[Backport to 14] fpbuiltin-max-error support

Changes were cherry-picked from the following commit: c6fe12b

This changes add SPIR-V translator support for the SPIR-V extension documented here: KhronosGroup/SPIRV-Registry#193. This extension adds one decoration to represent maximum error for FP operations and adds the related Capability. SPIRV Headers support for representing this in SPIR-V: KhronosGroup/SPIRV-Headers#363

intel/llvm#8134 added a new call-site attribute associated with FP builtin intrinsics. This attribute is named 'fpbuiltin-max-error'. Following example shows how this extension is supported in the translator. The input LLVM IR uses new LLVM builtin calls to represent FP operations. An attribute named 'fpbuiltin-max-error' is used to represent the max-error allowed in the FP operation. Example Input LLVM: %t6 = call float @llvm.fpbuiltin.sin.f32(float %f1) KhronosGroup#2 attributes KhronosGroup#2 = { "fpbuiltin-max-error"="2.5" }

This is translated into a SPIR-V instruction (for add/sub/mul/div/rem) and OpenCl extended instruction for other instructions. A decoration to represent the max-error is attached to the SPIR-V instruction.

SPIR-V code:
4 Decorate 97 FPMaxErrorDecorationINTEL 1075838976 6 ExtInst 2 97 1 sin 88

No new support is added to support translating this SPIR_V back to LLVM. Existing support is used. The decoration is translated back into named metadata associated with the LLVM instruction. This can be readily consumed by backends.

Based on input from @andykaylor, we emit attributes when the FP operation is translated back to a call to a builtin function and emit metadata otherwise.

Translated LLVM code for basic math functions (add/sub/mul/div/rem): %t6 = fmul float %f1, %f2, !fpbuiltin-max-error !7 !7 = !{!"2.500000"}

Translated LLVM code for other math functions:
%t6 = call spir_func float @_Z3sinf(float %f1) KhronosGroup#3
attributes KhronosGroup#3 = { "fpbuiltin-max-error"="4.000000" }

Author: MiloszSkobejko

include/LLVMSPIRVExtensions.inc

@@ -63,4 +63,5 @@ EXT(SPV_INTEL_masked_gather_scatter)
 EXT(SPV_INTEL_tensor_float32_conversion) // TODO: to remove old extension
 EXT(SPV_INTEL_tensor_float32_rounding)
 EXT(SPV_EXT_relaxed_printf_string_address_space)
+EXT(SPV_INTEL_fp_max_error)
 EXT(SPV_INTEL_cache_controls)

lib/SPIRV/SPIRVReader.cpp

@@ -3819,7 +3819,48 @@ void SPIRVToLLVM::transDecorationsToMetadata(SPIRVValue *BV, Value *V) {
     SetDecorationsMetadata(I);
 }
 
+namespace {
+
+static float convertSPIRVWordToFloat(SPIRVWord Spir) {
+  union {
+    float F;
+    SPIRVWord Spir;
+  } FPMaxError;
+  FPMaxError.Spir = Spir;
+  return FPMaxError.F;
+}
+
+static bool transFPMaxErrorDecoration(SPIRVValue *BV, Value *V,
+                                      LLVMContext *Context) {
+  SPIRVWord ID;
+  if (Instruction *I = dyn_cast<Instruction>(V))
+    if (BV->hasDecorate(DecorationFPMaxErrorDecorationINTEL, 0, &ID)) {
+      auto Literals =
+          BV->getDecorationLiterals(DecorationFPMaxErrorDecorationINTEL);
+      assert(Literals.size() == 1 &&
+             "FP Max Error decoration shall have 1 operand");
+      auto F = convertSPIRVWordToFloat(Literals[0]);
+      if (CallInst *CI = dyn_cast<CallInst>(I)) {
+        // Add attribute
+        auto A = llvm::Attribute::get(*Context, "fpbuiltin-max-error",
+                                      std::to_string(F));
+        CI->addFnAttr(A);
+      } else {
+        // Add metadata
+        MDNode *N =
+            MDNode::get(*Context, MDString::get(*Context, std::to_string(F)));
+        I->setMetadata("fpbuiltin-max-error", N);
+      }
+      return true;
+    }
+  return false;
+}
+} // namespace
+
 bool SPIRVToLLVM::transDecoration(SPIRVValue *BV, Value *V) {
+  if (transFPMaxErrorDecoration(BV, V, Context))
+    return true;
+
   if (!transAlign(BV, V))
     return false;
 

lib/SPIRV/SPIRVUtil.cpp

@@ -766,6 +766,8 @@ CallInst *mutateCallInst(
   auto NewCI = addCallInst(M, NewName, CI->getType(), Args, Attrs, CI, Mangle,
                            InstName, TakeFuncName);
   NewCI->setDebugLoc(CI->getDebugLoc());
+  NewCI->copyMetadata(*CI);
+  NewCI->setAttributes(CI->getAttributes());  
   LLVM_DEBUG(dbgs() << " => " << *NewCI << '\n');
   CI->replaceAllUsesWith(NewCI);
   CI->eraseFromParent();

lib/SPIRV/SPIRVWriter.cpp

@@ -105,6 +105,19 @@ using namespace llvm;
 using namespace SPIRV;
 using namespace OCLUtil;
 
