当前位置: 动力学知识库 > 问答 > 编程问答 >

java - Delay frame while encoding video file using google/grafika

问题描述:

I'm using google/grafika's examples to decode, transform and encode back to file a video clip. The transformation is downscaling and translating, it is done via shader stored in Texture2dProgram. My main activity is based on CameraCaptureActivity. The catch is I'm placing two videos on single texture at the same time, side by side. I would like to delay one of them for a given amount of frames. Also note that I don't need display preview while encoding.

My best idea so far was to change timestamps while advancing through frames. In TextureMovieEncoder, I'm sending information about frames, including timestamp in which they has to be placed in result video. It takes place in frameAvailiable(), where I'm sending information about two frames at once (left and right). The idea was to increase timestamp of one of them. The problem is that result video is distorted, so I don't know if my approach is feasible. TextureMovieEncoder is posted below.

package com.android.grafika;

import android.graphics.SurfaceTexture;

import android.opengl.EGLContext;

import android.opengl.GLES20;

import android.os.Handler;

import android.os.Looper;

import android.os.Message;

import android.util.Log;

import com.android.grafika.gles.EglCore;

import com.android.grafika.gles.FullFrameRect;

import com.android.grafika.gles.Texture2dProgram;

import com.android.grafika.gles.WindowSurface;

import java.io.File;

import java.io.IOException;

import java.lang.ref.WeakReference;

/**

* Encode a movie from frames rendered from an external texture image.

* <p>

* The object wraps an encoder running on a dedicated thread. The various control messages

* may be sent from arbitrary threads (typically the app UI thread). The encoder thread

* manages both sides of the encoder (feeding and draining); the only external input is

* the GL texture.

* <p>

* The design is complicated slightly by the need to create an EGL context that shares state

* with a view that gets restarted if (say) the device orientation changes. When the view

* in question is a GLSurfaceView, we don't have full control over the EGL context creation

* on that side, so we have to bend a bit backwards here.

* <p>

* To use:

* <ul>

* <li>create TextureMovieEncoder object

* <li>create an EncoderConfig

* <li>call TextureMovieEncoder#startRecording() with the config

* <li>call TextureMovieEncoder#setTextureId() with the texture object that receives frames

* <li>for each frame, after latching it with SurfaceTexture#updateTexImage(),

* call TextureMovieEncoder#frameAvailable().

* </ul>

*

* TODOO: tweak the API (esp. textureId) so it's less awkward for simple use cases.

*/

public class TextureMovieEncoder implements Runnable {

private static final String TAG = MainActivity.TAG;

private static final boolean VERBOSE = false;

private static final long timestampCorrection = 1000000000;

private long timestampCorected;

private static final int MSG_START_RECORDING = 0;

private static final int MSG_STOP_RECORDING = 1;

private static final int MSG_FRAME_AVAILABLE = 2;

private static final int MSG_SET_TEXTURE_ID = 3;

private static final int MSG_UPDATE_SHARED_CONTEXT = 4;

private static final int MSG_QUIT = 5;

private boolean measure_started = false;

private long startTime = -1;

private int cycle = 0;

private long handleFrameTime = 0;

private long last_timestamp = -1;

private float [] transform;

private long last_orig_timestamp = -1;

public long getFrame() {

return frame;

}

private long frame = 0;

private long average_diff = 0;

private long step = 40000000;

private long actTimestamp = 0;

private boolean shouldStop = false;

public void setmSpeedCallback(SpeedControlCallback mSpeedCallback) {

this.mSpeedCallback = mSpeedCallback;

}

private SpeedControlCallback mSpeedCallback;

// ----- accessed exclusively by encoder thread -----

private WindowSurface mInputWindowSurface;

private EglCore mEglCore;

private FullFrameRect mFullScreen;

private int mTextureId;

private VideoEncoderCore mVideoEncoder;

// ----- accessed by multiple threads -----

private volatile EncoderHandler mHandler;

private Object mReadyFence = new Object(); // guards ready/running

private boolean mReady;

private boolean mRunning;

/**

* Encoder configuration.

* <p>

* Object is immutable, which means we can safely pass it between threads without

* explicit synchronization (and don't need to worry about it getting tweaked out from

* under us).

* <p>

* TODO: make frame rate and iframe interval configurable? Maybe use builder pattern

* with reasonable defaults for those and bit rate.

