This flag was designed for two purpose:
- Grouping fragments into subsegments, achieving three level hierarchy:
segment < subsegment < fragment.
- Indicate whether to generate 'sidx' box in media segments (when the
value is set to a negative number).
There are no practical use case for the first purpose. Removing it to
simplify the code and reduce the confusion.
Introduce another flag --generate_sidx_in_media_segments for the second
purpose.
Change-Id: I4be7cd42662fb324c1158b978e05768ee49dd048
It was implemented to workaround Chromium's DTS
https://crbug.com/398130, but the workaround does not really work in
all situations.
Remove it now as we already have another workaround available.
Change-Id: I291f559d78120fb743a6679b7d927e5bbc5b6b4e
Previously, the text padder media handler would assume that text always
started at time zero. This would work for VOD but would result with a
large pad at the start of LIVE content.
To avoid this, the text padder will use a bias to test whether or not it
thinks the content starts at zero. Right now the bias is set to be 10
minutes, but will later be configurable with a command line flag.
10 minutes was used as LIVE content will have much larger values and VOD
content should have much lower values.
Issue: #416
Change-Id: I07af15a577392fb030e36f052085cd4e667700e8
Added the key frame field to the IsMediaSample mather. All
tests that do not care about a sample being a keyframe (or not)
have been updated to use "_".
Change-Id: I44180687c58c260b6856e683d647f532227b14d5
Updated all the stream data matcher we use in our unit tests
to allow us to use matchers in them. We are now able to use "_"
to ignore specific parameters.
With this we were able to replace the different version of
matchers for each stream data type with a single instance for
each type.
Includes updates to printing strings to the listener. Strings
now go through a "pretty" function to help make it easier to
read them in the output.
Change-Id: I146351b54fccd63ab9ec936877e6c6b30f9aa9fc
Make the text stream info factory method in media_handler_test_base
require the caller to specify the time scale.
Issue: #399
Change-Id: Ibdfb183e0aa3f4ff50edf6b58c4e9b966006c6d2
To ensure that every variable in a box is explicitly set
every variable has been assigned a default in the header.
Change-Id: Iaa806c4058ac6621a64363a00040fbd9903c6710
Text Samples with no payload should be ignored, so this adds a
test to check that samples with no payload get treated the
same as a gap.
As long as this case is true, using gaps in our other tests should
functionally be the same as using samples with no payload.
Change-Id: Ic16b240c43eda2514b537a2d938d4135638adc4e
Before we used the sample payload for each text sample as we were
focusing on the times rather than the contents.
As we look to add tests that rely on specific sample payloads, we
need to change the tests to explicitly set the payload for each
sample.
Change-Id: I24174686f46535cf6c2d59a18308101a3bb51c87
EPT (earliest presentation time) may be adjusted not to be lower than
the decoding timestamp (dts), but the adjustment should only be done
on the first file when there is one file per Representation per Period.
The second file and onwards should not be adjusted otherwise a GAP
would be created.
Closes#384
Closes b/78517422
Change-Id: I56771ad8fbbe6a87b832ec58854cfbf37d5f1817
In our text to mp4 tests, we were only checking if the times on
the samples lined-up with what we were expecting.
We want to check that text sample contents (ids, settings, and
payloads) were correctly merged into a single media sample. To
verify this, we now check if the sample ids appear in the
media sample.
Change-Id: Ica1a85a14e7b116275e3571332b2e90d7bc44c45
In our text to mp4 tests, we used the same sample id for each sample,
this changes it so that each sample (within a single test) has a
unique id.
This is done in preparation to look for the ids in the created
media samples.
Change-Id: I3215a6f09279af8f40e1ce8a959e0a522a811173
The previous text to mp4 webvtt pipeline was incomplete. It
did not insert ad cues and it could only insert a segment
after a sample ended.
Now the pipeline supports ad cue insert and segment insertion
mid text sample. This required the pipeline to use the text
chunker (to split samples and insert segments) and required
a major overhaul of the text to mp4 converter.
Before the converter came before the chunker. This meant that
the converter only expected to see stream info and text samples.
Moving the converter after the cue aligner and chunker means
that the convert had to be aware of segments and cues.
The general approach is the same, however the converter will
convert the samples per-segment as the chunker will introduce
duplicate samples if a sample spans across segments.
