add f16 and f16x2 intrinsics

Author: pgorlani

clang/include/clang/Basic/BuiltinsNVPTX.def

@@ -221,6 +221,8 @@ BUILTIN(__nvvm_cos_approx_f, "ff", "")
 // Tanh
 
 TARGET_BUILTIN(__nvvm_tanh_approx_f, "ff", "", AND(SM_80,PTX70))
+TARGET_BUILTIN(__nvvm_tanh_approx_f16, "hh", "", AND(SM_75, PTX70))
+TARGET_BUILTIN(__nvvm_tanh_approx_f16x2, "V2hV2h", "", AND(SM_75, PTX70))
 
 // Fma
 

llvm/include/llvm/IR/IntrinsicsNVVM.td

@@ -814,6 +814,10 @@ let TargetPrefix = "nvvm" in {
 
   def int_nvvm_tanh_approx_f : GCCBuiltin<"__nvvm_tanh_approx_f">,
        DefaultAttrsIntrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
+  def int_nvvm_tanh_approx_f16 : GCCBuiltin<"__nvvm_tanh_approx_f16">,
+      DefaultAttrsIntrinsic<[llvm_half_ty], [llvm_half_ty], [IntrNoMem]>;
+  def int_nvvm_tanh_approx_f16x2 : GCCBuiltin<"__nvvm_tanh_approx_f16x2">,
+      DefaultAttrsIntrinsic<[llvm_v2f16_ty], [llvm_v2f16_ty], [IntrNoMem]>;
 
 //
 // Fma

llvm/lib/Target/NVPTX/NVPTXIntrinsics.td

@@ -790,6 +790,10 @@ def INT_NVVM_COS_APPROX_F : F_MATH_1<"cos.approx.f32 \t$dst, $src0;",
 
 def INT_NVVM_TANH_APPROX_F : F_MATH_1<"tanh.approx.f32 \t$dst, $src0;",
    Float32Regs, Float32Regs, int_nvvm_tanh_approx_f>;
+def INT_NVVM_TANH_APPROX_F16 : F_MATH_1<"tanh.approx.f16 \t$dst, $src0;",
+  Float16Regs, Float16Regs, int_nvvm_tanh_approx_f16>;
+def INT_NVVM_TANH_APPROX_F16X2 : F_MATH_1<"tanh.approx.f16x2 \t$dst, $src0;",
+  Float16x2Regs, Float16x2Regs, int_nvvm_tanh_approx_f16x2>;
 
 //
 // Fma