From 782872541396e00a09b479a780d26af3123f6390 Mon Sep 17 00:00:00 2001
From: erbth <t.erbesdobler@team103.com>
Date: Tue, 9 Sep 2014 00:13:18 +0200
Subject: [PATCH] YUV data conversion of H.264 implementation (egfx): only
 convert invalid areas SIMD SSSE3 conversion in primitives compiling all
 primitives sources with optimization

and cleanup after last merge
---
 channels/drdynvc/client/dvcman.c              |   9 +-
 include/freerdp/codec/h264.h                  |   6 +-
 libfreerdp/codec/CMakeLists.txt               |  18 ---
 libfreerdp/codec/h264.c                       | 129 +++++-----------
 .../codec/test/Makefile.TestOpenH264ASM32     |  17 ---
 .../codec/test/Makefile.TestOpenH264ASM64     |  17 ---
 .../codec/test/Makefile.TestOpenH264SSSE3     |  14 --
 libfreerdp/codec/test/TestOpenH264            | Bin 15584 -> 0 bytes
 libfreerdp/codec/test/TestOpenH264ASM.c       |  92 ------------
 libfreerdp/codec/test/TestOpenH264ASM.h       |   7 -
 libfreerdp/primitives/CMakeLists.txt          |  15 +-
 libfreerdp/primitives/prim_YUV.c              | 138 +++++++++++-------
 libfreerdp/primitives/prim_YUV.h              |   1 +
 .../prim_YUV_opt.c}                           |  97 ++++++------
 winpr/libwinpr/utils/collections/StreamPool.c |   2 -
 15 files changed, 199 insertions(+), 363 deletions(-)
 delete mode 100644 libfreerdp/codec/test/Makefile.TestOpenH264ASM32
 delete mode 100644 libfreerdp/codec/test/Makefile.TestOpenH264ASM64
 delete mode 100644 libfreerdp/codec/test/Makefile.TestOpenH264SSSE3
 delete mode 100755 libfreerdp/codec/test/TestOpenH264
 delete mode 100644 libfreerdp/codec/test/TestOpenH264ASM.c
 delete mode 100644 libfreerdp/codec/test/TestOpenH264ASM.h
 rename libfreerdp/{codec/h264_ssse3.c => primitives/prim_YUV_opt.c} (80%)

diff --git a/channels/drdynvc/client/dvcman.c b/channels/drdynvc/client/dvcman.c
index 001717e14..f9e4873b8 100644
--- a/channels/drdynvc/client/dvcman.c
+++ b/channels/drdynvc/client/dvcman.c
@@ -486,7 +486,6 @@ int dvcman_receive_channel_data(IWTSVirtualChannelManager* pChannelMgr, UINT32 C
 	int status = 0;
 	DVCMAN_CHANNEL* channel;
 	UINT32 dataSize = Stream_GetRemainingLength(data);
-	wStream* s;
 
 	channel = (DVCMAN_CHANNEL*) dvcman_find_channel_by_id(pChannelMgr, ChannelId);
 
@@ -499,7 +498,7 @@ int dvcman_receive_channel_data(IWTSVirtualChannelManager* pChannelMgr, UINT32 C
 	if (channel->dvc_data)
 	{
 		/* Fragmented data */
-		if (Stream_GetPosition(channel->dvc_data) + dataSize > (UINT32) Stream_Length(channel->dvc_data))
+		if (Stream_GetPosition(channel->dvc_data) + dataSize > (UINT32) Stream_Capacity(channel->dvc_data))
 		{
 			CLOG_ERR("data exceeding declared length!");
 			Stream_Release(channel->dvc_data);
@@ -513,11 +512,9 @@ int dvcman_receive_channel_data(IWTSVirtualChannelManager* pChannelMgr, UINT32 C
 		{
 			Stream_SealLength(channel->dvc_data);
 			Stream_SetPosition(channel->dvc_data, 0);
-			s=channel->dvc_data;
+			status = channel->channel_callback->OnDataReceived(channel->channel_callback, channel->dvc_data);
+			Stream_Release(channel->dvc_data);
 			channel->dvc_data = NULL;
-
-			status = channel->channel_callback->OnDataReceived(channel->channel_callback, s);
-			Stream_Release(s);
 		}
 	}
 	else
diff --git a/include/freerdp/codec/h264.h b/include/freerdp/codec/h264.h
index d29a9e243..969914709 100644
--- a/include/freerdp/codec/h264.h
+++ b/include/freerdp/codec/h264.h
@@ -29,8 +29,7 @@ typedef struct _H264_CONTEXT H264_CONTEXT;
 typedef BOOL (*pfnH264SubsystemInit)(H264_CONTEXT* h264);
 typedef void (*pfnH264SubsystemUninit)(H264_CONTEXT* h264);
 
-typedef int (*pfnH264SubsystemDecompress)(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize,
-		BYTE* pDstData, DWORD DstFormat, int nDstStep, int nXDst, int nYDst, int nWidth, int nHeight);
+typedef int (*pfnH264SubsystemDecompress)(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize);
 
 struct _H264_CONTEXT_SUBSYSTEM
 {
@@ -50,6 +49,9 @@ struct _H264_CONTEXT
 	UINT32 width;
 	UINT32 height;
 	//int scanline;
+	
+	BYTE* pYUVData[3];
+	int iStride[3];
 
 /*
 <<<<<<< HEAD
diff --git a/libfreerdp/codec/CMakeLists.txt b/libfreerdp/codec/CMakeLists.txt
index f8ac3faa5..75999d262 100644
--- a/libfreerdp/codec/CMakeLists.txt
+++ b/libfreerdp/codec/CMakeLists.txt
@@ -101,24 +101,6 @@ if(WITH_LIBAVCODEC)
 	set(FREERDP_LIBAVCODEC_LIBS ${LIBAVCODEC_LIB} ${LIBAVUTIL_LIB})
 endif()
 
-if(WITH_LIBAVCODEC OR WITH_OPENH264)
-	if(WITH_H264_SSSE3)
-		add_definitions(-DWITH_H264_SSSE3)
-		set(${MODULE_PREFIX}_SRCS
-			${${MODULE_PREFIX}_SRCS}
-			h264_ssse3.c)
-		
-		if(CMAKE_COMPILER_IS_GNUCC)
-			set(OPTIMIZATION "${OPTIMIZATION} -msse2 -mssse3")
-		endif()
-		
-		if(MSVC)
-			set(OPTIMIZATION "${OPTIMIZATION} /arch:SSE2")
-		endif()
-		
-		set_property(SOURCE h264_ssse3.c PROPERTY COMPILE_FLAGS ${OPTIMIZATION})
-	endif()
-endif()
 
 add_complex_library(MODULE ${MODULE_NAME} TYPE "OBJECT"
 	MONOLITHIC ${MONOLITHIC_BUILD}
diff --git a/libfreerdp/codec/h264.c b/libfreerdp/codec/h264.c
index 77527a4de..5f8f688ab 100644
--- a/libfreerdp/codec/h264.c
+++ b/libfreerdp/codec/h264.c
@@ -28,39 +28,14 @@
 #include <freerdp/primitives.h>
 #include <freerdp/codec/h264.h>
 
-#ifdef WITH_LIBAVCODEC
-int h264_prepare_rgb_buffer(H264_CONTEXT* h264, int width, int height)
-{
-	UINT32 size;
+#include <sys/time.h>
 
-	h264->width = width;
-	h264->height = height;
-	h264->scanline = h264->width * 4;
-	size = h264->scanline * h264->height;
-
-	if (size > h264->size)
-	{
-		h264->size = size;
-
-		if (!h264->data)
-			h264->data = (BYTE*) _aligned_malloc(h264->size, 16);
-		else
-			h264->data = (BYTE*) _aligned_realloc(h264->data, h264->size, 16);
-	}
-
-	if (!h264->data)
-		return -1;
-
-	return 1;
-}
-#endif
 
