remove tmp fp32 var for gaussian_random (#46033)

Author: GuoxiaWang

python/paddle/fluid/initializer.py

@@ -374,68 +374,36 @@ def forward(self, var, block=None):
                                  ["uint16", "float16", "float32", "float64"],
                                  "guassian_random")
 
-        # to be compatible of fp16 initalizers
-        if var.dtype in [VarDesc.VarType.FP16, VarDesc.VarType.BF16]:
-            out_dtype = VarDesc.VarType.FP32
-            out_var = block.create_var(name=unique_name.generate(".".join(
-                ['normal_init', var.name, 'tmp'])),
-                                       shape=var.shape,
-                                       dtype=out_dtype,
-                                       type=VarDesc.VarType.LOD_TENSOR,
-                                       persistable=False)
-        else:
-            out_dtype = var.dtype
-            out_var = var
-
         if self._seed == 0:
             self._seed = block.program.random_seed
 
         if in_dygraph_mode():
             place = _current_expected_place()
             out_var = _C_ops.gaussian_random(var.shape, self._mean,
                                              self._std_dev, self._seed,
-                                             out_dtype, place)
-
-            if var.dtype in [VarDesc.VarType.FP16, VarDesc.VarType.BF16]:
-                var_tmp = _C_ops.cast(out_var, var.dtype)
-                var_tmp._share_underline_tensor_to(var)
-            else:
-                out_var._share_underline_tensor_to(var)
+                                             var.dtype, place)
+            out_var._share_underline_tensor_to(var)
             return None
 
         if _in_legacy_dygraph():
             out_var = _legacy_C_ops.gaussian_random(
-                'shape', var.shape, 'dtype', out_dtype, 'mean', self._mean,
+                'shape', var.shape, 'dtype', var.dtype, 'mean', self._mean,
                 'std', self._std_dev, 'seed', self._seed, 'use_mkldnn', False)
 
-            if var.dtype in [VarDesc.VarType.FP16, VarDesc.VarType.BF16]:
-                var_tmp = _legacy_C_ops.cast(out_var, 'in_dtype', out_var.dtype,
-                                             'out_dtype', var.dtype)
-                var_tmp._share_underline_tensor_to(var)
-            else:
-                out_var._share_underline_tensor_to(var)
+            out_var._share_underline_tensor_to(var)
             return None
         else:
             op = block.append_op(type="gaussian_random",
-                                 outputs={"Out": out_var},
+                                 outputs={"Out": var},
                                  attrs={
                                      "shape": var.shape,
-                                     "dtype": out_dtype,
+                                     "dtype": var.dtype,
                                      "mean": self._mean,
                                      "std": self._std_dev,
                                      "seed": self._seed,
                                      "use_mkldnn": False
                                  },
                                  stop_gradient=True)
-
-            if var.dtype in [VarDesc.VarType.FP16, VarDesc.VarType.BF16]:
-                block.append_op(type="cast",
-                                inputs={"X": out_var},
-                                outputs={"Out": var},
-                                attrs={
-                                    "in_dtype": out_var.dtype,
-                                    "out_dtype": var.dtype
-                                })
             var.op = op
             return op
 
@@ -695,7 +663,7 @@ def forward(self, var, block=None):
                                      outputs={"Out": out_var},
                                      attrs={
                                          "shape": out_var.shape,
-                                         "dtype": out_dtype,
+                                         "dtype": out_var.dtype,
                                          "mean": 0.0,
                                          "std": std,
                                          "seed": self._seed

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

@@ -245,7 +245,7 @@ def test_normal_initializer(self, dtype="float32"):
                                    name="param",
                                    initializer=initializer.NormalInitializer(
                                        2.3, 1.9, 123))
-        num_ops = 2 if (dtype == "float16" or dtype == "uint16") else 1
+        num_ops = 1
         self.assertEqual(len(block.ops), num_ops)
         init_op = block.ops[0]
         self.assertEqual(init_op.type, 'gaussian_random')
@@ -390,7 +390,6 @@ def test_xavier_initializer_fp16(self):
         """Test the Xavier initializer with float16
         """
         block = self.test_xavier_initializer_supplied_arguments("float16")
-        self.assertTrue(check_cast_op(block.ops[1]))
 
     def test_xavier_initializer_bf16(self):
         """Test the Xavier initializer with bfloat16
@@ -400,7 +399,6 @@ def test_xavier_initializer_bf16(self):
         self.assertEqual(len(block_uniform.ops), 1)
         block_gaussian = self.test_xavier_initializer_supplied_arguments(
             "uint16", False)
-        self.assertTrue(check_cast_op(block_gaussian.ops[1]))
 
 
 class TestMSRAInitializer(unittest.TestCase):

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

@@ -398,7 +398,7 @@ def test_normal_initializer(self, dtype="float32"):
                                    lod_level=0,
                                    name="param",
                                    initializer=initializer.Normal(2.3, 1.9))
-        num_ops = 2 if dtype in ["float16", "uint16"] else 1
+        num_ops = 1
         self.assertEqual(len(block.ops), num_ops)
         init_op = block.ops[0]
         self.assertEqual(init_op.type, 'gaussian_random')