Closes#362Closes#382
Change-Id: I0f54a40524c36a602ad3804a0da26e80851c92fd
In the text sample box (for mp4) there was a value called
"data_reference_index" that was never initialized. This meant
that it took on various values can caused different results
between runs.
Change-Id: I4b18ac97ec4700f6e651b14898ef250713a4253c
Renamed all the files called "webvtt_output_handler*" to
"webvtt_text_output_handler*" to better reflect the class name
in them.
Change-Id: I977bab362076974a124f263bcefff716ed8b6a0f
We don't want to allow any handler to be copyable or assigned-over
so this change enforces that for the webvtt output handler.
Change-Id: Ie0d59d6dbfb7a5e00bb4dd1422cd696d1a2d6072
Before, the webvtt output handler was written so that it could
share code between a segmented and non-segmented handler. As
we are not worried about that right now, this change simplifies
the handler to just be about segmented output.
Change-Id: I29dbc4e3a4ffbeb7ea10e23db489ee74b398a6c4
Instead of failing immediately, ignore unsupported audio codec when
parsing the source file, as there may be more than one stream in the
source file. This allows the supported streams to be packaged.
Closes#395.
Change-Id: I01005a93a19012c19065251647c9b06dd25c673a
The Cue class was from a previous WebVTT implementation and
is not used in the current implementation. It was missed when the
other classes were removed. This change removes it.
Change-Id: I661ab3fcd80b5e5ef98b5213746b341a4028d1a1
When originally implementing the webvtt parser, there was a
misunderstanding in what the BOM was suppose to be
(https://en.wikipedia.org/wiki/Byte_order_mark). This corrects the
misunderstanding.
Close#397
Change-Id: I250d392db228e5e9b86684614b57adc5d8a4e5fe
This is in preparation of supporting entitlement license API, where
common encryption server may return concatenated PSSHs directly.
Refactored ProtectionSystemSpecificInfo into a struct containing
concatenated PSSHs. This will make it easier to pass PSSHs around.
Also, most of the time, users of ProtectionSystemSpecificInfo do
not care what is in PSSH; so moved PSSH box parsing and building out
of ProtectionSystemSpecificInfo.
b/78171767
Change-Id: I1c4d5e7e23efd2f7d4b2b9704378323112e47f00
Packager uses ThreadedIO to write media segments and manifest /
playlists. There was a possibility that media segments write being
delayed and scheduled after updating manifest / playlists.
This CL fixes the race condition.
Also added a note on how segments can be synced to cloud storage to
avoid the race condition during file sync.
Also added a live WebM test.
Fixes#386.
Change-Id: Icf9c38cdec715fa3dc2836eab1511131e129fe41
To ensure that we can parse content with style and region blocks,
this change updates the parser to skip those blocks so that we
can still parse the cues from a file.
Full style and region support will be added later this year.
Issue #380
Change-Id: I11b8fd862a108c27a5c67b15d4703532b44a1214
Instead, the actual earliest presentation time is used except for
the first segment if there is an offset between presentation time
(pts) and decoding time (dts).
Chrome (as of v66) reports dts instead of pts in buffered ranges in
MSE API. To avoid breaking Chrome, the earliest_presentation_time
of the first segment is set to its dts as Chrome does not like negative
values for
adjusted dts = dts + Period@start (0 for the first period)
- presentationTimeOffset (earliest_presentation_time).
Fixes#303.
Change-Id: I5ca80e05d5570961400499436f2bcc01f06e69e0
The WebVtt Output Handler did not recognize cue events. This change
allows the handler to accept the events and tell muxer listener
about them.
Issue #362
Change-Id: I7c3318b72e539adc19af587c8e213fdb0af8290b
Created a media handler to come after parsers that will handle filling
in gaps between text samples. The padder takes a min duration, and if
the samples do not cover the min duration when flushed, one last empty
sample will be injected so that the samples will go up to the min duration.
Change-Id: I88605059664d09279676edac418ff3d4990d7556
Move the webvtt segmenter to the chunking directory so that it
can be converted to a general purpose text chunker.
Change-Id: I9ecd7ee39cb73070dab07b64f65ef24af1404813
Now that we have the end-to-end tests, we no longer need the webvtt pipeline
tests to verify that it is working.
Change-Id: I4ebec34e66eda67c40999d8802b447e2551e1fa6
This is called when reaching end of the file of the media. The duration
of the media is updated.
