diff --git a/BizHawk.Emulation/CPUs/MOS 6502X/Execute.cs b/BizHawk.Emulation/CPUs/MOS 6502X/Execute.cs
index 70ac5e2aa8..a24e87e50d 100644
--- a/BizHawk.Emulation/CPUs/MOS 6502X/Execute.cs	
+++ b/BizHawk.Emulation/CPUs/MOS 6502X/Execute.cs	
@@ -1,5 +1,4 @@
 //http://nesdev.parodius.com/6502_cpu.txt
-//TODO - correct brk/irq/nmi interrupting and prioritization
 //TODO - rename unofficial NOPs as DOPs? (see immediate instr tests)
 using System;
 
@@ -9,7 +8,7 @@ namespace BizHawk.Emulation.CPUs.M6502
 	{
 		static Uop[][] Microcode = new Uop[][]{
 			//0x00
-			/*BRK [implied]*/ new Uop[] { Uop.Fetch2, Uop.PushPCH, Uop.PushPCL, Uop.PushP_BRK, Uop.FetchPCLVector, Uop.FetchPCHVector, Uop.End },
+			/*BRK [implied]*/ new Uop[] { Uop.Fetch2, Uop.PushPCH, Uop.PushPCL, Uop.PushP_BRK, Uop.FetchPCLVector, Uop.FetchPCHVector, Uop.End_SuppressInterrupt },
 			/*ORA (addr,X) [indexed indirect READ]*/ new Uop[] { Uop.Fetch2, Uop.IdxInd_Stage3, Uop.IdxInd_Stage4, Uop.IdxInd_Stage5, Uop.IdxInd_Stage6_READ_ORA, Uop.End },
 			/*JAM*/ new Uop[] { Uop.End },
 			/*SLO* (addr,X) [indexed indirect RMW] [unofficial]*/ new Uop[] { Uop.Fetch2, Uop.End },
@@ -282,14 +281,16 @@ namespace BizHawk.Emulation.CPUs.M6502
 			/*ISB* addr,X [absolute indexed RMW X] [unofficial]*/ new Uop[] { Uop.Fetch2, Uop.AbsIdx_Stage3_X,  Uop.AbsIdx_Stage4, Uop.AbsIdx_RMW_Stage5, Uop.AbsIdx_RMW_Stage6_Unofficial, Uop.AbsIdx_RMW_Stage7, Uop.End },
 			//0x100
 			/*VOP_Fetch1*/ new Uop[] { Uop.Fetch1 },
-			/*VOP_RelativeStuff*/ new Uop[] { Uop.RelBranch_Stage3, Uop.End },
+			/*VOP_RelativeStuff*/ new Uop[] { Uop.RelBranch_Stage3, Uop.End_BranchSpecial },
 			/*VOP_RelativeStuff2*/ new Uop[] { Uop.RelBranch_Stage4, Uop.End },
+			/*VOP_RelativeStuff2*/ new Uop[] { Uop.End_SuppressInterrupt },
 			//i assume these are dummy fetches.... maybe theyre just nops? supposedly these take 7 cycles so thats the only way i can make sense of it
 			//one of them might be the next instruction's fetch, and whatever fetch follows it.
 			//the interrupt would then take place if necessary, using a cached PC. but im not so sure about that.
-			/*VOP_NMI*/ new Uop[] { Uop.FetchDummy, Uop.FetchDummy, Uop.PushPCH, Uop.PushPCL, Uop.PushP_NMI, Uop.FetchPCLVector, Uop.FetchPCHVector, Uop.End },
-			/*VOP_IRQ*/ new Uop[] { Uop.FetchDummy, Uop.FetchDummy, Uop.PushPCH, Uop.PushPCL, Uop.PushP_IRQ, Uop.FetchPCLVector, Uop.FetchPCHVector, Uop.End },
-			/*VOP_RESET*/ new Uop[] { Uop.FetchDummy, Uop.FetchDummy, Uop.PushDummy, Uop.PushDummy, Uop.PushP_Reset, Uop.FetchPCLVector, Uop.FetchPCHVector, Uop.End },
+			/*VOP_NMI*/ new Uop[] { Uop.FetchDummy, Uop.FetchDummy, Uop.PushPCH, Uop.PushPCL, Uop.PushP_NMI, Uop.FetchPCLVector, Uop.FetchPCHVector, Uop.End_SuppressInterrupt },
+			/*VOP_IRQ*/ new Uop[] { Uop.FetchDummy, Uop.FetchDummy, Uop.PushPCH, Uop.PushPCL, Uop.PushP_IRQ, Uop.FetchPCLVector, Uop.FetchPCHVector, Uop.End_SuppressInterrupt },
+			/*VOP_RESET*/ new Uop[] { Uop.FetchDummy, Uop.FetchDummy, Uop.PushDummy, Uop.PushDummy, Uop.PushP_Reset, Uop.FetchPCLVector, Uop.FetchPCHVector, Uop.End_SuppressInterrupt },
+			/*VOP_Fetch1_NoInterrupt*/  new Uop[] { Uop.Fetch1_Real },
 		};
 		
 		enum Uop
@@ -297,7 +298,7 @@ namespace BizHawk.Emulation.CPUs.M6502
 			//sometimes i used this as a marker for unsupported instructions, but it is very inconsistent
 			Unsupported,
 
-			Fetch1, Fetch2, Fetch3, 
+			Fetch1, Fetch1_Real, Fetch2, Fetch3, 
 			//used by instructions with no second opcode byte (6502 fetches a byte anyway but won't increment PC for these)
 			FetchDummy,
 
@@ -385,50 +386,45 @@ namespace BizHawk.Emulation.CPUs.M6502
 
 			End,
 			End_ISpecial, //same as end, but preserves the iflag set by the instruction
+			End_BranchSpecial,
+			End_SuppressInterrupt,
 		}
 
 		const int VOP_Fetch1 = 256;
 		const int VOP_RelativeStuff = 257;
 		const int VOP_RelativeStuff2 = 258;
-		const int VOP_NMI = 259;
-		const int VOP_IRQ = 260;
-		const int VOP_RESET = 261;
+		const int VOP_RelativeStuff3 = 259;
+		const int VOP_NMI = 260;
+		const int VOP_IRQ = 261;
+		const int VOP_RESET = 262;
+		const int VOP_Fetch1_NoInterrupt = 263;
 
+		//opcode bytes.. theoretically redundant with the temp variables? who knows.
 		int opcode;
-		byte opcode2, opcode3; //opcode bytes.. theoretically redundant with the temp variables? who knows.
+		byte opcode2, opcode3;
+
 		int ea, alu_temp; //cpu internal temp variables
 		int mi; //microcode index
-		//bool branch_taken; //only needed for the timing debug
 		bool iflag_pending; //iflag must be stored after it is checked in some cases (CLI and SEI).
 
