diff --git a/desmume/src/gfx3d.cpp b/desmume/src/gfx3d.cpp
index 9758915cf..89c311339 100644
--- a/desmume/src/gfx3d.cpp
+++ b/desmume/src/gfx3d.cpp
@@ -2454,7 +2454,7 @@ void gfx3d_savestate(std::ostream* os)
 	for(int i=0;i<4;i++)
 	{
 		OSWRITE(mtxStack[i].position);
-		for(int j=0;j<mtxStack[i].size*16+16;j++)
+		for(int j=0;j<mtxStack[i].size*16;j++)
 			OSWRITE(mtxStack[i].matrix[j]);
 	}
 }
@@ -2493,7 +2493,7 @@ bool gfx3d_loadstate(std::istream* is, int size)
 		for(int i=0;i<4;i++)
 		{
 			OSREAD(mtxStack[i].position);
-			for(int j=0;j<mtxStack[i].size*16+16;j++)
+			for(int j=0;j<mtxStack[i].size*16;j++)
 				OSREAD(mtxStack[i].matrix[j]);
 		}
 	}
diff --git a/desmume/src/wifi.cpp b/desmume/src/wifi.cpp
index 376994be6..be838f524 100644
--- a/desmume/src/wifi.cpp
+++ b/desmume/src/wifi.cpp
@@ -1362,8 +1362,6 @@ static pcap_if_t * WIFI_index_device(pcap_if_t *alldevs, int index) {
 
 bool SoftAP_Init()
 {
-	char errbuf[PCAP_ERRBUF_SIZE];
-	pcap_if_t *alldevs;
 
 	wifiMac.SoftAP.usecCounter = 0;
 
@@ -1372,6 +1370,9 @@ bool SoftAP_Init()
 	wifiMac.SoftAP.curPacketSending = FALSE;
 	
 #if 0
+	char errbuf[PCAP_ERRBUF_SIZE];
+	pcap_if_t *alldevs;
+
 	if(wifi_netEnabled)
 	{
 		if(desmume_pcap_findalldevs(&alldevs, errbuf) == -1)
diff --git a/desmume/src/windows/filter/scanline.cpp b/desmume/src/windows/filter/scanline.cpp
index 1dc6db3e0..38ee8bbab 100644
--- a/desmume/src/windows/filter/scanline.cpp
+++ b/desmume/src/windows/filter/scanline.cpp
@@ -27,6 +27,7 @@ typedef u64 uint64;
 extern CACHE_ALIGN u16 fadeOutColors[17][0x8000];
 
 extern int scanline_filter_a, scanline_filter_b;
+static int fac_a, fac_b;
 
 FORCEINLINE void ScanLine16( uint16 *lpDst, uint16 *lpSrc, unsigned int Width){
 	while(Width--){
@@ -42,6 +43,38 @@ FORCEINLINE void ScanLine16_2( uint16 *lpDst, uint16 *lpSrc, unsigned int Width)
 	}
 }
 
+FORCEINLINE void ScanLine32( uint32 *lpDst, uint32 *lpSrc, unsigned int Width){
+	while(Width--){
+		*lpDst++ = *lpSrc;
+		u8* u8dst = (u8*)lpDst;
+		u8* u8src = (u8*)lpSrc;
+		*u8dst++ = *u8src++ * fac_a / 16;
+		*u8dst++ = *u8src++ * fac_a / 16;
+		*u8dst++ = *u8src++ * fac_a / 16;
+		lpDst++; 
+		lpSrc++;
+	}
+}
+
+FORCEINLINE void ScanLine32_2( uint32 *lpDst, uint32 *lpSrc, unsigned int Width){
+	while(Width--){
+		u8* u8dst = (u8*)lpDst;
+		u8* u8src = (u8*)lpSrc;
+		*u8dst++ = *u8src++ * fac_a / 16;
+		*u8dst++ = *u8src++ * fac_a / 16;
+		*u8dst++ = *u8src++ * fac_a / 16;
+		u8dst++;
+		u8src = (u8*)lpSrc;
+		*u8dst++ = *u8src++ * fac_b / 16;
+		*u8dst++ = *u8src++ * fac_b / 16;
+		*u8dst++ = *u8src++ * fac_b / 16;
+		u8dst++; u8src++;
+		lpDst+=2; 
+		lpSrc++;
+	}
+}
+
+
 FORCEINLINE void DoubleLine32( uint32 *lpDst, uint32 *lpSrc, unsigned int Width){
 	while(Width--){
 		*lpDst++ = *lpSrc;
@@ -51,19 +84,20 @@ FORCEINLINE void DoubleLine32( uint32 *lpDst, uint32 *lpSrc, unsigned int Width)
 
 void RenderScanline( SSurface Src, SSurface Dst)
 {
-	uint16 *lpSrc;
+	fac_a = (16-scanline_filter_a);
+	fac_b = (16-scanline_filter_b);
 	unsigned int H;
 
 	const uint32 srcHeight = Src.Height;
 
 	const unsigned int srcPitch = Src.Pitch >> 1;
-	lpSrc = reinterpret_cast<uint16 *>(Src.Surface);
+	u32* lpSrc = (u32*)Src.Surface;
 
 	const unsigned int dstPitch = Dst.Pitch >> 1;
-	uint16 *lpDst = (uint16*)Dst.Surface;
+	u32 *lpDst = (u32*)Dst.Surface;
 	for (H = 0; H < srcHeight; H++, lpSrc += srcPitch)
-		ScanLine16 (lpDst, lpSrc, Src.Width), lpDst += dstPitch,
-		ScanLine16_2 (lpDst, lpSrc, Src.Width), lpDst += dstPitch;
+		ScanLine32 (lpDst, lpSrc, Src.Width), lpDst += dstPitch,
+		ScanLine32_2 (lpDst, lpSrc, Src.Width), lpDst += dstPitch;
 		//memset (lpDst, 0, 512*2), lpDst += dstPitch;
 }