3D Physics Rework, Other Stuff

-=-=-=-=-=-=-=-=-=-=-=-=-=-

3D Physics:
-Fixed "Bounce" parameter in 3D
-Fixed bug affecting Area (sometims it would not detect properly)
-Vehicle Body has seen heavy work
-Added Query API for doing space queries in 3D. Needs some docs though.
-Added JOINTS! Adapted Bullet Joints: and created easy gizmos for setting them up:
   -PinJoint
   -HingeJoint (with motor)
   -SliderJoint
   -ConeTwistJoint
   -Generic6DOFJoint
-Added OBJECT PICKING! based on the new query API. Any physics object now (Area or Body) has the following signals and virtual functions:
    -input_event (mouse or multitouch input over the body)
    -mouse_enter (mouse entered the body area)
    -mouse_exit (mouse exited body area)
   For Area it needs to be activated manually, as it isn't by default (ray goes thru).

Other:

-Begun working on Windows 8 (RT) port. Compiles but does not work yet.
-Added TheoraPlayer library for improved to-texture and portable video support.
-Fixed a few bugs in the renderer, collada importer, collada exporter, etc.

87 files changed, 51271 insertions, 0 deletions

diff --git a/drivers/theoraplayer/src/AVFoundation/TheoraVideoClip_AVFoundation.h b/drivers/theoraplayer/src/AVFoundation/TheoraVideoClip_AVFoundation.h
new file mode 100644
index 0000000000..abd898aa01
--- /dev/null
+++ b/drivers/theoraplayer/src/AVFoundation/TheoraVideoClip_AVFoundation.h
@@ -0,0 +1,47 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#if defined(__AVFOUNDATION) && !defined(_TheoraVideoClip_AVFoundation_h)
+#define _TheoraVideoClip_AVFoundation_h
+
+#include "TheoraAudioPacketQueue.h"
+#include "TheoraVideoClip.h"
+
+#ifndef AVFOUNDATION_CLASSES_DEFINED
+class AVAssetReader;
+class AVAssetReaderTrackOutput;
+#endif
+
+class TheoraVideoClip_AVFoundation : public TheoraVideoClip, public TheoraAudioPacketQueue
+{
+protected:
+	bool mLoaded;
+	int mFrameNumber;
+	AVAssetReader* mReader;
+	AVAssetReaderTrackOutput *mOutput, *mAudioOutput;
+	unsigned int mReadAudioSamples;
+	
+	void unload();
+	void doSeek();
+public:
+	TheoraVideoClip_AVFoundation(TheoraDataSource* data_source,
+								 TheoraOutputMode output_mode,
+								 int nPrecachedFrames,
+								 bool usePower2Stride);
+	~TheoraVideoClip_AVFoundation();
+	
+	bool _readData();
+	bool decodeNextFrame();
+	void _restart();
+	void load(TheoraDataSource* source);
+	float decodeAudio();
+	void decodedAudioCheck();
+	std::string getDecoderName() { return "AVFoundation"; }
+};
+
+#endif
diff --git a/drivers/theoraplayer/src/AVFoundation/TheoraVideoClip_AVFoundation.mm b/drivers/theoraplayer/src/AVFoundation/TheoraVideoClip_AVFoundation.mm
new file mode 100644
index 0000000000..8c3d2cc3b9
--- /dev/null
+++ b/drivers/theoraplayer/src/AVFoundation/TheoraVideoClip_AVFoundation.mm
@@ -0,0 +1,454 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#ifdef __AVFOUNDATION
+#define AVFOUNDATION_CLASSES_DEFINED
+#import <AVFoundation/AVFoundation.h>
+#include "TheoraAudioInterface.h"
+#include "TheoraDataSource.h"
+#include "TheoraException.h"
+#include "TheoraTimer.h"
+#include "TheoraUtil.h"
+#include "TheoraFrameQueue.h"
+#include "TheoraVideoFrame.h"
+#include "TheoraVideoManager.h"
+#include "TheoraVideoClip_AVFoundation.h"
+#include "TheoraPixelTransform.h"
+
+#ifdef _AVFOUNDATION_BGRX
+// a fast function developed to use kernel byte swapping calls to optimize alpha decoding.
+// In AVFoundation, BGRX mode conversion is prefered to YUV conversion because apple's YUV
+// conversion on iOS seems to run faster than libtheoraplayer's implementation
+// This may change in the future with more optimizations to libtheoraplayers's YUV conversion
+// code, making this function obsolete.
+static void bgrx2rgba(unsigned char* dest, int w, int h, struct TheoraPixelTransform* t)
+{
+	unsigned register int a;
+	unsigned int *dst = (unsigned int*) dest, *dstEnd;
+	unsigned char* src = t->raw;
+	int y, x, ax;
+	
+	for (y = 0; y < h; ++y, src += t->rawStride)
+	{
+		for (x = 0, ax = w * 4, dstEnd = dst + w; dst != dstEnd; x += 4, ax += 4, ++dst)
+		{
+            // use the full alpha range here because the Y channel has already been converted
+            // to RGB and that's in [0, 255] range.
+			a = src[ax];
+            *dst = (OSReadSwapInt32(src, x) >> 8) | (a << 24);
+		}
+	}
+}
+#endif
+
+static CVPlanarPixelBufferInfo_YCbCrPlanar getYUVStruct(void* src)
+{
+	CVPlanarPixelBufferInfo_YCbCrPlanar* bigEndianYuv = (CVPlanarPixelBufferInfo_YCbCrPlanar*) src;
+	CVPlanarPixelBufferInfo_YCbCrPlanar yuv;
+	yuv.componentInfoY.offset = OSSwapInt32(bigEndianYuv->componentInfoY.offset);
+	yuv.componentInfoY.rowBytes = OSSwapInt32(bigEndianYuv->componentInfoY.rowBytes);
+	yuv.componentInfoCb.offset = OSSwapInt32(bigEndianYuv->componentInfoCb.offset);
+	yuv.componentInfoCb.rowBytes = OSSwapInt32(bigEndianYuv->componentInfoCb.rowBytes);
+	yuv.componentInfoCr.offset = OSSwapInt32(bigEndianYuv->componentInfoCr.offset);
+	yuv.componentInfoCr.rowBytes = OSSwapInt32(bigEndianYuv->componentInfoCr.rowBytes);
+	return yuv;
+}
+
+TheoraVideoClip_AVFoundation::TheoraVideoClip_AVFoundation(TheoraDataSource* data_source,
+											   TheoraOutputMode output_mode,
+											   int nPrecachedFrames,
+											   bool usePower2Stride):
+	TheoraVideoClip(data_source, output_mode, nPrecachedFrames, usePower2Stride),
+	TheoraAudioPacketQueue()
+{
+	mLoaded = 0;
+	mReader = NULL;
+	mOutput = mAudioOutput = NULL;
+	mReadAudioSamples = mAudioFrequency = mNumAudioChannels = 0;
+}
+
+TheoraVideoClip_AVFoundation::~TheoraVideoClip_AVFoundation()
+{
+	unload();
+}
+
+void TheoraVideoClip_AVFoundation::unload()
+{
+	if (mOutput != NULL || mAudioOutput != NULL || mReader != NULL)
+	{
+		NSAutoreleasePool* pool = [[NSAutoreleasePool alloc] init];
+
+		if (mOutput != NULL)
+		{
+			[mOutput release];
+			mOutput = NULL;
+		}
+		
+		if (mAudioOutput)
+		{
+			[mAudioOutput release];
+			mAudioOutput = NULL;
+		}
+		
+		if (mReader != NULL)
+		{
+			[mReader release];
+			mReader = NULL;
+		}
+		
+		[pool release];
+	}
+}
+
+bool TheoraVideoClip_AVFoundation::_readData()
+{
+	return 1;
+}
+
+bool TheoraVideoClip_AVFoundation::decodeNextFrame()
+{
+	if (mReader == NULL || mEndOfFile) return 0;
+	AVAssetReaderStatus status = [mReader status];
+	if (status == AVAssetReaderStatusFailed)
+	{
+		// This can happen on iOS when you suspend the app... Only happens on the device, iOS simulator seems to work fine.
+		th_writelog("AVAssetReader reading failed, restarting...");
+
+		mSeekFrame = mTimer->getTime() * mFPS;
+		// just in case
+		if (mSeekFrame < 0) mSeekFrame = 0;
+		if (mSeekFrame > mDuration * mFPS - 1) mSeekFrame = mDuration * mFPS - 1;
+		_restart();
+		status = [mReader status];
+		if (status == AVAssetReaderStatusFailed)
+		{
+			th_writelog("AVAssetReader restart failed!");
+			return 0;
+		}
+		th_writelog("AVAssetReader restart succeeded!");
+	}
+
+	TheoraVideoFrame* frame = mFrameQueue->requestEmptyFrame();
+	if (!frame) return 0;
+
+	CMSampleBufferRef sampleBuffer = NULL;
+	NSAutoreleasePool* pool = NULL;
+	CMTime presentationTime;
+	
+	if (mAudioInterface) decodeAudio();
+	
+	if (status == AVAssetReaderStatusReading)
+	{
+		pool = [[NSAutoreleasePool alloc] init];
+		
+		while ((sampleBuffer = [mOutput copyNextSampleBuffer]))
+		{
+			presentationTime = CMSampleBufferGetOutputPresentationTimeStamp(sampleBuffer);
+			frame->mTimeToDisplay = (float) CMTimeGetSeconds(presentationTime);
+			frame->mIteration = mIteration;
+			frame->_setFrameNumber(mFrameNumber);
+			++mFrameNumber;
+			if (frame->mTimeToDisplay < mTimer->getTime() && !mRestarted && mFrameNumber % 16 != 0)
+			{
+				// %16 operation is here to prevent a playback halt during video playback if the decoder can't keep up with demand.
+#ifdef _DEBUG
+				th_writelog(mName + ": pre-dropped frame " + str(mFrameNumber - 1));
+#endif
+				++mNumDisplayedFrames;
+				++mNumDroppedFrames;
+				CMSampleBufferInvalidate(sampleBuffer);
+				CFRelease(sampleBuffer);
+				sampleBuffer = NULL;
+				continue; // drop frame
+			}
+
+			CVImageBufferRef imageBuffer = CMSampleBufferGetImageBuffer(sampleBuffer);
+			CVPixelBufferLockBaseAddress(imageBuffer, 0);
+			void *baseAddress = CVPixelBufferGetBaseAddress(imageBuffer);
+			
+			mStride = CVPixelBufferGetBytesPerRow(imageBuffer);
+			size_t width = CVPixelBufferGetWidth(imageBuffer);
+			size_t height = CVPixelBufferGetHeight(imageBuffer);
+
+			TheoraPixelTransform t;
+			memset(&t, 0, sizeof(TheoraPixelTransform));
+#ifdef _AVFOUNDATION_BGRX
+			if (mOutputMode == TH_BGRX || mOutputMode == TH_RGBA)
+			{
+				t.raw = (unsigned char*) baseAddress;
+				t.rawStride = mStride;
+			}
+			else
+#endif
+			{
+				CVPlanarPixelBufferInfo_YCbCrPlanar yuv = getYUVStruct(baseAddress);
+				
+				t.y = (unsigned char*) baseAddress + yuv.componentInfoY.offset;  t.yStride = yuv.componentInfoY.rowBytes;
+				t.u = (unsigned char*) baseAddress + yuv.componentInfoCb.offset; t.uStride = yuv.componentInfoCb.rowBytes;
+				t.v = (unsigned char*) baseAddress + yuv.componentInfoCr.offset; t.vStride = yuv.componentInfoCr.rowBytes;
+			}
+#ifdef _AVFOUNDATION_BGRX
+			if (mOutputMode == TH_RGBA)
+			{
+				for (int i = 0; i < 1000; ++i)
+					bgrx2rgba(frame->getBuffer(), mWidth / 2, mHeight, &t);
+				frame->mReady = true;
+			}
+			else
+#endif
+			frame->decode(&t);
+
+			CVPixelBufferUnlockBaseAddress(imageBuffer, 0);
+			CMSampleBufferInvalidate(sampleBuffer);
+			CFRelease(sampleBuffer);
+
+			break; // TODO - should this really be a while loop instead of an if block?
+		}
+	}
+	if (pool) [pool release];
+
+	if (!frame->mReady) // in case the frame wasn't used
+	{
+		frame->mInUse = 0;
+	}
+
+	if (sampleBuffer == NULL && mReader.status == AVAssetReaderStatusCompleted) // other cases could be app suspended
+	{
+		if (mAutoRestart)
+        {
+            ++mIteration;
+			_restart();
+        }
+		else
+		{
+			unload();
+			mEndOfFile = true;
+		}
+		return 0;
+	}
+	
+	
+	return 1;
+}
+
+void TheoraVideoClip_AVFoundation::_restart()
+{
+	mEndOfFile = false;
+	unload();
+	load(mStream);
+	mRestarted = true;
+}
+
+void TheoraVideoClip_AVFoundation::load(TheoraDataSource* source)
+{
+	mStream = source;
+	mFrameNumber = 0;
+	mEndOfFile = false;
+	TheoraFileDataSource* fileDataSource = dynamic_cast<TheoraFileDataSource*>(source);
+	std::string filename;
+	if (fileDataSource != NULL) filename = fileDataSource->getFilename();
+	else
+	{
+		TheoraMemoryFileDataSource* memoryDataSource = dynamic_cast<TheoraMemoryFileDataSource*>(source);
+		if (memoryDataSource != NULL) filename = memoryDataSource->getFilename();
+		else throw TheoraGenericException("Unable to load MP4 file");
+	}
+	
+	NSAutoreleasePool* pool = [[NSAutoreleasePool alloc] init];
+	NSString* path = [NSString stringWithUTF8String:filename.c_str()];
+	NSError* err;
+	NSURL *url = [NSURL fileURLWithPath:path];
+	AVAsset* asset = [[AVURLAsset alloc] initWithURL:url options:nil];
+	mReader = [[AVAssetReader alloc] initWithAsset:asset error:&err];
+	NSArray* tracks = [asset tracksWithMediaType:AVMediaTypeVideo];
+	if ([tracks count] == 0)
+		throw TheoraGenericException("Unable to open video file: " + filename);
+	AVAssetTrack *videoTrack = [tracks objectAtIndex:0];
+
+	NSArray* audioTracks = [asset tracksWithMediaType:AVMediaTypeAudio];
+	AVAssetTrack *audioTrack = audioTracks.count > 0 ? [audioTracks objectAtIndex:0] : NULL;
+	
+#ifdef _AVFOUNDATION_BGRX
+	bool yuv_output = (mOutputMode != TH_BGRX && mOutputMode != TH_RGBA);
+#else
+	bool yuv_output = true;
+#endif
+	
+	NSDictionary *videoOptions = [NSDictionary dictionaryWithObjectsAndKeys:[NSNumber numberWithInt:(yuv_output) ? kCVPixelFormatType_420YpCbCr8Planar : kCVPixelFormatType_32BGRA], kCVPixelBufferPixelFormatTypeKey, nil];
+
+	mOutput = [[AVAssetReaderTrackOutput alloc] initWithTrack:videoTrack outputSettings:videoOptions];
+	[mReader addOutput:mOutput];
+	if ([mOutput respondsToSelector:@selector(setAlwaysCopiesSampleData:)]) // Not supported on iOS versions older than 5.0
+		mOutput.alwaysCopiesSampleData = NO;
+
+	mFPS = videoTrack.nominalFrameRate;
+	mWidth = mSubFrameWidth = mStride = videoTrack.naturalSize.width;
+	mHeight = mSubFrameHeight = videoTrack.naturalSize.height;
+	mFrameDuration = 1.0f / mFPS;
+	mDuration = (float) CMTimeGetSeconds(asset.duration);
+	if (mFrameQueue == NULL)
+	{
+		mFrameQueue = new TheoraFrameQueue(this);
+		mFrameQueue->setSize(mNumPrecachedFrames);
+	}
+
+	if (mSeekFrame != -1)
+	{
+		mFrameNumber = mSeekFrame;
+		[mReader setTimeRange: CMTimeRangeMake(CMTimeMakeWithSeconds(mSeekFrame / mFPS, 1), kCMTimePositiveInfinity)];
+	}
+	if (audioTrack)
+	{
+		TheoraAudioInterfaceFactory* audio_factory = TheoraVideoManager::getSingleton().getAudioInterfaceFactory();
+		if (audio_factory)
+		{
+			NSDictionary *audioOptions = [NSDictionary dictionaryWithObjectsAndKeys:
+										  [NSNumber numberWithInt:kAudioFormatLinearPCM], AVFormatIDKey,
+										  [NSNumber numberWithBool:NO], AVLinearPCMIsNonInterleaved,
+										  [NSNumber numberWithBool:NO], AVLinearPCMIsBigEndianKey,
+										  [NSNumber numberWithBool:YES], AVLinearPCMIsFloatKey,
+										  [NSNumber numberWithInt:32], AVLinearPCMBitDepthKey,
+										  nil];
+
+			mAudioOutput = [[AVAssetReaderTrackOutput alloc] initWithTrack:audioTrack outputSettings:audioOptions];
+			[mReader addOutput:mAudioOutput];
+			if ([mAudioOutput respondsToSelector:@selector(setAlwaysCopiesSampleData:)]) // Not supported on iOS versions older than 5.0
+				mAudioOutput.alwaysCopiesSampleData = NO;
+			
+			NSArray* desclst = audioTrack.formatDescriptions;
+			CMAudioFormatDescriptionRef desc = (CMAudioFormatDescriptionRef) [desclst objectAtIndex:0];
+			const AudioStreamBasicDescription* audioDesc = CMAudioFormatDescriptionGetStreamBasicDescription(desc);
+			mAudioFrequency = (unsigned int) audioDesc->mSampleRate;
+			mNumAudioChannels = audioDesc->mChannelsPerFrame;
+			
+			if (mSeekFrame != -1)
+			{
+				mReadAudioSamples = mFrameNumber * (mAudioFrequency * mNumAudioChannels) / mFPS;
+			}
+			else mReadAudioSamples = 0;
+
+			if (mAudioInterface == NULL)
+				setAudioInterface(audio_factory->createInstance(this, mNumAudioChannels, mAudioFrequency));
+		}
+	}
+	
+#ifdef _DEBUG
+	else if (!mLoaded)
+	{
+		th_writelog("-----\nwidth: " + str(mWidth) + ", height: " + str(mHeight) + ", fps: " + str((int) getFPS()));
+		th_writelog("duration: " + strf(mDuration) + " seconds\n-----");
+	}
+#endif
+	[mReader startReading];
+	[pool release];
+	mLoaded = true;
+}
+ 
+void TheoraVideoClip_AVFoundation::decodedAudioCheck()
+{
+	if (!mAudioInterface || mTimer->isPaused()) return;
+	
+	mAudioMutex->lock();
+	flushAudioPackets(mAudioInterface);
+	mAudioMutex->unlock();
+}
+
+float TheoraVideoClip_AVFoundation::decodeAudio()
+{
+	if (mRestarted) return -1;
+
+	if (mReader == NULL || mEndOfFile) return 0;
+	AVAssetReaderStatus status = [mReader status];
+
+	if (mAudioOutput)
+	{
+		CMSampleBufferRef sampleBuffer = NULL;
+		NSAutoreleasePool* pool = NULL;
+		bool mutexLocked = 0;
+
+		float factor = 1.0f / (mAudioFrequency * mNumAudioChannels);
+		float videoTime = (float) mFrameNumber / mFPS;
+		float min = mFrameQueue->getSize() / mFPS + 1.0f;
+		
+		if (status == AVAssetReaderStatusReading)
+		{
+			pool = [[NSAutoreleasePool alloc] init];
+
+			// always buffer up of audio ahead of the frames
+			while (mReadAudioSamples * factor - videoTime < min)
+			{
+				if ((sampleBuffer = [mAudioOutput copyNextSampleBuffer]))
+				{
+					AudioBufferList audioBufferList;
+					
+					CMBlockBufferRef blockBuffer = NULL;
+					CMSampleBufferGetAudioBufferListWithRetainedBlockBuffer(sampleBuffer, NULL, &audioBufferList, sizeof(audioBufferList), NULL, NULL, 0, &blockBuffer);
+					
+					for (int y = 0; y < audioBufferList.mNumberBuffers; ++y)
+					{
+						AudioBuffer audioBuffer = audioBufferList.mBuffers[y];
+						float *frame = (float*) audioBuffer.mData;
+
+						if (!mutexLocked)
+						{
+							mAudioMutex->lock();
+							mutexLocked = 1;
+						}
+						addAudioPacket(frame, audioBuffer.mDataByteSize / (mNumAudioChannels * sizeof(float)), mAudioGain);
+						
+						mReadAudioSamples += audioBuffer.mDataByteSize / (sizeof(float));
+					}
+
+					CFRelease(blockBuffer);
+					CMSampleBufferInvalidate(sampleBuffer);
+					CFRelease(sampleBuffer);
+				}
+				else
+				{
+					[mAudioOutput release];
+					mAudioOutput = nil;
+					break;
+				}
+			}
+			[pool release];
+		}
+		if (mutexLocked) mAudioMutex->unlock();
+	}
+	
+	return -1;
+}
+
+void TheoraVideoClip_AVFoundation::doSeek()
+{
+#if _DEBUG
+	th_writelog(mName + " [seek]: seeking to frame " + str(mSeekFrame));
+#endif
+	int frame;
+	float time = mSeekFrame / getFPS();
+	mTimer->seek(time);
+	bool paused = mTimer->isPaused();
+	if (!paused) mTimer->pause(); // pause until seeking is done
+	
+	mEndOfFile = false;
+	mRestarted = false;
+	
+    resetFrameQueue();
+	unload();
+	load(mStream);
+
+	if (mAudioInterface)
+	{
+		mAudioMutex->lock();
+		destroyAllAudioPackets();
+		mAudioMutex->unlock();
+	}
+
+	if (!paused) mTimer->play();
+	mSeekFrame = -1;
+}
+#endif
diff --git a/drivers/theoraplayer/src/FFmpeg/TheoraVideoClip_FFmpeg.cpp b/drivers/theoraplayer/src/FFmpeg/TheoraVideoClip_FFmpeg.cpp
new file mode 100644
index 0000000000..fa3fd43a47
--- /dev/null
+++ b/drivers/theoraplayer/src/FFmpeg/TheoraVideoClip_FFmpeg.cpp
@@ -0,0 +1,439 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#ifdef __FFMPEG
+#include "TheoraAudioInterface.h"
+#include "TheoraDataSource.h"
+#include "TheoraException.h"
+#include "TheoraTimer.h"
+#include "TheoraUtil.h"
+#include "TheoraFrameQueue.h"
+#include "TheoraVideoFrame.h"
+#include "TheoraVideoManager.h"
+#include "TheoraVideoClip_FFmpeg.h"
+#include "TheoraPixelTransform.h"
+
+#define READ_BUFFER_SIZE 4096
+
+#ifdef __cplusplus
+#define __STDC_CONSTANT_MACROS
+#ifdef _STDINT_H
+#undef _STDINT_H
+#endif
+# include <stdint.h>
+#endif
+
+#define _FFMPEG_DEBUG
+
+extern "C"
+{
+#include <libavcodec/avcodec.h>
+#include <libavformat/avformat.h>
+#include "libavutil/avassert.h"
+}
+
+static bool ffmpegInitialised = 0;
+
+static int readFunction(void* data, uint8_t* buf, int buf_size)
+{
+#ifdef _FFMPEG_DEBUG
+	th_writelog("reading " + str(buf_size) + " bytes");
+#endif
+
+	TheoraDataSource* src = (TheoraDataSource*) data;
+	return src->read(buf, buf_size);
+}
+
+static int64_t seekFunction(void* data, int64_t offset, int whence)
+{
+#ifdef _FFMPEG_DEBUG
+	th_writelog("seeking: offset = " + str((long) offset) + ", whence = " + str(whence));
+#endif
+
+	TheoraDataSource* src = (TheoraDataSource*) data;
+    if (whence == AVSEEK_SIZE)
+		return src->size();
+	else if (whence == SEEK_SET)
+		src->seek((long) offset);
+	else if (whence == SEEK_END)
+		src->seek(src->size() - (long) offset);
+    return src->tell();
+}
+
+static void avlog_theoraplayer(void* p, int level, const char* fmt, va_list vargs)
+{
+	th_writelog(fmt);
+	static char logstr[2048];
+	vsprintf(logstr, fmt, vargs);
+	th_writelog("ffmpeg: " + std::string(logstr));
+}
+
+
+std::string text;
+
+static void _log(const char* s)
+{
+	text += s;
+//	th_writelog(text);
+//	text = "";
+}
+
+static void _log(const char c)
+{
+	char s[2] = {c, 0};
+	_log(s);
+}
+
+static const AVCodec *next_codec_for_id(enum AVCodecID id, const AVCodec *prev,
+                                        int encoder)
+{
+    while ((prev = av_codec_next(prev))) {
+        if (prev->id == id &&
+            (encoder ? av_codec_is_encoder(prev) : av_codec_is_decoder(prev)))
+            return prev;
+    }
+    return NULL;
+}
+
+static int compare_codec_desc(const void *a, const void *b)
+{
+    const AVCodecDescriptor **da = (const AVCodecDescriptor **) a;
+    const AVCodecDescriptor **db = (const AVCodecDescriptor **) b;
+	
+    return (*da)->type != (*db)->type ? (*da)->type - (*db)->type :
+	strcmp((*da)->name, (*db)->name);
+}
+
+static unsigned get_codecs_sorted(const AVCodecDescriptor ***rcodecs)
+{
+    const AVCodecDescriptor *desc = NULL;
+    const AVCodecDescriptor **codecs;
+    unsigned nb_codecs = 0, i = 0;
+	
+    while ((desc = avcodec_descriptor_next(desc)))
+        ++nb_codecs;
+    if (!(codecs = (const AVCodecDescriptor**) av_calloc(nb_codecs, sizeof(*codecs)))) {
+        av_log(NULL, AV_LOG_ERROR, "Out of memory\n");
+        exit(1);
+    }
+    desc = NULL;
+    while ((desc = avcodec_descriptor_next(desc)))
+        codecs[i++] = desc;
+    av_assert0(i == nb_codecs);
+    qsort(codecs, nb_codecs, sizeof(*codecs), compare_codec_desc);
+    *rcodecs = codecs;
+    return nb_codecs;
+}
+
+static char get_media_type_char(enum AVMediaType type)
+{
+    switch (type) {
+        case AVMEDIA_TYPE_VIDEO:    return 'V';
+        case AVMEDIA_TYPE_AUDIO:    return 'A';
+        case AVMEDIA_TYPE_DATA:     return 'D';
+        case AVMEDIA_TYPE_SUBTITLE: return 'S';
+        case AVMEDIA_TYPE_ATTACHMENT:return 'T';
+        default:                    return '?';
+    }
+}
+
+static void print_codecs_for_id(enum AVCodecID id, int encoder)
+{
+    const AVCodec *codec = NULL;
+	
+    _log(encoder ? "encoders" : "decoders");
+	
+    while ((codec = next_codec_for_id(id, codec, encoder)))
+        _log(codec->name);
+	
+    _log(")");
+}
+
+int show_codecs(void *optctx, const char *opt, const char *arg)
+{
+    const AVCodecDescriptor **codecs;
+    unsigned i, nb_codecs = get_codecs_sorted(&codecs);
+	
+	char tmp[1024];
+    th_writelog("Codecs:\n"
+           " D..... = Decoding supported\n"
+           " .E.... = Encoding supported\n"
+           " ..V... = Video codec\n"
+           " ..A... = Audio codec\n"
+           " ..S... = Subtitle codec\n"
+           " ...I.. = Intra frame-only codec\n"
+           " ....L. = Lossy compression\n"
+           " .....S = Lossless compression\n"
+           " -------\n");
+    for (i = 0; i < nb_codecs; ++i) {
+        const AVCodecDescriptor *desc = codecs[i];
+        const AVCodec *codec = NULL;
+		
+        _log(" ");
+        _log(avcodec_find_decoder(desc->id) ? "D" : ".");
+        _log(avcodec_find_encoder(desc->id) ? "E" : ".");
+		
+        _log(get_media_type_char(desc->type));
+        _log((desc->props & AV_CODEC_PROP_INTRA_ONLY) ? "I" : ".");
+        _log((desc->props & AV_CODEC_PROP_LOSSY)      ? "L" : ".");
+        _log((desc->props & AV_CODEC_PROP_LOSSLESS)   ? "S" : ".");
+		
+		
+        sprintf(tmp, " %-20s %s", desc->name, desc->long_name ? desc->long_name : "");
+		
+		_log(tmp);
+        /* print decoders/encoders when there's more than one or their
+         * names are different from codec name */
+        while ((codec = next_codec_for_id(desc->id, codec, 0))) {
+            if (strcmp(codec->name, desc->name)) {
+                print_codecs_for_id(desc->id, 0);
+                break;
+            }
+        }
+        codec = NULL;
+        while ((codec = next_codec_for_id(desc->id, codec, 1))) {
+            if (strcmp(codec->name, desc->name)) {
+                print_codecs_for_id(desc->id, 1);
+                break;
+            }
+        }
+		_log("\n");
+    }
+    av_free(codecs);
+	
+	av_log(0, 0, "%s", text.c_str());
+    return 0;
+}
+
+TheoraVideoClip_FFmpeg::TheoraVideoClip_FFmpeg(TheoraDataSource* data_source,
+														 TheoraOutputMode output_mode,
+														 int nPrecachedFrames,
+														 bool usePower2Stride):
+								 						 TheoraVideoClip(data_source, output_mode, nPrecachedFrames, usePower2Stride),
+														 TheoraAudioPacketQueue()
+{
+	mFormatContext = NULL;
+	mCodecContext = NULL;
+	mCodec = NULL;
+	mFrame = NULL;
+	mVideoStreamIndex = -1;
+}
+
+TheoraVideoClip_FFmpeg::~TheoraVideoClip_FFmpeg()
+{
+	unload();
+}
+
+void TheoraVideoClip_FFmpeg::load(TheoraDataSource* source)
+{
+	mVideoStreamIndex = -1;
+	mFrameNumber = 0;
+	AVDictionary* optionsDict = NULL;
+	
+	if (!ffmpegInitialised)
+	{
+#ifdef _FFMPEG_DEBUG
+		th_writelog("Initializing ffmpeg");
+#endif
+		th_writelog("avcodec version: " + str(avcodec_version()));
+		av_register_all();
+		av_log_set_level(AV_LOG_DEBUG);
+		av_log_set_callback(avlog_theoraplayer);
+		ffmpegInitialised = 1;
+		//show_codecs(0, 0, 0);
+	}
+	
+	mInputBuffer = (unsigned char*) av_malloc(READ_BUFFER_SIZE);
+	mAvioContext = avio_alloc_context(mInputBuffer, READ_BUFFER_SIZE, 0, source, &readFunction, NULL, &seekFunction);
+	
+#ifdef _FFMPEG_DEBUG
+	th_writelog(mName + ": avio context created");
+#endif
+
+	mFormatContext = avformat_alloc_context();
+#ifdef _FFMPEG_DEBUG
+	th_writelog(mName + ": avformat context created");
+#endif
+	mFormatContext->pb = mAvioContext;
+	
+	int err;
+	if ((err = avformat_open_input(&mFormatContext, "", NULL, NULL)) != 0)
+	{
+		th_writelog(mName + ": avformat input opening failed!");
+		th_writelog(mName + ": error_code: " + str(err));
+		return;
+	}
+	
+#ifdef _FFMPEG_DEBUG
+	th_writelog(mName + ": avformat input opened");
+#endif
+	
+	// Retrieve stream information
+	if (avformat_find_stream_info(mFormatContext, NULL) < 0)
+		return; // Couldn't find stream information
+	
+#ifdef _FFMPEG_DEBUG
+	th_writelog(mName + ": got stream info");
+#endif
+	
+	// Dump information about file onto standard error
+	//	av_dump_format(mFormatContext, 0, "", 0);
+	
+	// Find the first video stream
+	for (int i = 0; i < mFormatContext->nb_streams; ++i)
+	{
+		if(mFormatContext->streams[i]->codec->codec_type == AVMEDIA_TYPE_VIDEO)
+		{
+			mVideoStreamIndex = i;
+			break;
+		}
+	}
+	if (mVideoStreamIndex == -1)
+		return; // Didn't find a video stream
+
+#ifdef _FFMPEG_DEBUG
+	th_writelog(mName + ": Found video stream at index " + str(mVideoStreamIndex));
+#endif
+
+	// Get a pointer to the codec context for the video stream
+	mCodecContext = mFormatContext->streams[mVideoStreamIndex]->codec;
+	
+	// Find the decoder for the video stream
+	mCodec = avcodec_find_decoder(mCodecContext->codec_id);
+	if (mCodec == NULL)
+	{
+		th_writelog("Unsupported codec!");
+		return; // Codec not found
+	}
+	// Open codec
+	if(avcodec_open2(mCodecContext, mCodec, &optionsDict) < 0)
+		return; // Could not open codec
+	
+#ifdef _FFMPEG_DEBUG
+	th_writelog(mName + ": Codec opened");
+#endif
+
+	
+	mFrame = avcodec_alloc_frame();
+	
+#ifdef _FFMPEG_DEBUG
+	th_writelog(mName + ": Frame allocated");
+#endif
+		
+	//AVRational rational = mCodecContext->time_base;
+
+	mFPS = 25; //TODOOOOOO!!!
+	
+	mWidth = mStride = mCodecContext->width;
+	mHeight = mCodecContext->height;
+	mFrameDuration = 1.0f / mFPS;
+	mDuration = mFormatContext->duration / AV_TIME_BASE;
+	
+	if (mFrameQueue == NULL) // todo - why is this set in the backend class? it should be set in the base class, check other backends as well
+	{
+		mFrameQueue = new TheoraFrameQueue(this);
+		mFrameQueue->setSize(mNumPrecachedFrames);
+	}
+}
+
+void TheoraVideoClip_FFmpeg::unload()
+{
+	if (mInputBuffer)
+	{
+//		av_free(mInputBuffer);
+		mInputBuffer = NULL;
+	}
+	if (mAvioContext)
+	{
+		av_free(mAvioContext);
+		mAvioContext = NULL;
+	}
+	if (mFrame)
+	{
+		av_free(mFrame);
+		mFrame = NULL;
+	}
+	if (mCodecContext)
+	{
+		avcodec_close(mCodecContext);
+		mCodecContext = NULL;
+	}
+	if (mFormatContext)
+	{
+		avformat_close_input(&mFormatContext);
+		mFormatContext = NULL;
+	}
+}
+
+bool TheoraVideoClip_FFmpeg::_readData()
+{
+	return 1;
+}
+
+bool TheoraVideoClip_FFmpeg::decodeNextFrame()
+{
+	TheoraVideoFrame* frame = mFrameQueue->requestEmptyFrame();
+	if (!frame) return 0;
+
+	AVPacket packet;
+	int frameFinished;
+	
+	while (av_read_frame(mFormatContext, &packet) >= 0)
+	{
+		if (packet.stream_index == mVideoStreamIndex)
+		{
+			avcodec_decode_video2(mCodecContext, mFrame, &frameFinished, &packet);
+			
+			if (frameFinished)
+			{
+				TheoraPixelTransform t;
+				memset(&t, 0, sizeof(TheoraPixelTransform));
+
+				t.y = mFrame->data[0]; t.yStride = mFrame->linesize[0];
+				t.u = mFrame->data[1]; t.uStride = mFrame->linesize[1];
+				t.v = mFrame->data[2]; t.vStride = mFrame->linesize[2];
+				
+				frame->decode(&t);
+				frame->mTimeToDisplay = mFrameNumber / mFPS;
+				frame->mIteration = mIteration;
+				frame->_setFrameNumber(mFrameNumber++);
+
+				av_free_packet(&packet);
+				break;
+			}
+		}
+		av_free_packet(&packet);
+	}
+	return 1;
+}
+ 
+void TheoraVideoClip_FFmpeg::decodedAudioCheck()
+{
+	if (!mAudioInterface || mTimer->isPaused()) return;
+	
+	mAudioMutex->lock();
+	flushAudioPackets(mAudioInterface);
+	mAudioMutex->unlock();
+}
+
+float TheoraVideoClip_FFmpeg::decodeAudio()
+{
+	return -1;
+}
+
+void TheoraVideoClip_FFmpeg::doSeek()
+{
+
+}
+
+void TheoraVideoClip_FFmpeg::_restart()
+{
+
+}
+
+#endif
diff --git a/drivers/theoraplayer/src/FFmpeg/TheoraVideoClip_FFmpeg.h b/drivers/theoraplayer/src/FFmpeg/TheoraVideoClip_FFmpeg.h
new file mode 100644
index 0000000000..03f9a3d964
--- /dev/null
+++ b/drivers/theoraplayer/src/FFmpeg/TheoraVideoClip_FFmpeg.h
@@ -0,0 +1,53 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#if defined(__FFMPEG) && !defined(_TheoraVideoClip_FFmpeg_h)
+#define _TheoraVideoClip_FFmpeg_h
+
+#include "TheoraAudioPacketQueue.h"
+#include "TheoraVideoClip.h"
+
+struct AVFormatContext;
+struct AVCodecContext;
+struct AVCodec;
+struct AVFrame;
+struct AVIOContext;
+
+class TheoraVideoClip_FFmpeg : public TheoraVideoClip, public TheoraAudioPacketQueue
+{
+protected:
+	bool mLoaded;
+	
+	AVFormatContext* mFormatContext;
+	AVCodecContext* mCodecContext;
+	AVIOContext* mAvioContext;
+	AVCodec* mCodec;
+	AVFrame* mFrame;
+	unsigned char* mInputBuffer;
+	int mVideoStreamIndex;
+	int mFrameNumber;
+	
+	void unload();
+	void doSeek();
+public:
+	TheoraVideoClip_FFmpeg(TheoraDataSource* data_source,
+								 TheoraOutputMode output_mode,
+								 int nPrecachedFrames,
+								 bool usePower2Stride);
+	~TheoraVideoClip_FFmpeg();
+	
+	bool _readData();
+	bool decodeNextFrame();
+	void _restart();
+	void load(TheoraDataSource* source);
+	float decodeAudio();
+	void decodedAudioCheck();
+	std::string getDecoderName() { return "FFmpeg"; }
+};
+
+#endif
diff --git a/drivers/theoraplayer/src/Theora/TheoraVideoClip_Theora.cpp b/drivers/theoraplayer/src/Theora/TheoraVideoClip_Theora.cpp
new file mode 100644
index 0000000000..c4f070ec50
--- /dev/null
+++ b/drivers/theoraplayer/src/Theora/TheoraVideoClip_Theora.cpp
@@ -0,0 +1,703 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#ifdef __THEORA
+#include <memory.h>
+#include <algorithm>
+#include "TheoraVideoManager.h"
+#include "TheoraFrameQueue.h"
+#include "TheoraVideoFrame.h"
+#include "TheoraAudioInterface.h"
+#include "TheoraTimer.h"
+#include "TheoraDataSource.h"
+#include "TheoraUtil.h"
+#include "TheoraException.h"
+#include "TheoraVideoClip_Theora.h"
+#include "TheoraPixelTransform.h"
+
+TheoraVideoClip_Theora::TheoraVideoClip_Theora(TheoraDataSource* data_source,
+										TheoraOutputMode output_mode,
+										int nPrecachedFrames,
+										bool usePower2Stride):
+	TheoraVideoClip(data_source, output_mode, nPrecachedFrames, usePower2Stride),
+	TheoraAudioPacketQueue()
+{
+	mInfo.TheoraDecoder = NULL;
+	mInfo.TheoraSetup = NULL;
+	mVorbisStreams = mTheoraStreams = 0;
+	mReadAudioSamples = 0;
+	mLastDecodedFrameNumber = 0;
+}
+
+TheoraVideoClip_Theora::~TheoraVideoClip_Theora()
+{
+	if (mInfo.TheoraDecoder)
+	{
+		th_decode_free(mInfo.TheoraDecoder);
+		th_setup_free(mInfo.TheoraSetup);
+
+		if (mAudioInterface)
+		{
+			vorbis_dsp_clear(&mInfo.VorbisDSPState);
+			vorbis_block_clear(&mInfo.VorbisBlock);
+		}
+
+		ogg_stream_clear(&mInfo.TheoraStreamState);
+		th_comment_clear(&mInfo.TheoraComment);
+		th_info_clear(&mInfo.TheoraInfo);
+		
+		ogg_stream_clear(&mInfo.VorbisStreamState);
+		vorbis_comment_clear(&mInfo.VorbisComment);
+		vorbis_info_clear(&mInfo.VorbisInfo);
+		
+		ogg_sync_clear(&mInfo.OggSyncState);
+	}
+}
+
+bool TheoraVideoClip_Theora::_readData()
+{
+	int audio_eos = 0, serno;
+	float audio_time = 0;
+	float time = mTimer->getTime();
+	if (mRestarted) time = 0;
+	
+	for (;;)
+	{
+		char *buffer = ogg_sync_buffer(&mInfo.OggSyncState, 4096);
+		int bytes_read = mStream->read(buffer, 4096);
+		ogg_sync_wrote(&mInfo.OggSyncState, bytes_read);
+		
+		if (bytes_read < 4096)
+		{
+			if (bytes_read == 0)
+			{
+				if (!mAutoRestart) mEndOfFile = true;
+				return 0;
+			}
+		}
+		// when we fill the stream with enough pages, it'll start spitting out packets
+		// which contain keyframes, delta frames or audio data
+		while (ogg_sync_pageout(&mInfo.OggSyncState, &mInfo.OggPage) > 0)
+		{
+			serno = ogg_page_serialno(&mInfo.OggPage);
+			if (serno == mInfo.TheoraStreamState.serialno) ogg_stream_pagein(&mInfo.TheoraStreamState, &mInfo.OggPage);
+			if (mAudioInterface && serno == mInfo.VorbisStreamState.serialno)
+			{
+				ogg_int64_t g = ogg_page_granulepos(&mInfo.OggPage);
+				audio_time = (float) vorbis_granule_time(&mInfo.VorbisDSPState, g);
+				audio_eos = ogg_page_eos(&mInfo.OggPage);
+				ogg_stream_pagein(&mInfo.VorbisStreamState, &mInfo.OggPage);
+			}
+		}
+		if (!(mAudioInterface && !audio_eos && audio_time < time + 1.0f))
+			break;
+	}
+	return 1;
+}
+
+bool TheoraVideoClip_Theora::decodeNextFrame()
+{
+	if (mEndOfFile) return 0;
+	
+	TheoraVideoFrame* frame = mFrameQueue->requestEmptyFrame();
+	if (!frame) return 0; // max number of precached frames reached
+	bool should_restart = 0;
+	ogg_packet opTheora;
+	ogg_int64_t granulePos;
+	th_ycbcr_buffer buff;
+    int ret, nAttempts;
+	for (;;)
+	{
+        // ogg_stream_packetout can return -1 and the official docs suggest to do subsequent calls until it succeeds
+        // because the data is out of sync. still will limit the number of attempts just in case
+        for (ret = -1, nAttempts = 0; ret < 0 && nAttempts < 100; nAttempts++)
+        {
+            ret = ogg_stream_packetout(&mInfo.TheoraStreamState, &opTheora);
+        }
+		
+		if (ret > 0)
+		{
+			int status = th_decode_packetin(mInfo.TheoraDecoder, &opTheora, &granulePos);
+            if (status != 0 && status != TH_DUPFRAME) continue; // 0 means success
+
+			float time = (float) th_granule_time(mInfo.TheoraDecoder, granulePos);
+			unsigned long frame_number = (unsigned long) th_granule_frame(mInfo.TheoraDecoder, granulePos);
+			
+			if (time < mTimer->getTime() && !mRestarted && frame_number % 16 != 0)
+			{
+				// %16 operation is here to prevent a playback halt during video playback if the decoder can't keep up with demand.
+#ifdef _DEBUG
+				th_writelog(mName + ": pre-dropped frame " + str((int) frame_number));
+#endif
+				++mNumDroppedFrames;
+				continue; // drop frame
+			}
+			frame->mTimeToDisplay = time - mFrameDuration;
+			frame->mIteration     = mIteration;
+			frame->_setFrameNumber(frame_number);
+			mLastDecodedFrameNumber = frame_number;
+			th_decode_ycbcr_out(mInfo.TheoraDecoder, buff);
+			TheoraPixelTransform t;
+			memset(&t, 0, sizeof(TheoraPixelTransform));
+			
+			t.y = buff[0].data; t.yStride = buff[0].stride;
+			t.u = buff[1].data; t.uStride = buff[1].stride;
+			t.v = buff[2].data; t.vStride = buff[2].stride;
+			frame->decode(&t);
+			break;
+		}
+		else
+		{
+			if (!_readData())
+			{
+				frame->mInUse = 0;
+				should_restart = mAutoRestart;
+				break;
+			}
+		}
+	}
+	
+	if (mAudioInterface != NULL)
+	{
+		mAudioMutex->lock();
+		decodeAudio();
+		mAudioMutex->unlock();
+	}
+	if (should_restart)
+    {
+        ++mIteration;
+		_restart();
+	}
+	return 1;
+}
+
+void TheoraVideoClip_Theora::_restart()
+{
+	bool paused = mTimer->isPaused();
+	if (!paused) mTimer->pause();
+	long granule=0;
+	th_decode_ctl(mInfo.TheoraDecoder,TH_DECCTL_SET_GRANPOS,&granule,sizeof(granule));
+	th_decode_free(mInfo.TheoraDecoder);
+	mInfo.TheoraDecoder=th_decode_alloc(&mInfo.TheoraInfo,mInfo.TheoraSetup);
+	ogg_stream_reset(&mInfo.TheoraStreamState);
+	if (mAudioInterface)
+	{
+		// empty the DSP buffer
+		//float **pcm;
+		//int len = vorbis_synthesis_pcmout(&mInfo.VorbisDSPState,&pcm);
+		//if (len) vorbis_synthesis_read(&mInfo.VorbisDSPState,len);
+		ogg_packet opVorbis;
+		mReadAudioSamples = 0;
+		while (ogg_stream_packetout(&mInfo.VorbisStreamState,&opVorbis) > 0)
+		{
+			if (vorbis_synthesis(&mInfo.VorbisBlock,&opVorbis) == 0)
+				vorbis_synthesis_blockin(&mInfo.VorbisDSPState,&mInfo.VorbisBlock);
+		}
+		ogg_stream_reset(&mInfo.VorbisStreamState);
+	}
+	
+	ogg_sync_reset(&mInfo.OggSyncState);
+	mStream->seek(0);
+	ogg_int64_t granulePos = 0;
+	th_decode_ctl(mInfo.TheoraDecoder, TH_DECCTL_SET_GRANPOS, &granulePos, sizeof(granule));
+	
+	mEndOfFile = false;
+	
+	mRestarted = 1;
+	
+	if (!paused) mTimer->play();
+}
+
+void TheoraVideoClip_Theora::load(TheoraDataSource* source)
+{
+#ifdef _DEBUG
+	th_writelog("-----");
+#endif
+	mStream = source;
+	readTheoraVorbisHeaders();
+	
+	mInfo.TheoraDecoder = th_decode_alloc(&mInfo.TheoraInfo,mInfo.TheoraSetup);
+	
+	mWidth = mInfo.TheoraInfo.frame_width;
+	mHeight = mInfo.TheoraInfo.frame_height;
+    mSubFrameWidth	 = mInfo.TheoraInfo.pic_width;
+    mSubFrameHeight	 = mInfo.TheoraInfo.pic_height;
+    mSubFrameOffsetX = mInfo.TheoraInfo.pic_x;
+    mSubFrameOffsetY = mInfo.TheoraInfo.pic_y;
+    mStride = (mStride == 1) ? mStride = _nextPow2(getWidth()) : getWidth();
+	mFPS = mInfo.TheoraInfo.fps_numerator / (float) mInfo.TheoraInfo.fps_denominator;
+	
+#ifdef _DEBUG
+	th_writelog("width: " + str(mWidth) + ", height: " + str(mHeight) + ", fps: " + str((int) getFPS()));
+#endif
+	mFrameQueue = new TheoraFrameQueue(this);
+	mFrameQueue->setSize(mNumPrecachedFrames);
+	// find out the duration of the file by seeking to the end
+	// having ogg decode pages, extract the granule pos from
+	// the last theora page and seek back to beginning of the file
+	long streamSize = mStream->size(), seekPos;
+	for (int i = 1; i <= 50; ++i)
+	{
+		ogg_sync_reset(&mInfo.OggSyncState);
+		seekPos = streamSize - 4096 * i;
+		if (seekPos < 0) seekPos = 0;
+		mStream->seek(seekPos);
+		
+		char *buffer = ogg_sync_buffer(&mInfo.OggSyncState, 4096 * i);
+		int bytes_read = mStream->read(buffer, 4096 * i);
+		ogg_sync_wrote(&mInfo.OggSyncState, bytes_read);
+		ogg_sync_pageseek(&mInfo.OggSyncState, &mInfo.OggPage);
+		
+		for (;;)
+		{
+			int ret = ogg_sync_pageout(&mInfo.OggSyncState, &mInfo.OggPage);
+			if (ret == 0) break;
+			// if page is not a theora page, skip it
+			if (ogg_page_serialno(&mInfo.OggPage) != mInfo.TheoraStreamState.serialno) continue;
+			
+			ogg_int64_t granule = ogg_page_granulepos(&mInfo.OggPage);
+			if (granule >= 0)
+			{
+				mNumFrames = (int) th_granule_frame(mInfo.TheoraDecoder, granule) + 1;
+			}
+			else if (mNumFrames > 0)
+				++mNumFrames; // append delta frames at the end to get the exact numbe
+		}
+		if (mNumFrames > 0 || streamSize - 4096 * i < 0) break;
+		
+	}
+	if (mNumFrames < 0)
+		th_writelog("unable to determine file duration!");
+	else
+	{
+		mDuration = mNumFrames / mFPS;
+#ifdef _DEBUG
+		th_writelog("duration: " + strf(mDuration) + " seconds");
+#endif
+	}
+	// restore to beginning of stream.
+	ogg_sync_reset(&mInfo.OggSyncState);
+	mStream->seek(0);
+	
+	if (mVorbisStreams) // if there is no audio interface factory defined, even though the video
+		// clip might have audio, it will be ignored
+	{
+		vorbis_synthesis_init(&mInfo.VorbisDSPState, &mInfo.VorbisInfo);
+		vorbis_block_init(&mInfo.VorbisDSPState, &mInfo.VorbisBlock);
+		mNumAudioChannels = mInfo.VorbisInfo.channels;
+		mAudioFrequency = (int) mInfo.VorbisInfo.rate;
+
+		// create an audio interface instance if available
+		TheoraAudioInterfaceFactory* audio_factory = TheoraVideoManager::getSingleton().getAudioInterfaceFactory();
+		printf("**** audio factory is %p\n", audio_factory);
+		if (audio_factory) setAudioInterface(audio_factory->createInstance(this, mNumAudioChannels, mAudioFrequency));
+	}
+	
+	mFrameDuration = 1.0f / getFPS();
+#ifdef _DEBUG
+	th_writelog("-----");
+#endif
+}
+
+void TheoraVideoClip_Theora::readTheoraVorbisHeaders()
+{
+	ogg_packet tempOggPacket;
+	bool done = false;
+	bool decode_audio=TheoraVideoManager::getSingleton().getAudioInterfaceFactory() != NULL;
+	//init Vorbis/Theora Layer
+	//Ensure all structures get cleared out.
+	memset(&mInfo.OggSyncState, 0, sizeof(ogg_sync_state));
+	memset(&mInfo.OggPage, 0, sizeof(ogg_page));
+	memset(&mInfo.VorbisStreamState, 0, sizeof(ogg_stream_state));
+	memset(&mInfo.TheoraStreamState, 0, sizeof(ogg_stream_state));
+	memset(&mInfo.TheoraInfo, 0, sizeof(th_info));
+	memset(&mInfo.TheoraComment, 0, sizeof(th_comment));
+	memset(&mInfo.VorbisInfo, 0, sizeof(vorbis_info));
+	memset(&mInfo.VorbisDSPState, 0, sizeof(vorbis_dsp_state));
+	memset(&mInfo.VorbisBlock, 0, sizeof(vorbis_block));
+	memset(&mInfo.VorbisComment, 0, sizeof(vorbis_comment));
+	
+	ogg_sync_init(&mInfo.OggSyncState);
+	th_comment_init(&mInfo.TheoraComment);
+	th_info_init(&mInfo.TheoraInfo);
+	vorbis_info_init(&mInfo.VorbisInfo);
+	vorbis_comment_init(&mInfo.VorbisComment);
+	
+	while (!done)
+	{
+		char *buffer = ogg_sync_buffer(&mInfo.OggSyncState, 4096);
+		int bytes_read = mStream->read(buffer, 4096);
+		ogg_sync_wrote(&mInfo.OggSyncState, bytes_read);
+		
+		if (bytes_read == 0)
+			break;
+		
+		while (ogg_sync_pageout(&mInfo.OggSyncState, &mInfo.OggPage) > 0)
+		{
+			ogg_stream_state OggStateTest;
+			
+			//is this an initial header? If not, stop
+			if (!ogg_page_bos(&mInfo.OggPage))
+			{
+				//This is done blindly, because stream only accept themselves
+				if (mTheoraStreams) ogg_stream_pagein(&mInfo.TheoraStreamState, &mInfo.OggPage);
+				if (mVorbisStreams) ogg_stream_pagein(&mInfo.VorbisStreamState, &mInfo.OggPage);
+				
+				done=true;
+				break;
+			}
+			
+			ogg_stream_init(&OggStateTest, ogg_page_serialno(&mInfo.OggPage));
+			ogg_stream_pagein(&OggStateTest, &mInfo.OggPage);
+			ogg_stream_packetout(&OggStateTest, &tempOggPacket);
+			
+			//identify the codec
+			int ret;
+			if (!mTheoraStreams)
+			{
+				ret = th_decode_headerin(&mInfo.TheoraInfo, &mInfo.TheoraComment, &mInfo.TheoraSetup, &tempOggPacket);
+				
+				if (ret > 0)
+				{
+					//This is the Theora Header
+					memcpy(&mInfo.TheoraStreamState, &OggStateTest, sizeof(OggStateTest));
+					mTheoraStreams = 1;
+					continue;
+				}
+			}
+			if (decode_audio && !mVorbisStreams &&
+				vorbis_synthesis_headerin(&mInfo.VorbisInfo, &mInfo.VorbisComment, &tempOggPacket) >=0)
+			{
+				//This is vorbis header
+				memcpy(&mInfo.VorbisStreamState, &OggStateTest, sizeof(OggStateTest));
+				mVorbisStreams = 1;
+				continue;
+			}
+			//Hmm. I guess it's not a header we support, so erase it
+			ogg_stream_clear(&OggStateTest);
+		}
+	}
+	
+	while ((mTheoraStreams && (mTheoraStreams < 3)) ||
+		   (mVorbisStreams && (mVorbisStreams < 3)))
+	{
+		//Check 2nd'dary headers... Theora First
+		int iSuccess;
+		while (mTheoraStreams && mTheoraStreams < 3 &&
+			  (iSuccess = ogg_stream_packetout(&mInfo.TheoraStreamState, &tempOggPacket)))
+		{
+			if (iSuccess < 0)
+				throw TheoraGenericException("Error parsing Theora stream headers.");
+			if (!th_decode_headerin(&mInfo.TheoraInfo, &mInfo.TheoraComment, &mInfo.TheoraSetup, &tempOggPacket))
+				throw TheoraGenericException("invalid theora stream");
+			
+			++mTheoraStreams;
+		} //end while looking for more theora headers
+		
+		//look 2nd vorbis header packets
+		while (mVorbisStreams < 3 && (iSuccess = ogg_stream_packetout(&mInfo.VorbisStreamState, &tempOggPacket)))
+		{
+			if (iSuccess < 0)
+				throw TheoraGenericException("Error parsing vorbis stream headers");
+			
+			if (vorbis_synthesis_headerin(&mInfo.VorbisInfo, &mInfo.VorbisComment,&tempOggPacket))
+				throw TheoraGenericException("invalid stream");
+			
+			++mVorbisStreams;
+		} //end while looking for more vorbis headers
+		
+		//Not finished with Headers, get some more file data
+		if (ogg_sync_pageout(&mInfo.OggSyncState, &mInfo.OggPage) > 0)
+		{
+			if (mTheoraStreams) ogg_stream_pagein(&mInfo.TheoraStreamState, &mInfo.OggPage);
+			if (mVorbisStreams) ogg_stream_pagein(&mInfo.VorbisStreamState, &mInfo.OggPage);
+		}
+		else
+		{
+			char *buffer = ogg_sync_buffer(&mInfo.OggSyncState, 4096);
+			int bytes_read = mStream->read(buffer, 4096);
+			ogg_sync_wrote(&mInfo.OggSyncState, bytes_read);
+			
+			if (bytes_read == 0)
+				throw TheoraGenericException("End of file found prematurely");
+		}
+	} //end while looking for all headers
+	//	writelog("Vorbis Headers: " + str(mVorbisHeaders) + " Theora Headers : " + str(mTheoraHeaders));
+}
+
+void TheoraVideoClip_Theora::decodedAudioCheck()
+{
+	if (!mAudioInterface || mTimer->isPaused()) return;
+
+	mAudioMutex->lock();
+	flushAudioPackets(mAudioInterface);
+	mAudioMutex->unlock();
+}
+
+float TheoraVideoClip_Theora::decodeAudio()
+{
+	if (mRestarted) return -1;
+	
+	ogg_packet opVorbis;
+	float **pcm;
+	int len = 0;
+	float timestamp = -1;
+	bool read_past_timestamp = 0;
+	
+	float factor = 1.0f / mAudioFrequency;
+	float videoTime = (float) mLastDecodedFrameNumber / mFPS;
+	float min = mFrameQueue->getSize() / mFPS + 1.0f;
+
+	for (;;)
+	{
+		len = vorbis_synthesis_pcmout(&mInfo.VorbisDSPState, &pcm);
+		if (len == 0)
+		{
+			if (ogg_stream_packetout(&mInfo.VorbisStreamState, &opVorbis) > 0)
+			{
+				if (vorbis_synthesis(&mInfo.VorbisBlock, &opVorbis) == 0)
+				{
+					if (timestamp < 0 && opVorbis.granulepos >= 0)
+					{
+						timestamp = (float) vorbis_granule_time(&mInfo.VorbisDSPState, opVorbis.granulepos);
+					}
+					else if (timestamp >= 0) read_past_timestamp = 1;
+					vorbis_synthesis_blockin(&mInfo.VorbisDSPState, &mInfo.VorbisBlock);
+				}
+				continue;
+			}
+			else
+			{
+				float audioTime = mReadAudioSamples * factor;
+				// always buffer up of audio ahead of the frames
+				if (audioTime - videoTime < min)
+				{
+					if (!_readData()) break;
+				}
+				else
+					break;
+			}
+		}
+		addAudioPacket(pcm, len, mAudioGain);
+		mReadAudioSamples += len;
+		if (read_past_timestamp) timestamp += (float) len / mInfo.VorbisInfo.rate;
+		vorbis_synthesis_read(&mInfo.VorbisDSPState, len); // tell vorbis we read a number of samples
+	}
+	return timestamp;
+}
+
+long TheoraVideoClip_Theora::seekPage(long targetFrame, bool return_keyframe)
+{
+	int i,seek_min = 0, seek_max = (int) mStream->size();
+	long frame;
+	ogg_int64_t granule = 0;
+	
+	if (targetFrame == 0) mStream->seek(0);
+	for (i = (targetFrame == 0) ? 100 : 0; i < 100; ++i)
+	{
+		ogg_sync_reset(&mInfo.OggSyncState);
+		mStream->seek((seek_min + seek_max) / 2); // do a binary search
+		memset(&mInfo.OggPage, 0, sizeof(ogg_page));
+		ogg_sync_pageseek(&mInfo.OggSyncState, &mInfo.OggPage);
+		
+		for (;i < 1000;)
+		{
+			int ret = ogg_sync_pageout(&mInfo.OggSyncState, &mInfo.OggPage);
+			if (ret == 1)
+			{
+				int serno = ogg_page_serialno(&mInfo.OggPage);
+				if (serno == mInfo.TheoraStreamState.serialno)
+				{
+					granule = ogg_page_granulepos(&mInfo.OggPage);
+					if (granule >= 0)
+					{
+						frame = (long) th_granule_frame(mInfo.TheoraDecoder, granule);
+						if (frame < targetFrame && targetFrame - frame < 10)
+						{
+							// we're close enough, let's break this.
+							i = 1000;
+							break;
+						}
+						// we're not close enough, let's shorten the borders of the binary search
+						if (targetFrame - 1 > frame) seek_min = (seek_min + seek_max) / 2;
+						else				         seek_max = (seek_min + seek_max) / 2;
+						break;
+					}
+				}
+			}
+			else
+			{
+				char *buffer = ogg_sync_buffer(&mInfo.OggSyncState, 4096);
+				int bytes_read = mStream->read(buffer, 4096);
+				if (bytes_read == 0) break;
+				ogg_sync_wrote(&mInfo.OggSyncState, bytes_read);
+			}
+		}
+	}
+	if (return_keyframe) return (long) (granule >> mInfo.TheoraInfo.keyframe_granule_shift);
+	
+	ogg_sync_reset(&mInfo.OggSyncState);
+	memset(&mInfo.OggPage, 0, sizeof(ogg_page));
+	ogg_sync_pageseek(&mInfo.OggSyncState, &mInfo.OggPage);
+	if (targetFrame == 0) return -1;
+	mStream->seek((seek_min + seek_max) / 2); // do a binary search
+	return -1;
+}
+
+void TheoraVideoClip_Theora::doSeek()
+{
+#if _DEBUG
+	th_writelog(mName + " [seek]: seeking to frame " + str(mSeekFrame));
+#endif
+	int frame;
+	float time = mSeekFrame / getFPS();
+	mTimer->seek(time);
+	bool paused = mTimer->isPaused();
+	if (!paused) mTimer->pause(); // pause until seeking is done
+	
+	mEndOfFile = false;
+	mRestarted = false;
+	
+	resetFrameQueue();
+	// reset the video decoder.
+	ogg_stream_reset(&mInfo.TheoraStreamState);
+	th_decode_free(mInfo.TheoraDecoder);
+	mInfo.TheoraDecoder = th_decode_alloc(&mInfo.TheoraInfo, mInfo.TheoraSetup);
+	
+	if (mAudioInterface)
+	{
+		mAudioMutex->lock();
+		ogg_stream_reset(&mInfo.VorbisStreamState);
+		vorbis_synthesis_restart(&mInfo.VorbisDSPState);
+		destroyAllAudioPackets();
+	}
+	// first seek to desired frame, then figure out the location of the
+	// previous keyframe and seek to it.
+	// then by setting the correct time, the decoder will skip N frames untill
+	// we get the frame we want.
+	frame = (int) seekPage(mSeekFrame, 1); // find the keyframe nearest to the target frame
+#ifdef _DEBUG
+	//		th_writelog(mName + " [seek]: nearest keyframe for frame " + str(mSeekFrame) + " is frame: " + str(frame));
+#endif
+	seekPage(std::max(0, frame - 1), 0);
+	
+	ogg_packet opTheora;
+	ogg_int64_t granulePos;
+	bool granule_set = 0;
+	if (frame <= 1)
+	{
+		if (mInfo.TheoraInfo.version_major == 3 && mInfo.TheoraInfo.version_minor == 2 && mInfo.TheoraInfo.version_subminor == 0)
+			granulePos = 0;
+		else
+			granulePos = 1; // because of difference in granule interpretation in theora streams 3.2.0 and newer ones
+		th_decode_ctl(mInfo.TheoraDecoder, TH_DECCTL_SET_GRANPOS, &granulePos, sizeof(granulePos));
+		granule_set = 1;
+	}
+	
+	// now that we've found the keyframe that preceeds our desired frame, lets keep on decoding frames until we
+	// reach our target frame.
+	
+    int status, ret;
+	for (;mSeekFrame != 0;)
+	{
+		ret = ogg_stream_packetout(&mInfo.TheoraStreamState, &opTheora);
+		if (ret > 0)
+		{
+			if (!granule_set)
+			{
+				// theora decoder requires to set the granule pos after seek to be able to determine the current frame
+				if (opTheora.granulepos >= 0)
+				{
+					th_decode_ctl(mInfo.TheoraDecoder, TH_DECCTL_SET_GRANPOS, &opTheora.granulepos, sizeof(opTheora.granulepos));
+					granule_set = 1;
+				}
+				else continue; // ignore prev delta frames until we hit a keyframe
+			}
+			status = th_decode_packetin(mInfo.TheoraDecoder, &opTheora, &granulePos);
+            if (status != 0 && status != TH_DUPFRAME) continue;
+			frame = (int) th_granule_frame(mInfo.TheoraDecoder, granulePos);
+			if (frame >= mSeekFrame - 1) break;
+		}
+		else
+		{
+			if (!_readData())
+			{
+				th_writelog(mName + " [seek]: fineseeking failed, _readData failed!");
+				if (mAudioInterface) mAudioMutex->unlock();
+				return;
+			}
+		}
+	}
+#ifdef _DEBUG
+	//	th_writelog(mName + " [seek]: fineseeked to frame " + str(frame + 1) + ", requested: " + str(mSeekFrame));
+#endif
+	if (mAudioInterface)
+	{
+		// read audio data until we reach a timestamp. this usually takes only one iteration, but just in case let's
+		// wrap it in a loop
+		float timestamp;
+		for (;;)
+		{
+			timestamp = decodeAudio();
+			if (timestamp >= 0) break;
+			else _readData();
+		}
+		float rate = (float) mAudioFrequency * mNumAudioChannels;
+		float queued_time = getAudioPacketQueueLength();
+		// at this point there are only 2 possibilities: either we have too much packets and we have to delete
+		// the first N ones, or we don't have enough, so let's fill the gap with silence.
+ 		if (time > timestamp - queued_time)
+		{
+			while (mTheoraAudioPacketQueue != NULL)
+			{
+				if (time > timestamp - queued_time + mTheoraAudioPacketQueue->numSamples / rate)
+				{
+					queued_time -= mTheoraAudioPacketQueue->numSamples / rate;
+					destroyAudioPacket(popAudioPacket());
+				}
+				else
+				{
+					int n_trim = (int) ((timestamp - queued_time + mTheoraAudioPacketQueue->numSamples / rate - time) * rate);
+					if (mTheoraAudioPacketQueue->numSamples - n_trim <= 0)
+						destroyAudioPacket(popAudioPacket()); // if there's no data to be left, just destroy it
+					else
+					{
+						for (int i = n_trim, j = 0; i < mTheoraAudioPacketQueue->numSamples; ++i, ++j)
+							mTheoraAudioPacketQueue->pcm[j] = mTheoraAudioPacketQueue->pcm[i];
+						mTheoraAudioPacketQueue->numSamples -= n_trim;
+					}
+					break;
+				}
+			}
+		}
+		else
+		{
+			// expand the first packet with silence.
+			if (mTheoraAudioPacketQueue) // just in case!
+			{
+				int i, j, nmissing = (int) ((timestamp - queued_time - time) * rate);
+				if (nmissing > 0)
+				{
+					float* samples = new float[nmissing + mTheoraAudioPacketQueue->numSamples];
+					for (i = 0; i < nmissing; ++i) samples[i] = 0;
+					for (j = 0; i < nmissing + mTheoraAudioPacketQueue->numSamples; ++i, ++j)
+						samples[i] = mTheoraAudioPacketQueue->pcm[j];
+					delete [] mTheoraAudioPacketQueue->pcm;
+					mTheoraAudioPacketQueue->pcm = samples;
+				}
+			}
+		}
+		mLastDecodedFrameNumber = mSeekFrame;
+		mReadAudioSamples = (unsigned int) (timestamp * mAudioFrequency);
+		
+		mAudioMutex->unlock();
+	}
+	if (!paused) mTimer->play();
+	mSeekFrame = -1;
+}
+#endif
diff --git a/drivers/theoraplayer/src/Theora/TheoraVideoClip_Theora.h b/drivers/theoraplayer/src/Theora/TheoraVideoClip_Theora.h
new file mode 100644
index 0000000000..c64c183029
--- /dev/null
+++ b/drivers/theoraplayer/src/Theora/TheoraVideoClip_Theora.h
@@ -0,0 +1,64 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#if defined(__THEORA) && !defined(_TheoraVideoClip_Theora_h)
+#define _TheoraVideoClip_Theora_h
+
+#include <ogg/ogg.h>
+#include <vorbis/vorbisfile.h>
+#include <theora/theoradec.h>
+#include "TheoraAudioPacketQueue.h"
+#include "TheoraVideoClip.h"
+
+struct TheoraInfoStruct
+{
+	// ogg/vorbis/theora variables
+	ogg_sync_state   OggSyncState;
+	ogg_page         OggPage;
+	ogg_stream_state VorbisStreamState;
+	ogg_stream_state TheoraStreamState;
+	//Theora State
+	th_info        TheoraInfo;
+	th_comment     TheoraComment;
+	th_setup_info* TheoraSetup;
+	th_dec_ctx*    TheoraDecoder;
+	//Vorbis State
+	vorbis_info      VorbisInfo;
+	vorbis_dsp_state VorbisDSPState;
+	vorbis_block     VorbisBlock;
+	vorbis_comment   VorbisComment;
+};
+
+class TheoraVideoClip_Theora : public TheoraVideoClip, public TheoraAudioPacketQueue
+{
+protected:
+	TheoraInfoStruct mInfo; // a pointer is used to avoid having to include theora & vorbis headers
+	int mTheoraStreams, mVorbisStreams;	// Keeps track of Theora and Vorbis Streams
+
+	long seekPage(long targetFrame, bool return_keyframe);
+	void doSeek();
+	void readTheoraVorbisHeaders();
+	unsigned int mReadAudioSamples;
+	unsigned long mLastDecodedFrameNumber;
+public:
+	TheoraVideoClip_Theora(TheoraDataSource* data_source,
+						   TheoraOutputMode output_mode,
+						   int nPrecachedFrames,
+						   bool usePower2Stride);
+	~TheoraVideoClip_Theora();
+
+	bool _readData();
+	bool decodeNextFrame();
+	void _restart();
+	void load(TheoraDataSource* source);
+	float decodeAudio();
+	void decodedAudioCheck();
+	std::string getDecoderName() { return "Theora"; }
+};
+
+#endif
diff --git a/drivers/theoraplayer/src/TheoraAsync.cpp b/drivers/theoraplayer/src/TheoraAsync.cpp
new file mode 100644
index 0000000000..cc3b7a4bf5
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraAsync.cpp
@@ -0,0 +1,253 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#ifdef _WIN32
+#include <windows.h>
+#else
+#include <unistd.h>
+#include <pthread.h>
+#endif
+
+#include "TheoraAsync.h"
+#include "TheoraUtil.h"
+
+#ifdef _WINRT
+#include <wrl.h>
+#endif
+
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Mutex
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+TheoraMutex::TheoraMutex()
+{
+#ifdef _WIN32
+#ifndef _WINRT // WinXP does not have CreateTheoraMutexEx()
+	mHandle = CreateMutex(0, 0, 0);
+#else
+	mHandle = CreateMutexEx(NULL, NULL, 0, SYNCHRONIZE);
+#endif
+#else
+	mHandle = (pthread_mutex_t*)malloc(sizeof(pthread_mutex_t));
+	pthread_mutex_init((pthread_mutex_t*)mHandle, 0);
+#endif
+}
+
+TheoraMutex::~TheoraMutex()
+{
+#ifdef _WIN32
+	CloseHandle(mHandle);
+#else
+	pthread_mutex_destroy((pthread_mutex_t*)mHandle);
+	free((pthread_mutex_t*)mHandle);
+	mHandle = NULL;
+#endif
+}
+
+void TheoraMutex::lock()
+{
+#ifdef _WIN32
+	WaitForSingleObjectEx(mHandle, INFINITE, FALSE);
+#else
+	pthread_mutex_lock((pthread_mutex_t*)mHandle);
+#endif
+}
+
+void TheoraMutex::unlock()
+{
+#ifdef _WIN32
+	ReleaseMutex(mHandle);
+#else
+	pthread_mutex_unlock((pthread_mutex_t*)mHandle);
+#endif
+}
+	
+///////////////////////////////////////////////////////////////////////////////////////////////////
+// Thread
+///////////////////////////////////////////////////////////////////////////////////////////////////
+
+#ifdef _WINRT
+using namespace Windows::Foundation;
+using namespace Windows::System::Threading;
+#endif
+
+#ifdef _WIN32
+unsigned long WINAPI theoraAsyncCall(void* param)
+#else
+void* theoraAsyncCall(void* param)
+#endif
+{
+	TheoraThread* t = (TheoraThread*)param;
+	t->execute();
+#ifdef _WIN32
+	return 0;
+#else
+	pthread_exit(NULL);
+	return NULL;
+#endif
+}
+
+#ifdef _WINRT
+struct TheoraAsyncActionWrapper
+{
+public:
+	IAsyncAction^ mAsyncAction;
+	TheoraAsyncActionWrapper(IAsyncAction^ asyncAction)
+	{
+		mAsyncAction = asyncAction;
+	}
+};
+#endif
+	
+TheoraThread::TheoraThread() : mRunning(false), mId(0)
+{
+#ifndef _WIN32
+	mId = (pthread_t*)malloc(sizeof(pthread_t));
+#endif
+}
+
+TheoraThread::~TheoraThread()
+{
+	if (mRunning)
+	{
+		stop();
+	}
+	if (mId != NULL)
+	{
+#ifdef _WIN32
+#ifndef _WINRT
+		CloseHandle(mId);
+#else
+		delete mId;
+#endif
+#else
+		free((pthread_t*)mId);
+#endif
+		mId = NULL;
+	}
+}
+
+void TheoraThread::start()
+{
+	mRunning = true;
+#ifdef _WIN32
+#ifndef _WINRT
+	mId = CreateThread(0, 0, &theoraAsyncCall, this, 0, 0);
+#else
+	mId = new TheoraAsyncActionWrapper(ThreadPool::RunAsync(
+		ref new WorkItemHandler([&](IAsyncAction^ work_item)
+		{
+			execute();
+		}),
+		WorkItemPriority::Normal, WorkItemOptions::TimeSliced));
+#endif
+#else
+	pthread_create((pthread_t*)mId, NULL, &theoraAsyncCall, this);
+#endif
+}
+
+bool TheoraThread::isRunning()
+{
+	bool ret;
+	mRunningMutex.lock();
+	ret = mRunning;
+	mRunningMutex.unlock();
+	
+	return ret;
+}
+
+void TheoraThread::join()
+{
+	mRunningMutex.lock();
+	mRunning = false;
+	mRunningMutex.unlock();
+#ifdef _WIN32
+#ifndef _WINRT
+	WaitForSingleObject(mId, INFINITE);
+	if (mId != NULL)
+	{
+		CloseHandle(mId);
+		mId = NULL;
+	}
+#else
+	IAsyncAction^ action = ((TheoraAsyncActionWrapper*)mId)->mAsyncAction;
+	int i = 0;
+	while (action->Status != AsyncStatus::Completed &&
+		action->Status != AsyncStatus::Canceled &&
+		action->Status != AsyncStatus::Error &&
+		i < 100)
+	{
+		_psleep(50);
+		++i;
+	}
+	if (i >= 100)
+	{
+		i = 0;
+		action->Cancel();
+		while (action->Status != AsyncStatus::Completed &&
+			action->Status != AsyncStatus::Canceled &&
+			action->Status != AsyncStatus::Error &&
+			i < 100)
+		{
+			_psleep(50);
+			++i;
+		}
+	}
+#endif
+#else
+	pthread_join(*((pthread_t*)mId), 0);
+#endif
+}
+	
+void TheoraThread::resume()
+{
+#ifdef _WIN32
+#ifndef _WINRT
+	ResumeThread(mId);
+#else
+	// not available in WinRT
+#endif
+#endif
+}
+	
+void TheoraThread::pause()
+{
+#ifdef _WIN32
+#ifndef _WINRT
+	SuspendThread(mId);
+#else
+	// not available in WinRT
+#endif
+#endif
+}
+	
+void TheoraThread::stop()
+{
+	if (mRunning)
+	{
+		mRunningMutex.lock();
+		mRunning = false;
+		mRunningMutex.unlock();
+#ifdef _WIN32
+#ifndef _WINRT
+		TerminateThread(mId, 0);
+#else
+		((TheoraAsyncActionWrapper*)mId)->mAsyncAction->Cancel();
+#endif
+#elif defined(_ANDROID)
+		pthread_kill(*((pthread_t*)mId), 0);
+#else
+		pthread_cancel(*((pthread_t*)mId));
+#endif
+	}
+}
+	
diff --git a/drivers/theoraplayer/src/TheoraAudioInterface.cpp b/drivers/theoraplayer/src/TheoraAudioInterface.cpp
new file mode 100644
index 0000000000..a265cb57b5
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraAudioInterface.cpp
@@ -0,0 +1,21 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "TheoraAudioInterface.h"
+
+TheoraAudioInterface::TheoraAudioInterface(TheoraVideoClip* owner, int nChannels, int freq)
+{
+	mFreq = freq;
+	mNumChannels = nChannels;
+	mClip = owner;
+}
+
+TheoraAudioInterface::~TheoraAudioInterface()
+{
+	
+}
diff --git a/drivers/theoraplayer/src/TheoraAudioPacketQueue.cpp b/drivers/theoraplayer/src/TheoraAudioPacketQueue.cpp
new file mode 100644
index 0000000000..be5e1018f9
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraAudioPacketQueue.cpp
@@ -0,0 +1,126 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include <stdlib.h>
+#include "TheoraAudioPacketQueue.h"
+#include "TheoraAudioInterface.h"
+
+TheoraAudioPacketQueue::TheoraAudioPacketQueue()
+{
+	mTheoraAudioPacketQueue = NULL;
+}
+
+TheoraAudioPacketQueue::~TheoraAudioPacketQueue()
+{
+	destroyAllAudioPackets();
+}
+
+float TheoraAudioPacketQueue::getAudioPacketQueueLength()
+{
+	float len = 0;
+	for (TheoraAudioPacket* p = mTheoraAudioPacketQueue; p != NULL; p = p->next)
+		len += p->numSamples;
+	
+	return len / (mAudioFrequency * mNumAudioChannels);
+}
+
+void TheoraAudioPacketQueue::_addAudioPacket(float* data, int numSamples)
+{
+	TheoraAudioPacket* packet = new TheoraAudioPacket;
+	packet->pcm = data;
+	packet->numSamples = numSamples;
+	packet->next = NULL;
+
+
+	if (mTheoraAudioPacketQueue == NULL) mTheoraAudioPacketQueue = packet;
+	else
+	{
+		TheoraAudioPacket* last = mTheoraAudioPacketQueue;
+		for (TheoraAudioPacket* p = last; p != NULL; p = p->next)
+			last = p;
+		last->next = packet;
+	}
+}
+
+void TheoraAudioPacketQueue::addAudioPacket(float** buffer, int numSamples, float gain)
+{
+	float* data = new float[numSamples * mNumAudioChannels];
+	float* dataptr = data;
+	int i;
+	unsigned int j;
+	
+	if (gain < 1.0f)
+	{
+		// apply gain, let's attenuate the samples
+		for (i = 0; i < numSamples; ++i)
+			for (j = 0; j < mNumAudioChannels; j++, ++dataptr)
+				*dataptr = buffer[i][j] * gain;
+	}
+	else
+	{
+		// do a simple copy, faster then the above method, when gain is 1.0f
+		for (i = 0; i < numSamples; ++i)
+			for (j = 0; j < mNumAudioChannels; j++, ++dataptr)
+				*dataptr = buffer[j][i];
+	}
+		
+	_addAudioPacket(data, numSamples * mNumAudioChannels);
+}
+
+void TheoraAudioPacketQueue::addAudioPacket(float* buffer, int numSamples, float gain)
+{
+	float* data = new float[numSamples * mNumAudioChannels];
+	float* dataptr = data;
+	int i, numFloats = numSamples * mNumAudioChannels;
+	
+	if (gain < 1.0f)
+	{
+		// apply gain, let's attenuate the samples
+		for (i = 0; i < numFloats; ++i, dataptr++)
+			*dataptr = buffer[i] * gain;
+	}
+	else
+	{
+		// do a simple copy, faster then the above method, when gain is 1.0f
+		for (i = 0; i < numFloats; ++i, dataptr++)
+			*dataptr = buffer[i];
+	}
+	
+	_addAudioPacket(data, numFloats);
+}
+
+TheoraAudioPacket* TheoraAudioPacketQueue::popAudioPacket()
+{
+	if (mTheoraAudioPacketQueue == NULL) return NULL;
+	TheoraAudioPacket* p = mTheoraAudioPacketQueue;
+	mTheoraAudioPacketQueue = mTheoraAudioPacketQueue->next;
+	return p;
+}
+
+void TheoraAudioPacketQueue::destroyAudioPacket(TheoraAudioPacket* p)
+{
+	if (p == NULL) return;
+	delete [] p->pcm;
+	delete p;
+}
+
+void TheoraAudioPacketQueue::destroyAllAudioPackets()
+{
+	for (TheoraAudioPacket* p = popAudioPacket(); p != NULL; p = popAudioPacket())
+		destroyAudioPacket(p);
+}
+
+void TheoraAudioPacketQueue::flushAudioPackets(TheoraAudioInterface* audioInterface)
+{
+	
+	for (TheoraAudioPacket* p = popAudioPacket(); p != NULL; p = popAudioPacket())
+	{
+		audioInterface->insertData(p->pcm, p->numSamples);
+		destroyAudioPacket(p);
+	}
+}
\ No newline at end of file
diff --git a/drivers/theoraplayer/src/TheoraDataSource.cpp b/drivers/theoraplayer/src/TheoraDataSource.cpp
new file mode 100644
index 0000000000..6011dc6783
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraDataSource.cpp
@@ -0,0 +1,128 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include <stdio.h>
+#include <memory.h>
+#include "TheoraDataSource.h"
+#include "TheoraException.h"
+#include "TheoraVideoManager.h"
+#include "TheoraUtil.h"
+
+TheoraDataSource::~TheoraDataSource()
+{
+
+}
+
+TheoraFileDataSource::TheoraFileDataSource(std::string filename)
+{
+	mFilename = filename;
+	mFilePtr = NULL;
+}
+
+TheoraFileDataSource::~TheoraFileDataSource()
+{
+	if (mFilePtr)
+	{
+		fclose(mFilePtr);
+		mFilePtr = NULL;
+	}
+}
+
+void TheoraFileDataSource::openFile()
+{
+	if (mFilePtr == NULL)
+	{
+		mFilePtr=fopen(mFilename.c_str(), "rb");
+		if (!mFilePtr)
+        {
+            std::string msg = "Can't open video file: " + mFilename;
+            th_writelog(msg);
+            throw TheoraGenericException(msg);
+        }
+		fseek(mFilePtr, 0, SEEK_END);
+		mSize = ftell(mFilePtr);
+		fseek(mFilePtr, 0, SEEK_SET);
+	}
+}
+
+int TheoraFileDataSource::read(void* output, int nBytes)
+{
+	if (mFilePtr == NULL) openFile();
+	size_t n = fread(output, 1, nBytes, mFilePtr);
+	return (int) n;
+}
+
+void TheoraFileDataSource::seek(unsigned long byte_index)
+{
+	if (mFilePtr == NULL) openFile();
+	fseek(mFilePtr, byte_index, SEEK_SET);
+}
+
+unsigned long TheoraFileDataSource::size()
+{
+	if (mFilePtr == NULL) openFile();
+	return mSize;
+}
+
+unsigned long TheoraFileDataSource::tell()
+{
+	if (mFilePtr == NULL) return 0;
+	return ftell(mFilePtr);
+}
+
+TheoraMemoryFileDataSource::TheoraMemoryFileDataSource(std::string filename) :
+	mReadPointer(0),
+	mData(0)
+{
+	mFilename=filename;
+	FILE* f=fopen(filename.c_str(),"rb");
+	if (!f) throw TheoraGenericException("Can't open video file: "+filename);
+	fseek(f,0,SEEK_END);
+	mSize=ftell(f);
+	fseek(f,0,SEEK_SET);
+	mData=new unsigned char[mSize];
+	fread(mData,1,mSize,f);
+	fclose(f);
+}
+
+TheoraMemoryFileDataSource::TheoraMemoryFileDataSource(unsigned char* data, long size, const std::string& filename)
+{
+	mFilename = filename;
+	mData = data;
+	mSize = size;
+	mReadPointer = 0;
+}
+
+TheoraMemoryFileDataSource::~TheoraMemoryFileDataSource()
+{
+	if (mData) delete [] mData;
+}
+
+int TheoraMemoryFileDataSource::read(void* output, int nBytes)
+{
+	int n = (int) ((mReadPointer+nBytes <= mSize) ? nBytes : mSize - mReadPointer);
+	if (!n) return 0;
+	memcpy(output, mData + mReadPointer, n);
+	mReadPointer += n;
+	return n;
+}
+
+void TheoraMemoryFileDataSource::seek(unsigned long byte_index)
+{
+	mReadPointer=byte_index;
+}
+
+unsigned long TheoraMemoryFileDataSource::size()
+{
+	return mSize;
+}
+
+unsigned long TheoraMemoryFileDataSource::tell()
+{
+	return mReadPointer;
+}
diff --git a/drivers/theoraplayer/src/TheoraException.cpp b/drivers/theoraplayer/src/TheoraException.cpp
new file mode 100644
index 0000000000..4588a81397
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraException.cpp
@@ -0,0 +1,37 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "TheoraException.h"
+#include "TheoraUtil.h"
+#include "TheoraVideoManager.h"
+#include <stdio.h>
+
+_TheoraGenericException::_TheoraGenericException(const std::string& errorText, std::string type, std::string file, int line)
+{
+    mErrText = errorText;
+	int src = (int) file.find("src");
+	if (src >= 0) file = file.substr(src + 4, 1000);
+	mLineNumber = line;
+	mFile = file;
+}
+
+
+std::string _TheoraGenericException::repr()
+{
+	std::string text = getType();
+	if (text != "") text += ": ";
+
+	if (mFile != "") text += "[" + mFile + ":" + str(mLineNumber) + "] - ";
+
+	return text + getErrorText();
+}
+
+void _TheoraGenericException::writeOutput()
+{
+	th_writelog("----------------\nException Error!\n\n" + repr() + "\n----------------");
+}
diff --git a/drivers/theoraplayer/src/TheoraFrameQueue.cpp b/drivers/theoraplayer/src/TheoraFrameQueue.cpp
new file mode 100644
index 0000000000..f402144795
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraFrameQueue.cpp
@@ -0,0 +1,174 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "TheoraFrameQueue.h"
+#include "TheoraVideoFrame.h"
+#include "TheoraVideoManager.h"
+#include "TheoraUtil.h"
+
+
+TheoraFrameQueue::TheoraFrameQueue(TheoraVideoClip* parent)
+{
+	mParent = parent;
+}
+
+TheoraFrameQueue::~TheoraFrameQueue()
+{
+	foreach_l(TheoraVideoFrame*, mQueue)
+    {
+		delete (*it);
+    }
+	mQueue.clear();
+}
+
+TheoraVideoFrame* TheoraFrameQueue::createFrameInstance(TheoraVideoClip* clip)
+{
+	TheoraVideoFrame* frame = new TheoraVideoFrame(clip);
+	if (frame->getBuffer() == NULL) // This can happen if you run out of memory
+	{
+		delete frame;
+		return NULL;
+	}
+	return frame;
+}
+
+void TheoraFrameQueue::setSize(int n)
+{
+	mMutex.lock();
+	if (mQueue.size() > 0)
+	{
+		foreach_l (TheoraVideoFrame*, mQueue)
+        {
+			delete (*it);
+        }
+		mQueue.clear();
+	}
+	TheoraVideoFrame* frame;
+	for (int i = 0;i < n; ++i)
+	{
+		frame = createFrameInstance(mParent);
+		if (frame != NULL) mQueue.push_back(frame);
+		else
+		{
+			TheoraVideoManager::getSingleton().logMessage("TheoraFrameQueue: unable to create " + str(n) + " frames, out of memory. Created " + str((int) mQueue.size()) + " frames.");
+			break;
+		}
+	}
+	mMutex.unlock();
+}
+
+int TheoraFrameQueue::getSize()
+{
+	return (int) mQueue.size();
+}
+
+TheoraVideoFrame* TheoraFrameQueue::_getFirstAvailableFrame()
+{
+	TheoraVideoFrame* frame = mQueue.front();
+	if (frame->mReady) return frame;
+	else               return NULL;
+}
+
+TheoraVideoFrame* TheoraFrameQueue::getFirstAvailableFrame()
+{
+	mMutex.lock();
+	TheoraVideoFrame* frame = _getFirstAvailableFrame();
+	mMutex.unlock();
+	return frame;
+}
+
+void TheoraFrameQueue::clear()
+{
+	mMutex.lock();
+	foreach_l (TheoraVideoFrame*, mQueue)
+		(*it)->clear();
+	mMutex.unlock();
+}
+
+void TheoraFrameQueue::_pop(int n)
+{
+    for (int i = 0; i < n; ++i)
+    {
+        TheoraVideoFrame* first = mQueue.front();
+        first->clear();
+        mQueue.pop_front();
+        mQueue.push_back(first);
+    }
+}
+
+void TheoraFrameQueue::pop(int n)
+{
+	mMutex.lock();
+    _pop(n);
+	mMutex.unlock();
+}
+
+TheoraVideoFrame* TheoraFrameQueue::requestEmptyFrame()
+{
+	TheoraVideoFrame* frame = NULL;
+	mMutex.lock();
+	foreach_l (TheoraVideoFrame*, mQueue)
+	{
+		if (!(*it)->mInUse)
+		{
+			(*it)->mInUse = 1;
+			(*it)->mReady = 0;
+			frame = (*it);
+			break;
+		}
+	}
+	mMutex.unlock();
+	return frame;
+}
+
+int TheoraFrameQueue::getUsedCount()
+{
+	mMutex.lock();
+	int n=0;
+	foreach_l(TheoraVideoFrame*,mQueue)
+		if ((*it)->mInUse) ++n;
+	mMutex.unlock();
+	return n;
+}
+
+int TheoraFrameQueue::_getReadyCount()
+{
+	int n = 0;
+	foreach_l (TheoraVideoFrame*, mQueue)
+    if ((*it)->mReady) ++n;
+	return n;
+}
+
+
+int TheoraFrameQueue::getReadyCount()
+{
+	mMutex.lock();
+	int n = _getReadyCount();
+	mMutex.unlock();
+	return n;
+}
+
+bool TheoraFrameQueue::isFull()
+{
+	return getReadyCount() == mQueue.size();
+}
+
+void TheoraFrameQueue::lock()
+{
+	mMutex.lock();
+}
+
+void TheoraFrameQueue::unlock()
+{
+	mMutex.unlock();
+}
+
+std::list<TheoraVideoFrame*>& TheoraFrameQueue::_getFrameQueue()
+{
+    return mQueue;
+}
diff --git a/drivers/theoraplayer/src/TheoraTimer.cpp b/drivers/theoraplayer/src/TheoraTimer.cpp
new file mode 100644
index 0000000000..644d1c2ab7
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraTimer.cpp
@@ -0,0 +1,70 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "TheoraTimer.h"
+
+TheoraTimer::TheoraTimer()
+{
+	mTime = 0;
+	mPaused = 0;
+    mSpeed = 1.0f;
+}
+
+TheoraTimer::~TheoraTimer()
+{
+
+}
+
+void TheoraTimer::update(float timeDelta)
+{
+	if (!isPaused())
+	{
+		mTime += timeDelta * mSpeed;
+	}
+}
+
+float TheoraTimer::getTime()
+{
+	return mTime;
+}
+
+void TheoraTimer::pause()
+{
+	mPaused = true;
+}
+
+void TheoraTimer::play()
+{
+	mPaused = false;
+}
+
+
+bool TheoraTimer::isPaused()
+{
+	return mPaused;
+}
+
+void TheoraTimer::stop()
+{
+
+}
+
+void TheoraTimer::seek(float time)
+{
+	mTime = time;
+}
+
+void TheoraTimer::setSpeed(float speed)
+{
+    mSpeed = speed;
+}
+
+float TheoraTimer::getSpeed()
+{
+    return mSpeed;
+}
diff --git a/drivers/theoraplayer/src/TheoraUtil.cpp b/drivers/theoraplayer/src/TheoraUtil.cpp
new file mode 100644
index 0000000000..8f1ad0c9c1
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraUtil.cpp
@@ -0,0 +1,59 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include <stdio.h>
+#include <algorithm>
+#include <math.h>
+#include <map>
+#ifndef _WIN32
+#include <unistd.h>
+#include <pthread.h>
+#endif
+
+#include "TheoraUtil.h"
+#include "TheoraException.h"
+
+#ifdef _WIN32
+#include <windows.h>
+#pragma warning( disable: 4996 ) // MSVC++
+#endif
+
+std::string str(int i)
+{
+    char s[32];
+    sprintf(s, "%d", i);
+    return std::string(s);
+}
+
+std::string strf(float i)
+{
+    char s[32];
+    sprintf(s, "%.3f", i);
+    return std::string(s);
+}
+
+void _psleep(int miliseconds)
+{
+#ifdef _WIN32
+#ifndef _WINRT
+	Sleep(miliseconds);
+#else
+	WaitForSingleObjectEx(GetCurrentThread(), miliseconds, 0);
+#endif
+#else
+	usleep(miliseconds * 1000);
+#endif
+}
+
+
+int _nextPow2(int x)
+{
+	int y;
+	for (y = 1; y < x; y *= 2);
+	return y;
+}
diff --git a/drivers/theoraplayer/src/TheoraVideoClip.cpp b/drivers/theoraplayer/src/TheoraVideoClip.cpp
new file mode 100644
index 0000000000..3ee4b83370
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraVideoClip.cpp
@@ -0,0 +1,493 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "TheoraVideoClip.h"
+#include "TheoraVideoManager.h"
+#include "TheoraVideoFrame.h"
+#include "TheoraFrameQueue.h"
+#include "TheoraAudioInterface.h"
+#include "TheoraTimer.h"
+#include "TheoraDataSource.h"
+#include "TheoraUtil.h"
+#include "TheoraException.h"
+
+#include "core/os/memory.h"
+
+TheoraVideoClip::TheoraVideoClip(TheoraDataSource* data_source,
+								 TheoraOutputMode output_mode,
+								 int nPrecachedFrames,
+								 bool usePower2Stride):
+	mAudioInterface(NULL),
+	mNumDroppedFrames(0),
+	mNumDisplayedFrames(0),
+	mSeekFrame(-1),
+	mDuration(-1),
+	mNumFrames(-1),
+	mFPS(1),
+	mUseAlpha(0),
+	mFrameDuration(0),
+	mName(data_source->repr()),
+	mStride(usePower2Stride),
+	mSubFrameWidth(0),
+	mSubFrameHeight(0),
+	mSubFrameOffsetX(0),
+	mSubFrameOffsetY(0),
+	mAudioGain(1),
+	mRequestedOutputMode(output_mode),
+	mAutoRestart(0),
+	mEndOfFile(0),
+	mRestarted(0),
+	mIteration(0),
+    mPlaybackIteration(0),
+	mStream(0),
+	mThreadAccessCount(0),
+	mPriority(1),
+	mFirstFrameDisplayed(0),
+	mWaitingForCache(false),
+	mOutputMode(TH_UNDEFINED)
+{
+	mAudioMutex = NULL;
+	mThreadAccessMutex = new TheoraMutex();
+	mTimer = mDefaultTimer = new TheoraTimer();
+
+	mFrameQueue = NULL;
+	mAssignedWorkerThread = NULL;
+	mNumPrecachedFrames = nPrecachedFrames;
+	setOutputMode(output_mode);
+}
+
+TheoraVideoClip::~TheoraVideoClip()
+{
+	// wait untill a worker thread is done decoding the frame
+	mThreadAccessMutex->lock();
+
+	delete mDefaultTimer;
+
+	if (mStream) memdelete(mStream);
+
+	if (mFrameQueue) delete mFrameQueue;
+
+	if (mAudioInterface)
+	{
+		mAudioMutex->lock(); // ensure a thread isn't using this mutex
+		memdelete(mAudioInterface); // notify audio interface it's time to call it a day
+		mAudioMutex ->unlock();
+		delete mAudioMutex;
+	}
+	
+	mThreadAccessMutex->unlock();
+
+	delete mThreadAccessMutex;
+}
+
+TheoraTimer* TheoraVideoClip::getTimer()
+{
+	return mTimer;
+}
+
+void TheoraVideoClip::setTimer(TheoraTimer* timer)
+{
+	if (!timer) mTimer = mDefaultTimer;
+	else mTimer = timer;
+}
+
+void TheoraVideoClip::resetFrameQueue()
+{
+	mFrameQueue->clear();
+    mPlaybackIteration = mIteration = 0;
+}
+
+void TheoraVideoClip::restart()
+{
+	mEndOfFile = true; //temp, to prevent threads to decode while restarting
+	mThreadAccessMutex->lock();
+	_restart();
+	mTimer->seek(0);
+	mFirstFrameDisplayed = false;
+    resetFrameQueue();
+	mEndOfFile = false;
+	mRestarted = false;
+	mSeekFrame = -1;
+	mThreadAccessMutex->unlock();
+}
+
+void TheoraVideoClip::update(float timeDelta)
+{
+	if (mTimer->isPaused())
+	{
+		mTimer->update(0); // update timer in case there is some code that needs to execute each frame
+		return;
+	}
+	float time = mTimer->getTime(), speed = mTimer->getSpeed();
+    if (time + timeDelta * speed >= mDuration)
+    {
+        if (mAutoRestart && mRestarted)
+        {
+            float seekTime = time + timeDelta * speed;
+            for (;seekTime >= mDuration;)
+            {
+                seekTime -= mDuration;
+                ++mPlaybackIteration;
+            }
+
+            mTimer->seek(seekTime);
+        }
+        else
+        {
+            if (time != mDuration)
+            {
+                mTimer->update((mDuration - time) / speed);
+            }
+        }
+    }
+    else
+    {
+        mTimer->update(timeDelta);
+    }
+}
+
+float TheoraVideoClip::updateToNextFrame()
+{
+	TheoraVideoFrame* f = mFrameQueue->getFirstAvailableFrame();
+	if (!f) return 0;
+
+	float time = f->mTimeToDisplay - mTimer->getTime();
+	update(time);
+	return time;
+}
+
+TheoraFrameQueue* TheoraVideoClip::getFrameQueue()
+{
+	return mFrameQueue;
+}
+
+void TheoraVideoClip::popFrame()
+{
+	++mNumDisplayedFrames;
+	
+    // after transfering frame data to the texture, free the frame
+	// so it can be used again
+	if (!mFirstFrameDisplayed)
+	{
+		mFrameQueue->lock();
+		mFrameQueue->_pop(1);
+		mFirstFrameDisplayed = true;
+		mFrameQueue->unlock();
+	}
+	else
+	{
+		mFrameQueue->pop();
+	}
+}
+
+int TheoraVideoClip::getWidth()
+{
+	return mUseAlpha ? mWidth / 2 : mWidth;
+}
+
+int TheoraVideoClip::getHeight()
+{
+	return mHeight;
+}
+
+int TheoraVideoClip::getSubFrameWidth()
+{
+	return mUseAlpha ? mWidth / 2 : mSubFrameWidth;
+}
+
+int TheoraVideoClip::getSubFrameHeight()
+{
+	return mUseAlpha ? mHeight : mSubFrameHeight;
+}
+
+int TheoraVideoClip::getSubFrameOffsetX()
+{
+	return mUseAlpha ? 0 : mSubFrameOffsetX;
+}
+
+int TheoraVideoClip::getSubFrameOffsetY()
+{
+	return mUseAlpha ? 0 : mSubFrameOffsetY;
+}
+
+float TheoraVideoClip::getAbsPlaybackTime()
+{
+    return mTimer->getTime() + mPlaybackIteration * mDuration;
+}
+
+int TheoraVideoClip::discardOutdatedFrames(float absTime)
+{
+    int nReady = mFrameQueue->_getReadyCount();
+    // only drop frames if you have more frames to show. otherwise even the late frame will do..
