diff --git a/source/fceultra/boards/mmc5.cpp b/source/fceultra/boards/mmc5.cpp
index 771be44..125e562 100644
--- a/source/fceultra/boards/mmc5.cpp
+++ b/source/fceultra/boards/mmc5.cpp
@@ -50,7 +50,7 @@ static INLINE void MMC5BGVROM_BANK8(uint32 V) {if(CHRptr[0]){V&=CHRmask8[0];MMC5
 
 static uint8 PRGBanks[4];
 static uint8 WRAMPage;
-static uint8 CHRBanksA[8], CHRBanksB[4];
+static uint16 CHRBanksA[8], CHRBanksB[4];
 static uint8 WRAMMaskEnable[2];
 uint8 mmc5ABMode;                /* A=0, B=1 */
 
@@ -352,7 +352,8 @@ static DECLFW(Mapper5_write)
   if(A>=0x5120&&A<=0x5127)
   {
     mmc5ABMode = 0;
-    CHRBanksA[A&7]=V;
+    CHRBanksA[A&7]=V | ((MMC50x5130&0x3)<<8); //if we had a test case for this then we could test this, but it hasnt been verified
+	//CHRBanksA[A&7]=V;
     MMC5CHRA();
   }
   else switch(A)
@@ -416,6 +417,8 @@ static DECLFW(Mapper5_write)
                  }
                  ATFill=V;
                  break;
+    case 0x5130: MMC50x5130=V;break;
+
     case 0x5200: MMC5HackSPMode=V;break;
     case 0x5201: MMC5HackSPScroll=(V>>3)&0x1F;break;
     case 0x5202: MMC5HackSPPage=V&0x3F;break;
@@ -804,8 +807,8 @@ static void GenMMC5Reset(void)
 
 static SFORMAT MMC5_StateRegs[]={
         { PRGBanks, 4, "PRGB"},
-        { CHRBanksA, 8, "CHRA"},
-        { CHRBanksB, 4, "CHRB"},
+        { CHRBanksA, 16, "CHRA"},
+        { CHRBanksB, 8, "CHRB"},
         { &WRAMPage, 1, "WRMP"},
         { WRAMMaskEnable, 2, "WRME"},
         { &mmc5ABMode, 1, "ABMD"},
@@ -844,6 +847,7 @@ static void GenMMC5_Init(CartInfo *info, int wsize, int battery)
  AddExState(&MMC5HackSPMode, 1, 0, "SPLM");
  AddExState(&MMC5HackSPScroll, 1, 0, "SPLS");
  AddExState(&MMC5HackSPPage, 1, 0, "SPLP");
+ AddExState(&MMC50x5130, 1, 0, "5130");
 
  MMC5WRAMsize=wsize/8;
  BuildWRAMSizeTable();
diff --git a/source/fceultra/cart.cpp b/source/fceultra/cart.cpp
index 316d4ec..1a0f254 100644
--- a/source/fceultra/cart.cpp
+++ b/source/fceultra/cart.cpp
@@ -24,7 +24,7 @@
 #include <string.h>
 #include <stdlib.h>
 #include <stdio.h>
-
+#include <limits.h>
 #include "types.h"
 #include "fceu.h"
 #include "ppu.h"
@@ -293,7 +293,7 @@ void setchr8r(int r, unsigned int V)
 	if(CHRram[r])
 		PPUCHRRAM|=(255);
 	else
-		PPUCHRRAM&=~(255);
+		PPUCHRRAM=0; 
 }
 
 void setchr1(unsigned int A, unsigned int V)
@@ -640,7 +640,7 @@ void FCEU_SaveGameSave(CartInfo *LocalHWInfo)
 		std::string soot = FCEU_MakeFName(FCEUMKF_SAV,0,"sav");
 		if((sp=FCEUD_UTF8fopen(soot,"wb"))==NULL)
 		{
-			FCEU_PrintError("WRAM file \"%s\" cannot be written to.\n",soot);
+			FCEU_PrintError("WRAM file \"%s\" cannot be written to.\n",soot.c_str());
 		}
 		else
 		{
diff --git a/source/fceultra/fceu.h b/source/fceultra/fceu.h
index 0611bc4..83e2fdf 100644
--- a/source/fceultra/fceu.h
+++ b/source/fceultra/fceu.h
@@ -38,6 +38,7 @@ extern int MMC5Hack;
 extern uint8 *MMC5HackVROMPTR;
 extern uint8 MMC5HackCHRMode;
 extern uint8 MMC5HackSPMode;
+extern uint8 MMC50x5130;
 extern uint8 MMC5HackSPScroll;
 extern uint8 MMC5HackSPPage;
 
diff --git a/source/fceultra/file.cpp b/source/fceultra/file.cpp
index 6cc006e..2d9daaf 100644
--- a/source/fceultra/file.cpp
+++ b/source/fceultra/file.cpp
@@ -242,7 +242,7 @@ zpfail:
 		unzCloseCurrentFile(tz);
 		unzClose(tz);
 
-		FCEUFILE *fceufp=fceufp = new FCEUFILE();
+		FCEUFILE *fceufp = new FCEUFILE();
 		fceufp->stream = ms;
 		fceufp->size = size;
 		return fceufp;
@@ -259,7 +259,7 @@ FCEUFILE * FCEU_fopen(const char *path, const char *ipsfn, char *mode, char *ext
 
 	bool read = (std::string)mode == "rb";
 	bool write = (std::string)mode == "wb";
-	if(read&&write || (!read&&!write))
+	if((read&&write) || (!read&&!write))
 	{
 		FCEU_PrintError("invalid file open mode specified (only wb and rb are supported)");
 		return 0;
@@ -764,4 +764,4 @@ void FCEUARCHIVEFILEINFO::FilterByExtension(const char** ext)
 		this->erase(begin()+i);
 	ok: ;
 	}
-}
\ No newline at end of file
+}
diff --git a/source/fceultra/ines.cpp b/source/fceultra/ines.cpp
index 0cd45a8..84d81c6 100644
--- a/source/fceultra/ines.cpp
+++ b/source/fceultra/ines.cpp
@@ -753,6 +753,33 @@ int iNesSave(){
 	return 1;
 }
 
+int iNesSaveAs(char* name)
+{
+	//adelikat: TODO: iNesSave() and this have pretty much the same code, outsource the common code to a single function
+	FILE *fp;
+	
+	if(GameInfo->type != GIT_CART)return 0;
+	if(GameInterface!=iNESGI)return 0;
+
+	fp = fopen(name,"wb");
+	int x = 0;
+	if (!fp)
+		int x = 1;
+	if(fwrite(&head,1,16,fp)!=16)return 0;
+
+	if(head.ROM_type&4) 	/* Trainer */
+	{
+		fwrite(trainerpoo,512,1,fp);
+	}
+
+	fwrite(ROM,0x4000,ROM_size,fp);
+
+	if(head.VROM_size)fwrite(VROM,0x2000,head.VROM_size,fp);
+	fclose(fp);
+
+	return 1;
+}
+
 //para edit: added function below
 char *iNesShortFName() {
 	char *ret;
diff --git a/source/fceultra/ines.h b/source/fceultra/ines.h
index 91e6ac2..516428a 100644
--- a/source/fceultra/ines.h
+++ b/source/fceultra/ines.h
@@ -74,6 +74,7 @@ extern uint8 *VROM;
 extern uint32 VROM_size;
 extern uint32 ROM_size;
 extern int iNesSave(); //bbit Edited: line added
+extern int iNesSaveAs(char* name);
 extern char LoadedRomFName[2048]; //bbit Edited: line added
 
 //mbg merge 7/19/06 changed to c++ decl format
diff --git a/source/fceultra/input.cpp b/source/fceultra/input.cpp
index 15a4946..3fed7b9 100644
--- a/source/fceultra/input.cpp
+++ b/source/fceultra/input.cpp
@@ -202,6 +202,9 @@ static void UpdateGP(int w, void *data, int arg)
 		joy[2]= FCEU_LuaUsingJoypad(2) ? (FCEU_LuaReadJoypad(2) | (*(uint32 *)joyports[0].ptr >> 16)) : *(uint32 *)joyports[0].ptr >> 16;
 		if (FCEU_LuaReadJoypadFalse(2))
 			joy[2] &= FCEU_LuaReadJoypadFalse(2);
+		#else // without this, there seems to be no input at all without Lua
+		joy[0] = *(uint32 *)joyports[0].ptr;;
+		joy[2] = *(uint32 *)joyports[0].ptr >> 16;
 		#endif
 	}
 	else
@@ -215,6 +218,9 @@ static void UpdateGP(int w, void *data, int arg)
 		joy[3]= FCEU_LuaUsingJoypad(3) ? (FCEU_LuaReadJoypad(3) | (*(uint32 *)joyports[1].ptr >> 24)) : *(uint32 *)joyports[1].ptr >> 24;
 		if (FCEU_LuaReadJoypadFalse(3))
 			joy[3] &= FCEU_LuaReadJoypadFalse(3);
+		#else // same goes for the other two pads
+		joy[1] = *(uint32 *)joyports[1].ptr >> 8;
+		joy[3] = *(uint32 *)joyports[1].ptr >> 24;
 		#endif
 	}
 
