1 | /* Copyright 2018 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 | #ifndef TENSORFLOW_CORE_COMMON_RUNTIME_SCOPED_ALLOCATOR_MGR_H_ |
16 | #define TENSORFLOW_CORE_COMMON_RUNTIME_SCOPED_ALLOCATOR_MGR_H_ |
17 | |
18 | #include <string> |
19 | #include <unordered_map> |
20 | |
21 | #include "tensorflow/core/common_runtime/scoped_allocator.h" |
22 | #include "tensorflow/core/lib/core/refcount.h" |
23 | #include "tensorflow/core/lib/core/status.h" |
24 | #include "tensorflow/core/platform/mutex.h" |
25 | |
26 | namespace tensorflow { |
27 | class ScopedAllocatorMgr; |
28 | |
29 | // At most one of these exists per <device, step_id> pair. |
30 | // A Ref is held by every ScopedAllocator and also by the ScopedAllocatorMgr. |
31 | class ScopedAllocatorContainer : public core::RefCounted { |
32 | public: |
33 | // Establishes a reachable ScopedAllocator. |
34 | Status AddScopedAllocator( |
35 | const Tensor& backing_tensor, int32_t scope_id, |
36 | const std::string& scope_name, |
37 | const gtl::ArraySlice<ScopedAllocator::Field>& fields, |
38 | int32_t expected_call_count); |
39 | |
40 | ScopedAllocatorInstance* GetInstance(int32_t scope_id); |
41 | ScopedAllocator* GetAllocator(int32_t scope_id); |
42 | |
43 | // Retire the scope_id. |
44 | void Drop(int32_t scope_id, ScopedAllocator* sa); |
45 | |
46 | protected: |
47 | friend class ScopedAllocatorMgr; |
48 | ScopedAllocatorContainer(const ScopedAllocatorMgr* mgr, int64_t step_id) |
49 | : mgr_(mgr), step_id_(step_id) {} |
50 | ~ScopedAllocatorContainer(); |
51 | |
52 | private: |
53 | const ScopedAllocatorMgr* mgr_; |
54 | int64_t step_id_; |
55 | mutex mu_; |
56 | struct SAField { |
57 | int32 field_index; |
58 | union { |
59 | ScopedAllocator* scoped_allocator; |
60 | ScopedAllocatorInstance* instance; |
61 | }; |
62 | SAField(int32_t fi, ScopedAllocatorInstance* sai) |
63 | : field_index(fi), instance(sai) {} |
64 | SAField(int32_t fi, ScopedAllocator* sa) |
65 | : field_index(fi), scoped_allocator(sa) {} |
66 | SAField() |
67 | : field_index(ScopedAllocator::kBackingIndex), |
68 | scoped_allocator(nullptr) {} |
69 | }; |
70 | std::unordered_map<int32, SAField> allocators_ TF_GUARDED_BY(mu_); |
71 | }; |
72 | |
73 | // At most one of these exists per device. |
74 | class ScopedAllocatorMgr { |
75 | public: |
76 | explicit ScopedAllocatorMgr(const std::string& device_name) |
77 | : device_name_(device_name) {} |
78 | ~ScopedAllocatorMgr(); |
79 | |
80 | ScopedAllocatorContainer* GetContainer(int64_t step_id); |
81 | |
82 | // Establishes a reachable ScopedAllocator. |
83 | Status AddScopedAllocator( |
84 | const Tensor& backing_tensor, int64_t step_id, int32_t scope_id, |
85 | const std::string& scope_name, |
86 | const gtl::ArraySlice<ScopedAllocator::Field>& fields, |
87 | int32_t expected_call_count); |
88 | |
89 | void Cleanup(int64_t step_id); |
90 | |
91 | // Populate the bytes and offset members of Field. Instance allocaters get |
92 | // consecutive scope_id values following that of the base ScopedAllocator. |
93 | // Returns the total number of bytes required to be allocated in the |
94 | // backing tensor, for convenience. (The same value can be obtained |
95 | // by summing offset and bytes in the last field.) |
96 | static size_t PopulateFields(int32_t scope_id, |
97 | const gtl::ArraySlice<TensorShape>& shapes, |
98 | const DataType dtype, |
99 | std::vector<ScopedAllocator::Field>* fields); |
100 | |
101 | const std::string& device_name() const { return device_name_; } |
102 | |
103 | private: |
104 | std::string device_name_; |
105 | mutex mu_; |
106 | std::unordered_map<int64_t, ScopedAllocatorContainer*> per_step_map_ |
107 | TF_GUARDED_BY(mu_); |
108 | }; |
109 | |
110 | } // namespace tensorflow |
111 | #endif // TENSORFLOW_CORE_COMMON_RUNTIME_SCOPED_ALLOCATOR_MGR_H_ |
112 | |