+    if (nReady == 1) return 0;
+    float time = absTime;
+
+    int nPop = 0;
+    TheoraVideoFrame* frame;
+    float timeToDisplay;
+    
+    std::list<TheoraVideoFrame*>& queue = mFrameQueue->_getFrameQueue();
+    foreach_l (TheoraVideoFrame*, queue)
+    {
+        frame = *it;
+        if (!frame->mReady) break;
+        timeToDisplay = frame->mTimeToDisplay + frame->mIteration * mDuration;
+        if (time > timeToDisplay + mFrameDuration)
+        {
+            ++nPop;
+            if (nReady - nPop == 1) break; // always leave at least one in the queue
+        }
+        else break;
+    }
+    
+	if (nPop > 0)
+    {
+#ifdef _DEBUG
+        std::string log = getName() + ": dropped frame ";
+    
+        int i = nPop;
+        foreach_l (TheoraVideoFrame*, queue)
+        {
+            log += str((int) (*it)->getFrameNumber());
+            if (i-- > 1)
+            {
+                log += ", ";
+            }
+            else break;
+        }
+        th_writelog(log);
+#endif
+        mNumDroppedFrames += nPop;
+        mFrameQueue->_pop(nPop);
+	}
+    
+    return nPop;
+}
+
+TheoraVideoFrame* TheoraVideoClip::getNextFrame()
+{
+	TheoraVideoFrame* frame;
+    // if we are about to seek, then the current frame queue is invalidated
+	// (will be cleared when a worker thread does the actual seek)
+    if (mSeekFrame != -1) return NULL;
+
+    mFrameQueue->lock();
+	float time = getAbsPlaybackTime();
+    discardOutdatedFrames(time);
+    
+    frame = mFrameQueue->_getFirstAvailableFrame();
+    if (frame != NULL)
+    {
+        if (frame->mTimeToDisplay + frame->mIteration * mDuration > time && mFirstFrameDisplayed)
+        {
+            frame = NULL; // frame is ready but it's not yet time to display it, except when we haven't displayed any frames yet
+        }
+    }
+
+    mFrameQueue->unlock();
+	return frame;
+}
+
+std::string TheoraVideoClip::getName()
+{
+	return mName;
+}
+
+bool TheoraVideoClip::isBusy()
+{
+	return mAssignedWorkerThread || mOutputMode != mRequestedOutputMode;
+}
+
+TheoraOutputMode TheoraVideoClip::getOutputMode()
+{
+	return mOutputMode;
+}
+
+void TheoraVideoClip::setOutputMode(TheoraOutputMode mode)
+{
+	if (mode == TH_UNDEFINED) throw TheoraGenericException("Invalid output mode: TH_UNDEFINED for video: " + mName);
+	if (mOutputMode == mode) return;
+	mRequestedOutputMode = mode;
+	mUseAlpha = (mode == TH_RGBA   ||
+				 mode == TH_ARGB   ||
+				 mode == TH_BGRA   ||
+				 mode == TH_ABGR   ||
+				 mode == TH_GREY3A ||
+				 mode == TH_AGREY3 ||
+				 mode == TH_YUVA   ||
+				 mode == TH_AYUV);
+	if (mAssignedWorkerThread)
+	{
+		mThreadAccessMutex->lock();
+		// discard current frames and recreate them
+		mFrameQueue->setSize(mFrameQueue->getSize());
+		mThreadAccessMutex->unlock();
+
+	}
+	mOutputMode = mRequestedOutputMode;
+}
+
+float TheoraVideoClip::getTimePosition()
+{
+	return mTimer->getTime();
+}
+
+int TheoraVideoClip::getNumPrecachedFrames()
+{
+	return mFrameQueue->getSize();
+}
+
+void TheoraVideoClip::setNumPrecachedFrames(int n)
+{
+	if (mFrameQueue->getSize() != n)
+		mFrameQueue->setSize(n);
+}
+
+int TheoraVideoClip::_getNumReadyFrames()
+{
+	if (mSeekFrame != -1) return 0;
+	return mFrameQueue->_getReadyCount();
+}
+
+int TheoraVideoClip::getNumReadyFrames()
+{
+	if (mSeekFrame != -1) return 0; // we are about to seek, consider frame queue empty even though it will be emptied upon seek
+	return mFrameQueue->getReadyCount();
+}
+
+float TheoraVideoClip::getDuration()
+{
+	return mDuration;
+}
+
+float TheoraVideoClip::getFPS()
+{
+	return mFPS;
+}
+
+void TheoraVideoClip::play()
+{
+	mTimer->play();
+}
+
+void TheoraVideoClip::pause()
+{
+	mTimer->pause();
+}
+
+bool TheoraVideoClip::isPaused()
+{
+	return mTimer->isPaused();
+}
+
+bool TheoraVideoClip::isDone()
+{
+	return mEndOfFile && !mFrameQueue->getFirstAvailableFrame();
+}
+
+void TheoraVideoClip::stop()
+{
+	pause();
+    resetFrameQueue();
+	mFirstFrameDisplayed = false;
+	seek(0);
+}
+
+void TheoraVideoClip::setPlaybackSpeed(float speed)
+{
+	mTimer->setSpeed(speed);
+}
+
+float TheoraVideoClip::getPlaybackSpeed()
+{
+	return mTimer->getSpeed();
+}
+
+void TheoraVideoClip::seek(float time)
+{
+	seekToFrame((int) (time * getFPS()));
+}
+
+void TheoraVideoClip::seekToFrame(int frame)
+{
+	if      (frame < 0)          mSeekFrame = 0;
+	else if (frame > mNumFrames) mSeekFrame = mNumFrames;
+	else                         mSeekFrame = frame;
+
+	mFirstFrameDisplayed = false;
+	mEndOfFile = false;
+}
+
+void TheoraVideoClip::waitForCache(float desired_cache_factor, float max_wait_time)
+{
+	mWaitingForCache = true;
+	bool paused = mTimer->isPaused();
+	if (!paused) mTimer->pause();
+	int elapsed = 0;
+	int desired_num_precached_frames = (int) (desired_cache_factor * getNumPrecachedFrames());
+	while (getNumReadyFrames() < desired_num_precached_frames)
+	{
+		_psleep(10);
+		elapsed += 10;
+		if (elapsed >= max_wait_time * 1000) break;
+	}
+	if (!paused) mTimer->play();
+	mWaitingForCache = false;
+}
+
+float TheoraVideoClip::getPriority()
+{
+	return mPriority;
+}
+
+void TheoraVideoClip::setPriority(float priority)
+{
+	mPriority = priority;
+}
+
+float TheoraVideoClip::getPriorityIndex()
+{
+	float priority = (float) getNumReadyFrames();
+	if (mTimer->isPaused()) priority += getNumPrecachedFrames() / 2;
+	
+	return priority;
+}
+
+void TheoraVideoClip::setAudioInterface(TheoraAudioInterface* iface)
+{
+	mAudioInterface = iface;
+	if (iface && !mAudioMutex) mAudioMutex = new TheoraMutex;
+	if (!iface && mAudioMutex)
+	{
+		delete mAudioMutex;
+		mAudioMutex = NULL;
+	}
+}
+
+TheoraAudioInterface* TheoraVideoClip::getAudioInterface()
+{
+	return mAudioInterface;
+}
+
+void TheoraVideoClip::setAudioGain(float gain)
+{
+	if (gain > 1) mAudioGain=1;
+	if (gain < 0) mAudioGain=0;
+	else          mAudioGain=gain;
+}
+
+float TheoraVideoClip::getAudioGain()
+{
+	return mAudioGain;
+}
+
+void TheoraVideoClip::setAutoRestart(bool value)
+{
+	mAutoRestart = value;
+	if (value) mEndOfFile = false;
+}
diff --git a/drivers/theoraplayer/src/TheoraVideoFrame.cpp b/drivers/theoraplayer/src/TheoraVideoFrame.cpp
new file mode 100644
index 0000000000..b70253dabf
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraVideoFrame.cpp
@@ -0,0 +1,159 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include <memory.h>
+#include "TheoraPixelTransform.h"
+#include "TheoraVideoClip.h"
+#include "TheoraVideoFrame.h"
+#include "TheoraVideoManager.h"
+
+//#define YUV_TEST // uncomment this if you want to benchmark YUV decoding functions
+
+extern "C"
+{
+void decodeRGB  (struct TheoraPixelTransform* t);
+void decodeRGBA (struct TheoraPixelTransform* t);
+void decodeRGBX (struct TheoraPixelTransform* t);
+void decodeARGB (struct TheoraPixelTransform* t);
+void decodeXRGB (struct TheoraPixelTransform* t);
+void decodeBGR  (struct TheoraPixelTransform* t);
+void decodeBGRA (struct TheoraPixelTransform* t);
+void decodeBGRX (struct TheoraPixelTransform* t);
+void decodeABGR (struct TheoraPixelTransform* t);
+void decodeXBGR (struct TheoraPixelTransform* t);
+void decodeGrey (struct TheoraPixelTransform* t);
+void decodeGrey3(struct TheoraPixelTransform* t);
+void decodeGreyA(struct TheoraPixelTransform* t);
+void decodeGreyX(struct TheoraPixelTransform* t);
+void decodeAGrey(struct TheoraPixelTransform* t);
+void decodeXGrey(struct TheoraPixelTransform* t);
+void decodeYUV  (struct TheoraPixelTransform* t);
+void decodeYUVA (struct TheoraPixelTransform* t);
+void decodeYUVX (struct TheoraPixelTransform* t);
+void decodeAYUV (struct TheoraPixelTransform* t);
+void decodeXYUV (struct TheoraPixelTransform* t);
+}
+
+static void (*conversion_functions[])(struct TheoraPixelTransform*) = {0,
+	decodeRGB,
+	decodeRGBA,
+	decodeRGBX,
+	decodeARGB,
+	decodeXRGB,
+	decodeBGR,
+	decodeBGRA,
+	decodeBGRX,
+	decodeABGR,
+	decodeXBGR,
+	decodeGrey,
+	decodeGrey3,
+	decodeGreyA,
+	decodeGreyX,
+	decodeAGrey,
+	decodeXGrey,
+	decodeYUV,
+	decodeYUVA,
+	decodeYUVX,
+	decodeAYUV,
+	decodeXYUV
+};
+
+TheoraVideoFrame::TheoraVideoFrame(TheoraVideoClip* parent)
+{
+	mReady = mInUse = false;
+	mParent = parent;
+	mIteration = 0;
+	// number of bytes based on output mode
+	int bytemap[]={0, 3, 4, 4, 4, 4, 3, 4, 4, 4, 4, 1, 3, 4, 4, 4, 4, 3, 4, 4, 4, 4};
+	mBpp = bytemap[mParent->getOutputMode()];
+	unsigned int size = mParent->getStride() * mParent->mHeight * mBpp;
+	try
+	{
+		mBuffer = new unsigned char[size];
+	}
+	catch (std::bad_alloc)
+	{
+		mBuffer = NULL;
+		return;
+	}
+	memset(mBuffer, 255, size);
+}
+
+TheoraVideoFrame::~TheoraVideoFrame()
+{
+	if (mBuffer) delete [] mBuffer;
+}
+
+int TheoraVideoFrame::getWidth()
+{
+	return mParent->getWidth();
+}
+
+int TheoraVideoFrame::getStride()
+{
+	return mParent->mStride;
+}
+
+int TheoraVideoFrame::getHeight()
+{
+	return mParent->getHeight();
+}
+
+unsigned char* TheoraVideoFrame::getBuffer()
+{
+	return mBuffer;
+}
+
+void TheoraVideoFrame::decode(struct TheoraPixelTransform* t)
+{
+	if (t->raw != NULL)
+	{
+		int bufferStride = mParent->getWidth() * mBpp;
+		if (bufferStride == t->rawStride)
+		{
+			memcpy(mBuffer, t->raw, t->rawStride * mParent->getHeight());
+		}
+		else
+		{
+			unsigned char *buff = mBuffer, *src = t->raw;
+			int i, h = mParent->getHeight();
+			for (i = 0; i < h; ++i, buff += bufferStride, src += t->rawStride)
+			{
+				memcpy(buff, src, bufferStride);
+			}
+		}
+	}
+	else
+	{
+		t->out = mBuffer;
+		t->w = mParent->getWidth();
+		t->h = mParent->getHeight();
+        
+#ifdef YUV_TEST // when benchmarking yuv conversion functions during development, do a timed average
+        #define N 1000
+        clock_t time = clock();
+        for (int i = 0; i < N; ++i)
+        {
+            conversion_functions[mParent->getOutputMode()](t);
+        }
+        float diff = (clock() - time) * 1000.0f / CLOCKS_PER_SEC;
+        
+		char s[128];
+		sprintf(s, "%.2f", diff / N);
+        TheoraVideoManager::getSingleton().logMessage("YUV Decoding time: " + std::string(s) + " ms\n");
+#else
+		conversion_functions[mParent->getOutputMode()](t);
+#endif
+	}
+	mReady = true;
+}
+
+void TheoraVideoFrame::clear()
+{
+	mInUse = mReady = false;
+}
diff --git a/drivers/theoraplayer/src/TheoraVideoManager.cpp b/drivers/theoraplayer/src/TheoraVideoManager.cpp
new file mode 100644
index 0000000000..87696d12a9
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraVideoManager.cpp
@@ -0,0 +1,479 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "TheoraVideoManager.h"
+#include "TheoraWorkerThread.h"
+#include "TheoraVideoClip.h"
+#include "TheoraFrameQueue.h"
+#include "TheoraAudioInterface.h"
+#include "TheoraUtil.h"
+#include "TheoraDataSource.h"
+#include "TheoraException.h"
+#ifdef __THEORA
+	#include <theora/codec.h>
+	#include <vorbis/codec.h>
+	#include "TheoraVideoClip_Theora.h"
+#endif
+#ifdef __AVFOUNDATION
+	#include "TheoraVideoClip_AVFoundation.h"
+#endif
+#ifdef __FFMPEG
+	#include "TheoraVideoClip_FFmpeg.h"
+#endif
+#ifdef _ANDROID //libtheoraplayer addition for cpu feature detection
+	#include "cpu-features.h"
+#endif
+// declaring function prototype here so I don't have to put it in a header file
+// it only needs to be used by this plugin and called once
+extern "C"
+{
+	void initYUVConversionModule();
+}
+
+//#define _DECODING_BENCHMARK //uncomment to test average decoding time on a given device
+
+
+// --------------------------
+//#define _SCHEDULING_DEBUG
+#ifdef _SCHEDULING_DEBUG
+float gThreadDiagnosticTimer = 0;
+#endif
+// --------------------------
+
+#ifdef _DECODING_BENCHMARK
+void benchmark(TheoraVideoClip* clip)
+{
+	int nPrecached = 256;
+	int n = nPrecached;
+	char msg[1024];
+	clock_t t = clock();
+	while (n > 0)
+	{
+		clip->waitForCache(1.0f, 1000000);
+		n -= 32;
+		clip->getFrameQueue()->clear();
+	}
+	float diff = ((float) (clock() - t) * 1000.0f) / CLOCKS_PER_SEC;
+	sprintf(msg, "BENCHMARK: %s: Decoding %d frames took %.1fms (%.2fms average per frame)\n",clip->getName().c_str(), nPrecached, diff, diff / nPrecached);
+	TheoraVideoManager::getSingleton().logMessage(msg);
+	clip->seek(0);
+}
+#endif
+
+struct TheoraWorkCandidate
+{
+	TheoraVideoClip* clip;
+	float priority, queuedTime, workTime, entitledTime;
+};
+
+TheoraVideoManager* g_ManagerSingleton = NULL;
+
+void theora_writelog(std::string output)
+{
+	printf("%s\n", output.c_str());
+}
+
+void (*g_LogFuction)(std::string) = theora_writelog;
+
+void TheoraVideoManager::setLogFunction(void (*fn)(std::string))
+{
+	g_LogFuction = fn;
+}
+
+TheoraVideoManager* TheoraVideoManager::getSingletonPtr()
+{
+    return g_ManagerSingleton;
+}
+
+TheoraVideoManager& TheoraVideoManager::getSingleton()
+{  
+    return *g_ManagerSingleton;  
+}
+
+TheoraVideoManager::TheoraVideoManager(int num_worker_threads) : 
+	mDefaultNumPrecachedFrames(8)
+{
+	if (num_worker_threads < 1) throw TheoraGenericException("Unable to create TheoraVideoManager, at least one worker thread is reqired");
+
+	g_ManagerSingleton = this;
+
+	std::string msg = "Initializing Theora Playback Library (" + getVersionString() + ")\n";
+#ifdef __THEORA
+	msg += "  - libtheora version: " + std::string(th_version_string()) + "\n" +
+	       "  - libvorbis version: " +  std::string(vorbis_version_string()) + "\n";
+#endif
+#ifdef _ANDROID
+	uint64_t features = android_getCpuFeaturesExt();
+	char s[128];
+	sprintf(s, "  - Android: CPU Features: %u\n", (unsigned int) features);
+	msg += s;
+	if ((features & ANDROID_CPU_ARM_FEATURE_NEON) == 0)
+		msg += "  - Android: NEON features NOT SUPPORTED by CPU\n";
+	else
+		msg += "  - Android: Detected NEON CPU features\n";
+#endif
+
+#ifdef __AVFOUNDATION
+	msg += "  - using Apple AVFoundation classes.\n";
+#endif
+#ifdef __FFMPEG
+	msg += "  - using FFmpeg library.\n";
+#endif
+	
+	logMessage(msg + "------------------------------------");
+	mAudioFactory = NULL;
+	mWorkMutex = new TheoraMutex();
+
+	// for CPU based yuv2rgb decoding
+	initYUVConversionModule();
+
+	createWorkerThreads(num_worker_threads);
+}
+
+TheoraVideoManager::~TheoraVideoManager()
+{
+	destroyWorkerThreads();
+
+	mWorkMutex->lock();
+	ClipList::iterator ci;
+	for (ci = mClips.begin(); ci != mClips.end(); ++ci)
+		delete (*ci);
+	mClips.clear();
+	mWorkMutex->unlock();
+	delete mWorkMutex;
+}
+
+void TheoraVideoManager::logMessage(std::string msg)
+{
+	g_LogFuction(msg);
+}
+
+TheoraVideoClip* TheoraVideoManager::getVideoClipByName(std::string name)
+{
+	TheoraVideoClip* clip = NULL;
+	mWorkMutex->lock();
+
+	foreach(TheoraVideoClip*, mClips)
+	{
+		if ((*it)->getName() == name)
+		{
+			clip = *it;
+			break;
+		}
+	}
+	mWorkMutex->unlock();
+
+	return clip;
+}
+
+void TheoraVideoManager::setAudioInterfaceFactory(TheoraAudioInterfaceFactory* factory)
+{
+	mAudioFactory = factory;
+}
+
+TheoraAudioInterfaceFactory* TheoraVideoManager::getAudioInterfaceFactory()
+{
+	return mAudioFactory;
+}
+
+TheoraVideoClip* TheoraVideoManager::createVideoClip(std::string filename,
+													 TheoraOutputMode output_mode,
+													 int numPrecachedOverride,
+													 bool usePower2Stride)
+{
+	TheoraDataSource* src=new TheoraFileDataSource(filename);
+	return createVideoClip(src,output_mode,numPrecachedOverride,usePower2Stride);
+}
+
+TheoraVideoClip* TheoraVideoManager::createVideoClip(TheoraDataSource* data_source,
+													 TheoraOutputMode output_mode,
+													 int numPrecachedOverride,
+													 bool usePower2Stride)
+{
+	mWorkMutex->lock();
+
+	TheoraVideoClip* clip = NULL;
+	int nPrecached = numPrecachedOverride ? numPrecachedOverride : mDefaultNumPrecachedFrames;
+	logMessage("Creating video from data source: " + data_source->repr() + " [" + str(nPrecached) + " precached frames].");
+	
+#ifdef __AVFOUNDATION
+	TheoraFileDataSource* fileDataSource = dynamic_cast<TheoraFileDataSource*>(data_source);
+	std::string filename;
+	if (fileDataSource == NULL)
+	{
+		TheoraMemoryFileDataSource* memoryDataSource = dynamic_cast<TheoraMemoryFileDataSource*>(data_source);
+		if (memoryDataSource != NULL) filename = memoryDataSource->getFilename();
+		// if the user has his own data source, it's going to be a problem for AVAssetReader since it only supports reading from files...
+	}
+	else filename = fileDataSource->getFilename();
+
+	if (filename.size() > 4 && filename.substr(filename.size() - 4, filename.size()) == ".mp4")
+	{
+		clip = new TheoraVideoClip_AVFoundation(data_source, output_mode, nPrecached, usePower2Stride);
+	}
+#endif
+#if defined(__AVFOUNDATION) && defined(__THEORA)
+	else
+#endif
+#ifdef __THEORA
+		clip = new TheoraVideoClip_Theora(data_source, output_mode, nPrecached, usePower2Stride);
+#endif
+#ifdef __FFMPEG
+		clip = new TheoraVideoClip_FFmpeg(data_source, output_mode, nPrecached, usePower2Stride);
+#endif
+	clip->load(data_source);
+	clip->decodeNextFrame(); // ensure the first frame is always preloaded and have the main thread do it to prevent potential thread starvatio
+
+	mClips.push_back(clip);
+	mWorkMutex->unlock();
+	
+#ifdef _DECODING_BENCHMARK
+	benchmark(clip);
+#endif
+	return clip;
+}
+
+void TheoraVideoManager::destroyVideoClip(TheoraVideoClip* clip)
+{
+	if (clip)
+	{
+		th_writelog("Destroying video clip: " + clip->getName());
+		mWorkMutex->lock();
+		bool reported = 0;
+		while (clip->mAssignedWorkerThread)
+		{
+			if (!reported)
+			{
+				th_writelog(" - Waiting for WorkerThread to finish decoding in order to destroy");
+				reported = 1;
+			}
+			_psleep(1);
+		}
+		if (reported) th_writelog(" - WorkerThread done, destroying...");
+		
+		// erase the clip from the clip list
+		foreach (TheoraVideoClip*, mClips)
+		{
+			if ((*it) == clip)
+			{
+				mClips.erase(it);
+				break;
+			}
+		}
+		// remove all it's references from the work log
+		mWorkLog.remove(clip);
+
+		// delete the actual clip
+		delete clip;
+#ifdef _DEBUG
+		th_writelog("Destroyed video.");
+#endif
+		mWorkMutex->unlock();
+	}
+}
+
+TheoraVideoClip* TheoraVideoManager::requestWork(TheoraWorkerThread* caller)
+{
+	if (!mWorkMutex) return NULL;
+	mWorkMutex->lock();
+
+	TheoraVideoClip* selectedClip = NULL;
+	float maxQueuedTime = 0, totalAccessCount = 0, prioritySum = 0, diff, maxDiff = -1;
+	int nReadyFrames;
+	std::vector<TheoraWorkCandidate> candidates;
+	TheoraVideoClip* clip;
+	TheoraWorkCandidate candidate;
+
+	// first pass is for playing videos, but if no such videos are available for decoding
+	// paused videos are selected in the second pass.
+    // Note that paused videos that are waiting for cache are considered equal to playing
+    // videos in the scheduling context
+
+	for (int i = 0; i < 2 && candidates.size() == 0; ++i)
+	{
+		foreach (TheoraVideoClip*, mClips)
+		{
+			clip = *it;
+			if (clip->isBusy() || (i == 0 && clip->isPaused() && !clip->mWaitingForCache)) continue;
+			nReadyFrames = clip->getNumReadyFrames();
+			if (nReadyFrames == clip->getFrameQueue()->getSize()) continue;
+
+			candidate.clip = clip;
+			candidate.priority = clip->getPriority();
+			candidate.queuedTime = (float) nReadyFrames / (clip->getFPS() * clip->getPlaybackSpeed());
+			candidate.workTime = (float) clip->mThreadAccessCount;
+			
+			totalAccessCount += candidate.workTime;
+			if (maxQueuedTime < candidate.queuedTime) maxQueuedTime = candidate.queuedTime;
+
+			candidates.push_back(candidate);
+		}
+	}
+
+	// prevent division by zero
+	if (totalAccessCount == 0) totalAccessCount = 1;
+	if (maxQueuedTime == 0) maxQueuedTime = 1;
+
+	// normalize candidate values
+	foreach (TheoraWorkCandidate, candidates)
+	{
+		it->workTime /= totalAccessCount;
+		// adjust user priorities to favor clips that have fewer frames queued
+		it->priority *= 1.0f - (it->queuedTime / maxQueuedTime) * 0.5f;
+		prioritySum += it->priority;
+	}
+	foreach (TheoraWorkCandidate, candidates)
+	{
+		it->entitledTime = it->priority / prioritySum;
+	}
+
+	// now, based on how much access time has been given to each clip in the work log
+	// and how much time should be given to each clip based on calculated priorities,
+	// we choose a best suited clip for this worker thread to decode next
+	foreach (TheoraWorkCandidate, candidates)
+	{
+		diff = it->entitledTime - it->workTime;
+
+		if (maxDiff < diff)
+		{
+			maxDiff = diff;
+			selectedClip = it->clip;
+		}
+	}
+
+	if (selectedClip)
+	{
+		selectedClip->mAssignedWorkerThread = caller;
+		
+		int nClips = (int) mClips.size();
+		unsigned int maxWorkLogSize = (nClips - 1) * 50;
+
+		if (nClips > 1)
+		{
+			mWorkLog.push_front(selectedClip);
+			++selectedClip->mThreadAccessCount;
+		}
+		
+		TheoraVideoClip* c;
+		while (mWorkLog.size() > maxWorkLogSize)
+		{
+			c = mWorkLog.back();
+			mWorkLog.pop_back();
+			c->mThreadAccessCount--;
+		}
+#ifdef _SCHEDULING_DEBUG
+		if (mClips.size() > 1)
+		{
+			int accessCount = mWorkLog.size();
+			if (gThreadDiagnosticTimer > 2.0f)
+			{
+				gThreadDiagnosticTimer = 0;
+				std::string logstr = "-----\nTheora Playback Library debug CPU time analysis (" + str(accessCount) + "):\n";
+				int percent;
+				foreach (TheoraVideoClip*, mClips)
+				{
+					percent = ((float) (*it)->mThreadAccessCount / mWorkLog.size()) * 100.0f;
+					logstr += (*it)->getName() + " (" + str((*it)->getPriority()) + "): " + str((*it)->mThreadAccessCount) + ", " + str(percent) + "%\n";
+				}
+				logstr += "-----";
+				th_writelog(logstr);
+			}
+		}
+#endif
+	}
+
+	mWorkMutex->unlock();
+	return selectedClip;
+}
+
+void TheoraVideoManager::update(float timeDelta)
+{
+	mWorkMutex->lock();
+	foreach (TheoraVideoClip*, mClips)
+	{
+		(*it)->update(timeDelta);
+		(*it)->decodedAudioCheck();
+	}
+	mWorkMutex->unlock();
+#ifdef _SCHEDULING_DEBUG
+	gThreadDiagnosticTimer += timeDelta;
+#endif
+}
+
+int TheoraVideoManager::getNumWorkerThreads()
+{
+	return (int) mWorkerThreads.size();
+}
+
+void TheoraVideoManager::createWorkerThreads(int n)
+{
+	TheoraWorkerThread* t;
+	for (int i=0;i<n;++i)
+	{
+		t=new TheoraWorkerThread();
+		t->start();
+		mWorkerThreads.push_back(t);
+	}
+}
+
+void TheoraVideoManager::destroyWorkerThreads()
+{
+	foreach(TheoraWorkerThread*,mWorkerThreads)
+	{
+		(*it)->join();
+		delete (*it);
+	}
+	mWorkerThreads.clear();
+}
+
+void TheoraVideoManager::setNumWorkerThreads(int n)
+{
+	if (n == getNumWorkerThreads()) return;
+	if (n < 1) throw TheoraGenericException("Unable to change the number of worker threads in TheoraVideoManager, at least one worker thread is reqired");
+
+	th_writelog("changing number of worker threats to: "+str(n));
+
+	destroyWorkerThreads();
+	createWorkerThreads(n);
+}
+
+std::string TheoraVideoManager::getVersionString()
+{
+	int a, b, c;
+	getVersion(&a, &b, &c);
+	std::string out = str(a) + "." + str(b);
+	if (c != 0)
+	{
+		if (c < 0) out += " RC" + str(-c);
+		else       out += "." + str(c);
+	}
+	return out;
+}
+
+void TheoraVideoManager::getVersion(int* a, int* b, int* c) // TODO, return a struct instead of the current solution.
+{
+	*a = 1;
+	*b = 1;
+	*c = 0;
+}
+
+std::vector<std::string> TheoraVideoManager::getSupportedDecoders()
+{
+	std::vector<std::string> lst;
+#ifdef __THEORA
+	lst.push_back("Theora");
+#endif
+#ifdef __AVFOUNDATION
+	lst.push_back("AVFoundation");
+#endif
+#ifdef __FFMPEG
+	lst.push_back("FFmpeg");
+#endif
+	
+	return lst;
+}
diff --git a/drivers/theoraplayer/src/TheoraWorkerThread.cpp b/drivers/theoraplayer/src/TheoraWorkerThread.cpp
new file mode 100644
index 0000000000..cef8545b8d
--- /dev/null
+++ b/drivers/theoraplayer/src/TheoraWorkerThread.cpp
@@ -0,0 +1,49 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#ifdef _WIN32
+#pragma warning( disable: 4251 ) // MSVC++
+#endif
+#include "TheoraWorkerThread.h"
+#include "TheoraVideoManager.h"
+#include "TheoraVideoClip.h"
+#include "TheoraUtil.h"
+
+TheoraWorkerThread::TheoraWorkerThread() : TheoraThread()
+{
+	mClip = NULL;
+}
+
+TheoraWorkerThread::~TheoraWorkerThread()
+{
+
+}
+
+void TheoraWorkerThread::execute()
+{
+	while (isRunning())
+	{
+		mClip = TheoraVideoManager::getSingleton().requestWork(this);
+		if (!mClip)
+		{
+			_psleep(100);
+			continue;
+		}
+
+		mClip->mThreadAccessMutex->lock();
+		// if user requested seeking, do that then.
+		if (mClip->mSeekFrame >= 0) mClip->doSeek();
+
+		if (!mClip->decodeNextFrame())
+			_psleep(1); // this happens when the video frame queue is full.
+
+		mClip->mAssignedWorkerThread = NULL;
+		mClip->mThreadAccessMutex->unlock();
+		mClip = NULL;
+	}
+}
diff --git a/drivers/theoraplayer/src/YUV/C/yuv420_grey_c.c b/drivers/theoraplayer/src/YUV/C/yuv420_grey_c.c
new file mode 100644
index 0000000000..8af5dd1f58
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/C/yuv420_grey_c.c
@@ -0,0 +1,56 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "yuv_util.h"
+
+static void _decodeGrey3(struct TheoraPixelTransform* t, int stride, int nBytes)
+{
+	unsigned char *ySrc = t->y, *yLineEnd, *out = t->out;
+	unsigned int y;
+	for (y = 0; y < t->h; ++y, ySrc += t->yStride - t->w, out += stride-t->w * nBytes)
+		for (yLineEnd = ySrc + t->w; ySrc != yLineEnd; ++ySrc, out += nBytes)
+			out[0] = out[1] = out[2] = *ySrc;
+}
+
+void decodeGrey(struct TheoraPixelTransform* t)
+{
+	unsigned char *ySrc = t->y, *yLineEnd, *out = t->out;
+	unsigned int y;
+	for (y = 0; y < t->h; ++y, ySrc += t->yStride - t->w)
+		for (yLineEnd = ySrc + t->w; ySrc != yLineEnd; ++ySrc, ++out)
+			*out = *ySrc;
+
+}
+
+void decodeGrey3(struct TheoraPixelTransform* t)
+{
+	_decodeGrey3(t, t->w * 3, 3);
+}
+
+void decodeGreyA(struct TheoraPixelTransform* t)
+{
+	_decodeGrey3(t, t->w * 4, 4);
+	_decodeAlpha(incOut(t, 3), t->w * 4);
+}
+
+void decodeGreyX(struct TheoraPixelTransform* t)
+{
+	_decodeGrey3(t, t->w * 4, 4);
+}
+
+void decodeAGrey(struct TheoraPixelTransform* t)
+{
+	_decodeGrey3(incOut(t, 1), t->w * 4, 4);
+	_decodeAlpha(t, t->w * 4);
+}
+
+void decodeXGrey(struct TheoraPixelTransform* t)
+{
+	_decodeGrey3(incOut(t, 1), t->w * 4, 4);
+}
+
diff --git a/drivers/theoraplayer/src/YUV/C/yuv420_rgb_c.c b/drivers/theoraplayer/src/YUV/C/yuv420_rgb_c.c
new file mode 100644
index 0000000000..e981e75ead
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/C/yuv420_rgb_c.c
@@ -0,0 +1,358 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#ifdef _YUV_C
+#include "yuv_util.h"
+
+int YTable [256];
+int BUTable[256];
+int GUTable[256];
+int GVTable[256];
+int RVTable[256];
+
+#define CLIP_RGB_COLOR(dst, x) \
+	tmp = (x) >> 13;\
+	if ((tmp & ~0xFF) == 0) dst = tmp;\
+	else                    dst = (-tmp) >> 31;
+
+#define _decodeRGB(t, stride, nBytes, maxWidth, i1, i2, i3, j1, j2, j3)\
+	register int tmp;\
+	int nBytes2 = nBytes * 2, cv, cu, rgbY1, rgbY2, rgbY3, rgbY4, rV, gUV, bU, width = maxWidth == 0 ? t->w : maxWidth;\
+	unsigned int y;\
+	unsigned char *ySrcEven, *ySrcOdd, *yLineEnd, *uSrc, *vSrc, *out1, *out2;\
+	\
+	for (y = 0; y < t->h; y += 2)\
+	{\
+		ySrcEven = t->y + y * t->yStride;\
+		ySrcOdd  = t->y + (y + 1) * t->yStride;\
+		uSrc = t->u + y * t->uStride / 2;\
+		vSrc = t->v + y * t->vStride / 2;\
+		out1 = t->out + y * stride;\
+		out2 = t->out + (y + 1) * stride;\
+		\
+		for (yLineEnd = ySrcEven + width; ySrcEven != yLineEnd;)\
+		{\
+			cu = *uSrc; ++uSrc;\
+			cv = *vSrc; ++vSrc;\
+			rV   = RVTable[cv];\
+			gUV  = GUTable[cu] + GVTable[cv];\
+			bU   = BUTable[cu];\
+			\
+			rgbY1 = YTable[*ySrcEven]; ++ySrcEven;\
+			rgbY2 = YTable[*ySrcOdd];  ++ySrcOdd;\
+			rgbY3 = YTable[*ySrcEven]; ++ySrcEven;\
+			rgbY4 = YTable[*ySrcOdd];  ++ySrcOdd;\
+			\
+			CLIP_RGB_COLOR(out1[i1], rgbY1 + rV );\
+			CLIP_RGB_COLOR(out1[i2], rgbY1 - gUV);\
+			CLIP_RGB_COLOR(out1[i3], rgbY1 + bU );\
+			\
+			CLIP_RGB_COLOR(out2[i1], rgbY2 + rV );\
+			CLIP_RGB_COLOR(out2[i2], rgbY2 - gUV);\
+			CLIP_RGB_COLOR(out2[i3], rgbY2 + bU );\
+			\
+			CLIP_RGB_COLOR(out1[j1], rgbY3 + rV );\
+			CLIP_RGB_COLOR(out1[j2], rgbY3 - gUV);\
+			CLIP_RGB_COLOR(out1[j3], rgbY3 + bU );\
+			\
+			CLIP_RGB_COLOR(out2[j1], rgbY4 + rV );\
+			CLIP_RGB_COLOR(out2[j2], rgbY4 - gUV);\
+			CLIP_RGB_COLOR(out2[j3], rgbY4 + bU );\
+			\
+			out1 += nBytes2;  out2 += nBytes2;\
+		}\
+	}
+
+// The 'trick' with this function is that it skips decoding YUV pixels if the alpha value is 0, thus improving the decoding speed of a frame
+#define _decodeRGBA(t, stride, nBytes, maxWidth, i1, i2, i3, j1, j2, j3, aindex1, aindex2)\
+\
+	register int tmp;\
+	int nBytes2 = nBytes * 2, cv, cu, rgbY1, rgbY2, rgbY3, rgbY4, a1, a2, a3, a4, rV, gUV, bU, width = maxWidth == 0 ? t->w : maxWidth;\
+	int alphaStride = t->w;\
+	unsigned int y;\
+	unsigned char *ySrcEven, *ySrcOdd, *yLineEnd, *uSrc, *vSrc, *out1, *out2;\
+	\
+	for (y = 0; y < t->h; y += 2)\
+	{\
+		ySrcEven = t->y + y * t->yStride;\
+		ySrcOdd  = t->y + (y + 1) * t->yStride;\
+		uSrc = t->u + y * t->uStride / 2;\
+		vSrc = t->v + y * t->vStride / 2;\
+		out1 = t->out + y * stride;\
+		out2 = t->out + (y + 1) * stride;\
+		\
+		for (yLineEnd = ySrcEven + width; ySrcEven != yLineEnd;)\
+		{\
+			cu = *uSrc; ++uSrc;\
+			cv = *vSrc; ++vSrc;\
+			rV   = RVTable[cv];\
+			gUV  = GUTable[cu] + GVTable[cv];\
+			bU   = BUTable[cu];\
+			\
+			rgbY1 = YTable[*ySrcEven]; a1 = ySrcEven[alphaStride]; ++ySrcEven;\
+			rgbY2 = YTable[*ySrcOdd];  a2 = ySrcOdd [alphaStride];  ++ySrcOdd;\
+			rgbY3 = YTable[*ySrcEven]; a3 = ySrcEven[alphaStride]; ++ySrcEven;\
+			rgbY4 = YTable[*ySrcOdd];  a4 = ySrcOdd [alphaStride];  ++ySrcOdd;\
+			\
+			if (a1 > 16)\
+			{\
+				CLIP_RGB_COLOR(out1[i1], rgbY1 + rV );\
+				CLIP_RGB_COLOR(out1[i2], rgbY1 - gUV);\
+				CLIP_RGB_COLOR(out1[i3], rgbY1 + bU );\
+				out1[aindex1] = a1 >= 235 ? 255 : (unsigned char) (((a1 - 16) * 255) / 219);\
+			}\
+			else *((unsigned int*) out1) = 0;\
+			\
+			if (a2 > 16)\
+			{\
+				CLIP_RGB_COLOR(out2[i1], rgbY2 + rV );\
+				CLIP_RGB_COLOR(out2[i2], rgbY2 - gUV);\
+				CLIP_RGB_COLOR(out2[i3], rgbY2 + bU );\
+				out2[aindex1] = a2 >= 235 ? 255 : (unsigned char) (((a2 - 16) * 255) / 219);\
+			}\
+			else *((unsigned int*) out2) = 0;\
+			\
+			if (a3 > 16)\
+			{\
+				CLIP_RGB_COLOR(out1[j1], rgbY3 + rV );\
+				CLIP_RGB_COLOR(out1[j2], rgbY3 - gUV);\
+				CLIP_RGB_COLOR(out1[j3], rgbY3 + bU );\
+				out1[aindex2] = a3 >= 235 ? 255 : (unsigned char) (((a3 - 16) * 255) / 219);\
+			}\
+			else *((unsigned int*) &out1[4]) = 0;\
+			\
+			if (a4 > 16)\
+			{\
+				CLIP_RGB_COLOR(out2[j1], rgbY4 + rV );\
+				CLIP_RGB_COLOR(out2[j2], rgbY4 - gUV);\
+				CLIP_RGB_COLOR(out2[j3], rgbY4 + bU );\
+				out2[aindex2] = a4 >= 235 ? 255 : (unsigned char) (((a4 - 16) * 255) / 219);\
+			}\
+			else *((unsigned int*) &out2[4]) = 0;\
+			\
+			out1 += nBytes2;  out2 += nBytes2;\
+		}\
+	}\
+
+void decodeRGB(struct TheoraPixelTransform* t)
+{
+	_decodeRGB(t, t->w * 3, 3, 0, 0, 1, 2, 3, 4, 5);
+}
+
+void decodeRGBA(struct TheoraPixelTransform* t)
+{
+	_decodeRGBA(t, t->w * 4, 4, 0, 0, 1, 2, 4, 5, 6, 3, 7);
+// This is the old 2-phase version, leaving it here in case more debugging is needed
+//	_decodeRGB(t, t->w * 4, 4, 0, 0, 1, 2, 4, 5, 6);
+//	_decodeAlpha(incOut(t, 3), t->w * 4);
+}
+
+void decodeRGBX(struct TheoraPixelTransform* t)
+{
+	_decodeRGB(t, t->w * 4, 4, 0, 0, 1, 2, 4, 5, 6);
+}
+
+void decodeARGB(struct TheoraPixelTransform* t)
+{
+	_decodeRGBA(t, t->w * 4, 4, 0, 1, 2, 3, 5, 6, 7, 0, 4);
+// This is the old 2-phase version, leaving it here in case more debugging is needed
+//	_decodeRGB(t, t->w * 4, 4, 0, 1, 2, 3, 5, 6, 7);
+//	_decodeAlpha(t, t->w * 4);
+}
+
+void decodeXRGB(struct TheoraPixelTransform* t)
+{
+	_decodeRGB(t, t->w * 4, 4, 0, 1, 2, 3, 5, 6, 7);
+}
+
+void decodeBGR(struct TheoraPixelTransform* t)
+{
+	_decodeRGB(t, t->w * 3, 3, 0, 2, 1, 0, 5, 4, 3);
+}
+
+void decodeBGRA(struct TheoraPixelTransform* t)
+{
+	_decodeRGBA(t, t->w * 4, 4, 0, 2, 1, 0, 6, 5, 4, 3, 7);
+// This is the old 2-phase version, leaving it here in case more debugging is needed
+//	_decodeRGB(t, t->w * 4, 4, 0, 2, 1, 0, 6, 5, 4);
+//	_decodeAlpha(incOut(t, 3), t->w * 4);
+}
+
+void decodeBGRX(struct TheoraPixelTransform* t)
+{
+	_decodeRGB(t, t->w * 4, 4, 0, 2, 1, 0, 6, 5, 4);
+}
+
+void decodeABGR(struct TheoraPixelTransform* t)
+{
+	_decodeRGBA(t, t->w * 4, 4, 0, 3, 2, 1, 7, 6, 5, 0, 4);
+// This is the old 2-phase version, leaving it here in case more debugging is needed
+//	_decodeRGB(t, t->w * 4, 4, 0, 3, 2, 1, 7, 6, 5);
+//	_decodeAlpha(t, t->w * 4);
+}
+
+void decodeXBGR(struct TheoraPixelTransform* t)
+{
+	_decodeRGB(t, t->w * 4, 4, 0, 3, 2, 1, 7, 6, 5);
+}
+
+void initYUVConversionModule()
+{
+	//used to bring the table into the high side (scale up) so we
+	//can maintain high precision and not use floats (FIXED POINT)
+	
+	// this is the pseudocode for yuv->rgb conversion
+	//        r = 1.164*(*ySrc - 16) + 1.596*(cv - 128);
+	//        b = 1.164*(*ySrc - 16)                   + 2.018*(cu - 128);
+	//        g = 1.164*(*ySrc - 16) - 0.813*(cv - 128) - 0.391*(cu - 128);
+	
+    double scale = 1L << 13, temp;
+	
+	int i;
+	for (i = 0; i < 256; ++i)
+	{
+		temp = i - 128;
+		
+		YTable[i]  = (int)((1.164 * scale + 0.5) * (i - 16));	//Calc Y component
+		RVTable[i] = (int)((1.596 * scale + 0.5) * temp);		//Calc R component
+		GUTable[i] = (int)((0.391 * scale + 0.5) * temp);		//Calc G u & v components
+		GVTable[i] = (int)((0.813 * scale + 0.5) * temp);
+		BUTable[i] = (int)((2.018 * scale + 0.5) * temp);		//Calc B component
+	}
+}
+
+/*
+ * Below are the function versions of the above macros, use those for debugging, but leave the macros for maximum CPU execution speed
+ *
+ *
+ *
+ *
+
+void _decodeRGB(struct TheoraPixelTransform* t, int stride, int nBytes, int maxWidth, int i1, int i2, int i3, int j1, int j2, int j3)
+{
+	register int tmp;
+	int nBytes2 = nBytes * 2, cv, cu, rgbY1, rgbY2, rgbY3, rgbY4, rV, gUV, bU, width = maxWidth == 0 ? t->w : maxWidth;
+	unsigned int y;
+	unsigned char *ySrcEven, *ySrcOdd, *yLineEnd, *uSrc, *vSrc, *out1, *out2;
+	
+	for (y = 0; y < t->h; y += 2)
+	{
+		ySrcEven = t->y + y * t->yStride;
+		ySrcOdd  = t->y + (y + 1) * t->yStride;
+		uSrc = t->u + y * t->uStride / 2;
+		vSrc = t->v + y * t->vStride / 2;
+		out1 = t->out + y * stride;
+		out2 = t->out + (y + 1) * stride;
+		
+		for (yLineEnd = ySrcEven + width; ySrcEven != yLineEnd;)
+		{
+			cu = *uSrc; ++uSrc;
+			cv = *vSrc; ++vSrc;
+			rV   = RVTable[cv];
+			gUV  = GUTable[cu] + GVTable[cv];
+			bU   = BUTable[cu];
+			
+			rgbY1 = YTable[*ySrcEven]; ++ySrcEven;
+			rgbY2 = YTable[*ySrcOdd];  ++ySrcOdd;
+			rgbY3 = YTable[*ySrcEven]; ++ySrcEven;
+			rgbY4 = YTable[*ySrcOdd];  ++ySrcOdd;
+			
+			CLIP_RGB_COLOR(out1[i1], rgbY1 + rV );
+			CLIP_RGB_COLOR(out1[i2], rgbY1 - gUV);
+			CLIP_RGB_COLOR(out1[i3], rgbY1 + bU );
+			
+			CLIP_RGB_COLOR(out2[i1], rgbY2 + rV );
+			CLIP_RGB_COLOR(out2[i2], rgbY2 - gUV);
+			CLIP_RGB_COLOR(out2[i3], rgbY2 + bU );
+			
+			CLIP_RGB_COLOR(out1[j1], rgbY3 + rV );
+			CLIP_RGB_COLOR(out1[j2], rgbY3 - gUV);
+			CLIP_RGB_COLOR(out1[j3], rgbY3 + bU );
+			
+			CLIP_RGB_COLOR(out2[j1], rgbY4 + rV );
+			CLIP_RGB_COLOR(out2[j2], rgbY4 - gUV);
+			CLIP_RGB_COLOR(out2[j3], rgbY4 + bU );
+			
+			out1 += nBytes2;  out2 += nBytes2;
+		}
+	}
+}
+ 
+void _decodeRGBA(struct TheoraPixelTransform* t, int stride, int nBytes, int maxWidth, int i1, int i2, int i3, int j1, int j2, int j3, int aindex1, int aindex2)
+{
+	register int tmp;
+	int nBytes2 = nBytes * 2, cv, cu, rgbY1, rgbY2, rgbY3, rgbY4, a1, a2, a3, a4, rV, gUV, bU, width = maxWidth == 0 ? t->w : maxWidth;
+	int alphaStride = t->w;
+	unsigned int y;
+	unsigned char *ySrcEven, *ySrcOdd, *yLineEnd, *uSrc, *vSrc, *out1, *out2;
+	
+	for (y = 0; y < t->h; y += 2)
+	{
+		ySrcEven = t->y + y * t->yStride;
+		ySrcOdd  = t->y + (y + 1) * t->yStride;
+		uSrc = t->u + y * t->uStride / 2;
+		vSrc = t->v + y * t->vStride / 2;
+		out1 = t->out + y * stride;
+		out2 = t->out + (y + 1) * stride;
+		
+		for (yLineEnd = ySrcEven + width; ySrcEven != yLineEnd;)
+		{
+			cu = *uSrc; ++uSrc;
+			cv = *vSrc; ++vSrc;
+			rV   = RVTable[cv];
+			gUV  = GUTable[cu] + GVTable[cv];
+			bU   = BUTable[cu];
+			
+			rgbY1 = YTable[*ySrcEven]; a1 = ySrcEven[alphaStride]; ++ySrcEven;
+			rgbY2 = YTable[*ySrcOdd];  a2 = ySrcOdd [alphaStride];  ++ySrcOdd;
+			rgbY3 = YTable[*ySrcEven]; a3 = ySrcEven[alphaStride]; ++ySrcEven;
+			rgbY4 = YTable[*ySrcOdd];  a4 = ySrcOdd [alphaStride];  ++ySrcOdd;
+			
+			if (a1 >= 32)
+			{
+				CLIP_RGB_COLOR(out1[i1], rgbY1 + rV );
+				CLIP_RGB_COLOR(out1[i2], rgbY1 - gUV);
+				CLIP_RGB_COLOR(out1[i3], rgbY1 + bU );
+				out1[aindex1] = a1 > 224 ? 255 : a1;
+			}
+			else *((unsigned int*) out1) = 0;
+			
+			if (a2 >= 32)
+			{
+				CLIP_RGB_COLOR(out2[i1], rgbY2 + rV );
+				CLIP_RGB_COLOR(out2[i2], rgbY2 - gUV);
+				CLIP_RGB_COLOR(out2[i3], rgbY2 + bU );
+				out2[aindex1] = a2 > 224 ? 255 : a2;
+			}
+			else *((unsigned int*) out2) = 0;
+
+			
+			if (a3 >= 32)
+			{
+				CLIP_RGB_COLOR(out1[j1], rgbY3 + rV );
+				CLIP_RGB_COLOR(out1[j2], rgbY3 - gUV);
+				CLIP_RGB_COLOR(out1[j3], rgbY3 + bU );
+				out1[aindex2] = a3 > 224 ? 255 : a3;
+			}
+			else *((unsigned int*) &out1[4]) = 0;
+
+			if (a4 >= 32)
+			{
+				CLIP_RGB_COLOR(out2[j1], rgbY4 + rV );
+				CLIP_RGB_COLOR(out2[j2], rgbY4 - gUV);
+				CLIP_RGB_COLOR(out2[j3], rgbY4 + bU );
+				out2[aindex2] = a4 > 224 ? 255 : a4;
+			}
+			else *((unsigned int*) &out2[4]) = 0;
+
+			out1 += nBytes2;  out2 += nBytes2;
+		}
+	}
+}
+*/
+#endif
diff --git a/drivers/theoraplayer/src/YUV/C/yuv420_yuv_c.c b/drivers/theoraplayer/src/YUV/C/yuv420_yuv_c.c
new file mode 100644
index 0000000000..fea74eca71
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/C/yuv420_yuv_c.c
@@ -0,0 +1,86 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "yuv_util.h"
+
+static void _decodeYUV(struct TheoraPixelTransform* t, int stride, int nBytes, int maxWidth)
+{
+	int cv, cu, y1, y2, y3, y4, width = maxWidth == 0 ? t->w : maxWidth;
+	unsigned char *ySrcEven, *ySrcOdd, *yLineEnd, *uSrc, *vSrc, *out1, *out2;
+	unsigned int y;
+
+	for (y=0; y < t->h; y += 2)
+	{
+		ySrcEven = t->y + y * t->yStride;
+		ySrcOdd  = t->y + (y + 1) * t->yStride;
+		uSrc = t->u + y * t->uStride / 2;
+		vSrc = t->v + y * t->vStride / 2;
+		out1 = t->out + y * stride;
+		out2 = t->out + (y + 1) * stride;
+		
+		for (yLineEnd = ySrcEven + width; ySrcEven != yLineEnd;)
+		{
+			// EVEN columns
+			cu = *uSrc; ++uSrc;
+			cv = *vSrc; ++vSrc;
+			
+			y1 = *ySrcEven; ++ySrcEven;
+			y2 = *ySrcOdd;  ++ySrcOdd;
+			y3 = *ySrcEven; ++ySrcEven;
+			y4 = *ySrcOdd;  ++ySrcOdd;
+			
+			// EVEN columns
+			out1[0] = y1;
+			out1[1] = cu;
+			out1[2] = cv;
+			
+			out2[0] = y2;
+			out2[1] = cu;
+			out2[2] = cv;
+			
+			out1 += nBytes;  out2 += nBytes;
+			// ODD columns
+			out1[0] = y3;
+			out1[1] = cu;
+			out1[2] = cv;
+			
+			out2[0] = y4;
+			out2[1] = cu;
+			out2[2] = cv;
+			out1 += nBytes;  out2 += nBytes;
+		}
+	}
+}
+
+void decodeYUV(struct TheoraPixelTransform* t)
+{
+	_decodeYUV(t, t->w * 3, 3, 0);
+}
+
+void decodeYUVA(struct TheoraPixelTransform* t)
+{
+	_decodeYUV(t, t->w * 4, 4, 0);
+	_decodeAlpha(incOut(t, 3), t->w * 4);
+}
+
+void decodeYUVX(struct TheoraPixelTransform* t)
+{
+	_decodeYUV(t, t->w * 4, 4, 0);
+}
+
+void decodeAYUV(struct TheoraPixelTransform* t)
+{
+	_decodeYUV(incOut(t, 1), t->w * 4, 4, 0);
+	_decodeAlpha(t, t->w * 4);
+}
+
+void decodeXYUV(struct TheoraPixelTransform* t)
+{
+	_decodeYUV(incOut(t, 1), t->w * 4, 4, 0);
+}
+
diff --git a/drivers/theoraplayer/src/YUV/android/cpu-features.c b/drivers/theoraplayer/src/YUV/android/cpu-features.c
new file mode 100644
index 0000000000..623dc94e0e
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/android/cpu-features.c
@@ -0,0 +1,1095 @@
+/*
+ * Copyright (C) 2010 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+
+/* ChangeLog for this library:
+ *
+ * NDK r8d: Add android_setCpu().
+ *
+ * NDK r8c: Add new ARM CPU features: VFPv2, VFP_D32, VFP_FP16,
+ *          VFP_FMA, NEON_FMA, IDIV_ARM, IDIV_THUMB2 and iWMMXt.
+ *
+ *          Rewrite the code to parse /proc/self/auxv instead of
+ *          the "Features" field in /proc/cpuinfo.
+ *
+ *          Dynamically allocate the buffer that hold the content
+ *          of /proc/cpuinfo to deal with newer hardware.
+ *
+ * NDK r7c: Fix CPU count computation. The old method only reported the
+ *           number of _active_ CPUs when the library was initialized,
+ *           which could be less than the real total.
+ *
+ * NDK r5: Handle buggy kernels which report a CPU Architecture number of 7
+ *         for an ARMv6 CPU (see below).
+ *
+ *         Handle kernels that only report 'neon', and not 'vfpv3'
+ *         (VFPv3 is mandated by the ARM architecture is Neon is implemented)
+ *
+ *         Handle kernels that only report 'vfpv3d16', and not 'vfpv3'
+ *
+ *         Fix x86 compilation. Report ANDROID_CPU_FAMILY_X86 in
+ *         android_getCpuFamily().
+ *
+ * NDK r4: Initial release
+ */
+
+#if 0
+
+#ifdef _ANDROID
+#if defined(__le32__)
+
+// When users enter this, we should only provide interface and
+// libportable will give the implementations.
+
+#else // !__le32__
+
+#include <sys/system_properties.h>
+#include <pthread.h>
+#include "cpu-features.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <fcntl.h>
+#include <errno.h>
+
+static  pthread_once_t     g_once;
+static  int                g_inited;
+static  AndroidCpuFamily   g_cpuFamily;
+static  uint64_t           g_cpuFeatures;
+static  int                g_cpuCount;
+
+#ifdef __arm__
+static  uint32_t           g_cpuIdArm;
+#endif
+
+static const int  android_cpufeatures_debug = 0;
+
+#ifdef __arm__
+#  define DEFAULT_CPU_FAMILY  ANDROID_CPU_FAMILY_ARM
+#elif defined __i386__
+#  define DEFAULT_CPU_FAMILY  ANDROID_CPU_FAMILY_X86
+#else
+#  define DEFAULT_CPU_FAMILY  ANDROID_CPU_FAMILY_UNKNOWN
+#endif
+
+#define  D(...) \
+    do { \
+        if (android_cpufeatures_debug) { \
+            printf(__VA_ARGS__); fflush(stdout); \
+        } \
+    } while (0)
+
+#ifdef __i386__
+static __inline__ void x86_cpuid(int func, int values[4])
+{
+    int a, b, c, d;
+    /* We need to preserve ebx since we're compiling PIC code */
+    /* this means we can't use "=b" for the second output register */
+    __asm__ __volatile__ ( \
+      "push %%ebx\n"
+      "cpuid\n" \
+      "mov %%ebx, %1\n"
+      "pop %%ebx\n"
+      : "=a" (a), "=r" (b), "=c" (c), "=d" (d) \
+      : "a" (func) \
+    );
+    values[0] = a;
+    values[1] = b;
+    values[2] = c;
+    values[3] = d;
+}
+#endif
+
+/* Get the size of a file by reading it until the end. This is needed
+ * because files under /proc do not always return a valid size when
+ * using fseek(0, SEEK_END) + ftell(). Nor can they be mmap()-ed.
+ */
+static int
+get_file_size(const char* pathname)
+{
+    int fd, ret, result = 0;
+    char buffer[256];
+
+    fd = open(pathname, O_RDONLY);
+    if (fd < 0) {
+        D("Can't open %s: %s\n", pathname, strerror(errno));
+        return -1;
+    }
+
+    for (;;) {
+        int ret = read(fd, buffer, sizeof buffer);
+        if (ret < 0) {
+            if (errno == EINTR)
+                continue;
+            D("Error while reading %s: %s\n", pathname, strerror(errno));
+            break;
+        }
+        if (ret == 0)
+            break;
+
+        result += ret;
+    }
+    close(fd);
+    return result;
+}
+
+/* Read the content of /proc/cpuinfo into a user-provided buffer.
+ * Return the length of the data, or -1 on error. Does *not*
+ * zero-terminate the content. Will not read more
+ * than 'buffsize' bytes.
+ */
+static int
+read_file(const char*  pathname, char*  buffer, size_t  buffsize)
+{
+    int  fd, count;
+
+    fd = open(pathname, O_RDONLY);
+    if (fd < 0) {
+        D("Could not open %s: %s\n", pathname, strerror(errno));
+        return -1;
+    }
+    count = 0;
+    while (count < (int)buffsize) {
+        int ret = read(fd, buffer + count, buffsize - count);
+        if (ret < 0) {
+            if (errno == EINTR)
+                continue;
+            D("Error while reading from %s: %s\n", pathname, strerror(errno));
+            if (count == 0)
+                count = -1;
+            break;
+        }
+        if (ret == 0)
+            break;
+        count += ret;
+    }
+    close(fd);
+    return count;
+}
+
+/* Extract the content of a the first occurence of a given field in
+ * the content of /proc/cpuinfo and return it as a heap-allocated
+ * string that must be freed by the caller.
+ *
+ * Return NULL if not found
+ */
+static char*
+extract_cpuinfo_field(const char* buffer, int buflen, const char* field)
+{
+    int  fieldlen = strlen(field);
+    const char* bufend = buffer + buflen;
+    char* result = NULL;
+    int len, ignore;
+    const char *p, *q;
+
+    /* Look for first field occurence, and ensures it starts the line. */
+    p = buffer;
+    for (;;) {
+        p = memmem(p, bufend-p, field, fieldlen);
+        if (p == NULL)
+            goto EXIT;
+
+        if (p == buffer || p[-1] == '\n')
+            break;
+
+        p += fieldlen;
+    }
+
+    /* Skip to the first column followed by a space */
+    p += fieldlen;
+    p  = memchr(p, ':', bufend-p);
+    if (p == NULL || p[1] != ' ')
+        goto EXIT;
+
+    /* Find the end of the line */
+    p += 2;
+    q = memchr(p, '\n', bufend-p);
+    if (q == NULL)
+        q = bufend;
+
+    /* Copy the line into a heap-allocated buffer */
+    len = q-p;
+    result = malloc(len+1);
+    if (result == NULL)
+        goto EXIT;
+
+    memcpy(result, p, len);
+    result[len] = '\0';
+
+EXIT:
+    return result;
+}
+
+/* Checks that a space-separated list of items contains one given 'item'.
+ * Returns 1 if found, 0 otherwise.
+ */
+static int
+has_list_item(const char* list, const char* item)
+{
+    const char*  p = list;
+    int itemlen = strlen(item);
+
+    if (list == NULL)
+        return 0;
+
+    while (*p) {
+        const char*  q;
+
+        /* skip spaces */
+        while (*p == ' ' || *p == '\t')
+            p++;
+
+        /* find end of current list item */
+        q = p;
+        while (*q && *q != ' ' && *q != '\t')
+            q++;
+
+        if (itemlen == q-p && !memcmp(p, item, itemlen))
+            return 1;
+
+        /* skip to next item */
+        p = q;
+    }
+    return 0;
+}
+
+/* Parse a number starting from 'input', but not going further
+ * than 'limit'. Return the value into '*result'.
+ *
+ * NOTE: Does not skip over leading spaces, or deal with sign characters.
+ * NOTE: Ignores overflows.
+ *
+ * The function returns NULL in case of error (bad format), or the new
+ * position after the decimal number in case of success (which will always
+ * be <= 'limit').
+ */
+static const char*
+parse_number(const char* input, const char* limit, int base, int* result)
+{
+    const char* p = input;
+    int val = 0;
+    while (p < limit) {
+        int d = (*p - '0');
+        if ((unsigned)d >= 10U) {
+            d = (*p - 'a');
+            if ((unsigned)d >= 6U)
+              d = (*p - 'A');
+            if ((unsigned)d >= 6U)
+              break;
+            d += 10;
+        }
+        if (d >= base)
+          break;
+        val = val*base + d;
+        p++;
+    }
+    if (p == input)
+        return NULL;
+
+    *result = val;
+    return p;
+}
+
+static const char*
+parse_decimal(const char* input, const char* limit, int* result)
+{
+    return parse_number(input, limit, 10, result);
+}
+
+static const char*
+parse_hexadecimal(const char* input, const char* limit, int* result)
+{
+    return parse_number(input, limit, 16, result);
+}
+
+/* This small data type is used to represent a CPU list / mask, as read
+ * from sysfs on Linux. See http://www.kernel.org/doc/Documentation/cputopology.txt
+ *
+ * For now, we don't expect more than 32 cores on mobile devices, so keep
+ * everything simple.
+ */
+typedef struct {
+    uint32_t mask;
+} CpuList;
+
+static __inline__ void
+cpulist_init(CpuList* list) {
+    list->mask = 0;
+}
+
+static __inline__ void
+cpulist_and(CpuList* list1, CpuList* list2) {
+    list1->mask &= list2->mask;
+}
+
+static __inline__ void
+cpulist_set(CpuList* list, int index) {
+    if ((unsigned)index < 32) {
+        list->mask |= (uint32_t)(1U << index);
+    }
+}
+
+static __inline__ int
+cpulist_count(CpuList* list) {
+    return __builtin_popcount(list->mask);
+}
+
+/* Parse a textual list of cpus and store the result inside a CpuList object.
+ * Input format is the following:
+ * - comma-separated list of items (no spaces)
+ * - each item is either a single decimal number (cpu index), or a range made
+ *   of two numbers separated by a single dash (-). Ranges are inclusive.
+ *
+ * Examples:   0
+ *             2,4-127,128-143
+ *             0-1
+ */
+static void
+cpulist_parse(CpuList* list, const char* line, int line_len)
+{
+    const char* p = line;
+    const char* end = p + line_len;
+    const char* q;
+
+    /* NOTE: the input line coming from sysfs typically contains a
+     * trailing newline, so take care of it in the code below
+     */
+    while (p < end && *p != '\n')
+    {
+        int val, start_value, end_value;
+
+        /* Find the end of current item, and put it into 'q' */
+        q = memchr(p, ',', end-p);
+        if (q == NULL) {
+            q = end;
+        }
+
+        /* Get first value */
+        p = parse_decimal(p, q, &start_value);
+        if (p == NULL)
+            goto BAD_FORMAT;
+
+        end_value = start_value;
+
+        /* If we're not at the end of the item, expect a dash and
+         * and integer; extract end value.
+         */
+        if (p < q && *p == '-') {
+            p = parse_decimal(p+1, q, &end_value);
+            if (p == NULL)
+                goto BAD_FORMAT;
+        }
+
+        /* Set bits CPU list bits */
+        for (val = start_value; val <= end_value; val++) {
+            cpulist_set(list, val);
+        }
+
+        /* Jump to next item */
+        p = q;
+        if (p < end)
+            p++;
+    }
+
+BAD_FORMAT:
+    ;
+}
+
+/* Read a CPU list from one sysfs file */
+static void
+cpulist_read_from(CpuList* list, const char* filename)
+{
+    char   file[64];
+    int    filelen;
+
+    cpulist_init(list);
+
+    filelen = read_file(filename, file, sizeof file);
+    if (filelen < 0) {
+        D("Could not read %s: %s\n", filename, strerror(errno));
+        return;
+    }
+
+    cpulist_parse(list, file, filelen);
+}
+
+// See <asm/hwcap.h> kernel header.
+#define HWCAP_VFP       (1 << 6)
+#define HWCAP_IWMMXT    (1 << 9)
+#define HWCAP_NEON      (1 << 12)
+#define HWCAP_VFPv3     (1 << 13)
+#define HWCAP_VFPv3D16  (1 << 14)
+#define HWCAP_VFPv4     (1 << 16)
+#define HWCAP_IDIVA     (1 << 17)
+#define HWCAP_IDIVT     (1 << 18)
+
+#define AT_HWCAP 16
+
+#if defined(__arm__)
+/* Compute the ELF HWCAP flags.
+ */
+static uint32_t
+get_elf_hwcap(const char* cpuinfo, int cpuinfo_len)
+{
+  /* IMPORTANT:
+   *   Accessing /proc/self/auxv doesn't work anymore on all
+   *   platform versions. More specifically, when running inside
+   *   a regular application process, most of /proc/self/ will be
+   *   non-readable, including /proc/self/auxv. This doesn't
+   *   happen however if the application is debuggable, or when
+   *   running under the "shell" UID, which is why this was not
+   *   detected appropriately.
+   */
+#if 0
+    uint32_t result = 0;
+    const char filepath[] = "/proc/self/auxv";
+    int fd = open(filepath, O_RDONLY);
+    if (fd < 0) {
+        D("Could not open %s: %s\n", filepath, strerror(errno));
+        return 0;
+    }
+
+    struct { uint32_t tag; uint32_t value; } entry;
+
+    for (;;) {
+        int ret = read(fd, (char*)&entry, sizeof entry);
+        if (ret < 0) {
+            if (errno == EINTR)
+                continue;
+            D("Error while reading %s: %s\n", filepath, strerror(errno));
+            break;
+        }
+        // Detect end of list.
+        if (ret == 0 || (entry.tag == 0 && entry.value == 0))
+          break;
+        if (entry.tag == AT_HWCAP) {
+          result = entry.value;
+          break;
+        }
+    }
+    close(fd);
+    return result;
+#else
+    // Recreate ELF hwcaps by parsing /proc/cpuinfo Features tag.
+    uint32_t hwcaps = 0;
+
+    char* cpuFeatures = extract_cpuinfo_field(cpuinfo, cpuinfo_len, "Features");
+
+    if (cpuFeatures != NULL) {
+        D("Found cpuFeatures = '%s'\n", cpuFeatures);
+
+        if (has_list_item(cpuFeatures, "vfp"))
+            hwcaps |= HWCAP_VFP;
+        if (has_list_item(cpuFeatures, "vfpv3"))
+            hwcaps |= HWCAP_VFPv3;
+        if (has_list_item(cpuFeatures, "vfpv3d16"))
+            hwcaps |= HWCAP_VFPv3D16;
+        if (has_list_item(cpuFeatures, "vfpv4"))
+            hwcaps |= HWCAP_VFPv4;
+        if (has_list_item(cpuFeatures, "neon"))
+            hwcaps |= HWCAP_NEON;
+        if (has_list_item(cpuFeatures, "idiva"))
+            hwcaps |= HWCAP_IDIVA;
+        if (has_list_item(cpuFeatures, "idivt"))
+            hwcaps |= HWCAP_IDIVT;
+        if (has_list_item(cpuFeatures, "idiv"))
+            hwcaps |= HWCAP_IDIVA | HWCAP_IDIVT;
+        if (has_list_item(cpuFeatures, "iwmmxt"))
+            hwcaps |= HWCAP_IWMMXT;
+
+        free(cpuFeatures);
+    }
+    return hwcaps;
+#endif
+}
+#endif  /* __arm__ */
+
+/* Return the number of cpus present on a given device.
+ *
+ * To handle all weird kernel configurations, we need to compute the
+ * intersection of the 'present' and 'possible' CPU lists and count
+ * the result.
+ */
+static int
+get_cpu_count(void)
+{
+    CpuList cpus_present[1];
+    CpuList cpus_possible[1];
+
+    cpulist_read_from(cpus_present, "/sys/devices/system/cpu/present");
+    cpulist_read_from(cpus_possible, "/sys/devices/system/cpu/possible");
+
+    /* Compute the intersection of both sets to get the actual number of
+     * CPU cores that can be used on this device by the kernel.
+     */
+    cpulist_and(cpus_present, cpus_possible);
+
+    return cpulist_count(cpus_present);
+}
+
+static void
+android_cpuInitFamily(void)
+{
+#if defined(__arm__)
+    g_cpuFamily = ANDROID_CPU_FAMILY_ARM;
+#elif defined(__i386__)
+    g_cpuFamily = ANDROID_CPU_FAMILY_X86;
+#elif defined(__mips64)
+/* Needs to be before __mips__ since the compiler defines both */
+    g_cpuFamily = ANDROID_CPU_FAMILY_MIPS64;
+#elif defined(__mips__)
+    g_cpuFamily = ANDROID_CPU_FAMILY_MIPS;
+#elif defined(__aarch64__)
+    g_cpuFamily = ANDROID_CPU_FAMILY_ARM64;
+#elif defined(__x86_64__)
+    g_cpuFamily = ANDROID_CPU_FAMILY_X86_64;
+#else
+    g_cpuFamily = ANDROID_CPU_FAMILY_UNKNOWN;
+#endif
+}
+
+static void
+android_cpuInit(void)
+{
+    char* cpuinfo = NULL;
+    int   cpuinfo_len;
+
+    android_cpuInitFamily();
+
+    g_cpuFeatures = 0;
+    g_cpuCount    = 1;
+    g_inited      = 1;
+
+    cpuinfo_len = get_file_size("/proc/cpuinfo");
+    if (cpuinfo_len < 0) {
+      D("cpuinfo_len cannot be computed!");
+      return;
+    }
+    cpuinfo = malloc(cpuinfo_len);
+    if (cpuinfo == NULL) {
+      D("cpuinfo buffer could not be allocated");
+      return;
+    }
+    cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, cpuinfo_len);
+    D("cpuinfo_len is (%d):\n%.*s\n", cpuinfo_len,
+      cpuinfo_len >= 0 ? cpuinfo_len : 0, cpuinfo);
+
+    if (cpuinfo_len < 0)  /* should not happen */ {
+        free(cpuinfo);
+        return;
+    }
+
+    /* Count the CPU cores, the value may be 0 for single-core CPUs */
+    g_cpuCount = get_cpu_count();
+    if (g_cpuCount == 0) {
+        g_cpuCount = 1;
+    }
+
+    D("found cpuCount = %d\n", g_cpuCount);
+
+#ifdef __arm__
+    {
+        char*  features = NULL;
+        char*  architecture = NULL;
+
+        /* Extract architecture from the "CPU Architecture" field.
+         * The list is well-known, unlike the the output of
+         * the 'Processor' field which can vary greatly.
+         *
+         * See the definition of the 'proc_arch' array in
+         * $KERNEL/arch/arm/kernel/setup.c and the 'c_show' function in
+         * same file.
+         */
+        char* cpuArch = extract_cpuinfo_field(cpuinfo, cpuinfo_len, "CPU architecture");
+
+        if (cpuArch != NULL) {
+            char*  end;
+            long   archNumber;
+            int    hasARMv7 = 0;
+
+            D("found cpuArch = '%s'\n", cpuArch);
+
+            /* read the initial decimal number, ignore the rest */
+            archNumber = strtol(cpuArch, &end, 10);
+
+            /* Here we assume that ARMv8 will be upwards compatible with v7
+             * in the future. Unfortunately, there is no 'Features' field to
+             * indicate that Thumb-2 is supported.
+             */
+            if (end > cpuArch && archNumber >= 7) {
+                hasARMv7 = 1;
+            }
+
+            /* Unfortunately, it seems that certain ARMv6-based CPUs
+             * report an incorrect architecture number of 7!
+             *
+             * See http://code.google.com/p/android/issues/detail?id=10812
+             *
+             * We try to correct this by looking at the 'elf_format'
+             * field reported by the 'Processor' field, which is of the
+             * form of "(v7l)" for an ARMv7-based CPU, and "(v6l)" for
+             * an ARMv6-one.
+             */
+            if (hasARMv7) {
+                char* cpuProc = extract_cpuinfo_field(cpuinfo, cpuinfo_len,
+                                                      "Processor");
+                if (cpuProc != NULL) {
+                    D("found cpuProc = '%s'\n", cpuProc);
+                    if (has_list_item(cpuProc, "(v6l)")) {
+                        D("CPU processor and architecture mismatch!!\n");
+                        hasARMv7 = 0;
+                    }
+                    free(cpuProc);
+                }
+            }
+
+            if (hasARMv7) {
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_ARMv7;
+            }
+
+            /* The LDREX / STREX instructions are available from ARMv6 */
+            if (archNumber >= 6) {
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_LDREX_STREX;
+            }
+
+            free(cpuArch);
+        }
+
+        /* Extract the list of CPU features from ELF hwcaps */
+        uint32_t hwcaps = get_elf_hwcap(cpuinfo, cpuinfo_len);
+
+        if (hwcaps != 0) {
+            int has_vfp = (hwcaps & HWCAP_VFP);
+            int has_vfpv3 = (hwcaps & HWCAP_VFPv3);
+            int has_vfpv3d16 = (hwcaps & HWCAP_VFPv3D16);
+            int has_vfpv4 = (hwcaps & HWCAP_VFPv4);
+            int has_neon = (hwcaps & HWCAP_NEON);
+            int has_idiva = (hwcaps & HWCAP_IDIVA);
+            int has_idivt = (hwcaps & HWCAP_IDIVT);
+            int has_iwmmxt = (hwcaps & HWCAP_IWMMXT);
+
+            // The kernel does a poor job at ensuring consistency when
+            // describing CPU features. So lots of guessing is needed.
+
+            // 'vfpv4' implies VFPv3|VFP_FMA|FP16
+            if (has_vfpv4)
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3    |
+                                 ANDROID_CPU_ARM_FEATURE_VFP_FP16 |
+                                 ANDROID_CPU_ARM_FEATURE_VFP_FMA;
+
+            // 'vfpv3' or 'vfpv3d16' imply VFPv3. Note that unlike GCC,
+            // a value of 'vfpv3' doesn't necessarily mean that the D32
+            // feature is present, so be conservative. All CPUs in the
+            // field that support D32 also support NEON, so this should
+            // not be a problem in practice.
+            if (has_vfpv3 || has_vfpv3d16)
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3;
+
+            // 'vfp' is super ambiguous. Depending on the kernel, it can
+            // either mean VFPv2 or VFPv3. Make it depend on ARMv7.
+            if (has_vfp) {
+              if (g_cpuFeatures & ANDROID_CPU_ARM_FEATURE_ARMv7)
+                  g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3;
+              else
+                  g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv2;
+            }
+
+            // Neon implies VFPv3|D32, and if vfpv4 is detected, NEON_FMA
+            if (has_neon) {
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3 |
+                                 ANDROID_CPU_ARM_FEATURE_NEON |
+                                 ANDROID_CPU_ARM_FEATURE_VFP_D32;
+              if (has_vfpv4)
+                  g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_NEON_FMA;
+            }
+
+            // VFPv3 implies VFPv2 and ARMv7
+            if (g_cpuFeatures & ANDROID_CPU_ARM_FEATURE_VFPv3)
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv2 |
+                                 ANDROID_CPU_ARM_FEATURE_ARMv7;
+
+            if (has_idiva)
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_IDIV_ARM;
+            if (has_idivt)
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_IDIV_THUMB2;
+
+            if (has_iwmmxt)
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_iWMMXt;
+        }
+
+        /* Extract the cpuid value from various fields */
+        // The CPUID value is broken up in several entries in /proc/cpuinfo.
+        // This table is used to rebuild it from the entries.
+        static const struct CpuIdEntry {
+            const char* field;
+            char        format;
+            char        bit_lshift;
+            char        bit_length;
+        } cpu_id_entries[] = {
+            { "CPU implementer", 'x', 24, 8 },
+            { "CPU variant", 'x', 20, 4 },
+            { "CPU part", 'x', 4, 12 },
+            { "CPU revision", 'd', 0, 4 },
+        };
+        size_t i;
+        D("Parsing /proc/cpuinfo to recover CPUID\n");
+        for (i = 0;
+             i < sizeof(cpu_id_entries)/sizeof(cpu_id_entries[0]);
+             ++i) {
+            const struct CpuIdEntry* entry = &cpu_id_entries[i];
+            char* value = extract_cpuinfo_field(cpuinfo,
+                                                cpuinfo_len,
+                                                entry->field);
+            if (value == NULL)
+                continue;
+
+            D("field=%s value='%s'\n", entry->field, value);
+            char* value_end = value + strlen(value);
+            int val = 0;
+            const char* start = value;
+            const char* p;
+            if (value[0] == '0' && (value[1] == 'x' || value[1] == 'X')) {
+              start += 2;
+              p = parse_hexadecimal(start, value_end, &val);
+            } else if (entry->format == 'x')
+              p = parse_hexadecimal(value, value_end, &val);
+            else
+              p = parse_decimal(value, value_end, &val);
+
+            if (p > (const char*)start) {
+              val &= ((1 << entry->bit_length)-1);
+              val <<= entry->bit_lshift;
+              g_cpuIdArm |= (uint32_t) val;
+            }
+
+            free(value);
+        }
+
+        // Handle kernel configuration bugs that prevent the correct
+        // reporting of CPU features.
+        static const struct CpuFix {
+            uint32_t  cpuid;
+            uint64_t  or_flags;
+        } cpu_fixes[] = {
+            /* The Nexus 4 (Qualcomm Krait) kernel configuration
+             * forgets to report IDIV support. */
+            { 0x510006f2, ANDROID_CPU_ARM_FEATURE_IDIV_ARM |
+                          ANDROID_CPU_ARM_FEATURE_IDIV_THUMB2 },
+            { 0x510006f3, ANDROID_CPU_ARM_FEATURE_IDIV_ARM |
+                          ANDROID_CPU_ARM_FEATURE_IDIV_THUMB2 },
+        };
+        size_t n;
+        for (n = 0; n < sizeof(cpu_fixes)/sizeof(cpu_fixes[0]); ++n) {
+            const struct CpuFix* entry = &cpu_fixes[n];
+
+            if (g_cpuIdArm == entry->cpuid)
+                g_cpuFeatures |= entry->or_flags;
+        }
+
+    }
+#endif /* __arm__ */
+
+#ifdef __i386__
+    int regs[4];
+
+/* According to http://en.wikipedia.org/wiki/CPUID */
+#define VENDOR_INTEL_b  0x756e6547
+#define VENDOR_INTEL_c  0x6c65746e
+#define VENDOR_INTEL_d  0x49656e69
+
+    x86_cpuid(0, regs);
+    int vendorIsIntel = (regs[1] == VENDOR_INTEL_b &&
+                         regs[2] == VENDOR_INTEL_c &&
+                         regs[3] == VENDOR_INTEL_d);
+
+    x86_cpuid(1, regs);
+    if ((regs[2] & (1 << 9)) != 0) {
+        g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_SSSE3;
+    }
+    if ((regs[2] & (1 << 23)) != 0) {
+        g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_POPCNT;
+    }
+    if (vendorIsIntel && (regs[2] & (1 << 22)) != 0) {
+        g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_MOVBE;
+    }
+#endif
+
+    free(cpuinfo);
+}
+
+
+AndroidCpuFamily
+android_getCpuFamily(void)
+{
+    pthread_once(&g_once, android_cpuInit);
+    return g_cpuFamily;
+}
+
+
+uint64_t
+android_getCpuFeaturesExt(void)
+{
+    pthread_once(&g_once, android_cpuInit);
+    return g_cpuFeatures;
+}
+
+
+int
+android_getCpuCount(void)
+{
+    pthread_once(&g_once, android_cpuInit);
+    return g_cpuCount;
+}
+
+static void
+android_cpuInitDummy(void)
+{
+    g_inited = 1;
+}
+
+int
+android_setCpu(int cpu_count, uint64_t cpu_features)
+{
+    /* Fail if the library was already initialized. */
+    if (g_inited)
+        return 0;
+
+    android_cpuInitFamily();
+    g_cpuCount = (cpu_count <= 0 ? 1 : cpu_count);
+    g_cpuFeatures = cpu_features;
+    pthread_once(&g_once, android_cpuInitDummy);
+
+    return 1;
+}
+
+#ifdef __arm__
+uint32_t
+android_getCpuIdArm(void)
+{
+    pthread_once(&g_once, android_cpuInit);
+    return g_cpuIdArm;
+}
+
+int
+android_setCpuArm(int cpu_count, uint64_t cpu_features, uint32_t cpu_id)
+{
+    if (!android_setCpu(cpu_count, cpu_features))
+        return 0;
+
+    g_cpuIdArm = cpu_id;
+    return 1;
+}
+#endif  /* __arm__ */
+
+/*
+ * Technical note: Making sense of ARM's FPU architecture versions.
+ *
+ * FPA was ARM's first attempt at an FPU architecture. There is no Android
+ * device that actually uses it since this technology was already obsolete
+ * when the project started. If you see references to FPA instructions
+ * somewhere, you can be sure that this doesn't apply to Android at all.
+ *
+ * FPA was followed by "VFP", soon renamed "VFPv1" due to the emergence of
+ * new versions / additions to it. ARM considers this obsolete right now,
+ * and no known Android device implements it either.
+ *
+ * VFPv2 added a few instructions to VFPv1, and is an *optional* extension
+ * supported by some ARMv5TE, ARMv6 and ARMv6T2 CPUs. Note that a device
+ * supporting the 'armeabi' ABI doesn't necessarily support these.
+ *
+ * VFPv3-D16 adds a few instructions on top of VFPv2 and is typically used
+ * on ARMv7-A CPUs which implement a FPU. Note that it is also mandated
+ * by the Android 'armeabi-v7a' ABI. The -D16 suffix in its name means
+ * that it provides 16 double-precision FPU registers (d0-d15) and 32
+ * single-precision ones (s0-s31) which happen to be mapped to the same
+ * register banks.
+ *
+ * VFPv3-D32 is the name of an extension to VFPv3-D16 that provides 16
+ * additional double precision registers (d16-d31). Note that there are
+ * still only 32 single precision registers.
+ *
+ * VFPv3xD is a *subset* of VFPv3-D16 that only provides single-precision
+ * registers. It is only used on ARMv7-M (i.e. on micro-controllers) which
+ * are not supported by Android. Note that it is not compatible with VFPv2.
+ *
+ * NOTE: The term 'VFPv3' usually designate either VFPv3-D16 or VFPv3-D32
+ *       depending on context. For example GCC uses it for VFPv3-D32, but
+ *       the Linux kernel code uses it for VFPv3-D16 (especially in
+ *       /proc/cpuinfo). Always try to use the full designation when
+ *       possible.
+ *
+ * NEON, a.k.a. "ARM Advanced SIMD" is an extension that provides
+ * instructions to perform parallel computations on vectors of 8, 16,
+ * 32, 64 and 128 bit quantities. NEON requires VFPv32-D32 since all
+ * NEON registers are also mapped to the same register banks.
+ *
+ * VFPv4-D16, adds a few instructions on top of VFPv3-D16 in order to
+ * perform fused multiply-accumulate on VFP registers, as well as
+ * half-precision (16-bit) conversion operations.
+ *
+ * VFPv4-D32 is VFPv4-D16 with 32, instead of 16, FPU double precision
+ * registers.
+ *
+ * VPFv4-NEON is VFPv4-D32 with NEON instructions. It also adds fused
+ * multiply-accumulate instructions that work on the NEON registers.
+ *
+ * NOTE: Similarly, "VFPv4" might either reference VFPv4-D16 or VFPv4-D32
+ *       depending on context.
+ *
+ * The following information was determined by scanning the binutils-2.22
+ * sources:
+ *
+ * Basic VFP instruction subsets:
+ *
+ * #define FPU_VFP_EXT_V1xD 0x08000000     // Base VFP instruction set.
+ * #define FPU_VFP_EXT_V1   0x04000000     // Double-precision insns.
+ * #define FPU_VFP_EXT_V2   0x02000000     // ARM10E VFPr1.
+ * #define FPU_VFP_EXT_V3xD 0x01000000     // VFPv3 single-precision.
+ * #define FPU_VFP_EXT_V3   0x00800000     // VFPv3 double-precision.
+ * #define FPU_NEON_EXT_V1  0x00400000     // Neon (SIMD) insns.
+ * #define FPU_VFP_EXT_D32  0x00200000     // Registers D16-D31.
+ * #define FPU_VFP_EXT_FP16 0x00100000     // Half-precision extensions.
+ * #define FPU_NEON_EXT_FMA 0x00080000     // Neon fused multiply-add
+ * #define FPU_VFP_EXT_FMA  0x00040000     // VFP fused multiply-add
+ *
+ * FPU types (excluding NEON)
+ *
+ * FPU_VFP_V1xD (EXT_V1xD)
+ *    |
+ *    +--------------------------+
+ *    |                          |
+ * FPU_VFP_V1 (+EXT_V1)       FPU_VFP_V3xD (+EXT_V2+EXT_V3xD)
+ *    |                          |
+ *    |                          |
+ * FPU_VFP_V2 (+EXT_V2)       FPU_VFP_V4_SP_D16 (+EXT_FP16+EXT_FMA)
+ *    |
+ * FPU_VFP_V3D16 (+EXT_Vx3D+EXT_V3)
+ *    |
+ *    +--------------------------+
+ *    |                          |
+ * FPU_VFP_V3 (+EXT_D32)     FPU_VFP_V4D16 (+EXT_FP16+EXT_FMA)
+ *    |                          |
+ *    |                      FPU_VFP_V4 (+EXT_D32)
+ *    |
+ * FPU_VFP_HARD (+EXT_FMA+NEON_EXT_FMA)
+ *
+ * VFP architectures:
+ *
+ * ARCH_VFP_V1xD  (EXT_V1xD)
+ *   |
+ *   +------------------+
+ *   |                  |
+ *   |             ARCH_VFP_V3xD (+EXT_V2+EXT_V3xD)
+ *   |                  |
+ *   |             ARCH_VFP_V3xD_FP16 (+EXT_FP16)
+ *   |                  |
+ *   |             ARCH_VFP_V4_SP_D16 (+EXT_FMA)
+ *   |
+ * ARCH_VFP_V1 (+EXT_V1)
+ *   |
+ * ARCH_VFP_V2 (+EXT_V2)
+ *   |
+ * ARCH_VFP_V3D16 (+EXT_V3xD+EXT_V3)
+ *   |
+ *   +-------------------+
+ *   |                   |
+ *   |         ARCH_VFP_V3D16_FP16  (+EXT_FP16)
+ *   |
+ *   +-------------------+
+ *   |                   |
+ *   |         ARCH_VFP_V4_D16 (+EXT_FP16+EXT_FMA)
+ *   |                   |
+ *   |         ARCH_VFP_V4 (+EXT_D32)
+ *   |                   |
+ *   |         ARCH_NEON_VFP_V4 (+EXT_NEON+EXT_NEON_FMA)
+ *   |
+ * ARCH_VFP_V3 (+EXT_D32)
+ *   |
+ *   +-------------------+
+ *   |                   |
+ *   |         ARCH_VFP_V3_FP16 (+EXT_FP16)
+ *   |
+ * ARCH_VFP_V3_PLUS_NEON_V1 (+EXT_NEON)
+ *   |
+ * ARCH_NEON_FP16 (+EXT_FP16)
+ *
+ * -fpu=<name> values and their correspondance with FPU architectures above:
+ *
+ *   {"vfp",               FPU_ARCH_VFP_V2},
+ *   {"vfp9",              FPU_ARCH_VFP_V2},
+ *   {"vfp3",              FPU_ARCH_VFP_V3}, // For backwards compatbility.
+ *   {"vfp10",             FPU_ARCH_VFP_V2},
+ *   {"vfp10-r0",          FPU_ARCH_VFP_V1},
+ *   {"vfpxd",             FPU_ARCH_VFP_V1xD},
+ *   {"vfpv2",             FPU_ARCH_VFP_V2},
+ *   {"vfpv3",             FPU_ARCH_VFP_V3},
+ *   {"vfpv3-fp16",        FPU_ARCH_VFP_V3_FP16},
+ *   {"vfpv3-d16",         FPU_ARCH_VFP_V3D16},
+ *   {"vfpv3-d16-fp16",    FPU_ARCH_VFP_V3D16_FP16},
+ *   {"vfpv3xd",           FPU_ARCH_VFP_V3xD},
+ *   {"vfpv3xd-fp16",      FPU_ARCH_VFP_V3xD_FP16},
+ *   {"neon",              FPU_ARCH_VFP_V3_PLUS_NEON_V1},
+ *   {"neon-fp16",         FPU_ARCH_NEON_FP16},
+ *   {"vfpv4",             FPU_ARCH_VFP_V4},
+ *   {"vfpv4-d16",         FPU_ARCH_VFP_V4D16},
+ *   {"fpv4-sp-d16",       FPU_ARCH_VFP_V4_SP_D16},
+ *   {"neon-vfpv4",        FPU_ARCH_NEON_VFP_V4},
+ *
+ *
+ * Simplified diagram that only includes FPUs supported by Android:
+ * Only ARCH_VFP_V3D16 is actually mandated by the armeabi-v7a ABI,
+ * all others are optional and must be probed at runtime.
+ *
+ * ARCH_VFP_V3D16 (EXT_V1xD+EXT_V1+EXT_V2+EXT_V3xD+EXT_V3)
+ *   |
+ *   +-------------------+
+ *   |                   |
+ *   |         ARCH_VFP_V3D16_FP16  (+EXT_FP16)
+ *   |
+ *   +-------------------+
+ *   |                   |
+ *   |         ARCH_VFP_V4_D16 (+EXT_FP16+EXT_FMA)
+ *   |                   |
+ *   |         ARCH_VFP_V4 (+EXT_D32)
+ *   |                   |
+ *   |         ARCH_NEON_VFP_V4 (+EXT_NEON+EXT_NEON_FMA)
+ *   |
+ * ARCH_VFP_V3 (+EXT_D32)
+ *   |
+ *   +-------------------+
+ *   |                   |
+ *   |         ARCH_VFP_V3_FP16 (+EXT_FP16)
+ *   |
+ * ARCH_VFP_V3_PLUS_NEON_V1 (+EXT_NEON)
+ *   |
+ * ARCH_NEON_FP16 (+EXT_FP16)
+ *
+ */
+
+#endif // defined(__le32__)
+#endif
+
+#endif
diff --git a/drivers/theoraplayer/src/YUV/android/cpu-features.h b/drivers/theoraplayer/src/YUV/android/cpu-features.h
new file mode 100644
index 0000000000..12d3ad5645
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/android/cpu-features.h
@@ -0,0 +1,212 @@
+/*
+ * Copyright (C) 2010 The Android Open Source Project
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *  * Redistributions of source code must retain the above copyright
+ *    notice, this list of conditions and the following disclaimer.
+ *  * Redistributions in binary form must reproduce the above copyright
+ *    notice, this list of conditions and the following disclaimer in
+ *    the documentation and/or other materials provided with the
+ *    distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ * SUCH DAMAGE.
+ */
+#ifndef CPU_FEATURES_H
+#define CPU_FEATURES_H
+
+#include <sys/cdefs.h>
+#include <stdint.h>
+
+__BEGIN_DECLS
+
+typedef enum {
+    ANDROID_CPU_FAMILY_UNKNOWN = 0,
+    ANDROID_CPU_FAMILY_ARM,
+    ANDROID_CPU_FAMILY_X86,
+    ANDROID_CPU_FAMILY_MIPS,
+
+    ANDROID_CPU_FAMILY_MAX  /* do not remove */
+
+} AndroidCpuFamily;
+
+/* Return family of the device's CPU */
+extern AndroidCpuFamily   android_getCpuFamily(void);
+
+/* The list of feature flags for ARM CPUs that can be recognized by the
+ * library. Value details are:
+ *
+ *   VFPv2:
+ *     CPU supports the VFPv2 instruction set. Many, but not all, ARMv6 CPUs
+ *     support these instructions. VFPv2 is a subset of VFPv3 so this will
+ *     be set whenever VFPv3 is set too.
+ *
+ *   ARMv7:
+ *     CPU supports the ARMv7-A basic instruction set.
+ *     This feature is mandated by the 'armeabi-v7a' ABI.
+ *
+ *   VFPv3:
+ *     CPU supports the VFPv3-D16 instruction set, providing hardware FPU
+ *     support for single and double precision floating point registers.
+ *     Note that only 16 FPU registers are available by default, unless
+ *     the D32 bit is set too. This feature is also mandated by the
+ *     'armeabi-v7a' ABI.
+ *
+ *   VFP_D32:
+ *     CPU VFP optional extension that provides 32 FPU registers,
+ *     instead of 16. Note that ARM mandates this feature is the 'NEON'
+ *     feature is implemented by the CPU.
+ *
+ *   NEON:
+ *     CPU FPU supports "ARM Advanced SIMD" instructions, also known as
+ *     NEON. Note that this mandates the VFP_D32 feature as well, per the
+ *     ARM Architecture specification.
+ *
+ *   VFP_FP16:
+ *     Half-width floating precision VFP extension. If set, the CPU
+ *     supports instructions to perform floating-point operations on
+ *     16-bit registers. This is part of the VFPv4 specification, but
+ *     not mandated by any Android ABI.
+ *
+ *   VFP_FMA:
+ *     Fused multiply-accumulate VFP instructions extension. Also part of
+ *     the VFPv4 specification, but not mandated by any Android ABI.
+ *
+ *   NEON_FMA:
+ *     Fused multiply-accumulate NEON instructions extension. Optional
+ *     extension from the VFPv4 specification, but not mandated by any
+ *     Android ABI.
+ *
+ *   IDIV_ARM:
+ *     Integer division available in ARM mode. Only available
+ *     on recent CPUs (e.g. Cortex-A15).
+ *
+ *   IDIV_THUMB2:
+ *     Integer division available in Thumb-2 mode. Only available
+ *     on recent CPUs (e.g. Cortex-A15).
+ *
+ *   iWMMXt:
+ *     Optional extension that adds MMX registers and operations to an
+ *     ARM CPU. This is only available on a few XScale-based CPU designs
+ *     sold by Marvell. Pretty rare in practice.
+ *
+ * If you want to tell the compiler to generate code that targets one of
+ * the feature set above, you should probably use one of the following
+ * flags (for more details, see technical note at the end of this file):
+ *
+ *   -mfpu=vfp
+ *   -mfpu=vfpv2
+ *     These are equivalent and tell GCC to use VFPv2 instructions for
+ *     floating-point operations. Use this if you want your code to
+ *     run on *some* ARMv6 devices, and any ARMv7-A device supported
+ *     by Android.
+ *
+ *     Generated code requires VFPv2 feature.
+ *
+ *   -mfpu=vfpv3-d16
+ *     Tell GCC to use VFPv3 instructions (using only 16 FPU registers).
+ *     This should be generic code that runs on any CPU that supports the
+ *     'armeabi-v7a' Android ABI. Note that no ARMv6 CPU supports this.
+ *
+ *     Generated code requires VFPv3 feature.
+ *
+ *   -mfpu=vfpv3
+ *     Tell GCC to use VFPv3 instructions with 32 FPU registers.
+ *     Generated code requires VFPv3|VFP_D32 features.
+ *
+ *   -mfpu=neon
+ *     Tell GCC to use VFPv3 instructions with 32 FPU registers, and
+ *     also support NEON intrinsics (see <arm_neon.h>).
+ *     Generated code requires VFPv3|VFP_D32|NEON features.
+ *
+ *   -mfpu=vfpv4-d16
+ *     Generated code requires VFPv3|VFP_FP16|VFP_FMA features.
+ *
+ *   -mfpu=vfpv4
+ *     Generated code requires VFPv3|VFP_FP16|VFP_FMA|VFP_D32 features.
+ *
+ *   -mfpu=neon-vfpv4
+ *     Generated code requires VFPv3|VFP_FP16|VFP_FMA|VFP_D32|NEON|NEON_FMA
+ *     features.
+ *
+ *   -mcpu=cortex-a7
+ *   -mcpu=cortex-a15
+ *     Generated code requires VFPv3|VFP_FP16|VFP_FMA|VFP_D32|
+ *                             NEON|NEON_FMA|IDIV_ARM|IDIV_THUMB2
+ *     This flag implies -mfpu=neon-vfpv4.
+ *
+ *   -mcpu=iwmmxt
+ *     Allows the use of iWMMXt instrinsics with GCC.
+ */
+enum {
+    ANDROID_CPU_ARM_FEATURE_ARMv7       = (1 << 0),
+    ANDROID_CPU_ARM_FEATURE_VFPv3       = (1 << 1),
+    ANDROID_CPU_ARM_FEATURE_NEON        = (1 << 2),
+    ANDROID_CPU_ARM_FEATURE_LDREX_STREX = (1 << 3),
+    ANDROID_CPU_ARM_FEATURE_VFPv2       = (1 << 4),
+    ANDROID_CPU_ARM_FEATURE_VFP_D32     = (1 << 5),
+    ANDROID_CPU_ARM_FEATURE_VFP_FP16    = (1 << 6),
+    ANDROID_CPU_ARM_FEATURE_VFP_FMA     = (1 << 7),
+    ANDROID_CPU_ARM_FEATURE_NEON_FMA    = (1 << 8),
+    ANDROID_CPU_ARM_FEATURE_IDIV_ARM    = (1 << 9),
+    ANDROID_CPU_ARM_FEATURE_IDIV_THUMB2 = (1 << 10),
+    ANDROID_CPU_ARM_FEATURE_iWMMXt      = (1 << 11),
+};
+
+enum {
+    ANDROID_CPU_X86_FEATURE_SSSE3  = (1 << 0),
+    ANDROID_CPU_X86_FEATURE_POPCNT = (1 << 1),
+    ANDROID_CPU_X86_FEATURE_MOVBE  = (1 << 2),
+};
+
+// libtheoraplayer addition, renamed this to "Ext" as not to conflict with your own project if you've included cpu-features.c in it
+//extern uint64_t    android_getCpuFeaturesExt(void);
+#define android_getCpuFeaturesExt android_getCpuFeatures
+
+/* Return the number of CPU cores detected on this device. */
+extern int         android_getCpuCount(void);
+
+/* The following is used to force the CPU count and features
+ * mask in sandboxed processes. Under 4.1 and higher, these processes
+ * cannot access /proc, which is the only way to get information from
+ * the kernel about the current hardware (at least on ARM).
+ *
+ * It _must_ be called only once, and before any android_getCpuXXX
+ * function, any other case will fail.
+ *
+ * This function return 1 on success, and 0 on failure.
+ */
+extern int android_setCpu(int      cpu_count,
+                          uint64_t cpu_features);
+
+#ifdef __arm__
+/* Retrieve the ARM 32-bit CPUID value from the kernel.
+ * Note that this cannot work on sandboxed processes under 4.1 and
+ * higher, unless you called android_setCpuArm() before.
+ */
+extern uint32_t android_getCpuIdArm(void);
+
+/* An ARM-specific variant of android_setCpu() that also allows you
+ * to set the ARM CPUID field.
+ */
+extern int android_setCpuArm(int      cpu_count,
+                             uint64_t cpu_features,
+                             uint32_t cpu_id);
+#endif
+
+__END_DECLS
+
+#endif /* CPU_FEATURES_H */
diff --git a/drivers/theoraplayer/src/YUV/libyuv/LICENSE b/drivers/theoraplayer/src/YUV/libyuv/LICENSE
new file mode 100755
index 0000000000..c911747a6b
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/LICENSE
@@ -0,0 +1,29 @@
+Copyright 2011 The LibYuv Project Authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+  * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+
+  * Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in
+    the documentation and/or other materials provided with the
+    distribution.
+
+  * Neither the name of Google nor the names of its contributors may
+    be used to endorse or promote products derived from this software
+    without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/drivers/theoraplayer/src/YUV/libyuv/LICENSE_THIRD_PARTY b/drivers/theoraplayer/src/YUV/libyuv/LICENSE_THIRD_PARTY
new file mode 100755
index 0000000000..a71591e771
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/LICENSE_THIRD_PARTY
@@ -0,0 +1,8 @@
+This source tree contains third party source code which is governed by third
+party licenses. This file contains references to files which are under other
+licenses than the one provided in the LICENSE file in the root of the source
+tree.
+
+Files governed by third party licenses:
+source/x86inc.asm
+
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv.h
new file mode 100755
index 0000000000..3bebe642cc
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_H_  // NOLINT
+#define INCLUDE_LIBYUV_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/compare.h"
+#include "libyuv/convert.h"
+#include "libyuv/convert_argb.h"
+#include "libyuv/convert_from.h"
+#include "libyuv/convert_from_argb.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#include "libyuv/mjpeg_decoder.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/scale.h"
+#include "libyuv/scale_argb.h"
+#include "libyuv/scale_row.h"
+#include "libyuv/version.h"
+#include "libyuv/video_common.h"
+
+#endif  // INCLUDE_LIBYUV_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/basic_types.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/basic_types.h
new file mode 100755
index 0000000000..beb750ba65
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/basic_types.h
@@ -0,0 +1,118 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_  // NOLINT
+#define INCLUDE_LIBYUV_BASIC_TYPES_H_
+
+#include <stddef.h>  // for NULL, size_t
+
+#if defined(__ANDROID__) || (defined(_MSC_VER) && (_MSC_VER < 1600))
+#include <sys/types.h>  // for uintptr_t on x86
+#else
+#include <stdint.h>  // for uintptr_t
+#endif
+
+#ifndef GG_LONGLONG
+#ifndef INT_TYPES_DEFINED
+#define INT_TYPES_DEFINED
+#ifdef COMPILER_MSVC
+typedef unsigned __int64 uint64;
+typedef __int64 int64;
+#ifndef INT64_C
+#define INT64_C(x) x ## I64
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## UI64
+#endif
+#define INT64_F "I64"
+#else  // COMPILER_MSVC
+#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long uint64;  // NOLINT
+typedef long int64;  // NOLINT
+#ifndef INT64_C
+#define INT64_C(x) x ## L
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## UL
+#endif
+#define INT64_F "l"
+#else  // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long long uint64;  // NOLINT
+typedef long long int64;  // NOLINT
+#ifndef INT64_C
+#define INT64_C(x) x ## LL
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## ULL
+#endif
+#define INT64_F "ll"
+#endif  // __LP64__
+#endif  // COMPILER_MSVC
+typedef unsigned int uint32;
+typedef int int32;
+typedef unsigned short uint16;  // NOLINT
+typedef short int16;  // NOLINT
+typedef unsigned char uint8;
+typedef signed char int8;
+#endif  // INT_TYPES_DEFINED
+#endif  // GG_LONGLONG
+
+// Detect compiler is for x86 or x64.
+#if defined(__x86_64__) || defined(_M_X64) || \
+    defined(__i386__) || defined(_M_IX86)
+#define CPU_X86 1
+#endif
+// Detect compiler is for ARM.
+#if defined(__arm__) || defined(_M_ARM)
+#define CPU_ARM 1
+#endif
+
+#ifndef ALIGNP
+#ifdef __cplusplus
+#define ALIGNP(p, t) \
+    (reinterpret_cast<uint8*>(((reinterpret_cast<uintptr_t>(p) + \
+    ((t) - 1)) & ~((t) - 1))))
+#else
+#define ALIGNP(p, t) \
+    ((uint8*)((((uintptr_t)(p) + ((t) - 1)) & ~((t) - 1))))  /* NOLINT */
+#endif
+#endif
+
+#if !defined(LIBYUV_API)
+#if defined(_WIN32) || defined(__CYGWIN__)
+#if defined(LIBYUV_BUILDING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllexport)
+#elif defined(LIBYUV_USING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllimport)
+#else
+#define LIBYUV_API
+#endif  // LIBYUV_BUILDING_SHARED_LIBRARY
+#elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__APPLE__) && \
+    (defined(LIBYUV_BUILDING_SHARED_LIBRARY) || \
+    defined(LIBYUV_USING_SHARED_LIBRARY))
+#define LIBYUV_API __attribute__ ((visibility ("default")))
+#else
+#define LIBYUV_API
+#endif  // __GNUC__
+#endif  // LIBYUV_API
+
+#define LIBYUV_BOOL int
+#define LIBYUV_FALSE 0
+#define LIBYUV_TRUE 1
+
+// Visual C x86 or GCC little endian.
+#if defined(__x86_64__) || defined(_M_X64) || \
+  defined(__i386__) || defined(_M_IX86) || \
+  defined(__arm__) || defined(_M_ARM) || \
+  (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#define LIBYUV_LITTLE_ENDIAN
+#endif
+
+#endif  // INCLUDE_LIBYUV_BASIC_TYPES_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/compare.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/compare.h
new file mode 100755
index 0000000000..5dfac7c86a
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/compare.h
@@ -0,0 +1,73 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_COMPARE_H_  // NOLINT
+#define INCLUDE_LIBYUV_COMPARE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Compute a hash for specified memory. Seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed);
+
+// Sum Square Error - used to compute Mean Square Error or PSNR.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a,
+                             const uint8* src_b, int count);
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
+                                  const uint8* src_b, int stride_b,
+                                  int width, int height);
+
+static const int kMaxPsnr = 128;
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count);
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height);
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height);
+
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height);
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_COMPARE_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert.h
new file mode 100755
index 0000000000..1bd45c837f
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert.h
@@ -0,0 +1,254 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_H_
+
+#include "libyuv/basic_types.h"
+// TODO(fbarchard): Remove the following headers includes.
+#include "libyuv/convert_from.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert I444 to I420.
+LIBYUV_API
+int I444ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I422 to I420.
+LIBYUV_API
+int I422ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I411 to I420.
+LIBYUV_API
+int I411ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Copy I420 to I420.
+#define I420ToI420 I420Copy
+LIBYUV_API
+int I420Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert I400 (grey) to I420.
+LIBYUV_API
+int I400ToI420(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert NV12 to I420.
+LIBYUV_API
+int NV12ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert NV21 to I420.
+LIBYUV_API
+int NV21ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert YUY2 to I420.
+LIBYUV_API
+int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I420.
+LIBYUV_API
+int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert M420 to I420.
+LIBYUV_API
+int M420ToI420(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert Q420 to I420.
+LIBYUV_API
+int Q420ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// ARGB little endian (bgra in memory) to I420.
+LIBYUV_API
+int ARGBToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// BGRA little endian (argb in memory) to I420.
+LIBYUV_API
+int BGRAToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// ABGR little endian (rgba in memory) to I420.
+LIBYUV_API
+int ABGRToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// RGBA little endian (abgr in memory) to I420.
+LIBYUV_API
+int RGBAToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// RGB little endian (bgr in memory) to I420.
+LIBYUV_API
+int RGB24ToI420(const uint8* src_frame, int src_stride_frame,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height);
+
+// RGB big endian (rgb in memory) to I420.
+LIBYUV_API
+int RAWToI420(const uint8* src_frame, int src_stride_frame,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int width, int height);
+
+// RGB16 (RGBP fourcc) little endian to I420.
+LIBYUV_API
+int RGB565ToI420(const uint8* src_frame, int src_stride_frame,
+                 uint8* dst_y, int dst_stride_y,
+                 uint8* dst_u, int dst_stride_u,
+                 uint8* dst_v, int dst_stride_v,
+                 int width, int height);
+
+// RGB15 (RGBO fourcc) little endian to I420.
+LIBYUV_API
+int ARGB1555ToI420(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height);
+
+// RGB12 (R444 fourcc) little endian to I420.
+LIBYUV_API
+int ARGB4444ToI420(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height);
+
+#ifdef HAVE_JPEG
+// src_width/height provided by capture.
+// dst_width/height for clipping determine final size.
+LIBYUV_API
+int MJPGToI420(const uint8* sample, size_t sample_size,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int src_width, int src_height,
+               int dst_width, int dst_height);
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size,
+             int* width, int* height);
+#endif
+
+// Note Bayer formats (BGGR) To I420 are in format_conversion.h
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// "src_size" is needed to parse MJPG.
+// "dst_stride_y" number of bytes in a row of the dst_y plane.
+//   Normally this would be the same as dst_width, with recommended alignment
+//   to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected. The caller should
+//   allocate the I420 buffer according to rotation.
+// "dst_stride_u" number of bytes in a row of the dst_u plane.
+//   Normally this would be the same as (dst_width + 1) / 2, with
+//   recommended alignment to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected.
+// "crop_x" and "crop_y" are starting position for cropping.
+//   To center, crop_x = (src_width - dst_width) / 2
+//              crop_y = (src_height - dst_height) / 2
+// "src_width" / "src_height" is size of src_frame in pixels.
+//   "src_height" can be negative indicating a vertically flipped image source.
+// "crop_width" / "crop_height" is the size to crop the src to.
+//    Must be less than or equal to src_width/src_height
+//    Cropping parameters are pre-rotation.
+// "rotation" can be 0, 90, 180 or 270.
+// "format" is a fourcc. ie 'I420', 'YUY2'
+// Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
+LIBYUV_API
+int ConvertToI420(const uint8* src_frame, size_t src_size,
+                  uint8* dst_y, int dst_stride_y,
+                  uint8* dst_u, int dst_stride_u,
+                  uint8* dst_v, int dst_stride_v,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_argb.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_argb.h
new file mode 100755
index 0000000000..a18014ca2c
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_argb.h
@@ -0,0 +1,225 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_ARGB_H_
+
+#include "libyuv/basic_types.h"
+// TODO(fbarchard): Remove the following headers includes
+#include "libyuv/convert_from.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+
+// TODO(fbarchard): This set of functions should exactly match convert.h
+// Add missing Q420.
+// TODO(fbarchard): Add tests. Create random content of right size and convert
+// with C vs Opt and or to I420 and compare.
+// TODO(fbarchard): Some of these functions lack parameter setting.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Alias.
+#define ARGBToARGB ARGBCopy
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I411 to ARGB.
+LIBYUV_API
+int I411ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I400 (grey) to ARGB.
+LIBYUV_API
+int I400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Alias.
+#define YToARGB I400ToARGB_Reference
+
+// Convert I400 to ARGB. Reverse of ARGBToI400.
+LIBYUV_API
+int I400ToARGB_Reference(const uint8* src_y, int src_stride_y,
+                         uint8* dst_argb, int dst_stride_argb,
+                         int width, int height);
+
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert M420 to ARGB.
+LIBYUV_API
+int M420ToARGB(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// TODO(fbarchard): Convert Q420 to ARGB.
+// LIBYUV_API
+// int Q420ToARGB(const uint8* src_y, int src_stride_y,
+//                const uint8* src_yuy2, int src_stride_yuy2,
+//                uint8* dst_argb, int dst_stride_argb,
+//                int width, int height);
+
+// Convert YUY2 to ARGB.
+LIBYUV_API
+int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert UYVY to ARGB.
+LIBYUV_API
+int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// BGRA little endian (argb in memory) to ARGB.
+LIBYUV_API
+int BGRAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// ABGR little endian (rgba in memory) to ARGB.
+LIBYUV_API
+int ABGRToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// RGBA little endian (abgr in memory) to ARGB.
+LIBYUV_API
+int RGBAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Deprecated function name.
+#define BG24ToARGB RGB24ToARGB
+
+// RGB little endian (bgr in memory) to ARGB.
+LIBYUV_API
+int RGB24ToARGB(const uint8* src_frame, int src_stride_frame,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+// RGB big endian (rgb in memory) to ARGB.
+LIBYUV_API
+int RAWToARGB(const uint8* src_frame, int src_stride_frame,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// RGB16 (RGBP fourcc) little endian to ARGB.
+LIBYUV_API
+int RGB565ToARGB(const uint8* src_frame, int src_stride_frame,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// RGB15 (RGBO fourcc) little endian to ARGB.
+LIBYUV_API
+int ARGB1555ToARGB(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height);
+
+// RGB12 (R444 fourcc) little endian to ARGB.
+LIBYUV_API
+int ARGB4444ToARGB(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height);
+
+#ifdef HAVE_JPEG
+// src_width/height provided by capture
+// dst_width/height for clipping determine final size.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* dst_argb, int dst_stride_argb,
+               int src_width, int src_height,
+               int dst_width, int dst_height);
+#endif
+
+// Note Bayer formats (BGGR) to ARGB are in format_conversion.h.
+
+// Convert camera sample to ARGB with cropping, rotation and vertical flip.
+// "src_size" is needed to parse MJPG.
+// "dst_stride_argb" number of bytes in a row of the dst_argb plane.
+//   Normally this would be the same as dst_width, with recommended alignment
+//   to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected. The caller should
+//   allocate the I420 buffer according to rotation.
+// "dst_stride_u" number of bytes in a row of the dst_u plane.
+//   Normally this would be the same as (dst_width + 1) / 2, with
+//   recommended alignment to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected.
+// "crop_x" and "crop_y" are starting position for cropping.
+//   To center, crop_x = (src_width - dst_width) / 2
+//              crop_y = (src_height - dst_height) / 2
+// "src_width" / "src_height" is size of src_frame in pixels.
+//   "src_height" can be negative indicating a vertically flipped image source.
+// "crop_width" / "crop_height" is the size to crop the src to.