@@ -609,6 +615,7 @@ static void CloseRom(void);
 static void MovieSubtitleToggle(void);
 static void UndoRedoSavestate(void);
 static void FCEUI_DoExit(void);
+static void ToggleFullscreen(void);
 
 struct EMUCMDTABLE FCEUI_CommandTable[]=
 {
@@ -619,8 +626,6 @@ struct EMUCMDTABLE FCEUI_CommandTable[]=
 	{ EMUCMD_SCREENSHOT,					EMUCMDTYPE_MISC,	FCEUI_SaveSnapshot,	  0, 0, "Screenshot", EMUCMDFLAG_TASEDIT },
 	{ EMUCMD_HIDE_MENU_TOGGLE,				EMUCMDTYPE_MISC,	FCEUD_HideMenuToggle, 0, 0, "Hide Menu Toggle", EMUCMDFLAG_TASEDIT },
 	{ EMUCMD_EXIT,							EMUCMDTYPE_MISC,	FCEUI_DoExit,			  0, 0, "Exit", 0},
-	//adelikat: CaH4e3, perhaps finding the true cause should be on the person who made the change?
-	//Also, removing the windows only function from this table.  This is a core file and should stay compatible with the SDL build
 	{ EMUCMD_SPEED_SLOWEST,					EMUCMDTYPE_SPEED,	CommandEmulationSpeed, 0, 0, "Slowest Speed", 0 },
 	{ EMUCMD_SPEED_SLOWER,					EMUCMDTYPE_SPEED,	CommandEmulationSpeed, 0, 0, "Speed Down", 0 },
 	{ EMUCMD_SPEED_NORMAL,					EMUCMDTYPE_SPEED,	CommandEmulationSpeed, 0, 0, "Normal Speed", 0 },
@@ -671,7 +676,7 @@ struct EMUCMDTABLE FCEUI_CommandTable[]=
 	{ EMUCMD_MOVIE_PLAY_FROM_BEGINNING,		EMUCMDTYPE_MOVIE,	FCEUI_MoviePlayFromBeginning, 0, 0, "Play Movie From Beginning", 0 },
 	{ EMUCMD_MOVIE_STOP,					EMUCMDTYPE_MOVIE,	FCEUI_StopMovie, 0, 0, "Stop Movie", 0 },
 	{ EMUCMD_MOVIE_READONLY_TOGGLE,			EMUCMDTYPE_MOVIE,	FCEUI_MovieToggleReadOnly, 0, 0, "Toggle Read-Only", EMUCMDFLAG_TASEDIT },
-	{ EMUCMD_MOVIE_FRAME_DISPLAY_TOGGLE,	EMUCMDTYPE_MOVIE,	FCEUI_MovieToggleFrameDisplay, 0, 0, "Movie Frame Display Toggle", 0 },
+	{ EMUCMD_MOVIE_FRAME_DISPLAY_TOGGLE,	EMUCMDTYPE_MOVIE,	FCEUI_MovieToggleFrameDisplay, 0, 0, "Frame Display Toggle", 0 },
 
 	{ EMUCMD_MOVIE_INPUT_DISPLAY_TOGGLE,	EMUCMDTYPE_MISC,	FCEUI_ToggleInputDisplay, 0, 0, "Toggle Input Display", 0 },
 	{ EMUCMD_MOVIE_ICON_DISPLAY_TOGGLE,		EMUCMDTYPE_MISC,	FCEUD_ToggleStatusIcon, 0, 0, "Toggle Status Icon", 0 },
@@ -722,15 +727,12 @@ struct EMUCMDTABLE FCEUI_CommandTable[]=
 	{ EMUCMD_OPENROM,						EMUCMDTYPE_TOOL,	OpenRom,		  0, 0,  "Open ROM", 0},
 	{ EMUCMD_CLOSEROM,						EMUCMDTYPE_TOOL,	CloseRom,		  0, 0,	 "Close ROM", 0},
 	{ EMUCMD_MISC_DISPLAY_MOVIESUBTITLES,	EMUCMDTYPE_MISC,	MovieSubtitleToggle,0,0,"Toggle Movie Subtitles", 0},
-	{ EMUCMD_MISC_UNDOREDOSAVESTATE,		EMUCMDTYPE_MISC,	UndoRedoSavestate,  0,0,"Undo/Redo Savestate",    0}
+	{ EMUCMD_MISC_UNDOREDOSAVESTATE,		EMUCMDTYPE_MISC,	UndoRedoSavestate,  0,0,"Undo/Redo Savestate",    0},
+	{ EMUCMD_MISC_TOGGLEFULLSCREEN,			EMUCMDTYPE_MISC,	ToggleFullscreen, 0, 0, "Toggle Fullscreen",	  0}
 };
 
 #define NUM_EMU_CMDS		(sizeof(FCEUI_CommandTable)/sizeof(FCEUI_CommandTable[0]))
 
-// jabberwoocky my son, don't be aware lol
-// this is much mindfucking thing i ever seen here
-// even when i fixed it, there is a lot of possibilities to break all key input stuff with one
-// unarranged command enumerator.
 static int execcmd, i;
 
 void FCEUI_HandleEmuCommands(TestCommandState* testfn)
@@ -949,4 +951,18 @@ static void FCEUI_DoExit(void)
 #ifdef WIN32
 	DoFCEUExit();
 #endif
+}
+
+static void ToggleFullscreen(void)
+{
+#ifdef WIN32
+	extern int SetVideoMode(int fs);		//adelikat: Yeah, I know, hacky
+	extern void UpdateCheckedMenuItems();
+	
+	UpdateCheckedMenuItems();
+	changerecursive=1;
+	if(!SetVideoMode(fullscreen^1))
+		SetVideoMode(fullscreen);
+	changerecursive=0;
+#endif
 }
\ No newline at end of file
diff --git a/source/fceultra/input.h b/source/fceultra/input.h
index dd410e6..0b0779f 100644
--- a/source/fceultra/input.h
+++ b/source/fceultra/input.h
@@ -218,6 +218,7 @@ enum EMUCMD
 	//-----------------------------
 	EMUCMD_MISC_DISPLAY_MOVIESUBTITLES,
 	EMUCMD_MISC_UNDOREDOSAVESTATE,
+	EMUCMD_MISC_TOGGLEFULLSCREEN,
 	EMUCMD_MAX
 };
 
diff --git a/source/fceultra/movie.cpp b/source/fceultra/movie.cpp
index 9541fcb..735ed05 100644
--- a/source/fceultra/movie.cpp
+++ b/source/fceultra/movie.cpp
@@ -34,11 +34,12 @@
 
 #ifdef WIN32
 #include <windows.h>
+extern void AddRecentMovieFile(const char *filename);
 #endif
 
 using namespace std;
 
-#define MOVIE_VERSION           3
+#define MOVIE_VERSION           3 
 
 extern char FileBase[];
 extern bool AutoSS;		//Declared in fceu.cpp, keeps track if a auto-savestate has been made
@@ -105,7 +106,7 @@ void MovieData::clearRecordRange(int start, int len)
 void MovieData::insertEmpty(int at, int frames)
 {
 #ifndef GEKKO
-	if(at == -1)
+	if(at == -1) 
 	{
 		int currcount = records.size();
 		records.resize(records.size()+frames);
@@ -465,7 +466,7 @@ bool FCEUI_GetLagged(void)
 
 bool FCEUMOV_ShouldPause(void)
 {
-	if(pauseframe && currFrameCounter == pauseframe)
+	if(pauseframe && currFrameCounter == (pauseframe-1))	//adelikat: changed to pauseframe -1 to prevent an off by 1 error.  THis is probably the hackiest solution but I think it would cause some major restructuring to fix it properly.
 	{
 		pauseframe = 0; //only pause once!
 		return true;
@@ -529,7 +530,7 @@ static void LoadFM2_binarychunk(MovieData& movieData, std::istream* fp, int size
 }
 
