diff --git a/plugins/zzogl-pg/opengl/HostMemory.cpp b/plugins/zzogl-pg/opengl/HostMemory.cpp
index caec95a8fa..a19402ddc4 100644
--- a/plugins/zzogl-pg/opengl/HostMemory.cpp
+++ b/plugins/zzogl-pg/opengl/HostMemory.cpp
@@ -371,7 +371,7 @@
 		}
 	}
 
-__forceinline void _TransferLocalLocal(_writePixel_0 wp, _readPixel_0 rp, u32 widthlimit)
+/*__forceinline void _TransferLocalLocal(_writePixel_0 wp, _readPixel_0 rp, u32 widthlimit)
 {
 	u8* pSrcBuf = g_pbyGSMemory + gs.srcbuf.bp * 256;
 	u8* pDstBuf = g_pbyGSMemory + gs.dstbuf.bp * 256;
@@ -414,6 +414,96 @@ __forceinline void _TransferLocalLocal_4(_getPixelAddress_0 gsp, _getPixelAddres
 
 	assert((gs.imageWnew % 8) == 0);
 		
+	for(int i = gs.trxpos.sy, i2 = gs.trxpos.dy; i < gs.trxpos.sy + gs.imageHnew; ++i, ++i2) 
+	{
+		for(int j = gs.trxpos.sx, j2 = gs.trxpos.dx; j < gs.trxpos.sx + gs.imageWnew; j += 8, j2 += 8) 
+		{
+			// NOTE: the 2 conseq 4bit values are in NOT in the same byte 
+			u32 read = gsp(j%2048, i%2048, gs.srcbuf.bw);
+			u32 write = gdp(j2%2048, i2%2048, gs.dstbuf.bw);
+			pDstBuf[write] = (pDstBuf[write]&0xf0)|(pSrcBuf[read]&0x0f);
+	
+			read = gsp((j+1)%2048, i%2048, gs.srcbuf.bw);
+			write = gdp((j2+1)%2048, i2%2048, gs.dstbuf.bw);
+			pDstBuf[write] = (pDstBuf[write]&0x0f)|(pSrcBuf[read]&0xf0);
+	
+			read = gsp((j+2)%2048, i%2048, gs.srcbuf.bw);
+			write = gdp((j2+2)%2048, i2%2048, gs.dstbuf.bw);
+			pDstBuf[write] = (pDstBuf[write]&0xf0)|(pSrcBuf[read]&0x0f);
+	
+			read = gsp((j+3)%2048, i%2048, gs.srcbuf.bw);
+			write = gdp((j2+3)%2048, i2%2048, gs.dstbuf.bw);
+			pDstBuf[write] = (pDstBuf[write]&0x0f)|(pSrcBuf[read]&0xf0);
+	
+			read = gsp((j+2)%2048, i%2048, gs.srcbuf.bw);
+			write = gdp((j2+2)%2048, i2%2048, gs.dstbuf.bw);
+			pDstBuf[write] = (pDstBuf[write]&0xf0)|(pSrcBuf[read]&0x0f);
+	
+			read = gsp((j+3)%2048, i%2048, gs.srcbuf.bw);
+			write = gdp((j2+3)%2048, i2%2048, gs.dstbuf.bw);
+			pDstBuf[write] = (pDstBuf[write]&0x0f)|(pSrcBuf[read]&0xf0);
+	
+			read = gsp((j+2)%2048, i%2048, gs.srcbuf.bw);
+			write = gdp((j2+2)%2048, i2%2048, gs.dstbuf.bw);
+			pDstBuf[write] = (pDstBuf[write]&0xf0)|(pSrcBuf[read]&0x0f);
+	
+			read = gsp((j+3)%2048, i%2048, gs.srcbuf.bw);
+			write = gdp((j2+3)%2048, i2%2048, gs.dstbuf.bw);
+			pDstBuf[write] = (pDstBuf[write]&0x0f)|(pSrcBuf[read]&0xf0);
+		}
+	}
+}*/
+	
+__forceinline void _TransferLocalLocal()
+{
+	//ZZLog::Error_Log("TransferLocalLocal(0x%x, 0x%x)", gs.srcbuf.psm, gs.dstbuf.psm);
+	_writePixel_0 wp = writePixelFunction_0(gs.srcbuf.psm);
