From fa8251d31dd061abf3b2740b3700f288de38c28b Mon Sep 17 00:00:00 2001
From: "rajshekar.k" <rajshekar.k@iitdh.ac.in>
Date: Wed, 20 Jan 2021 11:50:21 +0530
Subject: [PATCH] support added for predicated instructions that are not
 executed

---
 src/emulator/pin/causalityTool.cpp            |  17 +-
 src/emulator/pin/encoding.h                   |   2 +
 src/simulator/config/config.xml               | 885 ++++++++++--------
 .../DynamicInstructionBuffer.java             |  12 +
 .../emulatorinterface/RunnableThread.java     |   3 +-
 .../communication/Encoding.java               |   2 +
 .../translator/visaHandler/IntegerALU.java    |   2 +-
 .../x86/instruction/ConditionalMove.java      |   8 +-
 .../x86/instruction/ConditionalSet.java       |   2 +-
 src/simulator/generic/Instruction.java        |  16 +
 .../pipeline/outoforder/ReorderBuffer.java    |   2 +-
 11 files changed, 534 insertions(+), 417 deletions(-)

diff --git a/src/emulator/pin/causalityTool.cpp b/src/emulator/pin/causalityTool.cpp
index 72d9b58..6c707dd 100755
--- a/src/emulator/pin/causalityTool.cpp
+++ b/src/emulator/pin/causalityTool.cpp
@@ -505,7 +505,7 @@ VOID BarrierInit(ADDRINT first_arg, ADDRINT val, UINT32 encode, THREADID tid) {
         }
 }
 /*** This function is called on every instruction ***/
-VOID printip(THREADID tid, VOID *ip, char *asmString) {
+VOID printip(THREADID tid, VOID *ip, char *asmString, bool predicateInsnNotExecuted) {
 
 	if(outOfROIPhase)
 	{
@@ -612,8 +612,17 @@ VOID printip(THREADID tid, VOID *ip, char *asmString) {
 		uint64_t nip = MASK & (uint64_t) ip;
 
 		if(traceMethod==SharedMemory) {
-			while (tst->analysisFn(tid, nip, INSTRUCTION, 1) == -1) {
-				PIN_Yield();
+			if(predicateInsnNotExecuted)
+			{
+				while (tst->analysisFn(tid, nip, PREDICATE_INSN_NOT_EXECUTED, 1) == -1) {
+					PIN_Yield();
+				}
+			}
+			else
+			{
+				while (tst->analysisFn(tid, nip, INSTRUCTION, 1) == -1) {
+					PIN_Yield();
+				}
 			}
 		} else if(traceMethod==File) {
 			while (tst->analysisFnAssembly(tid, nip, ASSEMBLY, asmString) == -1) {
@@ -674,7 +683,7 @@ VOID Instruction(INS ins, VOID *v) {
 		asmChar = (char *)asmString->c_str();
 	}
 
-	INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)printip, IARG_THREAD_ID, IARG_INST_PTR, IARG_PTR, asmChar, IARG_END);
+	INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)printip, IARG_THREAD_ID, IARG_INST_PTR, IARG_PTR, asmChar, IARG_EXECUTING, IARG_END);
 
 
 	UINT32 memOperands = INS_MemoryOperandCount(ins);
diff --git a/src/emulator/pin/encoding.h b/src/emulator/pin/encoding.h
index e5bfa6a..c135bc3 100755
--- a/src/emulator/pin/encoding.h
+++ b/src/emulator/pin/encoding.h
@@ -31,5 +31,7 @@
 #define PARENT_SPAWN 35
 #define CHILD_START 36
 
+#define PREDICATE_INSN_NOT_EXECUTED 37
+
 
 const char* findType(int type);
diff --git a/src/simulator/config/config.xml b/src/simulator/config/config.xml
index a3e0ddf..ffa48bb 100755
--- a/src/simulator/config/config.xml
+++ b/src/simulator/config/config.xml
@@ -19,9 +19,11 @@
 	Contributors:  Moksh Upadhyay, Abhishek Sagar, Prathmesh Kallurkar
 *****************************************************************************
 
-based on Intel® Core™ i7-7820X X-series Processor
-TDP = 112W
-52.5 mm x 45 mm
+based on Intel® Core™ i5-7200U (Kabylake) Processor
+https://en.wikichip.org/wiki/intel/core_i5/i5-7200u
+https://www.agner.org/optimize/instruction_tables.pdf
+TDP = 15W
+42 mm * 24 mm
 /-->
 <Configuration>
 	<Emulator>
@@ -45,13 +47,14 @@ TDP = 112W
 		<!-- If store packets in a file option is set to true, this parameter indicates the basename for the trace files -->
 		<!--One trace file is maintained for each store. The name of trace file for core n is basename_n.gz-->
 		<!--We do not allow overwriting of trace files. So if a tracefile with same name is pre-existing, kindly rename it-->
-		<BasenameForTraceFiles>/home/prathmesh/tejasupdate/Tejas-dram/test/helloworld_trace</BasenameForTraceFiles>
+		<BasenameForTraceFiles>/home/rajshekar/tmp/gcc_trace</BasenameForTraceFiles>
 		
 		<PinTool>/home/rajshekar/softwares/pin-97554/</PinTool>
-		<PinInstrumentor>/home/rajshekar/projects/nvms/workspace/Tejas/src/emulator/pin/obj-pin/causalityTool.so</PinInstrumentor>
+		<PinInstrumentor>/home/rajshekar/projects/tejas/workspace/Tejas/src/emulator/pin/obj-pin/causalityTool.so</PinInstrumentor>
 		<QemuTool>TODO/home/prathmesh/workspace/qemu/x86_64-linux-user/qemu-x86_64 /home/prathmesh/tmp/testQemu.o</QemuTool>
-		<ShmLibDirectory>/home/rajshekar/projects/nvms/workspace/Tejas/src/emulator/pin/obj-comm</ShmLibDirectory>
-		<KillEmulatorScript>/home/rajshekar/projects/nvms/workspace/Tejas/src/simulator/main/killAllDescendents.sh</KillEmulatorScript>
+		<ShmLibDirectory>/home/rajshekar/resources/tejas_configs/</ShmLibDirectory>
+		<GetBenchmarkPIDScript>/home/rajshekar_resources/tejas_configs/getBenchmarkPID.sh</GetBenchmarkPIDScript>
+		<KillEmulatorScript>/home/rajshekar/resources/tejas_configs/killAllDescendents.sh</KillEmulatorScript>
 	</Emulator>
 	