 /**
  * Dummy subsystem
  */
 
-static int dummy_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize,
-	BYTE* pDstData, DWORD DstFormat, int nDstStep, int nXDst, int nYDst, int nWidth, int nHeight)
+static int dummy_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize)
 {
 	return -1;
 }
@@ -107,13 +82,9 @@ static void openh264_trace_callback(H264_CONTEXT* h264, int level, const char* m
 
 static int openh264_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize)
 {
-	int srcStep[3];
-	prim_size_t roi;
-	BYTE* pYUVData[3];
 	DECODING_STATE state;
 	SBufferInfo sBufferInfo;
 	SSysMEMBuffer* pSystemBuffer;
-	primitives_t* prims = primitives_get();
 	H264_CONTEXT_OPENH264* sys = (H264_CONTEXT_OPENH264*) h264->pSystemData;
 
 	struct timeval T1,T2;
@@ -147,7 +118,7 @@ static int openh264_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSiz
 	 */
 
 	if (sBufferInfo.iBufferStatus != 1)
-		state = (*sys->pDecoder)->DecodeFrame2(sys->pDecoder, NULL, 0, pYUVData, &sBufferInfo);
+		state = (*sys->pDecoder)->DecodeFrame2(sys->pDecoder, NULL, 0, h264->pYUVData, &sBufferInfo);
 	
 	gettimeofday(&T2,NULL);
 	printf("OpenH264: decoding took: %u sec %u usec\n",(unsigned int)(T2.tv_sec-T1.tv_sec),(unsigned int)(T2.tv_usec-T1.tv_usec));
@@ -164,17 +135,19 @@ static int openh264_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSiz
 	if (state != 0)
 		return -1;
 
-	if (!h264->pYUVData[0] || !h264->pYUVData[1] || !h264->pYUVData[2])
-		return -1;
-
 	if (sBufferInfo.iBufferStatus != 1)
-		return -1;
+		return -2;
 
 	if (pSystemBuffer->iFormat != videoFormatI420)
 		return -1;
 
+	if (!h264->pYUVData[0] || !h264->pYUVData[1] || !h264->pYUVData[2])
+		return -1;
+
 	h264->iStride[0] = pSystemBuffer->iStride[0];
 	h264->iStride[1] = pSystemBuffer->iStride[1];
+	h264->iStride[2] = pSystemBuffer->iStride[1];
+
 	h264->width = pSystemBuffer->iWidth;
 	h264->height = pSystemBuffer->iHeight;
 
@@ -305,16 +278,11 @@ struct _H264_CONTEXT_LIBAVCODEC
 };
 typedef struct _H264_CONTEXT_LIBAVCODEC H264_CONTEXT_LIBAVCODEC;
 
-static int libavcodec_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize,
-	BYTE* pDstData, DWORD DstFormat, int nDstStep, int nXDst, int nYDst, int nWidth, int nHeight)
+static int libavcodec_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize)
 {
 	int status;
-	int srcStep[3];
 	int gotFrame = 0;
 	AVPacket packet;
-	prim_size_t roi;
-	const BYTE* pSrc[3];
-	primitives_t* prims = primitives_get();
 	H264_CONTEXT_LIBAVCODEC* sys = (H264_CONTEXT_LIBAVCODEC*) h264->pSystemData;
 
 	struct timeval T1,T2;
@@ -346,22 +314,19 @@ static int libavcodec_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcS
 
 	if (gotFrame)
 	{
-		if (h264_prepare_rgb_buffer(h264, sys->videoFrame->width, sys->videoFrame->height) < 0)
-			return -1;
+		h264->pYUVData[0] = sys->videoFrame->data[0];
+		h264->pYUVData[1] = sys->videoFrame->data[1];
+		h264->pYUVData[2] = sys->videoFrame->data[2];
 
-		roi.width = h264->width;
-		roi.height = h264->height;
+		h264->iStride[0] = sys->videoFrame->linesize[0];
+		h264->iStride[1] = sys->videoFrame->linesize[1];
+		h264->iStride[2] = sys->videoFrame->linesize[2];
 
-		pSrc[0] = sys->videoFrame->data[0];
-		pSrc[1] = sys->videoFrame->data[1];
-		pSrc[2] = sys->videoFrame->data[2];
-
-		srcStep[0] = sys->videoFrame->linesize[0];
-		srcStep[1] = sys->videoFrame->linesize[1];
-		srcStep[2] = sys->videoFrame->linesize[2];
-
-		prims->YUV420ToRGB_8u_P3AC4R(pSrc, srcStep, h264->data, h264->scanline, &roi);
+		h264->width = sys->videoFrame->width;
+		h264->height = sys->videoFrame->height;
 	}
+	else
+		return -2;
 
 	return 1;
 }
@@ -482,6 +447,8 @@ int h264_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize,
 	int* iStride;
 	int ret, i, cx, cy;
 	int UncompressedSize;
+	primitives_t *prims = primitives_get();
+	prim_size_t roi;
 	
 	struct timeval T1,T2;
 
@@ -489,24 +456,24 @@ int h264_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize,
 		return -1;
 
 #if 0
-	printf("h264_decompress: pSrcData=%p, SrcSize=%u, pDstData=%p, nDstStep=%d, nXDst=%d, nYDst=%d, nWidth=%d, nHeight=%d)\n",
-		pSrcData, SrcSize, *ppDstData, nDstStep, nXDst, nYDst, nWidth, nHeight);
+	printf("h264_decompress: pSrcData=%p, SrcSize=%u, pDstData=%p, nDstStep=%d, nDstHeight=%d, numRegionRects=%d\n",
+		pSrcData, SrcSize, *ppDstData, nDstStep, nDstHeight, numRegionRects);
 #endif
 
 	if (!(pDstData = *ppDstData))
 		return -1;
 
 
-<<<<<<< HEAD
-	if (h264->subsystem->Decompress(h264, pSrcData, SrcSize,
-			pDstData, DstFormat, nDstStep, nXDst, nYDst, nWidth, nHeight))
-		return -1;
+	if ((ret = h264->subsystem->Decompress(h264, pSrcData, SrcSize)) < 0)
+		return ret;
 
 
 	UncompressedSize = h264->width * h264->height * 4;
 	if (UncompressedSize > (nDstStep * nDstHeight))
 		return -1;
 
+	pYUVData = h264->pYUVData;
+	iStride = h264->iStride;
 
 	gettimeofday(&T1,NULL);
 	for (i = 0; i < numRegionRects; i++){
@@ -517,32 +484,18 @@ int h264_decompress(H264_CONTEXT* h264, BYTE* pSrcData, UINT32 SrcSize,
 		pDstPoint = pDstData + rect->top * nDstStep + rect->left * 4;
 		pYUVPoint[0] = pYUVData[0] + rect->top * iStride[0] + rect->left;
 
-		ret = rect->top/2 * iStride[1] + rect->left/2;
-		pYUVPoint[1] = pYUVData[1] + ret;
-		pYUVPoint[2] = pYUVData[2] + ret;
+		pYUVPoint[1] = pYUVData[1] + rect->top/2 * iStride[1] + rect->left/2;
+		pYUVPoint[2] = pYUVData[2] + rect->top/2 * iStride[2] + rect->left/2;
 
 #if 0
 		printf("regionRect: x: %d, y: %d, cx: %d, cy: %d\n",
 		       rect->left, rect->top, cx, cy);
 #endif
 
-#ifdef WITH_H264_SSSE3
-		freerdp_image_yuv420p_to_xrgb_ssse3(pDstPoint, pYUVPoint, cx, cy, iStride, nDstStep);
-#else
-/*		roi.width = h264->width;
-		roi.height = h264->height;
+		roi.width = cx;
+		roi.height = cy;
 
-		pSrc[0] = sys->videoFrame->data[0];
-		pSrc[1] = sys->videoFrame->data[1];
-		pSrc[2] = sys->videoFrame->data[2];
-
-		srcStep[0] = sys->videoFrame->linesize[0];
-		srcStep[1] = sys->videoFrame->linesize[1];
-		srcStep[2] = sys->videoFrame->linesize[2];
-
-		prims->YUV420ToRGB_8u_P3AC4R(pSrc, srcStep, h264->data, h264->scanline, &roi)
-		*/
-#endif
+		prims->YUV420ToRGB_8u_P3AC4R((const BYTE**) pYUVPoint, iStride, pDstPoint, nDstStep, &roi);
 	}
 	gettimeofday(&T2,NULL);
 	printf("converting took %u sec %u usec\n",(unsigned int)(T2.tv_sec-T1.tv_sec),(unsigned int)(T2.tv_usec-T1.tv_usec));
@@ -582,24 +535,12 @@ H264_CONTEXT* h264_context_new(BOOL Compressor)
 
 	h264 = (H264_CONTEXT*) calloc(1, sizeof(H264_CONTEXT));
 
-#ifdef WITH_H264_SSSE3
-	if(freerdp_check_ssse3()){
-		printf("SSSE3 seems to be not supported on this system, try without WITH_H264_SSSE3 ...");
-		return NULL;
-	}
-#endif
-
 	if (h264)
 	{
 		h264->Compressor = Compressor;
 
 		h264->subsystem = &g_Subsystem_dummy;
 
