[AMP OP&Test]stack & unstack ops fp16 bf16 support (#50999)

* stack fp16 & bf16 support

* unstack fp16 support

* unstack bf16 support

* append stack fp16 ut

* add unstack

* recover unstack cpu kernel

* fix some issue for unstack ut

* delete unuse var

* add check_place

* fix inference err

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

@@ -105,6 +105,47 @@ class TestStackOp_ZeroDim(TestStackOpBase):
         self.enable_cinn = False
 
 
+class TestStackFP16Op(TestStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+
+
+class TestStackFP16Op1(TestStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.num_inputs = 8
+
+
+class TestStackFP16Op2(TestStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.num_inputs = 10
+
+
+class TestStackFP16Op3(TestStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.axis = -1
+
+
+class TestStackFP16Op4(TestStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.axis = -4
+
+
+class TestStackFP16Op5(TestStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.axis = 1
+
+
+class TestStackFP16Op6(TestStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.axis = 3
+
+
 class TestStackBF16Op(OpTest):
     def initDefaultParameters(self):
         self.num_inputs = 4

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

@@ -15,9 +15,11 @@
 import unittest
 
 import numpy as np
-from eager_op_test import OpTest
+from eager_op_test import OpTest, convert_float_to_uint16
 
 import paddle
+from paddle import fluid
+from paddle.fluid import core
 
 
 class TestUnStackOpBase(OpTest):
@@ -64,6 +66,35 @@ class TestUnStackOpBase(OpTest):
         self.check_grad(['X'], self.get_y_names())
 
 
+class TestUnStackFP16Op(TestUnStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+
+
+class TestStackFP16Op3(TestUnStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.axis = -1
+
+
+class TestStackFP16Op4(TestUnStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.axis = -3
+
+
+class TestStackFP16Op5(TestUnStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.axis = 1
+
+
+class TestStackFP16Op6(TestUnStackOpBase):
+    def initParameters(self):
+        self.dtype = np.float16
+        self.axis = 2
+
+
 class TestStackOp3(TestUnStackOpBase):
     def initParameters(self):
         self.axis = -1
@@ -84,6 +115,71 @@ class TestStackOp6(TestUnStackOpBase):
         self.axis = 2
 
 
+@unittest.skipIf(
+    not core.is_compiled_with_cuda()
+    or not core.is_bfloat16_supported(core.CUDAPlace(0)),
+    "core is not compiled with CUDA and do not support bfloat16",
+)
+class TestUnStackBF16Op(OpTest):
+    def initDefaultParameters(self):
+        self.input_dim = (5, 6, 7)
+        self.axis = 0
+        self.dtype = np.uint16
+
+    def initParameters(self):
+        pass
+
+    def get_y_names(self):
+        y_names = []
+        for i in range(self.input_dim[self.axis]):
+            y_names.append(f'y{i}')
+        return y_names
+
+    def setUp(self):
+        self.initDefaultParameters()
+        self.initParameters()
+        self.op_type = 'unstack'
+        self.python_api = paddle.unstack
+        self.x = np.random.random(size=self.input_dim).astype(np.float32)
+        outs = np.split(self.x, self.input_dim[self.axis], self.axis)
+        new_shape = list(self.input_dim)
+        del new_shape[self.axis]
+        y_names = self.get_y_names()
+        tmp = []
+        tmp_names = []
+        for i in range(self.input_dim[self.axis]):
+            tmp.append(
+                (
+                    y_names[i],
+                    np.reshape(convert_float_to_uint16(outs[i]), new_shape),
+                )
+            )
+            tmp_names.append(y_names[i])
+
+        self.x = convert_float_to_uint16(self.x)
+        self.python_out_sig = tmp_names
+        self.inputs = {'X': self.x}
+        self.outputs = {'Y': tmp}
+        self.attrs = {'axis': self.axis, 'num': self.input_dim[self.axis]}
+
+    def test_check_output(self):
+        place = core.CUDAPlace(0)
+        self.check_output_with_place(place)
+
+    def test_check_grad(self):
+        with fluid.dygraph.guard():
+            x = paddle.to_tensor(self.inputs['X'])
+            x.stop_gradient = False
+            y = paddle.unstack(
+                x, axis=self.attrs['axis'], num=self.attrs['num']
+            )
+            dx = paddle.grad(y, x)[0].numpy()
+            dx_expected = convert_float_to_uint16(
+                np.ones(self.input_dim, np.float32)
+            )
+            np.testing.assert_array_equal(dx, dx_expected)
+
+
 class TestUnstackZeroInputOp(unittest.TestCase):
     def unstack_zero_input_static(self):