[cherry-pick] implement autotune python API(42299) (#42301)

* implement autotune python API

* fix doc

* fix windows error

* fix doc and enable auto-tuning when config is None

* fix windows error

* fix doc

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

@@ -15,12 +15,14 @@
 from __future__ import print_function
 import unittest
 import numpy as np
-
+import tempfile
+import warnings
+import json
 import paddle
 import paddle.nn as nn
 from paddle.io import Dataset, DataLoader, BatchSampler, SequenceSampler
-from paddle.fluid.reader import set_autotune_config
 import sys
+import os
 
 
 class RandomDataset(Dataset):
@@ -51,12 +53,21 @@ class TestAutoTune(unittest.TestCase):
         self.dataset = RandomDataset(10)
 
     def test_dataloader_use_autotune(self):
-        set_autotune_config(True, 1)
+        paddle.incubate.autotune.set_config(
+            config={"dataloader": {
+                "enable": True,
+                "tuning_steps": 1,
+            }})
         loader = DataLoader(
             self.dataset, batch_size=self.batch_size, num_workers=0)
 
     def test_dataloader_disable_autotune(self):
-        set_autotune_config(False)
+        config = {"dataloader": {"enable": False, "tuning_steps": 1}}
+        tfile = tempfile.NamedTemporaryFile(mode="w+", delete=False)
+        json.dump(config, tfile)
+        tfile.close()
+        paddle.incubate.autotune.set_config(tfile.name)
+        os.remove(tfile.name)
         loader = DataLoader(
             self.dataset, batch_size=self.batch_size, num_workers=2)
         if (sys.platform == 'darwin' or sys.platform == 'win32'):
@@ -65,12 +76,28 @@ class TestAutoTune(unittest.TestCase):
             self.assertEqual(loader.num_workers, 2)
 
     def test_distributer_batch_sampler_autotune(self):
-        set_autotune_config(True, 1)
+        paddle.incubate.autotune.set_config(
+            config={"dataloader": {
+                "enable": True,
+                "tuning_steps": 1,
+            }})
         batch_sampler = paddle.io.DistributedBatchSampler(
             self.dataset, batch_size=self.batch_size)
         loader = DataLoader(
             self.dataset, batch_sampler=batch_sampler, num_workers=2)
 
 
+class TestAutoTuneAPI(unittest.TestCase):
+    def test_set_config_warnings(self):
+        with warnings.catch_warnings(record=True) as w:
+            config = {"kernel": {"enable": 1, "tuning_range": True}}
+            tfile = tempfile.NamedTemporaryFile(mode="w+", delete=False)
+            json.dump(config, tfile)
+            tfile.close()
+            paddle.incubate.autotune.set_config(tfile.name)
+            os.remove(tfile.name)
+            self.assertTrue(len(w) == 2)
+
+
 if __name__ == '__main__':
     unittest.main()

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

@@ -16,6 +16,10 @@ import paddle
 import unittest
 import numpy
 import paddle.nn.functional as F
+import tempfile
+import warnings
+import json
+import os
 
 
 class SimpleNet(paddle.nn.Layer):
@@ -41,10 +45,18 @@ class SimpleNet(paddle.nn.Layer):
 class LayoutAutoTune(unittest.TestCase):
     def use_autoune(self):
         if paddle.is_compiled_with_cuda():
-            paddle.fluid.core.enable_layout_autotune()
+            paddle.incubate.autotune.set_config(
+                config={"layout": {
+                    "enable": True
+                }})
             return paddle.fluid.core.use_layout_autotune()
         else:
-            paddle.fluid.core.disable_layout_autotune()
+            config = {"layout": {"enable": False}}
+            tfile = tempfile.NamedTemporaryFile(mode="w+", delete=False)
+            json.dump(config, tfile)
+            tfile.close()
+            paddle.incubate.autotune.set_config(tfile.name)
+            os.remove(tfile.name)
             return paddle.fluid.core.use_layout_autotune()
 
     def train(self, data_format):
@@ -103,7 +115,6 @@ class LayoutAutoTune(unittest.TestCase):
     def test_flatten_op_transposer(self):
         if not self.use_autoune():
             return
-        paddle.fluid.core.enable_layout_autotune()
         conv = paddle.nn.Conv2D(3, 8, (3, 3))
         flatten = paddle.nn.Flatten(start_axis=1, stop_axis=2)
         data = paddle.rand([1, 3, 16, 14])
@@ -119,5 +130,20 @@ class LayoutAutoTune(unittest.TestCase):
         self.assertEqual(out.shape, [1, 112, 12])
 
