Implement ChunkingHandler
This handler is a multi-in multi-out handler. If more than one input is
provided, there should be one and only one video stream; also, all inputs
should come from the same thread and are synchronized.
There can be multiple chunking handler running in different threads or even
different processes, we use the "consistent chunking algorithm" to make sure
the chunks in different streams are aligned without explicit communcating
with each other - which is not efficient and often difficult.
Consistent Chunking Algorithm:
1. Find the consistent chunkable boundary
Let the timestamps for video frames be (t1, t2, t3, ...). Then a
consistent chunkable boundary is simply the first chunkable boundary after
(tk / N) != (tk-1 / N), where '/' denotes integer division, and N is the
intended chunk duration.
2. Chunk only at the consistent chunkable boundary
This algorithm will make sure the chunks from different video streams are
aligned if they have aligned GoPs. However, this algorithm will only work
for video streams. To be able to chunk non video streams at similar
positions as video streams, ChunkingHandler is designed to accept one video
input and multiple non video inputs, the non video inputs are chunked when
the video input is chunked. If the inputs are synchronized - which is true
if the inputs come from the same demuxer, the video and non video chunks
are aligned.
Change-Id: Id3bad51ab14f311efdb8713b6cd36d36cf9e4639
2017-02-07 18:58:47 +00:00
|
|
|
// Copyright 2017 Google Inc. All rights reserved.
|
|
|
|
//
|
|
|
|
// Use of this source code is governed by a BSD-style
|
|
|
|
// license that can be found in the LICENSE file or at
|
|
|
|
// https://developers.google.com/open-source/licenses/bsd
|
|
|
|
|
|
|
|
#include "packager/media/chunking/chunking_handler.h"
|
|
|
|
|
|
|
|
#include "packager/base/logging.h"
|
|
|
|
#include "packager/base/threading/platform_thread.h"
|
|
|
|
#include "packager/media/base/media_sample.h"
|
|
|
|
|
|
|
|
namespace {
|
|
|
|
int64_t kThreadIdUnset = -1;
|
|
|
|
int64_t kTimeStampToDispatchAllSamples = -1;
|
|
|
|
} // namespace
|
|
|
|
|
|
|
|
namespace shaka {
|
|
|
|
namespace media {
|
|
|
|
|
|
|
|
ChunkingHandler::ChunkingHandler(const ChunkingOptions& chunking_options)
|
|
|
|
: chunking_options_(chunking_options), thread_id_(kThreadIdUnset) {
|
|
|
|
CHECK_NE(chunking_options.segment_duration_in_seconds, 0u);
|
|
|
|
}
|
|
|
|
|
|
|
|
ChunkingHandler::~ChunkingHandler() {}
|
|
|
|
|
|
|
|
Status ChunkingHandler::InitializeInternal() {
|
|
|
|
segment_info_.resize(num_input_streams());
|
|
|
|
subsegment_info_.resize(num_input_streams());
|
|
|
|
time_scales_.resize(num_input_streams());
|
|
|
|
last_sample_end_timestamps_.resize(num_input_streams());
|
|
|
|
return Status::OK;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status ChunkingHandler::Process(std::unique_ptr<StreamData> stream_data) {
|
|
|
|
switch (stream_data->stream_data_type) {
|
|
|
|
case StreamDataType::kStreamInfo: {
|
|
|
|
// Make sure the inputs come from the same thread.
|
|
|
|
const int64_t thread_id =
|
|
|
|
static_cast<int64_t>(base::PlatformThread::CurrentId());
|
|
|
|
int64_t expected = kThreadIdUnset;
|
|
|
|
if (!thread_id_.compare_exchange_strong(expected, thread_id) &&
|
|
|
|
expected != thread_id) {
|
|
|
|
return Status(error::CHUNKING_ERROR,
|
|
|
|
"Inputs should come from the same thread.");
|
|
|
|
}
|
|
|
|
|
|
|
|
const auto time_scale = stream_data->stream_info->time_scale();
|
|
|
|
// The video stream is treated as the main stream. If there is only one
|
|
|
|
// stream, it is the main stream.