+		//tracks whether an interrupt condition has popped up recently.
+		//not sure if this is real or not but it helps with the branch_irq_hack
+		bool interrupt_pending; 
+		bool branch_irq_hack; //see Uop.RelBranch_Stage3 for more details
+
+		bool Interrupted
+		{
+			get
+			{
+				return NMI || (IRQ && !FlagI);
+			}
+		}
+
 		void FetchDummy()
 		{
 			DummyReadMemory(PC);
 		}
 
-		////timing debug
-		//int ctr = 0;
-		//int realOpcode = 0;
-		//public static byte[] CycTable = new byte[]
-		//{                             
-		///*0x00*/ 7,6,2,8,3,3,5,5,3,2,2,2,4,4,6,6,
-		///*0x10*/ 2,5,2,8,4,4,6,6,2,4,2,7,4,4,7,7,
-		///*0x20*/ 6,6,2,8,3,3,5,5,4,2,2,2,4,4,6,6,
-		///*0x30*/ 2,5,2,8,4,4,6,6,2,4,2,7,4,4,7,7,
-		///*0x40*/ 6,6,2,8,3,3,5,5,3,2,2,2,3,4,6,6,
-		///*0x50*/ 2,5,2,8,4,4,6,6,2,4,2,7,4,4,7,7,
-		///*0x60*/ 6,6,2,8,3,3,5,5,4,2,2,2,5,4,6,6,
-		///*0x70*/ 2,5,2,8,4,4,6,6,2,4,2,7,4,4,7,7,
-		///*0x80*/ 2,6,2,6,3,3,3,3,2,2,2,2,4,4,4,4,
-		///*0x90*/ 2,6,2,6,4,4,4,4,2,5,2,5,5,5,5,5,
-		///*0xA0*/ 2,6,2,6,3,3,3,3,2,2,2,2,4,4,4,4,
-		///*0xB0*/ 2,5,2,5,4,4,4,4,2,4,2,4,4,4,4,4,
-		///*0xC0*/ 2,6,2,8,3,3,5,5,2,2,2,2,4,4,6,6,
-		///*0xD0*/ 2,5,2,8,4,4,6,6,2,4,2,7,4,4,7,7,
-		///*0xE0*/ 2,6,3,8,3,3,5,5,2,2,2,2,4,4,6,6,
-		///*0xF0*/ 2,5,2,8,4,4,6,6,2,4,2,7,4,4,7,7,
-		//};
-
 		public void Execute(int cycles)
 		{
 			for (int i = 0; i < cycles; i++)
@@ -445,6 +441,8 @@ namespace BizHawk.Emulation.CPUs.M6502
 
 			TotalExecutedCycles++;
 
+			interrupt_pending |= Interrupted;
+
 		RETRY:
 			Uop uop = Microcode[opcode][mi];
 			switch (uop)
@@ -454,44 +452,34 @@ namespace BizHawk.Emulation.CPUs.M6502
 					{
 						bool my_iflag = FlagI;
 						FlagI = iflag_pending;
-						if (NMI)
+						if (!branch_irq_hack)
 						{
-							ea = NMIVector;
-							opcode = VOP_NMI;
-							NMI = false;
-							mi = 0;
-							goto RETRY;
+							interrupt_pending = false;
+							if (NMI)
+							{
+								ea = NMIVector;
+								opcode = VOP_NMI;
+								NMI = false;
+								mi = 0;
+								goto RETRY;
+							}
+							else if (IRQ && !my_iflag)
+							{
+								ea = IRQVector;
+								opcode = VOP_IRQ;
+								mi = 0;
+								goto RETRY;
+							}
 						}
-						else if (IRQ && !my_iflag)
-						{
-							ea = IRQVector;
-							opcode = VOP_IRQ;
-							mi = 0;
-							goto RETRY;
-						}
-#if TIMINGDEBUG
-						if (debug)
-						{
-							int ideal = CycTable[realOpcode] + (branch_taken ? 1 : 0);
-							int actual = TotalExecutedCycles - ctr;
-							Console.Write(" | ideal={0}", ideal);
-							Console.Write(" actual={0}", actual);
-							if (actual != ideal) Console.WriteLine(" !!!"); else Console.WriteLine();
-							Console.Write(State());
-						}
-						branch_taken = false;
-						opcode = ReadMemory(PC++);
-						realOpcode = opcode;
-						mi = -1;
-						ctr = TotalExecutedCycles;
-						break;
-#else
-						if(debug) Console.WriteLine(State());
-						opcode = ReadMemory(PC++);
-						mi = -1;
-						break;
-#endif
+						goto case Uop.Fetch1_Real;
 					}
+
+				case Uop.Fetch1_Real:
+					if (debug) Console.WriteLine(State());
+					branch_irq_hack = false;
+					opcode = ReadMemory(PC++);
+					mi = -1;
+					break;
 					
