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 #ifndef TENSORFLOW_STREAM_EXECUTOR_CUDA_CUDA_PLATFORM_H_ 17 #define TENSORFLOW_STREAM_EXECUTOR_CUDA_CUDA_PLATFORM_H_ 18 19 #include <memory> 20 #include "tensorflow/stream_executor/platform/port.h" 21 #include <vector> 22 23 #include "tensorflow/stream_executor/executor_cache.h" 24 #include "tensorflow/stream_executor/lib/statusor.h" 25 #include "tensorflow/stream_executor/multi_platform_manager.h" 26 #include "tensorflow/stream_executor/platform.h" 27 #include "tensorflow/stream_executor/platform/mutex.h" 28 #include "tensorflow/stream_executor/platform/port.h" 29 #include "tensorflow/stream_executor/platform/thread_annotations.h" 30 #include "tensorflow/stream_executor/stream_executor_internal.h" 31 #include "tensorflow/stream_executor/stream_executor_pimpl.h" 32 #include "tensorflow/stream_executor/trace_listener.h" 33 34 namespace perftools { 35 namespace gputools { 36 namespace cuda { 37 38 // Opaque and unique identifier for the CUDA platform plugin. 39 // This is needed so that plugins can refer to/identify this platform without 40 // instantiating a CudaPlatform object. 41 extern const Platform::Id kCudaPlatformId; 42 43 // Cuda-specific platform plugin, registered as a singleton value via module 44 // initializer. 45 class CudaPlatform : public Platform { 46 public: 47 CudaPlatform(); 48 ~CudaPlatform() override; 49 50 // CudaPlatform-specific functionality 51 // Returns the number of distinct buses / NUMA nodes on the machine. 52 int BusCount(); 53 54 // Returns the bus/NUMA node for the specified device ordinal. 55 int DeviceToBus(int device_ordinal); 56 57 // Returns the lowest-ordinal-number StreamExecutor on the specified bus. 58 port::StatusOr<StreamExecutor*> FirstExecutorForBus(int bus_ordinal); 59 60 // Platform interface implementation: 61 // Returns the same value as kCudaPlatform above. 62 Platform::Id id() const override; 63 64 // Returns -1 as a sentinel on internal failure (and logs the error). 65 int VisibleDeviceCount() const override; 66 67 const string& Name() const override; 68 69 port::StatusOr<StreamExecutor*> ExecutorForDevice(int ordinal) override; 70 71 port::StatusOr<StreamExecutor*> ExecutorForDeviceWithPluginConfig( 72 int ordinal, const PluginConfig& config) override; 73 74 port::StatusOr<StreamExecutor*> GetExecutor( 75 const StreamExecutorConfig& config) override; 76 77 port::StatusOr<std::unique_ptr<StreamExecutor>> GetUncachedExecutor( 78 const StreamExecutorConfig& config) override; 79 80 void RegisterTraceListener(std::unique_ptr<TraceListener> listener) override; 81 82 void UnregisterTraceListener(TraceListener* listener) override; 83 84 private: 85 // Determines the number of NUMA nodes and the assignment of executor to each. 86 void InspectNumaNodes(); 87 88 // This platform's name. 89 string name_; 90 91 // Cache of created executors. 92 ExecutorCache executor_cache_; 93 94 // The smallest NUMA node value for any device managed by this machine 95 // manager. Used, along with limit_numa_node_, to convert NUMA nodes into bus 96 // ordinals. The NUMA node space occupied by GPUs is assumed to be dense./ 97 int min_numa_node_; 98 99 // Larger than the NUMA node value for any device managed by this machine 100 // manager. 101 int limit_numa_node_; 102 103 SE_DISALLOW_COPY_AND_ASSIGN(CudaPlatform); 104 }; 105 106 } // namespace cuda 107 } // namespace gputools 108 } // namespace perftools 109 110 #endif // TENSORFLOW_STREAM_EXECUTOR_CUDA_CUDA_PLATFORM_H_ 111
