blob: 71f8f5764df6e2fcc316fc4666490e3c1c378d80 [file] [log] [blame]
/*
* Copyright (c) 2013 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "webrtc/video_engine/overuse_frame_detector.h"
#include <assert.h>
#include <math.h>
#include <algorithm>
#include <list>
#include <map>
#include "webrtc/base/checks.h"
#include "webrtc/base/exp_filter.h"
#include "webrtc/system_wrappers/interface/clock.h"
#include "webrtc/system_wrappers/interface/logging.h"
namespace webrtc {
// TODO(mflodman) Test different values for all of these to trigger correctly,
// avoid fluctuations etc.
namespace {
const int64_t kProcessIntervalMs = 5000;
// Weight factor to apply to the standard deviation.
const float kWeightFactor = 0.997f;
// Weight factor to apply to the average.
const float kWeightFactorMean = 0.98f;
// Delay between consecutive rampups. (Used for quick recovery.)
const int kQuickRampUpDelayMs = 10 * 1000;
// Delay between rampup attempts. Initially uses standard, scales up to max.
const int kStandardRampUpDelayMs = 40 * 1000;
const int kMaxRampUpDelayMs = 240 * 1000;
// Expontential back-off factor, to prevent annoying up-down behaviour.
const double kRampUpBackoffFactor = 2.0;
// Max number of overuses detected before always applying the rampup delay.
const int kMaxOverusesBeforeApplyRampupDelay = 4;
// The maximum exponent to use in VCMExpFilter.
const float kSampleDiffMs = 33.0f;
const float kMaxExp = 7.0f;
} // namespace
// TODO(asapersson): Remove this class. Not used.
Statistics::Statistics(const CpuOveruseOptions& options)
: sum_(0.0),
count_(0),
options_(options),
filtered_samples_(new rtc::ExpFilter(kWeightFactorMean)),
filtered_variance_(new rtc::ExpFilter(kWeightFactor)) {
Reset();
}
void Statistics::Reset() {
sum_ = 0.0;
count_ = 0;
filtered_variance_->Reset(kWeightFactor);
filtered_variance_->Apply(1.0f, InitialVariance());
}
void Statistics::AddSample(float sample_ms) {
sum_ += sample_ms;
++count_;
if (count_ < static_cast<uint32_t>(options_.min_frame_samples)) {
// Initialize filtered samples.
filtered_samples_->Reset(kWeightFactorMean);
filtered_samples_->Apply(1.0f, InitialMean());
return;
}
float exp = sample_ms / kSampleDiffMs;
exp = std::min(exp, kMaxExp);
filtered_samples_->Apply(exp, sample_ms);
filtered_variance_->Apply(exp, (sample_ms - filtered_samples_->filtered()) *
(sample_ms - filtered_samples_->filtered()));
}
float Statistics::InitialMean() const {
if (count_ == 0)
return 0.0;
return sum_ / count_;
}
float Statistics::InitialVariance() const {
// Start in between the underuse and overuse threshold.
float average_stddev = (options_.low_capture_jitter_threshold_ms +
options_.high_capture_jitter_threshold_ms) / 2.0f;
return average_stddev * average_stddev;
}
float Statistics::Mean() const { return filtered_samples_->filtered(); }
float Statistics::StdDev() const {
return sqrt(std::max(filtered_variance_->filtered(), 0.0f));
}
uint64_t Statistics::Count() const { return count_; }
// Class for calculating the average encode time.
