_at_estimated) && (_hits_above_estimated > SIGNIFICANCE_UP)) {
// increase the estimate
int change = fps - _estimated_fps;
change /= 2;
change = MAX(1, change);
_estimated_fps += change;
made_new_estimate();
}
return;
}
}
bool MainTimerSync::DeltaSmoother::fps_allows_smoothing(int64_t p_delta) {
_measurement_time += p_delta;
_measurement_frame_count++;
if (_measurement_frame_count == _measurement_end_frame) {
// only switch on or off if the estimate is complete
if (_estimate_complete) {
int64_t time_passed = _measurement_time - _measurement_start_time;
// average delta
time_passed /= MEASURE_FPS_OVER_NUM_FRAMES;
// estimate fps
if (time_passed) {
double fps = 1000000.0 / time_passed;
double ratio = fps / (double)_estimated_fps;
//print_line("ratio : " + String(Variant(ratio)));
if ((ratio > 0.95) && (ratio < 1.05)) {
_measurement_allows_smoothing = true;
} else {
_measurement_allows_smoothing = false;
}
}
} // estimate complete
// new start time for next iteration
_measurement_start_time = _measurement_time;
_measurement_end_frame += MEASURE_FPS_OVER_NUM_FRAMES;
}
return _measurement_allows_smoothing;
}
int64_t MainTimerSync::DeltaSmoother::smooth_delta(int64_t p_delta) {
// Conditions to disable smoothing.
// Note that vsync is a request, it cannot be relied on, the OS may override this.
// If the OS turns vsync on without vsync in the app, smoothing will not be enabled.
// If the OS turns vsync off with sync enabled in the app, the smoothing must detect this
// via the error metric and switch off.
if (!OS::get_singleton()->is_delta_smoothing_enabled() || !OS::get_singleton()->is_vsync_enabled() || Engine::get_singleton()->is_editor_hint()) {
return p_delta;
}
// Very important, ignore long deltas and pass them back unmodified.
// This is to deal with resuming after suspend for long periods.
if (p_delta > 1000000) {
return p_delta;
}
// keep a running guesstimate of the FPS, and turn off smoothing if
// conditions not close to the estimated FPS
if (!fps_allows_smoothing(p_delta)) {
return p_delta;
}
// we can't cope with negative deltas .. OS bug on some hardware
// and also very small deltas caused by vsync being off.
// This could possibly be part of a hiccup, this value isn't fixed in stone...
if (p_delta < 1000) {
return p_delta;
}
// note still some vsync off will still get through to this point...
// and we need to cope with it by not converging the estimator / and / or not smoothing
update_refresh_rate_estimator(p_delta);
// no smoothing until we know what the refresh rate is
if (!_estimate_complete) {
return p_delta;
}
// accumulate the time we have available to use
_leftover_time += p_delta;
// how many vsyncs units can we fit?
int64_t units = _leftover_time / _vsync_delta;
// a delta must include minimum 1 vsync
// (if it is less than that, it is either random error or we are no longer running at the vsync rate,
// in which case we should switch off delta smoothing, or re-estimate the refresh rate)
units = MAX(units, 1);
_leftover_time -= units * _vsync_delta;
// print_line("units " + itos(units) + ", leftover " + itos(_leftover_time/1000) + " ms");
return units * _vsync_delta;
}
/////////////////////////////////////
// returns the fraction of p_frame_slice required for the timer to overshoot
// before advance_core considers changing the physics_steps return from
// the typical values as defined by typical_physics_steps
float MainTimerSync::get_physics_jitter_fix() {
return Engine::get_singleton()->get_physics_jitter_fix();
}
// gets our best bet for the average number of physics steps per render frame
// return value: number of frames back this data is consistent
int MainTimerSync::get_average_physics_steps(float &p_min, float &p_max) {
p_min = typical_physics_steps[0];
p_max = p_min + 1;
for (int i = 1; i < CONTROL_STEPS; ++i) {
const float typical_lower = typical_physics_steps[i];
const float current_min = typical_lower / (i + 1);
if (current_min > p_max) {
return i; // bail out of further restrictions would void the interval
} else if (current_min > p_min) {
p_min = current_min;
}
const float current_max = (typical_lower + 1) / (i + 1);
if (current_max < p_min) {
return i;
} else if (current_max < p_max) {
p_max = current_max;
}
}
return CONTROL_STEPS;
}
// advance physics clock by p_idle_step, return appropriate number of steps to simulate
MainFrameTime MainTimerSync::advance_core(float p_frame_slice, int p_iterations_per_second, float p_idle_step) {
MainFrameTime ret;
ret.idle_step = p_idle_step;
// simple determination of number of physics iteration
time_accum += ret.idle_step;
ret.physics_steps = floor(time_accum * p_iterations_per_second);
int min_typical_steps = typical_physics_steps[0];
int max_typical_steps = min_typical_steps + 1;
// given the past recorded steps and typical steps to match, calculate bounds for this
// step to be typical