 				case Uop.Fetch2: opcode2 = ReadMemory(PC++); break;
 				case Uop.Fetch3: opcode3 = ReadMemory(PC++); break;
@@ -526,6 +514,16 @@ namespace BizHawk.Emulation.CPUs.M6502
 					S--;
 					break;
 				case Uop.FetchPCLVector:
+					if (ea == BRKVector && FlagB && NMI)
+					{
+						NMI = false;
+						ea = NMIVector;
+					}
+					if(ea == IRQVector && !FlagB && NMI)
+					{
+						NMI = false;
+						ea = NMIVector;
+					}
 					alu_temp = ReadMemory((ushort)ea);
 					break;
 				case Uop.FetchPCHVector:
@@ -652,18 +650,26 @@ namespace BizHawk.Emulation.CPUs.M6502
 						opcode = VOP_RelativeStuff;
 						mi = -1;
 					}
+
 					break;
 				case Uop.RelBranch_Stage3:
 					FetchDummy();
 					alu_temp = (byte)PC + (int)(sbyte)opcode2;
 					PC &= 0xFF00;
 					PC |= (ushort)((alu_temp&0xFF));
-					if(alu_temp.Bit(8))
+					if (alu_temp.Bit(8))
 					{
 						//we need to carry the add, and then we'll be ready to fetch the next instruction
 						opcode = VOP_RelativeStuff2;
 						mi = -1;
 					}
+					else
+					{
+						//to pass cpu_interrupts_v2/5-branch_delays_irq we need to handle a quirk here
+						//if we decide to interrupt in the next cycle, this condition will cause it to get deferred by one instruction
+						if(!interrupt_pending)
+							branch_irq_hack = true;
+					}
 					break;
 				case Uop.RelBranch_Stage4:
 					FetchDummy();
@@ -678,7 +684,7 @@ namespace BizHawk.Emulation.CPUs.M6502
 				case Uop.JSR: PC = (ushort)((ReadMemory((ushort)(PC)) << 8) + opcode2); break;
 				case Uop.PullP: 
 					P = ReadMemory((ushort)(S++ + 0x100));
-					FlagT = true;
+					FlagT = true; //force T always to remain true
 					break;
 				case Uop.PullPCL:
 					PC &= 0xFF00;
@@ -968,7 +974,7 @@ namespace BizHawk.Emulation.CPUs.M6502
 						P = ReadMemory((ushort)(S + 0x100));
 						iflag_pending = FlagI;
 						FlagI = my_iflag;
-						FlagT = true; //why?
+						FlagT = true; //force T always to remain true
 						break;
 					}
 
@@ -1239,11 +1245,18 @@ namespace BizHawk.Emulation.CPUs.M6502
 					mi = 0;
 					goto RETRY;
 
+				case Uop.End_SuppressInterrupt:
+					opcode = VOP_Fetch1_NoInterrupt;
+					mi = 0;
+					goto RETRY;
+
 				case Uop.End:
 					opcode = VOP_Fetch1;
 					mi = 0;
 					iflag_pending = FlagI;
 					goto RETRY;
+				case Uop.End_BranchSpecial:
+					goto case Uop.End;
 			}
 
 			mi++;
diff --git a/BizHawk.Emulation/CPUs/MOS 6502X/MOS6502X.cs b/BizHawk.Emulation/CPUs/MOS 6502X/MOS6502X.cs
index d65c362dac..65eff5243f 100644
--- a/BizHawk.Emulation/CPUs/MOS 6502X/MOS6502X.cs	
+++ b/BizHawk.Emulation/CPUs/MOS 6502X/MOS6502X.cs	
@@ -96,6 +96,8 @@ namespace BizHawk.Emulation.CPUs.M6502
 			ser.Sync("alu_temp", ref alu_temp);
 			ser.Sync("mi", ref mi);
 			ser.Sync("iflag_pending", ref iflag_pending);
+			ser.Sync("interrupt_pending", ref interrupt_pending);
+			ser.Sync("branch_irq_hack", ref branch_irq_hack);
 			ser.EndSection();
 		}
 
diff --git a/BizHawk.Emulation/Consoles/Nintendo/NES/APU.cs b/BizHawk.Emulation/Consoles/Nintendo/NES/APU.cs
index 27c0bc2aed..e517535364 100644
--- a/BizHawk.Emulation/Consoles/Nintendo/NES/APU.cs
+++ b/BizHawk.Emulation/Consoles/Nintendo/NES/APU.cs
@@ -47,11 +47,11 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				//4, 7, 14, 30, 60, 88, 118, 148, 188, 236, 354, 472, 708,  944, 1890, 3778 //PAL
 			};
 
-	
+
 			class PulseUnit
 			{
 				public PulseUnit(int unit) { this.unit = unit; }
-				public int unit; 
+				public int unit;
 
 				//reg0
 				int duty_cnt, env_loop, env_constant, env_cnt_value;
@@ -87,10 +87,11 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 
 				public bool IsLenCntNonZero() { return len_cnt > 0; }
 
+
 				public void WriteReg(int addr, byte val)
 				{
 					//Console.WriteLine("write pulse {0:X} {1:X}", addr, val);
-					switch(addr)
+					switch (addr)
 					{
 						case 0:
 							env_cnt_value = val & 0xF;
@@ -120,7 +121,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 
 							//allow the lenctr_en to kill the len_cnt
 							set_lenctr_en(lenctr_en);
-							
+
 							//serves as a useful note-on diagnostic
 							//if(unit==1) Console.WriteLine("{0} timer_reload_value: {1}", unit, timer_reload_value);
 							break;
@@ -163,7 +164,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 						if (swp_divider_counter == 0)
 						{
 							swp_divider_counter = sweep_divider_cnt + 1;
-						
+
 							//divider was clocked: process sweep pitch bend
 							if (sweep_shiftcount != 0 && !swp_silence)
 							{
@@ -180,7 +181,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 							sweep_reload = false;
 						}
 					}
-					
+
 					//env_loopdoubles as "halt length counter"
 					if (env_loop == 0 && len_cnt > 0)
 						len_cnt--;
@@ -196,7 +197,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 					}
 					else
 					{
-						if(env_divider != 0) env_divider--;
+						if (env_divider != 0) env_divider--;
 						if (env_divider == 0)
 						{
 							env_divider = (env_cnt_value + 1);
@@ -219,7 +220,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 					else env_output = env_counter;
 
 					if (timer_counter > 0) timer_counter--;
-					if (timer_counter == 0 && timer_raw_reload_value!=0)
+					if (timer_counter == 0 && timer_raw_reload_value != 0)
 					{
 						duty_step = (duty_step + 1) & 7;
 						duty_value = PULSE_DUTY[duty_cnt, duty_step] == 1;
@@ -233,8 +234,8 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 						if (swp_silence)
 							sample = 0;
 
