In (new) estimator of encoder cpu load, count max per input frame.

Bug: webrtc:9619
Change-Id: Ifc874fa3bd069e48a10fec57b673546aafd070e3
Reviewed-on: https://webrtc-review.googlesource.com/94143
Reviewed-by: Erik Språng <sprang@webrtc.org>
Commit-Queue: Niels Moller <nisse@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#24293}
This commit is contained in:
Niels Möller 2018-08-14 17:00:46 +02:00 committed by Commit Bot
parent a955849901
commit f5033ad201
2 changed files with 109 additions and 4 deletions

View File

@ -243,11 +243,13 @@ class SendProcessingUsage2 : public OveruseFrameDetector::ProcessingUsage {
int64_t last_capture_time_us) override {}
absl::optional<int> FrameSent(
uint32_t timestamp,
int64_t time_sent_in_us,
uint32_t /* timestamp */,
int64_t /* time_sent_in_us */,
int64_t capture_time_us,
absl::optional<int> encode_duration_us) override {
if (encode_duration_us) {
int duration_per_frame_us =
DurationPerInputFrame(capture_time_us, *encode_duration_us);
if (prev_time_us_ != -1) {
if (capture_time_us < prev_time_us_) {
// The weighting in AddSample assumes that samples are processed with
@ -257,7 +259,7 @@ class SendProcessingUsage2 : public OveruseFrameDetector::ProcessingUsage {
// bit forward in time.
capture_time_us = prev_time_us_;
}
AddSample(1e-6 * (*encode_duration_us),
AddSample(1e-6 * duration_per_frame_us,
1e-6 * (capture_time_us - prev_time_us_));
}
}
@ -287,12 +289,43 @@ class SendProcessingUsage2 : public OveruseFrameDetector::ProcessingUsage {
load_estimate_ = c * encode_time + exp(-e) * load_estimate_;
}
int64_t DurationPerInputFrame(int64_t capture_time_us,
int64_t encode_time_us) {
// Discard data on old frames; limit 2 seconds.
static constexpr int64_t kMaxAge = 2 * rtc::kNumMicrosecsPerSec;
for (auto it = max_encode_time_per_input_frame_.begin();
it != max_encode_time_per_input_frame_.end() &&
it->first < capture_time_us - kMaxAge;) {
it = max_encode_time_per_input_frame_.erase(it);
}
std::map<int64_t, int>::iterator it;
bool inserted;
std::tie(it, inserted) = max_encode_time_per_input_frame_.emplace(
capture_time_us, encode_time_us);
if (inserted) {
// First encoded frame for this input frame.
return encode_time_us;
}
if (encode_time_us <= it->second) {
// Shorter encode time than previous frame (unlikely). Count it as being
// done in parallel.
return 0;
}
// Record new maximum encode time, and return increase from previous max.
int increase = encode_time_us - it->second;
it->second = encode_time_us;
return increase;
}
int Value() override {
return static_cast<int>(100.0 * load_estimate_ + 0.5);
}
private:
const CpuOveruseOptions options_;
// Indexed by the capture timestamp, used as frame id.
std::map<int64_t, int> max_encode_time_per_input_frame_;
int64_t prev_time_us_ = -1;
double load_estimate_;
};

View File

@ -108,6 +108,36 @@ class OveruseFrameDetectorTest : public ::testing::Test,
}
}
virtual void InsertAndSendSimulcastFramesWithInterval(
int num_frames,
int interval_us,
int width,
int height,
// One element per layer
rtc::ArrayView<const int> delays_us) {
VideoFrame frame(I420Buffer::Create(width, height),
webrtc::kVideoRotation_0, 0);
uint32_t timestamp = 0;
while (num_frames-- > 0) {
frame.set_timestamp(timestamp);
int64_t capture_time_us = rtc::TimeMicros();
overuse_detector_->FrameCaptured(frame, capture_time_us);
int max_delay_us = 0;
for (int delay_us : delays_us) {
if (delay_us > max_delay_us) {
clock_.AdvanceTimeMicros(delay_us - max_delay_us);
max_delay_us = delay_us;
}
overuse_detector_->FrameSent(timestamp, rtc::TimeMicros(),
capture_time_us, delay_us);
}
overuse_detector_->CheckForOveruse(observer_);
clock_.AdvanceTimeMicros(interval_us - max_delay_us);
timestamp += interval_us * 90 / 1000;
}
}
virtual void InsertAndSendFramesWithRandomInterval(int num_frames,
int min_interval_us,
int max_interval_us,
@ -578,6 +608,27 @@ TEST_F(OveruseFrameDetectorTest, NoOveruseForRandomFrameIntervalWithReset) {
EXPECT_LE(UsagePercent(), InitialUsage() + 5);
}
// Models simulcast, with multiple encoded frames for each input frame.
// Load estimate should be based on the maximum encode time per input frame.
TEST_F(OveruseFrameDetectorTest, NoOveruseForSimulcast) {
overuse_detector_->SetOptions(options_);
EXPECT_CALL(mock_observer_, AdaptDown(_)).Times(0);
constexpr int kNumFrames = 500;
constexpr int kEncodeTimesUs[] = {
10 * rtc::kNumMicrosecsPerMillisec, 8 * rtc::kNumMicrosecsPerMillisec,
12 * rtc::kNumMicrosecsPerMillisec,
};
constexpr int kIntervalUs = 30 * rtc::kNumMicrosecsPerMillisec;
InsertAndSendSimulcastFramesWithInterval(kNumFrames, kIntervalUs, kWidth,
kHeight, kEncodeTimesUs);
// Average usage 40%. 12 ms / 30 ms.
EXPECT_GE(UsagePercent(), 35);
EXPECT_LE(UsagePercent(), 45);
}
// Tests using new cpu load estimator
class OveruseFrameDetectorTest2 : public OveruseFrameDetectorTest {
protected:
@ -905,4 +956,25 @@ TEST_F(OveruseFrameDetectorTest2, ToleratesOutOfOrderFrames) {
EXPECT_GE(UsagePercent(), InitialUsage());
}
// Models simulcast, with multiple encoded frames for each input frame.
// Load estimate should be based on the maximum encode time per input frame.
TEST_F(OveruseFrameDetectorTest2, NoOveruseForSimulcast) {
overuse_detector_->SetOptions(options_);
EXPECT_CALL(mock_observer_, AdaptDown(_)).Times(0);
constexpr int kNumFrames = 500;
constexpr int kEncodeTimesUs[] = {
10 * rtc::kNumMicrosecsPerMillisec, 8 * rtc::kNumMicrosecsPerMillisec,
12 * rtc::kNumMicrosecsPerMillisec,
};
constexpr int kIntervalUs = 30 * rtc::kNumMicrosecsPerMillisec;
InsertAndSendSimulcastFramesWithInterval(kNumFrames, kIntervalUs, kWidth,
kHeight, kEncodeTimesUs);
// Average usage 40%. 12 ms / 30 ms.
EXPECT_GE(UsagePercent(), 35);
EXPECT_LE(UsagePercent(), 45);
}
} // namespace webrtc