 	<!--Simulation Parameters-->
@@ -71,23 +74,23 @@ TDP = 112W
 		<DetachMemSysInsn>false</DetachMemSysInsn>
 		<PrintPowerStats>true</PrintPowerStats>
 		<Broadcast>false</Broadcast>
-		<pinpointsSim>true</pinpointsSim>
-		<pinpointsFile>/home/rajshekar/benchmarks/cpu2006/PinPoints/perlbench.ref.t.sorted</pinpointsFile>
-		<NumInsToIgnore>0</NumInsToIgnore>	<!--Ignores these many profilable instructions from the start of the program-->
-		<subsetSim>true</subsetSim>
-		<subsetSimSize>1000000000</subsetSimSize>
+		<pinpointsSim>false</pinpointsSim>
+		<pinpointsFile>/mnt/srishtistr0/scratch/rajshekar/tejas/PinPoints_working_directory/soplex.test.Data/t.sorted</pinpointsFile>
+		<NumInsToIgnore>00000000</NumInsToIgnore>	<!--Ignores these many profilable instructions from the start of the program-->
+		<subsetSim>false</subsetSim>
+		<subsetSimSize>2000000</subsetSimSize>
 		<markerFunctions>false</markerFunctions>
-		<startSimMarker>XXX_startInstrumentation</startSimMarker>
-		<endSimMarker>XXX_endInstrumentation</endSimMarker>
-		<NumCores>8</NumCores>
+		<startSimMarker>add</startSimMarker>
+		<endSimMarker>sub</endSimMarker>
+		<NumCores>2</NumCores>
 	</Simulation>
 
 	<!--System Parameters-->
 	<System>
 		<MainMemory>
 			<MemControllerToUse>SIMPLE</MemControllerToUse>       <!-- Set the value as DRAM to enable DRAM else use SIMPLE to disable DRAM -->
-			<MainMemoryLatency>100</MainMemoryLatency>		<!--The latency of main memory (in clock cycles)-->
-			<MainMemoryFrequency>2100</MainMemoryFrequency>		<!--Operating frequency of the main memory (in MHz)-->
+			<MainMemoryLatency>132</MainMemoryLatency>		<!--The latency of main memory (in clock cycles)-->
+			<MainMemoryFrequency>2400</MainMemoryFrequency>		<!--Operating frequency of the main memory (in MHz)-->
 			<MainMemoryPortType>FCFS</MainMemoryPortType>		<!--Type of access ports in the Main Memory (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
 			<MainMemoryAccessPorts>1</MainMemoryAccessPorts>	<!--Number of access ports in the Main Memory-->
 			<MainMemoryPortOccupancy>1</MainMemoryPortOccupancy>	<!--The occupancy of the Main Memory ports (in clock cycles)-->
@@ -104,8 +107,8 @@ TDP = 112W
 			
 		<!--Core Parameters-->
 		<Core>
-			<CoreNumber>0-7</CoreNumber>
-			<CoreFrequency>4200</CoreFrequency>		<!--Operating frequency of the core (in MHz)-->
+			<CoreNumber>0-1</CoreNumber>
+			<CoreFrequency>2500</CoreFrequency>		<!--Operating frequency of the core (in MHz)-->
 			<PipelineType>outOfOrder</PipelineType>	<!--inOrder,outOfOrder(set issue width for multi-issue in-order)-->
 			
 			<BranchPredictor>
@@ -130,27 +133,38 @@ TDP = 112W
 			</LSQ>
 
 			<ITLB>
-				<Size>128</Size>				<!--Maximum number of entries in the ITLB-->
+				<Size>128</Size>				<!--Maximum number of entries in the TLB-->
 				<Latency>1</Latency>			<!--In clock cycles-->
-				<MissPenalty>10</MissPenalty>			<!--In clock cycles-->
-				<PortType>UL</PortType>			<!--Type of access ports in the ITLB (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-				<AccessPorts>-1</AccessPorts>		<!--Number of access ports in the ITLB-->
-				<PortOccupancy>-1</PortOccupancy>		<!--The occupancy of the ITLB ports (in clock cycles)-->
+				<MissPenalty>-1</MissPenalty>			<!--In clock cycles; -1 indicates there is another TLB level below-->
+				<PortType>UL</PortType>			<!--Type of access ports in the TLB (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+				<AccessPorts>-1</AccessPorts>		<!--Number of access ports in the TLB-->
+				<PortOccupancy>-1</PortOccupancy>		<!--The occupancy of the TLB ports (in clock cycles)-->
 				<LeakageEnergy>0.00546275</LeakageEnergy>
 				<DynamicEnergy>0.06792852941</DynamicEnergy>
 			</ITLB>
 			
 			<DTLB>
-				<Size>64</Size>				<!--Maximum number of entries in the DTLB-->
+				<Size>64</Size>				<!--Maximum number of entries in the TLB-->
 				<Latency>1</Latency>			<!--In clock cycles-->
-				<MissPenalty>10</MissPenalty>			<!--In clock cycles-->
-				<PortType>UL</PortType>			<!--Type of access ports in the ITLB (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-				<AccessPorts>-1</AccessPorts>		<!--Number of access ports in the ITLB-->
-				<PortOccupancy>-1</PortOccupancy>		<!--The occupancy of the ITLB ports (in clock cycles)-->
+				<MissPenalty>-1</MissPenalty>			<!--In clock cycles; -1 indicates there is another TLB level below-->
+				<PortType>UL</PortType>			<!--Type of access ports in the TLB (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+				<AccessPorts>-1</AccessPorts>		<!--Number of access ports in the TLB-->
+				<PortOccupancy>-1</PortOccupancy>		<!--The occupancy of the TLB ports (in clock cycles)-->
 				<LeakageEnergy>0.00546275</LeakageEnergy>
 				<DynamicEnergy>0.06792852941</DynamicEnergy>
 			</DTLB>
 			
+			<STLB>							<!-- unified second level TLB -->
+				<Size>1536</Size>				<!--Maximum number of entries in the TLB-->
+				<Latency>9</Latency>			<!--In clock cycles-->
+				<MissPenalty>17</MissPenalty>			<!--In clock cycles; -1 indicates there is another TLB level below-->
+				<PortType>UL</PortType>			<!--Type of access ports in the TLB (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+				<AccessPorts>-1</AccessPorts>		<!--Number of access ports in the TLB-->
+				<PortOccupancy>-1</PortOccupancy>		<!--The occupancy of the TLB ports (in clock cycles)-->
+				<LeakageEnergy>0.00546275</LeakageEnergy>	<!-- TODO this number isn't right -->
+				<DynamicEnergy>0.06792852941</DynamicEnergy>	<!-- TODO this number isn't right -->
+			</STLB>
+			
 			<Decode>								<!--Instruction decode-->
 				<Width>6</Width>			
 				<LeakageEnergy>0.0598</LeakageEnergy>
@@ -188,7 +202,7 @@ TDP = 112W
 			</InstructionWindow>
 			
