Use getThreadGroupSize() in more places.

Have COpenCLKernel use CComputeShaderBase.

Author: scottp101

IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_list.h

@@ -32,7 +32,7 @@ Abstract:   Contains common patch structure definitions
 
 namespace iOpenCL
 {
-const uint32_t CURRENT_ICBE_VERSION = 1070;
+const uint32_t CURRENT_ICBE_VERSION = 1071;
 const uint32_t MAGIC_CL = 0x494E5443;      // 'I', 'N', 'T', 'C'
 const uint32_t INVALID_INDEX = 0xFFFFFFFF;
 

IGC/Compiler/CISACodeGen/ComputeShaderBase.cpp

@@ -51,23 +51,24 @@ namespace IGC
         uint threadGroupSize_Z)
     {
         const CodeGenContext* pCtx = GetContext();
+        const ModuleMetaData* MMD = pCtx->getModuleMetaData();
 
-        if (pCtx->getModuleMetaData()->csInfo.neededThreadIdLayout == ThreadIDLayout::QuadTile)
+        if (MMD->csInfo.neededThreadIdLayout == ThreadIDLayout::QuadTile)
         {
             m_ThreadIDLayout = ThreadIDLayout::QuadTile;
             return;
         }
 
         if ((numberOfTypedAccess >= numberOfUntypedAccess) &&
             threadGroupSize_Y % 4 == 0 &&
-            !pCtx->getModuleMetaData()->csInfo.disableLocalIdOrderOptimizations &&
+            !MMD->csInfo.disableLocalIdOrderOptimizations &&
             IGC_IS_FLAG_ENABLED(UseTiledCSThreadOrder)) {
             m_ThreadIDLayout = ThreadIDLayout::TileY;
             m_walkOrder = WO_YXZ;
         }
 
         bool needsLinearWalk =
-            pCtx->getModuleMetaData()->csInfo.neededThreadIdLayout == ThreadIDLayout::X;
+            MMD->csInfo.neededThreadIdLayout == ThreadIDLayout::X;
         if (needsLinearWalk)
         {
             m_ThreadIDLayout = ThreadIDLayout::X;

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -13647,14 +13647,9 @@ void EmitPass::emitThreadGroupBarrier(llvm::Instruction* inst)
     else if (m_currShader->GetShaderType() == ShaderType::OPENCL_SHADER) {
         Function* F = inst->getParent()->getParent();
         MetaDataUtils* pMdUtils = getAnalysis<MetaDataUtilsWrapper>().getMetaDataUtils();
-        FunctionInfoMetaDataHandle funcInfoMD = pMdUtils->getFunctionsInfoItem(F);
-        ThreadGroupSizeMetaDataHandle threadGroupSize = funcInfoMD->getThreadGroupSize();
-        if (threadGroupSize->hasValue())
-        {
-            int32_t sz = threadGroupSize->getXDim() * threadGroupSize->getYDim() * threadGroupSize->getZDim();
-            if (sz <= (int32_t)numLanes(m_SimdMode)) {
-                skipBarrierInstructionInCS = true;
-            }
+        uint32_t sz = IGCMetaDataHelper::getThreadGroupSize(*pMdUtils, F);
+        if (sz != 0 && sz <= numLanes(m_SimdMode)) {
+            skipBarrierInstructionInCS = true;
         }
     }
 

IGC/Compiler/CISACodeGen/FoldKnownWorkGroupSizes.cpp

@@ -95,27 +95,16 @@ void FoldKnownWorkGroupSizes::visitCallInst(llvm::CallInst& I)
     }
     else if (funcName.equals(WIFuncsAnalysis::GET_ENQUEUED_LOCAL_SIZE))
     {
-        auto itr = ctx->getMetaDataUtils()->findFunctionsInfoItem(I.getFunction());
+        auto Dims = IGCMetaDataHelper::getThreadGroupDims(
+            *ctx->getMetaDataUtils(),
+            I.getFunction());
 
-        //Check function exists in the metadata
-        if (itr == ctx->getMetaDataUtils()->end_FunctionsInfo())
-            return;
-
-        FunctionInfoMetaDataHandle funcMDHandle = itr->second;
-        ThreadGroupSizeMetaDataHandle tgMD = funcMDHandle->getThreadGroupSize();
-        //Check threadGroup has value
-        if (!tgMD->hasValue())
+        if (!Dims)
             return;
 
