Further optimize the JNI code for faster H264 decoding. Add an experimental RenderScript renderer.

jni/nv_avc_dec/nv_avc_dec.c

@@ -8,14 +8,20 @@
 #include <jni.h>
 #include <android/native_window_jni.h>
 
+// General decoder and renderer state
+AVPacket pkt;
 AVCodec* decoder;
 AVCodecContext* decoder_ctx;
 AVFrame* yuv_frame;
-AVFrame* rgb_frame;
 AVFrame* dec_frame;
 pthread_mutex_t mutex;
+
+// Color conversion and rendering
+AVFrame* rgb_frame;
 char* rgb_frame_buf;
 struct SwsContext* scaler_ctx;
+ANativeWindow* window;
+
 
 #define RENDER_PIX_FMT AV_PIX_FMT_RGBA
 #define BYTES_PER_PIXEL 4
@@ -32,6 +38,8 @@ struct SwsContext* scaler_ctx;
 #define BILINEAR_FILTERING      0x10
 // Uses a faster bilinear filtering with lower image quality
 #define FAST_BILINEAR_FILTERING 0x20
+// Disables color conversion (output is NV21)
+#define NO_COLOR_CONVERSION     0x40
 
 // This function must be called before
 // any other decoding functions
@@ -45,6 +53,8 @@ int nv_avc_init(int width, int height, int perf_lvl, int thread_count) {
 	av_log_set_level(AV_LOG_QUIET);
 	avcodec_register_all();
 	
+	av_init_packet(&pkt);
+
 	decoder = avcodec_find_decoder(AV_CODEC_ID_H264);
 	if (decoder == NULL) {
 		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
@@ -99,53 +109,55 @@ int nv_avc_init(int width, int height, int perf_lvl, int thread_count) {
 		return -1;
 	}
 	
-	rgb_frame = av_frame_alloc();
-	if (rgb_frame == NULL) {
-		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
-			"Couldn't allocate frame");
-		return -1;
-	}
+	if (!(perf_lvl & NO_COLOR_CONVERSION)) {
+		rgb_frame = av_frame_alloc();
+		if (rgb_frame == NULL) {
+			__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
+				"Couldn't allocate frame");
+			return -1;
+		}
 
-	rgb_frame_buf = (char*)av_malloc(width * height * BYTES_PER_PIXEL);
-	if (rgb_frame_buf == NULL) {
-		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
-			"Couldn't allocate picture");
-		return -1;
-	}
+		rgb_frame_buf = (char*)av_malloc(width * height * BYTES_PER_PIXEL);
+		if (rgb_frame_buf == NULL) {
+			__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
+				"Couldn't allocate picture");
+			return -1;
+		}
 
-	err = avpicture_fill((AVPicture*)rgb_frame,
-		rgb_frame_buf,
-		RENDER_PIX_FMT,
-		decoder_ctx->width,
-		decoder_ctx->height);
-	if (err < 0) {
-		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
-			"Couldn't fill picture");
-		return err;
-	}
+		err = avpicture_fill((AVPicture*)rgb_frame,
+			rgb_frame_buf,
+			RENDER_PIX_FMT,
+			decoder_ctx->width,
+			decoder_ctx->height);
+		if (err < 0) {
+			__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
+				"Couldn't fill picture");
+			return err;
+		}
 
-	if (perf_lvl & FAST_BILINEAR_FILTERING) {
-		filtering = SWS_FAST_BILINEAR;
-	}
-	else if (perf_lvl & BILINEAR_FILTERING) {
-		filtering = SWS_BILINEAR;
-	}
-	else {
-		filtering = SWS_BICUBIC;
-	}
+		if (perf_lvl & FAST_BILINEAR_FILTERING) {
+			filtering = SWS_FAST_BILINEAR;
+		}
+		else if (perf_lvl & BILINEAR_FILTERING) {
+			filtering = SWS_BILINEAR;
+		}
+		else {
+			filtering = SWS_BICUBIC;
+		}
 