class OveruseFrameDetector::EncodeTimeAvg {
public:
EncodeTimeAvg()
: kWeightFactor(0.5f),
kInitialAvgEncodeTimeMs(5.0f),
filtered_encode_time_ms_(new rtc::ExpFilter(kWeightFactor)) {
filtered_encode_time_ms_->Apply(1.0f, kInitialAvgEncodeTimeMs);
}
~EncodeTimeAvg() {}
void AddSample(float encode_time_ms, int64_t diff_last_sample_ms) {
float exp = diff_last_sample_ms / kSampleDiffMs;
exp = std::min(exp, kMaxExp);
filtered_encode_time_ms_->Apply(exp, encode_time_ms);
}
int Value() const {
return static_cast<int>(filtered_encode_time_ms_->filtered() + 0.5);
}
private:
const float kWeightFactor;
const float kInitialAvgEncodeTimeMs;
rtc::scoped_ptr<rtc::ExpFilter> filtered_encode_time_ms_;
};
// Class for calculating the processing usage on the send-side (the average
// processing time of a frame divided by the average time difference between
// captured frames).
class OveruseFrameDetector::SendProcessingUsage {
public:
explicit SendProcessingUsage(const CpuOveruseOptions& options)
: kWeightFactorFrameDiff(0.998f),
kWeightFactorProcessing(0.995f),
kInitialSampleDiffMs(40.0f),
kMaxSampleDiffMs(45.0f),
count_(0),
options_(options),
filtered_processing_ms_(new rtc::ExpFilter(kWeightFactorProcessing)),
filtered_frame_diff_ms_(new rtc::ExpFilter(kWeightFactorFrameDiff)) {
Reset();
}
~SendProcessingUsage() {}
void Reset() {
count_ = 0;
filtered_frame_diff_ms_->Reset(kWeightFactorFrameDiff);
filtered_frame_diff_ms_->Apply(1.0f, kInitialSampleDiffMs);
filtered_processing_ms_->Reset(kWeightFactorProcessing);
filtered_processing_ms_->Apply(1.0f, InitialProcessingMs());
}
void AddCaptureSample(float sample_ms) {
float exp = sample_ms / kSampleDiffMs;
exp = std::min(exp, kMaxExp);
filtered_frame_diff_ms_->Apply(exp, sample_ms);
}
void AddSample(float processing_ms, int64_t diff_last_sample_ms) {
++count_;
float exp = diff_last_sample_ms / kSampleDiffMs;
exp = std::min(exp, kMaxExp);
filtered_processing_ms_->Apply(exp, processing_ms);
}
int Value() const {
if (count_ < static_cast<uint32_t>(options_.min_frame_samples)) {
return static_cast<int>(InitialUsageInPercent() + 0.5f);
}
float frame_diff_ms = std::max(filtered_frame_diff_ms_->filtered(), 1.0f);
frame_diff_ms = std::min(frame_diff_ms, kMaxSampleDiffMs);
float encode_usage_percent =
100.0f * filtered_processing_ms_->filtered() / frame_diff_ms;
return static_cast<int>(encode_usage_percent + 0.5);
}
private:
float InitialUsageInPercent() const {
// Start in between the underuse and overuse threshold.
return (options_.low_encode_usage_threshold_percent +
options_.high_encode_usage_threshold_percent) / 2.0f;
}
float InitialProcessingMs() const {
return InitialUsageInPercent() * kInitialSampleDiffMs / 100;
}
const float kWeightFactorFrameDiff;
const float kWeightFactorProcessing;
const float kInitialSampleDiffMs;
const float kMaxSampleDiffMs;
uint64_t count_;
const CpuOveruseOptions options_;
rtc::scoped_ptr<rtc::ExpFilter> filtered_processing_ms_;
rtc::scoped_ptr<rtc::ExpFilter> filtered_frame_diff_ms_;
};
// Class for calculating the processing time of frames.
class OveruseFrameDetector::FrameQueue {
public:
FrameQueue() : last_processing_time_ms_(-1) {}
~FrameQueue() {}
// Called when a frame is captured.
// Starts the measuring of the processing time of the frame.
void Start(int64_t capture_time, int64_t now) {
const size_t kMaxSize = 90; // Allows for processing time of 1.5s at 60fps.
if (frame_times_.size() > kMaxSize) {
LOG(LS_WARNING) << "Max size reached, removed oldest frame.";
frame_times_.erase(frame_times_.begin());
}
if (frame_times_.find(capture_time) != frame_times_.end()) {
// Frame should not exist.
assert(false);
return;
}
frame_times_[capture_time] = now;
}
// Called when the processing of a frame has finished.
