Bundle BCR for two source inst for OCL

Add bundle conflict reduction for two source instructions for OCL.

Author: bcheng0127

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -4842,6 +4842,11 @@ namespace IGC
         {
             SaveOption(vISA_enableBCR, true);
         }
+        if (context->type == ShaderType::OPENCL_SHADER &&
+            m_program->m_Platform->supportTwoSrcBundleConflictReduction())
+        {
+            SaveOption(vISA_twoSrcBCR, true);
+        }
         if (context->type == ShaderType::OPENCL_SHADER &&
             m_program->m_Platform->supportDpasInstruction())
         {

IGC/Compiler/CISACodeGen/Platform.hpp

@@ -579,6 +579,11 @@ bool supportBfnInstruction() const
     return isProductChildOf(IGFX_XE_HP_SDV);
 }
 
+bool supportTwoSrcBundleConflictReduction() const
+{
+    return isProductChildOf(IGFX_ARROWLAKE);
+}
+
 bool supportDpasInstruction() const
 {
     return isProductChildOf(IGFX_XE_HP_SDV) && m_platformInfo.eProductFamily != IGFX_METEORLAKE &&

visa/GraphColor.cpp

@@ -806,6 +806,39 @@ void BankConflictPass::setupBankConflictsforDPAS(G4_INST *inst) {
   return;
 }
 
+
+void BankConflictPass::setupBundleConflictsforTwoSrcsInst(G4_INST *inst) {
+  vISA_ASSERT(inst->getNumSrc() == 2, "Only support two source operands instructions");
+
+  G4_Declare *dcls[2];
+  G4_Declare *opndDcls[2];
+  unsigned offset[2];
+
+  for (int i = 0; i < 2; i += 1) {
+    dcls[i] = nullptr;
+    opndDcls[i] = nullptr;
+
+    G4_Operand *src = inst->getSrc(i);
+    if (!src || !src->isSrcRegRegion() || src->isAreg()) {
+      // bank conflict not possible
+      continue;
+    }
+
+    dcls[i] = GetTopDclFromRegRegion(src);
+    opndDcls[i] = src->getBase()->asRegVar()->getDeclare();
+    offset[i] = (opndDcls[i]->getOffsetFromBase() + src->getLeftBound()) /
+                gra.kernel.numEltPerGRF<Type_UB>();
+  }
+
+  // Add potential bundle conflicts
+  if (dcls[0] && dcls[1]) {
+    gra.addBundleConflictDcl(dcls[0], dcls[1], offset[0] - offset[1]);
+    gra.addBundleConflictDcl(dcls[1], dcls[0], offset[1] - offset[0]);
+  }
+
+  return;
+}
+
 void BankConflictPass::setupBankConflictsforMad(G4_INST *inst) {
   BankConflict srcBC[3];
   unsigned offset[3];
@@ -1022,10 +1055,10 @@ void BankConflictPass::setupBankConflictsForBBTGL(G4_BB *bb,
       } else {
         setupBankConflictsforMad(inst);
       }
-    } else if (gra.forceBCR && !forGlobal &&
+    } else if ((gra.forceBCR || gra.twoSrcBundleBCR) && !forGlobal &&
                inst->getNumSrc() == 2) {
       threeSourceInstNum++;
-      setupBankConflictsforMad(inst);
+      setupBundleConflictsforTwoSrcsInst(inst);
     }
   }
 
@@ -10103,6 +10136,21 @@ bool GlobalRA::tryHybridRA() {
   copyMissingAlignment();
   BankConflictPass bc(*this, false);
 
+  if (twoSrcBundleBCR) {
+    LivenessAnalysis liveAnalysis(*this, G4_GRF | G4_INPUT);
+    liveAnalysis.computeLiveness();
+
+    if (!liveAnalysis.getNumSelectedVar()) {
+      return false;
+    }
+
+    RPE rpe(*this, &liveAnalysis);
+    rpe.run();
+    if (rpe.getMaxRP() >= kernel.getNumRegTotal() - 24) {
+      // No two src bundle conflict reduction if high register pressure
+      twoSrcBundleBCR = false;
+    }
+  }
 
