[Autobackout][FuncReg]Revert of change: 4c0d9b1

scottp101 · igcbot · commit 49ea47e7e569 · 2021-03-26T19:11:49.000+01:00
Use getThreadGroupSize() in more places.
Have COpenCLKernel use CComputeShaderBase.
diff --git a/IGC/AdaptorOCL/ocl_igc_shared/executable_format/patch_list.h b/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 = 1071;
+const uint32_t CURRENT_ICBE_VERSION = 1070;
 const uint32_t MAGIC_CL = 0x494E5443;      // 'I', 'N', 'T', 'C'
 const uint32_t INVALID_INDEX = 0xFFFFFFFF;
 
diff --git a/IGC/Compiler/CISACodeGen/ComputeShaderBase.cpp b/IGC/Compiler/CISACodeGen/ComputeShaderBase.cpp
@@ -51,24 +51,23 @@ namespace IGC
         uint threadGroupSize_Z)
     {
         const CodeGenContext* pCtx = GetContext();
-        const ModuleMetaData* MMD = pCtx->getModuleMetaData();
 
-        if (MMD->csInfo.neededThreadIdLayout == ThreadIDLayout::QuadTile)
+        if (pCtx->getModuleMetaData()->csInfo.neededThreadIdLayout == ThreadIDLayout::QuadTile)
         {
             m_ThreadIDLayout = ThreadIDLayout::QuadTile;
             return;
         }
 
         if ((numberOfTypedAccess >= numberOfUntypedAccess) &&
             threadGroupSize_Y % 4 == 0 &&
-            !MMD->csInfo.disableLocalIdOrderOptimizations &&
+            !pCtx->getModuleMetaData()->csInfo.disableLocalIdOrderOptimizations &&
             IGC_IS_FLAG_ENABLED(UseTiledCSThreadOrder)) {
             m_ThreadIDLayout = ThreadIDLayout::TileY;
             m_walkOrder = WO_YXZ;
         }
 
         bool needsLinearWalk =
-            MMD->csInfo.neededThreadIdLayout == ThreadIDLayout::X;
+            pCtx->getModuleMetaData()->csInfo.neededThreadIdLayout == ThreadIDLayout::X;
         if (needsLinearWalk)
         {
             m_ThreadIDLayout = ThreadIDLayout::X;
diff --git a/IGC/Compiler/CISACodeGen/EmitVISAPass.cpp b/IGC/Compiler/CISACodeGen/EmitVISAPass.cpp
@@ -13647,9 +13647,14 @@ void EmitPass::emitThreadGroupBarrier(llvm::Instruction* inst)
     else if (m_currShader->GetShaderType() == ShaderType::OPENCL_SHADER) {
         Function* F = inst->getParent()->getParent();
         MetaDataUtils* pMdUtils = getAnalysis<MetaDataUtilsWrapper>().getMetaDataUtils();
-        uint32_t sz = IGCMetaDataHelper::getThreadGroupSize(*pMdUtils, F);
-        if (sz != 0 && sz <= numLanes(m_SimdMode)) {
-            skipBarrierInstructionInCS = true;
+        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;
+            }
         }
     }
 
diff --git a/IGC/Compiler/CISACodeGen/FoldKnownWorkGroupSizes.cpp b/IGC/Compiler/CISACodeGen/FoldKnownWorkGroupSizes.cpp
@@ -95,16 +95,27 @@ void FoldKnownWorkGroupSizes::visitCallInst(llvm::CallInst& I)
     }
     else if (funcName.equals(WIFuncsAnalysis::GET_ENQUEUED_LOCAL_SIZE))
     {
-        auto Dims = IGCMetaDataHelper::getThreadGroupDims(
-            *ctx->getMetaDataUtils(),
-            I.getFunction());
+        auto itr = ctx->getMetaDataUtils()->findFunctionsInfoItem(I.getFunction());
 
-        if (!Dims)
+        //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())
             return;
 
         IRBuilder<> IRB(&I);
 
-        auto* CV = ConstantDataVector::get(I.getContext(), *Dims);
+        uint32_t Dims[] =
+        {
+            (uint32_t)tgMD->getXDim(),
+            (uint32_t)tgMD->getYDim(),
+            (uint32_t)tgMD->getZDim(),
+        };
+        auto* CV = ConstantDataVector::get(I.getContext(), Dims);
 
         auto* Dim = I.getArgOperand(0);
         auto* EE = IRB.CreateExtractElement(CV, Dim, "enqueuedLocalSize");
diff --git a/IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp b/IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.cpp
@@ -60,7 +60,7 @@ namespace IGC
 {
 