+namespace {
+
+static SPIRVWord convertFloatToSPIRVWord(float F) {
+  union {
+    float F;
+    SPIRVWord Spir;
+  } FPMaxError;
+  FPMaxError.F = F;
+  return FPMaxError.Spir;
+}
+
+} // namespace
+
 namespace SPIRV {
 
 static void foreachKernelArgMD(
@@ -3396,6 +3409,26 @@ bool LLVMToSPIRVBase::isKnownIntrinsic(Intrinsic::ID Id) {
   }
 }
 
+// Add decoration if needed
+SPIRVInstruction *addFPBuiltinDecoration(SPIRVModule *BM, IntrinsicInst *II,
+                                         SPIRVInstruction *I) {
+  const bool AllowFPMaxError =
+      BM->isAllowedToUseExtension(ExtensionID::SPV_INTEL_fp_max_error);
+  assert(II->getCalledFunction()->getName().startswith("llvm.fpbuiltin"));
+  // Add a new decoration for llvm.builtin intrinsics, if needed
+  if (AllowFPMaxError)
+    if (II->getAttributes().hasFnAttr("fpbuiltin-max-error")) {
+      double F = 0.0;
+      II->getAttributes()
+          .getFnAttr("fpbuiltin-max-error")
+          .getValueAsString()
+          .getAsDouble(F);
+      I->addDecorate(DecorationFPMaxErrorDecorationINTEL,
+                     convertFloatToSPIRVWord(F));
+    }
+  return I;
+}
+
 // Performs mapping of LLVM IR rounding mode to SPIR-V rounding mode
 // Value *V is metadata <rounding mode> argument of
 // llvm.experimental.constrained.* intrinsics
@@ -4090,6 +4123,8 @@ SPIRVValue *LLVMToSPIRVBase::transIntrinsicInst(IntrinsicInst *II,
   }
 
   default:
+    if (auto *BVar = transFPBuiltinIntrinsicInst(II, BB))
+      return BVar;
     if (BM->isUnknownIntrinsicAllowed(II))
       return BM->addCallInst(
           transFunctionDecl(II->getCalledFunction()),
@@ -4105,6 +4140,124 @@ SPIRVValue *LLVMToSPIRVBase::transIntrinsicInst(IntrinsicInst *II,
   return nullptr;
 }
 
