move compute shader state processing

move compute shader state processing

Author: scottp101

IGC/AdaptorOCL/OCL/sp/spp_g8.cpp

@@ -374,7 +374,7 @@ void dumpOCLCos(const IGC::COpenCLKernel *Kernel, const std::string &stateDebugM
     dumpName = dumpName.PostFix(kernelName);
 
     dumpName = dumpName.DispatchMode(Kernel->m_ShaderDispatchMode);
-    dumpName = dumpName.SIMDSize(Kernel->m_dispatchSize).Retry(context->m_retryManager.GetRetryId()).Extension("cos");
+    dumpName = dumpName.SIMDSize(Kernel->m_State.m_dispatchSize).Retry(context->m_retryManager.GetRetryId()).Extension("cos");
 
     auto dump = IGC::Debug::Dump(dumpName, IGC::Debug::DumpType::COS_TEXT);
 

IGC/Compiler/CISACodeGen/CISABuilder.cpp

@@ -35,7 +35,6 @@ CShader::CShader(Function *pFunc, CShaderProgram *pProgram, GenericShaderState &
     , entry(pFunc)
     , m_parent(pProgram)
     , encoder()
-    , m_BarrierNumber(0)
 {
     m_ctx = m_parent->GetContext();
     m_WI = nullptr;
@@ -53,7 +52,6 @@ CShader::CShader(Function *pFunc, CShaderProgram *pProgram, GenericShaderState &
     // set this to ture if there is any stateless access.
     m_HasGlobalStatelessMemoryAccess = false;
     m_HasConstantStatelessMemoryAccess = false;
-    m_HasDPAS = false;
 
     m_SavedSRetPtr = nullptr;
     m_FP = nullptr;
@@ -110,7 +108,7 @@ void CShader::InitEncoder(SIMDMode simdSize, bool canAbortOnSpill, ShaderDispatc
         m_SIMDSize = simdSize;
         m_numberInstance = 1;
     }
-    m_dispatchSize = simdSize;
+    m_State.m_dispatchSize = simdSize;
     globalSymbolMapping.clear();
     symbolMapping.clear();
     ccTupleMapping.clear();
@@ -138,7 +136,7 @@ void CShader::PreAnalysisPass()
                 const uint32_t GRFSize = getGRFSize();
                 IGC_ASSERT(0 < GRFSize);
 
-                m_ScratchSpaceSize = funcMDItr->second.privateMemoryPerWI * numLanes(m_dispatchSize);
+                m_ScratchSpaceSize = funcMDItr->second.privateMemoryPerWI * numLanes(m_State.m_dispatchSize);
                 m_ScratchSpaceSize = std::max(m_ScratchSpaceSize, m_ctx->getIntelScratchSpacePrivateMemoryMinimalSizePerThread());
 
                 // Round up to GRF-byte aligned.
@@ -1141,7 +1139,7 @@ CVariable* CShader::ImmToVariable(uint64_t immediate, VISA_Type type, bool isCod
         // src-variable is no longer a boolean, V-ISA cannot take boolean-src immed.
 
         CVariable* dst = GetNewVariable(
-            numLanes(m_dispatchSize), ISA_TYPE_BOOL, EALIGN_BYTE, CName::NONE);
+            numLanes(m_State.m_dispatchSize), ISA_TYPE_BOOL, EALIGN_BYTE, CName::NONE);
         // FIXME: We need to pop/push the encoder context
         //encoder.save();
         if (isCodePatchCandidate)
@@ -1550,7 +1548,7 @@ void CShader::GetSimdOffsetBase(CVariable*& pVar, bool dup)
     encoder.Cast(pVar, ImmToVariable(0x76543210, ISA_TYPE_V));
     encoder.Push();
 
-    if (m_dispatchSize >= SIMDMode::SIMD16)
+    if (m_State.m_dispatchSize >= SIMDMode::SIMD16)
     {
         encoder.SetSimdSize(SIMDMode::SIMD8);
         encoder.SetDstSubReg(8);
@@ -3043,7 +3041,7 @@ unsigned int CShader::EvaluateSIMDConstExpr(Value* C)
     {
         if (genInst->getIntrinsicID() == GenISAIntrinsic::GenISA_simdSize)
         {
-            return numLanes(m_dispatchSize);
+            return numLanes(m_State.m_dispatchSize);
 
         }
     }

IGC/Compiler/CISACodeGen/CShader.cpp

@@ -23,10 +23,10 @@ using namespace llvm;
 
 namespace IGC
 {
-CComputeShaderBase::CComputeShaderBase(llvm::Function *pFunc,
-                                       CShaderProgram *pProgram)
-    : m_State(*pFunc, *pProgram->GetContext()),
-      CShader(pFunc, pProgram, m_State) {}
+    CComputeShaderBase::CComputeShaderBase(Function* pFunc,
+        CShaderProgram* pProgram,
+        GenericShaderState& GState)
+    : CShader(pFunc, pProgram, GState) {}
 
     CComputeShaderBase::~CComputeShaderBase() {}
 

IGC/Compiler/CISACodeGen/ComputeShaderBase.cpp

@@ -19,10 +19,9 @@ namespace IGC
     class CComputeShaderBase : public CShader
     {
     public:
-        CComputeShaderBase(llvm::Function* pFunc, CShaderProgram* pProgram);
+        CComputeShaderBase(
+            llvm::Function* pFunc, CShaderProgram* pProgram, GenericShaderState& GState);
         virtual ~CComputeShaderBase();
-
-        GenericShaderState m_State;
     protected:
         // Determines if HW can handle auto generating local IDs with this
         // order

IGC/Compiler/CISACodeGen/ComputeShaderBase.hpp

@@ -388,7 +388,7 @@ void DebugInfoData::markVariableAsOutput(CShader *pShader, CVariable *pVariable)
 
     // Mark variable with "Output", to extend it's living time will be extended to the end of the function.
     pShader->GetEncoder().GetVISAKernel()->AddAttributeToVar(pVariable->visaGenVariable[0], "Output", 0, nullptr);
-    if (pShader->m_dispatchSize == SIMDMode::SIMD32 && pVariable->visaGenVariable[1])
+    if (pShader->m_State.m_dispatchSize == SIMDMode::SIMD32 && pVariable->visaGenVariable[1])
     {
         pShader->GetEncoder().GetVISAKernel()->AddAttributeToVar(pVariable->visaGenVariable[1], "Output", 0, nullptr);
     }
@@ -459,7 +459,7 @@ void DebugInfoData::transferMappings(const llvm::Function& F)
             {
                 unsigned int lower16Channels = (unsigned int)m_pShader->GetEncoder().GetVISAKernel()->getDeclarationID(CVar->visaGenVariable[0]);
                 unsigned int higher16Channels = 0;
-                if (numLanes(m_pShader->m_dispatchSize) == 32 && !CVar->IsUniform() && CVar->visaGenVariable[1])
+                if (numLanes(m_pShader->m_State.m_dispatchSize) == 32 && !CVar->IsUniform() && CVar->visaGenVariable[1])
                     higher16Channels = m_pShader->GetEncoder().GetVISAKernel()->getDeclarationID(CVar->visaGenVariable[1]);
                 CVarToVISADclId[CVar] = std::make_pair(lower16Channels, higher16Channels);
             }