Add dot2add and dot4add_*8packed intrinsics

docs/DXIL.rst

@@ -2248,6 +2248,9 @@ ID  Name                          Description
 159 CallShader                    Call a shader in the callable shader table supplied through the DispatchRays() API
 160 CreateHandleForLib            create resource handle from resource struct for library
 161 PrimitiveIndex                PrimitiveIndex for raytracing shaders
+162 Dot2AddHalf                   2D half dot product with accumulate to float
+163 Dot4AddI8Packed               signed dot product of 4 x i8 vectors packed into i32, with accumulate to i32
+164 Dot4AddU8Packed               unsigned dot product of 4 x u8 vectors packed into i32, with accumulate to i32
 === ============================= =======================================================================================================================================================================================================================
 
 

include/dxc/DXIL/DxilConstants.h

@@ -343,6 +343,11 @@ namespace DXIL {
     // Domain shader
     DomainLocation = 105, // DomainLocation
   
+    // Dot product with accumulate
+    Dot2AddHalf = 162, // 2D half dot product with accumulate to float
+    Dot4AddI8Packed = 163, // signed dot product of 4 x i8 vectors packed into i32, with accumulate to i32
+    Dot4AddU8Packed = 164, // unsigned dot product of 4 x u8 vectors packed into i32, with accumulate to i32
+  
     // Dot
     Dot2 = 54, // Two-dimensional vector dot-product
     Dot3 = 55, // Three-dimensional vector dot-product
@@ -549,8 +554,9 @@ namespace DXIL {
     NumOpCodes_Dxil_1_1 = 139,
     NumOpCodes_Dxil_1_2 = 141,
     NumOpCodes_Dxil_1_3 = 162,
+    NumOpCodes_Dxil_1_4 = 165,
   
-    NumOpCodes = 162 // exclusive last value of enumeration
+    NumOpCodes = 165 // exclusive last value of enumeration
   };
   // OPCODE-ENUM:END
 
@@ -592,6 +598,10 @@ namespace DXIL {
     // Domain shader
     DomainLocation,
   
+    // Dot product with accumulate
+    Dot2AddHalf,
+    Dot4AddPacked,
+  
     // Dot
     Dot2,
     Dot3,
@@ -756,8 +766,9 @@ namespace DXIL {
     NumOpClasses_Dxil_1_1 = 95,
     NumOpClasses_Dxil_1_2 = 97,
     NumOpClasses_Dxil_1_3 = 118,
+    NumOpClasses_Dxil_1_4 = 120,
   
-    NumOpClasses = 118 // exclusive last value of enumeration
+    NumOpClasses = 120 // exclusive last value of enumeration
   };
   // OPCODECLASS-ENUM:END
 

include/dxc/DXIL/DxilInstructions.h

@@ -5341,5 +5341,104 @@ struct DxilInst_PrimitiveIndex {
   // Metadata
   bool requiresUniformInputs() const { return false; }
 };
+
+/// This instruction 2D half dot product with accumulate to float
+struct DxilInst_Dot2AddHalf {
+  llvm::Instruction *Instr;
+  // Construction and identification
+  DxilInst_Dot2AddHalf(llvm::Instruction *pInstr) : Instr(pInstr) {}
+  operator bool() const {
+    return hlsl::OP::IsDxilOpFuncCallInst(Instr, hlsl::OP::OpCode::Dot2AddHalf);
+  }
+  // Validation support
+  bool isAllowed() const { return true; }
+  bool isArgumentListValid() const {
+    if (6 != llvm::dyn_cast<llvm::CallInst>(Instr)->getNumArgOperands()) return false;
+    return true;
+  }
+  // Metadata
+  bool requiresUniformInputs() const { return false; }
+  // Operand indexes
+  enum OperandIdx {
+    arg_acc = 1,
+    arg_ax = 2,
+    arg_ay = 3,
+    arg_bx = 4,
+    arg_by = 5,
+  };
+  // Accessors
+  llvm::Value *get_acc() const { return Instr->getOperand(1); }
+  void set_acc(llvm::Value *val) { Instr->setOperand(1, val); }
+  llvm::Value *get_ax() const { return Instr->getOperand(2); }
+  void set_ax(llvm::Value *val) { Instr->setOperand(2, val); }
+  llvm::Value *get_ay() const { return Instr->getOperand(3); }
+  void set_ay(llvm::Value *val) { Instr->setOperand(3, val); }
+  llvm::Value *get_bx() const { return Instr->getOperand(4); }
+  void set_bx(llvm::Value *val) { Instr->setOperand(4, val); }
+  llvm::Value *get_by() const { return Instr->getOperand(5); }
+  void set_by(llvm::Value *val) { Instr->setOperand(5, val); }
+};
+
+/// This instruction signed dot product of 4 x i8 vectors packed into i32, with accumulate to i32
+struct DxilInst_Dot4AddI8Packed {
+  llvm::Instruction *Instr;
+  // Construction and identification
+  DxilInst_Dot4AddI8Packed(llvm::Instruction *pInstr) : Instr(pInstr) {}
+  operator bool() const {
+    return hlsl::OP::IsDxilOpFuncCallInst(Instr, hlsl::OP::OpCode::Dot4AddI8Packed);
+  }
+  // Validation support
+  bool isAllowed() const { return true; }
+  bool isArgumentListValid() const {
+    if (4 != llvm::dyn_cast<llvm::CallInst>(Instr)->getNumArgOperands()) return false;
+    return true;
+  }
+  // Metadata
+  bool requiresUniformInputs() const { return false; }
+  // Operand indexes
+  enum OperandIdx {
+    arg_acc = 1,
+    arg_a = 2,
+    arg_b = 3,
+  };
+  // Accessors
+  llvm::Value *get_acc() const { return Instr->getOperand(1); }
+  void set_acc(llvm::Value *val) { Instr->setOperand(1, val); }
+  llvm::Value *get_a() const { return Instr->getOperand(2); }
+  void set_a(llvm::Value *val) { Instr->setOperand(2, val); }
+  llvm::Value *get_b() const { return Instr->getOperand(3); }
+  void set_b(llvm::Value *val) { Instr->setOperand(3, val); }
+};
+
+/// This instruction unsigned dot product of 4 x u8 vectors packed into i32, with accumulate to i32
+struct DxilInst_Dot4AddU8Packed {
+  llvm::Instruction *Instr;
+  // Construction and identification
+  DxilInst_Dot4AddU8Packed(llvm::Instruction *pInstr) : Instr(pInstr) {}
+  operator bool() const {
+    return hlsl::OP::IsDxilOpFuncCallInst(Instr, hlsl::OP::OpCode::Dot4AddU8Packed);
+  }
+  // Validation support
+  bool isAllowed() const { return true; }
+  bool isArgumentListValid() const {
+    if (4 != llvm::dyn_cast<llvm::CallInst>(Instr)->getNumArgOperands()) return false;
+    return true;
+  }
+  // Metadata
+  bool requiresUniformInputs() const { return false; }
+  // Operand indexes
+  enum OperandIdx {
+    arg_acc = 1,
+    arg_a = 2,
+    arg_b = 3,
+  };
+  // Accessors
+  llvm::Value *get_acc() const { return Instr->getOperand(1); }
+  void set_acc(llvm::Value *val) { Instr->setOperand(1, val); }
+  llvm::Value *get_a() const { return Instr->getOperand(2); }
+  void set_a(llvm::Value *val) { Instr->setOperand(2, val); }
+  llvm::Value *get_b() const { return Instr->getOperand(3); }
+  void set_b(llvm::Value *val) { Instr->setOperand(3, val); }
+};
 // INSTR-HELPER:END
 } // namespace hlsl

include/dxc/HlslIntrinsicOp.h

@@ -136,6 +136,9 @@ import hctdb_instrhelp
   IOP_determinant,
   IOP_distance,
   IOP_dot,
+  IOP_dot2add,
+  IOP_dot4add_i8packed,
+  IOP_dot4add_u8packed,
   IOP_dst,
   IOP_exp,
   IOP_exp2,

lib/DXIL/DxilOperations.cpp

@@ -311,6 +311,11 @@ const OP::OpCodeProperty OP::m_OpCodeProps[(unsigned)OP::OpCode::NumOpCodes] = {
 
