Small fixes to align with the original TAGE

- Fix to improve usefulness update function in LLBP
- Fixes to align the TAGE model with the original
  TAGE-SC-L.

Author: David Schall

analysis/mpki.ipynb

@@ -335,7 +335,7 @@
     "ax.set_xticklabels(_bms, rotation=10, horizontalalignment=\"center\", fontsize=fs-1)\n",
     "ax.tick_params(axis='x', which='major', pad=0)\n",
     "\n",
-    "ax.set_ylim(bottom=-0.0, top=0.48)\n",
+    "ax.set_ylim(bottom=-0.0, top=0.52)\n",
     "\n",
     "\n",
     "\n",

bpmodels/llbp/llbp.cc

@@ -278,13 +278,13 @@ bool LLBP::isUseful(bool taken) {
 }
 
 bool LLBP::llbpCorrect(bool taken) {
-    return llbp.isProvider && (taken == llbp.pred);
+    return llbp.hit && (taken == llbp.pred);
 }
 
 bool LLBP::primCorrect(bool taken) {
     return (scl_provider == STC) ? (sc_pred == taken) :
-                    HitBank ? (LongestMatchPred == taken) :
-                    AltBank ? (alttaken == taken) : (bim_pred == taken);
+                         HitBank ? (LongestMatchPred == taken) :
+                         AltBank ? (alttaken == taken) : (bim_pred == taken);
 }
 
 bool LLBP::tageCorrect(bool taken) {
@@ -293,7 +293,7 @@ bool LLBP::tageCorrect(bool taken) {
 }
 
 bool LLBP::llbpUseful(bool taken) {
-    return llbpCorrect(taken) && !primCorrect(taken);
+    return llbp.isProvider && (taken == llbp.pred) && !primCorrect(taken);
 }
 
 void LLBP::updateL2Usefulness(bool taken) {
@@ -311,13 +311,6 @@ void LLBP::updateL2Usefulness(bool taken) {
                 HitEntry->u++;
         }
     }
-
-
-    // Same for level 2. If it was the provider and was correct
-    // but level 1 not, it was useful.
-    if (llbp.hit && !llbp.shorter && llbp_correct && !prim_correct) {
-        ctrupdate(llbpEntry->replace, true, ReplCtrWidth);
-    }
 }
 
 
@@ -491,7 +484,6 @@ void LLBP::llbpUpdate(uint64_t pc, bool resolveDir, bool predDir) {
         // - If a pattern becomes low confident (incorrect prediction)
         //   the replacement counter is decreased
         if (llbpEntry->ctr == (resolveDir ? 1 : -2)) {
-        // else if (llbpEntry->ctr == (resolveDir ? 2 : -3)) {
             // entry became medium confident
             ctrupdate(HitContext->replace, true, CtxReplCtrWidth);
         }

bpmodels/tage/tage.cc

@@ -38,7 +38,7 @@ using namespace std;
 
 #define COND (stats.total >= 00000)
 
-#define USEPATH 1
+#define USEPATH
 
 namespace LLBP {
 
@@ -58,7 +58,7 @@ TageBase::TageBase(TageConfig cfg)
         Tbits(cfg.Tbits),
         uwidth(cfg.uwidth),
         cwidth(cfg.cwidth),
-        // size_use_alt(1 << (cfg.log_size_use_alt)),
+        size_use_alt(1 << (cfg.log_size_use_alt)),
         ghr(histbufferlen),
         disableInterleaving(cfg.disableInterleaving)
 {
@@ -214,7 +214,7 @@ int TageBase::predictorsize() {
 
     printf("nhist= %d; MInhist= %d; MAXHIST= %d; STORAGESIZE= %d, %dKB; "
             "NBENTRY= %d\n",
-            nhist, minhist, minhist, STORAGESIZE, STORAGESIZE / 1024, NBENTRY);
+            nhist, minhist, maxhist, STORAGESIZE, STORAGESIZE / 1024, NBENTRY);
 
     return (STORAGESIZE);
 }
@@ -271,7 +271,7 @@ int TageBase::gindex(unsigned int PC, int bank) {
 
 
 #ifdef USEPATH
-    int M = (m[bank] > 16) ? 16 : m[bank];
+    int M = (m[bank] > phistwidth) ? phistwidth : m[bank];
     index ^= F(phist, M, bank);
 #endif
 
@@ -317,6 +317,7 @@ bool TageBase::predict(uint64_t PC) {
 
     // 2. The TAGE prediction
     tagePredict(PC);
+    provider = tage_provider;
     return tage_pred;
 }
 