|
|
|
|
const bool is_main_stream =
|
|
|
|
main_stream_index_ == -1 &&
|
|
|
|
(stream_data->stream_info->stream_type() == kStreamVideo ||
|
|
|
|
num_input_streams() == 1);
|
|
|
|
if (is_main_stream) {
|
|
|
|
main_stream_index_ = stream_data->stream_index;
|
|
|
|
segment_duration_ =
|
|
|
|
chunking_options_.segment_duration_in_seconds * time_scale;
|
|
|
|
subsegment_duration_ =
|
|
|
|
chunking_options_.subsegment_duration_in_seconds * time_scale;
|
|
|
|
} else if (stream_data->stream_info->stream_type() == kStreamVideo) {
|
|
|
|
return Status(error::CHUNKING_ERROR,
|
|
|
|
"Only one video stream is allowed per chunking handler.");
|
|
|
|
}
|
|
|
|
time_scales_[stream_data->stream_index] = time_scale;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
case StreamDataType::kSegmentInfo:
|
|
|
|
VLOG(3) << "Drop existing segment info.";
|
|
|
|
return Status::OK;
|
|
|
|
case StreamDataType::kMediaSample: {
|
|
|
|
const int stream_index = stream_data->stream_index;
|
|
|
|
DCHECK_NE(time_scales_[stream_index], 0u)
|
|
|
|
<< "kStreamInfo should arrive before kMediaSample";
|
|
|
|
if (stream_index != main_stream_index_) {
|
|
|
|
if (!stream_data->media_sample->is_key_frame()) {
|
|
|
|
return Status(error::CHUNKING_ERROR,
|
|
|
|
"All non video samples should be key frames.");
|
|
|
|
}
|
|
|
|
// Cache non main stream samples, since we don't know yet whether these
|
|
|
|
// samples belong to the current or next segment.
|
|
|
|
non_main_samples_.push_back(std::move(stream_data));
|
|
|
|
// The streams are expected to be synchronized, so we don't expect to
|
|
|
|
// see a lot of samples before seeing video samples.
|
|
|
|
const size_t kMaxSamplesPerStreamBeforeVideoSample = 5u;
|
|
|
|
if (non_main_samples_.size() >
|
|
|
|
num_input_streams() * kMaxSamplesPerStreamBeforeVideoSample) {
|
|
|
|
return Status(error::CHUNKING_ERROR,
|
|
|
|
"Too many non video samples before video sample.");
|
|
|
|
}
|
|
|
|
return Status::OK;
|
|
|
|
}
|
|
|
|
|
|
|
|
const MediaSample* sample = stream_data->media_sample.get();
|
|
|
|
Status status = ProcessMediaSample(sample);
|
|
|
|
if (!status.ok())
|
|
|
|
return status;
|
|
|
|
// Discard samples before segment start.
|
|
|
|
if (!segment_info_[stream_index])
|
|
|
|
return Status::OK;
|
|
|
|
last_sample_end_timestamps_[stream_index] =
|
|
|
|
sample->dts() + sample->duration();
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
VLOG(3) << "Stream data type "
|
|
|
|
<< static_cast<int>(stream_data->stream_data_type) << " ignored.";
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
return Dispatch(std::move(stream_data));
|
|
|
|
}
|
|
|
|
|
2017-02-22 20:14:26 +00:00
|
|
|
Status ChunkingHandler::OnFlushRequest(int input_stream_index) {
|
Implement ChunkingHandler
This handler is a multi-in multi-out handler. If more than one input is
provided, there should be one and only one video stream; also, all inputs
should come from the same thread and are synchronized.
There can be multiple chunking handler running in different threads or even
different processes, we use the "consistent chunking algorithm" to make sure
the chunks in different streams are aligned without explicit communcating
with each other - which is not efficient and often difficult.
Consistent Chunking Algorithm:
1. Find the consistent chunkable boundary
Let the timestamps for video frames be (t1, t2, t3, ...). Then a
consistent chunkable boundary is simply the first chunkable boundary after
(tk / N) != (tk-1 / N), where '/' denotes integer division, and N is the
intended chunk duration.
2. Chunk only at the consistent chunkable boundary
This algorithm will make sure the chunks from different video streams are
aligned if they have aligned GoPs. However, this algorithm will only work
for video streams. To be able to chunk non video streams at similar
positions as video streams, ChunkingHandler is designed to accept one video
input and multiple non video inputs, the non video inputs are chunked when
the video input is chunked. If the inputs are synchronized - which is true
if the inputs come from the same demuxer, the video and non video chunks
are aligned.