+//    Must be less than or equal to src_width/src_height
+//    Cropping parameters are pre-rotation.
+// "rotation" can be 0, 90, 180 or 270.
+// "format" is a fourcc. ie 'I420', 'YUY2'
+// Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
+LIBYUV_API
+int ConvertToARGB(const uint8* src_frame, size_t src_size,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_ARGB_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_from.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_from.h
new file mode 100755
index 0000000000..b1cf57f7dc
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_from.h
@@ -0,0 +1,173 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_FROM_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/rotate.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// See Also convert.h for conversions from formats to I420.
+
+// I420Copy in convert to I420ToI420.
+
+LIBYUV_API
+int I420ToI422(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int I420ToI444(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int I420ToI411(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Copy to I400. Source can be I420, I422, I444, I400, NV12 or NV21.
+LIBYUV_API
+int I400Copy(const uint8* src_y, int src_stride_y,
+             uint8* dst_y, int dst_stride_y,
+             int width, int height);
+
+// TODO(fbarchard): I420ToM420
+// TODO(fbarchard): I420ToQ420
+
+LIBYUV_API
+int I420ToNV12(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+LIBYUV_API
+int I420ToNV21(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+LIBYUV_API
+int I420ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+LIBYUV_API
+int I420ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+LIBYUV_API
+int I420ToRGB24(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_frame, int dst_stride_frame,
+                int width, int height);
+
+LIBYUV_API
+int I420ToRAW(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              uint8* dst_frame, int dst_stride_frame,
+              int width, int height);
+
+LIBYUV_API
+int I420ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_u, int src_stride_u,
+                 const uint8* src_v, int src_stride_v,
+                 uint8* dst_frame, int dst_stride_frame,
+                 int width, int height);
+
+LIBYUV_API
+int I420ToARGB1555(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_frame, int dst_stride_frame,
+                   int width, int height);
+
+LIBYUV_API
+int I420ToARGB4444(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_frame, int dst_stride_frame,
+                   int width, int height);
+
+// Note Bayer formats (BGGR) To I420 are in format_conversion.h.
+
+// Convert I420 to specified format.
+// "dst_sample_stride" is bytes in a row for the destination. Pass 0 if the
+//    buffer has contiguous rows. Can be negative. A multiple of 16 is optimal.
+LIBYUV_API
+int ConvertFromI420(const uint8* y, int y_stride,
+                    const uint8* u, int u_stride,
+                    const uint8* v, int v_stride,
+                    uint8* dst_sample, int dst_sample_stride,
+                    int width, int height,
+                    uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_from_argb.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_from_argb.h
new file mode 100755
index 0000000000..f0343a77d3
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/convert_from_argb.h
@@ -0,0 +1,168 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy ARGB to ARGB.
+#define ARGBToARGB ARGBCopy
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Convert ARGB To BGRA. (alias)
+#define ARGBToBGRA BGRAToARGB
+LIBYUV_API
+int BGRAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert ARGB To ABGR. (alias)
+#define ARGBToABGR ABGRToARGB
+LIBYUV_API
+int ABGRToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert ARGB To RGBA.
+LIBYUV_API
+int ARGBToRGBA(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert ARGB To RGB24.
+LIBYUV_API
+int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height);
+
+// Convert ARGB To RAW.
+LIBYUV_API
+int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_rgb, int dst_stride_rgb,
+              int width, int height);
+
+// Convert ARGB To RGB565.
+LIBYUV_API
+int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert ARGB To ARGB1555.
+LIBYUV_API
+int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height);
+
+// Convert ARGB To ARGB4444.
+LIBYUV_API
+int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height);
+
+// Convert ARGB To I444.
+LIBYUV_API
+int ARGBToI444(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I422.
+LIBYUV_API
+int ARGBToI422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I420. (also in convert.h)
+LIBYUV_API
+int ARGBToI420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J420. (JPeg full range I420).
+LIBYUV_API
+int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I411.
+LIBYUV_API
+int ARGBToI411(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J400. (JPeg full range).
+LIBYUV_API
+int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               int width, int height);
+
+// Convert ARGB to I400.
+LIBYUV_API
+int ARGBToI400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Convert ARGB To NV12.
+LIBYUV_API
+int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+// Convert ARGB To NV21.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+// Convert ARGB To NV21.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+// Convert ARGB To YUY2.
+LIBYUV_API
+int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height);
+
+// Convert ARGB To UYVY.
+LIBYUV_API
+int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/cpu_id.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/cpu_id.h
new file mode 100755
index 0000000000..dc858a814a
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/cpu_id.h
@@ -0,0 +1,81 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_CPU_ID_H_  // NOLINT
+#define INCLUDE_LIBYUV_CPU_ID_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// TODO(fbarchard): Consider overlapping bits for different architectures.
+// Internal flag to indicate cpuid requires initialization.
+#define kCpuInit 0x1
+
+// These flags are only valid on ARM processors.
+static const int kCpuHasARM = 0x2;
+static const int kCpuHasNEON = 0x4;
+// 0x8 reserved for future ARM flag.
+
+// These flags are only valid on x86 processors.
+static const int kCpuHasX86 = 0x10;
+static const int kCpuHasSSE2 = 0x20;
+static const int kCpuHasSSSE3 = 0x40;
+static const int kCpuHasSSE41 = 0x80;
+static const int kCpuHasSSE42 = 0x100;
+static const int kCpuHasAVX = 0x200;
+static const int kCpuHasAVX2 = 0x400;
+static const int kCpuHasERMS = 0x800;
+static const int kCpuHasFMA3 = 0x1000;
+// 0x2000, 0x4000, 0x8000 reserved for future X86 flags.
+
+// These flags are only valid on MIPS processors.
+static const int kCpuHasMIPS = 0x10000;
+static const int kCpuHasMIPS_DSP = 0x20000;
+static const int kCpuHasMIPS_DSPR2 = 0x40000;
+
+// Internal function used to auto-init.
+LIBYUV_API
+int InitCpuFlags(void);
+
+// Internal function for parsing /proc/cpuinfo.
+LIBYUV_API
+int ArmCpuCaps(const char* cpuinfo_name);
+
+// Detect CPU has SSE2 etc.
+// Test_flag parameter should be one of kCpuHas constants above.
+// returns non-zero if instruction set is detected
+static __inline int TestCpuFlag(int test_flag) {
+  LIBYUV_API extern int cpu_info_;
+  return (cpu_info_ == kCpuInit ? InitCpuFlags() : cpu_info_) & test_flag;
+}
+
+// For testing, allow CPU flags to be disabled.
+// ie MaskCpuFlags(~kCpuHasSSSE3) to disable SSSE3.
+// MaskCpuFlags(-1) to enable all cpu specific optimizations.
+// MaskCpuFlags(0) to disable all cpu specific optimizations.
+LIBYUV_API
+void MaskCpuFlags(int enable_flags);
+
+// Low level cpuid for X86. Returns zeros on other CPUs.
+// eax is the info type that you want.
+// ecx is typically the cpu number, and should normally be zero.
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CPU_ID_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/format_conversion.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/format_conversion.h
new file mode 100755
index 0000000000..b18bf05343
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/format_conversion.h
@@ -0,0 +1,168 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_FORMATCONVERSION_H_  // NOLINT
+#define INCLUDE_LIBYUV_FORMATCONVERSION_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert Bayer RGB formats to I420.
+LIBYUV_API
+int BayerBGGRToI420(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_y, int dst_stride_y,
+                    uint8* dst_u, int dst_stride_u,
+                    uint8* dst_v, int dst_stride_v,
+                    int width, int height);
+
+LIBYUV_API
+int BayerGBRGToI420(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_y, int dst_stride_y,
+                    uint8* dst_u, int dst_stride_u,
+                    uint8* dst_v, int dst_stride_v,
+                    int width, int height);
+
+LIBYUV_API
+int BayerGRBGToI420(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_y, int dst_stride_y,
+                    uint8* dst_u, int dst_stride_u,
+                    uint8* dst_v, int dst_stride_v,
+                    int width, int height);
+
+LIBYUV_API
+int BayerRGGBToI420(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_y, int dst_stride_y,
+                    uint8* dst_u, int dst_stride_u,
+                    uint8* dst_v, int dst_stride_v,
+                    int width, int height);
+
+// Temporary API mapper.
+#define BayerRGBToI420(b, bs, f, y, ys, u, us, v, vs, w, h) \
+    BayerToI420(b, bs, y, ys, u, us, v, vs, w, h, f)
+
+LIBYUV_API
+int BayerToI420(const uint8* src_bayer, int src_stride_bayer,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height,
+                uint32 src_fourcc_bayer);
+
+// Convert I420 to Bayer RGB formats.
+LIBYUV_API
+int I420ToBayerBGGR(const uint8* src_y, int src_stride_y,
+                    const uint8* src_u, int src_stride_u,
+                    const uint8* src_v, int src_stride_v,
+                    uint8* dst_frame, int dst_stride_frame,
+                    int width, int height);
+
+LIBYUV_API
+int I420ToBayerGBRG(const uint8* src_y, int src_stride_y,
+                    const uint8* src_u, int src_stride_u,
+                    const uint8* src_v, int src_stride_v,
+                    uint8* dst_frame, int dst_stride_frame,
+                    int width, int height);
+
+LIBYUV_API
+int I420ToBayerGRBG(const uint8* src_y, int src_stride_y,
+                    const uint8* src_u, int src_stride_u,
+                    const uint8* src_v, int src_stride_v,
+                    uint8* dst_frame, int dst_stride_frame,
+                    int width, int height);
+
+LIBYUV_API
+int I420ToBayerRGGB(const uint8* src_y, int src_stride_y,
+                    const uint8* src_u, int src_stride_u,
+                    const uint8* src_v, int src_stride_v,
+                    uint8* dst_frame, int dst_stride_frame,
+                    int width, int height);
+
+// Temporary API mapper.
+#define I420ToBayerRGB(y, ys, u, us, v, vs, b, bs, f, w, h) \
+    I420ToBayer(y, ys, u, us, v, vs, b, bs, w, h, f)
+
+LIBYUV_API
+int I420ToBayer(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_frame, int dst_stride_frame,
+                int width, int height,
+                uint32 dst_fourcc_bayer);
+
+// Convert Bayer RGB formats to ARGB.
+LIBYUV_API
+int BayerBGGRToARGB(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+LIBYUV_API
+int BayerGBRGToARGB(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+LIBYUV_API
+int BayerGRBGToARGB(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+LIBYUV_API
+int BayerRGGBToARGB(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+// Temporary API mapper.
+#define BayerRGBToARGB(b, bs, f, a, as, w, h) BayerToARGB(b, bs, a, as, w, h, f)
+
+LIBYUV_API
+int BayerToARGB(const uint8* src_bayer, int src_stride_bayer,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height,
+                uint32 src_fourcc_bayer);
+
+// Converts ARGB to Bayer RGB formats.
+LIBYUV_API
+int ARGBToBayerBGGR(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_bayer, int dst_stride_bayer,
+                    int width, int height);
+
+LIBYUV_API
+int ARGBToBayerGBRG(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_bayer, int dst_stride_bayer,
+                    int width, int height);
+
+LIBYUV_API
+int ARGBToBayerGRBG(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_bayer, int dst_stride_bayer,
+                    int width, int height);
+
+LIBYUV_API
+int ARGBToBayerRGGB(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_bayer, int dst_stride_bayer,
+                    int width, int height);
+
+// Temporary API mapper.
+#define ARGBToBayerRGB(a, as, b, bs, f, w, h) ARGBToBayer(b, bs, a, as, w, h, f)
+
+LIBYUV_API
+int ARGBToBayer(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_bayer, int dst_stride_bayer,
+                int width, int height,
+                uint32 dst_fourcc_bayer);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_FORMATCONVERSION_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/mjpeg_decoder.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/mjpeg_decoder.h
new file mode 100755
index 0000000000..faffaea8fa
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/mjpeg_decoder.h
@@ -0,0 +1,201 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_MJPEG_DECODER_H_  // NOLINT
+#define INCLUDE_LIBYUV_MJPEG_DECODER_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+// NOTE: For a simplified public API use convert.h MJPGToI420().
+
+struct jpeg_common_struct;
+struct jpeg_decompress_struct;
+struct jpeg_source_mgr;
+
+namespace libyuv {
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+static const uint32 kUnknownDataSize = 0xFFFFFFFF;
+
+enum JpegSubsamplingType {
+  kJpegYuv420,
+  kJpegYuv422,
+  kJpegYuv411,
+  kJpegYuv444,
+  kJpegYuv400,
+  kJpegUnknown
+};
+
+struct SetJmpErrorMgr;
+
+// MJPEG ("Motion JPEG") is a pseudo-standard video codec where the frames are
+// simply independent JPEG images with a fixed huffman table (which is omitted).
+// It is rarely used in video transmission, but is common as a camera capture
+// format, especially in Logitech devices. This class implements a decoder for
+// MJPEG frames.
+//
+// See http://tools.ietf.org/html/rfc2435
+class LIBYUV_API MJpegDecoder {
+ public:
+  typedef void (*CallbackFunction)(void* opaque,
+                                   const uint8* const* data,
+                                   const int* strides,
+                                   int rows);
+
+  static const int kColorSpaceUnknown;
+  static const int kColorSpaceGrayscale;
+  static const int kColorSpaceRgb;
+  static const int kColorSpaceYCbCr;
+  static const int kColorSpaceCMYK;
+  static const int kColorSpaceYCCK;
+
+  MJpegDecoder();
+  ~MJpegDecoder();
+
+  // Loads a new frame, reads its headers, and determines the uncompressed
+  // image format.
+  // Returns LIBYUV_TRUE if image looks valid and format is supported.
+  // If return value is LIBYUV_TRUE, then the values for all the following
+  // getters are populated.
+  // src_len is the size of the compressed mjpeg frame in bytes.
+  LIBYUV_BOOL LoadFrame(const uint8* src, size_t src_len);
+
+  // Returns width of the last loaded frame in pixels.
+  int GetWidth();
+
+  // Returns height of the last loaded frame in pixels.
+  int GetHeight();
+
+  // Returns format of the last loaded frame. The return value is one of the
+  // kColorSpace* constants.
+  int GetColorSpace();
+
+  // Number of color components in the color space.
+  int GetNumComponents();
+
+  // Sample factors of the n-th component.
+  int GetHorizSampFactor(int component);
+
+  int GetVertSampFactor(int component);
+
+  int GetHorizSubSampFactor(int component);
+
+  int GetVertSubSampFactor(int component);
+
+  // Public for testability.
+  int GetImageScanlinesPerImcuRow();
+
+  // Public for testability.
+  int GetComponentScanlinesPerImcuRow(int component);
+
+  // Width of a component in bytes.
+  int GetComponentWidth(int component);
+
+  // Height of a component.
+  int GetComponentHeight(int component);
+
+  // Width of a component in bytes with padding for DCTSIZE. Public for testing.
+  int GetComponentStride(int component);
+
+  // Size of a component in bytes.
+  int GetComponentSize(int component);
+
+  // Call this after LoadFrame() if you decide you don't want to decode it
+  // after all.
+  LIBYUV_BOOL UnloadFrame();
+
+  // Decodes the entire image into a one-buffer-per-color-component format.
+  // dst_width must match exactly. dst_height must be <= to image height; if
+  // less, the image is cropped. "planes" must have size equal to at least
+  // GetNumComponents() and they must point to non-overlapping buffers of size
+  // at least GetComponentSize(i). The pointers in planes are incremented
+  // to point to after the end of the written data.
+  // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
+  LIBYUV_BOOL DecodeToBuffers(uint8** planes, int dst_width, int dst_height);
+
+  // Decodes the entire image and passes the data via repeated calls to a
+  // callback function. Each call will get the data for a whole number of
+  // image scanlines.
+  // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
+  LIBYUV_BOOL DecodeToCallback(CallbackFunction fn, void* opaque,
+                        int dst_width, int dst_height);
+
+  // The helper function which recognizes the jpeg sub-sampling type.
+  static JpegSubsamplingType JpegSubsamplingTypeHelper(
+     int* subsample_x, int* subsample_y, int number_of_components);
+
+ private:
+  struct Buffer {
+    const uint8* data;
+    int len;
+  };
+
+  struct BufferVector {
+    Buffer* buffers;
+    int len;
+    int pos;
+  };
+
+  // Methods that are passed to jpeglib.
+  static int fill_input_buffer(jpeg_decompress_struct* cinfo);
+  static void init_source(jpeg_decompress_struct* cinfo);
+  static void skip_input_data(jpeg_decompress_struct* cinfo,
+                              long num_bytes);  // NOLINT
+  static void term_source(jpeg_decompress_struct* cinfo);
+
+  static void ErrorHandler(jpeg_common_struct* cinfo);
+
+  void AllocOutputBuffers(int num_outbufs);
+  void DestroyOutputBuffers();
+
+  LIBYUV_BOOL StartDecode();
+  LIBYUV_BOOL FinishDecode();
+
+  void SetScanlinePointers(uint8** data);
+  LIBYUV_BOOL DecodeImcuRow();
+
+  int GetComponentScanlinePadding(int component);
+
+  // A buffer holding the input data for a frame.
+  Buffer buf_;
+  BufferVector buf_vec_;
+
+  jpeg_decompress_struct* decompress_struct_;
+  jpeg_source_mgr* source_mgr_;
+  SetJmpErrorMgr* error_mgr_;
+
+  // LIBYUV_TRUE iff at least one component has scanline padding. (i.e.,
+  // GetComponentScanlinePadding() != 0.)
+  LIBYUV_BOOL has_scanline_padding_;
+
+  // Temporaries used to point to scanline outputs.
+  int num_outbufs_;  // Outermost size of all arrays below.
+  uint8*** scanlines_;
+  int* scanlines_sizes_;
+  // Temporary buffer used for decoding when we can't decode directly to the
+  // output buffers. Large enough for just one iMCU row.
+  uint8** databuf_;
+  int* databuf_strides_;
+};
+
+}  // namespace libyuv
+
+#endif  //  __cplusplus
+#endif  // INCLUDE_LIBYUV_MJPEG_DECODER_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/planar_functions.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/planar_functions.h
new file mode 100755
index 0000000000..ac516c5ba5
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/planar_functions.h
@@ -0,0 +1,434 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  // NOLINT
+#define INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_
+
+#include "libyuv/basic_types.h"
+
+// TODO(fbarchard): Remove the following headers includes.
+#include "libyuv/convert.h"
+#include "libyuv/convert_argb.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data.
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Set a plane of data to a 32 bit value.
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value);
+
+// Copy I400.  Supports inverting.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+
+// Copy I422 to I422.
+#define I422ToI422 I422Copy
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Copy I444 to I444.
+#define I444ToI444 I444Copy
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I420 to I400. (calls CopyPlane ignoring u/v).
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define I420ToI420Mirror I420Mirror
+
+// I420 mirror.
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Alias
+#define I400ToI400Mirror I400Mirror
+
+// I400 mirror.  A single plane is mirrored horizontally.
+// Pass negative height to achieve 180 degree rotation.
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define ARGBToARGBMirror ARGBMirror
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// I422ToARGB is in convert_argb.h
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height);
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+// Draw a rectangle into I420.
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y, int width, int height,
+             int value_y, int value_u, int value_v);
+
+// Draw a rectangle into ARGB.
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height, uint32 value);
+
+// Convert ARGB to gray scale ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height);
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int x, int y, int width, int height);
+
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 4 signed ARGB values. -128 to 127 representing -2 to 2.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The next 4 coefficients apply to B, G, R, A and produce R of the output.
+// The last 4 coefficients apply to B, G, R, A and produce A of the output.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height);
+
+// Deprecated. Use ARGBColorMatrix instead.
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 3 signed ARGB values. -128 to 127 representing -1 to 1.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The last 4 coefficients apply to B, G, R, A and produce R of the output.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int x, int y, int width, int height);
+
+// Apply a luma/color table each ARGB pixel but preserve destination alpha.
+// Table contains 32768 values indexed by [Y][C] where 7 it 7 bit luma from
+// RGB (YJ style) and C is an 8 bit color component (R, G or B).
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma_rgb_table,
+                       int width, int height);
+
+// Apply a 3 term polynomial to ARGB values.
+// poly points to a 4x4 matrix.  The first row is constants.  The 2nd row is
+// coefficients for b, g, r and a.  The 3rd row is coefficients for b squared,
+// g squared, r squared and a squared.  The 4rd row is coefficients for b to
+// the 3, g to the 3, r to the 3 and a to the 3.  The values are summed and
+// result clamped to 0 to 255.
+// A polynomial approximation can be dirived using software such as 'R'.
+
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height);
+
+// Quantize a rectangle of ARGB. Alpha unaffected.
+// scale is a 16 bit fractional fixed point scaler between 0 and 65535.
+// interval_size should be a value between 1 and 255.
+// interval_offset should be a value between 0 and 255.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int x, int y, int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height);
+
+typedef void (*ARGBBlendRow)(const uint8* src_argb0, const uint8* src_argb1,
+                             uint8* dst_argb, int width);
+
+// Get function to Alpha Blend ARGB pixels and store to destination.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend();
+
+// Alpha Blend ARGB images and store to destination.
+// Alpha of destination is set to 255.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Multiply ARGB image by ARGB image. Shifted down by 8. Saturates to 255.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Add ARGB image with ARGB image. Saturates to 255.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height);
+
+// Subtract ARGB image (argb1) from ARGB image (argb0). Saturates to 0.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Convert I422 to YUY2.
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert I422 to UYVY.
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+// Convert MJPG to ARGB.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h, int dw, int dh);
+
+// Internal function - do not call directly.
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height);
+
+// Blur ARGB image.
+// dst_cumsum table of width * (height + 1) * 16 bytes aligned to
+//   16 byte boundary.
+// dst_stride32_cumsum is number of ints in a row (width * 4).
+// radius is number of pixels around the center.  e.g. 1 = 3x3. 2=5x5.
+// Blur is optimized for radius of 5 (11x11) or less.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius);
+
+// Multiply ARGB image by ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value);
+
+// Interpolate between two ARGB images using specified amount of interpolation
+// (0 to 255) and store to destination.
+// 'interpolation' is specified as 8 bit fraction where 0 means 100% src_argb0
+// and 255 means 1% src_argb0 and 99% src_argb1.
+// Internally uses ARGBScale bilinear filtering.
+// Caveat: This function will write up to 16 bytes beyond the end of dst_argb.
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation);
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR)
+#define LIBYUV_DISABLE_X86
+#endif
+
+// Row functions for copying a pixels from a source with a slope to a row
+// of destination. Useful for scaling, rotation, mirror, texture mapping.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+#define HAS_ARGBAFFINEROW_SSE2
+#endif  // LIBYUV_DISABLE_X86
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+// shuffler is 16 bytes and must be aligned.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height);
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height);
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Sobel ARGB effect w/ Sobel X, Sobel, Sobel Y in ARGB.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/rotate.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/rotate.h
new file mode 100755
index 0000000000..8af60b8955
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/rotate.h
@@ -0,0 +1,117 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Supported rotation.
+typedef enum RotationMode {
+  kRotate0 = 0,  // No rotation.
+  kRotate90 = 90,  // Rotate 90 degrees clockwise.
+  kRotate180 = 180,  // Rotate 180 degrees.
+  kRotate270 = 270,  // Rotate 270 degrees clockwise.
+
+  // Deprecated.
+  kRotateNone = 0,
+  kRotateClockwise = 90,
+  kRotateCounterClockwise = 270,
+} RotationModeEnum;
+
+// Rotate I420 frame.
+LIBYUV_API
+int I420Rotate(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int src_width, int src_height, enum RotationMode mode);
+
+// Rotate NV12 input and store in I420.
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
+                     const uint8* src_uv, int src_stride_uv,
+                     uint8* dst_y, int dst_stride_y,
+                     uint8* dst_u, int dst_stride_u,
+                     uint8* dst_v, int dst_stride_v,
+                     int src_width, int src_height, enum RotationMode mode);
+
+// Rotate a plane by 0, 90, 180, or 270.
+LIBYUV_API
+int RotatePlane(const uint8* src, int src_stride,
+                uint8* dst, int dst_stride,
+                int src_width, int src_height, enum RotationMode mode);
+
+// Rotate planes by 90, 180, 270. Deprecated.
+LIBYUV_API
+void RotatePlane90(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height);
+
+LIBYUV_API
+void RotatePlane180(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void RotatePlane270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void RotateUV90(const uint8* src, int src_stride,
+                uint8* dst_a, int dst_stride_a,
+                uint8* dst_b, int dst_stride_b,
+                int width, int height);
+
+// Rotations for when U and V are interleaved.
+// These functions take one input pointer and
+// split the data into two buffers while
+// rotating them. Deprecated.
+LIBYUV_API
+void RotateUV180(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+LIBYUV_API
+void RotateUV270(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+// The 90 and 270 functions are based on transposes.
+// Doing a transpose with reversing the read/write
+// order will result in a rotation by +- 90 degrees.
+// Deprecated.
+LIBYUV_API
+void TransposePlane(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void TransposeUV(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/rotate_argb.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/rotate_argb.h
new file mode 100755
index 0000000000..660ff5573e
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/rotate_argb.h
@@ -0,0 +1,33 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_ARGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/rotate.h"  // For RotationMode.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Rotate ARGB frame
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int src_width, int src_height, enum RotationMode mode);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_ARGB_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/row.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/row.h
new file mode 100755
index 0000000000..757020da86
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/row.h
@@ -0,0 +1,1694 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROW_H_
+
+#include <stdlib.h>  // For malloc.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
+
+#ifdef __cplusplus
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = reinterpret_cast<uint8*>(malloc((size) + 63));            \
+  uint8* var = reinterpret_cast<uint8*>                                        \
+      ((reinterpret_cast<intptr_t>(var##_mem) + 63) & ~63)
+#else
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = (uint8*)(malloc((size) + 63));               /* NOLINT */ \
+  uint8* var = (uint8*)(((intptr_t)(var##_mem) + 63) & ~63)       /* NOLINT */
+#endif
+
+#define free_aligned_buffer_64(var) \
+  free(var##_mem);  \
+  var = 0
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR)
+#define LIBYUV_DISABLE_X86
+#endif
+// True if compiling for SSSE3 as a requirement.
+#if defined(__SSSE3__) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 3))
+#define LIBYUV_SSSE3_ONLY
+#endif
+
+// Enable for NaCL pepper 33 for bundle and AVX2 support.
+//  #define NEW_BINUTILS
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+// Effects:
+#define HAS_ARGBADDROW_SSE2
+#define HAS_ARGBAFFINEROW_SSE2
+#define HAS_ARGBATTENUATEROW_SSSE3
+#define HAS_ARGBBLENDROW_SSSE3
+#define HAS_ARGBCOLORMATRIXROW_SSSE3
+#define HAS_ARGBCOLORTABLEROW_X86
+#define HAS_ARGBCOPYALPHAROW_SSE2
+#define HAS_ARGBCOPYYTOALPHAROW_SSE2
+#define HAS_ARGBGRAYROW_SSSE3
+#define HAS_ARGBLUMACOLORTABLEROW_SSSE3
+#define HAS_ARGBMIRRORROW_SSSE3
+#define HAS_ARGBMULTIPLYROW_SSE2
+#define HAS_ARGBPOLYNOMIALROW_SSE2
+#define HAS_ARGBQUANTIZEROW_SSE2
+#define HAS_ARGBSEPIAROW_SSSE3
+#define HAS_ARGBSHADEROW_SSE2
+#define HAS_ARGBSUBTRACTROW_SSE2
+#define HAS_ARGBTOUVROW_SSSE3
+#define HAS_ARGBUNATTENUATEROW_SSE2
+#define HAS_COMPUTECUMULATIVESUMROW_SSE2
+#define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+#define HAS_INTERPOLATEROW_SSE2
+#define HAS_INTERPOLATEROW_SSSE3
+#define HAS_RGBCOLORTABLEROW_X86
+#define HAS_SOBELROW_SSE2
+#define HAS_SOBELTOPLANEROW_SSE2
+#define HAS_SOBELXROW_SSE2
+#define HAS_SOBELXYROW_SSE2
+#define HAS_SOBELYROW_SSE2
+
+// Conversions:
+#define HAS_ABGRTOUVROW_SSSE3
+#define HAS_ABGRTOYROW_SSSE3
+#define HAS_ARGB1555TOARGBROW_SSE2
+#define HAS_ARGB4444TOARGBROW_SSE2
+#define HAS_ARGBSHUFFLEROW_SSE2
+#define HAS_ARGBSHUFFLEROW_SSSE3
+#define HAS_ARGBTOARGB1555ROW_SSE2
+#define HAS_ARGBTOARGB4444ROW_SSE2
+#define HAS_ARGBTOBAYERGGROW_SSE2
+#define HAS_ARGBTOBAYERROW_SSSE3
+#define HAS_ARGBTORAWROW_SSSE3
+#define HAS_ARGBTORGB24ROW_SSSE3
+#define HAS_ARGBTORGB565ROW_SSE2
+#define HAS_ARGBTOUV422ROW_SSSE3
+#define HAS_ARGBTOUV444ROW_SSSE3
+#define HAS_ARGBTOUVJROW_SSSE3
+#define HAS_ARGBTOYJROW_SSSE3
+#define HAS_ARGBTOYROW_SSSE3
+#define HAS_BGRATOUVROW_SSSE3
+#define HAS_BGRATOYROW_SSSE3
+#define HAS_COPYROW_ERMS
+#define HAS_COPYROW_SSE2
+#define HAS_COPYROW_X86
+#define HAS_HALFROW_SSE2
+#define HAS_I400TOARGBROW_SSE2
+#define HAS_I411TOARGBROW_SSSE3
+#define HAS_I422TOARGB1555ROW_SSSE3
+#define HAS_I422TOABGRROW_SSSE3
+#define HAS_I422TOARGB1555ROW_SSSE3
+#define HAS_I422TOARGB4444ROW_SSSE3
+#define HAS_I422TOARGBROW_SSSE3
+#define HAS_I422TOBGRAROW_SSSE3
+#define HAS_I422TORAWROW_SSSE3
+#define HAS_I422TORGB24ROW_SSSE3
+#define HAS_I422TORGB565ROW_SSSE3
+#define HAS_I422TORGBAROW_SSSE3
+#define HAS_I422TOUYVYROW_SSE2
+#define HAS_I422TOYUY2ROW_SSE2
+#define HAS_I444TOARGBROW_SSSE3
+#define HAS_MERGEUVROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#define HAS_MIRRORROW_SSSE3
+#define HAS_MIRRORROW_UV_SSSE3
+#define HAS_MIRRORUVROW_SSSE3
+#define HAS_NV12TOARGBROW_SSSE3
+#define HAS_NV12TORGB565ROW_SSSE3
+#define HAS_NV21TOARGBROW_SSSE3
+#define HAS_NV21TORGB565ROW_SSSE3
+#define HAS_RAWTOARGBROW_SSSE3
+#define HAS_RAWTOYROW_SSSE3
+#define HAS_RGB24TOARGBROW_SSSE3
+#define HAS_RGB24TOYROW_SSSE3
+#define HAS_RGB565TOARGBROW_SSE2
+#define HAS_RGBATOUVROW_SSSE3
+#define HAS_RGBATOYROW_SSSE3
+#define HAS_SETROW_X86
+#define HAS_SPLITUVROW_SSE2
+#define HAS_UYVYTOARGBROW_SSSE3
+#define HAS_UYVYTOUV422ROW_SSE2
+#define HAS_UYVYTOUVROW_SSE2
+#define HAS_UYVYTOYROW_SSE2
+#define HAS_YTOARGBROW_SSE2
+#define HAS_YUY2TOARGBROW_SSSE3
+#define HAS_YUY2TOUV422ROW_SSE2
+#define HAS_YUY2TOUVROW_SSE2
+#define HAS_YUY2TOYROW_SSE2
+#endif
+
+// GCC >= 4.7.0 required for AVX2.
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
+#if (__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
+#define GCC_HAS_AVX2 1
+#endif  // GNUC >= 4.7
+#endif  // __GNUC__
+
+// clang >= 3.4.0 required for AVX2.
+#if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
+#if (__clang_major__ > 3) || (__clang_major__ == 3 && (__clang_minor__ >= 4))
+#define CLANG_HAS_AVX2 1
+#endif  // clang >= 3.4
+#endif  // __clang__
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// The following are available on all x86 platforms, but
+// require VS2012, clang 3.4 or gcc 4.7.
+// The code supports NaCL but requires a new compiler and validator.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(VISUALC_HAS_AVX2) || \
+    defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+// Effects:
+#define HAS_ARGBPOLYNOMIALROW_AVX2
+#define HAS_ARGBSHUFFLEROW_AVX2
+#define HAS_ARGBCOPYALPHAROW_AVX2
+#define HAS_ARGBCOPYYTOALPHAROW_AVX2
+#endif
+
+// The following are require VS2012.
+// TODO(fbarchard): Port to gcc.
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+#define HAS_ARGBTOUVROW_AVX2
+#define HAS_ARGBTOYJROW_AVX2
+#define HAS_ARGBTOYROW_AVX2
+#define HAS_HALFROW_AVX2
+#define HAS_I422TOARGBROW_AVX2
+#define HAS_INTERPOLATEROW_AVX2
+#define HAS_MERGEUVROW_AVX2
+#define HAS_MIRRORROW_AVX2
+#define HAS_SPLITUVROW_AVX2
+#define HAS_UYVYTOUV422ROW_AVX2
+#define HAS_UYVYTOUVROW_AVX2
+#define HAS_UYVYTOYROW_AVX2
+#define HAS_YUY2TOUV422ROW_AVX2
+#define HAS_YUY2TOUVROW_AVX2
+#define HAS_YUY2TOYROW_AVX2
+
+// Effects:
+#define HAS_ARGBADDROW_AVX2
+#define HAS_ARGBATTENUATEROW_AVX2
+#define HAS_ARGBMIRRORROW_AVX2
+#define HAS_ARGBMULTIPLYROW_AVX2
+#define HAS_ARGBSUBTRACTROW_AVX2
+#define HAS_ARGBUNATTENUATEROW_AVX2
+#endif  // defined(VISUALC_HAS_AVX2)
+
+// The following are Yasm x86 only:
+// TODO(fbarchard): Port AVX2 to inline.
+#if !defined(LIBYUV_DISABLE_X86) && defined(HAVE_YASM)
+    (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__x86_64__) || defined(__i386__))
+#define HAS_MERGEUVROW_AVX2
+#define HAS_MERGEUVROW_MMX
+#define HAS_SPLITUVROW_AVX2
+#define HAS_SPLITUVROW_MMX
+#define HAS_UYVYTOYROW_AVX2
+#define HAS_UYVYTOYROW_MMX
+#define HAS_YUY2TOYROW_AVX2
+#define HAS_YUY2TOYROW_MMX
+#endif
+
+// The following are disabled when SSSE3 is available:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
+    !defined(LIBYUV_SSSE3_ONLY)
+#define HAS_ARGBBLENDROW_SSE2
+#define HAS_ARGBATTENUATEROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_ABGRTOUVROW_NEON
+#define HAS_ABGRTOYROW_NEON
+#define HAS_ARGB1555TOARGBROW_NEON
+#define HAS_ARGB1555TOUVROW_NEON
+#define HAS_ARGB1555TOYROW_NEON
+#define HAS_ARGB4444TOARGBROW_NEON
+#define HAS_ARGB4444TOUVROW_NEON
+#define HAS_ARGB4444TOYROW_NEON
+#define HAS_ARGBTOARGB1555ROW_NEON
+#define HAS_ARGBTOARGB4444ROW_NEON
+#define HAS_ARGBTOBAYERROW_NEON
+#define HAS_ARGBTOBAYERGGROW_NEON
+#define HAS_ARGBTORAWROW_NEON
+#define HAS_ARGBTORGB24ROW_NEON
+#define HAS_ARGBTORGB565ROW_NEON
+#define HAS_ARGBTOUV411ROW_NEON
+#define HAS_ARGBTOUV422ROW_NEON
+#define HAS_ARGBTOUV444ROW_NEON
+#define HAS_ARGBTOUVROW_NEON
+#define HAS_ARGBTOUVJROW_NEON
+#define HAS_ARGBTOYROW_NEON
+#define HAS_ARGBTOYJROW_NEON
+#define HAS_BGRATOUVROW_NEON
+#define HAS_BGRATOYROW_NEON
+#define HAS_COPYROW_NEON
+#define HAS_HALFROW_NEON
+#define HAS_I400TOARGBROW_NEON
+#define HAS_I411TOARGBROW_NEON
+#define HAS_I422TOABGRROW_NEON
+#define HAS_I422TOARGB1555ROW_NEON
+#define HAS_I422TOARGB4444ROW_NEON
+#define HAS_I422TOARGBROW_NEON
+#define HAS_I422TOBGRAROW_NEON
+#define HAS_I422TORAWROW_NEON
+#define HAS_I422TORGB24ROW_NEON
+#define HAS_I422TORGB565ROW_NEON
+#define HAS_I422TORGBAROW_NEON
+#define HAS_I422TOUYVYROW_NEON
+#define HAS_I422TOYUY2ROW_NEON
+#define HAS_I444TOARGBROW_NEON
+#define HAS_MERGEUVROW_NEON
+#define HAS_MIRRORROW_NEON
+#define HAS_MIRRORUVROW_NEON
+#define HAS_NV12TOARGBROW_NEON
+#define HAS_NV12TORGB565ROW_NEON
+#define HAS_NV21TOARGBROW_NEON
+#define HAS_NV21TORGB565ROW_NEON
+#define HAS_RAWTOARGBROW_NEON
+#define HAS_RAWTOUVROW_NEON
+#define HAS_RAWTOYROW_NEON
+#define HAS_RGB24TOARGBROW_NEON
+#define HAS_RGB24TOUVROW_NEON
+#define HAS_RGB24TOYROW_NEON
+#define HAS_RGB565TOARGBROW_NEON
+#define HAS_RGB565TOUVROW_NEON
+#define HAS_RGB565TOYROW_NEON
+#define HAS_RGBATOUVROW_NEON
+#define HAS_RGBATOYROW_NEON
+#define HAS_SETROW_NEON
+#define HAS_SPLITUVROW_NEON
+#define HAS_UYVYTOARGBROW_NEON
+#define HAS_UYVYTOUV422ROW_NEON
+#define HAS_UYVYTOUVROW_NEON
+#define HAS_UYVYTOYROW_NEON
+#define HAS_YTOARGBROW_NEON
+#define HAS_YUY2TOARGBROW_NEON
+#define HAS_YUY2TOUV422ROW_NEON
+#define HAS_YUY2TOUVROW_NEON
+#define HAS_YUY2TOYROW_NEON
+
+// Effects:
+#define HAS_ARGBADDROW_NEON
+#define HAS_ARGBATTENUATEROW_NEON
+#define HAS_ARGBBLENDROW_NEON
+#define HAS_ARGBCOLORMATRIXROW_NEON
+#define HAS_ARGBGRAYROW_NEON
+#define HAS_ARGBMIRRORROW_NEON
+#define HAS_ARGBMULTIPLYROW_NEON
+#define HAS_ARGBQUANTIZEROW_NEON
+#define HAS_ARGBSEPIAROW_NEON
+#define HAS_ARGBSHADEROW_NEON
+#define HAS_ARGBSUBTRACTROW_NEON
+#define HAS_SOBELROW_NEON
+#define HAS_SOBELTOPLANEROW_NEON
+#define HAS_SOBELXYROW_NEON
+#define HAS_SOBELXROW_NEON
+#define HAS_SOBELYROW_NEON
+#define HAS_INTERPOLATEROW_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__)
+#define HAS_COPYROW_MIPS
+#if defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_I422TOABGRROW_MIPS_DSPR2
+#define HAS_I422TOARGBROW_MIPS_DSPR2
+#define HAS_I422TOBGRAROW_MIPS_DSPR2
+#define HAS_INTERPOLATEROWS_MIPS_DSPR2
+#define HAS_MIRRORROW_MIPS_DSPR2
+#define HAS_MIRRORUVROW_MIPS_DSPR2
+#define HAS_SPLITUVROW_MIPS_DSPR2
+#endif
+#endif
+
+#if defined(_MSC_VER) && !defined(__CLR_VER)
+#define SIMD_ALIGNED(var) __declspec(align(16)) var
+typedef __declspec(align(16)) int16 vec16[8];
+typedef __declspec(align(16)) int32 vec32[4];
+typedef __declspec(align(16)) int8 vec8[16];
+typedef __declspec(align(16)) uint16 uvec16[8];
+typedef __declspec(align(16)) uint32 uvec32[4];
+typedef __declspec(align(16)) uint8 uvec8[16];
+typedef __declspec(align(32)) int16 lvec16[16];
+typedef __declspec(align(32)) int32 lvec32[8];
+typedef __declspec(align(32)) int8 lvec8[32];
+typedef __declspec(align(32)) uint16 ulvec16[16];
+typedef __declspec(align(32)) uint32 ulvec32[8];
+typedef __declspec(align(32)) uint8 ulvec8[32];
+
+#elif defined(__GNUC__)
+// Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
+#define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
+typedef int16 __attribute__((vector_size(16))) vec16;
+typedef int32 __attribute__((vector_size(16))) vec32;
+typedef int8 __attribute__((vector_size(16))) vec8;
+typedef uint16 __attribute__((vector_size(16))) uvec16;
+typedef uint32 __attribute__((vector_size(16))) uvec32;
+typedef uint8 __attribute__((vector_size(16))) uvec8;
+#else
+#define SIMD_ALIGNED(var) var
+typedef int16 vec16[8];
+typedef int32 vec32[4];
+typedef int8 vec8[16];
+typedef uint16 uvec16[8];
+typedef uint32 uvec32[4];
+typedef uint8 uvec8[16];
+#endif
+
+#if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__)
+#define OMITFP
+#else
+#define OMITFP __attribute__((optimize("omit-frame-pointer")))
+#endif
+
+// NaCL macros for GCC x86 and x64.
+
+// TODO(nfullagar): When pepper_33 toolchain is distributed, default to
+// NEW_BINUTILS and remove all BUNDLEALIGN occurances.
+#if defined(__native_client__)
+#define LABELALIGN ".p2align 5\n"
+#else
+#define LABELALIGN ".p2align 2\n"
+#endif
+#if defined(__native_client__) && defined(__x86_64__)
+#if defined(NEW_BINUTILS)
+#define BUNDLELOCK ".bundle_lock\n"
+#define BUNDLEUNLOCK ".bundle_unlock\n"
+#define BUNDLEALIGN "\n"
+#else
+#define BUNDLELOCK "\n"
+#define BUNDLEUNLOCK "\n"
+#define BUNDLEALIGN ".p2align 5\n"
+#endif
+#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
+#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
+#define MEMLEA(offset, base) #offset "(%q" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%q" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%q" #base ",%q" #index "," #scale ")"
+#define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
+#define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%%" #reg "\n" \
+    BUNDLEUNLOCK
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " %%" #reg ",(%%r15,%%r14)\n" \
+    BUNDLEUNLOCK
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%" #arg "\n" \
+    BUNDLEUNLOCK
+#else
+#define BUNDLEALIGN "\n"
+#define MEMACCESS(base) "(%" #base ")"
+#define MEMACCESS2(offset, base) #offset "(%" #base ")"
+#define MEMLEA(offset, base) #offset "(%" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%" #base ",%" #index "," #scale ")"
+#define MEMMOVESTRING(s, d)
+#define MEMSTORESTRING(reg, d)
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    #opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
+#endif
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width);
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width);
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width);
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width);
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width);
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width);
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width);
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width);
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width);
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_vu,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width);
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_vu,
+                          uint8* dst_rgb565,
+                          int width);
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width);
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width);
+
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Unaligned_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Unaligned_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_C(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_C(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_C(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_C(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_C(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_Any_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_Any_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_Any_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+
+void ARGBToUVRow_AVX2(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_AVX2(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb, int src_stride_argb,
+                                  uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                                 uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                            uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_Any_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_Any_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                           uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_Any_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV422Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV411Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUVRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_Any_NEON(const uint8* src_bgra, int src_stride_bgra,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_Any_NEON(const uint8* src_abgr, int src_stride_abgr,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_Any_NEON(const uint8* src_rgba, int src_stride_rgba,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_Any_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_Any_NEON(const uint8* src_raw, int src_stride_raw,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_Any_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                            uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_Any_NEON(const uint8* src_argb1555,
+                              int src_stride_argb1555,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_Any_NEON(const uint8* src_argb4444,
+                              int src_stride_argb4444,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVRow_C(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_C(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_C(const uint8* src_bgra, int src_stride_bgra,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_C(const uint8* src_abgr, int src_stride_abgr,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_C(const uint8* src_rgba, int src_stride_rgba,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGB24ToUVRow_C(const uint8* src_rgb24, int src_stride_rgb24,
+                    uint8* dst_u, uint8* dst_v, int width);
+void RAWToUVRow_C(const uint8* src_raw, int src_stride_raw,
+                  uint8* dst_u, uint8* dst_v, int width);
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width);
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb,
+                                    uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV422Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb,
+                                    uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width);
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width);
+void MirrorRow_C(const uint8* src, uint8* dst, int width);
+
+void MirrorUVRow_SSSE3(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                       int width);
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width);
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width);
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                   int width);
+
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width);
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int pix);
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix);
+void SplitUVRow_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
+                                     uint8* dst_v, int pix);
+void SplitUVRow_Any_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix);
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width);
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width);
+void MergeUVRow_Any_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count);
+void CopyRow_X86(const uint8* src, uint8* dst, int count);
+void CopyRow_NEON(const uint8* src, uint8* dst, int count);
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count);
+void CopyRow_C(const uint8* src, uint8* dst, int count);
+
+void ARGBCopyAlphaRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+
+void SetRow_X86(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                     int dst_stride, int height);
+void SetRow_NEON(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
+                      int dst_stride, int height);
+void SetRow_C(uint8* dst, uint32 v32, int count);
+void ARGBSetRows_C(uint8* dst, uint32 v32, int width, int dst_stride,
+                   int height);
+
+// ARGBShufflers for BGRAToARGB etc.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix);
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_C(const uint8* src_rgb, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void ARGB4444ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void RGB24ToARGBRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_Any_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+void RGB24ToARGBRow_Any_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_uv,
+                     uint8* dst_argb,
+                     int width);
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_vu,
+                       uint8* dst_argb,
+                       int width);
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_uv,
+                       uint8* dst_argb,
+                       int width);
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* dst_argb,
+                     int width);
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* dst_argb,
+                     int width);
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* dst_argb,
+                     int width);
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_bgra,
+                     int width);
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_abgr,
+                     int width);
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_rgba,
+                     int width);
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* dst_rgb24,
+                      int width);
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* dst_raw,
+                    int width);
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width);
+void YToARGBRow_C(const uint8* src_y,
+                  uint8* dst_argb,
+                  int width);
+void I422ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I444ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I422ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I411ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void NV12ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_uv,
+                         uint8* dst_argb,
+                         int width);
+void NV21ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_vu,
+                         uint8* dst_argb,
+                         int width);
+void NV12ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_uv,
+                           uint8* dst_argb,
+                           int width);
+void NV21ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_vu,
+                           uint8* dst_argb,
+                           int width);
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
+                         uint8* dst_argb,
+                         int width);
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
+                         uint8* dst_argb,
+                         int width);
+void I422ToBGRARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_bgra,
+                         int width);
+void I422ToABGRRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_abgr,
+                         int width);
+void I422ToRGBARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgba,
+                         int width);
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+// RGB24/RAW are unaligned.
+void I422ToRGB24Row_SSSE3(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb24,
+                          int width);
+void I422ToRAWRow_SSSE3(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_raw,
+                        int width);
+
+void I444ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void I422ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void I411ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_argb,
+                                   int width);
+void NV12ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_uv,
+                                   uint8* dst_argb,
+                                   int width);
+void NV21ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_vu,
+                                   uint8* dst_argb,
+                                   int width);
+void YUY2ToARGBRow_Unaligned_SSSE3(const uint8* src_yuy2,
+                                   uint8* dst_argb,
+                                   int width);
+void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
+                                   uint8* dst_argb,
+                                   int width);
+void I422ToBGRARow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_bgra,
+                                   int width);
+void I422ToABGRRow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_abgr,
+                                   int width);
+void I422ToRGBARow_Unaligned_SSSE3(const uint8* src_y,
+                                   const uint8* src_u,
+                                   const uint8* src_v,
+                                   uint8* dst_rgba,
+                                   int width);
+void I422ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I444ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I411ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void NV12ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_uv,
+                             uint8* dst_argb,
+                             int width);
+void NV21ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_vu,
+                             uint8* dst_argb,
+                             int width);
+void NV12ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_uv,
+                               uint8* dst_argb,
+                               int width);
+void NV21ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_vu,
+                               uint8* dst_argb,
+                               int width);
+void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2,
+                             uint8* dst_argb,
+                             int width);
+void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy,
+                             uint8* dst_argb,
+                             int width);
+void I422ToBGRARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_bgra,
+                             int width);
+void I422ToABGRRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_abgr,
+                             int width);
+void I422ToRGBARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_rgba,
+                             int width);
+void I422ToARGB4444Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToARGB1555Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_u,
+                               const uint8* src_v,
+                               uint8* dst_rgba,
+                               int width);
+// RGB24/RAW are unaligned.
+void I422ToRGB24Row_Any_SSSE3(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToRAWRow_Any_SSSE3(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void YToARGBRow_SSE2(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width);
+void YToARGBRow_NEON(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width);
+void YToARGBRow_Any_SSE2(const uint8* src_y,
+                         uint8* dst_argb,
+                         int width);
+void YToARGBRow_Any_NEON(const uint8* src_y,
+                         uint8* dst_argb,
+                         int width);
+
+// ARGB preattenuated alpha blend.
+void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1,
+                        uint8* dst_argb, int width);
+void ARGBBlendRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
+                    uint8* dst_argb, int width);
+
+// ARGB multiply images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBMultiplyRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+// ARGB add images.
+void ARGBAddRow_C(const uint8* src_argb, const uint8* src_argb1,
+                  uint8* dst_argb, int width);
+void ARGBAddRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+
+// ARGB subtract images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void I444ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I411ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToABGRRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGBARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGB24Row_Any_NEON(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRAWRow_Any_NEON(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+void I422ToARGB4444Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToARGB1555Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void NV12ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV21ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void NV21ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void YUY2ToARGBRow_Any_NEON(const uint8* src_yuy2,
+                            uint8* dst_argb,
+                            int width);
+void UYVYToARGBRow_Any_NEON(const uint8* src_uyvy,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+
+void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix);
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_C(const uint8* src_yuy2, int stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_AVX2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_SSE2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_NEON(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_NEON(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix);
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_NEON(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_C(const uint8* src_uyvy, int stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_AVX2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_SSE2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_NEON(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_NEON(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+
+void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+               uint8* dst_uv, int pix);
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix);
+
+void ARGBToBayerRow_C(const uint8* src_argb, uint8* dst_bayer,
+                      uint32 selector, int pix);
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix);
+void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix);
+void ARGBToBayerRow_Any_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                              uint32 selector, int pix);
+void ARGBToBayerRow_Any_NEON(const uint8* src_argb, uint8* dst_bayer,
+                             uint32 selector, int pix);
+void ARGBToBayerGGRow_C(const uint8* src_argb, uint8* dst_bayer,
+                        uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_Any_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                               uint32 /* selector */, int pix);
+void ARGBToBayerGGRow_Any_NEON(const uint8* src_argb, uint8* dst_bayer,
+                               uint32 /* selector */, int pix);
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_yuy2, int width);
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_uyvy, int width);
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+
+// Effects related row functions.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                int width);
+void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+
+// Inverse table for unattenuate, shared by C and SSE2.
+extern const uint32 fixed_invtbl8[256];
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBSepiaRow_C(uint8* dst_argb, int width);
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width);
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width);
+
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width);
+
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width);
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value);
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+
+// Used for blur.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width);
+
+void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft,
+                                 int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width);
+
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+
+// Used for I420Scale, ARGBScale, and ARGBInterpolate.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride_ptr,
+                      int width, int source_y_fraction);
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride_ptr, int width,
+                          int source_y_fraction);
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                ptrdiff_t src_stride_ptr, int width,
+                                int source_y_fraction);
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride_ptr, int width,
+                                   int source_y_fraction);
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride_ptr, int width,
+                                    int source_y_fraction);
+void InterpolateRow_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                              ptrdiff_t src_stride_ptr, int width,
+                              int source_y_fraction);
+void InterpolateRow_Any_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRows_Any_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride_ptr, int width,
+                                    int source_y_fraction);
+
+// Sobel images.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width);
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width);
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width);
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width);
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width);
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width);
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff);
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROW_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale.h
new file mode 100755
index 0000000000..592b8ed5fa
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale.h
@@ -0,0 +1,85 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Supported filtering.
+typedef enum FilterMode {
+  kFilterNone = 0,  // Point sample; Fastest.
+  kFilterLinear = 1,  // Filter horizontally only.
+  kFilterBilinear = 2,  // Faster than box, but lower quality scaling down.
+  kFilterBox = 3  // Highest quality.
+} FilterModeEnum;
+
+// Scale a YUV plane.
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering);
+
+// Scales a YUV 4:2:0 image from the src width and height to the
+// dst width and height.
+// If filtering is kFilterNone, a simple nearest-neighbor algorithm is
+// used. This produces basic (blocky) quality at the fastest speed.
+// If filtering is kFilterBilinear, interpolation is used to produce a better
+// quality image, at the expense of speed.
+// If filtering is kFilterBox, averaging is used to produce ever better
+// quality image, at further expense of speed.
+// Returns 0 if successful.
+
+LIBYUV_API
+int I420Scale(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              int src_width, int src_height,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int dst_width, int dst_height,
+              enum FilterMode filtering);
+
+#ifdef __cplusplus
+// Legacy API.  Deprecated.
+LIBYUV_API
+int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
+          int src_stride_y, int src_stride_u, int src_stride_v,
+          int src_width, int src_height,
+          uint8* dst_y, uint8* dst_u, uint8* dst_v,
+          int dst_stride_y, int dst_stride_u, int dst_stride_v,
+          int dst_width, int dst_height,
+          LIBYUV_BOOL interpolate);
+
+// Legacy API.  Deprecated.
+LIBYUV_API
+int ScaleOffset(const uint8* src_i420, int src_width, int src_height,
+                uint8* dst_i420, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate);
+
+// For testing, allow disabling of specialized scalers.
+LIBYUV_API
+void SetUseReferenceImpl(LIBYUV_BOOL use);
+#endif  // __cplusplus
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale_argb.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale_argb.h
new file mode 100755
index 0000000000..0c9b362575
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale_argb.h
@@ -0,0 +1,57 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ARGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"  // For FilterMode
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+int ARGBScale(const uint8* src_argb, int src_stride_argb,
+              int src_width, int src_height,
+              uint8* dst_argb, int dst_stride_argb,
+              int dst_width, int dst_height,
+              enum FilterMode filtering);
+
+// Clipped scale takes destination rectangle coordinates for clip values.
+LIBYUV_API
+int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
+                  int src_width, int src_height,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int dst_width, int dst_height,
+                  int clip_x, int clip_y, int clip_width, int clip_height,
+                  enum FilterMode filtering);
+
+// TODO(fbarchard): Implement this.
+// Scale with YUV conversion to ARGB and clipping.
+LIBYUV_API
+int YUVToARGBScaleClip(const uint8* src_y, int src_stride_y,
+                       const uint8* src_u, int src_stride_u,
+                       const uint8* src_v, int src_stride_v,
+                       uint32 src_fourcc,
+                       int src_width, int src_height,
+                       uint8* dst_argb, int dst_stride_argb,
+                       uint32 dst_fourcc,
+                       int dst_width, int dst_height,
+                       int clip_x, int clip_y, int clip_width, int clip_height,
+                       enum FilterMode filtering);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ARGB_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale_row.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale_row.h
new file mode 100644
index 0000000000..13eccc4d77
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/scale_row.h
@@ -0,0 +1,301 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ROW_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
+    defined(TARGET_IPHONE_SIMULATOR)
+#define LIBYUV_DISABLE_X86
+#endif
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_SCALEROWDOWN2_SSE2
+#define HAS_SCALEROWDOWN4_SSE2
+#define HAS_SCALEROWDOWN34_SSSE3
+#define HAS_SCALEROWDOWN38_SSSE3
+#define HAS_SCALEADDROWS_SSE2
+#define HAS_SCALEFILTERCOLS_SSSE3
+#define HAS_SCALECOLSUP2_SSE2
+#define HAS_SCALEARGBROWDOWN2_SSE2
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+#define HAS_SCALEARGBCOLS_SSE2
+#define HAS_SCALEARGBFILTERCOLS_SSSE3
+#define HAS_SCALEARGBCOLSUP2_SSE2
+#define HAS_FIXEDDIV_X86
+#define HAS_FIXEDDIV1_X86
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_SCALEROWDOWN2_NEON
+#define HAS_SCALEROWDOWN4_NEON
+#define HAS_SCALEROWDOWN34_NEON
+#define HAS_SCALEROWDOWN38_NEON
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+#define HAS_SCALEARGBROWDOWN2_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_SCALEROWDOWN2_MIPS_DSPR2
+#define HAS_SCALEROWDOWN4_MIPS_DSPR2
+#define HAS_SCALEROWDOWN34_MIPS_DSPR2
+#define HAS_SCALEROWDOWN38_MIPS_DSPR2
+#endif
+
+// Scale ARGB vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering);
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering);
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div);
+int FixedDiv_X86(int num, int div);
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div);
+int FixedDiv1_X86(int num, int div);
+#ifdef HAS_FIXEDDIV_X86
+#define FixedDiv FixedDiv_X86
+#define FixedDiv1 FixedDiv1_X86
+#else
+#define FixedDiv FixedDiv_C
+#define FixedDiv1 FixedDiv1_C
+#endif
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy);
+
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx);
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int, int);
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x, int dx);
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleAddRows_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height);
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx);
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x, int dx);
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int, int);
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x, int dx);
+
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width,
+                       int src_height);
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx);
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx);
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx);
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+// Row functions.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb, int src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, int src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+
+// ScaleRowDown2Box also used by planar functions
+// NEON downscalers with interpolation.
+
+// Note - not static due to reuse in convert for 444 to 420.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+//  to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+// 32x3 -> 12x1
+void ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ROW_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/version.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/version.h
new file mode 100755
index 0000000000..4881861866
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/version.h
@@ -0,0 +1,16 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
+#define INCLUDE_LIBYUV_VERSION_H_
+
+#define LIBYUV_VERSION 998
+
+#endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/video_common.h b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/video_common.h
new file mode 100755
index 0000000000..039efb96d1
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/include/libyuv/video_common.h
@@ -0,0 +1,182 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+// Common definitions for video, including fourcc and VideoFormat.
+
+#ifndef INCLUDE_LIBYUV_VIDEO_COMMON_H_  // NOLINT
+#define INCLUDE_LIBYUV_VIDEO_COMMON_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+// Definition of FourCC codes
+//////////////////////////////////////////////////////////////////////////////
+
+// Convert four characters to a FourCC code.
+// Needs to be a macro otherwise the OS X compiler complains when the kFormat*
+// constants are used in a switch.
+#ifdef __cplusplus
+#define FOURCC(a, b, c, d) ( \
+    (static_cast<uint32>(a)) | (static_cast<uint32>(b) << 8) | \
+    (static_cast<uint32>(c) << 16) | (static_cast<uint32>(d) << 24))
+#else
+#define FOURCC(a, b, c, d) ( \
+    ((uint32)(a)) | ((uint32)(b) << 8) | /* NOLINT */ \
+    ((uint32)(c) << 16) | ((uint32)(d) << 24))  /* NOLINT */
+#endif
+
+// Some pages discussing FourCC codes:
+//   http://www.fourcc.org/yuv.php
+//   http://v4l2spec.bytesex.org/spec/book1.htm
+//   http://developer.apple.com/quicktime/icefloe/dispatch020.html
+//   http://msdn.microsoft.com/library/windows/desktop/dd206750.aspx#nv12
+//   http://people.xiph.org/~xiphmont/containers/nut/nut4cc.txt
+
+// FourCC codes grouped according to implementation efficiency.
+// Primary formats should convert in 1 efficient step.
+// Secondary formats are converted in 2 steps.
+// Auxilliary formats call primary converters.
+enum FourCC {
+  // 9 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
+  FOURCC_I420 = FOURCC('I', '4', '2', '0'),
+  FOURCC_I422 = FOURCC('I', '4', '2', '2'),
+  FOURCC_I444 = FOURCC('I', '4', '4', '4'),
+  FOURCC_I411 = FOURCC('I', '4', '1', '1'),
+  FOURCC_I400 = FOURCC('I', '4', '0', '0'),
+  FOURCC_NV21 = FOURCC('N', 'V', '2', '1'),
+  FOURCC_NV12 = FOURCC('N', 'V', '1', '2'),
+  FOURCC_YUY2 = FOURCC('Y', 'U', 'Y', '2'),
+  FOURCC_UYVY = FOURCC('U', 'Y', 'V', 'Y'),
+
+  // 2 Secondary YUV formats: row biplanar.
+  FOURCC_M420 = FOURCC('M', '4', '2', '0'),
+  FOURCC_Q420 = FOURCC('Q', '4', '2', '0'),
+
+  // 9 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp.
+  FOURCC_ARGB = FOURCC('A', 'R', 'G', 'B'),
+  FOURCC_BGRA = FOURCC('B', 'G', 'R', 'A'),
+  FOURCC_ABGR = FOURCC('A', 'B', 'G', 'R'),
+  FOURCC_24BG = FOURCC('2', '4', 'B', 'G'),
+  FOURCC_RAW  = FOURCC('r', 'a', 'w', ' '),
+  FOURCC_RGBA = FOURCC('R', 'G', 'B', 'A'),
+  FOURCC_RGBP = FOURCC('R', 'G', 'B', 'P'),  // rgb565 LE.
+  FOURCC_RGBO = FOURCC('R', 'G', 'B', 'O'),  // argb1555 LE.
+  FOURCC_R444 = FOURCC('R', '4', '4', '4'),  // argb4444 LE.
+
+  // 4 Secondary RGB formats: 4 Bayer Patterns.
+  FOURCC_RGGB = FOURCC('R', 'G', 'G', 'B'),
+  FOURCC_BGGR = FOURCC('B', 'G', 'G', 'R'),
+  FOURCC_GRBG = FOURCC('G', 'R', 'B', 'G'),
+  FOURCC_GBRG = FOURCC('G', 'B', 'R', 'G'),
+
+  // 1 Primary Compressed YUV format.
+  FOURCC_MJPG = FOURCC('M', 'J', 'P', 'G'),
+
+  // 5 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
+  FOURCC_YV12 = FOURCC('Y', 'V', '1', '2'),
+  FOURCC_YV16 = FOURCC('Y', 'V', '1', '6'),
+  FOURCC_YV24 = FOURCC('Y', 'V', '2', '4'),
+  FOURCC_YU12 = FOURCC('Y', 'U', '1', '2'),  // Linux version of I420.
+  FOURCC_J420 = FOURCC('J', '4', '2', '0'),
+  FOURCC_J400 = FOURCC('J', '4', '0', '0'),
+
+  // 14 Auxiliary aliases.  CanonicalFourCC() maps these to canonical fourcc.
+  FOURCC_IYUV = FOURCC('I', 'Y', 'U', 'V'),  // Alias for I420.
+  FOURCC_YU16 = FOURCC('Y', 'U', '1', '6'),  // Alias for I422.
+  FOURCC_YU24 = FOURCC('Y', 'U', '2', '4'),  // Alias for I444.
+  FOURCC_YUYV = FOURCC('Y', 'U', 'Y', 'V'),  // Alias for YUY2.
+  FOURCC_YUVS = FOURCC('y', 'u', 'v', 's'),  // Alias for YUY2 on Mac.
+  FOURCC_HDYC = FOURCC('H', 'D', 'Y', 'C'),  // Alias for UYVY.
+  FOURCC_2VUY = FOURCC('2', 'v', 'u', 'y'),  // Alias for UYVY on Mac.
+  FOURCC_JPEG = FOURCC('J', 'P', 'E', 'G'),  // Alias for MJPG.
+  FOURCC_DMB1 = FOURCC('d', 'm', 'b', '1'),  // Alias for MJPG on Mac.
+  FOURCC_BA81 = FOURCC('B', 'A', '8', '1'),  // Alias for BGGR.
+  FOURCC_RGB3 = FOURCC('R', 'G', 'B', '3'),  // Alias for RAW.
+  FOURCC_BGR3 = FOURCC('B', 'G', 'R', '3'),  // Alias for 24BG.
+  FOURCC_CM32 = FOURCC(0, 0, 0, 32),  // Alias for BGRA kCMPixelFormat_32ARGB
+  FOURCC_CM24 = FOURCC(0, 0, 0, 24),  // Alias for RAW kCMPixelFormat_24RGB
+  FOURCC_L555 = FOURCC('L', '5', '5', '5'),  // Alias for RGBO.
+  FOURCC_L565 = FOURCC('L', '5', '6', '5'),  // Alias for RGBP.
+  FOURCC_5551 = FOURCC('5', '5', '5', '1'),  // Alias for RGBO.
+
+  // 1 Auxiliary compressed YUV format set aside for capturer.
+  FOURCC_H264 = FOURCC('H', '2', '6', '4'),
+
+  // Match any fourcc.
+  FOURCC_ANY  = 0xFFFFFFFF,
+};
+
+enum FourCCBpp {
+  // Canonical fourcc codes used in our code.
+  FOURCC_BPP_I420 = 12,
+  FOURCC_BPP_I422 = 16,
+  FOURCC_BPP_I444 = 24,
+  FOURCC_BPP_I411 = 12,
+  FOURCC_BPP_I400 = 8,
+  FOURCC_BPP_NV21 = 12,
+  FOURCC_BPP_NV12 = 12,
+  FOURCC_BPP_YUY2 = 16,
+  FOURCC_BPP_UYVY = 16,
+  FOURCC_BPP_M420 = 12,
+  FOURCC_BPP_Q420 = 12,
+  FOURCC_BPP_ARGB = 32,
+  FOURCC_BPP_BGRA = 32,
+  FOURCC_BPP_ABGR = 32,
+  FOURCC_BPP_RGBA = 32,
+  FOURCC_BPP_24BG = 24,
+  FOURCC_BPP_RAW  = 24,
+  FOURCC_BPP_RGBP = 16,
+  FOURCC_BPP_RGBO = 16,
+  FOURCC_BPP_R444 = 16,
+  FOURCC_BPP_RGGB = 8,
+  FOURCC_BPP_BGGR = 8,
+  FOURCC_BPP_GRBG = 8,
+  FOURCC_BPP_GBRG = 8,
+  FOURCC_BPP_YV12 = 12,
+  FOURCC_BPP_YV16 = 16,
+  FOURCC_BPP_YV24 = 24,
+  FOURCC_BPP_YU12 = 12,
+  FOURCC_BPP_J420 = 12,
+  FOURCC_BPP_J400 = 8,
+  FOURCC_BPP_MJPG = 0,  // 0 means unknown.
+  FOURCC_BPP_H264 = 0,
+  FOURCC_BPP_IYUV = 12,
+  FOURCC_BPP_YU16 = 16,
+  FOURCC_BPP_YU24 = 24,
+  FOURCC_BPP_YUYV = 16,
+  FOURCC_BPP_YUVS = 16,
+  FOURCC_BPP_HDYC = 16,
+  FOURCC_BPP_2VUY = 16,
+  FOURCC_BPP_JPEG = 1,
+  FOURCC_BPP_DMB1 = 1,
+  FOURCC_BPP_BA81 = 8,
+  FOURCC_BPP_RGB3 = 24,
+  FOURCC_BPP_BGR3 = 24,
+  FOURCC_BPP_CM32 = 32,
+  FOURCC_BPP_CM24 = 24,
+
+  // Match any fourcc.
+  FOURCC_BPP_ANY  = 0,  // 0 means unknown.
+};
+
+// Converts fourcc aliases into canonical ones.
+LIBYUV_API uint32 CanonicalFourCC(uint32 fourcc);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_VIDEO_COMMON_H_  NOLINT
diff --git a/drivers/theoraplayer/src/YUV/libyuv/libtheoraplayer-readme.txt b/drivers/theoraplayer/src/YUV/libyuv/libtheoraplayer-readme.txt
new file mode 100755
index 0000000000..680e4a1c36
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/libtheoraplayer-readme.txt
@@ -0,0 +1,15 @@
+libyuv's source code is here provided in minimalist distribution format
+with all source files not needed for compiling libtheoraplayer removed.
+
+- The project files were modified to fit libtheoraplayer's binary output
+  folder structure.
+- Some project files missing in the original source distibution were added to support
+  compiling the libtheoraplayer on those platforms.
+- Also, some code may have been changed to address certain compiler/platform
+  specific problems and is so indicated in the source code.
+
+libyuv is owned and maintained by the Google Inc. and this distribution
+is present here only for convenience and easier compilation of libtheoraplayer.
+
+If you want to use libyuv outside of libtheoraplayer, it is encouraged to use the
+original source distribution by Google Inc: https://code.google.com/p/libyuv/
\ No newline at end of file
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/compare.cc b/drivers/theoraplayer/src/YUV/libyuv/src/compare.cc
new file mode 100755
index 0000000000..9ea81b4e21
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/compare.cc
@@ -0,0 +1,325 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/compare.h"
+
+#include <float.h>
+#include <math.h>
+#ifdef _OPENMP
+#include <omp.h>
+#endif
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
+
+// This module is for Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))))
+#define HAS_HASHDJB2_SSE41
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);
+
+#if _MSC_VER >= 1700
+#define HAS_HASHDJB2_AVX2
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed);
+#endif
+
+#endif  // HAS_HASHDJB2_SSE41
+
+// hash seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
+  const int kBlockSize = 1 << 15;  // 32768;
+  int remainder;
+  uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) = HashDjb2_C;
+#if defined(HAS_HASHDJB2_SSE41)
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    HashDjb2_SSE = HashDjb2_SSE41;
+  }
+#endif
+#if defined(HAS_HASHDJB2_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HashDjb2_SSE = HashDjb2_AVX2;
+  }
+#endif
+
+  while (count >= (uint64)(kBlockSize)) {
+    seed = HashDjb2_SSE(src, kBlockSize, seed);
+    src += kBlockSize;
+    count -= kBlockSize;
+  }
+  remainder = (int)(count) & ~15;
+  if (remainder) {
+    seed = HashDjb2_SSE(src, remainder, seed);
+    src += remainder;
+    count -= remainder;
+  }
+  remainder = (int)(count) & 15;
+  if (remainder) {
+    seed = HashDjb2_C(src, remainder, seed);
+  }
+  return seed;
+}
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_SUMSQUAREERROR_NEON
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count);
+#endif
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_SUMSQUAREERROR_SSE2
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count);
+#endif
+// Visual C 2012 required for AVX2.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && _MSC_VER >= 1700
+#define HAS_SUMSQUAREERROR_AVX2
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count);
+#endif
+
+// TODO(fbarchard): Refactor into row function.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
+                             int count) {
+  // SumSquareError returns values 0 to 65535 for each squared difference.
+  // Up to 65536 of those can be summed and remain within a uint32.
+  // After each block of 65536 pixels, accumulate into a uint64.
+  const int kBlockSize = 65536;
+  int remainder = count & (kBlockSize - 1) & ~31;
+  uint64 sse = 0;
+  int i;
+  uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) =
+      SumSquareError_C;
+#if defined(HAS_SUMSQUAREERROR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SumSquareError = SumSquareError_NEON;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(src_a, 16) && IS_ALIGNED(src_b, 16)) {
+    // Note only used for multiples of 16 so count is not checked.
+    SumSquareError = SumSquareError_SSE2;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    // Note only used for multiples of 32 so count is not checked.
+    SumSquareError = SumSquareError_AVX2;
+  }
+#endif
+#ifdef _OPENMP
+#pragma omp parallel for reduction(+: sse)
+#endif
+  for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
+    sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
+  }
+  src_a += count & ~(kBlockSize - 1);
+  src_b += count & ~(kBlockSize - 1);
+  if (remainder) {
+    sse += SumSquareError(src_a, src_b, remainder);
+    src_a += remainder;
+    src_b += remainder;
+  }
+  remainder = count & 31;
+  if (remainder) {
+    sse += SumSquareError_C(src_a, src_b, remainder);
+  }
+  return sse;
+}
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
+                                  const uint8* src_b, int stride_b,
+                                  int width, int height) {
+  uint64 sse = 0;
+  int h;
+  // Coalesce rows.
+  if (stride_a == width &&
+      stride_b == width) {
+    width *= height;
+    height = 1;
+    stride_a = stride_b = 0;
+  }
+  for (h = 0; h < height; ++h) {
+    sse += ComputeSumSquareError(src_a, src_b, width);
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+  return sse;
+}
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count) {
+  double psnr;
+  if (sse > 0) {
+    double mse = (double)(count) / (double)(sse);
+    psnr = 10.0 * log10(255.0 * 255.0 * mse);
+  } else {
+    psnr = kMaxPsnr;      // Limit to prevent divide by 0
+  }
+
+  if (psnr > kMaxPsnr)
+    psnr = kMaxPsnr;
+
+  return psnr;
+}
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height) {
+  const uint64 samples = width * height;
+  const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a,
+                                                src_b, stride_b,
+                                                width, height);
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height) {
+  const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a,
+                                                  src_y_b, stride_y_b,
+                                                  width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const uint64 sse_u = ComputeSumSquareErrorPlane(src_u_a, stride_u_a,
+                                                  src_u_b, stride_u_b,
+                                                  width_uv, height_uv);
+  const uint64 sse_v = ComputeSumSquareErrorPlane(src_v_a, stride_v_a,
+                                                  src_v_b, stride_v_b,
+                                                  width_uv, height_uv);
+  const uint64 samples = width * height + 2 * (width_uv * height_uv);
+  const uint64 sse = sse_y + sse_u + sse_v;
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+static const int64 cc1 =  26634;  // (64^2*(.01*255)^2
+static const int64 cc2 = 239708;  // (64^2*(.03*255)^2
+
+static double Ssim8x8_C(const uint8* src_a, int stride_a,
+                        const uint8* src_b, int stride_b) {
+  int64 sum_a = 0;
+  int64 sum_b = 0;
+  int64 sum_sq_a = 0;
+  int64 sum_sq_b = 0;
+  int64 sum_axb = 0;
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    int j;
+    for (j = 0; j < 8; ++j) {
+      sum_a += src_a[j];
+      sum_b += src_b[j];
+      sum_sq_a += src_a[j] * src_a[j];
+      sum_sq_b += src_b[j] * src_b[j];
+      sum_axb += src_a[j] * src_b[j];
+    }
+
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+
+  {
+    const int64 count = 64;
+    // scale the constants by number of pixels
+    const int64 c1 = (cc1 * count * count) >> 12;
+    const int64 c2 = (cc2 * count * count) >> 12;
+
+    const int64 sum_a_x_sum_b = sum_a * sum_b;
+
+    const int64 ssim_n = (2 * sum_a_x_sum_b + c1) *
+                         (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
+
+    const int64 sum_a_sq = sum_a*sum_a;
+    const int64 sum_b_sq = sum_b*sum_b;
+
+    const int64 ssim_d = (sum_a_sq + sum_b_sq + c1) *
+                         (count * sum_sq_a - sum_a_sq +
+                          count * sum_sq_b - sum_b_sq + c2);
+
+    if (ssim_d == 0.0) {
+      return DBL_MAX;
+    }
+    return ssim_n * 1.0 / ssim_d;
+  }
+}
+
+// We are using a 8x8 moving window with starting location of each 8x8 window
+// on the 4x4 pixel grid. Such arrangement allows the windows to overlap
+// block boundaries to penalize blocking artifacts.
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height) {
+  int samples = 0;
+  double ssim_total = 0;
+  double (*Ssim8x8)(const uint8* src_a, int stride_a,
+                    const uint8* src_b, int stride_b) = Ssim8x8_C;
+
+  // sample point start with each 4x4 location
+  int i;
+  for (i = 0; i < height - 8; i += 4) {
+    int j;
+    for (j = 0; j < width - 8; j += 4) {
+      ssim_total += Ssim8x8(src_a + j, stride_a, src_b + j, stride_b);
+      samples++;
+    }
+
+    src_a += stride_a * 4;
+    src_b += stride_b * 4;
+  }
+
+  ssim_total /= samples;
+  return ssim_total;
+}
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height) {
+  const double ssim_y = CalcFrameSsim(src_y_a, stride_y_a,
+                                      src_y_b, stride_y_b, width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a,
+                                      src_u_b, stride_u_b,
+                                      width_uv, height_uv);
+  const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a,
+                                      src_v_b, stride_v_b,
+                                      width_uv, height_uv);
+  return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/compare_common.cc b/drivers/theoraplayer/src/YUV/libyuv/src/compare_common.cc
new file mode 100755
index 0000000000..c546b51829
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/compare_common.cc
@@ -0,0 +1,42 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse = 0u;
+  int i;
+  for (i = 0; i < count; ++i) {
+    int diff = src_a[i] - src_b[i];
+    sse += (uint32)(diff * diff);
+  }
+  return sse;
+}
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed) {
+  uint32 hash = seed;
+  int i;
+  for (i = 0; i < count; ++i) {
+    hash += (hash << 5) + src[i];
+  }
+  return hash;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/compare_neon.cc b/drivers/theoraplayer/src/YUV/libyuv/src/compare_neon.cc
new file mode 100755
index 0000000000..bb843a6ab8
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/compare_neon.cc
@@ -0,0 +1,64 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  volatile uint32 sse;
+  asm volatile (
+#ifdef _ANDROID
+	".fpu neon\n"
+#endif
+    "vmov.u8    q8, #0                         \n"
+    "vmov.u8    q10, #0                        \n"
+    "vmov.u8    q9, #0                         \n"
+    "vmov.u8    q11, #0                        \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"
+    "vld1.8     {q1}, [%1]!                    \n"
+    "subs       %2, %2, #16                    \n"
+    "vsubl.u8   q2, d0, d2                     \n"
+    "vsubl.u8   q3, d1, d3                     \n"
+    "vmlal.s16  q8, d4, d4                     \n"
+    "vmlal.s16  q9, d6, d6                     \n"
+    "vmlal.s16  q10, d5, d5                    \n"
+    "vmlal.s16  q11, d7, d7                    \n"
+    "bgt        1b                             \n"
+
+    "vadd.u32   q8, q8, q9                     \n"
+    "vadd.u32   q10, q10, q11                  \n"
+    "vadd.u32   q11, q8, q10                   \n"
+    "vpaddl.u32 q1, q11                        \n"
+    "vadd.u64   d0, d2, d3                     \n"
+    "vmov.32    %3, d0[0]                      \n"
+    : "+r"(src_a),
+      "+r"(src_b),
+      "+r"(count),
+      "=r"(sse)
+    :
+    : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+  return sse;
+}
+
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/compare_posix.cc b/drivers/theoraplayer/src/YUV/libyuv/src/compare_posix.cc
new file mode 100755
index 0000000000..ac361190e8
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/compare_posix.cc
@@ -0,0 +1,158 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse;
+  asm volatile (  // NOLINT
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10, 1) ",%1          \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psubusb   %%xmm2,%%xmm1                   \n"
+    "psubusb   %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpckhbw %%xmm5,%%xmm2                   \n"
+    "pmaddwd   %%xmm1,%%xmm1                   \n"
+    "pmaddwd   %%xmm2,%%xmm2                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "jg        1b                              \n"
+
+    "pshufd    $0xee,%%xmm0,%%xmm1             \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "pshufd    $0x1,%%xmm0,%%xmm1              \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0,%3                       \n"
+
+  : "+r"(src_a),      // %0
+    "+r"(src_b),      // %1
+    "+r"(count),      // %2
+    "=g"(sse)         // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );  // NOLINT
+  return sse;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  uint32 hash;
+  asm volatile (  // NOLINT
+    "movd      %2,%%xmm0                       \n"
+    "pxor      %%xmm7,%%xmm7                   \n"
+    "movdqa    %4,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "pmulld    %%xmm6,%%xmm0                   \n"
+    "movdqa    %5,%%xmm5                       \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm7,%%xmm3                   \n"
+    "pmulld    %%xmm5,%%xmm3                   \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpckhwd %%xmm7,%%xmm4                   \n"
+    "pmulld    %%xmm5,%%xmm4                   \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "punpckhbw %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm7,%%xmm2                   \n"
+    "pmulld    %%xmm5,%%xmm2                   \n"
+    "movdqa    %8,%%xmm5                       \n"
+    "punpckhwd %%xmm7,%%xmm1                   \n"
+    "pmulld    %%xmm5,%%xmm1                   \n"
+    "paddd     %%xmm4,%%xmm3                   \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "sub       $0x10,%1                        \n"
+    "paddd     %%xmm3,%%xmm1                   \n"
+    "pshufd    $0xe,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "pshufd    $0x1,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "jg        1b                              \n"
+    "movd      %%xmm0,%3                       \n"
+  : "+r"(src),        // %0
+    "+r"(count),      // %1
+    "+rm"(seed),      // %2
+    "=g"(hash)        // %3
+  : "m"(kHash16x33),  // %4
+    "m"(kHashMul0),   // %5
+    "m"(kHashMul1),   // %6
+    "m"(kHashMul2),   // %7
+    "m"(kHashMul3)    // %8
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );  // NOLINT
+  return hash;
+}
+#endif  // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/compare_win.cc b/drivers/theoraplayer/src/YUV/libyuv/src/compare_win.cc
new file mode 100755
index 0000000000..99831651f5
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/compare_win.cc
@@ -0,0 +1,232 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+__declspec(naked) __declspec(align(16))
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    pxor       xmm0, xmm0
+    pxor       xmm5, xmm5
+
+    align      4
+  wloop:
+    movdqa     xmm1, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, [edx]
+    lea        edx,  [edx + 16]
+    sub        ecx, 16
+    movdqa     xmm3, xmm1  // abs trick
+    psubusb    xmm1, xmm2
+    psubusb    xmm2, xmm3
+    por        xmm1, xmm2
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm1, xmm5
+    punpckhbw  xmm2, xmm5
+    pmaddwd    xmm1, xmm1
+    pmaddwd    xmm2, xmm2
+    paddd      xmm0, xmm1
+    paddd      xmm0, xmm2
+    jg         wloop
+
+    pshufd     xmm1, xmm0, 0xee
+    paddd      xmm0, xmm1
+    pshufd     xmm1, xmm0, 0x01
+    paddd      xmm0, xmm1
+    movd       eax, xmm0
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+// C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX.
+#pragma warning(disable: 4752)
+__declspec(naked) __declspec(align(16))
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    vpxor      ymm0, ymm0, ymm0  // sum
+    vpxor      ymm5, ymm5, ymm5  // constant 0 for unpck
+    sub        edx, eax
+
+    align      4
+  wloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + edx]
+    lea        eax,  [eax + 32]
+    sub        ecx, 32
+    vpsubusb   ymm3, ymm1, ymm2  // abs difference trick
+    vpsubusb   ymm2, ymm2, ymm1
+    vpor       ymm1, ymm2, ymm3
+    vpunpcklbw ymm2, ymm1, ymm5  // u16.  mutates order.
+    vpunpckhbw ymm1, ymm1, ymm5
+    vpmaddwd   ymm2, ymm2, ymm2  // square + hadd to u32.
+    vpmaddwd   ymm1, ymm1, ymm1
+    vpaddd     ymm0, ymm0, ymm1
+    vpaddd     ymm0, ymm0, ymm2
+    jg         wloop
+
+    vpshufd    ymm1, ymm0, 0xee  // 3, 2 + 1, 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpshufd    ymm1, ymm0, 0x01  // 1 + 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpermq     ymm1, ymm0, 0x02  // high + low lane.
+    vpaddd     ymm0, ymm0, ymm1
+    vmovd      eax, xmm0
+    vzeroupper
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+// 27: 66 0F 38 40 C6     pmulld      xmm0,xmm6
+// 44: 66 0F 38 40 DD     pmulld      xmm3,xmm5
+// 59: 66 0F 38 40 E5     pmulld      xmm4,xmm5
+// 72: 66 0F 38 40 D5     pmulld      xmm2,xmm5
+// 83: 66 0F 38 40 CD     pmulld      xmm1,xmm5
+#define pmulld(reg) _asm _emit 0x66 _asm _emit 0x0F _asm _emit 0x38 \
+    _asm _emit 0x40 _asm _emit reg
+
+__declspec(naked) __declspec(align(16))
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+
+    pxor       xmm7, xmm7        // constant 0 for unpck
+    movdqa     xmm6, kHash16x33
+
+    align      4
+  wloop:
+    movdqu     xmm1, [eax]       // src[0-15]
+    lea        eax, [eax + 16]
+    pmulld(0xc6)                 // pmulld      xmm0,xmm6  hash *= 33 ^ 16
+    movdqa     xmm5, kHashMul0
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm2, xmm7        // src[0-7]
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm3, xmm7        // src[0-3]
+    pmulld(0xdd)                 // pmulld     xmm3, xmm5
+    movdqa     xmm5, kHashMul1
+    movdqa     xmm4, xmm2
+    punpckhwd  xmm4, xmm7        // src[4-7]
+    pmulld(0xe5)                 // pmulld     xmm4, xmm5
+    movdqa     xmm5, kHashMul2
+    punpckhbw  xmm1, xmm7        // src[8-15]
+    movdqa     xmm2, xmm1
+    punpcklwd  xmm2, xmm7        // src[8-11]
+    pmulld(0xd5)                 // pmulld     xmm2, xmm5
+    movdqa     xmm5, kHashMul3
+    punpckhwd  xmm1, xmm7        // src[12-15]
+    pmulld(0xcd)                 // pmulld     xmm1, xmm5
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    sub        ecx, 16
+    paddd      xmm1, xmm3
+
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+__declspec(naked) __declspec(align(16))
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+    movdqa     xmm6, kHash16x33
+
+    align      4
+  wloop:
+    vpmovzxbd  xmm3, dword ptr [eax]  // src[0-3]
+    pmulld     xmm0, xmm6  // hash *= 33 ^ 16
+    vpmovzxbd  xmm4, dword ptr [eax + 4]  // src[4-7]
+    pmulld     xmm3, kHashMul0
+    vpmovzxbd  xmm2, dword ptr [eax + 8]  // src[8-11]
+    pmulld     xmm4, kHashMul1
+    vpmovzxbd  xmm1, dword ptr [eax + 12]  // src[12-15]
+    pmulld     xmm2, kHashMul2
+    lea        eax, [eax + 16]
+    pmulld     xmm1, kHashMul3
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    sub        ecx, 16
+    paddd      xmm1, xmm3
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/convert.cc b/drivers/theoraplayer/src/YUV/libyuv/src/convert.cc
new file mode 100755
index 0000000000..c8408dc798
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/convert.cc
@@ -0,0 +1,1491 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// Any I4xx To I420 format with mirroring.
+static int I4xxToI420(const uint8* src_y, int src_stride_y,
+                      const uint8* src_u, int src_stride_u,
+                      const uint8* src_v, int src_stride_v,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int src_y_width, int src_y_height,
+                      int src_uv_width, int src_uv_height) {
+  if (src_y_width == 0 || src_y_height == 0 ||
+      src_uv_width == 0 || src_uv_height == 0) {
+    return -1;
+  }
+  const int dst_y_width = Abs(src_y_width);
+  const int dst_y_height = Abs(src_y_height);
+  const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1);
+  const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1);
+  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
+             dst_y, dst_stride_y, dst_y_width, dst_y_height,
+             kFilterBilinear);
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
+             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
+             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  return 0;
+}
+
+// Copy I420 with optional flipping
+// TODO(fbarchard): Use Scale plane which supports mirroring, but ensure
+// is does row coalescing.
+LIBYUV_API
+int I420Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    const int halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  // Copy UV planes.
+  const int halfwidth = (width + 1) >> 1;
+  const int halfheight = (height + 1) >> 1;
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// 422 chroma is 1/2 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I422ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int src_uv_width = SUBSAMPLE(width, 1, 1);
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    src_uv_width, height);
+}
+
+// 444 chroma is 1x width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I444ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    width, height);
+}
+
+// 411 chroma is 1/4 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I411ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int src_uv_width = SUBSAMPLE(width, 3, 2);
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    src_uv_width, height);
+}
+
+// I400 is greyscale typically used in MJPG
+LIBYUV_API
+int I400ToI420(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_y || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  SetPlane(dst_u, dst_stride_u, halfwidth, halfheight, 128);
+  SetPlane(dst_v, dst_stride_v, halfwidth, halfheight, 128);
+  return 0;
+}
+
+static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
+                       uint8* dst, int dst_stride,
+                       int width, int height) {
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src, 16) &&
+      IS_ALIGNED(src_stride_0, 16) && IS_ALIGNED(src_stride_1, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (int y = 0; y < height - 1; y += 2) {
+    CopyRow(src, dst, width);
+    CopyRow(src + src_stride_0, dst + dst_stride, width);
+    src += src_stride_0 + src_stride_1;
+    dst += dst_stride * 2;
+  }
+  if (height & 1) {
+    CopyRow(src, dst, width);
+  }
+}
+
+// Support converting from FOURCC_M420
+// Useful for bandwidth constrained transports like USB 1.0 and 2.0 and for
+// easy conversion to I420.
+// M420 format description:
+// M420 is row biplanar 420: 2 rows of Y and 1 row of UV.
+// Chroma is half width / half height. (420)
+// src_stride_m420 is row planar. Normally this will be the width in pixels.
+//   The UV plane is half width, but 2 values, so src_stride_m420 applies to
+//   this as well as the two Y planes.
+static int X420ToI420(const uint8* src_y,
+                      int src_stride_y0, int src_stride_y1,
+                      const uint8* src_uv, int src_stride_uv,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int width, int height) {
+  if (!src_y || !src_uv ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    int halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
+    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // Coalesce rows.
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (src_stride_y0 == width &&
+      src_stride_y1 == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y0 = src_stride_y1 = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == halfwidth * 2 &&
+      dst_stride_u == halfwidth &&
+      dst_stride_v == halfwidth) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+  }
+  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) =
+      SplitUVRow_C;
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_uv, 16) && IS_ALIGNED(src_stride_uv, 16) &&
+          IS_ALIGNED(dst_u, 16) && IS_ALIGNED(dst_stride_u, 16) &&
+          IS_ALIGNED(dst_v, 16) && IS_ALIGNED(dst_stride_v, 16)) {
+        SplitUVRow = SplitUVRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+    SplitUVRow = SplitUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && halfwidth >= 16) {
+    SplitUVRow = SplitUVRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_Unaligned_MIPS_DSPR2;
+      if (IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) &&
+          IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) &&
+          IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) {
+        SplitUVRow = SplitUVRow_MIPS_DSPR2;
+      }
+    }
+  }
+#endif
+
+  if (dst_y) {
+    if (src_stride_y0 == src_stride_y1) {
+      CopyPlane(src_y, src_stride_y0, dst_y, dst_stride_y, width, height);
+    } else {
+      CopyPlane2(src_y, src_stride_y0, src_stride_y1, dst_y, dst_stride_y,
+                 width, height);
+    }
+  }
+
+  for (int y = 0; y < halfheight; ++y) {
+    // Copy a row of UV.
+    SplitUVRow(src_uv, dst_u, dst_v, halfwidth);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += src_stride_uv;
+  }
+  return 0;
+}
+
+// Convert NV12 to I420.
+LIBYUV_API
+int NV12ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_y, src_stride_y, src_stride_y,
+                    src_uv, src_stride_uv,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height);
+}
+
+// Convert NV21 to I420.  Same as NV12 but u and v pointers swapped.
+LIBYUV_API
+int NV21ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_y, src_stride_y, src_stride_y,
+                    src_vu, src_stride_vu,
+                    dst_y, dst_stride_y,
+                    dst_v, dst_stride_v,
+                    dst_u, dst_stride_u,
+                    width, height);
+}
+
+// Convert M420 to I420.
+LIBYUV_API
+int M420ToI420(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_m420, src_stride_m420, src_stride_m420 * 2,
+                    src_m420 + src_stride_m420 * 2, src_stride_m420 * 3,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height);
+}
+
+// Convert Q420 to I420.
+// Format is rows of YY/YUYV
+LIBYUV_API
+int Q420ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_y || !src_yuy2 ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    int halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
+    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // CopyRow for rows of just Y in Q420 copied to Y plane of I420.
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_X86)
+  if (IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  void (*YUY2ToUV422Row)(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+      int pix) = YUY2ToUV422Row_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) =
+      YUY2ToYRow_C;
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Unaligned_SSE2;
+      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
+        YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          YUY2ToYRow = YUY2ToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUV422Row = YUY2ToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    CopyRow(src_y, dst_y, width);
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+  }
+  return 0;
+}
+
+// Convert YUY2 to I420.
+LIBYUV_API
+int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  void (*YUY2ToUVRow)(const uint8* src_yuy2, int src_stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+  void (*YUY2ToYRow)(const uint8* src_yuy2,
+                     uint8* dst_y, int pix);
+  YUY2ToYRow = YUY2ToYRow_C;
+  YUY2ToUVRow = YUY2ToUVRow_C;
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    YUY2ToUVRow = YUY2ToUVRow_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUVRow = YUY2ToUVRow_Unaligned_SSE2;
+      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
+        YUY2ToUVRow = YUY2ToUVRow_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          YUY2ToYRow = YUY2ToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    YUY2ToUVRow = YUY2ToUVRow_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUVRow = YUY2ToUVRow_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUVRow = YUY2ToUVRow_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUVRow = YUY2ToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    YUY2ToUVRow(src_yuy2, src_stride_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    YUY2ToYRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y, width);
+    src_yuy2 += src_stride_yuy2 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    YUY2ToUVRow(src_yuy2, 0, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert UYVY to I420.
+LIBYUV_API
+int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  void (*UYVYToUVRow)(const uint8* src_uyvy, int src_stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+                     uint8* dst_y, int pix);
+  UYVYToYRow = UYVYToYRow_C;
+  UYVYToUVRow = UYVYToUVRow_C;
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    UYVYToUVRow = UYVYToUVRow_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUVRow = UYVYToUVRow_Unaligned_SSE2;
+      UYVYToYRow = UYVYToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16)) {
+        UYVYToUVRow = UYVYToUVRow_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          UYVYToYRow = UYVYToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    UYVYToUVRow = UYVYToUVRow_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUVRow = UYVYToUVRow_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (width >= 16) {
+      UYVYToUVRow = UYVYToUVRow_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUVRow = UYVYToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    UYVYToUVRow(src_uyvy, src_stride_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    UYVYToYRow(src_uyvy + src_stride_uyvy, dst_y + dst_stride_y, width);
+    src_uyvy += src_stride_uyvy * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    UYVYToUVRow(src_uyvy, 0, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert ARGB to I420.
+LIBYUV_API
+int ARGBToI420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_argb ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUVRow = ARGBToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          ARGBToYRow = ARGBToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVRow = ARGBToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVRow = ARGBToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    ARGBToUVRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+    src_argb += src_stride_argb * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ARGBToUVRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert BGRA to I420.
+LIBYUV_API
+int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_bgra ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+  void (*BGRAToUVRow)(const uint8* src_bgra0, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int width) = BGRAToUVRow_C;
+  void (*BGRAToYRow)(const uint8* src_bgra, uint8* dst_y, int pix) =
+      BGRAToYRow_C;
+#if defined(HAS_BGRATOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    BGRAToUVRow = BGRAToUVRow_Any_SSSE3;
+    BGRAToYRow = BGRAToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToUVRow = BGRAToUVRow_Unaligned_SSSE3;
+      BGRAToYRow = BGRAToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_bgra, 16) && IS_ALIGNED(src_stride_bgra, 16)) {
+        BGRAToUVRow = BGRAToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          BGRAToYRow = BGRAToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_BGRATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    BGRAToYRow = BGRAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      BGRAToYRow = BGRAToYRow_NEON;
+    }
+    if (width >= 16) {
+      BGRAToUVRow = BGRAToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        BGRAToUVRow = BGRAToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    BGRAToUVRow(src_bgra, src_stride_bgra, dst_u, dst_v, width);
+    BGRAToYRow(src_bgra, dst_y, width);
+    BGRAToYRow(src_bgra + src_stride_bgra, dst_y + dst_stride_y, width);
+    src_bgra += src_stride_bgra * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    BGRAToUVRow(src_bgra, 0, dst_u, dst_v, width);
+    BGRAToYRow(src_bgra, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert ABGR to I420.
+LIBYUV_API
+int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_abgr ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+  void (*ABGRToUVRow)(const uint8* src_abgr0, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int width) = ABGRToUVRow_C;
+  void (*ABGRToYRow)(const uint8* src_abgr, uint8* dst_y, int pix) =
+      ABGRToYRow_C;
+#if defined(HAS_ABGRTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    ABGRToYRow = ABGRToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_Unaligned_SSSE3;
+      ABGRToYRow = ABGRToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_abgr, 16) && IS_ALIGNED(src_stride_abgr, 16)) {
+        ABGRToUVRow = ABGRToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          ABGRToYRow = ABGRToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_ABGRTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ABGRToYRow = ABGRToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ABGRToYRow = ABGRToYRow_NEON;
+    }
+    if (width >= 16) {
+      ABGRToUVRow = ABGRToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ABGRToUVRow = ABGRToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    ABGRToUVRow(src_abgr, src_stride_abgr, dst_u, dst_v, width);
+    ABGRToYRow(src_abgr, dst_y, width);
+    ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width);
+    src_abgr += src_stride_abgr * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ABGRToUVRow(src_abgr, 0, dst_u, dst_v, width);
+    ABGRToYRow(src_abgr, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert RGBA to I420.
+LIBYUV_API
+int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_rgba ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgba = src_rgba + (height - 1) * src_stride_rgba;
+    src_stride_rgba = -src_stride_rgba;
+  }
+  void (*RGBAToUVRow)(const uint8* src_rgba0, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int width) = RGBAToUVRow_C;
+  void (*RGBAToYRow)(const uint8* src_rgba, uint8* dst_y, int pix) =
+      RGBAToYRow_C;
+#if defined(HAS_RGBATOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
+    RGBAToYRow = RGBAToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_Unaligned_SSSE3;
+      RGBAToYRow = RGBAToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_rgba, 16) && IS_ALIGNED(src_stride_rgba, 16)) {
+        RGBAToUVRow = RGBAToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          RGBAToYRow = RGBAToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_RGBATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGBAToYRow = RGBAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGBAToYRow = RGBAToYRow_NEON;
+    }
+    if (width >= 16) {
+      RGBAToUVRow = RGBAToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGBAToUVRow = RGBAToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    RGBAToUVRow(src_rgba, src_stride_rgba, dst_u, dst_v, width);
+    RGBAToYRow(src_rgba, dst_y, width);
+    RGBAToYRow(src_rgba + src_stride_rgba, dst_y + dst_stride_y, width);
+    src_rgba += src_stride_rgba * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    RGBAToUVRow(src_rgba, 0, dst_u, dst_v, width);
+    RGBAToYRow(src_rgba, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert RGB24 to I420.
+LIBYUV_API
+int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height) {
+  if (!src_rgb24 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+
+#if defined(HAS_RGB24TOYROW_NEON)
+  void (*RGB24ToUVRow)(const uint8* src_rgb24, int src_stride_rgb24,
+      uint8* dst_u, uint8* dst_v, int width) = RGB24ToUVRow_C;
+  void (*RGB24ToYRow)(const uint8* src_rgb24, uint8* dst_y, int pix) =
+      RGB24ToYRow_C;
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGB24ToYRow = RGB24ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB24ToYRow = RGB24ToYRow_NEON;
+    }
+    if (width >= 16) {
+      RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGB24ToUVRow = RGB24ToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_RGB24TOYROW_NEON
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+
+  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB24ToARGBRow_C;
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#endif
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_RGB24TOYROW_NEON
+
+  for (int y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB24TOYROW_NEON)
+    RGB24ToUVRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
+    RGB24ToYRow(src_rgb24, dst_y, width);
+    RGB24ToYRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
+#else
+    RGB24ToARGBRow(src_rgb24, row, width);
+    RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_rgb24 += src_stride_rgb24 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_RGB24TOYROW_NEON)
+    RGB24ToUVRow(src_rgb24, 0, dst_u, dst_v, width);
+    RGB24ToYRow(src_rgb24, dst_y, width);
+#else
+    RGB24ToARGBRow(src_rgb24, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_RGB24TOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+// Convert RAW to I420.
+LIBYUV_API
+int RAWToI420(const uint8* src_raw, int src_stride_raw,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int width, int height) {
+  if (!src_raw || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+
+#if defined(HAS_RAWTOYROW_NEON)
+  void (*RAWToUVRow)(const uint8* src_raw, int src_stride_raw,
+      uint8* dst_u, uint8* dst_v, int width) = RAWToUVRow_C;
+  void (*RAWToYRow)(const uint8* src_raw, uint8* dst_y, int pix) =
+      RAWToYRow_C;
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RAWToYRow = RAWToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToYRow = RAWToYRow_NEON;
+    }
+    if (width >= 16) {
+      RAWToUVRow = RAWToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RAWToUVRow = RAWToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_RAWTOYROW_NEON
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+
+  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RAWToARGBRow_C;
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_RAWTOYROW_NEON
+
+  for (int y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RAWTOYROW_NEON)
+    RAWToUVRow(src_raw, src_stride_raw, dst_u, dst_v, width);
+    RAWToYRow(src_raw, dst_y, width);
+    RAWToYRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
+#else
+    RAWToARGBRow(src_raw, row, width);
+    RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_raw += src_stride_raw * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_RAWTOYROW_NEON)
+    RAWToUVRow(src_raw, 0, dst_u, dst_v, width);
+    RAWToYRow(src_raw, dst_y, width);
+#else
+    RAWToARGBRow(src_raw, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_RAWTOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+// Convert RGB565 to I420.
+LIBYUV_API
+int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height) {
+  if (!src_rgb565 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565;
+    src_stride_rgb565 = -src_stride_rgb565;
+  }
+
+#if defined(HAS_RGB565TOYROW_NEON)
+  void (*RGB565ToUVRow)(const uint8* src_rgb565, int src_stride_rgb565,
+      uint8* dst_u, uint8* dst_v, int width) = RGB565ToUVRow_C;
+  void (*RGB565ToYRow)(const uint8* src_rgb565, uint8* dst_y, int pix) =
+      RGB565ToYRow_C;
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGB565ToYRow = RGB565ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToYRow = RGB565ToYRow_NEON;
+    }
+    if (width >= 16) {
+      RGB565ToUVRow = RGB565ToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGB565ToUVRow = RGB565ToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_RGB565TOYROW_NEON
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+
+  void (*RGB565ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB565ToARGBRow_C;
+#if defined(HAS_RGB565TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
+    }
+  }
+#endif
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_RGB565TOYROW_NEON
+
+  for (int y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB565TOYROW_NEON)
+    RGB565ToUVRow(src_rgb565, src_stride_rgb565, dst_u, dst_v, width);
+    RGB565ToYRow(src_rgb565, dst_y, width);
+    RGB565ToYRow(src_rgb565 + src_stride_rgb565, dst_y + dst_stride_y, width);
+#else
+    RGB565ToARGBRow(src_rgb565, row, width);
+    RGB565ToARGBRow(src_rgb565 + src_stride_rgb565, row + kRowSize, width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_rgb565 += src_stride_rgb565 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_RGB565TOYROW_NEON)
+    RGB565ToUVRow(src_rgb565, 0, dst_u, dst_v, width);
+    RGB565ToYRow(src_rgb565, dst_y, width);
+#else
+    RGB565ToARGBRow(src_rgb565, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_RGB565TOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+// Convert ARGB1555 to I420.
+LIBYUV_API
+int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height) {
+  if (!src_argb1555 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555;
+    src_stride_argb1555 = -src_stride_argb1555;
+  }
+
+#if defined(HAS_ARGB1555TOYROW_NEON)
+  void (*ARGB1555ToUVRow)(const uint8* src_argb1555, int src_stride_argb1555,
+      uint8* dst_u, uint8* dst_v, int width) = ARGB1555ToUVRow_C;
+  void (*ARGB1555ToYRow)(const uint8* src_argb1555, uint8* dst_y, int pix) =
+      ARGB1555ToYRow_C;
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGB1555ToUVRow = ARGB1555ToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGB1555ToUVRow = ARGB1555ToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_ARGB1555TOYROW_NEON
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+
+  void (*ARGB1555ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      ARGB1555ToARGBRow_C;
+#if defined(HAS_ARGB1555TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
+    }
+  }
+#endif
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_ARGB1555TOYROW_NEON
+
+  for (int y = 0; y < height - 1; y += 2) {
+#if defined(HAS_ARGB1555TOYROW_NEON)
+    ARGB1555ToUVRow(src_argb1555, src_stride_argb1555, dst_u, dst_v, width);
+    ARGB1555ToYRow(src_argb1555, dst_y, width);
+    ARGB1555ToYRow(src_argb1555 + src_stride_argb1555, dst_y + dst_stride_y,
+                   width);
+#else
+    ARGB1555ToARGBRow(src_argb1555, row, width);
+    ARGB1555ToARGBRow(src_argb1555 + src_stride_argb1555, row + kRowSize,
+                      width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_argb1555 += src_stride_argb1555 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_ARGB1555TOYROW_NEON)
+    ARGB1555ToUVRow(src_argb1555, 0, dst_u, dst_v, width);
+    ARGB1555ToYRow(src_argb1555, dst_y, width);
+#else
+    ARGB1555ToARGBRow(src_argb1555, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_ARGB1555TOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+// Convert ARGB4444 to I420.
+LIBYUV_API
+int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height) {
+  if (!src_argb4444 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444;
+    src_stride_argb4444 = -src_stride_argb4444;
+  }
+
+#if defined(HAS_ARGB4444TOYROW_NEON)
+  void (*ARGB4444ToUVRow)(const uint8* src_argb4444, int src_stride_argb4444,
+      uint8* dst_u, uint8* dst_v, int width) = ARGB4444ToUVRow_C;
+  void (*ARGB4444ToYRow)(const uint8* src_argb4444, uint8* dst_y, int pix) =
+      ARGB4444ToYRow_C;
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGB4444ToYRow = ARGB4444ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToYRow = ARGB4444ToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGB4444ToUVRow = ARGB4444ToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGB4444ToUVRow = ARGB4444ToUVRow_NEON;
+      }
+    }
+  }
+#else  // HAS_ARGB4444TOYROW_NEON
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+
+  void (*ARGB4444ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      ARGB4444ToARGBRow_C;
+#if defined(HAS_ARGB4444TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
+    }
+  }
+#endif
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif  // HAS_ARGBTOUVROW_SSSE3
+#endif  // HAS_ARGB4444TOYROW_NEON
+
+  for (int y = 0; y < height - 1; y += 2) {
+#if defined(HAS_ARGB4444TOYROW_NEON)
+    ARGB4444ToUVRow(src_argb4444, src_stride_argb4444, dst_u, dst_v, width);
+    ARGB4444ToYRow(src_argb4444, dst_y, width);
+    ARGB4444ToYRow(src_argb4444 + src_stride_argb4444, dst_y + dst_stride_y,
+                   width);
+#else
+    ARGB4444ToARGBRow(src_argb4444, row, width);
+    ARGB4444ToARGBRow(src_argb4444 + src_stride_argb4444, row + kRowSize,
+                      width);
+    ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+    ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+    src_argb4444 += src_stride_argb4444 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+#if defined(HAS_ARGB4444TOYROW_NEON)
+    ARGB4444ToUVRow(src_argb4444, 0, dst_u, dst_v, width);
+    ARGB4444ToYRow(src_argb4444, dst_y, width);
+#else
+    ARGB4444ToARGBRow(src_argb4444, row, width);
+    ARGBToUVRow(row, 0, dst_u, dst_v, width);
+    ARGBToYRow(row, dst_y, width);
+#endif
+  }
+#if !defined(HAS_ARGB4444TOYROW_NEON)
+  free_aligned_buffer_64(row);
+#endif
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/convert_argb.cc b/drivers/theoraplayer/src/YUV/libyuv/src/convert_argb.cc
new file mode 100755
index 0000000000..a8aab91478
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/convert_argb.cc
@@ -0,0 +1,901 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy ARGB with optional flipping
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height) {
+  if (!src_argb || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  CopyPlane(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+            width * 4, height);
+  return 0;
+}
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u == width &&
+      src_stride_v == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+  void (*I444ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I444ToARGBRow_C;
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I444ToARGBRow = I444ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I444ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I422ToARGBRow = I422ToARGBRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I411 to ARGB.