-						if (len_cnt==0) //length counter is 0
-						    sample = 0; //silenced
+						if (len_cnt == 0) //length counter is 0
+							sample = 0; //silenced
 					}
 					else
 						sample = 0; //duty cycle is 0, silenced.
@@ -297,13 +298,13 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 						case 0:
 							env_cnt_value = val & 0xF;
 							env_constant = (val >> 4) & 1;
-							env_loop = (val>>5)&1;
+							env_loop = (val >> 5) & 1;
 							break;
 						case 1:
 							break;
 						case 2:
 							period_cnt = NOISE_TABLE[val & 0xF];
-							mode_cnt = (val>>7)&1;
+							mode_cnt = (val >> 7) & 1;
 							//Console.WriteLine("noise period: {0}, vol: {1}", (val & 0xF), env_cnt_value);
 							break;
 						case 3:
@@ -350,7 +351,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				}
 				public void clock_length_and_sweep()
 				{
-				
+
 					if (len_cnt > 0 && env_loop == 0)
 						len_cnt--;
 				}
@@ -373,10 +374,10 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 						shift_register >>= 1;
 						shift_register &= ~(1 << 14);
 						shift_register |= (feedback << 14);
-						noise_bit = (shift_register & 1)!=0;
+						noise_bit = (shift_register & 1) != 0;
 					}
 
-					if (noise_bit || len_cnt==0) sample = 0;
+					if (noise_bit || len_cnt == 0) sample = 0;
 					else
 						sample = env_output;
 				}
@@ -463,21 +464,21 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 					//is clocked in frame counter.
 					if (en)
 					{
-						if(timer>0) timer--;
+						if (timer > 0) timer--;
 						if (timer == 0)
 						{
 							seq = (seq + 1) & 0x1F;
 							timer = timer_cnt_reload;
 						}
 						//if(CFG_DECLICK)
-						    //sample = TRIANGLE_TABLE[(seq+8)&0x1F];
+						//sample = TRIANGLE_TABLE[(seq+8)&0x1F];
 						//else
-							sample = TRIANGLE_TABLE[seq];
+						sample = TRIANGLE_TABLE[seq];
 					}
 
 				}
 
-				
+
 				public void clock_length_and_sweep()
 				{
 					//env_loopdoubles as "halt length counter"
@@ -487,7 +488,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 
 				public void clock_linear_counter()
 				{
-				//	Console.WriteLine("linear_counter: {0}", linear_counter);
+					//	Console.WriteLine("linear_counter: {0}", linear_counter);
 					if (halt_flag == 1)
 					{
 						linear_counter = linear_counter_reload;
@@ -543,7 +544,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 					ser.Sync("irq_enabled", ref irq_enabled);
 					ser.Sync("loop_flag", ref loop_flag);
 					ser.Sync("timer_reload", ref timer_reload);
-					
+
 					ser.Sync("timer", ref timer);
 					ser.Sync("user_address", ref user_address);
 					ser.Sync("user_length", ref user_length);
@@ -597,7 +598,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 						//apu.nes.LogLine("dmc out sample: {0}", out_deltacounter);
 					}
 
-					if (out_bits_remaining==0)
+					if (out_bits_remaining == 0)
 					{
 						out_bits_remaining = 7;
 						if (sample_length > 0)
@@ -619,7 +620,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 
 				public void set_lenctr_en(bool en)
 				{
-					if (!en) 
+					if (!en)
 						//disable playback
 						sample_length = 0;
 					else
@@ -688,12 +689,20 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 
 			public void SyncState(Serializer ser)
 			{
+				ser.Sync("irq_pending", ref irq_pending);
+				ser.Sync("dmc_irq", ref dmc_irq);
+				ser.Sync("pending_reg", ref pending_reg);
+				ser.Sync("pending_val", ref pending_val);
+
 				ser.Sync("sequencer_counter", ref sequencer_counter);
 				ser.Sync("sequencer_step", ref sequencer_step);
 				ser.Sync("sequencer_mode", ref sequencer_mode);
-				ser.Sync("sequencer_irq_inhibit", ref sequencer_irq_inhibit);
+				ser.Sync("sequencer_irq_inhibit;", ref sequencer_irq_inhibit);
 				ser.Sync("sequencer_irq", ref sequencer_irq);
-				ser.Sync("dmc_irq", ref dmc_irq);
+				ser.Sync("sequence_reset_pending", ref sequence_reset_pending);
+				ser.Sync("sequencer_irq_clear_pending", ref sequencer_irq_clear_pending);
+				ser.Sync("sequencer_irq_assert", ref sequencer_irq_assert);
+	
 				pulse[0].SyncState(ser);
 				pulse[1].SyncState(ser);
 				triangle.SyncState(ser);
@@ -705,115 +714,153 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 			TriangleUnit triangle = new TriangleUnit();
 			NoiseUnit noise = new NoiseUnit();
 			DMCUnit dmc;
-			bool sequencer_irq;
-			public bool dmc_irq;
+
+			bool irq_pending;
+			bool dmc_irq;
+			int pending_reg = -1;
+			byte pending_val = 0;
+
+			int sequencer_counter, sequencer_step, sequencer_mode, sequencer_irq_inhibit;
+			bool sequencer_irq, sequence_reset_pending, sequencer_irq_clear_pending, sequencer_irq_assert;
 
 			public void RunDMCFetch()
 			{
 				dmc.Fetch();
 			}
 
-			int sequencer_counter, sequencer_step, sequencer_mode, sequencer_irq_inhibit;
 			void sequencer_reset()
 			{
 				sequencer_counter = 0;
 
 				if (sequencer_mode == 1)
 				{
-					sequencer_step = 5;
-					sequencer_check();
+					sequencer_step = 0;
+					QuarterFrame();
+					HalfFrame();
 				}
 				else
-					sequencer_step = 1;
+					sequencer_step = 0;
 			}
 