Change-Id: Id3bad51ab14f311efdb8713b6cd36d36cf9e4639
2017-02-07 18:58:47 +00:00
|
|
|
if (segment_info_[input_stream_index]) {
|
|
|
|
Status status;
|
|
|
|
if (input_stream_index != main_stream_index_) {
|
|
|
|
status = DispatchNonMainSamples(kTimeStampToDispatchAllSamples);
|
|
|
|
if (!status.ok())
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
auto& segment_info = segment_info_[input_stream_index];
|
|
|
|
if (segment_info->start_timestamp != -1) {
|
|
|
|
segment_info->duration = last_sample_end_timestamps_[input_stream_index] -
|
|
|
|
segment_info->start_timestamp;
|
|
|
|
status = DispatchSegmentInfo(input_stream_index, std::move(segment_info));
|
|
|
|
if (!status.ok())
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
}
|
2017-02-22 20:14:26 +00:00
|
|
|
const int output_stream_index = input_stream_index;
|
|
|
|
return FlushDownstream(output_stream_index);
|
Implement ChunkingHandler
This handler is a multi-in multi-out handler. If more than one input is
provided, there should be one and only one video stream; also, all inputs
should come from the same thread and are synchronized.
There can be multiple chunking handler running in different threads or even
different processes, we use the "consistent chunking algorithm" to make sure
the chunks in different streams are aligned without explicit communcating
with each other - which is not efficient and often difficult.
Consistent Chunking Algorithm:
1. Find the consistent chunkable boundary
Let the timestamps for video frames be (t1, t2, t3, ...). Then a
consistent chunkable boundary is simply the first chunkable boundary after
(tk / N) != (tk-1 / N), where '/' denotes integer division, and N is the
intended chunk duration.
2. Chunk only at the consistent chunkable boundary
This algorithm will make sure the chunks from different video streams are
aligned if they have aligned GoPs. However, this algorithm will only work
for video streams. To be able to chunk non video streams at similar
positions as video streams, ChunkingHandler is designed to accept one video
input and multiple non video inputs, the non video inputs are chunked when
the video input is chunked. If the inputs are synchronized - which is true
if the inputs come from the same demuxer, the video and non video chunks
are aligned.
Change-Id: Id3bad51ab14f311efdb8713b6cd36d36cf9e4639
2017-02-07 18:58:47 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
Status ChunkingHandler::ProcessMediaSample(const MediaSample* sample) {
|
|
|
|
const bool is_key_frame = sample->is_key_frame();
|
|
|
|
const int64_t timestamp = sample->dts();
|
|
|
|
// Check if we need to terminate the current (sub)segment.
|
|
|
|
bool new_segment = false;
|
|
|
|
bool new_subsegment = false;
|
|
|
|
if (is_key_frame || !chunking_options_.segment_sap_aligned) {
|
|
|
|
const int64_t segment_index = timestamp / segment_duration_;
|
|
|
|
if (segment_index != current_segment_index_) {
|
|
|
|
current_segment_index_ = segment_index;
|
2017-02-24 01:17:47 +00:00
|
|
|
// Reset subsegment index.
|
|
|
|
current_subsegment_index_ = 0;
|
Implement ChunkingHandler
This handler is a multi-in multi-out handler. If more than one input is
provided, there should be one and only one video stream; also, all inputs
should come from the same thread and are synchronized.
There can be multiple chunking handler running in different threads or even
different processes, we use the "consistent chunking algorithm" to make sure
the chunks in different streams are aligned without explicit communcating
with each other - which is not efficient and often difficult.
Consistent Chunking Algorithm:
1. Find the consistent chunkable boundary
Let the timestamps for video frames be (t1, t2, t3, ...). Then a
consistent chunkable boundary is simply the first chunkable boundary after
(tk / N) != (tk-1 / N), where '/' denotes integer division, and N is the
intended chunk duration.
2. Chunk only at the consistent chunkable boundary
This algorithm will make sure the chunks from different video streams are
aligned if they have aligned GoPs. However, this algorithm will only work
for video streams. To be able to chunk non video streams at similar
positions as video streams, ChunkingHandler is designed to accept one video
input and multiple non video inputs, the non video inputs are chunked when
the video input is chunked. If the inputs are synchronized - which is true
if the inputs come from the same demuxer, the video and non video chunks
are aligned.