   // Raytracing object space uint System Values                                                                              void,     h,     f,     d,    i1,    i8,   i16,   i32,   i64,   udt,   obj ,  function attribute
   {  OC::PrimitiveIndex,          "PrimitiveIndex",           OCC::PrimitiveIndex,           "primitiveIndex",            { false, false, false, false, false, false, false,  true, false, false, false}, Attribute::ReadNone, },
+
+  // Dot product with accumulate                                                                                             void,     h,     f,     d,    i1,    i8,   i16,   i32,   i64,   udt,   obj ,  function attribute
+  {  OC::Dot2AddHalf,             "Dot2AddHalf",              OCC::Dot2AddHalf,              "dot2AddHalf",               { false, false,  true, false, false, false, false, false, false, false, false}, Attribute::ReadNone, },
+  {  OC::Dot4AddI8Packed,         "Dot4AddI8Packed",          OCC::Dot4AddPacked,            "dot4AddPacked",             { false, false, false, false, false, false, false,  true, false, false, false}, Attribute::ReadNone, },
+  {  OC::Dot4AddU8Packed,         "Dot4AddU8Packed",          OCC::Dot4AddPacked,            "dot4AddPacked",             { false, false, false, false, false, false, false,  true, false, false, false}, Attribute::ReadNone, },
 };
 // OPCODE-OLOADS:END
 
@@ -640,6 +645,11 @@ void OP::GetMinShaderModelAndMask(OpCode C, bool bWithTranslation,
     }
     return;
   }
+  // Instructions: Dot2AddHalf=162, Dot4AddI8Packed=163, Dot4AddU8Packed=164
+  if ((162 <= op && op <= 164)) {
+    major = 6;  minor = 4;
+    return;
+  }
   // OPCODE-SMMASK:END
 #undef SFLAG
 }
@@ -1030,6 +1040,11 @@ Function *OP::GetOpFunc(OpCode opCode, Type *pOverloadType) {
 
     // Raytracing object space uint System Values
   case OpCode::PrimitiveIndex:         A(pI32);     A(pI32); break;
+
+    // Dot product with accumulate
+  case OpCode::Dot2AddHalf:            A(pETy);     A(pI32); A(pETy); A(pF16); A(pF16); A(pF16); A(pF16); break;
+  case OpCode::Dot4AddI8Packed:        A(pI32);     A(pI32); A(pI32); A(pI32); A(pI32); break;
+  case OpCode::Dot4AddU8Packed:        A(pI32);     A(pI32); A(pI32); A(pI32); A(pI32); break;
   // OPCODE-OLOAD-FUNCS:END
   default: DXASSERT(false, "otherwise unhandled case"); break;
   }
@@ -1203,6 +1218,8 @@ llvm::Type *OP::GetOverloadType(OpCode opCode, llvm::Function *F) {
   case OpCode::DispatchRaysIndex:
   case OpCode::DispatchRaysDimensions:
   case OpCode::PrimitiveIndex:
+  case OpCode::Dot4AddI8Packed:
+  case OpCode::Dot4AddU8Packed:
     return IntegerType::get(m_Ctx, 32);
   case OpCode::CalculateLOD:
   case OpCode::DomainLocation:

lib/HLSL/DxilValidation.cpp

@@ -849,6 +849,9 @@ static bool ValidateOpcodeInProfile(DXIL::OpCode opcode,
   if ((145 <= op && op <= 146))
     return (major > 6 || (major == 6 && minor >= 3))
         && (SK == DXIL::ShaderKind::Library || SK == DXIL::ShaderKind::RayGeneration || SK == DXIL::ShaderKind::Intersection || SK == DXIL::ShaderKind::AnyHit || SK == DXIL::ShaderKind::ClosestHit || SK == DXIL::ShaderKind::Miss || SK == DXIL::ShaderKind::Callable);
+  // Instructions: Dot2AddHalf=162, Dot4AddI8Packed=163, Dot4AddU8Packed=164
+  if ((162 <= op && op <= 164))
+    return (major > 6 || (major == 6 && minor >= 4));
   return true;
   // VALOPCODESM-TEXT:END
 }
@@ -5074,6 +5077,9 @@ void GetValidationVersion(_Out_ unsigned *pMajor, _Out_ unsigned *pMinor) {
   // - Library support
   // - Raytracing support
   // - i64/f64 overloads for rawBufferLoad/Store
+  // 1.4 adds:
+  // - packed u8x4/i8x4 dot with accumulate to i32
+  // - half dot2 with accumulate to float
   *pMajor = 1;
   *pMinor = 4;
 }

lib/HLSL/HLOperationLower.cpp

@@ -4560,6 +4560,62 @@ Value *TranslateNoArgNoReturnPreserveOutput(CallInst *CI, IntrinsicOp IOP, OP::O
   return pResult;
 }
 