+LIBYUV_API
+int I411ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 4 == width &&
+      src_stride_v * 4 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+  void (*I411ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I411ToARGBRow_C;
+#if defined(HAS_I411TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I411ToARGBRow = I411ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I411ToARGBRow = I411ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I411ToARGBRow = I411ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I411TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I411ToARGBRow = I411ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I411ToARGBRow = I411ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I411ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I400 to ARGB.
+LIBYUV_API
+int I400ToARGB_Reference(const uint8* src_y, int src_stride_y,
+                         uint8* dst_argb, int dst_stride_argb,
+                         int width, int height) {
+  if (!src_y || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+  void (*YToARGBRow)(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) = YToARGBRow_C;
+#if defined(HAS_YTOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    YToARGBRow = YToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      YToARGBRow = YToARGBRow_SSE2;
+    }
+  }
+#elif defined(HAS_YTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YToARGBRow = YToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      YToARGBRow = YToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    YToARGBRow(src_y, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+  }
+  return 0;
+}
+
+// Convert I400 to ARGB.
+LIBYUV_API
+int I400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_y || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+  void (*I400ToARGBRow)(const uint8* src_y, uint8* dst_argb, int pix) =
+      I400ToARGBRow_C;
+#if defined(HAS_I400TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8) {
+    I400ToARGBRow = I400ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      I400ToARGBRow = I400ToARGBRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I400ToARGBRow = I400ToARGBRow_SSE2;
+      }
+    }
+  }
+#elif defined(HAS_I400TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I400ToARGBRow = I400ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I400ToARGBRow = I400ToARGBRow_NEON;
+    }
+  }
+#endif
+  for (int y = 0; y < height; ++y) {
+    I400ToARGBRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle table for converting BGRA to ARGB.
+static uvec8 kShuffleMaskBGRAToARGB = {
+  3u, 2u, 1u, 0u, 7u, 6u, 5u, 4u, 11u, 10u, 9u, 8u, 15u, 14u, 13u, 12u
+};
+
+// Shuffle table for converting ABGR to ARGB.
+static uvec8 kShuffleMaskABGRToARGB = {
+  2u, 1u, 0u, 3u, 6u, 5u, 4u, 7u, 10u, 9u, 8u, 11u, 14u, 13u, 12u, 15u
+};
+
+// Shuffle table for converting RGBA to ARGB.
+static uvec8 kShuffleMaskRGBAToARGB = {
+  1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u
+};
+
+// Convert BGRA to ARGB.
+LIBYUV_API
+int BGRAToARGB(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskBGRAToARGB),
+                     width, height);
+}
+
+// Convert ABGR to ARGB.
+LIBYUV_API
+int ABGRToARGB(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskABGRToARGB),
+                     width, height);
+}
+
+// Convert RGBA to ARGB.
+LIBYUV_API
+int RGBAToARGB(const uint8* src_rgba, int src_stride_rgba,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_rgba, src_stride_rgba,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskRGBAToARGB),
+                     width, height);
+}
+
+// Convert RGB24 to ARGB.
+LIBYUV_API
+int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  if (!src_rgb24 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb24 == width * 3 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_rgb24 = dst_stride_argb = 0;
+  }
+  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB24ToARGBRow_C;
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#elif defined(HAS_RGB24TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    RGB24ToARGBRow(src_rgb24, dst_argb, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert RAW to ARGB.
+LIBYUV_API
+int RAWToARGB(const uint8* src_raw, int src_stride_raw,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  if (!src_raw || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_argb = 0;
+  }
+  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RAWToARGBRow_C;
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#elif defined(HAS_RAWTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RAWToARGBRow = RAWToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToARGBRow = RAWToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    RAWToARGBRow(src_raw, dst_argb, width);
+    src_raw += src_stride_raw;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert RGB565 to ARGB.
+LIBYUV_API
+int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  if (!src_rgb565 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565;
+    src_stride_rgb565 = -src_stride_rgb565;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb565 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_rgb565 = dst_stride_argb = 0;
+  }
+  void (*RGB565ToARGBRow)(const uint8* src_rgb565, uint8* dst_argb, int pix) =
+      RGB565ToARGBRow_C;
+#if defined(HAS_RGB565TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
+    }
+  }
+#elif defined(HAS_RGB565TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    RGB565ToARGBRow(src_rgb565, dst_argb, width);
+    src_rgb565 += src_stride_rgb565;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB1555 to ARGB.
+LIBYUV_API
+int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height) {
+  if (!src_argb1555 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555;
+    src_stride_argb1555 = -src_stride_argb1555;
+  }
+  // Coalesce rows.
+  if (src_stride_argb1555 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb1555 = dst_stride_argb = 0;
+  }
+  void (*ARGB1555ToARGBRow)(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) = ARGB1555ToARGBRow_C;
+#if defined(HAS_ARGB1555TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
+    }
+  }
+#elif defined(HAS_ARGB1555TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGB1555ToARGBRow(src_argb1555, dst_argb, width);
+    src_argb1555 += src_stride_argb1555;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB4444 to ARGB.
+LIBYUV_API
+int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height) {
+  if (!src_argb4444 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444;
+    src_stride_argb4444 = -src_stride_argb4444;
+  }
+  // Coalesce rows.
+  if (src_stride_argb4444 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb4444 = dst_stride_argb = 0;
+  }
+  void (*ARGB4444ToARGBRow)(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) = ARGB4444ToARGBRow_C;
+#if defined(HAS_ARGB4444TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
+    }
+  }
+#elif defined(HAS_ARGB4444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGB4444ToARGBRow(src_argb4444, dst_argb, width);
+    src_argb4444 += src_stride_argb4444;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_y || !src_uv || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  void (*NV12ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV12ToARGBRow_C;
+#if defined(HAS_NV12TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    NV12ToARGBRow(src_y, src_uv, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_y || !src_uv || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  void (*NV21ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV21ToARGBRow_C;
+#if defined(HAS_NV21TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        NV21ToARGBRow = NV21ToARGBRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    NV21ToARGBRow(src_y, src_uv, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert M420 to ARGB.
+LIBYUV_API
+int M420ToARGB(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_m420 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  void (*NV12ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV12ToARGBRow_C;
+#if defined(HAS_NV12TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width);
+    NV12ToARGBRow(src_m420 + src_stride_m420, src_m420 + src_stride_m420 * 2,
+                  dst_argb + dst_stride_argb, width);
+    dst_argb += dst_stride_argb * 2;
+    src_m420 += src_stride_m420 * 3;
+  }
+  if (height & 1) {
+    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width);
+  }
+  return 0;
+}
+
+// Convert YUY2 to ARGB.
+LIBYUV_API
+int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_yuy2 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_argb = 0;
+  }
+  void (*YUY2ToARGBRow)(const uint8* src_yuy2, uint8* dst_argb, int pix) =
+      YUY2ToARGBRow_C;
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+  // Posix is 16, Windows is 8.
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        YUY2ToARGBRow = YUY2ToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_YUY2TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_NEON;
+    }
+  }
+#endif
+  for (int y = 0; y < height; ++y) {
+    YUY2ToARGBRow(src_yuy2, dst_argb, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert UYVY to ARGB.
+LIBYUV_API
+int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_uyvy || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_argb = 0;
+  }
+  void (*UYVYToARGBRow)(const uint8* src_uyvy, uint8* dst_argb, int pix) =
+      UYVYToARGBRow_C;
+#if defined(HAS_UYVYTOARGBROW_SSSE3)
+  // Posix is 16, Windows is 8.
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToARGBRow = UYVYToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        UYVYToARGBRow = UYVYToARGBRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_UYVYTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      UYVYToARGBRow = UYVYToARGBRow_NEON;
+    }
+  }
+#endif
+  for (int y = 0; y < height; ++y) {
+    UYVYToARGBRow(src_uyvy, dst_argb, width);
+    src_uyvy += src_stride_uyvy;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/convert_from.cc b/drivers/theoraplayer/src/YUV/libyuv/src/convert_from.cc
new file mode 100755
index 0000000000..1e10832856
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/convert_from.cc
@@ -0,0 +1,1196 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_from.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/convert.h"  // For I420Copy
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/video_common.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// I420 To any I4xx YUV format with mirroring.
+static int I420ToI4xx(const uint8* src_y, int src_stride_y,
+                      const uint8* src_u, int src_stride_u,
+                      const uint8* src_v, int src_stride_v,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int src_y_width, int src_y_height,
+                      int dst_uv_width, int dst_uv_height) {
+  if (src_y_width == 0 || src_y_height == 0 ||
+      dst_uv_width <= 0 || dst_uv_height <= 0) {
+    return -1;
+  }
+  const int dst_y_width = Abs(src_y_width);
+  const int dst_y_height = Abs(src_y_height);
+  const int src_uv_width = SUBSAMPLE(src_y_width, 1, 1);
+  const int src_uv_height = SUBSAMPLE(src_y_height, 1, 1);
+  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
+             dst_y, dst_stride_y, dst_y_width, dst_y_height,
+             kFilterBilinear);
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
+             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
+             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  return 0;
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 422 chroma is 1/2 width, 1x height
+LIBYUV_API
+int I420ToI422(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = (Abs(width) + 1) >> 1;
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 444 chroma is 1x width, 1x height
+LIBYUV_API
+int I420ToI444(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = Abs(width);
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 411 chroma is 1/4 width, 1x height
+LIBYUV_API
+int I420ToI411(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = (Abs(width) + 3) >> 2;
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// Copy to I400. Source can be I420,422,444,400,NV12,NV21
+LIBYUV_API
+int I400Copy(const uint8* src_y, int src_stride_y,
+             uint8* dst_y, int dst_stride_y,
+             int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_yuy2 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_yuy2 = 0;
+  }
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_yuy2 += dst_stride_yuy2;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+    I422ToYUY2Row(src_y + src_stride_y, src_u, src_v,
+                  dst_yuy2 + dst_stride_yuy2, width);
+    src_y += src_stride_y * 2;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_yuy2 += dst_stride_yuy2 * 2;
+  }
+  if (height & 1) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_uyvy == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_uyvy = 0;
+  }
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uyvy += dst_stride_uyvy;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+    I422ToUYVYRow(src_y + src_stride_y, src_u, src_v,
+                  dst_uyvy + dst_stride_uyvy, width);
+    src_y += src_stride_y * 2;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uyvy += dst_stride_uyvy * 2;
+  }
+  if (height & 1) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToNV12(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    int halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_uv = dst_uv + (halfheight - 1) * dst_stride_uv;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_uv = -dst_stride_uv;
+  }
+  // Coalesce rows.
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_u == halfwidth &&
+      src_stride_v == halfwidth &&
+      dst_stride_uv == halfwidth * 2) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_u = src_stride_v = dst_stride_uv = 0;
+  }
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_u, 16) && IS_ALIGNED(src_stride_u, 16) &&
+          IS_ALIGNED(src_v, 16) && IS_ALIGNED(src_stride_v, 16) &&
+          IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
+        MergeUVRow_ = MergeUVRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  for (int y = 0; y < halfheight; ++y) {
+    // Merge a row of U and V into a row of UV.
+    MergeUVRow_(src_u, src_v, dst_uv, halfwidth);
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uv += dst_stride_uv;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToNV21(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height) {
+  return I420ToNV12(src_y, src_stride_y,
+                    src_v, src_stride_v,
+                    src_u, src_stride_u,
+                    dst_y, src_stride_y,
+                    dst_vu, dst_stride_vu,
+                    width, height);
+}
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I422ToARGBRow = I422ToARGBRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to BGRA.
+LIBYUV_API
+int I420ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+#if defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_bgra, 16) && IS_ALIGNED(dst_stride_bgra, 16)) {
+        I422ToBGRARow = I422ToBGRARow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ABGR.
+LIBYUV_API
+int I420ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+#if defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_abgr, 16) && IS_ALIGNED(dst_stride_abgr, 16)) {
+        I422ToABGRRow = I422ToABGRRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGBA.
+LIBYUV_API
+int I420ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_rgba, 16) && IS_ALIGNED(dst_stride_rgba, 16)) {
+        I422ToRGBARow = I422ToRGBARow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB24.
+LIBYUV_API
+int I420ToRGB24(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_rgb24 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+  void (*I422ToRGB24Row)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGB24Row_C;
+#if defined(HAS_I422TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RAW.
+LIBYUV_API
+int I420ToRAW(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_raw, int dst_stride_raw,
+                int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_raw ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_raw = dst_raw + (height - 1) * dst_stride_raw;
+    dst_stride_raw = -dst_stride_raw;
+  }
+  void (*I422ToRAWRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRAWRow_C;
+#if defined(HAS_I422TORAWROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRAWRow = I422ToRAWRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRAWRow = I422ToRAWRow_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRAWRow = I422ToRAWRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRAWRow = I422ToRAWRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToRAWRow(src_y, src_u, src_v, dst_raw, width);
+    dst_raw += dst_stride_raw;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ARGB1555.
+LIBYUV_API
+int I420ToARGB1555(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_argb1555 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555;
+    dst_stride_argb1555 = -dst_stride_argb1555;
+  }
+  void (*I422ToARGB1555Row)(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* rgb_buf,
+                          int width) = I422ToARGB1555Row_C;
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, width);
+    dst_argb1555 += dst_stride_argb1555;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+
+// Convert I420 to ARGB4444.
+LIBYUV_API
+int I420ToARGB4444(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_argb4444 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444;
+    dst_stride_argb4444 = -dst_stride_argb4444;
+  }
+  void (*I422ToARGB4444Row)(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* rgb_buf,
+                          int width) = I422ToARGB4444Row_C;
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, width);
+    dst_argb4444 += dst_stride_argb4444;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB565.
+LIBYUV_API
+int I420ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_u, int src_stride_u,
+                 const uint8* src_v, int src_stride_v,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  if (!src_y || !src_u || !src_v || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+  void (*I422ToRGB565Row)(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* rgb_buf,
+                          int width) = I422ToRGB565Row_C;
+#if defined(HAS_I422TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
+    }
+  }
+#elif defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to specified format
+LIBYUV_API
+int ConvertFromI420(const uint8* y, int y_stride,
+                    const uint8* u, int u_stride,
+                    const uint8* v, int v_stride,
+                    uint8* dst_sample, int dst_sample_stride,
+                    int width, int height,
+                    uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  if (!y || !u|| !v || !dst_sample ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  int r = 0;
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      r = I420ToYUY2(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2,
+                     width, height);
+      break;
+    case FOURCC_UYVY:
+      r = I420ToUYVY(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2,
+                     width, height);
+      break;
+    case FOURCC_RGBP:
+      r = I420ToRGB565(y, y_stride,
+                       u, u_stride,
+                       v, v_stride,
+                       dst_sample,
+                       dst_sample_stride ? dst_sample_stride : width * 2,
+                       width, height);
+      break;
+    case FOURCC_RGBO:
+      r = I420ToARGB1555(y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         dst_sample,
+                         dst_sample_stride ? dst_sample_stride : width * 2,
+                         width, height);
+      break;
+    case FOURCC_R444:
+      r = I420ToARGB4444(y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         dst_sample,
+                         dst_sample_stride ? dst_sample_stride : width * 2,
+                         width, height);
+      break;
+    case FOURCC_24BG:
+      r = I420ToRGB24(y, y_stride,
+                      u, u_stride,
+                      v, v_stride,
+                      dst_sample,
+                      dst_sample_stride ? dst_sample_stride : width * 3,
+                      width, height);
+      break;
+    case FOURCC_RAW:
+      r = I420ToRAW(y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    dst_sample,
+                    dst_sample_stride ? dst_sample_stride : width * 3,
+                    width, height);
+      break;
+    case FOURCC_ARGB:
+      r = I420ToARGB(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_BGRA:
+      r = I420ToBGRA(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_ABGR:
+      r = I420ToABGR(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_RGBA:
+      r = I420ToRGBA(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_BGGR:
+      r = I420ToBayerBGGR(y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          dst_sample,
+                          dst_sample_stride ? dst_sample_stride : width,
+                          width, height);
+      break;
+    case FOURCC_GBRG:
+      r = I420ToBayerGBRG(y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          dst_sample,
+                          dst_sample_stride ? dst_sample_stride : width,
+                          width, height);
+      break;
+    case FOURCC_GRBG:
+      r = I420ToBayerGRBG(y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          dst_sample,
+                          dst_sample_stride ? dst_sample_stride : width,
+                          width, height);
+      break;
+    case FOURCC_RGGB:
+      r = I420ToBayerRGGB(y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          dst_sample,
+                          dst_sample_stride ? dst_sample_stride : width,
+                          width, height);
+      break;
+    case FOURCC_I400:
+      r = I400Copy(y, y_stride,
+                   dst_sample,
+                   dst_sample_stride ? dst_sample_stride : width,
+                   width, height);
+      break;
+    case FOURCC_NV12: {
+      uint8* dst_uv = dst_sample + width * height;
+      r = I420ToNV12(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     dst_uv,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     width, height);
+      break;
+    }
+    case FOURCC_NV21: {
+      uint8* dst_vu = dst_sample + width * height;
+      r = I420ToNV21(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     dst_vu,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     width, height);
+      break;
+    }
+    // TODO(fbarchard): Add M420 and Q420.
+    // Triplanar formats
+    // TODO(fbarchard): halfstride instead of halfwidth
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      int halfwidth = (width + 1) / 2;
+      int halfheight = (height + 1) / 2;
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV12) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + halfwidth * halfheight;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + halfwidth * halfheight;
+      }
+      r = I420Copy(y, y_stride,
+                   u, u_stride,
+                   v, v_stride,
+                   dst_sample, width,
+                   dst_u, halfwidth,
+                   dst_v, halfwidth,
+                   width, height);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      int halfwidth = (width + 1) / 2;
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV16) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + halfwidth * height;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + halfwidth * height;
+      }
+      r = I420ToI422(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, halfwidth,
+                     dst_v, halfwidth,
+                     width, height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV24) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + width * height;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + width * height;
+      }
+      r = I420ToI444(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, width,
+                     dst_v, width,
+                     width, height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (width + 3) / 4;
+      uint8* dst_u = dst_sample + width * height;
+      uint8* dst_v = dst_u + quarterwidth * height;
+      r = I420ToI411(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, quarterwidth,
+                     dst_v, quarterwidth,
+                     width, height);
+      break;
+    }
+
+    // Formats not supported - MJPG, biplanar, some rgb formats.
+    default:
+      return -1;  // unknown fourcc - return failure code.
+  }
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/convert_from_argb.cc b/drivers/theoraplayer/src/YUV/libyuv/src/convert_from_argb.cc
new file mode 100755
index 0000000000..41421fb30b
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/convert_from_argb.cc
@@ -0,0 +1,1096 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_from_argb.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGB little endian (bgra in memory) to I444
+LIBYUV_API
+int ARGBToI444(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u == width &&
+      dst_stride_v == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*ARGBToUV444Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) = ARGBToUV444Row_C;
+#if defined(HAS_ARGBTOUV444ROW_SSSE3)
+    if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+      ARGBToUV444Row = ARGBToUV444Row_Any_SSSE3;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV444Row = ARGBToUV444Row_Unaligned_SSSE3;
+        if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+          ARGBToUV444Row = ARGBToUV444Row_SSSE3;
+        }
+      }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    ARGBToUV444Row = ARGBToUV444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+      ARGBToUV444Row = ARGBToUV444Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToUV444Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to I422
+LIBYUV_API
+int ARGBToI422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) = ARGBToUV422Row_C;
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+      }
+    }
+  }
+#endif
+
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_NEON;
+      }
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToUV422Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to I411
+LIBYUV_API
+int ARGBToI411(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 4 == width &&
+      dst_stride_v * 4 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+  void (*ARGBToUV411Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) = ARGBToUV411Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 32) {
+      ARGBToUV411Row = ARGBToUV411Row_Any_NEON;
+      if (IS_ALIGNED(width, 32)) {
+        ARGBToUV411Row = ARGBToUV411Row_NEON;
+      }
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToUV411Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  if (!src_argb ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUVRow = ARGBToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          ARGBToYRow = ARGBToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVRow = ARGBToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVRow = ARGBToUVRow_NEON;
+      }
+    }
+  }
+#endif
+  int halfwidth = (width + 1) >> 1;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
+        MergeUVRow_ = MergeUVRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+
+  // Allocate a rows of uv.
+  align_buffer_64(row_u, ((halfwidth + 15) & ~15) * 2);
+  uint8* row_v = row_u + ((halfwidth + 15) & ~15);
+
+  for (int y = 0; y < height - 1; y += 2) {
+    ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+    MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+    ARGBToYRow(src_argb, dst_y, width);
+    ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+    src_argb += src_stride_argb * 2;
+    dst_y += dst_stride_y * 2;
+    dst_uv += dst_stride_uv;
+  }
+  if (height & 1) {
+    ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+    MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+    ARGBToYRow(src_argb, dst_y, width);
+  }
+  free_aligned_buffer_64(row_u);
+  return 0;
+}
+
+// Same as NV12 but U and V swapped.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  if (!src_argb ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_Unaligned_SSSE3;
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUVRow = ARGBToUVRow_SSSE3;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          ARGBToYRow = ARGBToYRow_SSSE3;
+        }
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVRow = ARGBToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVRow = ARGBToUVRow_NEON;
+      }
+    }
+  }
+#endif
+  int halfwidth = (width + 1) >> 1;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_uv, 16) && IS_ALIGNED(dst_stride_uv, 16)) {
+        MergeUVRow_ = MergeUVRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && halfwidth >= 32) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && halfwidth >= 16) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+
+  // Allocate a rows of uv.
+  align_buffer_64(row_u, ((halfwidth + 15) & ~15) * 2);
+  uint8* row_v = row_u + ((halfwidth + 15) & ~15);
+
+  for (int y = 0; y < height - 1; y += 2) {
+    ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+    MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
+    ARGBToYRow(src_argb, dst_y, width);
+    ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+    src_argb += src_stride_argb * 2;
+    dst_y += dst_stride_y * 2;
+    dst_uv += dst_stride_uv;
+  }
+  if (height & 1) {
+    ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+    MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
+    ARGBToYRow(src_argb, dst_y, width);
+  }
+  free_aligned_buffer_64(row_u);
+  return 0;
+}
+
+// Convert ARGB to YUY2.
+LIBYUV_API
+int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  if (!src_argb || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_yuy2 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_yuy2 = 0;
+  }
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) = ARGBToUV422Row_C;
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+      }
+    }
+  }
+#endif
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_NEON;
+      }
+    }
+  }
+#endif
+
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToUV422Row(src_argb, row_u, row_v, width);
+    ARGBToYRow(src_argb, row_y, width);
+    I422ToYUY2Row(row_y, row_u, row_v, dst_yuy2, width);
+    src_argb += src_stride_argb;
+    dst_yuy2 += dst_stride_yuy2;
+  }
+
+  free_aligned_buffer_64(row_y);
+  return 0;
+}
+
+// Convert ARGB to UYVY.
+LIBYUV_API
+int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  if (!src_argb || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_uyvy == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_uyvy = 0;
+  }
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) = ARGBToUV422Row_C;
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+      }
+    }
+  }
+#endif
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV422Row = ARGBToUV422Row_NEON;
+      }
+    }
+  }
+#endif
+
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#elif defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToUV422Row(src_argb, row_u, row_v, width);
+    ARGBToYRow(src_argb, row_y, width);
+    I422ToUYVYRow(row_y, row_u, row_v, dst_uyvy, width);
+    src_argb += src_stride_argb;
+    dst_uyvy += dst_stride_uyvy;
+  }
+
+  free_aligned_buffer_64(row_y);
+  return 0;
+}
+
+// Convert ARGB to I400.
+LIBYUV_API
+int ARGBToI400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_argb || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = 0;
+  }
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+  }
+  return 0;
+}
+
+// Shuffle table for converting ARGB to RGBA.
+static uvec8 kShuffleMaskARGBToRGBA = {
+  3u, 0u, 1u, 2u, 7u, 4u, 5u, 6u, 11u, 8u, 9u, 10u, 15u, 12u, 13u, 14u
+};
+
+// Convert ARGB to RGBA.
+LIBYUV_API
+int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  return ARGBShuffle(src_argb, src_stride_argb,
+                     dst_rgba, dst_stride_rgba,
+                     (const uint8*)(&kShuffleMaskARGBToRGBA),
+                     width, height);
+}
+
+// Convert ARGB To RGB24.
+LIBYUV_API
+int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height) {
+  if (!src_argb || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_rgb24 == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_rgb24 = 0;
+  }
+  void (*ARGBToRGB24Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRGB24Row_C;
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_SSSE3;
+    }
+  }
+#elif defined(HAS_ARGBTORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToRGB24Row(src_argb, dst_rgb24, width);
+    src_argb += src_stride_argb;
+    dst_rgb24 += dst_stride_rgb24;
+  }
+  return 0;
+}
+
+// Convert ARGB To RAW.
+LIBYUV_API
+int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_raw, int dst_stride_raw,
+              int width, int height) {
+  if (!src_argb || !dst_raw || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_raw == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_raw = 0;
+  }
+  void (*ARGBToRAWRow)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRAWRow_C;
+#if defined(HAS_ARGBTORAWROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRAWRow = ARGBToRAWRow_SSSE3;
+    }
+  }
+#elif defined(HAS_ARGBTORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRAWRow = ARGBToRAWRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToRAWRow(src_argb, dst_raw, width);
+    src_argb += src_stride_argb;
+    dst_raw += dst_stride_raw;
+  }
+  return 0;
+}
+
+// Convert ARGB To RGB565.
+LIBYUV_API
+int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_rgb565 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_rgb565 = 0;
+  }
+  void (*ARGBToRGB565Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRGB565Row_C;
+#if defined(HAS_ARGBTORGB565ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_SSE2;
+    }
+  }
+#elif defined(HAS_ARGBTORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToRGB565Row(src_argb, dst_rgb565, width);
+    src_argb += src_stride_argb;
+    dst_rgb565 += dst_stride_rgb565;
+  }
+  return 0;
+}
+
+// Convert ARGB To ARGB1555.
+LIBYUV_API
+int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height) {
+  if (!src_argb || !dst_argb1555 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb1555 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb1555 = 0;
+  }
+  void (*ARGBToARGB1555Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToARGB1555Row_C;
+#if defined(HAS_ARGBTOARGB1555ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_SSE2;
+    }
+  }
+#elif defined(HAS_ARGBTOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToARGB1555Row(src_argb, dst_argb1555, width);
+    src_argb += src_stride_argb;
+    dst_argb1555 += dst_stride_argb1555;
+  }
+  return 0;
+}
+
+// Convert ARGB To ARGB4444.
+LIBYUV_API
+int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height) {
+  if (!src_argb || !dst_argb4444 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb4444 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb4444 = 0;
+  }
+  void (*ARGBToARGB4444Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToARGB4444Row_C;
+#if defined(HAS_ARGBTOARGB4444ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_SSE2;
+    }
+  }
+#elif defined(HAS_ARGBTOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToARGB4444Row(src_argb, dst_argb4444, width);
+    src_argb += src_stride_argb;
+    dst_argb4444 += dst_stride_argb4444;
+  }
+  return 0;
+}
+
+// Convert ARGB to J420. (JPeg full range I420).
+LIBYUV_API
+int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  if (!src_argb ||
+      !dst_yj || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  void (*ARGBToUVJRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) =
+      ARGBToYJRow_C;
+#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_Unaligned_SSSE3;
+      ARGBToYJRow = ARGBToYJRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+        ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+        if (IS_ALIGNED(dst_yj, 16) && IS_ALIGNED(dst_stride_yj, 16)) {
+          ARGBToYJRow = ARGBToYJRow_SSSE3;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2) && defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVJRow = ARGBToUVJRow_NEON;
+      }
+    }
+  }
+#endif
+
+  for (int y = 0; y < height - 1; y += 2) {
+    ARGBToUVJRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_yj, width);
+    ARGBToYJRow(src_argb + src_stride_argb, dst_yj + dst_stride_yj, width);
+    src_argb += src_stride_argb * 2;
+    dst_yj += dst_stride_yj * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ARGBToUVJRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_yj, width);
+  }
+  return 0;
+}
+
+// Convert ARGB to J400.
+LIBYUV_API
+int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               int width, int height) {
+  if (!src_argb || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_yj = 0;
+  }
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) =
+      ARGBToYJRow_C;
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+          IS_ALIGNED(dst_yj, 16) && IS_ALIGNED(dst_stride_yj, 16)) {
+        ARGBToYJRow = ARGBToYJRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+  for (int y = 0; y < height; ++y) {
+    ARGBToYJRow(src_argb, dst_yj, width);
+    src_argb += src_stride_argb;
+    dst_yj += dst_stride_yj;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/convert_jpeg.cc b/drivers/theoraplayer/src/YUV/libyuv/src/convert_jpeg.cc
new file mode 100755
index 0000000000..bcb980f7f1
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/convert_jpeg.cc
@@ -0,0 +1,392 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#ifdef HAVE_JPEG
+struct I420Buffers {
+  uint8* y;
+  int y_stride;
+  uint8* u;
+  int u_stride;
+  uint8* v;
+  int v_stride;
+  int w;
+  int h;
+};
+
+static void JpegCopyI420(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I420Copy(data[0], strides[0],
+           data[1], strides[1],
+           data[2], strides[2],
+           dest->y, dest->y_stride,
+           dest->u, dest->u_stride,
+           dest->v, dest->v_stride,
+           dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I422ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I444ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI411ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I411ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I400ToI420(data[0], strides[0],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size,
+             int* width, int* height) {
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret) {
+    *width = mjpeg_decoder.GetWidth();
+    *height = mjpeg_decoder.GetHeight();
+  }
+  mjpeg_decoder.UnloadFrame();
+  return ret ? 0 : -1;  // -1 for runtime failure.
+}
+
+// MJPG (Motion JPeg) to I420
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToI420(const uint8* sample,
+               size_t sample_size,
+               uint8* y, int y_stride,
+               uint8* u, int u_stride,
+               uint8* v, int v_stride,
+               int w, int h,
+               int dw, int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != w ||
+              mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    I420Buffers bufs = { y, y_stride, u, u_stride, v, v_stride, dw, dh };
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() ==
+            MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegCopyI420, &bufs, dw, dh);
+    // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToI420, &bufs, dw, dh);
+    // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToI420, &bufs, dw, dh);
+    // YUV411
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToI420, &bufs, dw, dh);
+    // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToI420, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice. 411 is supported by libjpeg
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+
+#ifdef HAVE_JPEG
+struct ARGBBuffers {
+  uint8* argb;
+  int argb_stride;
+  int w;
+  int h;
+};
+
+static void JpegI420ToARGB(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I420ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I422ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I444ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI411ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I411ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I400ToARGB(data[0], strides[0],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+// MJPG (Motion JPeg) to ARGB
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample,
+               size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h,
+               int dw, int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != w ||
+              mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    ARGBBuffers bufs = { argb, argb_stride, dw, dh };
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() ==
+            MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToARGB, &bufs, dw, dh);
+    // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToARGB, &bufs, dw, dh);
+    // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToARGB, &bufs, dw, dh);
+    // YUV411
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToARGB, &bufs, dw, dh);
+    // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToARGB, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice. 411 is supported by libjpeg
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+#endif
+
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/convert_to_argb.cc b/drivers/theoraplayer/src/YUV/libyuv/src/convert_to_argb.cc
new file mode 100755
index 0000000000..1b228a7b4d
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/convert_to_argb.cc
@@ -0,0 +1,327 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/format_conversion.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToARGB(const uint8* sample, size_t sample_size,
+                  uint8* crop_argb, int argb_stride,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination crop_argb is same as source sample,
+  // also enable temporary buffer.
+  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_ARGB) ||
+      crop_argb == sample;
+  uint8* tmp_argb = crop_argb;
+  int tmp_argb_stride = argb_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (crop_argb == NULL || sample == NULL ||
+      src_width <= 0 || crop_width <= 0 ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  if (need_buf) {
+    int argb_size = crop_width * abs_crop_height * 4;
+    rotate_buffer = (uint8*)malloc(argb_size);
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    crop_argb = rotate_buffer;
+    argb_stride = crop_width;
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToARGB(src, aligned_src_width * 2,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToARGB(src, aligned_src_width * 2,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToARGB(src, src_width * 3,
+                      crop_argb, argb_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToARGB(src, src_width * 3,
+                    crop_argb, argb_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToARGB(src, src_width * 2,
+                       crop_argb, argb_stride,
+                       crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToARGB(src, src_width * 2,
+                         crop_argb, argb_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToARGB(src, src_width * 2,
+                         crop_argb, argb_stride,
+                         crop_width, inv_crop_height);
+      break;
+    // TODO(fbarchard): Support cropping Bayer by odd numbers
+    // by adjusting fourcc.
+    case FOURCC_BGGR:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerBGGRToARGB(src, src_width,
+                          crop_argb, argb_stride,
+                          crop_width, inv_crop_height);
+      break;
+
+    case FOURCC_GBRG:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerGBRGToARGB(src, src_width,
+                          crop_argb, argb_stride,
+                          crop_width, inv_crop_height);
+      break;
+
+    case FOURCC_GRBG:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerGRBGToARGB(src, src_width,
+                          crop_argb, argb_stride,
+                          crop_width, inv_crop_height);
+      break;
+
+    case FOURCC_RGGB:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerRGGBToARGB(src, src_width,
+                          crop_argb, argb_stride,
+                          crop_width, inv_crop_height);
+      break;
+
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToARGB(src, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      r = NV12ToARGB(src, src_width,
+                     src_uv, aligned_src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      // Call NV12 but with u and v parameters swapped.
+      r = NV21ToARGB(src, src_width,
+                     src_uv, aligned_src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToARGB(src, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+//    case FOURCC_Q420:
+//      src = sample + (src_width + aligned_src_width * 2) * crop_y + crop_x;
+//      src_uv = sample + (src_width + aligned_src_width * 2) * crop_y +
+//               src_width + crop_x * 2;
+//      r = Q420ToARGB(src, src_width * 3,
+//                    src_uv, src_width * 3,
+//                    crop_argb, argb_stride,
+//                    crop_width, inv_crop_height);
+//      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToARGB(src_y, src_width,
+                     src_u, src_width,
+                     src_v, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (src_width + 3) / 4;
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u = sample + src_width * abs_src_height +
+          quarterwidth * crop_y + crop_x / 4;
+      const uint8* src_v = sample + src_width * abs_src_height +
+          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
+      r = I411ToARGB(src_y, src_width,
+                     src_u, quarterwidth,
+                     src_v, quarterwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToARGB(sample, sample_size,
+                     crop_argb, argb_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = ARGBRotate(crop_argb, argb_stride,
+                     tmp_argb, tmp_argb_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/convert_to_i420.cc b/drivers/theoraplayer/src/YUV/libyuv/src/convert_to_i420.cc
new file mode 100755
index 0000000000..7b194fff72
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/convert_to_i420.cc
@@ -0,0 +1,383 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "libyuv/convert.h"
+
+#include "libyuv/format_conversion.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToI420(const uint8* sample,
+                  size_t sample_size,
+                  uint8* y, int y_stride,
+                  uint8* u, int u_stride,
+                  uint8* v, int v_stride,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_I420 &&
+      format != FOURCC_NV12 && format != FOURCC_NV21 &&
+      format != FOURCC_YU12 && format != FOURCC_YV12) || y == sample;
+  uint8* tmp_y = y;
+  uint8* tmp_u = u;
+  uint8* tmp_v = v;
+  int tmp_y_stride = y_stride;
+  int tmp_u_stride = u_stride;
+  int tmp_v_stride = v_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (!y || !u || !v || !sample ||
+      src_width <= 0 || crop_width <= 0  ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination y is same as source sample,
+  // also enable temporary buffer.
+  if (need_buf) {
+    int y_size = crop_width * abs_crop_height;
+    int uv_size = ((crop_width + 1) / 2) * ((abs_crop_height + 1) / 2);
+    rotate_buffer = (uint8*)malloc(y_size + uv_size * 2);
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    y = rotate_buffer;
+    u = y + y_size;
+    v = u + uv_size;
+    y_stride = crop_width;
+    u_stride = v_stride = ((crop_width + 1) / 2);
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToI420(src, aligned_src_width * 2,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToI420(src, aligned_src_width * 2,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToI420(src, src_width * 2,
+                       y, y_stride,
+                       u, u_stride,
+                       v, v_stride,
+                       crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToI420(src, src_width * 2,
+                         y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToI420(src, src_width * 2,
+                         y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToI420(src, src_width * 3,
+                      y, y_stride,
+                      u, u_stride,
+                      v, v_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToI420(src, src_width * 3,
+                    y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // TODO(fbarchard): Support cropping Bayer by odd numbers
+    // by adjusting fourcc.
+    case FOURCC_BGGR:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerBGGRToI420(src, src_width,
+                          y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          crop_width, inv_crop_height);
+      break;
+    case FOURCC_GBRG:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerGBRGToI420(src, src_width,
+                          y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          crop_width, inv_crop_height);
+      break;
+    case FOURCC_GRBG:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerGRBGToI420(src, src_width,
+                          y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGGB:
+      src = sample + (src_width * crop_y + crop_x);
+      r = BayerRGGBToI420(src, src_width,
+                          y, y_stride,
+                          u, u_stride,
+                          v, v_stride,
+                          crop_width, inv_crop_height);
+      break;
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToI420(src, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      r = NV12ToI420Rotate(src, src_width,
+                           src_uv, aligned_src_width,
+                           y, y_stride,
+                           u, u_stride,
+                           v, v_stride,
+                           crop_width, inv_crop_height, rotation);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      // Call NV12 but with u and v parameters swapped.
+      r = NV12ToI420Rotate(src, src_width,
+                           src_uv, aligned_src_width,
+                           y, y_stride,
+                           v, v_stride,
+                           u, u_stride,
+                           crop_width, inv_crop_height, rotation);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToI420(src, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_Q420:
+      src = sample + (src_width + aligned_src_width * 2) * crop_y + crop_x;
+      src_uv = sample + (src_width + aligned_src_width * 2) * crop_y +
+               src_width + crop_x * 2;
+      r = Q420ToI420(src, src_width * 3,
+                    src_uv, src_width * 3,
+                    y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    crop_width, inv_crop_height);
+      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420Rotate(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height, rotation);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToI420(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToI420(src_y, src_width,
+                     src_u, src_width,
+                     src_v, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (src_width + 3) / 4;
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u = sample + src_width * abs_src_height +
+          quarterwidth * crop_y + crop_x / 4;
+      const uint8* src_v = sample + src_width * abs_src_height +
+          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
+      r = I411ToI420(src_y, src_width,
+                     src_u, quarterwidth,
+                     src_v, quarterwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToI420(sample, sample_size,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = I420Rotate(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     tmp_y, tmp_y_stride,
+                     tmp_u, tmp_u_stride,
+                     tmp_v, tmp_v_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/cpu_id.cc b/drivers/theoraplayer/src/YUV/libyuv/src/cpu_id.cc
new file mode 100755
index 0000000000..f52bd95551
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/cpu_id.cc
@@ -0,0 +1,300 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/cpu_id.h"
+
+#ifdef _ANDROID //libtheoraplayer addition for cpu feature detection
+#include "cpu-features.h"
+#endif
+
+#ifdef _MSC_VER
+#include <intrin.h>  // For __cpuidex()
+#endif
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    !defined(__native_client__) && defined(_M_X64) && \
+    defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+#include <immintrin.h>  // For _xgetbv()
+#endif
+
+#if !defined(__native_client__)
+#include <stdlib.h>  // For getenv()
+#endif
+
+// For ArmCpuCaps() but unittested on all platforms
+#include <stdio.h>
+#include <string.h>
+
+#include "libyuv/basic_types.h"  // For CPU_X86
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// For functions that use the stack and have runtime checks for overflow,
+// use SAFEBUFFERS to avoid additional check.
+#if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+#define SAFEBUFFERS __declspec(safebuffers)
+#else
+#define SAFEBUFFERS
+#endif
+
+// Low level cpuid for X86. Returns zeros on other CPUs.
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__i386__) || defined(__x86_64__))
+LIBYUV_API
+void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
+#if defined(_MSC_VER)
+#if (_MSC_FULL_VER >= 160040219)
+  __cpuidex((int*)(cpu_info), info_eax, info_ecx);
+#elif defined(_M_IX86)
+  __asm {
+    mov        eax, info_eax
+    mov        ecx, info_ecx
+    mov        edi, cpu_info
+    cpuid
+    mov        [edi], eax
+    mov        [edi + 4], ebx
+    mov        [edi + 8], ecx
+    mov        [edi + 12], edx
+  }
+#else
+  if (info_ecx == 0) {
+    __cpuid((int*)(cpu_info), info_eax);
+  } else {
+    cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0;
+  }
+#endif
+#else  // defined(_MSC_VER)
+  uint32 info_ebx, info_edx;
+  asm volatile (  // NOLINT
+#if defined( __i386__) && defined(__PIC__)
+    // Preserve ebx for fpic 32 bit.
+    "mov %%ebx, %%edi                          \n"
+    "cpuid                                     \n"
+    "xchg %%edi, %%ebx                         \n"
+    : "=D" (info_ebx),
+#else
+    "cpuid                                     \n"
+    : "=b" (info_ebx),
+#endif  //  defined( __i386__) && defined(__PIC__)
+      "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx));
+  cpu_info[0] = info_eax;
+  cpu_info[1] = info_ebx;
+  cpu_info[2] = info_ecx;
+  cpu_info[3] = info_edx;
+#endif  // defined(_MSC_VER)
+}
+
+#if !defined(__native_client__)
+#define HAS_XGETBV
+// X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
+int TestOsSaveYmm() {
+  uint32 xcr0 = 0u;
+#if defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
+  xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
+#elif defined(_M_IX86)
+  __asm {
+    xor        ecx, ecx    // xcr 0
+    _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
+    mov        xcr0, eax
+  }
+#elif defined(__i386__) || defined(__x86_64__)
+  asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
+#endif  // defined(_MSC_VER)
+  return((xcr0 & 6) == 6);  // Is ymm saved?
+}
+#endif  // !defined(__native_client__)
+#else
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
+  cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
+}
+#endif
+
+// based on libvpx arm_cpudetect.c
+// For Arm, but public to allow testing on any CPU
+LIBYUV_API SAFEBUFFERS
+int ArmCpuCaps(const char* cpuinfo_name) {
+  char cpuinfo_line[512];
+  FILE* f = fopen(cpuinfo_name, "r");
+  if (!f) {
+    // Assume Neon if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasNEON;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
+    if (memcmp(cpuinfo_line, "Features", 8) == 0) {
+      char* p = strstr(cpuinfo_line, " neon");
+      if (p && (p[5] == ' ' || p[5] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+    }
+  }
+  fclose(f);
+  return 0;
+}
+
+#if defined(__mips__) && defined(__linux__)
+static int MipsCpuCaps(const char* search_string) {
+  char cpuinfo_line[512];
+  const char* file_name = "/proc/cpuinfo";
+  FILE* f = fopen(file_name, "r");
+  if (!f) {
+    // Assume DSP if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasMIPS_DSP;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) {
+    if (strstr(cpuinfo_line, search_string) != NULL) {
+      fclose(f);
+      return kCpuHasMIPS_DSP;
+    }
+  }
+  fclose(f);
+  return 0;
+}
+#endif
+
+// CPU detect function for SIMD instruction sets.
+LIBYUV_API
+int cpu_info_ = kCpuInit;  // cpu_info is not initialized yet.
+
+// Test environment variable for disabling CPU features. Any non-zero value
+// to disable. Zero ignored to make it easy to set the variable on/off.
+#if !defined(__native_client__) && !defined(_M_ARM)
+
+static LIBYUV_BOOL TestEnv(const char* name) {
+#ifndef _WINRT
+  const char* var = getenv(name);
+  if (var) {
+    if (var[0] != '0') {
+      return LIBYUV_TRUE;
+    }
+  }
+#endif
+  return LIBYUV_FALSE;
+}
+#else  // nacl does not support getenv().
+static LIBYUV_BOOL TestEnv(const char*) {
+  return LIBYUV_FALSE;
+}
+#endif
+
+LIBYUV_API SAFEBUFFERS
+int InitCpuFlags(void) {
+#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
+
+  uint32 cpu_info1[4] = { 0, 0, 0, 0 };
+  uint32 cpu_info7[4] = { 0, 0, 0, 0 };
+  CpuId(1, 0, cpu_info1);
+  CpuId(7, 0, cpu_info7);
+  cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
+              ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
+              ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
+              ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
+              ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
+              ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
+              kCpuHasX86;
+#ifdef HAS_XGETBV
+  if ((cpu_info1[2] & 0x18000000) == 0x18000000 &&  // AVX and OSSave
+      TestOsSaveYmm()) {  // Saves YMM.
+    cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
+                 kCpuHasAVX;
+  }
+#endif
+  // Environment variable overrides for testing.
+  if (TestEnv("LIBYUV_DISABLE_X86")) {
+    cpu_info_ &= ~kCpuHasX86;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE2")) {
+    cpu_info_ &= ~kCpuHasSSE2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
+    cpu_info_ &= ~kCpuHasSSSE3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE41")) {
+    cpu_info_ &= ~kCpuHasSSE41;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE42")) {
+    cpu_info_ &= ~kCpuHasSSE42;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX")) {
+    cpu_info_ &= ~kCpuHasAVX;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX2")) {
+    cpu_info_ &= ~kCpuHasAVX2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_ERMS")) {
+    cpu_info_ &= ~kCpuHasERMS;
+  }
+  if (TestEnv("LIBYUV_DISABLE_FMA3")) {
+    cpu_info_ &= ~kCpuHasFMA3;
+  }
+#elif defined(__mips__) && defined(__linux__)
+  // Linux mips parse text file for dsp detect.
+  cpu_info_ = MipsCpuCaps("dsp");  // set kCpuHasMIPS_DSP.
+#if defined(__mips_dspr2)
+  cpu_info_ |= kCpuHasMIPS_DSPR2;
+#endif
+  cpu_info_ |= kCpuHasMIPS;
+
+  if (getenv("LIBYUV_DISABLE_MIPS")) {
+    cpu_info_ &= ~kCpuHasMIPS;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSP")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSP;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSPR2;
+  }
+#elif defined(__arm__)
+// gcc -mfpu=neon defines __ARM_NEON__
+// __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
+// For Linux, /proc/cpuinfo can be tested but without that assume Neon.
+#if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__)
+#ifdef _ANDROID
+  cpu_info_ = ArmCpuCaps("/proc/cpuinfo"); // libtheoraplayer #ifdef addition, just in case, android gave us troubles
+#else
+  cpu_info_ = kCpuHasNEON;
+#endif
+#else
+  // Linux arm parse text file for neon detect.
+  cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
+#endif
+  cpu_info_ |= kCpuHasARM;
+  if (TestEnv("LIBYUV_DISABLE_NEON")) {
+    cpu_info_ &= ~kCpuHasNEON;
+  }
+#ifdef _ANDROID
+  // libtheoraplayer addition to disable NEON support on android devices that don't support it, once again, just in case	
+  if ((android_getCpuFeaturesExt() & ANDROID_CPU_ARM_FEATURE_NEON) == 0)
+  {
+ 	cpu_info_ = kCpuHasARM;
+  }
+#endif
+#endif  // __arm__
+  if (TestEnv("LIBYUV_DISABLE_ASM")) {
+    cpu_info_ = 0;
+  }
+  return cpu_info_;
+}
+
+LIBYUV_API
+void MaskCpuFlags(int enable_flags) {
+  cpu_info_ = InitCpuFlags() & enable_flags;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/format_conversion.cc b/drivers/theoraplayer/src/YUV/libyuv/src/format_conversion.cc
new file mode 100755
index 0000000000..a3daf96a98
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/format_conversion.cc
@@ -0,0 +1,552 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/format_conversion.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/video_common.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// generate a selector mask useful for pshufb
+static uint32 GenerateSelector(int select0, int select1) {
+  return (uint32)(select0) |
+         (uint32)((select1 + 4) << 8) |
+         (uint32)((select0 + 8) << 16) |
+         (uint32)((select1 + 12) << 24);
+}
+
+static int MakeSelectors(const int blue_index,
+                         const int green_index,
+                         const int red_index,
+                         uint32 dst_fourcc_bayer,
+                         uint32* index_map) {
+  // Now build a lookup table containing the indices for the four pixels in each
+  // 2x2 Bayer grid.
+  switch (dst_fourcc_bayer) {
+    case FOURCC_BGGR:
+      index_map[0] = GenerateSelector(blue_index, green_index);
+      index_map[1] = GenerateSelector(green_index, red_index);
+      break;
+    case FOURCC_GBRG:
+      index_map[0] = GenerateSelector(green_index, blue_index);
+      index_map[1] = GenerateSelector(red_index, green_index);
+      break;
+    case FOURCC_RGGB:
+      index_map[0] = GenerateSelector(red_index, green_index);
+      index_map[1] = GenerateSelector(green_index, blue_index);
+      break;
+    case FOURCC_GRBG:
+      index_map[0] = GenerateSelector(green_index, red_index);
+      index_map[1] = GenerateSelector(blue_index, green_index);
+      break;
+    default:
+      return -1;  // Bad FourCC
+  }
+  return 0;
+}
+
+// Converts 32 bit ARGB to Bayer RGB formats.
+LIBYUV_API
+int ARGBToBayer(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_bayer, int dst_stride_bayer,
+                int width, int height,
+                uint32 dst_fourcc_bayer) {
+  int y;
+  const int blue_index = 0;  // Offsets for ARGB format
+  const int green_index = 1;
+  const int red_index = 2;
+  uint32 index_map[2];
+  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) = ARGBToBayerRow_C;
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
+    }
+  }
+#elif defined(HAS_ARGBTOBAYERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_NEON;
+    }
+  }
+#endif
+  if (MakeSelectors(blue_index, green_index, red_index,
+                    dst_fourcc_bayer, index_map)) {
+    return -1;  // Bad FourCC
+  }
+
+  for (y = 0; y < height; ++y) {
+    ARGBToBayerRow(src_argb, dst_bayer, index_map[y & 1], width);
+    src_argb += src_stride_argb;
+    dst_bayer += dst_stride_bayer;
+  }
+  return 0;
+}
+
+#define AVG(a, b) (((a) + (b)) >> 1)
+
+static void BayerRowBG(const uint8* src_bayer0, int src_stride_bayer,
+                       uint8* dst_argb, int pix) {
+  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
+  uint8 g = src_bayer0[1];
+  uint8 r = src_bayer1[1];
+  int x;
+  for (x = 0; x < pix - 2; x += 2) {
+    dst_argb[0] = src_bayer0[0];
+    dst_argb[1] = AVG(g, src_bayer0[1]);
+    dst_argb[2] = AVG(r, src_bayer1[1]);
+    dst_argb[3] = 255U;
+    dst_argb[4] = AVG(src_bayer0[0], src_bayer0[2]);
+    dst_argb[5] = src_bayer0[1];
+    dst_argb[6] = src_bayer1[1];
+    dst_argb[7] = 255U;
+    g = src_bayer0[1];
+    r = src_bayer1[1];
+    src_bayer0 += 2;
+    src_bayer1 += 2;
+    dst_argb += 8;
+  }
+  dst_argb[0] = src_bayer0[0];
+  dst_argb[1] = AVG(g, src_bayer0[1]);
+  dst_argb[2] = AVG(r, src_bayer1[1]);
+  dst_argb[3] = 255U;
+  if (!(pix & 1)) {
+    dst_argb[4] = src_bayer0[0];
+    dst_argb[5] = src_bayer0[1];
+    dst_argb[6] = src_bayer1[1];
+    dst_argb[7] = 255U;
+  }
+}
+
+static void BayerRowRG(const uint8* src_bayer0, int src_stride_bayer,
+                       uint8* dst_argb, int pix) {
+  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
+  uint8 g = src_bayer0[1];
+  uint8 b = src_bayer1[1];
+  int x;
+  for (x = 0; x < pix - 2; x += 2) {
+    dst_argb[0] = AVG(b, src_bayer1[1]);
+    dst_argb[1] = AVG(g, src_bayer0[1]);
+    dst_argb[2] = src_bayer0[0];
+    dst_argb[3] = 255U;
+    dst_argb[4] = src_bayer1[1];
+    dst_argb[5] = src_bayer0[1];
+    dst_argb[6] = AVG(src_bayer0[0], src_bayer0[2]);
+    dst_argb[7] = 255U;
+    g = src_bayer0[1];
+    b = src_bayer1[1];
+    src_bayer0 += 2;
+    src_bayer1 += 2;
+    dst_argb += 8;
+  }
+  dst_argb[0] = AVG(b, src_bayer1[1]);
+  dst_argb[1] = AVG(g, src_bayer0[1]);
+  dst_argb[2] = src_bayer0[0];
+  dst_argb[3] = 255U;
+  if (!(pix & 1)) {
+    dst_argb[4] = src_bayer1[1];
+    dst_argb[5] = src_bayer0[1];
+    dst_argb[6] = src_bayer0[0];
+    dst_argb[7] = 255U;
+  }
+}
+
+static void BayerRowGB(const uint8* src_bayer0, int src_stride_bayer,
+                       uint8* dst_argb, int pix) {
+  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
+  uint8 b = src_bayer0[1];
+  int x;
+  for (x = 0; x < pix - 2; x += 2) {
+    dst_argb[0] = AVG(b, src_bayer0[1]);
+    dst_argb[1] = src_bayer0[0];
+    dst_argb[2] = src_bayer1[0];
+    dst_argb[3] = 255U;
+    dst_argb[4] = src_bayer0[1];
+    dst_argb[5] = AVG(src_bayer0[0], src_bayer0[2]);
+    dst_argb[6] = AVG(src_bayer1[0], src_bayer1[2]);
+    dst_argb[7] = 255U;
+    b = src_bayer0[1];
+    src_bayer0 += 2;
+    src_bayer1 += 2;
+    dst_argb += 8;
+  }
+  dst_argb[0] = AVG(b, src_bayer0[1]);
+  dst_argb[1] = src_bayer0[0];
+  dst_argb[2] = src_bayer1[0];
+  dst_argb[3] = 255U;
+  if (!(pix & 1)) {
+    dst_argb[4] = src_bayer0[1];
+    dst_argb[5] = src_bayer0[0];
+    dst_argb[6] = src_bayer1[0];
+    dst_argb[7] = 255U;
+  }
+}
+
+static void BayerRowGR(const uint8* src_bayer0, int src_stride_bayer,
+                       uint8* dst_argb, int pix) {
+  const uint8* src_bayer1 = src_bayer0 + src_stride_bayer;
+  uint8 r = src_bayer0[1];
+  int x;
+  for (x = 0; x < pix - 2; x += 2) {
+    dst_argb[0] = src_bayer1[0];
+    dst_argb[1] = src_bayer0[0];
+    dst_argb[2] = AVG(r, src_bayer0[1]);
+    dst_argb[3] = 255U;
+    dst_argb[4] = AVG(src_bayer1[0], src_bayer1[2]);
+    dst_argb[5] = AVG(src_bayer0[0], src_bayer0[2]);
+    dst_argb[6] = src_bayer0[1];
+    dst_argb[7] = 255U;
+    r = src_bayer0[1];
+    src_bayer0 += 2;
+    src_bayer1 += 2;
+    dst_argb += 8;
+  }
+  dst_argb[0] = src_bayer1[0];
+  dst_argb[1] = src_bayer0[0];
+  dst_argb[2] = AVG(r, src_bayer0[1]);
+  dst_argb[3] = 255U;
+  if (!(pix & 1)) {
+    dst_argb[4] = src_bayer1[0];
+    dst_argb[5] = src_bayer0[0];
+    dst_argb[6] = src_bayer0[1];
+    dst_argb[7] = 255U;
+  }
+}
+
+// Converts any Bayer RGB format to ARGB.
+LIBYUV_API
+int BayerToARGB(const uint8* src_bayer, int src_stride_bayer,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height,
+                uint32 src_fourcc_bayer) {
+  int y;
+  void (*BayerRow0)(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int pix);
+  void (*BayerRow1)(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int pix);
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  switch (src_fourcc_bayer) {
+    case FOURCC_BGGR:
+      BayerRow0 = BayerRowBG;
+      BayerRow1 = BayerRowGR;
+      break;
+    case FOURCC_GBRG:
+      BayerRow0 = BayerRowGB;
+      BayerRow1 = BayerRowRG;
+      break;
+    case FOURCC_GRBG:
+      BayerRow0 = BayerRowGR;
+      BayerRow1 = BayerRowBG;
+      break;
+    case FOURCC_RGGB:
+      BayerRow0 = BayerRowRG;
+      BayerRow1 = BayerRowGB;
+      break;
+    default:
+      return -1;    // Bad FourCC
+  }
+
+  for (y = 0; y < height - 1; y += 2) {
+    BayerRow0(src_bayer, src_stride_bayer, dst_argb, width);
+    BayerRow1(src_bayer + src_stride_bayer, -src_stride_bayer,
+              dst_argb + dst_stride_argb, width);
+    src_bayer += src_stride_bayer * 2;
+    dst_argb += dst_stride_argb * 2;
+  }
+  if (height & 1) {
+    BayerRow0(src_bayer, src_stride_bayer, dst_argb, width);
+  }
+  return 0;
+}
+
+// Converts any Bayer RGB format to ARGB.
+LIBYUV_API
+int BayerToI420(const uint8* src_bayer, int src_stride_bayer,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height,
+                uint32 src_fourcc_bayer) {
+  void (*BayerRow0)(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int pix);
+  void (*BayerRow1)(const uint8* src_bayer, int src_stride_bayer,
+                    uint8* dst_argb, int pix);
+
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    int halfheight;
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
+    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_Unaligned_SSSE3;
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+        ARGBToYRow = ARGBToYRow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+    if (width >= 16) {
+      ARGBToUVRow = ARGBToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUVRow = ARGBToUVRow_NEON;
+      }
+    }
+  }
+#endif
+
+  switch (src_fourcc_bayer) {
+    case FOURCC_BGGR:
+      BayerRow0 = BayerRowBG;
+      BayerRow1 = BayerRowGR;
+      break;
+    case FOURCC_GBRG:
+      BayerRow0 = BayerRowGB;
+      BayerRow1 = BayerRowRG;
+      break;
+    case FOURCC_GRBG:
+      BayerRow0 = BayerRowGR;
+      BayerRow1 = BayerRowBG;
+      break;
+    case FOURCC_RGGB:
+      BayerRow0 = BayerRowRG;
+      BayerRow1 = BayerRowGB;
+      break;
+    default:
+      return -1;  // Bad FourCC
+  }
+
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 15) & ~15;
+    align_buffer_64(row, kRowSize * 2);
+    int y;
+    for (y = 0; y < height - 1; y += 2) {
+      BayerRow0(src_bayer, src_stride_bayer, row, width);
+      BayerRow1(src_bayer + src_stride_bayer, -src_stride_bayer,
+                row + kRowSize, width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+      src_bayer += src_stride_bayer * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+      BayerRow0(src_bayer, src_stride_bayer, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+    }
+    free_aligned_buffer_64(row);
+  }
+  return 0;
+}
+
+// Convert I420 to Bayer.
+LIBYUV_API
+int I420ToBayer(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_bayer, int dst_stride_bayer,
+                int width, int height,
+                uint32 dst_fourcc_bayer) {
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) = ARGBToBayerRow_C;
+  const int blue_index = 0;  // Offsets for ARGB format
+  const int green_index = 1;
+  const int red_index = 2;
+  uint32 index_map[2];
+  // Negative height means invert the image.
+  if (height < 0) {
+    int halfheight;
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
+    }
+  }
+#elif defined(HAS_ARGBTOBAYERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_NEON;
+    }
+  }
+#endif
+
+  if (MakeSelectors(blue_index, green_index, red_index,
+                    dst_fourcc_bayer, index_map)) {
+    return -1;  // Bad FourCC
+  }
+  {
+    // Allocate a row of ARGB.
+    align_buffer_64(row, width * 4);
+    int y;
+    for (y = 0; y < height; ++y) {
+      I422ToARGBRow(src_y, src_u, src_v, row, width);
+      ARGBToBayerRow(row, dst_bayer, index_map[y & 1], width);
+      dst_bayer += dst_stride_bayer;
+      src_y += src_stride_y;
+      if (y & 1) {
+        src_u += src_stride_u;
+        src_v += src_stride_v;
+      }
+    }
+    free_aligned_buffer_64(row);
+  }
+  return 0;
+}
+
+#define MAKEBAYERFOURCC(BAYER)                                                 \
+LIBYUV_API                                                                     \
+int Bayer##BAYER##ToI420(const uint8* src_bayer, int src_stride_bayer,         \
+                         uint8* dst_y, int dst_stride_y,                       \
+                         uint8* dst_u, int dst_stride_u,                       \
+                         uint8* dst_v, int dst_stride_v,                       \
+                         int width, int height) {                              \
+  return BayerToI420(src_bayer, src_stride_bayer,                              \
+                     dst_y, dst_stride_y,                                      \
+                     dst_u, dst_stride_u,                                      \
+                     dst_v, dst_stride_v,                                      \
+                     width, height,                                            \
+                     FOURCC_##BAYER);                                          \
+}                                                                              \
+                                                                               \
+LIBYUV_API                                                                     \
+int I420ToBayer##BAYER(const uint8* src_y, int src_stride_y,                   \
+                       const uint8* src_u, int src_stride_u,                   \
+                       const uint8* src_v, int src_stride_v,                   \
+                       uint8* dst_bayer, int dst_stride_bayer,                 \
+                       int width, int height) {                                \
+  return I420ToBayer(src_y, src_stride_y,                                      \
+                     src_u, src_stride_u,                                      \
+                     src_v, src_stride_v,                                      \
+                     dst_bayer, dst_stride_bayer,                              \
+                     width, height,                                            \
+                     FOURCC_##BAYER);                                          \
+}                                                                              \
+                                                                               \
+LIBYUV_API                                                                     \
+int ARGBToBayer##BAYER(const uint8* src_argb, int src_stride_argb,             \
+                       uint8* dst_bayer, int dst_stride_bayer,                 \
+                       int width, int height) {                                \
+  return ARGBToBayer(src_argb, src_stride_argb,                                \
+                     dst_bayer, dst_stride_bayer,                              \
+                     width, height,                                            \
+                     FOURCC_##BAYER);                                          \
+}                                                                              \
+                                                                               \
+LIBYUV_API                                                                     \
+int Bayer##BAYER##ToARGB(const uint8* src_bayer, int src_stride_bayer,         \
+                         uint8* dst_argb, int dst_stride_argb,                 \
+                         int width, int height) {                              \
+  return BayerToARGB(src_bayer, src_stride_bayer,                              \
+                     dst_argb, dst_stride_argb,                                \
+                     width, height,                                            \
+                     FOURCC_##BAYER);                                          \
+}
+
+MAKEBAYERFOURCC(BGGR)
+MAKEBAYERFOURCC(GBRG)
+MAKEBAYERFOURCC(GRBG)
+MAKEBAYERFOURCC(RGGB)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/mjpeg_decoder.cc b/drivers/theoraplayer/src/YUV/libyuv/src/mjpeg_decoder.cc
new file mode 100755
index 0000000000..193b829ba9
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/mjpeg_decoder.cc
@@ -0,0 +1,558 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#ifdef HAVE_JPEG
+#include <assert.h>
+
+#if !defined(__pnacl__) && !defined(__CLR_VER) && !defined(COVERAGE_ENABLED) &&\
+    !defined(TARGET_IPHONE_SIMULATOR)
+// Must be included before jpeglib.
+#include <setjmp.h>
+#define HAVE_SETJMP
+#endif
+struct FILE;  // For jpeglib.h.
+
+// C++ build requires extern C for jpeg internals.
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <jpeglib.h>
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#include "libyuv/planar_functions.h"  // For CopyPlane().
+
+namespace libyuv {
+
+#ifdef HAVE_SETJMP
+struct SetJmpErrorMgr {
+  jpeg_error_mgr base;  // Must be at the top
+  jmp_buf setjmp_buffer;
+};
+#endif
+
+const int MJpegDecoder::kColorSpaceUnknown = JCS_UNKNOWN;
+const int MJpegDecoder::kColorSpaceGrayscale = JCS_GRAYSCALE;
+const int MJpegDecoder::kColorSpaceRgb = JCS_RGB;
+const int MJpegDecoder::kColorSpaceYCbCr = JCS_YCbCr;
+const int MJpegDecoder::kColorSpaceCMYK = JCS_CMYK;
+const int MJpegDecoder::kColorSpaceYCCK = JCS_YCCK;
+
+MJpegDecoder::MJpegDecoder()
+    : has_scanline_padding_(LIBYUV_FALSE),
+      num_outbufs_(0),
+      scanlines_(NULL),
+      scanlines_sizes_(NULL),
+      databuf_(NULL),
+      databuf_strides_(NULL) {
+  decompress_struct_ = new jpeg_decompress_struct;
+  source_mgr_ = new jpeg_source_mgr;
+#ifdef HAVE_SETJMP
+  error_mgr_ = new SetJmpErrorMgr;
+  decompress_struct_->err = jpeg_std_error(&error_mgr_->base);
+  // Override standard exit()-based error handler.
+  error_mgr_->base.error_exit = &ErrorHandler;
+#endif
+  decompress_struct_->client_data = NULL;
+  source_mgr_->init_source = &init_source;
+  source_mgr_->fill_input_buffer = &fill_input_buffer;
+  source_mgr_->skip_input_data = &skip_input_data;
+  source_mgr_->resync_to_restart = &jpeg_resync_to_restart;
+  source_mgr_->term_source = &term_source;
+  jpeg_create_decompress(decompress_struct_);
+  decompress_struct_->src = source_mgr_;
+  buf_vec_.buffers = &buf_;
+  buf_vec_.len = 1;
+}
+
+MJpegDecoder::~MJpegDecoder() {
+  jpeg_destroy_decompress(decompress_struct_);
+  delete decompress_struct_;
+  delete source_mgr_;
+#ifdef HAVE_SETJMP
+  delete error_mgr_;
+#endif
+  DestroyOutputBuffers();
+}
+
+LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) {
+  if (!ValidateJpeg(src, src_len)) {
+    return LIBYUV_FALSE;
+  }
+
+  buf_.data = src;
+  buf_.len = (int)(src_len);
+  buf_vec_.pos = 0;
+  decompress_struct_->client_data = &buf_vec_;
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_read_header, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (jpeg_read_header(decompress_struct_, TRUE) != JPEG_HEADER_OK) {
+    // ERROR: Bad MJPEG header
+    return LIBYUV_FALSE;
+  }
+  AllocOutputBuffers(GetNumComponents());
+  for (int i = 0; i < num_outbufs_; ++i) {
+    int scanlines_size = GetComponentScanlinesPerImcuRow(i);
+    if (scanlines_sizes_[i] != scanlines_size) {
+      if (scanlines_[i]) {
+        delete scanlines_[i];
+      }
+      scanlines_[i] = new uint8* [scanlines_size];
+      scanlines_sizes_[i] = scanlines_size;
+    }
+
+    // We allocate padding for the final scanline to pad it up to DCTSIZE bytes
+    // to avoid memory errors, since jpeglib only reads full MCUs blocks. For
+    // the preceding scanlines, the padding is not needed/wanted because the
+    // following addresses will already be valid (they are the initial bytes of
+    // the next scanline) and will be overwritten when jpeglib writes out that
+    // next scanline.
+    int databuf_stride = GetComponentStride(i);
+    int databuf_size = scanlines_size * databuf_stride;
+    if (databuf_strides_[i] != databuf_stride) {
+      if (databuf_[i]) {
+        delete databuf_[i];
+      }
+      databuf_[i] = new uint8[databuf_size];
+      databuf_strides_[i] = databuf_stride;
+    }
+
+    if (GetComponentStride(i) != GetComponentWidth(i)) {
+      has_scanline_padding_ = LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_TRUE;
+}
+
+static int DivideAndRoundUp(int numerator, int denominator) {
+  return (numerator + denominator - 1) / denominator;
+}
+
+static int DivideAndRoundDown(int numerator, int denominator) {
+  return numerator / denominator;
+}
+
+// Returns width of the last loaded frame.
+int MJpegDecoder::GetWidth() {
+  return decompress_struct_->image_width;
+}
+
+// Returns height of the last loaded frame.
+int MJpegDecoder::GetHeight() {
+  return decompress_struct_->image_height;
+}
+
+// Returns format of the last loaded frame. The return value is one of the
+// kColorSpace* constants.
+int MJpegDecoder::GetColorSpace() {
+  return decompress_struct_->jpeg_color_space;
+}
+
+// Number of color components in the color space.
+int MJpegDecoder::GetNumComponents() {
+  return decompress_struct_->num_components;
+}
+
+// Sample factors of the n-th component.
+int MJpegDecoder::GetHorizSampFactor(int component) {
+  return decompress_struct_->comp_info[component].h_samp_factor;
+}
+
+int MJpegDecoder::GetVertSampFactor(int component) {
+  return decompress_struct_->comp_info[component].v_samp_factor;
+}
+
+int MJpegDecoder::GetHorizSubSampFactor(int component) {
+  return decompress_struct_->max_h_samp_factor /
+      GetHorizSampFactor(component);
+}
+
+int MJpegDecoder::GetVertSubSampFactor(int component) {
+  return decompress_struct_->max_v_samp_factor /
+      GetVertSampFactor(component);
+}
+
+int MJpegDecoder::GetImageScanlinesPerImcuRow() {
+  return decompress_struct_->max_v_samp_factor * DCTSIZE;
+}
+
+int MJpegDecoder::GetComponentScanlinesPerImcuRow(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetImageScanlinesPerImcuRow(), vs);
+}
+
+int MJpegDecoder::GetComponentWidth(int component) {
+  int hs = GetHorizSubSampFactor(component);
+  return DivideAndRoundUp(GetWidth(), hs);
+}
+
+int MJpegDecoder::GetComponentHeight(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetHeight(), vs);
+}
+
+// Get width in bytes padded out to a multiple of DCTSIZE
+int MJpegDecoder::GetComponentStride(int component) {
+  return (GetComponentWidth(component) + DCTSIZE - 1) & ~(DCTSIZE - 1);
+}
+
+int MJpegDecoder::GetComponentSize(int component) {
+  return GetComponentWidth(component) * GetComponentHeight(component);
+}
+
+LIBYUV_BOOL MJpegDecoder::UnloadFrame() {
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_abort_decompress, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+// TODO(fbarchard): Allow rectangle to be specified: x, y, width, height.
+LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(
+    uint8** planes, int dst_width, int dst_height) {
+  if (dst_width != GetWidth() ||
+      dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // Compute amount of lines to skip to implement vertical crop.
+  // TODO(fbarchard): Ensure skip is a multiple of maximum component
+  // subsample. ie 2
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    // There is no API to skip lines in the output data, so we read them
+    // into the temp buffer.
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip. Must read it and then
+      // copy the parts we want into the destination.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip =
+            DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i) -
+                                rows_to_skip;
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        CopyPlane(databuf_[i] + data_to_skip, GetComponentStride(i),
+                  planes[i], GetComponentWidth(i),
+                  GetComponentWidth(i), scanlines_to_copy);
+        planes[i] += scanlines_to_copy * GetComponentWidth(i);
+      }
+      lines_left -= (GetImageScanlinesPerImcuRow() - skip);
+    }
+  }
+
+  // Read full MCUs but cropped horizontally
+  for (; lines_left > GetImageScanlinesPerImcuRow();
+         lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i);
+      CopyPlane(databuf_[i], GetComponentStride(i),
+                planes[i], GetComponentWidth(i),
+                GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy =
+          DivideAndRoundUp(lines_left, GetVertSubSampFactor(i));
+      CopyPlane(databuf_[i], GetComponentStride(i),
+                planes[i], GetComponentWidth(i),
+                GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+  return FinishDecode();
+}
+
+LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn, void* opaque,
+    int dst_width, int dst_height) {
+  if (dst_width != GetWidth() ||
+      dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // TODO(fbarchard): Compute amount of lines to skip to implement vertical crop
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        // Change our own data buffer pointers so we can pass them to the
+        // callback.
+        databuf_[i] += data_to_skip;
+      }
+      int scanlines_to_copy = GetImageScanlinesPerImcuRow() - skip;
+      (*fn)(opaque, databuf_, databuf_strides_, scanlines_to_copy);
+      // Now change them back.
+      for (int i = 0; i < num_outbufs_; ++i) {
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        databuf_[i] -= data_to_skip;
+      }
+      lines_left -= scanlines_to_copy;
+    }
+  }
+  // Read full MCUs until we get to the crop point.
+  for (; lines_left >= GetImageScanlinesPerImcuRow();
+         lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, GetImageScanlinesPerImcuRow());
+  }
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, lines_left);
+  }
+  return FinishDecode();
+}
+
+void MJpegDecoder::init_source(j_decompress_ptr cinfo) {
+  fill_input_buffer(cinfo);
+}
+
+boolean MJpegDecoder::fill_input_buffer(j_decompress_ptr cinfo) {
+  BufferVector* buf_vec = (BufferVector*)(cinfo->client_data);
+  if (buf_vec->pos >= buf_vec->len) {
+    assert(0 && "No more data");
+    // ERROR: No more data
+    return FALSE;
+  }
+  cinfo->src->next_input_byte = buf_vec->buffers[buf_vec->pos].data;
+  cinfo->src->bytes_in_buffer = buf_vec->buffers[buf_vec->pos].len;
+  ++buf_vec->pos;
+  return TRUE;
+}
+
+void MJpegDecoder::skip_input_data(j_decompress_ptr cinfo,
+                                   long num_bytes) {  // NOLINT
+  cinfo->src->next_input_byte += num_bytes;
+}
+
+void MJpegDecoder::term_source(j_decompress_ptr cinfo) {
+  // Nothing to do.
+}
+
+#ifdef HAVE_SETJMP
+void MJpegDecoder::ErrorHandler(j_common_ptr cinfo) {
+  // This is called when a jpeglib command experiences an error. Unfortunately
+  // jpeglib's error handling model is not very flexible, because it expects the
+  // error handler to not return--i.e., it wants the program to terminate. To
+  // recover from errors we use setjmp() as shown in their example. setjmp() is
+  // C's implementation for the "call with current continuation" functionality
+  // seen in some functional programming languages.
+  // A formatted message can be output, but is unsafe for release.
+#ifdef DEBUG
+  char buf[JMSG_LENGTH_MAX];
+  (*cinfo->err->format_message)(cinfo, buf);
+  // ERROR: Error in jpeglib: buf
+#endif
+
+  SetJmpErrorMgr* mgr = (SetJmpErrorMgr*)(cinfo->err);
+  // This rewinds the call stack to the point of the corresponding setjmp()
+  // and causes it to return (for a second time) with value 1.
+  longjmp(mgr->setjmp_buffer, 1);
+}
+#endif
+
+void MJpegDecoder::AllocOutputBuffers(int num_outbufs) {
+  if (num_outbufs != num_outbufs_) {
+    // We could perhaps optimize this case to resize the output buffers without
+    // necessarily having to delete and recreate each one, but it's not worth
+    // it.
+    DestroyOutputBuffers();
+
+    scanlines_ = new uint8** [num_outbufs];
+    scanlines_sizes_ = new int[num_outbufs];
+    databuf_ = new uint8* [num_outbufs];
+    databuf_strides_ = new int[num_outbufs];
+
+    for (int i = 0; i < num_outbufs; ++i) {
+      scanlines_[i] = NULL;
+      scanlines_sizes_[i] = 0;
+      databuf_[i] = NULL;
+      databuf_strides_[i] = 0;
+    }
+
+    num_outbufs_ = num_outbufs;
+  }
+}
+
+void MJpegDecoder::DestroyOutputBuffers() {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    delete [] scanlines_[i];
+    delete [] databuf_[i];
+  }
+  delete [] scanlines_;
+  delete [] databuf_;
+  delete [] scanlines_sizes_;
+  delete [] databuf_strides_;
+  scanlines_ = NULL;
+  databuf_ = NULL;
+  scanlines_sizes_ = NULL;
+  databuf_strides_ = NULL;
+  num_outbufs_ = 0;
+}
+
+// JDCT_IFAST and do_block_smoothing improve performance substantially.
+LIBYUV_BOOL MJpegDecoder::StartDecode() {
+  decompress_struct_->raw_data_out = TRUE;
+  decompress_struct_->dct_method = JDCT_IFAST;  // JDCT_ISLOW is default
+  decompress_struct_->dither_mode = JDITHER_NONE;
+  // Not applicable to 'raw':
+  decompress_struct_->do_fancy_upsampling = LIBYUV_FALSE;
+  // Only for buffered mode:
+  decompress_struct_->enable_2pass_quant = LIBYUV_FALSE;
+  // Blocky but fast:
+  decompress_struct_->do_block_smoothing = LIBYUV_FALSE;
+
+  if (!jpeg_start_decompress(decompress_struct_)) {
+    // ERROR: Couldn't start JPEG decompressor";
+    return LIBYUV_FALSE;
+  }
+  return LIBYUV_TRUE;
+}
+
+LIBYUV_BOOL MJpegDecoder::FinishDecode() {
+  // jpeglib considers it an error if we finish without decoding the whole
+  // image, so we call "abort" rather than "finish".
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+void MJpegDecoder::SetScanlinePointers(uint8** data) {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    uint8* data_i = data[i];
+    for (int j = 0; j < scanlines_sizes_[i]; ++j) {
+      scanlines_[i][j] = data_i;
+      data_i += GetComponentStride(i);
+    }
+  }
+}
+
+inline LIBYUV_BOOL MJpegDecoder::DecodeImcuRow() {
+  return (unsigned int)(GetImageScanlinesPerImcuRow()) ==
+      jpeg_read_raw_data(decompress_struct_,
+                         scanlines_,
+                         GetImageScanlinesPerImcuRow());
+}
+
+// The helper function which recognizes the jpeg sub-sampling type.
+JpegSubsamplingType MJpegDecoder::JpegSubsamplingTypeHelper(
+    int* subsample_x, int* subsample_y, int number_of_components) {
+  if (number_of_components == 3) {  // Color images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 2 && subsample_y[1] == 2 &&
+        subsample_x[2] == 2 && subsample_y[2] == 2) {
+      return kJpegYuv420;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 2 && subsample_y[1] == 1 &&
+        subsample_x[2] == 2 && subsample_y[2] == 1) {
+      return kJpegYuv422;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 1 && subsample_y[1] == 1 &&
+        subsample_x[2] == 1 && subsample_y[2] == 1) {
+      return kJpegYuv444;
+    }
+  } else if (number_of_components == 1) {  // Grey-scale images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1) {
+      return kJpegYuv400;
+    }
+  }
+  return kJpegUnknown;
+}
+
+}  // namespace libyuv
+#endif  // HAVE_JPEG
+
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/mjpeg_validate.cc b/drivers/theoraplayer/src/YUV/libyuv/src/mjpeg_validate.cc
new file mode 100755
index 0000000000..23d22d099b
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/mjpeg_validate.cc
@@ -0,0 +1,47 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Helper function to validate the jpeg appears intact.
+// TODO(fbarchard): Optimize case where SOI is found but EOI is not.
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) {
+  size_t i;
+  if (sample_size < 64) {
+    // ERROR: Invalid jpeg size: sample_size
+    return LIBYUV_FALSE;
+  }
+  if (sample[0] != 0xff || sample[1] != 0xd8) {  // Start Of Image
+    // ERROR: Invalid jpeg initial start code
+    return LIBYUV_FALSE;
+  }
+  for (i = sample_size - 2; i > 1;) {
+    if (sample[i] != 0xd9) {
+      if (sample[i] == 0xff && sample[i + 1] == 0xd9) {  // End Of Image
+        return LIBYUV_TRUE;  // Success: Valid jpeg.
+      }
+      --i;
+    }
+    --i;
+  }
+  // ERROR: Invalid jpeg end code not found. Size sample_size
+  return LIBYUV_FALSE;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/planar_functions.cc b/drivers/theoraplayer/src/YUV/libyuv/src/planar_functions.cc
new file mode 100755
index 0000000000..f0a8989051
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/planar_functions.cc
@@ -0,0 +1,2238 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/planar_functions.h"
+
+#include <string.h>  // for memset()
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  int y;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Copy I422.
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+  return 0;
+}
+
+// Copy I444.
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+  return 0;
+}
+
+// Copy I400.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Convert I420 to I400.
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror a plane of data.
+void MirrorPlane(const uint8* src_y, int src_stride_y,
+                 uint8* dst_y, int dst_stride_y,
+                 int width, int height) {
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    MirrorRow = MirrorRow_NEON;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16)) {
+    MirrorRow = MirrorRow_SSE2;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    MirrorRow = MirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
+    MirrorRow = MirrorRow_AVX2;
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    MirrorRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*YUY2ToUV422Row)(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      YUY2ToUV422Row_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) =
+      YUY2ToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Unaligned_SSE2;
+      YUY2ToYRow = YUY2ToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_yuy2, 16) && IS_ALIGNED(src_stride_yuy2, 16)) {
+        YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          YUY2ToYRow = YUY2ToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUV422Row = YUY2ToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*UYVYToUV422Row)(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      UYVYToUV422Row_C;
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) = UYVYToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 16) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUV422Row = UYVYToUV422Row_Unaligned_SSE2;
+      UYVYToYRow = UYVYToYRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_uyvy, 16) && IS_ALIGNED(src_stride_uyvy, 16)) {
+        UYVYToUV422Row = UYVYToUV422Row_SSE2;
+        if (IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+          UYVYToYRow = UYVYToYRow_SSE2;
+        }
+      }
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 32) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUV422Row = UYVYToUV422Row_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (width >= 16) {
+      UYVYToUV422Row = UYVYToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUV422Row = UYVYToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    UYVYToUV422Row(src_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    src_uyvy += src_stride_uyvy;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Mirror I400 with optional flipping
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+
+  MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror I420 with optional flipping
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  MirrorPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  MirrorPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+#if defined(HAS_ARGBMIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBMirrorRow = ARGBMirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
+    ARGBMirrorRow = ARGBMirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
+    ARGBMirrorRow = ARGBMirrorRow_NEON;
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    ARGBMirrorRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Get a blender that optimized for the CPU, alignment and pixel count.
+// As there are 6 blenders to choose from, the caller should try to use
+// the same blend function for all pixels if possible.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend() {
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = ARGBBlendRow_C;
+#if defined(HAS_ARGBBLENDROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBBlendRow = ARGBBlendRow_SSSE3;
+    return ARGBBlendRow;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBBlendRow = ARGBBlendRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBBlendRow = ARGBBlendRow_NEON;
+  }
+#endif
+  return ARGBBlendRow;
+}
+
+// Alpha Blend 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  int y;
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = GetARGBBlend();
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+
+  for (y = 0; y < height; ++y) {
+    ARGBBlendRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBMultiplyRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBMultiplyRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBMULTIPLYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_NEON;
+    }
+  }
+#endif
+
+  // Multiply plane
+  for (y = 0; y < height; ++y) {
+    ARGBMultiplyRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Add 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height) {
+  int y;
+  void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                     int width) = ARGBAddRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBADDROW_SSE2) && defined(_MSC_VER)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAddRow = ARGBAddRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBADDROW_SSE2) && !defined(_MSC_VER)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBAddRow = ARGBAddRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAddRow = ARGBAddRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBAddRow = ARGBAddRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBAddRow = ARGBAddRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_NEON;
+    }
+  }
+#endif
+
+  // Add plane
+  for (y = 0; y < height; ++y) {
+    ARGBAddRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Subtract 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBSubtractRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBSubtractRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSUBTRACTROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSubtractRow = ARGBSubtractRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_NEON;
+    }
+  }
+#endif
+
+  // Subtract plane
+  for (y = 0; y < height; ++y) {
+    ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  int y;
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_bgra == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_bgra = 0;
+  }
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_bgra, 16) && IS_ALIGNED(dst_stride_bgra, 16)) {
+        I422ToBGRARow = I422ToBGRARow_SSSE3;
+      }
+    }
+  }
+#elif defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  int y;
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_abgr == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_abgr = 0;
+  }
+#if defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#elif defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_abgr, 16) && IS_ALIGNED(dst_stride_abgr, 16)) {
+        I422ToABGRRow = I422ToABGRRow_SSSE3;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_rgba == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_rgba = 0;
+  }
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#elif defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_rgba, 16) && IS_ALIGNED(dst_stride_rgba, 16)) {
+        I422ToRGBARow = I422ToRGBARow_SSSE3;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV12ToRGB565Row)(const uint8* y_buf,
+                          const uint8* uv_buf,
+                          uint8* rgb_buf,
+                          int width) = NV12ToRGB565Row_C;
+  if (!src_y || !src_uv || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
+    }
+  }
+#elif defined(HAS_NV12TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_vu, int src_stride_vu,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV21ToRGB565Row)(const uint8* y_buf,
+                          const uint8* src_vu,
+                          uint8* rgb_buf,
+                          int width) = NV21ToRGB565Row_C;
+  if (!src_y || !src_vu || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV21TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_SSSE3;
+    }
+  }
+#elif defined(HAS_NV21TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV21ToRGB565Row(src_y, src_vu, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_vu += src_stride_vu;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value) {
+  int y;
+  uint32 v32 = value | (value << 8) | (value << 16) | (value << 24);
+  void (*SetRow)(uint8* dst, uint32 value, int pix) = SetRow_C;
+  // Coalesce rows.
+  if (dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    dst_stride_y = 0;
+  }
+#if defined(HAS_SETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) &&
+      IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    SetRow = SetRow_NEON;
+  }
+#endif
+#if defined(HAS_SETROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    SetRow = SetRow_X86;
+  }
+#endif
+
+  // Set plane
+  for (y = 0; y < height; ++y) {
+    SetRow(dst_y, v32, width);
+    dst_y += dst_stride_y;
+  }
+}
+
+// Draw a rectangle into I420
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y,
+             int width, int height,
+             int value_y, int value_u, int value_v) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  uint8* start_y = dst_y + y * dst_stride_y + x;
+  uint8* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
+  uint8* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
+  if (!dst_y || !dst_u || !dst_v ||
+      width <= 0 || height <= 0 ||
+      x < 0 || y < 0 ||
+      value_y < 0 || value_y > 255 ||
+      value_u < 0 || value_u > 255 ||
+      value_v < 0 || value_v > 255) {
+    return -1;
+  }
+
+  SetPlane(start_y, dst_stride_y, width, height, value_y);
+  SetPlane(start_u, dst_stride_u, halfwidth, halfheight, value_u);
+  SetPlane(start_v, dst_stride_v, halfwidth, halfheight, value_v);
+  return 0;
+}
+
+// Draw a rectangle into ARGB
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height,
+             uint32 value) {
+  if (!dst_argb ||
+      width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  dst_argb += dst_y * dst_stride_argb + dst_x * 4;
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_SETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBSetRows_NEON(dst_argb, value, width, dst_stride_argb, height);
+    return 0;
+  }
+#endif
+#if defined(HAS_SETROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBSetRows_X86(dst_argb, value, width, dst_stride_argb, height);
+    return 0;
+  }
+#endif
+  ARGBSetRows_C(dst_argb, value, width, dst_stride_argb, height);
+  return 0;
+}
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+// An unattenutated ARGB alpha blend uses the formula
+// p = a * f + (1 - a) * b
+// where
+//   p is output pixel
+//   f is foreground pixel
+//   b is background pixel
+//   a is alpha value from foreground pixel
+// An preattenutated ARGB alpha blend uses the formula
+// p = f + (1 - a) * b
+// where
+//   f is foreground pixel premultiplied by alpha
+
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBAttenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 4) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBAttenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBUnattenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                             int width) = ARGBUnattenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBUNATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBUNATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Neon version.
+
+  for (y = 0; y < height; ++y) {
+    ARGBUnattenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB to Grayed ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#elif defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#elif defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(dst, dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBSepiaRow)(uint8* dst_argb, int width) = ARGBSepiaRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSEPIAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBSepiaRow = ARGBSepiaRow_SSSE3;
+  }
+#elif defined(HAS_ARGBSEPIAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBSepiaRow = ARGBSepiaRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBSepiaRow(dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x4 matrix to each ARGB pixel.
+// Note: Normally for shading, but can be used to swizzle or invert.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBColorMatrixRow)(const uint8* src_argb, uint8* dst_argb,
+      const int8* matrix_argb, int width) = ARGBColorMatrixRow_C;
+  if (!src_argb || !dst_argb || !matrix_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORMATRIXROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_SSSE3;
+  }
+#elif defined(HAS_ARGBCOLORMATRIXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorMatrixRow(src_argb, dst_argb, matrix_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x3 matrix to each ARGB pixel.
+// Deprecated.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int dst_x, int dst_y, int width, int height) {
+  SIMD_ALIGNED(int8 matrix_argb[16]);
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !matrix_rgb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+
+  // Convert 4x3 7 bit matrix to 4x4 6 bit matrix.
+  matrix_argb[0] = matrix_rgb[0] / 2;
+  matrix_argb[1] = matrix_rgb[1] / 2;
+  matrix_argb[2] = matrix_rgb[2] / 2;
+  matrix_argb[3] = matrix_rgb[3] / 2;
+  matrix_argb[4] = matrix_rgb[4] / 2;
+  matrix_argb[5] = matrix_rgb[5] / 2;
+  matrix_argb[6] = matrix_rgb[6] / 2;
+  matrix_argb[7] = matrix_rgb[7] / 2;
+  matrix_argb[8] = matrix_rgb[8] / 2;
+  matrix_argb[9] = matrix_rgb[9] / 2;
+  matrix_argb[10] = matrix_rgb[10] / 2;
+  matrix_argb[11] = matrix_rgb[11] / 2;
+  matrix_argb[14] = matrix_argb[13] = matrix_argb[12] = 0;
+  matrix_argb[15] = 64;  // 1.0
+
+  return ARGBColorMatrix((const uint8*)(dst), dst_stride_argb,
+                         dst, dst_stride_argb,
+                         &matrix_argb[0], width, height);
+}
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                            int width) = ARGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBColorTableRow = ARGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*RGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                           int width) = RGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_RGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    RGBColorTableRow = RGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    RGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// ARGBQuantize is used to posterize art.
+// e.g. rgb / qvalue * qvalue + qvalue / 2
+// But the low levels implement efficiently with 3 parameters, and could be
+// used for other high level operations.
+// dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+// where scale is 1 / interval_size as a fixed point value.
+// The divide is replaces with a multiply by reciprocal fixed point multiply.
+// Caveat - although SSE2 saturates, the C function does not and should be used
+// with care if doing anything but quantization.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBQuantizeRow)(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) = ARGBQuantizeRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0 ||
+      interval_size < 1 || interval_size > 255) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBQUANTIZEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_SSE2;
+  }
+#elif defined(HAS_ARGBQUANTIZEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBQuantizeRow(dst, scale, interval_size, interval_offset, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8* row, int32* cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  int32* previous_cumsum = dst_cumsum;
+  if (!dst_cumsum || !src_argb || width <= 0 || height <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+  }
+#endif
+  memset(dst_cumsum, 0, width * sizeof(dst_cumsum[0]) * 4);  // 4 int per pixel.
+  for (y = 0; y < height; ++y) {
+    ComputeCumulativeSumRow(src_argb, dst_cumsum, previous_cumsum, width);
+    previous_cumsum = dst_cumsum;
+    dst_cumsum += dst_stride32_cumsum;
+    src_argb += src_stride_argb;
+  }
+  return 0;
+}
+
+// Blur ARGB image.
+// Caller should allocate CumulativeSum table of width * height * 16 bytes
+// aligned to 16 byte boundary. height can be radius * 2 + 2 to save memory
+// as the buffer is treated as circular.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8 *row, int32 *cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  void (*CumulativeSumToAverageRow)(const int32* topleft, const int32* botleft,
+      int width, int area, uint8* dst, int count) = CumulativeSumToAverageRow_C;
+  int32* cumsum_bot_row;
+  int32* max_cumsum_bot_row;
+  int32* cumsum_top_row;
+
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  if (radius > height) {
+    radius = height;
+  }
+  if (radius > (width / 2 - 1)) {
+    radius = width / 2 - 1;
+  }
+  if (radius <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+    CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2;
+  }
+#endif
+  // Compute enough CumulativeSum for first row to be blurred. After this
+  // one row of CumulativeSum is updated at a time.
+  ARGBComputeCumulativeSum(src_argb, src_stride_argb,
+                           dst_cumsum, dst_stride32_cumsum,
+                           width, radius);
+
+  src_argb = src_argb + radius * src_stride_argb;
+  cumsum_bot_row = &dst_cumsum[(radius - 1) * dst_stride32_cumsum];
+
+  max_cumsum_bot_row = &dst_cumsum[(radius * 2 + 2) * dst_stride32_cumsum];
+  cumsum_top_row = &dst_cumsum[0];
+
+  for (y = 0; y < height; ++y) {
+    int top_y = ((y - radius - 1) >= 0) ? (y - radius - 1) : 0;
+    int bot_y = ((y + radius) < height) ? (y + radius) : (height - 1);
+    int area = radius * (bot_y - top_y);
+    int boxwidth = radius * 4;
+    int x;
+    int n;
+
+    // Increment cumsum_top_row pointer with circular buffer wrap around.
+    if (top_y) {
+      cumsum_top_row += dst_stride32_cumsum;
+      if (cumsum_top_row >= max_cumsum_bot_row) {
+        cumsum_top_row = dst_cumsum;
+      }
+    }
+    // Increment cumsum_bot_row pointer with circular buffer wrap around and
+    // then fill in a row of CumulativeSum.
+    if ((y + radius) < height) {
+      const int32* prev_cumsum_bot_row = cumsum_bot_row;
+      cumsum_bot_row += dst_stride32_cumsum;
+      if (cumsum_bot_row >= max_cumsum_bot_row) {
+        cumsum_bot_row = dst_cumsum;
+      }
+      ComputeCumulativeSumRow(src_argb, cumsum_bot_row, prev_cumsum_bot_row,
+                              width);
+      src_argb += src_stride_argb;
+    }
+
+    // Left clipped.
+    for (x = 0; x < radius + 1; ++x) {
+      CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+      area += (bot_y - top_y);
+      boxwidth += 4;
+    }
+
+    // Middle unclipped.
+    n = (width - 1) - radius - x + 1;
+    CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                              boxwidth, area, &dst_argb[x * 4], n);
+
+    // Right clipped.
+    for (x += n; x <= width - 1; ++x) {
+      area -= (bot_y - top_y);
+      boxwidth -= 4;
+      CumulativeSumToAverageRow(cumsum_top_row + (x - radius - 1) * 4,
+                                cumsum_bot_row + (x - radius - 1) * 4,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+    }
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply ARGB image by a specified ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value) {
+  int y;
+  void (*ARGBShadeRow)(const uint8* src_argb, uint8* dst_argb,
+                       int width, uint32 value) = ARGBShadeRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0 || value == 0u) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHADEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBShadeRow = ARGBShadeRow_SSE2;
+  }
+#elif defined(HAS_ARGBSHADEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBShadeRow = ARGBShadeRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShadeRow(src_argb, dst_argb, width, value);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Interpolate 2 ARGB images by specified amount (0 to 255).
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation) {
+  int y;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
+          IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
+          IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 4) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && width >= 1 &&
+      IS_ALIGNED(src_argb0, 4) && IS_ALIGNED(src_stride_argb0, 4) &&
+      IS_ALIGNED(src_argb1, 4) && IS_ALIGNED(src_stride_argb1, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    ScaleARGBFilterRows = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    InterpolateRow(dst_argb, src_argb0, src_argb1 - src_argb0,
+                   width * 4, interpolation);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height) {
+  int y;
+  void (*ARGBShuffleRow)(const uint8* src_bgra, uint8* dst_argb,
+                         const uint8* shuffler, int pix) = ARGBShuffleRow_C;
+  if (!src_bgra || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_bgra == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_bgra = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHUFFLEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBShuffleRow = ARGBShuffleRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_bgra, 16) && IS_ALIGNED(src_stride_bgra, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        ARGBShuffleRow = ARGBShuffleRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 16) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBShuffleRow = ARGBShuffleRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 4) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShuffleRow(src_bgra, dst_argb, shuffler, width);
+    src_bgra += src_stride_bgra;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_argb, int dst_stride_argb,
+                        int width, int height,
+                        void (*SobelRow)(const uint8* src_sobelx,
+                                         const uint8* src_sobely,
+                                         uint8* dst, int width)) {
+  int y;
+  void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) = ARGBToBayerGGRow_C;
+  void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) = SobelYRow_C;
+  void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobely, int width) =
+      SobelXRow_C;
+  const int kEdge = 16;  // Extra pixels at start of row for extrude/align.
+  if (!src_argb  || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // ARGBToBayer used to select G channel from ARGB.
+#if defined(HAS_ARGBTOBAYERGGROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 8 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToBayerRow = ARGBToBayerGGRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerGGRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && width >= 8 &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
+    ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToBayerRow = ARGBToBayerGGRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToBayerRow = ARGBToBayerGGRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SOBELYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelYRow = SobelYRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelYRow = SobelYRow_NEON;
+  }
+#endif
+#if defined(HAS_SOBELXROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelXRow = SobelXRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelXRow = SobelXRow_NEON;
+  }
+#endif
+  {
+    // 3 rows with edges before/after.
+    const int kRowSize = (width + kEdge + 15) & ~15;
+    align_buffer_64(rows, kRowSize * 2 + (kEdge + kRowSize * 3 + kEdge));
+    uint8* row_sobelx = rows;
+    uint8* row_sobely = rows + kRowSize;
+    uint8* row_y = rows + kRowSize * 2;
+
+    // Convert first row.
+    uint8* row_y0 = row_y + kEdge;
+    uint8* row_y1 = row_y0 + kRowSize;
+    uint8* row_y2 = row_y1 + kRowSize;
+    ARGBToBayerRow(src_argb, row_y0, 0x0d090501, width);
+    row_y0[-1] = row_y0[0];
+    memset(row_y0 + width, row_y0[width - 1], 16);  // Extrude 16 for valgrind.
+    ARGBToBayerRow(src_argb, row_y1, 0x0d090501, width);
+    row_y1[-1] = row_y1[0];
+    memset(row_y1 + width, row_y1[width - 1], 16);
+    memset(row_y2 + width, 0, 16);
+
+    for (y = 0; y < height; ++y) {
+      // Convert next row of ARGB to Y.
+      if (y < (height - 1)) {
+        src_argb += src_stride_argb;
+      }
+      ARGBToBayerRow(src_argb, row_y2, 0x0d090501, width);
+      row_y2[-1] = row_y2[0];
+      row_y2[width] = row_y2[width - 1];
+
+      SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width);
+      SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, width);
+      SobelRow(row_sobelx, row_sobely, dst_argb, width);
+
+      // Cycle thru circular queue of 3 row_y buffers.
+      {
+        uint8* row_yt = row_y0;
+        row_y0 = row_y1;
+        row_y1 = row_y2;
+        row_y2 = row_yt;
+      }
+
+      dst_argb += dst_stride_argb;
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) = SobelRow_C;
+#if defined(HAS_SOBELROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    SobelRow = SobelRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    SobelRow = SobelRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelRow);
+}
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height) {
+  void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_, int width) = SobelToPlaneRow_C;
+#if defined(HAS_SOBELTOPLANEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
+    SobelToPlaneRow = SobelToPlaneRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELTOPLANEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    SobelToPlaneRow = SobelToPlaneRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y,
+                      width, height, SobelToPlaneRow);
+}
+
+// SobelXY ARGB effect.
+// Similar to Sobel, but also stores Sobel X in R and Sobel Y in B.  G = Sobel.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) = SobelXYRow_C;
+#if defined(HAS_SOBELXYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    SobelXYRow = SobelXYRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELXYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    SobelXYRow = SobelXYRow_NEON;
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelXYRow);
+}
+
+// Apply a 4x4 polynomial to each ARGB pixel.
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height) {
+  int y;
+  void (*ARGBPolynomialRow)(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) = ARGBPolynomialRow_C;
+  if (!src_argb || !dst_argb || !poly || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBPOLYNOMIALROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBPOLYNOMIALROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && TestCpuFlag(kCpuHasFMA3) &&
+      IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBPolynomialRow(src_argb, dst_argb, poly, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a lumacolortable to each ARGB pixel.
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma,
+                       int width, int height) {
+  int y;
+  void (*ARGBLumaColorTableRow)(const uint8* src_argb, uint8* dst_argb,
+      int width, const uint8* luma, const uint32 lumacoeff) =
+      ARGBLumaColorTableRow_C;
+  if (!src_argb || !dst_argb || !luma || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBLUMACOLORTABLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4)) {
+    ARGBLumaColorTableRow = ARGBLumaColorTableRow_SSSE3;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBLumaColorTableRow(src_argb, dst_argb, width, luma, 0x00264b0f);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy Alpha from one ARGB image to another.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBCopyAlphaRow)(const uint8* src_argb, uint8* dst_argb, int width) =
+      ARGBCopyAlphaRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16) &&
+      IS_ALIGNED(width, 8)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyAlphaRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy a planar Y channel to the alpha channel of a destination ARGB image.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height) {
+  int y;
+  void (*ARGBCopyYToAlphaRow)(const uint8* src_y, uint8* dst_argb, int width) =
+      ARGBCopyYToAlphaRow_C;
+  if (!src_y || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYYTOALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(src_y, 16) && IS_ALIGNED(src_stride_y, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16) &&
+      IS_ALIGNED(width, 8)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYYTOALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyYToAlphaRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/rotate.cc b/drivers/theoraplayer/src/YUV/libyuv/src/rotate.cc
new file mode 100755
index 0000000000..b052ac1dc4
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/rotate.cc
@@ -0,0 +1,1301 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/convert.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#if defined(__APPLE__) && defined(__i386__)
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".private_extern _" #name "                \n"                             \
+    ".align 4,0x90                             \n"                             \
+"_" #name ":                                   \n"
+#elif defined(__MINGW32__) || defined(__CYGWIN__) && defined(__i386__)
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".align 4,0x90                             \n"                             \
+"_" #name ":                                   \n"
+#else
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".align 4,0x90                             \n"                             \
+#name ":                                       \n"
+#endif
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_MIRRORROW_NEON
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
+#define HAS_MIRRORROW_UV_NEON
+void MirrorUVRow_NEON(const uint8* src, uint8* dst_a, uint8* dst_b, int width);
+#define HAS_TRANSPOSE_WX8_NEON
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride, int width);
+#define HAS_TRANSPOSE_UVWX8_NEON
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width);
+#endif  // defined(__ARM_NEON__)
+
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_TRANSPOSE_WX8_MIPS_DSPR2
+void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width);
+
+void TransposeWx8_FAST_MIPS_DSPR2(const uint8* src, int src_stride,
+                                  uint8* dst, int dst_stride, int width);
+#define HAS_TRANSPOSE_UVWx8_MIPS_DSPR2
+void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                               uint8* dst_a, int dst_stride_a,
+                               uint8* dst_b, int dst_stride_b,
+                               int width);
+#endif  // defined(__mips__)
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    defined(_M_IX86) && defined(_MSC_VER)
+#define HAS_TRANSPOSE_WX8_SSSE3
+__declspec(naked) __declspec(align(16))
+static void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                               uint8* dst, int dst_stride, int width) {
+  __asm {
+    push      edi
+    push      esi
+    push      ebp
+    mov       eax, [esp + 12 + 4]   // src
+    mov       edi, [esp + 12 + 8]   // src_stride
+    mov       edx, [esp + 12 + 12]  // dst
+    mov       esi, [esp + 12 + 16]  // dst_stride
+    mov       ecx, [esp + 12 + 20]  // width
+
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    align      4
+ convertloop:
+    movq      xmm0, qword ptr [eax]
+    lea       ebp, [eax + 8]
+    movq      xmm1, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm0, xmm1
+    movq      xmm2, qword ptr [eax]
+    movdqa    xmm1, xmm0
+    palignr   xmm1, xmm1, 8
+    movq      xmm3, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm2, xmm3
+    movdqa    xmm3, xmm2
+    movq      xmm4, qword ptr [eax]
+    palignr   xmm3, xmm3, 8
+    movq      xmm5, qword ptr [eax + edi]
+    punpcklbw xmm4, xmm5
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm5, xmm4
+    movq      xmm6, qword ptr [eax]
+    palignr   xmm5, xmm5, 8
+    movq      xmm7, qword ptr [eax + edi]
+    punpcklbw xmm6, xmm7
+    mov       eax, ebp
+    movdqa    xmm7, xmm6
+    palignr   xmm7, xmm7, 8
+    // Second round of bit swap.
