Implementation of vector compute extension

include/LLVMSPIRVExtensions.inc

@@ -19,3 +19,5 @@ EXT(SPV_INTEL_io_pipes)
 EXT(SPV_INTEL_inline_assembly)
 EXT(SPV_INTEL_arbitrary_precision_integers)
 EXT(SPV_INTEL_optimization_hints)
+EXT(SPV_INTEL_float_controls2)
+EXT(SPV_INTEL_vector_compute)

lib/SPIRV/CMakeLists.txt

@@ -8,6 +8,7 @@ add_llvm_library(LLVMSPIRVLib
   OCL21ToSPIRV.cpp
   OCLTypeToSPIRV.cpp
   OCLUtil.cpp
+  VectorComputeUtil.cpp
   SPIRVLowerBool.cpp
   SPIRVLowerConstExpr.cpp
   SPIRVLowerMemmove.cpp

lib/SPIRV/PreprocessMetadata.cpp

@@ -42,6 +42,7 @@
 #include "SPIRVInternal.h"
 #include "SPIRVMDBuilder.h"
 #include "SPIRVMDWalker.h"
+#include "VectorComputeUtil.h"
 
 #include "llvm/ADT/Triple.h"
 #include "llvm/IR/IRBuilder.h"
@@ -69,6 +70,8 @@ class PreprocessMetadata : public ModulePass {
   bool runOnModule(Module &M) override;
   void visit(Module *M);
   void preprocessOCLMetadata(Module *M, SPIRVMDBuilder *B, SPIRVMDWalker *W);
+  void preprocessVectorComputeMetadata(Module *M, SPIRVMDBuilder *B,
+                                       SPIRVMDWalker *W);
 
   static char ID;
 
@@ -100,6 +103,7 @@ void PreprocessMetadata::visit(Module *M) {
   SPIRVMDWalker W(*M);
 
   preprocessOCLMetadata(M, &B, &W);
+  preprocessVectorComputeMetadata(M, &B, &W);
 
   // Create metadata representing (empty so far) list
   // of OpExecutionMode instructions
@@ -243,6 +247,55 @@ void PreprocessMetadata::preprocessOCLMetadata(Module *M, SPIRVMDBuilder *B,
     B->eraseNamedMD(kSPIR2MD::FPContract);
 }
 
+void PreprocessMetadata::preprocessVectorComputeMetadata(Module *M,
+                                                         SPIRVMDBuilder *B,
+                                                         SPIRVMDWalker *W) {
+  using namespace VectorComputeUtil;
+
+  auto EM = B->addNamedMD(kSPIRVMD::ExecutionMode);
+
+  for (auto &F : *M) {
+    // Add VC float control execution modes
+    // RoundMode and FloatMode are always same for all types in VC
+    // While Denorm could be different for double, float and half
+    auto Attrs = F.getAttributes();
+    if (Attrs.hasFnAttribute(kVCMetadata::VCFloatControl)) {
+      SPIRVWord Mode = 0;
+      Attrs
+          .getAttribute(AttributeList::FunctionIndex,
+                        kVCMetadata::VCFloatControl)
+          .getValueAsString()
+          .getAsInteger(0, Mode);
+      spv::ExecutionMode ExecRoundMode =
+          VCRoundModeExecModeMap::map(getVCRoundMode(Mode));
+      spv::ExecutionMode ExecFloatMode =
+          VCFloatModeExecModeMap::map(getVCFloatMode(Mode));
+      VCFloatTypeSizeMap::foreach (
+          [&](VCFloatType FloatType, unsigned TargetWidth) {
+            EM.addOp().add(&F).add(ExecRoundMode).add(TargetWidth).done();
+            EM.addOp().add(&F).add(ExecFloatMode).add(TargetWidth).done();
+            EM.addOp()
+                .add(&F)
+                .add(VCDenormModeExecModeMap::map(
+                    getVCDenormPreserve(Mode, FloatType)))
+                .add(TargetWidth)
+                .done();
+          });
+    }
+    if (Attrs.hasFnAttribute(kVCMetadata::VCSLMSize)) {
+      SPIRVWord SLMSize = 0;
+      Attrs.getAttribute(AttributeList::FunctionIndex, kVCMetadata::VCSLMSize)
+          .getValueAsString()
+          .getAsInteger(0, SLMSize);
+      EM.addOp()
+          .add(&F)
+          .add(spv::ExecutionModeSharedLocalMemorySizeINTEL)
+          .add(SLMSize)
+          .done();
+    }
+  }
+}
+
 } // namespace SPIRV
 