+// Special half dot2 with accumulate to float
+Value *TranslateDot2Add(CallInst *CI, IntrinsicOp IOP, OP::OpCode opcode,
+                        HLOperationLowerHelper &helper,
+                        HLObjectOperationLowerHelper *pObjHelper,
+                        bool &Translated) {
+  hlsl::OP *hlslOP = &helper.hlslOP;
+  Value *src0 = CI->getArgOperand(HLOperandIndex::kTrinaryOpSrc0Idx);
+  Type *srcTy = src0->getType();
+  const unsigned vecSize = 2;
+  DXASSERT(srcTy->isVectorTy() && vecSize == srcTy->getVectorNumElements() &&
+           srcTy->getScalarType()->isHalfTy(),
+           "otherwise, unexpected input dimension or component type");
+
+  Value *src1 = CI->getArgOperand(HLOperandIndex::kTrinaryOpSrc1Idx);
+  DXASSERT(srcTy == src1->getType(), "otherwise, mismatched argument types");
+  Value *accArg = CI->getArgOperand(HLOperandIndex::kTrinaryOpSrc2Idx);
+  Type *accTy = accArg->getType();
+  DXASSERT(!accTy->isVectorTy() && accTy->isFloatTy(),
+           "otherwise, unexpected accumulator type");
+  IRBuilder<> Builder(CI);
+
+  Function *dxilFunc = hlslOP->GetOpFunc(opcode, accTy);
+  Constant *opArg = hlslOP->GetU32Const((unsigned)opcode);
+
+  SmallVector<Value *, 6> args;
+  args.emplace_back(opArg);
+  args.emplace_back(accArg);
+  for (unsigned i = 0; i < vecSize; i++)
+    args.emplace_back(Builder.CreateExtractElement(src0, i));
+  for (unsigned i = 0; i < vecSize; i++)
+    args.emplace_back(Builder.CreateExtractElement(src1, i));
+  return Builder.CreateCall(dxilFunc, args);
+}
+
+Value *TranslateDot4AddPacked(CallInst *CI, IntrinsicOp IOP, OP::OpCode opcode,
+                              HLOperationLowerHelper &helper,
+                              HLObjectOperationLowerHelper *pObjHelper,
+                              bool &Translated) {
+  hlsl::OP *hlslOP = &helper.hlslOP;
+  Value *src0 = CI->getArgOperand(HLOperandIndex::kTrinaryOpSrc0Idx);
+  Type *srcTy = src0->getType();
+  DXASSERT(!srcTy->isVectorTy() && srcTy->isIntegerTy(32),
+    "otherwise, unexpected vector support in high level intrinsic tempalte");
+  Value *src1 = CI->getArgOperand(HLOperandIndex::kTrinaryOpSrc1Idx);
+  DXASSERT(srcTy == src1->getType(), "otherwise, mismatched argument types");
+  Value *accArg = CI->getArgOperand(HLOperandIndex::kTrinaryOpSrc2Idx);
+  Type *accTy = accArg->getType();
+  DXASSERT(!accTy->isVectorTy() && accTy->isIntegerTy(32),
+    "otherwise, unexpected vector support in high level intrinsic tempalte");
+  IRBuilder<> Builder(CI);
+
+  Function *dxilFunc = hlslOP->GetOpFunc(opcode, accTy);
+  Constant *opArg = hlslOP->GetU32Const((unsigned)opcode);
+  return Builder.CreateCall(dxilFunc, { opArg, accArg, src0, src1 });
+}
+
 } // namespace
 
 // Lower table.
@@ -4710,6 +4766,9 @@ IntrinsicLower gLowerTable[static_cast<unsigned>(IntrinsicOp::Num_Intrinsics)] =
     {IntrinsicOp::IOP_determinant, EmptyLower, DXIL::OpCode::NumOpCodes},
     {IntrinsicOp::IOP_distance, TranslateDistance, DXIL::OpCode::NumOpCodes},
     {IntrinsicOp::IOP_dot, TranslateDot, DXIL::OpCode::NumOpCodes},
+    {IntrinsicOp::IOP_dot2add, TranslateDot2Add, DXIL::OpCode::Dot2AddHalf},
+    {IntrinsicOp::IOP_dot4add_i8packed, TranslateDot4AddPacked, DXIL::OpCode::Dot4AddI8Packed},
+    {IntrinsicOp::IOP_dot4add_u8packed, TranslateDot4AddPacked, DXIL::OpCode::Dot4AddU8Packed},
     {IntrinsicOp::IOP_dst, TranslateDst, DXIL::OpCode::NumOpCodes},
     {IntrinsicOp::IOP_exp, TranslateExp, DXIL::OpCode::NumOpCodes},
     {IntrinsicOp::IOP_exp2, TrivialUnaryOperation, DXIL::OpCode::Exp},

tools/clang/lib/Sema/gen_intrin_main_tables_15.h

@@ -807,37 +807,61 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args121[] =
 };
 