+LLVMToSPIRVBase::FPBuiltinType
+LLVMToSPIRVBase::getFPBuiltinType(IntrinsicInst *II, StringRef &OpName) {
+  StringRef Name = II->getCalledFunction()->getName();
+  if (!Name.startswith("llvm.fpbuiltin"))
+    return FPBuiltinType::UNKNOWN;
+  Name.consume_front("llvm.fpbuiltin.");
+  OpName = Name.split('.').first;
+  FPBuiltinType Type =
+      StringSwitch<FPBuiltinType>(OpName)
+          .Cases("fadd", "fsub", "fmul", "fdiv", "frem",
+                 FPBuiltinType::REGULAR_MATH)
+          .Cases("sin", "cos", "tan", FPBuiltinType::EXT_1OPS)
+          .Cases("sinh", "cosh", "tanh", FPBuiltinType::EXT_1OPS)
+          .Cases("asin", "acos", "atan", FPBuiltinType::EXT_1OPS)
+          .Cases("asinh", "acosh", "atanh", FPBuiltinType::EXT_1OPS)
+          .Cases("exp", "exp2", "exp10", "expm1", FPBuiltinType::EXT_1OPS)
+          .Cases("log", "log2", "log10", "log1p", FPBuiltinType::EXT_1OPS)
+          .Cases("sqrt", "rsqrt", "erf", "erfc", FPBuiltinType::EXT_1OPS)
+          .Cases("atan2", "pow", "hypot", "ldexp", FPBuiltinType::EXT_2OPS)
+          .Case("sincos", FPBuiltinType::EXT_3OPS)
+          .Default(FPBuiltinType::UNKNOWN);
+  return Type;
+}
+
+SPIRVValue *LLVMToSPIRVBase::transFPBuiltinIntrinsicInst(IntrinsicInst *II,
+                                                         SPIRVBasicBlock *BB) {
+  StringRef OpName;
+  auto FPBuiltinTypeVal = getFPBuiltinType(II, OpName);
+  if (FPBuiltinTypeVal == FPBuiltinType::UNKNOWN)
+    return nullptr;
+  switch (FPBuiltinTypeVal) {
+  case FPBuiltinType::REGULAR_MATH: {
+    auto BinOp = StringSwitch<Op>(OpName)
+                     .Case("fadd", OpFAdd)
+                     .Case("fsub", OpFSub)
+                     .Case("fmul", OpFMul)
+                     .Case("fdiv", OpFDiv)
+                     .Case("frem", OpFRem)
+                     .Default(OpUndef);
+    auto *BI = BM->addBinaryInst(BinOp, transType(II->getType()),
+                                 transValue(II->getArgOperand(0), BB),
+                                 transValue(II->getArgOperand(1), BB), BB);
+    return addFPBuiltinDecoration(BM, II, BI);
+  }
+  case FPBuiltinType::EXT_1OPS: {
+    if (!checkTypeForSPIRVExtendedInstLowering(II, BM))
+      break;
+    SPIRVType *STy = transType(II->getType());
+    std::vector<SPIRVValue *> Ops(1, transValue(II->getArgOperand(0), BB));
+    auto ExtOp = StringSwitch<SPIRVWord>(OpName)
+                     .Case("sin", OpenCLLIB::Sin)
+                     .Case("cos", OpenCLLIB::Cos)
+                     .Case("tan", OpenCLLIB::Tan)
+                     .Case("sinh", OpenCLLIB::Sinh)
+                     .Case("cosh", OpenCLLIB::Cosh)
+                     .Case("tanh", OpenCLLIB::Tanh)
+                     .Case("asin", OpenCLLIB::Asin)
+                     .Case("acos", OpenCLLIB::Acos)
+                     .Case("atan", OpenCLLIB::Atan)
+                     .Case("asinh", OpenCLLIB::Asinh)
+                     .Case("acosh", OpenCLLIB::Acosh)
+                     .Case("atanh", OpenCLLIB::Atanh)
+                     .Case("exp", OpenCLLIB::Exp)
+                     .Case("exp2", OpenCLLIB::Exp2)
+                     .Case("exp10", OpenCLLIB::Exp10)
+                     .Case("expm1", OpenCLLIB::Expm1)
+                     .Case("log", OpenCLLIB::Log)
+                     .Case("log2", OpenCLLIB::Log2)
+                     .Case("log10", OpenCLLIB::Log10)
+                     .Case("log1p", OpenCLLIB::Log1p)
+                     .Case("sqrt", OpenCLLIB::Sqrt)
+                     .Case("rsqrt", OpenCLLIB::Rsqrt)
+                     .Case("erf", OpenCLLIB::Erf)
+                     .Case("erfc", OpenCLLIB::Erfc)
+                     .Default(SPIRVWORD_MAX);
+    assert(ExtOp != SPIRVWORD_MAX);
+    auto *BI = BM->addExtInst(STy, BM->getExtInstSetId(SPIRVEIS_OpenCL), ExtOp,
+                              Ops, BB);
+    return addFPBuiltinDecoration(BM, II, BI);
+  }
+  case FPBuiltinType::EXT_2OPS: {
+    if (!checkTypeForSPIRVExtendedInstLowering(II, BM))
+      break;
+    SPIRVType *STy = transType(II->getType());
+    std::vector<SPIRVValue *> Ops{transValue(II->getArgOperand(0), BB),
+                                  transValue(II->getArgOperand(1), BB)};
+    auto ExtOp = StringSwitch<SPIRVWord>(OpName)
+                     .Case("atan2", OpenCLLIB::Atan2)
+                     .Case("hypot", OpenCLLIB::Hypot)
+                     .Case("pow", OpenCLLIB::Pow)
+                     .Case("ldexp", OpenCLLIB::Ldexp)
+                     .Default(SPIRVWORD_MAX);
+    assert(ExtOp != SPIRVWORD_MAX);
+    auto *BI = BM->addExtInst(STy, BM->getExtInstSetId(SPIRVEIS_OpenCL), ExtOp,
+                              Ops, BB);
+    return addFPBuiltinDecoration(BM, II, BI);
+  }
+  case FPBuiltinType::EXT_3OPS: {
+    if (!checkTypeForSPIRVExtendedInstLowering(II, BM))
+      break;
+    SPIRVType *STy = transType(II->getType());
+    std::vector<SPIRVValue *> Ops{transValue(II->getArgOperand(0), BB),
+                                  transValue(II->getArgOperand(1), BB),
+                                  transValue(II->getArgOperand(2), BB)};
+    auto ExtOp = StringSwitch<SPIRVWord>(OpName)
+                     .Case("sincos", OpenCLLIB::Sincos)
+                     .Default(SPIRVWORD_MAX);
+    assert(ExtOp != SPIRVWORD_MAX);
+    auto *BI = BM->addExtInst(STy, BM->getExtInstSetId(SPIRVEIS_OpenCL), ExtOp,
+                              Ops, BB);
+    return addFPBuiltinDecoration(BM, II, BI);
+  }
+  default:
+    return nullptr;
+  }
+  return nullptr;
+}
+
 SPIRVValue *LLVMToSPIRVBase::transFenceInst(FenceInst *FI,
                                             SPIRVBasicBlock *BB) {
   SPIRVWord MemorySemantics;