 //yuck... another custom text parser.
-static bool LoadFM2(MovieData& movieData, std::istream* fp, int size, bool stopAfterHeader)
+bool LoadFM2(MovieData& movieData, std::istream* fp, int size, bool stopAfterHeader)
 {
 #ifndef GEKKO
 	//first, look for an fcm signature
@@ -780,11 +781,11 @@ void MovieData::dumpSavestateTo(std::vector<char>* buf, int compressionLevel)
 }
 
 //begin playing an existing movie
-void FCEUI_LoadMovie(const char *fname, bool _read_only, bool tasedit, int _pauseframe)
+bool FCEUI_LoadMovie(const char *fname, bool _read_only, bool tasedit, int _pauseframe)
 {
 #ifndef GEKKO
 	if(!tasedit && !FCEU_IsValidUI(FCEUI_PLAYMOVIE))
-		return;
+		return true;	//adelikat: file did not fail to load, so let's return true here, just do nothing
 
 	assert(fname);
 
@@ -796,17 +797,22 @@ void FCEUI_LoadMovie(const char *fname, bool _read_only, bool tasedit, int _paus
 	//--------------
 
 	currMovieData = MovieData();
-
+	
 	strcpy(curMovieFilename, fname);
 	FCEUFILE *fp = FCEU_fopen(fname,0,"rb",0);
-	if (!fp) return;
+	if (!fp) return false;
 	if(fp->isArchive() && !_read_only) {
 		FCEU_PrintError("Cannot open a movie in read+write from an archive.");
-		return;
+		return true;	//adelikat: file did not fail to load, so return true (false is only for file not exist/unable to open errors
 	}
 
+#ifdef WIN32
+	//Add to the recent movie menu
+	AddRecentMovieFile(fname);
+#endif
+
 	LoadFM2(currMovieData, fp->stream, INT_MAX, false);
-	LoadSubtitles();
+	LoadSubtitles(currMovieData);
 	delete fp;
 
 	freshMovie = true;	//Movie has been loaded, so it must be unaltered
@@ -818,7 +824,7 @@ void FCEUI_LoadMovie(const char *fname, bool _read_only, bool tasedit, int _paus
 	if(currMovieData.savestate.size() != 0)
 	{
 		bool success = MovieData::loadSavestateFrom(&currMovieData.savestate);
-		if(!success) return;
+		if(!success) return true;	//adelikat: I guess return true here?  False is only for a bad movie filename, if it got this far the file was god?
 	}
 
 	//if there is no savestate, we won't have this crucial piece of information at the start of the movie.
@@ -852,7 +858,7 @@ void FCEUI_LoadMovie(const char *fname, bool _read_only, bool tasedit, int _paus
 		else
 			FCEU_DispMessage("Replay started Read+Write.");
 	}
-
+	
 	#ifdef CREATE_AVI
 	if(LoggingEnabled)
 	{
@@ -860,6 +866,9 @@ void FCEUI_LoadMovie(const char *fname, bool _read_only, bool tasedit, int _paus
 	    LoggingEnabled = 2;
 	}
 	#endif
+	return true;
+#else
+	return false;
 #endif
 }
 
@@ -1326,16 +1335,16 @@ bool FCEUI_MovieGetInfo(FCEUFILE* fp, MOVIE_INFO& info, bool skipFrameCount)
 }
 
 //This function creates an array of frame numbers and corresponding strings for displaying subtitles
-void LoadSubtitles(void)
+void LoadSubtitles(MovieData moviedata)
 {
 #ifndef GEKKO
 	extern std::vector<string> subtitles;
-	for(uint32 i=0;i<currMovieData.subtitles.size();i++)
+	for(uint32 i=0; i < moviedata.subtitles.size() ; i++)
 	{
-		std::string& subtitle = currMovieData.subtitles[i];
+		std::string& subtitle = moviedata.subtitles[i];
 		size_t splitat = subtitle.find_first_of(' ');
 		std::string key, value;
-
+		
 		//If we can't split them, then don't process this one
 		if(splitat == std::string::npos)
 			{
@@ -1381,6 +1390,16 @@ void FCEU_DisplaySubtitles(char *format, ...)
 #endif
 }
 
+void FCEUI_CreateMovieFile(std::string fn)
+{
+#ifndef GEKKO
+	MovieData md = currMovieData;							//Get current movie data
+	std::fstream* outf = FCEUD_UTF8_fstream(fn, "wb");		//open/create file
+	md.dump(outf,false);									//dump movie data
+	delete outf;											//clean up, delete file object
+#endif
+}
+
 void FCEUI_MakeBackupMovie(bool dispMessage)
 {
 #ifndef GEKKO
@@ -1411,13 +1430,13 @@ void FCEUI_MakeBackupMovie(bool dispMessage)
 		backupFn.append(".bak");		 //add extension
 
 		exist = CheckFileExists(backupFn.c_str());	//Check if file exists
-
-		if (!exist)
+		
+		if (!exist) 
 			break;						//Yeah yeah, I should use a do loop or something
 		else
 		{
 			backupFn = tempFn;			//Before we loop again, reset the filename
-
+			
 			if (backNum == 999)			//If 999 exists, we have overflowed, let's handle that
 			{
 				backupFn.append("-001.bak"); //We are going to simply overwrite 001.bak
@@ -1426,15 +1445,11 @@ void FCEUI_MakeBackupMovie(bool dispMessage)
 			}
 		}
 	}
-
-	MovieData md = currMovieData;								//Get current movie data
-	std::fstream* outf = FCEUD_UTF8_fstream(backupFn, "wb");	//open/create file
-	md.dump(outf,false);										//dump movie data
-	delete outf;												//clean up, delete file object
-
+	FCEUI_CreateMovieFile(backupFn);
+		
 	//TODO, decide if fstream successfully opened the file and print error message if it doesn't
 
-	if (dispMessage)	//If we should inform the user
+	if (dispMessage)	//If we should inform the user 
 	{
 		if (overflow)
 			FCEUI_DispMessage("Backup overflow, overwriting %s",backupFn.c_str()); //Inform user of overflow
@@ -1448,26 +1463,26 @@ bool CheckFileExists(const char* filename)
 {
 #ifndef GEKKO
 	//This function simply checks to see if the given filename exists
-	string checkFilename;
+	string checkFilename; 
 
 	if (filename)
 		checkFilename = filename;
-
+			
 	//Check if this filename exists
 	fstream test;
 	test.open(checkFilename.c_str(),fstream::in);
-
+		
 	if (test.fail())
 	{
 		test.close();
-		return false;
+		return false; 
 	}
 	else
 	{
 		test.close();
-		return true;
+		return true; 
 	}
 #else
 	return false;
 #endif
-}
+}
\ No newline at end of file
diff --git a/source/fceultra/movie.h b/source/fceultra/movie.h
index 7d84451..9dd5255 100644
--- a/source/fceultra/movie.h
+++ b/source/fceultra/movie.h
@@ -19,7 +19,7 @@ enum EMOVIE_FLAG
 	//an ARCHAIC flag which means the movie was recorded from a soft reset.
 	//WHY would you do this?? do not create any new movies with this flag
 	MOVIE_FLAG_FROM_RESET = (1<<1),
-
+	
 	MOVIE_FLAG_PAL = (1<<2),
 
 	//movie was recorded from poweron. the alternative is from a savestate (or from reset)
@@ -41,7 +41,6 @@ typedef struct
 	uint32 emu_version_used;				// 9813 = 0.98.13
 	MD5DATA md5_of_rom_used;
 	std::string name_of_rom_used;
-
 #ifndef GEKKO
 	std::vector<std::wstring> comments;
 #endif
@@ -93,7 +92,7 @@ class MovieRecord
 
 public:
 	ValueArray<uint8,4> joysticks;
-
+	
 	struct {
 		uint8 x,y,b,bogo;
 		uint64 zaphit;
@@ -133,7 +132,7 @@ public:
 	}
 
 	void clear();
-
+	
 	//a waste of memory in lots of cases..  maybe make it a pointer later?
 	std::vector<char> savestate;
 
