[Phi] Add swish yaml and final state api (PaddlePaddle#41479)

* add swish yaml and final state api

* skip mkldnn test

* fix grad mkldnn test

Author: chenwhql

python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py

@@ -113,6 +113,7 @@ def setUp(self):
         super(TestMKLDNNSwishDim2, self).setUp()
 
         self.attrs["use_mkldnn"] = True
+        self.check_eager = False
 
     def init_dtype(self):
         self.dtype = np.float32
@@ -284,6 +285,7 @@ def setUp(self):
         self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
         self.outputs = {'Out': out}
         self.attrs = {"use_mkldnn": True, "beta": beta}
+        self.check_eager = False
 
     def init_dtype(self):
         self.dtype = np.float32

python/paddle/fluid/tests/unittests/test_activation_op.py

@@ -2928,7 +2928,9 @@ def ref_swish(x):
 class TestSwish(TestActivation):
     def setUp(self):
         self.op_type = "swish"
+        self.python_api = paddle.nn.functional.swish
         self.init_dtype()
+        self.check_eager = True
 
         np.random.seed(1024)
         x = np.random.uniform(-1, 1, [10, 12]).astype(self.dtype)
@@ -2940,7 +2942,10 @@ def setUp(self):
     def test_check_grad(self):
         if self.dtype == np.float16:
             return
-        self.check_grad(['X'], 'Out')
+        check_eager = False
+        if hasattr(self, 'check_eager'):
+            check_eager = self.check_eager
+        self.check_grad(['X'], 'Out', check_eager=check_eager)
 
 
 class TestSwishAPI(unittest.TestCase):
@@ -2975,6 +2980,10 @@ def test_dygraph_api(self):
             self.assertEqual(np.allclose(out_ref, r.numpy()), True)
         paddle.enable_static()
 
+    def test_dygraph_final_state_api(self):
+        with _test_eager_guard():
+            self.test_dygraph_api()
+
     def test_fluid_api(self):
         paddle.enable_static()
         with fluid.program_guard(fluid.Program()):

python/paddle/nn/functional/activation.py

@@ -1175,8 +1175,9 @@ def swish(x, name=None):
             x = paddle.to_tensor(np.array([-2., 0., 1.]))
             out = F.swish(x) # [-0.238406, 0., 0.731059]
     """
-
-    if in_dynamic_mode():
+    if in_dygraph_mode():
+        return _C_ops.final_state_swish(x, 1.0)
+    if _in_legacy_dygraph():
         return _C_ops.swish(x, 'beta', 1.0)
 
     check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'], 'swish')

python/paddle/utils/code_gen/api.yaml

@@ -1732,6 +1732,17 @@
     data_type : x
   backward : sum_grad
 
+# The python API paddle.nn.functional.swish has no `bete` argument, it may be removed later
+- api : swish
+  args : (Tensor x, float beta=1.0)
+  output : Tensor(out)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : swish
+  backward : swish_grad
+
 # take_along_axis
 - api : take_along_axis
   args : (Tensor x, Tensor index, int axis)

python/paddle/utils/code_gen/backward.yaml

@@ -1317,6 +1317,16 @@
   kernel :
     func : sum_grad
 
+- backward_api : swish_grad
+  forward : swish (Tensor x, float beta=1.0) -> Tensor(out)
+  args : (Tensor x, Tensor out_grad, float bete=1.0)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : GeneralUnaryGradInferMeta
+    param : [x]
+  kernel :
+    func : swish_grad
+
 - backward_api : take_along_axis_grad
   forward : take_along_axis (Tensor x, Tensor index, int axis) -> Tensor(out)
   args : (Tensor x, Tensor index, Tensor out_grad, int axis)