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 // This is a registration-oriented interface for multiple platforms. It will 17 // replace the MachineManager singleton interface, as MachineManager does not 18 // currently support simultaneous use of multiple platforms. 19 // 20 // Usage: 21 // 22 // In your BUILD rule, add a dependency on a platform plugin that you'd like 23 // to use, such as: 24 // 25 // //perftools/gputools/executor/cuda:cuda_platform 26 // //perftools/gputools/executor/opencl:opencl_platform 27 // 28 // This will register platform plugins that can be discovered via this 29 // interface. Sample API usage: 30 // 31 // port::StatusOr<Platform*> platform_status = 32 // gpu::MultiPlatformManager::PlatformWithName("OpenCL"); 33 // if (!platform_status.ok()) { ... } 34 // Platform* platform = platform_status.ValueOrDie(); 35 // LOG(INFO) << platform->VisibleDeviceCount() << " devices visible"; 36 // if (platform->VisibleDeviceCount() <= 0) { return; } 37 // 38 // for (int i = 0; i < platform->VisibleDeviceCount(); ++i) { 39 // port::StatusOr<StreamExecutor*> executor_status = 40 // platform->ExecutorForDevice(i); 41 // if (!executor_status.ok()) { 42 // LOG(INFO) << "could not retrieve executor for device ordinal " << i 43 // << ": " << executor_status.status(); 44 // continue; 45 // } 46 // LOG(INFO) << "found usable executor: " << executor_status.ValueOrDie(); 47 // } 48 // 49 // A few things to note: 50 // - There is no standard formatting/practice for identifying the name of a 51 // platform. Ideally, a platform will list its registered name in its header 52 // or in other associated documentation. 53 // - Platform name lookup is case-insensitive. "OpenCL" or "opencl" (or even 54 // ("OpEnCl") would work correctly in the above example. 55 // 56 // And similarly, for standard interfaces (BLAS, RNG, etc.) you can add 57 // dependencies on support libraries, e.g.: 58 // 59 // //perftools/gputools/executor/cuda:pluton_blas_plugin 60 // //perftools/gputools/executor/cuda:cudnn_plugin 61 // //perftools/gputools/executor/cuda:cublas_plugin 62 // //perftools/gputools/executor/cuda:curand_plugin 63 64 #ifndef TENSORFLOW_STREAM_EXECUTOR_MULTI_PLATFORM_MANAGER_H_ 65 #define TENSORFLOW_STREAM_EXECUTOR_MULTI_PLATFORM_MANAGER_H_ 66 67 #include <functional> 68 #include <map> 69 #include <memory> 70 #include "tensorflow/stream_executor/platform/port.h" 71 72 #include "tensorflow/stream_executor/lib/status.h" 73 #include "tensorflow/stream_executor/lib/statusor.h" 74 #include "tensorflow/stream_executor/platform.h" 75 #include "tensorflow/stream_executor/platform/mutex.h" 76 #include "tensorflow/stream_executor/platform/port.h" 77 78 namespace perftools { 79 namespace gputools { 80 81 // Manages multiple platforms that may be present on the current machine. 82 class MultiPlatformManager { 83 public: 84 // Registers a platform object, returns an error status if the platform is 85 // already registered. The associated listener, if not null, will be used to 86 // trace events for ALL executors for that platform. 87 // Takes ownership of listener. 88 static port::Status RegisterPlatform(std::unique_ptr<Platform> platform); 89 90 // Retrieves the platform registered with the given platform name; e.g. 91 // "CUDA", "OpenCL", ... 92 // 93 // If the requested platform is not registered, an error status is returned. 94 // Ownership of the platform is NOT transferred to the caller -- 95 // the MultiPlatformManager owns the platforms in a singleton-like fashion. 96 static port::StatusOr<Platform*> PlatformWithName(const string& target); 97 98 // Retrieves the platform registered with the given platform ID, which 99 // is an opaque (but comparable) value. 100 // 101 // If the requested platform is not registered, an error status is returned. 102 // Ownership of the platform is NOT transferred to the caller -- 103 // the MultiPlatformManager owns the platforms in a singleton-like fashion. 104 static port::StatusOr<Platform*> PlatformWithId(const Platform::Id& id); 105 106 // Clears the set of registered platforms, primarily used for testing. 107 static void ClearPlatformRegistry(); 108 109 // Although the MultiPlatformManager "owns" its platforms, it holds them as 110 // undecorated pointers to prevent races during program exit (between this 111 // object's data and the underlying platforms (e.g., CUDA, OpenCL). 112 // Because certain platforms have unpredictable deinitialization 113 // times/sequences, it is not possible to strucure a safe deinitialization 114 // sequence. Thus, we intentionally "leak" allocated platforms to defer 115 // cleanup to the OS. This should be acceptable, as these are one-time 116 // allocations per program invocation. 117 // The MultiPlatformManager should be considered the owner 118 // of any platforms registered with it, and leak checking should be disabled 119 // during allocation of such Platforms, to avoid spurious reporting at program 120 // exit. 121 using PlatformMap = std::map<string, Platform*>; 122 123 // Provides access to the available set of platforms under a lock. 124 static port::Status WithPlatforms( 125 std::function<port::Status(PlatformMap*)> callback) { 126 mutex_lock lock(GetPlatformsMutex()); 127 return callback(GetPlatformMap()); 128 } 129 130 private: 131 // mutex that guards the platform map. 132 static mutex& GetPlatformsMutex() { 133 static mutex* platforms_mutex = new mutex; 134 return *platforms_mutex; 135 } 136 137 // TODO(b/22689637): Clean up these two maps; make sure they coexist nicely. 138 // TODO(b/22689637): Move this (whatever the final/"official" map is) to 139 // plugin_regstry.h, along with the associated functionality. 140 // Platform-name-to-object mapping. These platforms are registered via module 141 // initializers, and linkage determines which platforms are available to a 142 // given target. 143 static PlatformMap* GetPlatformMap() { 144 static PlatformMap* instance = new PlatformMap; 145 return instance; 146 } 147 148 // Holds a Platform::Id-to-object mapping. 149 // Unlike platforms_ above, this map does not own its contents. 150 static std::map<Platform::Id, Platform*>* GetPlatformByIdMap() { 151 using PlatformIdMap = std::map<Platform::Id, Platform*>; 152 static PlatformIdMap* instance = new PlatformIdMap; 153 return instance; 154 } 155 156 SE_DISALLOW_COPY_AND_ASSIGN(MultiPlatformManager); 157 }; 158 159 } // namespace gputools 160 } // namespace perftools 161 162 #endif // TENSORFLOW_STREAM_EXECUTOR_MULTI_PLATFORM_MANAGER_H_ 163