• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2020 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_TPU_KERNELS_TPU_PROGRAM_GROUP_H_
16 #define TENSORFLOW_CORE_TPU_KERNELS_TPU_PROGRAM_GROUP_H_
17 
18 #include <memory>
19 #include <vector>
20 
21 #include "absl/types/optional.h"
22 #include "tensorflow/compiler/tf2xla/host_compute_metadata.pb.h"
23 #include "tensorflow/compiler/tf2xla/xla_compiler.h"
24 #include "tensorflow/compiler/xla/client/compile_only_client.h"
25 #include "tensorflow/compiler/xla/service/computation_placer.h"
26 #include "tensorflow/compiler/xla/service/hlo.pb.h"
27 #include "tensorflow/compiler/xrt/xrt.pb.h"
28 #include "tensorflow/core/platform/macros.h"
29 #include "tensorflow/core/tpu/kernels/tpu_compile_op_support.h"
30 #include "tensorflow/core/tpu/kernels/tpu_executable_info.pb.h"
31 #include "tensorflow/core/tpu/kernels/tpu_mesh_state_interface.h"
32 #include "tensorflow/core/tpu/kernels/tpu_program_group_interface.h"
33 #include "tensorflow/core/tpu/tpu_ops_c_api.h"
34 #include "tensorflow/stream_executor/tpu/tpu_platform_interface.h"
35 
36 namespace tensorflow {
37 namespace tpu {
38 
39 class TpuAotCompilationOptions : public xla::AotCompilationOptions {
40  public:
TpuAotCompilationOptions(int64 replica_count)41   explicit TpuAotCompilationOptions(int64 replica_count)
42       : num_cores_(0), replica_count_(replica_count) {}
43 
44   // Returns the ID of the platform to which these options apply.
PlatformId()45   se::Platform::Id PlatformId() const override {
46     LOG(FATAL) << "Not implemented.";
47     return nullptr;
48   };
49 
set_num_cores(int64 tpu_cores)50   void set_num_cores(int64 tpu_cores) { num_cores_ = tpu_cores; }
replica_count()51   int64 replica_count() const override { return replica_count_; }
num_cores()52   int64 num_cores() const override { return num_cores_; }
53 
set_allow_separate_sharding_programs(bool allow)54   void set_allow_separate_sharding_programs(bool allow) {
55     allow_separate_sharding_programs_ = allow;
56   }
allow_separate_sharding_programs()57   bool allow_separate_sharding_programs() const {
58     return allow_separate_sharding_programs_;
59   }
60 
61   const std::vector<xla::HloModuleConfig::ShardableValueUpdatePair>
shardable_value_update_pairs()62   shardable_value_update_pairs() const {
63     return shardable_value_update_pairs_;
64   }
set_shardable_value_update_pairs(std::vector<xla::HloModuleConfig::ShardableValueUpdatePair> pairs)65   void set_shardable_value_update_pairs(
66       std::vector<xla::HloModuleConfig::ShardableValueUpdatePair> pairs) {
67     shardable_value_update_pairs_ = std::move(pairs);
68   }
69 
70  private:
71   int64 num_cores_;
72   int64 replica_count_;
73 
74   // Whether to allow the compiler to create separte sharding and unsharding
75   // programs, and modify the original program's input/output sharded size. This
76   // is used for XLA-chosen sharding on parameters without an on-device loop:
77   // the caller can invoke sharding first, then (repeatedly) invoke the sharded
78   // main program, and finally invoke the unsharding program when it needs the
79   // full output.
80   bool allow_separate_sharding_programs_ = false;
81 
82   // The list of input/output pairs in the main program that could be sharded.
83   std::vector<xla::HloModuleConfig::ShardableValueUpdatePair>
84       shardable_value_update_pairs_;
85 };
86 
87 class TpuProgramGroup : public TpuProgramGroupInterface {
88  public:
89   using Status = ::stream_executor::port::Status;
90 
91   // Compiles Mlir or TF function computation by lowering into HLO IR and
92   // returns TPU programs ready for execution.
93   static Status CompileAndBuild(
94       const TpuCompilationRequestProto& compilation_request,
95       const XLA_TpuMeshState* mesh_state,
96       TpuProgramGroupInterface* tpu_program_group_interface);
97 
98   // Compiles HLO IR and returns TPU programs ready for execution.
