Changes in code.

Change-Id: Icb748a042da2f2b8e5fdd901a2f86fa67289d8fa

Author: houjenko

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -3733,7 +3733,6 @@ namespace IGC
         {
             params.push_back(param_uptr("-debug", literal_deleter));
         }
-
         if (context->getModuleMetaData()->compOpt.FastVISACompile)
         {
             params.push_back(param_uptr("-fasterRA", literal_deleter));
@@ -3794,7 +3793,7 @@ namespace IGC
             }
         }
     }
-    void CEncoder::InitVISABuilderOptions(TARGET_PLATFORM VISAPlatform, bool canAbortOnSpill, bool hasStackCall, bool enableVISA_IR)
+    void CEncoder::InitVISABuilderOptions(TARGET_PLATFORM VISAPlatform, bool canAbortOnSpill, bool hasStackCall, bool enableVISA_IR, CODE_PATCH_T mode)
     {
         CodeGenContext* context = m_program->GetContext();
         bool KernelDebugEnable = false;
@@ -3836,6 +3835,7 @@ namespace IGC
             SaveOption(vISA_DumpCompilerStats, true);
         }
 
+
         if (IGC_IS_FLAG_ENABLED(ForceFFIDOverwrite)/*|| m_program->m_Platform->WaOverwriteFFID()*/)
         {
             unsigned int ffid[unsigned(ShaderType::END)] = {
@@ -4329,7 +4329,7 @@ namespace IGC
         }
     }
 
-    void CEncoder::InitEncoder(bool canAbortOnSpill, bool hasStackCall)
+    void CEncoder::InitEncoder(bool canAbortOnSpill, bool hasStackCall, VISAKernel* prevKernel, CODE_PATCH_T mode)
     {
         m_aliasesMap.clear();
         m_encoderState.m_SubSpanDestination = false;
@@ -4367,7 +4367,7 @@ namespace IGC
         V(CreateVISABuilder(vbuilder, builderMode, builderOpt, VISAPlatform, params.size(), params.data(),
             &m_vISAWaTable));
 
-        InitVISABuilderOptions(VISAPlatform, canAbortOnSpill, hasStackCall, builderOpt == VISA_BUILDER_BOTH);
+        InitVISABuilderOptions(VISAPlatform, canAbortOnSpill, hasStackCall, builderOpt == VISA_BUILDER_BOTH, mode);
 
         // Pass all build options to builder
         SetBuilderOptions(vbuilder);
@@ -4429,7 +4429,7 @@ namespace IGC
             asmName = "kernel.asm";
         }
 
-        V(vbuilder->AddKernel(vKernel, kernelName.c_str()));
+        V(vbuilder->AddKernel(vKernel, kernelName.c_str(), prevKernel));
         V(vKernel->AddKernelAttribute("OutputAsmPath", asmName.length(), asmName.c_str()));
 
         vMainKernel = vKernel;

IGC/Compiler/CISACodeGen/CISABuilder.hpp

@@ -130,9 +130,9 @@ namespace IGC
     class CEncoder
     {
     public:
-        void InitEncoder(bool canAbortOnSpill, bool hasStackCall);
+        void InitEncoder(bool canAbortOnSpill, bool hasStackCall, VISAKernel* prevkernel, CODE_PATCH_T mode);
         void InitBuildParams(llvm::SmallVector<std::unique_ptr< char, std::function<void(char*)>>, 10> & params);
-        void InitVISABuilderOptions(TARGET_PLATFORM VISAPlatform, bool canAbortOnSpill, bool hasStackCall, bool enableVISA_IR);
+        void InitVISABuilderOptions(TARGET_PLATFORM VISAPlatform, bool canAbortOnSpill, bool hasStackCall, bool enableVISA_IR, CODE_PATCH_T mode = CodePatch_NOT_HANDLED);
         SEncoderState CopyEncoderState();
         void SetEncoderState(SEncoderState& newState);
 