// Returns the processing time of the frame.
int End(int64_t capture_time, int64_t now) {
std::map<int64_t, int64_t>::iterator it = frame_times_.find(capture_time);
if (it == frame_times_.end()) {
return -1;
}
// Remove any old frames up to current.
// Old frames have been skipped by the capture process thread.
// TODO(asapersson): Consider measuring time from first frame in list.
last_processing_time_ms_ = now - (*it).second;
frame_times_.erase(frame_times_.begin(), ++it);
return last_processing_time_ms_;
}
void Reset() { frame_times_.clear(); }
int NumFrames() const { return static_cast<int>(frame_times_.size()); }
int last_processing_time_ms() const { return last_processing_time_ms_; }
private:
// Captured frames mapped by the capture time.
std::map<int64_t, int64_t> frame_times_;
int last_processing_time_ms_;
};
// TODO(asapersson): Remove this class. Not used.
// Class for calculating the capture queue delay change.
class OveruseFrameDetector::CaptureQueueDelay {
public:
CaptureQueueDelay()
: kWeightFactor(0.5f),
delay_ms_(0),
filtered_delay_ms_per_s_(new rtc::ExpFilter(kWeightFactor)) {
filtered_delay_ms_per_s_->Apply(1.0f, 0.0f);
}
~CaptureQueueDelay() {}
void FrameCaptured(int64_t now) {
const size_t kMaxSize = 200;
if (frames_.size() > kMaxSize) {
frames_.pop_front();
}
frames_.push_back(now);
}
void FrameProcessingStarted(int64_t now) {
if (frames_.empty()) {
return;
}
delay_ms_ = now - frames_.front();
frames_.pop_front();
}
void CalculateDelayChange(int64_t diff_last_sample_ms) {
if (diff_last_sample_ms <= 0) {
return;
}
float exp = static_cast<float>(diff_last_sample_ms) / kProcessIntervalMs;
exp = std::min(exp, kMaxExp);
filtered_delay_ms_per_s_->Apply(exp,
delay_ms_ * 1000.0f / diff_last_sample_ms);
ClearFrames();
}
void ClearFrames() {
frames_.clear();
}
int delay_ms() const {
return delay_ms_;
}
int Value() const {
return static_cast<int>(filtered_delay_ms_per_s_->filtered() + 0.5);
}
private:
const float kWeightFactor;
std::list<int64_t> frames_;
int delay_ms_;
rtc::scoped_ptr<rtc::ExpFilter> filtered_delay_ms_per_s_;
};
OveruseFrameDetector::OveruseFrameDetector(
Clock* clock,
const CpuOveruseOptions& options,
CpuOveruseObserver* observer,
CpuOveruseMetricsObserver* metrics_observer)
: options_(options),
observer_(observer),
metrics_observer_(metrics_observer),
clock_(clock),
next_process_time_(clock_->TimeInMilliseconds()),
num_process_times_(0),
capture_deltas_(options),
last_capture_time_(0),
last_overuse_time_(0),
checks_above_threshold_(0),
num_overuse_detections_(0),
last_rampup_time_(0),
in_quick_rampup_(false),
current_rampup_delay_ms_(kStandardRampUpDelayMs),
num_pixels_(0),
last_encode_sample_ms_(0),
encode_time_(new EncodeTimeAvg()),
usage_(new SendProcessingUsage(options)),
frame_queue_(new FrameQueue()),
last_sample_time_ms_(0),
capture_queue_delay_(new CaptureQueueDelay()) {
DCHECK(metrics_observer != nullptr);
// Make sure stats are initially up-to-date. This simplifies unit testing
// since we don't have to trigger an update using one of the methods which
// would also alter the overuse state.