 
+class TestAutoTuneAPI(unittest.TestCase):
+    def test_set_config_warnings(self):
+        with warnings.catch_warnings(record=True) as w:
+            config = {"layout": {"enable": 1}}
+            # On linux, we can open the file again to read the content
+            # without closing the file, but on windows system, there is
+            # no permission to open it again without closing it.
+            tfile = tempfile.NamedTemporaryFile(mode="w+", delete=False)
+            json.dump(config, tfile)
+            tfile.close()
+            paddle.incubate.autotune.set_config(tfile.name)
+            os.remove(tfile.name)
+            self.assertTrue(len(w) == 1)
+
+
 if __name__ == '__main__':
     unittest.main()

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

@@ -15,6 +15,10 @@
 import paddle
 import unittest
 import numpy as np
+import tempfile
+import warnings
+import json
+import os
 
 
 class SimpleNet(paddle.nn.Layer):
@@ -73,10 +77,13 @@ class TestAutoTune(unittest.TestCase):
         return expected_res
 
     def test_autotune(self):
-        paddle.fluid.core.disable_autotune()
+        paddle.incubate.autotune.set_config(
+            config={"kernel": {
+                "enable": False
+            }})
         self.assertEqual(self.get_flags("FLAGS_use_autotune"), False)
 
-        paddle.fluid.core.enable_autotune()
+        paddle.incubate.autotune.set_config(config={"kernel": {"enable": True}})
         self.assertEqual(self.get_flags("FLAGS_use_autotune"), True)
 
     def check_status(self, expected_res):
@@ -93,10 +100,16 @@ class TestDygraphAutoTuneStatus(TestAutoTune):
     def run_program(self, enable_autotune):
         self.set_flags(enable_autotune)
         if enable_autotune:
-            paddle.fluid.core.enable_autotune()
+            paddle.incubate.autotune.set_config(
+                config={"kernel": {
+                    "enable": True,
+                    "tuning_range": [1, 2]
+                }})
         else:
-            paddle.fluid.core.disable_autotune()
-        paddle.fluid.core.set_autotune_range(1, 2)
+            paddle.incubate.autotune.set_config(
+                config={"kernel": {
+                    "enable": False
+                }})
         x_var = paddle.uniform((1, 1, 8, 8), dtype='float32', min=-1., max=1.)
         net = SimpleNet()
         for i in range(3):
@@ -141,10 +154,18 @@ class TestStaticAutoTuneStatus(TestAutoTune):
 
         self.set_flags(enable_autotune)
         if enable_autotune:
-            paddle.fluid.core.enable_autotune()
+            config = {"kernel": {"enable": True, "tuning_range": [1, 2]}}
+            tfile = tempfile.NamedTemporaryFile(mode="w+", delete=False)
+            json.dump(config, tfile)
+            tfile.close()
+            paddle.incubate.autotune.set_config(tfile.name)
+            os.remove(tfile.name)
         else:
-            paddle.fluid.core.disable_autotune()
-        paddle.fluid.core.set_autotune_range(1, 2)
+            paddle.incubate.autotune.set_config(
+                config={"kernel": {
+                    "enable": False,
+                    "tuning_range": [1, 2]
+                }})
 
         for i in range(3):
             exe.run(program=main_program, feed={'X': x}, fetch_list=[loss])
@@ -166,5 +187,22 @@ class TestStaticAutoTuneStatus(TestAutoTune):
         self.func_disable_autotune()
 
 
+class TestAutoTuneAPI(unittest.TestCase):
+    def test_set_config_warnings(self):
+        with warnings.catch_warnings(record=True) as w:
+            config = {"kernel": {"enable": 1, "tuning_range": 1}}
+            tfile = tempfile.NamedTemporaryFile(mode="w+", delete=False)
+            json.dump(config, tfile)
+            tfile.close()
+            paddle.incubate.autotune.set_config(tfile.name)
+            os.remove(tfile.name)
+            self.assertTrue(len(w) == 2)
+
+    def test_set_config_attr(self):
+        paddle.incubate.autotune.set_config(config=None)
+        self.assertEqual(
+            paddle.get_flags("FLAGS_use_autotune")["FLAGS_use_autotune"], True)
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/incubate/__init__.py