 			<ROB>
-				<RetireWidth>4</RetireWidth>			<!--Instruction retire width-->
+				<RetireWidth>6</RetireWidth>			<!--Instruction retire width-->
 				<ROBSize>224</ROBSize>				<!--Maximum number of entries in the ROB-->
 				<LeakageEnergy>0.0058</LeakageEnergy>
 				<DynamicEnergy>0.0304</DynamicEnergy>
@@ -197,17 +211,17 @@ TDP = 112W
 			<RegisterFile>
 				<Integer>
 					<IntRegFileSize>180</IntRegFileSize>		<!--Maximum number of entries in the Integer register file-->
-					<IntArchRegNum>32</IntArchRegNum>		<!--Number of Integer architectural registers-->
+					<IntArchRegNum>64</IntArchRegNum>		<!--Number of Integer architectural registers-->
 					<LeakageEnergy>0.0108</LeakageEnergy>
 					<DynamicEnergy>0.0572</DynamicEnergy>
 				</Integer>
 				
-				<Float>
-					<FloatRegFileSize>144</FloatRegFileSize>		<!--Maximum number of entries in the Floating point register file-->
-					<FloatArchRegNum>32</FloatArchRegNum>		<!--Number of Floating point architectural registers-->
+				<Vector>
+					<VectorRegFileSize>168</VectorRegFileSize>		<!--Maximum number of entries in the Floating point register file-->
+					<VectorArchRegNum>64</VectorArchRegNum>		<!--Number of Floating point architectural registers-->
 					<LeakageEnergy>0.0075</LeakageEnergy>
 					<DynamicEnergy>0.0207</DynamicEnergy>
-				</Float>
+				</Vector>
 				
 			</RegisterFile>
 			
@@ -228,7 +242,7 @@ TDP = 112W
 			
 			<IntMul>
 				<Num>1</Num>
-				<Latency>3</Latency>
+				<Latency>4</Latency>
 				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
 				<PortNumber>1</PortNumber>
 				<LeakageEnergy>0.0271</LeakageEnergy>
@@ -237,8 +251,8 @@ TDP = 112W
 			
 			<IntDiv>
 				<Num>1</Num>
-				<Latency>21</Latency>
-				<ReciprocalOfThroughput>12</ReciprocalOfThroughput>
+				<Latency>50</Latency>
+				<ReciprocalOfThroughput>50</ReciprocalOfThroughput>
 				<PortNumber>0</PortNumber>
 				<LeakageEnergy>0.0271</LeakageEnergy>
 				<DynamicEnergy>0.6514</DynamicEnergy>			
@@ -264,8 +278,8 @@ TDP = 112W
 			
 			<FloatDiv>
 				<Num>1</Num>
-				<Latency>24</Latency>
-				<ReciprocalOfThroughput>12</ReciprocalOfThroughput>
+				<Latency>14</Latency>
+				<ReciprocalOfThroughput>4</ReciprocalOfThroughput>
 				<PortNumber>0</PortNumber>
 				<LeakageEnergy>0.0271</LeakageEnergy>
 				<DynamicEnergy>0.6514</DynamicEnergy>			
@@ -273,7 +287,7 @@ TDP = 112W
 			
 			<IntVectorALU>
 				<Num>3</Num>
-				<Latency>3</Latency>
+				<Latency>1</Latency>
 				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
 				<PortNumber>0</PortNumber>
 				<PortNumber>1</PortNumber>
@@ -284,7 +298,7 @@ TDP = 112W
 			
 			<IntVectorMul>
 				<Num>2</Num>
-				<Latency>3</Latency>
+				<Latency>5</Latency>
 				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
 				<PortNumber>0</PortNumber>
 				<PortNumber>1</PortNumber>
@@ -294,7 +308,7 @@ TDP = 112W
 			
 			<FloatVectorALU>
 				<Num>0</Num>
-				<Latency>3</Latency>
+				<Latency>1</Latency>
 				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
 				<PortNumber>0</PortNumber>
 				<PortNumber>1</PortNumber>
@@ -305,7 +319,7 @@ TDP = 112W
 			
 			<FloatVectorMul>
 				<Num>0</Num>
-				<Latency>3</Latency>
+				<Latency>5</Latency>
 				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
 				<PortNumber>0</PortNumber>
 				<PortNumber>1</PortNumber>
@@ -315,24 +329,59 @@ TDP = 112W
 			
 			<FMA>
 				<Num>2</Num>
-				<FPALULatency>4</FPALULatency>
-				<FPMulLatency>4</FPMulLatency>
-				<FPVectorALULatency>4</FPVectorALULatency>
-				<FPVectorMulLatency>4</FPVectorMulLatency>
-				<FMALatency>4</FMALatency>
-				<VectorFMALatency>4</VectorFMALatency>
-				<FPALUReciprocalOfThroughput>3</FPALUReciprocalOfThroughput>
-				<FPMulReciprocalOfThroughput>3</FPMulReciprocalOfThroughput>
-				<FPVectorALUReciprocalOfThroughput>3</FPVectorALUReciprocalOfThroughput>
-				<FPVectorMulReciprocalOfThroughput>3</FPVectorMulReciprocalOfThroughput>
-				<FMAReciprocalOfThroughput>3</FMAReciprocalOfThroughput>
-				<VectorFMAReciprocalOfThroughput>3</VectorFMAReciprocalOfThroughput>
+				<Latency>4</Latency>
+				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
 				<PortNumber>0</PortNumber>
 				<PortNumber>1</PortNumber>
 				<LeakageEnergy>0.0271</LeakageEnergy>
 				<DynamicEnergy>0.6514</DynamicEnergy>			
 			</FMA>
 			
+			<AES>
+				<Num>1</Num>
+				<Latency>4</Latency>
+				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
+				<PortNumber>0</PortNumber>
+				<LeakageEnergy>0.0271</LeakageEnergy>
+				<DynamicEnergy>0.6514</DynamicEnergy>			
+			</AES>
+			
+			<VectorString>
+				<Num>1</Num>
+				<Latency>9</Latency>
+				<ReciprocalOfThroughput>5</ReciprocalOfThroughput>
+				<PortNumber>0</PortNumber>
+				<LeakageEnergy>0.0271</LeakageEnergy>
+				<DynamicEnergy>0.6514</DynamicEnergy>			
+			</VectorString>
+			
+			<BitScan>
+				<Num>1</Num>
+				<Latency>3</Latency>
+				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
+				<PortNumber>1</PortNumber>
+				<LeakageEnergy>0.0271</LeakageEnergy>
+				<DynamicEnergy>0.6514</DynamicEnergy>			
+			</BitScan>
+			
+			<VectorShuffle>
+				<Num>1</Num>
+				<Latency>1</Latency>
+				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
+				<PortNumber>5</PortNumber>
+				<LeakageEnergy>0.0271</LeakageEnergy>
+				<DynamicEnergy>0.6514</DynamicEnergy>			
+			</VectorShuffle>
+			
+			<LEA>
+				<Num>1</Num>
+				<Latency>3</Latency>
+				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
+				<PortNumber>5</PortNumber>
+				<LeakageEnergy>0.0271</LeakageEnergy>
+				<DynamicEnergy>0.6514</DynamicEnergy>			
+			</LEA>
+			
 			<Branch>
 				<Num>2</Num>
 				<Latency>1</Latency>
@@ -343,16 +392,6 @@ TDP = 112W
 				<DynamicEnergy>0.6514</DynamicEnergy>			
 			</Branch>
 			