-#ifdef WITH_LIBAVCODEC
-		if (h264_prepare_rgb_buffer(h264, 256, 256) < 0)
-			return NULL;
-#endif
-
 		if (!h264_context_init(h264))
 		{
 			free(h264);
@@ -614,10 +555,6 @@ void h264_context_free(H264_CONTEXT* h264)
 {
 	if (h264)
 	{
-#ifdef WITH_LIBAVCODEC
-		_aligned_free(h264->data);
-#endif
-
 		h264->subsystem->Uninit(h264);
 
 		free(h264);
diff --git a/libfreerdp/codec/test/Makefile.TestOpenH264ASM32 b/libfreerdp/codec/test/Makefile.TestOpenH264ASM32
deleted file mode 100644
index 2a0308db4..000000000
--- a/libfreerdp/codec/test/Makefile.TestOpenH264ASM32
+++ /dev/null
@@ -1,17 +0,0 @@
-TestOpenH264ASM: TestOpenH264ASM.c.o h264.c.o h264_ssse3.asm.o h264.asm.o
-	gcc -o TestOpenH264ASM TestOpenH264ASM.c.o h264.c.o h264_ssse3.asm.o h264.asm.o -lwinpr
-
-h264_ssse3.asm.o: ../h264_ssse3_x32.asm
-	nasm -f elf32 -o h264_ssse3.asm.o ../h264_ssse3_x32.asm
-	
-h264.asm.o: ../h264_x32.asm
-	nasm -f elf32 -o h264.asm.o ../h264_x32.asm
-
-TestOpenH264ASM.c.o: TestOpenH264ASM.c
-	gcc -c -o TestOpenH264ASM.c.o TestOpenH264ASM.c
-
-h264.c.o: ../h264.c
-	gcc -c -o h264.c.o ../h264.c
-
-clean:
-	rm -f TestOpenH264ASM TestOpenH264ASM.c.o h264_ssse3.asm.o h264.c.o h264.asm.o
diff --git a/libfreerdp/codec/test/Makefile.TestOpenH264ASM64 b/libfreerdp/codec/test/Makefile.TestOpenH264ASM64
deleted file mode 100644
index 53e208b69..000000000
--- a/libfreerdp/codec/test/Makefile.TestOpenH264ASM64
+++ /dev/null
@@ -1,17 +0,0 @@
-TestOpenH264ASM: TestOpenH264ASM.c.o h264.c.o h264_ssse3.asm.o h264.asm.o
-	gcc -o TestOpenH264ASM TestOpenH264ASM.c.o h264.c.o h264_ssse3.asm.o h264.asm.o -lwinpr
-
-h264_ssse3.asm.o: ../h264_ssse3_x64.asm
-	nasm -f elf64 -o h264_ssse3.asm.o ../h264_ssse3_x64.asm
-	
-h264.asm.o: ../h264_x64.asm
-	nasm -f elf64 -o h264.asm.o ../h264_x64.asm
-
-TestOpenH264ASM.c.o: TestOpenH264ASM.c
-	gcc -c -o TestOpenH264ASM.c.o TestOpenH264ASM.c
-
-h264.c.o: ../h264.c
-	gcc -c -o h264.c.o ../h264.c
-
-clean:
-	rm -f TestOpenH264ASM TestOpenH264ASM.c.o h264_ssse3.asm.o h264.c.o h264.asm.o
diff --git a/libfreerdp/codec/test/Makefile.TestOpenH264SSSE3 b/libfreerdp/codec/test/Makefile.TestOpenH264SSSE3
deleted file mode 100644
index 7709e9423..000000000
--- a/libfreerdp/codec/test/Makefile.TestOpenH264SSSE3
+++ /dev/null
@@ -1,14 +0,0 @@
-TestOpenH264: TestOpenH264ASM.c.o h264.c.o h264_ssse3.c.o
-	gcc -o TestOpenH264 TestOpenH264ASM.c.o h264.c.o h264_ssse3.c.o -lwinpr
-
-h264_ssse3.c.o: ../h264_ssse3.c
-	gcc -c -O3 -o h264_ssse3.c.o ../h264_ssse3.c -mssse3
-
-TestOpenH264ASM.c.o: TestOpenH264ASM.c
-	gcc -c -o TestOpenH264ASM.c.o TestOpenH264ASM.c
-
-h264.c.o: ../h264.c
-	gcc -c -o h264.c.o ../h264.c
-
-clean:
-	rm -f TestOpenH264 TestOpenH264ASM.c.o h264_ssse3.asm.o h264.c.o h264.asm.o
diff --git a/libfreerdp/codec/test/TestOpenH264 b/libfreerdp/codec/test/TestOpenH264
deleted file mode 100755
index c92bd5af2190f0d681727a24c74e78bfb62ea1c6..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 15584
zcmcIre{@vUoxd{^AVkargc=d$iIbM>mJldlRID>F4_|bm6o}OpoP=Zo(Ik_d9}-$3
zamRS_c8Jnd_fWT<a=M;9w7PB9RzzxH0!ZSIwE=%Lr_gOwU}s3w01;C}W<THi?w!0D
zGJD!T_9pM$`~7@>-0%H;=e>8gs-6{Ai-pNxW#48*4d*#IWZW6+|2iiHR>pFflg(zc
z*hC-`a5*`LVk0Hvm@qBlTp?#eoI`1=0IW_UA>vsEg&d3^J*1fO8w*a6G~&1C7^I9D
zQ#e3+<ZGYc_syx|xQu&+9|O`Kp+_X?#f4s6=*f7P3?s-`mXFFt*DPUgmemQ0XeUL?
z&F^C|uM1!+PljqK-8I_~=$h^12^l|az)Z$c?>o>VKmXbw{hZ%WzkbQ0`3<%68tR+E
zE%RE+mdsnSs3hnwS;+k_fI%wa9d}o9)3QEhf}*^q<D#}C{@15mweZOm%ciVfws+p~
z@0GvV`tZ+pO$F~dToej%&A>GcmmQZ3Q;?u`z82S2xURvKkLzk&WdCcpGJ~jg4pV<1
zA7mgO^CW@zN?cUlNw}yy<kJ<nC{Q`5-ef%zHkocBkX`bF@*}y9ZpFfKoL3M=ev;pN
zKmO+}%hmUX`v2j+S!HLVvtI2ylRx<GXH<l&;aPB~9x}rMWHcV$0>Sa{ku3Q9EOw~u
z3a-qc0H7-~d<bB?d<(MJKbS@SSQh-cEOr(_Ucr?af&j+z=aVe>O5h5v%y0_8cy_+b
zf=|xk=iV&x4`#vN%Yxqp`vvSwva6^H5@VrE;5#z#+q2+S;8W47rMw*Hzk$hVYzpg|
z>>M+W)YBZdGtYJDIK#ZKslKT>KvThjk_F7`-O%W7@&-fIfsohBymvnAUFF+Q9}M{d
z%NweLL0^!0%hyzRYkdI$uc=tx;BWG+sb1gUGkD{3svGJzH2G@1bpfA`l?PV_{58H{
z&>vXgs}6+&zLf!A(AN}FTbQ@5zNtRD1oMI&VqR&lvAUtbUqc0{p+YQSb#)El;6_&G
zZ}v5@=5Q#;Y7CKJD9`|?(bpLCg;-rpgFonF!4QfPV0D`U^&ubI;0uN78-4z|+Um`y
