Intel Quick Sync Video Consumption by Applications

I wrote a few posts on hardware H.264 encoding (e.g. this and the latest one Applying Hardsubs to H.264 Video with GPU). A blog reader asked a question regarding availability of the mentioned Intel Quick Sync Video support with low end Intel x5-Z8300 CPU.

[…] Intel has advertised that the Cherry Trail CPUs support H264 encoding and / or QSV, but nowhere have I seen a demo of this being used […].
What did you use to encode the video? Is the QSV codec available in the x5-z8300 for possible 720p realtime encoding? I’d like to see this checked in regards to using software like FFmpeg with qsv_h264 -codec and OBS. […]

A picture below explains how applications are consuming Intel’s hardware video compression offering in Windows.

Intel QSV includes hardware implementation of the encoder and corresponding drivers which provide a frontend API to software. This includes a component which integrates the codec with Microsoft’s Media Foundation API. Applications are to choose between interfacing the codec using Windows API – this is the way stock Microsoft applications work, and this is the way I used for video encoding development mentioned on the blog. Other applications prefer to interface through Intel Media SDK, which is an alternate route ending up at the same hardware-backed services.

Intel x5-Z8300 system in question has H.264 video encoding support integrated into Windows API and the services can be consumed without additional Intel runtime and/or development kit. The codec, according to the benchmarks made earlier, is fast enough to handle real-time 720p video encoding nevertheless the device is a budget thing.

Bug in Media Foundation MPEG-4 File Source related to timestamping video frames of a fragmented MP4 file

Some recent update in Media Foundation platform introduced a new bug related to fragmented MP4 files and H.264 video. The bug shows up consistently with file versions:

  • mfplat.dll – 10.0.14393.351 (rs1_release_inmarket.161014-1755)    15-Oct-16 05:48
  • mfmp4srcsnk.dll – 10.0.14393.351 (rs1_release_inmarket.161014-1755)    15-Oct-16 05:45

The nature of the problem is that MPEG-4 File Source is incorrectly time stamping the data: frame time stamps are incorrect, they seems to be getting wrong durations and increments, then quickly jumps into future… and on playback this leads to unobvious playback freezes. As Media Foundation is used by Windows Media Player, Windows 10 Movies & TV Player, the bug is present there as well.

The original report is on MSDN Forums.

Presumably it is possible to roll certain Windows Update package back, or alternatively one has to wait for Microsoft to fix the problem and deliver a new update deploying the fix.

Applying Hardsubs to H.264 Video with GPU

Video adapters currently offer a range of services which enables transcoding of H.264 content with certain modifications (including but not limited to flexible overlays, scaling, mirroring, effects and filters) end-to end on GPU keeping data as DirectX Graphics Infrastructure resource at all processing stages.

Such specialized processing capabilities are pretty powerful compared to traditional CPU processing, especially taking into consideration performance of low end low power-consumption systems still equipped with contemporary GPU.

For a test, I transcoded video H.264/MP4 files of different resolutions applying a text overlay having a time stamp of video frame being processed. The overlay is complex enough to be  varying frame to frame, be a standard font with respective rasterization (using DirectWrite). The test re-encoded H.264 content maintaining bitrate without giving too much care for other encoding details (defaults used).

Hardsub Performance Test

The roughly made test was successful with two video GPUs:

  1. Intel HD Graphics 4600 (7th gen; Desktop system; Core i7-4790 CPU)
  2. Intel HD Graphics (8th gen; Ultramobile system; Atom x5-Z8300 CPU) – the system is actually a $200 worth budget Chinese tablet Cube iWork 10 Ultimate

The test failed on other GPUs:

  • Intel HD Graphics 4000 (7th gen; Mobile system; Core i7-3517U CPU)
  • NVIDIA GeForce GTX 750

The problem – as it looks without getting into details – seems to be the inability of Media Foundation APIs to fit Direct3D-enabled pipelines out of the box, such as because of lack of certain conversion. It looks like transcoding can be achieved, with just putting some more effort into it.