@@ -143,7 +142,7 @@ public:
 	void dumpBinary(MovieData* md, std::ostream* os, int index);
 	void parseJoy(std::istream* is, uint8& joystate);
 	void dumpJoy(std::ostream* os, uint8 joystate);
-
+	
 	static const char mnemonics[8];
 
 private:
@@ -154,7 +153,7 @@ class MovieData
 {
 public:
 	MovieData();
-
+	
 
 	int version;
 	int emuVersion;
@@ -179,7 +178,7 @@ public:
 	int ports[3];
 	//whether fourscore is enabled
 	bool fourscore;
-
+	
 	//----TasEdit stuff---
 	int greenZoneCount;
 	//----
@@ -219,7 +218,7 @@ public:
 	int dump(std::ostream* os, bool binary);
 	void clearRecordRange(int start, int len);
 	void insertEmpty(int at, int frames);
-
+	
 	static bool loadSavestateFrom(std::vector<char>* buf);
 	static void dumpSavestateTo(std::vector<char>* buf, int compressionLevel);
 	void TryDumpIncremental();
@@ -246,10 +245,11 @@ extern bool autoMovieBackup;
 //--------------------------------------------------
 bool CheckFileExists(const char* filename);	//Receives a filename (fullpath) and checks to see if that file exists
 void FCEUI_MakeBackupMovie(bool dispMessage);
+void FCEUI_CreateMovieFile(std::string fn);
 #ifndef GEKKO
 void FCEUI_SaveMovie(const char *fname, EMOVIE_FLAG flags, std::wstring author);
 #endif
-void FCEUI_LoadMovie(const char *fname, bool read_only, bool tasedit, int _stopframe);
+bool FCEUI_LoadMovie(const char *fname, bool read_only, bool tasedit, int _stopframe);
 void FCEUI_MoviePlayFromBeginning(void);
 void FCEUI_StopMovie(void);
 bool FCEUI_MovieGetInfo(FCEUFILE* fp, MOVIE_INFO& info, bool skipFrameCount = false);
@@ -263,8 +263,8 @@ std::string FCEUI_GetMovieName(void);
 void FCEUI_MovieToggleFrameDisplay();
 void FCEUI_ToggleInputDisplay(void);
 
-void LoadSubtitles(void);
+void LoadSubtitles(MovieData);
 void ProcessSubtitles(void);
 void FCEU_DisplaySubtitles(char *format, ...);
 
-#endif //__MOVIE_H_
+#endif //__MOVIE_H_
\ No newline at end of file
diff --git a/source/fceultra/ops.inc b/source/fceultra/ops.inc
index 0d66b63..1524d61 100644
--- a/source/fceultra/ops.inc
+++ b/source/fceultra/ops.inc
@@ -343,8 +343,9 @@ case 0x6B: {
 /* ASR */
 case 0x4B: LD_IM(AND;LSRA);
 
-/* ATX(OAL) Is this(OR with $EE) correct? */
-case 0xAB: LD_IM(_A|=0xEE;AND;_X=_A);
+/* ATX(OAL) Is this(OR with $EE) correct? Blargg did some test
+   and found the constant to be OR with is $FF for NES */
+case 0xAB: LD_IM(_A|=0xFF;AND;_X=_A);
 
 /* AXS */ 
 case 0xCB: LD_IM(AXS);
diff --git a/source/fceultra/ppu.cpp b/source/fceultra/ppu.cpp
index 6d9965e..d440083 100644
--- a/source/fceultra/ppu.cpp
+++ b/source/fceultra/ppu.cpp
@@ -49,6 +49,7 @@
 #define SpriteON  (PPU[1]&0x10)   //Show Sprite
 #define ScreenON  (PPU[1]&0x08)   //Show screen
 #define PPUON    (PPU[1]&0x18)		//PPU should operate
+#define GRAYSCALE (PPU[1]&0x01) //Grayscale (AND palette entries with 0x30)
 
 #define SpriteLeft8 (PPU[1]&0x04)
 #define BGLeft8 (PPU[1]&0x02)
@@ -99,6 +100,24 @@ struct BITREVLUT {
 };
 BITREVLUT<uint8,8> bitrevlut;
 
+struct PPUSTATUS
+{
+    int sl;
+    int cycle, end_cycle;
+};
+struct SPRITE_READ
+{
+    int num;
+    int count;
+    int fetch;
+    int found;
+    int found_pos[8];
+    int ret;
+    int last;
+    int mode;
+};
+struct SPRITE_READ spr_read = { 0 };
+
 //uses the internal counters concept at http://nesdev.icequake.net/PPU%20addressing.txt
 struct PPUREGS {
 	uint32 fv;//3
@@ -110,13 +129,17 @@ struct PPUREGS {
 	uint32 s;//1
 	uint32 par;//8
 	uint32 ar;//2
-
-	uint32 _fv, _v, _h, _vt, _ht;
-
+     
+    uint32 _fv, _v, _h, _vt, _ht;
+    
+    struct PPUSTATUS status;
+        
 	PPUREGS()
 		: fv(0), v(0), h(0), vt(0), ht(0), fh(0), s(0), par(0), ar(0)
 		,  _fv(0), _v(0), _h(0), _vt(0), _ht(0)
-	{}
+    { status.cycle = 0; status.end_cycle = 341;
+      status.sl = 241; 
+    }
 
 	void install_latches() {
 		fv = _fv;
@@ -127,9 +150,6 @@ struct PPUREGS {
 	}
 
 	void install_h_latches() {
-		if(ht!=_ht || h != _h) {
-			int zzz=9;
-		}
 		ht = _ht;
 		h = _h;
 	}
@@ -261,6 +281,7 @@ uint8 *MMC5HackExNTARAMPtr=0;
 uint8 *MMC5HackVROMPTR=0;
 uint8 MMC5HackCHRMode=0;
 uint8 MMC5HackSPMode=0;   
+uint8 MMC50x5130=0;
 uint8 MMC5HackSPScroll=0; 
 uint8 MMC5HackSPPage=0;
 
@@ -291,8 +312,8 @@ static uint32 scanlines_per_frame;
 uint8 PPU[4];
 uint8 PPUSPL;
 uint8 NTARAM[0x800],PALRAM[0x20],SPRAM[0x100],SPRBUF[0x100];
-
-
+uint8 UPALRAM[0x03]; //for 0x4/0x8/0xC addresses in palette, the ones in
+                     //0x20 are 0 to not break fceu rendering.
 
 
 #define MMC5SPRVRAMADR(V)      &MMC5SPRVPage[(V)>>10][(V)]
@@ -320,6 +341,7 @@ uint8* FCEUPPU_GetCHR(uint32 vadr, uint32 refreshaddr) {
 		if(MMC5HackCHRMode==1) {
 			uint8 *C = MMC5HackVROMPTR;
 			C += (((MMC5HackExNTARAMPtr[refreshaddr & 0x3ff]) & 0x3f & MMC5HackVROMMask) << 12) + (vadr & 0xfff);
+			C += (MMC50x5130&0x3)<<18; //11-jun-2009 for kuja_killer
 			return C;
 		} else {
 			return MMC5BGVRAMADR(vadr);
@@ -343,23 +365,29 @@ int FCEUPPU_GetAttr(int ntnum, int xt, int yt) {
 inline void FFCEUX_PPUWrite_Default(uint32 A, uint8 V) {
 	uint32 tmp = A;
 
-	if(tmp>=0x3F00)
-		{
-			// hmmm....
-			if(!(tmp&0xf))
-				PALRAM[0x00]=PALRAM[0x04]=PALRAM[0x08]=PALRAM[0x0C]=V&0x3F;
-			else if(tmp&3) PALRAM[(tmp&0x1f)]=V&0x3f;
-		}
-		else if(tmp<0x2000)
-		{
-			if(PPUCHRRAM&(1<<(tmp>>10)))
-				VPage[tmp>>10][tmp]=V;
-		}   
-		else
-		{
-			if(PPUNTARAM&(1<<((tmp&0xF00)>>10)))
-				vnapage[((tmp&0xF00)>>10)][tmp&0x3FF]=V;
-		}
+    if(tmp<0x2000)
+    {
+        if(PPUCHRRAM&(1<<(tmp>>10)))
+            VPage[tmp>>10][tmp]=V;
+    }   
+    else if (tmp<0x3F00)
+    {
+        if(PPUNTARAM&(1<<((tmp&0xF00)>>10)))
+            vnapage[((tmp&0xF00)>>10)][tmp&0x3FF]=V;
+    }
+    else
+    {
+        if (!(tmp & 3))
+        {
+            if (!(tmp & 0xC))
+                  PALRAM[0x00] = PALRAM[0x04] = 
+                  PALRAM[0x08] = PALRAM[0x0C] = V & 0x3F;
+            else
+                UPALRAM[((tmp & 0xC) >> 2) - 1] = V & 0x3F;
+        }
+        else
+            PALRAM[tmp & 0x1F] = V & 0x3F;
+    }
 }
 
 uint8 FFCEUX_PPURead_Default(uint32 A) {
@@ -369,10 +397,27 @@ uint8 FFCEUX_PPURead_Default(uint32 A) {
 	{
 		return VPage[tmp>>10][tmp];
 	}
-	else
+	else if (tmp < 0x3F00)
 	{   
 		return vnapage[(tmp>>10)&0x3][tmp&0x3FF];
 	}
+    else
+    {
+        uint8 ret;
+        if (!(tmp & 3))
+        {
+            if (!(tmp & 0xC))
+                ret = PALRAM[0x00];
+            else
+                ret = UPALRAM[((tmp & 0xC) >> 2) - 1];
+        }
+        else
+            ret = PALRAM[tmp & 0x1F];
+            
+        if (GRAYSCALE)
+            ret &= 0x30;
+        return ret;
+    }
 }
 