         IRBuilder<> IRB(&I);
 
-        uint32_t Dims[] =
-        {
-            (uint32_t)tgMD->getXDim(),
-            (uint32_t)tgMD->getYDim(),
-            (uint32_t)tgMD->getZDim(),
-        };
-        auto* CV = ConstantDataVector::get(I.getContext(), Dims);
+        auto* CV = ConstantDataVector::get(I.getContext(), *Dims);
 
         auto* Dim = I.getArgOperand(0);
         auto* EE = IRB.CreateExtractElement(CV, Dim, "enqueuedLocalSize");

IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp

@@ -60,7 +60,7 @@ namespace IGC
 {
 
     COpenCLKernel::COpenCLKernel(const OpenCLProgramContext* ctx, Function* pFunc, CShaderProgram* pProgram) :
-        CShader(pFunc, pProgram)
+        CComputeShaderBase(pFunc, pProgram)
     {
         m_HasTID = false;
         m_HasGlobalSize = false;
@@ -140,6 +140,21 @@ namespace IGC
         }
     }
 
+    bool COpenCLKernel::hasWorkGroupWalkOrder()
+    {
+        const CodeGenContext* pCtx = GetContext();
+        const ModuleMetaData* MMD = pCtx->getModuleMetaData();
+        if (auto I = MMD->FuncMD.find(entry); I != MMD->FuncMD.end())
+        {
+            auto& FMD = I->second;
+            auto& Order = FMD.workGroupWalkOrder;
+            if (Order.dim0 != 0 || Order.dim1 != 0 || Order.dim2 != 0)
+                return true;
+        }
+
+        return false;
+    }
+
     SOpenCLKernelInfo::SResourceInfo COpenCLKernel::getResourceInfo(int argNo)
     {
         CodeGenContext* pCtx = GetContext();

IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.hpp

@@ -23,7 +23,7 @@ IN THE SOFTWARE.
 ============================= end_copyright_notice ===========================*/
 
 #pragma once
-#include "Compiler/CISACodeGen/ShaderCodeGen.hpp"
+#include "Compiler/CISACodeGen/ComputeShaderBase.hpp"
 
 namespace IGC
 {
@@ -33,7 +33,7 @@ namespace IGC
 namespace IGC
 {
 
-    class COpenCLKernel : public CShader
+    class COpenCLKernel : public CComputeShaderBase
     {
     public:
         friend class CShaderProgram;
@@ -129,6 +129,8 @@ namespace IGC
         OpenCLProgramContext* m_Context;
 
         void ClearKernelInfo();
+    private:
+        bool hasWorkGroupWalkOrder();
     };
 
 }

IGC/Compiler/MetaDataApi/IGCMetaDataHelper.cpp

@@ -78,6 +78,42 @@ void IGCMetaDataHelper::removeFunction(
     }
 }
 
+llvm::Optional<std::array<uint32_t, 3>>
+IGCMetaDataHelper::getThreadGroupDims(
+    MetaDataUtils& mdUtils,
+    llvm::Function* pKernelFunc)
+{
+    auto finfo = mdUtils.findFunctionsInfoItem(pKernelFunc);
+    if (finfo == mdUtils.end_FunctionsInfo())
+        return llvm::None;
+
+    auto& FI = finfo->second;
+
+    if (!FI->getThreadGroupSize()->hasValue())
+        return llvm::None;
+
+    auto Dims = FI->getThreadGroupSize();
+
+    std::array<uint32_t, 3> A {
+        (uint32_t)Dims->getXDim(),
+        (uint32_t)Dims->getYDim(),
+        (uint32_t)Dims->getZDim()
+    };
+
+    return A;
+}
+
+uint32_t IGCMetaDataHelper::getThreadGroupSize(MetaDataUtils& mdUtils, llvm::Function* pKernelFunc)
+{
+    auto Dims = IGCMD::IGCMetaDataHelper::getThreadGroupDims(mdUtils, pKernelFunc);
+    if (!Dims)
+        return 0;
+
+    auto& Vals = *Dims;
+
+    return Vals[0] * Vals[1] * Vals[2];
+}
+
 uint32_t IGCMetaDataHelper::getThreadGroupSizeHint(MetaDataUtils& mdUtils, llvm::Function* pKernelFunc)
 {
     FunctionInfoMetaDataHandle finfo = mdUtils.getFunctionsInfoItem(pKernelFunc);
@@ -89,17 +125,4 @@ uint32_t IGCMetaDataHelper::getThreadGroupSizeHint(MetaDataUtils& mdUtils, llvm:
             finfo->getThreadGroupSizeHint()->getZDim();
     }
     return size;
-}
-
-uint32_t IGCMetaDataHelper::getThreadGroupSize(MetaDataUtils& mdUtils, llvm::Function* pKernelFunc)
-{
-    FunctionInfoMetaDataHandle finfo = mdUtils.getFunctionsInfoItem(pKernelFunc);
-    uint32_t size = 0;
-    if (finfo->getThreadGroupSize()->hasValue())
-    {
-        size = finfo->getThreadGroupSize()->getXDim() *
-            finfo->getThreadGroupSize()->getYDim() *
-            finfo->getThreadGroupSize()->getZDim();
-    }
-    return size;
 }

