[HLSL] Implement dot2add intrinsic (#131237)

Resolves #99221 
Key points: For SPIRV backend, it decompose into a `dot` followed a
`add`.

- [x] Implement dot2add clang builtin,
- [x] Link dot2add clang builtin with hlsl_intrinsics.h
- [x] Add sema checks for dot2add to CheckHLSLBuiltinFunctionCall in
SemaHLSL.cpp
- [x] Add codegen for dot2add to EmitHLSLBuiltinExpr in CGBuiltin.cpp
- [x] Add codegen tests to clang/test/CodeGenHLSL/builtins/dot2add.hlsl
- [x] Add sema tests to clang/test/SemaHLSL/BuiltIns/dot2add-errors.hlsl
- [x] Create the int_dx_dot2add intrinsic in IntrinsicsDirectX.td
- [x] Create the DXILOpMapping of int_dx_dot2add to 162 in DXIL.td
- [x] Create the dot2add.ll and dot2add_errors.ll tests in
llvm/test/CodeGen/DirectX/

clang/include/clang/Basic/Builtins.td

@@ -4891,6 +4891,12 @@ def HLSLDotProduct : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "void(...)";
 }
 
+def HLSLDot2Add : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_dot2add"];
+  let Attributes = [NoThrow, Const];
+  let Prototype = "float(_ExtVector<2, _Float16>, _ExtVector<2, _Float16>, float)";
+}
+
 def HLSLDot4AddI8Packed : LangBuiltin<"HLSL_LANG"> {
   let Spellings = ["__builtin_hlsl_dot4add_i8packed"];
   let Attributes = [NoThrow, Const];

clang/lib/CodeGen/CGHLSLBuiltins.cpp

@@ -380,6 +380,19 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
         getDotProductIntrinsic(CGM.getHLSLRuntime(), VecTy0->getElementType()),
         ArrayRef<Value *>{Op0, Op1}, nullptr, "hlsl.dot");
   }
+  case Builtin::BI__builtin_hlsl_dot2add: {
+    llvm::Triple::ArchType Arch = CGM.getTarget().getTriple().getArch();
+    assert(Arch == llvm::Triple::dxil &&
+           "Intrinsic dot2add is only allowed for dxil architecture");
+    Value *A = EmitScalarExpr(E->getArg(0));
+    Value *B = EmitScalarExpr(E->getArg(1));
+    Value *C = EmitScalarExpr(E->getArg(2));
+
+    Intrinsic::ID ID = llvm ::Intrinsic::dx_dot2add;
+    return Builder.CreateIntrinsic(
+        /*ReturnType=*/C->getType(), ID, ArrayRef<Value *>{A, B, C}, nullptr,
+        "dx.dot2add");
+  }
   case Builtin::BI__builtin_hlsl_dot4add_i8packed: {
     Value *A = EmitScalarExpr(E->getArg(0));
     Value *B = EmitScalarExpr(E->getArg(1));

clang/lib/Headers/hlsl/hlsl_intrinsic_helpers.h

@@ -45,6 +45,14 @@ distance_vec_impl(vector<T, N> X, vector<T, N> Y) {
   return length_vec_impl(X - Y);
 }
 
+constexpr float dot2add_impl(half2 a, half2 b, float c) {
+#if defined(__DIRECTX__)
+  return __builtin_hlsl_dot2add(a, b, c);
+#else
+  return dot(a, b) + c;
+#endif
+}
+
 template <typename T> constexpr T reflect_impl(T I, T N) {
   return I - 2 * N * I * N;
 }

clang/lib/Headers/hlsl/hlsl_intrinsics.h

@@ -175,6 +175,21 @@ const inline float distance(__detail::HLSL_FIXED_VECTOR<float, N> X,
   return __detail::distance_vec_impl(X, Y);
 }
 
+//===----------------------------------------------------------------------===//
+// dot2add builtins
+//===----------------------------------------------------------------------===//
+
+/// \fn float dot2add(half2 A, half2 B, float C)
+/// \brief Dot product of 2 vector of type half and add a float scalar value.
+/// \param A The first input value to dot product.
+/// \param B The second input value to dot product.
+/// \param C The input value added to the dot product.
+
+_HLSL_AVAILABILITY(shadermodel, 6.4)
+const inline float dot2add(half2 A, half2 B, float C) {
+  return __detail::dot2add_impl(A, B, C);
+}
+
 //===----------------------------------------------------------------------===//
 // fmod builtins
 //===----------------------------------------------------------------------===//

