From 6c650445483d81748a16ffda075afd0943f7e1c7 Mon Sep 17 00:00:00 2001
From: yangjianfengo1 <125249383+yangjianfengo1@users.noreply.github.com>
Date: Wed, 8 Mar 2023 14:48:52 +0800
Subject: [PATCH] AMP tile_op & Test (#51193)

* tile_op

* fix bfloat16 x
---
 .../fluid/tests/unittests/test_tile_op.py     | 48 ++++++++++++++++++-
 python/paddle/tensor/manipulation.py          | 10 +++-
 2 files changed, 56 insertions(+), 2 deletions(-)

diff --git a/python/paddle/fluid/tests/unittests/test_tile_op.py b/python/paddle/fluid/tests/unittests/test_tile_op.py
index b95e8c4af7..c954cd7796 100644
--- a/python/paddle/fluid/tests/unittests/test_tile_op.py
+++ b/python/paddle/fluid/tests/unittests/test_tile_op.py
@@ -17,7 +17,7 @@ import unittest
 import gradient_checker
 import numpy as np
 from decorator_helper import prog_scope
-from eager_op_test import OpTest
+from eager_op_test import OpTest, convert_float_to_uint16
 
 import paddle
 import paddle.fluid as fluid
@@ -196,6 +196,52 @@ class TestTileOpInteger(OpTest):
         self.check_output()
 
 
+class TestTileOpFloat16(OpTest):
+    def setUp(self):
+        self.op_type = "tile"
+        self.dtype = np.float16
+        self.__class__.op_type = self.op_type
+        self.python_api = paddle.tile
+        self.inputs = {
+            'X': np.random.uniform(10, size=(100, 4, 5)).astype(self.dtype)
+        }
+        self.attrs = {'repeat_times': [2, 1, 4]}
+        output = np.tile(self.inputs['X'], (2, 1, 4))
+        self.outputs = {'Out': output}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+@unittest.skipIf(
+    not core.is_compiled_with_cuda()
+    or not core.is_bfloat16_supported(core.CUDAPlace(0)),
+    "core is not complied with CUDA and not support the bfloat16",
+)
+class TestWhereOpBFloat16(OpTest):
+    def setUp(self):
+        self.op_type = 'tile'
+        self.dtype = np.uint16
+        self.__class__.op_type = self.op_type
+        self.python_api = paddle.tile
+        x = np.random.uniform(10, size=(100, 4, 5)).astype(np.float32)
+        output = np.tile(x, (2, 1, 4))
+        self.inputs = {'X': convert_float_to_uint16(x)}
+        self.attrs = {'repeat_times': [2, 1, 4]}
+        self.outputs = {'Out': convert_float_to_uint16(output)}
+
+    def test_check_output(self):
+        place = core.CUDAPlace(0)
+        self.check_output_with_place(place)
+
+    def test_check_grad(self):
+        place = core.CUDAPlace(0)
+        self.check_grad_with_place(place, ['X'], 'Out')
+
+
 # Situation 5: input x is Bool
 class TestTileOpBoolean(OpTest):
     def setUp(self):
diff --git a/python/paddle/tensor/manipulation.py b/python/paddle/tensor/manipulation.py
index bf0a52e572..801fcf5b6e 100644
--- a/python/paddle/tensor/manipulation.py
+++ b/python/paddle/tensor/manipulation.py
@@ -3156,7 +3156,15 @@ def tile(x, repeat_times, name=None):
         check_variable_and_dtype(
             x,
             'x',
-            ['bool', 'float16', 'float32', 'float64', 'int32', 'int64'],
+            [
+                'bool',
+                'float16',
+                'bfloat16',
+                'float32',
+                'float64',
+                'int32',
+                'int64',
+            ],
             'tile',
         )
         if convert_dtype(x.dtype) == 'bool' and not x.stop_gradient:
-- 
GitLab