-			<Load>
-				<Num>2</Num>
-				<Latency>1</Latency>
-				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
-				<PortNumber>2</PortNumber>
-				<PortNumber>3</PortNumber>
-				<LeakageEnergy>0.0271</LeakageEnergy>
-				<DynamicEnergy>0.6514</DynamicEnergy>			
-			</Load>
-			
 			<LoadAGU>
 				<Num>2</Num>
 				<Latency>1</Latency>
@@ -363,14 +402,15 @@ TDP = 112W
 				<DynamicEnergy>0.6514</DynamicEnergy>			
 			</LoadAGU>
 			
-			<Store>
-				<Num>1</Num>
+			<Load>
+				<Num>2</Num>
 				<Latency>1</Latency>
 				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
-				<PortNumber>4</PortNumber>
+				<PortNumber>2</PortNumber>
+				<PortNumber>3</PortNumber>
 				<LeakageEnergy>0.0271</LeakageEnergy>
 				<DynamicEnergy>0.6514</DynamicEnergy>			
-			</Store>
+			</Load>
 			
 			<StoreAGU>
 				<Num>1</Num>
@@ -381,50 +421,14 @@ TDP = 112W
 				<DynamicEnergy>0.6514</DynamicEnergy>			
 			</StoreAGU>
 			
-			<AES>
+			<Store>
 				<Num>1</Num>
-				<Latency>24</Latency>
-				<ReciprocalOfThroughput>12</ReciprocalOfThroughput>
-				<PortNumber>0</PortNumber>
+				<Latency>1</Latency>
+				<ReciprocalOfThroughput>1</ReciprocalOfThroughput>
+				<PortNumber>4</PortNumber>
 				<LeakageEnergy>0.0271</LeakageEnergy>
 				<DynamicEnergy>0.6514</DynamicEnergy>			
-			</AES>
-			
-			<VectorString>
-				<Num>1</Num>
-				<Latency>24</Latency>
-				<ReciprocalOfThroughput>12</ReciprocalOfThroughput>
-				<PortNumber>0</PortNumber>
-				<LeakageEnergy>0.0271</LeakageEnergy>
-				<DynamicEnergy>0.6514</DynamicEnergy>			
-			</VectorString>
-			
-			<BitScan>
-				<Num>1</Num>
-				<Latency>24</Latency>
-				<ReciprocalOfThroughput>12</ReciprocalOfThroughput>
-				<PortNumber>1</PortNumber>
-				<LeakageEnergy>0.0271</LeakageEnergy>
-				<DynamicEnergy>0.6514</DynamicEnergy>			
-			</BitScan>
-			
-			<VectorShuffle>
-				<Num>1</Num>
-				<Latency>24</Latency>
-				<ReciprocalOfThroughput>12</ReciprocalOfThroughput>
-				<PortNumber>5</PortNumber>
-				<LeakageEnergy>0.0271</LeakageEnergy>
-				<DynamicEnergy>0.6514</DynamicEnergy>			
-			</VectorShuffle>
-			
-			<LEA>
-				<Num>1</Num>
-				<Latency>24</Latency>
-				<ReciprocalOfThroughput>12</ReciprocalOfThroughput>
-				<PortNumber>5</PortNumber>
-				<LeakageEnergy>0.0271</LeakageEnergy>
-				<DynamicEnergy>0.6514</DynamicEnergy>			
-			</LEA>
+			</Store>
 			
 			<ResultsBroadcastBus>
 				<LeakageEnergy>0.0239</LeakageEnergy>
@@ -460,24 +464,24 @@ TDP = 112W
 		</Core>
 		
 		<SharedCaches>
-			<Cache name="L3" type="L3Cache_8M_16"/><!--Intel® Core™ i9-7960X X-series Processor actually has 11M L3. Reducing to 8 to keep it a power of 2 (required by Tejas)-->
+			<Cache name="L3" type="L3Cache_4M_8"/><!--Intel® Core™ i7-7820X X-series Processor actually has 3M L3 12-way. Changing to 4M 8-way to keep it a power of 2 (required by Tejas)-->
 			<Cache name="D1" type="Directory1"/>
 		</SharedCaches>
 		
-		<Interconnect>BUS</Interconnect>
+		<Interconnect>NOC</Interconnect>
 		
 		<NOC>
-			<NocConfigFile>/home/rajshekar/projects/nvms/configurations/config_1052_8core_skylake_NocConfig.txt</NocConfigFile>
+			<NocConfigFile>/home/rajshekar/resources/tejas_configs/config_2core_kabylake_NocConfig.txt</NocConfigFile>
 			<NocSelScheme>STATIC</NocSelScheme>
 			<NocNumberOfBuffers>4</NocNumberOfBuffers>
 			<NocPortType>FCFS</NocPortType>
 			<NocAccessPorts>4</NocAccessPorts>
 			<NocPortOccupancy>1</NocPortOccupancy>
 			<NocLatency>1</NocLatency>
-			<NocOperatingFreq>2000</NocOperatingFreq>
+			<NocOperatingFreq>2400</NocOperatingFreq>
 			<NocTopology>TORUS</NocTopology>     <!--NOCTopology-->
 			<NocRoutingAlgorithm>SIMPLE</NocRoutingAlgorithm>
-			<NocLatencyBetweenNOCElements>4</NocLatencyBetweenNOCElements>
+			<NocLatencyBetweenNOCElements>2</NocLatencyBetweenNOCElements>
 			<NocRouterArbiter>RR_ARBITER</NocRouterArbiter>
 			<TechPoint>90</TechPoint>
 			<NocConnection>ELECTRICAL</NocConnection>
@@ -539,7 +543,8 @@ TDP = 112W
 <!--Give all the library elements here-->
 <Library>
 	<UnifiedCache_32K_8>
-		<Frequency>4200</Frequency>			<!-- private caches take frequency of containing core -->
+		<AMAT>-1</AMAT>
+		<Frequency>4000</Frequency>			<!-- private caches take frequency of containing core -->
 		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
 		<BlockSize>64</BlockSize>			<!--In bytes-->
 		<Associativity>8</Associativity>
@@ -547,306 +552,372 @@ TDP = 112W
 		<ReadLatency>3</ReadLatency>				<!--In clock cycles-->
 		<WriteLatency>3</WriteLatency>				<!--In clock cycles-->
 		<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>1</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-			<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>16</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
-			<LeakageEnergy>0.1092</LeakageEnergy>
-			<ReadDynamicEnergy>0.33964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.33964264705</WriteDynamicEnergy>
-			<CacheType>Unified</CacheType>  <!--Instruction,Data,Unified-->
-		</UnifiedCache_32K_8>
+		<ReadPorts>0</ReadPorts>
+		<WritePorts>0</WritePorts>
+		<ReadWritePorts>1</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>None</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>16</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<LeakageEnergy>0.1092</LeakageEnergy>
+		<ReadDynamicEnergy>0.33964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.33964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--Instruction,Data,Unified-->
+	</UnifiedCache_32K_8>
 		