Change-Id: Id3bad51ab14f311efdb8713b6cd36d36cf9e4639
2017-02-07 18:58:47 +00:00
|
|
|
new_segment = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (!new_segment && subsegment_duration_ > 0 &&
|
|
|
|
(is_key_frame || !chunking_options_.subsegment_sap_aligned)) {
|
|
|
|
const int64_t subsegment_index =
|
|
|
|
(timestamp - segment_info_[main_stream_index_]->start_timestamp) /
|
|
|
|
subsegment_duration_;
|
|
|
|
if (subsegment_index != current_subsegment_index_) {
|
|
|
|
current_subsegment_index_ = subsegment_index;
|
|
|
|
new_subsegment = true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Status status;
|
|
|
|
if (new_segment || new_subsegment) {
|
|
|
|
// Dispatch the samples before |timestamp| - See the implemention on how we
|
|
|
|
// determine if a sample is before |timestamp|..
|
|
|
|
status.Update(DispatchNonMainSamples(timestamp));
|
|
|
|
}
|
|
|
|
|
|
|
|
if (new_segment) {
|
|
|
|
status.Update(DispatchSegmentInfoForAllStreams());
|
|
|
|
segment_info_[main_stream_index_]->start_timestamp = timestamp;
|
|
|
|
}
|
|
|
|
if (subsegment_duration_ > 0 && (new_segment || new_subsegment)) {
|
|
|
|
status.Update(DispatchSubsegmentInfoForAllStreams());
|
|
|
|
subsegment_info_[main_stream_index_]->start_timestamp = timestamp;
|
|
|
|
}
|
|
|
|
if (!status.ok())
|
|
|
|
return status;
|
|
|
|
|
|
|
|
// Dispatch non-main samples for the next segment.
|
|
|
|
return DispatchNonMainSamples(kTimeStampToDispatchAllSamples);
|
|
|
|
}
|
|
|
|
|
|
|
|
Status ChunkingHandler::DispatchNonMainSamples(int64_t timestamp_threshold) {
|
|
|
|
Status status;
|
|
|
|
while (status.ok() && !non_main_samples_.empty()) {
|
|
|
|
DCHECK_EQ(non_main_samples_.front()->stream_data_type,
|
|
|
|
StreamDataType::kMediaSample);
|
|
|
|
const int stream_index = non_main_samples_.front()->stream_index;
|
|
|
|
const MediaSample* sample = non_main_samples_.front()->media_sample.get();
|
|
|
|
// If the portion of the sample before |timestamp_threshold| is bigger than
|
|
|
|
// the other portion, we consider it part of the current segment.
|
|
|
|
const int64_t timestamp = sample->dts() + sample->duration() / 2;
|
|
|
|
const bool stop =
|
|
|
|
(timestamp_threshold != kTimeStampToDispatchAllSamples &&
|
|
|
|
(static_cast<double>(timestamp) / time_scales_[stream_index]) >
|
|
|
|
(static_cast<double>(timestamp_threshold) /
|
|
|
|
time_scales_[main_stream_index_]));
|
|
|
|
VLOG(3) << "Sample ts: " << sample->dts() << " "
|
|
|
|
<< " duration: " << sample->duration()
|
|
|
|
<< " scale: " << time_scales_[stream_index] << "\n"
|
|
|
|
<< " threshold: " << timestamp_threshold
|
|
|
|
<< " scale: " << time_scales_[main_stream_index_]
|
|
|
|
<< (stop ? " stop "
|
|
|
|
: (segment_info_[stream_index] ? " dispatch "
|
|
|
|
: " discard "));
|
|
|
|
if (stop)
|
|
|
|
break;
|
|
|
|
// Only dispatch samples if the segment has started, otherwise discard
|
|
|
|
// them.
|
|
|
|
if (segment_info_[stream_index]) {
|
|
|
|
if (segment_info_[stream_index]->start_timestamp == -1)
|
|
|
|
segment_info_[stream_index]->start_timestamp = sample->dts();
|
|
|
|
if (subsegment_info_[stream_index] &&
|
|
|
|
subsegment_info_[stream_index]->start_timestamp == -1) {
|
|
|
|
subsegment_info_[stream_index]->start_timestamp = sample->dts();
|
|
|
|
}
|
|
|
|
last_sample_end_timestamps_[stream_index] =
|
|
|
|
sample->dts() + sample->duration();
|
|
|
|
status.Update(Dispatch(std::move(non_main_samples_.front())));
|
|
|
|
}
|
|
|
|
non_main_samples_.pop_front();
|
|
|
|
}
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status ChunkingHandler::DispatchSegmentInfoForAllStreams() {
|
|
|
|
Status status;
|
2017-02-11 06:57:32 +00:00
|
|
|
for (int i = 0; i < static_cast<int>(segment_info_.size()) && status.ok();
|
|
|
|
++i) {
|
Implement ChunkingHandler
This handler is a multi-in multi-out handler. If more than one input is
provided, there should be one and only one video stream; also, all inputs
should come from the same thread and are synchronized.