clang/test/CodeGenHLSL/builtins/dot2add.hlsl

@@ -0,0 +1,135 @@
+// RUN: %clang_cc1 -finclude-default-header -fnative-half-type -triple \
+// RUN:   dxil-pc-shadermodel6.3-compute %s -emit-llvm -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-DXIL
+// RUN: %clang_cc1 -finclude-default-header -fnative-half-type -triple \
+// RUN:   spirv-pc-vulkan-compute %s -emit-llvm -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-SPIRV
+
+// Test basic lowering to runtime function call.
+
+// CHECK-LABEL: define {{.*}}test_default_parameter_type
+float test_default_parameter_type(half2 p1, half2 p2, float p3) {
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_float_arg2_type
+float test_float_arg2_type(half2 p1, float2 p2, float p3) {
+  // CHECK:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %{{.*}} to <2 x half>
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_float_arg1_type
+float test_float_arg1_type(float2 p1, half2 p2, float p3) {
+  // CHECK:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %{{.*}} to <2 x half>
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_double_arg3_type
+float test_double_arg3_type(half2 p1, half2 p2, double p3) {
+  // CHECK:  %conv = fptrunc reassoc nnan ninf nsz arcp afn double %{{.*}} to float
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_float_arg1_arg2_type
+float test_float_arg1_arg2_type(float2 p1, float2 p2, float p3) {
+  // CHECK:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %{{.*}} to <2 x half>
+  // CHECK:  %conv1 = fptrunc reassoc nnan ninf nsz arcp afn <2 x float> %{{.*}} to <2 x half>
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_double_arg1_arg2_type
+float test_double_arg1_arg2_type(double2 p1, double2 p2, float p3) {
+  // CHECK:  %conv = fptrunc reassoc nnan ninf nsz arcp afn <2 x double> %{{.*}} to <2 x half>
+  // CHECK:  %conv1 = fptrunc reassoc nnan ninf nsz arcp afn <2 x double> %{{.*}} to <2 x half>
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_int16_arg1_arg2_type
+float test_int16_arg1_arg2_type(int16_t2 p1, int16_t2 p2, float p3) {
+  // CHECK:  %conv = sitofp <2 x i16> %{{.*}} to <2 x half>
+  // CHECK:  %conv1 = sitofp <2 x i16> %{{.*}} to <2 x half>
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_int32_arg1_arg2_type
+float test_int32_arg1_arg2_type(int32_t2 p1, int32_t2 p2, float p3) {
+  // CHECK:  %conv = sitofp <2 x i32> %{{.*}} to <2 x half>
+  // CHECK:  %conv1 = sitofp <2 x i32> %{{.*}} to <2 x half>
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_int64_arg1_arg2_type
+float test_int64_arg1_arg2_type(int64_t2 p1, int64_t2 p2, float p3) {
+  // CHECK:  %conv = sitofp <2 x i64> %{{.*}} to <2 x half>
+  // CHECK:  %conv1 = sitofp <2 x i64> %{{.*}} to <2 x half>
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}
+
+// CHECK-LABEL: define {{.*}}test_bool_arg1_arg2_type
+float test_bool_arg1_arg2_type(bool2 p1, bool2 p2, float p3) {
+  // CHECK:  %loadedv = trunc <2 x i32> %{{.*}} to <2 x i1>
+  // CHECK:  %conv = uitofp <2 x i1> %loadedv to <2 x half>
+  // CHECK:  %loadedv1 = trunc <2 x i32> %{{.*}} to <2 x i1>
+  // CHECK:  %conv2 = uitofp <2 x i1> %loadedv1 to <2 x half>
+  // CHECK-SPIRV:  %[[MUL:.*]] = call reassoc nnan ninf nsz arcp afn half @llvm.spv.fdot.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}})
+  // CHECK-SPIRV:  %[[CONV:.*]] = fpext reassoc nnan ninf nsz arcp afn half %[[MUL]] to float
+  // CHECK-SPIRV:  %[[C:.*]] = load float, ptr %c.addr.i, align 4
+  // CHECK-SPIRV:  %[[RES:.*]] = fadd reassoc nnan ninf nsz arcp afn float %[[CONV]], %[[C]]
+  // CHECK-DXIL:  %[[RES:.*]] = call {{.*}} float @llvm.dx.dot2add.v2f16(<2 x half> %{{.*}}, <2 x half> %{{.*}}, float %{{.*}})
+  // CHECK:  ret float %[[RES]]
+  return dot2add(p1, p2, p3);
+}

clang/test/SemaHLSL/BuiltIns/dot2add-errors.hlsl

@@ -0,0 +1,13 @@
+// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.6-library %s -fnative-half-type -emit-llvm-only -disable-llvm-passes -verify
+
+float test_too_few_arg() {
+  return dot2add();
+  // expected-error@-1 {{no matching function for call to 'dot2add'}}
+  // expected-note@hlsl/hlsl_intrinsics.h:* {{candidate function not viable: requires 3 arguments, but 0 were provided}}
+}
+
+float test_too_many_arg(half2 p1, half2 p2, float p3) {
+  return dot2add(p1, p2, p3, p1);
+  // expected-error@-1 {{no matching function for call to 'dot2add'}}
+  // expected-note@hlsl/hlsl_intrinsics.h:* {{candidate function not viable: requires 3 arguments, but 4 were provided}}
+}