-			//21477272 master clock
-			//1789772 cpu clock (master / 12)
-			//240 apu clock (master / 89490) = (cpu / 7457)
+			//these figures are not valid for PAL. they must be recalculated with nintendulator's values above
+			//these values (the NTSC at least) are derived from nintendulator. they are all 2 higher than the specifications, due to some shortcoming in the emulation
+			//this is probably a hint that we're doing something a little wrong but making up for it with curcuitous chaos in other ways
+			static int[][] sequencer_lut = new int[][]{
+				new int[]{7458,14914,22372,29830},
+				new int[]{7458,14914,22372,29830,37282}
+			};
+
 			void sequencer_tick()
 			{
 				sequencer_counter++;
-				//this figure is not valid for PAL. it must be recalculated
-				if (sequencer_counter != 7457) return;
-				sequencer_counter = 0;
-				sequencer_step++;
+				if (sequence_reset_pending)
+				{
+					sequencer_reset();
+					sequence_reset_pending = false;
+				}
+				if (sequencer_lut[sequencer_mode][sequencer_step] != sequencer_counter)
+					return;
 				sequencer_check();
 			}
 
 			public void SyncIRQ()
 			{
-				nes.irq_apu = sequencer_irq | dmc_irq;
-				//if (nes.irq_apu) Console.WriteLine("apu irq");
-				nes.sync_irq();
+				irq_pending = sequencer_irq | dmc_irq;
 			}
 
 			void sequencer_check()
 			{
 				//Console.WriteLine("sequencer mode {0} step {1}", sequencer_mode, sequencer_step);
+				bool quarter, half, reset;
 				switch (sequencer_mode)
 				{
 					case 0: //4-step
-						pulse[0].clock_env();
-						pulse[1].clock_env();
-						triangle.clock_linear_counter();
-						noise.clock_env();
-						if (sequencer_step == 2 || sequencer_step == 4)
+						quarter = true;
+						half = (sequencer_step == 1 || sequencer_step == 3);
+						reset = sequencer_step == 3;
+						if (reset && sequencer_irq_inhibit == 0)
 						{
-							pulse[0].clock_length_and_sweep();
-							pulse[1].clock_length_and_sweep();
-							triangle.clock_length_and_sweep();
-							noise.clock_length_and_sweep();
-						}
-						if (sequencer_step == 4)
-						{
-							if (sequencer_irq_inhibit == 0)
-							{
-								sequencer_irq = true;
-								SyncIRQ();
-							}
-							sequencer_step = 1;
+							//Console.WriteLine("{0} {1,5} set irq_assert", nes.Frame, sequencer_counter);
+							sequencer_irq_assert = true;
 						}
 						break;
+
 					case 1: //5-step
-						if (sequencer_step != 4)
-						{
-							pulse[0].clock_env();
-							pulse[1].clock_env();
-							triangle.clock_linear_counter();
-							noise.clock_env();
-						}
-						if (sequencer_step == 2 || sequencer_step == 5)
-						{
-							pulse[0].clock_length_and_sweep();
-							pulse[1].clock_length_and_sweep();
-							triangle.clock_length_and_sweep();
-							noise.clock_length_and_sweep();
-						}
-						if (sequencer_step == 5)
-							sequencer_step = 1;
+						quarter = sequencer_step != 3;
+						half = (sequencer_step == 1 || sequencer_step == 4);
+						reset = sequencer_step == 4;
 						break;
+
+					default:
+						throw new InvalidOperationException();
 				}
+
+				if (reset)
+				{
+					sequencer_counter = 0;
+					sequencer_step = 0;
+				}
+				else sequencer_step++;
+
+				if (quarter) QuarterFrame();
+				if (half) HalfFrame();
 			}
 
+			void HalfFrame()
+			{
+				pulse[0].clock_length_and_sweep();
+				pulse[1].clock_length_and_sweep();
+				triangle.clock_length_and_sweep();
+				noise.clock_length_and_sweep();
+			}
+
+			void QuarterFrame()
+			{
+				pulse[0].clock_env();
+				pulse[1].clock_env();
+				triangle.clock_linear_counter();
+				noise.clock_env();
+			}
 