 static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args122[] =
+{
+    {"dot2add", AR_QUAL_OUT, 3, LITEMPLATE_SCALAR, 3, LICOMPTYPE_FLOAT, 1, 1},
+    {"a", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT16, 1, 2},
+    {"b", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT16, 1, 2},
+    {"b", AR_QUAL_IN, 3, LITEMPLATE_SCALAR, 3, LICOMPTYPE_FLOAT, 1, 1},
+};
+
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args123[] =
+{
+    {"dot4add_i8packed", AR_QUAL_OUT, 3, LITEMPLATE_SCALAR, 3, LICOMPTYPE_INT, 1, 1},
+    {"a", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 1, LICOMPTYPE_UINT, 1, 1},
+    {"b", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 1, LICOMPTYPE_UINT, 1, 1},
+    {"b", AR_QUAL_IN, 3, LITEMPLATE_SCALAR, 3, LICOMPTYPE_INT, 1, 1},
+};
+
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args124[] =
+{
+    {"dot4add_u8packed", AR_QUAL_OUT, 3, LITEMPLATE_SCALAR, 3, LICOMPTYPE_UINT, 1, 1},
+    {"a", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 1, LICOMPTYPE_UINT, 1, 1},
+    {"b", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 1, LICOMPTYPE_UINT, 1, 1},
+    {"b", AR_QUAL_IN, 3, LITEMPLATE_SCALAR, 3, LICOMPTYPE_UINT, 1, 1},
+};
+
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args125[] =
 {
     {"dst", AR_QUAL_OUT, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_NUMERIC, 1, 4},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_NUMERIC, 1, 4},
     {"b", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_NUMERIC, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args123[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args126[] =
 {
     {"exp", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args124[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args127[] =
 {
     {"exp2", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args125[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args128[] =
 {
     {"f16tof32", AR_QUAL_OUT, 0, LITEMPLATE_ANY, 0, LICOMPTYPE_FLOAT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_UINT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args126[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args129[] =
 {
     {"f32tof16", AR_QUAL_OUT, 0, LITEMPLATE_ANY, 0, LICOMPTYPE_UINT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args127[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args130[] =
 {
     {"faceforward", AR_QUAL_OUT, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
     {"N", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
@@ -845,25 +869,25 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args127[] =
     {"Ng", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args128[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args131[] =
 {
     {"firstbithigh", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 0, LICOMPTYPE_UINT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_ANY_INT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args129[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args132[] =
 {
     {"firstbitlow", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 0, LICOMPTYPE_UINT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_ANY_INT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args130[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args133[] =
 {
     {"floor", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args131[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args134[] =
 {
     {"fma", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_DOUBLE_ONLY, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_DOUBLE_ONLY, IA_R, IA_C},
@@ -871,64 +895,64 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args131[] =
     {"c", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_DOUBLE_ONLY, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args132[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args135[] =
 {
     {"fmod", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"b", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args133[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args136[] =
 {
     {"frac", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args134[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args137[] =
 {
     {"frexp", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT, IA_R, IA_C},
     {"exp", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args135[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args138[] =
 {
     {"fwidth", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args136[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args139[] =
 {
     {"isfinite", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 0, LICOMPTYPE_BOOL, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args137[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args140[] =
 {
     {"isinf", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 0, LICOMPTYPE_BOOL, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args138[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args141[] =
 {
     {"isnan", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 0, LICOMPTYPE_BOOL, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args139[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args142[] =
 {
     {"ldexp", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"exp", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args140[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args143[] =
 {
     {"length", AR_QUAL_OUT, 0, LITEMPLATE_SCALAR, 1, LICOMPTYPE_FLOAT_LIKE, 1, 1},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args141[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args144[] =
 {
     {"lerp", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
@@ -936,7 +960,7 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args141[] =
     {"s", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args142[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args145[] =
 {
     {"lit", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"l", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 1, LICOMPTYPE_FLOAT_LIKE, 1, 1},
@@ -944,25 +968,25 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args142[] =
     {"m", AR_QUAL_IN, 3, LITEMPLATE_SCALAR, 1, LICOMPTYPE_FLOAT_LIKE, 1, 1},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args143[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args146[] =
 {
     {"log", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args144[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args147[] =
 {
     {"log10", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args145[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args148[] =
 {
     {"log2", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args146[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args149[] =
 {
     {"mad", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
@@ -970,28 +994,28 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args146[] =
     {"c", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args147[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args150[] =
 {
     {"max", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"b", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args148[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args151[] =
 {
     {"min", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"b", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args149[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args152[] =
 {
     {"modf", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"ip", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args150[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args153[] =
 {
     {"msad4", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_UINT, 1, 4},
     {"reference", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 1, LICOMPTYPE_UINT, 1, 1},
@@ -999,102 +1023,102 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args150[] =
     {"accum", AR_QUAL_IN, 3, LITEMPLATE_VECTOR, 3, LICOMPTYPE_UINT, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args151[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args154[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_SCALAR, 0, LICOMPTYPE_NUMERIC, 1, 1},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 0, LICOMPTYPE_NUMERIC, 1, 1},
     {"b", AR_QUAL_IN, 2, LITEMPLATE_SCALAR, 0, LICOMPTYPE_NUMERIC, 1, 1},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args152[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args155[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C2},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 0, LICOMPTYPE_NUMERIC, 1, 1},
     {"b", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C2},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args153[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args156[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_R2, IA_C2},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_SCALAR, 0, LICOMPTYPE_NUMERIC, 1, 1},
     {"b", AR_QUAL_IN, 2, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_R2, IA_C2},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args154[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args157[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C},
     {"b", AR_QUAL_IN, 2, LITEMPLATE_SCALAR, 0, LICOMPTYPE_NUMERIC, 1, 1},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args155[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args158[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_SCALAR, 0, LICOMPTYPE_NUMERIC, 1, 1},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C},
     {"b", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args156[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args159[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C2},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C},
     {"b", AR_QUAL_IN | AR_QUAL_COLMAJOR, 2, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_C, IA_C2},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args157[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args160[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"b", AR_QUAL_IN, 2, LITEMPLATE_SCALAR, 0, LICOMPTYPE_NUMERIC, 1, 1},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args158[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args161[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_R},
     {"a", AR_QUAL_IN | AR_QUAL_ROWMAJOR, 1, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"b", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 0, LICOMPTYPE_NUMERIC, 1, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args159[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args162[] =
 {
     {"mul", AR_QUAL_OUT, 0, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_R, IA_C2},
     {"a", AR_QUAL_IN | AR_QUAL_ROWMAJOR, 1, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_R, IA_C},
     {"b", AR_QUAL_IN | AR_QUAL_COLMAJOR, 2, LITEMPLATE_MATRIX, 0, LICOMPTYPE_NUMERIC, IA_C, IA_C2},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args160[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args163[] =
 {
     {"normalize", AR_QUAL_OUT, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args161[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args164[] =
 {
     {"pow", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"y", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args162[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args165[] =
 {
     {"radians", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args163[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args166[] =
 {
     {"rcp", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_ANY_FLOAT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_ANY_FLOAT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args164[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args167[] =
 {
     {"reflect", AR_QUAL_OUT, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
     {"i", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
     {"n", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args165[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args168[] =
 {
     {"refract", AR_QUAL_OUT, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
     {"i", AR_QUAL_IN, 1, LITEMPLATE_VECTOR, 1, LICOMPTYPE_FLOAT_LIKE, 1, IA_C},
@@ -1102,43 +1126,43 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args165[] =
     {"ri", AR_QUAL_IN, 3, LITEMPLATE_SCALAR, 3, LICOMPTYPE_FLOAT_LIKE, 1, 1},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args166[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args169[] =
 {
     {"reversebits", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_ANY_INT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_ANY_INT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args167[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args170[] =
 {
     {"round", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args168[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args171[] =
 {
     {"rsqrt", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args169[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args172[] =
 {
     {"saturate", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_ANY_FLOAT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_ANY_FLOAT, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args170[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args173[] =
 {
     {"sign", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 0, LICOMPTYPE_INT, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_NUMERIC, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args171[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args174[] =
 {
     {"sin", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args172[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args175[] =
 {
     {"sincos", 0, 0, LITEMPLATE_VOID, 0, LICOMPTYPE_VOID, 0, 0},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
@@ -1146,13 +1170,13 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args172[] =
     {"c", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args173[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args176[] =
 {
     {"sinh", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args174[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args177[] =
 {
     {"smoothstep", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
@@ -1160,44 +1184,44 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args174[] =
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args175[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args178[] =
 {
     {"source_mark", 0, 0, LITEMPLATE_VOID, 0, LICOMPTYPE_VOID, 0, 0},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args176[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args179[] =
 {
     {"sqrt", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args177[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args180[] =
 {
     {"step", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"a", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args178[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args181[] =
 {
     {"tan", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args179[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args182[] =
 {
     {"tanh", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args180[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args183[] =
 {
     {"tex1D", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER1D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_SCALAR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 1},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args181[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args184[] =
 {
     {"tex1D", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER1D, 1, 1},
@@ -1206,14 +1230,14 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args181[] =
     {"ddy", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 1},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args182[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args185[] =
 {
     {"tex1Dbias", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER1D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args183[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args186[] =
 {
     {"tex1Dgrad", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER1D, 1, 1},
@@ -1222,28 +1246,28 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args183[] =
     {"ddy", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 1},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args184[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args187[] =
 {
     {"tex1Dlod", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER1D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args185[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args188[] =
 {
     {"tex1Dproj", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER1D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args186[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args189[] =
 {
     {"tex2D", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER2D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 2},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args187[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args190[] =
 {
     {"tex2D", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER2D, 1, 1},
@@ -1252,14 +1276,14 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args187[] =
     {"ddy", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 2},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args188[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args191[] =
 {
     {"tex2Dbias", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER2D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args189[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args192[] =
 {
     {"tex2Dgrad", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER2D, 1, 1},
@@ -1268,28 +1292,28 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args189[] =
     {"ddy", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 2},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args190[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args193[] =
 {
     {"tex2Dlod", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER2D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args191[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args194[] =
 {
     {"tex2Dproj", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER2D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args192[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args195[] =
 {
     {"tex3D", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER3D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 3},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args193[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args196[] =
 {
     {"tex3D", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER3D, 1, 1},
@@ -1298,14 +1322,14 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args193[] =
     {"ddy", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 3},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args194[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args197[] =
 {
     {"tex3Dbias", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER3D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args195[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args198[] =
 {
     {"tex3Dgrad", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER3D, 1, 1},
@@ -1314,28 +1338,28 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args195[] =
     {"ddy", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 3},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args196[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args199[] =
 {
     {"tex3Dlod", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER3D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args197[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args200[] =
 {
     {"tex3Dproj", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLER3D, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args198[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args201[] =
 {
     {"texCUBE", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLERCUBE, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 3},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args199[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args202[] =
 {
     {"texCUBE", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLERCUBE, 1, 1},
@@ -1344,14 +1368,14 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args199[] =
     {"ddy", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 3},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args200[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args203[] =
 {
     {"texCUBEbias", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLERCUBE, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args201[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args204[] =
 {
     {"texCUBEgrad", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLERCUBE, 1, 1},
@@ -1360,27 +1384,27 @@ static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args201[] =
     {"ddy", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 3},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args202[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args205[] =
 {
     {"texCUBElod", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLERCUBE, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args203[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args206[] =
 {
     {"texCUBEproj", AR_QUAL_OUT, 0, LITEMPLATE_VECTOR, 0, LICOMPTYPE_FLOAT_LIKE, 1, 4},
     {"s", AR_QUAL_IN, 1, LITEMPLATE_OBJECT, 1, LICOMPTYPE_SAMPLERCUBE, 1, 1},
     {"x", AR_QUAL_IN, 2, LITEMPLATE_VECTOR, 2, LICOMPTYPE_FLOAT_LIKE, 1, 4},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args204[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args207[] =
 {
     {"transpose", AR_QUAL_OUT, 1, LITEMPLATE_MATRIX, 1, LICOMPTYPE_ANY, IA_C, IA_R},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_MATRIX, 1, LICOMPTYPE_ANY, IA_R, IA_C},
 };
 
