Separate the different SPS fixups by decoder. Go back to the old way of doing bitstream restrictions because the new way seems to be broken. TI OMAP4's Ducati decoder works now :)

AndroidManifest.xml

@@ -1,8 +1,8 @@
 <?xml version="1.0" encoding="utf-8"?>
 <manifest xmlns:android="http://schemas.android.com/apk/res/android"
     package="com.limelight"
-    android:versionCode="9"
-    android:versionName="2.1.3" >
+    android:versionCode="11"
+    android:versionName="2.1.5" >
 
     <uses-sdk
         android:minSdkVersion="16"

assets/controllers.json

@@ -0,0 +1,12 @@
+{
+   "SupportedControllers" : {
+      "Gamepad" : {},
+      "Remote" : "false",
+      "SecondScreen" : {
+         "DPad" : "false",
+         "AnalogSticks" : "0",
+         "DigitalButtons" : "0",
+         "Mouse" : "false"
+      }
+   }
+}

src/com/limelight/Game.java

@@ -34,7 +34,7 @@ import android.view.WindowManager;
 import android.widget.Toast;
 
 
-public class Game extends Activity implements OnGenericMotionListener, OnTouchListener, NvConnectionListener {
+public class Game extends Activity implements SurfaceHolder.Callback, OnGenericMotionListener, OnTouchListener, NvConnectionListener {
 	private int lastMouseX = Integer.MIN_VALUE;
 	private int lastMouseY = Integer.MIN_VALUE;
 	private int lastButtonState = 0;
@@ -52,6 +52,10 @@ public class Game extends Activity implements OnGenericMotionListener, OnTouchLi
 	private NvConnection conn;
 	private SpinnerDialog spinner;
 	private boolean displayedFailureDialog = false;
+	private boolean connecting = false;
+	private boolean connected = false;
+	
+	private int drFlags = 0;
 	
 	public static final String PREFS_FILE_NAME = "gameprefs";
 	
@@ -101,7 +105,6 @@ public class Game extends Activity implements OnGenericMotionListener, OnTouchLi
 		
 		// Read the stream preferences
 		SharedPreferences prefs = getSharedPreferences(PREFS_FILE_NAME, Context.MODE_MULTI_PROCESS);
-		int drFlags = 0;
 		switch (prefs.getInt(Game.DECODER_PREF_STRING, Game.DEFAULT_DECODER)) {
 		case Game.FORCE_SOFTWARE_DECODER:
 			drFlags |= VideoDecoderRenderer.FLAG_FORCE_SOFTWARE_DECODING;
@@ -130,8 +133,9 @@ public class Game extends Activity implements OnGenericMotionListener, OnTouchLi
 				new StreamConfiguration(width, height, refreshRate));
 		keybTranslator = new KeyboardTranslator(conn);
 		controllerHandler = new ControllerHandler(conn);
-		conn.start(PlatformBinding.getDeviceName(), sv.getHolder(), drFlags,
-				PlatformBinding.getAudioRenderer(), new ConfigurableDecoderRenderer());
+		
+		// The connection will be started when the surface gets created
+		sh.addCallback(this);
 	}
 	
 	private void checkDataConnection()
@@ -420,6 +424,7 @@ public class Game extends Activity implements OnGenericMotionListener, OnTouchLi
 			displayedFailureDialog = true;
 			Dialog.displayDialog(this, "Connection Error", "Starting "+stage.getName()+" failed", true);
 			conn.stop();
+			connecting = false;
 		}
 	}
 
@@ -430,6 +435,7 @@ public class Game extends Activity implements OnGenericMotionListener, OnTouchLi
 			e.printStackTrace();
 			Dialog.displayDialog(this, "Connection Terminated", "The connection failed unexpectedly", true);
 			conn.stop();
+			connected = false;
 		}
 	}
 
@@ -438,6 +444,9 @@ public class Game extends Activity implements OnGenericMotionListener, OnTouchLi
 		spinner.dismiss();
 		spinner = null;
 		
+		connecting = false;
+		connected = true;
+		
 		hideSystemUi();
 	}
 
@@ -460,4 +469,25 @@ public class Game extends Activity implements OnGenericMotionListener, OnTouchLi
 			}
 		});
 	}
+
+	@Override
+	public void surfaceChanged(SurfaceHolder holder, int format, int width, int height) {
+	}
+
+	@Override
+	public void surfaceCreated(SurfaceHolder holder) {
+		if (!connected && !connecting) {
+			connecting = true;
+			conn.start(PlatformBinding.getDeviceName(), holder, drFlags,
+					PlatformBinding.getAudioRenderer(), new ConfigurableDecoderRenderer());
+		}
+	}
+
+	@Override
+	public void surfaceDestroyed(SurfaceHolder holder) {
+		if (connected) {
+			conn.stop();
+			connected = false;
+		}
+	}
 }