z3o<I|qOrQZi3L$p^-ZBV=JnN9hpHj5J{U9v05#RJJ3QsfmU|bLEG${f>!0Q*t62Of
zHvS|gDWWa>&xmOrvamOyOY<0o@XUHlS5~$X?M|WJQN?4LtLTzJ_6M3PY0i^jRLIlZ
zC`=eE(ccIcOBly<O}Kf^EimEcxwO!Po7Zi{gv&lk8O$}|=6Sf(gq!E~G7~Q6bh6_#
z;k3TUpqX&<1ijLPo7blMOgPP*GE|vx^E|TAgwt9kL$e8=E<waCCY;tr8QM&^dHvHO
zL%H!9w3HS(l#9tWxp>aC3QHC_aJOTxkSrp4BbiAIBoT^c5=~8;=;QP>qNzy}F;3?Z
zO--5T;<S}$YQn@$PG6V_nwl=LozrKDo=9{Xr%w}2-6YY>>5qt}CQVdv`j13YQzlk&
z`Y6%Vgb637-z1uvE>X(q14L7kB@|BYC7PD-L;<H?BAS{m!8rX3qN&Le!<PUQ{fuZD
z9*F@?KT9+<U7`>4f=|lzBM)f$M_S~>@X9siJ+ZD?PNwyA$9U9p@d_u4-vK)#+qY9)
zYtae+d5yxf_E^Z8?BkWb-5mD(6b)6W&D()Ui;U)L`pHe!bqDt#Kdv;gZa=Y-1NPZ%
zgzX_p%6E_2Ag}2MwC>ZlYnDDuKdc?O6o&p@lGWOehNS*SLci^{1l(casg=--KLSVB
z9iH$3Aj`Rf)c#uJw%n^B5V|(y!qm$B19RY=elpp&F7ZRCBp$|X`*t$1T6=V87xLoL
z@#|6VDCu03`iHQo>D_y&Q|_bgvwtI&i1-%N;U3<bTF__qvzHVWAN?ws)O59Lf7>^6
zS^Ts??J=n11{K>!EwZ2e@(~s<hY5%e7}ReKfqpHj4(uiVz9b|yy(eBxoT%ELj@|{d
zM;(F;OFTi6(aNr<I;8E;Iu}`^6+?R;ZEI`Od%rl{nUkj<=)N#X@9iF)U3@_A>mJB0
z?$h6-WXD9P?SWdSW^0exweIdwD=b(JXweE`&C+Y%%{r`*i(M|)$K@)Aw3?%u<xTtU
zX<r=eD2S*-mPcEMlK(cF*`M&ChI`bVF$|0z_4xtRmDZ_l-$!m!*<QYgvUOJO)H=gm
zRC3p=?0Uw^^}bijF!c7eL5;-rVVvmddr)mlss4aeQ7t`gM<JO*%ohW0PG955MWRd7
zr#oESmTM0gdzB=4r`ksuk?rS@k=EHxgEnRRdTzVD1(=SyP&@P^#oc<Z-lIjiAP3v@
zcZ(0|{phPCKw>+DSU1vb9N8{p4n)=WT>4*J`mla@e;YcG(a-9>G|HewVpdeMefKal
z`$1Vix$V1SL~_O`QUfWmUO|zVeK%S`9pI^$zR02K-7b9;*~C)WBri}S#HqWAE;INL
z^_3dZ?8kt!?=Il9VV5^~*@4h<=umxSh$6?{6g0rK$q7KnbU2miKOxh>D`jVR@)7pP
zwn%H7h4S`ac|8Wu3t%I6v`|m{Gb6o+(jOb?7^VMUr28rTmXRKy^lVfcuR!h5Lo{f=
zzwRK7VZHTYr1c`(Jc;uLuX?ogB8{O(Chb8xCI@X%^`aq&2~E+@=zY$P)}fAyc!zqz
zz8l`F7ZJrHu{K>DMd}1kjX(?I=cC-HI#M5d3Uz4di^SWyM|0uG<FMPIo+FX&?hkXL
zm7{u}C1$}8%lV?ebGijO55+o63YA>p?3m5b`q;5_n+v(f88~o`@-0B(EQajl14x{i
z-1i1ADrPGMy<O9P7yk@>^yL${M`s+Hu25GVCg{js1=YNdHb|&4qq}|MFErXXf8Szo
zsV-L&h5|?1$p-Hd73gg3WS7r(B=iIcMb%MCzhp|+A)Ru_X*kqRxxf4#-Jj_rlX_<_
zqUem<4DSYrb7XJXW!~wd5bkUze#*O-3Dea$GLO?NOzX2T3*Swh=seibPRb?v4CN@)
zT^OoI_O7}thi{DKKqcwm#%ZpL%Y12w5bHT&i9Bf!VB~o}<@t}C)lS9Kq7YU`d(;t#
zu)zh<cJ2fXE0{!mr7Ez(P~sKAXogisBkBnCk6YWx{Vj)#+|VAWoO|MvkT$A!HM(q-
zrnh$K;TQGFU3$gyZe88!)+>I}qaNo@N7Vu?ihs2*%Iv>V=V~>*QB|R=dijc~4n)=G
z!FVxRxvTplD;8Nk`nwEfyhk0TG9-#&3`Pnt)#SonR2`01%<WtRPmk;8^zL`R)Yt8x
zOu*0K8SS7fuwtt;myjB{@i(aT%G%oueVWBO)E-nzv{JEP(mbxe%QNjayzN6)T;y<8
z5h^VgdELId5?94xjGG?)9W<(i&+3(N-Ks^@cn)p@7TF5wPWv|65=1L^^088kF>KZ0
z!HXyywHLJ~s`l^IotzwqsKcm~;ak<8*tfkUHv`xgRf)-n<`^^>dTn?Bs)#XsP{Jq{
zvXR5;Id)h*i|Y&<oDLZC4SOsWo(i8vEs@3%OQdy#g-;DywhQJ`O9*SV8XpX%L<A=m
zpM)tWMNH^W&k7<kFuOxN)7E;93Ei98TF*=l&!ZUyQ*{-3voSCHc?5kKIa`r4R(d#D
zIigcv8Xe5RJuds;qeCQ*NoON!O4DDajjjI7T26G*uGX!$-^Zg)4}lRb$ijJhKqV6k
zvD?PXLW0J~!GhL0kYZm2c67qE1qyAwu8taZqwPGC4`@)@_-ffsnMYT`F(c*7wg2h>
zHI{lj+D_S`lNa2o9uJCKU!xvv-Z;2WTDfz&158@WuxpzBIt>66w{WaK_t{@@=Gb4M
zu8kzKzY-HGy!y5QIVYyLPAjpEErHrA*o4`C#oZD;*f0ZssT<LzfmTp-ju?`Aj1f$^
zLXm8A(HQH-ST^jD2dM!~U4x6(9f=!WbjeXE9p)UYN4okVH!lbEi<CL#N@ICXYS5<O
zozirwbd+;6+#9WUTkK&{o#$s~lSx|TQH{5%ihbLE(6%;XKlKh|&|WQQ9i7VDQH;B>
zlQQ;7Zy7tO{aq=ucR_`3u(8Hduam<{Z^bGU9@eqx)GG!&y4ugB#KI1)X^C$g+DAX2
z_R|Cz{{{I#gD}!M%0e^X#b~s0$mr(OyI$knQXLiRS!AdH!&gX8A!!m)VdRa2_~App
ziT2a4$M{QUlF9biUdkghbwB+;i*E(A;3ziKPh*d5{8n;(*LAN;f5%nn(%*NjDc48%
z4}nUL{`zW7|8l<qHJB`O>mEm8d_Hn<>!&NCQydjh>|`S!*(#!I9R>Oaq4{kWtaWz(
zgryUa-Hf&iHX`gq<Wgz|Qg!yNZEbLeFFvl$N}F3h<BnpN9Qn`^ZDsDLFQ@y0)usR5
z6*+9tkEoWI>&T_<4|6O@S8;FWlxN@yEB<|_XPWiV4`~DCDt=pSKkw?iJ86I70W84o
z&IgkfH76=+VjjKLQQ&$F>ps(JlI4-tEp=|Y|2hRuROp{-`YeYB<;gAYK9$pcw8DPR
zsq*fV)(Xo&`%#bXb||hrw6p?$9j;$_0l%$1mL5&t;=tmy#ZgGMW}E;@i7N+@ARdP&
zd_pkTVZ`>?ZpM~(F1_7@`f7i}^*d@74+@;?w!4$0ytoNs9{nx#hOac+T1Vjmj?SZB
z^L6&mjwQZCZHeENbuN1oRdCT2V3tP?<kZ>A*o6OvqX6ff=W#;N?+*aGSZ6;Oud~;m
ztg}~tMKWi(%=^GjPB>-geGWe83m@US0J-5h`=!%prd@6D6zQ{+tF1h8(8#S><kCVh
z+#iM6c%5p0{}OKg|AFwyPl4Beg<Cc91ryNgpe|iN>itiU+Pa4xV<F)usY}Qi9zYzn
zE<t6h0>Mj3+?o`3r(;Q^$8z%@K$WqNU3T|~>ouxFmoft5hj0x+>r|aRFp^7ZKmx<)
zZl4}Rc|NPNpY8_mIkAo%06GXH!EGf#eGKYDPNB92(b9v5?CrmVb@j1N&bjPcE#YQQ
zO&HXPy5;sJM=2MD67q%LCqPoJZZhO1ORD`CRb!9I2jowl!;4FBIB=d^JBMt~aCjJc
zg9CIs6~6x*YUalBNOwUw_AgrGV8P879;tPn@3Gh)vtr&lXZLqen?TIJlPVBES>gkO
zvy%qZS(kmdxXykIb?|4@<KUshBFr0g+absYowM)KS^Myf$nF@*`)4Y|;J{$V_DI5_