-	scaler_ctx = sws_getContext(decoder_ctx->width,
-		decoder_ctx->height,
-		decoder_ctx->pix_fmt,
-		decoder_ctx->width,
-		decoder_ctx->height,
-		RENDER_PIX_FMT,
-		filtering,
-		NULL, NULL, NULL);
-	if (scaler_ctx == NULL) {
-		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
-			"Couldn't get scaler context");
-		return -1;
+		scaler_ctx = sws_getContext(decoder_ctx->width,
+			decoder_ctx->height,
+			decoder_ctx->pix_fmt,
+			decoder_ctx->width,
+			decoder_ctx->height,
+			RENDER_PIX_FMT,
+			filtering,
+			NULL, NULL, NULL);
+		if (scaler_ctx == NULL) {
+			__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
+				"Couldn't get scaler context");
+			return -1;
+		}
 	}
 
 	return 0;
@@ -179,14 +191,15 @@ void nv_avc_destroy(void) {
 		av_free(rgb_frame_buf);
 		rgb_frame_buf = NULL;
 	}
+	if (window) {
+		ANativeWindow_release(window);
+		window = NULL;
+	}
 	pthread_mutex_destroy(&mutex);
 }
 
-void nv_avc_redraw(JNIEnv *env, jobject surface) {
-	ANativeWindow* window;
-	ANativeWindow_Buffer buffer;
-	AVFrame *our_yuv_frame;
-	int err;
+static AVFrame* dequeue_new_frame(void) {
+	AVFrame *our_yuv_frame = NULL;
 
 	pthread_mutex_lock(&mutex);
 
@@ -196,77 +209,149 @@ void nv_avc_redraw(JNIEnv *env, jobject surface) {
 		// responsible for freeing it when we're done
 		our_yuv_frame = yuv_frame;
 		yuv_frame = NULL;
+	}
 
-		// The remaining processing can be done without the mutex
-		pthread_mutex_unlock(&mutex);
+	pthread_mutex_unlock(&mutex);
 
-		// Convert the YUV image to RGB
-		err = sws_scale(scaler_ctx,
-			our_yuv_frame->data,
-			our_yuv_frame->linesize,
-			0,
-			decoder_ctx->height,
-			rgb_frame->data,
-			rgb_frame->linesize);
-		if (err != decoder_ctx->height) {
-			__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
-					"Scaling failed");
-			goto free_frame_and_return;
-		}
+	return our_yuv_frame;
+}
 
-		window = ANativeWindow_fromSurface(env, surface);
-		if (window == NULL) {
-			__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
-					"Failed to get window from surface");
-			goto free_frame_and_return;
-		}
+static int update_rgb_frame(void) {
+	AVFrame *our_yuv_frame;
+	int err;
 