-static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args205[] =
+static const HLSL_INTRINSIC_ARGUMENT g_Intrinsics_Args208[] =
 {
     {"trunc", AR_QUAL_OUT, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
     {"x", AR_QUAL_IN, 1, LITEMPLATE_ANY, 1, LICOMPTYPE_FLOAT_LIKE, IA_R, IA_C},
@@ -1510,90 +1534,93 @@ static const HLSL_INTRINSIC g_Intrinsics[] =
     {(UINT)hlsl::IntrinsicOp::IOP_determinant, false, true, -1, 2, g_Intrinsics_Args119},
     {(UINT)hlsl::IntrinsicOp::IOP_distance, false, true, -1, 3, g_Intrinsics_Args120},
     {(UINT)hlsl::IntrinsicOp::IOP_dot, false, true, -1, 3, g_Intrinsics_Args121},
-    {(UINT)hlsl::IntrinsicOp::IOP_dst, false, true, -1, 3, g_Intrinsics_Args122},
-    {(UINT)hlsl::IntrinsicOp::IOP_exp, false, true, -1, 2, g_Intrinsics_Args123},
-    {(UINT)hlsl::IntrinsicOp::IOP_exp2, false, true, -1, 2, g_Intrinsics_Args124},
-    {(UINT)hlsl::IntrinsicOp::IOP_f16tof32, false, true, -1, 2, g_Intrinsics_Args125},
-    {(UINT)hlsl::IntrinsicOp::IOP_f32tof16, false, true, -1, 2, g_Intrinsics_Args126},
-    {(UINT)hlsl::IntrinsicOp::IOP_faceforward, false, true, -1, 4, g_Intrinsics_Args127},
-    {(UINT)hlsl::IntrinsicOp::IOP_firstbithigh, false, true, 0, 2, g_Intrinsics_Args128},
-    {(UINT)hlsl::IntrinsicOp::IOP_firstbitlow, false, true, -1, 2, g_Intrinsics_Args129},
-    {(UINT)hlsl::IntrinsicOp::IOP_floor, false, true, -1, 2, g_Intrinsics_Args130},
-    {(UINT)hlsl::IntrinsicOp::IOP_fma, false, true, -1, 4, g_Intrinsics_Args131},
-    {(UINT)hlsl::IntrinsicOp::IOP_fmod, false, true, -1, 3, g_Intrinsics_Args132},
-    {(UINT)hlsl::IntrinsicOp::IOP_frac, false, true, -1, 2, g_Intrinsics_Args133},
-    {(UINT)hlsl::IntrinsicOp::IOP_frexp, false, false, -1, 3, g_Intrinsics_Args134},
-    {(UINT)hlsl::IntrinsicOp::IOP_fwidth, false, true, -1, 2, g_Intrinsics_Args135},
-    {(UINT)hlsl::IntrinsicOp::IOP_isfinite, false, true, -1, 2, g_Intrinsics_Args136},
-    {(UINT)hlsl::IntrinsicOp::IOP_isinf, false, true, -1, 2, g_Intrinsics_Args137},
-    {(UINT)hlsl::IntrinsicOp::IOP_isnan, false, true, -1, 2, g_Intrinsics_Args138},
-    {(UINT)hlsl::IntrinsicOp::IOP_ldexp, false, true, -1, 3, g_Intrinsics_Args139},
-    {(UINT)hlsl::IntrinsicOp::IOP_length, false, true, -1, 2, g_Intrinsics_Args140},
-    {(UINT)hlsl::IntrinsicOp::IOP_lerp, false, true, -1, 4, g_Intrinsics_Args141},
-    {(UINT)hlsl::IntrinsicOp::IOP_lit, false, true, -1, 4, g_Intrinsics_Args142},
-    {(UINT)hlsl::IntrinsicOp::IOP_log, false, true, -1, 2, g_Intrinsics_Args143},
-    {(UINT)hlsl::IntrinsicOp::IOP_log10, false, true, -1, 2, g_Intrinsics_Args144},
-    {(UINT)hlsl::IntrinsicOp::IOP_log2, false, true, -1, 2, g_Intrinsics_Args145},
-    {(UINT)hlsl::IntrinsicOp::IOP_mad, false, true, -1, 4, g_Intrinsics_Args146},
-    {(UINT)hlsl::IntrinsicOp::IOP_max, false, true, -1, 3, g_Intrinsics_Args147},
-    {(UINT)hlsl::IntrinsicOp::IOP_min, false, true, -1, 3, g_Intrinsics_Args148},
-    {(UINT)hlsl::IntrinsicOp::IOP_modf, false, false, -1, 3, g_Intrinsics_Args149},
-    {(UINT)hlsl::IntrinsicOp::IOP_msad4, false, true, -1, 4, g_Intrinsics_Args150},
-    {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args151},
-    {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args152},
-    {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args153},
+    {(UINT)hlsl::IntrinsicOp::IOP_dot2add, false, true, -1, 4, g_Intrinsics_Args122},
+    {(UINT)hlsl::IntrinsicOp::IOP_dot4add_i8packed, false, true, -1, 4, g_Intrinsics_Args123},
+    {(UINT)hlsl::IntrinsicOp::IOP_dot4add_u8packed, false, true, -1, 4, g_Intrinsics_Args124},
+    {(UINT)hlsl::IntrinsicOp::IOP_dst, false, true, -1, 3, g_Intrinsics_Args125},
+    {(UINT)hlsl::IntrinsicOp::IOP_exp, false, true, -1, 2, g_Intrinsics_Args126},
+    {(UINT)hlsl::IntrinsicOp::IOP_exp2, false, true, -1, 2, g_Intrinsics_Args127},
+    {(UINT)hlsl::IntrinsicOp::IOP_f16tof32, false, true, -1, 2, g_Intrinsics_Args128},
+    {(UINT)hlsl::IntrinsicOp::IOP_f32tof16, false, true, -1, 2, g_Intrinsics_Args129},
+    {(UINT)hlsl::IntrinsicOp::IOP_faceforward, false, true, -1, 4, g_Intrinsics_Args130},
+    {(UINT)hlsl::IntrinsicOp::IOP_firstbithigh, false, true, 0, 2, g_Intrinsics_Args131},
+    {(UINT)hlsl::IntrinsicOp::IOP_firstbitlow, false, true, -1, 2, g_Intrinsics_Args132},
+    {(UINT)hlsl::IntrinsicOp::IOP_floor, false, true, -1, 2, g_Intrinsics_Args133},
+    {(UINT)hlsl::IntrinsicOp::IOP_fma, false, true, -1, 4, g_Intrinsics_Args134},
+    {(UINT)hlsl::IntrinsicOp::IOP_fmod, false, true, -1, 3, g_Intrinsics_Args135},
+    {(UINT)hlsl::IntrinsicOp::IOP_frac, false, true, -1, 2, g_Intrinsics_Args136},
+    {(UINT)hlsl::IntrinsicOp::IOP_frexp, false, false, -1, 3, g_Intrinsics_Args137},
+    {(UINT)hlsl::IntrinsicOp::IOP_fwidth, false, true, -1, 2, g_Intrinsics_Args138},
+    {(UINT)hlsl::IntrinsicOp::IOP_isfinite, false, true, -1, 2, g_Intrinsics_Args139},
+    {(UINT)hlsl::IntrinsicOp::IOP_isinf, false, true, -1, 2, g_Intrinsics_Args140},
+    {(UINT)hlsl::IntrinsicOp::IOP_isnan, false, true, -1, 2, g_Intrinsics_Args141},
+    {(UINT)hlsl::IntrinsicOp::IOP_ldexp, false, true, -1, 3, g_Intrinsics_Args142},
+    {(UINT)hlsl::IntrinsicOp::IOP_length, false, true, -1, 2, g_Intrinsics_Args143},
+    {(UINT)hlsl::IntrinsicOp::IOP_lerp, false, true, -1, 4, g_Intrinsics_Args144},
+    {(UINT)hlsl::IntrinsicOp::IOP_lit, false, true, -1, 4, g_Intrinsics_Args145},
+    {(UINT)hlsl::IntrinsicOp::IOP_log, false, true, -1, 2, g_Intrinsics_Args146},
+    {(UINT)hlsl::IntrinsicOp::IOP_log10, false, true, -1, 2, g_Intrinsics_Args147},
+    {(UINT)hlsl::IntrinsicOp::IOP_log2, false, true, -1, 2, g_Intrinsics_Args148},
+    {(UINT)hlsl::IntrinsicOp::IOP_mad, false, true, -1, 4, g_Intrinsics_Args149},
+    {(UINT)hlsl::IntrinsicOp::IOP_max, false, true, -1, 3, g_Intrinsics_Args150},
+    {(UINT)hlsl::IntrinsicOp::IOP_min, false, true, -1, 3, g_Intrinsics_Args151},
+    {(UINT)hlsl::IntrinsicOp::IOP_modf, false, false, -1, 3, g_Intrinsics_Args152},
+    {(UINT)hlsl::IntrinsicOp::IOP_msad4, false, true, -1, 4, g_Intrinsics_Args153},
     {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args154},
     {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args155},
     {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args156},
     {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args157},
     {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args158},
     {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args159},
-    {(UINT)hlsl::IntrinsicOp::IOP_normalize, false, true, -1, 2, g_Intrinsics_Args160},
-    {(UINT)hlsl::IntrinsicOp::IOP_pow, false, true, -1, 3, g_Intrinsics_Args161},
-    {(UINT)hlsl::IntrinsicOp::IOP_radians, false, true, -1, 2, g_Intrinsics_Args162},
-    {(UINT)hlsl::IntrinsicOp::IOP_rcp, false, true, -1, 2, g_Intrinsics_Args163},
-    {(UINT)hlsl::IntrinsicOp::IOP_reflect, false, true, -1, 3, g_Intrinsics_Args164},
-    {(UINT)hlsl::IntrinsicOp::IOP_refract, false, true, -1, 4, g_Intrinsics_Args165},
-    {(UINT)hlsl::IntrinsicOp::IOP_reversebits, false, true, -1, 2, g_Intrinsics_Args166},
-    {(UINT)hlsl::IntrinsicOp::IOP_round, false, true, -1, 2, g_Intrinsics_Args167},
-    {(UINT)hlsl::IntrinsicOp::IOP_rsqrt, false, true, -1, 2, g_Intrinsics_Args168},
-    {(UINT)hlsl::IntrinsicOp::IOP_saturate, false, true, -1, 2, g_Intrinsics_Args169},
-    {(UINT)hlsl::IntrinsicOp::IOP_sign, false, true, -1, 2, g_Intrinsics_Args170},
-    {(UINT)hlsl::IntrinsicOp::IOP_sin, false, true, -1, 2, g_Intrinsics_Args171},
-    {(UINT)hlsl::IntrinsicOp::IOP_sincos, false, false, -1, 4, g_Intrinsics_Args172},