 INITIALIZE_PASS(PreprocessMetadata, "preprocess-metadata",

lib/SPIRV/SPIRVReader.cpp

@@ -50,6 +50,7 @@
 #include "SPIRVType.h"
 #include "SPIRVUtil.h"
 #include "SPIRVValue.h"
+#include "VectorComputeUtil.h"
 
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/Analysis/LoopInfo.h"
@@ -108,7 +109,7 @@ const static char *Restrict = "restrict";
 const static char *Pipe = "pipe";
 } // namespace kOCLTypeQualifierName
 
-static bool isOpenCLKernel(SPIRVFunction *BF) {
+static bool isKernel(SPIRVFunction *BF) {
   return BF->getModule()->isEntryPoint(ExecutionModelKernel, BF->getId());
 }
 
@@ -1530,14 +1531,34 @@ Value *SPIRVToLLVM::transValueWithoutDecoration(SPIRVValue *BV, Function *F,
       assert(BB && "Invalid BB");
       return mapValue(BV, new AllocaInst(Ty, 0, BV->getName(), BB));
     }
-    auto AddrSpace = SPIRSPIRVAddrSpaceMap::rmap(BS);
+    SPIRAddressSpace AddrSpace;
+
+    bool IsVectorCompute =
+        BVar->hasDecorate(DecorationVectorComputeVariableINTEL);
+    if (IsVectorCompute) {
+      AddrSpace = VectorComputeUtil::getVCGlobalVarAddressSpace(BS);
+      if (!Initializer)
+        Initializer = UndefValue::get(Ty);
+    } else
+      AddrSpace = SPIRSPIRVAddrSpaceMap::rmap(BS);
+
     auto LVar = new GlobalVariable(*M, Ty, IsConst, LinkageTy, Initializer,
                                    BV->getName(), 0,
                                    GlobalVariable::NotThreadLocal, AddrSpace);
     LVar->setUnnamedAddr((IsConst && Ty->isArrayTy() &&
                           Ty->getArrayElementType()->isIntegerTy(8))
                              ? GlobalValue::UnnamedAddr::Global
                              : GlobalValue::UnnamedAddr::None);
+
+    if (IsVectorCompute) {
+      LVar->addAttribute(kVCMetadata::VCGlobalVariable);
+      SPIRVWord Offset;
+      if (BVar->hasDecorate(DecorationGlobalVariableOffsetINTEL, 0, &Offset))
+        LVar->addAttribute(kVCMetadata::VCByteOffset, utostr(Offset));
+      if (BVar->hasDecorate(DecorationVolatile))
+        LVar->addAttribute(kVCMetadata::VCVolatile);
+    }
+
     SPIRVBuiltinVariableKind BVKind;
     if (BVar->isBuiltin(&BVKind))
       BuiltinGVMap[LVar] = BVKind;
@@ -2403,7 +2424,7 @@ Function *SPIRVToLLVM::transFunction(SPIRVFunction *BF) {
   if (Loc != FuncMap.end())
     return Loc->second;
 
-  auto IsKernel = BM->isEntryPoint(ExecutionModelKernel, BF->getId());
+  auto IsKernel = isKernel(BF);
   auto Linkage = IsKernel ? GlobalValue::ExternalLinkage : transLinkageType(BF);
   FunctionType *FT = dyn_cast<FunctionType>(transType(BF->getFunctionType()));
   Function *F = cast<Function>(
@@ -3217,7 +3238,7 @@ bool SPIRVToLLVM::transFPContractMetadata() {
   bool ContractOff = false;
   for (unsigned I = 0, E = BM->getNumFunctions(); I != E; ++I) {
     SPIRVFunction *BF = BM->getFunction(I);
-    if (!isOpenCLKernel(BF))
+    if (!isKernel(BF))
       continue;
     if (BF->getExecutionMode(ExecutionModeContractionOff)) {
       ContractOff = true;
@@ -3295,6 +3316,7 @@ bool SPIRVToLLVM::transMetadata() {
     assert(F && "Invalid translated function");
 
     transOCLMetadata(BF);
+    transVectorComputeMetadata(BF);
 
     if (F->getCallingConv() != CallingConv::SPIR_KERNEL)
       continue;
@@ -3435,6 +3457,73 @@ bool SPIRVToLLVM::transOCLMetadata(SPIRVFunction *BF) {
   return true;
 }
 