zMYlK<k8O>kK(j1$c+e5C#8PLsY{3x!OVZwc6MqKK6TkgaOjVIXWeZ~8U*}ruTIYJe
z<)ul-6wV#2qfla>{TTeEq0!*zr|~$Tb<UeKQDL6WqR|P6fX6+0$T94};K3ZW)`1ym
zsLWG*(xrb@uAkSszp|C<e~7;@ggo0n3C&t?jGmu(^a~#Syj%awm7IPY6=l(G9t@n;
zFbm~s(FR9>CtB+$^dO~pq9Mmz4+>IBF3Y)*&wAW5<7N?0PlI?u{w<}+A5wzh=4O8&
z<g2}b9)i_3Db>M7Wo~IpQEAzI#Y%lh3HkjGDMeu==&PY9jQE>*kmHmIEHUYaHu@Ai
zCl5D-lugw^rRYB6KTk3A1PP6b^3|&=Tx*tVN_9vns=Y-iT6mvw!)+AhG2g?!z~<1#
z`lb!aJ$D)Y&kr>=&)?MG-!OlD&|6envT@;(MQqc0Him0Jh|}EM$W|`SV@0*Fo5vO|
zMw+p^T^^+pkL49trQ-HA)HmW`df*nNzP1KgLyTSt_?uSwYC^XtEd+g2*=$5Lf~?^*
z*`iS6Z-SE{a&eq`%Hxj&(J^@-c}h?$gTXOjZLH0LKP&!h_yd{tSIW}=+F;0h9kZnW
zJ!MJ%Yhmesu#WM9%ZiJ#W8~i0#qo*Y<>AUV#7tzH<yzaV^v;GJXcnAKCc6-O(1rVn
z9!9w^R&PeEApR1u6Y(}I%T<U6vAnh+ehwSKU5NdkCzHL1N4`iV>HYOI{Ol^g1UL(E
zDdM$=S0Zjh+>E#jaR*|C;o6V50P!%<2=bY><p-;nrKP}f?UYHm+hNlI-Yi`AC14XY
z%P&}wUwEheicPs~?Dm<r{@u46bAWO=H{yoDWbzRG+I8g@JZ@cn<wPxFv%O-2HIg!H
zH6YH#kH=yHq+G2bLV0b&H5L2=X1?`zT!8p5;`%lCKT6|oOYz?Z?m*pn$zN&T_B{&;
zP<e)M)u2vaGV@jVxr>W{Yxp@G{7+ElM+uO2tpCOHAv=q}|LHmGqNolfpUT@}<uas3
z{;q}GMYI(j+>OcIZ-@|o3-}kn=VL_LQOEc{0sjQ}PP6_G$N2P1A?JKDd4rk1*5FgS
z9tS@c{Q1Op=NCMfpUXT*jsRDHFUtY3EWhAM>$3d9Cv0lI(wU>?&)qhmJiqkuiSGQe
z$RtmGvn@X-zifGasVjf(vU~-S&{zgbMxcD$s0Rg)yQn&3pn6&hpb)rD;CHr7czmKa
z-?<_(=}GGowoc016#DS|o<f(1Ie{Bo<smbEQ{YoXEZ>jPewKoKFG}zFDah|dt~E&h
zT>&L$3c7zz6;DySP>|)5ZdO^HKpFe|QqoVcVslc&u<PQXaH5mP*h=zHCHQjwTPq46
z-*puVIeJWQ4m5-)3^=NIEbFvO=*#hnXTv<u_Y6GbI(a1VzDWix%Tda?W8X-~{IONw
z<>?agm}faHvSS!doDK=TF!TQ&$?tSBGx}h%lS8LOd|t#8^9=lJBL2FFmxy?Ui0>6~
zoroV6@#7-?7ZLxvh<A(lu!uhp@hK5!`gh0j<+mtv-M;np)lJHxlAB5v&Rbk;Bv=Ly
z+<6O$hXi9K!Oe}K>h*|20VCchQ+Pq`3pBG5>_U7cu4U!(Le(1twV^3ovOZjo&w%P{
zSqV`atMRJ5q;_)?>=|(=U~sVh2C-(R5nk{DzJ_X25UJ*d5G&zt<V!-n7Nq&x`;vg4
zzaTI1ZNv-z>PDY;V=Z(gWoUS-0|C5d=NdBoU=74zy}Gf!23h-I5VsPZi(wIO+*wJD
zzp)Wt3uIR3-%Qi^roJo34YN8yAqYT@Zvz~gUnG4a5;TUS{w%8#6y|dQD4ZI*$@xJd
z3y?7DFSR;3YpKw8aBA$9sjq=WbBfHr%IXBgSd{=`oErN^AXm`RjvF+lcE8k@^OPbc
zmZe$r1<dY7g61fxFXyo;p}$8ksEkz4GJiRr+zX6ErM{dO3&q4J=Rt_@Am<rLKY#?y
zK~i7NJ0UT#-690#JS*j7{1Eswr%77QYh6NL5&D__3z+=?393t}FXz!dp)coAnSW;a
zAIs2ZVqTUNCFgO{r?o0mpI(hpok;s~UF(_*5Fy`)Qg_+z5}>u#tREBeV@&Aty_^(A
zOwU3n$oOfb&HD0t(qXZXtu=o%VbUWT6rROp)|czB=Q_?#`$mtDO#gQRBYj!_Ro6K|
zq5lhg+Si(c^jl)PkTC1Z?=R(dnN&V=$jpB?m}Y%VadL4@;SqZ*BY)X`2A8oI64XXw
zxTN+S^83|H`AnMjuy&Ka+?Vvv0f>+(pGnj23G!d&PhSvF82Bso{{&g;ds1Jn!*QWM
zOvVw!a7g(h+u;;==KSS%P5ln8mXbH9vGT}r3})!deb9i=-z61AEcIkOlA+&GRK+Db
ziVOwmr$LS7FX;<NkS$q$`s{<k4xumSIS%3`^OZFHBBb({N<hR?zbpe6)yZ{;MZXVa
z-e%E0k$<jK5X=mc+sWvnTyjT=p-?5}0s2j{7?<>4mW$Zq<v)ro3Wa#afXU3?kQ)DP
zBk!!v%a}HdcDc@)nKpPL*Uxl32kW8J4DqxLE2m-`51+`Y#>FQwxjvbV*-$HT-ATtM
zGr69m<9ST36X`fsQaS&d7`Ab~=OnfbD}tP_P2#}wQ{VNc<Kuk)$?X(izI2*JdD@2F
zCFesk(}pih<vf>;UzM7_((&m`&P(a|49qiCW{9V4_!3;sBW9)zUuetuAssJF{r*YE
zuVHdLr{iPa&8Fk>*WMV49*EwS&t_fW90QS}ahHSdc+)@-vv67P^?PQyih+|IXNH|`
z3%of4|Bk>jf8TE4_Ghx4VxA%YC_IeIf~PZ?_52)g%C|4W{?EDm<?YoAe7th~F$?~o
zuphHKk;01=erJs4b;>s(@B;!rgV=&!(a#J2<@8|8gEQH#47>mZBs(t@IJt!%2|Etp
z<N2AbUFdsz3;J#E9K((<!|40!@!E^N)3)FwgIrIfd^K=7y^vYY0fEcs22#GA+qt|y
zKLK39$%4%O`D5VIP~N3eo!kU}HiguS9G6zd;=@_+KW4$tXTkHZ2vENCl_UkZCd}kG
zhth7PLSF9oB=aVLAGq4dA+Zf)cLG<WkSV@D3*N}>m;^;~YZm;+9KU>i`)L;WUuVG&
zXTeW$JfA%;%0=rTg>wS`*J(~p^Jj%feGQAE!jbgbLf}--Qwp7&af4vHfltM*TF%3A
zT&(8u={AV+Wx+QCS1uQ4PXR9&r@!swcJf)N@Mpfz{f)q9&v0^tKf6SVp4m>dUXwa^
zAjDapz_*L>EAc-GJM`Ht1)bmVyj0*-B47S&5xl7=IGvA>`zRsKt^-c>Rx#blA^xlt
zsc&;!N{z*<gq^Mof4-YV{vqJwou}BEMgA#aKhw`=xqLo*Tl5pBaOs6Cc7B@$e?!=z
z@02O<XXS|B;rQj}JKoD;haY_~9qC{-flx3M#+$Vo9A3C{`6{oceDxX}^cXwO;SDu<
zaWtX{Cq3jr4in_{*80608vN_48@#n4e<0|s4!7WSO=EL|4==k)mXzMS5SGUg;>S9S
z$2>SH5ejT(#)F~SaAPB#wJ=fqgh!g_8ecGU56*MYTPWA+3V2U%s1VgO)Hk#HEARD&
z{N7b}EMptIB(j1YIhEH6${lWO=17ohd68qYm-M~-i4sro7k9i;R;+SWs9yDMHywZB
zRV31Pc<x!|@_6rAv0}Bl#=FL~%%hS7J^M=eTAL2=V>Ha`zW;7lMfq|@*@@F4M)u^$
z=J3OYALZH3#4zK~iPLF5U6NHACuoclF6r{(FiTA^Y@Dx=M_DqY43D{Ycn3#M^t^O1
z2j^`}y(r7UAUbQ4Zq9t1CY@tAVVvDb=kRko>Cj~danjks+s0;i8bq)1`2w}g-ulMs
z4L)y;zj-rN@uG#L&D2L)0vp!<pC_2oiq6j}dCTuXqt(_odGU_6mfKIy-#Ay4j`2fM
za{Qzpk>af_x`A}IW}~m>AusQQqL2D$Oz{GFz0gl}$l9RayRo{dmJVAP=aZ;O`2$#K
z%qV$Ap@nhti6vfCIA7oFZNhL4^Nt+G(Jt}+R-Bv4C>R}cGR{t=GmL{&>6me<DjgF?
er_$=V$+zjU16`>DUc3?b@heeGe!eT+=>G!h9cxDb

