Add max pre-decode queue size threshold for pacing

When pacing is enabled for the low latency rendering path,
frames are sent to the decoder in regular intervals. In case of a
jitter, these frames intervals could add up to create a large latency.
Hence, disable frame pacing if the pre-decode queue grows beyond the
threshold. The threshold for when to disable frame pacing is set
through a field trial. The default value is high enough so that
the behavior is not changed unless the field trial is specified.

Bug: chromium:1237402
Change-Id: I901fd579f68da286eca3d654118f60d3c55e21ce
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/228241
Reviewed-by: Ilya Nikolaevskiy <ilnik@webrtc.org>
Commit-Queue: Johannes Kron <kron@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#34705}
This commit is contained in:
Johannes Kron
2021-08-10 16:56:12 +02:00
committed by WebRTC LUCI CQ
parent 5653c95ca2
commit 2ddc39e2b9
7 changed files with 119 additions and 28 deletions

View File

@ -63,7 +63,11 @@ FrameBuffer::FrameBuffer(Clock* clock,
last_log_non_decoded_ms_(-kLogNonDecodedIntervalMs),
add_rtt_to_playout_delay_(
webrtc::field_trial::IsEnabled("WebRTC-AddRttToPlayoutDelay")),
rtt_mult_settings_(RttMultExperiment::GetRttMultValue()) {
rtt_mult_settings_(RttMultExperiment::GetRttMultValue()),
zero_playout_delay_max_decode_queue_size_("max_decode_queue_size",
kMaxFramesBuffered) {
ParseFieldTrial({&zero_playout_delay_max_decode_queue_size_},
field_trial::FindFullName("WebRTC-ZeroPlayoutDelay"));
callback_checker_.Detach();
}
@ -212,7 +216,11 @@ int64_t FrameBuffer::FindNextFrame(int64_t now_ms) {
if (frame->RenderTime() == -1) {
frame->SetRenderTime(timing_->RenderTimeMs(frame->Timestamp(), now_ms));
}
wait_ms = timing_->MaxWaitingTime(frame->RenderTime(), now_ms);
bool too_many_frames_queued =
frames_.size() > zero_playout_delay_max_decode_queue_size_ ? true
: false;
wait_ms = timing_->MaxWaitingTime(frame->RenderTime(), now_ms,
too_many_frames_queued);
// This will cause the frame buffer to prefer high framerate rather
// than high resolution in the case of the decoder not decoding fast

View File

@ -25,6 +25,7 @@
#include "modules/video_coding/jitter_estimator.h"
#include "modules/video_coding/utility/decoded_frames_history.h"
#include "rtc_base/event.h"
#include "rtc_base/experiments/field_trial_parser.h"
#include "rtc_base/experiments/rtt_mult_experiment.h"
#include "rtc_base/numerics/sequence_number_util.h"
#include "rtc_base/synchronization/mutex.h"
@ -188,6 +189,13 @@ class FrameBuffer {
// rtt_mult experiment settings.
const absl::optional<RttMultExperiment::Settings> rtt_mult_settings_;
// Maximum number of frames in the decode queue to allow pacing. If the
// queue grows beyond the max limit, pacing will be disabled and frames will
// be pushed to the decoder as soon as possible. This only has an effect
// when the low-latency rendering path is active, which is indicated by
// the frame's render time == 0.
FieldTrialParameter<unsigned> zero_playout_delay_max_decode_queue_size_;
};
} // namespace video_coding

View File

@ -56,7 +56,8 @@ class VCMTimingFake : public VCMTiming {
}
int64_t MaxWaitingTime(int64_t render_time_ms,
int64_t now_ms) const override {
int64_t now_ms,
bool too_many_frames_queued) const override {
return render_time_ms - now_ms - kDecodeTime;
}

View File

@ -140,7 +140,8 @@ VCMEncodedFrame* VCMReceiver::FrameForDecoding(uint16_t max_wait_time_ms,
uint16_t new_max_wait_time =
static_cast<uint16_t>(VCM_MAX(available_wait_time, 0));
uint32_t wait_time_ms = rtc::saturated_cast<uint32_t>(
timing_->MaxWaitingTime(render_time_ms, clock_->TimeInMilliseconds()));
timing_->MaxWaitingTime(render_time_ms, clock_->TimeInMilliseconds(),
/*too_many_frames_queued=*/false));
if (new_max_wait_time < wait_time_ms) {
// We're not allowed to wait until the frame is supposed to be rendered,
// waiting as long as we're allowed to avoid busy looping, and then return