     COpenCLKernel::COpenCLKernel(const OpenCLProgramContext* ctx, Function* pFunc, CShaderProgram* pProgram) :
-        CComputeShaderBase(pFunc, pProgram)
+        CShader(pFunc, pProgram)
     {
         m_HasTID = false;
         m_HasGlobalSize = false;
@@ -140,21 +140,6 @@ 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();
diff --git a/IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.hpp b/IGC/Compiler/CISACodeGen/OpenCLKernelCodeGen.hpp
@@ -23,7 +23,7 @@ IN THE SOFTWARE.
 ============================= end_copyright_notice ===========================*/
 
 #pragma once
-#include "Compiler/CISACodeGen/ComputeShaderBase.hpp"
+#include "Compiler/CISACodeGen/ShaderCodeGen.hpp"
 
 namespace IGC
 {
@@ -33,7 +33,7 @@ namespace IGC
 namespace IGC
 {
 
-    class COpenCLKernel : public CComputeShaderBase
+    class COpenCLKernel : public CShader
     {
     public:
         friend class CShaderProgram;
@@ -129,8 +129,6 @@ namespace IGC
         OpenCLProgramContext* m_Context;
 
         void ClearKernelInfo();
-    private:
-        bool hasWorkGroupWalkOrder();
     };
 
 }
diff --git a/IGC/Compiler/MetaDataApi/IGCMetaDataHelper.cpp b/IGC/Compiler/MetaDataApi/IGCMetaDataHelper.cpp
@@ -78,42 +78,6 @@ 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);
@@ -125,4 +89,17 @@ 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;
 }
diff --git a/IGC/Compiler/MetaDataApi/IGCMetaDataHelper.h b/IGC/Compiler/MetaDataApi/IGCMetaDataHelper.h
@@ -41,8 +41,6 @@ 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);
         };
diff --git a/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncResolution.cpp b/IGC/Compiler/Optimizer/OpenCLPasses/WIFuncs/WIFuncResolution.cpp
@@ -56,11 +56,23 @@ WIFuncResolution::WIFuncResolution() : FunctionPass(ID), m_implicitArgs()
 Constant* WIFuncResolution::getKnownWorkGroupSize(
     IGCMD::MetaDataUtils* MDUtils, llvm::Function& F) const
 {
-    auto Dims = IGCMD::IGCMetaDataHelper::getThreadGroupDims(*MDUtils, &F);
-    if (!Dims)
+    auto finfo = MDUtils->findFunctionsInfoItem(&F);
+    if (finfo == MDUtils->end_FunctionsInfo())
         return nullptr;
 
-    return ConstantDataVector::get(F.getContext(), *Dims);
+    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;
 }
 
 bool WIFuncResolution::runOnFunction(Function& F)
diff --git a/IGC/common/MDFrameWork.h b/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
diff --git a/visa/BinaryEncodingIGA.cpp b/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() && "expect at least one non-label inst in second BB");
+        assert(iter != secondBB->end() && "execpt at least one non-label inst in second BB");
         kernel.fg.builder->getJitInfo()->offsetToSkipPerThreadDataLoad =
             (uint32_t)(*iter)->getGenOffset();
     }

