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 EXPERIMENTAL_BRAIN_TPU_1VM_MINIEXECUTOR_TPU_MESH_STATE_INTERFACE_H_ 16 #define EXPERIMENTAL_BRAIN_TPU_1VM_MINIEXECUTOR_TPU_MESH_STATE_INTERFACE_H_ 17 18 #include <string> 19 20 #include "tensorflow/core/framework/resource_mgr.h" 21 #include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h" 22 #include "tensorflow/core/tpu/tpu_api.h" 23 #include "tensorflow/core/tpu/tpu_ops_c_api.h" 24 25 namespace tensorflow { 26 27 class TpuMeshCommonState; 28 29 namespace tpu { 30 31 const char kTpuMeshStateInterfaceResourceName[] = "tpu_mesh_common_state"; 32 33 class TpuMeshStateInterface : public tensorflow::ResourceBase { 34 public: TpuMeshStateInterface(XLA_TpuMeshState * handle)35 explicit TpuMeshStateInterface(XLA_TpuMeshState* handle) 36 : mesh_state_(handle) { 37 } 38 ~TpuMeshStateInterface()39 ~TpuMeshStateInterface() override { 40 if (mesh_state_ != nullptr) { 41 OpsApiFn()->TpuMeshState_FreeFn(mesh_state_); 42 } 43 } 44 Create()45 static TpuMeshStateInterface* Create() { 46 XLA_TpuMeshState* state = nullptr; 47 if (OpsApiFn()->TpuMeshState_CreateFn != nullptr) { 48 state = OpsApiFn()->TpuMeshState_CreateFn(); 49 } 50 return new TpuMeshStateInterface(state); 51 } 52 data()53 const XLA_TpuMeshState* data() const { return mesh_state_; } 54 mesh_common_state()55 tensorflow::TpuMeshCommonState* mesh_common_state() const { 56 if (mesh_state_ == nullptr) { 57 return nullptr; 58 } 59 return static_cast<tensorflow::TpuMeshCommonState*>( 60 OpsApiFn()->TpuMeshState_MeshCommonStateFn(mesh_state_)); 61 } 62 63 // Returns whether we should include the device assignment as a static field 64 // to the TPU program. This also determines whether we should include the 65 // device assignment as part of the compilation cache key. NeedsStaticDeviceAssignment(const TPUCompileMetadataProto & metadata,TpuCoreTypeEnum tpu_core_type)66 bool NeedsStaticDeviceAssignment( 67 const TPUCompileMetadataProto& metadata, 68 TpuCoreTypeEnum tpu_core_type) const { 69 if (mesh_state_ == nullptr) { 70 return false; 71 } 72 // Static device assignment enables XLA to perform certain optimization when 73 // all cores are used in the replicated computation. 74 return metadata.num_cores_per_replica() * metadata.num_replicas() == 75 OpsApiFn()->TpuTopology_AvailableCoreCountFn(mesh_state_, 76 tpu_core_type); 77 } 78 DebugString()79 string DebugString() const override { return "TpuMeshStateInterface"; } 80 81 private: 82 XLA_TpuMeshState* mesh_state_; 83 }; 84 85 } // namespace tpu 86 } // namespace tensorflow 87 88 #endif // EXPERIMENTAL_BRAIN_TPU_1VM_MINIEXECUTOR_TPU_MESH_STATE_INTERFACE_H_ 89