 
@@ -419,6 +464,191 @@ static DECLFR(A2002)
 	return ret;
 }
 
+static DECLFR(A2004)
+{
+    if (newppu)
+    {
+        if ((ppur.status.sl < 241) && PPUON)
+        {
+            /* from cycles 0 to 63, the
+             * 32 byte OAM buffer gets init
+             * to 0xFF */
+            if (ppur.status.cycle < 64)
+                return spr_read.ret = 0xFF;
+            else 
+            {
+                for (int i = spr_read.last; 
+                     i != ppur.status.cycle; ++i)
+                {
+                    if (i < 256)
+                    {
+                        switch (spr_read.mode)
+                        {
+                            case 0:
+                                if (spr_read.count < 2)
+                                    spr_read.ret = (PPU[3] & 0xF8) 
+                                    + (spr_read.count << 2);
+                                else
+                                    spr_read.ret = spr_read.count << 2;
+                                spr_read.found_pos[spr_read.found] = 
+                                    spr_read.ret;
+
+                                spr_read.ret = SPRAM[spr_read.ret];
+
+                                if (i & 1) //odd cycle
+                                {
+                                    //see if in range
+                                    if ( !((ppur.status.sl - 1 - 
+                                            spr_read.ret) 
+                                            & ~(Sprite16 ? 0xF : 0x7)) )
+
+                                    {
+                                        ++spr_read.found;
+                                        spr_read.fetch = 1;
+                                        spr_read.mode = 1;
+                                    }
+                                    else 
+                                    {
+                                        if (++spr_read.count == 64)
+                                        {
+                                            spr_read.mode = 4;
+                                            spr_read.count = 0;
+                                        }
+                                        else if (spr_read.found == 8)
+                                        {
+                                            spr_read.fetch = 0;
+                                            spr_read.mode = 2;
+                                        }
+                                    }
+                                }
+                                break;
+                            case 1: //sprite is in range fetch next 3 bytes
+                                if (i & 1)
+                                {
+                                    ++spr_read.fetch;
+                                    if (spr_read.fetch == 4)
+                                    {
+                                        spr_read.fetch = 1;
+                                        if (++spr_read.count == 64)
+                                        {
+                                            spr_read.count = 0;
+                                            spr_read.mode = 4;
+                                        }
+                                        else if (spr_read.found == 8)
+                                        {
+                                            spr_read.fetch = 0;
+                                            spr_read.mode = 2;
+                                        }
+                                        else
+                                            spr_read.mode = 0;
+                                    }
+                                }
+
+                                if (spr_read.count < 2)
+                                    spr_read.ret = (PPU[3] & 0xF8) 
+                                        + (spr_read.count << 2);
+                                else
+                                    spr_read.ret = spr_read.count << 2;
+
+                                spr_read.ret = SPRAM[spr_read.ret | 
+                                    spr_read.fetch];
+                                break;
+                            case 2: //8th sprite fetched 
+                                spr_read.ret = SPRAM[(spr_read.count << 2) 
+                                    | spr_read.fetch];
+                                if (i & 1)
+                                {
+                                    if ( !((ppur.status.sl - 1 - 
+                                              SPRAM[((spr_read.count << 2)
+                                                    | spr_read.fetch)]) 
+                                            & ~((Sprite16) ? 0xF : 0x7)) ) 
+                                    {
+                                        spr_read.fetch = 1;
+                                        spr_read.mode = 3;
+                                    }
+                                    else
+                                    {
+                                        if (++spr_read.count == 64)
+                                        {
+                                            spr_read.count = 0;
+                                            spr_read.mode = 4;
+                                        }
+                                        spr_read.fetch = 
+                                            (spr_read.fetch + 1) & 3;
+                                    }
+                                }
+                                spr_read.ret = spr_read.count;
+                                break;
+                            case 3: //9th sprite overflow detected
+                                spr_read.ret = SPRAM[spr_read.count 
+                                               | spr_read.fetch];
+                                if (i & 1)
+                                {
+                                    if (++spr_read.fetch == 4)
+                                    {
+                                        spr_read.count = (spr_read.count 
+                                                + 1) & 63;
+                                        spr_read.mode = 4;
+                                    }
+                                }
+                                break;
+                            case 4: //read OAM[n][0] until hblank
+                                if (i & 1)
+                                    spr_read.count = 
+                                        (spr_read.count + 1) & 63;
+                                spr_read.fetch = 0;
+                                spr_read.ret = SPRAM[spr_read.count << 2];
+                                break;
+                        }
+                    }
+                    else if (i < 320)
+                    {
+                        spr_read.ret = (i & 0x38) >> 3;
+                        if (spr_read.found < (spr_read.ret + 1))
+                        {
+                            if (spr_read.num)
+                            {
+                                spr_read.ret = SPRAM[252];
+                                spr_read.num = 0;
+                            }
+                            else 
+                                spr_read.ret = 0xFF;
+                        }
+                        else if ((i & 7) < 4)
+                        { 
+                            spr_read.ret = 
+                                SPRAM[spr_read.found_pos[spr_read.ret] 
+                                      | spr_read.fetch++];
+                            if (spr_read.fetch == 4)
+                                spr_read.fetch = 0;
+                        }
+                        else
+                            spr_read.ret = SPRAM[spr_read.found_pos
+                                                 [spr_read.ret | 3]];
+                    }
+                    else 
+                    {
+                        if (!spr_read.found)
+                            spr_read.ret = SPRAM[252];
+                        else
+                            spr_read.ret = SPRAM[spr_read.found_pos[0]];
+                        break;
+                    }
+                }
+                spr_read.last = ppur.status.cycle;
+                return spr_read.ret;
+            }
+        }
+        else
+            return SPRAM[PPU[3]];
+    }
+    else
+    {
+        FCEUPPU_LineUpdate();
+        return PPUGenLatch;
+    }
+}
+
 static DECLFR(A200x)  /* Not correct for $2004 reads. */
 {
 	FCEUPPU_LineUpdate();
@@ -456,14 +686,34 @@ static DECLFR(A2007)
 	uint32 tmp=RefreshAddr&0x3FFF;
 
 	if(newppu) {
-		//mbg
-		ret = VRAMBuffer;
-		RefreshAddr = ppur.get_2007access();
-		VRAMBuffer = CALL_PPUREAD(RefreshAddr);
+        ret = VRAMBuffer;
+		RefreshAddr = ppur.get_2007access() & 0x3FFF;
+        if ((RefreshAddr & 0x3F00) == 0x3F00)
+        {
+            //if it is in the palette range bypass the 
+            //delayed read, and what gets filled in the temp
+            //buffer is the address - 0x1000, also
+            //if grayscale is set then the return is AND with 0x30
+            //to get a gray color reading
+            if (!(tmp & 3))
+            {
+                if (!(tmp & 0xC))
+                    ret = PALRAM[0x00];
+                else
+                    ret = UPALRAM[((tmp & 0xC) >> 2) - 1];
+            }
+            else
+                ret = PALRAM[tmp & 0x1F];
+            if (GRAYSCALE)
+                ret &= 0x30;
+            VRAMBuffer = CALL_PPUREAD(RefreshAddr - 0x1000);
+        }
+        else
+		    VRAMBuffer = CALL_PPUREAD(RefreshAddr);
 		ppur.increment2007(INC32!=0);
 		RefreshAddr = ppur.get_2007access();
-		return ret;
-	} else {
+        return ret;
+    } else {
 		FCEUPPU_LineUpdate();
 
