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 //   //third_party/tensorflow/stream_executor/cuda:cuda_platform
26 //   //third_party/tensorflow/stream_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 //      se::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 //    //third_party/tensorflow/stream_executor/cuda:pluton_blas_plugin
60 //    //third_party/tensorflow/stream_executor/cuda:cudnn_plugin
61 //    //third_party/tensorflow/stream_executor/cuda:cublas_plugin
62 //    //third_party/tensorflow/stream_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 <vector>
71 
72 #include "absl/strings/string_view.h"
73 #include "tensorflow/stream_executor/lib/initialize.h"
74 #include "tensorflow/stream_executor/lib/status.h"
75 #include "tensorflow/stream_executor/lib/statusor.h"
76 #include "tensorflow/stream_executor/platform.h"
77 #include "tensorflow/stream_executor/platform/port.h"
78 
79 namespace stream_executor {
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 platform.
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", ...) or id (an opaque, comparable value provided by the
92   // Platform's Id() method).
93   //
94   // If the platform has not already been initialized, it will be initialized
95   // with a default set of parameters.
96   //
97   // If the requested platform is not registered, an error status is returned.
98   // Ownership of the platform is NOT transferred to the caller --
99   // the MultiPlatformManager owns the platforms in a singleton-like fashion.
100   static port::StatusOr<Platform*> PlatformWithName(absl::string_view target);
101   static port::StatusOr<Platform*> PlatformWithId(const Platform::Id& id);
102 
103   // Retrieves the platform registered with the given platform name (e.g.
104   // "CUDA", "OpenCL", ...) or id (an opaque, comparable value provided by the
105   // Platform's Id() method).
106   //
107   // The platform will be initialized with the given options. If the platform
108   // was already initialized, an error will be returned.
109   //
110   // If the requested platform is not registered, an error status is returned.
111   // Ownership of the platform is NOT transferred to the caller --
112   // the MultiPlatformManager owns the platforms in a singleton-like fashion.
113   static port::StatusOr<Platform*> InitializePlatformWithName(
114       absl::string_view target, const std::map<string, string>& options);
115 
116   static port::StatusOr<Platform*> InitializePlatformWithId(
117       const Platform::Id& id, const std::map<string, string>& options);
118 
119   static std::vector<Platform*> AllPlatforms();
120 
121   // Although the MultiPlatformManager "owns" its platforms, it holds them as
122   // undecorated pointers to prevent races during program exit (between this
123   // object's data and the underlying platforms (e.g., CUDA, OpenCL).
124   // Because certain platforms have unpredictable deinitialization
125   // times/sequences, it is not possible to strucure a safe deinitialization
126   // sequence. Thus, we intentionally "leak" allocated platforms to defer
127   // cleanup to the OS. This should be acceptable, as these are one-time
128   // allocations per program invocation.
129   // The MultiPlatformManager should be considered the owner
130   // of any platforms registered with it, and leak checking should be disabled
131   // during allocation of such Platforms, to avoid spurious reporting at program
132   // exit.
133 
134   // Interface for a listener that gets notfied at certain events.
135   class Listener {
136    public:
137     virtual ~Listener() = default;
138     // Callback that is invoked when a Platform is registered.
139     virtual void PlatformRegistered(Platform* platform) = 0;
140   };
141   // Registers a listeners to receive notifications about certain events.
142   // Precondition: No Platform has been registered yet.
143   static port::Status RegisterListener(std::unique_ptr<Listener> listener);
144 };
145 
146 }  // namespace stream_executor
147 
148 // multi_platform_manager.cc will define these instances.
149 //
150 // Registering a platform:
151 // REGISTER_MODULE_INITIALIZER_SEQUENCE(my_platform, multi_platform_manager);
152 // REGISTER_MODULE_INITIALIZER_SEQUENCE(multi_platform_manager_listener,
153 // my_platform);
154 //
155 // Registering a listener:
156 // REGISTER_MODULE_INITIALIZER_SEQUENCE(my_listener,
157 // multi_platform_manager_listener);
158 DECLARE_MODULE_INITIALIZER(multi_platform_manager);
159 DECLARE_MODULE_INITIALIZER(multi_platform_manager_listener);
160 
161 #endif  // TENSORFLOW_STREAM_EXECUTOR_MULTI_PLATFORM_MANAGER_H_
162