+    punpcklwd xmm0, xmm2
+    punpcklwd xmm1, xmm3
+    movdqa    xmm2, xmm0
+    movdqa    xmm3, xmm1
+    palignr   xmm2, xmm2, 8
+    palignr   xmm3, xmm3, 8
+    punpcklwd xmm4, xmm6
+    punpcklwd xmm5, xmm7
+    movdqa    xmm6, xmm4
+    movdqa    xmm7, xmm5
+    palignr   xmm6, xmm6, 8
+    palignr   xmm7, xmm7, 8
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    punpckldq xmm0, xmm4
+    movq      qword ptr [edx], xmm0
+    movdqa    xmm4, xmm0
+    palignr   xmm4, xmm4, 8
+    movq      qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    punpckldq xmm2, xmm6
+    movdqa    xmm6, xmm2
+    palignr   xmm6, xmm6, 8
+    movq      qword ptr [edx], xmm2
+    punpckldq xmm1, xmm5
+    movq      qword ptr [edx + esi], xmm6
+    lea       edx, [edx + 2 * esi]
+    movdqa    xmm5, xmm1
+    movq      qword ptr [edx], xmm1
+    palignr   xmm5, xmm5, 8
+    punpckldq xmm3, xmm7
+    movq      qword ptr [edx + esi], xmm5
+    lea       edx, [edx + 2 * esi]
+    movq      qword ptr [edx], xmm3
+    movdqa    xmm7, xmm3
+    palignr   xmm7, xmm7, 8
+    sub       ecx, 8
+    movq      qword ptr [edx + esi], xmm7
+    lea       edx, [edx + 2 * esi]
+    jg        convertloop
+
+    pop       ebp
+    pop       esi
+    pop       edi
+    ret
+  }
+}
+
+#define HAS_TRANSPOSE_UVWX8_SSE2
+__declspec(naked) __declspec(align(16))
+static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                                uint8* dst_a, int dst_stride_a,
+                                uint8* dst_b, int dst_stride_b,
+                                int w) {
+  __asm {
+    push      ebx
+    push      esi
+    push      edi
+    push      ebp
+    mov       eax, [esp + 16 + 4]   // src
+    mov       edi, [esp + 16 + 8]   // src_stride
+    mov       edx, [esp + 16 + 12]  // dst_a
+    mov       esi, [esp + 16 + 16]  // dst_stride_a
+    mov       ebx, [esp + 16 + 20]  // dst_b
+    mov       ebp, [esp + 16 + 24]  // dst_stride_b
+    mov       ecx, esp
+    sub       esp, 4 + 16
+    and       esp, ~15
+    mov       [esp + 16], ecx
+    mov       ecx, [ecx + 16 + 28]  // w
+
+    align      4
+ convertloop:
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    movdqa    xmm0, [eax]
+    movdqa    xmm1, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm0  // use xmm7 as temp register.
+    punpcklbw xmm0, xmm1
+    punpckhbw xmm7, xmm1
+    movdqa    xmm1, xmm7
+    movdqa    xmm2, [eax]
+    movdqa    xmm3, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm2
+    punpcklbw xmm2, xmm3
+    punpckhbw xmm7, xmm3
+    movdqa    xmm3, xmm7
+    movdqa    xmm4, [eax]
+    movdqa    xmm5, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm4
+    punpcklbw xmm4, xmm5
+    punpckhbw xmm7, xmm5
+    movdqa    xmm5, xmm7
+    movdqa    xmm6, [eax]
+    movdqa    xmm7, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    [esp], xmm5  // backup xmm5
+    neg       edi
+    movdqa    xmm5, xmm6   // use xmm5 as temp register.
+    punpcklbw xmm6, xmm7
+    punpckhbw xmm5, xmm7
+    movdqa    xmm7, xmm5
+    lea       eax, [eax + 8 * edi + 16]
+    neg       edi
+    // Second round of bit swap.
+    movdqa    xmm5, xmm0
+    punpcklwd xmm0, xmm2
+    punpckhwd xmm5, xmm2
+    movdqa    xmm2, xmm5
+    movdqa    xmm5, xmm1
+    punpcklwd xmm1, xmm3
+    punpckhwd xmm5, xmm3
+    movdqa    xmm3, xmm5
+    movdqa    xmm5, xmm4
+    punpcklwd xmm4, xmm6
+    punpckhwd xmm5, xmm6
+    movdqa    xmm6, xmm5
+    movdqa    xmm5, [esp]  // restore xmm5
+    movdqa    [esp], xmm6  // backup xmm6
+    movdqa    xmm6, xmm5    // use xmm6 as temp register.
+    punpcklwd xmm5, xmm7
+    punpckhwd xmm6, xmm7
+    movdqa    xmm7, xmm6
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    movdqa    xmm6, xmm0
+    punpckldq xmm0, xmm4
+    punpckhdq xmm6, xmm4
+    movdqa    xmm4, xmm6
+    movdqa    xmm6, [esp]  // restore xmm6
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [ebx], xmm0
+    movlpd    qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm4
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm2   // use xmm0 as the temp register.
+    punpckldq xmm2, xmm6
+    movlpd    qword ptr [edx], xmm2
+    movhpd    qword ptr [ebx], xmm2
+    punpckhdq xmm0, xmm6
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm1   // use xmm0 as the temp register.
+    punpckldq xmm1, xmm5
+    movlpd    qword ptr [edx], xmm1
+    movhpd    qword ptr [ebx], xmm1
+    punpckhdq xmm0, xmm5
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm3   // use xmm0 as the temp register.
+    punpckldq xmm3, xmm7
+    movlpd    qword ptr [edx], xmm3
+    movhpd    qword ptr [ebx], xmm3
+    punpckhdq xmm0, xmm7
+    sub       ecx, 8
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    jg        convertloop
+
+    mov       esp, [esp + 16]
+    pop       ebp
+    pop       edi
+    pop       esi
+    pop       ebx
+    ret
+  }
+}
+#elif !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__)))
+#define HAS_TRANSPOSE_WX8_SSSE3
+static void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                               uint8* dst, int dst_stride, int width) {
+  asm volatile (
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    ".p2align  2                                 \n"
+  "1:                                            \n"
+    "movq       (%0),%%xmm0                      \n"
+    "movq       (%0,%3),%%xmm1                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm1,%%xmm0                    \n"
+    "movq       (%0),%%xmm2                      \n"
+    "movdqa     %%xmm0,%%xmm1                    \n"
+    "palignr    $0x8,%%xmm1,%%xmm1               \n"
+    "movq       (%0,%3),%%xmm3                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm3,%%xmm2                    \n"
+    "movdqa     %%xmm2,%%xmm3                    \n"
+    "movq       (%0),%%xmm4                      \n"
+    "palignr    $0x8,%%xmm3,%%xmm3               \n"
+    "movq       (%0,%3),%%xmm5                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm5,%%xmm4                    \n"
+    "movdqa     %%xmm4,%%xmm5                    \n"
+    "movq       (%0),%%xmm6                      \n"
+    "palignr    $0x8,%%xmm5,%%xmm5               \n"
+    "movq       (%0,%3),%%xmm7                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm7,%%xmm6                    \n"
+    "neg        %3                               \n"
+    "movdqa     %%xmm6,%%xmm7                    \n"
+    "lea        0x8(%0,%3,8),%0                  \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    "neg        %3                               \n"
+     // Second round of bit swap.
+    "punpcklwd  %%xmm2,%%xmm0                    \n"
+    "punpcklwd  %%xmm3,%%xmm1                    \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "palignr    $0x8,%%xmm2,%%xmm2               \n"
+    "palignr    $0x8,%%xmm3,%%xmm3               \n"
+    "punpcklwd  %%xmm6,%%xmm4                    \n"
+    "punpcklwd  %%xmm7,%%xmm5                    \n"
+    "movdqa     %%xmm4,%%xmm6                    \n"
+    "movdqa     %%xmm5,%%xmm7                    \n"
+    "palignr    $0x8,%%xmm6,%%xmm6               \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    "punpckldq  %%xmm4,%%xmm0                    \n"
+    "movq       %%xmm0,(%1)                      \n"
+    "movdqa     %%xmm0,%%xmm4                    \n"
+    "palignr    $0x8,%%xmm4,%%xmm4               \n"
+    "movq       %%xmm4,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "punpckldq  %%xmm6,%%xmm2                    \n"
+    "movdqa     %%xmm2,%%xmm6                    \n"
+    "movq       %%xmm2,(%1)                      \n"
+    "palignr    $0x8,%%xmm6,%%xmm6               \n"
+    "punpckldq  %%xmm5,%%xmm1                    \n"
+    "movq       %%xmm6,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "movdqa     %%xmm1,%%xmm5                    \n"
+    "movq       %%xmm1,(%1)                      \n"
+    "palignr    $0x8,%%xmm5,%%xmm5               \n"
+    "movq       %%xmm5,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "punpckldq  %%xmm7,%%xmm3                    \n"
+    "movq       %%xmm3,(%1)                      \n"
+    "movdqa     %%xmm3,%%xmm7                    \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    "sub        $0x8,%2                          \n"
+    "movq       %%xmm7,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "jg         1b                               \n"
+    : "+r"(src),    // %0
+      "+r"(dst),    // %1
+      "+r"(width)   // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "r"((intptr_t)(dst_stride))   // %4
+    : "memory", "cc"
+  #if defined(__SSE2__)
+      , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  #endif
+  );
+}
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(__i386__)
+#define HAS_TRANSPOSE_UVWX8_SSE2
+extern "C" void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                                    uint8* dst_a, int dst_stride_a,
+                                    uint8* dst_b, int dst_stride_b,
+                                    int w);
+  asm (
+    DECLARE_FUNCTION(TransposeUVWx8_SSE2)
+    "push   %ebx                               \n"
+    "push   %esi                               \n"
+    "push   %edi                               \n"
+    "push   %ebp                               \n"
+    "mov    0x14(%esp),%eax                    \n"
+    "mov    0x18(%esp),%edi                    \n"
+    "mov    0x1c(%esp),%edx                    \n"
+    "mov    0x20(%esp),%esi                    \n"
+    "mov    0x24(%esp),%ebx                    \n"
+    "mov    0x28(%esp),%ebp                    \n"
+    "mov    %esp,%ecx                          \n"
+    "sub    $0x14,%esp                         \n"
+    "and    $0xfffffff0,%esp                   \n"
+    "mov    %ecx,0x10(%esp)                    \n"
+    "mov    0x2c(%ecx),%ecx                    \n"
+
+"1:                                            \n"
+    "movdqa (%eax),%xmm0                       \n"
+    "movdqa (%eax,%edi,1),%xmm1                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm0,%xmm7                        \n"
+    "punpcklbw %xmm1,%xmm0                     \n"
+    "punpckhbw %xmm1,%xmm7                     \n"
+    "movdqa %xmm7,%xmm1                        \n"
+    "movdqa (%eax),%xmm2                       \n"
+    "movdqa (%eax,%edi,1),%xmm3                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm2,%xmm7                        \n"
+    "punpcklbw %xmm3,%xmm2                     \n"
+    "punpckhbw %xmm3,%xmm7                     \n"
+    "movdqa %xmm7,%xmm3                        \n"
+    "movdqa (%eax),%xmm4                       \n"
+    "movdqa (%eax,%edi,1),%xmm5                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm4,%xmm7                        \n"
+    "punpcklbw %xmm5,%xmm4                     \n"
+    "punpckhbw %xmm5,%xmm7                     \n"
+    "movdqa %xmm7,%xmm5                        \n"
+    "movdqa (%eax),%xmm6                       \n"
+    "movdqa (%eax,%edi,1),%xmm7                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm5,(%esp)                       \n"
+    "neg    %edi                               \n"
+    "movdqa %xmm6,%xmm5                        \n"
+    "punpcklbw %xmm7,%xmm6                     \n"
+    "punpckhbw %xmm7,%xmm5                     \n"
+    "movdqa %xmm5,%xmm7                        \n"
+    "lea    0x10(%eax,%edi,8),%eax             \n"
+    "neg    %edi                               \n"
+    "movdqa %xmm0,%xmm5                        \n"
+    "punpcklwd %xmm2,%xmm0                     \n"
+    "punpckhwd %xmm2,%xmm5                     \n"
+    "movdqa %xmm5,%xmm2                        \n"
+    "movdqa %xmm1,%xmm5                        \n"
+    "punpcklwd %xmm3,%xmm1                     \n"
+    "punpckhwd %xmm3,%xmm5                     \n"
+    "movdqa %xmm5,%xmm3                        \n"
+    "movdqa %xmm4,%xmm5                        \n"
+    "punpcklwd %xmm6,%xmm4                     \n"
+    "punpckhwd %xmm6,%xmm5                     \n"
+    "movdqa %xmm5,%xmm6                        \n"
+    "movdqa (%esp),%xmm5                       \n"
+    "movdqa %xmm6,(%esp)                       \n"
+    "movdqa %xmm5,%xmm6                        \n"
+    "punpcklwd %xmm7,%xmm5                     \n"
+    "punpckhwd %xmm7,%xmm6                     \n"
+    "movdqa %xmm6,%xmm7                        \n"
+    "movdqa %xmm0,%xmm6                        \n"
+    "punpckldq %xmm4,%xmm0                     \n"
+    "punpckhdq %xmm4,%xmm6                     \n"
+    "movdqa %xmm6,%xmm4                        \n"
+    "movdqa (%esp),%xmm6                       \n"
+    "movlpd %xmm0,(%edx)                       \n"
+    "movhpd %xmm0,(%ebx)                       \n"
+    "movlpd %xmm4,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm4,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm2,%xmm0                        \n"
+    "punpckldq %xmm6,%xmm2                     \n"
+    "movlpd %xmm2,(%edx)                       \n"
+    "movhpd %xmm2,(%ebx)                       \n"
+    "punpckhdq %xmm6,%xmm0                     \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm1,%xmm0                        \n"
+    "punpckldq %xmm5,%xmm1                     \n"
+    "movlpd %xmm1,(%edx)                       \n"
+    "movhpd %xmm1,(%ebx)                       \n"
+    "punpckhdq %xmm5,%xmm0                     \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm3,%xmm0                        \n"
+    "punpckldq %xmm7,%xmm3                     \n"
+    "movlpd %xmm3,(%edx)                       \n"
+    "movhpd %xmm3,(%ebx)                       \n"
+    "punpckhdq %xmm7,%xmm0                     \n"
+    "sub    $0x8,%ecx                          \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "jg     1b                                 \n"
+    "mov    0x10(%esp),%esp                    \n"
+    "pop    %ebp                               \n"
+    "pop    %edi                               \n"
+    "pop    %esi                               \n"
+    "pop    %ebx                               \n"
+#if defined(__native_client__)
+    "pop    %ecx                               \n"
+    "and    $0xffffffe0,%ecx                   \n"
+    "jmp    *%ecx                              \n"
+#else
+    "ret                                       \n"
+#endif
+);
+#elif !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
+    defined(__x86_64__)
+// 64 bit version has enough registers to do 16x8 to 8x16 at a time.
+#define HAS_TRANSPOSE_WX8_FAST_SSSE3
+static void TransposeWx8_FAST_SSSE3(const uint8* src, int src_stride,
+                                    uint8* dst, int dst_stride, int width) {
+  asm volatile (
+  // Read in the data from the source pointer.
+  // First round of bit swap.
+  ".p2align  2                                 \n"
+"1:                                            \n"
+  "movdqa     (%0),%%xmm0                      \n"
+  "movdqa     (%0,%3),%%xmm1                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpcklbw  %%xmm1,%%xmm0                    \n"
+  "punpckhbw  %%xmm1,%%xmm8                    \n"
+  "movdqa     (%0),%%xmm2                      \n"
+  "movdqa     %%xmm0,%%xmm1                    \n"
+  "movdqa     %%xmm8,%%xmm9                    \n"
+  "palignr    $0x8,%%xmm1,%%xmm1               \n"
+  "palignr    $0x8,%%xmm9,%%xmm9               \n"
+  "movdqa     (%0,%3),%%xmm3                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm2,%%xmm10                   \n"
+  "punpcklbw  %%xmm3,%%xmm2                    \n"
+  "punpckhbw  %%xmm3,%%xmm10                   \n"
+  "movdqa     %%xmm2,%%xmm3                    \n"
+  "movdqa     %%xmm10,%%xmm11                  \n"
+  "movdqa     (%0),%%xmm4                      \n"
+  "palignr    $0x8,%%xmm3,%%xmm3               \n"
+  "palignr    $0x8,%%xmm11,%%xmm11             \n"
+  "movdqa     (%0,%3),%%xmm5                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm4,%%xmm12                   \n"
+  "punpcklbw  %%xmm5,%%xmm4                    \n"
+  "punpckhbw  %%xmm5,%%xmm12                   \n"
+  "movdqa     %%xmm4,%%xmm5                    \n"
+  "movdqa     %%xmm12,%%xmm13                  \n"
+  "movdqa     (%0),%%xmm6                      \n"
+  "palignr    $0x8,%%xmm5,%%xmm5               \n"
+  "palignr    $0x8,%%xmm13,%%xmm13             \n"
+  "movdqa     (%0,%3),%%xmm7                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm6,%%xmm14                   \n"
+  "punpcklbw  %%xmm7,%%xmm6                    \n"
+  "punpckhbw  %%xmm7,%%xmm14                   \n"
+  "neg        %3                               \n"
+  "movdqa     %%xmm6,%%xmm7                    \n"
+  "movdqa     %%xmm14,%%xmm15                  \n"
+  "lea        0x10(%0,%3,8),%0                 \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  "neg        %3                               \n"
+   // Second round of bit swap.
+  "punpcklwd  %%xmm2,%%xmm0                    \n"
+  "punpcklwd  %%xmm3,%%xmm1                    \n"
+  "movdqa     %%xmm0,%%xmm2                    \n"
+  "movdqa     %%xmm1,%%xmm3                    \n"
+  "palignr    $0x8,%%xmm2,%%xmm2               \n"
+  "palignr    $0x8,%%xmm3,%%xmm3               \n"
+  "punpcklwd  %%xmm6,%%xmm4                    \n"
+  "punpcklwd  %%xmm7,%%xmm5                    \n"
+  "movdqa     %%xmm4,%%xmm6                    \n"
+  "movdqa     %%xmm5,%%xmm7                    \n"
+  "palignr    $0x8,%%xmm6,%%xmm6               \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "punpcklwd  %%xmm10,%%xmm8                   \n"
+  "punpcklwd  %%xmm11,%%xmm9                   \n"
+  "movdqa     %%xmm8,%%xmm10                   \n"
+  "movdqa     %%xmm9,%%xmm11                   \n"
+  "palignr    $0x8,%%xmm10,%%xmm10             \n"
+  "palignr    $0x8,%%xmm11,%%xmm11             \n"
+  "punpcklwd  %%xmm14,%%xmm12                  \n"
+  "punpcklwd  %%xmm15,%%xmm13                  \n"
+  "movdqa     %%xmm12,%%xmm14                  \n"
+  "movdqa     %%xmm13,%%xmm15                  \n"
+  "palignr    $0x8,%%xmm14,%%xmm14             \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  // Third round of bit swap.
+  // Write to the destination pointer.
+  "punpckldq  %%xmm4,%%xmm0                    \n"
+  "movq       %%xmm0,(%1)                      \n"
+  "movdqa     %%xmm0,%%xmm4                    \n"
+  "palignr    $0x8,%%xmm4,%%xmm4               \n"
+  "movq       %%xmm4,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm6,%%xmm2                    \n"
+  "movdqa     %%xmm2,%%xmm6                    \n"
+  "movq       %%xmm2,(%1)                      \n"
+  "palignr    $0x8,%%xmm6,%%xmm6               \n"
+  "punpckldq  %%xmm5,%%xmm1                    \n"
+  "movq       %%xmm6,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "movdqa     %%xmm1,%%xmm5                    \n"
+  "movq       %%xmm1,(%1)                      \n"
+  "palignr    $0x8,%%xmm5,%%xmm5               \n"
+  "movq       %%xmm5,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm7,%%xmm3                    \n"
+  "movq       %%xmm3,(%1)                      \n"
+  "movdqa     %%xmm3,%%xmm7                    \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "movq       %%xmm7,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm12,%%xmm8                   \n"
+  "movq       %%xmm8,(%1)                      \n"
+  "movdqa     %%xmm8,%%xmm12                   \n"
+  "palignr    $0x8,%%xmm12,%%xmm12             \n"
+  "movq       %%xmm12,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm14,%%xmm10                  \n"
+  "movdqa     %%xmm10,%%xmm14                  \n"
+  "movq       %%xmm10,(%1)                     \n"
+  "palignr    $0x8,%%xmm14,%%xmm14             \n"
+  "punpckldq  %%xmm13,%%xmm9                   \n"
+  "movq       %%xmm14,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "movdqa     %%xmm9,%%xmm13                   \n"
+  "movq       %%xmm9,(%1)                      \n"
+  "palignr    $0x8,%%xmm13,%%xmm13             \n"
+  "movq       %%xmm13,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm15,%%xmm11                  \n"
+  "movq       %%xmm11,(%1)                     \n"
+  "movdqa     %%xmm11,%%xmm15                  \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  "sub        $0x10,%2                         \n"
+  "movq       %%xmm15,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "jg         1b                               \n"
+  : "+r"(src),    // %0
+    "+r"(dst),    // %1
+    "+r"(width)   // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "r"((intptr_t)(dst_stride))   // %4
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+    "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",  "xmm14",  "xmm15"
+);
+}
+
+#define HAS_TRANSPOSE_UVWX8_SSE2
+static void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                                uint8* dst_a, int dst_stride_a,
+                                uint8* dst_b, int dst_stride_b,
+                                int w) {
+  asm volatile (
+  // Read in the data from the source pointer.
+  // First round of bit swap.
+  ".p2align  2                                 \n"
+"1:                                            \n"
+  "movdqa     (%0),%%xmm0                      \n"
+  "movdqa     (%0,%4),%%xmm1                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpcklbw  %%xmm1,%%xmm0                    \n"
+  "punpckhbw  %%xmm1,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm1                    \n"
+  "movdqa     (%0),%%xmm2                      \n"
+  "movdqa     (%0,%4),%%xmm3                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm2,%%xmm8                    \n"
+  "punpcklbw  %%xmm3,%%xmm2                    \n"
+  "punpckhbw  %%xmm3,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm3                    \n"
+  "movdqa     (%0),%%xmm4                      \n"
+  "movdqa     (%0,%4),%%xmm5                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm4,%%xmm8                    \n"
+  "punpcklbw  %%xmm5,%%xmm4                    \n"
+  "punpckhbw  %%xmm5,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm5                    \n"
+  "movdqa     (%0),%%xmm6                      \n"
+  "movdqa     (%0,%4),%%xmm7                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm6,%%xmm8                    \n"
+  "punpcklbw  %%xmm7,%%xmm6                    \n"
+  "neg        %4                               \n"
+  "lea        0x10(%0,%4,8),%0                 \n"
+  "punpckhbw  %%xmm7,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm7                    \n"
+  "neg        %4                               \n"
+   // Second round of bit swap.
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "movdqa     %%xmm1,%%xmm9                    \n"
+  "punpckhwd  %%xmm2,%%xmm8                    \n"
+  "punpckhwd  %%xmm3,%%xmm9                    \n"
+  "punpcklwd  %%xmm2,%%xmm0                    \n"
+  "punpcklwd  %%xmm3,%%xmm1                    \n"
+  "movdqa     %%xmm8,%%xmm2                    \n"
+  "movdqa     %%xmm9,%%xmm3                    \n"
+  "movdqa     %%xmm4,%%xmm8                    \n"
+  "movdqa     %%xmm5,%%xmm9                    \n"
+  "punpckhwd  %%xmm6,%%xmm8                    \n"
+  "punpckhwd  %%xmm7,%%xmm9                    \n"
+  "punpcklwd  %%xmm6,%%xmm4                    \n"
+  "punpcklwd  %%xmm7,%%xmm5                    \n"
+  "movdqa     %%xmm8,%%xmm6                    \n"
+  "movdqa     %%xmm9,%%xmm7                    \n"
+  // Third round of bit swap.
+  // Write to the destination pointer.
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpckldq  %%xmm4,%%xmm0                    \n"
+  "movlpd     %%xmm0,(%1)                      \n"  // Write back U channel
+  "movhpd     %%xmm0,(%2)                      \n"  // Write back V channel
+  "punpckhdq  %%xmm4,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm2,%%xmm8                    \n"
+  "punpckldq  %%xmm6,%%xmm2                    \n"
+  "movlpd     %%xmm2,(%1)                      \n"
+  "movhpd     %%xmm2,(%2)                      \n"
+  "punpckhdq  %%xmm6,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm1,%%xmm8                    \n"
+  "punpckldq  %%xmm5,%%xmm1                    \n"
+  "movlpd     %%xmm1,(%1)                      \n"
+  "movhpd     %%xmm1,(%2)                      \n"
+  "punpckhdq  %%xmm5,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm3,%%xmm8                    \n"
+  "punpckldq  %%xmm7,%%xmm3                    \n"
+  "movlpd     %%xmm3,(%1)                      \n"
+  "movhpd     %%xmm3,(%2)                      \n"
+  "punpckhdq  %%xmm7,%%xmm8                    \n"
+  "sub        $0x8,%3                          \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "jg         1b                               \n"
+  : "+r"(src),    // %0
+    "+r"(dst_a),  // %1
+    "+r"(dst_b),  // %2
+    "+r"(w)   // %3
+  : "r"((intptr_t)(src_stride)),    // %4
+    "r"((intptr_t)(dst_stride_a)),  // %5
+    "r"((intptr_t)(dst_stride_b))   // %6
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+    "xmm8", "xmm9"
+);
+}
+#endif
+#endif
+
+static void TransposeWx8_C(const uint8* src, int src_stride,
+                           uint8* dst, int dst_stride,
+                           int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst[0] = src[0 * src_stride];
+    dst[1] = src[1 * src_stride];
+    dst[2] = src[2 * src_stride];
+    dst[3] = src[3 * src_stride];
+    dst[4] = src[4 * src_stride];
+    dst[5] = src[5 * src_stride];
+    dst[6] = src[6 * src_stride];
+    dst[7] = src[7 * src_stride];
+    ++src;
+    dst += dst_stride;
+  }
+}
+
+static void TransposeWxH_C(const uint8* src, int src_stride,
+                           uint8* dst, int dst_stride,
+                           int width, int height) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst[i * dst_stride + j] = src[j * src_stride + i];
+    }
+  }
+}
+
+LIBYUV_API
+void TransposePlane(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  int i = height;
+  void (*TransposeWx8)(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride,
+                       int width) = TransposeWx8_C;
+#if defined(HAS_TRANSPOSE_WX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeWx8 = TransposeWx8_NEON;
+  }
+#endif
+#if defined(HAS_TRANSPOSE_WX8_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    TransposeWx8 = TransposeWx8_SSSE3;
+  }
+#endif
+#if defined(HAS_TRANSPOSE_WX8_FAST_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) &&
+      IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+    TransposeWx8 = TransposeWx8_FAST_SSSE3;
+  }
+#endif
+#if defined(HAS_TRANSPOSE_WX8_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    if (IS_ALIGNED(width, 4) &&
+        IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+      TransposeWx8 = TransposeWx8_FAST_MIPS_DSPR2;
+    } else {
+      TransposeWx8 = TransposeWx8_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  // Work across the source in 8x8 tiles
+  while (i >= 8) {
+    TransposeWx8(src, src_stride, dst, dst_stride, width);
+    src += 8 * src_stride;    // Go down 8 rows.
+    dst += 8;                 // Move over 8 columns.
+    i -= 8;
+  }
+
+  TransposeWxH_C(src, src_stride, dst, dst_stride, width, i);
+}
+
+LIBYUV_API
+void RotatePlane90(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height) {
+  // Rotate by 90 is a transpose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Rotate by 270 is a transpose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane180(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    MirrorRow = MirrorRow_NEON;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    MirrorRow = MirrorRow_SSE2;
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    MirrorRow = MirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
+    MirrorRow = MirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_MIRRORROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst, 4) && IS_ALIGNED(dst_stride, 4)) {
+    MirrorRow = MirrorRow_MIPS_DSPR2;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86) && IS_ALIGNED(width, 4)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    MirrorRow(src, row, width);  // Mirror first row into a buffer
+    src += src_stride;
+    MirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    dst += dst_stride;
+    CopyRow(row, dst_bot, width);  // Copy first mirrored row into last
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+static void TransposeUVWx8_C(const uint8* src, int src_stride,
+                             uint8* dst_a, int dst_stride_a,
+                             uint8* dst_b, int dst_stride_b,
+                             int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst_a[0] = src[0 * src_stride + 0];
+    dst_b[0] = src[0 * src_stride + 1];
+    dst_a[1] = src[1 * src_stride + 0];
+    dst_b[1] = src[1 * src_stride + 1];
+    dst_a[2] = src[2 * src_stride + 0];
+    dst_b[2] = src[2 * src_stride + 1];
+    dst_a[3] = src[3 * src_stride + 0];
+    dst_b[3] = src[3 * src_stride + 1];
+    dst_a[4] = src[4 * src_stride + 0];
+    dst_b[4] = src[4 * src_stride + 1];
+    dst_a[5] = src[5 * src_stride + 0];
+    dst_b[5] = src[5 * src_stride + 1];
+    dst_a[6] = src[6 * src_stride + 0];
+    dst_b[6] = src[6 * src_stride + 1];
+    dst_a[7] = src[7 * src_stride + 0];
+    dst_b[7] = src[7 * src_stride + 1];
+    src += 2;
+    dst_a += dst_stride_a;
+    dst_b += dst_stride_b;
+  }
+}
+
+static void TransposeUVWxH_C(const uint8* src, int src_stride,
+                             uint8* dst_a, int dst_stride_a,
+                             uint8* dst_b, int dst_stride_b,
+                             int width, int height) {
+  int i;
+  for (i = 0; i < width * 2; i += 2) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst_a[j + ((i >> 1) * dst_stride_a)] = src[i + (j * src_stride)];
+      dst_b[j + ((i >> 1) * dst_stride_b)] = src[i + (j * src_stride) + 1];
+    }
+  }
+}
+
+LIBYUV_API
+void TransposeUV(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  int i = height;
+  void (*TransposeUVWx8)(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) = TransposeUVWx8_C;
+#if defined(HAS_TRANSPOSE_UVWX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeUVWx8 = TransposeUVWx8_NEON;
+  }
+#elif defined(HAS_TRANSPOSE_UVWX8_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(width, 8) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+    TransposeUVWx8 = TransposeUVWx8_SSE2;
+  }
+#elif defined(HAS_TRANSPOSE_UVWx8_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+    TransposeUVWx8 = TransposeUVWx8_MIPS_DSPR2;
+  }
+#endif
+
+  // Work through the source in 8x8 tiles.
+  while (i >= 8) {
+    TransposeUVWx8(src, src_stride,
+                   dst_a, dst_stride_a,
+                   dst_b, dst_stride_b,
+                   width);
+    src += 8 * src_stride;    // Go down 8 rows.
+    dst_a += 8;               // Move over 8 columns.
+    dst_b += 8;               // Move over 8 columns.
+    i -= 8;
+  }
+
+  TransposeUVWxH_C(src, src_stride,
+                   dst_a, dst_stride_a,
+                   dst_b, dst_stride_b,
+                   width, i);
+}
+
+LIBYUV_API
+void RotateUV90(const uint8* src, int src_stride,
+                uint8* dst_a, int dst_stride_a,
+                uint8* dst_b, int dst_stride_b,
+                int width, int height) {
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+
+  TransposeUV(src, src_stride,
+              dst_a, dst_stride_a,
+              dst_b, dst_stride_b,
+              width, height);
+}
+
+LIBYUV_API
+void RotateUV270(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  dst_a += dst_stride_a * (width - 1);
+  dst_b += dst_stride_b * (width - 1);
+  dst_stride_a = -dst_stride_a;
+  dst_stride_b = -dst_stride_b;
+
+  TransposeUV(src, src_stride,
+              dst_a, dst_stride_a,
+              dst_b, dst_stride_b,
+              width, height);
+}
+
+// Rotate 180 is a horizontal and vertical flip.
+LIBYUV_API
+void RotateUV180(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  int i;
+  void (*MirrorRowUV)(const uint8* src, uint8* dst_u, uint8* dst_v, int width) =
+      MirrorUVRow_C;
+#if defined(HAS_MIRRORUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    MirrorRowUV = MirrorUVRow_NEON;
+  }
+#elif defined(HAS_MIRRORROW_UV_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16)) {
+    MirrorRowUV = MirrorUVRow_SSSE3;
+  }
+#elif defined(HAS_MIRRORUVROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+    MirrorRowUV = MirrorUVRow_MIPS_DSPR2;
+  }
+#endif
+
+  dst_a += dst_stride_a * (height - 1);
+  dst_b += dst_stride_b * (height - 1);
+
+  for (i = 0; i < height; ++i) {
+    MirrorRowUV(src, dst_a, dst_b, width);
+    src += src_stride;
+    dst_a -= dst_stride_a;
+    dst_b -= dst_stride_b;
+  }
+}
+
+LIBYUV_API
+int RotatePlane(const uint8* src, int src_stride,
+                uint8* dst, int dst_stride,
+                int width, int height,
+                enum RotationMode mode) {
+  if (!src || width <= 0 || height == 0 || !dst) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane(src, src_stride,
+                dst, dst_stride,
+                width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90(src, src_stride,
+                    dst, dst_stride,
+                    width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src, src_stride,
+                     dst, dst_stride,
+                     width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src, src_stride,
+                     dst, dst_stride,
+                     width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int I420Rotate(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return I420Copy(src_y, src_stride_y,
+                      src_u, src_stride_u,
+                      src_v, src_stride_v,
+                      dst_y, dst_stride_y,
+                      dst_u, dst_stride_u,
+                      dst_v, dst_stride_v,
+                      width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y,
+                    dst_y, dst_stride_y,
+                    width, height);
+      RotatePlane90(src_u, src_stride_u,
+                    dst_u, dst_stride_u,
+                    halfwidth, halfheight);
+      RotatePlane90(src_v, src_stride_v,
+                    dst_v, dst_stride_v,
+                    halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotatePlane270(src_u, src_stride_u,
+                     dst_u, dst_stride_u,
+                     halfwidth, halfheight);
+      RotatePlane270(src_v, src_stride_v,
+                     dst_v, dst_stride_v,
+                     halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotatePlane180(src_u, src_stride_u,
+                     dst_u, dst_stride_u,
+                     halfwidth, halfheight);
+      RotatePlane180(src_v, src_stride_v,
+                     dst_v, dst_stride_v,
+                     halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
+                     const uint8* src_uv, int src_stride_uv,
+                     uint8* dst_y, int dst_stride_y,
+                     uint8* dst_u, int dst_stride_u,
+                     uint8* dst_v, int dst_stride_v,
+                     int width, int height,
+                     enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_uv || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return NV12ToI420(src_y, src_stride_y,
+                        src_uv, src_stride_uv,
+                        dst_y, dst_stride_y,
+                        dst_u, dst_stride_u,
+                        dst_v, dst_stride_v,
+                        width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y,
+                    dst_y, dst_stride_y,
+                    width, height);
+      RotateUV90(src_uv, src_stride_uv,
+                 dst_u, dst_stride_u,
+                 dst_v, dst_stride_v,
+                 halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotateUV270(src_uv, src_stride_uv,
+                  dst_u, dst_stride_u,
+                  dst_v, dst_stride_v,
+                  halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotateUV180(src_uv, src_stride_uv,
+                  dst_u, dst_stride_u,
+                  dst_v, dst_stride_v,
+                  halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/rotate_argb.cc b/drivers/theoraplayer/src/YUV/libyuv/src/rotate_argb.cc
new file mode 100755
index 0000000000..ab0f9ce070
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/rotate_argb.cc
@@ -0,0 +1,209 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/convert.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGBScale has a function to copy pixels to a row, striding each source
+// pixel by a constant.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || \
+    (defined(__x86_64__) && !defined(__native_client__)) || defined(__i386__))
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+void ScaleARGBRowDownEven_SSE2(const uint8* src_ptr, int src_stride,
+                               int src_stepx,
+                               uint8* dst_ptr, int dst_width);
+#endif
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+void ScaleARGBRowDownEven_NEON(const uint8* src_ptr, int src_stride,
+                               int src_stepx,
+                               uint8* dst_ptr, int dst_width);
+#endif
+
+void ScaleARGBRowDownEven_C(const uint8* src_ptr, int,
+                            int src_stepx,
+                            uint8* dst_ptr, int dst_width);
+
+static void ARGBTranspose(const uint8* src, int src_stride,
+                          uint8* dst, int dst_stride,
+                          int width, int height) {
+  int i;
+  int src_pixel_step = src_stride >> 2;
+  void (*ScaleARGBRowDownEven)(const uint8* src_ptr, int src_stride,
+      int src_step, uint8* dst_ptr, int dst_width) = ScaleARGBRowDownEven_C;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4) &&  // Width of dest.
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2;
+  }
+#elif defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4) &&  // Width of dest.
+      IS_ALIGNED(src, 4)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON;
+  }
+#endif
+
+  for (i = 0; i < width; ++i) {  // column of source to row of dest.
+    ScaleARGBRowDownEven(src, 0, src_pixel_step, dst, height);
+    dst += dst_stride;
+    src += 4;
+  }
+}
+
+void ARGBRotate90(const uint8* src, int src_stride,
+                  uint8* dst, int dst_stride,
+                  int width, int height) {
+  // Rotate by 90 is a ARGBTranspose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Rotate by 270 is a ARGBTranspose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate180(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width * 4);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_ARGBMIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ARGBMirrorRow = ARGBMirrorRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
+    ARGBMirrorRow = ARGBMirrorRow_AVX2;
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 4)) {
+    ARGBMirrorRow = ARGBMirrorRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width * 4, 32)) {
+    CopyRow = CopyRow_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    CopyRow = CopyRow_X86;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width * 4, 32) &&
+      IS_ALIGNED(src, 16) && IS_ALIGNED(src_stride, 16) &&
+      IS_ALIGNED(dst, 16) && IS_ALIGNED(dst_stride, 16)) {
+    CopyRow = CopyRow_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    ARGBMirrorRow(src, row, width);  // Mirror first row into a buffer
+    ARGBMirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    CopyRow(row, dst_bot, width * 4);  // Copy first mirrored row into last
+    src += src_stride;
+    dst += dst_stride;
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height,
+               enum RotationMode mode) {
+  if (!src_argb || width <= 0 || height == 0 || !dst_argb) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return ARGBCopy(src_argb, src_stride_argb,
+                      dst_argb, dst_stride_argb,
+                      width, height);
+    case kRotate90:
+      ARGBRotate90(src_argb, src_stride_argb,
+                   dst_argb, dst_stride_argb,
+                   width, height);
+      return 0;
+    case kRotate270:
+      ARGBRotate270(src_argb, src_stride_argb,
+                    dst_argb, dst_stride_argb,
+                    width, height);
+      return 0;
+    case kRotate180:
+      ARGBRotate180(src_argb, src_stride_argb,
+                    dst_argb, dst_stride_argb,
+                    width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/rotate_mips.cc b/drivers/theoraplayer/src/YUV/libyuv/src/rotate_mips.cc
new file mode 100755
index 0000000000..04d5a663f7
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/rotate_mips.cc
@@ -0,0 +1,486 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride,
+                             int width) {
+   __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "sll              $t2, %[src_stride], 0x1          \n" // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n" // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n" // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+//dst + dst_stride word aligned
+    "1:                                                  \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "sw               $s0, 0(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "sw               $s1, 4(%[dst])                   \n"
+      "bnez             %[width], 1b                     \n"
+      " addu            %[dst], %[dst], %[dst_stride]    \n"
+      "b                2f                               \n"
+//dst + dst_stride unaligned
+   "11:                                                  \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "swr              $s0, 0(%[dst])                   \n"
+      "swl              $s0, 3(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "swr              $s1, 4(%[dst])                   \n"
+      "swl              $s1, 7(%[dst])                   \n"
+      "bnez             %[width], 11b                    \n"
+       "addu             %[dst], %[dst], %[dst_stride]   \n"
+    "2:                                                  \n"
+      ".set pop                                          \n"
+      :[src] "+r" (src),
+       [dst] "+r" (dst),
+       [width] "+r" (width)
+      :[src_stride] "r" (src_stride),
+       [dst_stride] "r" (dst_stride)
+      : "t0", "t1",  "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1"
+  );
+}
+
+void TransposeWx8_FAST_MIPS_DSPR2(const uint8* src, int src_stride,
+                                  uint8* dst, int dst_stride,
+                                  int width) {
+  __asm__ __volatile__ (
+      ".set noat                                         \n"
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz             %[width], 2f                     \n"
+      " sll             $t2, %[src_stride], 0x1          \n"  // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+
+      "srl              $AT, %[width], 0x2               \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+//dst + dst_stride word aligned
+      "1:                                                \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+// t0 = | 30 | 20 | 10 | 00 |
+// t1 = | 31 | 21 | 11 | 01 |
+// t8 = | 32 | 22 | 12 | 02 |
+// t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 21 | 01 | 20 | 00 |
+  // s1 = | 23 | 03 | 22 | 02 |
+  // s2 = | 31 | 11 | 30 | 10 |
+  // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+  // s4 = | 03 | 02 | 01 | 00 |
+  // s5 = | 23 | 22 | 21 | 20 |
+  // s6 = | 13 | 12 | 11 | 10 |
+  // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+// t0 = | 34 | 24 | 14 | 04 |
+// t1 = | 35 | 25 | 15 | 05 |
+// t8 = | 36 | 26 | 16 | 06 |
+// t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 25 | 05 | 24 | 04 |
+  // s1 = | 27 | 07 | 26 | 06 |
+  // s2 = | 35 | 15 | 34 | 14 |
+  // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+  // t0 = | 07 | 06 | 05 | 04 |
+  // t1 = | 27 | 26 | 25 | 24 |
+  // t8 = | 17 | 16 | 15 | 14 |
+  // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "sw              $s4, 0(%[dst])                    \n"
+      "sw              $t0, 4(%[dst])                    \n"
+      "sw              $s6, 0($s0)                       \n"
+      "sw              $t8, 4($s0)                       \n"
+      "sw              $s5, 0($s1)                       \n"
+      "sw              $t1, 4($s1)                       \n"
+      "sw              $s7, 0($s2)                       \n"
+      "sw              $t9, 4($s2)                       \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 1b                          \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "b                2f                               \n"
+//dst + dst_stride unaligned
+      "11:                                               \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+// t0 = | 30 | 20 | 10 | 00 |
+// t1 = | 31 | 21 | 11 | 01 |
+// t8 = | 32 | 22 | 12 | 02 |
+// t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 21 | 01 | 20 | 00 |
+  // s1 = | 23 | 03 | 22 | 02 |
+  // s2 = | 31 | 11 | 30 | 10 |
+  // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+  // s4 = | 03 | 02 | 01 | 00 |
+  // s5 = | 23 | 22 | 21 | 20 |
+  // s6 = | 13 | 12 | 11 | 10 |
+  // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+// t0 = | 34 | 24 | 14 | 04 |
+// t1 = | 35 | 25 | 15 | 05 |
+// t8 = | 36 | 26 | 16 | 06 |
+// t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 25 | 05 | 24 | 04 |
+  // s1 = | 27 | 07 | 26 | 06 |
+  // s2 = | 35 | 15 | 34 | 14 |
+  // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+  // t0 = | 07 | 06 | 05 | 04 |
+  // t1 = | 27 | 26 | 25 | 24 |
+  // t8 = | 17 | 16 | 15 | 14 |
+  // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "swr              $s4, 0(%[dst])                   \n"
+      "swl              $s4, 3(%[dst])                   \n"
+      "swr              $t0, 4(%[dst])                   \n"
+      "swl              $t0, 7(%[dst])                   \n"
+      "swr              $s6, 0($s0)                      \n"
+      "swl              $s6, 3($s0)                      \n"
+      "swr              $t8, 4($s0)                      \n"
+      "swl              $t8, 7($s0)                      \n"
+      "swr              $s5, 0($s1)                      \n"
+      "swl              $s5, 3($s1)                      \n"
+      "swr              $t1, 4($s1)                      \n"
+      "swl              $t1, 7($s1)                      \n"
+      "swr              $s7, 0($s2)                      \n"
+      "swl              $s7, 3($s2)                      \n"
+      "swr              $t9, 4($s2)                      \n"
+      "swl              $t9, 7($s2)                      \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 11b                         \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "2:                                                \n"
+      ".set pop                                          \n"
+      ".set at                                           \n"
+      :[src] "+r" (src),
+       [dst] "+r" (dst),
+       [width] "+r" (width)
+      :[src_stride] "r" (src_stride),
+       [dst_stride] "r" (dst_stride)
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1", "s2", "s3", "s4",
+        "s5", "s6", "s7"
+  );
+}
+
+void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                               uint8* dst_a, int dst_stride_a,
+                               uint8* dst_b, int dst_stride_b,
+                               int width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz            %[width], 2f                      \n"
+      " sll            $t2, %[src_stride], 0x1           \n" // src_stride x 2
+      "sll             $t4, %[src_stride], 0x2           \n" // src_stride x 4
+      "sll             $t9, %[src_stride], 0x3           \n" // src_stride x 8
+      "addu            $t3, $t2, %[src_stride]           \n"
+      "addu            $t5, $t4, %[src_stride]           \n"
+      "addu            $t6, $t2, $t4                     \n"
+      "subu            $t7, $t9, %[src_stride]           \n"
+      "srl             $t1, %[width], 1                  \n"
+
+// check word aligment for dst_a, dst_b, dst_stride_a and dst_stride_b
+      "andi            $t0, %[dst_a], 0x3                \n"
+      "andi            $t8, %[dst_b], 0x3                \n"
+      "or              $t0, $t0, $t8                     \n"
+      "andi            $t8, %[dst_stride_a], 0x3         \n"
+      "andi            $s5, %[dst_stride_b], 0x3         \n"
+      "or              $t8, $t8, $s5                     \n"
+      "or              $t0, $t0, $t8                     \n"
+      "bnez            $t0, 11f                          \n"
+      " nop                                              \n"
+// dst + dst_stride word aligned (both, a & b dst addresses)
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
+
+      "sw              $s3, 0($s5)                       \n"
+      "sw              $s4, 0($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
+      "sw              $s3, 0(%[dst_a])                  \n"
+      "sw              $s4, 0(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
+      "sw              $s3, 4($s5)                       \n"
+      "sw              $s4, 4($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "sw              $s3, 4(%[dst_a])                  \n"
+      "sw              $s4, 4(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 1b                           \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+      "b               2f                                \n"
+      " nop                                              \n"
+
+// dst_a or dst_b or dst_stride_a or dst_stride_b not word aligned
+   "11:                                                  \n"
+      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
+
+      "swr             $s3, 0($s5)                       \n"
+      "swl             $s3, 3($s5)                       \n"
+      "swr             $s4, 0($s6)                       \n"
+      "swl             $s4, 3($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
+      "swr             $s3, 0(%[dst_a])                  \n"
+      "swl             $s3, 3(%[dst_a])                  \n"
+      "swr             $s4, 0(%[dst_b])                  \n"
+      "swl             $s4, 3(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
+
+      "swr             $s3, 4($s5)                       \n"
+      "swl             $s3, 7($s5)                       \n"
+      "swr             $s4, 4($s6)                       \n"
+      "swl             $s4, 7($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "swr             $s3, 4(%[dst_a])                  \n"
+      "swl             $s3, 7(%[dst_a])                  \n"
+      "swr             $s4, 4(%[dst_b])                  \n"
+      "swl             $s4, 7(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 11b                          \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+
+      "2:                                                \n"
+      ".set pop                                          \n"
+      : [src] "+r" (src),
+        [dst_a] "+r" (dst_a),
+        [dst_b] "+r" (dst_b),
+        [width] "+r" (width),
+        [src_stride] "+r" (src_stride)
+      : [dst_stride_a] "r" (dst_stride_a),
+        [dst_stride_b] "r" (dst_stride_b)
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1", "s2", "s3",
+        "s4", "s5", "s6"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/rotate_neon.cc b/drivers/theoraplayer/src/YUV/libyuv/src/rotate_neon.cc
new file mode 100755
index 0000000000..274c4109cd
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/rotate_neon.cc
@@ -0,0 +1,412 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+static uvec8 kVTbl4x4Transpose =
+  { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 };
+
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride,
+                       int width) {
+  const uint8* src_temp = NULL;
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+#ifdef _ANDROID
+				".fpu neon\n"
+#endif
+    "sub         %5, #8                        \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    ".p2align  2                               \n"
+    "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      "vld1.8      {d0}, [%0], %2              \n"
+      "vld1.8      {d1}, [%0], %2              \n"
+      "vld1.8      {d2}, [%0], %2              \n"
+      "vld1.8      {d3}, [%0], %2              \n"
+      "vld1.8      {d4}, [%0], %2              \n"
+      "vld1.8      {d5}, [%0], %2              \n"
+      "vld1.8      {d6}, [%0], %2              \n"
+      "vld1.8      {d7}, [%0]                  \n"
+
+      "vtrn.8      d1, d0                      \n"
+      "vtrn.8      d3, d2                      \n"
+      "vtrn.8      d5, d4                      \n"
+      "vtrn.8      d7, d6                      \n"
+
+      "vtrn.16     d1, d3                      \n"
+      "vtrn.16     d0, d2                      \n"
+      "vtrn.16     d5, d7                      \n"
+      "vtrn.16     d4, d6                      \n"
+
+      "vtrn.32     d1, d5                      \n"
+      "vtrn.32     d0, d4                      \n"
+      "vtrn.32     d3, d7                      \n"
+      "vtrn.32     d2, d6                      \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+
+      "mov         %0, %3                      \n"
+
+      "vst1.8      {d1}, [%0], %4              \n"
+      "vst1.8      {d0}, [%0], %4              \n"
+      "vst1.8      {d3}, [%0], %4              \n"
+      "vst1.8      {d2}, [%0], %4              \n"
+      "vst1.8      {d5}, [%0], %4              \n"
+      "vst1.8      {d4}, [%0], %4              \n"
+      "vst1.8      {d7}, [%0], %4              \n"
+      "vst1.8      {d6}, [%0]                  \n"
+
+      "add         %1, #8                      \n"  // src += 8
+      "add         %3, %3, %4, lsl #3          \n"  // dst += 8 * dst_stride
+      "subs        %5,  #8                     \n"  // w   -= 8
+      "bge         1b                          \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %5, #8                        \n"
+    "beq         4f                            \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %5, #2                        \n"
+    "blt         3f                            \n"
+
+    "cmp         %5, #4                        \n"
+    "blt         2f                            \n"
+
+    // 4x8 block
+    "mov         %0, %1                        \n"
+    "vld1.32     {d0[0]}, [%0], %2             \n"
+    "vld1.32     {d0[1]}, [%0], %2             \n"
+    "vld1.32     {d1[0]}, [%0], %2             \n"
+    "vld1.32     {d1[1]}, [%0], %2             \n"
+    "vld1.32     {d2[0]}, [%0], %2             \n"
+    "vld1.32     {d2[1]}, [%0], %2             \n"
+    "vld1.32     {d3[0]}, [%0], %2             \n"
+    "vld1.32     {d3[1]}, [%0]                 \n"
+
+    "mov         %0, %3                        \n"
+
+    "vld1.8      {q3}, [%6]                    \n"
+
+    "vtbl.8      d4, {d0, d1}, d6              \n"
+    "vtbl.8      d5, {d0, d1}, d7              \n"
+    "vtbl.8      d0, {d2, d3}, d6              \n"
+    "vtbl.8      d1, {d2, d3}, d7              \n"
+
+    // TODO(frkoenig): Rework shuffle above to
+    // write out with 4 instead of 8 writes.
+    "vst1.32     {d4[0]}, [%0], %4             \n"
+    "vst1.32     {d4[1]}, [%0], %4             \n"
+    "vst1.32     {d5[0]}, [%0], %4             \n"
+    "vst1.32     {d5[1]}, [%0]                 \n"
+
+    "add         %0, %3, #4                    \n"
+    "vst1.32     {d0[0]}, [%0], %4             \n"
+    "vst1.32     {d0[1]}, [%0], %4             \n"
+    "vst1.32     {d1[0]}, [%0], %4             \n"
+    "vst1.32     {d1[1]}, [%0]                 \n"
+
+    "add         %1, #4                        \n"  // src += 4
+    "add         %3, %3, %4, lsl #2            \n"  // dst += 4 * dst_stride
+    "subs        %5,  #4                       \n"  // w   -= 4
+    "beq         4f                            \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %5, #2                        \n"
+    "blt         3f                            \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov         %0, %1                        \n"
+    "vld1.16     {d0[0]}, [%0], %2             \n"
+    "vld1.16     {d1[0]}, [%0], %2             \n"
+    "vld1.16     {d0[1]}, [%0], %2             \n"
+    "vld1.16     {d1[1]}, [%0], %2             \n"
+    "vld1.16     {d0[2]}, [%0], %2             \n"
+    "vld1.16     {d1[2]}, [%0], %2             \n"
+    "vld1.16     {d0[3]}, [%0], %2             \n"
+    "vld1.16     {d1[3]}, [%0]                 \n"
+
+    "vtrn.8      d0, d1                        \n"
+
+    "mov         %0, %3                        \n"
+
+    "vst1.64     {d0}, [%0], %4                \n"
+    "vst1.64     {d1}, [%0]                    \n"
+
+    "add         %1, #2                        \n"  // src += 2
+    "add         %3, %3, %4, lsl #1            \n"  // dst += 2 * dst_stride
+    "subs        %5,  #2                       \n"  // w   -= 2
+    "beq         4f                            \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    "vld1.8      {d0[0]}, [%1], %2             \n"
+    "vld1.8      {d0[1]}, [%1], %2             \n"
+    "vld1.8      {d0[2]}, [%1], %2             \n"
+    "vld1.8      {d0[3]}, [%1], %2             \n"
+    "vld1.8      {d0[4]}, [%1], %2             \n"
+    "vld1.8      {d0[5]}, [%1], %2             \n"
+    "vld1.8      {d0[6]}, [%1], %2             \n"
+    "vld1.8      {d0[7]}, [%1]                 \n"
+
+    "vst1.64     {d0}, [%3]                    \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),          // %0
+      "+r"(src),               // %1
+      "+r"(src_stride),        // %2
+      "+r"(dst),               // %3
+      "+r"(dst_stride),        // %4
+      "+r"(width)              // %5
+    : "r"(&kVTbl4x4Transpose)  // %6
+    : "memory", "cc", "q0", "q1", "q2", "q3"
+  );
+}
+
+static uvec8 kVTbl4x4TransposeDi =
+  { 0,  8,  1,  9,  2, 10,  3, 11,  4, 12,  5, 13,  6, 14,  7, 15 };
+
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) {
+  const uint8* src_temp = NULL;
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %7, #8                        \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    ".p2align  2                               \n"
+    "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      "vld2.8      {d0,  d1},  [%0], %2        \n"
+      "vld2.8      {d2,  d3},  [%0], %2        \n"
+      "vld2.8      {d4,  d5},  [%0], %2        \n"
+      "vld2.8      {d6,  d7},  [%0], %2        \n"
+      "vld2.8      {d16, d17}, [%0], %2        \n"
+      "vld2.8      {d18, d19}, [%0], %2        \n"
+      "vld2.8      {d20, d21}, [%0], %2        \n"
+      "vld2.8      {d22, d23}, [%0]            \n"
+
+      "vtrn.8      q1, q0                      \n"
+      "vtrn.8      q3, q2                      \n"
+      "vtrn.8      q9, q8                      \n"
+      "vtrn.8      q11, q10                    \n"
+
+      "vtrn.16     q1, q3                      \n"
+      "vtrn.16     q0, q2                      \n"
+      "vtrn.16     q9, q11                     \n"
+      "vtrn.16     q8, q10                     \n"
+
+      "vtrn.32     q1, q9                      \n"
+      "vtrn.32     q0, q8                      \n"
+      "vtrn.32     q3, q11                     \n"
+      "vtrn.32     q2, q10                     \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+      "vrev16.8    q8, q8                      \n"
+      "vrev16.8    q9, q9                      \n"
+      "vrev16.8    q10, q10                    \n"
+      "vrev16.8    q11, q11                    \n"
+
+      "mov         %0, %3                      \n"
+
+      "vst1.8      {d2},  [%0], %4             \n"
+      "vst1.8      {d0},  [%0], %4             \n"
+      "vst1.8      {d6},  [%0], %4             \n"
+      "vst1.8      {d4},  [%0], %4             \n"
+      "vst1.8      {d18}, [%0], %4             \n"
+      "vst1.8      {d16}, [%0], %4             \n"
+      "vst1.8      {d22}, [%0], %4             \n"
+      "vst1.8      {d20}, [%0]                 \n"
+
+      "mov         %0, %5                      \n"
+
+      "vst1.8      {d3},  [%0], %6             \n"
+      "vst1.8      {d1},  [%0], %6             \n"
+      "vst1.8      {d7},  [%0], %6             \n"
+      "vst1.8      {d5},  [%0], %6             \n"
+      "vst1.8      {d19}, [%0], %6             \n"
+      "vst1.8      {d17}, [%0], %6             \n"
+      "vst1.8      {d23}, [%0], %6             \n"
+      "vst1.8      {d21}, [%0]                 \n"
+
+      "add         %1, #8*2                    \n"  // src   += 8*2
+      "add         %3, %3, %4, lsl #3          \n"  // dst_a += 8 * dst_stride_a
+      "add         %5, %5, %6, lsl #3          \n"  // dst_b += 8 * dst_stride_b
+      "subs        %7,  #8                     \n"  // w     -= 8
+      "bge         1b                          \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %7, #8                        \n"
+    "beq         4f                            \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %7, #2                        \n"
+    "blt         3f                            \n"
+
+    "cmp         %7, #4                        \n"
+    "blt         2f                            \n"
+
+    //TODO(frkoenig): Clean this up
+    // 4x8 block
+    "mov         %0, %1                        \n"
+    "vld1.64     {d0}, [%0], %2                \n"
+    "vld1.64     {d1}, [%0], %2                \n"
+    "vld1.64     {d2}, [%0], %2                \n"
+    "vld1.64     {d3}, [%0], %2                \n"
+    "vld1.64     {d4}, [%0], %2                \n"
+    "vld1.64     {d5}, [%0], %2                \n"
+    "vld1.64     {d6}, [%0], %2                \n"
+    "vld1.64     {d7}, [%0]                    \n"
+
+    "vld1.8      {q15}, [%8]                   \n"
+
+    "vtrn.8      q0, q1                        \n"
+    "vtrn.8      q2, q3                        \n"
+
+    "vtbl.8      d16, {d0, d1}, d30            \n"
+    "vtbl.8      d17, {d0, d1}, d31            \n"
+    "vtbl.8      d18, {d2, d3}, d30            \n"
+    "vtbl.8      d19, {d2, d3}, d31            \n"
+    "vtbl.8      d20, {d4, d5}, d30            \n"
+    "vtbl.8      d21, {d4, d5}, d31            \n"
+    "vtbl.8      d22, {d6, d7}, d30            \n"
+    "vtbl.8      d23, {d6, d7}, d31            \n"
+
+    "mov         %0, %3                        \n"
+
+    "vst1.32     {d16[0]},  [%0], %4           \n"
+    "vst1.32     {d16[1]},  [%0], %4           \n"
+    "vst1.32     {d17[0]},  [%0], %4           \n"
+    "vst1.32     {d17[1]},  [%0], %4           \n"
+
+    "add         %0, %3, #4                    \n"
+    "vst1.32     {d20[0]}, [%0], %4            \n"
+    "vst1.32     {d20[1]}, [%0], %4            \n"
+    "vst1.32     {d21[0]}, [%0], %4            \n"
+    "vst1.32     {d21[1]}, [%0]                \n"
+
+    "mov         %0, %5                        \n"
+
+    "vst1.32     {d18[0]}, [%0], %6            \n"
+    "vst1.32     {d18[1]}, [%0], %6            \n"
+    "vst1.32     {d19[0]}, [%0], %6            \n"
+    "vst1.32     {d19[1]}, [%0], %6            \n"
+
+    "add         %0, %5, #4                    \n"
+    "vst1.32     {d22[0]},  [%0], %6           \n"
+    "vst1.32     {d22[1]},  [%0], %6           \n"
+    "vst1.32     {d23[0]},  [%0], %6           \n"
+    "vst1.32     {d23[1]},  [%0]               \n"
+
+    "add         %1, #4*2                      \n"  // src   += 4 * 2
+    "add         %3, %3, %4, lsl #2            \n"  // dst_a += 4 * dst_stride_a
+    "add         %5, %5, %6, lsl #2            \n"  // dst_b += 4 * dst_stride_b
+    "subs        %7,  #4                       \n"  // w     -= 4
+    "beq         4f                            \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %7, #2                        \n"
+    "blt         3f                            \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov         %0, %1                        \n"
+    "vld2.16     {d0[0], d2[0]}, [%0], %2      \n"
+    "vld2.16     {d1[0], d3[0]}, [%0], %2      \n"
+    "vld2.16     {d0[1], d2[1]}, [%0], %2      \n"
+    "vld2.16     {d1[1], d3[1]}, [%0], %2      \n"
+    "vld2.16     {d0[2], d2[2]}, [%0], %2      \n"
+    "vld2.16     {d1[2], d3[2]}, [%0], %2      \n"
+    "vld2.16     {d0[3], d2[3]}, [%0], %2      \n"
+    "vld2.16     {d1[3], d3[3]}, [%0]          \n"
+
+    "vtrn.8      d0, d1                        \n"
+    "vtrn.8      d2, d3                        \n"
+
+    "mov         %0, %3                        \n"
+
+    "vst1.64     {d0}, [%0], %4                \n"
+    "vst1.64     {d2}, [%0]                    \n"
+
+    "mov         %0, %5                        \n"
+
+    "vst1.64     {d1}, [%0], %6                \n"
+    "vst1.64     {d3}, [%0]                    \n"
+
+    "add         %1, #2*2                      \n"  // src   += 2 * 2
+    "add         %3, %3, %4, lsl #1            \n"  // dst_a += 2 * dst_stride_a
+    "add         %5, %5, %6, lsl #1            \n"  // dst_b += 2 * dst_stride_b
+    "subs        %7,  #2                       \n"  // w     -= 2
+    "beq         4f                            \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    "vld2.8      {d0[0], d1[0]}, [%1], %2      \n"
+    "vld2.8      {d0[1], d1[1]}, [%1], %2      \n"
+    "vld2.8      {d0[2], d1[2]}, [%1], %2      \n"
+    "vld2.8      {d0[3], d1[3]}, [%1], %2      \n"
+    "vld2.8      {d0[4], d1[4]}, [%1], %2      \n"
+    "vld2.8      {d0[5], d1[5]}, [%1], %2      \n"
+    "vld2.8      {d0[6], d1[6]}, [%1], %2      \n"
+    "vld2.8      {d0[7], d1[7]}, [%1]          \n"
+
+    "vst1.64     {d0}, [%3]                    \n"
+    "vst1.64     {d1}, [%5]                    \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),            // %0
+      "+r"(src),                 // %1
+      "+r"(src_stride),          // %2
+      "+r"(dst_a),               // %3
+      "+r"(dst_stride_a),        // %4
+      "+r"(dst_b),               // %5
+      "+r"(dst_stride_b),        // %6
+      "+r"(width)                // %7
+    : "r"(&kVTbl4x4TransposeDi)  // %8
+    : "memory", "cc",
+      "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/row_any.cc b/drivers/theoraplayer/src/YUV/libyuv/src/row_any.cc
new file mode 100755
index 0000000000..90c6a3ff5f
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/row_any.cc
@@ -0,0 +1,542 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// TODO(fbarchard): Consider 'any' functions handling any quantity of pixels.
+// TODO(fbarchard): Consider 'any' functions handling odd alignment.
+// YUV to RGB does multiple of 8 with SIMD and remainder with C.
+#define YANY(NAMEANY, I420TORGB_SIMD, I420TORGB_C, UV_SHIFT, BPP, MASK)        \
+    void NAMEANY(const uint8* y_buf,                                           \
+                 const uint8* u_buf,                                           \
+                 const uint8* v_buf,                                           \
+                 uint8* rgb_buf,                                               \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      I420TORGB_SIMD(y_buf, u_buf, v_buf, rgb_buf, n);                         \
+      I420TORGB_C(y_buf + n,                                                   \
+                  u_buf + (n >> UV_SHIFT),                                     \
+                  v_buf + (n >> UV_SHIFT),                                     \
+                  rgb_buf + n * BPP, width & MASK);                            \
+    }
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_Unaligned_SSSE3, I444ToARGBRow_C,
+     0, 4, 7)
+YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_Unaligned_SSSE3, I422ToARGBRow_C,
+     1, 4, 7)
+YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_Unaligned_SSSE3, I411ToARGBRow_C,
+     2, 4, 7)
+YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_Unaligned_SSSE3, I422ToBGRARow_C,
+     1, 4, 7)
+YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_Unaligned_SSSE3, I422ToABGRRow_C,
+     1, 4, 7)
+YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_Unaligned_SSSE3, I422ToRGBARow_C,
+     1, 4, 7)
+// I422ToRGB565Row_SSSE3 is unaligned.
+YANY(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, I422ToARGB4444Row_C,
+     1, 2, 7)
+YANY(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, I422ToARGB1555Row_C,
+     1, 2, 7)
+YANY(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, I422ToRGB565Row_C,
+     1, 2, 7)
+// I422ToRGB24Row_SSSE3 is unaligned.
+YANY(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, I422ToRGB24Row_C, 1, 3, 7)
+YANY(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, I422ToRAWRow_C, 1, 3, 7)
+YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15)
+YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
+#endif  // HAS_I422TOARGBROW_SSSE3
+#ifdef HAS_I422TOARGBROW_AVX2
+YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
+#endif  // HAS_I422TOARGBROW_AVX2
+#ifdef HAS_I422TOARGBROW_NEON
+YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7)
+YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7)
+YANY(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, I411ToARGBRow_C, 2, 4, 7)
+YANY(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, I422ToBGRARow_C, 1, 4, 7)
+YANY(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, I422ToABGRRow_C, 1, 4, 7)
+YANY(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, I422ToRGBARow_C, 1, 4, 7)
+YANY(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, I422ToRGB24Row_C, 1, 3, 7)
+YANY(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, I422ToRAWRow_C, 1, 3, 7)
+YANY(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, I422ToARGB4444Row_C,
+     1, 2, 7)
+YANY(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, I422ToARGB1555Row_C,
+     1, 2, 7)
+YANY(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, I422ToRGB565Row_C, 1, 2, 7)
+YANY(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, I422ToYUY2Row_C, 1, 2, 15)
+YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15)
+#endif  // HAS_I422TOARGBROW_NEON
+#undef YANY
+
+// Wrappers to handle odd width
+#define NV2NY(NAMEANY, NV12TORGB_SIMD, NV12TORGB_C, UV_SHIFT, BPP)             \
+    void NAMEANY(const uint8* y_buf,                                           \
+                 const uint8* uv_buf,                                          \
+                 uint8* rgb_buf,                                               \
+                 int width) {                                                  \
+      int n = width & ~7;                                                      \
+      NV12TORGB_SIMD(y_buf, uv_buf, rgb_buf, n);                               \
+      NV12TORGB_C(y_buf + n,                                                   \
+                  uv_buf + (n >> UV_SHIFT),                                    \
+                  rgb_buf + n * BPP, width & 7);                               \
+    }
+
+#ifdef HAS_NV12TOARGBROW_SSSE3
+NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_Unaligned_SSSE3, NV12ToARGBRow_C,
+      0, 4)
+NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_Unaligned_SSSE3, NV21ToARGBRow_C,
+      0, 4)
+#endif  // HAS_NV12TOARGBROW_SSSE3
+#ifdef HAS_NV12TOARGBROW_NEON
+NV2NY(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, NV12ToARGBRow_C, 0, 4)
+NV2NY(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, NV21ToARGBRow_C, 0, 4)
+#endif  // HAS_NV12TOARGBROW_NEON
+#ifdef HAS_NV12TORGB565ROW_SSSE3
+NV2NY(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, NV12ToRGB565Row_C,
+      0, 2)
+NV2NY(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, NV21ToRGB565Row_C,
+      0, 2)
+#endif  // HAS_NV12TORGB565ROW_SSSE3
+#ifdef HAS_NV12TORGB565ROW_NEON
+NV2NY(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, NV12ToRGB565Row_C, 0, 2)
+NV2NY(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, NV21ToRGB565Row_C, 0, 2)
+#endif  // HAS_NV12TORGB565ROW_NEON
+#undef NVANY
+
+#define RGBANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP)          \
+    void NAMEANY(const uint8* src,                                             \
+                 uint8* dst,                                                   \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      ARGBTORGB_SIMD(src, dst, n);                                             \
+      ARGBTORGB_C(src + n * SBPP, dst + n * BPP, width & MASK);                \
+    }
+
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+RGBANY(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, ARGBToRGB24Row_C,
+       15, 4, 3)
+RGBANY(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, ARGBToRAWRow_C,
+       15, 4, 3)
+RGBANY(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, ARGBToRGB565Row_C,
+       3, 4, 2)
+RGBANY(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, ARGBToARGB1555Row_C,
+       3, 4, 2)
+RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C,
+       3, 4, 2)
+#endif
+#if defined(HAS_I400TOARGBROW_SSE2)
+RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_Unaligned_SSE2, I400ToARGBRow_C,
+       7, 1, 4)
+#endif
+#if defined(HAS_YTOARGBROW_SSE2)
+RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
+       7, 1, 4)
+RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_Unaligned_SSSE3, YUY2ToARGBRow_C,
+       15, 2, 4)
+RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_Unaligned_SSSE3, UYVYToARGBRow_C,
+       15, 2, 4)
+// These require alignment on ARGB, so C is used for remainder.
+RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C,
+       15, 3, 4)
+RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C,
+       15, 3, 4)
+RGBANY(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, RGB565ToARGBRow_C,
+       7, 2, 4)
+RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
+       7, 2, 4)
+RGBANY(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, ARGB4444ToARGBRow_C,
+       7, 2, 4)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_NEON)
+RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 7, 4, 3)
+RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 7, 4, 3)
+RGBANY(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, ARGBToRGB565Row_C,
+       7, 4, 2)
+RGBANY(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, ARGBToARGB1555Row_C,
+       7, 4, 2)
+RGBANY(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, ARGBToARGB4444Row_C,
+       7, 4, 2)
+RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C,
+       7, 1, 4)
+RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C,
+       7, 1, 4)
+RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C,
+       7, 2, 4)
+RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C,
+       7, 2, 4)
+#endif
+#undef RGBANY
+
+// ARGB to Bayer does multiple of 4 pixels, SSSE3 aligned src, unaligned dst.
+#define BAYERANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP)        \
+    void NAMEANY(const uint8* src,                                             \
+                 uint8* dst, uint32 selector,                                  \
+                 int width) {                                                  \
+      int n = width & ~MASK;                                                   \
+      ARGBTORGB_SIMD(src, dst, selector, n);                                   \
+      ARGBTORGB_C(src + n * SBPP, dst + n * BPP, selector, width & MASK);      \
+    }
+
+#if defined(HAS_ARGBTOBAYERROW_SSSE3)
+BAYERANY(ARGBToBayerRow_Any_SSSE3, ARGBToBayerRow_SSSE3, ARGBToBayerRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERROW_NEON)
+BAYERANY(ARGBToBayerRow_Any_NEON, ARGBToBayerRow_NEON, ARGBToBayerRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_SSE2)
+BAYERANY(ARGBToBayerGGRow_Any_SSE2, ARGBToBayerGGRow_SSE2, ARGBToBayerGGRow_C,
+         7, 4, 1)
+#endif
+#if defined(HAS_ARGBTOBAYERGGROW_NEON)
+BAYERANY(ARGBToBayerGGRow_Any_NEON, ARGBToBayerGGRow_NEON, ARGBToBayerGGRow_C,
+         7, 4, 1)
+#endif
+
+#undef BAYERANY
+
+// RGB/YUV to Y does multiple of 16 with SIMD and last 16 with SIMD.
+#define YANY(NAMEANY, ARGBTOY_SIMD, SBPP, BPP, NUM)                            \
+    void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) {             \
+      ARGBTOY_SIMD(src_argb, dst_y, width - NUM);                              \
+      ARGBTOY_SIMD(src_argb + (width - NUM) * SBPP,                            \
+                   dst_y + (width - NUM) * BPP, NUM);                          \
+    }
+
+#ifdef HAS_ARGBTOYROW_AVX2
+YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 4, 1, 32)
+YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 4, 1, 32)
+YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32)
+YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32)
+#endif
+#ifdef HAS_ARGBTOYROW_SSSE3
+YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_Unaligned_SSSE3, 4, 1, 16)
+#endif
+#ifdef HAS_BGRATOYROW_SSSE3
+YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_Unaligned_SSSE3, 4, 1, 16)
+YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_Unaligned_SSE2, 2, 1, 16)
+YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_Unaligned_SSE2, 2, 1, 16)
+#endif
+#ifdef HAS_ARGBTOYJROW_SSSE3
+YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_Unaligned_SSSE3, 4, 1, 16)
+#endif
+#ifdef HAS_ARGBTOYROW_NEON
+YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 4, 1, 8)
+YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 4, 1, 8)
+YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 4, 1, 8)
+YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 4, 1, 8)
+YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 4, 1, 8)
+YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 3, 1, 8)
+YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, 3, 1, 8)
+YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 2, 1, 8)
+YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 2, 1, 8)
+YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 2, 1, 8)
+YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 2, 1, 16)
+YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 2, 1, 16)
+YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 3, 4, 8)
+YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 3, 4, 8)
+YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 2, 4, 8)
+YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 2, 4, 8)
+YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 2, 4, 8)
+#endif
+#undef YANY
+
+#define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK)                \
+    void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) {             \
+      int n = width & ~MASK;                                                   \
+      ARGBTOY_SIMD(src_argb, dst_y, n);                                        \
+      ARGBTOY_C(src_argb + n * SBPP,                                           \
+                dst_y  + n * BPP, width & MASK);                               \
+    }
+
+// Attenuate is destructive so last16 method can not be used due to overlap.
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+YANY(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, ARGBAttenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+YANY(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, ARGBAttenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+YANY(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, ARGBUnattenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+YANY(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, ARGBAttenuateRow_C,
+     4, 4, 7)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+YANY(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, ARGBUnattenuateRow_C,
+     4, 4, 7)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_NEON
+YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
+     4, 4, 7)
+#endif
+#undef YANY
+
+// RGB/YUV to UV does multiple of 16 with SIMD and remainder with C.
+#define UVANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK)                     \
+    void NAMEANY(const uint8* src_argb, int src_stride_argb,                   \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_argb, src_stride_argb, dst_u, dst_v, n);                \
+      ANYTOUV_C(src_argb  + n * BPP, src_stride_argb,                          \
+                dst_u + (n >> 1),                                              \
+                dst_v + (n >> 1),                                              \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+UVANY(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, ARGBToUVRow_C, 4, 31)
+UVANY(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, YUY2ToUVRow_C, 2, 31)
+UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_Unaligned_SSSE3, ARGBToUVRow_C, 4, 15)
+UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_Unaligned_SSSE3, ARGBToUVJRow_C,
+      4, 15)
+UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_Unaligned_SSSE3, BGRAToUVRow_C, 4, 15)
+UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_Unaligned_SSSE3, ABGRToUVRow_C, 4, 15)
+UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_Unaligned_SSSE3, RGBAToUVRow_C, 4, 15)
+UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_Unaligned_SSE2, YUY2ToUVRow_C, 2, 15)
+UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_Unaligned_SSE2, UYVYToUVRow_C, 2, 15)
+#endif
+#ifdef HAS_ARGBTOUVROW_NEON
+UVANY(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, ARGBToUVRow_C, 4, 15)
+UVANY(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, ARGBToUVJRow_C, 4, 15)
+UVANY(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, BGRAToUVRow_C, 4, 15)
+UVANY(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, ABGRToUVRow_C, 4, 15)
+UVANY(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, RGBAToUVRow_C, 4, 15)
+UVANY(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, RGB24ToUVRow_C, 3, 15)
+UVANY(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, RAWToUVRow_C, 3, 15)
+UVANY(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, RGB565ToUVRow_C, 2, 15)
+UVANY(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, ARGB1555ToUVRow_C, 2, 15)
+UVANY(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, ARGB4444ToUVRow_C, 2, 15)
+UVANY(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, YUY2ToUVRow_C, 2, 15)
+UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15)
+#endif
+#undef UVANY
+
+#define UV422ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK, SHIFT)           \
+    void NAMEANY(const uint8* src_uv,                                          \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                   \
+      ANYTOUV_C(src_uv  + n * BPP,                                             \
+                dst_u + (n >> SHIFT),                                          \
+                dst_v + (n >> SHIFT),                                          \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_ARGBTOUV444ROW_SSSE3
+UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_Unaligned_SSSE3,
+         ARGBToUV444Row_C, 4, 15, 0)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_AVX2
+UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2,
+         YUY2ToUV422Row_C, 2, 31, 1)
+UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2,
+         UYVYToUV422Row_C, 2, 31, 1)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_Unaligned_SSSE3,
+         ARGBToUV422Row_C, 4, 15, 1)
+UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_Unaligned_SSE2,
+         YUY2ToUV422Row_C, 2, 15, 1)
+UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_Unaligned_SSE2,
+         UYVYToUV422Row_C, 2, 15, 1)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_NEON
+UV422ANY(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON,
+         ARGBToUV444Row_C, 4, 7, 0)
+UV422ANY(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON,
+         ARGBToUV422Row_C, 4, 15, 1)
+UV422ANY(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON,
+         ARGBToUV411Row_C, 4, 31, 2)
+UV422ANY(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON,
+         YUY2ToUV422Row_C, 2, 15, 1)
+UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON,
+         UYVYToUV422Row_C, 2, 15, 1)
+#endif
+#undef UV422ANY
+
+#define SPLITUVROWANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK)                  \
+    void NAMEANY(const uint8* src_uv,                                          \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_uv, dst_u, dst_v, n);                                   \
+      ANYTOUV_C(src_uv + n * 2,                                                \
+                dst_u + n,                                                     \
+                dst_v + n,                                                     \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_SPLITUVROW_SSE2
+SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_Unaligned_SSE2, SplitUVRow_C, 15)
+#endif
+#ifdef HAS_SPLITUVROW_AVX2
+SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31)
+#endif
+#ifdef HAS_SPLITUVROW_NEON
+SPLITUVROWANY(SplitUVRow_Any_NEON, SplitUVRow_NEON, SplitUVRow_C, 15)
+#endif
+#ifdef HAS_SPLITUVROW_MIPS_DSPR2
+SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_Unaligned_MIPS_DSPR2,
+              SplitUVRow_C, 15)
+#endif
+#undef SPLITUVROWANY
+
+#define MERGEUVROW_ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK)                 \
+    void NAMEANY(const uint8* src_u, const uint8* src_v,                       \
+                 uint8* dst_uv, int width) {                                   \
+      int n = width & ~MASK;                                                   \
+      ANYTOUV_SIMD(src_u, src_v, dst_uv, n);                                   \
+      ANYTOUV_C(src_u + n,                                                     \
+                src_v + n,                                                     \
+                dst_uv + n * 2,                                                \
+                width & MASK);                                                 \
+    }
+
+#ifdef HAS_MERGEUVROW_SSE2
+MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_Unaligned_SSE2, MergeUVRow_C, 15)
+#endif
+#ifdef HAS_MERGEUVROW_AVX2
+MERGEUVROW_ANY(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, MergeUVRow_C, 31)
+#endif
+#ifdef HAS_MERGEUVROW_NEON
+MERGEUVROW_ANY(MergeUVRow_Any_NEON, MergeUVRow_NEON, MergeUVRow_C, 15)
+#endif
+#undef MERGEUVROW_ANY
+
+#define MATHROW_ANY(NAMEANY, ARGBMATH_SIMD, ARGBMATH_C, MASK)                  \
+    void NAMEANY(const uint8* src_argb0, const uint8* src_argb1,               \
+                 uint8* dst_argb, int width) {                                 \
+      int n = width & ~MASK;                                                   \
+      ARGBMATH_SIMD(src_argb0, src_argb1, dst_argb, n);                        \
+      ARGBMATH_C(src_argb0 + n * 4,                                            \
+                 src_argb1 + n * 4,                                            \
+                 dst_argb + n * 4,                                             \
+                 width & MASK);                                                \
+    }
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+MATHROW_ANY(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, ARGBMultiplyRow_C,
+            3)
+#endif
+#ifdef HAS_ARGBADDROW_SSE2
+MATHROW_ANY(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, ARGBAddRow_C, 3)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+MATHROW_ANY(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, ARGBSubtractRow_C,
+            3)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+MATHROW_ANY(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, ARGBMultiplyRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBADDROW_AVX2
+MATHROW_ANY(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, ARGBAddRow_C, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+MATHROW_ANY(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, ARGBSubtractRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+MATHROW_ANY(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, ARGBMultiplyRow_C,
+            7)
+#endif
+#ifdef HAS_ARGBADDROW_NEON
+MATHROW_ANY(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, ARGBAddRow_C, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+MATHROW_ANY(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, ARGBSubtractRow_C,
+            7)
+#endif
+#undef MATHROW_ANY
+
+// Shuffle may want to work in place, so last16 method can not be used.
+#define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK)                \
+    void NAMEANY(const uint8* src_argb, uint8* dst_argb,                       \
+                 const uint8* shuffler, int width) {                           \
+      int n = width & ~MASK;                                                   \
+      ARGBTOY_SIMD(src_argb, dst_argb, shuffler, n);                           \
+      ARGBTOY_C(src_argb + n * SBPP,                                           \
+                dst_argb  + n * BPP, shuffler, width & MASK);                  \
+    }
+
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+YANY(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2,
+     ARGBShuffleRow_C, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_Unaligned_SSSE3,
+     ARGBShuffleRow_C, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+YANY(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2,
+     ARGBShuffleRow_C, 4, 4, 15)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_NEON
+YANY(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON,
+     ARGBShuffleRow_C, 4, 4, 3)
+#endif
+#undef YANY
+
+// Interpolate may want to work in place, so last16 method can not be used.
+#define NANY(NAMEANY, TERP_SIMD, TERP_C, SBPP, BPP, MASK)                      \
+    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
+                 ptrdiff_t src_stride_ptr, int width,                          \
+                 int source_y_fraction) {                                      \
+      int n = width & ~MASK;                                                   \
+      TERP_SIMD(dst_ptr, src_ptr, src_stride_ptr,                              \
+                n, source_y_fraction);                                         \
+      TERP_C(dst_ptr + n * BPP,                                                \
+             src_ptr + n * SBPP, src_stride_ptr,                               \
+             width & MASK, source_y_fraction);                                 \
+    }
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+NANY(InterpolateRow_Any_AVX2, InterpolateRow_AVX2,
+     InterpolateRow_C, 1, 1, 32)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSSE3
+NANY(InterpolateRow_Any_SSSE3, InterpolateRow_Unaligned_SSSE3,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSE2
+NANY(InterpolateRow_Any_SSE2, InterpolateRow_Unaligned_SSE2,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_NEON
+NANY(InterpolateRow_Any_NEON, InterpolateRow_NEON,
+     InterpolateRow_C, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
+NANY(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2,
+     InterpolateRow_C, 1, 1, 3)
+#endif
+#undef NANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/row_common.cc b/drivers/theoraplayer/src/YUV/libyuv/src/row_common.cc
new file mode 100755
index 0000000000..135bdc9084
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/row_common.cc
@@ -0,0 +1,2247 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include <string.h>  // For memcpy and memset.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// llvm x86 is poor at ternary operator, so use branchless min/max.
+
+#define USE_BRANCHLESS 1
+#if USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return ((-(v) >> 31) & (v));
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (((255 - (v)) >> 31) | (v)) & 255;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  int m = v >> 31;
+  return (v + m) ^ m;
+}
+#else  // USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return (v < 0) ? 0 : v;
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (v > 255) ? 255 : v;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  return (v < 0) ? -v : v;
+}
+#endif  // USE_BRANCHLESS
+
+#ifdef LIBYUV_LITTLE_ENDIAN
+#define WRITEWORD(p, v) *(uint32*)(p) = v
+#else
+static inline void WRITEWORD(uint8* p, uint32 v) {
+  p[0] = (uint8)(v & 255);
+  p[1] = (uint8)((v >> 8) & 255);
+  p[2] = (uint8)((v >> 16) & 255);
+  p[3] = (uint8)((v >> 24) & 255);
+}
+#endif
+
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb24[0];
+    uint8 g = src_rgb24[1];
+    uint8 r = src_rgb24[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb24 += 3;
+  }
+}
+
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 r = src_raw[0];
+    uint8 g = src_raw[1];
+    uint8 b = src_raw[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_raw += 3;
+  }
+}
+
+void RGB565ToARGBRow_C(const uint8* src_rgb565, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 2) | (g >> 4);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb565 += 2;
+  }
+}
+
+void ARGB1555ToARGBRow_C(const uint8* src_argb1555, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 a = src_argb1555[1] >> 7;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 3) | (g >> 2);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = -a;
+    dst_argb += 4;
+    src_argb1555 += 2;
+  }
+}
+
+void ARGB4444ToARGBRow_C(const uint8* src_argb4444, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    uint8 a = src_argb4444[1] >> 4;
+    dst_argb[0] = (b << 4) | b;
+    dst_argb[1] = (g << 4) | g;
+    dst_argb[2] = (r << 4) | r;
+    dst_argb[3] = (a << 4) | a;
+    dst_argb += 4;
+    src_argb4444 += 2;
+  }
+}
+
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = b;
+    dst_rgb[1] = g;
+    dst_rgb[2] = r;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = r;
+    dst_rgb[1] = g;
+    dst_rgb[2] = b;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 2;
+    uint8 r1 = src_argb[6] >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
+              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 3;
+    uint8 r1 = src_argb[6] >> 3;
+    uint8 a1 = src_argb[7] >> 7;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | (a1 << 31);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15);
+  }
+}
+
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    uint8 b1 = src_argb[4] >> 4;
+    uint8 g1 = src_argb[5] >> 4;
+    uint8 r1 = src_argb[6] >> 4;
+    uint8 a1 = src_argb[7] >> 4;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | (a1 << 28);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12);
+  }
+}
+
+static __inline int RGBToY(uint8 r, uint8 g, uint8 b) {
+  return (66 * r + 129 * g +  25 * b + 0x1080) >> 8;
+}
+
+static __inline int RGBToU(uint8 r, uint8 g, uint8 b) {
+  return (112 * b - 74 * g - 38 * r + 0x8080) >> 8;
+}
+static __inline int RGBToV(uint8 r, uint8 g, uint8 b) {
+  return (112 * r - 94 * g - 18 * b + 0x8080) >> 8;
+}
+
+#define MAKEROWY(NAME, R, G, B, BPP) \
+void NAME ## ToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) {       \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToY(src_argb0[R], src_argb0[G], src_argb0[B]);               \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVRow_C(const uint8* src_rgb0, int src_stride_rgb,              \
+                       uint8* dst_u, uint8* dst_v, int width) {                \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = (src_rgb0[B] + src_rgb0[B + BPP] +                              \
+               src_rgb1[B] + src_rgb1[B + BPP]) >> 2;                          \
+    uint8 ag = (src_rgb0[G] + src_rgb0[G + BPP] +                              \
+               src_rgb1[G] + src_rgb1[G + BPP]) >> 2;                          \
+    uint8 ar = (src_rgb0[R] + src_rgb0[R + BPP] +                              \
+               src_rgb1[R] + src_rgb1[R + BPP]) >> 2;                          \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = (src_rgb0[B] + src_rgb1[B]) >> 1;                               \
+    uint8 ag = (src_rgb0[G] + src_rgb1[G]) >> 1;                               \
+    uint8 ar = (src_rgb0[R] + src_rgb1[R]) >> 1;                               \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+  }                                                                            \
+}
+
+MAKEROWY(ARGB, 2, 1, 0, 4)
+MAKEROWY(BGRA, 1, 2, 3, 4)
+MAKEROWY(ABGR, 0, 1, 2, 4)
+MAKEROWY(RGBA, 3, 2, 1, 4)
+MAKEROWY(RGB24, 2, 1, 0, 3)
+MAKEROWY(RAW, 0, 1, 2, 3)
+#undef MAKEROWY
+
+// JPeg uses a variation on BT.601-1 full range
+// y =  0.29900 * r + 0.58700 * g + 0.11400 * b
+// u = -0.16874 * r - 0.33126 * g + 0.50000 * b  + center
+// v =  0.50000 * r - 0.41869 * g - 0.08131 * b  + center
+// BT.601 Mpeg range uses:
+// b 0.1016 * 255 = 25.908 = 25
+// g 0.5078 * 255 = 129.489 = 129
+// r 0.2578 * 255 = 65.739 = 66
+// JPeg 8 bit Y (not used):
+// b 0.11400 * 256 = 29.184 = 29
+// g 0.58700 * 256 = 150.272 = 150
+// r 0.29900 * 256 = 76.544 = 77
+// JPeg 7 bit Y:
+// b 0.11400 * 128 = 14.592 = 15
+// g 0.58700 * 128 = 75.136 = 75
+// r 0.29900 * 128 = 38.272 = 38
+// JPeg 8 bit U:
+// b  0.50000 * 255 = 127.5 = 127
+// g -0.33126 * 255 = -84.4713 = -84
+// r -0.16874 * 255 = -43.0287 = -43
+// JPeg 8 bit V:
+// b -0.08131 * 255 = -20.73405 = -20
+// g -0.41869 * 255 = -106.76595 = -107
+// r  0.50000 * 255 = 127.5 = 127
+
+static __inline int RGBToYJ(uint8 r, uint8 g, uint8 b) {
+  return (38 * r + 75 * g +  15 * b + 64) >> 7;
+}
+
+static __inline int RGBToUJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * b - 84 * g - 43 * r + 0x8080) >> 8;
+}
+static __inline int RGBToVJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * r - 107 * g - 20 * b + 0x8080) >> 8;
+}
+
+#define AVGB(a, b) (((a) + (b) + 1) >> 1)
+
+#define MAKEROWYJ(NAME, R, G, B, BPP) \
+void NAME ## ToYJRow_C(const uint8* src_argb0, uint8* dst_y, int width) {      \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToYJ(src_argb0[R], src_argb0[G], src_argb0[B]);              \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVJRow_C(const uint8* src_rgb0, int src_stride_rgb,             \
+                        uint8* dst_u, uint8* dst_v, int width) {               \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = AVGB(AVGB(src_rgb0[B], src_rgb1[B]),                            \
+                    AVGB(src_rgb0[B + BPP], src_rgb1[B + BPP]));               \
+    uint8 ag = AVGB(AVGB(src_rgb0[G], src_rgb1[G]),                            \
+                    AVGB(src_rgb0[G + BPP], src_rgb1[G + BPP]));               \
+    uint8 ar = AVGB(AVGB(src_rgb0[R], src_rgb1[R]),                            \
+                    AVGB(src_rgb0[R + BPP], src_rgb1[R + BPP]));               \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = AVGB(src_rgb0[B], src_rgb1[B]);                                 \
+    uint8 ag = AVGB(src_rgb0[G], src_rgb1[G]);                                 \
+    uint8 ar = AVGB(src_rgb0[R], src_rgb1[R]);                                 \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+  }                                                                            \
+}
+
+MAKEROWYJ(ARGB, 2, 1, 0, 4)
+#undef MAKEROWYJ
+
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    b = (b << 3) | (b >> 2);
+    g = (g << 2) | (g >> 4);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_rgb565 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    b = (b << 3) | (b >> 2);
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb1555 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    b = (b << 4) | b;
+    g = (g << 4) | g;
+    r = (r << 4) | r;
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb4444 += 2;
+    dst_y += 1;
+  }
+}
+
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b1 = src_rgb565[2] & 0x1f;
+    uint8 g1 = (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3);
+    uint8 r1 = src_rgb565[3] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b3 = next_rgb565[2] & 0x1f;
+    uint8 g3 = (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3);
+    uint8 r3 = next_rgb565[3] >> 3;
+    uint8 b = (b0 + b1 + b2 + b3);  // 565 * 4 = 787.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 787 -> 888.
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_rgb565 += 4;
+    next_rgb565 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b = (b0 + b2);  // 565 * 2 = 676.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 676 -> 888
+    g = (g << 1) | (g >> 6);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b1 = src_argb1555[2] & 0x1f;
+    uint8 g1 = (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3);
+    uint8 r1 = (src_argb1555[3] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = (next_argb1555[1] & 0x7c) >> 2;
+    uint8 b3 = next_argb1555[2] & 0x1f;
+    uint8 g3 = (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3);
+    uint8 r3 = (next_argb1555[3] & 0x7c) >> 2;
+    uint8 b = (b0 + b1 + b2 + b3);  // 555 * 4 = 777.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 777 -> 888.
+    g = (g << 1) | (g >> 6);
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb1555 += 4;
+    next_argb1555 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = next_argb1555[1] >> 3;
+    uint8 b = (b0 + b2);  // 555 * 2 = 666.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb4444 = src_argb4444 + src_stride_argb4444;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b1 = src_argb4444[2] & 0x0f;
+    uint8 g1 = src_argb4444[2] >> 4;
+    uint8 r1 = src_argb4444[3] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b3 = next_argb4444[2] & 0x0f;
+    uint8 g3 = next_argb4444[2] >> 4;
+    uint8 r3 = next_argb4444[3] & 0x0f;
+    uint8 b = (b0 + b1 + b2 + b3);  // 444 * 4 = 666.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb4444 += 4;
+    next_argb4444 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b = (b0 + b2);  // 444 * 2 = 555.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 3) | (b >> 2);  // 555 -> 888.
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 8;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8] + src_argb[12]) >> 2;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9] + src_argb[13]) >> 2;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10] + src_argb[14]) >> 2;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 16;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if ((width & 3) == 3) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8]) / 3;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9]) / 3;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10]) / 3;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]);
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = src_argb[3];
+    dst_argb += 4;
+    src_argb += 4;
+  }
+}
+
+// Convert a row of image to Sepia tone.
+void ARGBSepiaRow_C(uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int sb = (b * 17 + g * 68 + r * 35) >> 7;
+    int sg = (b * 22 + g * 88 + r * 45) >> 7;
+    int sr = (b * 24 + g * 98 + r * 50) >> 7;
+    // b does not over flow. a is preserved from original.
+    dst_argb[0] = sb;
+    dst_argb[1] = clamp255(sg);
+    dst_argb[2] = clamp255(sr);
+    dst_argb += 4;
+  }
+}
+
+// Apply color matrix to a row of image. Matrix is signed.
+// TODO(fbarchard): Consider adding rounding (+32).
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = src_argb[0];
+    int g = src_argb[1];
+    int r = src_argb[2];
+    int a = src_argb[3];
+    int sb = (b * matrix_argb[0] + g * matrix_argb[1] +
+              r * matrix_argb[2] + a * matrix_argb[3]) >> 6;
+    int sg = (b * matrix_argb[4] + g * matrix_argb[5] +
+              r * matrix_argb[6] + a * matrix_argb[7]) >> 6;
+    int sr = (b * matrix_argb[8] + g * matrix_argb[9] +
+              r * matrix_argb[10] + a * matrix_argb[11]) >> 6;
+    int sa = (b * matrix_argb[12] + g * matrix_argb[13] +
+              r * matrix_argb[14] + a * matrix_argb[15]) >> 6;
+    dst_argb[0] = Clamp(sb);
+    dst_argb[1] = Clamp(sg);
+    dst_argb[2] = Clamp(sr);
+    dst_argb[3] = Clamp(sa);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int a = dst_argb[3];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb[3] = table_argb[a * 4 + 3];
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb += 4;
+  }
+}
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+    dst_argb[1] = (g * scale >> 16) * interval_size + interval_offset;
+    dst_argb[2] = (r * scale >> 16) * interval_size + interval_offset;
+    dst_argb += 4;
+  }
+}
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 24
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value) {
+  const uint32 b_scale = REPEAT8(value & 0xff);
+  const uint32 g_scale = REPEAT8((value >> 8) & 0xff);
+  const uint32 r_scale = REPEAT8((value >> 16) & 0xff);
+  const uint32 a_scale = REPEAT8(value >> 24);
+
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb[0]);
+    const uint32 g = REPEAT8(src_argb[1]);
+    const uint32 r = REPEAT8(src_argb[2]);
+    const uint32 a = REPEAT8(src_argb[3]);
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 16
+
+void ARGBMultiplyRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb0[0]);
+    const uint32 g = REPEAT8(src_argb0[1]);
+    const uint32 r = REPEAT8(src_argb0[2]);
+    const uint32 a = REPEAT8(src_argb0[3]);
+    const uint32 b_scale = src_argb1[0];
+    const uint32 g_scale = src_argb1[1];
+    const uint32 r_scale = src_argb1[2];
+    const uint32 a_scale = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define SHADE(f, v) clamp255(v + f)
+
+void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_add = src_argb1[0];
+    const int g_add = src_argb1[1];
+    const int r_add = src_argb1[2];
+    const int a_add = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_add);
+    dst_argb[1] = SHADE(g, g_add);
+    dst_argb[2] = SHADE(r, r_add);
+    dst_argb[3] = SHADE(a, a_add);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+#define SHADE(f, v) clamp0(f - v)
+
+void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_sub = src_argb1[0];
+    const int g_sub = src_argb1[1];
+    const int r_sub = src_argb1[2];
+    const int a_sub = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_sub);
+    dst_argb[1] = SHADE(g, g_sub);
+    dst_argb[2] = SHADE(r, r_sub);
+    dst_argb[3] = SHADE(a, a_sub);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+// Sobel functions which mimics SSSE3.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i];
+    int b = src_y1[i];
+    int c = src_y2[i];
+    int a_sub = src_y0[i + 2];
+    int b_sub = src_y1[i + 2];
+    int c_sub = src_y2[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobelx[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i + 0];
+    int b = src_y0[i + 1];
+    int c = src_y0[i + 2];
+    int a_sub = src_y1[i + 0];
+    int b_sub = src_y1[i + 1];
+    int c_sub = src_y1[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobely[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_argb[0] = (uint8)(s);
+    dst_argb[1] = (uint8)(s);
+    dst_argb[2] = (uint8)(s);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_y[i] = (uint8)(s);
+  }
+}
+
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int g = clamp255(r + b);
+    dst_argb[0] = (uint8)(b);
+    dst_argb[1] = (uint8)(g);
+    dst_argb[2] = (uint8)(r);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
+  // Copy a Y to RGB.
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = src_y[0];
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    ++src_y;
+  }
+}
+
+// C reference code that mimics the YUV assembly.
+
+#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
+
+#define UB 127 /* min(63,(int8)(2.018 * 64)) */
+#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+static __inline void YuvPixel(uint8 y, uint8 u, uint8 v,
+                              uint8* b, uint8* g, uint8* r) {
+  int32 y1 = ((int32)(y) - 16) * YG;
+  *b = Clamp((int32)((u * UB + v * VB) - (BB) + y1) >> 6);
+  *g = Clamp((int32)((u * UG + v * VG) - (BG) + y1) >> 6);
+  *r = Clamp((int32)((u * UR + v * VR) - (BR) + y1) >> 6);
+}
+
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+// C mimic assembly.
+// TODO(fbarchard): Remove subsampling from Neon.
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 u = (src_u[0] + src_u[1] + 1) >> 1;
+    uint8 v = (src_v[0] + src_v[1] + 1) >> 1;
+    YuvPixel(src_y[0], u, v, rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], u, v, rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 2;
+    src_v += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+#else
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+#endif
+// Also used for 420
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* rgb_buf,
+                      int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 4, rgb_buf + 5);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* rgb_buf,
+                    int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 4, rgb_buf + 3);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+  }
+}
+
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    b1 = b1 >> 4;
+    g1 = g1 >> 4;
+    r1 = r1 >> 4;
+    *(uint32*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | 0xf000f000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb4444 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    *(uint16*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        0xf000;
+  }
+}
+
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb1555,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 3;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | 0x80008000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb1555 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        0x8000;
+  }
+}
+
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    YuvPixel(src_y[2], src_u[0], src_v[0],
+             rgb_buf + 8, rgb_buf + 9, rgb_buf + 10);
+    rgb_buf[11] = 255;
+    YuvPixel(src_y[3], src_u[0], src_v[0],
+             rgb_buf + 12, rgb_buf + 13, rgb_buf + 14);
+    rgb_buf[15] = 255;
+    src_y += 4;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 16;  // Advance 4 pixels.
+  }
+  if (width & 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* usrc_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], usrc_v[0], usrc_v[1],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    usrc_v += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+
+    YuvPixel(src_y[1], src_vu[1], src_vu[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+
+    src_y += 2;
+    src_vu += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* usrc_v,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1], &b0, &g0, &r0);
+    YuvPixel(src_y[1], usrc_v[0], usrc_v[1], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    usrc_v += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], usrc_v[0], usrc_v[1], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* vsrc_u,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], vsrc_u[1], vsrc_u[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    vsrc_u += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_yuy2 += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_uyvy += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 7, rgb_buf + 6, rgb_buf + 5);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+  }
+}
+
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 6, rgb_buf + 5, rgb_buf + 4);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 6, rgb_buf + 7);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+  }
+}
+
+void YToARGBRow_C(const uint8* src_y, uint8* rgb_buf, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], 128, 128,
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], 128, 128,
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], 128, 128,
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void MirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  src += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst[x] = src[0];
+    dst[x + 1] = src[-1];
+    src -= 2;
+  }
+  if (width & 1) {
+    dst[width - 1] = src[0];
+  }
+}
+
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  src_uv += (width - 1) << 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[-2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[-2 + 1];
+    src_uv -= 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  const uint32* src32 = (const uint32*)(src);
+  uint32* dst32 = (uint32*)(dst);
+  src32 += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst32[x] = src32[0];
+    dst32[x + 1] = src32[-1];
+    src32 -= 2;
+  }
+  if (width & 1) {
+    dst32[width - 1] = src32[0];
+  }
+}
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[3];
+    src_uv += 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_uv[0] = src_u[x];
+    dst_uv[1] = src_v[x];
+    dst_uv[2] = src_u[x + 1];
+    dst_uv[3] = src_v[x + 1];
+    dst_uv += 4;
+  }
+  if (width & 1) {
+    dst_uv[0] = src_u[width - 1];
+    dst_uv[1] = src_v[width - 1];
+  }
+}
+
+void CopyRow_C(const uint8* src, uint8* dst, int count) {
+  memcpy(dst, src, count);
+}
+
+void SetRow_C(uint8* dst, uint32 v8, int count) {
+#ifdef _MSC_VER
+  // VC will generate rep stosb.
+  int x;
+  for (x = 0; x < count; ++x) {
+    dst[x] = v8;
+  }
+#else
+  memset(dst, v8, count);
+#endif
+}
+
+void ARGBSetRows_C(uint8* dst, uint32 v32, int width,
+                 int dst_stride, int height) {
+  int y;
+  for (y = 0; y < height; ++y) {
+    uint32* d = (uint32*)(dst);
+    int x;
+    for (x = 0; x < width; ++x) {
+      d[x] = v32;
+    }
+    dst += dst_stride;
+  }
+}
+
+// Filter 2 rows of YUY2 UV's (422) into U and V (420).
+void YUY2ToUVRow_C(const uint8* src_yuy2, int src_stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values, filtering 2 rows of YUY2.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1;
+    dst_v[0] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1;
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 UV's (422) into U and V (422).
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_yuy2[1];
+    dst_v[0] = src_yuy2[3];
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 Y's (422) into Y (420/422).
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_yuy2[0];
+    dst_y[x + 1] = src_yuy2[2];
+    src_yuy2 += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_yuy2[0];
+  }
+}
+
+// Filter 2 rows of UYVY UV's (422) into U and V (420).
+void UYVYToUVRow_C(const uint8* src_uyvy, int src_stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_uyvy[0] + src_uyvy[src_stride_uyvy + 0] + 1) >> 1;
+    dst_v[0] = (src_uyvy[2] + src_uyvy[src_stride_uyvy + 2] + 1) >> 1;
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY UV's (422) into U and V (422).
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_uyvy[0];
+    dst_v[0] = src_uyvy[2];
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY Y's (422) into Y (420/422).
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_uyvy[1];
+    dst_y[x + 1] = src_uyvy[3];
+    src_uyvy += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_uyvy[1];
+  }
+}
+
+#define BLEND(f, b, a) (((256 - a) * b) >> 8) + f
+
+// Blend src_argb0 over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb0 or src_argb1.
+// This code mimics the SSSE3 version for better testability.
+void ARGBBlendRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                    uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+
+    fb = src_argb0[4 + 0];
+    fg = src_argb0[4 + 1];
+    fr = src_argb0[4 + 2];
+    a = src_argb0[4 + 3];
+    bb = src_argb1[4 + 0];
+    bg = src_argb1[4 + 1];
+    br = src_argb1[4 + 2];
+    dst_argb[4 + 0] = BLEND(fb, bb, a);
+    dst_argb[4 + 1] = BLEND(fg, bg, a);
+    dst_argb[4 + 2] = BLEND(fr, br, a);
+    dst_argb[4 + 3] = 255u;
+    src_argb0 += 8;
+    src_argb1 += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+  }
+}
+#undef BLEND
+#define ATTENUATE(f, a) (a | (a << 8)) * (f | (f << 8)) >> 24
+
+// Multiply source RGB by alpha and store to destination.
+// This code mimics the SSSE3 version for better testability.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+    b = src_argb[4];
+    g = src_argb[5];
+    r = src_argb[6];
+    a = src_argb[7];
+    dst_argb[4] = ATTENUATE(b, a);
+    dst_argb[5] = ATTENUATE(g, a);
+    dst_argb[6] = ATTENUATE(r, a);
+    dst_argb[7] = a;
+    src_argb += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    const uint32 b = src_argb[0];
+    const uint32 g = src_argb[1];
+    const uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+  }
+}
+#undef ATTENUATE
+
+// Divide source RGB by alpha and store to destination.
+// b = (b * 255 + (a / 2)) / a;
+// g = (g * 255 + (a / 2)) / a;
+// r = (r * 255 + (a / 2)) / a;
+// Reciprocal method is off by 1 on some values. ie 125
+// 8.8 fixed point inverse table with 1.0 in upper short and 1 / a in lower.