*/

public static class EncoderConfig {

final File mOutputFile;

final int mWidth;

final int mHeight;

final int mBitRate;

final EGLContext mEglContext;

public EncoderConfig(File outputFile, int width, int height, int bitRate,

EGLContext sharedEglContext) {

mOutputFile = outputFile;

mWidth = width;

mHeight = height;

mBitRate = bitRate;

mEglContext = sharedEglContext;

}

@Override

public String toString() {

return "EncoderConfig: " + mWidth + "x" + mHeight + " @" + mBitRate +

" to '" + mOutputFile.toString() + "' ctxt=" + mEglContext;

}

}

/**

* Tells the video recorder to start recording. (Call from non-encoder thread.)

* <p>

* Creates a new thread, which will create an encoder using the provided configuration.

* <p>

* Returns after the recorder thread has started and is ready to accept Messages. The

* encoder may not yet be fully configured.

*/

public void startRecording(EncoderConfig config) {

Log.d(TAG, "Encoder: startRecording()");

synchronized (mReadyFence) {

if (mRunning) {

Log.w(TAG, "Encoder thread already running");

return;

}

mRunning = true;

new Thread(this, "TextureMovieEncoder").start();

while (!mReady) {

try {

mReadyFence.wait();

} catch (InterruptedException ie) {

// ignore

}

}

}

mHandler.sendMessage(mHandler.obtainMessage(MSG_START_RECORDING, config));

}

/**

* Tells the video recorder to stop recording. (Call from non-encoder thread.)

* <p>

* Returns immediately; the encoder/muxer may not yet be finished creating the movie.

* <p>

* TODO: have the encoder thread invoke a callback on the UI thread just before it shuts down

* so we can provide reasonable status UI (and let the caller know that movie encoding

* has completed).

*/

public void stopRecording() {

//mHandler.sendMessage(mHandler.obtainMessage(MSG_STOP_RECORDING));

//mHandler.sendMessage(mHandler.obtainMessage(MSG_QUIT));

// We don't know when these will actually finish (or even start). We don't want to

// delay the UI thread though, so we return immediately.

shouldStop = true;

Log.d(TAG, "Shout down flag set up.");

}

/**

* Returns true if recording has been started.

*/

public boolean isRecording() {

synchronized (mReadyFence) {

return mRunning;

}

}

/**

* Tells the video recorder to refresh its EGL surface. (Call from non-encoder thread.)

*/

public void updateSharedContext(EGLContext sharedContext) {

mHandler.sendMessage(mHandler.obtainMessage(MSG_UPDATE_SHARED_CONTEXT, sharedContext));

}

/**

* Tells the video recorder that a new frame is available. (Call from non-encoder thread.)

* <p>

* This function sends a message and returns immediately. This isn't sufficient -- we

* don't want the caller to latch a new frame until we're done with this one -- but we

* can get away with it so long as the input frame rate is reasonable and the encoder

* thread doesn't stall.

* <p>

* TODO: either block here until the texture has been rendered onto the encoder surface,

* or have a separate "block if still busy" method that the caller can execute immediately

* before it calls updateTexImage(). The latter is preferred because we don't want to

* stall the caller while this thread does work.

*/

public void frameAvailable(SurfaceTexture st) {

synchronized (mReadyFence) {

if (!mReady) {

return;

}

}

transform = new float[16]; // TODOO - avoid alloc every frame

st.getTransformMatrix(transform);

long timestamp = st.getTimestamp();

// if first frame

if (last_timestamp < 0) {

if (!measure_started) {

startTime = System.currentTimeMillis();

measure_started = true;

}

last_timestamp = timestamp;

last_orig_timestamp = timestamp;

}

else {

// HARDCODED FRAME NUMBER :(

// if playback finished or frame number reached

if ((frame == 200) || shouldStop) {

if (measure_started) {

long stopTime = System.currentTimeMillis();

long elapsedTime = stopTime - startTime;

Log.d(TAG, "Rendering time: " + (double)elapsedTime * 0.001 + "[s]");

Log.d(TAG, "HandlingFrame time: " + (double)(stopTime - handleFrameTime) * 0.001 + "[s]");

measure_started = false;

}

mHandler.sendMessage(mHandler.obtainMessage(MSG_STOP_RECORDING));

mHandler.sendMessage(mHandler.obtainMessage(MSG_QUIT));

return;

}

else if (timestamp == 0) {

// Seeing this after device is toggled off/on with power button. The

// first frame back has a zero timestamp.

