From f324b0ef9c0a59dcb691c51057ee9c2ce37460b3 Mon Sep 17 00:00:00 2001
From: Armin Novak <armin.novak@thincast.com>
Date: Tue, 14 Feb 2017 12:12:24 +0100
Subject: [PATCH] Use aligned malloc/free for GFX surfaces/buffers

---
 libfreerdp/codec/h264.c                        | 31 +++++++++++++-------------
 libfreerdp/gdi/gfx.c                           |  4 ++--
 libfreerdp/primitives/test/TestPrimitivesYUV.c | 19 +++++++++++-----
 3 files changed, 31 insertions(+), 23 deletions(-)

diff --git a/libfreerdp/codec/h264.c b/libfreerdp/codec/h264.c
index 270cd3b..4b9d7c4 100644
--- a/libfreerdp/codec/h264.c
+++ b/libfreerdp/codec/h264.c
@@ -1600,17 +1600,17 @@ INT32 avc420_compress(H264_CONTEXT* h264, const BYTE* pSrcData, DWORD SrcFormat,
 	nWidth = (nSrcWidth + 1) & ~1;
 	nHeight = (nSrcHeight + 1) & ~1;
 
-	if (!(pYUVData[0] = (BYTE*) malloc(nWidth * nHeight)))
+	if (!(pYUVData[0] = (BYTE*) _aligned_malloc(nWidth * nHeight, 16)))
 		return -1;
 
 	iStride[0] = nWidth;
 
-	if (!(pYUVData[1] = (BYTE*) malloc(nWidth * nHeight)))
+	if (!(pYUVData[1] = (BYTE*) _aligned_malloc(nWidth * nHeight, 16)))
 		goto error_1;
 
 	iStride[1] = nWidth / 2;
 
-	if (!(pYUVData[2] = (BYTE*) malloc(nWidth * nHeight)))
+	if (!(pYUVData[2] = (BYTE*) _aligned_malloc(nWidth * nHeight, 16)))
 		goto error_2;
 
 	iStride[2] = nWidth / 2;
@@ -1619,13 +1619,13 @@ INT32 avc420_compress(H264_CONTEXT* h264, const BYTE* pSrcData, DWORD SrcFormat,
 	prims->RGBToYUV420_8u_P3AC4R(pSrcData, SrcFormat, nSrcStep, pYUVData, iStride,
 	                             &roi);
 	status = h264->subsystem->Compress(h264, ppDstData, pDstSize, 0);
-	free(pYUVData[2]);
+	_aligned_free(pYUVData[2]);
 	pYUVData[2] = NULL;
 error_2:
-	free(pYUVData[1]);
+	_aligned_free(pYUVData[1]);
 	pYUVData[1] = NULL;
 error_1:
-	free(pYUVData[0]);
+	_aligned_free(pYUVData[0]);
 	pYUVData[0] = NULL;
 	return status;
 }
@@ -1737,15 +1737,14 @@ static BOOL avc444_combine_yuv(H264_CONTEXT* h264,
 	{
 		for (x = 0; x < 3; x++)
 		{
-			BYTE* ppYUVTmpData;
 			piDstStride[x] = piMainStride[0];
 			piDstSize[x] = piDstStride[x] * padDstHeight;
-			ppYUVTmpData = realloc(ppYUVDstData[x], piDstSize[x]);
+			_aligned_free(ppYUVDstData[x]);
+			ppYUVDstData[x] = _aligned_malloc(piDstSize[x], 16);
 
-			if (!ppYUVTmpData)
+			if (!ppYUVDstData[x])
 				goto fail;
 