+#define T(a) 0x01000000 + (0x10000 / a)
+const uint32 fixed_invtbl8[256] = {
+  0x01000000, 0x0100ffff, T(0x02), T(0x03), T(0x04), T(0x05), T(0x06), T(0x07),
+  T(0x08), T(0x09), T(0x0a), T(0x0b), T(0x0c), T(0x0d), T(0x0e), T(0x0f),
+  T(0x10), T(0x11), T(0x12), T(0x13), T(0x14), T(0x15), T(0x16), T(0x17),
+  T(0x18), T(0x19), T(0x1a), T(0x1b), T(0x1c), T(0x1d), T(0x1e), T(0x1f),
+  T(0x20), T(0x21), T(0x22), T(0x23), T(0x24), T(0x25), T(0x26), T(0x27),
+  T(0x28), T(0x29), T(0x2a), T(0x2b), T(0x2c), T(0x2d), T(0x2e), T(0x2f),
+  T(0x30), T(0x31), T(0x32), T(0x33), T(0x34), T(0x35), T(0x36), T(0x37),
+  T(0x38), T(0x39), T(0x3a), T(0x3b), T(0x3c), T(0x3d), T(0x3e), T(0x3f),
+  T(0x40), T(0x41), T(0x42), T(0x43), T(0x44), T(0x45), T(0x46), T(0x47),
+  T(0x48), T(0x49), T(0x4a), T(0x4b), T(0x4c), T(0x4d), T(0x4e), T(0x4f),
+  T(0x50), T(0x51), T(0x52), T(0x53), T(0x54), T(0x55), T(0x56), T(0x57),
+  T(0x58), T(0x59), T(0x5a), T(0x5b), T(0x5c), T(0x5d), T(0x5e), T(0x5f),
+  T(0x60), T(0x61), T(0x62), T(0x63), T(0x64), T(0x65), T(0x66), T(0x67),
+  T(0x68), T(0x69), T(0x6a), T(0x6b), T(0x6c), T(0x6d), T(0x6e), T(0x6f),
+  T(0x70), T(0x71), T(0x72), T(0x73), T(0x74), T(0x75), T(0x76), T(0x77),
+  T(0x78), T(0x79), T(0x7a), T(0x7b), T(0x7c), T(0x7d), T(0x7e), T(0x7f),
+  T(0x80), T(0x81), T(0x82), T(0x83), T(0x84), T(0x85), T(0x86), T(0x87),
+  T(0x88), T(0x89), T(0x8a), T(0x8b), T(0x8c), T(0x8d), T(0x8e), T(0x8f),
+  T(0x90), T(0x91), T(0x92), T(0x93), T(0x94), T(0x95), T(0x96), T(0x97),
+  T(0x98), T(0x99), T(0x9a), T(0x9b), T(0x9c), T(0x9d), T(0x9e), T(0x9f),
+  T(0xa0), T(0xa1), T(0xa2), T(0xa3), T(0xa4), T(0xa5), T(0xa6), T(0xa7),
+  T(0xa8), T(0xa9), T(0xaa), T(0xab), T(0xac), T(0xad), T(0xae), T(0xaf),
+  T(0xb0), T(0xb1), T(0xb2), T(0xb3), T(0xb4), T(0xb5), T(0xb6), T(0xb7),
+  T(0xb8), T(0xb9), T(0xba), T(0xbb), T(0xbc), T(0xbd), T(0xbe), T(0xbf),
+  T(0xc0), T(0xc1), T(0xc2), T(0xc3), T(0xc4), T(0xc5), T(0xc6), T(0xc7),
+  T(0xc8), T(0xc9), T(0xca), T(0xcb), T(0xcc), T(0xcd), T(0xce), T(0xcf),
+  T(0xd0), T(0xd1), T(0xd2), T(0xd3), T(0xd4), T(0xd5), T(0xd6), T(0xd7),
+  T(0xd8), T(0xd9), T(0xda), T(0xdb), T(0xdc), T(0xdd), T(0xde), T(0xdf),
+  T(0xe0), T(0xe1), T(0xe2), T(0xe3), T(0xe4), T(0xe5), T(0xe6), T(0xe7),
+  T(0xe8), T(0xe9), T(0xea), T(0xeb), T(0xec), T(0xed), T(0xee), T(0xef),
+  T(0xf0), T(0xf1), T(0xf2), T(0xf3), T(0xf4), T(0xf5), T(0xf6), T(0xf7),
+  T(0xf8), T(0xf9), T(0xfa), T(0xfb), T(0xfc), T(0xfd), T(0xfe), 0x01000100 };
+#undef T
+
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    const uint32 ia = fixed_invtbl8[a] & 0xffff;  // 8.8 fixed point
+    b = (b * ia) >> 8;
+    g = (g * ia) >> 8;
+    r = (r * ia) >> 8;
+    // Clamping should not be necessary but is free in assembly.
+    dst_argb[0] = clamp255(b);
+    dst_argb[1] = clamp255(g);
+    dst_argb[2] = clamp255(r);
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width) {
+  int32 row_sum[4] = {0, 0, 0, 0};
+  int x;
+  for (x = 0; x < width; ++x) {
+    row_sum[0] += row[x * 4 + 0];
+    row_sum[1] += row[x * 4 + 1];
+    row_sum[2] += row[x * 4 + 2];
+    row_sum[3] += row[x * 4 + 3];
+    cumsum[x * 4 + 0] = row_sum[0]  + previous_cumsum[x * 4 + 0];
+    cumsum[x * 4 + 1] = row_sum[1]  + previous_cumsum[x * 4 + 1];
+    cumsum[x * 4 + 2] = row_sum[2]  + previous_cumsum[x * 4 + 2];
+    cumsum[x * 4 + 3] = row_sum[3]  + previous_cumsum[x * 4 + 3];
+  }
+}
+
+void CumulativeSumToAverageRow_C(const int32* tl, const int32* bl,
+                                int w, int area, uint8* dst, int count) {
+  float ooa = 1.0f / area;
+  int i;
+  for (i = 0; i < count; ++i) {
+    dst[0] = (uint8)((bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * ooa);
+    dst[1] = (uint8)((bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * ooa);
+    dst[2] = (uint8)((bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) * ooa);
+    dst[3] = (uint8)((bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) * ooa);
+    dst += 4;
+    tl += 4;
+    bl += 4;
+  }
+}
+
+// Copy pixels from rotated source to destination row with a slope.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width) {
+  int i;
+  // Render a row of pixels from source into a buffer.
+  float uv[2];
+  uv[0] = uv_dudv[0];
+  uv[1] = uv_dudv[1];
+  for (i = 0; i < width; ++i) {
+    int x = (int)(uv[0]);
+    int y = (int)(uv[1]);
+    *(uint32*)(dst_argb) =
+        *(const uint32*)(src_argb + y * src_argb_stride +
+                                         x * 4);
+    dst_argb += 4;
+    uv[0] += uv_dudv[2];
+    uv[1] += uv_dudv[3];
+  }
+}
+
+// Blend 2 rows into 1 for conversions such as I422ToI420.
+void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+               uint8* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+// C version 2x2 -> 2x1.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride,
+                      int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
+void ARGBToBayerRow_C(const uint8* src_argb,
+                      uint8* dst_bayer, uint32 selector, int pix) {
+  int index0 = selector & 0xff;
+  int index1 = (selector >> 8) & 0xff;
+  // Copy a row of Bayer.
+  int x;
+  for (x = 0; x < pix - 1; x += 2) {
+    dst_bayer[0] = src_argb[index0];
+    dst_bayer[1] = src_argb[index1];
+    src_argb += 8;
+    dst_bayer += 2;
+  }
+  if (pix & 1) {
+    dst_bayer[0] = src_argb[index0];
+  }
+}
+
+// Select G channel from ARGB.  e.g.  GGGGGGGG
+void ARGBToBayerGGRow_C(const uint8* src_argb,
+                        uint8* dst_bayer, uint32 selector, int pix) {
+  // Copy a row of G.
+  int x;
+  for (x = 0; x < pix - 1; x += 2) {
+    dst_bayer[0] = src_argb[1];
+    dst_bayer[1] = src_argb[5];
+    src_argb += 8;
+    dst_bayer += 2;
+  }
+  if (pix & 1) {
+    dst_bayer[0] = src_argb[1];
+  }
+}
+
+// Use first 4 shuffler values to reorder ARGB channels.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix) {
+  int index0 = shuffler[0];
+  int index1 = shuffler[1];
+  int index2 = shuffler[2];
+  int index3 = shuffler[3];
+  // Shuffle a row of ARGB.
+  int x;
+  for (x = 0; x < pix; ++x) {
+    // To support in-place conversion.
+    uint8 b = src_argb[index0];
+    uint8 g = src_argb[index1];
+    uint8 r = src_argb[index2];
+    uint8 a = src_argb[index3];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = src_y[1];
+    dst_frame[3] = src_v[0];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = src_y[0];  // duplicate last y
+    dst_frame[3] = src_v[0];
+  }
+}
+
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = src_y[1];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = src_y[0];  // duplicate last y
+  }
+}
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(HAS_I422TOARGBROW_SSSE3)
+// row_win.cc has asm version, but GCC uses 2 step wrapper.
+#if !defined(_MSC_VER) && (defined(__x86_64__) || defined(__i386__))
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* rgb_buf,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToRGB565Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+#endif  // !defined(_MSC_VER) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* rgb_buf,
+                             int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToARGB1555Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* rgb_buf,
+                             int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, width);
+  ARGBToARGB4444Row_SSE2(row, rgb_buf, width);
+  free_aligned_buffer_64(row);
+}
+
+void NV12ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_uv,
+                           uint8* dst_rgb565,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  NV12ToARGBRow_SSSE3(src_y, src_uv, row, width);
+  ARGBToRGB565Row_SSE2(row, dst_rgb565, width);
+  free_aligned_buffer_64(row);
+}
+
+void NV21ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_vu,
+                           uint8* dst_rgb565,
+                           int width) {
+  // Allocate a row of ARGB.
+  align_buffer_64(row, width * 4);
+  NV21ToARGBRow_SSSE3(src_y, src_vu, row, width);
+  ARGBToRGB565Row_SSE2(row, dst_rgb565, width);
+  free_aligned_buffer_64(row);
+}
+
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
+                         uint8* dst_argb,
+                         int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  YUY2ToUV422Row_SSE2(src_yuy2, row_u, row_v, width);
+  YUY2ToYRow_SSE2(src_yuy2, row_y, width);
+  I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void YUY2ToARGBRow_Unaligned_SSSE3(const uint8* src_yuy2,
+                                   uint8* dst_argb,
+                                   int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  YUY2ToUV422Row_Unaligned_SSE2(src_yuy2, row_u, row_v, width);
+  YUY2ToYRow_Unaligned_SSE2(src_yuy2, row_y, width);
+  I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
+                         uint8* dst_argb,
+                         int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  UYVYToUV422Row_SSE2(src_uyvy, row_u, row_v, width);
+  UYVYToYRow_SSE2(src_uyvy, row_y, width);
+  I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
+                                   uint8* dst_argb,
+                                   int width) {
+  // Allocate a rows of yuv.
+  align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+  uint8* row_u = row_y + ((width + 63) & ~63);
+  uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+  UYVYToUV422Row_Unaligned_SSE2(src_uyvy, row_u, row_v, width);
+  UYVYToYRow_Unaligned_SSE2(src_uyvy, row_y, width);
+  I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
+  free_aligned_buffer_64(row_y);
+}
+
+#endif  // defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
+#endif  // !defined(LIBYUV_DISABLE_X86)
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    float b = (float)(src_argb[0]);
+    float g = (float)(src_argb[1]);
+    float r = (float)(src_argb[2]);
+    float a = (float)(src_argb[3]);
+    float b2 = b * b;
+    float g2 = g * g;
+    float r2 = r * r;
+    float a2 = a * a;
+    float db = poly[0] + poly[4] * b;
+    float dg = poly[1] + poly[5] * g;
+    float dr = poly[2] + poly[6] * r;
+    float da = poly[3] + poly[7] * a;
+    float b3 = b2 * b;
+    float g3 = g2 * g;
+    float r3 = r2 * r;
+    float a3 = a2 * a;
+    db += poly[8] * b2;
+    dg += poly[9] * g2;
+    dr += poly[10] * r2;
+    da += poly[11] * a2;
+    db += poly[12] * b3;
+    dg += poly[13] * g3;
+    dr += poly[14] * r3;
+    da += poly[15] * a3;
+
+    dst_argb[0] = Clamp((int32)(db));
+    dst_argb[1] = Clamp((int32)(dg));
+    dst_argb[2] = Clamp((int32)(dr));
+    dst_argb[3] = Clamp((int32)(da));
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff) {
+  uint32 bc = lumacoeff & 0xff;
+  uint32 gc = (lumacoeff >> 8) & 0xff;
+  uint32 rc = (lumacoeff >> 16) & 0xff;
+
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    const uint8* luma1;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+    luma1 = ((src_argb[4] * bc + src_argb[5] * gc +
+              src_argb[6] * rc) & 0x7F00u) + luma;
+    dst_argb[4] = luma1[src_argb[4]];
+    dst_argb[5] = luma1[src_argb[5]];
+    dst_argb[6] = luma1[src_argb[6]];
+    dst_argb[7] = src_argb[7];
+    src_argb += 8;
+    dst_argb += 8;
+  }
+  if (width & 1) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+  }
+}
+
+void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[3];
+    dst[7] = src[7];
+    dst += 8;
+    src += 8;
+  }
+  if (width & 1) {
+    dst[3] = src[3];
+  }
+}
+
+void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[0];
+    dst[7] = src[1];
+    dst += 8;
+    src += 2;
+  }
+  if (width & 1) {
+    dst[3] = src[0];
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/row_mips.cc b/drivers/theoraplayer/src/YUV/libyuv/src/row_mips.cc
new file mode 100755
index 0000000000..4435c55c5c
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/row_mips.cc
@@ -0,0 +1,991 @@
+/*
+ *  Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__)
+
+#ifdef HAS_COPYROW_MIPS
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
+  __asm__ __volatile__ (
+    ".set      noreorder                         \n"
+    ".set      noat                              \n"
+    "slti      $at, %[count], 8                  \n"
+    "bne       $at ,$zero, $last8                \n"
+    "xor       $t8, %[src], %[dst]               \n"
+    "andi      $t8, $t8, 0x3                     \n"
+
+    "bne       $t8, $zero, unaligned             \n"
+    "negu      $a3, %[dst]                       \n"
+    // make dst/src aligned
+    "andi      $a3, $a3, 0x3                     \n"
+    "beq       $a3, $zero, $chk16w               \n"
+    // word-aligned now count is the remining bytes count
+    "subu     %[count], %[count], $a3            \n"
+
+    "lwr       $t8, 0(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"
+    "swr       $t8, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+
+    // Now the dst/src are mutually word-aligned with word-aligned addresses
+    "$chk16w:                                    \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, chk8w              \n"
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"
+    // t0 is the "past the end" address
+
+    // When in the loop we exercise "pref 30,x(a1)", the a1+x should not be past
+    // the "t0-32" address
+    // This means: for x=128 the last "safe" a1 address is "t0-160"
+    // Alternatively, for x=64 the last "safe" a1 address is "t0-96"
+    // we will use "pref 30,128(a1)", so "t0-160" is the limit
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line of src
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $loop16w                     \n"
+    "nop                                         \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$loop16w:                                    \n"
+    "pref      0, 96(%[src])                     \n"
+    "lw        $t0, 0(%[src])                    \n"
+    "bgtz      $v1, $skip_pref30_96              \n"  // skip
+    "lw        $t1, 4(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"  // continue
+    "$skip_pref30_96:                            \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    //  bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lw        $t0, 32(%[src])                   \n"
+    "bgtz      $v1, $skip_pref30_128             \n"  // skip pref 30,128(a1)
+    "lw        $t1, 36(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"  // set dest, addr 128
+    "$skip_pref30_128:                           \n"
+    "lw        $t2, 40(%[src])                   \n"
+    "lw        $t3, 44(%[src])                   \n"
+    "lw        $t4, 48(%[src])                   \n"
+    "lw        $t5, 52(%[src])                   \n"
+    "lw        $t6, 56(%[src])                   \n"
+    "lw        $t7, 60(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst], %[dst], 64                \n"  // adding 64 to dest
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bne       %[dst], $a3, $loop16w             \n"
+    " addiu    %[src], %[src], 64                \n"  // adding 64 to src
+    "move      %[count], $t8                     \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "chk8w:                                      \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count past 32-bytes
+    "beq       %[count], $t8, chk1w              \n"
+    // count=t8,no 32-byte chunk
+    " nop                                        \n"
+
+    "lw        $t0, 0(%[src])                    \n"
+    "lw        $t1, 4(%[src])                    \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "chk1w:                                      \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, $last8             \n"
+    " subu     $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+    // copying in words (4-byte chunks)
+    "$wordCopy_loop:                             \n"
+    "lw        $t3, 0(%[src])                    \n"
+    // the first t3 may be equal t0 ... optimize?
+    "addiu     %[src], %[src],4                  \n"
+    "addiu     %[dst], %[dst],4                  \n"
+    "bne       %[dst], $a3,$wordCopy_loop        \n"
+    " sw       $t3, -4(%[dst])                   \n"
+
+    // For the last (<8) bytes
+    "$last8:                                     \n"
+    "blez      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n"  // a3 -last dst address
+    "$last8loop:                                 \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst], $a3, $last8loop           \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "leave:                                      \n"
+    "  j       $ra                               \n"
+    "  nop                                       \n"
+
+    //
+    // UNALIGNED case
+    //
+
+    "unaligned:                                  \n"
+    // got here with a3="negu a1"
+    "andi      $a3, $a3, 0x3                     \n"  // a1 is word aligned?
+    "beqz      $a3, $ua_chk16w                   \n"
+    " subu     %[count], %[count], $a3           \n"
+    // bytes left after initial a3 bytes
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"  // a3 may be 1, 2 or 3
+    "swr       $v1, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+    // below the dst will be word aligned (NOTE1)
+    "$ua_chk16w:                                 \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, ua_chk8w           \n"
+    // if a2==t8, no 64-byte chunks
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"  // t0 "past the end"
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line  addr 32
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // safe, as we have at least 64 bytes ahead
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $ua_loop16w                  \n"
+    // skip "pref 30,64(a1)" for too short arrays
+    " nop                                        \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$ua_loop16w:                                \n"
+    "pref      0, 96(%[src])                     \n"
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "bgtz      $v1, $ua_skip_pref30_96           \n"
+    " lwl      $t1, 7(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"
+    // continue setting up the dest, addr 96
+    "$ua_skip_pref30_96:                         \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    // bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lwr       $t0, 32(%[src])                   \n"
+    "lwl       $t0, 35(%[src])                   \n"
+    "lwr       $t1, 36(%[src])                   \n"
+    "bgtz      $v1, ua_skip_pref30_128           \n"
+    " lwl      $t1, 39(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"
+    // continue setting up the dest, addr 128
+    "ua_skip_pref30_128:                         \n"
+
+    "lwr       $t2, 40(%[src])                   \n"
+    "lwl       $t2, 43(%[src])                   \n"
+    "lwr       $t3, 44(%[src])                   \n"
+    "lwl       $t3, 47(%[src])                   \n"
+    "lwr       $t4, 48(%[src])                   \n"
+    "lwl       $t4, 51(%[src])                   \n"
+    "lwr       $t5, 52(%[src])                   \n"
+    "lwl       $t5, 55(%[src])                   \n"
+    "lwr       $t6, 56(%[src])                   \n"
+    "lwl       $t6, 59(%[src])                   \n"
+    "lwr       $t7, 60(%[src])                   \n"
+    "lwl       $t7, 63(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst],%[dst],64                  \n"  // adding 64 to dest
+    "sgtu      $v1,%[dst],$t9                    \n"
+    "bne       %[dst],$a3,$ua_loop16w            \n"
+    " addiu    %[src],%[src],64                  \n"  // adding 64 to src
+    "move      %[count],$t8                      \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "ua_chk8w:                                   \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count
+    "beq       %[count], $t8, $ua_chk1w          \n"
+    // when count==t8, no 32-byte chunk
+
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "lwl       $t1, 7(%[src])                    \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "$ua_chk1w:                                  \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, ua_smallCopy       \n"
+    "subu      $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+
+    // copying in words (4-byte chunks)
+    "$ua_wordCopy_loop:                          \n"
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addiu     %[src], %[src], 4                 \n"
+    "addiu     %[dst], %[dst], 4                 \n"
+    // note: dst=a1 is word aligned here, see NOTE1
+    "bne       %[dst], $a3, $ua_wordCopy_loop    \n"
+    " sw       $v1,-4(%[dst])                    \n"
+
+    // Now less than 4 bytes (value in count) left to copy
+    "ua_smallCopy:                               \n"
+    "beqz      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n" // a3 = last dst address
+    "$ua_smallCopy_loop:                         \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst],$a3,$ua_smallCopy_loop     \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "j         $ra                               \n"
+    " nop                                        \n"
+    ".set      at                                \n"
+    ".set      reorder                           \n"
+       : [dst] "+r" (dst), [src] "+r" (src)
+       : [count] "r" (count)
+       : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7",
+       "t8", "t9", "a3", "v1", "at"
+  );
+}
+#endif  // HAS_COPYROW_MIPS
+
+// MIPS DSPR2 functions
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_dsp) && \
+    (__mips_dsp_rev >= 2)
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int width) {
+  __asm__ __volatile__ (
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+    "srl             $t4, %[width], 4              \n"  // multiplies of 16
+    "blez            $t4, 2f                       \n"
+    " andi           %[width], %[width], 0xf       \n"  // residual
+
+    ".p2align        2                             \n"
+  "1:                                              \n"
+    "addiu           $t4, $t4, -1                  \n"
+    "lw              $t0, 0(%[src_uv])             \n"  // V1 | U1 | V0 | U0
+    "lw              $t1, 4(%[src_uv])             \n"  // V3 | U3 | V2 | U2
+    "lw              $t2, 8(%[src_uv])             \n"  // V5 | U5 | V4 | U4
+    "lw              $t3, 12(%[src_uv])            \n"  // V7 | U7 | V6 | U6
+    "lw              $t5, 16(%[src_uv])            \n"  // V9 | U9 | V8 | U8
+    "lw              $t6, 20(%[src_uv])            \n"  // V11 | U11 | V10 | U10
+    "lw              $t7, 24(%[src_uv])            \n"  // V13 | U13 | V12 | U12
+    "lw              $t8, 28(%[src_uv])            \n"  // V15 | U15 | V14 | U14
+    "addiu           %[src_uv], %[src_uv], 32      \n"
+    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
+    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
+    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
+    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
+    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
+    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
+    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
+    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
+    "sw              $t9, 0(%[dst_v])              \n"
+    "sw              $t0, 0(%[dst_u])              \n"
+    "sw              $t1, 4(%[dst_v])              \n"
+    "sw              $t2, 4(%[dst_u])              \n"
+    "sw              $t3, 8(%[dst_v])              \n"
+    "sw              $t5, 8(%[dst_u])              \n"
+    "sw              $t6, 12(%[dst_v])             \n"
+    "sw              $t7, 12(%[dst_u])             \n"
+    "addiu           %[dst_v], %[dst_v], 16        \n"
+    "bgtz            $t4, 1b                       \n"
+    " addiu          %[dst_u], %[dst_u], 16        \n"
+
+    "beqz            %[width], 3f                  \n"
+    " nop                                          \n"
+
+  "2:                                              \n"
+    "lbu             $t0, 0(%[src_uv])             \n"
+    "lbu             $t1, 1(%[src_uv])             \n"
+    "addiu           %[src_uv], %[src_uv], 2       \n"
+    "addiu           %[width], %[width], -1        \n"
+    "sb              $t0, 0(%[dst_u])              \n"
+    "sb              $t1, 0(%[dst_v])              \n"
+    "addiu           %[dst_u], %[dst_u], 1         \n"
+    "bgtz            %[width], 2b                  \n"
+    " addiu          %[dst_v], %[dst_v], 1         \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+     : [src_uv] "+r" (src_uv),
+       [width] "+r" (width),
+       [dst_u] "+r" (dst_u),
+       [dst_v] "+r" (dst_v)
+     :
+     : "t0", "t1", "t2", "t3",
+     "t4", "t5", "t6", "t7", "t8", "t9"
+  );
+}
+
+void SplitUVRow_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
+                                     uint8* dst_v, int width) {
+  __asm__ __volatile__ (
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+    "srl             $t4, %[width], 4              \n"  // multiplies of 16
+    "blez            $t4, 2f                       \n"
+    " andi           %[width], %[width], 0xf       \n"  // residual
+
+    ".p2align        2                             \n"
+  "1:                                              \n"
+    "addiu           $t4, $t4, -1                  \n"
+    "lwr             $t0, 0(%[src_uv])             \n"
+    "lwl             $t0, 3(%[src_uv])             \n"  // V1 | U1 | V0 | U0
+    "lwr             $t1, 4(%[src_uv])             \n"
+    "lwl             $t1, 7(%[src_uv])             \n"  // V3 | U3 | V2 | U2
+    "lwr             $t2, 8(%[src_uv])             \n"
+    "lwl             $t2, 11(%[src_uv])            \n"  // V5 | U5 | V4 | U4
+    "lwr             $t3, 12(%[src_uv])            \n"
+    "lwl             $t3, 15(%[src_uv])            \n"  // V7 | U7 | V6 | U6
+    "lwr             $t5, 16(%[src_uv])            \n"
+    "lwl             $t5, 19(%[src_uv])            \n"  // V9 | U9 | V8 | U8
+    "lwr             $t6, 20(%[src_uv])            \n"
+    "lwl             $t6, 23(%[src_uv])            \n"  // V11 | U11 | V10 | U10
+    "lwr             $t7, 24(%[src_uv])            \n"
+    "lwl             $t7, 27(%[src_uv])            \n"  // V13 | U13 | V12 | U12
+    "lwr             $t8, 28(%[src_uv])            \n"
+    "lwl             $t8, 31(%[src_uv])            \n"  // V15 | U15 | V14 | U14
+    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
+    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
+    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
+    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
+    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
+    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
+    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
+    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
+    "addiu           %[src_uv], %[src_uv], 32      \n"
+    "swr             $t9, 0(%[dst_v])              \n"
+    "swl             $t9, 3(%[dst_v])              \n"
+    "swr             $t0, 0(%[dst_u])              \n"
+    "swl             $t0, 3(%[dst_u])              \n"
+    "swr             $t1, 4(%[dst_v])              \n"
+    "swl             $t1, 7(%[dst_v])              \n"
+    "swr             $t2, 4(%[dst_u])              \n"
+    "swl             $t2, 7(%[dst_u])              \n"
+    "swr             $t3, 8(%[dst_v])              \n"
+    "swl             $t3, 11(%[dst_v])             \n"
+    "swr             $t5, 8(%[dst_u])              \n"
+    "swl             $t5, 11(%[dst_u])             \n"
+    "swr             $t6, 12(%[dst_v])             \n"
+    "swl             $t6, 15(%[dst_v])             \n"
+    "swr             $t7, 12(%[dst_u])             \n"
+    "swl             $t7, 15(%[dst_u])             \n"
+    "addiu           %[dst_u], %[dst_u], 16        \n"
+    "bgtz            $t4, 1b                       \n"
+    " addiu          %[dst_v], %[dst_v], 16        \n"
+
+    "beqz            %[width], 3f                  \n"
+    " nop                                          \n"
+
+  "2:                                              \n"
+    "lbu             $t0, 0(%[src_uv])             \n"
+    "lbu             $t1, 1(%[src_uv])             \n"
+    "addiu           %[src_uv], %[src_uv], 2       \n"
+    "addiu           %[width], %[width], -1        \n"
+    "sb              $t0, 0(%[dst_u])              \n"
+    "sb              $t1, 0(%[dst_v])              \n"
+    "addiu           %[dst_u], %[dst_u], 1         \n"
+    "bgtz            %[width], 2b                  \n"
+    " addiu          %[dst_v], %[dst_v], 1         \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+     : [src_uv] "+r" (src_uv),
+       [width] "+r" (width),
+       [dst_u] "+r" (dst_u),
+       [dst_v] "+r" (dst_v)
+     :
+     : "t0", "t1", "t2", "t3",
+     "t4", "t5", "t6", "t7", "t8", "t9"
+  );
+}
+
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width) {
+  __asm__ __volatile__ (
+    ".set push                             \n"
+    ".set noreorder                        \n"
+
+    "srl       $t4, %[width], 4            \n"  // multiplies of 16
+    "andi      $t5, %[width], 0xf          \n"
+    "blez      $t4, 2f                     \n"
+    " addu     %[src], %[src], %[width]    \n"  // src += width
+
+    ".p2align  2                           \n"
+   "1:                                     \n"
+    "lw        $t0, -16(%[src])            \n"  // |3|2|1|0|
+    "lw        $t1, -12(%[src])            \n"  // |7|6|5|4|
+    "lw        $t2, -8(%[src])             \n"  // |11|10|9|8|
+    "lw        $t3, -4(%[src])             \n"  // |15|14|13|12|
+    "wsbh      $t0, $t0                    \n"  // |2|3|0|1|
+    "wsbh      $t1, $t1                    \n"  // |6|7|4|5|
+    "wsbh      $t2, $t2                    \n"  // |10|11|8|9|
+    "wsbh      $t3, $t3                    \n"  // |14|15|12|13|
+    "rotr      $t0, $t0, 16                \n"  // |0|1|2|3|
+    "rotr      $t1, $t1, 16                \n"  // |4|5|6|7|
+    "rotr      $t2, $t2, 16                \n"  // |8|9|10|11|
+    "rotr      $t3, $t3, 16                \n"  // |12|13|14|15|
+    "addiu     %[src], %[src], -16         \n"
+    "addiu     $t4, $t4, -1                \n"
+    "sw        $t3, 0(%[dst])              \n"  // |15|14|13|12|
+    "sw        $t2, 4(%[dst])              \n"  // |11|10|9|8|
+    "sw        $t1, 8(%[dst])              \n"  // |7|6|5|4|
+    "sw        $t0, 12(%[dst])             \n"  // |3|2|1|0|
+    "bgtz      $t4, 1b                     \n"
+    " addiu    %[dst], %[dst], 16          \n"
+    "beqz      $t5, 3f                     \n"
+    " nop                                  \n"
+
+   "2:                                     \n"
+    "lbu       $t0, -1(%[src])             \n"
+    "addiu     $t5, $t5, -1                \n"
+    "addiu     %[src], %[src], -1          \n"
+    "sb        $t0, 0(%[dst])              \n"
+    "bgez      $t5, 2b                     \n"
+    " addiu    %[dst], %[dst], 1           \n"
+
+   "3:                                     \n"
+    ".set pop                              \n"
+      : [src] "+r" (src), [dst] "+r" (dst)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4", "t5"
+  );
+}
+
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width) {
+  int x = 0;
+  int y = 0;
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "addu            $t4, %[width], %[width]      \n"
+    "srl             %[x], %[width], 4            \n"
+    "andi            %[y], %[width], 0xf          \n"
+    "blez            %[x], 2f                     \n"
+    " addu           %[src_uv], %[src_uv], $t4    \n"
+
+    ".p2align        2                            \n"
+   "1:                                            \n"
+    "lw              $t0, -32(%[src_uv])          \n"  // |3|2|1|0|
+    "lw              $t1, -28(%[src_uv])          \n"  // |7|6|5|4|
+    "lw              $t2, -24(%[src_uv])          \n"  // |11|10|9|8|
+    "lw              $t3, -20(%[src_uv])          \n"  // |15|14|13|12|
+    "lw              $t4, -16(%[src_uv])          \n"  // |19|18|17|16|
+    "lw              $t6, -12(%[src_uv])          \n"  // |23|22|21|20|
+    "lw              $t7, -8(%[src_uv])           \n"  // |27|26|25|24|
+    "lw              $t8, -4(%[src_uv])           \n"  // |31|30|29|28|
+
+    "rotr            $t0, $t0, 16                 \n"  // |1|0|3|2|
+    "rotr            $t1, $t1, 16                 \n"  // |5|4|7|6|
+    "rotr            $t2, $t2, 16                 \n"  // |9|8|11|10|
+    "rotr            $t3, $t3, 16                 \n"  // |13|12|15|14|
+    "rotr            $t4, $t4, 16                 \n"  // |17|16|19|18|
+    "rotr            $t6, $t6, 16                 \n"  // |21|20|23|22|
+    "rotr            $t7, $t7, 16                 \n"  // |25|24|27|26|
+    "rotr            $t8, $t8, 16                 \n"  // |29|28|31|30|
+    "precr.qb.ph     $t9, $t0, $t1                \n"  // |0|2|4|6|
+    "precrq.qb.ph    $t5, $t0, $t1                \n"  // |1|3|5|7|
+    "precr.qb.ph     $t0, $t2, $t3                \n"  // |8|10|12|14|
+    "precrq.qb.ph    $t1, $t2, $t3                \n"  // |9|11|13|15|
+    "precr.qb.ph     $t2, $t4, $t6                \n"  // |16|18|20|22|
+    "precrq.qb.ph    $t3, $t4, $t6                \n"  // |17|19|21|23|
+    "precr.qb.ph     $t4, $t7, $t8                \n"  // |24|26|28|30|
+    "precrq.qb.ph    $t6, $t7, $t8                \n"  // |25|27|29|31|
+    "addiu           %[src_uv], %[src_uv], -32    \n"
+    "addiu           %[x], %[x], -1               \n"
+    "swr             $t4, 0(%[dst_u])             \n"
+    "swl             $t4, 3(%[dst_u])             \n"  // |30|28|26|24|
+    "swr             $t6, 0(%[dst_v])             \n"
+    "swl             $t6, 3(%[dst_v])             \n"  // |31|29|27|25|
+    "swr             $t2, 4(%[dst_u])             \n"
+    "swl             $t2, 7(%[dst_u])             \n"  // |22|20|18|16|
+    "swr             $t3, 4(%[dst_v])             \n"
+    "swl             $t3, 7(%[dst_v])             \n"  // |23|21|19|17|
+    "swr             $t0, 8(%[dst_u])             \n"
+    "swl             $t0, 11(%[dst_u])            \n"  // |14|12|10|8|
+    "swr             $t1, 8(%[dst_v])             \n"
+    "swl             $t1, 11(%[dst_v])            \n"  // |15|13|11|9|
+    "swr             $t9, 12(%[dst_u])            \n"
+    "swl             $t9, 15(%[dst_u])            \n"  // |6|4|2|0|
+    "swr             $t5, 12(%[dst_v])            \n"
+    "swl             $t5, 15(%[dst_v])            \n"  // |7|5|3|1|
+    "addiu           %[dst_v], %[dst_v], 16       \n"
+    "bgtz            %[x], 1b                     \n"
+    " addiu          %[dst_u], %[dst_u], 16       \n"
+    "beqz            %[y], 3f                     \n"
+    " nop                                         \n"
+    "b               2f                           \n"
+    " nop                                         \n"
+
+   "2:                                            \n"
+    "lbu             $t0, -2(%[src_uv])           \n"
+    "lbu             $t1, -1(%[src_uv])           \n"
+    "addiu           %[src_uv], %[src_uv], -2     \n"
+    "addiu           %[y], %[y], -1               \n"
+    "sb              $t0, 0(%[dst_u])             \n"
+    "sb              $t1, 0(%[dst_v])             \n"
+    "addiu           %[dst_u], %[dst_u], 1        \n"
+    "bgtz            %[y], 2b                     \n"
+    " addiu          %[dst_v], %[dst_v], 1        \n"
+
+   "3:                                            \n"
+    ".set pop                                     \n"
+      : [src_uv] "+r" (src_uv),
+        [dst_u] "+r" (dst_u),
+        [dst_v] "+r" (dst_v),
+        [x] "=&r" (x),
+        [y] "+r" (y)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4",
+      "t5", "t7", "t8", "t9"
+  );
+}
+
+// Convert (4 Y and 2 VU) I422 and arrange RGB values into
+// t5 = | 0 | B0 | 0 | b0 |
+// t4 = | 0 | B1 | 0 | b1 |
+// t9 = | 0 | G0 | 0 | g0 |
+// t8 = | 0 | G1 | 0 | g1 |
+// t2 = | 0 | R0 | 0 | r0 |
+// t1 = | 0 | R1 | 0 | r1 |
+#define I422ToTransientMipsRGB                                                 \
+      "lw                $t0, 0(%[y_buf])       \n"                            \
+      "lhu               $t1, 0(%[u_buf])       \n"                            \
+      "lhu               $t2, 0(%[v_buf])       \n"                            \
+      "preceu.ph.qbr     $t1, $t1               \n"                            \
+      "preceu.ph.qbr     $t2, $t2               \n"                            \
+      "preceu.ph.qbra    $t3, $t0               \n"                            \
+      "preceu.ph.qbla    $t0, $t0               \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t3, $t3, $s4          \n"                            \
+      "subu.ph           $t0, $t0, $s4          \n"                            \
+      "mul.ph            $t3, $t3, $s0          \n"                            \
+      "mul.ph            $t0, $t0, $s0          \n"                            \
+      "shll.ph           $t4, $t1, 0x7          \n"                            \
+      "subu.ph           $t4, $t4, $t1          \n"                            \
+      "mul.ph            $t6, $t1, $s1          \n"                            \
+      "mul.ph            $t1, $t2, $s2          \n"                            \
+      "addq_s.ph         $t5, $t4, $t3          \n"                            \
+      "addq_s.ph         $t4, $t4, $t0          \n"                            \
+      "shra.ph           $t5, $t5, 6            \n"                            \
+      "shra.ph           $t4, $t4, 6            \n"                            \
+      "addiu             %[u_buf], 2            \n"                            \
+      "addiu             %[v_buf], 2            \n"                            \
+      "addu.ph           $t6, $t6, $t1          \n"                            \
+      "mul.ph            $t1, $t2, $s3          \n"                            \
+      "addu.ph           $t9, $t6, $t3          \n"                            \
+      "addu.ph           $t8, $t6, $t0          \n"                            \
+      "shra.ph           $t9, $t9, 6            \n"                            \
+      "shra.ph           $t8, $t8, 6            \n"                            \
+      "addu.ph           $t2, $t1, $t3          \n"                            \
+      "addu.ph           $t1, $t1, $t0          \n"                            \
+      "shra.ph           $t2, $t2, 6            \n"                            \
+      "shra.ph           $t1, $t1, 6            \n"                            \
+      "subu.ph           $t5, $t5, $s5          \n"                            \
+      "subu.ph           $t4, $t4, $s5          \n"                            \
+      "subu.ph           $t9, $t9, $s5          \n"                            \
+      "subu.ph           $t8, $t8, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "shll_s.ph         $t5, $t5, 8            \n"                            \
+      "shll_s.ph         $t4, $t4, 8            \n"                            \
+      "shll_s.ph         $t9, $t9, 8            \n"                            \
+      "shll_s.ph         $t8, $t8, 8            \n"                            \
+      "shll_s.ph         $t2, $t2, 8            \n"                            \
+      "shll_s.ph         $t1, $t1, 8            \n"                            \
+      "shra.ph           $t5, $t5, 8            \n"                            \
+      "shra.ph           $t4, $t4, 8            \n"                            \
+      "shra.ph           $t9, $t9, 8            \n"                            \
+      "shra.ph           $t8, $t8, 8            \n"                            \
+      "shra.ph           $t2, $t2, 8            \n"                            \
+      "shra.ph           $t1, $t1, 8            \n"                            \
+      "addu.ph           $t5, $t5, $s5          \n"                            \
+      "addu.ph           $t4, $t4, $s5          \n"                            \
+      "addu.ph           $t9, $t9, $s5          \n"                            \
+      "addu.ph           $t8, $t8, $s5          \n"                            \
+      "addu.ph           $t2, $t2, $s5          \n"                            \
+      "addu.ph           $t1, $t1, $s5          \n"
+
+void I422ToARGBRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128| // clipping
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+// Arranging into argb format
+    "precr.qb.ph       $t4, $t8, $t4          \n"  // |G1|g1|B1|b1|
+    "precr.qb.ph       $t5, $t9, $t5          \n"  // |G0|g0|B0|b0|
+    "addiu             %[width], -4           \n"
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |G1|B1|G0|B0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |g1|b1|g0|b0|
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "or                $t1, $t1, $s6          \n"  // |ff|R1|ff|R0|
+    "or                $t2, $t2, $s6          \n"  // |ff|r1|ff|r0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|r1|g1|b1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|R1|G1|B1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|r0|g0|b0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|R0|G0|B0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToABGRRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|
+
+    ".p2align          2                       \n"
+   "1:                                         \n"
+      I422ToTransientMipsRGB
+// Arranging into abgr format
+    "precr.qb.ph      $t0, $t8, $t1           \n"  // |G1|g1|R1|r1|
+    "precr.qb.ph      $t3, $t9, $t2           \n"  // |G0|g0|R0|r0|
+    "precrq.qb.ph     $t8, $t0, $t3           \n"  // |G1|R1|G0|R0|
+    "precr.qb.ph      $t9, $t0, $t3           \n"  // |g1|r1|g0|r0|
+
+    "precr.qb.ph       $t2, $t4, $t5          \n"  // |B1|b1|B0|b0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |B1|0 |B0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |b1|0 |b0|
+    "or                $t1, $t1, $s6          \n"  // |ff|B1|ff|B0|
+    "or                $t2, $t2, $s6          \n"  // |ff|b1|ff|b0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|b1|g1|r1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|B1|G1|R1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|b0|g0|r0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|B0|G0|R0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToBGRARow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74 |74 |
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff              \n"
+    "ori               $s6, 0xff              \n"  // |00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+      // Arranging into bgra format
+    "precr.qb.ph       $t4, $t4, $t8          \n"  // |B1|b1|G1|g1|
+    "precr.qb.ph       $t5, $t5, $t9          \n"  // |B0|b0|G0|g0|
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |B1|G1|B0|G0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |b1|g1|b0|g0|
+
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "sll               $t1, $t1, 8            \n"  // |R1|0 |R0|0 |
+    "sll               $t2, $t2, 8            \n"  // |r1|0 |r0|0 |
+    "or                $t1, $t1, $s6          \n"  // |R1|ff|R0|ff|
+    "or                $t2, $t2, $s6          \n"  // |r1|ff|r0|ff|
+    "precrq.ph.w       $t0, $t9, $t2          \n"  // |b1|g1|r1|ff|
+    "precrq.ph.w       $t3, $t8, $t1          \n"  // |B1|G1|R1|ff|
+    "sll               $t1, $t1, 16           \n"
+    "sll               $t2, $t2, 16           \n"
+    "packrl.ph         $t2, $t9, $t2          \n"  // |b0|g0|r0|ff|
+    "packrl.ph         $t1, $t8, $t1          \n"  // |B0|G0|R0|ff|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                ptrdiff_t src_stride, int dst_width,
+                                int source_y_fraction) {
+    int y0_fraction = 256 - source_y_fraction;
+    const uint8* src_ptr1 = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+     ".set push                                           \n"
+     ".set noreorder                                      \n"
+
+     "replv.ph          $t0, %[y0_fraction]               \n"
+     "replv.ph          $t1, %[source_y_fraction]         \n"
+
+    ".p2align           2                                 \n"
+   "1:                                                    \n"
+     "lw                $t2, 0(%[src_ptr])                \n"
+     "lw                $t3, 0(%[src_ptr1])               \n"
+     "lw                $t4, 4(%[src_ptr])                \n"
+     "lw                $t5, 4(%[src_ptr1])               \n"
+     "muleu_s.ph.qbl    $t6, $t2, $t0                     \n"
+     "muleu_s.ph.qbr    $t7, $t2, $t0                     \n"
+     "muleu_s.ph.qbl    $t8, $t3, $t1                     \n"
+     "muleu_s.ph.qbr    $t9, $t3, $t1                     \n"
+     "muleu_s.ph.qbl    $t2, $t4, $t0                     \n"
+     "muleu_s.ph.qbr    $t3, $t4, $t0                     \n"
+     "muleu_s.ph.qbl    $t4, $t5, $t1                     \n"
+     "muleu_s.ph.qbr    $t5, $t5, $t1                     \n"
+     "addq.ph           $t6, $t6, $t8                     \n"
+     "addq.ph           $t7, $t7, $t9                     \n"
+     "addq.ph           $t2, $t2, $t4                     \n"
+     "addq.ph           $t3, $t3, $t5                     \n"
+     "shra.ph           $t6, $t6, 8                       \n"
+     "shra.ph           $t7, $t7, 8                       \n"
+     "shra.ph           $t2, $t2, 8                       \n"
+     "shra.ph           $t3, $t3, 8                       \n"
+     "precr.qb.ph       $t6, $t6, $t7                     \n"
+     "precr.qb.ph       $t2, $t2, $t3                     \n"
+     "addiu             %[src_ptr], %[src_ptr], 8         \n"
+     "addiu             %[src_ptr1], %[src_ptr1], 8       \n"
+     "addiu             %[dst_width], %[dst_width], -8    \n"
+     "sw                $t6, 0(%[dst_ptr])                \n"
+     "sw                $t2, 4(%[dst_ptr])                \n"
+     "bgtz              %[dst_width], 1b                  \n"
+     " addiu            %[dst_ptr], %[dst_ptr], 8         \n"
+
+     ".set pop                                            \n"
+  : [dst_ptr] "+r" (dst_ptr),
+    [src_ptr1] "+r" (src_ptr1),
+    [src_ptr] "+r" (src_ptr),
+    [dst_width] "+r" (dst_width)
+  : [source_y_fraction] "r" (source_y_fraction),
+    [y0_fraction] "r" (y0_fraction),
+    [src_stride] "r" (src_stride)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+#endif  // __mips_dsp_rev >= 2
+
+#endif  // defined(__mips__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/row_neon.cc b/drivers/theoraplayer/src/YUV/libyuv/src/row_neon.cc
new file mode 100755
index 0000000000..68e380051b
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/row_neon.cc
@@ -0,0 +1,2847 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+// Read 8 Y, 4 U and 4 V from 422
+#define READYUV422                                                             \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.32    {d2[0]}, [%1]!                 \n"                             \
+    "vld1.32    {d2[1]}, [%2]!                 \n"
+
+// Read 8 Y, 2 U and 2 V from 422
+#define READYUV411                                                             \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.16    {d2[0]}, [%1]!                 \n"                             \
+    "vld1.16    {d2[1]}, [%2]!                 \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vzip.u8    d2, d3                         \n"
+
+// Read 8 Y, 8 U and 8 V from 444
+#define READYUV444                                                             \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vld1.8     {d3}, [%2]!                    \n"                             \
+    "vpaddl.u8  q1, q1                         \n"                             \
+    "vrshrn.u16 d2, q1, #1                     \n"
+
+// Read 8 Y, and set 4 U and 4 V to 128
+#define READYUV400                                                             \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vmov.u8    d2, #128                       \n"
+
+// Read 8 Y and 4 UV from NV12
+#define READNV12                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 Y and 4 VU from NV21
+#define READNV21                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d3, d2                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 YUY2
+#define READYUY2                                                               \
+    "vld2.8     {d0, d2}, [%0]!                \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 UYVY
+#define READUYVY                                                               \
+    "vld2.8     {d2, d3}, [%0]!                \n"                             \
+    "vmov.u8    d0, d3                         \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+#define YUV422TORGB                                                            \
+    "veor.u8    d2, d26                        \n"/*subtract 128 from u and v*/\
+    "vmull.s8   q8, d2, d24                    \n"/*  u/v B/R component      */\
+    "vmull.s8   q9, d2, d25                    \n"/*  u/v G component        */\
+    "vmov.u8    d1, #0                         \n"/*  split odd/even y apart */\
+    "vtrn.u8    d0, d1                         \n"                             \
+    "vsub.s16   q0, q0, q15                    \n"/*  offset y               */\
+    "vmul.s16   q0, q0, q14                    \n"                             \
+    "vadd.s16   d18, d19                       \n"                             \
+    "vqadd.s16  d20, d0, d16                   \n" /* B */                     \
+    "vqadd.s16  d21, d1, d16                   \n"                             \
+    "vqadd.s16  d22, d0, d17                   \n" /* R */                     \
+    "vqadd.s16  d23, d1, d17                   \n"                             \
+    "vqadd.s16  d16, d0, d18                   \n" /* G */                     \
+    "vqadd.s16  d17, d1, d18                   \n"                             \
+    "vqshrun.s16 d0, q10, #6                   \n" /* B */                     \
+    "vqshrun.s16 d1, q11, #6                   \n" /* G */                     \
+    "vqshrun.s16 d2, q8, #6                    \n" /* R */                     \
+    "vmovl.u8   q10, d0                        \n"/*  set up for reinterleave*/\
+    "vmovl.u8   q11, d1                        \n"                             \
+    "vmovl.u8   q8, d2                         \n"                             \
+    "vtrn.u8    d20, d21                       \n"                             \
+    "vtrn.u8    d22, d23                       \n"                             \
+    "vtrn.u8    d16, d17                       \n"                             \
+    "vmov.u8    d21, d16                       \n"
+
+static vec8 kUVToRB  = { 127, 127, 127, 127, 102, 102, 102, 102,
+                         0, 0, 0, 0, 0, 0, 0, 0 };
+static vec8 kUVToG = { -25, -25, -25, -25, -52, -52, -52, -52,
+                       0, 0, 0, 0, 0, 0, 0, 0 };
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+#ifdef _ANDROID
+				".fpu neon\n"
+#endif
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV444
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV411
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d19, #255                      \n"
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_bgra),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_abgr),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d19, #255                      \n"
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgba),  // %3
+      "+r"(width)      // %4
+    : "r"(&kUVToRB),   // %5
+      "r"(&kUVToG)     // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),      // %0
+      "+r"(src_u),      // %1
+      "+r"(src_v),      // %2
+      "+r"(dst_rgb24),  // %3
+      "+r"(width)       // %4
+    : "r"(&kUVToRB),    // %5
+      "r"(&kUVToG)      // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_raw),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTORGB565                                                           \
+    "vshr.u8    d20, d20, #3                   \n"  /* B                    */ \
+    "vshr.u8    d21, d21, #2                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #11                  \n"  /* R                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q0, q0, q10                    \n"  /* BGR                  */
+
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    ARGBTORGB565
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_rgb565),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB1555                                                         \
+    "vshr.u8    q10, q10, #3                   \n"  /* B                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vshr.u8    d23, d23, #7                   \n"  /* A                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vmovl.u8   q11, d23                       \n"  /* A                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #10                  \n"  /* R                    */ \
+    "vshl.u16   q11, q11, #15                  \n"  /* A                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q1, q10, q11                   \n"  /* RA                   */ \
+    "vorr       q0, q0, q1                     \n"  /* BGRA                 */
+
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB1555
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb1555),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB4444                                                         \
+    "vshr.u8    d20, d20, #4                   \n"  /* B                    */ \
+    "vbic.32    d21, d21, d4                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #4                   \n"  /* R                    */ \
+    "vbic.32    d23, d23, d4                   \n"  /* A                    */ \
+    "vorr       d0, d20, d21                   \n"  /* BG                   */ \
+    "vorr       d1, d22, d23                   \n"  /* RA                   */ \
+    "vzip.u8    d0, d1                         \n"  /* BGRA                 */
+
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%5]                    \n"
+    "vld1.8     {d25}, [%6]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB4444
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb4444),  // %3
+      "+r"(width)     // %4
+    : "r"(&kUVToRB),  // %5
+      "r"(&kUVToG)    // %6
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void YToARGBRow_NEON(const uint8* src_y,
+                     uint8* dst_argb,
+                     int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%3]                    \n"
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV400
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+    "vmov.u8    d23, #255                      \n"
+  "1:                                          \n"
+    "vld1.8     {d20}, [%0]!                   \n"
+    "vmov       d21, d20                       \n"
+    "vmov       d22, d20                       \n"
+    "subs       %2, %2, #8                     \n"
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    :
+    : "cc", "memory", "d20", "d21", "d22", "d23"
+  );
+}
+
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%4]                    \n"
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%4]                    \n"
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%4]                    \n"
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%4]                    \n"
+    "vld1.8     {d25}, [%5]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : "r"(&kUVToRB),   // %4
+      "r"(&kUVToG)     // %5
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%3]                    \n"
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUY2
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_yuy2),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vld1.8     {d24}, [%3]                    \n"
+    "vld1.8     {d25}, [%4]                    \n"
+    "vmov.u8    d26, #128                      \n"
+    "vmov.u16   q14, #74                       \n"
+    "vmov.u16   q15, #16                       \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READUYVY
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_uyvy),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : "r"(&kUVToRB),   // %3
+      "r"(&kUVToG)     // %4
+    : "cc", "memory", "q0", "q1", "q2", "q3",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pairs of UV
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vst1.8     {q0}, [%1]!                    \n"  // store U
+    "vst1.8     {q1}, [%2]!                    \n"  // store V
+    "bgt        1b                             \n"
+    : "+r"(src_uv),  // %0
+      "+r"(dst_u),   // %1
+      "+r"(dst_v),   // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load U
+    "vld1.8     {q1}, [%1]!                    \n"  // load V
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vst2.u8    {q0, q1}, [%2]!                \n"  // store 16 pairs of UV
+    "bgt        1b                             \n"
+    :
+      "+r"(src_u),   // %0
+      "+r"(src_v),   // %1
+      "+r"(dst_uv),  // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 32
+    "subs       %2, %2, #32                    \n"  // 32 processed per loop
+    "vst1.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 32
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2  // Output registers
+  :                     // Input registers
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SetRow8 writes 'count' bytes using a 32 bit value repeated.
+void SetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+    "1:                                        \n"
+    "subs      %1, %1, #16                     \n"  // 16 bytes per loop
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "q0"
+  );
+}
+
+// TODO(fbarchard): Make fully assembler
+// SetRow32 writes 'count' words using a 32 bit value repeated.
+void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
+                      int dst_stride, int height) {
+  for (int y = 0; y < height; ++y) {
+    SetRow_NEON(dst, v32, width << 2);
+    dst += dst_stride;
+  }
+}
+
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2                     \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #16                        \n"  // 16 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r12, #-16                      \n"
+    "add        %0, %0, %3, lsl #1             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {d0, d1}, [%0], r12            \n"  // src -= 16
+    "subs       %3, #8                         \n"  // 8 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // dst += 8
+    "vst1.8     {d1}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_u),   // %1
+    "+r"(dst_v),   // %2
+    "+r"(width)    // %3
+  :
+  : "cc", "memory", "r12", "q0"
+  );
+}
+
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2, lsl #2             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #4                         \n"  // 4 pixels per loop.
+    "vrev64.32  q0, q0                         \n"
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),   // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+#define RGB565TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxGGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB RRRRRxxx */ \
+    "vshl.u8    d6, d6, #2                     \n"  /* G GGGGGG00 upper 6   */ \
+    "vshr.u8    d1, d1, #3                     \n"  /* R 000RRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #6                     \n"  /* G 000000GG lower 2   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB1555TOARGB                                                         \
+    "vshrn.u16  d7, q0, #8                     \n"  /* A Arrrrrxx           */ \
+    "vshr.u8    d6, d7, #2                     \n"  /* R xxxRRRRR           */ \
+    "vshrn.u16  d5, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vmovn.u16  d4, q0                         \n"  /* B xxxBBBBB           */ \
+    "vshr.u8    d7, d7, #7                     \n"  /* A 0000000A           */ \
+    "vneg.s8    d7, d7                         \n"  /* A AAAAAAAA upper 8   */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* R RRRRR000 upper 5   */ \
+    "vshr.u8    q1, q3, #5                     \n"  /* R,A 00000RRR lower 3 */ \
+    "vshl.u8    q0, q2, #3                     \n"  /* B,G BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,G 00000BBB lower 3 */ \
+    "vorr.u8    q1, q1, q3                     \n"  /* R,A                  */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,G                  */ \
+
+// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
+#define RGB555TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB xRRRRRxx */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* G GGGGG000 upper 5   */ \
+    "vshr.u8    d1, d1, #2                     \n"  /* R 00xRRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #5                     \n"  /* G 00000GGG lower 3   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB4444TOARGB                                                         \
+    "vuzp.u8    d0, d1                         \n"  /* d0 BG, d1 RA         */ \
+    "vshl.u8    q2, q0, #4                     \n"  /* B,R BBBB0000         */ \
+    "vshr.u8    q1, q0, #4                     \n"  /* G,A 0000GGGG         */ \
+    "vshr.u8    q0, q2, #4                     \n"  /* B,R 0000BBBB         */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,R BBBBBBBB         */ \
+    "vshl.u8    q2, q1, #4                     \n"  /* G,A GGGG0000         */ \
+    "vorr.u8    q1, q1, q2                     \n"  /* G,A GGGGGGGG         */ \
+    "vswp.u8    d1, d2                         \n"  /* B,R,G,A -> B,G,R,A   */
+
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RGB24.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_rgb24),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RAW.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_raw),   // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of YUY2.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    "vst1.8     {q0}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of UYVY.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    "vst1.8     {q1}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 U.
+    "vst1.8     {d3}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 U.
+    "vst1.8     {d2}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_yuy2
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row YUY2.
+    "vrhadd.u8  d1, d1, d5                     \n"  // average rows of U
+    "vrhadd.u8  d3, d3, d7                     \n"  // average rows of V
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 U.
+    "vst1.8     {d3}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),     // %0
+    "+r"(stride_yuy2),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_uyvy
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row UYVY.
+    "vrhadd.u8  d0, d0, d4                     \n"  // average rows of U
+    "vrhadd.u8  d2, d2, d6                     \n"  // average rows of V
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 U.
+    "vst1.8     {d2}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),     // %0
+    "+r"(stride_uyvy),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load row 1 16 pixels.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vld1.8     {q1}, [%1]!                    \n"  // load row 2 16 pixels.
+    "vrhadd.u8  q0, q1                         \n"  // average row 1 and 2
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),         // %0
+    "+r"(src_uv_stride),  // %1
+    "+r"(dst_uv),         // %2
+    "+r"(pix)             // %3
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Select 2 channels from ARGB on alternating pixels.  e.g.  BGBGBGBG
+void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                         uint32 selector, int pix) {
+  asm volatile (
+    "vmov.u32   d6[0], %3                      \n"  // selector
+  "1:                                          \n"
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 8 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vtbl.8     d4, {d0, d1}, d6               \n"  // look up 4 pixels
+    "vtbl.8     d5, {d2, d3}, d6               \n"  // look up 4 pixels
+    "vtrn.u32   d4, d5                         \n"  // combine 8 pixels
+    "vst1.8     {d4}, [%1]!                    \n"  // store 8.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_bayer),  // %1
+    "+r"(pix)         // %2
+  : "r"(selector)     // %3
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Select G channels from ARGB.  e.g.  GGGGGGGG
+void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 /*selector*/, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load row 8 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 G's.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_bayer),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    "vld1.8     {q2}, [%3]                     \n"  // shuffler
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 4 pixels.
+    "subs       %2, %2, #4                     \n"  // 4 processed per loop
+    "vtbl.8     d2, {d0, d1}, d4               \n"  // look up 2 first pixels
+    "vtbl.8     d3, {d0, d1}, d5               \n"  // look up 2 next pixels
+    "vst1.8     {q1}, [%1]!                    \n"  // store 4.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 Ys
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 Us
+    "vld1.8     {d3}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 YUY2/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_yuy2),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld2.8     {d1, d3}, [%0]!                \n"  // load 16 Ys
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 Us
+    "vld1.8     {d2}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 UYVY/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_uyvy),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTORGB565
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_rgb565),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+                            int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB1555
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb1555),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB4444
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb4444),  // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// 8x1 pixels.
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.u8    d24, #112                      \n"  // UB / VR 0.875 coefficient
+    "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
+    "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
+    "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
+    "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlsl.u8   q2, d1, d25                    \n"  // G
+    "vmlsl.u8   q2, d2, d26                    \n"  // R
+    "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
+
+    "vmull.u8   q3, d2, d24                    \n"  // R
+    "vmlsl.u8   q3, d1, d28                    \n"  // G
+    "vmlsl.u8   q3, d0, d27                    \n"  // B
+    "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
+
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 32x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 32.
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%0]!     \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%0]!     \n"  // load last 8 ARGB pixels.
+    "vpaddl.u8  q4, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q5, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q6, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vpadd.u16  d0, d0, d1                     \n"  // B 16 shorts -> 8 shorts.
+    "vpadd.u16  d1, d8, d9                     \n"  // B
+    "vpadd.u16  d2, d2, d3                     \n"  // G 16 shorts -> 8 shorts.
+    "vpadd.u16  d3, d10, d11                   \n"  // G
+    "vpadd.u16  d4, d4, d5                     \n"  // R 16 shorts -> 8 shorts.
+    "vpadd.u16  d5, d12, d13                   \n"  // R
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %3, %3, #32                    \n"  // 32 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#define RGBTOUV(QB, QG, QR) \
+    "vmul.s16   q8, " #QB ", q10               \n"  /* B                    */ \
+    "vmls.s16   q8, " #QG ", q11               \n"  /* G                    */ \
+    "vmls.s16   q8, " #QR ", q12               \n"  /* R                    */ \
+    "vadd.u16   q8, q8, q15                    \n"  /* +128 -> unsigned     */ \
+    "vmul.s16   q9, " #QR ", q10               \n"  /* R                    */ \
+    "vmls.s16   q9, " #QG ", q14               \n"  /* G                    */ \
+    "vmls.s16   q9, " #QB ", q13               \n"  /* B                    */ \
+    "vadd.u16   q9, q9, q15                    \n"  /* +128 -> unsigned     */ \
+    "vqshrn.u16  d0, q8, #8                    \n"  /* 16 bit to 8 bit U    */ \
+    "vqshrn.u16  d1, q9, #8                    \n"  /* 16 bit to 8 bit V    */
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): Subsample match C code.
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
+    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
+    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
+    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
+    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
+    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
+    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
+    "vrshr.u16  q2, q2, #1                     \n"
+    "vrshr.u16  q3, q3, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q3, q2, q1)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_stride_bgra),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_stride_abgr),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
+    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
+    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_stride_rgba),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_raw
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(src_stride_rgb565),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(src_stride_argb1555),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(src_stride_argb4444),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_y),       // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // R
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // R
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // B
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_NEON(uint8* dst_ptr,
+                         const uint8* src_ptr, ptrdiff_t src_stride,
+                         int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp        %4, #0                         \n"
+    "beq        100f                           \n"
+    "add        %2, %1                         \n"
+    "cmp        %4, #64                        \n"
+    "beq        75f                            \n"
+    "cmp        %4, #128                       \n"
+    "beq        50f                            \n"
+    "cmp        %4, #192                       \n"
+    "beq        25f                            \n"
+
+    "vdup.8     d5, %4                         \n"
+    "rsb        %4, #256                       \n"
+    "vdup.8     d4, %4                         \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    "vld1.8     {q0}, [%1]!                    \n"
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vmull.u8   q13, d0, d4                    \n"
+    "vmull.u8   q14, d1, d4                    \n"
+    "vmlal.u8   q13, d2, d5                    \n"
+    "vmlal.u8   q14, d3, d5                    \n"
+    "vrshrn.u16 d0, q13, #8                    \n"
+    "vrshrn.u16 d1, q14, #8                    \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        1b                             \n"
+    "b          99f                            \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    "vld1.8     {q0}, [%1]!                    \n"
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        25b                            \n"
+    "b          99f                            \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    "vld1.8     {q0}, [%1]!                    \n"
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        50b                            \n"
+    "b          99f                            \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    "vld1.8     {q1}, [%1]!                    \n"
+    "vld1.8     {q0}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        75b                            \n"
+    "b          99f                            \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    "vld1.8     {q0}, [%1]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        100b                           \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14"
+  );
+}
+
+// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
+void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "subs       %3, #8                         \n"
+    "blt        89f                            \n"
+    // Blend 8 pixels.
+  "8:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
+    "bge        8b                             \n"
+
+  "89:                                         \n"
+    "adds       %3, #8-1                       \n"
+    "blt        99f                            \n"
+
+    // Blend 1 pixels.
+  "1:                                          \n"
+    "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+    "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+    "subs       %3, %3, #1                     \n"  // 1 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
+    "bge        1b                             \n"
+
+  "99:                                         \n"
+
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12"
+  );
+}
+
+// Attenuate 8 pixels at a time.
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    // Attenuate 8 pixels.
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d0, d3                    \n"  // b * a
+    "vmull.u8   q11, d1, d3                    \n"  // g * a
+    "vmull.u8   q12, d2, d3                    \n"  // r * a
+    "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
+    "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
+    "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "q0", "q1", "q10", "q11", "q12"
+  );
+}
+
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "vdup.u16   q8, %2                         \n"
+    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
+    "vdup.u16   q9, %3                         \n"  // interval multiply.
+    "vdup.u16   q10, %4                        \n"  // interval add
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
+    "vmovl.u8   q1, d2                         \n"
+    "vmovl.u8   q2, d4                         \n"
+    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
+    "vqdmulh.s16 q1, q1, q8                    \n"  // g
+    "vqdmulh.s16 q2, q2, q8                    \n"  // r
+    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
+    "vmul.u16   q1, q1, q9                     \n"  // g
+    "vmul.u16   q2, q2, q9                     \n"  // r
+    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
+    "vadd.u16   q1, q1, q10                    \n"  // g
+    "vadd.u16   q2, q2, q10                    \n"  // r
+    "vqmovn.u16 d0, q0                         \n"
+    "vqmovn.u16 d2, q1                         \n"
+    "vqmovn.u16 d4, q2                         \n"
+    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
+  );
+}
+
+// Shade 8 pixels at a time by specified value.
+// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
+// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "vdup.u32   q0, %3                         \n"  // duplicate scale value.
+    "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
+    "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
+    "vmovl.u8   q11, d22                       \n"
+    "vmovl.u8   q12, d24                       \n"
+    "vmovl.u8   q13, d26                       \n"
+    "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
+    "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
+    "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
+    "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
+    "vqmovn.u16 d20, q10                       \n"
+    "vqmovn.u16 d22, q11                       \n"
+    "vqmovn.u16 d24, q12                       \n"
+    "vqmovn.u16 d26, q13                       \n"
+    "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "cc", "memory", "q0", "q10", "q11", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToYJ but stores ARGB.
+// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
+    "vmov       d1, d0                         \n"  // G
+    "vmov       d2, d0                         \n"  // R
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d20, #17                       \n"  // BB coefficient
+    "vmov.u8    d21, #68                       \n"  // BG coefficient
+    "vmov.u8    d22, #35                       \n"  // BR coefficient
+    "vmov.u8    d24, #22                       \n"  // GB coefficient
+    "vmov.u8    d25, #88                       \n"  // GG coefficient
+    "vmov.u8    d26, #45                       \n"  // GR coefficient
+    "vmov.u8    d28, #24                       \n"  // BB coefficient
+    "vmov.u8    d29, #98                       \n"  // BG coefficient
+    "vmov.u8    d30, #50                       \n"  // BR coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
+    "vmlal.u8   q2, d1, d21                    \n"  // G
+    "vmlal.u8   q2, d2, d22                    \n"  // R
+    "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
+    "vmlal.u8   q3, d1, d25                    \n"  // G
+    "vmlal.u8   q3, d2, d26                    \n"  // R
+    "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
+    "vmlal.u8   q8, d1, d29                    \n"  // G
+    "vmlal.u8   q8, d2, d30                    \n"  // R
+    "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
+    "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
+    "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
+    "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),  // %0
+    "+r"(width)      // %1
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
+// needs to saturate.  Consider doing a non-saturating version.
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width) {
+  asm volatile (
+    "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
+    "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
+    "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
+    "vmovl.u8   q9, d18                        \n"  // g
+    "vmovl.u8   q10, d20                       \n"  // r
+    "vmovl.u8   q15, d22                       \n"  // a
+    "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
+    "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
+    "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
+    "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
+    "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
+    "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
+    "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
+    "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
+    "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
+    "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
+    "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q15, d0[3]                 \n"  // B += A * Matrix B
+    "vmul.s16   q5, q15, d1[3]                 \n"  // G += A * Matrix G
+    "vmul.s16   q6, q15, d2[3]                 \n"  // R += A * Matrix R
+    "vmul.s16   q7, q15, d3[3]                 \n"  // A += A * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
+    "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
+    "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
+    "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
+    "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "r"(matrix_argb)  // %3
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q0, d0, d1                     \n"  // multiply B
+    "vmull.u8   q1, d2, d3                     \n"  // multiply G
+    "vmull.u8   q2, d4, d5                     \n"  // multiply R
+    "vmull.u8   q3, d6, d7                     \n"  // multiply A
+    "vrshrn.u16 d0, q0, #8                     \n"  // 16 bit to 8 bit B
+    "vrshrn.u16 d1, q1, #8                     \n"  // 16 bit to 8 bit G
+    "vrshrn.u16 d2, q2, #8                     \n"  // 16 bit to 8 bit R
+    "vrshrn.u16 d3, q3, #8                     \n"  // 16 bit to 8 bit A
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_NEON
+
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   q0, q0, q2                     \n"  // add B, G
+    "vqadd.u8   q1, q1, q3                     \n"  // add R, A
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
+    "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0}, [%0]!                    \n"  // load 8 sobelx.
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d0, d0, d1                     \n"  // add
+    "vmov.u8    d1, d0                         \n"
+    "vmov.u8    d2, d0                         \n"
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into plane.
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    // 16 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"  // load 16 sobelx.
+    "vld1.8     {q1}, [%1]!                    \n"  // load 16 sobely.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vqadd.u8   q0, q0, q1                     \n"  // add
+    "vst1.8     {q0}, [%2]!                    \n"  // store 16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d2}, [%0]!                    \n"  // load 8 sobelx.
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d1, d0, d2                     \n"  // add
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0}, [%0],%5                  \n"  // top
+    "vld1.8     {d1}, [%0],%6                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    "vld1.8     {d2}, [%1],%5                  \n"  // center * 2
+    "vld1.8     {d3}, [%1],%6                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vld1.8     {d2}, [%2],%5                  \n"  // bottom
+    "vld1.8     {d3}, [%2],%6                  \n"
+    "subs       %4, %4, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    "vst1.8     {d0}, [%3]!                    \n"  // store 8 sobelx
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  : "r"(2),            // %5
+    "r"(6)             // %6
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0}, [%0],%4                  \n"  // left
+    "vld1.8     {d1}, [%1],%4                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    "vld1.8     {d2}, [%0],%4                  \n"  // center * 2
+    "vld1.8     {d3}, [%1],%4                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vld1.8     {d2}, [%0],%5                  \n"  // right
+    "vld1.8     {d3}, [%1],%5                  \n"
+    "subs       %3, %3, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 sobely
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  : "r"(1),            // %4
+    "r"(6)             // %5
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/row_posix.cc b/drivers/theoraplayer/src/YUV/libyuv/src/row_posix.cc
new file mode 100755
index 0000000000..106fda5689
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/row_posix.cc
@@ -0,0 +1,6443 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+// Constants for ARGB
+static vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+static vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// Constants for BGRA
+static vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR
+static vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+static vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+
+// Shuffle table for converting RGB24 to ARGB.
+static uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#if defined(TESTING) && defined(__x86_64__)
+void TestRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    ".p2align  5                               \n"
+    "mov       %%eax,%%eax                     \n"
+    "mov       %%ebx,%%ebx                     \n"
+    "mov       %%ecx,%%ecx                     \n"
+    "mov       %%edx,%%edx                     \n"
+    "mov       %%esi,%%esi                     \n"
+    "mov       %%edi,%%edi                     \n"
+    "mov       %%ebp,%%ebp                     \n"
+    "mov       %%esp,%%esp                     \n"
+    ".p2align  5                               \n"
+    "mov       %%r8d,%%r8d                     \n"
+    "mov       %%r9d,%%r9d                     \n"
+    "mov       %%r10d,%%r10d                   \n"
+    "mov       %%r11d,%%r11d                   \n"
+    "mov       %%r12d,%%r12d                   \n"
+    "mov       %%r13d,%%r13d                   \n"
+    "mov       %%r14d,%%r14d                   \n"
+    "mov       %%r15d,%%r15d                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%rax),%%eax                   \n"
+    "lea       (%%rbx),%%ebx                   \n"
+    "lea       (%%rcx),%%ecx                   \n"
+    "lea       (%%rdx),%%edx                   \n"
+    "lea       (%%rsi),%%esi                   \n"
+    "lea       (%%rdi),%%edi                   \n"
+    "lea       (%%rbp),%%ebp                   \n"
+    "lea       (%%rsp),%%esp                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%r8),%%r8d                    \n"
+    "lea       (%%r9),%%r9d                    \n"
+    "lea       (%%r10),%%r10d                  \n"
+    "lea       (%%r11),%%r11d                  \n"
+    "lea       (%%r12),%%r12d                  \n"
+    "lea       (%%r13),%%r13d                  \n"
+    "lea       (%%r14),%%r14d                  \n"
+    "lea       (%%r15),%%r15d                  \n"
+
+    ".p2align  5                               \n"
+    "lea       0x10(%%rax),%%eax               \n"
+    "lea       0x10(%%rbx),%%ebx               \n"
+    "lea       0x10(%%rcx),%%ecx               \n"
+    "lea       0x10(%%rdx),%%edx               \n"
+    "lea       0x10(%%rsi),%%esi               \n"
+    "lea       0x10(%%rdi),%%edi               \n"
+    "lea       0x10(%%rbp),%%ebp               \n"
+    "lea       0x10(%%rsp),%%esp               \n"
+    ".p2align  5                               \n"
+    "lea       0x10(%%r8),%%r8d                \n"
+    "lea       0x10(%%r9),%%r9d                \n"
+    "lea       0x10(%%r10),%%r10d              \n"
+    "lea       0x10(%%r11),%%r11d              \n"
+    "lea       0x10(%%r12),%%r12d              \n"
+    "lea       0x10(%%r13),%%r13d              \n"
+    "lea       0x10(%%r14),%%r14d              \n"
+    "lea       0x10(%%r15),%%r15d              \n"
+
+    ".p2align  5                               \n"
+    "add       0x10,%%eax                      \n"
+    "add       0x10,%%ebx                      \n"
+    "add       0x10,%%ecx                      \n"
+    "add       0x10,%%edx                      \n"
+    "add       0x10,%%esi                      \n"
+    "add       0x10,%%edi                      \n"
+    "add       0x10,%%ebp                      \n"
+    "add       0x10,%%esp                      \n"
+    ".p2align  5                               \n"
+    "add       0x10,%%r8d                      \n"
+    "add       0x10,%%r9d                      \n"
+    "add       0x10,%%r10d                     \n"
+    "add       0x10,%%r11d                     \n"
+    "add       0x10,%%r12d                     \n"
+    "add       0x10,%%r13d                     \n"
+    "add       0x10,%%r14d                     \n"
+    "add       0x10,%%r15d                     \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // TESTING
+
+#ifdef HAS_I400TOARGBROW_SSE2
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
+                                  int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_I400TOARGBROW_SSE2
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRGB24ToARGB)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRAWToARGB)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x20802080,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xa,%%xmm4                     \n"
+    "psrlw     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm1,0x00,1,0,2)           //  movdqa  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm2,0x10,1,0,2)           //  movdqa  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x42004200,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "movdqa    %%xmm3,%%xmm4                   \n"
+    "psrlw     $0x6,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psllw     $0x1,%%xmm1                     \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm1,0x00,1,0,2)           //  movdqa  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm2,0x10,1,0,2)           //  movdqa  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0xf0f0f0f,%%eax                \n"
+    "movd      %%eax,%%xmm4                    \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x4,%%xmm5                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "psllw     $0x4,%%xmm1                     \n"
+    "psrlw     $0x4,%%xmm3                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,2)           //  movdqa  %%xmm0,(%1,%0,2)
+    MEMOPMEM(movdqa,xmm1,0x10,1,0,2)           //  movdqa  %%xmm1,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRGB24)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRAW)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psrld     $0x1b,%%xmm3                    \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1a,%%xmm4                    \n"
+    "pslld     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0xb,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pslld     $0x8,%%xmm0                     \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x5,%%xmm2                     \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1b,%%xmm4                    \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x5,%%xmm5                     \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "pslld     $0xa,%%xmm6                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "pslld     $0xf,%%xmm7                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x6,%%xmm2                     \n"
+    "psrld     $0x9,%%xmm3                     \n"
+    "pand      %%xmm7,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm6,%%xmm3                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMACCESS2(0x8,1) ",%1        \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm4,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm3,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "psrlq     $0x4,%%xmm0                     \n"
+    "psrlq     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#ifdef HAS_ARGBTOYROW_SSSE3
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+#ifdef HAS_ARGBTOYJROW_SSSE3
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),  // %3
+    "m"(kAddYJ64)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),  // %3
+    "m"(kAddYJ64)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOYJROW_SSSE3
+
+#ifdef HAS_ARGBTOUVROW_SSSE3
+// TODO(fbarchard): pass xmm constants to single block of assembly.
+// fpic on GCC 4.2 for OSX runs out of GPR registers. "m" effectively takes
+// 3 registers - ebx, ebp and eax. "m" can be passed with 3 normal registers,
+// or 4 if stack frame is disabled. Doing 2 assembly blocks is a work around
+// and considered unsafe.
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps    %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+// TODO(fbarchard): Share code with ARGBToUVRow_SSSE3.
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToUJ),  // %0
+    "m"(kARGBToVJ),  // %1
+    "m"(kAddUVJ128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),         // %0
+    "m"(kARGBToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                  uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToUJ),         // %0
+    "m"(kARGBToVJ),         // %1
+    "m"(kAddUVJ128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb))
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                          int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,2,1)           //  movdqa  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),        // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6"
+#endif
+  );
+}
+
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_u,
+                                    uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,2,1)           //  movdqu  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),        // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6"
+#endif
+  );
+}
+
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kARGBToU),  // %0
+    "m"(kARGBToV),  // %1
+    "m"(kAddUV128)  // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kBGRAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kBGRAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kBGRAToU),         // %0
+    "m"(kBGRAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_bgra)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kBGRAToU),         // %0
+    "m"(kBGRAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_bgra)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kABGRToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kABGRToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kRGBAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kRGBAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kABGRToU),         // %0
+    "m"(kABGRToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_abgr)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kABGRToU),         // %0
+    "m"(kABGRToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_abgr)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kRGBAToU),         // %0
+    "m"(kRGBAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(pavgb,0x00,0,4,1,xmm0)            //  pavgb   (%0,%4,1),%%xmm0
+    MEMOPREG(pavgb,0x10,0,4,1,xmm1)            //  pavgb   0x10(%0,%4,1),%%xmm1
+    MEMOPREG(pavgb,0x20,0,4,1,xmm2)            //  pavgb   0x20(%0,%4,1),%%xmm2
+    MEMOPREG(pavgb,0x30,0,4,1,xmm6)            //  pavgb   0x30(%0,%4,1),%%xmm6
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_rgba))
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kRGBAToU),         // %0
+    "m"(kRGBAToV),         // %1
+    "m"(kAddUV128)         // %2
+  );
+  asm volatile (
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)           //  movdqu  (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)           //  movdqu  0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)           //  movdqu  0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)           //  movdqu  0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_rgba)) // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOUVROW_SSSE3
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+#define UB 127 /* min(63,(int8)(2.018 * 64)) */
+#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
+
+struct {
+  vec8 kUVToB;  // 0
+  vec8 kUVToG;  // 16
+  vec8 kUVToR;  // 32
+  vec16 kUVBiasB;  // 48
+  vec16 kUVBiasG;  // 64
+  vec16 kUVBiasR;  // 80
+  vec16 kYSub16;  // 96
+  vec16 kYToRgb;  // 112
+  vec8 kVUToB;  // 128
+  vec8 kVUToG;  // 144
+  vec8 kVUToR;  // 160
+} static SIMD_ALIGNED(kYuvConstants) = {
+  { UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR },
+  { 16, 16, 16, 16, 16, 16, 16, 16 },
+  { YG, YG, YG, YG, YG, YG, YG, YG },
+  { VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR }
+};
+
+
+// Read 8 UV from 411
+#define READYUV444                                                             \
+    "movq       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"
+
+// Read 4 UV from 422, upsample to 8 UV
+#define READYUV422                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Read 2 UV from 411, upsample to 8 UV
+#define READYUV411                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    BUNDLEALIGN                                                                \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x2, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
+    "punpckldq  %%xmm0,%%xmm0                                   \n"
+
+// Read 4 UV from NV12, upsample to 8 UV
+#define READNV12                                                               \
+    "movq       " MEMACCESS([uv_buf]) ",%%xmm0                  \n"            \
+    "lea        " MEMLEA(0x8, [uv_buf]) ",%[uv_buf]             \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Convert 8 pixels: 8 UV and 8 Y
+#define YUVTORGB                                                               \
+    "movdqa     %%xmm0,%%xmm1                                   \n"            \
+    "movdqa     %%xmm0,%%xmm2                                   \n"            \
+    "pmaddubsw  " MEMACCESS([kYuvConstants]) ",%%xmm0           \n"            \
+    "pmaddubsw  " MEMACCESS2(16, [kYuvConstants]) ",%%xmm1      \n"            \
+    "pmaddubsw  " MEMACCESS2(32, [kYuvConstants]) ",%%xmm2      \n"            \
+    "psubw      " MEMACCESS2(48, [kYuvConstants]) ",%%xmm0      \n"            \
+    "psubw      " MEMACCESS2(64, [kYuvConstants]) ",%%xmm1      \n"            \
+    "psubw      " MEMACCESS2(80, [kYuvConstants]) ",%%xmm2      \n"            \
+    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
+    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
+    "punpcklbw  %%xmm4,%%xmm3                                   \n"            \
+    "psubsw     " MEMACCESS2(96, [kYuvConstants]) ",%%xmm3      \n"            \
+    "pmullw     " MEMACCESS2(112, [kYuvConstants]) ",%%xmm3     \n"            \
+    "paddsw     %%xmm3,%%xmm0                                   \n"            \
+    "paddsw     %%xmm3,%%xmm1                                   \n"            \
+    "paddsw     %%xmm3,%%xmm2                                   \n"            \
+    "psraw      $0x6,%%xmm0                                     \n"            \
+    "psraw      $0x6,%%xmm1                                     \n"            \
+    "psraw      $0x6,%%xmm2                                     \n"            \
+    "packuswb   %%xmm0,%%xmm0                                   \n"            \
+    "packuswb   %%xmm1,%%xmm1                                   \n"            \
+    "packuswb   %%xmm2,%%xmm2                                   \n"
+
+// Convert 8 pixels: 8 VU and 8 Y
+#define YVUTORGB                                                               \
+    "movdqa     %%xmm0,%%xmm1                                   \n"            \
+    "movdqa     %%xmm0,%%xmm2                                   \n"            \
+    "pmaddubsw  " MEMACCESS2(128, [kYuvConstants]) ",%%xmm0     \n"            \
+    "pmaddubsw  " MEMACCESS2(144, [kYuvConstants]) ",%%xmm1     \n"            \
+    "pmaddubsw  " MEMACCESS2(160, [kYuvConstants]) ",%%xmm2     \n"            \
+    "psubw      " MEMACCESS2(48, [kYuvConstants]) ",%%xmm0      \n"            \
+    "psubw      " MEMACCESS2(64, [kYuvConstants]) ",%%xmm1      \n"            \
+    "psubw      " MEMACCESS2(80, [kYuvConstants]) ",%%xmm2      \n"            \
+    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
+    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
+    "punpcklbw  %%xmm4,%%xmm3                                   \n"            \
+    "psubsw     " MEMACCESS2(96, [kYuvConstants]) ",%%xmm3      \n"            \
+    "pmullw     " MEMACCESS2(112, [kYuvConstants]) ",%%xmm3     \n"            \
+    "paddsw     %%xmm3,%%xmm0                                   \n"            \
+    "paddsw     %%xmm3,%%xmm1                                   \n"            \
+    "paddsw     %%xmm3,%%xmm2                                   \n"            \
+    "psraw      $0x6,%%xmm0                                     \n"            \
+    "psraw      $0x6,%%xmm1                                     \n"            \
+    "psraw      $0x6,%%xmm2                                     \n"            \
+    "packuswb   %%xmm0,%%xmm0                                   \n"            \
+    "packuswb   %%xmm1,%%xmm1                                   \n"            \
+    "packuswb   %%xmm2,%%xmm2                                   \n"
+
+void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "   \n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb]  \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                                 const uint8* u_buf,
+                                 const uint8* v_buf,
+                                 uint8* dst_rgb24,
+                                 int width) {
+// fpic 32 bit gcc 4.2 on OSX runs out of GPR regs.
+#if defined(__i386__)
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
+  :: [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24));
+#endif
+
+  asm volatile (
+#if !defined(__i386__)
+    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
+#endif
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_rgb24]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_rgb24]) "\n"
+    "lea       " MEMLEA(0x18,[dst_rgb24]) ",%[dst_rgb24] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgb24]"+r"(dst_rgb24),  // %[dst_rgb24]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)
+#if !defined(__i386__)
+    , [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
+#endif
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_raw,
+                               int width) {
+// fpic 32 bit gcc 4.2 on OSX runs out of GPR regs.
+#if defined(__i386__)
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
+  :: [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW));
+#endif
+
+  asm volatile (
+#if !defined(__i386__)
+    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
+#endif
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_raw]) " \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_raw]) "\n"
+    "lea       " MEMLEA(0x18,[dst_raw]) ",%[dst_raw] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_raw]"+r"(dst_raw),  // %[dst_raw]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)
+#if !defined(__i386__)
+    , [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW)
+#endif
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV411
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YVUTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV411
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* uv_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* uv_buf,
+                                          uint8* dst_argb,
+                                          int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YVUTORGB
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS([dst_argb]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x20,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+  // Does not use r14.
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_bgra,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm5," MEMACCESS([dst_bgra]) "\n"
+    "movdqa    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_abgr,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm2                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2," MEMACCESS([dst_abgr]) "\n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "\n"
+    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_rgba,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm2,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm5," MEMACCESS([dst_rgba]) "\n"
+    "movdqa    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "\n"
+    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_bgra,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm5," MEMACCESS([dst_bgra]) "\n"
+    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x20,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_abgr,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm2                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2," MEMACCESS([dst_abgr]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,[dst_abgr]) "\n"
+    "lea       " MEMLEA(0x20,[dst_abgr]) ",%[dst_abgr] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void OMITFP I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
+                                          const uint8* u_buf,
+                                          const uint8* v_buf,
+                                          uint8* dst_rgba,
+                                          int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm2,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm5,%%xmm0                   \n"
+    "punpcklwd %%xmm1,%%xmm5                   \n"
+    "punpckhwd %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm5," MEMACCESS([dst_rgba]) "\n"
+    "movdqu    %%xmm0," MEMACCESS2(0x10,[dst_rgba]) "\n"
+    "lea       " MEMLEA(0x20,[dst_rgba]) ",%[dst_rgba] \n"
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+#ifdef HAS_YTOARGBROW_SSE2
+void YToARGBRow_SSE2(const uint8* y_buf,
+                     uint8* dst_argb,
+                     int width) {
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "mov       $0x00100010,%%eax               \n"
+    "movd      %%eax,%%xmm3                    \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "mov       $0x004a004a,%%eax               \n"
+    "movd      %%eax,%%xmm2                    \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    LABELALIGN
+  "1:                                          \n"
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "psubusw   %%xmm3,%%xmm0                   \n"
+    "pmullw    %%xmm2,%%xmm0                   \n"
+    "psrlw     $6, %%xmm0                      \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+
+    // Step 2: Weave into ARGB
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "por       %%xmm4,%%xmm1                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(y_buf),     // %0
+    "+r"(dst_argb),  // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc", "eax"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+#endif  // HAS_YTOARGBROW_SSE2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %3,%%xmm5                       \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqa,0x00,0,2,1,xmm0)           //  movdqa  (%0,%2),%%xmm0
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kShuffleMirror) // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_SSE2
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqu,0x00,0,2,1,xmm0)           //  movdqu  (%0,%2),%%xmm0
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufd    $0x4e,%%xmm0,%%xmm0             \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1)",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %4,%%xmm1                       \n"
+    "lea       " MEMLEA4(-0x10,0,3,2) ",%0       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(-0x10,0) ",%0            \n"
+    "pshufb    %%xmm1,%%xmm0                   \n"
+    "sub       $8,%3                           \n"
+    "movlpd    %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movhpd,xmm0,0x00,1,2,1)           //  movhpd    %%xmm0,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src),      // %0
+    "+r"(dst_u),    // %1
+    "+r"(dst_v),    // %2
+    "+r"(temp_width)  // %3
+  : "m"(kShuffleMirrorUV)  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static uvec8 kARGBShuffleMirror = {
+  12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u
+};
+
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "lea       " MEMLEA4(-0x10,0,2,4) ",%0     \n"
+    "movdqa    %3,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kARGBShuffleMirror)  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_SSSE3
+
+#ifdef HAS_SPLITUVROW_SSE2
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb    %%xmm5,%%xmm5                    \n"
+    "psrlw      $0x8,%%xmm5                      \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqa     " MEMACCESS(0) ",%%xmm0          \n"
+    "movdqa     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
+    "lea        " MEMLEA(0x20,0) ",%0            \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "pand       %%xmm5,%%xmm0                    \n"
+    "pand       %%xmm5,%%xmm1                    \n"
+    "packuswb   %%xmm1,%%xmm0                    \n"
+    "psrlw      $0x8,%%xmm2                      \n"
+    "psrlw      $0x8,%%xmm3                      \n"
+    "packuswb   %%xmm3,%%xmm2                    \n"
+    "movdqa     %%xmm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(movdqa,xmm2,0x00,1,2,1)             // movdqa     %%xmm2,(%1,%2)
+    "lea        " MEMLEA(0x10,1) ",%1            \n"
+    "sub        $0x10,%3                         \n"
+    "jg         1b                               \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix) {
+  asm volatile (
+    "pcmpeqb    %%xmm5,%%xmm5                    \n"
+    "psrlw      $0x8,%%xmm5                      \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu     " MEMACCESS(0) ",%%xmm0          \n"
+    "movdqu     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
+    "lea        " MEMLEA(0x20,0) ",%0            \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "pand       %%xmm5,%%xmm0                    \n"
+    "pand       %%xmm5,%%xmm1                    \n"
+    "packuswb   %%xmm1,%%xmm0                    \n"
+    "psrlw      $0x8,%%xmm2                      \n"
+    "psrlw      $0x8,%%xmm3                      \n"
+    "packuswb   %%xmm3,%%xmm2                    \n"
+    "movdqu     %%xmm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(movdqu,xmm2,0x00,1,2,1)             //  movdqu     %%xmm2,(%1,%2)
+    "lea        " MEMLEA(0x10,1) ",%1            \n"
+    "sub        $0x10,%3                         \n"
+    "jg         1b                               \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_SSE2
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0           \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)             //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm2                     \n"
+    "punpcklbw %%xmm1,%%xmm0                     \n"
+    "punpckhbw %%xmm1,%%xmm2                     \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "           \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x10,2) "     \n"
+    "lea       " MEMLEA(0x20,2) ",%2             \n"
+    "sub       $0x10,%3                          \n"
+    "jg        1b                                \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)             //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm2                     \n"
+    "punpcklbw %%xmm1,%%xmm0                     \n"
+    "punpckhbw %%xmm1,%%xmm2                     \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "     \n"
+    "lea       " MEMLEA(0x20,2) ",%2             \n"
+    "sub       $0x10,%3                          \n"
+    "jg        1b                                \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_COPYROW_SSE2
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_COPYROW_SSE2
+
+#ifdef HAS_COPYROW_X86
+void CopyRow_X86(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "shr       $0x2,%2                         \n"
+    "rep movsl " MEMMOVESTRING(0,1) "          \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+#endif  // HAS_COPYROW_X86
+
+#ifdef HAS_COPYROW_ERMS
+// Unaligned Multiple of 1.
+void CopyRow_ERMS(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep movsb " MEMMOVESTRING(0,1) "          \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+#endif  // HAS_COPYROW_ERMS
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqa    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm1         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm2   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpckhwd %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm2,%%xmm2                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqa    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd " MEMACCESS(0) ",%%ymm1         \n"
+    "vpmovzxbd " MEMACCESS2(0x8,0) ",%%ymm2    \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "vpslld    $0x18,%%ymm1,%%ymm1             \n"
+    "vpslld    $0x18,%%ymm2,%%ymm2             \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+void SetRow_X86(uint8* dst, uint32 v32, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "shr       $0x2,%1                         \n"
+    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
+    : "+D"(dst),       // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v32)         // %2
+    : "memory", "cc");
+}
+
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                   int dst_stride, int height) {
+  for (int y = 0; y < height; ++y) {
+    size_t width_tmp = (size_t)(width);
+    uint32* d = (uint32*)(dst);
+    asm volatile (
+      "rep stosl " MEMSTORESTRING(eax,0) "     \n"
+      : "+D"(d),         // %0
+        "+c"(width_tmp)  // %1
+      : "a"(v32)         // %2
+      : "memory", "cc");
+    dst += dst_stride;
+  }
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_SSE2
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2,
+                                int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x1,%3                         \n"
+    "je        91f                             \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop until destination pointer is aligned.
+  "10:                                         \n"
+    "test      $0xf,%2                         \n"
+    "je        19f                             \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+    "add       $1-4,%3                         \n"
+    "jl        49f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "41:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       41b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+
+// Blend 8 pixels at a time
+// Shuffle table for reversing the bytes.
+
+// Same as SSE2, but replaces
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3,0F5h  // 8 alpha words
+//    pshuflw    xmm3, xmm3,0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x1,%3                         \n"
+    "je        91f                             \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop until destination pointer is aligned.
+  "10:                                         \n"
+    "test      $0xf,%2                         \n"
+    "je        19f                             \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+    "add       $1-4,%3                         \n"
+    "jl        49f                             \n"
+    "test      $0xf,%0                         \n"
+    "jne       41f                             \n"
+    "test      $0xf,%1                         \n"
+    "jne       41f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "40:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+    "jmp       49f                             \n"
+
+    // 4 pixel unaligned loop.
+    LABELALIGN
+  "41:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       41b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x1,%3                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  : "m"(kShuffleAlpha)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x8,%%xmm5                     \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "pshufhw   $0xff,%%xmm0,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pshufhw   $0xff,%%xmm1,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha
+static uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+};
+static uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+};
+// Attenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "pslld     $0x18,%%xmm3                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpcklbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm1,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "punpckhbw %%xmm2,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pand      %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  : "m"(kShuffleAlpha0),  // %3
+    "m"(kShuffleAlpha1)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+// aligned to 16 bytes
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8)  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrld     $0x18,%%xmm2                    \n"
+    "psrld     $0x18,%%xmm3                    \n"
+    "packuswb  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm3                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "m"(kARGBToYJ),   // %3
+    "m"(kAddYJ64)     // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone
+static vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %2,%%xmm2                       \n"
+    "movdqa    %3,%%xmm3                       \n"
+    "movdqa    %4,%%xmm4                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm6                   \n"
+    "phaddw    %%xmm6,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm5                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm5                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "psrld     $0x18,%%xmm6                    \n"
+    "psrld     $0x18,%%xmm1                    \n"
+    "packuswb  %%xmm1,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm5                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "punpckhwd %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%1                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),      // %0
+    "+r"(width)          // %1
+  : "m"(kARGBToSepiaB),  // %2
+    "m"(kARGBToSepiaG),  // %3
+    "m"(kARGBToSepiaR)   // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  asm volatile (
+    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
+    "pshufd    $0x00,%%xmm5,%%xmm2             \n"
+    "pshufd    $0x55,%%xmm5,%%xmm3             \n"
+    "pshufd    $0xaa,%%xmm5,%%xmm4             \n"
+    "pshufd    $0xff,%%xmm5,%%xmm5             \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm7                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddsw   %%xmm7,%%xmm0                   \n"
+    "phaddsw   %%xmm1,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm0                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm1                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm1                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "punpcklwd %%xmm1,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm6                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb),      // %1
+    "+r"(width)          // %2
+  : "r"(matrix_argb)     // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+// aligned to 16 bytes
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "movd      %2,%%xmm2                       \n"
+    "movd      %3,%%xmm3                       \n"
+    "movd      %4,%%xmm4                       \n"
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshufd    $0x44,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "pshufd    $0x44,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0x40,%%xmm4,%%xmm4             \n"
+    "pshufd    $0x44,%%xmm4,%%xmm4             \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "pslld     $0x18,%%xmm6                    \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "pmullw    %%xmm3,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm7         \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm6,%%xmm7                   \n"
+    "paddw     %%xmm4,%%xmm0                   \n"
+    "paddw     %%xmm4,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "sub       $0x4,%1                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+// Aligned to 16 bytes.
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "movd      %3,%%xmm2                       \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm2                  \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(value)       // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpckhbw %%xmm5,%%xmm3                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels, 4 pixels at a time.
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psubusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "sub       %0,%3                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm1    \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    MEMOPREG(movq,0x02,0,1,1,xmm2)             //  movq      0x2(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,0,2,1,xmm2)             //  movq      (%0,%2,1),%%xmm2
+    MEMOPREG(movq,0x02,0,2,1,xmm3)             //  movq      0x2(%0,%2,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x8,%4                         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm0,0x00,0,3,1)             //  movq      %%xmm0,(%0,%3,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    "movq      " MEMACCESS2(0x1,0) ",%%xmm1    \n"
+    MEMOPREG(movq,0x01,0,1,1,xmm2)             //  movq      0x1(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    BUNDLEALIGN
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm2    \n"
+    MEMOPREG(movq,0x02,0,1,1,xmm3)             //  movq      0x2(%0,%1,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x8,%3                         \n"
+    BUNDLEALIGN
+    MEMOPMEM(movq,xmm0,0x00,0,2,1)             //  movq      %%xmm0,(%0,%2,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm2                   \n"
+    "punpckhbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm1                   \n"
+    "punpckhwd %%xmm2,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklwd %%xmm0,%%xmm3                   \n"
+    "punpckhwd %%xmm0,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm1," MEMACCESS(2) "         \n"
+    "movdqa    %%xmm2," MEMACCESS2(0x10,2) "   \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x20,2) "   \n"
+    "movdqa    %%xmm0," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,1,1,xmm1)           //  movdqa    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "paddusb   %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "punpckhbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "punpcklbw %%xmm2,%%xmm4                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "punpcklwd %%xmm3,%%xmm6                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "punpcklwd %%xmm0,%%xmm7                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "sub       $0x10,%3                        \n"
+    "movdqa    %%xmm6," MEMACCESS(2) "         \n"
+    "movdqa    %%xmm4," MEMACCESS2(0x10,2) "   \n"
+    "movdqa    %%xmm7," MEMACCESS2(0x20,2) "   \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value, inclusive of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  asm volatile (
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "test      $0xf,%1                         \n"
+    "jne       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "punpckhwd %%xmm1,%%xmm3                   \n"
+    "punpckhbw %%xmm1,%%xmm4                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "punpcklwd %%xmm1,%%xmm4                   \n"
+    "punpckhwd %%xmm1,%%xmm5                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqa    " MEMACCESS(2) ",%%xmm2         \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x10,2) ",%%xmm3   \n"
+    "paddd     %%xmm0,%%xmm3                   \n"
+    "paddd     %%xmm4,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x20,2) ",%%xmm4   \n"
+    "paddd     %%xmm0,%%xmm4                   \n"
+    "paddd     %%xmm5,%%xmm0                   \n"
+    "movdqa    " MEMACCESS2(0x30,2) ",%%xmm5   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm5                   \n"
+    "movdqa    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "movdqa    %%xmm4," MEMACCESS2(0x20,1) "   \n"
+    "movdqa    %%xmm5," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+  : "+r"(row),  // %0
+    "+r"(cumsum),  // %1
+    "+r"(previous_cumsum),  // %2
+    "+r"(width)  // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  asm volatile (
+    "movd      %5,%%xmm5                       \n"
+    "cvtdq2ps  %%xmm5,%%xmm5                   \n"
+    "rcpss     %%xmm5,%%xmm4                   \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "cmpl      $0x80,%5                        \n"
+    "ja        40f                             \n"
+
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrld     $0x10,%%xmm6                    \n"
+    "cvtdq2ps  %%xmm6,%%xmm6                   \n"
+    "addps     %%xmm6,%%xmm5                   \n"
+    "mulps     %%xmm4,%%xmm5                   \n"
+    "cvtps2dq  %%xmm5,%%xmm5                   \n"
+    "packssdw  %%xmm5,%%xmm5                   \n"
+
+  // 4 pixel small loop                        \n"
+    LABELALIGN
+  "4:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       4b                              \n"
+    "jmp       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqa    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqa    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm1,%%xmm1                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm1                   \n"
+    "cvtdq2ps  %%xmm2,%%xmm2                   \n"
+    "cvtdq2ps  %%xmm3,%%xmm3                   \n"
+    "mulps     %%xmm4,%%xmm2                   \n"
+    "mulps     %%xmm4,%%xmm3                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "cvtps2dq  %%xmm1,%%xmm1                   \n"
+    "cvtps2dq  %%xmm2,%%xmm2                   \n"
+    "cvtps2dq  %%xmm3,%%xmm3                   \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    BUNDLEALIGN
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(topleft),  // %0
+    "+r"(botleft),  // %1
+    "+r"(dst),      // %2
+    "+rm"(count)    // %3
+  : "r"((intptr_t)(width)),  // %4
+    "rm"(area)     // %5
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* src_dudv, int width) {
+  intptr_t src_argb_stride_temp = src_argb_stride;
+  intptr_t temp = 0;
+  asm volatile (
+    "movq      " MEMACCESS(3) ",%%xmm2         \n"
+    "movq      " MEMACCESS2(0x08,3) ",%%xmm7   \n"
+    "shl       $0x10,%1                        \n"
+    "add       $0x4,%1                         \n"
+    "movd      %1,%%xmm5                       \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    "pshufd    $0x44,%%xmm7,%%xmm7             \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm0                   \n"
+    "movlhps   %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm7,%%xmm4                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"  // x, y float to int first 2
+    "cvttps2dq %%xmm3,%%xmm1                   \n"  // x, y float to int next 2
+    "packssdw  %%xmm1,%%xmm0                   \n"  // x, y as 8 shorts
+    "pmaddwd   %%xmm5,%%xmm0                   \n"  // off = x * 4 + y * stride
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm1                   \n"
+    "addps     %%xmm4,%%xmm2                   \n"
+    "movq      %%xmm1," MEMACCESS(2) "         \n"
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm0                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "sub       $0x4,%4                         \n"
+    "movq      %%xmm0," MEMACCESS2(0x08,2) "   \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%4                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "pmaddwd   %%xmm5,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm2                   \n"
+    "movd      %%xmm0,%k1                      \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    "sub       $0x1,%4                         \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x04,2) ",%2           \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_stride_temp),  // %1
+    "+r"(dst_argb),  // %2
+    "+r"(src_dudv),  // %3
+    "+rm"(width),    // %4
+    "+r"(temp)   // %5
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm2)
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "pmaddubsw %%xmm5,%%xmm0                   \n"
+    "pmaddubsw %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm0)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm2)           //  movdqa    (%1,%4,1),%%xmm2
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm2                   \n"
+    "psubw     %%xmm1,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm2                   \n"
+    "paddw     %%xmm3,%%xmm3                   \n"
+    "pmulhw    %%xmm5,%%xmm2                   \n"
+    "pmulhw    %%xmm5,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)           //  movdqa    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm1)           //  movdqa    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqa,0x00,1,4,1,xmm0)           //  movdqa    (%1,%4,1),%%xmm0
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,1,0,1)           //  movdqa    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride, int dst_width,
+                                    int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "pmaddubsw %%xmm5,%%xmm0                   \n"
+    "pmaddubsw %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm5"
+#endif
+  );
+}
+#endif   // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride, int dst_width,
+                                   int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)           //  movdqu    (%1,%4,1),%%xmm2
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm2                   \n"
+    "psubw     %%xmm1,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm2                   \n"
+    "paddw     %%xmm3,%%xmm3                   \n"
+    "pmulhw    %%xmm5,%%xmm2                   \n"
+    "pmulhw    %%xmm5,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)           //  movdqu    (%1,%4,1),%%xmm0
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    BUNDLEALIGN
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+#ifdef HAS_HALFROW_SSE2
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb     (%0,%3),%%xmm0
+    "sub       $0x10,%2                        \n"
+    MEMOPMEM(movdqa,xmm0,0x00,0,1,1)           //  movdqa    %%xmm0,(%0,%1)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "jg        1b                              \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_uv),  // %1
+    "+r"(pix)      // %2
+  : "r"((intptr_t)(src_uv_stride))  // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+      , "xmm0"
+#endif
+  );
+}
+#endif  // HAS_HALFROW_SSE2
+
+#ifdef HAS_ARGBTOBAYERROW_SSSE3
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix) {
+  asm volatile (
+    // NaCL caveat - assumes movd is from GPR
+    "movd      %3,%%xmm5                       \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    "sub       $0x8,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_bayer), // %1
+    "+r"(pix)        // %2
+  : "g"(selector)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOBAYERROW_SSSE3
+
+#ifdef HAS_ARGBTOBAYERGGROW_SSE2
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 selector, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrld     $0x8,%%xmm0                     \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x8,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_bayer), // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBTOBAYERGGROW_SSE2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  asm volatile (
+    "movdqa    " MEMACCESS(3) ",%%xmm5         \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix) {
+  asm volatile (
+    "movdqa    " MEMACCESS(3) ",%%xmm5         \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSSE3
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm5    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpshufb   %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpshufb   %%ymm5,%%ymm1,%%ymm1            \n"
+    "sub       $0x10,%2                        \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "mov       " MEMACCESS(4) ",%k2            \n"
+    "cmp       $0x3000102,%k2                  \n"
+    "je        3012f                           \n"
+    "cmp       $0x10203,%k2                    \n"
+    "je        123f                            \n"
+    "cmp       $0x30201,%k2                    \n"
+    "je        321f                            \n"
+    "cmp       $0x2010003,%k2                  \n"
+    "je        2103f                           \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(4) ",%2             \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS(1) "            \n"
+    "movzb     " MEMACCESS2(0x1,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x1,1) "       \n"
+    BUNDLEALIGN
+    "movzb     " MEMACCESS2(0x2,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x2,1) "       \n"
+    "movzb     " MEMACCESS2(0x3,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x3,1) "       \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "sub       $0x1,%3                         \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "123:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x1b,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        123b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "321:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x39,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x39,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        321b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "2103:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x93,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x93,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        2103b                           \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "3012:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0xc6,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0xc6,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        3012b                           \n"
+
+  "99:                                         \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+d"(pixel_temp),  // %2
+    "+r"(pix)         // %3
+  : "r"(shuffler)      // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSE2
+
+#ifdef HAS_I422TOYUY2ROW_SSE2
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub       %1,%2                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm1                     \n"
+    "punpcklbw %%xmm2,%%xmm0                     \n"
+    "punpckhbw %%xmm2,%%xmm1                     \n"
+    "movdqu    %%xmm0," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+#endif  // HAS_I422TOYUY2ROW_SSE2
+
+#ifdef HAS_I422TOUYVYROW_SSE2
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "movdqa    %%xmm2,%%xmm1                     \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "punpcklbw %%xmm0,%%xmm1                     \n"
+    "punpckhbw %%xmm0,%%xmm2                     \n"
+    "movdqu    %%xmm1," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+#endif  // HAS_I422TOUYVYROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "pxor      %%xmm3,%%xmm3                   \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm3,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm4                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm0   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm4   \n"
+    "addps     " MEMACCESS(3) ",%%xmm0         \n"
+    "addps     " MEMACCESS(3) ",%%xmm4         \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm1,%%xmm2                   \n"
+    "mulps     %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm2,%%xmm1                   \n"
+    "mulps     %%xmm6,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm2   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm6   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm1   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm5   \n"
+    "addps     %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm6,%%xmm4                   \n"
+    "addps     %%xmm1,%%xmm0                   \n"
+    "addps     %%xmm5,%%xmm4                   \n"
+    "cvttps2dq %%xmm0,%%xmm0                   \n"
+    "cvttps2dq %%xmm4,%%xmm4                   \n"
+    "packuswb  %%xmm4,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "sub       $0x2,%2                         \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm4     \n"
+    "vbroadcastf128 " MEMACCESS2(0x10,3) ",%%ymm5 \n"
+    "vbroadcastf128 " MEMACCESS2(0x20,3) ",%%ymm6 \n"
+    "vbroadcastf128 " MEMACCESS2(0x30,3) ",%%ymm7 \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd   " MEMACCESS(0) ",%%ymm0       \n"  // 2 ARGB pixels
+    "lea         " MEMLEA(0x8,0) ",%0          \n"
+    "vcvtdq2ps   %%ymm0,%%ymm0                 \n"  // X 8 floats
+    "vmulps      %%ymm0,%%ymm0,%%ymm2          \n"  // X * X
+    "vmulps      %%ymm7,%%ymm0,%%ymm3          \n"  // C3 * X
+    "vfmadd132ps %%ymm5,%%ymm4,%%ymm0          \n"  // result = C0 + C1 * X
+    "vfmadd231ps %%ymm6,%%ymm2,%%ymm0          \n"  // result += C2 * X * X
+    "vfmadd231ps %%ymm3,%%ymm2,%%ymm0          \n"  // result += C3 * X * X * X
+    "vcvttps2dq  %%ymm0,%%ymm0                 \n"
+    "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+    "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+    "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
+    "sub         $0x2,%2                       \n"
+    "vmovq       %%xmm0," MEMACCESS(1) "       \n"
+    "lea         " MEMLEA(0x8,1) ",%1          \n"
+    "jg          1b                            \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+#if defined(__SSE2__)
+// TODO(fbarchard): declare ymm usage when applicable.
