Refactor SPV_INTEL_device_side_avc_motion_estimation builtins

This is a preparation for SPIRV Translator switchover to vanilla version.
Vanilla translator produces slightly different form of SPIRV builtins
which is described here:
https://github.com/KhronosGroup/SPIRV-LLVM-Translator/blob/master/docs/SPIRVRepresentationInLLVM.rst

The plan is to support both versions until translator switchover is
done. Once vanilla translator is enabled, old version of SPIRV builtins
will get cleaned up.

Author: aratajew

IGC/AdaptorOCL/PreprocessSPVIR.cpp

@@ -41,67 +41,6 @@ PreprocessSPVIR::PreprocessSPVIR() : ModulePass(ID)
     initializePreprocessSPVIRPass(*PassRegistry::getPassRegistry());
 }
 
-uint64_t PreprocessSPVIR::parseSampledImageTy(StructType* sampledImageTy)
-{
-    std::string Name = sampledImageTy->getName().str();
-    std::smatch match;
-    std::regex reg("spirv.SampledImage._void_([0-6])_([0-2])_([0-1])_([0-1])_([0-2])_([0-9]+)_([0-2])");
-
-    if (std::regex_match(Name, match, reg))
-    {
-        IGC_ASSERT(match.size() == 8);
-        uint64_t ImageType = 0;
-
-        uint64_t Dim = stoi(match[1]);
-        ImageType |= Dim << 59;
-        uint64_t Depth = stoi(match[2]);
-        ImageType |= Depth << 58;
-        uint64_t Arrayed = stoi(match[3]);
-        ImageType |= Arrayed << 57;
-        uint64_t MS = stoi(match[4]);
-        ImageType |= MS << 56;
-        uint64_t Sampled = stoi(match[5]);
-        ImageType |= Sampled << 62;
-        uint64_t Access = stoi(match[6]);
-        ImageType |= Access << 54;
-
-        return ImageType;
-    }
-    else
-    {
-        IGC_ASSERT_MESSAGE(0, "SampledImage opaque type is incorrect!");
-        return 0;
-    }
-}
-
-Value* PreprocessSPVIR::getWidenImageCoordsArg(Value* Coords)
-{
-    Type* coordsType = Coords->getType();
-    Value* newCoords = nullptr;
-    if (!isa<IGCLLVM::FixedVectorType>(coordsType))
-    {
-        Value* undef = UndefValue::get(IGCLLVM::FixedVectorType::get(coordsType, 4));
-        newCoords = m_Builder->CreateInsertElement(undef, Coords, ConstantInt::get(m_Builder->getInt32Ty(), 0));
-    }
-    else if (cast<IGCLLVM::FixedVectorType>(coordsType)->getNumElements() < 4)
-    {
-        SmallVector<Constant*, 4> shuffleIdx;
-        for (uint64_t i = 0; i < cast<IGCLLVM::FixedVectorType>(coordsType)->getNumElements(); i++)
-            shuffleIdx.push_back(ConstantInt::get(m_Builder->getInt32Ty(), i));
-
-        for (uint64_t i = cast<IGCLLVM::FixedVectorType>(coordsType)->getNumElements(); i < 4; i++)
-            shuffleIdx.push_back(ConstantInt::get(m_Builder->getInt32Ty(), 0));
-
-        newCoords = m_Builder->CreateShuffleVector(Coords, UndefValue::get(coordsType), ConstantVector::get(shuffleIdx));
-    }
-    else
-    {
-        // coordinate argument has already a desired width
-        newCoords = Coords;
-    }
-    return newCoords;
-}
-
 void PreprocessSPVIR::createCallAndReplace(CallInst& oldCallInst, StringRef newFuncName, std::vector<Value*>& args )
 {
     Function* F = oldCallInst.getCalledFunction();
@@ -120,137 +59,6 @@ void PreprocessSPVIR::createCallAndReplace(CallInst& oldCallInst, StringRef newF
     oldCallInst.replaceAllUsesWith(newCall);
 }
 