   LocalRA lra(bc, *this);
   if (lra.localRA()) {

visa/GraphColor.h

@@ -71,6 +71,7 @@ class BankConflictPass {
 
   void setupBankConflictsforTwoGRFs(G4_INST *inst);
   void setupBankConflictsforMad(G4_INST *inst);
+  void setupBundleConflictsforTwoSrcsInst(G4_INST *inst);
   void setupBankConflictsForBB(G4_BB *bb, unsigned &threeSourceInstNum,
                                unsigned &sendInstNum, unsigned numRegLRA,
                                unsigned &internalConflict);
@@ -1487,6 +1488,7 @@ class GlobalRA {
   bool useLocalRA = false;
   bool favorBCR = false;
   bool forceBCR = false;
+  bool twoSrcBundleBCR = false;
   uint32_t nextSpillOffset = 0;
   uint32_t scratchOffset = 0;
 
@@ -1783,6 +1785,12 @@ class GlobalRA {
     allocVar(dcl).bundleConflicts.clear();
   }
 
+  void clearAllBundleConflictDcl() {
+    for (auto dcl : kernel.Declares) {
+      clearBundleConflictDcl(dcl);
+    }
+  }
+
   const std::vector<BundleConflict> &
   getBundleConflicts(const G4_Declare *dcl) const {
     return getVar(dcl).bundleConflicts;
@@ -1921,6 +1929,7 @@ class GlobalRA {
       verifyAugmentation = std::make_unique<VerifyAugmentation>();
     }
     forceBCR = kernel.getOption(vISA_forceBCR);
+    twoSrcBundleBCR = kernel.getOption(vISA_twoSrcBCR);
     // 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

@@ -359,7 +359,7 @@ bool LocalRA::localRAPass(bool doRoundRobin, bool doSplitLLR) {
         gra.useHybridRAwithSpill && !doRoundRobin);
   }
 
-  if (needGlobalRA && doRoundRobin) {
+  if (needGlobalRA && (doRoundRobin || gra.twoSrcBundleBCR)) {
     undoLocalRAAssignments(true);
   }
 
@@ -392,7 +392,7 @@ bool LocalRA::localRA() {
   bool reduceBCInRR = false;
 
   if (builder.getOption(vISA_LocalBankConflictReduction) &&
-      builder.hasBankCollision()) {
+      (builder.hasBankCollision() || gra.twoSrcBundleBCR)) {
     reduceBCInRR = bc.setupBankConflictsForKernel(
         doRoundRobin, reduceBCInTAandFF, numRegLRA, highInternalConflict);
   }
@@ -428,7 +428,7 @@ bool LocalRA::localRA() {
   }
 
   if (!doRoundRobin) {
-    if (gra.forceBCR && doBCR) {
+    if ((gra.forceBCR || gra.twoSrcBundleBCR) && doBCR) {
       RA_TRACE(std::cout << "\t--first-fit BCR RA\n");
       needGlobalRA = localRAPass(false, doSplitLLR);
     }
@@ -442,6 +442,8 @@ bool LocalRA::localRA() {
         globalLRSize = 0;
       }
       specialAlign();
+      gra.clearAllBundleConflictDcl();
+      gra.twoSrcBundleBCR = false;
       needGlobalRA = localRAPass(false, doSplitLLR);
     }
     gra.favorBCR |= doBCR && kernel.useAutoGRFSelection() &&

visa/include/VISAOptionsDefs.h

@@ -330,6 +330,7 @@ DEF_VISA_OPTION(vISA_AbortOnSpillThreshold, ET_INT32, "-abortOnSpill", UNUSED,
                 0)
 DEF_VISA_OPTION(vISA_enableBCR, ET_BOOL, "-enableBCR", UNUSED, false)
 DEF_VISA_OPTION(vISA_forceBCR, ET_BOOL, "-forceBCR", UNUSED, false)
+DEF_VISA_OPTION(vISA_twoSrcBCR, ET_BOOL, "-twoSrcBCR", UNUSED, false)
 DEF_VISA_OPTION(vISA_NewAugmentation, ET_BOOL_TRUE, "-newaugmentation",
                 "USAGE: -newaugmentation "
                 "enable using augmentation with holes",