-    {(UINT)hlsl::IntrinsicOp::IOP_sinh, false, true, -1, 2, g_Intrinsics_Args173},
-    {(UINT)hlsl::IntrinsicOp::IOP_smoothstep, false, true, -1, 4, g_Intrinsics_Args174},
-    {(UINT)hlsl::IntrinsicOp::IOP_source_mark, false, false, -1, 1, g_Intrinsics_Args175},
-    {(UINT)hlsl::IntrinsicOp::IOP_sqrt, false, true, -1, 2, g_Intrinsics_Args176},
-    {(UINT)hlsl::IntrinsicOp::IOP_step, false, true, -1, 3, g_Intrinsics_Args177},
-    {(UINT)hlsl::IntrinsicOp::IOP_tan, false, true, -1, 2, g_Intrinsics_Args178},
-    {(UINT)hlsl::IntrinsicOp::IOP_tanh, false, true, -1, 2, g_Intrinsics_Args179},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex1D, true, false, -1, 3, g_Intrinsics_Args180},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex1D, true, false, -1, 5, g_Intrinsics_Args181},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex1Dbias, true, false, -1, 3, g_Intrinsics_Args182},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex1Dgrad, true, false, -1, 5, g_Intrinsics_Args183},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex1Dlod, true, false, -1, 3, g_Intrinsics_Args184},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex1Dproj, true, false, -1, 3, g_Intrinsics_Args185},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex2D, true, false, -1, 3, g_Intrinsics_Args186},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex2D, true, false, -1, 5, g_Intrinsics_Args187},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex2Dbias, true, false, -1, 3, g_Intrinsics_Args188},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex2Dgrad, true, false, -1, 5, g_Intrinsics_Args189},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex2Dlod, true, false, -1, 3, g_Intrinsics_Args190},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex2Dproj, true, false, -1, 3, g_Intrinsics_Args191},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex3D, true, false, -1, 3, g_Intrinsics_Args192},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex3D, true, false, -1, 5, g_Intrinsics_Args193},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex3Dbias, true, false, -1, 3, g_Intrinsics_Args194},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex3Dgrad, true, false, -1, 5, g_Intrinsics_Args195},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex3Dlod, true, false, -1, 3, g_Intrinsics_Args196},
-    {(UINT)hlsl::IntrinsicOp::IOP_tex3Dproj, true, false, -1, 3, g_Intrinsics_Args197},
-    {(UINT)hlsl::IntrinsicOp::IOP_texCUBE, true, false, -1, 3, g_Intrinsics_Args198},
-    {(UINT)hlsl::IntrinsicOp::IOP_texCUBE, true, false, -1, 5, g_Intrinsics_Args199},
-    {(UINT)hlsl::IntrinsicOp::IOP_texCUBEbias, true, false, -1, 3, g_Intrinsics_Args200},
-    {(UINT)hlsl::IntrinsicOp::IOP_texCUBEgrad, true, false, -1, 5, g_Intrinsics_Args201},
-    {(UINT)hlsl::IntrinsicOp::IOP_texCUBElod, true, false, -1, 3, g_Intrinsics_Args202},
-    {(UINT)hlsl::IntrinsicOp::IOP_texCUBEproj, true, false, -1, 3, g_Intrinsics_Args203},
-    {(UINT)hlsl::IntrinsicOp::IOP_transpose, false, true, -1, 2, g_Intrinsics_Args204},
-    {(UINT)hlsl::IntrinsicOp::IOP_trunc, false, true, -1, 2, g_Intrinsics_Args205},
+    {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args160},
+    {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args161},
+    {(UINT)hlsl::IntrinsicOp::IOP_mul, false, true, -1, 3, g_Intrinsics_Args162},
+    {(UINT)hlsl::IntrinsicOp::IOP_normalize, false, true, -1, 2, g_Intrinsics_Args163},
+    {(UINT)hlsl::IntrinsicOp::IOP_pow, false, true, -1, 3, g_Intrinsics_Args164},
+    {(UINT)hlsl::IntrinsicOp::IOP_radians, false, true, -1, 2, g_Intrinsics_Args165},
+    {(UINT)hlsl::IntrinsicOp::IOP_rcp, false, true, -1, 2, g_Intrinsics_Args166},
+    {(UINT)hlsl::IntrinsicOp::IOP_reflect, false, true, -1, 3, g_Intrinsics_Args167},
+    {(UINT)hlsl::IntrinsicOp::IOP_refract, false, true, -1, 4, g_Intrinsics_Args168},
+    {(UINT)hlsl::IntrinsicOp::IOP_reversebits, false, true, -1, 2, g_Intrinsics_Args169},
+    {(UINT)hlsl::IntrinsicOp::IOP_round, false, true, -1, 2, g_Intrinsics_Args170},
+    {(UINT)hlsl::IntrinsicOp::IOP_rsqrt, false, true, -1, 2, g_Intrinsics_Args171},
+    {(UINT)hlsl::IntrinsicOp::IOP_saturate, false, true, -1, 2, g_Intrinsics_Args172},
+    {(UINT)hlsl::IntrinsicOp::IOP_sign, false, true, -1, 2, g_Intrinsics_Args173},
+    {(UINT)hlsl::IntrinsicOp::IOP_sin, false, true, -1, 2, g_Intrinsics_Args174},
+    {(UINT)hlsl::IntrinsicOp::IOP_sincos, false, false, -1, 4, g_Intrinsics_Args175},
+    {(UINT)hlsl::IntrinsicOp::IOP_sinh, false, true, -1, 2, g_Intrinsics_Args176},
+    {(UINT)hlsl::IntrinsicOp::IOP_smoothstep, false, true, -1, 4, g_Intrinsics_Args177},
+    {(UINT)hlsl::IntrinsicOp::IOP_source_mark, false, false, -1, 1, g_Intrinsics_Args178},
+    {(UINT)hlsl::IntrinsicOp::IOP_sqrt, false, true, -1, 2, g_Intrinsics_Args179},
+    {(UINT)hlsl::IntrinsicOp::IOP_step, false, true, -1, 3, g_Intrinsics_Args180},
+    {(UINT)hlsl::IntrinsicOp::IOP_tan, false, true, -1, 2, g_Intrinsics_Args181},
+    {(UINT)hlsl::IntrinsicOp::IOP_tanh, false, true, -1, 2, g_Intrinsics_Args182},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex1D, true, false, -1, 3, g_Intrinsics_Args183},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex1D, true, false, -1, 5, g_Intrinsics_Args184},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex1Dbias, true, false, -1, 3, g_Intrinsics_Args185},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex1Dgrad, true, false, -1, 5, g_Intrinsics_Args186},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex1Dlod, true, false, -1, 3, g_Intrinsics_Args187},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex1Dproj, true, false, -1, 3, g_Intrinsics_Args188},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex2D, true, false, -1, 3, g_Intrinsics_Args189},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex2D, true, false, -1, 5, g_Intrinsics_Args190},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex2Dbias, true, false, -1, 3, g_Intrinsics_Args191},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex2Dgrad, true, false, -1, 5, g_Intrinsics_Args192},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex2Dlod, true, false, -1, 3, g_Intrinsics_Args193},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex2Dproj, true, false, -1, 3, g_Intrinsics_Args194},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex3D, true, false, -1, 3, g_Intrinsics_Args195},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex3D, true, false, -1, 5, g_Intrinsics_Args196},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex3Dbias, true, false, -1, 3, g_Intrinsics_Args197},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex3Dgrad, true, false, -1, 5, g_Intrinsics_Args198},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex3Dlod, true, false, -1, 3, g_Intrinsics_Args199},
+    {(UINT)hlsl::IntrinsicOp::IOP_tex3Dproj, true, false, -1, 3, g_Intrinsics_Args200},
+    {(UINT)hlsl::IntrinsicOp::IOP_texCUBE, true, false, -1, 3, g_Intrinsics_Args201},
+    {(UINT)hlsl::IntrinsicOp::IOP_texCUBE, true, false, -1, 5, g_Intrinsics_Args202},
+    {(UINT)hlsl::IntrinsicOp::IOP_texCUBEbias, true, false, -1, 3, g_Intrinsics_Args203},
+    {(UINT)hlsl::IntrinsicOp::IOP_texCUBEgrad, true, false, -1, 5, g_Intrinsics_Args204},
+    {(UINT)hlsl::IntrinsicOp::IOP_texCUBElod, true, false, -1, 3, g_Intrinsics_Args205},
+    {(UINT)hlsl::IntrinsicOp::IOP_texCUBEproj, true, false, -1, 3, g_Intrinsics_Args206},
+    {(UINT)hlsl::IntrinsicOp::IOP_transpose, false, true, -1, 2, g_Intrinsics_Args207},
+    {(UINT)hlsl::IntrinsicOp::IOP_trunc, false, true, -1, 2, g_Intrinsics_Args208},
 };
 