 		ret=VRAMBuffer;
@@ -478,7 +728,7 @@ static DECLFR(A2007)
 			{
 				VRAMBuffer=VPage[tmp>>10][tmp];
 			}
-			else
+			else if (tmp < 0x3F00)
 			{   
 				VRAMBuffer=vnapage[(tmp>>10)&0x3][tmp&0x3FF];
 			}
@@ -525,7 +775,7 @@ static DECLFW(B2001)
 	if(V&0xE0)
 		deemp=V>>5;
 }
-
+//
 static DECLFW(B2002)
 {
 	PPUGenLatch=V;
@@ -542,21 +792,32 @@ static DECLFW(B2003)
 static DECLFW(B2004)
 {
 	//printf("Wr: %04x:$%02x\n",A,V);
-
-	PPUGenLatch=V;
-	if(PPUSPL>=8) 
-	{
-		if(PPU[3]>=8)
-			SPRAM[PPU[3]]=V;
-	}
-	else
-	{   
-		//printf("$%02x:$%02x\n",PPUSPL,V);
-		SPRAM[PPUSPL]=V;
-	}
-	PPU[3]++;
-	PPUSPL++;
-
+    PPUGenLatch=V;
+    if (newppu)
+    {
+        //the attribute upper bits are not connected
+        //so AND them out on write, since reading them
+        //should return 0 in those bits.
+        if ((PPU[3] & 3) == 2)
+            V &= 0xE3;
+        SPRAM[PPU[3]] = V;
+        PPU[3] = (PPU[3] + 1) & 0xFF; 
+    }  
+    else
+    {
+        if(PPUSPL>=8) 
+        {
+            if(PPU[3]>=8)
+                SPRAM[PPU[3]]=V;
+        }
+        else
+        {   
+            //printf("$%02x:$%02x\n",PPUSPL,V);
+            SPRAM[PPUSPL]=V;
+        }
+        PPU[3]++;
+        PPUSPL++;
+    }
 }
 
 static DECLFW(B2005)
@@ -618,15 +879,8 @@ static DECLFW(B2006)
 		ppur._ht = V&31;
 
 		ppur.install_latches();
-
-		if(RefreshAddr==0x18DE) {
-			int zzz=9;
-		}
 	}
 
-	if(ppur._fv == 1) {
-		int zzz=9;
-	}
 	vtoggle^=1;
 }
 
@@ -635,7 +889,7 @@ static DECLFW(B2007)
 	uint32 tmp=RefreshAddr&0x3FFF;
 
 	if(newppu) {
-		RefreshAddr = ppur.get_2007access();
+		RefreshAddr = ppur.get_2007access() & 0x3FFF;
 		CALL_PPUWRITE(RefreshAddr,V);
 		//printf("%04x ",RefreshAddr);
 		ppur.increment2007(INC32!=0);
@@ -644,13 +898,6 @@ static DECLFW(B2007)
 	else
 	{
 		//printf("%04x ",tmp);
-		if(tmp==0x2679)
-		{
-			int zzz=9;
-		}
-				if(tmp == 0x3f13 ) {
-			int zzz=9;
-		}
 		PPUGenLatch=V;
 		if(tmp>=0x3F00)
 		{
@@ -1539,7 +1786,6 @@ void FCEUPPU_Reset(void)
 	vtoggle = 0;
 	ppudead = 2;
 	kook = 0;
-
 	//	XOffset=0;
 }
 
@@ -1548,7 +1794,8 @@ void FCEUPPU_Power(void)
 	int x;
 
 	memset(NTARAM,0x00,0x800);
-	memset(PALRAM,0x00,0x20); 
+	memset(PALRAM,0x00,0x20);
+    memset(UPALRAM,0x00,0x03);
 	memset(SPRAM,0x00,0x100); 
 	FCEUPPU_Reset();
 
@@ -1562,7 +1809,7 @@ void FCEUPPU_Power(void)
 		BWrite[x+2]=B2002;
 		ARead[x+3]=A200x;
 		BWrite[x+3]=B2003;
-		ARead[x+4]=A200x; //A2004;
+		ARead[x+4]=A2004; //A2004;
 		BWrite[x+4]=B2004;
 		ARead[x+5]=A200x;
 		BWrite[x+5]=B2005;
@@ -1753,11 +2000,13 @@ int pputime=0;
 int totpputime=0;
 const int kLineTime=341;
 const int kFetchTime=2;
-int idleSynch = 0;
+int idleSynch = 1;
 
 void runppu(int x) {
 	//pputime+=x;
 	//if(cputodo<200) return;
+    ppur.status.cycle = (ppur.status.cycle + x) % 
+                           ppur.status.end_cycle;
 	X6502_Run(x);
 	//pputime -= cputodo<<2;
 }
@@ -1774,24 +2023,30 @@ struct BGData {
 
 				RefreshAddr = ppur.get_atread();
 				at = CALL_PPUREAD(RefreshAddr);
-				runppu(kFetchTime);
-
+				
 				//modify at to get appropriate palette shift
 				if(ppur.vt&2) at >>= 4;
 				if(ppur.ht&2) at >>= 2;
 				at &= 0x03;
 				at <<= 2;
-
-				ppur.par = nt;
+                //horizontal scroll clocked at cycle 3 and then
+                //vertical scroll at 251
+                runppu(1);
+                if (PPUON)
+                {
+			        ppur.increment_hsc();
+                    if (ppur.status.cycle == 251) 
+                        ppur.increment_vs();
+                }
+                runppu(1);
+				
+                ppur.par = nt;
 				RefreshAddr = ppur.get_ptread();
 				pt[0] = CALL_PPUREAD(RefreshAddr);
 				runppu(kFetchTime);
 				RefreshAddr |= 8;
 				pt[1] = CALL_PPUREAD(RefreshAddr);
 				runppu(kFetchTime);
-
-				if(PPUON)
-					ppur.increment_hsc();
 			}
 		};
 
@@ -1802,15 +2057,24 @@ struct BGData {
 int framectr=0;
 int FCEUX_PPU_Loop(int skip) {
 	//262 scanlines
-
-	if(ppudead)
-	{
-		memset(XBuf, 0x80, 256*240);
-		runppu(262*kLineTime);
-		ppudead--;
-		goto finish;
-	}
-
+    if (ppudead)
+    {
+        /* not quite emulating all the NES power up behavior
+         * since it is known that the NES ignores writes to some
+         * register before around a full frame, but no games
+         * should write to those regs during that time, it needs
+         * to wait for vblank  */
+        ppur.status.sl = 241;
+        if (PAL)
+            runppu(70*kLineTime);
+        else
+            runppu(20*kLineTime);
+        ppur.status.sl = 0;
+        runppu(242*kLineTime);
+        ppudead = 0;
+        goto finish;
+    }
+   
 	{
 		PPU_status |= 0x80;
 		ppuphase = PPUPHASE_VBL;
@@ -1818,16 +2082,20 @@ int FCEUX_PPU_Loop(int skip) {
 		//Not sure if this is correct.  According to Matt Conte and my own tests, it is.
 		//Timing is probably off, though.  
 		//NOTE:  Not having this here breaks a Super Donkey Kong game. 
-		//PPU[3]=PPUSPL=0;       
-
+		PPU[3]=PPUSPL=0;       
 		const int delay = 20; //fceu used 12 here but I couldnt get it to work in marble madness and pirates.
-		runppu(delay); //X6502_Run(12);
+        
+        ppur.status.sl = 241; //for sprite reads
+		
+        runppu(delay); //X6502_Run(12);
 		if(VBlankON) TriggerNMI();
-		runppu(20*(kLineTime)-delay);
-
+        if (PAL)
+            runppu(70*(kLineTime)-delay);
+        else
+		    runppu(20*(kLineTime)-delay);
+        
 		//this seems to run just before the dummy scanline begins
-		PPU_status&=0x1f;
-
+		PPU_status = 0;
 		//this early out caused metroid to fail to boot. I am leaving it here as a reminder of what not to do
 		//if(!PPUON) { runppu(kLineTime*242); goto finish; }
 
@@ -1837,8 +2105,8 @@ int FCEUX_PPU_Loop(int skip) {
 		//rendering data for the first time in a frame (this update won't happen if 
 		//all rendering is disabled via 2001.3 and 2001.4).
 