+	our_yuv_frame = dequeue_new_frame();
+	if (our_yuv_frame == NULL) {
+		return 0;
+	}
+
+	// Convert the YUV image to RGB
+	err = sws_scale(scaler_ctx,
+		our_yuv_frame->data,
+		our_yuv_frame->linesize,
+		0,
+		decoder_ctx->height,
+		rgb_frame->data,
+		rgb_frame->linesize);
+
+	av_frame_free(&our_yuv_frame);
+
+	if (err != decoder_ctx->height) {
+		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
+				"Scaling failed");
+		return 0;
+	}
+
+	return 1;
+}
+
+static int render_rgb_to_buffer(char* buffer, int size) {
+	int err;
+
+	// Draw the frame to the buffer
+	err = avpicture_layout((AVPicture*)rgb_frame,
+		RENDER_PIX_FMT,
+		decoder_ctx->width,
+		decoder_ctx->height,
+		buffer,
+		size);
+	if (err < 0) {
+		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
+			"Picture fill failed");
+		return 0;
+	}
+
+	return 1;
+}
+
+int nv_avc_get_raw_frame(char* buffer, int size) {
+	AVFrame *our_yuv_frame;
+	int err;
+
+	our_yuv_frame = dequeue_new_frame();
+	if (our_yuv_frame == NULL) {
+		return 0;
+	}
+
+	err = avpicture_layout((AVPicture*)our_yuv_frame,
+		decoder_ctx->pix_fmt,
+		decoder_ctx->width,
+		decoder_ctx->height,
+		buffer,
+		size);
+
+	av_frame_free(&our_yuv_frame);
+
+	return (err >= 0);
+}
+
+int nv_avc_get_rgb_frame(char* buffer, int size) {
+	return (update_rgb_frame() && render_rgb_to_buffer(buffer, size));
+}
+
+int nv_avc_set_render_target(JNIEnv *env, jobject surface) {
+	// Release the old window
+	if (window) {
+		ANativeWindow_release(window);
+		window = NULL;
+	}
+
+	// If no new surface was supplied, we're done
+	if (surface == NULL) {
+		return 1;
+	}
+
+	// Get a window from the surface
+	window = ANativeWindow_fromSurface(env, surface);
+	if (window == NULL) {
+		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
+				"Failed to get window from surface");
+		return 0;
+	}
+
+	return 1;
+}
+
+int nv_avc_redraw(void) {
+	ANativeWindow_Buffer buffer;
+	int ret = 0;
+
+	// Check if there's a new frame
+	if (update_rgb_frame()) {
 		// Lock down a render buffer
 		if (ANativeWindow_lock(window, &buffer, NULL) >= 0) {
 			// Draw the frame to the buffer
-			err = avpicture_layout((AVPicture*)rgb_frame,
-				RENDER_PIX_FMT,
-				decoder_ctx->width,
-				decoder_ctx->height,
-				buffer.bits,
+			if (render_rgb_to_buffer(buffer.bits, 
 				decoder_ctx->width *
 				decoder_ctx->height *
-				BYTES_PER_PIXEL);
-			if (err < 0) {
-				__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
-					"Picture fill failed");
+				BYTES_PER_PIXEL)) {
+				// A new frame will be drawn
+				ret = 1;
 			}
 
 			// Draw the frame to the surface
 			ANativeWindow_unlockAndPost(window);
 		}
-
-		ANativeWindow_release(window);
-		
-	free_frame_and_return:
-		av_frame_free(&our_yuv_frame);
-	}
-	else {
-		pthread_mutex_unlock(&mutex);
 	}
+	
+	return ret;
+}
+
+int nv_avc_get_input_padding_size(void) {
+	return FF_INPUT_BUFFER_PADDING_SIZE;
 }
 
 // packets must be decoded in order
+// indata must be inlen + FF_INPUT_BUFFER_PADDING_SIZE in length
 int nv_avc_decode(unsigned char* indata, int inlen) {
 	int err;
-	AVPacket pkt;
-	int got_pic;
+	int got_pic = 0;
 
-	err = av_new_packet(&pkt, inlen);
-	if (err < 0) {
-		__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
-			"Failed to allocate packet");
-		return err;
-	}
-
-	memcpy(pkt.data, indata, inlen);
+	pkt.data = indata;
+	pkt.size = inlen;
 
 	while (pkt.size > 0) {
+		got_pic = 0;
 		err = avcodec_decode_video2(
 			decoder_ctx,
 			dec_frame,
@@ -275,29 +360,28 @@ int nv_avc_decode(unsigned char* indata, int inlen) {
 		if (err < 0) {
 			__android_log_write(ANDROID_LOG_ERROR, "NVAVCDEC",
 				"Decode failed");
-			pthread_mutex_unlock(&mutex);
+			got_pic = 0;
 			break;
 		}
 
-		if (got_pic) {
-			pthread_mutex_lock(&mutex);
-
-			// Only clone this frame if the last frame was taken.
-			// This saves on extra copies for frames that don't get
-			// rendered.
-			if (yuv_frame == NULL) {
-				// Clone a new frame
-				yuv_frame = av_frame_clone(dec_frame);
-			}
-
-			pthread_mutex_unlock(&mutex);
-		}
-
 		pkt.size -= err;
 		pkt.data += err;
 	}
 	
