Add a field for per thread data prolog offset

Change-Id: Ifad94dcd5aa55715c44d47089ed31a70fb4e5f7f

Author: weiyu-chen

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -4623,6 +4623,8 @@ void CEncoder::Compile()
 
     pOutput->m_scratchSpaceUsedByGtpin = jitInfo->numBytesScratchGtpin;
 
+    pOutput->m_offsetToSkipPerThreadDataLoad = jitInfo->offsetToSkipPerThreadDataLoad;
+
     COMPILER_TIME_END(m_program->GetContext(), TIME_CG_vISAEmitPass);
 }
 

IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp

@@ -1667,6 +1667,7 @@ void COpenCLKernel::FillKernel()
     }
 
     m_kernelInfo.m_executionEnivronment.HasGlobalAtomics = GetHasGlobalAtomics();
+    m_kernelInfo.m_threadPayload.OffsetToSkipPerThreadDataLoad = ProgramOutput()->m_offsetToSkipPerThreadDataLoad;
 
 }
 

IGC/Compiler/CodeGenPublic.h

@@ -108,6 +108,7 @@ namespace IGC
         void*           m_funcRelocationTable;
         unsigned int    m_funcRelocationTableSize;
         unsigned int    m_funcRelocationTableEntries;
+        unsigned int    m_offsetToSkipPerThreadDataLoad = 0;
 
         void Destroy()
         {

visa/BinaryEncodingIGA.cpp

@@ -813,6 +813,19 @@ void BinaryEncodingIGA::DoAll()
     {
         inst.second->setGenOffset(inst.first->getPC());
     }
+    if (kernel.fg.builder->getHasPerThreadProlog())
+    {
+        // per thread data load is in the first BB
+#if 0
+        assert(kernel.fg.getNumBB() > 1 && "expect at least one prolog BB");
+        auto secondBB = *(std::next(kernel.fg.BBs.begin()));
+        assert(secondBB->size() > 0 && "expect at least one inst in second BB");
+        auto startInst = *(secondBB->begin());
+        kernel.fg.builder->getJitInfo()->offsetToSkipPerThreadDataLoad = (uint32_t)startInst->getGenOffset();
+#else
+        kernel.fg.builder->getJitInfo()->offsetToSkipPerThreadDataLoad = 8 * 16;
+#endif
+    }
 }
 
 SendDescArg BinaryEncodingIGA::getIGASendDescArg(G4_INST* sendInst) const

visa/BuildIR.h

@@ -507,6 +507,7 @@ class IR_Builder {
     bool hasNullReturnSampler = false;
 
     int perThreadInputSize = 0;
+    bool hasPerThreadProlog = false;
 
 public:
     PreDefinedVars preDefVars;
@@ -607,6 +608,8 @@ class IR_Builder {
 
     uint32_t getPerThreadInputSize() const { return perThreadInputSize; }
     void setPerThreadInputSize(uint32_t val) { perThreadInputSize = val; }
+    bool getHasPerThreadProlog() const { return hasPerThreadProlog; }
+    void setHasPerThreadProlog() { hasPerThreadProlog = true; }
 
     bool isOpndAligned( G4_Operand *opnd, unsigned short &offset, int align_byte );
 

visa/FlowGraph.cpp

@@ -595,9 +595,9 @@ void FlowGraph::constructFlowGraph(INST_LIST& instlist)
     // create the entry block of the flow graph
     //
     G4_BB* fstartBB = NULL;
-    G4_BB* curr_BB = entryBB = fstartBB = beginBB(labelMap, instlist.front());
+    G4_BB* curr_BB = fstartBB = beginBB(labelMap, instlist.front());
 
-    kernelInfo = new (mem)FuncInfo(UINT_MAX, entryBB, NULL);
+    kernelInfo = new (mem)FuncInfo(UINT_MAX, curr_BB, NULL);
 
     std::vector<G4_BB*> subroutineStartBB; // needed by handleExit()
 