 //
@@ -5757,7 +5784,7 @@ static const UINT g_uAppendStructuredBufferMethodsCount = 2;
 static const UINT g_uBufferMethodsCount = 3;
 static const UINT g_uByteAddressBufferMethodsCount = 9;
 static const UINT g_uConsumeStructuredBufferMethodsCount = 2;
-static const UINT g_uIntrinsicsCount = 206;
+static const UINT g_uIntrinsicsCount = 209;
 static const UINT g_uRWBufferMethodsCount = 3;
 static const UINT g_uRWByteAddressBufferMethodsCount = 29;
 static const UINT g_uRWStructuredBufferMethodsCount = 5;

tools/clang/test/CodeGenHLSL/quick-test/sm_6_4_dot2_1.hlsl

@@ -0,0 +1,11 @@
+// RUN: %dxc -E main -T ps_6_4 -enable-16bit-types %s | FileCheck %s
+
+// CHECK:       ; Note: shader requires additional functionality:
+// CHECK-NEXT:  ;       Use native low precision
+
+// CHECK: call float @dx.op.dot2AddHalf.f32(i32 162,
+
+float2 main(half4 inputs : Inputs0, float acc : Acc0) : SV_Target {
+  acc = dot2add(inputs.xy, inputs.zw, acc);
+  return acc;
+}