-	av_free_packet(&pkt);
+	// Only copy the picture at the end of decoding the packet
+	if (got_pic) {
+		pthread_mutex_lock(&mutex);
+
+		// Only clone this frame if the last frame was taken.
+		// This saves on extra copies for frames that don't get
+		// rendered.
+		if (yuv_frame == NULL) {
+			// Clone a new frame
+			yuv_frame = av_frame_clone(dec_frame);
+		}
+
+		pthread_mutex_unlock(&mutex);
+	}
 
 	return err < 0 ? err : 0;
 }

jni/nv_avc_dec/nv_avc_dec.h

@@ -2,5 +2,12 @@
 
 int nv_avc_init(int width, int height, int perf_lvl, int thread_count);
 void nv_avc_destroy(void);
-void nv_avc_redraw(JNIEnv *env, jobject surface);
+
+int nv_avc_get_raw_frame(char* buffer, int size);
+
+int nv_avc_get_rgb_frame(char* buffer, int size);
+int nv_avc_set_render_target(JNIEnv *env, jobject surface);
+int nv_avc_redraw(void);
+
+int nv_avc_get_input_padding_size(void);
 int nv_avc_decode(unsigned char* indata, int inlen);

jni/nv_avc_dec/nv_avc_dec_jni.c

@@ -6,7 +6,7 @@
 // This function must be called before
 // any other decoding functions
 JNIEXPORT jint JNICALL
-Java_com_limelight_nvstream_av_video_AvcDecoder_init(JNIEnv *env, jobject this, jint width,
+Java_com_limelight_nvstream_av_video_cpu_AvcDecoder_init(JNIEnv *env, jobject this, jint width,
 	jint height, jint perflvl, jint threadcount)
 {
 	return nv_avc_init(width, height, perflvl, threadcount);
@@ -15,20 +15,69 @@ Java_com_limelight_nvstream_av_video_AvcDecoder_init(JNIEnv *env, jobject this,
 // This function must be called after
 // decoding is finished
 JNIEXPORT void JNICALL
-Java_com_limelight_nvstream_av_video_AvcDecoder_destroy(JNIEnv *env, jobject this) {
+Java_com_limelight_nvstream_av_video_cpu_AvcDecoder_destroy(JNIEnv *env, jobject this) {
 	nv_avc_destroy();
 }
 
+// fills the output buffer with a raw YUV frame
+JNIEXPORT jboolean JNICALL
+Java_com_limelight_nvstream_av_video_cpu_AvcDecoder_getRawFrame(
+	JNIEnv *env, jobject this, // JNI parameters
+	jbyteArray outdata, jint outlen) // Output data
+{
+	jint ret;
+	jbyte* jni_output_data;
+
+	jni_output_data = (*env)->GetByteArrayElements(env, outdata, 0);
+
+	ret = nv_avc_get_raw_frame(jni_output_data, outlen);
+
+	(*env)->ReleaseByteArrayElements(env, outdata, jni_output_data, 0);
+
+	return ret != 0 ? JNI_TRUE : JNI_FALSE;
+}
+
+// fills the output buffer with an RGB frame
+JNIEXPORT jboolean JNICALL
+Java_com_limelight_nvstream_av_video_cpu_AvcDecoder_getRgbFrame(
+	JNIEnv *env, jobject this, // JNI parameters
+	jbyteArray outdata, jint outlen) // Output data
+{
+	jint ret;
+	jbyte* jni_output_data;
+
+	jni_output_data = (*env)->GetByteArrayElements(env, outdata, 0);
+
+	ret = nv_avc_get_rgb_frame(jni_output_data, outlen);
+
+	(*env)->ReleaseByteArrayElements(env, outdata, jni_output_data, 0);
+
+	return ret != 0 ? JNI_TRUE : JNI_FALSE;
+}
+
+// This function sets the rendering target for redraw
+JNIEXPORT jboolean JNICALL
+Java_com_limelight_nvstream_av_video_cpu_AvcDecoder_setRenderTarget(JNIEnv *env, jobject this, jobject surface) {
+	return nv_avc_set_render_target(env, surface) != 0 ? JNI_TRUE : JNI_FALSE;
+}
+
 // This function redraws the surface
-JNIEXPORT void JNICALL
-Java_com_limelight_nvstream_av_video_AvcDecoder_redraw(JNIEnv *env, jobject this, jobject surface) {
-	nv_avc_redraw(env, surface);
+JNIEXPORT jboolean JNICALL
+Java_com_limelight_nvstream_av_video_cpu_AvcDecoder_redraw(JNIEnv *env, jobject this) {
+	return nv_avc_redraw() != 0 ? JNI_TRUE : JNI_FALSE;
+}
+
+// This function returns the required input buffer padding
+JNIEXPORT jint JNICALL
+Java_com_limelight_nvstream_av_video_cpu_AvcDecoder_getInputPaddingSize(JNIEnv *env, jobject this) {
+	return nv_avc_get_input_padding_size();
 }
 