As of now, Intel offers their 9th generation GPUs and the ones being tested are hardware of a few yeas in age…

Compared to real time performance of 100% (meaning that it takes one second to process one second of video of given metrics), both systems managed to do the transcoding relatively efficiently. With a roughly built test having a bottleneck at applying overlay, taking place serially in single thread, both systems showed performance sufficient to convert 1920×1080@60 video faster than in real time and without maxing CPU out.

Intel’s seventh generation desktop GPU managed to do the job way much faster.

It is interesting that even cheap tablet can process a Full HD video stream loading CPU less than 40%. Basically, the performance is sufficient for doing real time video processing (including using external web camera like Logitech C930E) with certain processing in 1080p resolution using budget grade hardware.

Re-encoding Performance

When there is no necessity to keep the real time processing pace, the cheap tablet showed the ability to do GPU processing on 2K video, which is also good news for those who wants to apply budget hardware to high resolution material.

Apparently, the key factor is ability of the process to keep data in video hardware. As Intel GPU H.264 abilities scale well when used for concurrent multi-stream processing, the performance numbers promise great performance recording video in several formats at a time: raw video, video with overlay, scaled down etc.

The table below gives more numbers for the tests concluded:

Re-encoding Performance Numbers

As mentioned above, the overlaying part itself is a single threaded bottleneck and presumably it is a reserve to be used to cut elapsed time down even further.

Another interesting observation is that while ultramobile system still uses much of CPU time (which is okay – it’s not a powerful system by design), the desktop GPU has minimal impact on CPU while doing pretty complicated task.

Screen recording using Desktop Duplication API and hardware H.264 encoder

The application takes advantage of three powerful Windows APIs at a time:

MediaFoundationDesktopRecorder initializes a desktop duplication session and sends obtained desktop images to H.264 video encoder producing a standard MP4 recording. Optionally, it can add an audio track capturing data from one of the standard inputs.

The best performance is achieved when used with hardware H.264 encoder: not only the performance of hardware encoder is better, but additionally desktop images are transferred to the encoder efficiently, without being copied through system memory. With respective hardware, recording is pretty efficient.

There are certain limitations: duplication API is Windows 8+, encoder availability depends on hardware and OS versions. The application let API pick encoder automatically and in worth case scenario falls back to software encoder, which is typically a performance hit.

MediaFoundationDesktopRecorder UI

When started, the application prints initial information, esp. regarding availability of devices, and appends as actions and events take place.

The application uses configuration file with the same name and location as the application, and .INI extension. Changes to the configuration file take effect when the application is restarted.

The application registers Win+F5, Win+F8 hotkeys globally to start/stop recording when the application is in background (that is, when user interacts with another application).

The application generates .MP4 files in the directory of its own location. There will be a video track, and optionally one additional audio track – depending on settings. Video is taken from one of the monitors, and audio – from one of the available standard audio input devices.

The application also generates log files at one the locations:

  • C:\ProgramData\MediaFoundationDesktopRecorder.log
  • C:\Users\$(UserName)\AppData\Local\MediaFoundationDesktopRecorder.log (in case the first path above is inaccessible, esp. due to insufficient permissions)

Configuration

The configuration .INI file might contain a few settings that set up and alter the behavoir of the application:

[Input]
;Video Adapter Description=NVIDIA GeForce GTX 750
Video Output Device Name=\\.\DISPLAY2
;Audio Friendly Name=Stereo Mix (Realtek High Definition Audio)

When started, the application enumerates (“found video…”, “found audio…”) available video and audio inputs. These discoveries are compared against configuration file settings in order to identify monitor for recording, and possibly audio input device.

Default behavior is to take first available monitor, which happens when settings do not instruct otherwise. By default, no audio is recorded. Audio is recorded and added to resulting file if input device is provided explicitly.

The application also prints which devices are taken for further recording (“using adapter…”).

[Format]
;Video Frame Rate=30000
;Video Frame Rate Denominator=1001
Video Bitrate=4096000
Video Texture Pool Capacity=24
Video Throttle=70
Audio Bitrate=192000

Default behavior is to identify monitor’s refresh rate and produce output file with video at the same frame rate. Video Frame Rate and Video Frame Rate Denominator settings offer an override to target file frame rate. With the former value only, it is the frame rate. With both values they define a ratio, e.g. values of 30000 and 1001 result in 29.97 fps file.