UpdateCpuOveruseMetrics();
processing_thread_.DetachFromThread();
}
OveruseFrameDetector::~OveruseFrameDetector() {
}
int OveruseFrameDetector::CaptureQueueDelayMsPerS() const {
rtc::CritScope cs(&crit_);
return capture_queue_delay_->delay_ms();
}
int OveruseFrameDetector::LastProcessingTimeMs() const {
rtc::CritScope cs(&crit_);
return frame_queue_->last_processing_time_ms();
}
int OveruseFrameDetector::FramesInQueue() const {
rtc::CritScope cs(&crit_);
return frame_queue_->NumFrames();
}
void OveruseFrameDetector::UpdateCpuOveruseMetrics() {
metrics_.capture_jitter_ms = static_cast<int>(capture_deltas_.StdDev() + 0.5);
metrics_.avg_encode_time_ms = encode_time_->Value();
metrics_.encode_usage_percent = usage_->Value();
metrics_.capture_queue_delay_ms_per_s = capture_queue_delay_->Value();
metrics_observer_->CpuOveruseMetricsUpdated(metrics_);
}
int64_t OveruseFrameDetector::TimeUntilNextProcess() {
DCHECK(processing_thread_.CalledOnValidThread());
return next_process_time_ - clock_->TimeInMilliseconds();
}
bool OveruseFrameDetector::FrameSizeChanged(int num_pixels) const {
if (num_pixels != num_pixels_) {
return true;
}
return false;
}
bool OveruseFrameDetector::FrameTimeoutDetected(int64_t now) const {
if (last_capture_time_ == 0) {
return false;
}
return (now - last_capture_time_) > options_.frame_timeout_interval_ms;
}
void OveruseFrameDetector::ResetAll(int num_pixels) {
num_pixels_ = num_pixels;
capture_deltas_.Reset();
usage_->Reset();
frame_queue_->Reset();
capture_queue_delay_->ClearFrames();
last_capture_time_ = 0;
num_process_times_ = 0;
UpdateCpuOveruseMetrics();
}
void OveruseFrameDetector::FrameCaptured(int width,
int height,
int64_t capture_time_ms) {
rtc::CritScope cs(&crit_);
int64_t now = clock_->TimeInMilliseconds();
if (FrameSizeChanged(width * height) || FrameTimeoutDetected(now)) {
ResetAll(width * height);
}
if (last_capture_time_ != 0) {
capture_deltas_.AddSample(now - last_capture_time_);
usage_->AddCaptureSample(now - last_capture_time_);
}
last_capture_time_ = now;
capture_queue_delay_->FrameCaptured(now);
if (options_.enable_extended_processing_usage) {
frame_queue_->Start(capture_time_ms, now);
}
UpdateCpuOveruseMetrics();
}
void OveruseFrameDetector::FrameProcessingStarted() {
rtc::CritScope cs(&crit_);
capture_queue_delay_->FrameProcessingStarted(clock_->TimeInMilliseconds());
}
void OveruseFrameDetector::FrameEncoded(int encode_time_ms) {
rtc::CritScope cs(&crit_);
int64_t now = clock_->TimeInMilliseconds();
if (last_encode_sample_ms_ != 0) {
int64_t diff_ms = now - last_encode_sample_ms_;
encode_time_->AddSample(encode_time_ms, diff_ms);
}
last_encode_sample_ms_ = now;
if (!options_.enable_extended_processing_usage) {
AddProcessingTime(encode_time_ms);
}
UpdateCpuOveruseMetrics();
}
void OveruseFrameDetector::FrameSent(int64_t capture_time_ms) {
rtc::CritScope cs(&crit_);
if (!options_.enable_extended_processing_usage) {
return;
}
int delay_ms = frame_queue_->End(capture_time_ms,
clock_->TimeInMilliseconds());
if (delay_ms > 0) {
AddProcessingTime(delay_ms);
}
UpdateCpuOveruseMetrics();
}
void OveruseFrameDetector::AddProcessingTime(int elapsed_ms) {
int64_t now = clock_->TimeInMilliseconds();
if (last_sample_time_ms_ != 0) {
int64_t diff_ms = now - last_sample_time_ms_;
usage_->AddSample(elapsed_ms, diff_ms);
}
last_sample_time_ms_ = now;
}
int32_t OveruseFrameDetector::Process() {
DCHECK(processing_thread_.CalledOnValidThread());
int64_t now = clock_->TimeInMilliseconds();
// Used to protect against Process() being called too often.