99   static Status CompileAndBuild(
100       const xrt::XLAComputation& xrt_computation_proto,
101       const XLA_TpuMeshState* mesh_state,
102       TpuProgramGroupInterface* tpu_program_group_interface);
103 
104   // Initializes `TpuProgramGroup` object with `xla_tpu_programs`.
105   void Initialize(absl::Span<XLA_TpuProgram* const> xla_tpu_programs);
106 
107   TpuProgramGroup() = default;
108   TpuProgramGroup(TpuProgramGroup&& other);
109   TpuProgramGroup& operator=(TpuProgramGroup&&) = delete;
110 
111   bool has_sharding_program() const override;
112 
113   size_t program_count() const override;
114 
115   int64_t program_size() const override;
116 
117   bool LogProgramMemorySummary() override;
118 
119   void UnloadAndDestroyPrograms() override;
120 
121   Status LogCompilationStats(const TpuCompilationCacheKey& key,
122                              absl::Duration duration) override;
123 
124   const std::vector<bool>& may_modify_variables_list() const override;
125   void set_may_modify_variables(const std::vector<bool>& may_modify_variables);
126   bool may_modify_variables(int index) const override;
127 
128   const std::vector<XLA_TpuProgram*>& tpu_programs() const;
129   std::vector<XLA_TpuProgram*> tpu_programs(TpuProgramShardingType type) const;
130   const XLA_TpuProgram* tpu_program(int index) const;
131   void set_tpu_programs(absl::Span<XLA_TpuProgram* const> tpu_programs);
132 
133   const TPUExecutableInfoProto& executable_info(int index) const;
134 
135   const TPUHostTransferInfoProto& host_transfer_info(int index) const;
136   void set_hlo_metadatas(absl::Span<const xla::HloProto> hlo_metadatas);
137   const xla::HloProto* hlo_metadata(int index) const;
138   absl::Span<const xla::HloProto* const> hlo_metadatas() const override;
139 
140   // Deserializes `GetTpuProgramResponse` protos from remote cache.
141   Status DeserializeFromRpcResponseProtos(
142       const std::vector<TpuSerializedProto>& rpc_response_protos);
143 
144   // Serializes executable proto from the TPU program for the given core
145   // `index`.
146   Status SerializeExecutable(int index,
147                              TpuExecutableSerializedProto* executable) const;
148 
149   // Serializes compiler metadata of the TPU program for the given core `index`.
150   Status SerializeCompilerMetadata(
151       int index, CompilerMetadataSerializedProto* compiler_metadata) const;
152 
153   // Serializes host compute metadata of the TPU program for the given core
154   // `index`.
155   Status SerializeHostComputeMetadata(
156       int index,
157       HostComputeMetadataSerializedProto* host_compute_metadata) const;
158 
159  private:
160   TPUExecutableInfoProto ConstructExecutableInfo(
161       const XLA_TpuProgram* tpu_program);
162   TPUHostTransferInfoProto ConstructHostTransferInfo(
163       const XLA_TpuProgram* tpu_program);
164   xla::HloProto ConstructHloMetadata(const XLA_TpuProgram* tpu_program);
165 
166   // Update `hlo_metadatas__ptrs_` array from `hlo_metadatas_`. This needs to be
167   // called on `hlo_metadatas_` change(s).
168   void RefreshHloMetadatasPtrs();
169 
170   std::vector<bool> may_modify_variables_;
171 
172   std::vector<XLA_TpuProgram*> tpu_programs_;  // Not owned.
173   std::vector<TPUExecutableInfoProto> executable_infos_;
174   std::vector<TPUHostTransferInfoProto> host_transfer_infos_;
175 
176   // To be consistent with the TpuProgramGroupInterface::hlo_metadatas()
177   // signature, we store HloProto values in hlo_metadatas_ when
178   // set_hlo_metadata(...) is called, and return their pointers from
179   // hlo_metadatas_ptrs_ when hlo_metadatas() is called. hlo_metadata_ptrs_ is
180   // refreshed whenever hlo_metadatas_ is set or the object is moved.
181   std::vector<xla::HloProto> hlo_metadatas_;  // Owned.
182   std::vector<const xla::HloProto*> hlo_metadatas_ptrs_;
183 
184   TF_DISALLOW_COPY_AND_ASSIGN(TpuProgramGroup);
185 };
186 
187 }  // namespace tpu
188 }  // namespace tensorflow
189 
190 #endif  // TENSORFLOW_CORE_TPU_KERNELS_TPU_PROGRAM_GROUP_H_
191