1 /* Copyright 2019 Google LLC. 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 // Context is the user-facing context class. 17 18 #ifndef RUY_RUY_CONTEXT_H_ 19 #define RUY_RUY_CONTEXT_H_ 20 21 #include <cstdint> 22 23 namespace ruy { 24 25 class Ctx; 26 class CtxImpl; 27 class ThreadPool; 28 enum class Path : std::uint8_t; 29 enum class Tuning; 30 enum class PerformanceAdvisory; 31 32 // A Context holds runtime information used by Ruy. It holds runtime resources 33 // such as the workers thread pool and the allocator (which holds buffers for 34 // temporary data), as well as runtime options controlling which Paths are 35 // enabled (typically based on which instruction sets are detected) and how 36 // many threads to use. 37 class Context final { 38 public: 39 Context(); 40 ~Context(); 41 42 // Returns the Path enum value that corresponds to the code path used by 43 // the last ruy::Mul with this Context. 44 Path last_used_path() const; 45 46 // Control of whether to use kernels tuned for in-order or out-of-order CPU 47 // cores. The default is auto-detection, so these methods should only be used 48 // to override that auto-detection if it's not working as intended or for 49 // testing. 50 Tuning explicit_tuning() const; 51 void set_explicit_tuning(Tuning value); 52 53 // The thread pool held by this context to dispatch a ruy::Mul to worker 54 // threads. 55 // 56 // By default, threads may spin-wait for a few milliseconds before reverting 57 // to passive wait. This can be controlled by 58 // `mutable_thread_pool()->set_spin_milliseconds(value)`. 59 const ThreadPool& thread_pool() const; 60 ThreadPool* mutable_thread_pool(); 61 62 // Controls the maximum number of threads to be used by ruy::Mul with this 63 // Context. The number of threads in the pool will be that value minus one, 64 // as the remaining portion of the work is done directly on the calling 65 // thread. 66 // 67 // This defaults to 1. Multi-threading in ruy is always opt-in. There is 68 // no auto-detection of hardware concurrency. That is on purpose, ruy focuses 69 // on mobile applications where such concepts are difficult to define 70 // (e.g. ARM big.LITTLE). 71 int max_num_threads() const; 72 void set_max_num_threads(int value); 73 74 // Returns true of the last ruy::Mul using this Context flagged the specified 75 // `advisory`. This is reset by each ruy::Mul call. 76 bool performance_advisory(PerformanceAdvisory advisory) const; 77 78 // When using Matrix::set_cache_policy(), this Context will keep a cache of 79 // pre-packed matrix data. This function clears that cache. 80 void ClearPrepackedCache(); 81 82 // Override auto-detection of supported code paths. 83 // 84 // Passing `paths == Path::kNone` means reverting to the default behavior. 85 // This will trigger auto-detection on the next use. 86 // 87 // Other values will override auto-detection with the explicitly provided set 88 // of paths. 89 // 90 // Paths in kNonArchPaths are always implicitly supported. 91 void set_runtime_enabled_paths(Path paths); 92 93 // Returns the set of Path's that are available. 94 Path get_runtime_enabled_paths(); 95 96 private: 97 CtxImpl* const impl_; 98 99 const Ctx& ctx() const; 100 Ctx* mutable_ctx(); 101 102 friend const Ctx* get_ctx(const Context*); 103 friend Ctx* get_ctx(Context*); 104 105 // Disallow copy 106 Context(const Context&) = delete; 107 }; 108 109 } // end namespace ruy 110 111 #endif // RUY_RUY_CONTEXT_H_ 112