Refactor code for retrieving the number of accumulators so it returns correct value for TGL.

Change-Id: I629799518e5317712239f73a21069a3eb74a3f63

Author: weiyu-chen

visa/FlowGraph.cpp

@@ -3094,7 +3094,7 @@ void G4_Kernel::setKernelParameters()
     }
 
 
-    // Set the number of Acc
+    // Set the number of Acc. They are in the unit of GRFs (i.e., 1 accumulator is the same size as 1 GRF)
     unsigned overrideNumAcc = m_options->getuInt32Option(vISA_numGeneralAcc);
     if (overrideNumAcc > 0)
     {
@@ -3110,10 +3110,7 @@ void G4_Kernel::setKernelParameters()
             numAcc = 2;
         }
 
-        if (numRegTotal == 256)
-        {
-            numAcc *= 2;
-        }
+
     }
 }
 

visa/FlowGraph.h

@@ -1471,7 +1471,6 @@ class G4_Kernel
     void setNumSWSBTokens(int nSWSBs) { numSWSBTokens = nSWSBs; }
     uint32_t getNumSWSBTokens() const { return numSWSBTokens; }
 
-    void setNumAcc(int nAcc) { numAcc = nAcc; }
     uint32_t getNumAcc() const { return numAcc; }
 
     void setAsmCount(int count) { asmInstCount = count; }

visa/HWCapsOpen.inc

@@ -357,12 +357,6 @@
         return getGenxPlatform() >= GENX_TGLLP;
     }
 
-    // 1 means one SIMD16 float accumulator
-    uint32_t getNumACC() const
-    {
-        return 1;
-    }
-
     bool fuseURBMessage() const
     {
         return isGen12LP();

visa/HWConformity.cpp

@@ -4657,7 +4657,7 @@ struct AccAssignment
 
 void HWConformity::multiAccSubstitution(G4_BB* bb)
 {
-    int numGeneralAcc = builder.getNumACC();
+    int numGeneralAcc = kernel.getNumAcc();
 
     std::vector<AccInterval*> intervals;
 

visa/LocalScheduler/SWSB_G4IR.cpp

@@ -325,8 +325,8 @@ SBFootprint* G4_BB_SB::getFootprintForFlag(G4_Operand* opnd,
     LB = (unsigned short)(opnd->getLeftBound() + subRegOff * 16) * bitToBytes;
     RB = (unsigned short)(opnd->getRightBound() + subRegOff * 16) * bitToBytes;
 
-    LB += (builder.kernel.getNumRegTotal() + builder.getNumACC()) * G4_GRF_REG_NBYTES;
-    RB += (builder.kernel.getNumRegTotal() + builder.getNumACC()) * G4_GRF_REG_NBYTES;
+    LB += (builder.kernel.getNumRegTotal() + builder.kernel.getNumAcc()) * G4_GRF_REG_NBYTES;
+    RB += (builder.kernel.getNumRegTotal() + builder.kernel.getNumAcc()) * G4_GRF_REG_NBYTES;
 
     void* allocedMem = mem.alloc(sizeof(SBFootprint));
     SBFootprint* footprint = nullptr;
@@ -1008,8 +1008,8 @@ void SWSB::SWSBGenerator()
     }
 
     //Note that getNumFlagRegisters() treat each 16 bits as a flag register
-    LiveGRFBuckets LB(mem, kernel.getNumRegTotal() + fg.builder->getNumACC() + fg.builder->getNumFlagRegisters(), kernel);
-    LiveGRFBuckets globalSendsLB(mem, kernel.getNumRegTotal() + fg.builder->getNumACC() + fg.builder->getNumFlagRegisters(), kernel);
+    LiveGRFBuckets LB(mem, kernel.getNumRegTotal() + kernel.getNumAcc() + fg.builder->getNumFlagRegisters(), kernel);
+    LiveGRFBuckets globalSendsLB(mem, kernel.getNumRegTotal() + kernel.getNumAcc() + fg.builder->getNumFlagRegisters(), kernel);
 
     SWSBDepDistanceGenerator(p, LB, globalSendsLB);
 
@@ -4110,8 +4110,8 @@ void G4_BB_SB::getLiveBucketsFromFootprint(SBFootprint* firstFootprint, SBBucket
         }
         else if (footprint->fType == FLAG_T)
         {
-            startBucket = footprint->LeftB + aregOffset + builder.getNumACC();
-            endBucket = footprint->RightB + aregOffset + builder.getNumACC();
+            startBucket = footprint->LeftB + aregOffset + builder.kernel.getNumAcc();
+            endBucket = footprint->RightB + aregOffset + builder.kernel.getNumAcc();
         }
 
         for (int j = startBucket; j < endBucket + 1; j++)