//

// MPEG4Writer thinks this is cause to abort() in native code, so it's very

// important that we just ignore the frame.

Log.w(TAG, "HEY: got SurfaceTexture with timestamp of zero");

return;

}

// this is workaround for duplicated timestamp

// might cause troubles with some videos

else if ((timestamp == last_orig_timestamp)) {

return;

}

else {

frame++;

mHandler.sendMessage(mHandler.obtainMessage(MSG_FRAME_AVAILABLE,

(int) (actTimestamp >> 32), (int) actTimestamp, transform));

timestampCorected = actTimestamp + timestampCorrection;

mHandler.sendMessage(mHandler.obtainMessage(MSG_FRAME_AVAILABLE,

(int) (timestampCorected >> 32), (int) timestampCorected, transform));

actTimestamp += step;

}

last_orig_timestamp = timestamp;

}

}

/**

* Calculates 'average' diffrence between frames.

* Result is based on first 50 frames.

* Shuld be called in frameAvailiable.

*

* @param timestamp actual frame timestamp

*/

private void calcAndShowAverageDiff(long timestamp) {

if ((frame < 50) && (frame > 0)) {

average_diff += timestamp - last_timestamp;

last_timestamp = timestamp;

}

if (frame == 50) {

average_diff /= frame;

Log.d(TAG, "Average timestamp difference: " + Long.toString(average_diff));

}

}

/**

* Tells the video recorder what texture name to use. This is the external texture that

* we're receiving camera previews in. (Call from non-encoder thread.)

* <p>

* TODOO: do something less clumsy

*/

public void setTextureId(int id) {

synchronized (mReadyFence) {

if (!mReady) {

return;

}

}

mHandler.sendMessage(mHandler.obtainMessage(MSG_SET_TEXTURE_ID, id, 0, null));

}

/**

* Encoder thread entry point. Establishes Looper/Handler and waits for messages.

* <p>

* @see java.lang.Thread#run()

*/

@Override

public void run() {

// Establish a Looper for this thread, and define a Handler for it.

Looper.prepare();

synchronized (mReadyFence) {

mHandler = new EncoderHandler(this);

mReady = true;

mReadyFence.notify();

}

Looper.loop();

Log.d(TAG, "Encoder thread exiting");

synchronized (mReadyFence) {

mReady = mRunning = false;

mHandler = null;

}

}

/**

* Handles encoder state change requests. The handler is created on the encoder thread.

*/

private static class EncoderHandler extends Handler {

private WeakReference<TextureMovieEncoder> mWeakEncoder;

public EncoderHandler(TextureMovieEncoder encoder) {

mWeakEncoder = new WeakReference<TextureMovieEncoder>(encoder);

}

@Override // runs on encoder thread

public void handleMessage(Message inputMessage) {

int what = inputMessage.what;

Object obj = inputMessage.obj;

TextureMovieEncoder encoder = mWeakEncoder.get();

if (encoder == null) {

Log.w(TAG, "EncoderHandler.handleMessage: encoder is null");

return;

}

switch (what) {

case MSG_START_RECORDING:

encoder.handleStartRecording((EncoderConfig) obj);

break;

case MSG_STOP_RECORDING:

encoder.handleStopRecording();

break;

case MSG_FRAME_AVAILABLE:

long timestamp = (((long) inputMessage.arg1) << 32) |

(((long) inputMessage.arg2) & 0xffffffffL);

encoder.handleFrameAvailable((float[]) obj, timestamp);

break;

case MSG_SET_TEXTURE_ID:

encoder.handleSetTexture(inputMessage.arg1);

break;

case MSG_UPDATE_SHARED_CONTEXT:

encoder.handleUpdateSharedContext((EGLContext) inputMessage.obj);

break;

case MSG_QUIT:

Looper.myLooper().quit();

break;

default:

throw new RuntimeException("Unhandled msg what=" + what);

}

}

}

/**

* Starts recording.

*/

private void handleStartRecording(EncoderConfig config) {

Log.d(TAG, "handleStartRecording " + config);

prepareEncoder(config.mEglContext, config.mWidth, config.mHeight, config.mBitRate,

config.mOutputFile);

}

/**

* Handles notification of an available frame.

* <p>

* The texture is rendered onto the encoder's input surface, along with a moving

* box (just because we can).