-		<ICache_32K_8>
-			<Frequency>4200</Frequency>			<!-- private caches take frequency of containing core -->
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>8</Associativity>
-			<Size>32768</Size>					<!--In Bytes-->
-			<ReadLatency>4</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>4</WriteLatency>				<!--In clock cycles-->
-			<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>4</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-			<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>16</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
-			<LeakageEnergy>0.1092</LeakageEnergy>
-			<ReadDynamicEnergy>0.33964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.33964264705</WriteDynamicEnergy>
-			<CacheType>Instruction</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</ICache_32K_8>
+	<ICache_32K_8>
+		<AMAT>-1</AMAT>
+		<Frequency>2500</Frequency>			<!-- private caches take frequency of containing core -->
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>8</Associativity>
+		<Size>32768</Size>					<!--In Bytes-->
+		<ReadLatency>4</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>4</WriteLatency>				<!--In clock cycles-->
+		<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>0</ReadPorts>
+		<WritePorts>0</WritePorts>
+		<ReadWritePorts>4</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>Power4</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>16</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<LeakageEnergy>0.1092</LeakageEnergy>
+		<ReadDynamicEnergy>0.33964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.33964264705</WriteDynamicEnergy>
+		<CacheType>Instruction</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</ICache_32K_8>
 		
-		<L1Cache_32K_8>
-			<Frequency>4200</Frequency>			<!-- private caches take frequency of containing core -->
-			<WriteMode>WT</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>8</Associativity>
-			<Size>32768</Size>					<!--In Bytes-->
-			<ReadLatency>4</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>4</WriteLatency>				<!--In clock cycles-->
-			<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>2</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-		<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>16</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->			
-			<LeakageEnergy>0.1092</LeakageEnergy>
-			<ReadDynamicEnergy>0.33964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.33964264705</WriteDynamicEnergy>
-			<CacheType>Data</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</L1Cache_32K_8>
+	<L1Cache_32K_8>
+		<AMAT>-1</AMAT>
+		<Frequency>2500</Frequency>			<!-- private caches take frequency of containing core -->
+		<WriteMode>WT</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>8</Associativity>
+		<Size>32768</Size>					<!--In Bytes-->
+		<ReadLatency>4</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>4</WriteLatency>				<!--In clock cycles-->
+		<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>1</ReadPorts>
+		<WritePorts>1</WritePorts>
+		<ReadWritePorts>0</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>2</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>Power4</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>16</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->			
+		<LeakageEnergy>0.1092</LeakageEnergy>
+		<ReadDynamicEnergy>0.33964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.33964264705</WriteDynamicEnergy>
+		<CacheType>Data</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L1Cache_32K_8>
 		
-		<L2Cache_256K_8>
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>8</Associativity>
-			<Size>262144</Size>					<!--In Bytes-->
-			<ReadLatency>12</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>12</WriteLatency>				<!--In clock cycles-->
-			<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>1</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-			<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>256</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->			
-			<LeakageEnergy>0.1592</LeakageEnergy>
-			<ReadDynamicEnergy>0.43964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.43964264705</WriteDynamicEnergy>
-			<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</L2Cache_256K_8>
+	<L2Cache_256K_8>
+		<AMAT>-1</AMAT>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>8</Associativity>
+		<Size>262144</Size>					<!--In Bytes-->
+		<ReadLatency>12</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>12</WriteLatency>				<!--In clock cycles-->
+		<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>0</ReadPorts>
+		<WritePorts>0</WritePorts>
+		<ReadWritePorts>1</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>None</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>256</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->			
+		<LeakageEnergy>0.1592</LeakageEnergy>
+		<ReadDynamicEnergy>0.43964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.43964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L2Cache_256K_8>
 		
-		<L2Cache_256K_4>
-			<Frequency>4200</Frequency>			<!-- private caches take frequency of containing core -->
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>4</Associativity>
-			<Size>262144</Size>					<!--In Bytes-->
-			<ReadLatency>12</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>12</WriteLatency>				<!--In clock cycles-->
-			<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>1</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-			<Coherence>D1</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>Power4</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>256</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->			
-			<LeakageEnergy>0.1592</LeakageEnergy>
-			<ReadDynamicEnergy>0.43964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.43964264705</WriteDynamicEnergy>
-			<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</L2Cache_256K_4>
+	<L2Cache_256K_4>
+		<AMAT>-1</AMAT>
+		<Frequency>2500</Frequency>			<!-- private caches take frequency of containing core -->
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>4</Associativity>
+		<Size>262144</Size>					<!--In Bytes-->
+		<ReadLatency>12</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>12</WriteLatency>				<!--In clock cycles-->
+		<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>1</ReadPorts>
+		<WritePorts>1</WritePorts>
+		<ReadWritePorts>0</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>Power4</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>256</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->			
+		<LeakageEnergy>0.1592</LeakageEnergy>
+		<ReadDynamicEnergy>0.43964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.43964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L2Cache_256K_4>
 		
-		<L3Cache_1M_8>
-			<Frequency>2000</Frequency>
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>8</Associativity>
-			<Size>1048576</Size>					<!--In Bytes-->
-			<ReadLatency>60</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>60</WriteLatency>				<!--In clock cycles-->
-			<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>1</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-			<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>8</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
-			<LeakageEnergy>0.1892</LeakageEnergy>
-			<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
-			<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</L3Cache_1M_8>
+	<L2Cache_1M_16>
+		<AMAT>-1</AMAT>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>16</Associativity>
+		<Size>1048576</Size>					<!--In Bytes-->
+		<Latency>12</Latency>				<!--In clock cycles-->
+		<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>1</ReadPorts>
+		<WritePorts>1</WritePorts>
+		<ReadWritePorts>0</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>D1</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>Power4</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>256</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->			
+		<LeakageEnergy>0.1592</LeakageEnergy>
+		<ReadDynamicEnergy>0.43964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.43964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L2Cache_1M_16>
 		
-		<L3Cache_8M_16>
-			<Frequency>2000</Frequency>
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>16</Associativity>
-			<Size>8388608</Size>					<!--In Bytes-->
-			<ReadLatency>60</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>60</WriteLatency>				<!--In clock cycles-->
-			<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>1</ReadWritePorts>
-			<PortReadOccupancy>5</PortReadOccupancy>
-			<PortWriteOccupancy>5</PortWriteOccupancy>
-			<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>Power4</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>8</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
-			<LeakageEnergy>0.1892</LeakageEnergy>
-			<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
-			<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</L3Cache_8M_16>
+	<L3Cache_1M_8>
+		<AMAT>-1</AMAT>
+		<Frequency>2000</Frequency>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>8</Associativity>
+		<Size>1048576</Size>					<!--In Bytes-->
+		<ReadLatency>60</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>60</WriteLatency>				<!--In clock cycles-->
+		<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>0</ReadPorts>
+		<WritePorts>0</WritePorts>
+		<ReadWritePorts>1</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>None</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>8</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<LeakageEnergy>0.1892</LeakageEnergy>
+		<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L3Cache_1M_8>
 		