-// Transform:
-// %SampledImage = call spir_func %spirv.SampledImage._void_1_0_0_0_0_0_0 addrspace(1)* @_Z20__spirv_SampledImage14ocl_image2d_ro11ocl_sampler(
-//                      %opencl.image2d_ro_t addrspace(1)* %image, %opencl.sampler_t addrspace(2)* %sampler)
-// call spir_func <4 x float> @_Z38__spirv_ImageSampleExplicitLod_Rfloat4PU3AS140__spirv_SampledImage__void_1_0_0_0_0_0_0Dv2_fif(
-//                      %spirv.SampledImage._void_1_0_0_0_0_0_0 addrspace(1)*% SampledImage, <2 x float> %coords, i32 2, float 0.000000e+00)
-//
-// ---->
-//
-// %0 = ptrtoint %opencl.image2d_ro_t addrspace(1)* %image to i64
-// %1 = ptrtoint % opencl.sampler_t addrspace(2) * %sampler to i64
-// %2 = insertelement <3 x i64> undef, i64 %0, i32 0
-// %3 = insertelement <3 x i64> %2, i64 576460752303423488, i32 1
-// %4 = insertelement <3 x i64> %3, i64 %1, i32 2
-// %5 = shufflevector <2 x float> %coords, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 0>
-// %6 = call <4 x float> @__intel_sample_image_lod_fcoords_Rfloat4(<3 x i64> %4, <4 x float> %5, i32 2, float 0.000000e+00)
-void PreprocessSPVIR::visitImageSampleExplicitLod(CallInst& CI)
-{
-    m_Builder->SetInsertPoint(&CI);
-
-    enum {
-        SAMPLED_IMAGE_INDEX,
-        COORDS_INDEX,
-        IMAGE_OPERANDS_INDEX,
-        LOD_INDEX,
-        DX_INDEX = LOD_INDEX,
-        DY_INDEX
-    };
-
-    auto callSampledImage = cast<CallInst>(CI.getArgOperand(0));
-    auto sampledImagePtr = cast<PointerType>(callSampledImage->getType());
-    auto sampledImageTy = cast<StructType>(sampledImagePtr->getElementType());
-
-    std::vector<Value*> args;
-    for (size_t i = 0; i < CI.getNumArgOperands(); i++)
-        args.push_back(CI.getArgOperand(i));
-
-    // Khronos SPIRV-LLVM Translator represents OpTypeSampledImage as a global pointer to opaque type:
-    //  %spirv.SampledImage._{SampledType}_{Dim}_{Depth}_{Arrayed}_{MS}_{Sampled}_{Format}_{Access}.
-    // Such representation cannot be optimally handled in BiFModule itself, since it would require to temporary
-    // allocate global memory and then clear this up once image/sampler intrinsics are resolved.
-    // It's easier to transform it into <3 x i64> vector containing:
-    //   -image,
-    //   -image type which comes from parsing opaque type
-    //   -sampler.
-    Value* unifiedSampledImage = UndefValue::get(IGCLLVM::FixedVectorType::get(m_Builder->getInt64Ty(), 3));
-
-    Value* imageAsInt = m_Builder->CreatePtrToInt(callSampledImage->getArgOperand(0), m_Builder->getInt64Ty());
-    Value* samplerAsInt = m_Builder->CreatePtrToInt(callSampledImage->getArgOperand(1), m_Builder->getInt64Ty());
-
-    unifiedSampledImage = m_Builder->CreateInsertElement(unifiedSampledImage, imageAsInt, m_Builder->getInt32(0));
-    uint64_t imageType = parseSampledImageTy(sampledImageTy);
-    unifiedSampledImage = m_Builder->CreateInsertElement(unifiedSampledImage, m_Builder->getInt64(imageType), m_Builder->getInt32(1));
-    unifiedSampledImage = m_Builder->CreateInsertElement(unifiedSampledImage, samplerAsInt, m_Builder->getInt32(2));
-
-    args[SAMPLED_IMAGE_INDEX] = unifiedSampledImage;
-
-    // Widen coords argument, since __intel_sample_image always takes coordinates as 4-elements vector type
-    Value* coords = getWidenImageCoordsArg(args[COORDS_INDEX]);
-    args[COORDS_INDEX] = coords;
-
-    // Prepare a new builtin name
-    std::string unifiedImageSampleName = "__intel_sample_image";
-    auto imageOperand = cast<ConstantInt>(args[IMAGE_OPERANDS_INDEX]);
-    switch (imageOperand->getZExtValue())
-    {
-    case 2: // Lod
-    {
-        unifiedImageSampleName += "_lod";