if (now < next_process_time_)
return 0;
int64_t diff_ms = now - next_process_time_ + kProcessIntervalMs;
next_process_time_ = now + kProcessIntervalMs;
rtc::CritScope cs(&crit_);
++num_process_times_;
capture_queue_delay_->CalculateDelayChange(diff_ms);
UpdateCpuOveruseMetrics();
if (num_process_times_ <= options_.min_process_count) {
return 0;
}
if (IsOverusing()) {
// If the last thing we did was going up, and now have to back down, we need
// to check if this peak was short. If so we should back off to avoid going
// back and forth between this load, the system doesn't seem to handle it.
bool check_for_backoff = last_rampup_time_ > last_overuse_time_;
if (check_for_backoff) {
if (now - last_rampup_time_ < kStandardRampUpDelayMs ||
num_overuse_detections_ > kMaxOverusesBeforeApplyRampupDelay) {
// Going up was not ok for very long, back off.
current_rampup_delay_ms_ *= kRampUpBackoffFactor;
if (current_rampup_delay_ms_ > kMaxRampUpDelayMs)
current_rampup_delay_ms_ = kMaxRampUpDelayMs;
} else {
// Not currently backing off, reset rampup delay.
current_rampup_delay_ms_ = kStandardRampUpDelayMs;
}
}
last_overuse_time_ = now;
in_quick_rampup_ = false;
checks_above_threshold_ = 0;
++num_overuse_detections_;
if (observer_ != NULL)
observer_->OveruseDetected();
} else if (IsUnderusing(now)) {
last_rampup_time_ = now;
in_quick_rampup_ = true;
if (observer_ != NULL)
observer_->NormalUsage();
}
int rampup_delay =
in_quick_rampup_ ? kQuickRampUpDelayMs : current_rampup_delay_ms_;
LOG(LS_VERBOSE) << " Frame stats: capture avg: " << capture_deltas_.Mean()
<< " capture stddev " << capture_deltas_.StdDev()
<< " encode usage " << usage_->Value()
<< " overuse detections " << num_overuse_detections_
<< " rampup delay " << rampup_delay;
return 0;
}
bool OveruseFrameDetector::IsOverusing() {
bool overusing = false;
if (options_.enable_capture_jitter_method) {
overusing = capture_deltas_.StdDev() >=
options_.high_capture_jitter_threshold_ms;
} else if (options_.enable_encode_usage_method) {
overusing = usage_->Value() >= options_.high_encode_usage_threshold_percent;
}
if (overusing) {
++checks_above_threshold_;
} else {
checks_above_threshold_ = 0;
}
return checks_above_threshold_ >= options_.high_threshold_consecutive_count;
}
bool OveruseFrameDetector::IsUnderusing(int64_t time_now) {
int delay = in_quick_rampup_ ? kQuickRampUpDelayMs : current_rampup_delay_ms_;
if (time_now < last_rampup_time_ + delay)
return false;
bool underusing = false;
if (options_.enable_capture_jitter_method) {
underusing = capture_deltas_.StdDev() <
options_.low_capture_jitter_threshold_ms;
} else if (options_.enable_encode_usage_method) {
underusing = usage_->Value() < options_.low_encode_usage_threshold_percent;
}
return underusing;
}
} // namespace webrtc