Frame rate reduction is a good way to reduce encoding complexity and overall graphics subsystem load.

Bitrate values define respective bitrates for the encoded content.

Details

As recording goes, the application grabs new desktop snapshots and sends them to encoder. There are no specific expectations about frame rate stability and reduction in case of overload of graphics subsystem. When the complexity is excessive, it is expected that some frames might be lost without breaking the entire playability of the output file.

The application provides additional information when it creates a file, for example:

Using Direct3D 11 at feature level D3D_FEATURE_LEVEL_11_0
Using Desktop Duplication mode: Resolution 1680 x 1050, Refresh Rate 59954/1000, Format DXGI_FORMAT_B8G8R8A8_UNORM
Using path “D:\Projects\...\Output\20160707-070707.mp4”
Using video transform Direct3D 11 Aware, Category MFT_CATEGORY_VIDEO_PROCESSOR, Input MFVideoFormat_ARGB32, Output MFVideoFormat_NV12
Using video transform NVIDIA H.264 Encoder MFT, Direct3D 11 Aware, Category MFT_CATEGORY_VIDEO_ENCODER, Input MFVideoFormat_NV12, Output MFVideoFormat_H264
Started writing…
PPP frames written (QQQ frame timeouts, RRR early frame skips, SSS late frame skips)
Stopped writing
Output file size is TTT bytes

When started the application might experience a condition when certain hardware resource is no longer available, e.g. the desktop itself is locked by user. The application will close the file, and attempt to automatically restart recording into new file. The attempts keep going until user explicitly stops recording.

The application does NOT do the following (among things it could):

  • the application is limited to record from one monitor only; to record from two at a time it is possible to start several instances however the produced result will not be synchronized
  • the application does not provide options to record single window image, to cut a section of monitor image or to scale image down
  • the application does not offer choices for video encoders (e.g. there are two or more hardware H.264 encoders), it will always use encoder picked by the system
  • the application only offers bitrate setting for video encoding
  • the application does not provide flexibility in audio encoding settings, it also expects that audio device is available throughout the entire recording session (esp. is not unplugged as recording goes)

References (Informational)

Download links

Untweetable Video

Two H.264 MP4 files, close one to another. The files are playable, a sort of: Windows desktop players (except Media Foundation based), QuickTime, Android and iOS devices play them. The files are not flawless but make sense, hence the glitches.

There is a problem with the second file, which is rejected by Twitter “Your media file could not be processed.”. The file is treated well by browser, and uploaded to remote server. Twitter is unable to convert it on its backend.

 

Reference Signal Source: RGB32/ARGB32 Subtypes, Media Foundation Media Source for Video

An update for Reference signal source for DirectShow DLLs:

  • the source is doing more accurately RGB subtypes and allows specification whether you want MEDIASUBSTYPE_RGB32 or MEDIASUBSTYPE_ARGB32
  • additionally the DLL implements Microsoft Media Foundation Media Source for the video stream

A more detailed description follows.

RGB32 and ARGB32 are very close and share the same byte structure, and due to minimal support of alpha channel with video, these are having the difference mostly in counterpart support in other applications, like for example and specifically hardware-assisted H.264 encoders whcih are taking alpha-enabled variant.

IVideoSourceFilter::SetMediaType method takes vCompression argument which defines the subtype. RegisterSources sample code shows how the method is used when exposing reference signal as video capture device.

Similar IVideoMediaSource::SetMediaType methods is applicable to Media Foundation counterpart (see below).

Both implementation only offer the given subtype as default, but in the same time both accept the other variant as well if an application or peer connection is trying to re-agree the media type. Same applies to changing resolution etc. The sources are flexible to take different video format if anyone is requesting it.