-		if(PPUON)
-			ppur.install_latches();
+		//if(PPUON)
+		//	ppur.install_latches();
 
 		uint8 oams[2][64][7];
 		int oamcounts[2]={0,0};
@@ -1847,11 +2115,19 @@ int FCEUX_PPU_Loop(int skip) {
 
 		//capture the initial xscroll
 		//int xscroll = ppur.fh;
-
 		//render 241 scanlines (including 1 dummy at beginning)
 		for(int sl=0;sl<241;sl++) {
-			int yp = sl-1;
+            spr_read.num = 1;
+            spr_read.found = 0;
+            spr_read.fetch = 1;
+            spr_read.count = 0;
+            spr_read.last = 64;
+            spr_read.mode = 0;
+            memset(spr_read.found_pos, 0, sizeof(spr_read.found_pos));
 
+            ppur.status.sl = sl;
+
+			int yp = sl-1;
 			ppuphase = PPUPHASE_BG;
 
 			if(sl != 0) {
@@ -1875,7 +2151,7 @@ int FCEUX_PPU_Loop(int skip) {
 			for(int xt=0;xt<32;xt++) {
 				bgdata.main[xt+2].Read();
 
-				//ok, we're also going to draw here.
+                //ok, we're also going to draw here.
 				//unless we're on the first dummy scanline
 				if(sl != 0) {
 					int xstart = xt<<3;
@@ -1887,7 +2163,6 @@ int FCEUX_PPU_Loop(int skip) {
 					//check all the conditions that can cause things to render in these 8px
 					bool renderspritenow = SpriteON && rendersprites && (xt>0 || SpriteLeft8);
 					bool renderbgnow = ScreenON && renderbg && (xt>0 || BGLeft8);
-
 					for(int xp=0;xp<8;xp++,rasterpos++) {
 
 						//bg pos is different from raster pos due to its offsetability.
@@ -1987,14 +2262,19 @@ int FCEUX_PPU_Loop(int skip) {
 			//FV is clocked by the PPU's horizontal blanking impulse, and therefore will increment every scanline.
 			//well, according to (which?) tests, maybe at the end of hblank. 
 			//but, according to what it took to get crystalis working, it is at the beginning of hblank.
-			if(PPUON && sl != 0)
-				ppur.increment_vs();
+            
+            //this is done at cycle 251 
+            //rendering scanline, it doesn't need to be scanline 0, 
+            //because on the first scanline when the increment is 0, the vs_scroll is reloaded.
+			//if(PPUON && sl != 0)
+			//	ppur.increment_vs();
 
 			//todo - think about clearing oams to a predefined value to force deterministic behavior
 
 			//so.. this is the end of hblank. latch horizontal scroll values
-			if(PPUON && sl != 0)
-				ppur.install_h_latches();
+            //do it cycle at 251 
+			 if(PPUON && sl != 0)
+				ppur.install_h_latches(); 
 
 			ppuphase = PPUPHASE_OBJ;
 
@@ -2035,7 +2315,23 @@ int FCEUX_PPU_Loop(int skip) {
 				patternAddress += line&7;
 
 				//garbage nametable fetches
-				if(realSprite) runppu(kFetchTime);
+                //reset the scroll counter, happens at cycle 304
+                if (realSprite)
+                {
+                    if ((sl == 0) && PPUON)
+                    {
+                        if (ppur.status.cycle == 304)
+                        {
+                            runppu(1);
+                            ppur.install_latches();
+                            runppu(1);
+                        }
+                        else
+                            runppu(kFetchTime);
+                    }
+                    else
+                        runppu(kFetchTime);
+                }
 
 				if(((PPU[0]&0x38)!=0x18) && s == 2 && SpriteON ) {
 					//(The MMC3 scanline counter is based entirely on PPU A12, triggered on rising edges (after the line remains low for a sufficiently long period of time))
@@ -2050,15 +2346,16 @@ int FCEUX_PPU_Loop(int skip) {
 
 				if(realSprite) runppu(kFetchTime);
 
+
 				//pattern table fetches
 				RefreshAddr = patternAddress;
 				oam[4] = CALL_PPUREAD(RefreshAddr);
 				if(realSprite) runppu(kFetchTime);
+
 				RefreshAddr += 8;
 				oam[5] = CALL_PPUREAD(RefreshAddr);
 				if(realSprite) runppu(kFetchTime);
 
-
 				//hflip
 				if(!(oam[2]&0x40)) {
 					oam[4] = bitrevlut[oam[4]];
@@ -2077,29 +2374,32 @@ int FCEUX_PPU_Loop(int skip) {
 			//same nametable address that points to the 3rd tile to be rendered on the 
 			//screen (or basically, the first nametable address that will be accessed when 
 			//the PPU is fetching background data on the next scanline).
-			//(not implemented yet)
+			//(not implemented yet) 
 			runppu(kFetchTime);
+            if (sl == 0)
+            {
+                if (idleSynch && PPUON && !PAL)
+                    ppur.status.end_cycle = 340;
+                else
+                    ppur.status.end_cycle = 341;
+                idleSynch ^= 1;
+            }
+            else
+                ppur.status.end_cycle = 341;
 			runppu(kFetchTime);
 
-
-			//After memory access 170, the PPU simply rests for 4 cycles (or the 
+            //After memory access 170, the PPU simply rests for 4 cycles (or the 
 			//equivelant of half a memory access cycle) before repeating the whole 
 			//pixel/scanline rendering process. If the scanline being rendered is the very 
 			//first one on every second frame, then this delay simply doesn't exist.
-			if(sl==0 && idleSynch==0)
-			{}
-			else 
-				runppu(1);
-		}
-		
-		idleSynch ++;
-		if(idleSynch==2) idleSynch = 0;
+            if (ppur.status.end_cycle == 341)
+                runppu(1);
+        }
 
 		if(MMC5Hack && PPUON) MMC5_hb(240);
 
 		//idle for one line
 		runppu(kLineTime);
-
 		framectr++;
 
 	}
diff --git a/source/fceultra/pputile.inc b/source/fceultra/pputile.inc
index 7eacb59..733b1d6 100644
--- a/source/fceultra/pputile.inc
+++ b/source/fceultra/pputile.inc
@@ -76,6 +76,7 @@ pshift[1]<<=8;
 		C = MMC5HackVROMPTR;
 		C += (((MMC5HackExNTARAMPtr[RefreshAddr & 0x3ff]) & 0x3f & 
 			MMC5HackVROMMask) << 12) + (vadr & 0xfff);
+		C += (MMC50x5130&0x3)<<18; //11-jun-2009 for kuja_killer
 	#elif defined(PPUT_MMC5)
 		C=MMC5BGVRAMADR(vadr);
 	#else
diff --git a/source/fceultra/sound.cpp b/source/fceultra/sound.cpp
index 2f5f37e..1da02c9 100644
--- a/source/fceultra/sound.cpp
+++ b/source/fceultra/sound.cpp
@@ -89,25 +89,41 @@ static int32 sqacc[2];
 static int32 lengthcount[4]; 
 static const uint8 lengthtable[0x20]=
 {
- 0x5*2,0x7f*2,0xA*2,0x1*2,0x14*2,0x2*2,0x28*2,0x3*2,0x50*2,0x4*2,0x1E*2,0x5*2,0x7*2,0x6*2,0x0E*2,0x7*2,
- 0x6*2,0x08*2,0xC*2,0x9*2,0x18*2,0xa*2,0x30*2,0xb*2,0x60*2,0xc*2,0x24*2,0xd*2,0x8*2,0xe*2,0x10*2,0xf*2
+	10,254, 20,  2, 40,  4, 80,  6, 160,  8, 60, 10, 14, 12, 26, 14,
+	12, 16, 24, 18, 48, 20, 96, 22, 192, 24, 72, 26, 16, 28, 32, 30
 };
 
-static const uint32 NoiseFreqTable[0x10]=
+
+static const uint32 NoiseFreqTableNTSC[0x10] = 
 {
- 2,4,8,0x10,0x20,0x30,0x40,0x50,0x65,0x7f,0xbe,0xfe,0x17d,0x1fc,0x3f9,0x7f2
+	4, 8, 16, 32, 64, 96, 128, 160, 202, 
+	254, 380, 508, 762, 1016, 2034, 4068
 };
 