 // packets must be decoded in order
+// the input buffer must have proper padding
 // returns 0 on success, < 0 on error
 JNIEXPORT jint JNICALL
-Java_com_limelight_nvstream_av_video_AvcDecoder_decode(
+Java_com_limelight_nvstream_av_video_cpu_AvcDecoder_decode(
 	JNIEnv *env, jobject this, // JNI parameters
 	jbyteArray indata, jint inoff, jint inlen)
 {

src/com/limelight/nvstream/NvConnection.java

@@ -150,7 +150,7 @@ public class NvConnection {
 	private boolean startVideoStream() throws IOException
 	{
 		videoStream = new NvVideoStream(hostAddr, listener, controlStream);
-		videoStream.startVideoStream(video, drFlags);
+		videoStream.startVideoStream(activity, video, drFlags);
 		return true;
 	}
 	

src/com/limelight/nvstream/NvVideoStream.java

@@ -17,10 +17,11 @@ import com.limelight.nvstream.av.AvRtpPacket;
 import com.limelight.nvstream.av.ConnectionStatusListener;
 import com.limelight.nvstream.av.video.AvVideoDepacketizer;
 import com.limelight.nvstream.av.video.AvVideoPacket;
-import com.limelight.nvstream.av.video.CpuDecoderRenderer;
 import com.limelight.nvstream.av.video.DecoderRenderer;
 import com.limelight.nvstream.av.video.MediaCodecDecoderRenderer;
+import com.limelight.nvstream.av.video.cpu.CpuDecoderRenderer;
 
+import android.content.Context;
 import android.os.Build;
 import android.view.SurfaceHolder;
 
@@ -129,7 +130,7 @@ public class NvVideoStream {
 		rtp = new DatagramSocket(RTP_PORT);
 	}
 	
-	public void setupDecoderRenderer(SurfaceHolder renderTarget, int drFlags) {		
+	public void setupDecoderRenderer(Context context, SurfaceHolder renderTarget, int drFlags) {		
 		if (Build.HARDWARE.equals("goldfish")) {
 			// Emulator - don't render video (it's slow!)
 			decrend = null;
@@ -144,14 +145,14 @@ public class NvVideoStream {
 		}
 		
 		if (decrend != null) {
-			decrend.setup(1280, 720, renderTarget, drFlags);
+			decrend.setup(context, 1280, 720, renderTarget, drFlags);
 		}
 	}
 
-	public void startVideoStream(final SurfaceHolder surface, int drFlags) throws IOException
+	public void startVideoStream(Context context, SurfaceHolder surface, int drFlags) throws IOException
 	{
 		// Setup the decoder and renderer
-		setupDecoderRenderer(surface, drFlags);
+		setupDecoderRenderer(context, surface, drFlags);
 		
 		// Open RTP sockets and start session
 		setupRtpSession();

src/com/limelight/nvstream/av/video/DecoderRenderer.java

@@ -2,12 +2,13 @@ package com.limelight.nvstream.av.video;
 
 import com.limelight.nvstream.av.AvDecodeUnit;
 
+import android.content.Context;
 import android.view.SurfaceHolder;
 
 public interface DecoderRenderer {
 	public static int FLAG_PREFER_QUALITY = 0x1;
 	
