diff --git a/src/win32/Direct3D.cpp b/src/win32/Direct3D.cpp
index 115e6e65..be3dd016 100644
--- a/src/win32/Direct3D.cpp
+++ b/src/win32/Direct3D.cpp
@@ -353,25 +353,43 @@ void Direct3DDisplay::render()
 		DXTRACE_ERR_MSGBOX( _T("Can not lock texture"), hr );
 		return;
 	} else {
-		if( !theApp.filterFunction ) {
-			copyImage(
-				pix,
-				lr.pBits,
-				(systemColorDepth == 32) ? theApp.sizeX
-				: theApp.sizeX + 1, // TODO: workaround results in one pixel black border at right side
-				theApp.sizeY,
-				lr.Pitch,
-				systemColorDepth
-				);
-		} else {
-			u32 pitch = theApp.filterWidth * (systemColorDepth>>3) + 4;
-			theApp.filterFunction( pix + pitch,
+		unsigned short pitch = theApp.sizeX * ( systemColorDepth >> 3 ) + 4;
+		if( theApp.filterFunction ) {
+			// pixel filter enabled
+			theApp.filterFunction(
+				pix + pitch,
 				pitch,
 				(u8*)theApp.delta,
 				(u8*)lr.pBits,
 				lr.Pitch,
-				theApp.filterWidth,
-				theApp.filterHeight);
+				theApp.sizeX,
+				theApp.sizeY
+				);
+		} else {
+			// pixel filter disabled
+			switch( systemColorDepth )
+			{
+			case 32:
+				cpyImg32(
+					(unsigned char *)lr.pBits,
+					lr.Pitch,
+					pix + pitch,
+					pitch,
+					theApp.sizeX,
+					theApp.sizeY
+					);
+				break;
+			case 16:
+				cpyImg16(
+					(unsigned char *)lr.pBits,
+					lr.Pitch,
+					pix + pitch,
+					pitch,
+					theApp.sizeX,
+					theApp.sizeY
+					);
+				break;
+			}
 		}
 		emulatedImage->UnlockRect( 0 );
 	}
diff --git a/src/win32/DirectDraw.cpp b/src/win32/DirectDraw.cpp
index 3789d4e8..d97f351d 100644
--- a/src/win32/DirectDraw.cpp
+++ b/src/win32/DirectDraw.cpp
@@ -569,7 +569,7 @@ void DirectDrawDisplay::render()
     if(hret == DD_OK) {
       hret = ddsOffscreen->Restore();
 
-      if(hret == DD_OK) {
+	  if(hret == DD_OK) {
         hret = ddsOffscreen->Lock(NULL,
                                   &ddsDesc,
                                   DDLOCK_WRITEONLY|
@@ -580,46 +580,49 @@ void DirectDrawDisplay::render()
                                   NULL);
 
       }
-    }
+	}
   }
 
   if(hret == DD_OK) {
-    if(theApp.filterFunction) {
-      if(systemColorDepth == 16)
-        (*theApp.filterFunction)(pix+theApp.filterWidth*2+4,
-                                 theApp.filterWidth*2+4,
-                                 (u8*)theApp.delta,
-                                 (u8*)ddsDesc.lpSurface,
-                                 ddsDesc.lPitch,
-                                 theApp.filterWidth,
-                                 theApp.filterHeight);
-      else
-        (*theApp.filterFunction)(pix+theApp.filterWidth*4+4,
-                                 theApp.filterWidth*4+4,
-                                 (u8*)theApp.delta,
-                                 (u8*)ddsDesc.lpSurface,
-                                 ddsDesc.lPitch,
-                                 theApp.filterWidth,
-                                 theApp.filterHeight);
-
-    } else {
-      int copyX = 240;
-      int copyY = 160;
-
-      if(theApp.cartridgeType == 1) {
-        if(gbBorderOn) {
-          copyX = 256;
-          copyY = 224;
-        } else {
-          copyX = 160;
-          copyY = 144;
-        }
-      }
-	  if( systemColorDepth == 16 ) {
-		  copyX++; // TODO: workaround results in one pixel black border at right side
+	  unsigned short pitch = theApp.sizeX * ( systemColorDepth >> 3 ) + 4;
+	  if( theApp.filterFunction ) {
+		  // pixel filter enabled
+		  theApp.filterFunction(
+			  pix + pitch,
+			  pitch,
+			  (u8*)theApp.delta,
+			  (u8*)ddsDesc.lpSurface,
+			  ddsDesc.lPitch,
+			  theApp.sizeX,
+			  theApp.sizeY
+			  );
+	  } else {
+		  // pixel filter disabled
+		  switch( systemColorDepth )
+		  {
+		  case 32:
+			  cpyImg32(				  
+				  (unsigned char *)ddsDesc.lpSurface,
+				  ddsDesc.lPitch,
+				  pix + pitch,
+				  pitch,
+				  theApp.sizeX,
+				  theApp.sizeY
+				  );
+			  break;
+		  case 16:
+			  cpyImg16(
+				  (unsigned char *)ddsDesc.lpSurface,
+				  ddsDesc.lPitch,
+				  pix + pitch,
+				  pitch,
+				  theApp.sizeX,
+				  theApp.sizeY
+				  );
+			  break;
+		  }
 	  }
-	  copyImage( pix, ddsDesc.lpSurface, copyX, copyY, ddsDesc.lPitch, systemColorDepth );
-    }
+
     if(theApp.showSpeed && (theApp.videoOption > VIDEO_4X || theApp.skin != NULL)) {
       char buffer[30];
       if(theApp.showSpeed == 1)
diff --git a/src/win32/Display.h b/src/win32/Display.h
index f0b353f4..894623c0 100644
--- a/src/win32/Display.h
+++ b/src/win32/Display.h
@@ -19,6 +19,8 @@
 
