Bpu add percep perf

.gitignore

@@ -69,4 +69,7 @@ AGENTS.md
 
 microbench/build/
 microbench/output/
-microbench/dramsim3*
+microbench/dramsim3*
+NEMU
+
+workrecord.txt

configs/example/kmhv3.py

@@ -105,9 +105,15 @@ def setKmhV3Params(args, system):
             cpu.branchPred.mbtb.enabled = True
             cpu.branchPred.tage.enabled = True
             cpu.branchPred.ittage.enabled = True
-            cpu.branchPred.mgsc.enabled = False
+            cpu.branchPred.mgsc.enabled = True
             cpu.branchPred.ras.enabled = True
 
+            # 64KB prediction
+            cpu.branchPred.mgsc.enablePerceptionPred = True
+            cpu.branchPred.mgsc.percepTableEntryNum = 128
+            cpu.branchPred.mgsc.percepTableWidth = 8
+            cpu.branchPred.mgsc.gbhrLen = 512
+            cpu.branchPred.mgsc.percepThres = 500
         # l1 cache per core
         if args.caches:
             cpu.icache.size = '64kB'

src/cpu/pred/BranchPredictor.py

@@ -1041,6 +1041,7 @@ class BTBTAGE(TimedBaseBTBPredictor):
     cxx_header = "cpu/pred/btb/btb_tage.hh"
 
     needMoreHistories = Param.Bool(True, "BTBTAGE needs more histories")
+    needGBHR = Param.Bool(False,"BTBTAGE don't need GBHR")
     enableSC = Param.Bool(False, "Enable SC or not")    # TODO: BTBTAGE doesn't support SC
     updateOnRead = Param.Bool(True, "Enable update on read, no need to save tage meta in FTQ")
     numPredictors = Param.Unsigned(8, "Number of TAGE predictors")
@@ -1077,6 +1078,7 @@ class BTBITTAGE(TimedBaseBTBPredictor):
     cxx_header = "cpu/pred/btb/btb_ittage.hh"
 
     needMoreHistories = Param.Bool(True, "BTBITTAGE needs more histories")
+    needGBHR = Param.Bool(False,"BTBITTAGE don't need GBHR")
     numPredictors = Param.Unsigned(5, "Number of TAGE predictors")
     tableSizes = VectorParam.Unsigned([256]*2 + [512]*3, "the ITTAGE T0~Tn length")
     TTagBitSizes = VectorParam.Unsigned([9]*5, "the T0~Tn entry's tag bit size")
@@ -1094,6 +1096,7 @@ class BTBMGSC(TimedBaseBTBPredictor):
     cxx_header = "cpu/pred/btb/btb_mgsc.hh"
 
     needMoreHistories = Param.Bool(True, "MGSC needs more histories")
+    needGBHR = Param.Bool(True,"MGSC needs GBHR")
 
     bwTableNum = Param.Unsigned(2, "Num global backward branch GEHL tables")
     bwHistLen = VectorParam.Int([4, 8], "Global backward branch GEHL history lengths")
@@ -1123,6 +1126,11 @@ class BTBMGSC(TimedBaseBTBPredictor):
     biasTableNum = Param.Unsigned(1, "Num bias tables")
     biasTableIdxWidth = Param.Unsigned(11, "Log number of bias entries")
 
+    percepTableEntryNum = Param.Unsigned(4096, "Num of entries in perception preditor weight table")
+    gbhrLen = Param.Unsigned(23, "Length of gbhr in perception predictor")
+    percepTableWidth = Param.Unsigned(8, "Width of weight in perception predictor table")
+    percepThres = Param.Unsigned(1472, "Threshold of perception predictor") # gbhrLen * (1<<(percepTableWidth - 2))
+
     thresholdTablelogSize = Param.Unsigned(6,
         "Log size of update threshold counters tables")
 
@@ -1152,6 +1160,7 @@ class BTBMGSC(TimedBaseBTBPredictor):
     enablePTable = Param.Bool(True, "Enable P (path) table")
     enableBiasTable = Param.Bool(True, "Enable Bias table")
     enablePCThreshold = Param.Bool(False, "Enable PC-indexed threshold table")
+    enablePerceptionPred = Param.Bool(True, "Enable perception predictor")
 
     numDelay = 2
 

src/cpu/pred/btb/btb_ittage.cc

@@ -9,8 +9,8 @@
 #include "base/trace.hh"
 #include "cpu/o3/dyn_inst.hh"
 #include "debug/DecoupleBP.hh"
-#include "debug/DecoupleBPVerbose.hh"
 #include "debug/DecoupleBPUseful.hh"
+#include "debug/DecoupleBPVerbose.hh"
 #include "debug/ITTAGE.hh"
 #include "debug/ITTAGEHistory.hh"
 
@@ -32,6 +32,7 @@ numTablesToAlloc(p.numTablesToAlloc),
 ittageStats(this, p.numPredictors)
 {
     this->needMoreHistories = p.needMoreHistories;
+    this->needGBHR = p.needGBHR;
     DPRINTF(ITTAGE, "BTBITTAGE constructor numBr=%d\n", numBr);
     tageTable.resize(numPredictors);
     tableIndexBits.resize(numPredictors);
@@ -242,7 +243,7 @@ BTBITTAGE::update(const FetchTarget &stream)
     auto updateTagFoldedHist = meta->tagFoldedHist;
     auto updateAltTagFoldedHist = meta->altTagFoldedHist;
     auto updateIndexFoldedHist = meta->indexFoldedHist;
-    
+
     // update each branch
     for (auto &btb_entry : all_entries_to_update) {
         bool this_indirect_actual_taken = stream.exeTaken && stream.exeBranchInfo == btb_entry;