IGC/Compiler/MetaDataApi/IGCMetaDataHelper.h

@@ -41,6 +41,8 @@ namespace IGC {
 
             // In OCL, thread group size (hint) is given by kernel attributes reqd_work_group_size and work_group_size_hint.
             // Return thread group size (hint) if present; return 0 otherwise.
+            static llvm::Optional<std::array<uint32_t, 3>>
+                getThreadGroupDims(MetaDataUtils& mdUtils, llvm::Function* pKernelFunc);
             static uint32_t getThreadGroupSize(MetaDataUtils& mdUtils, llvm::Function* pKernelFunc);
             static uint32_t getThreadGroupSizeHint(MetaDataUtils& mdUtils, llvm::Function* pKernelFunc);
         };

IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncResolution.cpp

@@ -56,23 +56,11 @@ WIFuncResolution::WIFuncResolution() : FunctionPass(ID), m_implicitArgs()
 Constant* WIFuncResolution::getKnownWorkGroupSize(
     IGCMD::MetaDataUtils* MDUtils, llvm::Function& F) const
 {
-    auto finfo = MDUtils->findFunctionsInfoItem(&F);
-    if (finfo == MDUtils->end_FunctionsInfo())
+    auto Dims = IGCMD::IGCMetaDataHelper::getThreadGroupDims(*MDUtils, &F);
+    if (!Dims)
         return nullptr;
 
-    auto& FI = finfo->second;
-    if (FI->getThreadGroupSize()->hasValue())
-    {
-        uint32_t Dims[] =
-        {
-            (uint32_t)FI->getThreadGroupSize()->getXDim(),
-            (uint32_t)FI->getThreadGroupSize()->getYDim(),
-            (uint32_t)FI->getThreadGroupSize()->getZDim(),
-        };
-        return ConstantDataVector::get(F.getContext(), Dims);
-    }
-
-    return nullptr;
+    return ConstantDataVector::get(F.getContext(), *Dims);
 }
 
 bool WIFuncResolution::runOnFunction(Function& F)

IGC/common/MDFrameWork.h

@@ -175,9 +175,9 @@ namespace IGC
 
     struct WorkGroupWalkOrderMD
     {
-    int dim0 = 0;
-    int dim1 = 0;
-    int dim2 = 0;
+        int dim0 = 0;
+        int dim1 = 0;
+        int dim2 = 0;
     };
 
     struct FuncArgMD

visa/BinaryEncodingIGA.cpp

@@ -853,7 +853,7 @@ void BinaryEncodingIGA::Encode()
         auto secondBB = *(std::next(kernel.fg.begin()));
         auto iter = std::find_if(secondBB->begin(), secondBB->end(),
             [](G4_INST* inst) { return !inst->isLabel();});
-        assert(iter != secondBB->end() && "execpt at least one non-label inst in second BB");
+        assert(iter != secondBB->end() && "expect at least one non-label inst in second BB");
         kernel.fg.builder->getJitInfo()->offsetToSkipPerThreadDataLoad =
             (uint32_t)(*iter)->getGenOffset();
     }