src/com/limelight/binding/video/MediaCodecDecoderRenderer.java

@@ -24,23 +24,27 @@ public class MediaCodecDecoderRenderer implements VideoDecoderRenderer {
 	private MediaCodec videoDecoder;
 	private Thread rendererThread;
 	private int redrawRate;
-	private boolean needsSpsFixup;
+	private boolean needsSpsBitstreamFixup;
+	private boolean needsSpsNumRefFixup;
 	private boolean fastInputQueueing;
 	
 	public static final List<String> blacklistedDecoderPrefixes;
-	public static final List<String> spsFixupDecoderPrefixes;
+	public static final List<String> spsFixupBitsreamFixupDecoderPrefixes;
+	public static final List<String> spsFixupNumRefFixupDecoderPrefixes;
 	public static final List<String> fastInputQueueingPrefixes;
 	
 	static {
 		blacklistedDecoderPrefixes = new LinkedList<String>();
 		
-		// TI's decoder technically supports high profile but doesn't work for some reason
-		blacklistedDecoderPrefixes.add("omx.TI");
+		// Nothing here right now :)
 	}
 	
 	static {
-		spsFixupDecoderPrefixes = new LinkedList<String>();
-		spsFixupDecoderPrefixes.add("omx.nvidia");
+		spsFixupBitsreamFixupDecoderPrefixes = new LinkedList<String>();
+		spsFixupBitsreamFixupDecoderPrefixes.add("omx.nvidia");
+		
+		spsFixupNumRefFixupDecoderPrefixes = new LinkedList<String>();
+		spsFixupNumRefFixupDecoderPrefixes.add("omx.TI");
 	}
 	
 	static {
@@ -128,9 +132,13 @@ public class MediaCodecDecoderRenderer implements VideoDecoderRenderer {
 		MediaCodecInfo safeDecoder = findSafeDecoder();
 		if (safeDecoder != null) {
 			videoDecoder = MediaCodec.createByCodecName(safeDecoder.getName());
-			needsSpsFixup = isDecoderInList(spsFixupDecoderPrefixes, safeDecoder.getName());
-			if (needsSpsFixup) {
-				LimeLog.info("Decoder "+safeDecoder.getName()+" needs SPS fixup");
+			needsSpsBitstreamFixup = isDecoderInList(spsFixupBitsreamFixupDecoderPrefixes, safeDecoder.getName());
+			needsSpsNumRefFixup = isDecoderInList(spsFixupNumRefFixupDecoderPrefixes, safeDecoder.getName());
+			if (needsSpsBitstreamFixup) {
+				LimeLog.info("Decoder "+safeDecoder.getName()+" needs SPS bitstream restrictions fixup");
+			}
+			if (needsSpsNumRefFixup) {
+				LimeLog.info("Decoder "+safeDecoder.getName()+" needs SPS ref num fixup");
 			}
 			fastInputQueueing = isDecoderInList(fastInputQueueingPrefixes, safeDecoder.getName());
 			if (fastInputQueueing) {
@@ -139,7 +147,8 @@ public class MediaCodecDecoderRenderer implements VideoDecoderRenderer {
 		}
 		else {
 			videoDecoder = MediaCodec.createDecoderByType("video/avc");
-			needsSpsFixup = false;
+			needsSpsBitstreamFixup = false;
+			needsSpsNumRefFixup = false;
 			fastInputQueueing = false;
 		}
 		
@@ -256,39 +265,49 @@ public class MediaCodecDecoderRenderer implements VideoDecoderRenderer {
 			// Clear old input data
 			buf.clear();
 
-			// The SPS that comes in the current H264 bytestream doesn't set bitstream_restriction_flag
-			// or max_dec_frame_buffering which increases decoding latency on Tegra.
-			// We manually modify the SPS here to speed-up decoding if the decoder was flagged as needing it.
-			if (needsSpsFixup) {
+			if (needsSpsBitstreamFixup || needsSpsNumRefFixup) {
 				ByteBufferDescriptor header = decodeUnit.getBufferList().get(0);
-				// Check for SPS NALU type
 				if (header.data[header.offset+4] == 0x67) {
-					int spsLength;
+					// TI OMAP4 requires a reference frame count of 1 to decode successfully
+					if (needsSpsNumRefFixup) {
+						LimeLog.info("Fixing up num ref frames");
+						this.replace(header, 80, 9, new byte[] {0x40}, 3);
+					}
 