tools/clang/test/CodeGenHLSL/quick-test/sm_6_4_dot2_2.hlsl

@@ -0,0 +1,8 @@
+// RUN: %dxc -E main -T ps_6_3 -enable-16bit-types %s | FileCheck %s
+
+// CHECK: Opcode Dot2AddHalf not valid in shader model ps_6_3
+
+float2 main(half4 inputs : Inputs0, float acc : Acc0) : SV_Target {
+  acc = dot2add(inputs.xy, inputs.zw, acc);
+  return acc;
+}

tools/clang/test/CodeGenHLSL/quick-test/sm_6_4_dot4_1.hlsl

@@ -0,0 +1,13 @@
+// RUN: %dxc -E main -T ps_6_4 %s | FileCheck %s
+
+// CHECK: call i32 @dx.op.dot4AddPacked.i32(i32 164,
+// CHECK: call i32 @dx.op.dot4AddPacked.i32(i32 163,
+
+float2 f2;
+
+float2 main(uint4 inputs : Inputs0, uint acc0 : Acc0, int acc1 : Acc1) : SV_Target {
+  uint acc = 0;
+  acc += dot4add_u8packed(inputs.x, inputs.y, acc0) * 2;
+  acc += dot4add_i8packed(inputs.z, inputs.w, acc1);
+  return f2 * acc;
+}

tools/clang/test/CodeGenHLSL/quick-test/sm_6_4_dot4_2.hlsl

@@ -0,0 +1,14 @@
+// RUN: %dxc -E main -T ps_6_3 %s | FileCheck %s
+
+// TODO: Check intrinsic earlier in front end
+// CHECK: Opcode Dot4AddU8Packed not valid in shader model ps_6_3
+// CHECK: Opcode Dot4AddI8Packed not valid in shader model ps_6_3
+
+float2 f2;
+
+float2 main(uint4 inputs : Inputs0, uint acc0 : Acc0, int acc1 : Acc1) : SV_Target {
+  uint acc = 0;
+  acc += dot4add_u8packed(inputs.x, inputs.y, acc0) * 2;
+  acc += dot4add_i8packed(inputs.z, inputs.w, acc1);
+  return f2 * acc;
+}

tools/clang/tools/dxcompiler/dxcdisassembler.cpp

@@ -1016,7 +1016,10 @@ static const char *OpCodeSignatures[] = {
   "(THit,HitKind,Attributes)",  // ReportHit
   "(ShaderIndex,Parameter)",  // CallShader
   "(Resource)",  // CreateHandleForLib
-  "()"  // PrimitiveIndex
+  "()",  // PrimitiveIndex
+  "(acc,ax,ay,bx,by)",  // Dot2AddHalf
+  "(acc,a,b)",  // Dot4AddI8Packed
+  "(acc,a,b)"  // Dot4AddU8Packed
 };
 // OPCODE-SIGS:END
 

utils/hct/gen_intrin_main.txt

@@ -297,6 +297,12 @@ float<3,4> [[rn]] ObjectToWorld3x4();
 float<3,4> [[rn]] WorldToObject3x4();
 float<4,3> [[rn]] ObjectToWorld4x3();
 float<4,3> [[rn]] WorldToObject4x3();
+
+// Packed dot products with accumulate:
+$type3 [[rn]] dot4add_u8packed(in uint a, in $type1 b, in uint b);
+$type3 [[rn]] dot4add_i8packed(in uint a, in $type1 b, in int b);
+$type3 [[rn]] dot2add(in float16_t<2> a, in $type1 b, in float b);
+
 } namespace
 
 namespace StreamMethods {

utils/hct/hctdb.py

@@ -374,6 +374,9 @@ class db_dxil(object):
             self.name_idx[i].category = "Library create handle from resource struct (like HL intrinsic)"
             self.name_idx[i].shader_model = 6,3
             self.name_idx[i].shader_model_translated = 6,0
+        for i in "Dot4AddU8Packed,Dot4AddI8Packed,Dot2AddHalf".split(","):
+            self.name_idx[i].category = "Dot product with accumulate"
+            self.name_idx[i].shader_model = 6,4
 
     def populate_llvm_instructions(self):
         # Add instructions that map to LLVM instructions.
@@ -1309,6 +1312,33 @@ class db_dxil(object):
         self.set_op_count_for_version(1, 3, next_op_idx)
         assert next_op_idx == 162, "next operation index is %d rather than 162 and thus opcodes are broken" % next_op_idx
 
+        self.add_dxil_op("Dot2AddHalf", next_op_idx, "Dot2AddHalf", "2D half dot product with accumulate to float", "f", "rn", [
+            db_dxil_param(0, "$o", "", "accumulated result"),
+            db_dxil_param(2, "$o", "acc", "input accumulator"),
+            db_dxil_param(3, "h", "ax", "the first component of the first vector"),
+            db_dxil_param(4, "h", "ay", "the second component of the first vector"),
+            db_dxil_param(5, "h", "bx", "the first component of the second vector"),
+            db_dxil_param(6, "h", "by", "the second component of the second vector")])
+        next_op_idx += 1
+
+        self.add_dxil_op("Dot4AddI8Packed", next_op_idx, "Dot4AddPacked", "signed dot product of 4 x i8 vectors packed into i32, with accumulate to i32", "i", "rn", [
+            db_dxil_param(0, "i32", "", "accumulated result"),
+            db_dxil_param(2, "i32", "acc", "input accumulator"),
+            db_dxil_param(3, "i32", "a", "first packed 4 x i8 for dot product"),
+            db_dxil_param(4, "i32", "b", "second packed 4 x i8 for dot product")])
+        next_op_idx += 1
+
+        self.add_dxil_op("Dot4AddU8Packed", next_op_idx, "Dot4AddPacked", "unsigned dot product of 4 x u8 vectors packed into i32, with accumulate to i32", "i", "rn", [
+            db_dxil_param(0, "i32", "", "accumulated result"),
+            db_dxil_param(2, "i32", "acc", "input accumulator"),
+            db_dxil_param(3, "i32", "a", "first packed 4 x u8 for dot product"),
+            db_dxil_param(4, "i32", "b", "second packed 4 x u8 for dot product")])
+        next_op_idx += 1
+
+        # End of DXIL 1.4 opcodes.
+        self.set_op_count_for_version(1, 4, next_op_idx)
+        assert next_op_idx == 165, "next operation index is %d rather than 165 and thus opcodes are broken" % next_op_idx
+
         # Set interesting properties.
         self.build_indices()
         for i in "CalculateLOD,DerivCoarseX,DerivCoarseY,DerivFineX,DerivFineY,Sample,SampleBias,SampleCmp,TextureGather,TextureGatherCmp".split(","):