 			public void WriteReg(int addr, byte val)
+			{
+				pending_reg = addr;
+				pending_val = val;
+			}
+
+			void _WriteReg(int addr, byte val)
 			{
 				//Console.WriteLine("{0:X4} = {1:X2}", addr, val);
 				switch (addr)
 				{
-					case 0x4000: case 0x4001: case 0x4002: case 0x4003:
+					case 0x4000:
+					case 0x4001:
+					case 0x4002:
+					case 0x4003:
 						pulse[0].WriteReg(addr - 0x4000, val);
 						break;
-					case 0x4004: case 0x4005: case 0x4006: case 0x4007:
+					case 0x4004:
+					case 0x4005:
+					case 0x4006:
+					case 0x4007:
 						pulse[1].WriteReg(addr - 0x4004, val);
 						break;
-					case 0x4008: case 0x4009: case 0x400A: case 0x400B:
+					case 0x4008:
+					case 0x4009:
+					case 0x400A:
+					case 0x400B:
 						triangle.WriteReg(addr - 0x4008, val);
 						break;
-					case 0x400C: case 0x400D: case 0x400E: case 0x400F:
+					case 0x400C:
+					case 0x400D:
+					case 0x400E:
+					case 0x400F:
 						noise.WriteReg(addr - 0x400C, val);
 						break;
-					case 0x4010: case 0x4011: case 0x4012: case 0x4013:
+					case 0x4010:
+					case 0x4011:
+					case 0x4012:
+					case 0x4013:
 						dmc.WriteReg(addr - 0x4010, val);
 						break;
 					case 0x4015:
@@ -824,14 +871,14 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 						dmc.set_lenctr_en(val.Bit(4));
 						break;
 					case 0x4017:
-						sequencer_mode = (val>>7)&1;
+						//Console.WriteLine("apu 4017 = {0:X2}", val);
+						sequencer_mode = (val >> 7) & 1;
 						sequencer_irq_inhibit = (val >> 6) & 1;
 						if (sequencer_irq_inhibit == 1)
 						{
-							sequencer_irq = false;
-							SyncIRQ();
+							sequencer_irq_clear_pending = true;
 						}
-						sequencer_reset();
+						sequence_reset_pending = true;
 						break;
 				}
 			}
@@ -841,36 +888,32 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				switch (addr)
 				{
 					case 0x4015:
-					{
-						//notice a missing bit here. should properly emulate with empty bus
-						//if an interrupt flag was set at the same moment of the read, it will read back as 1 but it will not be cleared. 
-						int dmc_nonzero = dmc.IsLenCntNonZero() ? 1 : 0;
-						int noise_nonzero = noise.IsLenCntNonZero() ? 1 : 0;
-						int tri_nonzero = triangle.IsLenCntNonZero() ? 1 : 0;
-						int pulse1_nonzero = pulse[1].IsLenCntNonZero() ? 1 : 0;
-						int pulse0_nonzero = pulse[0].IsLenCntNonZero() ? 1 : 0;
-						int ret = ((dmc_irq?1:0) << 7) | ((sequencer_irq?1:0) << 6) | (dmc_nonzero << 4) | (noise_nonzero << 3) | (tri_nonzero<<2) | (pulse1_nonzero<<1) | (pulse0_nonzero);
-						sequencer_irq = false;
-						SyncIRQ();
-						return (byte)ret;
-					}
+						{
+							//notice a missing bit here. should properly emulate with empty bus
+							//if an interrupt flag was set at the same moment of the read, it will read back as 1 but it will not be cleared. 
+							int dmc_nonzero = dmc.IsLenCntNonZero() ? 1 : 0;
+							int noise_nonzero = noise.IsLenCntNonZero() ? 1 : 0;
+							int tri_nonzero = triangle.IsLenCntNonZero() ? 1 : 0;
+							int pulse1_nonzero = pulse[1].IsLenCntNonZero() ? 1 : 0;
+							int pulse0_nonzero = pulse[0].IsLenCntNonZero() ? 1 : 0;
+							int ret = ((dmc_irq ? 1 : 0) << 7) | ((sequencer_irq ? 1 : 0) << 6) | (dmc_nonzero << 4) | (noise_nonzero << 3) | (tri_nonzero << 2) | (pulse1_nonzero << 1) | (pulse0_nonzero);
+							//Console.WriteLine("{0} {1,5} $4015 clear irq, was at {2}", nes.Frame, sequencer_counter, sequencer_irq);
+							sequencer_irq = false;
+							SyncIRQ();
+							return (byte)ret;
+						}
 					default:
 						//don't return 0xFF here or SMB will break
 						return 0x00;
 				}
 			}
 
-			public void Run(int cycles)
-			{
-				for (int i = 0; i < cycles; i++)
-					RunOne();
-			}
-
 			public void DiscardSamples()
 			{
 				metaspu.buffer.clear();
 			}
 
+			int toggle = 0;
 			public void RunOne()
 			{
 				pulse[0].Run();
@@ -883,12 +926,42 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				mix += pulse[0].sample;
 				mix += pulse[1].sample;
 				mix += triangle.sample;
-				mix += noise.sample>>1;
+				mix += noise.sample >> 1;
 				mix += dmc.sample;
 
 				EmitSample(mix);
 
+				//this (and the similar line below) is a crude hack
+				//we should be generating logic to suppress the $4015 clear when the assert signal is set instead
+				//be sure to test "apu_test" if you mess with this
+				sequencer_irq |= sequencer_irq_assert;
+
+				if (toggle == 0)
+				{
+					//handle sequencer irq clear signal
+					sequencer_irq_assert = false;
+					if (sequencer_irq_clear_pending)
+					{
+						//Console.WriteLine("{0} {1,5} $4017 clear irq (delayed)", nes.Frame, sequencer_counter);
+						sequencer_irq_clear_pending = false;
+						sequencer_irq = false;
+						SyncIRQ();
+					}
+
+					//handle writes from the odd clock cycle
+					if (pending_reg != -1) _WriteReg(pending_reg, pending_val);
+					pending_reg = -1;
+					toggle = 1;
+
+					//latch whatever irq logic we had and send to cpu
+					nes.irq_apu = irq_pending;
+				}
+				else toggle = 0;
+
 				sequencer_tick();
+				sequencer_irq |= sequencer_irq_assert;
+				SyncIRQ();
+
 				//since the units run concurrently, the APU frame sequencer is ran last because
 				//it can change the ouput values of the pulse/triangle channels
 				//we want the changes to affect it on the *next* cycle.
@@ -909,9 +982,9 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 					samp = 0;
 				else
 					panic_sample = samp;
-				
+
 				int this_samp = samp;
-				const double kMixRate = 44100.0/1789772.0;
+				const double kMixRate = 44100.0 / 1789772.0;
 				const double kInvMixRate = (1 / kMixRate);
 				timer += kMixRate;
 				accumulate += samp;
@@ -922,7 +995,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				timer -= 1;
 				double ratio = (timer / kMixRate);
 				double fractional = (this_samp - last_hwsamp) * ratio;
-				double factional_remainder = (this_samp - last_hwsamp) * (1-ratio);
+				double factional_remainder = (this_samp - last_hwsamp) * (1 - ratio);
 				accumulate += fractional;
 
 				accumulate *= 436; //32768/(15*4) -- adjust later for other sound channels
@@ -936,7 +1009,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 			}
 
 			MetaspuSoundProvider metaspu = new MetaspuSoundProvider(ESynchMethod.ESynchMethod_V);
-            public int MaxVolume { get; set; } // not supported
+			public int MaxVolume { get; set; } // not supported
 
 			void ISoundProvider.GetSamples(short[] samples)
 			{
@@ -954,14 +1027,14 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 			{
 				//Console.WriteLine("a: {0} with todo: {1}",squeue.Count,samples.Length/2);
 