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "movzb     " MEMACCESS2(-0x1,0) ",%1       \n"
+    MEMOPARG(movzb,0x03,3,1,4,1) "             \n"  // movzb 0x3(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x1,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  uintptr_t pixel_temp = 0u;
+  uintptr_t table_temp = 0u;
+  asm volatile (
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0x8,%%xmm4                     \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm0         \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "phaddw    %%xmm0,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS(2) ",%0             \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS(3) "            \n"
+    "movzb     " MEMACCESS2(0x1,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x1,3) "       \n"
+    "movzb     " MEMACCESS2(0x2,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x2,3) "       \n"
+    "movzb     " MEMACCESS2(0x3,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x3,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x4,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x4,3) "       \n"
+    BUNDLEALIGN
+    "movzb     " MEMACCESS2(0x5,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x5,3) "       \n"
+    "movzb     " MEMACCESS2(0x6,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x6,3) "       \n"
+    "movzb     " MEMACCESS2(0x7,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x7,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x8,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x8,3) "       \n"
+    "movzb     " MEMACCESS2(0x9,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x9,3) "       \n"
+    "movzb     " MEMACCESS2(0xa,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xa,3) "       \n"
+    "movzb     " MEMACCESS2(0xb,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xb,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+
+    "movzb     " MEMACCESS2(0xc,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xc,3) "       \n"
+    "movzb     " MEMACCESS2(0xd,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xd,3) "       \n"
+    "movzb     " MEMACCESS2(0xe,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xe,3) "       \n"
+    "movzb     " MEMACCESS2(0xf,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xf,3) "       \n"
+    "sub       $0x4,%4                         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "lea       " MEMLEA(0x10,3) ",%3           \n"
+    "jg        1b                              \n"
+  : "+d"(pixel_temp),  // %0
+    "+a"(table_temp),  // %1
+    "+r"(src_argb),    // %2
+    "+r"(dst_argb),    // %3
+    "+rm"(width)       // %4
+  : "r"(luma),         // %5
+    "rm"(lumacoeff)    // %6
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/row_win.cc b/drivers/theoraplayer/src/YUV/libyuv/src/row_win.cc
new file mode 100755
index 0000000000..f13e4d7ae5
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/row_win.cc
@@ -0,0 +1,7284 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef HAS_ARGBTOYROW_SSSE3
+
+// Constants for ARGB.
+static const vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static const vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+
+static const vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static const vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static const vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static const vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
+
+// vpshufb for vphaddw + vpackuswb packed to shorts.
+static const lvec8 kShufARGBToUV_AVX = {
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+};
+
+// Constants for BGRA.
+static const vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static const vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static const vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR.
+static const vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static const vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static const vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static const vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static const vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static const vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static const uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+static const vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static const uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static const uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+
+// Shuffle table for converting RGB24 to ARGB.
+static const uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static const uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static const uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static const uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+
+// Duplicates gray value 3 times and fills in alpha opaque.
+__declspec(naked) __declspec(align(16))
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]        // src_y
+    mov        edx, [esp + 8]        // dst_argb
+    mov        ecx, [esp + 12]       // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    pslld      xmm5, 24
+
+    align      4
+  convertloop:
+    movq       xmm0, qword ptr [eax]
+    lea        eax,  [eax + 8]
+    punpcklbw  xmm0, xmm0
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0
+    punpckhwd  xmm1, xmm1
+    por        xmm0, xmm5
+    por        xmm1, xmm5
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb,
+                                  int pix) {
+  __asm {
+    mov        eax, [esp + 4]        // src_y
+    mov        edx, [esp + 8]        // dst_argb
+    mov        ecx, [esp + 12]       // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    pslld      xmm5, 24
+
+    align      4
+  convertloop:
+    movq       xmm0, qword ptr [eax]
+    lea        eax,  [eax + 8]
+    punpcklbw  xmm0, xmm0
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0
+    punpckhwd  xmm1, xmm1
+    por        xmm0, xmm5
+    por        xmm1, xmm5
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_rgb24
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRGB24ToARGB
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqa    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqa    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqa    [edx + 16], xmm1
+    por       xmm3, xmm5
+    sub       ecx, 16
+    movdqa    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
+                        int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_raw
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRAWToARGB
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqa    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqa    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqa    [edx + 16], xmm1
+    por       xmm3, xmm5
+    sub       ecx, 16
+    movdqa    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    jg        convertloop
+    ret
+  }
+}
+
+// pmul method to replicate bits.
+// Math to replicate bits:
+// (v << 8) | (v << 3)
+// v * 256 + v * 8
+// v * (256 + 8)
+// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
+// 20 instructions.
+__declspec(naked) __declspec(align(16))
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                          int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    pcmpeqb   xmm4, xmm4       // generate mask 0x07e007e0 for Green
+    psllw     xmm4, 10
+    psrlw     xmm4, 5
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_rgb565
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgr565
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    pand      xmm1, xmm3    // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    pand      xmm0, xmm4    // G in middle 6 bits
+    pmulhuw   xmm0, xmm6    // << 5 * (256 + 4)
+    por       xmm0, xmm7    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqa    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+// 24 instructions
+__declspec(naked) __declspec(align(16))
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    movdqa    xmm4, xmm3       // generate mask 0x03e003e0 for Green
+    psrlw     xmm4, 6
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_argb1555
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of 1555
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    psllw     xmm1, 1       // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pand      xmm1, xmm3
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // G in middle 5 bits
+    psraw     xmm2, 8       // A
+    pmulhuw   xmm0, xmm6    // << 6 * (256 + 8)
+    pand      xmm2, xmm7
+    por       xmm0, xmm2    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqa    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+// 18 instructions.
+__declspec(naked) __declspec(align(16))
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x0f0f0f0f  // generate mask 0x0f0f0f0f
+    movd      xmm4, eax
+    pshufd    xmm4, xmm4, 0
+    movdqa    xmm5, xmm4       // 0xf0f0f0f0 for high nibbles
+    pslld     xmm5, 4
+    mov       eax, [esp + 4]   // src_argb4444
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgra4444
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // mask low nibbles
+    pand      xmm2, xmm5    // mask high nibbles
+    movdqa    xmm1, xmm0
+    movdqa    xmm3, xmm2
+    psllw     xmm1, 4
+    psrlw     xmm3, 4
+    por       xmm0, xmm1
+    por       xmm2, xmm3
+    movdqa    xmm1, xmm0
+    punpcklbw xmm0, xmm2
+    punpckhbw xmm1, xmm2
+    movdqa    [eax * 2 + edx], xmm0  // store 4 pixels of ARGB
+    movdqa    [eax * 2 + edx + 16], xmm1  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRGB24
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRAW
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    psrld     xmm3, 27
+    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    psrld     xmm4, 26
+    pslld     xmm4, 5
+    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pslld     xmm5, 11
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    pslld     xmm0, 8       // R
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 5       // G
+    psrad     xmm0, 16      // R
+    pand      xmm1, xmm3    // B
+    pand      xmm2, xmm4    // G
+    pand      xmm0, xmm5    // R
+    por       xmm1, xmm2    // BG
+    por       xmm0, xmm1    // BGR
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+// TODO(fbarchard): Improve sign extension/packing.
+__declspec(naked) __declspec(align(16))
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0x0000001f
+    psrld     xmm4, 27
+    movdqa    xmm5, xmm4       // generate mask 0x000003e0
+    pslld     xmm5, 5
+    movdqa    xmm6, xmm4       // generate mask 0x00007c00
+    pslld     xmm6, 10
+    pcmpeqb   xmm7, xmm7       // generate mask 0xffff8000
+    pslld     xmm7, 15
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    movdqa    xmm3, xmm0    // R
+    psrad     xmm0, 16      // A
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 6       // G
+    psrld     xmm3, 9       // R
+    pand      xmm0, xmm7    // A
+    pand      xmm1, xmm4    // B
+    pand      xmm2, xmm5    // G
+    pand      xmm3, xmm6    // R
+    por       xmm0, xmm1    // BA
+    por       xmm2, xmm3    // GR
+    por       xmm0, xmm2    // BGRA
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB1555
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0xf000f000
+    psllw     xmm4, 12
+    movdqa    xmm3, xmm4       // generate mask 0x00f000f0
+    psrlw     xmm3, 8
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0
+    pand      xmm0, xmm3    // low nibble
+    pand      xmm1, xmm4    // high nibble
+    psrl      xmm0, 4
+    psrl      xmm1, 8
+    por       xmm0, xmm1
+    packuswb  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB4444
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+__declspec(naked) __declspec(align(16))
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kARGBToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+__declspec(naked) __declspec(align(16))
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    paddw      xmm2, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked) __declspec(align(32))
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToY
+    vbroadcastf128 ymm5, kAddY16
+    vmovdqa    ymm6, kPermdARGBToY_AVX
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    vpaddb     ymm0, ymm0, ymm5
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYROW_AVX2
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked) __declspec(align(32))
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToYJ
+    vbroadcastf128 ymm5, kAddYJ64
+    vmovdqa    ymm6, kPermdARGBToY_AVX
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpaddw     ymm0, ymm0, ymm5  // Add .5 for rounding.
+    vpaddw     ymm2, ymm2, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYJROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kARGBToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    paddw      xmm0, xmm5
+    paddw      xmm2, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kBGRAToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kBGRAToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kABGRToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kABGRToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kRGBAToY
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm5, kAddY16
+    movdqa     xmm4, kRGBAToY
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToUJ
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm5, kAddUVJ128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm1, xmm5
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+__declspec(naked) __declspec(align(32))
+void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    vbroadcastf128 ymm5, kAddUV128
+    vbroadcastf128 ymm6, kARGBToV
+    vbroadcastf128 ymm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 32x2 argb pixels to 16x1 */
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    vpavgb     ymm2, ymm2, [eax + esi + 64]
+    vpavgb     ymm3, ymm3, [eax + esi + 96]
+    lea        eax,  [eax + 128]
+    vshufps    ymm4, ymm0, ymm1, 0x88
+    vshufps    ymm0, ymm0, ymm1, 0xdd
+    vpavgb     ymm0, ymm0, ymm4  // mutated by vshufps
+    vshufps    ymm4, ymm2, ymm3, 0x88
+    vshufps    ymm2, ymm2, ymm3, 0xdd
+    vpavgb     ymm2, ymm2, ymm4  // mutated by vshufps
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 32 different pixels, its 16 pixels of U and 16 of V
+    vpmaddubsw ymm1, ymm0, ymm7  // U
+    vpmaddubsw ymm3, ymm2, ymm7
+    vpmaddubsw ymm0, ymm0, ymm6  // V
+    vpmaddubsw ymm2, ymm2, ymm6
+    vphaddw    ymm1, ymm1, ymm3  // mutates
+    vphaddw    ymm0, ymm0, ymm2
+    vpsraw     ymm1, ymm1, 8
+    vpsraw     ymm0, ymm0, 8
+    vpacksswb  ymm0, ymm1, ymm0  // mutates
+    vpermq     ymm0, ymm0, 0xd8  // For vpacksswb
+    vpshufb    ymm0, ymm0, kShufARGBToUV_AVX  // For vshufps + vphaddw
+    vpaddb     ymm0, ymm0, ymm5  // -> unsigned
+
+    // step 3 - store 16 U and 16 V values
+    sub         ecx, 32
+    vextractf128 [edx], ymm0, 0 // U
+    vextractf128 [edx + edi], ymm0, 1 // V
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOUVROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kARGBToUJ
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm5, kAddUVJ128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm1, xmm5
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* convert to U and V */
+    movdqa     xmm0, [eax]          // U
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm7
+    pmaddubsw  xmm1, xmm7
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm3, xmm7
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx,  16
+    movdqa     [edx], xmm0
+
+    movdqa     xmm0, [eax]          // V
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm6
+    pmaddubsw  xmm1, xmm6
+    pmaddubsw  xmm2, xmm6
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    lea        eax,  [eax + 64]
+    movdqa     [edx + edi], xmm0
+    lea        edx,  [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* convert to U and V */
+    movdqu     xmm0, [eax]          // U
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm7
+    pmaddubsw  xmm1, xmm7
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm3, xmm7
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    sub        ecx,  16
+    movdqu     [edx], xmm0
+
+    movdqu     xmm0, [eax]          // V
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm6
+    pmaddubsw  xmm1, xmm6
+    pmaddubsw  xmm2, xmm6
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    lea        eax,  [eax + 64]
+    movdqu     [edx + edi], xmm0
+    lea        edx,  [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb0,
+                                    uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm7, kARGBToU
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kBGRAToU
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kBGRAToU
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kABGRToU
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kABGRToU
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kRGBAToU
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+    pavgb      xmm0, [eax + esi]
+    pavgb      xmm1, [eax + esi + 16]
+    pavgb      xmm2, [eax + esi + 32]
+    pavgb      xmm3, [eax + esi + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                                 uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm7, kRGBAToU
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm5, kAddUV128
+    sub        edi, edx             // stride from u to v
+
+    align      4
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    sub        ecx, 16
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+#define YG 74 /* (int8)(1.164 * 64 + 0.5) */
+
+#define UB 127 /* min(63,(int8)(2.018 * 64)) */
+#define UG -25 /* (int8)(-0.391 * 64 - 0.5) */
+#define UR 0
+
+#define VB 0
+#define VG -52 /* (int8)(-0.813 * 64 - 0.5) */
+#define VR 102 /* (int8)(1.596 * 64 + 0.5) */
+
+// Bias
+#define BB UB * 128 + VB * 128
+#define BG UG * 128 + VG * 128
+#define BR UR * 128 + VR * 128
+
+#ifdef HAS_I422TOARGBROW_AVX2
+
+static const lvec8 kUVToB_AVX = {
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB,
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
+};
+static const lvec8 kUVToR_AVX = {
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR,
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
+};
+static const lvec8 kUVToG_AVX = {
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
+};
+static const lvec16 kYToRgb_AVX = {
+  YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG
+};
+static const lvec16 kYSub16_AVX = {
+  16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16
+};
+static const lvec16 kUVBiasB_AVX = {
+  BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB
+};
+static const lvec16 kUVBiasG_AVX = {
+  BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG
+};
+static const lvec16 kUVBiasR_AVX = {
+  BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR
+};
+
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_AVX2(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+    vpxor      ymm4, ymm4, ymm4
+
+    align      4
+ convertloop:
+    vmovq      xmm0, qword ptr [esi]          //  U
+    vmovq      xmm1, qword ptr [esi + edi]    //  V
+    lea        esi,  [esi + 8]
+    vpunpcklbw ymm0, ymm0, ymm1               // UV
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm0, ymm0, ymm0              // UVUV
+    vpmaddubsw ymm2, ymm0, kUVToB_AVX        // scale B UV
+    vpmaddubsw ymm1, ymm0, kUVToG_AVX        // scale G UV
+    vpmaddubsw ymm0, ymm0, kUVToR_AVX        // scale R UV
+    vpsubw     ymm2, ymm2, kUVBiasB_AVX      // unbias back to signed
+    vpsubw     ymm1, ymm1, kUVBiasG_AVX
+    vpsubw     ymm0, ymm0, kUVBiasR_AVX
+
+    // Step 2: Find Y contribution to 16 R,G,B values
+    vmovdqu    xmm3, [eax]                  // NOLINT
+    lea        eax, [eax + 16]
+    vpermq     ymm3, ymm3, 0xd8
+    vpunpcklbw ymm3, ymm3, ymm4
+    vpsubsw    ymm3, ymm3, kYSub16_AVX
+    vpmullw    ymm3, ymm3, kYToRgb_AVX
+    vpaddsw    ymm2, ymm2, ymm3           // B += Y
+    vpaddsw    ymm1, ymm1, ymm3           // G += Y
+    vpaddsw    ymm0, ymm0, ymm3           // R += Y
+    vpsraw     ymm2, ymm2, 6
+    vpsraw     ymm1, ymm1, 6
+    vpsraw     ymm0, ymm0, 6
+    vpackuswb  ymm2, ymm2, ymm2           // B
+    vpackuswb  ymm1, ymm1, ymm1           // G
+    vpackuswb  ymm0, ymm0, ymm0           // R
+
+    // Step 3: Weave into ARGB
+    vpunpcklbw ymm2, ymm2, ymm1           // BG
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklbw ymm0, ymm0, ymm5           // RA
+    vpermq     ymm0, ymm0, 0xd8
+    vpunpcklwd ymm1, ymm2, ymm0           // BGRA first 8 pixels
+    vpunpckhwd ymm2, ymm2, ymm0           // BGRA next 8 pixels
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_I422TOARGBROW_AVX2
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+
+static const vec8 kUVToB = {
+  UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB
+};
+
+static const vec8 kUVToR = {
+  UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR
+};
+
+static const vec8 kUVToG = {
+  UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG
+};
+
+static const vec8 kVUToB = {
+  VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB, VB, UB,
+};
+
+static const vec8 kVUToR = {
+  VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR, VR, UR,
+};
+
+static const vec8 kVUToG = {
+  VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+};
+
+static const vec16 kYToRgb = { YG, YG, YG, YG, YG, YG, YG, YG };
+static const vec16 kYSub16 = { 16, 16, 16, 16, 16, 16, 16, 16 };
+static const vec16 kUVBiasB = { BB, BB, BB, BB, BB, BB, BB, BB };
+static const vec16 kUVBiasG = { BG, BG, BG, BG, BG, BG, BG, BG };
+static const vec16 kUVBiasR = { BR, BR, BR, BR, BR, BR, BR, BR };
+
+// TODO(fbarchard): Read that does half size on Y and treats 420 as 444.
+
+// Read 8 UV from 444.
+#define READYUV444 __asm {                                                     \
+    __asm movq       xmm0, qword ptr [esi] /* U */                /* NOLINT */ \
+    __asm movq       xmm1, qword ptr [esi + edi] /* V */          /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+  }
+
+// Read 4 UV from 422, upsample to 8 UV.
+#define READYUV422 __asm {                                                     \
+    __asm movd       xmm0, [esi]          /* U */                              \
+    __asm movd       xmm1, [esi + edi]    /* V */                              \
+    __asm lea        esi,  [esi + 4]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Read 2 UV from 411, upsample to 8 UV.
+#define READYUV411 __asm {                                                     \
+    __asm movzx      ebx, word ptr [esi]        /* U */           /* NOLINT */ \
+    __asm movd       xmm0, ebx                                                 \
+    __asm movzx      ebx, word ptr [esi + edi]  /* V */           /* NOLINT */ \
+    __asm movd       xmm1, ebx                                                 \
+    __asm lea        esi,  [esi + 2]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+    __asm punpckldq  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Read 4 UV from NV12, upsample to 8 UV.
+#define READNV12 __asm {                                                       \
+    __asm movq       xmm0, qword ptr [esi] /* UV */               /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Convert 8 pixels: 8 UV and 8 Y.
+#define YUVTORGB __asm {                                                       \
+    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm movdqa     xmm2, xmm0                                                \
+    __asm pmaddubsw  xmm0, kUVToB        /* scale B UV */                      \
+    __asm pmaddubsw  xmm1, kUVToG        /* scale G UV */                      \
+    __asm pmaddubsw  xmm2, kUVToR        /* scale R UV */                      \
+    __asm psubw      xmm0, kUVBiasB      /* unbias back to signed */           \
+    __asm psubw      xmm1, kUVBiasG                                            \
+    __asm psubw      xmm2, kUVBiasR                                            \
+    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
+    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm punpcklbw  xmm3, xmm4                                                \
+    __asm psubsw     xmm3, kYSub16                                             \
+    __asm pmullw     xmm3, kYToRgb                                             \
+    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
+    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
+    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
+    __asm psraw      xmm0, 6                                                   \
+    __asm psraw      xmm1, 6                                                   \
+    __asm psraw      xmm2, 6                                                   \
+    __asm packuswb   xmm0, xmm0           /* B */                              \
+    __asm packuswb   xmm1, xmm1           /* G */                              \
+    __asm packuswb   xmm2, xmm2           /* R */                              \
+  }
+
+// Convert 8 pixels: 8 VU and 8 Y.
+#define YVUTORGB __asm {                                                       \
+    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm movdqa     xmm2, xmm0                                                \
+    __asm pmaddubsw  xmm0, kVUToB        /* scale B UV */                      \
+    __asm pmaddubsw  xmm1, kVUToG        /* scale G UV */                      \
+    __asm pmaddubsw  xmm2, kVUToR        /* scale R UV */                      \
+    __asm psubw      xmm0, kUVBiasB      /* unbias back to signed */           \
+    __asm psubw      xmm1, kUVBiasG                                            \
+    __asm psubw      xmm2, kUVBiasR                                            \
+    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
+    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm punpcklbw  xmm3, xmm4                                                \
+    __asm psubsw     xmm3, kYSub16                                             \
+    __asm pmullw     xmm3, kYToRgb                                             \
+    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
+    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
+    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
+    __asm psraw      xmm0, 6                                                   \
+    __asm psraw      xmm1, 6                                                   \
+    __asm psraw      xmm2, 6                                                   \
+    __asm packuswb   xmm0, xmm0           /* B */                              \
+    __asm packuswb   xmm1, xmm1           /* G */                              \
+    __asm packuswb   xmm2, xmm2           /* R */                              \
+  }
+
+// 8 pixels, dest aligned 16.
+// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I444ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV444
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* dst_rgb24,
+                          int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb24
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    movdqa     xmm5, kShuffleMaskARGBToRGB24_0
+    movdqa     xmm6, kShuffleMaskARGBToRGB24
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+    pshufb     xmm0, xmm5           // Pack into first 8 and last 4 bytes.
+    pshufb     xmm1, xmm6           // Pack into first 12 bytes.
+    palignr    xmm1, xmm0, 12       // last 4 bytes of xmm0 + 12 from xmm1
+    movq       qword ptr [edx], xmm0  // First 8 bytes
+    movdqu     [edx + 8], xmm1      // Last 16 bytes. = 24 bytes, 8 RGB pixels.
+    lea        edx,  [edx + 24]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRAWRow_SSSE3(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_raw,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // raw
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    movdqa     xmm5, kShuffleMaskARGBToRAW_0
+    movdqa     xmm6, kShuffleMaskARGBToRAW
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+    pshufb     xmm0, xmm5           // Pack into first 8 and last 4 bytes.
+    pshufb     xmm1, xmm6           // Pack into first 12 bytes.
+    palignr    xmm1, xmm0, 12       // last 4 bytes of xmm0 + 12 from xmm1
+    movq       qword ptr [edx], xmm0  // First 8 bytes
+    movdqu     [edx + 8], xmm1      // Last 16 bytes. = 24 bytes, 8 RGB pixels.
+    lea        edx,  [edx + 24]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest unaligned.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToRGB565Row_SSSE3(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb565_buf,
+                           int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb565
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+    pcmpeqb    xmm5, xmm5       // generate mask 0x0000001f
+    psrld      xmm5, 27
+    pcmpeqb    xmm6, xmm6       // generate mask 0x000007e0
+    psrld      xmm6, 26
+    pslld      xmm6, 5
+    pcmpeqb    xmm7, xmm7       // generate mask 0xfffff800
+    pslld      xmm7, 11
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RRGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm2           // RR
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRR first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRR next 4 pixels
+
+    // Step 3b: RRGB -> RGB565
+    movdqa     xmm3, xmm0    // B  first 4 pixels of argb
+    movdqa     xmm2, xmm0    // G
+    pslld      xmm0, 8       // R
+    psrld      xmm3, 3       // B
+    psrld      xmm2, 5       // G
+    psrad      xmm0, 16      // R
+    pand       xmm3, xmm5    // B
+    pand       xmm2, xmm6    // G
+    pand       xmm0, xmm7    // R
+    por        xmm3, xmm2    // BG
+    por        xmm0, xmm3    // BGR
+    movdqa     xmm3, xmm1    // B  next 4 pixels of argb
+    movdqa     xmm2, xmm1    // G
+    pslld      xmm1, 8       // R
+    psrld      xmm3, 3       // B
+    psrld      xmm2, 5       // G
+    psrad      xmm1, 16      // R
+    pand       xmm3, xmm5    // B
+    pand       xmm2, xmm6    // G
+    pand       xmm1, xmm7    // R
+    por        xmm3, xmm2    // BG
+    por        xmm1, xmm3    // BGR
+    packssdw   xmm0, xmm1
+    sub        ecx, 8
+    movdqu     [edx], xmm0   // store 8 pixels of RGB565
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// Similar to I420 but duplicate UV once more.
+__declspec(naked) __declspec(align(16))
+void I411ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]   // Y
+    mov        esi, [esp + 12 + 8]   // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ecx, [esp + 12 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV411  // modifies EBX
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // VU
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YVUTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I444ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV444
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void I422ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, unaligned.
+// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// Similar to I420 but duplicate UV once more.
+__declspec(naked) __declspec(align(16))
+void I411ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]   // Y
+    mov        esi, [esp + 12 + 8]   // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ecx, [esp + 12 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV411  // modifies EBX
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV12ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* uv_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels, dest aligned 16.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked) __declspec(align(16))
+void NV21ToARGBRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* uv_buf,
+                                   uint8* dst_argb,
+                                   int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // VU
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READNV12
+    YVUTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm0, xmm1           // BG
+    punpcklbw  xmm2, xmm5           // RA
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm2           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm2           // BGRA next 4 pixels
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToBGRARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_bgra,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // bgra
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into BGRA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm0           // GB
+    punpcklbw  xmm5, xmm2           // AR
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // BGRA first 4 pixels
+    punpckhwd  xmm0, xmm1           // BGRA next 4 pixels
+    movdqa     [edx], xmm5
+    movdqa     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToBGRARow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_bgra,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // bgra
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into BGRA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm0           // GB
+    punpcklbw  xmm5, xmm2           // AR
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // BGRA first 4 pixels
+    punpckhwd  xmm0, xmm1           // BGRA next 4 pixels
+    movdqu     [edx], xmm5
+    movdqu     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToABGRRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_abgr,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // abgr
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm2, xmm1           // RG
+    punpcklbw  xmm0, xmm5           // BA
+    movdqa     xmm1, xmm2
+    punpcklwd  xmm2, xmm0           // RGBA first 4 pixels
+    punpckhwd  xmm1, xmm0           // RGBA next 4 pixels
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToABGRRow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_abgr,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // abgr
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into ARGB
+    punpcklbw  xmm2, xmm1           // RG
+    punpcklbw  xmm0, xmm5           // BA
+    movdqa     xmm1, xmm2
+    punpcklwd  xmm2, xmm0           // RGBA first 4 pixels
+    punpckhwd  xmm1, xmm0           // RGBA next 4 pixels
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToRGBARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_rgba,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgba
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RGBA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm2           // GR
+    punpcklbw  xmm5, xmm0           // AB
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // RGBA first 4 pixels
+    punpckhwd  xmm0, xmm1           // RGBA next 4 pixels
+    movdqa     [edx], xmm5
+    movdqa     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToRGBARow_Unaligned_SSSE3(const uint8* y_buf,
+                                   const uint8* u_buf,
+                                   const uint8* v_buf,
+                                   uint8* dst_rgba,
+                                   int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgba
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pxor       xmm4, xmm4
+
+    align      4
+ convertloop:
+    READYUV422
+    YUVTORGB
+
+    // Step 3: Weave into RGBA
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+    punpcklbw  xmm1, xmm2           // GR
+    punpcklbw  xmm5, xmm0           // AB
+    movdqa     xmm0, xmm5
+    punpcklwd  xmm5, xmm1           // RGBA first 4 pixels
+    punpckhwd  xmm0, xmm1           // RGBA next 4 pixels
+    movdqu     [edx], xmm5
+    movdqu     [edx + 16], xmm0
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+#ifdef HAS_YTOARGBROW_SSE2
+__declspec(naked) __declspec(align(16))
+void YToARGBRow_SSE2(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) {
+  __asm {
+    pxor       xmm5, xmm5
+    pcmpeqb    xmm4, xmm4           // generate mask 0xff000000
+    pslld      xmm4, 24
+    mov        eax, 0x00100010
+    movd       xmm3, eax
+    pshufd     xmm3, xmm3, 0
+    mov        eax, 0x004a004a       // 74
+    movd       xmm2, eax
+    pshufd     xmm2, xmm2,0
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+    align      4
+ convertloop:
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    movq       xmm0, qword ptr [eax]
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm5           // 0.Y
+    psubusw    xmm0, xmm3
+    pmullw     xmm0, xmm2
+    psrlw      xmm0, 6
+    packuswb   xmm0, xmm0           // G
+
+    // Step 2: Weave into ARGB
+    punpcklbw  xmm0, xmm0           // GG
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm1           // BGRA next 4 pixels
+    por        xmm0, xmm4
+    por        xmm1, xmm4
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_YTOARGBROW_SSE2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    movdqa    xmm5, kShuffleMirror
+    lea       eax, [eax - 16]
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax + ecx]
+    pshufb    xmm0, xmm5
+    sub       ecx, 16
+    movdqa    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec8 kShuffleMirror_AVX2 = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u,
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    vmovdqa   ymm5, kShuffleMirror_AVX2
+    lea       eax, [eax - 32]
+
+    align      4
+ convertloop:
+    vmovdqu   ymm0, [eax + ecx]
+    vpshufb   ymm0, ymm0, ymm5
+    vpermq    ymm0, ymm0, 0x4e  // swap high and low halfs
+    sub       ecx, 32
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSE2
+// SSE2 version has movdqu so it can be used on unaligned buffers when SSSE3
+// version can not.
+__declspec(naked) __declspec(align(16))
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 16]
+
+    align      4
+ convertloop:
+    movdqu    xmm0, [eax + ecx]
+    movdqa    xmm1, xmm0        // swap bytes
+    psllw     xmm0, 8
+    psrlw     xmm1, 8
+    por       xmm0, xmm1
+    pshuflw   xmm0, xmm0, 0x1b  // swap words
+    pshufhw   xmm0, xmm0, 0x1b
+    pshufd    xmm0, xmm0, 0x4e  // swap qwords
+    sub       ecx, 16
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static const uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+
+__declspec(naked) __declspec(align(16))
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  __asm {
+    push      edi
+    mov       eax, [esp + 4 + 4]   // src
+    mov       edx, [esp + 4 + 8]   // dst_u
+    mov       edi, [esp + 4 + 12]  // dst_v
+    mov       ecx, [esp + 4 + 16]  // width
+    movdqa    xmm1, kShuffleMirrorUV
+    lea       eax, [eax + ecx * 2 - 16]
+    sub       edi, edx
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufb    xmm0, xmm1
+    sub       ecx, 8
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [edx + edi], xmm0
+    lea       edx, [edx + 8]
+    jg        convertloop
+
+    pop       edi
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static const uvec8 kARGBShuffleMirror = {
+  12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u, 4u, 5u, 6u, 7u, 0u, 1u, 2u, 3u
+};
+
+__declspec(naked) __declspec(align(16))
+void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 16 + ecx * 4]  // last 4 pixels.
+    movdqa    xmm5, kARGBShuffleMirror
+
+    align      4
+ convertloop:
+    movdqa    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufb    xmm0, xmm5
+    sub       ecx, 4
+    movdqa    [edx], xmm0
+    lea       edx, [edx + 16]
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked) __declspec(align(16))
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 32]
+    vmovdqa   ymm5, kARGBShuffleMirror_AVX2
+
+    align      4
+ convertloop:
+    vpermd    ymm0, ymm5, [eax + ecx * 4]  // permute dword order
+    sub       ecx, 8
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
+#ifdef HAS_SPLITUVROW_SSE2
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    movdqa     xmm3, xmm1
+    pand       xmm0, xmm5   // even bytes
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm3, 8
+    packuswb   xmm2, xmm3
+    movdqa     [edx], xmm0
+    movdqa     [edx + edi], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    movdqa     xmm3, xmm1
+    pand       xmm0, xmm5   // even bytes
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm3, 8
+    packuswb   xmm2, xmm3
+    movdqu     [edx], xmm0
+    movdqu     [edx + edi], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_SPLITUVROW_AVX2
+__declspec(naked) __declspec(align(16))
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm2, ymm0, 8      // odd bytes
+    vpsrlw     ymm3, ymm1, 8
+    vpand      ymm0, ymm0, ymm5   // even bytes
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1
+    vpackuswb  ymm2, ymm2, ymm3
+    vpermq     ymm0, ymm0, 0xd8
+    vpermq     ymm2, ymm2, 0xd8
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + edi], ymm2
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SPLITUVROW_AVX2
+
+#ifdef HAS_MERGEUVROW_SSE2
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]      // read 16 U's
+    movdqa     xmm1, [eax + edx]  // and 16 V's
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, xmm0
+    punpcklbw  xmm0, xmm1       // first 8 UV pairs
+    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    movdqa     [edi], xmm0
+    movdqa     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
+                               uint8* dst_uv, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]      // read 16 U's
+    movdqu     xmm1, [eax + edx]  // and 16 V's
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, xmm0
+    punpcklbw  xmm0, xmm1       // first 8 UV pairs
+    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_AVX2
+__declspec(naked) __declspec(align(16))
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]           // read 32 U's
+    vmovdqu    ymm1, [eax + edx]     // and 32 V's
+    lea        eax,  [eax + 32]
+    vpunpcklbw ymm2, ymm0, ymm1      // low 16 UV pairs. mutated qqword 0,2
+    vpunpckhbw ymm0, ymm0, ymm1      // high 16 UV pairs. mutated qqword 1,3
+    vperm2i128 ymm1, ymm2, ymm0, 0x20  // low 128 of ymm2 and low 128 of ymm0
+    vperm2i128 ymm2, ymm2, ymm0, 0x31  // high 128 of ymm2 and high 128 of ymm0
+    vmovdqu    [edi], ymm1
+    vmovdqu    [edi + 32], ymm2
+    lea        edi, [edi + 64]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_AVX2
+
+#ifdef HAS_COPYROW_SSE2
+// CopyRow copys 'count' bytes using a 16 byte load/store, 32 bytes at time.
+__declspec(naked) __declspec(align(16))
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_COPYROW_SSE2
+
+// Unaligned Multiple of 1.
+__declspec(naked) __declspec(align(16))
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, esi
+    mov        edx, edi
+    mov        esi, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    rep movsb
+    mov        edi, edx
+    mov        esi, eax
+    ret
+  }
+}
+
+#ifdef HAS_COPYROW_X86
+__declspec(naked) __declspec(align(16))
+void CopyRow_X86(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, esi
+    mov        edx, edi
+    mov        esi, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    shr        ecx, 2
+    rep movsd
+    mov        edi, edx
+    mov        esi, eax
+    ret
+  }
+}
+#endif  // HAS_COPYROW_X86
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+    align      4
+  convertloop:
+    movdqa     xmm2, [eax]
+    movdqa     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqa     xmm4, [edx]
+    movdqa     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+    align      4
+  convertloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + 32]
+    lea        eax, [eax + 64]
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [eax]  // 8 Y's
+    lea        eax, [eax + 8]
+    punpcklbw  xmm2, xmm2
+    punpckhwd  xmm3, xmm2
+    punpcklwd  xmm2, xmm2
+    movdqa     xmm4, [edx]
+    movdqa     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+__declspec(naked) __declspec(align(16))
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+    align      4
+  convertloop:
+    vpmovzxbd  ymm1, qword ptr [eax]
+    vpmovzxbd  ymm2, qword ptr [eax + 8]
+    lea        eax, [eax + 16]
+    vpslld     ymm1, ymm1, 24
+    vpslld     ymm2, ymm2, 24
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+// SetRow8 writes 'count' bytes using a 32 bit value repeated.
+__declspec(naked) __declspec(align(16))
+void SetRow_X86(uint8* dst, uint32 v32, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v32
+    mov        ecx, [esp + 12]  // count
+    shr        ecx, 2
+    rep stosd
+    mov        edi, edx
+    ret
+  }
+}
+
+// SetRow32 writes 'count' words using a 32 bit value repeated.
+__declspec(naked) __declspec(align(16))
+void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
+                   int dst_stride, int height) {
+  __asm {
+    push       esi
+    push       edi
+    push       ebp
+    mov        edi, [esp + 12 + 4]   // dst
+    mov        eax, [esp + 12 + 8]   // v32
+    mov        ebp, [esp + 12 + 12]  // width
+    mov        edx, [esp + 12 + 16]  // dst_stride
+    mov        esi, [esp + 12 + 20]  // height
+    lea        ecx, [ebp * 4]
+    sub        edx, ecx             // stride - width * 4
+
+    align      4
+  convertloop:
+    mov        ecx, ebp
+    rep stosd
+    add        edi, edx
+    sub        esi, 1
+    jg         convertloop
+
+    pop        ebp
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_AVX2
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_AVX2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // even bytes are Y
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_AVX2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // odd bytes are Y
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    sub        ecx, 32
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+    vzeroupper
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_AVX2
+
+#ifdef HAS_YUY2TOYROW_SSE2
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_SSE2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
+                               uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_SSE2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
+                               uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                                uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
+                                   uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+    align      4
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 1
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+
+    sub        ecx, 1
+    je         convertloop1     // only 1 pixel?
+    jl         convertloop1b
+
+    // 1 pixel loop until destination pointer is aligned.
+  alignloop1:
+    test       edx, 15          // aligned?
+    je         alignloop1b
+    movd       xmm3, [eax]
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        alignloop1
+
+  alignloop1b:
+    add        ecx, 1 - 4
+    jl         convertloop4b
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static const uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+// Same as SSE2, but replaces:
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+//    pshuflw    xmm3, xmm3, 0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+// Blend 8 pixels at a time.
+
+__declspec(naked) __declspec(align(16))
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 0x0001
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+
+    sub        ecx, 1
+    je         convertloop1     // only 1 pixel?
+    jl         convertloop1b
+
+    // 1 pixel loop until destination pointer is aligned.
+  alignloop1:
+    test       edx, 15          // aligned?
+    je         alignloop1b
+    movd       xmm3, [eax]
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        alignloop1
+
+  alignloop1b:
+    add        ecx, 1 - 4
+    jl         convertloop4b
+
+    test       eax, 15          // unaligned?
+    jne        convertuloop4
+    test       esi, 15          // unaligned?
+    jne        convertuloop4
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqa     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqa     xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqa     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+    jmp        convertloop4b
+
+    // 4 pixel unaligned loop.
+  convertuloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertuloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    pcmpeqb    xmm5, xmm5       // generate mask 0x00ffffff
+    psrld      xmm5, 8
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]      // read 4 pixels
+    punpcklbw  xmm0, xmm0       // first 2
+    pshufhw    xmm2, xmm0, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm0, xmm2       // rgb * a
+    movdqa     xmm1, [eax]      // read 4 pixels
+    punpckhbw  xmm1, xmm1       // next 2 pixels
+    pshufhw    xmm2, xmm1, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqa     xmm2, [eax]      // alphas
+    lea        eax, [eax + 16]
+    psrlw      xmm0, 8
+    pand       xmm2, xmm4
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    pand       xmm0, xmm5       // keep original alphas
+    por        xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+};
+static const uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+};
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm3, xmm3       // generate mask 0xff000000
+    pslld      xmm3, 24
+    movdqa     xmm4, kShuffleAlpha0
+    movdqa     xmm5, kShuffleAlpha1
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    pshufb     xmm0, xmm4       // isolate first 2 alphas
+    movdqu     xmm1, [eax]      // read 4 pixels
+    punpcklbw  xmm1, xmm1       // first 2 pixel rgbs
+    pmulhuw    xmm0, xmm1       // rgb * a
+    movdqu     xmm1, [eax]      // read 4 pixels
+    pshufb     xmm1, xmm5       // isolate next 2 alphas
+    movdqu     xmm2, [eax]      // read 4 pixels
+    punpckhbw  xmm2, xmm2       // next 2 pixel rgbs
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqu     xmm2, [eax]      // mask original alpha
+    lea        eax, [eax + 16]
+    pand       xmm2, xmm3
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm2       // copy original alpha
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const ulvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
+  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u,
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u,
+  14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u,
+};
+__declspec(naked) __declspec(align(16))
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm4, kShuffleAlpha_AVX2
+    vpcmpeqb   ymm5, ymm5, ymm5 // generate mask 0xff000000
+    vpslld     ymm5, ymm5, 24
+
+    align      4
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpshufb    ymm2, ymm0, ymm4  // low 4 alphas
+    vpshufb    ymm3, ymm1, ymm4  // high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * a
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * a
+    vpand      ymm6, ymm6, ymm5  // isolate alpha
+    vpsrlw     ymm0, ymm0, 8
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vpor       ymm0, ymm0, ymm6  // copy original alpha
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb0
+    mov        edx, [esp + 8 + 8]   // dst_argb
+    mov        ecx, [esp + 8 + 12]  // width
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 3]  // first alpha
+    movzx      edi, byte ptr [eax + 7]  // second alpha
+    punpcklbw  xmm0, xmm0       // first 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words.  1, a, a, a
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm0, xmm2       // rgb * a
+
+    movdqu     xmm1, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 11]  // third alpha
+    movzx      edi, byte ptr [eax + 15]  // forth alpha
+    punpckhbw  xmm1, xmm1       // next 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm1, xmm2       // rgb * a
+    lea        eax, [eax + 16]
+
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const ulvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15,
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15,
+};
+// TODO(fbarchard): Enable USE_GATHER for future hardware if faster.
+// USE_GATHER is not on by default, due to being a slow instruction.
+#ifdef USE_GATHER
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm4, kUnattenShuffleAlpha_AVX2
+
+    align      4
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0xffffffff for gather.
+    vpsrld     ymm2, ymm6, 24    // alpha in low 8 bits.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpgatherdd ymm3, [ymm2 * 4 + fixed_invtbl8], ymm5  // ymm5 cleared.  1, a
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm4  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm4  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#else  // USE_GATHER
+__declspec(naked) __declspec(align(16))
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vmovdqa    ymm5, kUnattenShuffleAlpha_AVX2
+
+    push       esi
+    push       edi
+
+    align      4
+ convertloop:
+    // replace VPGATHER
+    movzx      esi, byte ptr [eax + 3]                 // alpha0
+    movzx      edi, byte ptr [eax + 7]                 // alpha1
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a0]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a1]
+    movzx      esi, byte ptr [eax + 11]                // alpha2
+    movzx      edi, byte ptr [eax + 15]                // alpha3
+    vpunpckldq xmm6, xmm0, xmm1                        // [1,a1,1,a0]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a2]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a3]
+    movzx      esi, byte ptr [eax + 19]                // alpha4
+    movzx      edi, byte ptr [eax + 23]                // alpha5
+    vpunpckldq xmm7, xmm2, xmm3                        // [1,a3,1,a2]
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a4]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a5]
+    movzx      esi, byte ptr [eax + 27]                // alpha6
+    movzx      edi, byte ptr [eax + 31]                // alpha7
+    vpunpckldq xmm0, xmm0, xmm1                        // [1,a5,1,a4]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a6]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a7]
+    vpunpckldq xmm2, xmm2, xmm3                        // [1,a7,1,a6]
+    vpunpcklqdq xmm3, xmm6, xmm7                       // [1,a3,1,a2,1,a1,1,a0]
+    vpunpcklqdq xmm0, xmm0, xmm2                       // [1,a7,1,a6,1,a5,1,a4]
+    vinserti128 ymm3, ymm3, xmm0, 1 // [1,a7,1,a6,1,a5,1,a4,1,a3,1,a2,1,a1,1,a0]
+    // end of VPGATHER
+
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm5  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm5  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    sub        ecx, 8
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // USE_GATHER
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels.
+__declspec(naked) __declspec(align(16))
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* width */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm0, xmm1
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 G bytes
+    movdqa     xmm2, [eax]  // A
+    movdqa     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm2, 24
+    psrld      xmm3, 24
+    packuswb   xmm2, xmm3
+    packuswb   xmm2, xmm2   // 8 A bytes
+    movdqa     xmm3, xmm0   // Weave into GG, GA, then GGGA
+    punpcklbw  xmm0, xmm0   // 8 GG words
+    punpcklbw  xmm3, xmm2   // 8 GA words
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm3   // GGGA first 4
+    punpckhwd  xmm1, xmm3   // GGGA next 4
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone.
+static const vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static const vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static const vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+__declspec(naked) __declspec(align(16))
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* dst_argb */
+    mov        ecx, [esp + 8]   /* width */
+    movdqa     xmm2, kARGBToSepiaB
+    movdqa     xmm3, kARGBToSepiaG
+    movdqa     xmm4, kARGBToSepiaR
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // B
+    movdqa     xmm6, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm6, xmm2
+    phaddw     xmm0, xmm6
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 B values
+    movdqa     xmm5, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 G values
+    punpcklbw  xmm0, xmm5   // 8 BG values
+    movdqa     xmm5, [eax]  // R
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 R values
+    movdqa     xmm6, [eax]  // A
+    movdqa     xmm1, [eax + 16]
+    psrld      xmm6, 24
+    psrld      xmm1, 24
+    packuswb   xmm6, xmm1
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm5, xmm6   // 8 RA values
+    movdqa     xmm1, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm5   // BGRA first 4
+    punpckhwd  xmm1, xmm5   // BGRA next 4
+    sub        ecx, 8
+    movdqa     [eax], xmm0
+    movdqa     [eax + 16], xmm1
+    lea        eax, [eax + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+// TODO(fbarchard): packuswbs only use half of the reg. To make RGBA, combine R
+// and B into a high and low, then G/A, unpackl/hbw and then unpckl/hwd.
+__declspec(naked) __declspec(align(16))
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* matrix_argb */
+    movdqu     xmm5, [ecx]
+    pshufd     xmm2, xmm5, 0x00
+    pshufd     xmm3, xmm5, 0x55
+    pshufd     xmm4, xmm5, 0xaa
+    pshufd     xmm5, xmm5, 0xff
+    mov        ecx, [esp + 16]  /* width */
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // B
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm7, xmm2
+    movdqa     xmm6, [eax]  // G
+    movdqa     xmm1, [eax + 16]
+    pmaddubsw  xmm6, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddsw    xmm0, xmm7   // B
+    phaddsw    xmm6, xmm1   // G
+    psraw      xmm0, 6      // B
+    psraw      xmm6, 6      // G
+    packuswb   xmm0, xmm0   // 8 B values
+    packuswb   xmm6, xmm6   // 8 G values
+    punpcklbw  xmm0, xmm6   // 8 BG values
+    movdqa     xmm1, [eax]  // R
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm7, xmm4
+    phaddsw    xmm1, xmm7   // R
+    movdqa     xmm6, [eax]  // A
+    movdqa     xmm7, [eax + 16]
+    pmaddubsw  xmm6, xmm5
+    pmaddubsw  xmm7, xmm5
+    phaddsw    xmm6, xmm7   // A
+    psraw      xmm1, 6      // R
+    psraw      xmm6, 6      // A
+    packuswb   xmm1, xmm1   // 8 R values
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm1, xmm6   // 8 RA values
+    movdqa     xmm6, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm1   // BGRA first 4
+    punpckhwd  xmm6, xmm1   // BGRA next 4
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm6
+    lea        eax, [eax + 32]
+    lea        edx, [edx + 32]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  __asm {
+    mov        eax, [esp + 4]    /* dst_argb */
+    movd       xmm2, [esp + 8]   /* scale */
+    movd       xmm3, [esp + 12]  /* interval_size */
+    movd       xmm4, [esp + 16]  /* interval_offset */
+    mov        ecx, [esp + 20]   /* width */
+    pshuflw    xmm2, xmm2, 040h
+    pshufd     xmm2, xmm2, 044h
+    pshuflw    xmm3, xmm3, 040h
+    pshufd     xmm3, xmm3, 044h
+    pshuflw    xmm4, xmm4, 040h
+    pshufd     xmm4, xmm4, 044h
+    pxor       xmm5, xmm5  // constant 0
+    pcmpeqb    xmm6, xmm6  // generate mask 0xff000000
+    pslld      xmm6, 24
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]  // read 4 pixels
+    punpcklbw  xmm0, xmm5   // first 2 pixels
+    pmulhuw    xmm0, xmm2   // pixel * scale >> 16
+    movdqa     xmm1, [eax]  // read 4 pixels
+    punpckhbw  xmm1, xmm5   // next 2 pixels
+    pmulhuw    xmm1, xmm2
+    pmullw     xmm0, xmm3   // * interval_size
+    movdqa     xmm7, [eax]  // read 4 pixels
+    pmullw     xmm1, xmm3
+    pand       xmm7, xmm6   // mask alpha
+    paddw      xmm0, xmm4   // + interval_size / 2
+    paddw      xmm1, xmm4
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm7
+    sub        ecx, 4
+    movdqa     [eax], xmm0
+    lea        eax, [eax + 16]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    movd       xmm2, [esp + 16]  // value
+    punpcklbw  xmm2, xmm2
+    punpcklqdq xmm2, xmm2
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]      // read 4 pixels
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0       // first 2
+    punpckhbw  xmm1, xmm1       // next 2
+    pmulhuw    xmm0, xmm2       // argb * value
+    pmulhuw    xmm1, xmm2       // argb * value
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    movdqu     xmm2, [esi]        // read 4 pixels from src_argb1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm0, xmm0         // first 2
+    punpckhbw  xmm1, xmm1         // next 2
+    punpcklbw  xmm2, xmm5         // first 2
+    punpckhbw  xmm3, xmm5         // next 2
+    pmulhuw    xmm0, xmm2         // src_argb0 * src_argb1 first 2
+    pmulhuw    xmm1, xmm3         // src_argb0 * src_argb1 next 2
+    lea        eax, [eax + 16]
+    lea        esi, [esi + 16]
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+// TODO(fbarchard): Port this to posix, neon and other math functions.
+__declspec(naked) __declspec(align(16))
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    sub        ecx, 4
+    jl         convertloop49
+
+    align      4
+ convertloop4:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jge        convertloop4
+
+ convertloop49:
+    add        ecx, 4 - 1
+    jl         convertloop19
+
+ convertloop1:
+    movd       xmm0, [eax]        // read 1 pixels from src_argb0
+    lea        eax, [eax + 4]
+    movd       xmm1, [esi]        // read 1 pixels from src_argb1
+    lea        esi, [esi + 4]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        convertloop1
+
+ convertloop19:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    psubusb    xmm0, xmm1         // src_argb0 - src_argb1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpxor      ymm5, ymm5, ymm5     // constant 0
+
+    align      4
+ convertloop:
+    vmovdqu    ymm1, [eax]        // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vmovdqu    ymm3, [esi]        // read 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vpunpcklbw ymm0, ymm1, ymm1   // low 4
+    vpunpckhbw ymm1, ymm1, ymm1   // high 4
+    vpunpcklbw ymm2, ymm3, ymm5   // low 4
+    vpunpckhbw ymm3, ymm3, ymm5   // high 4
+    vpmulhuw   ymm0, ymm0, ymm2   // src_argb0 * src_argb1 low 4
+    vpmulhuw   ymm1, ymm1, ymm3   // src_argb0 * src_argb1 high 4
+    vpackuswb  ymm0, ymm0, ymm1
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_AVX2
+
+#ifdef HAS_ARGBADDROW_AVX2
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpaddusb   ymm0, ymm0, [esi]        // add 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_AVX2
+
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+// Subtract 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked) __declspec(align(16))
+void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    align      4
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpsubusb   ymm0, ymm0, [esi]        // src_argb0 - src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_AVX2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+__declspec(naked) __declspec(align(16))
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_y0
+    mov        esi, [esp + 8 + 8]   // src_y1
+    mov        edi, [esp + 8 + 12]  // src_y2
+    mov        edx, [esp + 8 + 16]  // dst_sobelx
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        esi, eax
+    sub        edi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    movq       xmm2, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + edi]      // read 8 pixels from src_y2[0]
+    movq       xmm3, qword ptr [eax + edi + 2]  // read 8 pixels from src_y2[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+__declspec(naked) __declspec(align(16))
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_y0
+    mov        esi, [esp + 4 + 8]   // src_y1
+    mov        edx, [esp + 4 + 12]  // dst_sobely
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+    align      4
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + 1]        // read 8 pixels from src_y0[1]
+    movq       xmm2, qword ptr [eax + esi + 1]  // read 8 pixels from src_y1[1]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    movq       xmm3, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+__declspec(naked) __declspec(align(16))
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+    pslld      xmm5, 24             // 0xff000000
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm2, xmm0             // GG
+    punpcklbw  xmm2, xmm0             // First 8
+    punpckhbw  xmm0, xmm0             // Next 8
+    movdqa     xmm1, xmm2             // GGGG
+    punpcklwd  xmm1, xmm2             // First 4
+    punpckhwd  xmm2, xmm2             // Next 4
+    por        xmm1, xmm5             // GGGA
+    por        xmm2, xmm5
+    movdqa     xmm3, xmm0             // GGGG
+    punpcklwd  xmm3, xmm0             // Next 4
+    punpckhwd  xmm0, xmm0             // Last 4
+    por        xmm3, xmm5             // GGGA
+    por        xmm0, xmm5
+    sub        ecx, 16
+    movdqa     [edx], xmm1
+    movdqa     [edx + 16], xmm2
+    movdqa     [edx + 32], xmm3
+    movdqa     [edx + 48], xmm0
+    lea        edx, [edx + 64]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+__declspec(naked) __declspec(align(16))
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+__declspec(naked) __declspec(align(16))
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+
+    align      4
+ convertloop:
+    movdqa     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqa     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    movdqa     xmm2, xmm0
+    paddusb    xmm2, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm3, xmm0             // XA
+    punpcklbw  xmm3, xmm5
+    punpckhbw  xmm0, xmm5
+    movdqa     xmm4, xmm1             // YS
+    punpcklbw  xmm4, xmm2
+    punpckhbw  xmm1, xmm2
+    movdqa     xmm6, xmm4             // YSXA
+    punpcklwd  xmm6, xmm3             // First 4
+    punpckhwd  xmm4, xmm3             // Next 4
+    movdqa     xmm7, xmm1             // YSXA
+    punpcklwd  xmm7, xmm0             // Next 4
+    punpckhwd  xmm1, xmm0             // Last 4
+    sub        ecx, 16
+    movdqa     [edx], xmm6
+    movdqa     [edx + 16], xmm4
+    movdqa     [edx + 32], xmm7
+    movdqa     [edx + 48], xmm1
+    lea        edx, [edx + 64]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+// Consider float CumulativeSum.
+// Consider calling CumulativeSum one row at time as needed.
+// Consider circular CumulativeSum buffer of radius * 2 + 1 height.
+// Convert cumulative sum for an area to an average for 1 pixel.
+// topleft is pointer to top left of CumulativeSum buffer for area.
+// botleft is pointer to bottom left of CumulativeSum buffer.
+// width is offset from left to right of area in CumulativeSum buffer measured
+//   in number of ints.
+// area is the number of pixels in the area being averaged.
+// dst points to pixel to store result to.
+// count is number of averaged pixels to produce.
+// Does 4 pixels at a time, requires CumulativeSum pointers to be 16 byte
+// aligned.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  __asm {
+    mov        eax, topleft  // eax topleft
+    mov        esi, botleft  // esi botleft
+    mov        edx, width
+    movd       xmm5, area
+    mov        edi, dst
+    mov        ecx, count
+    cvtdq2ps   xmm5, xmm5
+    rcpss      xmm4, xmm5  // 1.0f / area
+    pshufd     xmm4, xmm4, 0
+    sub        ecx, 4
+    jl         l4b
+
+    cmp        area, 128  // 128 pixels will not overflow 15 bits.
+    ja         l4
+
+    pshufd     xmm5, xmm5, 0        // area
+    pcmpeqb    xmm6, xmm6           // constant of 65536.0 - 1 = 65535.0
+    psrld      xmm6, 16
+    cvtdq2ps   xmm6, xmm6
+    addps      xmm5, xmm6           // (65536.0 + area - 1)
+    mulps      xmm5, xmm4           // (65536.0 + area - 1) * 1 / area
+    cvtps2dq   xmm5, xmm5           // 0.16 fixed point
+    packssdw   xmm5, xmm5           // 16 bit shorts
+
+    // 4 pixel loop small blocks.
+    align      4
+  s4:
+    // top left
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    packssdw   xmm0, xmm1  // pack 4 pixels into 2 registers
+    packssdw   xmm2, xmm3
+
+    pmulhuw    xmm0, xmm5
+    pmulhuw    xmm2, xmm5
+
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        s4
+
+    jmp        l4b
+
+    // 4 pixel loop
+    align      4
+  l4:
+    // top left
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + 32]
+    movdqa     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    cvtdq2ps   xmm0, xmm0   // Average = Sum * 1 / Area
+    cvtdq2ps   xmm1, xmm1
+    mulps      xmm0, xmm4
+    mulps      xmm1, xmm4
+    cvtdq2ps   xmm2, xmm2
+    cvtdq2ps   xmm3, xmm3
+    mulps      xmm2, xmm4
+    mulps      xmm3, xmm4
+    cvtps2dq   xmm0, xmm0
+    cvtps2dq   xmm1, xmm1
+    cvtps2dq   xmm2, xmm2
+    cvtps2dq   xmm3, xmm3
+    packssdw   xmm0, xmm1
+    packssdw   xmm2, xmm3
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    movdqa     xmm0, [eax]
+    psubd      xmm0, [eax + edx * 4]
+    lea        eax, [eax + 16]
+    psubd      xmm0, [esi]
+    paddd      xmm0, [esi + edx * 4]
+    lea        esi, [esi + 16]
+    cvtdq2ps   xmm0, xmm0
+    mulps      xmm0, xmm4
+    cvtps2dq   xmm0, xmm0
+    packssdw   xmm0, xmm0
+    packuswb   xmm0, xmm0
+    movd       dword ptr [edi], xmm0
+    lea        edi, [edi + 4]
+    sub        ecx, 1
+    jge        l1
+  l1b:
+  }
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  __asm {
+    mov        eax, row
+    mov        edx, cumsum
+    mov        esi, previous_cumsum
+    mov        ecx, width
+    pxor       xmm0, xmm0
+    pxor       xmm1, xmm1
+
+    sub        ecx, 4
+    jl         l4b
+    test       edx, 15
+    jne        l4b
+
+    // 4 pixel loop
+    align      4
+  l4:
+    movdqu     xmm2, [eax]  // 4 argb pixels 16 bytes.
+    lea        eax, [eax + 16]
+    movdqa     xmm4, xmm2
+
+    punpcklbw  xmm2, xmm1
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm2, xmm1
+    punpckhwd  xmm3, xmm1
+
+    punpckhbw  xmm4, xmm1
+    movdqa     xmm5, xmm4
+    punpcklwd  xmm4, xmm1
+    punpckhwd  xmm5, xmm1
+
+    paddd      xmm0, xmm2
+    movdqa     xmm2, [esi]  // previous row above.
+    paddd      xmm2, xmm0
+
+    paddd      xmm0, xmm3
+    movdqa     xmm3, [esi + 16]
+    paddd      xmm3, xmm0
+
+    paddd      xmm0, xmm4
+    movdqa     xmm4, [esi + 32]
+    paddd      xmm4, xmm0
+
+    paddd      xmm0, xmm5
+    movdqa     xmm5, [esi + 48]
+    lea        esi, [esi + 64]
+    paddd      xmm5, xmm0
+
+    movdqa     [edx], xmm2
+    movdqa     [edx + 16], xmm3
+    movdqa     [edx + 32], xmm4
+    movdqa     [edx + 48], xmm5
+
+    lea        edx, [edx + 64]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    movd       xmm2, dword ptr [eax]  // 1 argb pixel 4 bytes.
+    lea        eax, [eax + 4]
+    punpcklbw  xmm2, xmm1
+    punpcklwd  xmm2, xmm1
+    paddd      xmm0, xmm2
+    movdqu     xmm2, [esi]
+    lea        esi, [esi + 16]
+    paddd      xmm2, xmm0
+    movdqu     [edx], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 1
+    jge        l1
+
+ l1b:
+  }
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+__declspec(naked) __declspec(align(16))
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 12]  // src_argb
+    mov        esi, [esp + 16]  // stride
+    mov        edx, [esp + 20]  // dst_argb
+    mov        ecx, [esp + 24]  // pointer to uv_dudv
+    movq       xmm2, qword ptr [ecx]  // uv
+    movq       xmm7, qword ptr [ecx + 8]  // dudv
+    mov        ecx, [esp + 28]  // width
+    shl        esi, 16          // 4, stride
+    add        esi, 4
+    movd       xmm5, esi
+    sub        ecx, 4
+    jl         l4b
+
+    // setup for 4 pixel loop
+    pshufd     xmm7, xmm7, 0x44  // dup dudv
+    pshufd     xmm5, xmm5, 0  // dup 4, stride
+    movdqa     xmm0, xmm2    // x0, y0, x1, y1
+    addps      xmm0, xmm7
+    movlhps    xmm2, xmm0
+    movdqa     xmm4, xmm7
+    addps      xmm4, xmm4    // dudv *= 2
+    movdqa     xmm3, xmm2    // x2, y2, x3, y3
+    addps      xmm3, xmm4
+    addps      xmm4, xmm4    // dudv *= 4
+
+    // 4 pixel loop
+    align      4
+  l4:
+    cvttps2dq  xmm0, xmm2    // x, y float to int first 2
+    cvttps2dq  xmm1, xmm3    // x, y float to int next 2
+    packssdw   xmm0, xmm1    // x, y as 8 shorts
+    pmaddwd    xmm0, xmm5    // offsets = x * 4 + y * stride.
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       xmm1, [eax + esi]  // read pixel 0
+    movd       xmm6, [eax + edi]  // read pixel 1
+    punpckldq  xmm1, xmm6     // combine pixel 0 and 1
+    addps      xmm2, xmm4    // x, y += dx, dy first 2
+    movq       qword ptr [edx], xmm1
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    movd       xmm6, [eax + esi]  // read pixel 2
+    movd       xmm0, [eax + edi]  // read pixel 3
+    punpckldq  xmm6, xmm0     // combine pixel 2 and 3
+    addps      xmm3, xmm4    // x, y += dx, dy next 2
+    sub        ecx, 4
+    movq       qword ptr 8[edx], xmm6
+    lea        edx, [edx + 16]
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+    align      4
+  l1:
+    cvttps2dq  xmm0, xmm2    // x, y float to int
+    packssdw   xmm0, xmm0    // x, y as shorts
+    pmaddwd    xmm0, xmm5    // offset = x * 4 + y * stride
+    addps      xmm2, xmm7    // x, y += dx, dy
+    movd       esi, xmm0
+    movd       xmm0, [eax + esi]  // copy a pixel
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jge        l1
+  l1b:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    sub        edi, esi
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    vmovd      xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    vmovd      xmm5, eax  // low fraction 128..1
+    vpunpcklbw xmm5, xmm5, xmm0
+    vpunpcklwd xmm5, xmm5, xmm5
+    vpxor      ymm0, ymm0, ymm0
+    vpermd     ymm5, ymm0, ymm5
+
+    align      4
+  xloop:
+    vmovdqu    ymm0, [esi]
+    vmovdqu    ymm2, [esi + edx]
+    vpunpckhbw ymm1, ymm0, ymm2  // mutates
+    vpunpcklbw ymm0, ymm0, ymm2  // mutates
+    vpmaddubsw ymm0, ymm0, ymm5
+    vpmaddubsw ymm1, ymm1, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm1, ymm1, 7
+    vpackuswb  ymm0, ymm0, ymm1  // unmutates
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    vmovdqu    ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    vmovdqu    ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi + edx]
+    sub        ecx, 32
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    vmovdqu    ymm0, [esi + edx]
+    vpavgb     ymm0, ymm0, [esi]
+    vpavgb     ymm0, ymm0, [esi]
+    sub        ecx, 32
+    vmovdqu     [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    rep movsb
+
+  xloop99:
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_AVX2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    movd       xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    movd       xmm5, eax  // low fraction 128..1
+    punpcklbw  xmm5, xmm0
+    punpcklwd  xmm5, xmm5
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  xloop:
+    movdqa     xmm0, [esi]
+    movdqa     xmm2, [esi + edx]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm2
+    punpckhbw  xmm1, xmm2
+    pmaddubsw  xmm0, xmm5
+    pmaddubsw  xmm1, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm1, 7
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqa     xmm1, [esi]
+    movdqa     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqa     xmm0, [esi]
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 256.  Blend 100 / 0.
+    cmp        eax, 64
+    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
+    cmp        eax, 128
+    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    cmp        eax, 192
+    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
+
+    movd       xmm5, eax            // xmm5 = y fraction
+    punpcklbw  xmm5, xmm5
+    psrlw      xmm5, 1
+    punpcklwd  xmm5, xmm5
+    punpckldq  xmm5, xmm5
+    punpcklqdq xmm5, xmm5
+    pxor       xmm4, xmm4
+
+    align      4
+  xloop:
+    movdqa     xmm0, [esi]  // row0
+    movdqa     xmm2, [esi + edx]  // row1
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    psubw      xmm2, xmm0  // row1 - row0
+    psubw      xmm3, xmm1
+    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
+    paddw      xmm3, xmm3
+    pmulhw     xmm2, xmm5  // scale diff
+    pmulhw     xmm3, xmm5
+    paddw      xmm0, xmm2  // sum rows
+    paddw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqa     xmm0, [esi]
+    movdqa     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqa     xmm1, [esi]
+    movdqa     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqa     xmm0, [esi]
+    sub        ecx, 16
+    movdqa     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                                    ptrdiff_t src_stride, int dst_width,
+                                    int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    movd       xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    movd       xmm5, eax  // low fraction 128..1
+    punpcklbw  xmm5, xmm0
+    punpcklwd  xmm5, xmm5
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  xloop:
+    movdqu     xmm0, [esi]
+    movdqu     xmm2, [esi + edx]
+    movdqu     xmm1, xmm0
+    punpcklbw  xmm0, xmm2
+    punpckhbw  xmm1, xmm2
+    pmaddubsw  xmm0, xmm5
+    pmaddubsw  xmm1, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm1, 7
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqu     xmm0, [esi]
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked) __declspec(align(16))
+void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride, int dst_width,
+                                   int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 256.  Blend 100 / 0.
+    cmp        eax, 64
+    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
+    cmp        eax, 128
+    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    cmp        eax, 192
+    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
+
+    movd       xmm5, eax            // xmm5 = y fraction
+    punpcklbw  xmm5, xmm5
+    psrlw      xmm5, 1
+    punpcklwd  xmm5, xmm5
+    punpckldq  xmm5, xmm5
+    punpcklqdq xmm5, xmm5
+    pxor       xmm4, xmm4
+
+    align      4
+  xloop:
+    movdqu     xmm0, [esi]  // row0
+    movdqu     xmm2, [esi + edx]  // row1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    psubw      xmm2, xmm0  // row1 - row0
+    psubw      xmm3, xmm1
+    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
+    paddw      xmm3, xmm3
+    pmulhw     xmm2, xmm5  // scale diff
+    pmulhw     xmm3, xmm5
+    paddw      xmm0, xmm2  // sum rows
+    paddw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+    align      4
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+    align      4
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+    align      4
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+    align      4
+  xloop100:
+    movdqu     xmm0, [esi]
+    sub        ecx, 16
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+__declspec(naked) __declspec(align(16))
+void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // src_uv_stride
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    sub        edi, eax
+
+    align      4
+  convertloop:
+    movdqa     xmm0, [eax]
+    pavgb      xmm0, [eax + edx]
+    sub        ecx, 16
+    movdqa     [eax + edi], xmm0
+    lea        eax,  [eax + 16]
+    jg         convertloop
+    pop        edi
+    ret
+  }
+}
+
+#ifdef HAS_HALFROW_AVX2
+__declspec(naked) __declspec(align(16))
+void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
+                  uint8* dst_uv, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // src_uv_stride
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    sub        edi, eax
+
+    align      4
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vpavgb     ymm0, ymm0, [eax + edx]
+    sub        ecx, 32
+    vmovdqu    [eax + edi], ymm0
+    lea        eax,  [eax + 32]
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_HALFROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
+                          uint32 selector, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_bayer
+    movd       xmm5, [esp + 12]  // selector
+    mov        ecx, [esp + 16]   // pix
+    pshufd     xmm5, xmm5, 0
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    punpckldq  xmm0, xmm1
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+    ret
+  }
+}
+
+// Specialized ARGB to Bayer that just isolates G channel.
+__declspec(naked) __declspec(align(16))
+void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
+                           uint32 selector, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_bayer
+                                 // selector
+    mov        ecx, [esp + 16]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x000000ff
+    psrld      xmm5, 24
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm0, 8  // Move green to bottom.
+    psrld      xmm1, 8
+    pand       xmm0, xmm5
+    pand       xmm1, xmm5
+    packssdw   xmm0, xmm1
+    packuswb   xmm0, xmm1
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+    ret
+  }
+}
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_argb
+    mov        ecx, [esp + 12]   // shuffler
+    movdqa     xmm5, [ecx]
+    mov        ecx, [esp + 16]   // pix
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                    const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_argb
+    mov        ecx, [esp + 12]   // shuffler
+    movdqa     xmm5, [ecx]
+    mov        ecx, [esp + 16]   // pix
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    sub        ecx, 8
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]     // src_argb
+    mov        edx, [esp + 8]     // dst_argb
+    mov        ecx, [esp + 12]    // shuffler
+    vbroadcastf128 ymm5, [ecx]    // same shuffle in high as low.
+    mov        ecx, [esp + 16]    // pix
+
+    align      4
+  wloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vpshufb    ymm0, ymm0, ymm5
+    vpshufb    ymm1, ymm1, ymm5
+    sub        ecx, 16
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    jg         wloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+__declspec(naked) __declspec(align(16))
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    push       ebx
+    push       esi
+    mov        eax, [esp + 8 + 4]    // src_argb
+    mov        edx, [esp + 8 + 8]    // dst_argb
+    mov        esi, [esp + 8 + 12]   // shuffler
+    mov        ecx, [esp + 8 + 16]   // pix
+    pxor       xmm5, xmm5
+
+    mov        ebx, [esi]   // shuffler
+    cmp        ebx, 0x03000102
+    je         shuf_3012
+    cmp        ebx, 0x00010203
+    je         shuf_0123
+    cmp        ebx, 0x00030201
+    je         shuf_0321
+    cmp        ebx, 0x02010003
+    je         shuf_2103
+
+  // TODO(fbarchard): Use one source pointer and 3 offsets.
+  shuf_any1:
+    movzx      ebx, byte ptr [esi]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx], bl
+    movzx      ebx, byte ptr [esi + 1]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 1], bl
+    movzx      ebx, byte ptr [esi + 2]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 2], bl
+    movzx      ebx, byte ptr [esi + 3]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 3], bl
+    lea        eax, [eax + 4]
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jg         shuf_any1
+    jmp        shuf99
+
+    align      4
+  shuf_0123:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 01Bh   // 1B = 00011011 = 0x0123 = BGRAToARGB
+    pshuflw    xmm0, xmm0, 01Bh
+    pshufhw    xmm1, xmm1, 01Bh
+    pshuflw    xmm1, xmm1, 01Bh
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_0123
+    jmp        shuf99
+
+    align      4
+  shuf_0321:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 039h   // 39 = 00111001 = 0x0321 = RGBAToARGB
+    pshuflw    xmm0, xmm0, 039h
+    pshufhw    xmm1, xmm1, 039h
+    pshuflw    xmm1, xmm1, 039h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_0321
+    jmp        shuf99
+
+    align      4
+  shuf_2103:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 093h   // 93 = 10010011 = 0x2103 = ARGBToRGBA
+    pshuflw    xmm0, xmm0, 093h
+    pshufhw    xmm1, xmm1, 093h
+    pshuflw    xmm1, xmm1, 093h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_2103
+    jmp        shuf99
+
+    align      4
+  shuf_3012:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 0C6h   // C6 = 11000110 = 0x3012 = ABGRToARGB
+    pshuflw    xmm0, xmm0, 0C6h
+    pshufhw    xmm1, xmm1, 0C6h
+    pshuflw    xmm1, xmm1, 0C6h
+    packuswb   xmm0, xmm1
+    sub        ecx, 4
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         shuf_3012
+
+  shuf99:
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// YUY2 - Macro-pixel = 2 image pixels
+// Y0U0Y1V0....Y2U2Y3V2...Y4U4Y5V4....
+
+// UYVY - Macro-pixel = 2 image pixels
+// U0Y0V0Y1
+
+__declspec(naked) __declspec(align(16))
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm2 // YUYV
+    punpckhbw  xmm1, xmm2
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm1
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked) __declspec(align(16))
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+    align      4
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    movdqa     xmm1, xmm2
+    lea        eax, [eax + 16]
+    punpcklbw  xmm1, xmm0 // UYVY
+    punpckhbw  xmm2, xmm0
+    movdqu     [edi], xmm1
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+__declspec(naked) __declspec(align(16))
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* src_argb */
+    mov        edx, [esp + 4 + 8]   /* dst_argb */
+    mov        esi, [esp + 4 + 12]  /* poly */
+    mov        ecx, [esp + 4 + 16]  /* width */
+    pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
+
+    // 2 pixel loop.
+    align      4
+ convertloop:
+//    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
+//    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
+    movq       xmm0, qword ptr [eax]  // BGRABGRA
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm3
+    movdqa     xmm4, xmm0
+    punpcklwd  xmm0, xmm3  // pixel 0
+    punpckhwd  xmm4, xmm3  // pixel 1
+    cvtdq2ps   xmm0, xmm0  // 4 floats
+    cvtdq2ps   xmm4, xmm4
+    movdqa     xmm1, xmm0  // X
+    movdqa     xmm5, xmm4
+    mulps      xmm0, [esi + 16]  // C1 * X
+    mulps      xmm4, [esi + 16]
+    addps      xmm0, [esi]  // result = C0 + C1 * X
+    addps      xmm4, [esi]
+    movdqa     xmm2, xmm1
+    movdqa     xmm6, xmm5
+    mulps      xmm2, xmm1  // X * X
+    mulps      xmm6, xmm5
+    mulps      xmm1, xmm2  // X * X * X
+    mulps      xmm5, xmm6
+    mulps      xmm2, [esi + 32]  // C2 * X * X
+    mulps      xmm6, [esi + 32]
+    mulps      xmm1, [esi + 48]  // C3 * X * X * X
+    mulps      xmm5, [esi + 48]
+    addps      xmm0, xmm2  // result += C2 * X * X
+    addps      xmm4, xmm6
+    addps      xmm0, xmm1  // result += C3 * X * X * X
+    addps      xmm4, xmm5
+    cvttps2dq  xmm0, xmm0
+    cvttps2dq  xmm4, xmm4
+    packuswb   xmm0, xmm4
+    packuswb   xmm0, xmm0
+    sub        ecx, 2
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]   /* poly */
+    vbroadcastf128 ymm4, [ecx]       // C0
+    vbroadcastf128 ymm5, [ecx + 16]  // C1
+    vbroadcastf128 ymm6, [ecx + 32]  // C2
+    vbroadcastf128 ymm7, [ecx + 48]  // C3
+    mov        ecx, [esp + 16]  /* width */
+
+    // 2 pixel loop.
+    align      4
+ convertloop:
+    vpmovzxbd   ymm0, qword ptr [eax]  // 2 BGRA pixels
+    lea         eax, [eax + 8]
+    vcvtdq2ps   ymm0, ymm0        // X 8 floats
+    vmulps      ymm2, ymm0, ymm0  // X * X
+    vmulps      ymm3, ymm0, ymm7  // C3 * X
+    vfmadd132ps ymm0, ymm4, ymm5  // result = C0 + C1 * X
+    vfmadd231ps ymm0, ymm2, ymm6  // result += C2 * X * X
+    vfmadd231ps ymm0, ymm2, ymm3  // result += C3 * X * X * X
+    vcvttps2dq  ymm0, ymm0
+    vpackusdw   ymm0, ymm0, ymm0  // b0g0r0a0_00000000_b0g0r0a0_00000000
+    vpermq      ymm0, ymm0, 0xd8  // b0g0r0a0_b0g0r0a0_00000000_00000000
+    vpackuswb   xmm0, xmm0, xmm0  // bgrabgra_00000000_00000000_00000000
+    sub         ecx, 2
+    vmovq       qword ptr [edx], xmm0
+    lea         edx, [edx + 8]
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+__declspec(naked) __declspec(align(16))
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+    align      4
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    movzx      edx, byte ptr [eax - 4 + 3]
+    movzx      edx, byte ptr [esi + edx * 4 + 3]
+    mov        byte ptr [eax - 4 + 3], dl
+    dec        ecx
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+__declspec(naked) __declspec(align(16))
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+    align      4
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    dec        ecx
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+__declspec(naked) __declspec(align(16))
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   /* src_argb */
+    mov        edi, [esp + 8 + 8]   /* dst_argb */
+    mov        ecx, [esp + 8 + 12]  /* width */
+    movd       xmm2, dword ptr [esp + 8 + 16]  // luma table
+    movd       xmm3, dword ptr [esp + 8 + 20]  // lumacoeff
+    pshufd     xmm2, xmm2, 0
+    pshufd     xmm3, xmm3, 0
+    pcmpeqb    xmm4, xmm4        // generate mask 0xff00ff00
+    psllw      xmm4, 8
+    pxor       xmm5, xmm5
+
+    // 4 pixel loop.
+    align      4
+  convertloop:
+    movdqu     xmm0, qword ptr [eax]      // generate luma ptr
+    pmaddubsw  xmm0, xmm3
+    phaddw     xmm0, xmm0
+    pand       xmm0, xmm4  // mask out low bits
+    punpcklwd  xmm0, xmm5
+    paddd      xmm0, xmm2  // add table base
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi], dl
+    movzx      edx, byte ptr [eax + 1]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 1], dl
+    movzx      edx, byte ptr [eax + 2]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 2], dl
+    movzx      edx, byte ptr [eax + 3]  // copy alpha.
+    mov        byte ptr [edi + 3], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 4]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 4], dl
+    movzx      edx, byte ptr [eax + 5]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 5], dl
+    movzx      edx, byte ptr [eax + 6]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 6], dl
+    movzx      edx, byte ptr [eax + 7]  // copy alpha.
+    mov        byte ptr [edi + 7], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 8]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 8], dl
+    movzx      edx, byte ptr [eax + 9]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 9], dl
+    movzx      edx, byte ptr [eax + 10]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 10], dl
+    movzx      edx, byte ptr [eax + 11]  // copy alpha.
+    mov        byte ptr [edi + 11], dl
+
+    movd       esi, xmm0
+
+    movzx      edx, byte ptr [eax + 12]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 12], dl
+    movzx      edx, byte ptr [eax + 13]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 13], dl
+    movzx      edx, byte ptr [eax + 14]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 14], dl
+    movzx      edx, byte ptr [eax + 15]  // copy alpha.
+    mov        byte ptr [edi + 15], dl
+
+    sub        ecx, 4
+    lea        eax, [eax + 16]
+    lea        edi, [edi + 16]
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/row_x86.asm b/drivers/theoraplayer/src/YUV/libyuv/src/row_x86.asm
new file mode 100755
index 0000000000..0cb326f8e5
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/row_x86.asm
@@ -0,0 +1,146 @@
+;
+; Copyright 2012 The LibYuv Project Authors. All rights reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+%ifdef __YASM_VERSION_ID__
+%if __YASM_VERSION_ID__ < 01020000h
+%error AVX2 is supported only by yasm 1.2.0 or later.
+%endif
+%endif
+%include "x86inc.asm"
+
+SECTION .text
+
+; cglobal numeric constants are parameters, gpr regs, mm regs
+
+; void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix)
+
+%macro YUY2TOYROW 2-3
+cglobal %1ToYRow%3, 3, 3, 3, src_yuy2, dst_y, pix
+%ifidn %1,YUY2
+    pcmpeqb    m2, m2, m2        ; generate mask 0x00ff00ff
+    psrlw      m2, m2, 8
+%endif
+
+    ALIGN      4
+.convertloop:
+    mov%2      m0, [src_yuy2q]
+    mov%2      m1, [src_yuy2q + mmsize]
+    lea        src_yuy2q, [src_yuy2q + mmsize * 2]
+%ifidn %1,YUY2
+    pand       m0, m0, m2   ; YUY2 even bytes are Y
+    pand       m1, m1, m2
+%else
+    psrlw      m0, m0, 8    ; UYVY odd bytes are Y
+    psrlw      m1, m1, 8
+%endif
+    packuswb   m0, m0, m1
+%if cpuflag(AVX2)
+    vpermq     m0, m0, 0xd8
+%endif
+    sub        pixd, mmsize
+    mov%2      [dst_yq], m0
+    lea        dst_yq, [dst_yq + mmsize]
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+; TODO(fbarchard): Remove MMX.  Add SSSE3 pshufb version.
+INIT_MMX MMX
+YUY2TOYROW YUY2,a,
+YUY2TOYROW YUY2,u,_Unaligned
+YUY2TOYROW UYVY,a,
+YUY2TOYROW UYVY,u,_Unaligned
+INIT_XMM SSE2
+YUY2TOYROW YUY2,a,
+YUY2TOYROW YUY2,u,_Unaligned
+YUY2TOYROW UYVY,a,
+YUY2TOYROW UYVY,u,_Unaligned
+INIT_YMM AVX2
+YUY2TOYROW YUY2,a,
+YUY2TOYROW UYVY,a,
+
+; void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix)
+
+%macro SplitUVRow 1-2
+cglobal SplitUVRow%2, 4, 4, 5, src_uv, dst_u, dst_v, pix
+    pcmpeqb    m4, m4, m4        ; generate mask 0x00ff00ff
+    psrlw      m4, m4, 8
+    sub        dst_vq, dst_uq
+
+    ALIGN      4
+.convertloop:
+    mov%1      m0, [src_uvq]
+    mov%1      m1, [src_uvq + mmsize]
+    lea        src_uvq, [src_uvq + mmsize * 2]
+    psrlw      m2, m0, 8         ; odd bytes
+    psrlw      m3, m1, 8
+    pand       m0, m0, m4        ; even bytes
+    pand       m1, m1, m4
+    packuswb   m0, m0, m1
+    packuswb   m2, m2, m3
+%if cpuflag(AVX2)
+    vpermq     m0, m0, 0xd8
+    vpermq     m2, m2, 0xd8
+%endif
+    mov%1      [dst_uq], m0
+    mov%1      [dst_uq + dst_vq], m2
+    lea        dst_uq, [dst_uq + mmsize]
+    sub        pixd, mmsize
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+INIT_MMX MMX
+SplitUVRow a,
+SplitUVRow u,_Unaligned
+INIT_XMM SSE2
+SplitUVRow a,
+SplitUVRow u,_Unaligned
+INIT_YMM AVX2
+SplitUVRow a,
+
+; void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+;                      int width);
+
+%macro MergeUVRow_ 1-2
+cglobal MergeUVRow_%2, 4, 4, 3, src_u, src_v, dst_uv, pix
+    sub        src_vq, src_uq
+
+    ALIGN      4
+.convertloop:
+    mov%1      m0, [src_uq]
+    mov%1      m1, [src_vq]
+    lea        src_uq, [src_uq + mmsize]
+    punpcklbw  m2, m0, m1       // first 8 UV pairs
+    punpckhbw  m0, m0, m1       // next 8 UV pairs
+%if cpuflag(AVX2)
+    vperm2i128 m1, m2, m0, 0x20  // low 128 of ymm2 and low 128 of ymm0
+    vperm2i128 m2, m2, m0, 0x31  // high 128 of ymm2 and high 128 of ymm0
+    mov%1      [dst_uvq], m1
+    mov%1      [dst_uvq + mmsize], m2
+%else
+    mov%1      [dst_uvq], m2
+    mov%1      [dst_uvq + mmsize], m0
+%endif
+    lea        dst_uvq, [dst_uvq + mmsize * 2]
+    sub        pixd, mmsize
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+INIT_MMX MMX
+MergeUVRow_ a,
+MergeUVRow_ u,_Unaligned
+INIT_XMM SSE2
+MergeUVRow_ a,
+MergeUVRow_ u,_Unaligned
+INIT_YMM AVX2
+MergeUVRow_ a,
+
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/scale.cc b/drivers/theoraplayer/src/YUV/libyuv/src/scale.cc
new file mode 100755
index 0000000000..b3893cc00c
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/scale.cc
@@ -0,0 +1,926 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyPlane
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Remove this macro if OVERREAD is safe.
+#define AVOID_OVERREAD 1
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+
+// Scale plane, 1/2
+// This is an optimized version for scaling down a plane to 1/2 of
+// its original size.
+
+static void ScalePlaneDown2(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+    filtering == kFilterNone ? ScaleRowDown2_C :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_C :
+        ScaleRowDown2Box_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering ? ScaleRowDown2Box_NEON : ScaleRowDown2_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Unaligned_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Unaligned_SSE2 :
+        ScaleRowDown2Box_Unaligned_SSE2);
+    if (IS_ALIGNED(src_ptr, 16) &&
+        IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
+          ScaleRowDown2Box_SSE2);
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_MIPS_DSPR2 : ScaleRowDown2_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane, 1/4
+// This is an optimized version for scaling down a plane to 1/4 of
+// its original size.
+
+static void ScalePlaneDown4(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_C : ScaleRowDown4_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_NEON : ScaleRowDown4_NEON;
+  }
+#elif defined(HAS_SCALEROWDOWN4_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) &&
+      IS_ALIGNED(dst_width, 8) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_SSE2 : ScaleRowDown4_SSE2;
+  }
+#elif defined(HAS_SCALEROWDOWN4_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_MIPS_DSPR2 : ScaleRowDown4_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane down, 3/4
+
+static void ScalePlaneDown34(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_C;
+    ScaleRowDown34_1 = ScaleRowDown34_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+
+// Scale plane, 3/8
+// This is an optimized version for scaling down a plane to 3/8
+// of its original size.
+//
+// Uses box filter arranges like this
+// aaabbbcc -> abc
+// aaabbbcc    def
+// aaabbbcc    ghi
+// dddeeeff
+// dddeeeff
+// dddeeeff
+// ggghhhii
+// ggghhhii
+// Boxes are 3x3, 2x3, 3x2 and 2x2
+
+static void ScalePlaneDown38(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_C;
+    ScaleRowDown38_2 = ScaleRowDown38_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_C;
+  }
+#if defined(HAS_SCALEROWDOWN38_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_NEON;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_SSSE3;
+    }
+  }
+#elif defined(HAS_SCALEROWDOWN38_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static __inline uint32 SumBox(int iboxwidth, int iboxheight,
+                              ptrdiff_t src_stride, const uint8* src_ptr) {
+  uint32 sum = 0u;
+  int y;
+  assert(iboxwidth > 0);
+  assert(iboxheight > 0);
+  for (y = 0; y < iboxheight; ++y) {
+    int x;
+    for (x = 0; x < iboxwidth; ++x) {
+      sum += src_ptr[x];
+    }
+    src_ptr += src_stride;
+  }
+  return sum;
+}
+
+static void ScalePlaneBoxRow_C(int dst_width, int boxheight,
+                               int x, int dx, ptrdiff_t src_stride,
+                               const uint8* src_ptr, uint8* dst_ptr) {
+  int i;
+  int boxwidth;
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumBox(boxwidth, boxheight, src_stride, src_ptr + ix) /
+        (boxwidth * boxheight);
+  }
+}
+
+static __inline uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static void ScaleAddCols2_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = (dx >> 16);
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (minboxwidth * boxheight);
+  scaletbl[1] = 65536 / ((minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = (x >> 16) - ix;
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols1_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int boxwidth = (dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+// Scale plane down to any dimensions, with interpolation.
+// (boxfilter).
+//
+// Same method as SimpleScale, which is fixed point, outputting
+// one pixel of destination using fixed point (16.16) to step
+// through source, sampling a box of pixel with simple
+// averaging.
+static void ScalePlaneBox(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  // TODO(fbarchard): Remove this and make AddRows handle boxheight 1.
+  if (!IS_ALIGNED(src_width, 16) || dst_height * 2 > src_height) {
+    uint8* dst = dst_ptr;
+    int j;
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint8* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = (y >> 16) - iy;
+      ScalePlaneBoxRow_C(dst_width, boxheight,
+                         x, dx, src_stride,
+                         src, dst);
+      dst += dst_stride;
+    }
+    return;
+  }
+  {
+    // Allocate a row buffer of uint16.
+    align_buffer_64(row16, src_width * 2);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint16* src_ptr, uint8* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_C: ScaleAddCols1_C;
+    void (*ScaleAddRows)(const uint8* src_ptr, ptrdiff_t src_stride,
+        uint16* dst_ptr, int src_width, int src_height) = ScaleAddRows_C;
+
+#if defined(HAS_SCALEADDROWS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) &&
+#ifdef AVOID_OVERREAD
+        IS_ALIGNED(src_width, 16) &&
+#endif
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+      ScaleAddRows = ScaleAddRows_SSE2;
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint8* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > (src_height << 16)) {
+        y = (src_height << 16);
+      }
+      boxheight = (y >> 16) - iy;
+      ScaleAddRows(src, src_stride, (uint16*)(row16),
+                 src_width, boxheight);
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint16*)(row16),
+                 dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row16);
+  }
+}
+
+// Scale plane down with bilinear interpolation.
+void ScalePlaneBilinearDown(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_C : ScaleFilterCols_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && src_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+// Scale up down with bilinear interpolation.
+void ScalePlaneBilinearUp(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr,
+                          enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+       int dst_width, int x, int dx) =
+       filtering ? ScaleFilterCols_C : ScaleCols_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleFilterCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 15) & ~15;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale Plane to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScalePlaneSimple(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_ptr, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride,
+              dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+// Scale a plane.
+// This function dispatches to a specialized scaler based on scale factor.
+
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical(src_height,
+                       dst_width, dst_height,
+                       src_stride, dst_stride, src, dst,
+                       0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+               filtering != kFilterBilinear) {
+      // optimized, 1/4
+      ScalePlaneDown4(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox(src_width, src_height, dst_width, dst_height,
+                  src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown(src_width, src_height, dst_width, dst_height,
+                           src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
+                   src_stride, dst_stride, src, dst);
+}
+
+// Scale an I420 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int I420Scale(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              int src_width, int src_height,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height,
+             dst_y, dst_stride_y, dst_width, dst_height,
+             filtering);
+  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
+             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+             filtering);
+  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
+             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+             filtering);
+  return 0;
+}
+
+// Deprecated api
+LIBYUV_API
+int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
+          int src_stride_y, int src_stride_u, int src_stride_v,
+          int src_width, int src_height,
+          uint8* dst_y, uint8* dst_u, uint8* dst_v,
+          int dst_stride_y, int dst_stride_u, int dst_stride_v,
+          int dst_width, int dst_height,
+          LIBYUV_BOOL interpolate) {
+  return I420Scale(src_y, src_stride_y,
+                   src_u, src_stride_u,
+                   src_v, src_stride_v,
+                   src_width, src_height,
+                   dst_y, dst_stride_y,
+                   dst_u, dst_stride_u,
+                   dst_v, dst_stride_v,
+                   dst_width, dst_height,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+// Deprecated api
+LIBYUV_API
+int ScaleOffset(const uint8* src, int src_width, int src_height,
+                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate) {
+  // Chroma requires offset to multiple of 2.
+  int dst_yoffset_even = dst_yoffset & ~1;
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  int aheight = dst_height - dst_yoffset_even * 2;  // actual output height
+  const uint8* src_y = src;
+  const uint8* src_u = src + src_width * src_height;
+  const uint8* src_v = src + src_width * src_height +
+                             src_halfwidth * src_halfheight;
+  uint8* dst_y = dst + dst_yoffset_even * dst_width;
+  uint8* dst_u = dst + dst_width * dst_height +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  if (!src || src_width <= 0 || src_height <= 0 ||
+      !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset_even < 0 ||
+      dst_yoffset_even >= dst_height) {
+    return -1;
+  }
+  return I420Scale(src_y, src_width,
+                   src_u, src_halfwidth,
+                   src_v, src_halfwidth,
+                   src_width, src_height,
+                   dst_y, dst_width,
+                   dst_u, dst_halfwidth,
+                   dst_v, dst_halfwidth,
+                   dst_width, aheight,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/scale_argb.cc b/drivers/theoraplayer/src/YUV/libyuv/src/scale_argb.cc
new file mode 100755
index 0000000000..e339cd7c79
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/scale_argb.cc
@@ -0,0 +1,809 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// ScaleARGB ARGB, 1/2
+// This is an optimized version for scaling down a ARGB to 1/2 of
+// its original size.
+static void ScaleARGBDown2(int src_width, int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint8* src_argb, uint8* dst_argb,
+                           int x, int dx, int y, int dy,
+                           enum FilterMode filtering) {
+  int j;
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) =
+    filtering == kFilterNone ? ScaleARGBRowDown2_C :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C :
+        ScaleARGBRowDown2Box_C);
+  assert(dx == 65536 * 2);  // Test scale factor of 2.
+  assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
+  // Advance to odd row, even column.
+  if (filtering == kFilterBilinear) {
+    src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  } else {
+    src_argb += (y >> 16) * src_stride + ((x >> 16) - 1) * 4;
+  }
+
+#if defined(HAS_SCALEARGBROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 :
+        ScaleARGBRowDown2Box_SSE2);
+  }
+#elif defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(row_stride, 4)) {
+    ScaleARGBRowDown2 = filtering ? ScaleARGBRowDown2Box_NEON :
+        ScaleARGBRowDown2_NEON;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDown2(src_argb, src_stride, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+}
+
+// ScaleARGB ARGB, 1/4
+// This is an optimized version for scaling down a ARGB to 1/4 of
+// its original size.
+static void ScaleARGBDown4Box(int src_width, int src_height,
+                              int dst_width, int dst_height,
+                              int src_stride, int dst_stride,
+                              const uint8* src_argb, uint8* dst_argb,
+                              int x, int dx, int y, int dy) {
+  int j;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (dst_width * 2 * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
+    uint8* dst_argb, int dst_width) = ScaleARGBRowDown2Box_C;
+  // Advance to odd row, even column.
+  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  assert(dx == 65536 * 4);  // Test scale factor of 4.
+  assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
+#if defined(HAS_SCALEARGBROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
+      IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ScaleARGBRowDown2 = ScaleARGBRowDown2Box_SSE2;
+  }
+#elif defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(row_stride, 4)) {
+    ScaleARGBRowDown2 = ScaleARGBRowDown2Box_NEON;
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDown2(src_argb, src_stride, row, dst_width * 2);
+    ScaleARGBRowDown2(src_argb + src_stride * 2, src_stride,
+                      row + kRowSize, dst_width * 2);
+    ScaleARGBRowDown2(row, kRowSize, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+// ScaleARGB ARGB Even
+// This is an optimized version for scaling down a ARGB to even
+// multiple of its original size.
+static void ScaleARGBDownEven(int src_width, int src_height,
+                              int dst_width, int dst_height,
+                              int src_stride, int dst_stride,
+                              const uint8* src_argb, uint8* dst_argb,
+                              int x, int dx, int y, int dy,
+                              enum FilterMode filtering) {
+  int j;
+  int col_step = dx >> 16;
+  int row_stride = (dy >> 16) * src_stride;
+  void (*ScaleARGBRowDownEven)(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_step, uint8* dst_argb, int dst_width) =
+      filtering ? ScaleARGBRowDownEvenBox_C : ScaleARGBRowDownEven_C;
+  assert(IS_ALIGNED(src_width, 2));
+  assert(IS_ALIGNED(src_height, 2));
+  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_SSE2 :
+        ScaleARGBRowDownEven_SSE2;
+  }
+#elif defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 4) &&
+      IS_ALIGNED(src_argb, 4)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_NEON :
+        ScaleARGBRowDownEven_NEON;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDownEven(src_argb, src_stride, col_step, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+}
+
+// Scale ARGB down with bilinear interpolation.
+static void ScaleARGBBilinearDown(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  int src_stride, int dst_stride,
+                                  const uint8* src_argb, uint8* dst_argb,
+                                  int x, int dx, int y, int dy,
+                                  enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBFilterCols64_C : ScaleARGBFilterCols_C;
+  int64 xlast = x + (int64)(dst_width - 1) * dx;
+  int64 xl = (dx >= 0) ? x : xlast;
+  int64 xr = (dx >= 0) ? xlast : x;
+  int clip_src_width;
+  xl = (xl >> 16) & ~3;  // Left edge aligned.
+  xr = (xr >> 16) + 1;  // Right most pixel used.  Bilinear uses 2 pixels.
+  xr = (xr + 1 + 3) & ~3;  // 1 beyond 4 pixel aligned right most pixel.
+  if (xr > src_width) {
+    xr = src_width;
+  }
+  clip_src_width = (int)(xr - xl) * 4;  // Width aligned to 4.
+  src_argb += xl * 4;
+  x -= (int)(xl << 16);
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && clip_src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && clip_src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && clip_src_width >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && clip_src_width >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && clip_src_width >= 4 &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(clip_src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row of ARGB.
+  {
+    align_buffer_64(row, clip_src_width * 4);
+
+    const int max_y = (src_height - 1) << 16;
+    if (y > max_y) {
+      y = max_y;
+    }
+    for (j = 0; j < dst_height; ++j) {
+      int yi = y >> 16;
+      const uint8* src = src_argb + yi * src_stride;
+      if (filtering == kFilterLinear) {
+        ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(row, src, src_stride, clip_src_width, yf);
+        ScaleARGBFilterCols(dst_argb, row, dst_width, x, dx);
+      }
+      dst_argb += dst_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale ARGB up with bilinear interpolation.
+static void ScaleARGBBilinearUp(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint8* src_argb, uint8* dst_argb,
+                                int x, int dx, int y, int dy,
+                                enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
+  const int max_y = (src_height - 1) << 16;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 8) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 1 &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+  }
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBCols_SSE2;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBFilterCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  {
+    int yi = y >> 16;
+    const uint8* src = src_argb + yi * src_stride;
+
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (dst_width * 4 + 15) & ~15;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleARGBFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_argb + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
+      }
+      dst_argb += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+#ifdef YUVSCALEUP
+// Scale YUV to ARGB up with bilinear interpolation.
+static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
+                                     int dst_width, int dst_height,
+                                     int src_stride_y,
+                                     int src_stride_u,
+                                     int src_stride_v,
+                                     int dst_stride_argb,
+                                     const uint8* src_y,
+                                     const uint8* src_u,
+                                     const uint8* src_v,
+                                     uint8* dst_argb,
+                                     int x, int dx, int y, int dy,
+                                     enum FilterMode filtering) {
+  int j;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        I422ToARGBRow = I422ToARGBRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && src_width >= 16) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width >= 8) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width >= 8) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width >= 4) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width >= 1 &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+  }
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBCols_SSE2;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBFilterCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  const int max_y = (src_height - 1) << 16;
+  if (y > max_y) {
+    y = max_y;
+  }
+  const int kYShift = 1;  // Shift Y by 1 to convert Y plane to UV coordinate.
+  int yi = y >> 16;
+  int uv_yi = yi >> kYShift;
+  const uint8* src_row_y = src_y + yi * src_stride_y;
+  const uint8* src_row_u = src_u + uv_yi * src_stride_u;
+  const uint8* src_row_v = src_v + uv_yi * src_stride_v;
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (dst_width * 4 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+
+  // Allocate 1 row of ARGB for source conversion.
+  align_buffer_64(argb_row, src_width * 4);
+
+  uint8* rowptr = row;
+  int rowstride = kRowSize;
+  int lasty = yi;
+
+  // TODO(fbarchard): Convert first 2 rows of YUV to ARGB.
+  ScaleARGBFilterCols(rowptr, src_row_y, dst_width, x, dx);
+  if (src_height > 1) {
+    src_row_y += src_stride_y;
+    if (yi & 1) {
+      src_row_u += src_stride_u;
+      src_row_v += src_stride_v;
+    }
+  }
+  ScaleARGBFilterCols(rowptr + rowstride, src_row_y, dst_width, x, dx);
+  if (src_height > 2) {
+    src_row_y += src_stride_y;
+    if (!(yi & 1)) {
+      src_row_u += src_stride_u;
+      src_row_v += src_stride_v;
+    }
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    yi = y >> 16;
+    if (yi != lasty) {
+      if (y > max_y) {
+        y = max_y;
+        yi = y >> 16;
+        uv_yi = yi >> kYShift;
+        src_row_y = src_y + yi * src_stride_y;
+        src_row_u = src_u + uv_yi * src_stride_u;
+        src_row_v = src_v + uv_yi * src_stride_v;
+      }
+      if (yi != lasty) {
+        // TODO(fbarchard): Convert the clipped region of row.
+        I422ToARGBRow(src_row_y, src_row_u, src_row_v, argb_row, src_width);
+        ScaleARGBFilterCols(rowptr, argb_row, dst_width, x, dx);
+        rowptr += rowstride;
+        rowstride = -rowstride;
+        lasty = yi;
+        src_row_y += src_stride_y;
+        if (yi & 1) {
+          src_row_u += src_stride_u;
+          src_row_v += src_stride_v;
+        }
+      }
+    }
+    if (filtering == kFilterLinear) {
+      InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
+    }
+    dst_argb += dst_stride_argb;
+    y += dy;
+  }
+  free_aligned_buffer_64(row);
+  free_aligned_buffer_64(row_argb);
+}
+#endif
+
+// Scale ARGB to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScaleARGBSimple(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_argb, uint8* dst_argb,
+                            int x, int dx, int y, int dy) {
+  int j;
+  void (*ScaleARGBCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBCols64_C : ScaleARGBCols_C;
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBCols = ScaleARGBCols_SSE2;
+  }
+#endif
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8) &&
+        IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+        IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+      ScaleARGBCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBCols(dst_argb, src_argb + (y >> 16) * src_stride,
+                  dst_width, x, dx);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// ScaleARGB a ARGB.
+// This function in turn calls a scaling function
+// suitable for handling the desired resolutions.
+static void ScaleARGB(const uint8* src, int src_stride,
+                      int src_width, int src_height,
+                      uint8* dst, int dst_stride,
+                      int dst_width, int dst_height,
+                      int clip_x, int clip_y, int clip_width, int clip_height,
+                      enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // ARGB does not support box filter yet, but allow the user to pass it.
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative src_height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  if (clip_x) {
+    int64 clipf = (int64)(clip_x) * dx;
+    x += (clipf & 0xffff);
+    src += (clipf >> 16) * 4;
+    dst += clip_x * 4;
+  }
+  if (clip_y) {
+    int64 clipf = (int64)(clip_y) * dy;
+    y += (clipf & 0xffff);
+    src += (clipf >> 16) * src_stride;
+    dst += clip_y * dst_stride;
+  }
+
+  // Special case for integer step values.
+  if (((dx | dy) & 0xffff) == 0) {
+    if (!dx || !dy) {  // 1 pixel wide and/or tall.
+      filtering = kFilterNone;
+    } else {
+      // Optimized even scale down. ie 2, 4, 6, 8, 10x.
+      if (!(dx & 0x10000) && !(dy & 0x10000)) {
+        if (dx == 0x20000) {
+          // Optimized 1/2 downsample.
+          ScaleARGBDown2(src_width, src_height,
+                         clip_width, clip_height,
+                         src_stride, dst_stride, src, dst,
+                         x, dx, y, dy, filtering);
+          return;
+        }
+        if (dx == 0x40000 && filtering == kFilterBox) {
+          // Optimized 1/4 box downsample.
+          ScaleARGBDown4Box(src_width, src_height,
+                            clip_width, clip_height,
+                            src_stride, dst_stride, src, dst,
+                            x, dx, y, dy);
+          return;
+        }
+        ScaleARGBDownEven(src_width, src_height,
+                          clip_width, clip_height,
+                          src_stride, dst_stride, src, dst,
+                          x, dx, y, dy, filtering);
+        return;
+      }
+      // Optimized odd scale down. ie 3, 5, 7, 9x.
+      if ((dx & 0x10000) && (dy & 0x10000)) {
+        filtering = kFilterNone;
+        if (dx == 0x10000 && dy == 0x10000) {
+          // Straight copy.
+          ARGBCopy(src + (y >> 16) * src_stride + (x >> 16) * 4, src_stride,
+                   dst, dst_stride, clip_width, clip_height);
+          return;
+        }
+      }
+    }
+  }
+  if (dx == 0x10000 && (x & 0xffff) == 0) {
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical(src_height,
+                       clip_width, clip_height,
+                       src_stride, dst_stride, src, dst,
+                       x, y, dy, 4, filtering);
+    return;
+  }
+  if (filtering && dy < 65536) {
+    ScaleARGBBilinearUp(src_width, src_height,
+                        clip_width, clip_height,
+                        src_stride, dst_stride, src, dst,
+                        x, dx, y, dy, filtering);
+    return;
+  }
+  if (filtering) {
+    ScaleARGBBilinearDown(src_width, src_height,
+                          clip_width, clip_height,
+                          src_stride, dst_stride, src, dst,
+                          x, dx, y, dy, filtering);
+    return;
+  }
+  ScaleARGBSimple(src_width, src_height, clip_width, clip_height,
+                  src_stride, dst_stride, src, dst,
+                  x, dx, y, dy);
+}
+
+LIBYUV_API
+int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
+                  int src_width, int src_height,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int dst_width, int dst_height,
+                  int clip_x, int clip_y, int clip_width, int clip_height,
+                  enum FilterMode filtering) {
+  if (!src_argb || src_width == 0 || src_height == 0 ||
+      !dst_argb || dst_width <= 0 || dst_height <= 0 ||
+      clip_x < 0 || clip_y < 0 ||
+      (clip_x + clip_width) > dst_width ||
+      (clip_y + clip_height) > dst_height) {
+    return -1;
+  }
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
+            dst_argb, dst_stride_argb, dst_width, dst_height,
+            clip_x, clip_y, clip_width, clip_height, filtering);
+  return 0;
+}
+
+// Scale an ARGB image.