 #pragma once
 
+#include <memory.h>
+
 enum DISPLAY_TYPE {
   DIRECT_DRAW = 0,
   DIRECT_3D = 1,
@@ -43,46 +45,27 @@ class IDisplay {
   virtual int selectFullScreenMode(GUID **) = 0;
 };
 
-inline void copyImage( void *source, void *destination, unsigned int width, unsigned int height, unsigned int destinationPitch, unsigned int colorDepth )
+inline void cpyImg32( unsigned char *dst, unsigned int dstPitch, unsigned char *src, unsigned int srcPitch, unsigned short width, unsigned short height )
 {
 	// fast, iterative C version
-	register unsigned int lineSize;
-	register unsigned char *src, *dst;
-	switch(colorDepth)
-	{
-	case 16:
-		lineSize = width<<1;
-		src = ((unsigned char*)source) + lineSize + 4;
-		dst = (unsigned char*)destination;
-		do {
-			MoveMemory( dst, src, lineSize );
-				src+=lineSize;
-				dst+=lineSize;
-			src += 2;
-			dst += (destinationPitch - lineSize);
-		} while ( --height);
-		break;
-	case 32:
-		lineSize = width<<2;
-		src = ((unsigned char*)source) + lineSize + 4;
-		dst = (unsigned char*)destination;
-		do {
-			MoveMemory( dst, src, lineSize );
-				src+=lineSize;
-				dst+=lineSize;
-			src += 4;
-			dst += (destinationPitch - lineSize);
-		} while ( --height);
-		break;
-	}
+	// copies an width*height array of visible pixels from src to dst
+	// srcPitch and dstPitch are the number of garbage bytes after a scanline
+	register unsigned short lineSize = width<<2;
 
-	// compact but slow C version
-	//unsigned int nBytesPerPixel = colorDepth>>3;
-	//unsigned int i, x, y, srcPitch = (width+1) * nBytesPerPixel;
-	//unsigned char * src = ((unsigned char*)source)+srcPitch;
-	//unsigned char * dst = (unsigned char*)destination;
-	//for (y=0;y<height;y++) //Width
-		//for (x=0;x<width;x++) //Height
-			//for (i=0;i<nBytesPerPixel;i++) //Byte# Of Pixel
-				//*(dst+i+(x*nBytesPerPixel)+(y*destinationPitch)) = *(src+i+(x*nBytesPerPixel)+(y*srcPitch));
+	while( height-- ) {
+		memcpy( dst, src, lineSize );
+		src += srcPitch;
+		dst += dstPitch;
+	}
+}
+
+inline void cpyImg16( unsigned char *dst, unsigned int dstPitch, unsigned char *src, unsigned int srcPitch, unsigned short width, unsigned short height )
+{
+	register unsigned short lineSize = width<<1;
+
+	while( height-- ) {
+		memcpy( dst, src, lineSize );
+		src += srcPitch;
+		dst += dstPitch;
+	}
 }