lgc: add CreateFDot2 to expose amdgcn_fdot2 to the client

lgc/builder/ArithBuilder.cpp

@@ -1260,6 +1260,24 @@ Value *BuilderImpl::CreateMsad4(Value *src, Value *ref, Value *accum, const Twin
   return result;
 }
 
+
+// =====================================================================================================================
+// Create "fdot2" operation, returning an 32-bit float result of the sum of dot2 of 2 half vec2 and a float scalar.
+//
+// @param a : Vector of 2xhalf A.
+// @param b : Vector of 2xhalf B.
+// @param scalar : A float scalar.
+// @param clamp : Whether the dot2 result should be clamped.
+Value *BuilderImpl::CreateFDot2(Value *a, Value *b, Value *scalar, Value *clamp, const Twine &instName) {
+  assert(a->getType()->getScalarType()->isHalfTy() && b->getType()->getScalarType()->isHalfTy());
+  assert(scalar->getType()->isFloatTy());
+  assert(clamp->getType()->isIntegerTy() && clamp->getType()->getIntegerBitWidth() == 1);
+
+  Value *result = CreateIntrinsic(scalar->getType(), Intrinsic::amdgcn_fdot2, {a, b, scalar, clamp});
+  result->setName(instName);
+  return result;
+}
+
 // =====================================================================================================================
 // Create "fmix" operation, returning ( 1 - A ) * X + A * Y. Result would be FP scalar or vector value.
 // Returns scalar, if and only if "pX", "pY" and "pA" are all scalars.

lgc/builder/BuilderRecorder.cpp

@@ -170,6 +170,8 @@ StringRef BuilderRecorder::getCallName(BuilderOpcode opcode) {
     return "fmix";
   case BuilderOpcode::Msad4:
     return "msad4";
+  case BuilderOpcode::FDot2:
+    return "fdot2";
   case BuilderOpcode::LoadBufferDesc:
     return "load.buffer.desc";
   case BuilderOpcode::GetDescStride:
@@ -1039,6 +1041,17 @@ Value *Builder::CreateMsad4(Value *src, Value *ref, Value *accum, const Twine &i
   return record(BuilderOpcode::Msad4, src->getType(), {src, ref, accum}, instName);
 }
 
+// =====================================================================================================================
+// Create "fdot2" operation, returning an 32-bit float result of the sum of dot2 of 2 half vec2 and a float scalar.
+//
+// @param a : Vector of 2xhalf A.
+// @param b : Vector of 2xhalf B.
+// @param scalar : A float scalar.
+// @param clamp : Whether the dot2 result should be clamped.
+Value *Builder::CreateFDot2(Value *a, Value *b, Value *scalar, Value *clamp, const Twine &instName) {
+  return record(BuilderOpcode::FDot2, scalar->getType(), {a, b, scalar, clamp}, instName);
+}
+
 // =====================================================================================================================
 // Create a load of a buffer descriptor.
 //
@@ -2017,6 +2030,7 @@ Instruction *Builder::record(BuilderOpcode opcode, Type *resultTy, ArrayRef<Valu
     case BuilderOpcode::FindSMsb:
     case BuilderOpcode::CountLeadingSignBits:
     case BuilderOpcode::Msad4:
+    case BuilderOpcode::FDot2:
     case BuilderOpcode::Fma:
     case BuilderOpcode::FpTruncWithRounding:
     case BuilderOpcode::Fract:

lgc/builder/BuilderRecorder.h

@@ -103,6 +103,7 @@ enum BuilderOpcode : unsigned {
   CountLeadingSignBits,
   FMix,
   Msad4,
+  FDot2,
 
   // Descriptor
   LoadBufferDesc,

lgc/builder/BuilderReplayer.cpp

@@ -396,6 +396,10 @@ Value *BuilderReplayer::processCall(unsigned opcode, CallInst *call) {
     return m_builder->CreateMsad4(args[0], args[1], args[2]);
   }
 
+  case BuilderOpcode::FDot2: {
+    return m_builder->CreateFDot2(args[0], args[1], args[2], args[3]);
+  }
+
   // Replayer implementations of DescBuilder methods
   case BuilderOpcode::LoadBufferDesc: {
     return m_builder->CreateLoadBufferDesc(cast<ConstantInt>(args[0])->getZExtValue(),  // descSet

lgc/include/lgc/builder/BuilderImpl.h

@@ -231,6 +231,10 @@ class BuilderImpl : public BuilderDefs {
   // Create "Masked Sum of Absolute Differences" operation.
   llvm::Value *CreateMsad4(llvm::Value *src, llvm::Value *ref, llvm::Value *accum, const llvm::Twine &instName = "");
 
+  // Create fdot2_f16 + f32 operation.
+  llvm::Value *CreateFDot2(llvm::Value *a, llvm::Value *b, llvm::Value *scalar, llvm::Value *clamp,
+                           const llvm::Twine &instName = "");
+
   // Create "fmix" operation.
   llvm::Value *createFMix(llvm::Value *x, llvm::Value *y, llvm::Value *a, const llvm::Twine &instName = "");
 

lgc/interface/lgc/Builder.h

@@ -837,6 +837,15 @@ class Builder : public BuilderDefs {
   // @param accum : A 32-bit unsigned integer, providing an existing accumulation.
   llvm::Value *CreateMsad4(llvm::Value *src, llvm::Value *ref, llvm::Value *accum, const llvm::Twine &instName = "");
 
+  // Create "fdot2" operation, returning an 32-bit float result of the sum of dot2 of 2 half vec2 and a float scalar.
+  //
+  // @param a : Vector of 2xhalf A.
+  // @param b : Vector of 2xhalf B.
+  // @param scalar : A float scalar.
+  // @param clamp : Whether the dot2 result should be clamped.
+  llvm::Value *CreateFDot2(llvm::Value *a, llvm::Value *b, llvm::Value *scalar, llvm::Value *clamp,
+                           const llvm::Twine &instName = "");
+
   // Create "fmix" operation, returning ( 1 - A ) * X + A * Y. Result would be FP scalar or vector value.
   // Returns scalar, if and only if "pX", "pY" and "pA" are all scalars.
   // Returns vector, if "pX" and "pY" are vector but "pA" is a scalar, under such condition, "pA" will be splatted.