-				for (int i = 0; i < samples.Length/2; i++)
+				for (int i = 0; i < samples.Length / 2; i++)
 				{
 					int samp = 0;
 					if (squeue.Count != 0)
 						samp = squeue.Dequeue();
-					
-					samples[i*2+0] = (short)(samp);
-					samples[i*2+1] = (short)(samp);
+
+					samples[i * 2 + 0] = (short)(samp);
+					samples[i * 2 + 1] = (short)(samp);
 					//bw.Write((short)samp);
 				}
 			}
diff --git a/BizHawk.Emulation/Consoles/Nintendo/NES/Core.cs b/BizHawk.Emulation/Consoles/Nintendo/NES/Core.cs
index aa6ced4e41..436f654044 100644
--- a/BizHawk.Emulation/Consoles/Nintendo/NES/Core.cs
+++ b/BizHawk.Emulation/Consoles/Nintendo/NES/Core.cs
@@ -23,7 +23,10 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 		CartInfo cart; //the current cart prototype. should be moved into the board, perhaps
 		INESBoard board; //the board hardware that is currently driving things
 		public bool SoundOn = true;
-		bool _irq_apu, _irq_cart;
+		int sprdma_countdown; //used to 
+		bool _irq_apu, _irq_cart; //various irq signals that get merged to the cpu irq pin
+
+		//irq state management
 		public bool irq_apu { get { return _irq_apu; } set { _irq_apu = value; sync_irq(); } }
 		public bool irq_cart { get { return _irq_cart; } set { _irq_cart = value; sync_irq(); } }
 		void sync_irq()
@@ -95,10 +98,6 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 		static ByteBuffer cpu_sequence_NTSC = new ByteBuffer(new byte[]{3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3});
 		static ByteBuffer cpu_sequence_PAL = new ByteBuffer(new byte[]{4,3,3,3,3,4,3,3,3,3,4,3,3,3,3,4,3,3,3,3,4,3,3,3,3,4,3,3,3,3,4,3,3,3,3,4,3,3,3,3});
 		public int cpu_step, cpu_stepcounter, cpu_deadcounter;
-		public void DmaTimingHack(int amount)
-		{
-			cpu_deadcounter += amount;
-		}
 		protected void RunCpuOne()
 		{
 			cpu_stepcounter++;
@@ -107,10 +106,24 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				cpu_step++;
 				cpu_step &= 31;
 				cpu_stepcounter = 0;
-				if (cpu_deadcounter == 0)
+
+				if (sprdma_countdown > 0)
+				{
+					sprdma_countdown--;
+					if (sprdma_countdown == 0)
+					{
+						//its weird that this is 514.. normally itd be 512 (and people would say its wrong) or 513 (and people would say its right)
+						//but 514 passes test 4-irq_and_dma
+						cpu_deadcounter = 514;
+					}
+				}
+
+				if(cpu_deadcounter>0)
+					cpu_deadcounter--;
+				else
 					cpu.ExecuteOne();
-				else cpu_deadcounter--;
-				if (SoundOn) apu.RunOne(); //THIS ISNT SAFE!!!!!!!!!
+
+				if (SoundOn) apu.RunOne(); //THIS ISNT SAFE!!!!!!!!! SOUND MUST ALWAYS RUN!!!!
 				ppu.PostCpuInstructionOne();
 			}
 		}
@@ -176,10 +189,10 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 		{
 			//read joystick port
 			//many todos here
-            lagged = false;
+			lagged = false;
 			byte ret;
 			if(addr == 0x4016)
-                ret = ports[0].Read();
+				ret = ports[0].Read();
 			else ret = ports[1].Read();
 			return ret;
 		}
@@ -193,7 +206,9 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				WriteMemory(0x2004, db);
 				addr++;
 			}
-			DmaTimingHack(513);
+			//schedule a sprite dma event for beginning 1 cycle in the future.
+			//this receives 2 because thats just the way it works out.
+			sprdma_countdown = 2;
 		}
 
 		/// <summary>
diff --git a/BizHawk.Emulation/Consoles/Nintendo/NES/NES.cs b/BizHawk.Emulation/Consoles/Nintendo/NES/NES.cs
index beb09dd396..8c2b464414 100644
--- a/BizHawk.Emulation/Consoles/Nintendo/NES/NES.cs
+++ b/BizHawk.Emulation/Consoles/Nintendo/NES/NES.cs
@@ -597,9 +597,11 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 			ser.Sync("cpu_accumulate", ref cpu_accumulate);
 			ser.Sync("_irq_apu", ref _irq_apu);
 			ser.Sync("_irq_cart", ref _irq_cart);
+			ser.Sync("sprdma_countdown", ref sprdma_countdown);
+			ser.Sync("cpu_step", ref cpu_step);
+			ser.Sync("cpu_stepcounter", ref cpu_stepcounter);
+			ser.Sync("cpu_deadcounter", ref cpu_deadcounter);
 			sync_irq();
-			//string inp = GetControllersAsMnemonic();  TODO sorry bout that
-			//ser.SyncFixedString("input", ref inp, 32);
 			board.SyncState(ser);
 			ppu.SyncState(ser);
 			apu.SyncState(ser);
diff --git a/BizHawk.Emulation/Consoles/Nintendo/NES/PPU.cs b/BizHawk.Emulation/Consoles/Nintendo/NES/PPU.cs
index 33f39b5569..96fe6dff0b 100644
--- a/BizHawk.Emulation/Consoles/Nintendo/NES/PPU.cs
+++ b/BizHawk.Emulation/Consoles/Nintendo/NES/PPU.cs
@@ -95,6 +95,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				ser.Sync("PPUGenLatch", ref PPUGenLatch);
 				ser.Sync("vtoggle", ref vtoggle);
 				ser.Sync("VRAMBuffer", ref VRAMBuffer);
+				ser.Sync("ppu_addr_temp", ref ppu_addr_temp);
 
 				ser.Sync("OAM", ref OAM, false);
 				ser.Sync("PALRAM", ref PALRAM, false);
diff --git a/BizHawk.Emulation/Consoles/Nintendo/NES/PPU.run.cs b/BizHawk.Emulation/Consoles/Nintendo/NES/PPU.run.cs
index 3787d9c818..1dc84848ae 100644
--- a/BizHawk.Emulation/Consoles/Nintendo/NES/PPU.run.cs
+++ b/BizHawk.Emulation/Consoles/Nintendo/NES/PPU.run.cs
@@ -27,39 +27,67 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 
 			public short[] xbuf = new short[256*240];
 