@@ -837,7 +837,7 @@ void FlowGraph::constructFlowGraph(INST_LIST& instlist)
     // Ensure each block other than entry starts with a label.
     for (auto bb : BBs)
     {
-        if (bb != entryBB && !bb->empty())
+        if (bb != getEntryBB() && !bb->empty())
         {
             G4_INST *inst = bb->front();
             if (inst->isLabel())
@@ -1833,7 +1833,7 @@ void FlowGraph::removeUnreachableBlocks()
     //
     // assign DFS based pre/rpost ids to all blocks in the main program
     //
-    doDFS(entryBB, preId);
+    doDFS(getEntryBB(), preId);
 
     for (BB_LIST_ITER it = BBs.begin(), itEnd = BBs.end(); it != itEnd; ++it)
     {
@@ -1972,7 +1972,7 @@ void FlowGraph::removeRedundantLabels()
     for (BB_LIST_ITER it = BBs.begin(); it != BBs.end();)
     {
         G4_BB* bb = *it;
-        if (bb == entryBB)
+        if (bb == getEntryBB())
         {
             it++;
             continue;
@@ -4818,7 +4818,7 @@ void FlowGraph::findBackEdges()
     unsigned postID = 0;
     backEdges.clear();
 
-    DFSTraverse(entryBB, preId, postID, kernelInfo);
+    DFSTraverse(getEntryBB(), preId, postID, kernelInfo);
 
     for (auto fn : funcInfoTable)
     {
@@ -4860,7 +4860,9 @@ void FlowGraph::findNaturalLoops()
                     loopBody.insert(loopBlock->BBBeforeCall());
                 }
             }
-            else {
+            else 
+            {
+                auto entryBB = getEntryBB();
                 for (auto predBB : loopBlock->Preds)
                 {
                     if (!predBB->isInNaturalLoop())

visa/FlowGraph.h

@@ -657,7 +657,6 @@ class FlowGraph
 
 private:
 
-    G4_BB* entryBB;                             // entry block
     unsigned traversalNum;                      // used for flow graph traversals
     unsigned numBBId;                            // number of basic blocks
     bool     reducible;                            // reducibility of the graph
@@ -859,7 +858,8 @@ class FlowGraph
 
     void preprocess(INST_LIST& instlist);
 
-    FlowGraph(INST_LIST_NODE_ALLOCATOR& alloc, G4_Kernel* kernel, Mem_Manager& m) : entryBB(NULL), traversalNum(0), numBBId(0), reducible(true),
+    FlowGraph(INST_LIST_NODE_ALLOCATOR& alloc, G4_Kernel* kernel, Mem_Manager& m) : 
+      traversalNum(0), numBBId(0), reducible(true),
       doIPA(false), hasStackCalls(false), isStackCallFunc(false), autoLabelId(0),
       pKernel(kernel), mem(m), instListAlloc(alloc),
       builder(NULL), kernelInfo(NULL), globalOpndHT(m), framePtrDcl(NULL), stackPtrDcl(NULL),
@@ -961,8 +961,7 @@ class FlowGraph
     void resetLocalDataFlowData();
 
     unsigned getNumBB() const      {return numBBId;}
-    G4_BB* getEntryBB()        {return entryBB;}
-    void setEntryBB(G4_BB *entry) {entryBB = entry;}
+    G4_BB* getEntryBB()        {return BBs.front();}
 
     void doFilescopeVarLayout(IR_Builder& builder, DECLARE_LIST& declares, unsigned& fileScopeFrameOffset);
     void addFrameSetupDeclares(IR_Builder& builder, PhyRegPool& regPool);

visa/include/JitterDataStruct.h

@@ -74,6 +74,8 @@ typedef struct _CM_JIT_INFO {
     void* freeGRFInfo;
     unsigned int freeGRFInfoSize;
     unsigned char numBytesScratchGtpin;
+
+    uint32_t offsetToSkipPerThreadDataLoad = 0; 
 } FINALIZER_INFO;
 
 #endif // _CM_JITTERDATASTRUCT_