diff --git a/libfreerdp/codec/test/TestOpenH264ASM.c b/libfreerdp/codec/test/TestOpenH264ASM.c
deleted file mode 100644
index 040b1650d..000000000
--- a/libfreerdp/codec/test/TestOpenH264ASM.c
+++ /dev/null
@@ -1,92 +0,0 @@
-#include <stdio.h>
-#include <stdlib.h>
-#include <sys/time.h>
-
-#include <winpr/crt.h>
-
-#include "TestOpenH264ASM.h"
-
-#define WIDTH 1920
-#define HEIGHT 1080
-
-#define SSSE3 1
-
-
-int main(void){
-	int i,j,k;
-	int ret;
-	unsigned char *pDstData_c,*pDstData_asm,*pSrcData[3];
-	int nSrcStep[2];
-	
-#if SSSE3
-	if(freerdp_check_ssse3()){
-		fprintf(stderr,"ssse3 not supported!\n");
-		return EXIT_FAILURE;
-	}
-#endif
-	
-	struct timeval t1,t2,t3;
-	
-	pSrcData[0]=malloc(1984*HEIGHT*sizeof(char));
-	pSrcData[1]=malloc(1984*HEIGHT/4*sizeof(char));
-	pSrcData[2]=malloc(1984*HEIGHT/4*sizeof(char));
-	pDstData_asm=_aligned_malloc(WIDTH*(HEIGHT+1)*4*sizeof(char),16);
-	pDstData_c=malloc(WIDTH*(HEIGHT+1)*4*sizeof(char));
-	
-	memset(pDstData_asm,0xFF,WIDTH*(HEIGHT+1)*4*sizeof(char));
-	memset(pDstData_c,0xFF,WIDTH*(HEIGHT+1)*4*sizeof(char));
-	
-	for(i=0;i<WIDTH*HEIGHT;i++){
-		pSrcData[0][i]=i%255;
-		pSrcData[1][i/4]=pSrcData[0][i];
-		pSrcData[2][i/4]=255-pSrcData[0][i];
-	}
-	
-	nSrcStep[0]=1984;
-	nSrcStep[1]=992;
-	
-	gettimeofday(&t1,NULL);
-#if SSSE3
-		ret=freerdp_image_yuv420p_to_xrgb_ssse3(pDstData_asm,pSrcData,WIDTH,HEIGHT,nSrcStep,WIDTH*4);
-#else
-		ret=freerdp_image_yuv_to_xrgb_asm(pDstData_asm,pSrcData,WIDTH,HEIGHT,nSrcStep,WIDTH*4);
-#endif
-	gettimeofday(&t2,NULL);
-		freerdp_image_copy_yuv420p_to_xrgb(pDstData_c,WIDTH*4,0,0,WIDTH,HEIGHT,pSrcData,nSrcStep,0,0);
-	gettimeofday(&t3,NULL);
-	
-	printf("in asm (0x%08X) it took %u sec %u usec,\nin c %u sec %u usec.\n",ret,(int)(t2.tv_sec-t1.tv_sec),(int)(t2.tv_usec-t1.tv_usec),
-		(int)(t3.tv_sec-t2.tv_sec),(int)(t3.tv_usec-t2.tv_usec));
-	
-	printf("in asm the result was %X %X %X\n in c %X %X %X.\n",pDstData_asm[0],pDstData_asm[1],pDstData_asm[2],
-		pDstData_c[0],pDstData_c[1],pDstData_c[2]);
-	
-	/*k=0;
-	for(i=0;i<HEIGHT+1;i++){
-		for(j=0;j<WIDTH;j++){
-			printf("%08X:%08X ",((unsigned int*)pDstData_asm)[k],((unsigned int*)pDstData_c)[k]);
-			k++;
-		}
-		puts("\n");
-	}*/
-	
-	k=1;
-	for(i=0;i<(WIDTH*HEIGHT*4);i++){
-		if(pDstData_c[i]!=pDstData_asm[i]){
-			k=0;
-			printf("MISSMATCH at %d: %2X != %2X\n",i,(unsigned char)pDstData_asm[i],(unsigned char)pDstData_c[i]);
-			break;
-		}
-	}
-	
-	if(k)
-		printf("everything OK\n");
-	
-	free(pSrcData[0]);
-	free(pSrcData[1]);
-	free(pSrcData[2]);
-	free(pDstData_c);
-	_aligned_free(pDstData_asm);
-	
-	return 0;
-}
diff --git a/libfreerdp/codec/test/TestOpenH264ASM.h b/libfreerdp/codec/test/TestOpenH264ASM.h
deleted file mode 100644
index 9125ba524..000000000
--- a/libfreerdp/codec/test/TestOpenH264ASM.h
+++ /dev/null
@@ -1,7 +0,0 @@
-int freerdp_image_copy_yuv420p_to_xrgb(unsigned char* pDstData, int nDstStep, int nXDst, int nYDst,
-		int nWidth, int nHeight, unsigned char* pSrcData[3], int nSrcStep[2], int nXSrc, int nYSrc);
-
-extern int freerdp_image_yuv_to_xrgb_asm(unsigned char *pDstData,unsigned char **pSrcData,int nWidth,int nHeight,int *istride,int scanline);
-
-extern int freerdp_check_ssse3();
-extern int freerdp_image_yuv420p_to_xrgb_ssse3(unsigned char *pDstData,unsigned char **pSrcData,int nWidth,int nHeight,int *istride,int scanline);
\ No newline at end of file
diff --git a/libfreerdp/primitives/CMakeLists.txt b/libfreerdp/primitives/CMakeLists.txt
index 2c4ef7414..6f9d245c2 100644
--- a/libfreerdp/primitives/CMakeLists.txt
+++ b/libfreerdp/primitives/CMakeLists.txt
@@ -40,6 +40,7 @@ set(${MODULE_PREFIX}_OPT_SRCS
 	prim_set_opt.c
 	prim_shift_opt.c
 	prim_sign_opt.c
+	prim_YUV_opt.c
 	prim_YCoCg_opt.c)
 