-		<L3Cache_12M_16>
-			<Frequency>2000</Frequency>
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>16</Associativity>
-			<Size>12582912</Size>					<!--In Bytes-->
-			<ReadLatency>60</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>60</WriteLatency>				<!--In clock cycles-->
-			<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>1</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-			<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>8</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
-			<LeakageEnergy>0.1892</LeakageEnergy>
-			<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
-			<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</L3Cache_12M_16>
+	<L3Cache_8M_16><!--Intel® Core™ i7-7820X X-series Processor actually has 11M L3. Reducing to 8 to keep it a power of 2 (required by Tejas)-->
+		<AMAT>-1</AMAT>
+		<Frequency>4000</Frequency>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>16</Associativity>
+		<Size>8388608</Size>					<!--In Bytes-->
+		<ReadLatency>77</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>77</WriteLatency>				<!--In clock cycles-->
+		<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>1</ReadPorts>
+		<WritePorts>1</WritePorts>
+		<ReadWritePorts>1</ReadWritePorts>
+		<PortReadOccupancy>2</PortReadOccupancy>
+		<PortWriteOccupancy>2</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>Power4</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>8</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<LeakageEnergy>0.1892</LeakageEnergy>
+		<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L3Cache_8M_16>
 		
-		<L3Cache_22M_16>
-			<Frequency>2000</Frequency>
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>16</Associativity>
-			<Size>23068672</Size>					<!--In Bytes-->
-			<ReadLatency>44</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>44</WriteLatency>				<!--In clock cycles-->
-			<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>1</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-			<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>8</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
-			<LeakageEnergy>0.1892</LeakageEnergy>
-			<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
-			<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</L3Cache_22M_16>
+	<L3Cache_4M_8><!--Intel® Core™ i5-7200U Processor actually has 3M 12-way L3. Changing to 4M 8-way to keep it a power of 2 (required by Tejas)-->
+		<AMAT>-1</AMAT>
+		<Frequency>2500</Frequency>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>8</Associativity>
+		<Size>4194304</Size>					<!--In Bytes-->
+		<ReadLatency>44</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>44</WriteLatency>				<!--In clock cycles-->
+		<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>1</ReadPorts>
+		<WritePorts>1</WritePorts>
+		<ReadWritePorts>1</ReadWritePorts>
+		<PortReadOccupancy>2</PortReadOccupancy>
+		<PortWriteOccupancy>2</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>Power4</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>8</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<LeakageEnergy>0.1892</LeakageEnergy>
+		<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L3Cache_4M_8>
 		
-		<L3Cache_16M_16>
-			<Frequency>2000</Frequency>
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<BlockSize>64</BlockSize>			<!--In bytes-->
-			<Associativity>16</Associativity>
-			<Size>16777216</Size>					<!--In Bytes-->
-			<ReadLatency>44</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>44</WriteLatency>				<!--In clock cycles-->
-			<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>1</ReadWritePorts>
-			<PortReadOccupancy>5</PortReadOccupancy>
-			<PortWriteOccupancy>5</PortWriteOccupancy>
-			<Coherence>None</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>8</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
-			<LeakageEnergy>0.1892</LeakageEnergy>
-			<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
-			<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
-			<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</L3Cache_16M_16>
+	<L3Cache_12M_16>
+		<AMAT>-1</AMAT>
+		<Frequency>2000</Frequency>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>16</Associativity>
+		<Size>12582912</Size>					<!--In Bytes-->
+		<ReadLatency>60</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>60</WriteLatency>				<!--In clock cycles-->
+		<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>0</ReadPorts>
+		<WritePorts>0</WritePorts>
+		<ReadWritePorts>1</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>None</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>8</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<LeakageEnergy>0.1892</LeakageEnergy>
+		<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L3Cache_12M_16>
 		
-		<Directory1>
-			<Frequency>2000</Frequency>
-			<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
-			<LastLevel>N</LastLevel>			<!--Whether this is the last level in the hierarchy or not (Y for yes, N for no)-->
-			<BlockSize>64</BlockSize>			<!--In bytes (this should be same as the block size of the Caches between those you want coherence)-->
-			<Associativity>64</Associativity>
-			<NumEntries>65536</NumEntries>
-			<ReadLatency>5</ReadLatency>				<!--In clock cycles-->
-			<WriteLatency>5</WriteLatency>				<!--In clock cycles-->
-			<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
-			<ReadPorts>0</ReadPorts>
-			<WritePorts>0</WritePorts>
-			<ReadWritePorts>2</ReadWritePorts>
-			<PortReadOccupancy>1</PortReadOccupancy>
-			<PortWriteOccupancy>1</PortWriteOccupancy>
-			<Coherence>N</Coherence>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<Prefetcher>None</Prefetcher>
-			<NumBuses>1</NumBuses>			<!--Coherence of upper level (N : None, S : Snoopy, D : Directory)-->
-			<MSHRSize>16</MSHRSize>
-			<BusOccupancy>0</BusOccupancy>
-			<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
-			<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
-			<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
-			<CacheType>Unified</CacheType>
-			<LeakageEnergy>.1092</LeakageEnergy>
-			<ReadDynamicEnergy>.3396</ReadDynamicEnergy>
-			<WriteDynamicEnergy>.3396</WriteDynamicEnergy>
-			<IsDirectory>true</IsDirectory>
-			<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
-		</Directory1>
+	<L3Cache_22M_16>
+		<AMAT>-1</AMAT>
+		<Frequency>2000</Frequency>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>16</Associativity>
+		<Size>23068672</Size>					<!--In Bytes-->
+		<ReadLatency>44</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>44</WriteLatency>				<!--In clock cycles-->
+		<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>0</ReadPorts>
+		<WritePorts>0</WritePorts>
+		<ReadWritePorts>1</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>None</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>8</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<LeakageEnergy>0.1892</LeakageEnergy>
+		<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L3Cache_22M_16>
+		
+	<L3Cache_16M_16>
+		<AMAT>-1</AMAT>
+		<Frequency>2000</Frequency>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<BlockSize>64</BlockSize>			<!--In bytes-->
+		<Associativity>16</Associativity>
+		<Size>16777216</Size>					<!--In Bytes-->
+		<ReadLatency>44</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>44</WriteLatency>				<!--In clock cycles-->
+		<PortType>UL</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>0</ReadPorts>
+		<WritePorts>0</WritePorts>
+		<ReadWritePorts>1</ReadWritePorts>
+		<PortReadOccupancy>5</PortReadOccupancy>
+		<PortWriteOccupancy>5</PortWriteOccupancy>
+		<Coherence>None</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>None</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>8</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<LeakageEnergy>0.1892</LeakageEnergy>
+		<ReadDynamicEnergy>0.60964264705</ReadDynamicEnergy>
+		<WriteDynamicEnergy>0.60964264705</WriteDynamicEnergy>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</L3Cache_16M_16>
+		
+	<Directory1>
+		<AMAT>-1</AMAT>
+		<Frequency>2400</Frequency>
+		<WriteMode>WB</WriteMode>			<!--Write-back (WB) or write-through (WT)-->
+		<LastLevel>N</LastLevel>			<!--Whether this is the last level in the hierarchy or not (Y for yes, N for no)-->
+		<BlockSize>64</BlockSize>			<!--In bytes (this should be same as the block size of the Caches between those you want coherence)-->
+		<Associativity>64</Associativity>
+		<NumEntries>65536</NumEntries>
+		<ReadLatency>5</ReadLatency>				<!--In clock cycles-->
+		<WriteLatency>5</WriteLatency>				<!--In clock cycles-->
+		<PortType>FCFS</PortType>				<!--Type of access ports in the Cache (UL : Unlimited; FCFS : First Come First Serve; PR : Priority port)-->
+		<ReadPorts>0</ReadPorts>
+		<WritePorts>0</WritePorts>
+		<ReadWritePorts>2</ReadWritePorts>
+		<PortReadOccupancy>1</PortReadOccupancy>
+		<PortWriteOccupancy>1</PortWriteOccupancy>
+		<Coherence>N</Coherence>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<Prefetcher>None</Prefetcher>
+		<NumBuses>1</NumBuses>			<!--Coherence of level (N : None, S : Snoopy, D : Directory)-->
+		<MSHRSize>16</MSHRSize>
+		<BusOccupancy>0</BusOccupancy>
+		<Nuca>NONE</Nuca>						<!--NUCA type (S_NUCA, D_NUCA, NONE)-->
+		<ONuca>NONE</ONuca>						<!--ONUCA type (BCAST, TSI, NONE)-->
+		<NucaMapping>S</NucaMapping>            <!--Valid for NUCA; S: Set-Associative A: Address-Mapped -->
+		<CacheType>Unified</CacheType>
+		<LeakageEnergy>.1092</LeakageEnergy>
+		<ReadDynamicEnergy>.3396</ReadDynamicEnergy>
+		<WriteDynamicEnergy>.3396</WriteDynamicEnergy>
+		<IsDirectory>true</IsDirectory>
+		<CacheType>Unified</CacheType>  <!--I : Instruction, D : Data, U : Unified-->
+	</Directory1>
 		
