BCR tuning

BCR tuning

Author: bcheng0127

visa/GraphColor.cpp

@@ -1022,7 +1022,7 @@ void BankConflictPass::setupBankConflictsForBBTGL(G4_BB *bb,
       } else {
         setupBankConflictsforMad(inst);
       }
-    } else if (gra.kernel.getOption(vISA_forceBCR) && !forGlobal &&
+    } else if (gra.forceBCR && !forGlobal &&
                inst->getNumSrc() == 2) {
       threeSourceInstNum++;
       setupBankConflictsforMad(inst);
@@ -3499,7 +3499,7 @@ bool Augmentation::markNonDefaultMaskDef() {
         nonDefaultMaskDefFound = true;
       }
 
-      if (kernel.getOption(vISA_forceBCR) &&
+      if (gra.favorBCR &&
           gra.getBankConflict(dcl) != BANK_CONFLICT_NONE) {
         gra.setAugmentationMask(dcl, AugmentationMasks::NonDefault);
         nonDefaultMaskDefFound = true;
@@ -6660,7 +6660,6 @@ bool GraphColor::assignColors(ColorHeuristic colorHeuristicGRF,
 
       bool failed_alloc = false;
       G4_Declare *dcl = lrVar->getDeclare();
-
       if (!(noIndirForceSpills && liveAnalysis.isAddressSensitive(lr_id)) &&
           forceSpill &&
           (dcl->getRegFile() == G4_GRF || dcl->getRegFile() == G4_FLAG) &&
@@ -6723,7 +6722,7 @@ bool GraphColor::assignColors(ColorHeuristic colorHeuristicGRF,
         // pass) then abort on spill
         //
         if ((heuristic == ROUND_ROBIN ||
-             (doBankConflict && !kernel.getOption(vISA_forceBCR))) &&
+             (doBankConflict && !gra.forceBCR)) &&
             (lr->getRegKind() == G4_GRF || lr->getRegKind() == G4_FLAG)) {
           return false;
         } else if (kernel.fg.isPseudoDcl(dcl)) {
@@ -7345,7 +7344,7 @@ bool GraphColor::regAlloc(bool doBankConflictReduction,
           return false;
         }
 
-        if (!kernel.getOption(vISA_forceBCR)) {
+        if (!gra.forceBCR) {
           if (!success && doBankConflictReduction) {
             resetTemporaryRegisterAssignments();
             assignColors(FIRST_FIT);
@@ -10113,6 +10112,16 @@ bool GlobalRA::tryHybridRA() {
     return true;
   }
 
+  //Skip hybridRA if BCR is needed
+  if (favorBCR) {
+    lra.undoLocalRAAssignments(true);
+    // Restore alignment in case LRA modified it
+    copyAlignment();
+    //reset favorBCR
+    favorBCR = forceBCR;
+    return false;
+  }
+
   if (useHybridRAwithSpill) {
     insertPhyRegDecls();
   } else {
@@ -10433,14 +10442,14 @@ std::pair<bool, bool> GlobalRA::bankConflict() {
   bool doBankConflictReduction = false, highInternalConflict = false;
   if (builder.getOption(vISA_LocalBankConflictReduction) &&
       builder.hasBankCollision()) {
-    bool reduceBCInRR = false;
-    bool reduceBCInTAandFF = false;
+    bool reduceBC = false;
+    bool threeSouceBCR = false;
     BankConflictPass bc(*this, true);
 
-    reduceBCInRR = bc.setupBankConflictsForKernel(
-        true, reduceBCInTAandFF, SECOND_HALF_BANK_START_GRF * 2,
+    reduceBC = bc.setupBankConflictsForKernel(
+        true, threeSouceBCR, SECOND_HALF_BANK_START_GRF * 2,
         highInternalConflict);
-    doBankConflictReduction = reduceBCInRR && reduceBCInTAandFF;
+    doBankConflictReduction = reduceBC && threeSouceBCR;
   }
   return std::make_pair(doBankConflictReduction, highInternalConflict);
 }
@@ -10771,11 +10780,8 @@ GlobalRA::insertSpillCode(bool enableSpillSpaceCompression,
 
 bool GlobalRA::rerunGRAIter(bool rerunGRA)
 {
-  if (getIterNo() == 0 && (rerunGRA || kernel.getOption(vISA_forceBCR))) {
-    if (kernel.getOption(vISA_forceBCR)) {
-      // FIXME: We shouldn't modify options. Use local bool flag instead.
-      kernel.getOptions()->setOption(vISA_forceBCR, false);
-    }
+  if (getIterNo() == 0 && (rerunGRA || forceBCR)) {
+    forceBCR = false;
     return true;
   }
   return false;
@@ -10993,6 +10999,7 @@ int GlobalRA::coloringRegAlloc() {
     if (!fastCompile) {
       rpe.run();
       writeVerboseRPEStats(rpe);
+      favorBCR |= doBankConflictReduction && (rpe.getMaxRP() < kernel.getNumRegTotal() - 16);
     }
     GraphColor coloring(liveAnalysis, false, forceSpill);
 

visa/GraphColor.h

@@ -1485,6 +1485,8 @@ class GlobalRA {
   bool useFastRA = false;
   bool useHybridRAwithSpill = false;
   bool useLocalRA = false;
+  bool favorBCR = false;
+  bool forceBCR = false;
   uint32_t nextSpillOffset = 0;
   uint32_t scratchOffset = 0;
 
@@ -1918,7 +1920,7 @@ class GlobalRA {
     if (kernel.getOptions()->getOption(vISA_VerifyAugmentation)) {
       verifyAugmentation = std::make_unique<VerifyAugmentation>();
     }
-
+    favorBCR = forceBCR = kernel.getOption(vISA_forceBCR);
     // Set callWA condition.
     //    Call return ip and mask need wa only for non-entry functions. As call
     //    WA also needs a temp, we conservatively add WA for

visa/LocalRA.cpp

@@ -428,7 +428,7 @@ bool LocalRA::localRA() {
   }
 
   if (!doRoundRobin) {
-    if (kernel.getOption(vISA_forceBCR) && doBCR) {
+    if (gra.forceBCR && doBCR) {
       RA_TRACE(std::cout << "\t--first-fit BCR RA\n");
       needGlobalRA = localRAPass(false, doSplitLLR);
     }
@@ -444,6 +444,7 @@ bool LocalRA::localRA() {
       specialAlign();
       needGlobalRA = localRAPass(false, doSplitLLR);
     }
+    gra.favorBCR |= doBCR;
   }
 
   if (needGlobalRA == false) {