@@ -29,6 +29,7 @@ from .tensor import segment_max
 from .tensor import segment_min
 from .passes import fuse_resnet_unit_pass
 import paddle.incubate.autograd
+import paddle.incubate.autotune
 
 from . import nn  #noqa: F401
 

python/paddle/incubate/autotune.py

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+#     http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import json
+import warnings
+from paddle.fluid import core
+
+__all__ = ['set_config']
+
+
+def set_config(config=None):
+    r"""
+    Set the configuration for kernel, layout and dataloader auto-tuning.
+
+    1. kernel: When it is enable, exhaustive search method will be used to select
+    and cache the best algorithm for the operator in the tuning iteration. Tuning
+    parameters are as follows:
+
+    - enable(bool): Whether to enable kernel tuning.
+    - tuning_range(list): Start and end iteration for auto-tuning. Default: [1, 10].
+
+    2. layout: When it is enable, the best data layout such as NCHW or NHWC will be
+    determined based on the device and data type. When the origin layout setting is
+    not best, layout transformation will be automaticly performed to improve model
+    performance. Layout auto-tuning only supports dygraph mode currently. Tuning
+    parameters are as follows:
+
+    - enable(bool): Whether to enable layout tuning.
+
+    3. dataloader: When it is enable, the best num_workers will be selected to replace
+    the origin dataloader setting. Tuning parameters are as follows:
+
+    - enable(bool): Whether to enable dataloader tuning.
+
+    Args:
+        config (dict|str|None, optional): Configuration for auto-tuning. If it is a
+            dictionary, the key is the tuning type, and the value is a dictionary
+            of the corresponding tuning parameters. If it is a string, the path of
+            a json file will be specified and the tuning configuration will be set
+            by the the json file. Default: None, auto-tuning for kernel, layout and
+            dataloader will be enabled.
+
+    Examples:
+        .. code-block:: python
+            :name: auto-tuning
+
+            import paddle
+            import json
+
+            # config is a dict.
+            config = {
+                "kernel": {
+                    "enable": True,
+                    "tuning_range": [1, 5],
+                },
+                "layout": {
+                    "enable": True,
+                },
+                "dataloader": {
+                    "enable": True,
+                }
+            }
+            paddle.incubate.autotune.set_config(config)
+
+            # config is the path of json file.
+            config_json = json.dumps(config)
+            with open('config.json', 'w') as json_file:
+                json_file.write(config_json)
+            paddle.incubate.autotune.set_config('config.json')
+
+    """
+    if config is None:
+        core.enable_autotune()
+        core.enable_layout_autotune()
+        paddle.fluid.reader.set_autotune_config(use_autotune=True)
+        return
+
+    config_dict = {}
+    if isinstance(config, dict):
+        config_dict = config
+    elif isinstance(config, str):
+        try:
+            with open(config, 'r') as filehandle:
+                config_dict = json.load(filehandle)
+        except Exception as e:
+            print('Load config error: {}'.format(e))
+            warnings.warn("Use default configuration for auto-tuning.")
+
+    if "kernel" in config_dict:
+        kernel_config = config_dict["kernel"]
+        if "enable" in kernel_config:
+            if isinstance(kernel_config['enable'], bool):
+                if kernel_config['enable']:
+                    core.enable_autotune()
+                else:
+                    core.disable_autotune()
+            else:
+                warnings.warn(
+                    "The auto-tuning configuration of the kernel is incorrect."
+                    "The `enable` should be bool. Use default parameter instead."
+                )
+        if "tuning_range" in kernel_config:
+            if isinstance(kernel_config['tuning_range'], list):
+                tuning_range = kernel_config['tuning_range']
+                assert len(tuning_range) == 2
+                core.set_autotune_range(tuning_range[0], tuning_range[1])
+            else:
+                warnings.warn(
+                    "The auto-tuning configuration of the kernel is incorrect."
+                    "The `tuning_range` should be list. Use default parameter instead."
+                )
+    if "layout" in config_dict:
+        layout_config = config_dict["layout"]
+        if "enable" in layout_config:
+            if isinstance(layout_config['enable'], bool):
+                if layout_config['enable']:
+                    core.enable_layout_autotune()
+                else:
+                    core.disable_layout_autotune()
+            else:
+                warnings.warn(
+                    "The auto-tuning configuration of the layout is incorrect."
+                    "The `enable` should be bool. Use default parameter instead."
+                )
+    if "dataloader" in config_dict:
+        dataloader_config = config_dict["dataloader"]
+        use_autoune = False
+        if "enable" in dataloader_config:
+            if isinstance(dataloader_config['enable'], bool):
+                use_autoune = dataloader_config['enable']
+            else:
+                warnings.warn(
+                    "The auto-tuning configuration of the dataloader is incorrect."
+                    "The `enable` should be bool. Use default parameter instead."
+                )
+        if "tuning_steps" in dataloader_config:
+            if isinstance(dataloader_config['tuning_steps'], int):
+                paddle.fluid.reader.set_autotune_config(
+                    use_autoune, dataloader_config['tuning_steps'])
+            else:
+                warnings.warn(
+                    "The auto-tuning configuration of the dataloader is incorrect."
+                    "The `tuning_steps` should be int. Use default parameter instead."
+                )
+                paddle.fluid.reader.set_autotune_config(use_autoune)