@@ -233,6 +233,7 @@ namespace IGC
         void File(std::string& s);
         void PredAdd(CVariable* flag, CVariable* dst, CVariable* src0, CVariable* src1);
         void DebugLinePlaceholder();
+        void SetPrevKernel(VISAKernel* kernel);
         void SetCurrentInst(llvm::Instruction *inst);
 
         inline void Jump(uint label);
@@ -532,6 +533,9 @@ namespace IGC
         // encoder states
         SEncoderState m_encoderState;
 
+        // directly patch code from the recorded simd
+        CODE_PATCH_T m_codePatch;
+
         llvm::DenseMap<SAlias, CVariable*, SAliasMapInfo> m_aliasesMap;
 
         // vISA needs its own Wa-table as some of the W/A are applicable

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -358,6 +358,18 @@ void EmitPass::CreateKernelShaderMap(CodeGenContext* ctx, MetaDataUtils* pMdUtil
     }
 }
 
+CODE_PATCH_T EmitPass::GetCodePatchMode(
+    SIMDMode fromSimd, ShaderDispatchMode fromDispatch,
+    SIMDMode toSimd,   ShaderDispatchMode toDispatch)
+{
+    switch (fromSimd)
+    {
+    default:
+        return CodePatch_NOT_HANDLED;
+    }
+    return CodePatch_NOT_HANDLED;
+}
+
 bool EmitPass::runOnFunction(llvm::Function& F)
 {
     m_pCtx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
@@ -440,8 +452,27 @@ bool EmitPass::runOnFunction(llvm::Function& F)
         {
             return false;
         }
+
+        m_prevShader = m_pCtx->m_prevShader;
+        VISAKernel* prevKernel = nullptr;
+        CODE_PATCH_T mode = CodePatch_NOT_HANDLED;
+        if (m_prevShader)
+        {
+            mode = GetCodePatchMode(
+                m_prevShader->m_SIMDSize,
+                m_prevShader->m_ShaderDispatchMode,
+                m_currShader->m_SIMDSize,
+                m_currShader->m_ShaderDispatchMode
+                );
+            if (mode != CodePatch_NOT_HANDLED)
+            {
+                prevKernel = m_prevShader->GetEncoder().GetVISAKernel();
+            }
+        }
+
         // call builder after pre-analysis pass where scratchspace offset to VISA is calculated
-        m_encoder->InitEncoder(m_canAbortOnSpill, hasStackCall);
+        m_encoder->InitEncoder(m_canAbortOnSpill, hasStackCall, prevKernel, mode);
+
         initDefaultRoundingMode();
         m_currShader->PreCompile();
         if (hasStackCall)
@@ -687,6 +718,7 @@ bool EmitPass::runOnFunction(llvm::Function& F)
         Function* currHead = m_FGA ? m_FGA->getGroupHead(&F) : &F;
         bool compileWithSymbolTable = (currHead == uniqueEntry);
         m_encoder->Compile(compileWithSymbolTable);
+        m_pCtx->m_prevShader = m_currShader;
         // if we are doing stack-call, do the following:
         // - Hard-code a large scratch-space for visa
         if (hasStackCall)

IGC/Compiler/CISACodeGen/EmitVISAPass.hpp

@@ -77,6 +77,10 @@ namespace IGC
         virtual bool runOnFunction(llvm::Function& F) override;
         virtual llvm::StringRef getPassName() const  override { return "EmitPass"; }
 
+        CODE_PATCH_T GetCodePatchMode(
+                SIMDMode fromSimd, ShaderDispatchMode fromDispatch,
+                SIMDMode toSimd,   ShaderDispatchMode toDispatch);
+
         void CreateKernelShaderMap(CodeGenContext* ctx, IGC::IGCMD::MetaDataUtils* pMdUtils, llvm::Function& F);
 
         void Frc(const SSource& source, const DstModifier& modifier);
@@ -534,6 +538,7 @@ namespace IGC
         ShaderDispatchMode m_ShaderMode;
         CShaderProgram::KernelShaderMap& m_shaders;
         CShader* m_currShader;
+        CShader* m_prevShader;
         CEncoder* m_encoder;
         const llvm::DataLayout* m_DL;
         CoalescingEngine* m_CE;

