Android MediaCodec stuff
Last update: 2016-06-08

This page is about the Android MediaCodec class, which can be used to encode and decode audio and video data. It includes a collection of sample code and answers to frequently-asked questions.

As of Marshmallow (API 23), the official documentation is quite detailed and very useful. (This is a huge step up from what was there when this page was created, and a significant advancement over the semi-reasonable docs in API 21.) The information on that page should be your primary source of information. The code here is expected to work with API 18+, for broad compatibility. If you're specifically targeting Lollipop or Marshmallow, you have options available that aren't shown here.


The MediaCodec class first became available in Android 4.1 (API 16). It was added to allow direct access to the media codecs on the device. As such, it provides a rather "raw" interface. While the MediaCodec class exists in both Java and C++ sources, only the former is public.

In Android 4.3 (API 18), MediaCodec was expanded to include a way to provide input through a Surface (via the createInputSurface method). This allows input to come from camera preview or OpenGL ES rendering. Android 4.3 was also the first release with MediaCodec tests in CTS, which helps ensure consistent behavior between devices.

Android 4.3 also introduced MediaMuxer, which allows the output of the AVC codec (a raw H.264 elementary stream) to be converted to .MP4 format, with or without an associated audio stream.

Android 5.0 (API 21) introduced "asynchronous mode", which allows an app to provide a callback method that executes as buffers become available. The bits of code linked from this page don't take advantage of this, because they target API 18+.

Basic Usage

All uses of the "synchronous" MediaCodec API follow a basic pattern:

A single instance of MediaCodec handles one specific type of data (e.g. MP3 audio or H.264 video), and may encode or decode. It operates on "raw" data, so any file headers (e.g. ID3 tags) must be stripped off. It does not talk to any higher-level system components; it will not play your audio out through the speaker or receive a stream of video over a network. It just takes buffers of data in and spits buffers of data out. (You can use MediaExtractor to strip the wrappers off in most situations.)

Some codecs are very particular about their buffers. They may need to have a particular memory alignment, or have a certain minimum or maximum size, or it may be important to have a certain number of them available. To accommodate the wide range of possibilities, buffer allocation is performed by the codecs themselves, rather than the application. You do not hand a buffer with data to MediaCodec. You ask it for a buffer, and if one is available, you copy the data in.

This may seem contrary to "zero-copy" principles, but in most cases there will be fewer copies because the codec won't have to copy or adjust the data to meet its requirements. In some cases you may be able to use the buffer directly, e.g. read data from disk or network directly into the buffer, so a copy won't be necessary.

Input to MediaCodec must be done in "access units". When encoding H.264 video that means one frame, when decoding that means a single NAL unit. However, it behaves like a stream in the sense that you can't submit a single chunk and expect a chunk to appear shortly from the output. In practice, the codec may want to have several buffers queued up before producing any output.

It is strongly recommended that you start with sample code, rather than trying to figure it out from the documentation.

Samples (requires 4.3, API 18)

Generates a movie using OpenGL ES. Uses MediaCodec to encode the movie in an H.264 elementary stream, and MediaMuxer to convert the stream to a .MP4 file.

This was written as if it were a CTS test, but is not part of CTS. It should be straightforward to adapt the code to other environments. (requires 4.3, API 18)

Records video from the camera preview and encodes it as an MP4 file. Uses MediaCodec to encode the movie in an H.264 elementary stream, and MediaMuxer to convert the stream to a .MP4 file. As an added bonus, demonstrates the use of a GLES fragment shader to modify the video as it's being recorded.

This was written as if it were a CTS test, but is not part of CTS. It should be straightforward to adapt the code to other environments.

Android Breakout game recorder patch (requires 4.3, API 18)

This is a patch for Android Breakout v1.0.2 that adds game recording. While the game is playing at 60fps at full screen resolution, a 30fps 720p recording is made with the AVC codec. The file is saved in the app-private data area, e.g. /data/data/com.faddensoft.breakout/files/video.mp4.

This is essentially the same as, but it's part of a full app rather than an isolated CTS test. One key difference is in the EGL setup, which is done in a way that allows textures to be shared between the display and video contexts. WARNING: this code has a race condition. See this bug report for details and suggested fix.