-			ppYUVDstData[x] = ppYUVTmpData;
 			memset(ppYUVDstData[x], 0, piDstSize[x]);
 		}
 	}
@@ -1766,9 +1765,9 @@ static BOOL avc444_combine_yuv(H264_CONTEXT* h264,
 
 	return TRUE;
 fail:
-	free(ppYUVDstData[0]);
-	free(ppYUVDstData[1]);
-	free(ppYUVDstData[2]);
+	_aligned_free(ppYUVDstData[0]);
+	_aligned_free(ppYUVDstData[1]);
+	_aligned_free(ppYUVDstData[2]);
 	ppYUVDstData[0] = NULL;
 	ppYUVDstData[1] = NULL;
 	ppYUVDstData[2] = NULL;
@@ -1976,9 +1975,9 @@ void h264_context_free(H264_CONTEXT* h264)
 	if (h264)
 	{
 		h264->subsystem->Uninit(h264);
-		free(h264->pYUV444Data[0]);
-		free(h264->pYUV444Data[1]);
-		free(h264->pYUV444Data[2]);
+		_aligned_free(h264->pYUV444Data[0]);
+		_aligned_free(h264->pYUV444Data[1]);
+		_aligned_free(h264->pYUV444Data[2]);
 		free(h264);
 	}
 }
diff --git a/libfreerdp/gdi/gfx.c b/libfreerdp/gdi/gfx.c
index 97c723c..6ea833f 100644
--- a/libfreerdp/gdi/gfx.c
+++ b/libfreerdp/gdi/gfx.c
@@ -700,7 +700,7 @@ static UINT gdi_CreateSurface(RdpgfxClientContext* context,
 	}
 
 	surface->scanline = gfx_align_scanline(surface->width * 4, 16);
-	surface->data = (BYTE*) calloc(1, surface->scanline * surface->height);
+	surface->data = (BYTE*) _aligned_malloc(surface->scanline * surface->height, 16);
 
 	if (!surface->data)
 	{
@@ -731,7 +731,7 @@ static UINT gdi_DeleteSurface(RdpgfxClientContext* context,
 	{
 		region16_uninit(&surface->invalidRegion);
 		codecs = surface->codecs;
-		free(surface->data);
+		_aligned_free(surface->data);
 		free(surface);
 	}
 
diff --git a/libfreerdp/primitives/test/TestPrimitivesYUV.c b/libfreerdp/primitives/test/TestPrimitivesYUV.c
index 6759936..1b79893 100644
--- a/libfreerdp/primitives/test/TestPrimitivesYUV.c
+++ b/libfreerdp/primitives/test/TestPrimitivesYUV.c
@@ -140,18 +140,19 @@ static void* set_padding(size_t size, size_t padding)
 {
 	size_t halfPad = (padding + 1) / 2;
 	BYTE* psrc;
-	BYTE* src = calloc(1, size + 2 * halfPad);
+	BYTE* src = _aligned_malloc(size + 2 * halfPad, 16);
 
 	if (!src)
 		return NULL;
 
 	memset(&src[0], 'A', halfPad);
+	memset(&src[halfPad], 0, size);
 	memset(&src[halfPad + size], 'A', halfPad);
 	psrc = &src[halfPad];
 
 	if (!check_padding(psrc, size, padding, "init"))
 	{
-		free(src);
+		_aligned_free(src);
 		return NULL;
 	}
 
@@ -166,7 +167,7 @@ static void free_padding(void* src, size_t padding)
 		return;
 
 	ptr = ((BYTE*)src) - (padding + 1) / 2;
-	free(ptr);
+	_aligned_free(ptr);
 }
 
 /* Create 2 pseudo YUV420 frames of same size.
@@ -369,7 +370,7 @@ static BOOL TestPrimitiveYUV(primitives_t* prims, prim_size_t roi, BOOL use444)
 	BYTE* rgb_dst = NULL;
 	size_t size;
 	size_t uvsize, uvwidth;
-	size_t padding = 10000;
+	size_t padding = 100 * 16;
 	size_t stride;
 	const UINT32 formats[] =
 	{
@@ -567,7 +568,15 @@ int TestPrimitivesYUV(int argc, char* argv[])
 	for (x = 0; x < 10; x++)
 	{
 		prim_size_t roi;
-		get_size(large, &roi.width, &roi.height);
+
+		if (argc > 1)
+		{
+			roi.width = 1920;
+			roi.height = 1080;
+		}
+		else
+			get_size(large, &roi.width, &roi.height);
+
 		printf("-------------------- GENERIC ------------------------\n");
 
 		if (!TestPrimitiveYUV(generic, roi, TRUE))
-- 
2.7.4