diff -up libcrystalhd-20120405/linux_lib/libcrystalhd/libcrystalhd_int_if.cpp.nosse2 libcrystalhd-20120405/linux_lib/libcrystalhd/libcrystalhd_int_if.cpp
--- libcrystalhd-20120405/linux_lib/libcrystalhd/libcrystalhd_int_if.cpp.nosse2	2011-03-14 21:02:54.000000000 +0100
+++ libcrystalhd-20120405/linux_lib/libcrystalhd/libcrystalhd_int_if.cpp	2012-08-03 20:04:23.700362339 +0200
@@ -33,7 +33,9 @@
 #include "libcrystalhd_int_if.h"
 #include "libcrystalhd_fwcmds.h"
 
+#if __SSE2__ || !defined __GNUC__
 #include <emmintrin.h>
+#endif
 
 #define SV_MAX_LINE_SZ 128
 #define PCI_GLOBAL_CONTROL MISC2_GLOBAL_CTRL
@@ -1425,11 +1427,14 @@ BC_STATUS DtsCopyNV12(DTS_LIB_CONTEXT *C
 }
 
 // TODO: add sse2 detection
+#if __SSE2__
 static bool gSSE2 = true; // most of the platforms will have it anyway:
+#endif
 // 64 bits: no test necessary
 // mac: no test necessary
 // linux/windows: we might have to do the test.
 
+#if __SSE2__
 static void fast_memcpy(uint8_t *dst, const uint8_t *src, uint32_t count)
 {
 	// tested
@@ -1466,6 +1471,9 @@ static void fast_memcpy(uint8_t *dst, co
 	while (count --)
 		*dst++ = *src++;
 }
+#else
+#define fast_memcpy(a,b,c) memcpy(a,b,c)
+#endif
 
 // this is not good.
 // if we have 3 buffers, we cannot assume V is after U
@@ -1504,6 +1512,7 @@ static BC_STATUS DtsCopy422ToUYVY(uint8_
 
 	for (__y = 0; __y < height; __y++)
 	{
+#if __SSE2__
 		if (gSSE2)
 		{
 			if (((((uintptr_t) dstY) & 0xf) == 0) && ((((uintptr_t) srcY) & 0xf) == 0))
@@ -1529,6 +1538,7 @@ static BC_STATUS DtsCopy422ToUYVY(uint8_
 				}
 			}
 		}
+#endif
 
 		while (x < srcWidth-1)
 		{
@@ -1548,6 +1558,7 @@ static BC_STATUS DtsCopy422ToUYVY(uint8_
 // convert to NV12
 static BC_STATUS DtsCopy422ToNV12(uint8_t *dstY, uint8_t *dstUV, const uint8_t *srcY, uint32_t srcWidth, uint32_t dstWidth, uint32_t height, uint32_t strideY, uint32_t strideUV)
 {
+#if __SSE2__
 	// tested
 	uint32_t x, __y;
 
@@ -1667,6 +1678,9 @@ static BC_STATUS DtsCopy422ToNV12(uint8_
 		dstY += strideY;
 	}
 	return BC_STS_SUCCESS;
+#else
+	return BC_STS_INV_ARG;
+#endif
 }
 
 
@@ -1681,6 +1695,7 @@ static BC_STATUS DtsCopy420ToYV12(uint8_
 
 static BC_STATUS DtsCopy420ToYUY2(uint8_t *dstY, uint8_t *dstUV, const uint8_t *srcY, const uint8_t *srcUV, uint32_t srcWidth, uint32_t dstWidth, uint32_t height, uint32_t strideY, uint32_t strideUV)
 {
+#if __SSE2__
 	// TODO, test this
 	uint32_t x, __y;
 
@@ -1836,10 +1851,14 @@ static BC_STATUS DtsCopy420ToYUY2(uint8_
 	}
 
 	return BC_STS_SUCCESS;
+#else
+	return BC_STS_INV_ARG;
+#endif
 }
 
 static BC_STATUS DtsCopy420ToUYVY(uint8_t *dstY, uint8_t *dstUV, const uint8_t *srcY, const uint8_t *srcUV, uint32_t srcWidth, uint32_t dstWidth, uint32_t height, uint32_t strideY, uint32_t strideUV)
 {
+#if __SSE2__
 	// TODO, test this
 	uint32_t x, __y;
 
@@ -1993,6 +2012,9 @@ static BC_STATUS DtsCopy420ToUYVY(uint8_
 	}
 
 	return BC_STS_SUCCESS;
+#else
+	return BC_STS_INV_ARG;
+#endif
 }
 
 static BC_STATUS DtsCopy420ToNV12(uint8_t *dstY, uint8_t *dstUV, const uint8_t *srcY, const uint8_t *srcUV, uint32_t srcWidth, uint32_t dstWidth, uint32_t height, uint32_t strideY, uint32_t strideUV)