Another approach is to render each game frame to an FBO texture, and then render a full-screen quad with that twice (once for the display, once for the video). This could be faster for games that are expensive to render. Examples of both approaches can be found in Grafika's RecordFBOActivity class. (requires 4.3, API 18)

CTS test. There are three tests that do essentially the same thing, but in different ways. Each test will: The generation, encoding, decoding, and checking are near-simultaneous: frames are generated and fed to the encoder, and data from the encoder is fed to the decoder as soon as it becomes available.

The three tests are:

  1. Buffer-to-buffer. Buffers are software-generated YUV frames in ByteBuffer objects, and decoded to the same. This is the slowest (and least portable) approach, but it allows the application to examine and modify the YUV data.
  2. Buffer-to-surface. Encoding is again done from software-generated YUV data in ByteBuffers, but this time decoding is done to a Surface. Output is checked with OpenGL ES, using glReadPixels().
  3. Surface-to-surface. Frames are generated with OpenGL ES onto an input Surface, and decoded onto a Surface. This is the fastest approach, but may involve conversions between YUV and RGB.
Each test is run at three different resolutions: 720p (1280x720), QCIF (176x144), and QVGA (320x240).

The buffer-to-buffer and buffer-to-surface tests can be built with Android 4.1 (API 16). However, because the CTS tests did not exist until Android 4.3, a number of devices shipped with broken implementations.

NOTE: the setByteBuffer() usage may not be strictly correct, as it doesn't set "csd-1".

(For an example that uses the Android 5.x asynchronous API, see mstorsjo's android-decodeencodetest project.) (requires 4.3, API 18)

CTS test. The test does the following: The middle decode-edit-encode pass performs decoding and encoding near-simultaneously, streaming frames directly from the decoder to the encoder. The initial generation and final verification are done as separate passes on video data held in memory.

Each test is run at three different resolutions: 720p (1280x720), QCIF (176x144), and QVGA (320x240).

No software-interpreted YUV buffers are used. Everything goes through Surface. There will be conversions between RGB and YUV at certain points; how many and where they happen depends on how the drivers are implemented.

NOTE: for this and the other CTS tests, you can see related classes by editing the class URL. For example, to see InputSurface and OutputSurface, remove "" from the URL, yielding this directory link. (requires 4.1, API 16) (requires 4.2, API 17)

Extracts the first 10 frames of video from a .mp4 file and saves them to individual PNG files in /sdcard/. Uses MediaExtractor to extract the CSD data and feed individual access units into a MediaCodec decoder. The frames are decoded to a Surface created from SurfaceTexture, rendered (off-screen) into a pbuffer, extracted with glReadPixels(), and saved to a PNG file with Bitmap#compress().

Decoding the frame and copying it into a ByteBuffer with glReadPixels() takes about 8ms on the Nexus 5, easily fast enough to keep pace with 30fps input, but the additional steps required to save it to disk as a PNG are expensive (about half a second). The cost of saving a frame breaks down roughly like this (which you can get by modifying the test to extract full-size frames from 720p video on a Nexus 5, observing the total time required to save 10 frames, and doing successive runs with later stages removed; or by instrumenting with android.os.Trace and using systrace):

In theory, a Surface from the API 19 ImageReader class could be passed to the MediaCodec decoder, allowing direct access to the YUV data. As of Android 4.4, the MediaCodec decoder formats are not supported by ImageReader.

One possible way to speed up transfer of RGB data would be to copy the data asynchronously through a PBO, but in the current implementation the transfer time is dwarfed by the subsequent PNG activity, so there's little value in doing so.

The two versions of the source code function identically. One was written against EGL 1.0, the other EGL 1.4. EGL 1.4 is a little easier to work with and has some features that other examples use, but if you want your app to work on Android 4.1 you can't use it.

This was written as if it were a CTS test, but is not part of CTS. It should be straightforward to adapt the code to other environments. (NOTE: if you're having trouble with timeouts in awaitNewImage(), see this article.) (requires 4.1, API 16)

CTS test. Tests decoding pre-recorded audio streams. (requires 4.1, API 16)

CTS test. Tests encoding of audio streams. (requires 4.3, API 18)

CTS test. Uses MediaMuxer to clone the audio track, video track, and audio+video tracks from input to output.

screenrecord (uses non-public native APIs)

You can access the source code for screenrecord, a developer shell command introduced in Android 4.4 (API 19). It demonstrates the use of the native equivalents of MediaCodec and MediaMuxer in a pure-native command. v1.1 uses GLES and the native equivalent of SurfaceTexture.