View File

@ -209,14 +209,19 @@ int VCMTiming::RequiredDecodeTimeMs() const {
}
int64_t VCMTiming::MaxWaitingTime(int64_t render_time_ms,
int64_t now_ms) const {
int64_t now_ms,
bool too_many_frames_queued) const {
MutexLock lock(&mutex_);
if (render_time_ms == 0 && zero_playout_delay_min_pacing_->us() > 0) {
// `render_time_ms` == 0 indicates that the frame should be decoded and
// rendered as soon as possible. However, the decoder can be choked if too
// many frames are sent at ones. Therefore, limit the interframe delay to
// `zero_playout_delay_min_pacing_`.
// many frames are sent at once. Therefore, limit the interframe delay to
// |zero_playout_delay_min_pacing_| unless too many frames are queued in
// which case the frames are sent to the decoder at once.
if (too_many_frames_queued) {
return 0;
}
int64_t earliest_next_decode_start_time =
last_decode_scheduled_ts_ + zero_playout_delay_min_pacing_->ms();
int64_t max_wait_time_ms = now_ms >= earliest_next_decode_start_time

View File

@ -82,8 +82,15 @@ class VCMTiming {
virtual int64_t RenderTimeMs(uint32_t frame_timestamp, int64_t now_ms) const;
// Returns the maximum time in ms that we can wait for a frame to become
// complete before we must pass it to the decoder.
virtual int64_t MaxWaitingTime(int64_t render_time_ms, int64_t now_ms) const;
// complete before we must pass it to the decoder. render_time_ms==0 indicates
// that the frames should be processed as quickly as possible, with possibly
// only a small delay added to make sure that the decoder is not overloaded.
// In this case, the parameter too_many_frames_queued is used to signal that
// the decode queue is full and that the frame should be decoded as soon as
// possible.
virtual int64_t MaxWaitingTime(int64_t render_time_ms,
int64_t now_ms,
bool too_many_frames_queued) const;
// Returns the current target delay which is required delay + decode time +
// render delay.