 	</Library>
 	
diff --git a/src/simulator/emulatorinterface/DynamicInstructionBuffer.java b/src/simulator/emulatorinterface/DynamicInstructionBuffer.java
index f81788c..cfccfdd 100755
--- a/src/simulator/emulatorinterface/DynamicInstructionBuffer.java
+++ b/src/simulator/emulatorinterface/DynamicInstructionBuffer.java
@@ -44,6 +44,8 @@ public class DynamicInstructionBuffer implements Encoding
 	
 	private long ip;
 	
+	private boolean predicateInsnNotExecuted;
+	
 	public DynamicInstructionBuffer() 
 	{
 		memRead = new long[64];
@@ -59,6 +61,8 @@ public class DynamicInstructionBuffer implements Encoding
 		
 		branchAddress = -1;
 		
+		predicateInsnNotExecuted = false;
+		
 		ip = arrayListPacket.get(0).ip;
 		
 		int numAssemblyPackets = 0;
@@ -100,6 +104,10 @@ public class DynamicInstructionBuffer implements Encoding
 					branchTaken = false;
 					branchAddress = p.tgt;
 					break;
+					
+				case (PREDICATE_INSN_NOT_EXECUTED):
+					predicateInsnNotExecuted = true;
+					break;
 				
 				case (ASSEMBLY):
 					numAssemblyPackets++;
@@ -190,4 +198,8 @@ public class DynamicInstructionBuffer implements Encoding
 		boolean readAuthenticBranch = branchInformationRead==false && branchAddress!=-1;
 		return readAuthenticBranch;
 	}
+
+	public boolean isPredicateInsnNotExecuted() {
+		return predicateInsnNotExecuted;
+	}
 }
\ No newline at end of file
diff --git a/src/simulator/emulatorinterface/RunnableThread.java b/src/simulator/emulatorinterface/RunnableThread.java
index 5a718cd..abc07a2 100755
--- a/src/simulator/emulatorinterface/RunnableThread.java
+++ b/src/simulator/emulatorinterface/RunnableThread.java
@@ -204,6 +204,7 @@ public class RunnableThread implements Encoding, Runnable {
 									iNew.setDestinationOperandMemValue(Long.parseLong(splited[i+2]));
 									iNew.setBranchTargetAddress(Long.parseLong(splited[i+3]));
 									iNew.setBranchTaken(Boolean.parseBoolean(splited[i+4]));
+									iNew.setPredicateAndNotExecuted(Boolean.parseBoolean(splited[i+7]));
 									//iNew.setSerialNo(Long.parseLong(splited[i+6]));
 								}
 							}
@@ -833,7 +834,7 @@ public class RunnableThread implements Encoding, Runnable {
 			
 		}
 		
-		if (pnew.value!=INSTRUCTION && !(pnew.value>6 && pnew.value<26) && pnew.value!=Encoding.ASSEMBLY ) {
+		if (pnew.value!=INSTRUCTION && !(pnew.value>6 && pnew.value<26) && pnew.value!=Encoding.ASSEMBLY && pnew.value!=Encoding.PREDICATE_INSN_NOT_EXECUTED ) {
 			// just append the packet to outstanding packetList for current instruction pointer
 			thread.packetList.add(pnew);
 			
diff --git a/src/simulator/emulatorinterface/communication/Encoding.java b/src/simulator/emulatorinterface/communication/Encoding.java
index 812a80b..7cda502 100755
--- a/src/simulator/emulatorinterface/communication/Encoding.java
+++ b/src/simulator/emulatorinterface/communication/Encoding.java
@@ -41,5 +41,7 @@ public interface Encoding {
 	static final int CPL_SWITCH = 34;
 	static final int PARENT_SPAWN = 35;
 	static final int CHILD_START = 36;
+	
+	static final int PREDICATE_INSN_NOT_EXECUTED = 37;
 