+	_readPixel_0 rp = readPixelFunction_0(gs.dstbuf.psm);
+	u8* pSrcBuf = g_pbyGSMemory + gs.srcbuf.bp * 256;
+	u8* pDstBuf = g_pbyGSMemory + gs.dstbuf.bp * 256;
+	u32 widthlimit = 4;
+	
+	if (PSMT_BITMODE(gs.srcbuf.psm) == 0) widthlimit = 2;
+	if ((gs.imageWnew & widthlimit) != 0) return;
+	
+	for(int i = gs.trxpos.sy, i2 = gs.trxpos.dy; i < gs.trxpos.sy+gs.imageHnew; i++, i2++) 
+	{
+		for(int j = gs.trxpos.sx, j2 = gs.trxpos.dx; j < gs.trxpos.sx+gs.imageWnew; j+=widthlimit, j2+=widthlimit)
+		{
+			wp(pDstBuf, j2%2048, i2%2048,
+					rp(pSrcBuf, j%2048, i%2048, gs.srcbuf.bw), gs.dstbuf.bw);
+				
+			//if (widthlimit > 1) 
+			{
+				wp(pDstBuf, (j2+1)%2048, i2%2048,
+					rp(pSrcBuf, (j+1)%2048, i%2048, gs.srcbuf.bw), gs.dstbuf.bw);
+				
+				if (widthlimit > 2) 
+				{
+					wp(pDstBuf, (j2+2)%2048, i2%2048,
+						rp(pSrcBuf, (j+2)%2048, i%2048, gs.srcbuf.bw), gs.dstbuf.bw);
+				
+					//if (widthlimit > 3) 
+					{
+						wp(pDstBuf, (j2+3)%2048, i2%2048,
+							rp(pSrcBuf, (j+3)%2048, i%2048, gs.srcbuf.bw), gs.dstbuf.bw);
+					}
+				}
+			}
+		}
+	}
+}
+
+__forceinline void _TransferLocalLocal_4()
+{
+	//ZZLog::Error_Log("TransferLocalLocal_4(0x%x, 0x%x)", gs.srcbuf.psm, gs.dstbuf.psm);
+	_getPixelAddress_0 gsp = getPixelFunction_0(gs.srcbuf.psm);
+	_getPixelAddress_0 gdp = getPixelFunction_0(gs.dstbuf.psm);
+	u8* pSrcBuf = g_pbyGSMemory + gs.srcbuf.bp * 256;
+	u8* pDstBuf = g_pbyGSMemory + gs.dstbuf.bp * 256;
+
+	assert((gs.imageWnew % 8) == 0);
+		
 	for(int i = gs.trxpos.sy, i2 = gs.trxpos.dy; i < gs.trxpos.sy + gs.imageHnew; ++i, ++i2) 
 	{
 		for(int j = gs.trxpos.sx, j2 = gs.trxpos.dx; j < gs.trxpos.sx + gs.imageWnew; j += 8, j2 += 8) 
@@ -492,7 +582,15 @@ __forceinline void _TransferLocalLocal_4(_getPixelAddress_0 gsp, _getPixelAddres
 			}
 		}
 	 
-		switch (gs.srcbuf.psm) 
+		if (PSMT_BITMODE(gs.srcbuf.psm) != 4)
+		{
+			_TransferLocalLocal();
+		}
+		else
+		{
+			_TransferLocalLocal_4();
+		}
+		/*switch (gs.srcbuf.psm) 
 		{
 			case PSMCT32:
 				if (gs.dstbuf.psm == PSMCT32)
@@ -695,7 +793,7 @@ __forceinline void _TransferLocalLocal_4(_getPixelAddress_0 gsp, _getPixelAddres
 						break;
 				}
 				break;
-		}
+		}*/
 
 		g_MemTargs.ClearRange(dststart, dstend);
 
diff --git a/plugins/zzogl-pg/opengl/Mem.h b/plugins/zzogl-pg/opengl/Mem.h
index 2cc7627c58..9f7ef546c0 100644
--- a/plugins/zzogl-pg/opengl/Mem.h
+++ b/plugins/zzogl-pg/opengl/Mem.h
@@ -266,6 +266,27 @@ static __forceinline u32 getPixelAddress16SZ_0(int x, int y, u32 bw)
 	return word;
 }
 
