1# Copyright 2015 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================== 15 16"""Class to represent a device.""" 17 18from tensorflow.python import tf2 19from tensorflow.python.framework import device_spec 20 21if tf2.enabled(): 22 DeviceSpec = device_spec.DeviceSpecV2 23else: 24 DeviceSpec = device_spec.DeviceSpecV1 25 26 27def check_valid(spec): 28 """Check that a device spec is valid. 29 30 Args: 31 spec: a string. 32 33 Raises: 34 An exception if the spec is invalid. 35 """ 36 # Construct a DeviceSpec. It will assert a failure if spec is invalid. 37 DeviceSpec.from_string(spec) 38 39 40def is_device_spec(obj): 41 """Abstract away the fact that DeviceSpecV2 is the base class.""" 42 return isinstance(obj, device_spec.DeviceSpecV2) 43 44 45def canonical_name(device): 46 """Returns a canonical name for the given `DeviceSpec` or device name.""" 47 if device is None: 48 return "" 49 if is_device_spec(device): 50 return device.to_string() 51 else: 52 device = DeviceSpec.from_string(device) 53 return device.to_string() 54 55 56# Performance caches 57_cached_mergers = {} 58_string_merge_cache = {} 59 60 61def merge_device(spec): 62 """Returns a device function that merges devices specifications. 63 64 This can be used to merge partial specifications of devices. The 65 innermost setting for a device field takes precedence. For example: 66 67 with tf.device(merge_device("/device:GPU:0")) 68 # Nodes created here have device "/device:GPU:0" 69 with tf.device(merge_device("/job:worker")): 70 # Nodes created here have device "/job:worker/device:GPU:0" 71 with tf.device(merge_device("/device:CPU:0")): 72 # Nodes created here have device "/job:worker/device:CPU:0" 73 with tf.device(merge_device("/job:ps")): 74 # Nodes created here have device "/job:ps/device:CPU:0" 75 76 Args: 77 spec: A `DeviceSpec` or a device spec string (partially) describing the 78 device that should be used for all nodes created in the scope of 79 the returned device function's with block. 80 81 Returns: 82 A MergeDevice object with the above-described behavior. 83 84 Raises: 85 ValueError: if the spec was not valid. 86 """ 87 88 if isinstance(spec, MergeDevice): 89 return spec 90 91 merger = _cached_mergers.get(spec) 92 if merger: 93 return merger 94 merger = MergeDevice(spec) 95 # No locking needed, since updates are stateless. 96 _cached_mergers[spec] = merger 97 return merger 98 99 100class MergeDevice(object): 101 """Wraps a device specification (DeviceSpec or str) with merge functionality. 102 103 When called, this class will merge a node_def with its own spec. It also 104 exposes a `shortcut_string_merge` method which can significantly improve 105 performance of device placement. 106 """ 107 108 __slots__ = ["_spec"] 109 110 def __init__(self, spec): 111 if isinstance(spec, device_spec.DeviceSpecV2): 112 self._spec = spec 113 elif isinstance(spec, device_spec.DeviceSpecV1): 114 # Capture a snapshot of spec. 115 self._spec = spec.__class__.from_string(spec.to_string()) 116 else: 117 self._spec = DeviceSpec.from_string(spec) 118 119 def __call__(self, node_def): 120 # In general a user may create a device function which takes into account 121 # arbitrary properties of an op. (For instance dynamically placing ops based 122 # on type.) So even though the standard DeviceSpec route only uses the 123 # device attribute, we take an entire node_def to maintain a consistent 124 # signature with general device functions. 125 current_device = DeviceSpec.from_string(node_def.device or "") 126 return self._spec.make_merged_spec(current_device) 127 128 def shortcut_string_merge(self, node_def): 129 """Merge a node def without materializing a full DeviceSpec object. 130 131 Often a device merge is invoked in order to generate a string which can be 132 passed into the c api. In such a case, we can cache the 133 node_def.device -> merge_result_string 134 135 map, and in most cases avoid: 136 - Materializing a copy of self._spec (In the case of DeviceSpecV1) 137 - Materializing a DeviceSpec for node_def.device 138 - A DeviceSpec.merge_from invocation 139 140 In practice the cache hit rate for this function is very high, because the 141 number of invocations when iterating through the device stack is much 142 larger than the number of devices. 143 144 Args: 145 node_def: An Operation (or Operation-like) to merge device constraints 146 with self._spec 147 148 Returns: 149 A string containing the merged device specification. 150 """ 151 device = node_def.device or "" 152 153 merge_key = (self._spec, device) 154 result = _string_merge_cache.get(merge_key) 155 if result is None: 156 # This update is not atomic, however because the merge is stateless 157 # we don't need to lock when updating the cache. 158 result = self.__call__(node_def).to_string() 159 _string_merge_cache[merge_key] = result 160 161 return result 162 163 def __repr__(self): 164 return "{} (spec: {})".format( 165 super(MergeDevice, self).__repr__(), self._spec.to_string()) 166 167 @property 168 def is_null_merge(self): 169 """Indicate whether the wrapped spec is empty. 170 171 In the degenerate case where self._spec is an empty specification, a caller 172 may wish to skip a merge step entirely. (However this class does not have 173 enough information to make that determination.) 174 175 Returns: 176 A boolean indicating whether a device merge will be trivial. 177 """ 178 return not bool(self._spec.to_string()) 179