IGC/Compiler/CodeGenPublic.h

@@ -752,6 +752,9 @@ namespace IGC
         std::vector<unsigned> m_indexableTempSize;
         bool         m_highPsRegisterPressure = 0;
 
+        // Record previous simd for code patching
+        CShader* m_prevShader = nullptr;
+
         // For IR dump after pass
         unsigned     m_numPasses = 0;
         bool m_threadCombiningOptDone = false;

visa/BuildCISAIR.h

@@ -79,7 +79,7 @@ class CISA_IR_Builder : public VISABuilder
         const char* flags[],
         const PWA_TABLE pWaTable = nullptr);
     static int DestroyBuilder(CISA_IR_Builder *builder);
-    VISA_BUILDER_API virtual int AddKernel(VISAKernel *& kernel, const char* kernelName);
+    VISA_BUILDER_API virtual int AddKernel(VISAKernel *& kernel, const char* kernelName, VISAKernel* prevKernel = nullptr);
     VISA_BUILDER_API virtual int AddFunction(VISAFunction *& function, const char* functionName);
     VISA_BUILDER_API virtual int Compile(const char * isaFileNameint, std::ostream * os = nullptr, bool emit_visa_only = false);
 

visa/BuildCISAIRImpl.cpp

@@ -361,7 +361,7 @@ int CISA_IR_Builder::ClearAsmTextStreams()
     return VISA_FAILURE;
 }
 
-int CISA_IR_Builder::AddKernel(VISAKernel *& kernel, const char* kernelName)
+int CISA_IR_Builder::AddKernel(VISAKernel *& kernel, const char* kernelName, VISAKernel* prevKernel)
 {
 
     if( kernel != NULL )
@@ -377,7 +377,7 @@ int CISA_IR_Builder::AddKernel(VISAKernel *& kernel, const char* kernelName)
     m_kernel->setIsKernel(true);
     m_kernels.push_back(kerneltemp);
     m_kernel->setVersion((unsigned char)this->m_header.major_version, (unsigned char)this->m_header.minor_version);
-    m_kernel->InitializeKernel(kernelName);
+    m_kernel->InitializeKernel(kernelName, prevKernel);
     m_kernel->SetGTPinInit(getGtpinInit());
     this->m_kernel_count++;
 

visa/FlowGraph.cpp

@@ -806,7 +806,6 @@ void FlowGraph::constructFlowGraph(INST_LIST& instlist)
 
     if (builder->getOption(vISA_DumpDotAll))
     {
-        pKernel->dumpDotFile("beforeRemoveRedundantLabels");
     }
 
     removeRedundantLabels();

visa/FlowGraph.h

@@ -675,6 +675,7 @@ class FlowGraph
     bool     isStackCallFunc;                    // indicates the function itself is a STACK_CALL function
     unsigned int autoLabelId;
     G4_Kernel* pKernel;                         // back pointer to the kernel object
+    G4_Kernel* m_prevKernel;                    // back pointer to the previous kernel object
 
     // map each BB to its local RA GRF usage summary, populated in local RA
     std::map<G4_BB*, PhyRegSummary*> bbLocalRAMap;
@@ -852,6 +853,11 @@ class FlowGraph
         return pKernel;
     }
 
+    G4_Kernel* getPrevKernel()
+    {
+        return m_prevKernel;
+    }
+
     void mergeFReturns();
 
     G4_Declare*& getFramePtrDcl()                   {return framePtrDcl;}
@@ -926,6 +932,7 @@ class FlowGraph
 
     // functions for structure analysis
     G4_Kernel *getKernel() const { return pKernel; }
+    G4_Kernel* getPrevKernel() const { return m_prevKernel; }
     G4_Label* insertEndif( G4_BB* bb, unsigned char execSize, bool createLabel );
     void setJIPForEndif( G4_INST* endif, G4_INST* target, G4_BB* targetBB);
     void convertGotoToJmpi(G4_INST *gotoInst)
@@ -963,10 +970,10 @@ class FlowGraph
     FlowGraph(const FlowGraph&) = delete;
     FlowGraph& operator=(const FlowGraph&) = delete;
 