lib/SPIRV/SPIRVWriter.h

@@ -108,6 +108,16 @@ class LLVMToSPIRVBase {
   bool transWorkItemBuiltinCallsToVariables();
   bool isKnownIntrinsic(Intrinsic::ID Id);
   SPIRVValue *transIntrinsicInst(IntrinsicInst *Intrinsic, SPIRVBasicBlock *BB);
+  enum class FPBuiltinType {
+    REGULAR_MATH,
+    EXT_1OPS,
+    EXT_2OPS,
+    EXT_3OPS,
+    UNKNOWN
+  };
+  FPBuiltinType getFPBuiltinType(IntrinsicInst *II, StringRef &);
+  SPIRVValue *transFPBuiltinIntrinsicInst(IntrinsicInst *II,
+                                          SPIRVBasicBlock *BB);
   SPIRVValue *transFenceInst(FenceInst *FI, SPIRVBasicBlock *BB);
   SPIRVValue *transCallInst(CallInst *Call, SPIRVBasicBlock *BB);
   SPIRVValue *transDirectCallInst(CallInst *Call, SPIRVBasicBlock *BB);

lib/SPIRV/libSPIRV/SPIRVDecorate.h

@@ -178,6 +178,8 @@ class SPIRVDecorate : public SPIRVDecorateGeneric {
     case internal::DecorationInitModeINTEL:
     case internal::DecorationImplementInCSRINTEL:
       return ExtensionID::SPV_INTEL_global_variable_decorations;
+    case DecorationFPMaxErrorDecorationINTEL:
+      return ExtensionID::SPV_INTEL_fp_max_error;
     case internal::DecorationCacheControlLoadINTEL:
     case internal::DecorationCacheControlStoreINTEL:
       return ExtensionID::SPV_INTEL_cache_controls;

lib/SPIRV/libSPIRV/SPIRVEnum.h

@@ -470,6 +470,8 @@ template <> inline void SPIRVMap<Decoration, SPIRVCapVec>::init() {
                {internal::CapabilityCacheControlsINTEL});
   ADD_VEC_INIT(internal::DecorationCacheControlStoreINTEL,
                {internal::CapabilityCacheControlsINTEL});
+  ADD_VEC_INIT(DecorationFPMaxErrorDecorationINTEL,
+               {CapabilityFPMaxErrorINTEL});
 }
 
 template <> inline void SPIRVMap<BuiltIn, SPIRVCapVec>::init() {

lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h

@@ -179,6 +179,7 @@ template <> inline void SPIRVMap<Decoration, std::string>::init() {
   add(DecorationMediaBlockIOINTEL, "MediaBlockIOINTEL");
   add(DecorationAliasScopeINTEL, "AliasScopeINTEL");
   add(DecorationNoAliasINTEL, "NoAliasINTEL");
+  add(DecorationFPMaxErrorDecorationINTEL, "FPMaxErrorDecorationINTEL");
 
   // From spirv_internal.hpp
   add(internal::DecorationCallableFunctionINTEL, "CallableFunctionINTEL");
@@ -601,6 +602,7 @@ template <> inline void SPIRVMap<Capability, std::string>::init() {
   add(CapabilityDebugInfoModuleINTEL, "DebugInfoModuleINTEL");
   add(CapabilitySplitBarrierINTEL, "SplitBarrierINTEL");
   add(CapabilityGroupUniformArithmeticKHR, "GroupUniformArithmeticKHR");
+  add(CapabilityFPMaxErrorINTEL, "FPMaxErrorINTEL");
 
   // From spirv_internal.hpp
   add(internal::CapabilityFPGADSPControlINTEL, "FPGADSPControlINTEL");