Fixes#340.
Change-Id: I446f2d341b02125d4a7d8c958bda269b5403cb9c
We have an assert that ensures that the end time is greater than
the start time for any cue. However we never checked that cues
had a non-zero duration when parsing them.
We will throw away cues with a duration of zero (and print a
warning message) as they are not spec compliant.
Closes: #335
Change-Id: I404e8f3a5a8d43eff75a2554db3e38e8d340f421
In the video captured by Android's default camera app, meta box is written as
as Box instead of FullBox specified in the spec.
Closes#319.
Change-Id: I526492fdd505d5929c5161cb1ed1503b724de7e9
Looks like Safari does not like v0 tenc box.
- Use v1 tenc box for cbcs and cens protection_scheme as required by
CENCv3 spec.
- Set crypt_byte_block and skip_byte_block to 0 for full sample
encryption cbcs and cens.
Fixes#326
Change-Id: I5581cd856fffc4ff104d950f3ca19b9337d57a78
When a gap is found in the text stream, the WebVtt to Mp4 converter will
now output the special empty vtt cue.
Change-Id: I8be88c6b7589aa120a2215e1e4b8e98031fe326d
Closes: #324
- Add empty lines between different types of renditions to improve
readability.
- Group variants with the same audio/text group together, as it is
where the Adaptation occurs.
- Write master playlist after writing media playlists. This makes
more sense and it is also necessary to have the bandwidth of
the last iframe playist segment correctly computed.
- For fMP4, I-Frame segment must include the 'moof' header.
- Fix a problem that hls_iframe_playlist_name is not passed to
MuxerListenerFactory.
Issue: #287
Change-Id: Icf37c5de1dc29f85ae3f419cbc3264d04ca491a4
Adding all the code needed to allow the webvtt to mp4 converter to
send empty vtt cues. This change does everything except writing the
mp4 box.
Bug: #324
Change-Id: I16188e6357632b2ed06f7e9bab7844f093266696
Instead of always tracking segments with an index, this change makes a struct
that will act as the segment and track which segments are in it. Each
segment will store all the samples in the order they were given, it will
avoid any sorting.
Closes: 72867775
Change-Id: Ic5829161510fe8f3320d960c3bc4a276c26ff3be
Create WebVtt test that shows that the current implementation of the
webvtt segmenter does not respect the input order of cues.
Bug: 72867775
Change-Id: I811b3a93c10650e0cf9290c9c0c1680f562deb30
According to the HLS+WebVTT spec, if there is no text in a segment
a webvtt file with no cues can be added to the manifest. By outputting
empty segments it allows the accumulated duration in the Master Playlist
to better represent the duration of the text stream.
Spec Reference: //tools.ietf.org/html/draft-pantos-http-live-streaming-23
Section 3.5. WebVTT
Bug: #205
Change-Id: I5de01200fd9fa99c57949c773e8ee926b0f6ba8a
Use a mock muxer listener so that we can verify that the webvtt
output handlers will write to a manifest correctly.
Change-Id: I0294a998bfaf06a6d8f7d4c287fb014839b38f73
Connected the WebVtt Pipeline together to test that the parser,
segmenter, and output all work together as expected.
Change-Id: I07138f7b5b1318f84c27c5b607d8df207c57ddb3
- Update curl to 7.57.0
- Roll clang, which is needed due to MacOS / XCode update
- Fixes and suppress compilation errors due to clang update
Fixes#285
Change-Id: Ibac3288c641861605c3c0500d34d27373e6eecfe
This change introduces handlers to output WebVtt text files. There is
only one output but there is a common base to support others.
WebVttOutputHandler which handles all communication with other handlers
and WebVttSegmentedOutputHandler is responsible for listening for events
and choosing when and where to write cues and headers.
Bug: 36138902
Change-Id: I2b13a94262554398e66fee8cf024aa21041ddbab
Instead of having multiple char readers, one for strings and
one for files, just have one for files and use memory files
when testing.
Change-Id: Id1a2230046ba540ddf69ca10edb3edc74d2419b6
To better align with the Chunking Handler, the Text Segmenter
now outputs the SegmentInfo at the end of the segment rather
than the start.
Change-Id: If69ab951947d00779b4b63a52c4b6662bbdc4c0d
Implemented a MediaHandler that takes text samples and creates
media samples. The data in each media sample is the MP4 box for
non-overlapping cues.