* <p>

* @param transform The texture transform, from SurfaceTexture.

* @param timestampNanos The frame's timestamp, from SurfaceTexture.

*/

private void handleFrameAvailable(float[] transform, long timestampNanos) {

if (VERBOSE) Log.d(TAG, "handleFrameAvailable tr=" + transform);

if (cycle == 1) {

mVideoEncoder.drainEncoder(false);

mFullScreen.drawFrame(mTextureId, transform, 1.0f);

}

else {

mFullScreen.drawFrame(mTextureId, transform, -1.0f);

}

mInputWindowSurface.setPresentationTime(timestampNanos);

mInputWindowSurface.swapBuffers();

if (cycle == 1) {

mSpeedCallback.setCanRelease(true);

cycle = 0;

} else

cycle++;

}

/**

* Handles a request to stop encoding.

*/

private void handleStopRecording() {

Log.d(TAG, "handleStopRecording");

mVideoEncoder.drainEncoder(true);

releaseEncoder();

}

/**

* Sets the texture name that SurfaceTexture will use when frames are received.

*/

private void handleSetTexture(int id) {

//Log.d(TAG, "handleSetTexture " + id);

mTextureId = id;

}

/**

* Tears down the EGL surface and context we've been using to feed the MediaCodec input

* surface, and replaces it with a new one that shares with the new context.

* <p>

* This is useful if the old context we were sharing with went away (maybe a GLSurfaceView

* that got torn down) and we need to hook up with the new one.

*/

private void handleUpdateSharedContext(EGLContext newSharedContext) {

Log.d(TAG, "handleUpdatedSharedContext " + newSharedContext);

// Release the EGLSurface and EGLContext.

mInputWindowSurface.releaseEglSurface();

mFullScreen.release(false);

mEglCore.release();

// Create a new EGLContext and recreate the window surface.

mEglCore = new EglCore(newSharedContext, EglCore.FLAG_RECORDABLE);

mInputWindowSurface.recreate(mEglCore);

mInputWindowSurface.makeCurrent();

// Create new programs and such for the new context.

mFullScreen = new FullFrameRect(

new Texture2dProgram(Texture2dProgram.ProgramType.TEXTURE_SBS));

}

private void prepareEncoder(EGLContext sharedContext, int width, int height, int bitRate,

File outputFile) {

try {

mVideoEncoder = new VideoEncoderCore(width, height, bitRate, outputFile);

} catch (IOException ioe) {

throw new RuntimeException(ioe);

}

mEglCore = new EglCore(sharedContext, EglCore.FLAG_RECORDABLE);

mInputWindowSurface = new WindowSurface(mEglCore, mVideoEncoder.getInputSurface(), true);

mInputWindowSurface.makeCurrent();

mFullScreen = new FullFrameRect(

new Texture2dProgram(Texture2dProgram.ProgramType.TEXTURE_SBS));

}

private void releaseEncoder() {

mVideoEncoder.release();

if (mInputWindowSurface != null) {

mInputWindowSurface.release();

mInputWindowSurface = null;

}

if (mFullScreen != null) {

mFullScreen.release(false);

mFullScreen = null;

}

if (mEglCore != null) {

mEglCore.release();

mEglCore = null;

}

}

/**

* Draws a box, with position offset.

*/

private void drawBox(int posn) {

final int width = mInputWindowSurface.getWidth();

int xpos = (posn * 4) % (width - 50);

GLES20.glEnable(GLES20.GL_SCISSOR_TEST);

GLES20.glScissor(xpos, 0, 100, 100);

GLES20.glClearColor(1.0f, 0.0f, 1.0f, 1.0f);

GLES20.glClear(GLES20.GL_COLOR_BUFFER_BIT);

GLES20.glDisable(GLES20.GL_SCISSOR_TEST);

}

}

Is my idea viable? Or is there better/correct way to delay one of the videos?

网友答案:

It appears that my first idea with switching timestamps was invalid.

Following Fadden's suggestion, I've succesfuly created delay by using two decoders. I've modified code of grafika's MoviePlayer so it contains two pairs of extractor-decoder. Decoders has separate output textures. Extracting loops are running in separate threads. I thought that this approach will cause some heavy drop in performance, but it appears that it didn't happen, performance is still acceptable for my needs.

分享给朋友:
您可能感兴趣的文章:
随机阅读: