group instructions together for performance

Change-Id: I94b286e910729d4cdc33880f3c4e71849a117886

Author: bcheng0127

visa/Gen4_IR.cpp

@@ -8201,3 +8201,34 @@ G4_INST* G4_InstMath::cloneInst()
         src0, src1, getMathCtrl(), option, getLineNo(), getCISAOff(), getSrcFilename());
 }
 
+
+bool G4_INST::isRMWCombinCandidate(G4_INST *next)
+{
+    if (isSend() || next->isSend())
+    {
+        return false;
+    }
+
+    G4_Operand* dst = getDst();
+    G4_Operand* nextDst = next->getDst();
+    //The current instruction must has Integer macro dst as well
+    if (nextDst &&
+        nextDst->isGen12IntegerMacroDst() &&
+        dst->getType() == nextDst->getType())
+    {
+
+        unsigned short lastLB = (unsigned short)dst->getLinearizedStart();
+        unsigned short lastRB = (unsigned short)dst->getLinearizedEnd();
+        unsigned short nextLB = (unsigned short)nextDst->getLinearizedStart();
+        unsigned short nextRB = (unsigned short)nextDst->getLinearizedEnd();
+
+        if (lastLB / G4_GRF_REG_NBYTES == nextLB / G4_GRF_REG_NBYTES &&
+            lastRB / G4_GRF_REG_NBYTES == nextRB / G4_GRF_REG_NBYTES &&
+            (lastLB > nextRB || nextLB > lastRB))
+        {
+            return true;
+        }
+    }
+
+    return false;
+}

visa/Gen4_IR.hpp

@@ -957,6 +957,7 @@ class G4_INST
     }
 
     bool mayExpandToAccMacro() const;
+    bool isRMWCombinCandidate(G4_INST* next);
 
     Gen4_Operand_Number getSrcOperandNum(int srcPos) const
     {
@@ -2326,6 +2327,25 @@ class G4_Operand
                 return type;
         }
     }
+    bool isGen12IntegerMacroDst()
+    {
+        if (isNullReg())
+        {
+            return false;
+        }
+
+        if (type == Type_UW ||
+            type == Type_W ||
+            type == Type_B ||
+            type == Type_UB ||
+            type == Type_D ||
+            type == Type_UD)
+        {
+            return true;
+        }
+
+        return false;
+    }
 };
 
 class G4_VarBase

visa/LocalScheduler/SWSB_G4IR.cpp

@@ -4199,12 +4199,18 @@ void G4_BB_SB::getGRFFootprintForIndirect(SBNode* node,
 void G4_BB_SB::getGRFBuckets(SBNode* node,
     SBFootprint* footprint,
     Gen4_Operand_Number opndNum,
-    std::vector<SBBucketDescr>& BDvec)
+    std::vector<SBBucketDescr>& BDvec,
+    bool GRFOnly)
 {
     SBFootprint* curFootprint = footprint;
 
     while (curFootprint != nullptr)
     {
+        if (GRFOnly && (curFootprint->fType != GRF_T))
+        {
+            continue;
+        }
+
         int aregOffset = totalGRFNum;
         int startingBucket = curFootprint->LeftB / G4_GRF_REG_NBYTES;
         int endingBucket = curFootprint->RightB / G4_GRF_REG_NBYTES;
@@ -4273,7 +4279,7 @@ void G4_BB_SB::getGRFBucketsForOperands(SBNode* node,
         {
             continue;
         }
-        getGRFBuckets(node, footprint, opndNum, BDvec);
+        getGRFBuckets(node, footprint, opndNum, BDvec, GRFOnly);
     }
 
     return;
@@ -4451,7 +4457,6 @@ void G4_BB_SB::SBDDD(G4_BB* bb,
             getGRFBucketDescrs(node, liveBDvec, false);
         }
 
-
         // For ALU instructions without GRF usage
         if (distanceHonourInstruction(curInst))
         {

visa/LocalScheduler/SWSB_G4IR.h

@@ -136,6 +136,23 @@ namespace vISA
 
         void setOffset(unsigned short o) { offset = o; }
 
+        bool hasOverlap(SBFootprint* liveFootprint) const
+        {
+            SBFootprint* curFootprint2Ptr = liveFootprint;
+            while (curFootprint2Ptr)
+            {
+                // Negative of no overlap: !(LeftB > curFootprint2Ptr->RightB || RightB < curFootprint2Ptr->LeftB)
+                if (fType == curFootprint2Ptr->fType &&
+                    LeftB <= curFootprint2Ptr->RightB && RightB >= curFootprint2Ptr->LeftB)
+                {
+                    return true;
+                }
+                curFootprint2Ptr = curFootprint2Ptr->next;
+            }
+
+            return false;
+        }
+
         bool hasOverlap(SBFootprint *liveFootprint, unsigned short &internalOffset) const
         {
             SBFootprint *curFootprint2Ptr = liveFootprint;
@@ -984,14 +1001,11 @@ namespace vISA
         SBFootprint* getFootprintForFlag(G4_Operand* opnd,
             Gen4_Operand_Number opnd_num,
             G4_INST* inst);
-        void getGRFBuckets(SBNode *node,
-            SBFootprint* footprint,
-            Gen4_Operand_Number opndNum,
-            std::vector<SBBucketDescr>& BDvec);
         bool getFootprintForOperand(SBNode *node,
             G4_INST *inst,
             G4_Operand* opnd,
             Gen4_Operand_Number opnd_num);
+        void getGRFBuckets(SBNode* node, SBFootprint* footprint, Gen4_Operand_Number opndNum, std::vector<SBBucketDescr>& BDvec, bool GRFOnly);
         bool getGRFFootPrintOperands(SBNode *node,
             G4_INST *inst,
             Gen4_Operand_Number first_opnd,