As per WebVtt in Mp4, all cues must be non-overlapping. This handler
takes care of grouping and dividing cues.
Bug: 36138902
Change-Id: I0c1d27964180c14a22cb200591f70e46e04a651f
This change creates the webvtt segmenter which will be used for
segmented text webvtt for HLS. It takes a stream of text samples
and injects segment info messages based on the segment duration.
It is possible for samples to extend between segments, if so the
sample will appear twice (once in each segment).
Change-Id: Iae0134ee61cf269948026086520b6d3f8ce3785b
Took the WebVTT Media Parser and created the WebVTT Parser
that will take in a file and output a stream of cues that
will later be passed to another Media Handler that takes in
cues and chunks them.
Bug: 36138902
Change-Id: Ic77813fe19678e85d500269e69f46917510ab7ec
Seeing some failures on some platforms when compiled with clang
disabled:
GYP_DEFINES="clang=0" gclient runhooks
Several changes to make it work:
1. Mark packager code with packager_code=1 in GYP definitions.
2. Disable a few checks in non-packager code, which we do not have
direct control: dangling-else, deprecated-declarations,
unused-function
3. Fix the relevant errors in packager code.
4. Revert HAVE_STROPTS_H in curl config which is not available in
all linux distributions.
Fixes#286Fixes#293
Change-Id: I729b41f99403c5ad9487c6cc4a7dc06f6323cef8
The error is logged only once to avoid log spamming.
Added definitions for the unsupported stream types for easy look up.
Change-Id: I097e2f05759bc84ef03f264cfabd2fb20da7c711
The spec allows having more than one 'mdat' boxes even if it is not
used.
This is happening on some mp4 files in the wild:
http://www.sample-videos.com/.
Fixes#298.
Change-Id: I729cc94bee095560b5c4b3ee8511323f25f7ad5a
Implemented E-AC3 MPEG-2 Stream Encryption Format for HTTP Live
Streaming specified in https://goo.gl/1sgcwY
Issue #279
Change-Id: I36c1a05e3d0529ff810eaf52bdca45414baa93eb
Audio samples per frame for AC3 was not specified correctly earlier.
Also the number of channels is not correct if LFE channel is on.
Issue #165
Change-Id: Ibf20aa4c7aec43c07ec7cd394d631c537cb387dd
Implemented MPEG-2 Stream Encryption Format for HTTP Live Streaming
specified in https://goo.gl/N7Tvqi.
This change also moved ProgramMapTableWriter creation from TsWriter
to TsSegment.
Issue #165
Change-Id: Ia89dd16a5e6405706dd3d85ec6b6de580f7b13a7
Use std::shared_ptr<const uint8_t> with a custom deleter to
represent MediaSample::data_ instead of std::vector<uint8_t>.
MediaSample::data_ can be shared by multiple MediaSamples and it is
immutable. A new data instance must be created if the clients want to
modify the underlying data. The new data instance can be transferred
to MediaSample using provided MediaSample::TransferData function.
This avoids unnecessary data copying.
Change-Id: Ib59785a9e19d0abb3283179b12eb6779ee922f79
In prep for changes to Trick Play, we want to make all messages
copy on write so that if the same message is sent to multiple
handlers, it is not possible for one handler to change the data
another handler is using.
Change-Id: I554166ca11c532412e4dfced5603972ca24dc2bb
Previously packaging completes successfully without any error
or warning.
With the fix, an error will be reported if write fails. It may
appear as "Cannot close file error" as we use threaded IO, which
could delay the error reporting until Close() call, so the user
of the File API needs to make sure Close() returns successfully.
Also fixed a deadlock in threaded_io_file if internal_file->Write
fails.
Fixes#160
Change-Id: I17f945150fb4021d2dcdbe784e557673f53ca583
Implemented the other end of the webvtt timestamp parsing as it will
be needed when we write out WebVtt files. Also changed the name of
the parse function to be more inline with the writing function.
Change-Id: I1f36ddbbf80028732b4cb1b15e871dec17767f63
Created a handler called OriginHandler that can be used by the Job class
(previously called RemuxJob). Origin handlers represent the start (or
origin) of a pipeline (chain of handlers).
Change-Id: Ibd748ae0a932b6e0ebb879ea292fcb83c548214b
Derive file size from media ranges instead.