+bool SPIRVToLLVM::transVectorComputeMetadata(SPIRVFunction *BF) {
+  using namespace VectorComputeUtil;
+  Function *F = static_cast<Function *>(getTranslatedValue(BF));
+  assert(F && "Invalid translated function");
+
+  if (BF->hasDecorate(DecorationStackCallINTEL))
+    F->addFnAttr(kVCMetadata::VCStackCall);
+
+  bool IsVectorCompute = BF->hasDecorate(DecorationVectorComputeFunctionINTEL);
+  if (!IsVectorCompute)
+    return true;
+  F->addFnAttr(kVCMetadata::VCFunction);
+
+  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
+       ++I) {
+    auto ArgNo = I->getArgNo();
+    SPIRVFunctionParameter *BA = BF->getArgument(ArgNo);
+    SPIRVWord Kind;
+    if (BA->hasDecorate(DecorationFuncParamIOKind, 0, &Kind)) {
+      Attribute Attr = Attribute::get(*Context, kVCMetadata::VCArgumentIOKind,
+                                      std::to_string(Kind));
+      F->addAttribute(ArgNo + 1, Attr);
+    }
+  }
+
+  // Do not add float control if there is no any
+  bool IsVCFloatControl = false;
+  unsigned FloatControl = 0;
+  // RoundMode and FloatMode are always same for all types in Cm
+  // While Denorm could be different for double, float and half
+  VCRoundModeExecModeMap::foreach ([&](VCRoundMode VCRM, ExecutionMode EM) {
+    if (BF->getExecutionMode(EM)) {
+      IsVCFloatControl = true;
+      FloatControl |= getVCFloatControl(VCRM);
+    }
+  });
+  VCFloatModeExecModeMap::foreach ([&](VCFloatMode VCFM, ExecutionMode EM) {
+    if (BF->getExecutionMode(EM)) {
+      IsVCFloatControl = true;
+      FloatControl |= getVCFloatControl(VCFM);
+    }
+  });
+  VCDenormModeExecModeMap::foreach ([&](VCDenormMode VCDM, ExecutionMode EM) {
+    auto ExecModes = BF->getExecutionModeRange(EM);
+    for (auto It = ExecModes.first; It != ExecModes.second; It++) {
+      IsVCFloatControl = true;
+      unsigned TargetWidth = (*It).second->getLiterals()[0];
+      VCFloatType FloatType = VCFloatTypeSizeMap::rmap(TargetWidth);
+      FloatControl |= getVCFloatControl(VCDM, FloatType);
+    }
+  });
+  if (IsVCFloatControl) {
+    Attribute Attr = Attribute::get(*Context, kVCMetadata::VCFloatControl,
+                                    std::to_string(FloatControl));
+    F->addAttribute(AttributeList::FunctionIndex, Attr);
+  }
+
+  if (auto EM = BF->getExecutionMode(ExecutionModeSharedLocalMemorySizeINTEL)) {
+    unsigned int SLMSize = EM->getLiterals()[0];
+    Attribute Attr = Attribute::get(*Context, kVCMetadata::VCSLMSize,
+                                    std::to_string(SLMSize));
+    F->addAttribute(AttributeList::FunctionIndex, Attr);
+  }
+
+  return true;
+}
+
 bool SPIRVToLLVM::transAlign(SPIRVValue *BV, Value *V) {
   if (auto AL = dyn_cast<AllocaInst>(V)) {
     SPIRVWord Align = 0;

lib/SPIRV/SPIRVReader.h

@@ -104,6 +104,7 @@ class SPIRVToLLVM {
   bool transFPContractMetadata();
   bool transMetadata();
   bool transOCLMetadata(SPIRVFunction *BF);
+  bool transVectorComputeMetadata(SPIRVFunction *BF);
   Value *transAsmINTEL(SPIRVAsmINTEL *BA);
   CallInst *transAsmCallINTEL(SPIRVAsmCallINTEL *BI, Function *F,
                               BasicBlock *BB);