+LIBYUV_API
+int ARGBScale(const uint8* src_argb, int src_stride_argb,
+              int src_width, int src_height,
+              uint8* dst_argb, int dst_stride_argb,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  if (!src_argb || src_width == 0 || src_height == 0 ||
+      !dst_argb || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
+            dst_argb, dst_stride_argb, dst_width, dst_height,
+            0, 0, dst_width, dst_height, filtering);
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/scale_argb_neon.cc b/drivers/theoraplayer/src/YUV/libyuv/src/scale_argb_neon.cc
new file mode 100755
index 0000000000..c0b5433239
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/scale_argb_neon.cc
@@ -0,0 +1,145 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
+                            uint8* dst, int dst_width) {
+  asm volatile (
+#ifdef _ANDROID

+	".fpu neon\n"

+#endif
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    "vld2.32    {q0, q1}, [%0]!                \n"
+    "vld2.32    {q2, q3}, [%0]!                \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    "vst1.8     {q3}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
+    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    "vrshrn.u16 d2, q2, #2                     \n"
+    "vrshrn.u16 d3, q3, #2                     \n"
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t, int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %3, lsl #2                \n"
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "vld1.32    {d0[0]}, [%0], r12             \n"
+    "vld1.32    {d0[1]}, [%0], r12             \n"
+    "vld1.32    {d1[0]}, [%0], r12             \n"
+    "vld1.32    {d1[1]}, [%0], r12             \n"
+    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(dst_width)    // %2
+  : "r"(src_stepx)     // %3
+  : "memory", "cc", "r12", "q0"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov       r12, %4, lsl #2                 \n"
+    "add       %1, %1, %0                      \n"
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "vld1.8    {d0}, [%0], r12                 \n"  // Read 4 2x2 blocks -> 2x1
+    "vld1.8    {d1}, [%1], r12                 \n"
+    "vld1.8    {d2}, [%0], r12                 \n"
+    "vld1.8    {d3}, [%1], r12                 \n"
+    "vld1.8    {d4}, [%0], r12                 \n"
+    "vld1.8    {d5}, [%1], r12                 \n"
+    "vld1.8    {d6}, [%0], r12                 \n"
+    "vld1.8    {d7}, [%1], r12                 \n"
+    "vaddl.u8  q0, d0, d1                      \n"
+    "vaddl.u8  q1, d2, d3                      \n"
+    "vaddl.u8  q2, d4, d5                      \n"
+    "vaddl.u8  q3, d6, d7                      \n"
+    "vswp.8    d1, d2                          \n"  // ab_cd -> ac_bd
+    "vswp.8    d5, d6                          \n"  // ef_gh -> eg_fh
+    "vadd.u16  q0, q0, q1                      \n"  // (a+b)_(c+d)
+    "vadd.u16  q2, q2, q3                      \n"  // (e+f)_(g+h)
+    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
+    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
+    "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(src_stride),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(dst_width)    // %3
+  : "r"(src_stepx)     // %4
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/scale_common.cc b/drivers/theoraplayer/src/YUV/libyuv/src/scale_common.cc
new file mode 100644
index 0000000000..6ed8bfaf97
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/scale_common.cc
@@ -0,0 +1,772 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// CPU agnostic row functions
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+// Filter rows 0 and 1 together, 3 : 1
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Filter rows 1 and 2 together, 1 : 1
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+// (1-f)a + fb can be replaced with a + f(b-a)
+#define BLENDER(a, b, f) (uint8)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+// 8x3 -> 3x1
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+// 8x2 -> 3x1
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleAddRows_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  int x;
+  assert(src_width > 0);
+  assert(src_height > 0);
+  for (x = 0; x < src_width; ++x) {
+    const uint8* s = src_ptr + x;
+    unsigned int sum = 0u;
+    int y;
+    for (y = 0; y < src_height; ++y) {
+      sum += s[0];
+      s += src_stride;
+    }
+    // TODO(fbarchard): Consider limitting height to 256 to avoid overflow.
+    dst_ptr[x] = sum < 65535u ? sum : 65535u;
+  }
+}
+
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[1];
+    dst[1] = src[3];
+    src += 4;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[1];
+  }
+}
+
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1;
+    dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1;
+    dst_argb[2] = (src_argb[2] + src_argb[6] + 1) >> 1;
+    dst_argb[3] = (src_argb[3] + src_argb[7] + 1) >> 1;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[0];
+    dst[1] = src[src_stepx];
+    src += src_stepx * 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += src_stepx * 4;
+    dst_argb += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[1] = dst[0] = src[0];
+    src += 1;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+// Mimics SSSE3 blender
+#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7
+#define BLENDERC(a, b, f, s) (uint32)( \
+    BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
+#define BLENDER(a, b, f) \
+    BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \
+    BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
+
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+#undef BLENDER1
+#undef BLENDERC
+#undef BLENDER
+
+// Scale plane vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher bpp.
+  int dst_width_bytes = dst_width * bpp;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(bpp >= 1 && bpp <= 4);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * bpp;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSE2;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSE2;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_Unaligned_SSSE3;
+      if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
+          IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
+        InterpolateRow = InterpolateRow_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width_bytes >= 32) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && dst_width_bytes >= 16) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width_bytes >= 4 &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_bytes, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering) {
+  if (src_width < 0) {
+    src_width = -src_width;
+  }
+  if (src_height < 0) {
+    src_height = -src_height;
+  }
+  if (filtering == kFilterBox) {
+    // If scaling both axis to 0.5 or larger, switch from Box to Bilinear.
+    if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) {
+      filtering = kFilterBilinear;
+    }
+    // If scaling to larger, switch from Box to Bilinear.
+    if (dst_width >= src_width || dst_height >= src_height) {
+      filtering = kFilterBilinear;
+    }
+  }
+  if (filtering == kFilterBilinear) {
+    if (src_height == 1) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to Linear.
+    if (dst_height == src_height || dst_height * 3 == src_height) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Remove 1 pixel wide filter restriction, which is to
+    // avoid reading 2 pixels horizontally that causes memory exception.
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+  }
+  if (filtering == kFilterLinear) {
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to None.
+    if (dst_width == src_width || dst_width * 3 == src_width) {
+      filtering = kFilterNone;
+    }
+  }
+  return filtering;
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div) {
+  return (int)(((int64)(num) << 16) / div);
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div) {
+  return (int)((((int64)(num) << 16) - 0x00010001) /
+                          (div - 1));
+}
+
+#define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s)
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy) {
+  assert(x != NULL);
+  assert(y != NULL);
+  assert(dx != NULL);
+  assert(dy != NULL);
+  assert(src_width != 0);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  // Check for 1 pixel and avoid FixedDiv overflow.
+  if (dst_width == 1 && src_width >= 32768) {
+    dst_width = src_width;
+  }
+  if (dst_height == 1 && src_height >= 32768) {
+    dst_height = src_height;
+  }
+  if (filtering == kFilterBox) {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = 0;
+    *y = 0;
+  } else if (filtering == kFilterBilinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    if (dst_height <= src_height) {
+      *dy = FixedDiv(src_height,  dst_height);
+      *y = CENTERSTART(*dy, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_height > 1) {
+      *dy = FixedDiv1(src_height, dst_height);
+      *y = 0;
+    }
+  } else if (filtering == kFilterLinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    *dy = FixedDiv(src_height, dst_height);
+    *y = *dy >> 1;
+  } else {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = CENTERSTART(*dx, 0);
+    *y = CENTERSTART(*dy, 0);
+  }
+  // Negative src_width means horizontally mirror.
+  if (src_width < 0) {
+    *x += (dst_width - 1) * *dx;
+    *dx = -*dx;
+    // src_width = -src_width;   // Caller must do this.
+  }
+}
+#undef CENTERSTART
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/scale_mips.cc b/drivers/theoraplayer/src/YUV/libyuv/src/scale_mips.cc
new file mode 100755
index 0000000000..4572f4504e
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/scale_mips.cc
@@ -0,0 +1,653 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC MIPS DSPR2
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__(
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+
+    "srl            $t9, %[dst_width], 4           \n"  // iterations -> by 16
+    "beqz           $t9, 2f                        \n"
+    " nop                                          \n"
+
+    ".p2align       2                              \n"
+  "1:                                              \n"
+    "lw             $t0, 0(%[src_ptr])             \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])             \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])             \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])            \n"  // |15|14|13|12|
+    "lw             $t4, 16(%[src_ptr])            \n"  // |19|18|17|16|
+    "lw             $t5, 20(%[src_ptr])            \n"  // |23|22|21|20|
+    "lw             $t6, 24(%[src_ptr])            \n"  // |27|26|25|24|
+    "lw             $t7, 28(%[src_ptr])            \n"  // |31|30|29|28|
+    // TODO(fbarchard): Use odd pixels instead of even.
+    "precr.qb.ph    $t8, $t1, $t0                  \n"  // |6|4|2|0|
+    "precr.qb.ph    $t0, $t3, $t2                  \n"  // |14|12|10|8|
+    "precr.qb.ph    $t1, $t5, $t4                  \n"  // |22|20|18|16|
+    "precr.qb.ph    $t2, $t7, $t6                  \n"  // |30|28|26|24|
+    "addiu          %[src_ptr], %[src_ptr], 32     \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sw             $t8, 0(%[dst])                 \n"
+    "sw             $t0, 4(%[dst])                 \n"
+    "sw             $t1, 8(%[dst])                 \n"
+    "sw             $t2, 12(%[dst])                \n"
+    "bgtz           $t9, 1b                        \n"
+    " addiu         %[dst], %[dst], 16             \n"
+
+  "2:                                              \n"
+    "andi           $t9, %[dst_width], 0xf         \n"  // residue
+    "beqz           $t9, 3f                        \n"
+    " nop                                          \n"
+
+  "21:                                             \n"
+    "lbu            $t0, 0(%[src_ptr])             \n"
+    "addiu          %[src_ptr], %[src_ptr], 2      \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sb             $t0, 0(%[dst])                 \n"
+    "bgtz           $t9, 21b                       \n"
+    " addiu         %[dst], %[dst], 1              \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  const uint8* t = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "srl            $t9, %[dst_width], 3          \n"  // iterations -> step 8
+    "bltz           $t9, 2f                       \n"
+    " nop                                         \n"
+
+    ".p2align       2                             \n"
+  "1:                                             \n"
+    "lw             $t0, 0(%[src_ptr])            \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])            \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])            \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])           \n"  // |15|14|13|12|
+    "lw             $t4, 0(%[t])                  \n"  // |19|18|17|16|
+    "lw             $t5, 4(%[t])                  \n"  // |23|22|21|20|
+    "lw             $t6, 8(%[t])                  \n"  // |27|26|25|24|
+    "lw             $t7, 12(%[t])                 \n"  // |31|30|29|28|
+    "addiu          $t9, $t9, -1                  \n"
+    "srl            $t8, $t0, 16                  \n"  // |X|X|3|2|
+    "ins            $t0, $t4, 16, 16              \n"  // |17|16|1|0|
+    "ins            $t4, $t8, 0, 16               \n"  // |19|18|3|2|
+    "raddu.w.qb     $t0, $t0                      \n"  // |17+16+1+0|
+    "raddu.w.qb     $t4, $t4                      \n"  // |19+18+3+2|
+    "shra_r.w       $t0, $t0, 2                   \n"  // |t0+2|>>2
+    "shra_r.w       $t4, $t4, 2                   \n"  // |t4+2|>>2
+    "srl            $t8, $t1, 16                  \n"  // |X|X|7|6|
+    "ins            $t1, $t5, 16, 16              \n"  // |21|20|5|4|
+    "ins            $t5, $t8, 0, 16               \n"  // |22|23|7|6|
+    "raddu.w.qb     $t1, $t1                      \n"  // |21+20+5+4|
+    "raddu.w.qb     $t5, $t5                      \n"  // |23+22+7+6|
+    "shra_r.w       $t1, $t1, 2                   \n"  // |t1+2|>>2
+    "shra_r.w       $t5, $t5, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t2, 16                  \n"  // |X|X|11|10|
+    "ins            $t2, $t6, 16, 16              \n"  // |25|24|9|8|
+    "ins            $t6, $t8, 0, 16               \n"  // |27|26|11|10|
+    "raddu.w.qb     $t2, $t2                      \n"  // |25+24+9+8|
+    "raddu.w.qb     $t6, $t6                      \n"  // |27+26+11+10|
+    "shra_r.w       $t2, $t2, 2                   \n"  // |t2+2|>>2
+    "shra_r.w       $t6, $t6, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t3, 16                  \n"  // |X|X|15|14|
+    "ins            $t3, $t7, 16, 16              \n"  // |29|28|13|12|
+    "ins            $t7, $t8, 0, 16               \n"  // |31|30|15|14|
+    "raddu.w.qb     $t3, $t3                      \n"  // |29+28+13+12|
+    "raddu.w.qb     $t7, $t7                      \n"  // |31+30+15+14|
+    "shra_r.w       $t3, $t3, 2                   \n"  // |t3+2|>>2
+    "shra_r.w       $t7, $t7, 2                   \n"  // |t7+2|>>2
+    "addiu          %[src_ptr], %[src_ptr], 16    \n"
+    "addiu          %[t], %[t], 16                \n"
+    "sb             $t0, 0(%[dst])                \n"
+    "sb             $t4, 1(%[dst])                \n"
+    "sb             $t1, 2(%[dst])                \n"
+    "sb             $t5, 3(%[dst])                \n"
+    "sb             $t2, 4(%[dst])                \n"
+    "sb             $t6, 5(%[dst])                \n"
+    "sb             $t3, 6(%[dst])                \n"
+    "sb             $t7, 7(%[dst])                \n"
+    "bgtz           $t9, 1b                       \n"
+    " addiu         %[dst], %[dst], 8             \n"
+
+  "2:                                             \n"
+    "andi           $t9, %[dst_width], 0x7        \n"  // x = residue
+    "beqz           $t9, 3f                       \n"
+    " nop                                         \n"
+
+    "21:                                          \n"
+    "lwr            $t1, 0(%[src_ptr])            \n"
+    "lwl            $t1, 3(%[src_ptr])            \n"
+    "lwr            $t2, 0(%[t])                  \n"
+    "lwl            $t2, 3(%[t])                  \n"
+    "srl            $t8, $t1, 16                  \n"
+    "ins            $t1, $t2, 16, 16              \n"
+    "ins            $t2, $t8, 0, 16               \n"
+    "raddu.w.qb     $t1, $t1                      \n"
+    "raddu.w.qb     $t2, $t2                      \n"
+    "shra_r.w       $t1, $t1, 2                   \n"
+    "shra_r.w       $t2, $t2, 2                   \n"
+    "sb             $t1, 0(%[dst])                \n"
+    "sb             $t2, 1(%[dst])                \n"
+    "addiu          %[src_ptr], %[src_ptr], 4     \n"
+    "addiu          $t9, $t9, -2                  \n"
+    "addiu          %[t], %[t], 4                 \n"
+    "bgtz           $t9, 21b                      \n"
+    " addiu         %[dst], %[dst], 2             \n"
+
+  "3:                                             \n"
+    ".set pop                                     \n"
+
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst), [t] "+r" (t)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                    \n"
+      ".set noreorder                               \n"
+
+      "srl            $t9, %[dst_width], 3          \n"
+      "beqz           $t9, 2f                       \n"
+      " nop                                         \n"
+
+      ".p2align       2                             \n"
+     "1:                                            \n"
+      "lw             $t1, 0(%[src_ptr])            \n"  // |3|2|1|0|
+      "lw             $t2, 4(%[src_ptr])            \n"  // |7|6|5|4|
+      "lw             $t3, 8(%[src_ptr])            \n"  // |11|10|9|8|
+      "lw             $t4, 12(%[src_ptr])           \n"  // |15|14|13|12|
+      "lw             $t5, 16(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw             $t6, 20(%[src_ptr])           \n"  // |23|22|21|20|
+      "lw             $t7, 24(%[src_ptr])           \n"  // |27|26|25|24|
+      "lw             $t8, 28(%[src_ptr])           \n"  // |31|30|29|28|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |6|4|2|0|
+      "precr.qb.ph    $t2, $t4, $t3                 \n"  // |14|12|10|8|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |22|20|18|16|
+      "precr.qb.ph    $t6, $t8, $t7                 \n"  // |30|28|26|24|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |12|8|4|0|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |28|24|20|16|
+      "addiu          %[src_ptr], %[src_ptr], 32    \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sw             $t1, 0(%[dst])                \n"
+      "sw             $t5, 4(%[dst])                \n"
+      "bgtz           $t9, 1b                       \n"
+      " addiu         %[dst], %[dst], 8             \n"
+
+    "2:                                             \n"
+      "andi           $t9, %[dst_width], 7          \n"  // residue
+      "beqz           $t9, 3f                       \n"
+      " nop                                         \n"
+
+    "21:                                            \n"
+      "lbu            $t1, 0(%[src_ptr])            \n"
+      "addiu          %[src_ptr], %[src_ptr], 4     \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sb             $t1, 0(%[dst])                \n"
+      "bgtz           $t9, 21b                      \n"
+      " addiu         %[dst], %[dst], 1             \n"
+
+    "3:                                             \n"
+      ".set pop                                     \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst)
+      : [dst_width] "r" (dst_width)
+      : "t1", "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  const uint8* s2 = s1 + stride;
+  const uint8* s3 = s2 + stride;
+
+  __asm__ __volatile__ (
+      ".set push                                  \n"
+      ".set noreorder                             \n"
+
+      "srl           $t9, %[dst_width], 1         \n"
+      "andi          $t8, %[dst_width], 1         \n"
+
+      ".p2align      2                            \n"
+     "1:                                          \n"
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "lw            $t4, 4(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw            $t5, 4(%[s1])                \n"  // |23|22|21|20|
+      "lw            $t6, 4(%[s2])                \n"  // |27|26|25|24|
+      "lw            $t7, 4(%[s3])                \n"  // |31|30|29|28|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "raddu.w.qb    $t4, $t4                     \n"  // |19 + 18 + 17 + 16|
+      "raddu.w.qb    $t5, $t5                     \n"  // |23 + 22 + 21 + 20|
+      "raddu.w.qb    $t6, $t6                     \n"  // |27 + 26 + 25 + 24|
+      "raddu.w.qb    $t7, $t7                     \n"  // |31 + 30 + 29 + 28|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "add           $t4, $t4, $t5                \n"
+      "add           $t6, $t6, $t7                \n"
+      "add           $t4, $t4, $t6                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "shra_r.w      $t4, $t4, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+      "sb            $t4, 1(%[dst])               \n"
+      "addiu         %[src_ptr], %[src_ptr], 8    \n"
+      "addiu         %[s1], %[s1], 8              \n"
+      "addiu         %[s2], %[s2], 8              \n"
+      "addiu         %[s3], %[s3], 8              \n"
+      "addiu         $t9, $t9, -1                 \n"
+      "bgtz          $t9, 1b                      \n"
+      " addiu        %[dst], %[dst], 2            \n"
+      "beqz          $t8, 2f                      \n"
+      " nop                                       \n"
+
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+
+      "2:                                         \n"
+      ".set pop                                   \n"
+
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [s3] "+r" (s3)
+      : [dst_width] "r" (dst_width)
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                          \n"
+      ".set noreorder                                     \n"
+      ".p2align        2                                  \n"
+    "1:                                                   \n"
+      "lw              $t1, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw              $t2, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw              $t3, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw              $t4, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw              $t5, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw              $t6, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw              $t7, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw              $t8, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "precrq.qb.ph    $t0, $t2, $t4                      \n"  // |7|5|15|13|
+      "precrq.qb.ph    $t9, $t6, $t8                      \n"  // |23|21|31|30|
+      "addiu           %[dst_width], %[dst_width], -24    \n"
+      "ins             $t1, $t1, 8, 16                    \n"  // |3|1|0|X|
+      "ins             $t4, $t0, 8, 16                    \n"  // |X|15|13|12|
+      "ins             $t5, $t5, 8, 16                    \n"  // |19|17|16|X|
+      "ins             $t8, $t9, 8, 16                    \n"  // |X|31|29|28|
+      "addiu           %[src_ptr], %[src_ptr], 32         \n"
+      "packrl.ph       $t0, $t3, $t0                      \n"  // |9|8|7|5|
+      "packrl.ph       $t9, $t7, $t9                      \n"  // |25|24|23|21|
+      "prepend         $t1, $t2, 8                        \n"  // |4|3|1|0|
+      "prepend         $t3, $t4, 24                       \n"  // |15|13|12|11|
+      "prepend         $t5, $t6, 8                        \n"  // |20|19|17|16|
+      "prepend         $t7, $t8, 24                       \n"  // |31|29|28|27|
+      "sw              $t1, 0(%[dst])                     \n"
+      "sw              $t0, 4(%[dst])                     \n"
+      "sw              $t3, 8(%[dst])                     \n"
+      "sw              $t5, 12(%[dst])                    \n"
+      "sw              $t9, 16(%[dst])                    \n"
+      "sw              $t7, 20(%[dst])                    \n"
+      "bnez            %[dst_width], 1b                   \n"
+      " addiu          %[dst], %[dst], 24                 \n"
+      ".set pop                                           \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "repl.ph           $t3, 3                          \n"  // 0x00030003
+
+     ".p2align           2                               \n"
+    "1:                                                  \n"
+      "lw                $t0, 0(%[src_ptr])              \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])  \n"  // |T3|T2|T1|T0|
+      "rotr              $t2, $t0, 8                     \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                     \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t4, $t2, $t3                   \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t3                   \n"  // |T0*3|T3*3|
+      "andi              $t0, $t2, 0xFFFF                \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                        \n"
+      "raddu.w.qb        $t1, $t1                        \n"
+      "shra_r.w          $t0, $t0, 1                     \n"
+      "shra_r.w          $t1, $t1, 1                     \n"
+      "preceu.ph.qbr     $t2, $t2                        \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                        \n"  // |0|T2|0|T1|
+      "rotr              $t2, $t2, 16                    \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                    \n"  // |0|T1|0|T2|
+      "addu.ph           $t2, $t2, $t4                   \n"
+      "addu.ph           $t6, $t6, $t5                   \n"
+      "sll               $t5, $t0, 1                     \n"
+      "add               $t0, $t5, $t0                   \n"
+      "shra_r.ph         $t2, $t2, 2                     \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "shll.ph           $t4, $t2, 1                     \n"
+      "addq.ph           $t4, $t4, $t2                   \n"
+      "addu              $t0, $t0, $t1                   \n"
+      "addiu             %[src_ptr], %[src_ptr], 4       \n"
+      "shra_r.w          $t0, $t0, 2                     \n"
+      "addu.ph           $t6, $t6, $t4                   \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "srl               $t1, $t6, 16                    \n"
+      "addiu             %[dst_width], %[dst_width], -3  \n"
+      "sb                $t1, 0(%[d])                    \n"
+      "sb                $t0, 1(%[d])                    \n"
+      "sb                $t6, 2(%[d])                    \n"
+      "bgtz              %[dst_width], 1b                \n"
+      " addiu            %[d], %[d], 3                   \n"
+    "3:                                                  \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                           \n"
+      ".set noreorder                                      \n"
+      "repl.ph           $t2, 3                            \n"  // 0x00030003
+
+      ".p2align          2                                 \n"
+    "1:                                                    \n"
+      "lw                $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])    \n"  // |T3|T2|T1|T0|
+      "rotr              $t4, $t0, 8                       \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                       \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t3, $t4, $t2                     \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t2                     \n"  // |T0*3|T3*3|
+      "andi              $t0, $t4, 0xFFFF                  \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                  \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                          \n"
+      "raddu.w.qb        $t1, $t1                          \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "shra_r.w          $t1, $t1, 1                       \n"
+      "preceu.ph.qbr     $t4, $t4                          \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                          \n"  // |0|T2|0|T1|
+      "rotr              $t4, $t4, 16                      \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                      \n"  // |0|T1|0|T2|
+      "addu.ph           $t4, $t4, $t3                     \n"
+      "addu.ph           $t6, $t6, $t5                     \n"
+      "shra_r.ph         $t6, $t6, 2                       \n"
+      "shra_r.ph         $t4, $t4, 2                       \n"
+      "addu.ph           $t6, $t6, $t4                     \n"
+      "addiu             %[src_ptr], %[src_ptr], 4         \n"
+      "shra_r.ph         $t6, $t6, 1                       \n"
+      "addu              $t0, $t0, $t1                     \n"
+      "addiu             %[dst_width], %[dst_width], -3    \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "srl               $t1, $t6, 16                      \n"
+      "sb                $t1, 0(%[d])                      \n"
+      "sb                $t0, 1(%[d])                      \n"
+      "sb                $t6, 2(%[d])                      \n"
+      "bgtz              %[dst_width], 1b                  \n"
+      " addiu            %[d], %[d], 3                     \n"
+    "3:                                                    \n"
+      ".set pop                                            \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                     \n"
+      ".set noreorder                                \n"
+
+      ".p2align   2                                  \n"
+    "1:                                              \n"
+      "lw         $t0, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw         $t1, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw         $t2, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw         $t3, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw         $t4, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw         $t5, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw         $t6, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw         $t7, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "wsbh       $t0, $t0                           \n"  // |2|3|0|1|
+      "wsbh       $t6, $t6                           \n"  // |26|27|24|25|
+      "srl        $t0, $t0, 8                        \n"  // |X|2|3|0|
+      "srl        $t3, $t3, 16                       \n"  // |X|X|15|14|
+      "srl        $t5, $t5, 16                       \n"  // |X|X|23|22|
+      "srl        $t7, $t7, 16                       \n"  // |X|X|31|30|
+      "ins        $t1, $t2, 24, 8                    \n"  // |8|6|5|4|
+      "ins        $t6, $t5, 0, 8                     \n"  // |26|27|24|22|
+      "ins        $t1, $t0, 0, 16                    \n"  // |8|6|3|0|
+      "ins        $t6, $t7, 24, 8                    \n"  // |30|27|24|22|
+      "prepend    $t2, $t3, 24                       \n"  // |X|15|14|11|
+      "ins        $t4, $t4, 16, 8                    \n"  // |19|16|17|X|
+      "ins        $t4, $t2, 0, 16                    \n"  // |19|16|14|11|
+      "addiu      %[src_ptr], %[src_ptr], 32         \n"
+      "addiu      %[dst_width], %[dst_width], -12    \n"
+      "addiu      $t8,%[dst_width], -12              \n"
+      "sw         $t1, 0(%[dst])                     \n"
+      "sw         $t4, 4(%[dst])                     \n"
+      "sw         $t6, 8(%[dst])                     \n"
+      "bgez       $t8, 1b                            \n"
+      " addiu     %[dst], %[dst], 12                 \n"
+      ".set pop                                      \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* t = src_ptr + stride;
+  const int c = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[t])                      \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[t])                      \n"  // |T7|T6|T5|T4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t4, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "packrl.ph       $t5, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t4, $t4                          \n"  // S7+S6+T7+T6
+      "raddu.w.qb      $t5, $t5                          \n"  // T5+T4+S5+S4
+      "precrq.qb.ph    $t6, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t6, $t6, $t6                     \n"  // |S3|T3|S3|T3|
+      "srl             $t4, $t4, 2                       \n"  // t4 / 4
+      "srl             $t6, $t6, 16                      \n"  // |0|0|S3|T3|
+      "raddu.w.qb      $t6, $t6                          \n"  // 0+0+S3+T3
+      "addu            $t6, $t5, $t6                     \n"
+      "mul             $t6, $t6, %[c]                    \n"  // t6 * 0x2AAA
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "raddu.w.qb      $t0, $t0                          \n"  // S2+S1+S0+0
+      "raddu.w.qb      $t2, $t2                          \n"  // T2+T1+T0+0
+      "addu            $t0, $t0, $t2                     \n"
+      "mul             $t0, $t0, %[c]                    \n"  // t0 * 0x2AAA
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[t], %[t], 8                     \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t4, -1(%[dst_ptr])               \n"
+      "sb              $t6, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [t] "+r" (t),
+        [dst_width] "+r" (dst_width)
+      : [c] "r" (c)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  stride += stride;
+  const uint8* s2 = src_ptr + stride;
+  const int c1 = 0x1C71;
+  const int c2 = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[s1])                     \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[s1])                     \n"  // |T7|T6|T5|T4|
+      "lw              $t4, 0(%[s2])                     \n"  // |R3|R2|R1|R0|
+      "lw              $t5, 4(%[s2])                     \n"  // |R7|R6|R5|R4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t6, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "raddu.w.qb      $t6, $t6                          \n"  // S7+S6+T7+T6
+      "packrl.ph       $t7, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t7, $t7                          \n"  // T5+T4+S5+S4
+      "sll             $t8, $t5, 16                      \n"  // |R5|R4|0|0|
+      "raddu.w.qb      $t8, $t8                          \n"  // R5+R4
+      "addu            $t7, $t7, $t8                     \n"
+      "srl             $t8, $t5, 16                      \n"  // |0|0|R7|R6|
+      "raddu.w.qb      $t8, $t8                          \n"  // R7 + R6
+      "addu            $t6, $t6, $t8                     \n"
+      "mul             $t6, $t6, %[c2]                   \n"  // t6 * 0x2AAA
+      "precrq.qb.ph    $t8, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t8, $t8, $t4                     \n"  // |S3|T3|R3|R1|
+      "srl             $t8, $t8, 8                       \n"  // |0|S3|T3|R3|
+      "raddu.w.qb      $t8, $t8                          \n"  // S3 + T3 + R3
+      "addu            $t7, $t7, $t8                     \n"
+      "mul             $t7, $t7, %[c1]                   \n"  // t7 * 0x1C71
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "sll             $t4, $t4, 8                       \n"  // |R2|R1|R0|0|
+      "raddu.w.qb      $t0, $t0                          \n"
+      "raddu.w.qb      $t2, $t2                          \n"
+      "raddu.w.qb      $t4, $t4                          \n"
+      "addu            $t0, $t0, $t2                     \n"
+      "addu            $t0, $t0, $t4                     \n"
+      "mul             $t0, $t0, %[c1]                   \n"  // t0 * 0x1C71
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[s1], %[s1], 8                   \n"
+      "addiu           %[s2], %[s2], 8                   \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t7, $t7, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t6, -1(%[dst_ptr])               \n"
+      "sb              $t7, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [dst_width] "+r" (dst_width)
+      : [c1] "r" (c1), [c2] "r" (c2)
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/scale_neon.cc b/drivers/theoraplayer/src/YUV/libyuv/src/scale_neon.cc
new file mode 100755
index 0000000000..a9df93c055
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/scale_neon.cc
@@ -0,0 +1,699 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
+
+// NEON downscalers with interpolation.
+// Provided by Fritz Koenig
+
+// Read 32x1 throw away even pixels, and write 16x1.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  asm volatile (
+#ifdef _ANDROID
+	".fpu neon\n"
+#endif
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    "vld2.8     {q0, q1}, [%0]!                \n"
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1"              // Clobber List
+  );
+}
+
+// Read 32x2 average down and write 16x1.
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+#ifdef _ANDROID
+	".fpu neon\n"
+#endif
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
+    "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
+    "vpaddl.u8  q1, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent + row1
+    "vpadal.u8  q1, q3                         \n"
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "q0", "q1", "q2", "q3"     // Clobber List
+  );
+}
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+#ifdef _ANDROID
+	".fpu neon\n"
+#endif
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n" // src line 0
+    "subs       %2, %2, #8                     \n" // 8 processed per loop
+    "vst1.8     {d2}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1", "memory", "cc"
+  );
+}
+
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  asm volatile (
+#ifdef _ANDROID
+	".fpu neon\n"
+#endif
+    "add        r4, %0, %3                     \n"
+    "add        r5, r4, %3                     \n"
+    "add        %3, r5, %3                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {q0}, [%0]!                    \n"   // load up 16x4
+    "vld1.8     {q1}, [r4]!                    \n"
+    "vld1.8     {q2}, [r5]!                    \n"
+    "vld1.8     {q3}, [%3]!                    \n"
+    "subs       %2, %2, #4                     \n"
+    "vpaddl.u8  q0, q0                         \n"
+    "vpadal.u8  q0, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"
+    "vpadal.u8  q0, q3                         \n"
+    "vpaddl.u16 q0, q0                         \n"
+    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
+    "vmovn.u16  d0, q0                         \n"
+    "vst1.32    {d0[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(src_stride)         // %3
+  : "r4", "r5", "q0", "q1", "q2", "q3", "memory", "cc"
+  );
+}
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+// to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+#ifdef _ANDROID
+				".fpu neon\n"
+#endif
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "subs       %2, %2, #24                  \n"
+    "vmov       d2, d3                       \n" // order d0, d1, d2
+    "vst3.8     {d0, d1, d2}, [%1]!          \n"
+    "bgt        1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "d0", "d1", "d2", "d3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+
+    // filter src line 0 with src line 1
+    // expand chars to shorts to allow for room
+    // when adding lines together
+    "vmovl.u8     q8, d4                       \n"
+    "vmovl.u8     q9, d5                       \n"
+    "vmovl.u8     q10, d6                      \n"
+    "vmovl.u8     q11, d7                      \n"
+
+    // 3 * line_0 + line_1
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vmlal.u8     q9, d1, d24                  \n"
+    "vmlal.u8     q10, d2, d24                 \n"
+    "vmlal.u8     q11, d3, d24                 \n"
+
+    // (3 * line_0 + line_1) >> 2
+    "vqrshrn.u16  d0, q8, #2                   \n"
+    "vqrshrn.u16  d1, q9, #2                   \n"
+    "vqrshrn.u16  d2, q10, #2                  \n"
+    "vqrshrn.u16  d3, q11, #2                  \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q8, d1                       \n"
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vqrshrn.u16  d0, q8, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q8, d2                       \n"
+    "vmlal.u8     q8, d3, d24                  \n"
+    "vqrshrn.u16  d2, q8, #2                   \n"
+
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+    // average src line 0 with src line 1
+    "vrhadd.u8    q0, q0, q2                   \n"
+    "vrhadd.u8    q1, q1, q3                   \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q3, d1                       \n"
+    "vmlal.u8     q3, d0, d24                  \n"
+    "vqrshrn.u16  d0, q3, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q3, d2                       \n"
+    "vmlal.u8     q3, d3, d24                  \n"
+    "vqrshrn.u16  d2, q3, #2                   \n"
+
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
+  );
+}
+
+#define HAS_SCALEROWDOWN38_NEON
+static uvec8 kShuf38 =
+  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
+static uvec8 kShuf38_2 =
+  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
+static vec16 kMult38_Div6 =
+  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
+    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
+static vec16 kMult38_Div9 =
+  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
+    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vld1.8     {q3}, [%3]                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
+    "subs       %2, %2, #12                    \n"
+    "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
+    "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
+    "vst1.8     {d4}, [%1]!                    \n"
+    "vst1.32    {d5[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(&kShuf38)           // %3
+  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
+  );
+}
+
+// 32x3 -> 12x1
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vld1.16    {q13}, [%4]                    \n"
+    "vld1.8     {q14}, [%5]                    \n"
+    "vld1.8     {q15}, [%6]                    \n"
+    "add        r4, %0, %3, lsl #1             \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    "vld4.8       {d16, d17, d18, d19}, [r4]!  \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+    "vtrn.u8      d16, d17                     \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+    "vtrn.u8      d18, d19                     \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+    "vpaddl.u8    q8, q8                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+    "vpaddl.u8    d19, d19                     \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     q0, q8                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+    "vadd.u16     d4, d19                      \n"
+
+    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+    //             + s[6 + st * 1] + s[7 + st * 1]
+    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+    "vqrdmulh.s16 q2, q2, q13                  \n"
+    "vmovn.u16    d4, q2                       \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+    "vmovl.u8     q9, d18                      \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+    "vadd.u16     q1, q9                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q15                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    "vst1.8       {d3}, [%1]!                  \n"
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  : "r"(&kMult38_Div6),     // %4
+    "r"(&kShuf38_2),        // %5
+    "r"(&kMult38_Div9)      // %6
+  : "r4", "q0", "q1", "q2", "q3", "q8", "q9",
+    "q13", "q14", "q15", "memory", "cc"
+  );
+}
+
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vld1.16    {q13}, [%4]                    \n"
+    "vld1.8     {q14}, [%5]                    \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+
+    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+    "vqrshrn.u16  d4, q2, #2                   \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q13                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    "vst1.8       {d3}, [%1]!                  \n"
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),       // %0
+    "+r"(dst_ptr),       // %1
+    "+r"(dst_width),     // %2
+    "+r"(src_stride)     // %3
+  : "r"(&kMult38_Div6),  // %4
+    "r"(&kShuf38_2)      // %5
+  : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
+  );
+}
+
+// 16x2 -> 16x1
+void ScaleFilterRows_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr, ptrdiff_t src_stride,
+                          int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp          %4, #0                       \n"
+    "beq          100f                         \n"
+    "add          %2, %1                       \n"
+    "cmp          %4, #64                      \n"
+    "beq          75f                          \n"
+    "cmp          %4, #128                     \n"
+    "beq          50f                          \n"
+    "cmp          %4, #192                     \n"
+    "beq          25f                          \n"
+
+    "vdup.8       d5, %4                       \n"
+    "rsb          %4, #256                     \n"
+    "vdup.8       d4, %4                       \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    "vld1.8       {q0}, [%1]!                  \n"
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vmull.u8     q13, d0, d4                  \n"
+    "vmull.u8     q14, d1, d4                  \n"
+    "vmlal.u8     q13, d2, d5                  \n"
+    "vmlal.u8     q14, d3, d5                  \n"
+    "vrshrn.u16   d0, q13, #8                  \n"
+    "vrshrn.u16   d1, q14, #8                  \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          1b                           \n"
+    "b            99f                          \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    "vld1.8       {q0}, [%1]!                  \n"
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          25b                          \n"
+    "b            99f                          \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    "vld1.8       {q0}, [%1]!                  \n"
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          50b                          \n"
+    "b            99f                          \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    "vld1.8       {q1}, [%1]!                  \n"
+    "vld1.8       {q0}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          75b                          \n"
+    "b            99f                          \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    "vld1.8       {q0}, [%1]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          100b                         \n"
+
+  "99:                                         \n"
+    "vst1.8       {d1[7]}, [%0]                \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
+  );
+}
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    "vld2.32    {q0, q1}, [%0]!                \n"
+    "vld2.32    {q2, q3}, [%0]!                \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    "vst1.8     {q3}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
+    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
+    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    "vrshrn.u16 d2, q2, #2                     \n"
+    "vrshrn.u16 d3, q3, #2                     \n"
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %3, lsl #2                \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.32    {d0[0]}, [%0], r12             \n"
+    "vld1.32    {d0[1]}, [%0], r12             \n"
+    "vld1.32    {d1[0]}, [%0], r12             \n"
+    "vld1.32    {d1[1]}, [%0], r12             \n"
+    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(dst_width)    // %2
+  : "r"(src_stepx)     // %3
+  : "memory", "cc", "r12", "q0"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %4, lsl #2                \n"
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "vld1.8     {d0}, [%0], r12                \n"  // Read 4 2x2 blocks -> 2x1
+    "vld1.8     {d1}, [%1], r12                \n"
+    "vld1.8     {d2}, [%0], r12                \n"
+    "vld1.8     {d3}, [%1], r12                \n"
+    "vld1.8     {d4}, [%0], r12                \n"
+    "vld1.8     {d5}, [%1], r12                \n"
+    "vld1.8     {d6}, [%0], r12                \n"
+    "vld1.8     {d7}, [%1], r12                \n"
+    "vaddl.u8   q0, d0, d1                     \n"
+    "vaddl.u8   q1, d2, d3                     \n"
+    "vaddl.u8   q2, d4, d5                     \n"
+    "vaddl.u8   q3, d6, d7                     \n"
+    "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
+    "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
+    "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
+    "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
+    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
+    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
+    "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(src_stride),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(dst_width)    // %3
+  : "r"(src_stepx)     // %4
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
+  );
+}
+
+#endif  // __ARM_NEON__
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/scale_posix.cc b/drivers/theoraplayer/src/YUV/libyuv/src/scale_posix.cc
new file mode 100644
index 0000000000..352e667822
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/scale_posix.cc
@@ -0,0 +1,1315 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// GCC versions of row functions are verbatim conversions from Visual C.
+// Generated using gcc disassembly on Visual C object file:
+// objdump -D yuvscaler.obj >yuvscaler.txt
+
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm2)           //  movdqa  (%0,%3,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,3,1,xmm3)           //  movdqa  0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x18,%%xmm5                    \n"
+    "pslld     $0x10,%%xmm5                    \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  intptr_t stridex3 = 0;
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0x8,%%xmm7                     \n"
+    "lea       " MEMLEA4(0x00,4,4,2) ",%3      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqa,0x00,0,4,1,xmm2)           //  movdqa  (%0,%4,1),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,4,1,xmm3)           //  movdqa  0x10(%0,%4,1),%%xmm3
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    MEMOPREG(movdqa,0x00,0,4,2,xmm2)           //  movdqa  (%0,%4,2),%%xmm2
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,4,2,xmm3)           //  movdqa  0x10(%0,%4,2),%%xmm3
+    MEMOPREG(movdqa,0x00,0,3,1,xmm4)           //  movdqa  (%0,%3,1),%%xmm4
+    MEMOPREG(movdqa,0x10,0,3,1,xmm5)           //  movdqa  0x10(%0,%3,1),%%xmm5
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm4,%%xmm2                   \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm5,%%xmm3                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pand      %%xmm7,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width),   // %2
+    "+r"(stridex3)     // %3
+  : "r"((intptr_t)(src_stride))    // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm3                       \n"
+    "movdqa    %1,%%xmm4                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kShuf0),  // %0
+    "m"(kShuf1),  // %1
+    "m"(kShuf2)   // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "palignr   $0x8,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm3,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movq      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "movq      %%xmm2," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm7)           //  movdqa  (%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x10,0,3,1,xmm7)           //  movdqa  0x10(%0,%3),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "m"(kMadd21)     // %4
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm7)           //  movdqa  (%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqa,0x10,0,3,1,xmm7)           //  movdqa  0x10(%0,%3,1),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+    : "+r"(src_ptr),   // %0
+      "+r"(dst_ptr),   // %1
+      "+r"(dst_width)  // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "m"(kMadd21)     // %4
+    : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movd      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "lea       " MEMLEA(0xc,1) ",%1            \n"
+    "sub       $0xc,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "m"(kShuf38a),   // %3
+    "m"(kShuf38b)    // %4
+  : "memory", "cc"
+#if defined(__SSE2__)
+      , "xmm0", "xmm1", "xmm4", "xmm5"
+#endif
+  );
+}
+
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "movdqa    %3,%%xmm5                       \n"
+  :
+  : "m"(kShufAb0),   // %0
+    "m"(kShufAb1),   // %1
+    "m"(kShufAb2),   // %2
+    "m"(kScaleAb2)   // %3
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(pavgb,0x00,0,3,1,xmm0)            //  pavgb   (%0,%3,1),%%xmm0
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pshufb    %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "paddusw   %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "paddusw   %%xmm0,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "sub       $0x6,%2                         \n"
+    "movd      %%xmm1," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm1                    \n"
+    "movd      %%xmm1," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width)    // %2
+  : "r"((intptr_t)(src_stride))  // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+  :
+  : "m"(kShufAc),    // %0
+    "m"(kShufAc3),   // %1
+    "m"(kScaleAc33)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqa,0x00,0,3,1,xmm6)           //  movdqa  (%0,%3,1),%%xmm6
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqa,0x00,0,3,2,xmm6)           //  movdqa  (%0,%3,2),%%xmm6
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "pshufb    %%xmm3,%%xmm7                   \n"
+    "paddusw   %%xmm7,%%xmm6                   \n"
+    "pmulhuw   %%xmm4,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "sub       $0x6,%2                         \n"
+    "movd      %%xmm6," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm6                    \n"
+    "movd      %%xmm6," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+#endif
+  );
+}
+
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width, int src_height) {
+  int tmp_height = 0;
+  intptr_t tmp_src = 0;
+  asm volatile (
+    "pxor      %%xmm4,%%xmm4                   \n"
+    "sub       $0x1,%5                         \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "mov       %0,%3                           \n"
+    "add       %6,%0                           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "mov       %5,%2                           \n"
+    "test      %2,%2                           \n"
+    "je        3f                              \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm2         \n"
+    "add       %6,%0                           \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "paddusw   %%xmm2,%%xmm0                   \n"
+    "paddusw   %%xmm3,%%xmm1                   \n"
+    "sub       $0x1,%2                         \n"
+    "jg        2b                              \n"
+
+    LABELALIGN
+  "3:                                          \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x10,3) ",%0           \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x10,%4                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(tmp_height),  // %2
+    "+r"(tmp_src),     // %3
+    "+r"(src_width),   // %4
+    "+rm"(src_height)  // %5
+  : "rm"((intptr_t)(src_stride))  // %6
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+
+// Bilinear column filtering. SSSE3 version.
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0, temp_pixel = 0;
+  asm volatile (
+    "movd      %6,%%xmm2                       \n"
+    "movd      %7,%%xmm3                       \n"
+    "movl      $0x04040000,%k2                 \n"
+    "movd      %k2,%%xmm5                      \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "subl      $0x2,%5                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm1                     \n"
+    BUNDLEALIGN
+    MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
+    "movd      %k2,%%xmm4                      \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "punpcklwd %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %w2," MEMACCESS(0) "            \n"
+    "lea       " MEMLEA(0x2,0) ",%0            \n"
+    "sub       $0x2,%5                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "addl      $0x1,%5                         \n"
+    "jl        99f                             \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %b2," MEMACCESS(0) "            \n"
+  "99:                                         \n"
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+a"(temp_pixel),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1),          // %4
+    "+rm"(dst_width)   // %5
+  : "rm"(x),           // %6
+    "rm"(dx)           // %7
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "sub       $0x20,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+r"(dst_width)    // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "shufps    $0xdd,%%xmm1,%%xmm0             \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    BUNDLEALIGN
+    MEMOPREG(movdqa,0x00,0,3,1,xmm2)           //  movdqa   (%0,%3,1),%%xmm2
+    MEMOPREG(movdqa,0x10,0,3,1,xmm3)           //  movdqa   0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movd      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,0,1,2,xmm2)             //  movd      (%0,%1,2),%%xmm2
+    MEMOPREG(movd,0x00,0,4,1,xmm3)             //  movd      (%0,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "punpckldq %%xmm3,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm0                  \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(src_stepx_x4),  // %1
+    "+r"(dst_argb),      // %2
+    "+r"(dst_width),     // %3
+    "+r"(src_stepx_x12)  // %4
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+// Blends four 2x2 to 4x1.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride, int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  intptr_t row1 = (intptr_t)(src_stride);
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    "lea       " MEMLEA4(0x00,0,5,1) ",%5      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movhps,0x00,0,1,1,xmm0)           //  movhps    (%0,%1,1),%%xmm0
+    MEMOPREG(movq,0x00,0,1,2,xmm1)             //  movq      (%0,%1,2),%%xmm1
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,0,4,1,xmm1)           //  movhps    (%0,%4,1),%%xmm1
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "movq      " MEMACCESS(5) ",%%xmm2         \n"
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,5,1,1,xmm2)           //  movhps    (%5,%1,1),%%xmm2
+    MEMOPREG(movq,0x00,5,1,2,xmm3)             //  movq      (%5,%1,2),%%xmm3
+    MEMOPREG(movhps,0x00,5,4,1,xmm3)           //  movhps    (%5,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,5,1,4) ",%5      \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),       // %0
+    "+r"(src_stepx_x4),   // %1
+    "+r"(dst_argb),       // %2
+    "+rm"(dst_width),     // %3
+    "+r"(src_stepx_x12),  // %4
+    "+r"(row1)            // %5
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3"
+#endif
+  );
+}
+
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    "pshufd    $0x11,%%xmm3,%%xmm0             \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x5,%%xmm3,%%xmm0              \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "cmp       $0x0,%4                         \n"
+    "jl        99f                             \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    LABELALIGN
+  "40:                                         \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "pextrw    $0x7,%%xmm2,%k1                 \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm1)             //  movd      (%3,%0,4),%%xmm1
+    MEMOPREG(movd,0x00,3,1,4,xmm4)             //  movd      (%3,%1,4),%%xmm4
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "punpckldq %%xmm4,%%xmm1                   \n"
+    "punpcklqdq %%xmm1,%%xmm0                  \n"
+    "sub       $0x4,%4                         \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "test      $0x2,%4                         \n"
+    "je        29f                             \n"
+    BUNDLEALIGN
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x8,2) ",%2            \n"
+  "29:                                         \n"
+    "test      $0x1,%4                         \n"
+    "je        99f                             \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+  "99:                                         \n"
+  : "+a"(x0),          // %0
+    "+d"(x1),          // %1
+    "+r"(dst_argb),    // %2
+    "+r"(src_argb),    // %3
+    "+r"(dst_width)    // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+#endif
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqa    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpckldq %%xmm0,%%xmm0                   \n"
+    "punpckhdq %%xmm1,%%xmm1                   \n"
+    "sub       $0x8,%2                         \n"
+    "movdqa    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqa    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+r"(dst_width)    // %2
+  :
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm5                       \n"
+  :
+  : "m"(kShuffleColARGB),  // %0
+    "m"(kShuffleFractions)  // %1
+  );
+
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "sub       $0x2,%2                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "psrlw     $0x9,%%xmm1                     \n"
+    BUNDLEALIGN
+    MEMOPREG(movhps,0x00,1,4,4,xmm0)           //  movhps    (%1,%4,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x2,%2                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "add       $0x1,%2                         \n"
+    "jl        99f                             \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    BUNDLEALIGN
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(0) "         \n"
+
+    LABELALIGN
+  "99:                                         \n"
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+rm"(dst_width),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1)           // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc"
+#if defined(__native_client__) && defined(__x86_64__)
+    , "r14"
+#endif
+#if defined(__SSE2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+#endif
+  );
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "sub       $0x10001,%%eax                  \n"
+    "sbb       $0x0,%%edx                      \n"
+    "sub       $0x1,%1                         \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/scale_win.cc b/drivers/theoraplayer/src/YUV/libyuv/src/scale_win.cc
new file mode 100644
index 0000000000..840b9738da
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/scale_win.cc
@@ -0,0 +1,1320 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// Reads 32 pixels, throws half away and writes 16 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x1 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 32 pixels, throws half away and writes 16 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x1 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 16x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+    align      4
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    sub        ecx, 16
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Point samples 32 pixels to 8 pixels.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff0000
+    psrld      xmm5, 24
+    pslld      xmm5, 16
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm0, 8
+    packuswb   xmm0, xmm0
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x4 rectangle to 8x1.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_ptr
+    mov        esi, [esp + 8 + 8]    // src_stride
+    mov        edx, [esp + 8 + 12]   // dst_ptr
+    mov        ecx, [esp + 8 + 16]   // dst_width
+    lea        edi, [esi + esi * 2]  // src_stride * 3
+    pcmpeqb    xmm7, xmm7            // generate mask 0x00ff00ff
+    psrlw      xmm7, 8
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, [eax + esi * 2]
+    movdqa     xmm3, [eax + esi * 2 + 16]
+    movdqa     xmm4, [eax + edi]
+    movdqa     xmm5, [eax + edi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm2, xmm4
+    pavgb      xmm3, xmm5
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm7
+    pand       xmm3, xmm7
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqa     xmm2, xmm0            // average columns (16 to 8 pixels)
+    psrlw      xmm0, 8
+    pand       xmm2, xmm7
+    pavgw      xmm0, xmm2
+    packuswb   xmm0, xmm0
+
+    sub        ecx, 8
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Point samples 32 pixels to 24 pixels.
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm3, kShuf0
+    movdqa     xmm4, kShuf1
+    movdqa     xmm5, kShuf2
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm1
+    palignr    xmm1, xmm0, 8
+    pshufb     xmm0, xmm3
+    pshufb     xmm1, xmm4
+    pshufb     xmm2, xmm5
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + 8], xmm1
+    movq       qword ptr [edx + 16], xmm2
+    lea        edx, [edx + 24]
+    sub        ecx, 24
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 24x1
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+// Register usage:
+// xmm0 src_row 0
+// xmm1 src_row 1
+// xmm2 shuf 0
+// xmm3 shuf 1
+// xmm4 shuf 2
+// xmm5 madd 0
+// xmm6 madd 1
+// xmm7 kRound34
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]           // pixels 0..7
+    movdqa     xmm1, [eax + esi]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqa     xmm0, [eax + 16]      // pixels 16..23
+    movdqa     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    sub        ecx, 24
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx + 24]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Note that movdqa+palign may be better than movdqu.
+// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 8 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]           // pixels 0..7
+    movdqa     xmm1, [eax + esi]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqa     xmm0, [eax + 16]      // pixels 16..23
+    movdqa     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    sub        ecx, 24
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx+24]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 3/8 point sampler
+
+// Scale 32 pixels to 12
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm4, kShuf38a
+    movdqa     xmm5, kShuf38b
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // 16 pixels -> 0,1,2,3,4,5
+    movdqa     xmm1, [eax + 16]      // 16 pixels -> 6,7,8,9,10,11
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm4
+    pshufb     xmm1, xmm5
+    paddusb    xmm0, xmm1
+
+    sub        ecx, 12
+    movq       qword ptr [edx], xmm0  // write 12 pixels
+    movhlps    xmm1, xmm0
+    movd       [edx + 8], xmm1
+    lea        edx, [edx + 12]
+    jg         xloop
+
+    ret
+  }
+}
+
+// Scale 16x3 pixels to 6x1 with interpolation
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAc
+    movdqa     xmm3, kShufAc3
+    movdqa     xmm4, kScaleAc33
+    pxor       xmm5, xmm5
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // sum up 3 rows into xmm0/1
+    movdqa     xmm6, [eax + esi]
+    movhlps    xmm1, xmm0
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+    movdqa     xmm6, [eax + esi * 2]
+    lea        eax, [eax + 16]
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+
+    movdqa     xmm6, xmm0            // 8 pixels -> 0,1,2 of xmm6
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    pshufb     xmm6, xmm2
+
+    movdqa     xmm7, xmm1            // 8 pixels -> 3,4,5 of xmm6
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    pshufb     xmm7, xmm3
+    paddusw    xmm6, xmm7
+
+    pmulhuw    xmm6, xmm4            // divide by 9,9,6, 9,9,6
+    packuswb   xmm6, xmm6
+
+    sub        ecx, 6
+    movd       [edx], xmm6           // write 6 pixels
+    psrlq      xmm6, 16
+    movd       [edx + 2], xmm6
+    lea        edx, [edx + 6]
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Scale 16x2 pixels to 6x1 with interpolation
+__declspec(naked) __declspec(align(16))
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAb0
+    movdqa     xmm3, kShufAb1
+    movdqa     xmm4, kShufAb2
+    movdqa     xmm5, kScaleAb2
+
+    align      4
+  xloop:
+    movdqa     xmm0, [eax]           // average 2 rows into xmm0
+    pavgb      xmm0, [eax + esi]
+    lea        eax, [eax + 16]
+
+    movdqa     xmm1, xmm0            // 16 pixels -> 0,1,2,3,4,5 of xmm1
+    pshufb     xmm1, xmm2
+    movdqa     xmm6, xmm0
+    pshufb     xmm6, xmm3
+    paddusw    xmm1, xmm6
+    pshufb     xmm0, xmm4
+    paddusw    xmm1, xmm0
+
+    pmulhuw    xmm1, xmm5            // divide by 3,3,2, 3,3,2
+    packuswb   xmm1, xmm1
+
+    sub        ecx, 6
+    movd       [edx], xmm1           // write 6 pixels
+    psrlq      xmm1, 16
+    movd       [edx + 2], xmm1
+    lea        edx, [edx + 6]
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 16xN bytes and produces 16 shorts at a time.
+// TODO(fbarchard): Make this handle 4xN bytes for any width ARGB.
+__declspec(naked) __declspec(align(16))
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width,
+                       int src_height) {
+  __asm {
+    push       esi
+    push       edi
+    push       ebx
+    push       ebp
+    mov        esi, [esp + 16 + 4]   // src_ptr
+    mov        edx, [esp + 16 + 8]   // src_stride
+    mov        edi, [esp + 16 + 12]  // dst_ptr
+    mov        ecx, [esp + 16 + 16]  // dst_width
+    mov        ebx, [esp + 16 + 20]  // height
+    pxor       xmm4, xmm4
+    dec        ebx
+
+    align      4
+  xloop:
+    // first row
+    movdqa     xmm0, [esi]
+    lea        eax, [esi + edx]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    lea        esi, [esi + 16]
+    mov        ebp, ebx
+    test       ebp, ebp
+    je         ydone
+
+    // sum remaining rows
+    align      4
+  yloop:
+    movdqa     xmm2, [eax]       // read 16 pixels
+    lea        eax, [eax + edx]  // advance to next row
+    movdqa     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    paddusw    xmm0, xmm2        // sum 16 words
+    paddusw    xmm1, xmm3
+    sub        ebp, 1
+    jg         yloop
+
+    align      4
+  ydone:
+    movdqa     [edi], xmm0
+    movdqa     [edi + 16], xmm1
+    lea        edi, [edi + 32]
+
+    sub        ecx, 16
+    jg         xloop
+
+    pop        ebp
+    pop        ebx
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Bilinear column filtering. SSSE3 version.
+// TODO(fbarchard): Port to Neon
+// TODO(fbarchard): Switch the following:
+//    xor        ebx, ebx
+//    mov        bx, word ptr [esi + eax]  // 2 source x0 pixels
+// To
+//    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+// when drmemory bug fixed.
+// https://code.google.com/p/drmemory/issues/detail?id=1396
+
+__declspec(naked) __declspec(align(16))
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        edi, [esp + 12 + 4]    // dst_ptr
+    mov        esi, [esp + 12 + 8]    // src_ptr
+    mov        ecx, [esp + 12 + 12]   // dst_width
+    movd       xmm2, [esp + 12 + 16]  // x
+    movd       xmm3, [esp + 12 + 20]  // dx
+    mov        eax, 0x04040000      // shuffle to line up fractions with pixel.
+    movd       xmm5, eax
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+    align      4
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movzx      ebx, word ptr [esi + edx]  // 2 source x1 pixels
+    movd       xmm4, ebx
+    pshufb     xmm1, xmm5           // 0011
+    punpcklwd  xmm0, xmm4
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits, 2 pixels.
+    movd       ebx, xmm0
+    mov        [edi], bx
+    lea        edi, [edi + 2]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+    align      4
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm2, 9              // 7 bit fractions.
+    pshufb     xmm2, xmm5           // 0011
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // 16 bit
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits
+    movd       ebx, xmm0
+    mov        [edi], bl
+
+    align      4
+ xloop99:
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Reads 16 pixels, duplicates them and writes 32 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_ptr
+    mov        eax, [esp + 8]    // src_ptr
+    mov        ecx, [esp + 12]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0
+    punpckhbw  xmm1, xmm1
+    sub        ecx, 32
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Reads 8 pixels, throws half away and writes 4 even pixels (0, 2, 4, 6)
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    shufps     xmm0, xmm1, 0xdd
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x1 rectangle to 4x1.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x2 rectangle to 4x1.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_argb
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_argb
+    mov        ecx, [esp + 4 + 16]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    movdqa     xmm1, [eax + 16]
+    movdqa     xmm2, [eax + esi]
+    movdqa     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_argb
+                                     // src_stride ignored
+    mov        ebx, [esp + 8 + 12]   // src_stepx
+    mov        edx, [esp + 8 + 16]   // dst_argb
+    mov        ecx, [esp + 8 + 20]   // dst_width
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+    align      4
+  wloop:
+    movd       xmm0, [eax]
+    movd       xmm1, [eax + ebx]
+    punpckldq  xmm0, xmm1
+    movd       xmm2, [eax + ebx * 2]
+    movd       xmm3, [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    punpckldq  xmm2, xmm3
+    punpcklqdq xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        edi
+    pop        ebx
+    ret
+  }
+}
+
+// Blends four 2x2 to 4x1.
+// Alignment requirement: dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]    // src_argb
+    mov        esi, [esp + 12 + 8]    // src_stride
+    mov        ebx, [esp + 12 + 12]   // src_stepx
+    mov        edx, [esp + 12 + 16]   // dst_argb
+    mov        ecx, [esp + 12 + 20]   // dst_width
+    lea        esi, [eax + esi]       // row1 pointer
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+    align      4
+  wloop:
+    movq       xmm0, qword ptr [eax]  // row0 4 pairs
+    movhps     xmm0, qword ptr [eax + ebx]
+    movq       xmm1, qword ptr [eax + ebx * 2]
+    movhps     xmm1, qword ptr [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    movq       xmm2, qword ptr [esi]  // row1 4 pairs
+    movhps     xmm2, qword ptr [esi + ebx]
+    movq       xmm3, qword ptr [esi + ebx * 2]
+    movhps     xmm3, qword ptr [esi + edi]
+    lea        esi,  [esi + ebx * 4]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    sub        ecx, 4
+    movdqa     [edx], xmm0
+    lea        edx, [edx + 16]
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Column scaling unfiltered. SSE2 version.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  __asm {
+    push       edi
+    push       esi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+
+    pshufd     xmm2, xmm2, 0         // x0 x0 x0 x0
+    pshufd     xmm0, xmm3, 0x11      // dx  0 dx  0
+    paddd      xmm2, xmm0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 2
+    pshufd     xmm0, xmm3, 0x05      // dx * 2, dx * 2, 0, 0
+    paddd      xmm2, xmm0            // x3 x2 x1 x0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 4
+    pshufd     xmm3, xmm3, 0         // dx * 4, dx * 4, dx * 4, dx * 4
+
+    pextrw     eax, xmm2, 1          // get x0 integer.
+    pextrw     edx, xmm2, 3          // get x1 integer.
+
+    cmp        ecx, 0
+    jle        xloop99
+    sub        ecx, 4
+    jl         xloop49
+
+    // 4 Pixel loop.
+    align      4
+ xloop4:
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    pextrw     edx, xmm2, 7           // get x3 integer.
+    paddd      xmm2, xmm3             // x += dx
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movd       xmm1, [esi + eax * 4]  // 1 source x2 pixels
+    movd       xmm4, [esi + edx * 4]  // 1 source x3 pixels
+    pextrw     eax, xmm2, 1           // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3           // get x1 integer. next iteration.
+    punpckldq  xmm1, xmm4             // x2 x3
+    punpcklqdq xmm0, xmm1             // x0 x1 x2 x3
+    sub        ecx, 4                 // 4 pixels
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    jge        xloop4
+
+    align      4
+ xloop49:
+    test       ecx, 2
+    je         xloop29
+
+    // 2 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+
+ xloop29:
+    test       ecx, 1
+    je         xloop99
+
+    // 1 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x2 pixels
+    movd       dword ptr [edi], xmm0
+    align      4
+ xloop99:
+
+    pop        esi
+    pop        edi
+    ret
+  }
+}
+
+// Bilinear row filtering combines 2x1 -> 1x1. SSSE3 version.
+// TODO(fbarchard): Port to Neon
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+__declspec(naked) __declspec(align(16))
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+    movdqa     xmm4, kShuffleColARGB
+    movdqa     xmm5, kShuffleFractions
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+    align      4
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movhps     xmm0, qword ptr [esi + edx * 4]  // 2 source x1 pixels
+    pshufb     xmm1, xmm5           // 0000000011111111
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // argb_argb 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // argb 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // argb_argb 8 bits, 2 pixels.
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+    align      4
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    psrlw      xmm2, 9              // 7 bit fractions.
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    pshufb     xmm2, xmm5           // 00000000
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // argb 16 bit, 1 pixel.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0           // argb 8 bits, 1 pixel.
+    movd       [edi], xmm0
+
+    align      4
+ xloop99:
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+__declspec(naked) __declspec(align(16))
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_argb
+    mov        eax, [esp + 8]    // src_argb
+    mov        ecx, [esp + 12]   // dst_width
+
+    align      4
+  wloop:
+    movdqa     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpckldq  xmm0, xmm0
+    punpckhdq  xmm1, xmm1
+    sub        ecx, 8
+    movdqa     [edx], xmm0
+    movdqa     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    jg         wloop
+
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked) __declspec(align(16))
+int FixedDiv_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    idiv       dword ptr [esp + 8]
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked) __declspec(align(16))
+int FixedDiv1_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    mov        ecx, [esp + 8]    // denom
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    sub        eax, 0x00010001
+    sbb        edx, 0
+    sub        ecx, 1
+    idiv       ecx
+    ret
+  }
+}
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/video_common.cc b/drivers/theoraplayer/src/YUV/libyuv/src/video_common.cc
new file mode 100755
index 0000000000..efbedf46e2
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/video_common.cc
@@ -0,0 +1,64 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ARRAY_SIZE(x) (int)(sizeof(x) / sizeof(x[0]))
+
+struct FourCCAliasEntry {
+  uint32 alias;
+  uint32 canonical;
+};
+
+static const struct FourCCAliasEntry kFourCCAliases[] = {
+  {FOURCC_IYUV, FOURCC_I420},
+  {FOURCC_YU16, FOURCC_I422},
+  {FOURCC_YU24, FOURCC_I444},
+  {FOURCC_YUYV, FOURCC_YUY2},
+  {FOURCC_YUVS, FOURCC_YUY2},  // kCMPixelFormat_422YpCbCr8_yuvs
+  {FOURCC_HDYC, FOURCC_UYVY},
+  {FOURCC_2VUY, FOURCC_UYVY},  // kCMPixelFormat_422YpCbCr8
+  {FOURCC_JPEG, FOURCC_MJPG},  // Note: JPEG has DHT while MJPG does not.
+  {FOURCC_DMB1, FOURCC_MJPG},
+  {FOURCC_BA81, FOURCC_BGGR},
+  {FOURCC_RGB3, FOURCC_RAW },
+  {FOURCC_BGR3, FOURCC_24BG},
+  {FOURCC_CM32, FOURCC_BGRA},  // kCMPixelFormat_32ARGB
+  {FOURCC_CM24, FOURCC_RAW },  // kCMPixelFormat_24RGB
+  {FOURCC_L555, FOURCC_RGBO},  // kCMPixelFormat_16LE555
+  {FOURCC_L565, FOURCC_RGBP},  // kCMPixelFormat_16LE565
+  {FOURCC_5551, FOURCC_RGBO},  // kCMPixelFormat_16LE5551
+};
+// TODO(fbarchard): Consider mapping kCMPixelFormat_32BGRA to FOURCC_ARGB.
+//  {FOURCC_BGRA, FOURCC_ARGB},  // kCMPixelFormat_32BGRA
+
+LIBYUV_API
+uint32 CanonicalFourCC(uint32 fourcc) {
+  int i;
+  for (i = 0; i < ARRAY_SIZE(kFourCCAliases); ++i) {
+    if (kFourCCAliases[i].alias == fourcc) {
+      return kFourCCAliases[i].canonical;
+    }
+  }
+  // Not an alias, so return it as-is.
+  return fourcc;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/drivers/theoraplayer/src/YUV/libyuv/src/x86inc.asm b/drivers/theoraplayer/src/YUV/libyuv/src/x86inc.asm
new file mode 100755
index 0000000000..cb5c32df3a
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/src/x86inc.asm
@@ -0,0 +1,1136 @@
+;*****************************************************************************
+;* x86inc.asm: x264asm abstraction layer
+;*****************************************************************************
+;* Copyright (C) 2005-2012 x264 project
+;*
+;* Authors: Loren Merritt <lorenm@u.washington.edu>
+;*          Anton Mitrofanov <BugMaster@narod.ru>
+;*          Jason Garrett-Glaser <darkshikari@gmail.com>
+;*          Henrik Gramner <hengar-6@student.ltu.se>
+;*
+;* Permission to use, copy, modify, and/or distribute this software for any
+;* purpose with or without fee is hereby granted, provided that the above
+;* copyright notice and this permission notice appear in all copies.
+;*
+;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+;*****************************************************************************
+
+; This is a header file for the x264ASM assembly language, which uses
+; NASM/YASM syntax combined with a large number of macros to provide easy
+; abstraction between different calling conventions (x86_32, win64, linux64).
+; It also has various other useful features to simplify writing the kind of
+; DSP functions that are most often used in x264.
+
+; Unlike the rest of x264, this file is available under an ISC license, as it
+; has significant usefulness outside of x264 and we want it to be available
+; to the largest audience possible.  Of course, if you modify it for your own
+; purposes to add a new feature, we strongly encourage contributing a patch
+; as this feature might be useful for others as well.  Send patches or ideas
+; to x264-devel@videolan.org .
+
+; Local changes for libyuv:
+; remove %define program_name and references in labels
+; rename cpus to uppercase
+
+%define WIN64  0
+%define UNIX64 0
+%if ARCH_X86_64
+    %ifidn __OUTPUT_FORMAT__,win32
+        %define WIN64  1
+    %elifidn __OUTPUT_FORMAT__,win64
+        %define WIN64  1
+    %else
+        %define UNIX64 1
+    %endif
+%endif
+
+%ifdef PREFIX
+    %define mangle(x) _ %+ x
+%else
+    %define mangle(x) x
+%endif
+
+; Name of the .rodata section.
+; Kludge: Something on OS X fails to align .rodata even given an align attribute,
+; so use a different read-only section.
+%macro SECTION_RODATA 0-1 16
+    %ifidn __OUTPUT_FORMAT__,macho64
+        SECTION .text align=%1
+    %elifidn __OUTPUT_FORMAT__,macho
+        SECTION .text align=%1
+        fakegot:
+    %elifidn __OUTPUT_FORMAT__,aout
+        section .text
+    %else
+        SECTION .rodata align=%1
+    %endif
+%endmacro
+
+; aout does not support align=
+%macro SECTION_TEXT 0-1 16
+    %ifidn __OUTPUT_FORMAT__,aout
+        SECTION .text
+    %else
+        SECTION .text align=%1
+    %endif
+%endmacro
+
+%if WIN64
+    %define PIC
+%elif ARCH_X86_64 == 0
+; x86_32 doesn't require PIC.
+; Some distros prefer shared objects to be PIC, but nothing breaks if
+; the code contains a few textrels, so we'll skip that complexity.
+    %undef PIC
+%endif
+%ifdef PIC
+    default rel
+%endif
+
+; Always use long nops (reduces 0x90 spam in disassembly on x86_32)
+CPU amdnop
+
+; Macros to eliminate most code duplication between x86_32 and x86_64:
+; Currently this works only for leaf functions which load all their arguments
+; into registers at the start, and make no other use of the stack. Luckily that
+; covers most of x264's asm.
+
+; PROLOGUE:
+; %1 = number of arguments. loads them from stack if needed.
+; %2 = number of registers used. pushes callee-saved regs if needed.
+; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
+; %4 = list of names to define to registers
+; PROLOGUE can also be invoked by adding the same options to cglobal
+
+; e.g.
+; cglobal foo, 2,3,0, dst, src, tmp
+; declares a function (foo), taking two args (dst and src) and one local variable (tmp)
+
+; TODO Some functions can use some args directly from the stack. If they're the
+; last args then you can just not declare them, but if they're in the middle
+; we need more flexible macro.
+
+; RET:
+; Pops anything that was pushed by PROLOGUE, and returns.
+
+; REP_RET:
+; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
+; which are slow when a normal ret follows a branch.
+
+; registers:
+; rN and rNq are the native-size register holding function argument N
+; rNd, rNw, rNb are dword, word, and byte size
+; rNh is the high 8 bits of the word size
+; rNm is the original location of arg N (a register or on the stack), dword
+; rNmp is native size
+
+%macro DECLARE_REG 2-3
+    %define r%1q %2
+    %define r%1d %2d
+    %define r%1w %2w
+    %define r%1b %2b
+    %define r%1h %2h
+    %if %0 == 2
+        %define r%1m  %2d
+        %define r%1mp %2
+    %elif ARCH_X86_64 ; memory
+        %define r%1m [rsp + stack_offset + %3]
+        %define r%1mp qword r %+ %1m
+    %else
+        %define r%1m [esp + stack_offset + %3]
+        %define r%1mp dword r %+ %1m
+    %endif
+    %define r%1  %2
+%endmacro
+
+%macro DECLARE_REG_SIZE 3
+    %define r%1q r%1
+    %define e%1q r%1
+    %define r%1d e%1
+    %define e%1d e%1
+    %define r%1w %1
+    %define e%1w %1
+    %define r%1h %3
+    %define e%1h %3
+    %define r%1b %2
+    %define e%1b %2
+%if ARCH_X86_64 == 0
+    %define r%1  e%1
+%endif
+%endmacro
+
+DECLARE_REG_SIZE ax, al, ah
+DECLARE_REG_SIZE bx, bl, bh
+DECLARE_REG_SIZE cx, cl, ch
+DECLARE_REG_SIZE dx, dl, dh
+DECLARE_REG_SIZE si, sil, null
+DECLARE_REG_SIZE di, dil, null
+DECLARE_REG_SIZE bp, bpl, null
+
+; t# defines for when per-arch register allocation is more complex than just function arguments
+
+%macro DECLARE_REG_TMP 1-*
+    %assign %%i 0
+    %rep %0
+        CAT_XDEFINE t, %%i, r%1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro DECLARE_REG_TMP_SIZE 0-*
+    %rep %0
+        %define t%1q t%1 %+ q
+        %define t%1d t%1 %+ d
+        %define t%1w t%1 %+ w
+        %define t%1h t%1 %+ h
+        %define t%1b t%1 %+ b
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
+
+%if ARCH_X86_64
+    %define gprsize 8
+%else
+    %define gprsize 4
+%endif
+
+%macro PUSH 1
+    push %1
+    %assign stack_offset stack_offset+gprsize
+%endmacro
+
+%macro POP 1
+    pop %1
+    %assign stack_offset stack_offset-gprsize
+%endmacro
+
+%macro PUSH_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            PUSH r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro POP_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            pop r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro LOAD_IF_USED 1-*
+    %rep %0
+        %if %1 < num_args
+            mov r%1, r %+ %1 %+ mp
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro SUB 2
+    sub %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset+(%2)
+    %endif
+%endmacro
+
+%macro ADD 2
+    add %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset-(%2)
+    %endif
+%endmacro
+
+%macro movifnidn 2
+    %ifnidn %1, %2
+        mov %1, %2
+    %endif
+%endmacro
+
+%macro movsxdifnidn 2
+    %ifnidn %1, %2
+        movsxd %1, %2
+    %endif
+%endmacro
+
+%macro ASSERT 1
+    %if (%1) == 0
+        %error assert failed
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS 0-*
+    %ifdef n_arg_names
+        %assign %%i 0
+        %rep n_arg_names
+            CAT_UNDEF arg_name %+ %%i, q
+            CAT_UNDEF arg_name %+ %%i, d
+            CAT_UNDEF arg_name %+ %%i, w
+            CAT_UNDEF arg_name %+ %%i, h
+            CAT_UNDEF arg_name %+ %%i, b
+            CAT_UNDEF arg_name %+ %%i, m
+            CAT_UNDEF arg_name %+ %%i, mp
+            CAT_UNDEF arg_name, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+
+    %xdefine %%stack_offset stack_offset
+    %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine
+    %assign %%i 0
+    %rep %0
+        %xdefine %1q r %+ %%i %+ q
+        %xdefine %1d r %+ %%i %+ d
+        %xdefine %1w r %+ %%i %+ w
+        %xdefine %1h r %+ %%i %+ h
+        %xdefine %1b r %+ %%i %+ b
+        %xdefine %1m r %+ %%i %+ m
+        %xdefine %1mp r %+ %%i %+ mp
+        CAT_XDEFINE arg_name, %%i, %1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+    %xdefine stack_offset %%stack_offset
+    %assign n_arg_names %0
+%endmacro
+
+%if WIN64 ; Windows x64 ;=================================================
+
+DECLARE_REG 0,  rcx
+DECLARE_REG 1,  rdx
+DECLARE_REG 2,  R8
+DECLARE_REG 3,  R9
+DECLARE_REG 4,  R10, 40
+DECLARE_REG 5,  R11, 48
+DECLARE_REG 6,  rax, 56
+DECLARE_REG 7,  rdi, 64
+DECLARE_REG 8,  rsi, 72
+DECLARE_REG 9,  rbx, 80
+DECLARE_REG 10, rbp, 88
+DECLARE_REG 11, R12, 96
+DECLARE_REG 12, R13, 104
+DECLARE_REG 13, R14, 112
+DECLARE_REG 14, R15, 120
+
+%macro PROLOGUE 2-4+ 0 ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
+    %if mmsize == 8
+        %assign xmm_regs_used 0
+    %else
+        WIN64_SPILL_XMM %3
+    %endif
+    LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS %4
+%endmacro
+
+%macro WIN64_SPILL_XMM 1
+    %assign xmm_regs_used %1
+    ASSERT xmm_regs_used <= 16
+    %if xmm_regs_used > 6
+        SUB rsp, (xmm_regs_used-6)*16+16
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa [rsp + (%%i-6)*16+(~stack_offset&8)], xmm %+ %%i
+        %endrep
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM_INTERNAL 1
+    %if xmm_regs_used > 6
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa xmm %+ %%i, [%1 + (%%i-6)*16+(~stack_offset&8)]
+        %endrep
+        add %1, (xmm_regs_used-6)*16+16
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM 1
+    WIN64_RESTORE_XMM_INTERNAL %1
+    %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16
+    %assign xmm_regs_used 0
+%endmacro
+
+%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32
+
+%macro RET 0
+    WIN64_RESTORE_XMM_INTERNAL rsp
+    POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%elif ARCH_X86_64 ; *nix x64 ;=============================================
+
+DECLARE_REG 0,  rdi
+DECLARE_REG 1,  rsi
+DECLARE_REG 2,  rdx
+DECLARE_REG 3,  rcx
+DECLARE_REG 4,  R8
+DECLARE_REG 5,  R9
+DECLARE_REG 6,  rax, 8
+DECLARE_REG 7,  R10, 16
+DECLARE_REG 8,  R11, 24
+DECLARE_REG 9,  rbx, 32
+DECLARE_REG 10, rbp, 40
+DECLARE_REG 11, R12, 48
+DECLARE_REG 12, R13, 56
+DECLARE_REG 13, R14, 64
+DECLARE_REG 14, R15, 72
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 9, 10, 11, 12, 13, 14
+    LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS %4
+%endmacro
+
+%define has_epilogue regs_used > 9 || mmsize == 32
+
+%macro RET 0
+    POP_IF_USED 14, 13, 12, 11, 10, 9
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%else ; X86_32 ;==============================================================
+
+DECLARE_REG 0, eax, 4
+DECLARE_REG 1, ecx, 8
+DECLARE_REG 2, edx, 12
+DECLARE_REG 3, ebx, 16
+DECLARE_REG 4, esi, 20
+DECLARE_REG 5, edi, 24
+DECLARE_REG 6, ebp, 28
+%define rsp esp
+
+%macro DECLARE_ARG 1-*
+    %rep %0
+        %define r%1m [esp + stack_offset + 4*%1 + 4]
+        %define r%1mp dword r%1m
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    %if regs_used > 7
+        %assign regs_used 7
+    %endif
+    ASSERT regs_used >= num_args
+    PUSH_IF_USED 3, 4, 5, 6
+    LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
+    DEFINE_ARGS %4
+%endmacro
+
+%define has_epilogue regs_used > 3 || mmsize == 32
+
+%macro RET 0
+    POP_IF_USED 6, 5, 4, 3
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%endif ;======================================================================
+
+%if WIN64 == 0
+%macro WIN64_SPILL_XMM 1
+%endmacro
+%macro WIN64_RESTORE_XMM 1
+%endmacro
+%endif
+
+%macro REP_RET 0
+    %if has_epilogue
+        RET
+    %else
+        rep ret
+    %endif
+%endmacro
+
+%macro TAIL_CALL 2 ; callee, is_nonadjacent
+    %if has_epilogue
+        call %1
+        RET
+    %elif %2
+        jmp %1
+    %endif
+%endmacro
+
+;=============================================================================
+; arch-independent part
+;=============================================================================
+
+%assign function_align 16
+
+; Begin a function.
+; Applies any symbol mangling needed for C linkage, and sets up a define such that
+; subsequent uses of the function name automatically refer to the mangled version.
+; Appends cpuflags to the function name if cpuflags has been specified.
+%macro cglobal 1-2+ ; name, [PROLOGUE args]
+%if %0 == 1
+    cglobal_internal %1 %+ SUFFIX
+%else
+    cglobal_internal %1 %+ SUFFIX, %2
+%endif
+%endmacro
+%macro cglobal_internal 1-2+
+    %ifndef cglobaled_%1
+        %xdefine %1 mangle(%1)
+        %xdefine %1.skip_prologue %1 %+ .skip_prologue
+        CAT_XDEFINE cglobaled_, %1, 1
+    %endif
+    %xdefine current_function %1
+    %ifidn __OUTPUT_FORMAT__,elf
+        global %1:function hidden
+    %else
+        global %1
+    %endif
+    align function_align
+    %1:
+    RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
+    %assign stack_offset 0
+    %if %0 > 1
+        PROLOGUE %2
+    %endif
+%endmacro
+
+%macro cextern 1
+    %xdefine %1 mangle(%1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+; like cextern, but without the prefix
+%macro cextern_naked 1
+    %xdefine %1 mangle(%1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+%macro const 2+
+    %xdefine %1 mangle(%1)
+    global %1
+    %1: %2
+%endmacro
+
+; This is needed for ELF, otherwise the GNU linker assumes the stack is
+; executable by default.
+%ifidn __OUTPUT_FORMAT__,elf
+SECTION .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+%ifidn __OUTPUT_FORMAT__,elf32
+section .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+%ifidn __OUTPUT_FORMAT__,elf64
+section .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+
+; cpuflags
+
+%assign cpuflags_MMX      (1<<0)
+%assign cpuflags_MMX2     (1<<1) | cpuflags_MMX
+%assign cpuflags_3dnow    (1<<2) | cpuflags_MMX
+%assign cpuflags_3dnow2   (1<<3) | cpuflags_3dnow
+%assign cpuflags_SSE      (1<<4) | cpuflags_MMX2
+%assign cpuflags_SSE2     (1<<5) | cpuflags_SSE
+%assign cpuflags_SSE2slow (1<<6) | cpuflags_SSE2
+%assign cpuflags_SSE3     (1<<7) | cpuflags_SSE2
+%assign cpuflags_SSSE3    (1<<8) | cpuflags_SSE3
+%assign cpuflags_SSE4     (1<<9) | cpuflags_SSSE3
+%assign cpuflags_SSE42    (1<<10)| cpuflags_SSE4
+%assign cpuflags_AVX      (1<<11)| cpuflags_SSE42
+%assign cpuflags_xop      (1<<12)| cpuflags_AVX
+%assign cpuflags_fma4     (1<<13)| cpuflags_AVX
+%assign cpuflags_AVX2     (1<<14)| cpuflags_AVX
+%assign cpuflags_fma3     (1<<15)| cpuflags_AVX
+
+%assign cpuflags_cache32  (1<<16)
+%assign cpuflags_cache64  (1<<17)
+%assign cpuflags_slowctz  (1<<18)
+%assign cpuflags_lzcnt    (1<<19)
+%assign cpuflags_misalign (1<<20)
+%assign cpuflags_aligned  (1<<21) ; not a cpu feature, but a function variant
+%assign cpuflags_atom     (1<<22)
+%assign cpuflags_bmi1     (1<<23)
+%assign cpuflags_bmi2     (1<<24)|cpuflags_bmi1
+%assign cpuflags_tbm      (1<<25)|cpuflags_bmi1
+
+%define    cpuflag(x) ((cpuflags & (cpuflags_ %+ x)) == (cpuflags_ %+ x))
+%define notcpuflag(x) ((cpuflags & (cpuflags_ %+ x)) != (cpuflags_ %+ x))
+
+; Takes up to 2 cpuflags from the above list.
+; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
+; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
+%macro INIT_CPUFLAGS 0-2
+    %if %0 >= 1
+        %xdefine cpuname %1
+        %assign cpuflags cpuflags_%1
+        %if %0 >= 2
+            %xdefine cpuname %1_%2
+            %assign cpuflags cpuflags | cpuflags_%2
+        %endif
+        %xdefine SUFFIX _ %+ cpuname
+        %if cpuflag(AVX)
+            %assign AVX_enabled 1
+        %endif
+        %if mmsize == 16 && notcpuflag(SSE2)
+            %define mova movaps
+            %define movu movups
+            %define movnta movntps
+        %endif
+        %if cpuflag(aligned)
+            %define movu mova
+        %elifidn %1, SSE3
+            %define movu lddqu
+        %endif
+    %else
+        %xdefine SUFFIX
+        %undef cpuname
+        %undef cpuflags
+    %endif
+%endmacro
+
+; merge MMX and SSE*
+
+%macro CAT_XDEFINE 3
+    %xdefine %1%2 %3
+%endmacro
+
+%macro CAT_UNDEF 2
+    %undef %1%2
+%endmacro
+
+%macro INIT_MMX 0-1+
+    %assign AVX_enabled 0
+    %define RESET_MM_PERMUTATION INIT_MMX %1
+    %define mmsize 8
+    %define num_mmregs 8
+    %define mova movq
+    %define movu movq
+    %define movh movd
+    %define movnta movntq
+    %assign %%i 0
+    %rep 8
+    CAT_XDEFINE m, %%i, mm %+ %%i
+    CAT_XDEFINE nmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    %rep 8
+    CAT_UNDEF m, %%i
+    CAT_UNDEF nmm, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_XMM 0-1+
+    %assign AVX_enabled 0
+    %define RESET_MM_PERMUTATION INIT_XMM %1
+    %define mmsize 16
+    %define num_mmregs 8
+    %if ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %define movh movq
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, xmm %+ %%i
+    CAT_XDEFINE nxmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_YMM 0-1+
+    %assign AVX_enabled 1
+    %define RESET_MM_PERMUTATION INIT_YMM %1
+    %define mmsize 32
+    %define num_mmregs 8
+    %if ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova vmovaps
+    %define movu vmovups
+    %undef movh
+    %define movnta vmovntps
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, ymm %+ %%i
+    CAT_XDEFINE nymm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+INIT_XMM
+
+; I often want to use macros that permute their arguments. e.g. there's no
+; efficient way to implement butterfly or transpose or dct without swapping some
+; arguments.
+;
+; I would like to not have to manually keep track of the permutations:
+; If I insert a permutation in the middle of a function, it should automatically
+; change everything that follows. For more complex macros I may also have multiple
+; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
+;
+; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
+; permutes its arguments. It's equivalent to exchanging the contents of the
+; registers, except that this way you exchange the register names instead, so it
+; doesn't cost any cycles.
+
+%macro PERMUTE 2-* ; takes a list of pairs to swap
+%rep %0/2
+    %xdefine tmp%2 m%2
+    %xdefine ntmp%2 nm%2
+    %rotate 2
+%endrep
+%rep %0/2
+    %xdefine m%1 tmp%2
+    %xdefine nm%1 ntmp%2
+    %undef tmp%2
+    %undef ntmp%2
+    %rotate 2
+%endrep
+%endmacro
+
+%macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs)
+%rep %0-1
+%ifdef m%1
+    %xdefine tmp m%1
+    %xdefine m%1 m%2
+    %xdefine m%2 tmp
+    CAT_XDEFINE n, m%1, %1
+    CAT_XDEFINE n, m%2, %2
+%else
+    ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here.
+    ; Be careful using this mode in nested macros though, as in some cases there may be
+    ; other copies of m# that have already been dereferenced and don't get updated correctly.
+    %xdefine %%n1 n %+ %1
+    %xdefine %%n2 n %+ %2
+    %xdefine tmp m %+ %%n1
+    CAT_XDEFINE m, %%n1, m %+ %%n2
+    CAT_XDEFINE m, %%n2, tmp
+    CAT_XDEFINE n, m %+ %%n1, %%n1
+    CAT_XDEFINE n, m %+ %%n2, %%n2
+%endif
+    %undef tmp
+    %rotate 1
+%endrep
+%endmacro
+
+; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
+; calls to that function will automatically load the permutation, so values can
+; be returned in mmregs.
+%macro SAVE_MM_PERMUTATION 0-1
+    %if %0
+        %xdefine %%f %1_m
+    %else
+        %xdefine %%f current_function %+ _m
+    %endif
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE %%f, %%i, m %+ %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro LOAD_MM_PERMUTATION 1 ; name to load from
+    %ifdef %1_m0
+        %assign %%i 0
+        %rep num_mmregs
+            CAT_XDEFINE m, %%i, %1_m %+ %%i
+            CAT_XDEFINE n, m %+ %%i, %%i
+        %assign %%i %%i+1
+        %endrep
+    %endif
+%endmacro
+
+; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
+%macro call 1
+    call_internal %1, %1 %+ SUFFIX
+%endmacro
+%macro call_internal 2
+    %xdefine %%i %1
+    %ifndef cglobaled_%1
+        %ifdef cglobaled_%2
+            %xdefine %%i %2
+        %endif
+    %endif
+    call %%i
+    LOAD_MM_PERMUTATION %%i
+%endmacro
+
+; Substitutions that reduce instruction size but are functionally equivalent
+%macro add 2
+    %ifnum %2
+        %if %2==128
+            sub %1, -128
+        %else
+            add %1, %2
+        %endif
+    %else
+        add %1, %2
+    %endif
+%endmacro
+
+%macro sub 2
+    %ifnum %2
+        %if %2==128
+            add %1, -128
+        %else
+            sub %1, %2
+        %endif
+    %else
+        sub %1, %2
+    %endif
+%endmacro
+
+;=============================================================================
+; AVX abstraction layer
+;=============================================================================
+
+%assign i 0
+%rep 16
+    %if i < 8
+        CAT_XDEFINE sizeofmm, i, 8
+    %endif
+    CAT_XDEFINE sizeofxmm, i, 16
+    CAT_XDEFINE sizeofymm, i, 32
+%assign i i+1
+%endrep
+%undef i
+
+%macro CHECK_AVX_INSTR_EMU 3-*
+    %xdefine %%opcode %1
+    %xdefine %%dst %2
+    %rep %0-2
+        %ifidn %%dst, %3
+            %error non-AVX emulation of ``%%opcode'' is not supported
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
+;%4 == number of operands given
+;%5+: operands
+%macro RUN_AVX_INSTR 6-7+
+    %ifid %6
+        %define %%sizeofreg sizeof%6
+    %elifid %5
+        %define %%sizeofreg sizeof%5
+    %else
+        %define %%sizeofreg mmsize
+    %endif
+    %if %%sizeofreg==32
+        %if %4>=3
+            v%1 %5, %6, %7
+        %else
+            v%1 %5, %6
+        %endif
+    %else
+        %if %%sizeofreg==8
+            %define %%regmov movq
+        %elif %2
+            %define %%regmov movaps
+        %else
+            %define %%regmov movdqa
+        %endif
+
+        %if %4>=3+%3
+            %ifnidn %5, %6
+                %if AVX_enabled && %%sizeofreg==16
+                    v%1 %5, %6, %7
+                %else
+                    CHECK_AVX_INSTR_EMU {%1 %5, %6, %7}, %5, %7
+                    %%regmov %5, %6
+                    %1 %5, %7
+                %endif
+            %else
+                %1 %5, %7
+            %endif
+        %elif %4>=3
+            %1 %5, %6, %7
+        %else
+            %1 %5, %6
+        %endif
+    %endif
+%endmacro
+
+; 3arg AVX ops with a memory arg can only have it in src2,
+; whereas SSE emulation of 3arg prefers to have it in src1 (i.e. the mov).
+; So, if the op is symmetric and the wrong one is memory, swap them.
+%macro RUN_AVX_INSTR1 8
+    %assign %%swap 0
+    %if AVX_enabled
+        %ifnid %6
+            %assign %%swap 1
+        %endif
+    %elifnidn %5, %6
+        %ifnid %7
+            %assign %%swap 1
+        %endif
+    %endif
+    %if %%swap && %3 == 0 && %8 == 1
+        RUN_AVX_INSTR %1, %2, %3, %4, %5, %7, %6
+    %else
+        RUN_AVX_INSTR %1, %2, %3, %4, %5, %6, %7
+    %endif
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
+;%4 == 1 if symmetric (i.e. doesn't matter which src arg is which), 0 if not
+%macro AVX_INSTR 4
+    %macro %1 2-9 fnord, fnord, fnord, %1, %2, %3, %4
+        %ifidn %3, fnord
+            RUN_AVX_INSTR %6, %7, %8, 2, %1, %2
+        %elifidn %4, fnord
+            RUN_AVX_INSTR1 %6, %7, %8, 3, %1, %2, %3, %9
+        %elifidn %5, fnord
+            RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4
+        %else
+            RUN_AVX_INSTR %6, %7, %8, 5, %1, %2, %3, %4, %5
+        %endif
+    %endmacro
+%endmacro
+
+AVX_INSTR addpd, 1, 0, 1
+AVX_INSTR addps, 1, 0, 1
+AVX_INSTR addsd, 1, 0, 1
+AVX_INSTR addss, 1, 0, 1
+AVX_INSTR addsubpd, 1, 0, 0
+AVX_INSTR addsubps, 1, 0, 0
+AVX_INSTR andpd, 1, 0, 1
+AVX_INSTR andps, 1, 0, 1
+AVX_INSTR andnpd, 1, 0, 0
+AVX_INSTR andnps, 1, 0, 0
+AVX_INSTR blendpd, 1, 0, 0
+AVX_INSTR blendps, 1, 0, 0
+AVX_INSTR blendvpd, 1, 0, 0
+AVX_INSTR blendvps, 1, 0, 0
+AVX_INSTR cmppd, 1, 0, 0
+AVX_INSTR cmpps, 1, 0, 0
+AVX_INSTR cmpsd, 1, 0, 0
+AVX_INSTR cmpss, 1, 0, 0
+AVX_INSTR cvtdq2ps, 1, 0, 0
+AVX_INSTR cvtps2dq, 1, 0, 0
+AVX_INSTR divpd, 1, 0, 0
+AVX_INSTR divps, 1, 0, 0
+AVX_INSTR divsd, 1, 0, 0
+AVX_INSTR divss, 1, 0, 0
+AVX_INSTR dppd, 1, 1, 0
+AVX_INSTR dpps, 1, 1, 0
+AVX_INSTR haddpd, 1, 0, 0
+AVX_INSTR haddps, 1, 0, 0
+AVX_INSTR hsubpd, 1, 0, 0
+AVX_INSTR hsubps, 1, 0, 0
+AVX_INSTR maxpd, 1, 0, 1
+AVX_INSTR maxps, 1, 0, 1
+AVX_INSTR maxsd, 1, 0, 1
+AVX_INSTR maxss, 1, 0, 1
+AVX_INSTR minpd, 1, 0, 1
+AVX_INSTR minps, 1, 0, 1
+AVX_INSTR minsd, 1, 0, 1
+AVX_INSTR minss, 1, 0, 1
+AVX_INSTR movhlps, 1, 0, 0
+AVX_INSTR movlhps, 1, 0, 0
+AVX_INSTR movsd, 1, 0, 0
+AVX_INSTR movss, 1, 0, 0
+AVX_INSTR mpsadbw, 0, 1, 0
+AVX_INSTR mulpd, 1, 0, 1
+AVX_INSTR mulps, 1, 0, 1
+AVX_INSTR mulsd, 1, 0, 1
+AVX_INSTR mulss, 1, 0, 1
+AVX_INSTR orpd, 1, 0, 1
+AVX_INSTR orps, 1, 0, 1
+AVX_INSTR pabsb, 0, 0, 0
+AVX_INSTR pabsw, 0, 0, 0
+AVX_INSTR pabsd, 0, 0, 0
+AVX_INSTR packsswb, 0, 0, 0
+AVX_INSTR packssdw, 0, 0, 0
+AVX_INSTR packuswb, 0, 0, 0
+AVX_INSTR packusdw, 0, 0, 0
+AVX_INSTR paddb, 0, 0, 1
+AVX_INSTR paddw, 0, 0, 1
+AVX_INSTR paddd, 0, 0, 1
+AVX_INSTR paddq, 0, 0, 1
+AVX_INSTR paddsb, 0, 0, 1
+AVX_INSTR paddsw, 0, 0, 1
+AVX_INSTR paddusb, 0, 0, 1
+AVX_INSTR paddusw, 0, 0, 1
+AVX_INSTR palignr, 0, 1, 0
+AVX_INSTR pand, 0, 0, 1
+AVX_INSTR pandn, 0, 0, 0
+AVX_INSTR pavgb, 0, 0, 1
+AVX_INSTR pavgw, 0, 0, 1
+AVX_INSTR pblendvb, 0, 0, 0
+AVX_INSTR pblendw, 0, 1, 0
+AVX_INSTR pcmpestri, 0, 0, 0
+AVX_INSTR pcmpestrm, 0, 0, 0
+AVX_INSTR pcmpistri, 0, 0, 0
+AVX_INSTR pcmpistrm, 0, 0, 0
+AVX_INSTR pcmpeqb, 0, 0, 1
+AVX_INSTR pcmpeqw, 0, 0, 1
+AVX_INSTR pcmpeqd, 0, 0, 1
+AVX_INSTR pcmpeqq, 0, 0, 1
+AVX_INSTR pcmpgtb, 0, 0, 0
+AVX_INSTR pcmpgtw, 0, 0, 0
+AVX_INSTR pcmpgtd, 0, 0, 0
+AVX_INSTR pcmpgtq, 0, 0, 0
+AVX_INSTR phaddw, 0, 0, 0
+AVX_INSTR phaddd, 0, 0, 0
+AVX_INSTR phaddsw, 0, 0, 0
+AVX_INSTR phsubw, 0, 0, 0
+AVX_INSTR phsubd, 0, 0, 0
+AVX_INSTR phsubsw, 0, 0, 0
+AVX_INSTR pmaddwd, 0, 0, 1
+AVX_INSTR pmaddubsw, 0, 0, 0
+AVX_INSTR pmaxsb, 0, 0, 1
+AVX_INSTR pmaxsw, 0, 0, 1
+AVX_INSTR pmaxsd, 0, 0, 1
+AVX_INSTR pmaxub, 0, 0, 1
+AVX_INSTR pmaxuw, 0, 0, 1
+AVX_INSTR pmaxud, 0, 0, 1
+AVX_INSTR pminsb, 0, 0, 1
+AVX_INSTR pminsw, 0, 0, 1
+AVX_INSTR pminsd, 0, 0, 1
+AVX_INSTR pminub, 0, 0, 1
+AVX_INSTR pminuw, 0, 0, 1
+AVX_INSTR pminud, 0, 0, 1
+AVX_INSTR pmovmskb, 0, 0, 0
+AVX_INSTR pmulhuw, 0, 0, 1
+AVX_INSTR pmulhrsw, 0, 0, 1
+AVX_INSTR pmulhw, 0, 0, 1
+AVX_INSTR pmullw, 0, 0, 1
+AVX_INSTR pmulld, 0, 0, 1
+AVX_INSTR pmuludq, 0, 0, 1
+AVX_INSTR pmuldq, 0, 0, 1
+AVX_INSTR por, 0, 0, 1
+AVX_INSTR psadbw, 0, 0, 1
+AVX_INSTR pshufb, 0, 0, 0
+AVX_INSTR pshufd, 0, 1, 0
+AVX_INSTR pshufhw, 0, 1, 0
+AVX_INSTR pshuflw, 0, 1, 0
+AVX_INSTR psignb, 0, 0, 0
+AVX_INSTR psignw, 0, 0, 0
+AVX_INSTR psignd, 0, 0, 0
+AVX_INSTR psllw, 0, 0, 0
+AVX_INSTR pslld, 0, 0, 0
+AVX_INSTR psllq, 0, 0, 0
+AVX_INSTR pslldq, 0, 0, 0
+AVX_INSTR psraw, 0, 0, 0
+AVX_INSTR psrad, 0, 0, 0
+AVX_INSTR psrlw, 0, 0, 0
+AVX_INSTR psrld, 0, 0, 0
+AVX_INSTR psrlq, 0, 0, 0
+AVX_INSTR psrldq, 0, 0, 0
+AVX_INSTR psubb, 0, 0, 0
+AVX_INSTR psubw, 0, 0, 0
+AVX_INSTR psubd, 0, 0, 0
+AVX_INSTR psubq, 0, 0, 0
+AVX_INSTR psubsb, 0, 0, 0
+AVX_INSTR psubsw, 0, 0, 0
+AVX_INSTR psubusb, 0, 0, 0
+AVX_INSTR psubusw, 0, 0, 0
+AVX_INSTR ptest, 0, 0, 0
+AVX_INSTR punpckhbw, 0, 0, 0
+AVX_INSTR punpckhwd, 0, 0, 0
+AVX_INSTR punpckhdq, 0, 0, 0
+AVX_INSTR punpckhqdq, 0, 0, 0
+AVX_INSTR punpcklbw, 0, 0, 0
+AVX_INSTR punpcklwd, 0, 0, 0
+AVX_INSTR punpckldq, 0, 0, 0
+AVX_INSTR punpcklqdq, 0, 0, 0
+AVX_INSTR pxor, 0, 0, 1
+AVX_INSTR shufps, 1, 1, 0
+AVX_INSTR subpd, 1, 0, 0
+AVX_INSTR subps, 1, 0, 0
+AVX_INSTR subsd, 1, 0, 0
+AVX_INSTR subss, 1, 0, 0
+AVX_INSTR unpckhpd, 1, 0, 0
+AVX_INSTR unpckhps, 1, 0, 0
+AVX_INSTR unpcklpd, 1, 0, 0
+AVX_INSTR unpcklps, 1, 0, 0
+AVX_INSTR xorpd, 1, 0, 1
+AVX_INSTR xorps, 1, 0, 1
+
+; 3DNow instructions, for sharing code between AVX, SSE and 3DN
+AVX_INSTR pfadd, 1, 0, 1
+AVX_INSTR pfsub, 1, 0, 0
+AVX_INSTR pfmul, 1, 0, 1
+
+; base-4 constants for shuffles
+%assign i 0
+%rep 256
+    %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
+    %if j < 10
+        CAT_XDEFINE q000, j, i
+    %elif j < 100
+        CAT_XDEFINE q00, j, i
+    %elif j < 1000
+        CAT_XDEFINE q0, j, i
+    %else
+        CAT_XDEFINE q, j, i
+    %endif
+%assign i i+1
+%endrep
+%undef i
+%undef j
+
+%macro FMA_INSTR 3
+    %macro %1 4-7 %1, %2, %3
+        %if cpuflag(xop)
+            v%5 %1, %2, %3, %4
+        %else
+            %6 %1, %2, %3
+            %7 %1, %4
+        %endif
+    %endmacro
+%endmacro
+
+FMA_INSTR  pmacsdd,  pmulld, paddd
+FMA_INSTR  pmacsww,  pmullw, paddw
+FMA_INSTR pmadcswd, pmaddwd, paddd
+
+; tzcnt is equivalent to "rep bsf" and is backwards-compatible with bsf.
+; This lets us use tzcnt without bumping the yasm version requirement yet.
+%define tzcnt rep bsf
diff --git a/drivers/theoraplayer/src/YUV/libyuv/yuv_libyuv.c b/drivers/theoraplayer/src/YUV/libyuv/yuv_libyuv.c
new file mode 100755
index 0000000000..3712a3b6d6
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/yuv_libyuv.c
@@ -0,0 +1,72 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#ifdef _YUV_LIBYUV
+#include <libyuv.h>
+#include "yuv_util.h"
+#include "yuv_libyuv.h"
+
+void decodeRGB(struct TheoraPixelTransform* t)
+{
+	I420ToRAW(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 3, t->w, t->h);
+}
+
+void decodeRGBA(struct TheoraPixelTransform* t)
+{
+	I420ToABGR(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 4, t->w, t->h);
+	_decodeAlpha(incOut(t, 3), t->w * 4);
+}
+
+void decodeRGBX(struct TheoraPixelTransform* t)
+{
+    I420ToABGR(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 4, t->w, t->h);
+}
+
+void decodeARGB(struct TheoraPixelTransform* t)
+{
+    I420ToBGRA(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 4, t->w, t->h);
+	_decodeAlpha(t, t->w * 4);
+}
+
+void decodeXRGB(struct TheoraPixelTransform* t)
+{
+    I420ToBGRA(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 4, t->w, t->h);
+}
+
+void decodeBGR(struct TheoraPixelTransform* t)
+{
+	I420ToRGB24(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 3, t->w, t->h);
+}
+
+void decodeBGRA(struct TheoraPixelTransform* t)
+{
+	I420ToARGB(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 4, t->w, t->h);
+	_decodeAlpha(incOut(t, 3), t->w * 4);
+}
+
+void decodeBGRX(struct TheoraPixelTransform* t)
+{
+	I420ToARGB(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 4, t->w, t->h);
+}
+
+void decodeABGR(struct TheoraPixelTransform* t)
+{
+    I420ToRGBA(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 4, t->w, t->h);
+	_decodeAlpha(t, t->w * 4);
+}
+
+void decodeXBGR(struct TheoraPixelTransform* t)
+{
+	I420ToRGBA(t->y, t->yStride, t->u, t->uStride, t->v, t->vStride, t->out, t->w * 4, t->w, t->h);
+}
+
+void initYUVConversionModule()
+{
+
+}
+#endif
diff --git a/drivers/theoraplayer/src/YUV/libyuv/yuv_libyuv.h b/drivers/theoraplayer/src/YUV/libyuv/yuv_libyuv.h
new file mode 100755
index 0000000000..f621af0c5f
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/libyuv/yuv_libyuv.h
@@ -0,0 +1,14 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#ifndef _YUV_LIBYUV_h
+#define _YUV_LIBYUV_h
+
+#include "TheoraPixelTransform.h"
+
+#endif
diff --git a/drivers/theoraplayer/src/YUV/yuv_util.c b/drivers/theoraplayer/src/YUV/yuv_util.c
new file mode 100644
index 0000000000..f5bf3e5f9e
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/yuv_util.c
@@ -0,0 +1,39 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#include "yuv_util.h"
+
+struct TheoraPixelTransform* incOut(struct TheoraPixelTransform* t, int n)
+{
+	// used for XRGB, XBGR and similar
+	t->out += n;
+	return t;
+}
+
+void _decodeAlpha(struct TheoraPixelTransform* t, int stride)
+{
+	int width = t->w;
+	unsigned char *ySrc, *yLineEnd, *out;
+	int luma;
+	unsigned int y;
+	for (y = 0; y < t->h; y++)
+	{
+		ySrc = t->y + y * t->yStride + width;
+		out = t->out + y * stride;
+		
+		for (yLineEnd = ySrc + width; ySrc != yLineEnd; ++ySrc, out += 4)
+		{
+			luma = (*ySrc);
+            // because in YCbCr specification, luma values are in the range of [16, 235]
+            // account for 'footroom' and 'headroom' ranges while using luma values as alpha channel
+            if (luma <= 16)       *out = 0;
+            else if (luma >= 235) *out = 255;
+            else                  *out = (unsigned char) (((luma - 16) * 255) / 219);
+		}
+	}
+}
diff --git a/drivers/theoraplayer/src/YUV/yuv_util.h b/drivers/theoraplayer/src/YUV/yuv_util.h
new file mode 100644
index 0000000000..1f9d76634a
--- /dev/null
+++ b/drivers/theoraplayer/src/YUV/yuv_util.h
@@ -0,0 +1,17 @@
+/************************************************************************************
+This source file is part of the Theora Video Playback Library
+For latest info, see http://libtheoraplayer.googlecode.com
+*************************************************************************************
+Copyright (c) 2008-2014 Kresimir Spes (kspes@cateia.com)
+This program is free software; you can redistribute it and/or modify it under
+the terms of the BSD license: http://opensource.org/licenses/BSD-3-Clause
+*************************************************************************************/
+#ifndef _YUV_UTIL_h
+#define _YUV_UTIL_h
+
+#include "TheoraPixelTransform.h"
+
+struct TheoraPixelTransform* incOut(struct TheoraPixelTransform* t, int n);
+void _decodeAlpha(struct TheoraPixelTransform* t, int stride);
+
+#endif