@@ -332,7 +333,7 @@ void TageBase::calcIndicesAndTags(uint64_t PC) {
     }
 
 
-    int T = (PC) % nbankhigh;
+    int T = (PC ^ (phist & ((1 << m[born]) - 1))) % nbankhigh;
     T = disableInterleaving ? 1 : T;
 
     for (int i = born; i <= nhist; i++)
@@ -342,7 +343,7 @@ void TageBase::calcIndicesAndTags(uint64_t PC) {
             T = T % nbankhigh;
         }
 
-    T = (PC) % nbanklow;
+    T = (PC ^ (phist & ((1 << m[1]) - 1))) % nbanklow;
     T = disableInterleaving ? 1 : T;
 
     for (int i = 1; i <= born - 1; i++)
@@ -435,25 +436,35 @@ void TageBase::tagePredict(uint64_t PC) {
 
 void TageBase::updateHistory(const uint64_t pc, const bool taken,
                                 const OpType opType, const uint64_t target) {
-    int historyBits = taken;
+
+    bool indirect = (opType == OPTYPE_CALL_INDIRECT_COND) ||
+                    (opType == OPTYPE_CALL_INDIRECT_UNCOND) ||
+                    (opType == OPTYPE_JMP_INDIRECT_COND) ||
+                    (opType == OPTYPE_JMP_INDIRECT_UNCOND) ||
+                    (opType == OPTYPE_RET_COND) ||
+                    (opType == OPTYPE_RET_UNCOND);
+
+    int tbits = indirect ? nHistBits+1 : nHistBits;
+
+    int historyBits = taken ? 1 : 0;
     if (nHistBits > 1) {
         historyBits ^= pc ^ (pc >> 2);
     }
-    if (takenHistory) {
-        // don't update the history if the branch was not taken
-        if (!taken) {
-            return;
-        }
-        historyBits = (pc >> 2) ^ (target >> 3);
-    }
-
     ghr_l = (ghr_l << 1) | (taken & 1);
+    int PATH = pc ^ (pc >> 2) ^ (pc >> 4);
 
-    for (int t = 0; t < nHistBits; t++) {
+    for (int t = 0; t < tbits; t++) {
         // update  history
+
         bool brDir = (historyBits & 1);
         updateGHist(brDir);
         historyBits >>= 1;
+#ifdef USEPATH
+        int PATHBIT = (PATH & 127);
+        PATH >>= 1;
+        phist = (phist << 1) ^ PATHBIT;
+        phist &= (1<<phistwidth)-1;
+#endif
     }
 }
 
@@ -469,7 +480,7 @@ void TageBase::updateGHist(const bool bit) {
 
 
 int TageBase::idxChooser() {
-    bool add1 = (altConf == LowConf);
+    bool add1 = (altConf != LowConf);
     return ((((HitBank - 1) / 8) << 1) + add1) % (size_use_alt - 1);
 }
 
@@ -481,8 +492,9 @@ unsigned TageBase::chooseProvider() {
     // low) and USE_ALT_ON_NA is positive  use the alternate prediction
 
     // If the longest is somehow certain use its prediction.
-    if (tageConf != LowConf)
+    if (tageConf != LowConf) {
         return LONGEST;
+    }
 
     // Use on low confidence if the USE_ALT_ON_NA is negative
     if (use_alt_on_na[idxChooser()] < 0) {
@@ -547,7 +559,7 @@ TageBase::allocateTables(int nalloc, uint64_t PC, bool resolveDir) {
 
     // T is the number of entries additionally allocated to at
     // least one entry per missprediction.
-    int T = nalloc -1;
+    int T = nalloc - 1;
 
     int A = 1;
     if ((MYRANDOM() & 127) < 32) A = 2;

bpmodels/tage/tage.h

@@ -122,15 +122,14 @@ class TageBase : public BasePredictor {
 
     // the counter(s) to chose between longest match and alternate prediction on
     // TAGE when weak counters
-    static const int log_size_use_alt = 4;
     unsigned tageConf; // Different confidence levels
     unsigned baseConf;
     unsigned altConf;
     int provVal;
     // For chooser between alt and longest match
     const int alt_width = 5;
-    static const int size_use_alt = (1 << (log_size_use_alt));
-    int8_t use_alt_on_na[size_use_alt];
+    const int size_use_alt;
+    int8_t use_alt_on_na[10]; // 10 is not the actual size, but the maximum.
 
     int TICK;  // for the reset of the u counter
 
@@ -317,7 +316,7 @@ class TageBase : public BasePredictor {
 struct TageConfig {
 
     int nhist = 36;
-    int minhist = 4;
+    int minhist = 6;
     int maxhist = 3000;
     int LogG = 10;
     int LogB = 13;
@@ -366,5 +365,4 @@ class Tage64k : public TageBase {
 
 
 
-
 };  // namespace LLBP