 }
diff --git a/src/simulator/emulatorinterface/translator/visaHandler/IntegerALU.java b/src/simulator/emulatorinterface/translator/visaHandler/IntegerALU.java
index 1b0bde8..c489cbb 100755
--- a/src/simulator/emulatorinterface/translator/visaHandler/IntegerALU.java
+++ b/src/simulator/emulatorinterface/translator/visaHandler/IntegerALU.java
@@ -9,7 +9,7 @@ public class IntegerALU implements DynamicInstructionHandler
 	public int handle(int microOpIndex, 
 			Instruction microOp, DynamicInstructionBuffer dynamicInstructionBuffer) 
 	{
-		//nothing to be done in such cases
+		microOp.setPredicateAndNotExecuted(dynamicInstructionBuffer.isPredicateInsnNotExecuted());
 		return ++microOpIndex;
 	}
 }
diff --git a/src/simulator/emulatorinterface/translator/x86/instruction/ConditionalMove.java b/src/simulator/emulatorinterface/translator/x86/instruction/ConditionalMove.java
index 206d184..e6b9a88 100755
--- a/src/simulator/emulatorinterface/translator/x86/instruction/ConditionalMove.java
+++ b/src/simulator/emulatorinterface/translator/x86/instruction/ConditionalMove.java
@@ -44,7 +44,9 @@ public class ConditionalMove implements X86StaticInstructionHandler
 			(operand2.isIntegerRegisterOperand() || operand2.isImmediateOperand()) &&
 		    (operand3==null))
 		{
-			instructionArrayList.appendInstruction(Instruction.getIntALUInstruction(operand2, Registers.getEFlagsRegister(), operand1));
+			Operand temp = Registers.getTempIntReg(tempRegisterNum);
+			instructionArrayList.appendInstruction(Instruction.getIntALUInstruction(Registers.getEFlagsRegister(), operand1, temp));
+			instructionArrayList.appendInstruction(Instruction.getIntALUInstruction(operand2, temp, operand1));
 		}
 		
 		//if operand1 = register and operand2 = memory - load
@@ -52,8 +54,10 @@ public class ConditionalMove implements X86StaticInstructionHandler
 				 operand2.isMemoryOperand() && 
 				 operand3==null)
 		{
+			Operand temp = Registers.getTempIntReg(tempRegisterNum);
+			instructionArrayList.appendInstruction(Instruction.getIntALUInstruction(Registers.getEFlagsRegister(), operand1, temp));
 			Operand sourceOperand = OperandTranslator.processSourceMemoryOperand(operand2, instructionArrayList, tempRegisterNum, true);
-			instructionArrayList.appendInstruction(Instruction.getIntALUInstruction(sourceOperand, Registers.getEFlagsRegister(), operand1));
+			instructionArrayList.appendInstruction(Instruction.getIntALUInstruction(sourceOperand, temp, operand1));
 		}
 		
 //		//if operand1 = memory and operand2 = memory - store
diff --git a/src/simulator/emulatorinterface/translator/x86/instruction/ConditionalSet.java b/src/simulator/emulatorinterface/translator/x86/instruction/ConditionalSet.java
index e82bd02..f002892 100755
--- a/src/simulator/emulatorinterface/translator/x86/instruction/ConditionalSet.java
+++ b/src/simulator/emulatorinterface/translator/x86/instruction/ConditionalSet.java
@@ -19,7 +19,7 @@ public class ConditionalSet implements X86StaticInstructionHandler
 				operand2==null && operand3==null)
 		{
 			instructionArrayList.appendInstruction(Instruction.getIntALUInstruction(Registers.getEFlagsRegister(), 
-					Operand.getImmediateOperand(), operand1));
+					operand1, operand1));
 		}
 		
 		else if(operand1.isMemoryOperand())
diff --git a/src/simulator/generic/Instruction.java b/src/simulator/generic/Instruction.java
index b2a7e6b..f5c7a80 100755
--- a/src/simulator/generic/Instruction.java
+++ b/src/simulator/generic/Instruction.java
@@ -44,11 +44,14 @@ public class Instruction implements Serializable
 	private long serialNo;
 	private int threadID;
 	
+	private boolean isPredicateAndNotExecuted;
+	
 	public Instruction()
 	{
 		this.sourceOperand1 = null;
 		this.sourceOperand2 = null;
 		this.destinationOperand = null;
+		isPredicateAndNotExecuted = false;
 	}
 	
 	public void clear()
@@ -57,6 +60,7 @@ public class Instruction implements Serializable
 		this.sourceOperand1 = null;
 		this.sourceOperand2 = null;
 		this.destinationOperand = null;
+		isPredicateAndNotExecuted = false;
 	}
 	
 	public Instruction(OperationType type, Operand sourceOperand1,
@@ -66,6 +70,7 @@ public class Instruction implements Serializable
 		this.sourceOperand1 = sourceOperand1;
 		this.sourceOperand2 = sourceOperand2;
 		this.destinationOperand = destinationOperand;
+		isPredicateAndNotExecuted = false;
 	}
 	
 	private void set(OperationType type, Operand sourceOperand1,
@@ -75,6 +80,7 @@ public class Instruction implements Serializable
 		this.sourceOperand1 = sourceOperand1;
 		this.sourceOperand2 = sourceOperand2;
 		this.destinationOperand = destinationOperand;
+		isPredicateAndNotExecuted = false;
 	}
 	
 //	/* our clone constructor */
@@ -123,6 +129,8 @@ public class Instruction implements Serializable
 		
 		this.serialNo = sourceInstruction.serialNo;
 		this.threadID = sourceInstruction.threadID;
+		
+		this.isPredicateAndNotExecuted = sourceInstruction.isPredicateAndNotExecuted;
 	}
 	
 	public static Instruction getIntALUInstruction(Operand sourceOperand1, Operand sourceOperand2, Operand destinationOperand)
@@ -492,6 +500,14 @@ public class Instruction implements Serializable
 		this.destinationOperandMemValue = destinationOperandMemValue;
 	}
 	
+	public boolean isPredicateAndNotExecuted() {
+		return isPredicateAndNotExecuted;
+	}
+
+	public void setPredicateAndNotExecuted(boolean isPredicateAndNotExecuted) {
+		this.isPredicateAndNotExecuted = isPredicateAndNotExecuted;
+	}
+
 	/**
 	 * strInstruction method returns the instruction information in a string.
 	 * @return String describing the instruction
diff --git a/src/simulator/pipeline/outoforder/ReorderBuffer.java b/src/simulator/pipeline/outoforder/ReorderBuffer.java
index 0158791..686f39f 100755
--- a/src/simulator/pipeline/outoforder/ReorderBuffer.java
+++ b/src/simulator/pipeline/outoforder/ReorderBuffer.java
@@ -448,7 +448,7 @@ public class ReorderBuffer extends SimulationElement{
             else bw.write(" null null null");
             bw.write(" "+tmp.getSourceOperand1MemValue()+" "+
             tmp.getSourceOperand2MemValue()+" "+tmp.getDestinationOperandMemValue()+" "+
-            tmp.getBranchTargetAddress()+" "+tmp.isBranchTaken()+" "+tmp.getThreadID()+" "+tmp.getSerialNo()+"\n");
+            tmp.getBranchTargetAddress()+" "+tmp.isBranchTaken()+" "+tmp.getThreadID()+" "+tmp.getSerialNo()+" "+tmp.isPredicateAndNotExecuted()+"\n");
         }
 		catch(Exception e)
 		{