Also fix subsegment range reporting for WebM (internal variable,
no output affected).
Change-Id: I5f8152dff4c2cd5fbae5550992b86a669e278f7b
- Changed OnMediaEnd() to take (sub)segment ranges.
- Define a structure to pass the ranges.
- This requires muxers to "cache" the segment ranges but in most cases,
they already do it, e.g. 'sidx' for mp4.
Change-Id: I16b974c1f0f54ca658cc6dbe605efff84a3f52eb
- Remove the TS key rotation not supported check as there is nothing
to be done for key rotation in TS for SAMPLE-AES.
- Fix IV updated problem in new segments even if crypto period does
not change.
- Avoid duplicate EXT-X-KEY tags if it does not change.
- Make EXT-X-DISCONTINUITY-SEQUENCE match with number of removed
EXT-X-DISCONTINUITY.
- Added end to end test for HLS with key rotation.
Change-Id: I73cb82e9f5575fcdf63ee643228efe78e6766302
Moved all the timestamp parsing code into its own class with its
own unit tests. This is to make testing easier.
Change-Id: I624472baba51dfa1254a9f4bae55ebb79e310855
This addresses playready broken with clear lead = 0.
clear lead != 0 is still broken with playready on Edge. We think it is
likely an Edge bug, which does not support multiple sample entry boxes,
thus does not support clear lead.
b/37913785
Change-Id: I7adb77a913dccf669153b03b31b4a1e1c98d1cb0
- Implemented according to v1.0 spec @ https://www.webmproject.org/vp9/mp4/
- v0 is no longer supported
Change-Id: I189c813d788400beda797eea7da943a83dfa7d79
- Added flag --mp4_include_pssh_in_stream, default to true. If it is
set to false, the encrypted mp4 stream will not include pssh.
- Align BytesOfProtectedData to multiple of 16 bytes for cenc.
- Set TrackHeader flag to kTrackEnabled | kTrackInMovie |
kTrackInPreview
- Move mvex to after trak, required by HLS
- Add cmfc/cmfs compatible brands except for avc3/hev1, where CMAF
requires single initialization switching set which is not supported.
- Set duration to 0 in tkhd, mdhd, mvhd.
Also updated major_brand and compatible brands:
- Set major_brand to isom (iso-bmff media file format) and made dash
a compatible brand
- Replaced compatible brand iso6 with iso8 since we use sthd for text
tracks
Fixes b/36278260
Change-Id: I3cc5dd5aa1621714d517fe02fe3841d19a1a07f6
Added a flag --strip_parameter_set_nalus. When enabled, parameter
set NAL units, SPS/PPS for H264 and SPS/PPS/VPS for H265, are stripped
from frames when converting NAL byte stream (AnnexB stream) to NAL
unit stream, which generates avc1/hvc1; otherwise they are not
stripped, and avc3/hev1 is generated.
Parameter set NAL units should not be stripped if they are varying
in the frames otherwise the frames may fail to be decoded.
The flag is enabled by default as we don't usually see varying
SPS/PPS/VPS and it is more space efficient with them stripped.
Set --strip_parameter_set_nalus=false to disable the flag if there
are varying SPS/PPS/VPS in the frames. This addresses #206.
Change-Id: I34bde6f33069f9f77d51a510b39f58a0f0e141aa
- Disable subsample encryption for VP8 in ISO-BMFF
- Apply block alignment to all subsamples for VP9 in ISO-BMFF,
instead of just superframes.
Change-Id: I8dd31cc16e87abc4d538330eaff9acb0509497df
Also moved MediaHandler output validation to Initialize instead.
This CL also addresses #122 with consistent chunking.
Change-Id: I60c0da6d1b33421d7828bcb827d18899e71884ce
- WebVttMediaParser uses WebVttSampleConverter to generate non
overlapping media samples.
- The media samples contains ISO BMFF boxes.
- Add kCodecWebVtt to signal that the media is webvtt and
the samples will be in ISO BMFF boxes.
Change-Id: I639902cdba7b04af75428bc20622e26b8203cfb2
- Also sets up the packaging and verify it works.
Some of the changes are temporary to get the integration going.
Change-Id: I0cf6c379d185e157808acabb9ef58ff93d4a39ae
The spec actually requires all NAL units to be escaped, so there
is no need to escape it again.