-	public void setup(int width, int height, SurfaceHolder renderTarget, int drFlags);
+	public void setup(Context context, int width, int height, SurfaceHolder renderTarget, int drFlags);
 	
 	public void start();
 	

src/com/limelight/nvstream/av/video/MediaCodecDecoderRenderer.java

@@ -8,6 +8,7 @@ import com.limelight.nvstream.av.AvByteBufferDescriptor;
 import com.limelight.nvstream.av.AvDecodeUnit;
 
 import android.annotation.TargetApi;
+import android.content.Context;
 import android.media.MediaCodec;
 import android.media.MediaCodecInfo;
 import android.media.MediaCodecList;
@@ -73,7 +74,7 @@ public class MediaCodecDecoderRenderer implements DecoderRenderer {
 	}
 	
 	@Override
-	public void setup(int width, int height, SurfaceHolder renderTarget, int drFlags) {
+	public void setup(Context context, int width, int height, SurfaceHolder renderTarget, int drFlags) {
 		videoDecoder = MediaCodec.createByCodecName(findSafeDecoder().getName());
 		MediaFormat videoFormat = MediaFormat.createVideoFormat("video/avc", width, height);
 

src/com/limelight/nvstream/av/video/cpu/AvcDecoder.java

@@ -1,4 +1,4 @@
-package com.limelight.nvstream.av.video;
+package com.limelight.nvstream.av.video.cpu;
 
 import android.view.Surface;
 
@@ -27,9 +27,20 @@ public class AvcDecoder {
 	public static final int BILINEAR_FILTERING = 0x10;
 	/** Uses a faster bilinear filtering with lower image quality */
 	public static final int FAST_BILINEAR_FILTERING = 0x20;
+	/** Disables color conversion (output is NV21) */
+	public static final int NO_COLOR_CONVERSION = 0x40;
 	
 	public static native int init(int width, int height, int perflvl, int threadcount);
 	public static native void destroy();
-	public static native void redraw(Surface surface);
+	
+	// Rendering API when NO_COLOR_CONVERSION == 0
+	public static native boolean setRenderTarget(Surface surface);
+	public static native boolean getRgbFrame(byte[] rgbFrame, int bufferSize);
+	public static native boolean redraw();
+
+	// Rendering API when NO_COLOR_CONVERSION == 1
+	public static native boolean getRawFrame(byte[] yuvFrame, int bufferSize);
+	
+	public static native int getInputPaddingSize();
 	public static native int decode(byte[] indata, int inoff, int inlen);
 }

src/com/limelight/nvstream/av/video/cpu/CpuDecoderRenderer.java

@@ -1,4 +1,4 @@
-package com.limelight.nvstream.av.video;
+package com.limelight.nvstream.av.video.cpu;
 
 import java.io.BufferedReader;
 import java.io.File;
@@ -6,19 +6,24 @@ import java.io.FileReader;
 import java.io.IOException;
 import java.nio.ByteBuffer;
 
-import android.view.Surface;
+import android.content.Context;
 import android.view.SurfaceHolder;
 
 import com.limelight.nvstream.av.AvByteBufferDescriptor;
 import com.limelight.nvstream.av.AvDecodeUnit;
+import com.limelight.nvstream.av.video.DecoderRenderer;
 
 public class CpuDecoderRenderer implements DecoderRenderer {
 
-	private Surface renderTarget;
-	private ByteBuffer decoderBuffer;
 	private Thread rendererThread;
 	private int targetFps;
 	
+	private static final int DECODER_BUFFER_SIZE = 92*1024;
+	private ByteBuffer decoderBuffer;
+	
+	private RsRenderer rsRenderer;
+	private byte[] frameBuffer;
+	
 	// Only sleep if the difference is above this value
 	private static final int WAIT_CEILING_MS = 8;
 	
@@ -76,8 +81,7 @@ public class CpuDecoderRenderer implements DecoderRenderer {
 	}
 	
 	@Override
-	public void setup(int width, int height, SurfaceHolder renderTarget, int drFlags) {
-		this.renderTarget = renderTarget.getSurface();
+	public void setup(Context context, int width, int height, SurfaceHolder renderTarget, int drFlags) {
 		this.targetFps = 30;
 		
 		int perfLevel = findOptimalPerformanceLevel();
@@ -111,6 +115,12 @@ public class CpuDecoderRenderer implements DecoderRenderer {
 			break;
 		}
 		