bool update_typical = false;
for (int i = 0; i < CONTROL_STEPS - 1; ++i) {
int steps_left_to_match_typical = typical_physics_steps[i + 1] - accumulated_physics_steps[i];
if (steps_left_to_match_typical > max_typical_steps ||
steps_left_to_match_typical + 1 < min_typical_steps) {
update_typical = true;
break;
}
if (steps_left_to_match_typical > min_typical_steps) {
min_typical_steps = steps_left_to_match_typical;
}
if (steps_left_to_match_typical + 1 < max_typical_steps) {
max_typical_steps = steps_left_to_match_typical + 1;
}
}
// try to keep it consistent with previous iterations
if (ret.physics_steps < min_typical_steps) {
const int max_possible_steps = floor((time_accum)*p_iterations_per_second + get_physics_jitter_fix());
if (max_possible_steps < min_typical_steps) {
ret.physics_steps = max_possible_steps;
update_typical = true;
} else {
ret.physics_steps = min_typical_steps;
}
} else if (ret.physics_steps > max_typical_steps) {
const int min_possible_steps = floor((time_accum)*p_iterations_per_second - get_physics_jitter_fix());
if (min_possible_steps > max_typical_steps) {
ret.physics_steps = min_possible_steps;
update_typical = true;
} else {
ret.physics_steps = max_typical_steps;
}
}
time_accum -= ret.physics_steps * p_frame_slice;
// keep track of accumulated step counts
for (int i = CONTROL_STEPS - 2; i >= 0; --i) {
accumulated_physics_steps[i + 1] = accumulated_physics_steps[i] + ret.physics_steps;
}
accumulated_physics_steps[0] = ret.physics_steps;
if (update_typical) {
for (int i = CONTROL_STEPS - 1; i >= 0; --i) {
if (typical_physics_steps[i] > accumulated_physics_steps[i]) {
typical_physics_steps[i] = accumulated_physics_steps[i];
} else if (typical_physics_steps[i] < accumulated_physics_steps[i] - 1) {
typical_physics_steps[i] = accumulated_physics_steps[i] - 1;
}
}
}
return ret;
}
// calls advance_core, keeps track of deficit it adds to animaption_step, make sure the deficit sum stays close to zero
MainFrameTime MainTimerSync::advance_checked(float p_frame_slice, int p_iterations_per_second, float p_idle_step) {
if (fixed_fps != -1) {
p_idle_step = 1.0 / fixed_fps;
}
// compensate for last deficit
p_idle_step += time_deficit;
MainFrameTime ret = advance_core(p_frame_slice, p_iterations_per_second, p_idle_step);
// we will do some clamping on ret.idle_step and need to sync those changes to time_accum,
// that's easiest if we just remember their fixed difference now
const double idle_minus_accum = ret.idle_step - time_accum;
// first, least important clamping: keep ret.idle_step consistent with typical_physics_steps.
// this smoothes out the idle steps and culls small but quick variations.
{
float min_average_physics_steps, max_average_physics_steps;
int consistent_steps = get_average_physics_steps(min_average_physics_steps, max_average_physics_steps);
if (consistent_steps > 3) {
ret.clamp_idle(min_average_physics_steps * p_frame_slice, max_average_physics_steps * p_frame_slice);
}
}
// second clamping: keep abs(time_deficit) < jitter_fix * frame_slise
float max_clock_deviation = get_physics_jitter_fix() * p_frame_slice;
ret.clamp_idle(p_idle_step - max_clock_deviation, p_idle_step + max_clock_deviation);
// last clamping: make sure time_accum is between 0 and p_frame_slice for consistency between physics and idle
ret.clamp_idle(idle_minus_accum, idle_minus_accum + p_frame_slice);
// restore time_accum
time_accum = ret.idle_step - idle_minus_accum;
// track deficit
time_deficit = p_idle_step - ret.idle_step;
// p_frame_slice is 1.0 / iterations_per_sec
// i.e. the time in seconds taken by a physics tick
ret.interpolation_fraction = time_accum / p_frame_slice;
return ret;
}
// determine wall clock step since last iteration
float MainTimerSync::get_cpu_idle_step() {
uint64_t cpu_ticks_elapsed = current_cpu_ticks_usec - last_cpu_ticks_usec;
last_cpu_ticks_usec = current_cpu_ticks_usec;
cpu_ticks_elapsed = _delta_smoother.smooth_delta(cpu_ticks_elapsed);
return cpu_ticks_elapsed / 1000000.0;
}
MainTimerSync::MainTimerSync() :
last_cpu_ticks_usec(0),
current_cpu_ticks_usec(0),
time_accum(0),
time_deficit(0),
fixed_fps(0) {
for (int i = CONTROL_STEPS - 1; i >= 0; --i) {
typical_physics_steps[i] = i;
accumulated_physics_steps[i] = i;
}
}
// start the clock
void MainTimerSync::init(uint64_t p_cpu_ticks_usec) {
current_cpu_ticks_usec = last_cpu_ticks_usec = p_cpu_ticks_usec;
}
// set measured wall clock time
void MainTimerSync::set_cpu_ticks_usec(uint64_t p_cpu_ticks_usec) {
current_cpu_ticks_usec = p_cpu_ticks_usec;
}
void MainTimerSync::set_fixed_fps(int p_fixed_fps) {
fixed_fps = p_fixed_fps;
}
// advance one frame, return timesteps to take
MainFrameTime MainTimerSync::advance(float p_frame_slice, int p_iterations_per_second) {
float cpu_idle_step = get_cpu_idle_step();
return advance_checked(p_frame_slice, p_iterations_per_second, cpu_idle_step);
}