• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2017 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 #ifndef TENSORFLOW_COMPILER_TF2XLA_XLA_RESOURCE_H_
17 #define TENSORFLOW_COMPILER_TF2XLA_XLA_RESOURCE_H_
18 
19 #include <memory>
20 
21 #include "absl/strings/string_view.h"
22 #include "tensorflow/compiler/xla/client/xla_builder.h"
23 #include "tensorflow/compiler/xla/xla_data.pb.h"
24 #include "tensorflow/core/framework/tensor_shape.h"
25 #include "tensorflow/core/framework/types.pb.h"
26 #include "tensorflow/core/lib/core/status.h"
27 
28 namespace tensorflow {
29 
30 // Represents a resource, such as a Variable or TensorArray.
31 class XlaResource {
32  public:
33   enum Kind {
34     kInvalid,
35     kVariable,
36     kTensorArray,
37     kStack,
38   };
39   static absl::string_view KindToString(Kind kind);
40 
41   // Creates a new Stack resource.
42   static std::unique_ptr<XlaResource> CreateStack(string name, DataType type,
43                                                   int64 max_size);
44 
45   // Creates a new TensorArray resource.
46   static std::unique_ptr<XlaResource> CreateTensorArray(
47       string name, DataType type, TensorShape shape, xla::XlaOp initial_value,
48       int64 max_array_size);
49 
50   XlaResource(Kind kind, int arg_num, string name, DataType type,
51               TensorShape shape, const xla::XlaOp& initial_value,
52               int64 max_array_size,
53               const std::set<string>& tensor_array_gradients,
54               bool tensor_array_multiple_writes_aggregate);
55 
56   XlaResource(const XlaResource&) = delete;
57   XlaResource(XlaResource&&) = delete;
58   XlaResource& operator=(const XlaResource&) = delete;
59   XlaResource& operator=(XlaResource&&) = delete;
60 
kind()61   Kind kind() const { return kind_; }
62 
63   // If this resource is visible externally to the computation, what was its
64   // argument number?
65   // < 0 means "not visible externally".
arg_num()66   int arg_num() const { return arg_num_; }
67 
68   // A descriptive name for the resource, used in error messages.
name()69   const string& name() const { return name_; }
70 
71   // Current type and value of the resource. Uninitialized resources are
72   // represented by a default (zero) handle and type DT_INVALID.
73   // While the type of a resource is notionally fixed during execution, when
74   // a resource is first initialized we do not yet know its type, so we keep
75   // track of its type dynamically.
type()76   DataType type() const { return type_; }
77 
78   // Shape of the resource. For an uninitialized resource, this is ignored.
79   // For a Variable, this is the shape of the value. For a TensorArray or Stack
80   // this is the shape of each entry in the TensorArray/Stack.
shape()81   const TensorShape& shape() const { return shape_; }
82 
value()83   const xla::XlaOp& value() const { return value_; }
84 
85   // Value of the resource at computation entry. Used to detect which
86   // variables have new values that need to be written back.
initial_value()87   const xla::XlaOp& initial_value() const { return initial_value_; }
88 
89   // An xla shape that indicates how this resource variable is represented on
90   // device.
representation_shape()91   const absl::optional<xla::Shape>& representation_shape() const {
92     return representation_shape_;
93   }
94 
95   // A variable is initialized if it has a value.
initialized()96   bool initialized() const { return value_.valid(); }
97 
98   // Sets the type and shape of the resource. The type and shape of a resource
99   // must not change once the variable has been initialized.
100   Status SetTypeAndShape(DataType type, const TensorShape& shape);
101 
102   // Sets the current value of the resource. Returns an error if the type is not
103   // set to a valid value.
104   Status SetValue(const xla::XlaOp& value);
105 
106   // Sets the current value of the resource to an all-zero value.
107   Status SetZeroValue(xla::XlaBuilder* builder);
108 
109   // Sets the representational shape of the resource on device.
SetRepresentationShape(const xla::Shape & shape)110   void SetRepresentationShape(const xla::Shape& shape) {
111     representation_shape_ = absl::make_optional(shape);
112   }
113 
114   // Looks up the gradient for `source`, or creates it if it does not already
115   // exist. The call target must be an initialized TensorArray resource. A
116   // TensorArray can have multiple named gradients; see the operator
117   // documentation for TensorArrayGradV3 for details.
118   Status GetOrCreateTensorArrayGradient(const string& source,
119                                         xla::XlaBuilder* builder,
120                                         XlaResource** gradient_out);
121 
122   // Packs a resource into a single XLA value `pack`, suitable for use as
123   // an XlaCompiler::Argument. For non-TensorArrays or TensorArrays without
124   // gradients, sets `*pack` to `value`.
125   // For TensorArrays with gradients, packs the value and its gradient values in
126   // a tuple; the gradients values are packed in order by source name.
127   Status Pack(xla::XlaOp* pack, xla::XlaBuilder* builder) const;
128 
129   // Updates the resource with values from `pack`. If `gradient_sources` is
130   // non-empty, treats `pack` as a tuple that represents a TensorArray and
131   // its gradients, and unpacks and updates the gradient resources.
132   // If `reset_initial_values` is true, sets the initial_values as well as the
133   // values.
134   // Opposite of Pack().
135   Status SetFromPack(const std::set<string>& gradient_sources,
136                      const xla::XlaOp& pack, xla::XlaBuilder* builder);
137 
138   // TensorArray and Stack specific fields
139   // TODO(phawkins): refactor this code to use subclasses, rather than putting
140   // kind-specific fields in XlaResource.
141 
142   // 'max_array_size' stores the expected size of the TensorArray or Stack.
143   // We need to store this since sometimes TensorArrays must be initialized
144   // lazily since we do not know the element shape at construction time.
145   // Used by both TensorArrays and Stacks.
max_array_size()146   int64 max_array_size() const { return max_array_size_; }
set_max_array_size(int64 size)147   void set_max_array_size(int64 size) { max_array_size_ = size; }
148 
tensor_array_multiple_writes_aggregate()149   bool tensor_array_multiple_writes_aggregate() const {
150     return tensor_array_multiple_writes_aggregate_;
151   }
152 
153   // 'tensor_array_gradient' is a map from TensorArrayGradV3 'source' attributes
154   // to an XlaResource containing the gradient TensorArrays. We store a pointer
155   // here since there should only be one gradient TensorArray per 'source'
156   // string, irrespective of the number of calls to TensorArrayGrad. The map
157   // is ordered since values are packed into tuples by Pack() sorted by name
158   // order.
tensor_array_gradients()159   const std::map<string, std::unique_ptr<XlaResource>>& tensor_array_gradients()
160       const {
161     return tensor_array_gradients_;
162   }
163 
164  private:
165   const Kind kind_;
166   const int arg_num_;
167   const string name_;
168 
169   DataType type_;
170   TensorShape shape_;
171   xla::XlaOp value_;
172   xla::XlaOp initial_value_;
173 
174   // An xla shape that indicates how this resource variable is represented on
175   // device.
176   absl::optional<xla::Shape> representation_shape_;
177 
178   int64 max_array_size_ = -1;
179   bool tensor_array_multiple_writes_aggregate_ = false;
180 
181   std::map<string, std::unique_ptr<XlaResource>> tensor_array_gradients_;
182 };
183 
184 }  // namespace tensorflow
185 
186 #endif  // TENSORFLOW_COMPILER_TF2XLA_XLA_RESOURCE_H_
187