This is FOR REFERENCE ONLY. Non-public APIs are very likely to break between releases and are not guaranteed to have consistent behavior across different devices.

Grafika (requires 4.3, API 18)

Test application exercising various graphics & media features. Examples include recording and displaying camera preview, recording OpenGL ES rendering, decoding multiple videos simultaneously, and the use of SurfaceView, GLSurfaceView, and TextureView. Highly unstable. Tons of fun. Unlike most of the samples here, Grafika is a complete application, so it's easier to try things yourself. The EGL/GLES code is also more refined, and better suited for inclusion in other projects.


Q1. How do I play the video streams created by MediaCodec with the "video/avc" codec?

A1. The stream created is a raw H.264 elementary stream. The Totem Movie Player for Linux may work, but many other players won't touch them. You need to use the MediaMuxer class to create an MP4 file instead. See the EncodeAndMuxTest sample.


Q2. Why does my call to MediaCodec.configure() fail with an IllegalStateException when I try to create an encoder?

A2. This is usually because you haven't specified all of the mandatory keys required by the encoder. See this stackoverflow item for an example.


Q3. My video decoder is configured but won't accept data. What's wrong?

A3. A common mistake is neglecting to set the Codec-Specific Data, mentioned briefly in the documentation, through the keys "csd-0" and "csd-1". This is a bunch of raw data with things like Sequence Parameter Set and Picture Parameter Set; all you usually need to know is that the MediaCodec encoder generates them and the MediaCodec decoder wants them.

If you are feeding the output of the encoder to the decoder, you will note that the first packet you get from the encoder has the BUFFER_FLAG_CODEC_CONFIG flag set. You need to make sure you propagate this flag to the decoder, so that the first buffer the decoder receives does the setup. Alternatively, you can set the CSD data in the MediaFormat, and pass this into the decoder via configure(). You can see examples of both approaches in the EncodeDecodeTest sample.

If you're not sure how to set this up, you should probably be using MediaExtractor, which will handle it all for you.


Q4. Can I stream data into the decoder?

A4. Yes and no. The decoder takes a stream of "access units", which may not be a stream of bytes. For the video decoder, this means you need to preserve the "packet boundaries" established by the encoder (e.g. NAL units for H.264 video). For example, see how the VideoChunks class in the DecodeEditEncodeTest sample operates. You can't just read arbitrary chunks of the file and pass them in.


Q5. I'm encoding the output of the camera through a YUV preview buffer. Why do the colors look wrong?

A5. The color formats for the camera output and the MediaCodec encoder input are different. Camera supports YV12 (planar YUV 4:2:0) and NV21 (semi-planar YUV 4:2:0). The MediaCodec encoders support one or more of:

I420 has the same general data layout as YV12, but the Cr and Cb planes are reversed. Same with NV12 vs. NV21. So if you try to hand YV12 buffers from the camera to an encoder expecting something else, you'll see some odd color effects, like in these images.

As of Android 4.4 (API 19), there is still no common input format. Nvidia Tegra 3 devices like the Nexus 7 (2012), and Samsung Exynos devices like the Nexus 10, want COLOR_FormatYUV420Planar. Qualcomm Adreno devices like the Nexus 4, Nexus 5, and Nexus 7 (2013) want COLOR_FormatYUV420SemiPlanar. TI OMAP devices like the Galaxy Nexus want COLOR_TI_FormatYUV420PackedSemiPlanar. (This is based on the format that is returned first when the AVC codec is queried.)

A more portable, and more efficient, approach is to use the API 18 Surface input API, demonstrated in the CameraToMpegTest sample. The down side of this is that you have to operate in RGB rather than YUV, which is a problem for image processing software. If you can implement the image manipulation in a fragment shader, perhaps by converting between RGB and YUV before and after your computations, you can take advantage of code execution on the GPU.

