Add LLVM Legalization for tir.erf

Author: Gokulnath Srinivasan

src/target/llvm/intrin_rule_llvm.cc

@@ -241,6 +241,18 @@ TVM_REGISTER_OP("tir.atanh")
       PrimExpr one = make_const(x.dtype(), 1.0);
       return (log(one + x) - log(one - x)) * make_const(x.dtype(), 0.5);
     });
+    
+TVM_REGISTER_OP("tir.erf").set_attr<FLegalize>("llvm.FLegalize", [](const PrimExpr& e) -> PrimExpr {
+  using tir::make_const;
+  const tir::CallNode* call = e.as<tir::CallNode>();
+  ICHECK(call != nullptr) << "Invalid call node in erf legalization";
+  const PrimExpr& x = call->args[0];
+  PrimExpr sqrt_pi = sqrt(make_const(x.dtype(), M_PI));
+  PrimExpr coeff = make_const(x.dtype(), 2.0) / sqrt_pi;
+  PrimExpr x_cubed = x * x * x;
+  PrimExpr inner = x + make_const(x.dtype(), 11.0 / 123.0) * x_cubed;
+  return tanh(coeff * inner);
+});
 
 TVM_REGISTER_OP("tir.clz").set_attr<FLegalize>("llvm.FLegalize", [](const PrimExpr& e) -> PrimExpr {
   const tir::CallNode* call = e.as<tir::CallNode>();

tests/python/relax/test_frontend_onnx.py

@@ -2804,7 +2804,7 @@ def test_symbolic_shape_deduction():
 
     @R.function
     def expected(
-        data: R.Tensor(("batch", "seq"), dtype="float32")
+        data: R.Tensor(("batch", "seq"), dtype="float32"),
     ) -> R.Tensor(dtype="float32", ndim=1):
         batch = T.int64()
         seq = T.int64()