llvm/include/llvm/IR/IntrinsicsDirectX.td

@@ -100,6 +100,10 @@ def int_dx_udot :
     DefaultAttrsIntrinsic<[LLVMVectorElementType<0>],
     [llvm_anyint_ty, LLVMScalarOrSameVectorWidth<0, LLVMVectorElementType<0>>],
     [IntrNoMem, Commutative] >;
+def int_dx_dot2add : 
+    DefaultAttrsIntrinsic<[llvm_float_ty], 
+    [llvm_anyfloat_ty, LLVMMatchType<0>, llvm_float_ty], 
+    [IntrNoMem, Commutative]>;
 def int_dx_dot4add_i8packed : DefaultAttrsIntrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
 def int_dx_dot4add_u8packed : DefaultAttrsIntrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
 

llvm/lib/Target/DirectX/DXIL.td

@@ -1077,6 +1077,17 @@ def RawBufferStore : DXILOp<140, rawBufferStore> {
   let stages = [Stages<DXIL1_2, [all_stages]>];
 }
 
+def Dot2AddHalf : DXILOp<162, dot2AddHalf> {
+  let Doc = "dot product of 2 vectors of half having size = 2, returns "
+            "float";
+  let intrinsics = [IntrinSelect<int_dx_dot2add>];
+  let arguments = [FloatTy, HalfTy, HalfTy, HalfTy, HalfTy];
+  let result = FloatTy;
+  let overloads = [Overloads<DXIL1_0, []>];
+  let stages = [Stages<DXIL1_0, [all_stages]>];
+  let attributes = [Attributes<DXIL1_0, [ReadNone]>];
+}
+
 def Dot4AddI8Packed : DXILOp<163, dot4AddPacked> {
   let Doc = "signed dot product of 4 x i8 vectors packed into i32, with "
             "accumulate to i32";

llvm/lib/Target/DirectX/DXILOpLowering.cpp

@@ -54,10 +54,8 @@ static SmallVector<Value *> populateOperands(Value *Arg, IRBuilder<> &Builder) {
   return ExtractedElements;
 }
 
-static SmallVector<Value *> argVectorFlatten(CallInst *Orig,
-                                             IRBuilder<> &Builder) {
-  // Note: arg[NumOperands-1] is a pointer and is not needed by our flattening.
-  unsigned NumOperands = Orig->getNumOperands() - 1;
+static SmallVector<Value *>
+argVectorFlatten(CallInst *Orig, IRBuilder<> &Builder, unsigned NumOperands) {
   assert(NumOperands > 0);
   Value *Arg0 = Orig->getOperand(0);
   [[maybe_unused]] auto *VecArg0 = dyn_cast<FixedVectorType>(Arg0->getType());
@@ -75,6 +73,12 @@ static SmallVector<Value *> argVectorFlatten(CallInst *Orig,
   return NewOperands;
 }
 
+static SmallVector<Value *> argVectorFlatten(CallInst *Orig,
+                                             IRBuilder<> &Builder) {
+  // Note: arg[NumOperands-1] is a pointer and is not needed by our flattening.
+  return argVectorFlatten(Orig, Builder, Orig->getNumOperands() - 1);
+}
+
 namespace {
 class OpLowerer {
   Module &M;
@@ -168,6 +172,13 @@ public:
         }
       } else if (IsVectorArgExpansion) {
         Args = argVectorFlatten(CI, OpBuilder.getIRB());
+      } else if (F.getIntrinsicID() == Intrinsic::dx_dot2add) {
+        // arg[NumOperands-1] is a pointer and is not needed by our flattening.
+        // arg[NumOperands-2] also does not need to be flattened because it is a
+        // scalar.
+        unsigned NumOperands = CI->getNumOperands() - 2;
+        Args.push_back(CI->getArgOperand(NumOperands));
+        Args.append(argVectorFlatten(CI, OpBuilder.getIRB(), NumOperands));
       } else {
         Args.append(CI->arg_begin(), CI->arg_end());
       }

llvm/test/CodeGen/DirectX/dot2add.ll

@@ -0,0 +1,8 @@
+; RUN: opt -S -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-compute %s | FileCheck %s
+
+define noundef float @dot2add_simple(<2 x half> noundef %a, <2 x half> noundef %b, float %c) {
+entry:
+; CHECK: call float @dx.op.dot2AddHalf(i32 162, float %c, half %0, half %1, half %2, half %3)
+  %ret = call float @llvm.dx.dot2add(<2 x half> %a, <2 x half> %b, float %c)
+  ret float %ret
+}