The other big new thing is Media Foundation API Media Source which generates reference signal as well. There is no option to set it up as a virtual camera because the API does not offer extensibility of the kind, however the source can be used to generate test content via Media Foundation and the code remains pretty simple. I am publishing MfGenerate code snippet which demonstrates the necessary steps to create an MP4 file with video, with desired properties.

A frame from generated 4096x2304 content in Windows 10 player

As Media Foundation offers H.265 (HEVC) and fragmented MP4 options, they can also be easily used with the source to generate test footage.

The code does the following steps:

  1. Creates a media source (commented out lines show alternate steps to create a media source from a file)
  2. Creates a source reader from media source
  3. Builds an H.264 media type from raw video media type
  4. Creates and configures a sink writer, which is instructed to do its magic setting up H.264 encoder (a side note – the code produces 4096×2304 video, however it is only possible once hardware encoder is enabled; software encoder was rejecting the media type)
  5. Implements a loop of reading frames until they run out feeding them into encoder/writer

High level APIs are simple (similar to DirectShow), which is untrue for the internals (similar to DirectShow; even more so).

Media Foundation source is video only for now.

MF media source is supposed to be seekable (not really tested; not really testable with topoedit), and allows zero duration to produce infinite feed. Duration is not necessarily taken from property, it can also be specified with overwritten presentation descriptor attribute. The video format can also be set up through stream descriptor media type handler.

Download links

Update – Connecting MF Media Source to MFCaptureD3D Sample application

To quickly connect MF media source to Windows SDK MFCaptureD3D Sample application, add #import and a few code lines replacing the source around CPreview::SetDevice as shown on the image below:

MFCaptureD3D update for custom media source

Calling convention violator broke streaming loop pretty far away

A really nasty problem coming from MainConcept AVC/H.264 SDK Encoder was destroying media streaming pipeline. SDK is somewhat old (9.7.9.5738) and the problem might be already fixed, or might be not. The problem is a good example of how a small bug could become a big pain.

The problem was coming up in 64-bit Release builds only. Win32 build? OK. Debug build where you can step things through? No problem.

The bug materialized in GDCL MP4 Demultiplexer filter streaming (Demultiplexer filter in the pipeline below) generating media samples with incorrect time stamps.

Pipeline

Initial start and stop time are okay, and further go as _I64_MIN (incorrect).

Clipbrd3

The problem appears to be SSE optimization and x64 calling convention related. This explains why it’s only 64-bit Release build suffering from the issue. MS compiler decided to use XMM7 register for dRate variable in this code fragment:

REFERENCE_TIME tStart, tStop;
double dRate;
m_pParser->GetSeekingParams(&tStart, &tStop, &dRate);

[...]

for(; ; )
{
    [...]

    tSampleStart = REFERENCE_TIME(tSampleStart / dRate);
    tSampleEnd = REFERENCE_TIME(tSampleEnd / dRate);

dRate is the only floating point thing here and it’s clear why the compiler optimized the variable into register: no other floating point activity around.

However sample delivery goes pretty deep into other functions and modules reaching MainConcept H.264 encoder. One of its functions is violating x64 calling convention and does not preserve XMM6+ register values. OOPS! Everything is about working right, but after media sample delivery dRate value is destroyed and further media samples receive incorrect time stamps.

It is not really a problem of MP4 demultiplexer, of course, however media sample delivery might involve a long delivery chain where any violator would break streaming loop. In the same time, it is not really a big expense to de-optimize the floating point math in the demultiplexer for those a few time stamp adjustment operations. A volatile specifier breaks compiler optimization and makes the loop resistant to SSE2 register violators:

// HOTFIX: Volatile specifier is not really necessary here but it fixes a nasty problem with MainConcept AVC SDK violating x64 calling convention;
//         MS compiler might choose to keep dRate in XMM6 register and the value would be destroyed by the violating call leading to incorrect 
//         further streaming (wrong time stamps)
volatile DOUBLE dRate;
m_pParser->GetSeekingParams(&tStart, &tStop, (DOUBLE*) &dRate);

This makes H.264 this build of encoding SDK unstable and the problem is hopefully already fixed. The SDK indeed gave other troubles on specific architectures leading to undefined behavior.