 add_definitions(-DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE})
@@ -55,11 +56,11 @@ endif()
 
 if(WITH_SSE2)
 	if(CMAKE_COMPILER_IS_GNUCC)
-		set(OPTIMIZATION "${OPTIMIZATION} -msse2 -mssse3 -Wdeclaration-after-statement")
+		set(OPTIMIZATION "${OPTIMIZATION} -msse2 -mssse3 -O2 -Wdeclaration-after-statement")
 	endif()
 
 	if(MSVC)
-		set(OPTIMIZATION "${OPTIMIZATION} /arch:SSE2")
+		set(OPTIMIZATION "${OPTIMIZATION} /arch:SSE2 /O2")
 	endif()
 elseif(WITH_NEON)
 	if(CMAKE_COMPILER_IS_GNUCC)
@@ -70,6 +71,16 @@ endif()
 
 set_property(SOURCE ${${MODULE_PREFIX}_OPT_SRCS} PROPERTY COMPILE_FLAGS ${OPTIMIZATION})
 
+# always compile with optimization
+if(CMAKE_COMPILER_IS_GNUCC)
+	set_property(SOURCE ${${MODULE_PREFIX}_SRCS} PROPERTY COMPILE_FLAGS "-O2")
+endif()
+
+if(MSVC)
+	set_property(SOURCE ${${MODULE_PREFIX}_SRCS} PROPERTY COMPILE_FLAGS "/O2")
+endif()
+
+
 set(${MODULE_PREFIX}_SRCS ${${MODULE_PREFIX}_SRCS} ${${MODULE_PREFIX}_OPT_SRCS})
 
 add_complex_library(MODULE ${MODULE_NAME} TYPE "OBJECT"
diff --git a/libfreerdp/primitives/prim_YUV.c b/libfreerdp/primitives/prim_YUV.c
index c57b122b8..0425c9e8f 100644
--- a/libfreerdp/primitives/prim_YUV.c
+++ b/libfreerdp/primitives/prim_YUV.c
@@ -44,24 +44,40 @@ pstatus_t general_YUV420ToRGB_8u_P3AC4R(const BYTE* pSrc[3], int srcStep[3],
 	int Up48, Up475;
 	int Vp403, Vp120;
 	BYTE* pRGB = pDst;
+	int nWidth, nHeight;
+	int last_line, last_column;
 
 	pY = pSrc[0];
 	pU = pSrc[1];
 	pV = pSrc[2];
+	
+	last_column = roi->width & 0x01;
+	last_line = roi->height & 0x01;
+	
+	nWidth = (roi->width + 1) & ~0x0001;
+	nHeight = (roi->height + 1) & ~0x0001;
 
-	halfWidth = roi->width / 2;
-	halfHeight = roi->height / 2;
+	halfWidth = nWidth / 2;
+	halfHeight = nHeight / 2;
 
-	srcPad[0] = (srcStep[0] - roi->width);
+	srcPad[0] = (srcStep[0] - nWidth);
 	srcPad[1] = (srcStep[1] - halfWidth);
 	srcPad[2] = (srcStep[2] - halfWidth);
 
-	dstPad = (dstStep - (roi->width * 4));
+	dstPad = (dstStep - (nWidth * 4));
 
-	for (y = 0; y < halfHeight; y++)
+	for (y = 0; y < halfHeight; )
 	{
-		for (x = 0; x < halfWidth; x++)
+		y++;
+		if (y == halfHeight)
+			last_line = last_line << 1;
+
+		for (x = 0; x < halfWidth; )
 		{
+			x++;
+			if (x == halfWidth)
+				last_column = last_column << 1;
+
 			U = *pU++;
 			V = *pV++;
 
@@ -105,32 +121,41 @@ pstatus_t general_YUV420ToRGB_8u_P3AC4R(const BYTE* pSrc[3], int srcStep[3],
 
 			/* 2nd pixel */
 
-			Y = *pY++;
-			Yp = Y << 8;
+			if (!(last_column & 0x02))
+			{
+				Y = *pY++;
+				Yp = Y << 8;
 
-			R = (Yp + Vp403) >> 8;
-			G = (Yp - Up48 - Vp120) >> 8;
-			B = (Yp + Up475) >> 8;
+				R = (Yp + Vp403) >> 8;
+				G = (Yp - Up48 - Vp120) >> 8;
+				B = (Yp + Up475) >> 8;
 
-			if (R < 0)
-				R = 0;
-			else if (R > 255)
-				R = 255;
+				if (R < 0)
+					R = 0;
+				else if (R > 255)
+					R = 255;
 
-			if (G < 0)
-				G = 0;
-			else if (G > 255)
-				G = 255;
+				if (G < 0)
+					G = 0;
+				else if (G > 255)
+					G = 255;
 
-			if (B < 0)
-				B = 0;
-			else if (B > 255)
-				B = 255;
+				if (B < 0)
+					B = 0;
+				else if (B > 255)
+					B = 255;
 
-			*pRGB++ = (BYTE) B;
-			*pRGB++ = (BYTE) G;
-			*pRGB++ = (BYTE) R;
-			*pRGB++ = 0xFF;
+				*pRGB++ = (BYTE) B;
+				*pRGB++ = (BYTE) G;
+				*pRGB++ = (BYTE) R;
+				*pRGB++ = 0xFF;
+			}
+			else
+			{
+				pY++;
+				pRGB += 4;
+				last_column = last_column >> 1;
+			}
 		}
 
 		pY += srcPad[0];
@@ -138,8 +163,12 @@ pstatus_t general_YUV420ToRGB_8u_P3AC4R(const BYTE* pSrc[3], int srcStep[3],
 		pV -= halfWidth;
 		pRGB += dstPad;
 
-		for (x = 0; x < halfWidth; x++)
+		for (x = 0; x < halfWidth; )
 		{
+			x++;
+			if (x == halfWidth)
+				last_column = last_column << 1;
+
 			U = *pU++;
 			V = *pV++;
 
@@ -183,32 +212,41 @@ pstatus_t general_YUV420ToRGB_8u_P3AC4R(const BYTE* pSrc[3], int srcStep[3],
 
 			/* 4th pixel */
 
-			Y = *pY++;
-			Yp = Y << 8;
+			if(!(last_column & 0x02))
+			{
+				Y = *pY++;
+				Yp = Y << 8;
 
-			R = (Yp + Vp403) >> 8;
-			G = (Yp - Up48 - Vp120) >> 8;
-			B = (Yp + Up475) >> 8;
+				R = (Yp + Vp403) >> 8;
+				G = (Yp - Up48 - Vp120) >> 8;
+				B = (Yp + Up475) >> 8;
 
-			if (R < 0)
-				R = 0;
-			else if (R > 255)
-				R = 255;
+				if (R < 0)
+					R = 0;
+				else if (R > 255)
+					R = 255;
 
-			if (G < 0)
-				G = 0;
-			else if (G > 255)
-				G = 255;
+				if (G < 0)
+					G = 0;
+				else if (G > 255)
+					G = 255;
 
-			if (B < 0)
-				B = 0;
-			else if (B > 255)
-				B = 255;
+				if (B < 0)
+					B = 0;
+				else if (B > 255)
+					B = 255;
 
-			*pRGB++ = (BYTE) B;
-			*pRGB++ = (BYTE) G;
-			*pRGB++ = (BYTE) R;
-			*pRGB++ = 0xFF;
+				*pRGB++ = (BYTE) B;
+				*pRGB++ = (BYTE) G;
+				*pRGB++ = (BYTE) R;
+				*pRGB++ = 0xFF;
+			}
+			else
+			{
+				pY++;
+				pRGB += 4;
+				last_column = last_column >> 1;
+			}
 		}
 
 		pY += srcPad[0];
@@ -223,6 +261,8 @@ pstatus_t general_YUV420ToRGB_8u_P3AC4R(const BYTE* pSrc[3], int srcStep[3],
 void primitives_init_YUV(primitives_t* prims)
 {
 	prims->YUV420ToRGB_8u_P3AC4R = general_YUV420ToRGB_8u_P3AC4R;
+	
+	primitives_init_YUV_opt(prims);
 }
 
 void primitives_deinit_YUV(primitives_t* prims)
diff --git a/libfreerdp/primitives/prim_YUV.h b/libfreerdp/primitives/prim_YUV.h
index 12f796b61..99428ada6 100644
--- a/libfreerdp/primitives/prim_YUV.h
+++ b/libfreerdp/primitives/prim_YUV.h
@@ -22,6 +22,7 @@
 pstatus_t general_yCbCrToRGB_16s8u_P3AC4R(const INT16* pSrc[3], int srcStep, BYTE* pDst, int dstStep, const prim_size_t* roi);
 
 void primitives_init_YUV(primitives_t* prims);
+void primitives_init_YUV_opt(primitives_t* prims);
 void primitives_deinit_YUV(primitives_t* prims);
 