View File

@ -36,7 +36,7 @@ TEST(ReceiverTimingTest, JitterDelay) {
timing.set_render_delay(0);
uint32_t wait_time_ms = timing.MaxWaitingTime(
timing.RenderTimeMs(timestamp, clock.TimeInMilliseconds()),
clock.TimeInMilliseconds());
clock.TimeInMilliseconds(), /*too_many_frames_queued=*/false);
// First update initializes the render time. Since we have no decode delay
// we get wait_time_ms = renderTime - now - renderDelay = jitter.
EXPECT_EQ(jitter_delay_ms, wait_time_ms);
@ -48,7 +48,7 @@ TEST(ReceiverTimingTest, JitterDelay) {
timing.UpdateCurrentDelay(timestamp);
wait_time_ms = timing.MaxWaitingTime(
timing.RenderTimeMs(timestamp, clock.TimeInMilliseconds()),
clock.TimeInMilliseconds());
clock.TimeInMilliseconds(), /*too_many_frames_queued=*/false);
// Since we gradually increase the delay we only get 100 ms every second.
EXPECT_EQ(jitter_delay_ms - 10, wait_time_ms);
@ -57,7 +57,7 @@ TEST(ReceiverTimingTest, JitterDelay) {
timing.UpdateCurrentDelay(timestamp);
wait_time_ms = timing.MaxWaitingTime(
timing.RenderTimeMs(timestamp, clock.TimeInMilliseconds()),
clock.TimeInMilliseconds());
clock.TimeInMilliseconds(), /*too_many_frames_queued=*/false);
EXPECT_EQ(jitter_delay_ms, wait_time_ms);
// Insert frames without jitter, verify that this gives the exact wait time.
@ -70,7 +70,7 @@ TEST(ReceiverTimingTest, JitterDelay) {
timing.UpdateCurrentDelay(timestamp);
wait_time_ms = timing.MaxWaitingTime(
timing.RenderTimeMs(timestamp, clock.TimeInMilliseconds()),
clock.TimeInMilliseconds());
clock.TimeInMilliseconds(), /*too_many_frames_queued=*/false);
EXPECT_EQ(jitter_delay_ms, wait_time_ms);
// Add decode time estimates for 1 second.
@ -85,7 +85,7 @@ TEST(ReceiverTimingTest, JitterDelay) {
timing.UpdateCurrentDelay(timestamp);
wait_time_ms = timing.MaxWaitingTime(
timing.RenderTimeMs(timestamp, clock.TimeInMilliseconds()),
clock.TimeInMilliseconds());
clock.TimeInMilliseconds(), /*too_many_frames_queued=*/false);
EXPECT_EQ(jitter_delay_ms, wait_time_ms);
const int kMinTotalDelayMs = 200;
@ -97,7 +97,7 @@ TEST(ReceiverTimingTest, JitterDelay) {
timing.set_render_delay(kRenderDelayMs);
wait_time_ms = timing.MaxWaitingTime(
timing.RenderTimeMs(timestamp, clock.TimeInMilliseconds()),
clock.TimeInMilliseconds());
clock.TimeInMilliseconds(), /*too_many_frames_queued=*/false);
// We should at least have kMinTotalDelayMs - decodeTime (10) - renderTime
// (10) to wait.
EXPECT_EQ(kMinTotalDelayMs - kDecodeTimeMs - kRenderDelayMs, wait_time_ms);
@ -140,16 +140,26 @@ TEST(ReceiverTimingTest, MaxWaitingTimeIsZeroForZeroRenderTime) {
for (int i = 0; i < 10; ++i) {
clock.AdvanceTimeMilliseconds(kTimeDeltaMs);
int64_t now_ms = clock.TimeInMilliseconds();
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), 0);
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
0);
}
// Another frame submitted at the same time also returns a negative max
// waiting time.
int64_t now_ms = clock.TimeInMilliseconds();
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), 0);
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
0);
// MaxWaitingTime should be less than zero even if there's a burst of frames.
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), 0);
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), 0);
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), 0);
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
0);
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
0);
EXPECT_LT(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
0);
}
TEST(ReceiverTimingTest, MaxWaitingTimeZeroDelayPacingExperiment) {
@ -168,27 +178,38 @@ TEST(ReceiverTimingTest, MaxWaitingTimeZeroDelayPacingExperiment) {
for (int i = 0; i < 10; ++i) {
clock.AdvanceTimeMilliseconds(kTimeDeltaMs);
int64_t now_ms = clock.TimeInMilliseconds();
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), 0);
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
0);
timing.SetLastDecodeScheduledTimestamp(now_ms);
}
// Another frame submitted at the same time is paced according to the field
// trial setting.
int64_t now_ms = clock.TimeInMilliseconds();
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), kMinPacingMs);
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
kMinPacingMs);
// If there's a burst of frames, the wait time is calculated based on next
// decode time.
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), kMinPacingMs);
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), kMinPacingMs);
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
kMinPacingMs);
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
kMinPacingMs);
// Allow a few ms to pass, this should be subtracted from the MaxWaitingTime.
constexpr int64_t kTwoMs = 2;
clock.AdvanceTimeMilliseconds(kTwoMs);
now_ms = clock.TimeInMilliseconds();
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms),
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
kMinPacingMs - kTwoMs);
// A frame is decoded at the current time, the wait time should be restored to
// pacing delay.
timing.SetLastDecodeScheduledTimestamp(now_ms);
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms), kMinPacingMs);
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
kMinPacingMs);
}
TEST(ReceiverTimingTest, DefaultMaxWaitingTimeUnaffectedByPacingExperiment) {
@ -206,16 +227,56 @@ TEST(ReceiverTimingTest, DefaultMaxWaitingTimeUnaffectedByPacingExperiment) {
int64_t render_time_ms = now_ms + 30;
// Estimate the internal processing delay from the first frame.
int64_t estimated_processing_delay =
(render_time_ms - now_ms) - timing.MaxWaitingTime(render_time_ms, now_ms);
(render_time_ms - now_ms) -
timing.MaxWaitingTime(render_time_ms, now_ms,
/*too_many_frames_queued=*/false);
EXPECT_GT(estimated_processing_delay, 0);
// Any other frame submitted at the same time should be scheduled according to
// its render time.
for (int i = 0; i < 5; ++i) {
render_time_ms += kTimeDeltaMs;
EXPECT_EQ(timing.MaxWaitingTime(render_time_ms, now_ms),
EXPECT_EQ(timing.MaxWaitingTime(render_time_ms, now_ms,
/*too_many_frames_queued=*/false),
render_time_ms - now_ms - estimated_processing_delay);
}
}
TEST(ReceiverTiminTest, MaxWaitingTimeReturnsZeroIfTooManyFramesQueuedIsTrue) {
// The minimum pacing is enabled by a field trial and active if the RTP
// playout delay header extension is set to min==0.
constexpr int64_t kMinPacingMs = 3;
test::ScopedFieldTrials override_field_trials(
"WebRTC-ZeroPlayoutDelay/min_pacing:3ms/");
constexpr int64_t kStartTimeUs = 3.15e13; // About one year in us.
constexpr int64_t kTimeDeltaMs = 1000.0 / 60.0;
constexpr int64_t kZeroRenderTimeMs = 0;
SimulatedClock clock(kStartTimeUs);
VCMTiming timing(&clock);
timing.Reset();
// MaxWaitingTime() returns zero for evenly spaced video frames.
for (int i = 0; i < 10; ++i) {
clock.AdvanceTimeMilliseconds(kTimeDeltaMs);
int64_t now_ms = clock.TimeInMilliseconds();
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
0);
timing.SetLastDecodeScheduledTimestamp(now_ms);
}
// Another frame submitted at the same time is paced according to the field
// trial setting.
int64_t now_ms = clock.TimeInMilliseconds();
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/false),
kMinPacingMs);
// MaxWaitingTime returns 0 even if there's a burst of frames if
// too_many_frames_queued is set to true.
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/true),
0);
EXPECT_EQ(timing.MaxWaitingTime(kZeroRenderTimeMs, now_ms,
/*too_many_frames_queued=*/true),
0);
}
} // namespace webrtc