Note that the MediaCodec decoders may produce data in ByteBuffers using one of the above formats or in a proprietary format. For example, devices based on Qualcomm SoCs commonly use OMX_QCOM_COLOR_FormatYUV420PackedSemiPlanar32m (#2141391876 / 0x7FA30C04).

Surface input uses COLOR_FormatSurface, also known as OMX_COLOR_FormatAndroidOpaque (#2130708361 / 0x7F000789). For the full list, see OMX_COLOR_FORMATTYPE in OMX_IVCommon.h.

As of API 21 you can work with an Image object instead. See the MediaCodec getInputImage() and getOutputImage() calls.


Q6. What's this EGL_RECORDABLE_ANDROID flag?

A6. That tells EGL that the surface it creates must be compatible with the video codecs. Without this flag, EGL might use a buffer format that MediaCodec can't understand.


Q7. Can I use the ImageReader class with MediaCodec?

A7. Maybe. The ImageReader class, added in Android 4.4 (API 19), provides a handy way to access data in a YUV surface. Unfortunately, as of API 19 it only works with buffers from Camera (though that may have been corrected in API 21, when MediaCodec added getOutputImage()). Also, there was no corresponding ImageWriter class for creating content until API 23.


Q8. Do I have to set a presentation time stamp when encoding video?

A8. Yes. It appears that some devices will drop frames or encode them at low quality if the presentation time stamp isn't set to a reasonable value (see this stackoverflow item).

Remember that the time required by MediaCodec is in microseconds. Most timestamps passed around in Java code are in milliseconds or nanoseconds.


Q9. Most of the examples require API 18. I'm coding for API 16. Is there something I should know?

A9. Yes. Some key features aren't available until API 18, and some basic features are more difficult to use in API 16.

If you're decoding video, things don't change much. As you can see from the two implementations of ExtractMpegFramesTest, the newer version of EGL isn't available, but for many applications that won't matter.

If you're encoding video, things are much worse. Three key points:

  1. The MediaCodec encoders don't accept input from a Surface, so you have to provide the data as raw YUV frames.
  2. The layout of the YUV frames varies from device to device, and in some cases you have to check for specific vendors by name to handle certain qirks.
  3. Some devices may not advertise support for any usable YUV formats (i.e. they're internal-use only).
  4. The MediaMuxer class doesn't exist, so there's no way to convert the H.264 stream to something that MediaPlayer (or many desktop players) will accept. You have to use a 3rd-party library (perhaps mp4parser).
  5. When the MediaMuxer class was introduced in API 18, the behavior of MediaCodec encoders was changed to emit INFO_OUTPUT_FORMAT_CHANGED at the start, so that you have a convenient MediaFormat to feed to the muxer. On older versions of Android, this does not happen.

This stackoverflow item has additional links and commentary.

The CTS tests for MediaCodec were introduced with API 18 (Android 4.3), which in practice means that's the first release where the basic features are likely to work consistently across devices. In particular, pre-4.3 devices have been known to drop the last frame or scramble PTS values when decoding.


Q10. Can I use MediaCodec in the AOSP emulator?

A10. Maybe. The emulator provides a software AVC codec that lacks certain features, notably input from a Surface (although it appears that this may now be fixed in Android 5.0 "Lollipop"). Developing on a physical device will likely be less frustrating.


Q11. Why is the output messed up (all zeroes, too short, etc)?

A11. The most common mistake is failing to adjust the ByteBuffer position and limit values. As of API 19, MediaCodec does not do this for you.

You need to do something like:

  int bufIndex = codec.dequeueOutputBuffer(info, TIMEOUT);
  ByteBuffer outputData = outputBuffers[bufIndex];
  if (info.size != 0) {
      outputData.limit(info.offset + info.size);

On the input side, you want to call clear() on the buffer before copying data into it.


Q12. Why am I seeing storeMetaDataInBuffers failures in the log?

A12. They look like this (example from a Nexus 5):

E OMXNodeInstance: OMX_SetParameter() failed for StoreMetaDataInBuffers: 0x8000101a
E ACodec  : [] storeMetaDataInBuffers (output) failed w/ err -2147483648

You can ignore them, they're harmless.


Further Assistance

Please post all questions to stackoverflow with the android tag (and, for MediaCodec issues, the mediacodec tag as well). Comments or feature requests for the framework or CTS tests should be made on the AOSP bug tracker.