Original file line number	Diff line number	Diff line change
`@@ -32,7 +32,7 @@ Abstract: Contains common patch structure definitions`
`32`	`32`
`33`	`33`	`namespace iOpenCL`
`34`	`34`	`{`
`35`		`-const uint32_t CURRENT_ICBE_VERSION = 1071;`
	`35`	`+const uint32_t CURRENT_ICBE_VERSION = 1070;`
`36`	`36`	`const uint32_t MAGIC_CL = 0x494E5443; // 'I', 'N', 'T', 'C'`
`37`	`37`	`const uint32_t INVALID_INDEX = 0xFFFFFFFF;`
`38`	`38`
Original file line number	Diff line number	Diff line change
`@@ -51,24 +51,23 @@ namespace IGC`
`51`	`51`	`uint threadGroupSize_Z)`
`52`	`52`	`{`
`53`	`53`	`const CodeGenContext* pCtx = GetContext();`
`54`		`- const ModuleMetaData* MMD = pCtx->getModuleMetaData();`
`55`	`54`
`56`		`- if (MMD->csInfo.neededThreadIdLayout == ThreadIDLayout::QuadTile)`
	`55`	`+ if (pCtx->getModuleMetaData()->csInfo.neededThreadIdLayout == ThreadIDLayout::QuadTile)`
`57`	`56`	`{`
`58`	`57`	`m_ThreadIDLayout = ThreadIDLayout::QuadTile;`
`59`	`58`	`return;`
`60`	`59`	`}`
`61`	`60`
`62`	`61`	`if ((numberOfTypedAccess >= numberOfUntypedAccess) &&`
`63`	`62`	`threadGroupSize_Y % 4 == 0 &&`
`64`		`- !MMD->csInfo.disableLocalIdOrderOptimizations &&`
	`63`	`+ !pCtx->getModuleMetaData()->csInfo.disableLocalIdOrderOptimizations &&`
`65`	`64`	`IGC_IS_FLAG_ENABLED(UseTiledCSThreadOrder)) {`
`66`	`65`	`m_ThreadIDLayout = ThreadIDLayout::TileY;`
`67`	`66`	`m_walkOrder = WO_YXZ;`
`68`	`67`	`}`
`69`	`68`
`70`	`69`	`bool needsLinearWalk =`
`71`		`- MMD->csInfo.neededThreadIdLayout == ThreadIDLayout::X;`
	`70`	`+ pCtx->getModuleMetaData()->csInfo.neededThreadIdLayout == ThreadIDLayout::X;`
`72`	`71`	`if (needsLinearWalk)`
`73`	`72`	`{`
`74`	`73`	`m_ThreadIDLayout = ThreadIDLayout::X;`
Original file line number	Diff line number	Diff line change
`@@ -60,7 +60,7 @@ namespace IGC`
`60`	`60`	`{`
`61`	`61`
`62`	`62`	`COpenCLKernel::COpenCLKernel(const OpenCLProgramContext* ctx, Function* pFunc, CShaderProgram* pProgram) :`
`63`		`- CComputeShaderBase(pFunc, pProgram)`
	`63`	`+ CShader(pFunc, pProgram)`
`64`	`64`	`{`
`65`	`65`	`m_HasTID = false;`
`66`	`66`	`m_HasGlobalSize = false;`
`@@ -140,21 +140,6 @@ namespace IGC`
`140`	`140`	`}`
`141`	`141`	`}`
`142`	`142`
`143`		`- bool COpenCLKernel::hasWorkGroupWalkOrder()`
`144`		`- {`
`145`		`- const CodeGenContext* pCtx = GetContext();`
`146`		`- const ModuleMetaData* MMD = pCtx->getModuleMetaData();`
`147`		`- if (auto I = MMD->FuncMD.find(entry); I != MMD->FuncMD.end())`
`148`		`- {`
`149`		`- auto& FMD = I->second;`
`150`		`- auto& Order = FMD.workGroupWalkOrder;`
`151`		`- if (Order.dim0 != 0 \|\| Order.dim1 != 0 \|\| Order.dim2 != 0)`
`152`		`- return true;`
`153`		`- }`
`154`		`-`
`155`		`- return false;`
`156`		`- }`
`157`		`-`
`158`	`143`	`SOpenCLKernelInfo::SResourceInfo COpenCLKernel::getResourceInfo(int argNo)`
`159`	`144`	`{`
`160`	`145`	`CodeGenContext* pCtx = GetContext();`
Original file line number	Diff line number	Diff line change
`@@ -23,7 +23,7 @@ IN THE SOFTWARE.`
`23`	`23`	`============================= end_copyright_notice ===========================*/`
`24`	`24`
`25`	`25`	`#pragma once`
`26`		`-#include "Compiler/CISACodeGen/ComputeShaderBase.hpp"`
	`26`	`+#include "Compiler/CISACodeGen/ShaderCodeGen.hpp"`
`27`	`27`
`28`	`28`	`namespace IGC`
`29`	`29`	`{`
`@@ -33,7 +33,7 @@ namespace IGC`
`33`	`33`	`namespace IGC`
`34`	`34`	`{`
`35`	`35`
`36`		`- class COpenCLKernel : public CComputeShaderBase`
	`36`	`+ class COpenCLKernel : public CShader`
`37`	`37`	`{`
`38`	`38`	`public:`
`39`	`39`	`friend class CShaderProgram;`
`@@ -129,8 +129,6 @@ namespace IGC`
`129`	`129`	`OpenCLProgramContext* m_Context;`
`130`	`130`
`131`	`131`	`void ClearKernelInfo();`
`132`		`- private:`
`133`		`- bool hasWorkGroupWalkOrder();`
`134`	`132`	`};`
`135`	`133`
`136`	`134`	`}`