+static const uint32 NoiseFreqTablePAL[0x10] = 
+{
+	4, 7, 14, 30, 60, 88, 118, 148, 188, 
+	236, 354, 472, 708,  944, 1890, 3778
+};
+
+static const uint32 *NoiseFreqTable = NoiseFreqTableNTSC;
+
 static const uint32 NTSCDMCTable[0x10]=
 {
  428,380,340,320,286,254,226,214,
  190,160,142,128,106, 84 ,72,54
 };
 
+/* Previous values for PAL DMC was value - 1,
+ * I am not certain if this is if FCEU handled
+ * PAL differently or not, the NTSC values are right,
+ * so I am assuming that the current value is handled
+ * the same way NTSC is handled. */
+
 static const uint32 PALDMCTable[0x10]=
 {
- 397, 353, 315, 297, 265, 235, 209, 198, 
- 176, 148, 131, 118, 98, 78, 66, 50, 
+	398, 354, 316, 298, 276, 236, 210, 198, 
+	176, 148, 132, 118,  98,  78,  66,  50
 };
 
 // $4010        -        Frequency
@@ -575,7 +591,7 @@ static INLINE void RDoSQ(int x)		//Int x decides if this is Square Wave 1 or 2
    
    //Modify Square wave volume based on channel volume modifiers
    //adelikat: Note: the formulat x = x * y /100 does not yield exact results, but is "close enough" and avoids the need for using double vales or implicit cohersion which are slower (we need speed here)
-   ampx = x ? FSettings.Square1Volume : FSettings.Square2Volume; // TODO OPTIMIZE ME!
+   ampx = x ? FSettings.Square2Volume : FSettings.Square1Volume; // TODO OPTIMIZE ME!
    if (ampx != 256) amp = (amp * ampx) / 256; // CaH4e3: fixed - setting up maximum volume for square2 caused complete mute square2 channel
          
    amp<<=24;
@@ -837,7 +853,9 @@ static void RDoTriangleNoisePCMLQ(void)
     if(noiseacc<=0)
     {
      rea2:
-     noiseacc+=NoiseFreqTable[PSG[0xE]&0xF]<<(16+2);
+        //used to added <<(16+2) when the noise table
+        //values were half.
+     noiseacc+=NoiseFreqTable[PSG[0xE]&0xF]<<(16+1);
      nreg=(nreg<<1)+(((nreg>>nshift)^(nreg>>14))&1);
      nreg&=0x7fff;
      noiseout=amptab[(nreg>>0xe)];
@@ -876,7 +894,9 @@ static void RDoTriangleNoisePCMLQ(void)
      if(noiseacc<=0)
      {
       area2:
-      noiseacc+=NoiseFreqTable[PSG[0xE]&0xF]<<(16+2);
+         //used to be added <<(16+2) when the noise table
+         //values were half.
+      noiseacc+=NoiseFreqTable[PSG[0xE]&0xF]<<(16+1);
       nreg=(nreg<<1)+(((nreg>>nshift)^(nreg>>14))&1);
       nreg&=0x7fff;
       noiseout=amptab[(nreg>>0xe)];
@@ -927,7 +947,7 @@ static void RDoNoise(void)
    if(!wlcount[3])
    {
     uint8 feedback;
-    wlcount[3]=NoiseFreqTable[PSG[0xE]&0xF]<<1;
+    wlcount[3]=NoiseFreqTable[PSG[0xE]&0xF];
     feedback=((nreg>>8)&1)^((nreg>>14)&1);
     nreg=(nreg<<1)+feedback;
     nreg&=0x7fff;
@@ -942,7 +962,7 @@ static void RDoNoise(void)
    if(!wlcount[3])
    {
     uint8 feedback;
-    wlcount[3]=NoiseFreqTable[PSG[0xE]&0xF]<<1;
+    wlcount[3]=NoiseFreqTable[PSG[0xE]&0xF];
     feedback=((nreg>>13)&1)^((nreg>>14)&1);
     nreg=(nreg<<1)+feedback;
     nreg&=0x7fff;
@@ -1063,25 +1083,32 @@ due to that whole MegaMan 2 Game Genie thing.
 
 void FCEUSND_Reset(void)
 {
-        int x;
-
+	int x;
+	
 	IRQFrameMode=0x0;
-        fhcnt=fhinc;
-        fcnt=0;
+	fhcnt=fhinc;
+	fcnt=0;
+	nreg=1;
 
-        nreg=1;
-        for(x=0;x<2;x++)
+	if (PAL)
+		NoiseFreqTable = NoiseFreqTablePAL;
+	else
+		NoiseFreqTable = NoiseFreqTableNTSC;
+
+	for(x=0;x<2;x++)
 	{
-         wlcount[x]=2048;
-	 if(nesincsize) // lq mode
-	  sqacc[x]=((uint32)2048<<17)/nesincsize;
-	 else
-	  sqacc[x]=1;
-	 sweepon[x]=0;
-	 curfreq[x]=0;
+		wlcount[x]=2048;
+		if(nesincsize) // lq mode
+			sqacc[x]=((uint32)2048<<17)/nesincsize;
+		else
+			sqacc[x]=1;
+		sweepon[x]=0;
+		curfreq[x]=0;
 	}
-  wlcount[2]=1;  //2048;
-        wlcount[3]=2048;
+	
+	wlcount[2]=1;  //2048;
+	wlcount[3]=2048;
+
 	DMCHaveDMA=DMCHaveSample=0;
 	SIRQStat=0x00;
 
diff --git a/source/fceultra/state.cpp b/source/fceultra/state.cpp
index 7afb7b9..65cbee1 100644
--- a/source/fceultra/state.cpp
+++ b/source/fceultra/state.cpp
@@ -332,7 +332,7 @@ static bool ReadStateChunks(std::istream* is, int32 totalsize)
 	return ret;
 }
 
-int CurrentState=1;
+int CurrentState=0;
 extern int geniestage;
 
 
diff --git a/source/fceultra/types.h b/source/fceultra/types.h
index ebc8a0b..56b23c9 100644
--- a/source/fceultra/types.h
+++ b/source/fceultra/types.h
@@ -23,7 +23,7 @@
 #define __FCEU_TYPES
 
 #define FCEU_NAME "FCEUX"
-#define FCEU_VERSION_STRING "2.1.0a"
+#define FCEU_VERSION_STRING "2.1.1-interim"
 #define FCEU_VERSION_NUMERIC 20100
 #define FCEU_NAME_AND_VERSION FCEU_NAME " " FCEU_VERSION_STRING
 
diff --git a/source/fceultra/utils/md5.cpp b/source/fceultra/utils/md5.cpp
index 6471c3c..f4ea950 100644
--- a/source/fceultra/utils/md5.cpp
+++ b/source/fceultra/utils/md5.cpp
@@ -237,8 +237,8 @@ char *md5_asciistr(MD5DATA& md5)
 
  for(x=0;x<16;x++)
  {
-  str[x*2]=trans[digest[x]&0x0F];
-  str[x*2+1]=trans[digest[x]>>4];
+  str[x*2]=trans[digest[x]>>4];
+  str[x*2+1]=trans[digest[x]&0x0F];
  }
  return(str);
 }
diff --git a/source/fceultra/utils/memorystream.h b/source/fceultra/utils/memorystream.h
index 8c13dcd..7c6a343 100644
--- a/source/fceultra/utils/memorystream.h
+++ b/source/fceultra/utils/memorystream.h
@@ -185,7 +185,7 @@ private:
 			throw new std::runtime_error("memory_streambuf is not expandable");
 
 		size_t newcapacity;
-		if(upto == 0)
+		if(upto == -1)
 			newcapacity = capacity + capacity/2 + 2;
 		else
 			newcapacity = std::max(upto,capacity);
diff --git a/source/fceultra/x6502.cpp b/source/fceultra/x6502.cpp
index 1920f26..1e03bd0 100644
--- a/source/fceultra/x6502.cpp
+++ b/source/fceultra/x6502.cpp
@@ -29,7 +29,6 @@
 #endif
 
 #include "x6502abbrev.h"
-
 X6502 X;
 uint32 timestamp;
 void (*MapIRQHook)(int a);
@@ -574,4 +573,4 @@ const uint8 optype[256] = {
 /*0xD0*/	0,4,0,0,0,5,5,0,0,6,0,0,0,7,7,0,
 /*0xE0*/	0,1,0,0,2,2,2,0,0,0,0,0,3,3,3,0,
 /*0xF0*/	0,4,0,0,0,5,5,0,0,6,0,0,0,7,7,0
-};
\ No newline at end of file
+};