-			void Read_bgdata(ref BGDataRecord bgdata) {
-				int addr = ppur.get_ntread();
-				bgdata.nt = ppubus_read(addr, true);
-				runppu(kFetchTime);
-
-				addr = ppur.get_atread();
-				byte at = ppubus_read(addr, true);
-
-				//modify at to get appropriate palette shift
-				if((ppur.vt&2)!=0) at >>= 4;
-				if((ppur.ht&2)!=0) at >>= 2;
-				at &= 0x03;
-				at <<= 2;
-			
-				bgdata.at = at;
-
-				//horizontal scroll clocked at cycle 3 and then
-				//vertical scroll at 251
-				runppu(1);
-				if (reg_2001.PPUON)
+			int ppu_addr_temp;
+			void Read_bgdata(ref BGDataRecord bgdata)
+			{
+				for (int i = 0; i < 8; i++)
+					Read_bgdata(i,ref bgdata);
+			}
+			void Read_bgdata(int cycle, ref BGDataRecord bgdata)
+			{
+				switch (cycle)
 				{
-					ppur.increment_hsc();
-					if (ppur.status.cycle == 251)
-						ppur.increment_vs();
-				}
-				runppu(1);
+					case 0:
+						ppu_addr_temp = ppur.get_ntread();
+						bgdata.nt = ppubus_read(ppu_addr_temp, true);
+						runppu(1);
+						break;
+					case 1:
+						runppu(1);
+						break;
+					case 2:
+						{
+							ppu_addr_temp = ppur.get_atread();
+							byte at = ppubus_read(ppu_addr_temp, true);
 
-				addr = ppur.get_ptread(bgdata.nt);
-				bgdata.pt_0 = ppubus_read(addr, true);
-				runppu(kFetchTime);
-				addr |= 8;
-				bgdata.pt_1 = ppubus_read(addr, true);
-				runppu(kFetchTime);
+							//modify at to get appropriate palette shift
+							if ((ppur.vt & 2) != 0) at >>= 4;
+							if ((ppur.ht & 2) != 0) at >>= 2;
+							at &= 0x03;
+							at <<= 2;
+							bgdata.at = at;
+
+							//horizontal scroll clocked at cycle 3 and then
+							//vertical scroll at 251
+							runppu(1);
+							if (reg_2001.PPUON)
+							{
+								ppur.increment_hsc();
+								if (ppur.status.cycle == 251)
+									ppur.increment_vs();
+							}
+							break;
+						}
+					case 3:
+						runppu(1);
+						break;
+					case 4:
+						ppu_addr_temp = ppur.get_ptread(bgdata.nt);
+						bgdata.pt_0 = ppubus_read(ppu_addr_temp, true);
+						runppu(1);
+						break;
+					case 5:
+						runppu(1);
+						break;
+					case 6:
+						ppu_addr_temp |= 8;
+						bgdata.pt_1 = ppubus_read(ppu_addr_temp, true);
+						runppu(1);
+						break;
+					case 7:
+						runppu(1);
+						break;
+				} //switch(cycle)
 			}
 
 			unsafe struct TempOAM
@@ -124,12 +152,6 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 				{
 					ppur.status.cycle = 0;
 
-					//if (!reg_2001.PPUON)
-					//{
-					//    runppu(kLineTime);
-					//    continue;
-					//}
-
 					ppur.status.sl = sl;
 
 					int yp = sl - 1;
@@ -150,12 +172,12 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 					//two of those tiles were read in the last scanline.
 					for (int xt = 0; xt < 32; xt++)
 					{
-						Read_bgdata(ref bgdata[xt + 2]);
-
 						//ok, we're also going to draw here.
 						//unless we're on the first dummy scanline
 						if (sl != 0)
 						{
+
+
 							int xstart = xt << 3;
 							oamcount = oamcounts[renderslot];
 							int target = (yp << 8) + xstart;
@@ -167,6 +189,10 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 
 							for (int xp = 0; xp < 8; xp++, rasterpos++)
 							{
+								//process the current clock's worth of bg data fetching
+								//this needs to be split into 8 pieces or else exact sprite 0 hitting wont work due to the cpu not running while the sprite renders below
+								Read_bgdata(xp, ref bgdata[xt + 2]);
+
 								//bg pos is different from raster pos due to its offsetability.
 								//so adjust for that here
 								int bgpos = rasterpos + ppur.fh;
@@ -187,18 +213,27 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 								}
 								else
 								{
+									if (!renderspritenow)
+									{
+										//according to qeed's doc, use palette 0 or $2006's value if it is & 0x3Fxx
+										int addr = ppur.get_2007access();
+										if ((addr & 0x3F00) == 0x3F00)
+										{
+											pixel = addr & 0x1F;
+										}
+									}
 									pixelcolor = PALRAM[pixel];
-									pixelcolor |= 0x8000;
+									pixelcolor |= 0x8000; //whats this? i think its a flag to indicate a hidden background to be used by the canvas filling logic later
 								}
 
 								if (!nes.CoreInputComm.NES_ShowBG)
-									pixelcolor = 0x8000;
+									pixelcolor = 0x8000; //whats this? i think its a flag to indicate a hidden background to be used by the canvas filling logic later
 
 								//look for a sprite to be drawn
 								bool havepixel = false;
 								for (int s = 0; s < oamcount; s++)
 								{
-									TempOAM *oam = &oams[(renderslot<<6)+s];
+									TempOAM* oam = &oams[(renderslot << 6) + s];
 									{
 										int x = oam->oam[3];
 										if (rasterpos >= x && rasterpos < x + 8)
@@ -241,7 +276,7 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 												if ((pixel & 3) != 0)
 												{
 													drawsprite = false;
-												}	
+												}
 											}
 
 											if (drawsprite && nes.CoreInputComm.NES_ShowOBJ)
@@ -258,12 +293,20 @@ namespace BizHawk.Emulation.Consoles.Nintendo
 
 								}//oamcount loop
 
+
+								if (reg_2001.color_disable)
+									pixelcolor &= 0x30;
+
 								xbuf[target] = PaletteAdjustPixel(pixelcolor);
 
 								target++;
 
 							} //loop across 8 pixels
 						} //scanline != 0
+						else
+						{
+							Read_bgdata(ref bgdata[xt + 2]);
+						}
 					} //loop across 32 tiles