[CUDA][CodeGen] Fix cuda codegen's fp16 inf literal (#12581)

* Fix cuda codegen's fp16 inf literal

* add relay testcase

src/target/source/codegen_cuda.cc

@@ -1197,8 +1197,10 @@ inline void PrintConst(const FloatImmNode* op, std::ostream& os, CodeGenCUDA* p)
       break;
     }
     case 16: {
-      os << "__float2half_rn";
-      os << '(' << std::scientific << op->value << 'f' << ')';
+      os << "__float2half_rn" << '(';
+      FloatImm const_f32 = FloatImm(DataType::Float(32), op->value);
+      PrintConst(const_f32.get(), os, p);
+      os << ')';
       break;
     }
     default:

tests/python/relay/test_op_level3.py

@@ -1344,7 +1344,7 @@ def verify_gather_nd(xshape, yshape, y_data, batch_dims=0, indices_dtype="int32"
     verify_gather_nd((2, 2, 2), (2, 2, 1), [[[1], [0]], [[0], [1]]], 1, indices_dtype="uint32")
 
 
-def _verify_infiniteness_ops(relay_op, ref_op):
+def _verify_infiniteness_ops(relay_op, ref_op, target="llvm", dev=None):
     for dtype in ["float32", "float16", "float16", "int32", "int16"]:
         shape = (2, 8, 8)
         x = relay.var("x", relay.TensorType(shape, dtype))
@@ -1359,17 +1359,25 @@ def _verify_infiniteness_ops(relay_op, ref_op):
             ] = np.infty
             data.ravel()[np.random.choice(data.size, int(data.size * 0.5), replace=False)] = np.nan
 
-        op_res = create_executor().evaluate(y, {x: data})
+        op_res = create_executor(target=target, device=dev).evaluate(y, {x: data})
         ref_res = ref_op(data)
         np.testing.assert_allclose(op_res.numpy(), ref_res, rtol=0.01)
 
 
+@tvm.testing.requires_gpu
 def test_isfinite():
-    _verify_infiniteness_ops(relay.isfinite, np.isfinite)
+    for target, dev in tvm.testing.enabled_targets():
+        if target not in ["llvm", "cuda"]:
+            continue
+        _verify_infiniteness_ops(relay.isfinite, np.isfinite, target=target, dev=dev)
 
 
+@tvm.testing.requires_gpu
 def test_isinf():
-    _verify_infiniteness_ops(relay.isinf, np.isinf)
+    for target, dev in tvm.testing.enabled_targets():
+        if target not in ["llvm", "cuda"]:
+            continue
+        _verify_infiniteness_ops(relay.isinf, np.isinf, target=target, dev=dev)
 
 
 def test_unravel_index(target, dev, executor_kind):