+static __forceinline _getPixelAddress_0 getPixelFunction_0(u32 psm)
+{
+	switch(psm)
+	{
+		case PSMCT32: return getPixelAddress32_0;
+		case PSMCT24: return getPixelAddress24_0;
+		case PSMCT16: return getPixelAddress16_0;
+		case PSMCT16S: return getPixelAddress16S_0;
+		case PSMT8: return getPixelAddress8_0;
+		case PSMT4: return getPixelAddress4_0;
+		case PSMT8H: return getPixelAddress8H_0;
+		case PSMT4HL: return getPixelAddress4HL_0;
+		case PSMT4HH: return getPixelAddress4HH_0;
+		case PSMT32Z: return getPixelAddress32Z_0;
+		case PSMT24Z: return getPixelAddress24Z_0;
+		case PSMT16Z: return getPixelAddress16Z_0;
+		case PSMT16SZ: return getPixelAddress16SZ_0;
+		default:  return getPixelAddress32_0;
+	}
+}
+
 #define getPixelAddress_0(psm,x,y,bw) getPixelAddress##psm##_0(x,y,bw)
 #define getPixelAddress(psm,x,y,bp,bw) getPixelAddress##psm##(x,y,bp,bw)
 
@@ -511,6 +532,26 @@ static __forceinline void writePixel16SZ_0(void* pmem, int x, int y, u32 pixel,
 	((u16*)pmem)[getPixelAddress16SZ_0(x, y, bw)] = pixel;
 }
 
+static __forceinline _writePixel_0 writePixelFunction_0(u32 psm)
+{
+	switch(psm)
+	{
+		case PSMCT32: return writePixel32_0;
+		case PSMCT24: return writePixel24_0;
+		case PSMCT16: return writePixel16_0;
+		case PSMCT16S: return writePixel16S_0;
+		case PSMT8: return writePixel8_0;
+		case PSMT4: return writePixel4_0;
+		case PSMT8H: return writePixel8H_0;
+		case PSMT4HL: return writePixel4HL_0;
+		case PSMT4HH: return writePixel4HH_0;
+		case PSMT32Z: return writePixel32Z_0;
+		case PSMT24Z: return writePixel24Z_0;
+		case PSMT16Z: return writePixel16Z_0;
+		case PSMT16SZ: return writePixel16SZ_0;
+		default:  return writePixel32_0;
+	}
+}
 
 ///////////////
 
@@ -589,4 +630,24 @@ static __forceinline u32 readPixel16SZ_0(const void* pmem, int x, int y, u32 bw)
 	return ((const u16*)pmem)[getPixelAddress16SZ_0(x, y, bw)];
 }
 
+static __forceinline _readPixel_0 readPixelFunction_0(u32 psm)
+{
+	switch(psm)
+	{
+		case PSMCT32: return readPixel32_0;
+		case PSMCT24: return readPixel24_0;
+		case PSMCT16: return readPixel16_0;
+		case PSMCT16S: return readPixel16S_0;
+		case PSMT8: return readPixel8_0;
+		case PSMT4: return readPixel4_0;
+		case PSMT8H: return readPixel8H_0;
+		case PSMT4HL: return readPixel4HL_0;
+		case PSMT4HH: return readPixel4HH_0;
+		case PSMT32Z: return readPixel32Z_0;
+		case PSMT24Z: return readPixel24Z_0;
+		case PSMT16Z: return readPixel16Z_0;
+		case PSMT16SZ: return readPixel16SZ_0;
+		default:  return readPixel32_0;
+	}
+}
 #endif /* __MEM_H__ */