-    FlowGraph(INST_LIST_NODE_ALLOCATOR& alloc, G4_Kernel* kernel, Mem_Manager& m) :
+    FlowGraph(INST_LIST_NODE_ALLOCATOR& alloc, G4_Kernel* kernel, G4_Kernel* prevKernel, Mem_Manager& m) :
       traversalNum(0), numBBId(0), reducible(true),
       doIPA(false), hasStackCalls(false), isStackCallFunc(false), autoLabelId(0),
-      pKernel(kernel), mem(m), instListAlloc(alloc),
+      pKernel(kernel), m_prevKernel(prevKernel), mem(m), instListAlloc(alloc),
       kernelInfo(NULL), builder(NULL), globalOpndHT(m), framePtrDcl(NULL),
       stackPtrDcl(NULL), scratchRegDcl(NULL), pseudoVCEDcl(NULL) {}
 
@@ -1461,15 +1468,15 @@ class G4_Kernel
     unsigned char minor_version;
 
     G4_Kernel(INST_LIST_NODE_ALLOCATOR& alloc,
-              Mem_Manager &m, Options *options, unsigned char major, unsigned char minor)
+              Mem_Manager &m, Options *options, unsigned char major, unsigned char minor, G4_Kernel *prevKernel)
               : m_options(options), RAType(RA_Type::UNKNOWN_RA),
               asmInstCount(0), kernelID(0), tokenInstructionCount(0), tokenReuseCount(0),
               AWTokenReuseCount(0), ARTokenReuseCount(0), AATokenReuseCount(0),
               mathInstCount(0), syncInstCount(0),mathReuseCount(0),
               ARSyncInstCount(0), AWSyncInstCount(0), ARSyncAllCount(0), AWSyncAllCount(0),
               prunedDepEdges(0), prunedGlobalEdgeNum(0), prunedDiffBBEdgeNum(0), prunedDiffBBSameTokenEdgeNum(0),
               bank_good_num(0), bank_ok_num(0),
-              bank_bad_num(0), fg(alloc, this, m), major_version(major), minor_version(minor)
+              bank_bad_num(0), fg(alloc, this, prevKernel, m), major_version(major), minor_version(minor)
     {
         ASSERT_USER(
             major < COMMON_ISA_MAJOR_VER ||

visa/VISAKernel.h

@@ -97,6 +97,7 @@ class VISAKernelImpl : public VISAFunction
         m_forward_label_count = 0;
         m_jitInfo = NULL;
         errorMessage[0] = '\0';
+        m_prevKernel = NULL;
         m_kernel = NULL;
         m_builder = NULL;
         m_globalMem = NULL;
@@ -134,7 +135,7 @@ class VISAKernelImpl : public VISAFunction
     virtual ~VISAKernelImpl();
 
     void *operator new(size_t sz, vISA::Mem_Manager& m){ return m.alloc(sz); };
-    int InitializeKernel(const char *kernel_name);
+    int InitializeKernel(const char *kernel_name, VISAKernel* prevKernel = nullptr);
     int CISABuildPreDefinedDecls();
     void setVersion(unsigned char major_ver, unsigned char minor_ver){
         m_major_version = major_ver;
@@ -838,6 +839,8 @@ class VISAKernelImpl : public VISAFunction
 
     void computeAndEmitDebugInfo(std::list<VISAKernelImpl*>& functions);
 
+    vISA::G4_Kernel* GetVISAPrevKernel() { return m_prevKernel; }
+
 private:
     void setDefaultVariableName(Common_ISA_Var_Class Ty, const char *&varName);
     void dumpDebugFormatFile(std::vector<vISA::DebugInfoFormat>& debugSymbols, std::string filename);
@@ -964,6 +967,7 @@ class VISAKernelImpl : public VISAFunction
     char errorMessage[MAX_ERROR_MSG_LEN];
 
     VISA_BUILDER_OPTION mBuildOption;
+    vISA::G4_Kernel* m_prevKernel;
     vISA::G4_Kernel* m_kernel;
     CISA_IR_Builder* m_CISABuilder;
     vISA::IR_Builder* m_builder;