-					switch (header.length) {
-					case 26:
-						LimeLog.info("Modifying SPS (26)");
-						buf.put(header.data, header.offset, 24);
-						buf.put((byte) 0x11);
-						buf.put((byte) 0xe3);
-						buf.put((byte) 0x06);
-						buf.put((byte) 0x50);
-						spsLength = header.length + 2;
-						break;
-					case 27:
-						LimeLog.info("Modifying SPS (27)");
-						buf.put(header.data, header.offset, 25);
-						buf.put((byte) 0x04);
-						buf.put((byte) 0x78);
-						buf.put((byte) 0xc1);
-						buf.put((byte) 0x94);
-						spsLength = header.length + 2;
-						break;
-					default:
-						LimeLog.warning("Unknown SPS of length "+header.length);
+					// The SPS that comes in the current H264 bytestream doesn't set bitstream_restriction_flag
+					// or max_dec_frame_buffering which increases decoding latency on Tegra.
+					// We manually modify the SPS here to speed-up decoding if the decoder was flagged as needing it.
+					int spsLength;
+					if (needsSpsBitstreamFixup) {
+						switch (header.length) {
+						case 26:
+							LimeLog.info("Adding bitstream restrictions to SPS (26)");
+							buf.put(header.data, header.offset, 24);
+							buf.put((byte) 0x11);
+							buf.put((byte) 0xe3);
+							buf.put((byte) 0x06);
+							buf.put((byte) 0x50);
+							spsLength = header.length + 2;
+							break;
+						case 27:
+							LimeLog.info("Adding bitstream restrictions to SPS (27)");
+							buf.put(header.data, header.offset, 25);
+							buf.put((byte) 0x04);
+							buf.put((byte) 0x78);
+							buf.put((byte) 0xc1);
+							buf.put((byte) 0x94);
+							spsLength = header.length + 2;
+							break;
+						default:
+							LimeLog.warning("Unknown SPS of length "+header.length);
+							buf.put(header.data, header.offset, header.length);
+							spsLength = header.length;
+							break;
+						}
+					}
+					else {
 						buf.put(header.data, header.offset, header.length);
 						spsLength = header.length;
-						break;
 					}
 
 					videoDecoder.queueInputBuffer(inputIndex,
@@ -316,4 +335,83 @@ public class MediaCodecDecoderRenderer implements VideoDecoderRenderer {
 	public int getCapabilities() {
 		return fastInputQueueing ? VideoDecoderRenderer.CAPABILITY_DIRECT_SUBMIT : 0;
 	}
+
+	/**
+	 * Replace bits in array 
+	 * @param source array in which bits should be replaced
+	 * @param srcOffset offset in bits where replacement should take place
+	 * @param srcLength length in bits of data that should be replaced
+	 * @param data data array with the the replacement data
+	 * @param dataLength length of replacement data in bits
+	 */
+	public void replace(ByteBufferDescriptor source, int srcOffset, int srcLength, byte[] data, int dataLength) {
+		//Add 7 to always round up
+		int length = (source.length*8-srcLength+dataLength+7)/8;
+
+		int bitOffset = srcOffset%8;
+		int byteOffset = srcOffset/8;
+
+		byte dest[] = null;
+		int offset = 0;
+		if (length>source.length) {
+			dest = new byte[length];
+
+			//Copy the first bytes
+			System.arraycopy(source.data, source.offset, dest, offset, byteOffset);
+		} else {
+			dest = source.data;
+			offset = source.offset;
+		}
+
+		int byteLength = (bitOffset+dataLength+7)/8;
+		int bitTrailing = 8 - (srcOffset+dataLength) % 8;
+		for (int i=0;i<byteLength;i++) {
+			byte result = 0;
+			if (i != 0)
+				result = (byte) (data[i-1] << 8-bitOffset);
+			else if (bitOffset > 0)
+				result = (byte) (source.data[byteOffset+source.offset] & (0xFF << 8-bitOffset));
+
+			if (i == 0 || i != byteLength-1) {
+				byte moved = (byte) ((data[i]&0xFF) >>> bitOffset);
+				result |= moved;
+			}
+
+			if (i == byteLength-1 && bitTrailing > 0) {
+				int sourceOffset = srcOffset+srcLength/8;
+				int bitMove = (dataLength-srcLength)%8;
+				if (bitMove<0) {
+					result |= (byte) (source.data[sourceOffset+source.offset] << -bitMove & (0xFF >>> bitTrailing));
+					result |= (byte) (source.data[sourceOffset+1+source.offset] << -bitMove & (0xFF >>> 8+bitMove));
+				} else {
+					byte moved = (byte) ((source.data[sourceOffset+source.offset]&0xFF) >>> bitOffset);
+					result |= moved;
+				}
+			}
+
+			dest[i+byteOffset+offset] = result;
+		}
+
+		//Source offset
+		byteOffset += srcLength/8;
+		bitOffset = (srcOffset+dataLength-srcLength)%8;
+
+		//Offset in destination
+		int destOffset = (srcOffset+dataLength)/8;
+
+		for (int i=1;i<source.length-byteOffset;i++) {
+			int diff = destOffset >= byteOffset-1?i:source.length-byteOffset-i;
+
+			byte result = 0;
+			result = (byte) (source.data[byteOffset+diff-1+source.offset] << 8-bitOffset);
+			byte moved = (byte) ((source.data[byteOffset+diff+source.offset]&0xFF) >>> bitOffset);
+			result ^= moved;
+
+			dest[diff+destOffset+offset] = result;
+		}
+
+		source.data = dest;
+		source.offset = offset;
+		source.length = length;
+	}
 }