There can be multiple chunking handler running in different threads or even
different processes, we use the "consistent chunking algorithm" to make sure
the chunks in different streams are aligned without explicit communcating
with each other - which is not efficient and often difficult.
Consistent Chunking Algorithm:
1. Find the consistent chunkable boundary
Let the timestamps for video frames be (t1, t2, t3, ...). Then a
consistent chunkable boundary is simply the first chunkable boundary after
(tk / N) != (tk-1 / N), where '/' denotes integer division, and N is the
intended chunk duration.
2. Chunk only at the consistent chunkable boundary
This algorithm will make sure the chunks from different video streams are
aligned if they have aligned GoPs. However, this algorithm will only work
for video streams. To be able to chunk non video streams at similar
positions as video streams, ChunkingHandler is designed to accept one video
input and multiple non video inputs, the non video inputs are chunked when
the video input is chunked. If the inputs are synchronized - which is true
if the inputs come from the same demuxer, the video and non video chunks
are aligned.
Change-Id: Id3bad51ab14f311efdb8713b6cd36d36cf9e4639
2017-02-07 18:58:47 +00:00
|
|
|
if (segment_info_[i] && segment_info_[i]->start_timestamp != -1) {
|
|
|
|
segment_info_[i]->duration =
|
|
|
|
last_sample_end_timestamps_[i] - segment_info_[i]->start_timestamp;
|
|
|
|
status.Update(DispatchSegmentInfo(i, std::move(segment_info_[i])));
|
|
|
|
}
|
|
|
|
segment_info_[i].reset(new SegmentInfo);
|
|
|
|
subsegment_info_[i].reset();
|
|
|
|
}
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
|
|
|
Status ChunkingHandler::DispatchSubsegmentInfoForAllStreams() {
|
|
|
|
Status status;
|
2017-02-11 06:57:32 +00:00
|
|
|
for (int i = 0; i < static_cast<int>(subsegment_info_.size()) && status.ok();
|
|
|
|
++i) {
|
Implement ChunkingHandler
This handler is a multi-in multi-out handler. If more than one input is
provided, there should be one and only one video stream; also, all inputs
should come from the same thread and are synchronized.
There can be multiple chunking handler running in different threads or even
different processes, we use the "consistent chunking algorithm" to make sure
the chunks in different streams are aligned without explicit communcating
with each other - which is not efficient and often difficult.
Consistent Chunking Algorithm:
1. Find the consistent chunkable boundary
Let the timestamps for video frames be (t1, t2, t3, ...). Then a
consistent chunkable boundary is simply the first chunkable boundary after
(tk / N) != (tk-1 / N), where '/' denotes integer division, and N is the
intended chunk duration.
2. Chunk only at the consistent chunkable boundary
This algorithm will make sure the chunks from different video streams are
aligned if they have aligned GoPs. However, this algorithm will only work
for video streams. To be able to chunk non video streams at similar
positions as video streams, ChunkingHandler is designed to accept one video
input and multiple non video inputs, the non video inputs are chunked when
the video input is chunked. If the inputs are synchronized - which is true
if the inputs come from the same demuxer, the video and non video chunks
are aligned.
Change-Id: Id3bad51ab14f311efdb8713b6cd36d36cf9e4639
2017-02-07 18:58:47 +00:00
|
|
|
if (subsegment_info_[i] && subsegment_info_[i]->start_timestamp != -1) {
|
|
|
|
subsegment_info_[i]->duration =
|
|
|
|
last_sample_end_timestamps_[i] - subsegment_info_[i]->start_timestamp;
|
|
|
|
status.Update(DispatchSegmentInfo(i, std::move(subsegment_info_[i])));
|
|
|
|
}
|
|
|
|
subsegment_info_[i].reset(new SegmentInfo);
|
|
|
|
subsegment_info_[i]->is_subsegment = true;
|
|
|
|
}
|
|
|
|
return status;
|
|
|
|
}
|
|
|
|
|
|
|
|
} // namespace media
|
|
|
|
} // namespace shaka
|