NAL byte stream is one demarcation method; NAL unit stream is another
demarcation method. Regardless of the demarcation method used, the
requirement for NAL unit is the same. See the last paragraph of
7.4.1 NAL unit semantics on the requirement on emulation prevention bytes.
Change-Id: Icc63fcc5cf965632e331f5af5f673164c7c1663a
CENCv3 recommends only encrypting video data in slice NALs. Slice
data partition NALs should not be encrypted.
In the code, differentiate is_vcl and is_video_slice. They are the
same for H265; for H264, vcl NALs include slice data partition NALs
but video_slice NALs do not.
Change-Id: I91f4bdd76d25f0eac50e39aed350ebce3f667121
- Logic for splitting up WebVTT cues into MP4 samples.
- For any cue intervals that overlap, a new sample must be created.
Change-Id: Icf7478f42d8c5790bf6f404d512c1251dd34c405
- Deprecated command line flags --profile and --single_segment.
'segment_template' in stream descriptors implies live profile
and non-single segment.
- Added flag --generate_static_mpd_for_live_profile to generate
static mpd for live profile; if not set, dynamic mpd will be
generated.
Close#142
Change-Id: I78879297ed118f0f246c4753a16ad125bd6b5e4f
1. Allow encryption pattern of 0:0, which is treated as 1:0;
2. Remove the special handling of pattern 1:0 in pattern cryptor
which may not always be correct;
3. Allow senc/saio/saiz boxes to be absent.
Change-Id: I372e61182ec577107e39cb601c3aed80616b036d
The original code creates temporary file in the current directory,
which may not always be writable.
Change-Id: Icc278dd5db6be6206ae29defbb7423bae39cc978
Cues used to be generated in the end of the file; if http range
request is not supported, clients have to download the whole
file to get to the cues.
This CL updated TwoPassSingleSegmentSegmenter to write cues
after writing webm header.
This CL also updates libwebm dependency to latest.
Closes#159
Change-Id: Ic73548e1b872e6b13a37627707e7d0ff3e556877
The original code accumulates sample durations in seconds (double)
and then generates new segment / cluster by comparing the accumulated
value with configured value.
There may be loss of precision when accumulating values in double.
For example, with a GOP of 96 frames and frame rate of 24 fps; the
accumulated frame duration for 96 frames is 3.999999999999, which is
very close to 4.0 but not 4.0. Problem would arise if segment
duration is set to 4. The created segments would have a duration of
8 seconds instead of 4 seconds.
The new code accumulates sample durations in uint64_t relative to
input time scale; the configured segment duration will be converted to
timescale for comparison. This avoids loss of precision.
Change-Id: I3ae24be82a7ce45f923a6f90fea495b8b6b2e7ef
- Also regenerated test content to insert an iframe every one
second: ffmpeg -i source.webm -auto-alt-ref 1 -g 30 out.webm
This makes sure the content is actually encrypted with one
second clear lead in end to end test.
Change-Id: I488143c148e2d6c45ba1586f6d0d835dc46db86e
- Add decrypt_config in media sample and stream.
- Update subsamples when converting NAL unit to bytestream.
Change-Id: I7b8975a453f81b22cf74bee3c9a58b7e458cbaae
1. Replace STLDeleter with unique_ptr in container, which is
supported in C++11
2. Replace string_as_array with &string[0] which is guaranteed
to work in C++11
Change-Id: I7f39c0e51fc8a3fcbb41313094a0ca6b33db7bf3
In H264, there may be multiple consecutive video slice NAL units
in the same frame. The original code assigns a new access unit
for every video slice NAL unit, which is incorrect.
Fixes#134.
Change-Id: I4d44271df48cb08867ddd02f7494fb3573af3356
Mimic iso-bmff behavior: either all the samples in a segment are encrypted
or all the samples are clear.
Change-Id: I03bdbbf5a4b690f4d87c4dceb0f8155e6fae941e
The right way to detect the presence of a keyframe in a BlockGroup
is to look for the absence of the ReferenceBlock element [1].
Closes b/30433206
[1] https://matroska.org/technical/specs/index.html#ReferenceBlock
Change-Id: I2b3c8704cbabcd5cbee38448e5528425865561a4
KongQun Yang
Aaron Vaage
Haoming Chen
Rintaro Kuroiwa
Haoming Chen
Zajcev Evgeny
Kyle Alexander
Thomas Inskip