 #endif /* FREERDP_PRIMITIVES_YUV_H */
diff --git a/libfreerdp/codec/h264_ssse3.c b/libfreerdp/primitives/prim_YUV_opt.c
similarity index 80%
rename from libfreerdp/codec/h264_ssse3.c
rename to libfreerdp/primitives/prim_YUV_opt.c
index 1774856b4..4b5cea145 100644
--- a/libfreerdp/codec/h264_ssse3.c
+++ b/libfreerdp/primitives/prim_YUV_opt.c
@@ -1,32 +1,32 @@
 /** function for converting YUV420p data to the RGB format (but without any special upconverting)
  * It's completely written in nasm-x86-assembly for intel processors supporting SSSE3 and higher.
- * The target scanline (6th parameter) must be a multiple of 16.
- * iStride[0] must be (target scanline) / 4 or bigger and iStride[1] the next multiple of four
- * of the half of iStride[0] or bigger
+ * The target dstStep (6th parameter) must be a multiple of 16.
+ * srcStep[0] must be (target dstStep) / 4 or bigger and srcStep[1] the next multiple of four
+ * of the half of srcStep[0] or bigger
  */
 
 #include <stdio.h>
 
-#include <emmintrin.h>
-//#include <immintrin.h>
-#include <tmmintrin.h>
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
 
 #include <winpr/sysinfo.h>
 #include <winpr/crt.h>
-
-int freerdp_check_ssse3()
-{
-	if(IsProcessorFeaturePresentEx(PF_EX_SSSE3)&&IsProcessorFeaturePresent(PF_SSE3_INSTRUCTIONS_AVAILABLE))
-		return 0;
-	
-	return 1;
-}
+#include <freerdp/types.h>
+#include <freerdp/primitives.h>
 
 
-int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidth,int nHeight,int *iStride,int scanline)
+#ifdef WITH_SSE2
+
+#include <emmintrin.h>
+#include <tmmintrin.h>
+
+pstatus_t ssse3_YUV420ToRGB_8u_P3AC4R(const BYTE **pSrc, int *srcStep,
+		BYTE *pDst, int dstStep, const prim_size_t *roi)
 {
 	char last_line,last_column;
-	int i,VaddDst,VaddY,VaddUV;
+	int i,nWidth,nHeight,VaddDst,VaddY,VaddU,VaddV;
 	
 	BYTE *UData,*VData,*YData;
 	
@@ -37,9 +37,12 @@ int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidt
 	buffer=_aligned_malloc(4*16,16);
 	
 	
-	YData=pSrcData[0];
-	UData=pSrcData[1];
-	VData=pSrcData[2];
+	YData=(BYTE *)pSrc[0];
+	UData=(BYTE *)pSrc[1];
+	VData=(BYTE *)pSrc[2];
+	
+	nWidth=roi->width;
+	nHeight=roi->height;
 	
 	
 	if((last_column=nWidth&3)){
@@ -48,7 +51,7 @@ int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidt
 			case 2: r7=_mm_set_epi32(0,0,0xFFFFFFFF,0xFFFFFFFF); break;
 			case 3: r7=_mm_set_epi32(0,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF); break;
 		}
-		_mm_store_si128(buffer+48,r7);
+		_mm_store_si128(buffer+3,r7);
 		last_column=1;
 	}
 	
@@ -61,10 +64,10 @@ int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidt
 	nHeight=nHeight>>1;
 	
 	
-	VaddDst=(scanline<<1)-(nWidth<<4);
-	VaddY=(iStride[0]<<1)-(nWidth<<2);
-	VaddUV=iStride[1]-(((nWidth<<1)+2)&0xFFFC);
-	
+	VaddDst=(dstStep<<1)-(nWidth<<4);
+	VaddY=(srcStep[0]<<1)-(nWidth<<2);
+	VaddU=srcStep[1]-(((nWidth<<1)+2)&0xFFFC);
+	VaddV=srcStep[2]-(((nWidth<<1)+2)&0xFFFC);
 	
 	
 	while(nHeight-- >0){
@@ -129,7 +132,7 @@ int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidt
 				r1=_mm_add_epi32(r1,r6);
 				r7=_mm_add_epi32(r7,r6);
 				
-				_mm_store_si128(buffer+16,r7);
+				_mm_store_si128(buffer+1,r7);
 				
 /* Now we've prepared U-data. Preparing V-data is actually the same, just with other coefficients */
 				r2=_mm_cvtsi32_si128(*(UINT32 *)VData);
@@ -153,7 +156,7 @@ int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidt
 				r2=_mm_add_epi32(r2,r6);
 				r7=_mm_add_epi32(r7,r6);
 				
-				_mm_store_si128(buffer+32,r7);
+				_mm_store_si128(buffer+2,r7);
 				
 				
 				
@@ -170,8 +173,8 @@ int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidt
 				
 				_mm_store_si128(buffer,r4);
 			}else{
-				r1=_mm_load_si128(buffer+16);
-				r2=_mm_load_si128(buffer+32);
+				r1=_mm_load_si128(buffer+1);
+				r2=_mm_load_si128(buffer+2);
 				r0=_mm_load_si128(buffer);
 			}
 			
@@ -220,17 +223,17 @@ int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidt
 			
 			
 			if(last_column&0x02){
-				r6=_mm_load_si128(buffer+48);
+				r6=_mm_load_si128(buffer+3);
 				r4=_mm_and_si128(r4,r6);
-				r5=_mm_lddqu_si128((__m128i *)pDstData);
+				r5=_mm_lddqu_si128((__m128i *)pDst);
 				r6=_mm_andnot_si128(r6,r5);
 				r4=_mm_or_si128(r4,r6);
 			}
-			_mm_storeu_si128((__m128i *)pDstData,r4);
+			_mm_storeu_si128((__m128i *)pDst,r4);
 			
 			//Y data processing in secound line
 			if(!(last_line&0x02)){
-				r4=_mm_cvtsi32_si128(*(UINT32 *)(YData+iStride[0]));
+				r4=_mm_cvtsi32_si128(*(UINT32 *)(YData+srcStep[0]));
 				r7=_mm_set_epi32(0x80800380,0x80800280,0x80800180,0x80800080);
 				r4=_mm_shuffle_epi8(r4,r7);
 				
@@ -271,28 +274,40 @@ int freerdp_image_yuv420p_to_xrgb_ssse3(BYTE *pDstData,BYTE **pSrcData,int nWidt
 				
 				
 				if(last_column&0x02){
-					r6=_mm_load_si128(buffer+48);
+					r6=_mm_load_si128(buffer+3);
 					r4=_mm_and_si128(r4,r6);
-					r5=_mm_lddqu_si128((__m128i *)(pDstData+scanline));
+					r5=_mm_lddqu_si128((__m128i *)(pDst+dstStep));
 					r6=_mm_andnot_si128(r6,r5);
 					r4=_mm_or_si128(r4,r6);
 					
 					last_column=last_column>>1;
 				}
-				_mm_storeu_si128((__m128i *)(pDstData+scanline),r4);
+				_mm_storeu_si128((__m128i *)(pDst+dstStep),r4);
 			}
 			
-			pDstData+=16;
+			pDst+=16;
 			YData+=4;
 			
 		}while(i<nWidth);
 		
-		pDstData+=VaddDst;
+		pDst+=VaddDst;
 		YData+=VaddY;
-		UData+=VaddUV;
-		VData+=VaddUV;
+		UData+=VaddU;
+		VData+=VaddV;
 	}
 		
 	_aligned_free(buffer);
-	return 0;
-}
\ No newline at end of file
+	
+	return PRIMITIVES_SUCCESS;
+}
+#endif
+
+void primitives_init_YUV_opt(primitives_t *prims)
+{
+#ifdef WITH_SSE2
+	if(IsProcessorFeaturePresentEx(PF_EX_SSSE3)&&IsProcessorFeaturePresent(PF_SSE3_INSTRUCTIONS_AVAILABLE))
+	{
+		prims->YUV420ToRGB_8u_P3AC4R=ssse3_YUV420ToRGB_8u_P3AC4R;
+	}
+#endif
+}
diff --git a/winpr/libwinpr/utils/collections/StreamPool.c b/winpr/libwinpr/utils/collections/StreamPool.c
index 696ecd971..c95875fbe 100644
--- a/winpr/libwinpr/utils/collections/StreamPool.c
+++ b/winpr/libwinpr/utils/collections/StreamPool.c
@@ -155,8 +155,6 @@ wStream* StreamPool_Take(wStreamPool* pool, size_t size)
 
 		Stream_SetPosition(s, 0);
 		Stream_EnsureCapacity(s, size);
-
-		Stream_SetLength(s,size);
 	}
 
 	s->pool = pool;