+		// Create and initialize the RenderScript intrinsic we'll be using
+		rsRenderer = new RsRenderer(context, width, height, renderTarget.getSurface());
+		
+		// Allocate the frame buffer that the RGBA frame will be copied into
+		frameBuffer = new byte[width*height*4];
+		
 		// If the user wants quality, we'll remove the low IQ flags
 		if ((drFlags & DecoderRenderer.FLAG_PREFER_QUALITY) != 0) {
 			// Make sure the loop filter is enabled
@@ -127,7 +137,7 @@ public class CpuDecoderRenderer implements DecoderRenderer {
 			throw new IllegalStateException("AVC decoder initialization failure: "+err);
 		}
 		
-		decoderBuffer = ByteBuffer.allocate(92*1024);
+		decoderBuffer = ByteBuffer.allocate(DECODER_BUFFER_SIZE + AvcDecoder.getInputPaddingSize());
 		
 		System.out.println("Using software decoding (performance level: "+perfLevel+")");
 	}
@@ -152,7 +162,9 @@ public class CpuDecoderRenderer implements DecoderRenderer {
 					}
 
 					nextFrameTime = computePresentationTimeMs(targetFps);
-					AvcDecoder.redraw(renderTarget);
+					if (AvcDecoder.getRgbFrame(frameBuffer, frameBuffer.length)) {
+						rsRenderer.render(frameBuffer);
+					}
 				}
 			}
 		};
@@ -175,6 +187,10 @@ public class CpuDecoderRenderer implements DecoderRenderer {
 
 	@Override
 	public void release() {
+		if (rsRenderer != null) {
+			rsRenderer.release();
+		}
+		
 		AvcDecoder.destroy();
 	}
 
@@ -183,7 +199,7 @@ public class CpuDecoderRenderer implements DecoderRenderer {
 		byte[] data;
 		
 		// Use the reserved decoder buffer if this decode unit will fit
-		if (decodeUnit.getDataLength() <= decoderBuffer.limit()) {
+		if (decodeUnit.getDataLength() <= DECODER_BUFFER_SIZE) {
 			decoderBuffer.clear();
 			
 			for (AvByteBufferDescriptor bbd : decodeUnit.getBufferList()) {
@@ -193,7 +209,7 @@ public class CpuDecoderRenderer implements DecoderRenderer {
 			data = decoderBuffer.array();
 		}
 		else {
-			data = new byte[decodeUnit.getDataLength()];
+			data = new byte[decodeUnit.getDataLength()+AvcDecoder.getInputPaddingSize()];
 			
 			int offset = 0;
 			for (AvByteBufferDescriptor bbd : decodeUnit.getBufferList()) {

src/com/limelight/nvstream/av/video/cpu/RsRenderer.java

@@ -0,0 +1,36 @@
+package com.limelight.nvstream.av.video.cpu;
+
+import android.content.Context;
+import android.renderscript.Allocation;
+import android.renderscript.Element;
+import android.renderscript.RenderScript;
+import android.renderscript.Type;
+import android.view.Surface;
+
+public class RsRenderer {
+	private RenderScript rs;
+	private Allocation renderBuffer;
+	
+	public RsRenderer(Context context, int width, int height, Surface renderTarget) {
+		rs = RenderScript.create(context);
+
+		Type.Builder tb = new Type.Builder(rs, Element.RGBA_8888(rs));
+		tb.setX(width);
+		tb.setY(height);
+		Type bufferType = tb.create();
+
+		renderBuffer = Allocation.createTyped(rs, bufferType, Allocation.USAGE_SCRIPT | Allocation.USAGE_IO_OUTPUT);
+		renderBuffer.setSurface(renderTarget);
+	}
+	
+	public void release() {
+		renderBuffer.setSurface(null);
+		renderBuffer.destroy();
+		rs.destroy();
+	}
+	
+	public void render(byte[] rgbData) {
+		renderBuffer.copyFrom(rgbData);
+		renderBuffer.ioSend();
+	}
+}