-
-        Type* scalarCoordsType = coords->getType()->getScalarType();
-        if (scalarCoordsType->isFloatTy())
-            unifiedImageSampleName += "_fcoords";
-        else if (scalarCoordsType->isIntegerTy())
-            unifiedImageSampleName += "_icoords";
-        else
-            IGC_ASSERT_MESSAGE(0, "Unsupported coordinates type of ImageSampleExplicitLod builtin.");
-    }
-    break;
-    case 4: // Grad
-    {
-        unifiedImageSampleName += "_grad";
-        Type* dxType = args[DX_INDEX]->getType();
-        Type* dyType = args[DY_INDEX]->getType();
-        IGC_ASSERT(dxType == dyType);
-
-        Type* scalarDxType = dxType->getScalarType();
-        if (scalarDxType->isFloatTy())
-            unifiedImageSampleName += "_float";
-        else if(scalarDxType->isIntegerTy())
-            unifiedImageSampleName += "_uint";
-        else
-            IGC_ASSERT_MESSAGE(0, "Unsupported dx/dy types of ImageSampleExplicitLod builtin.");
-
-        if (auto *VT = dyn_cast<IGCLLVM::FixedVectorType>(dxType))
-        {
-            unifiedImageSampleName += std::to_string(VT->getNumElements());
-        }
-    }
-    break;
-    default:
-        IGC_ASSERT_MESSAGE(0, "Unsupported image operand!");
-        break;
-    }
-
-   Type* retType = CI.getType();
-   auto *VT = dyn_cast<IGCLLVM::FixedVectorType>(retType);
-   if (VT && VT->getNumElements() == 4)
-   {
-        Type* retScalarType = retType->getScalarType();
-        if (retScalarType->isIntegerTy())
-            unifiedImageSampleName += "_Ruint4";
-        else if (retScalarType->isFloatTy())
-            unifiedImageSampleName += "_Rfloat4";
-        else if (retScalarType->isHalfTy())
-            unifiedImageSampleName += "_Rhalf4";
-        else
-            IGC_ASSERT_MESSAGE(0, "Unsupported return type of ImageSampleExplicitLod builtin.");
-    }
-    else
-        IGC_ASSERT_MESSAGE(0, "ImageSampleExplicitLod builtin must return vector 4-elements type.");
-
-    // Create function call to __intel_sample_image_{Lod|Grad}_{i|f}coords_R{returnType}
-    createCallAndReplace(CI, unifiedImageSampleName, args);
-    CI.eraseFromParent();
-
-    if (callSampledImage->use_empty())
-        callSampledImage->eraseFromParent();
-
-    m_changed = true;
-}
-
 // Replace functions like:
 //  i32 @_Z18__spirv_ocl_printfPU3AS2c(i8 addrspace(2)*)
 //  i32 @_Z18__spirv_ocl_printfPU3AS2ci(i8 addrspace(2)*, i32)
@@ -364,11 +172,7 @@ void PreprocessSPVIR::visitCallInst(CallInst& CI)
     StringRef Name = F->getName();
     if (!isSPVIR(Name)) return;
 
-    if (Name.contains("ImageSampleExplicitLod"))
-    {
-        visitImageSampleExplicitLod(CI);
-    }
-    else if (Name.contains("printf"))
+    if (Name.contains("printf"))
     {
         visitOpenCLEISPrintf(CI);
     }

IGC/AdaptorOCL/PreprocessSPVIR.h

@@ -34,15 +34,12 @@ namespace IGC
 
         virtual bool runOnModule(llvm::Module& F) override;
         void visitCallInst(llvm::CallInst& CI);
-        void visitImageSampleExplicitLod(llvm::CallInst& CI);
         void visitOpenCLEISPrintf(llvm::CallInst& CI);
     private:
         bool hasArrayArg(llvm::Function& F);
         void processBuiltinsWithArrayArguments(llvm::Function& F);
         void processBuiltinsWithArrayArguments();
-        llvm::Value* getWidenImageCoordsArg(llvm::Value* Coords);
         void createCallAndReplace(llvm::CallInst& oldCallInst, llvm::StringRef newFuncName, std::vector<llvm::Value*>& args);
-        uint64_t parseSampledImageTy(llvm::StructType* SampledImageTy);
         bool isSPVIR(llvm::StringRef funcName);
 
         IGCLLVM::Module* m_Module = nullptr;