1 //===- AsyncRuntime.cpp - Async runtime reference implementation ----------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements basic Async runtime API for supporting Async dialect
10 // to LLVM dialect lowering.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "mlir/ExecutionEngine/AsyncRuntime.h"
15 
16 #ifdef MLIR_ASYNCRUNTIME_DEFINE_FUNCTIONS
17 
18 #include <atomic>
19 #include <cassert>
20 #include <condition_variable>
21 #include <functional>
22 #include <iostream>
23 #include <mutex>
24 #include <thread>
25 #include <vector>
26 
27 using namespace mlir::runtime;
28 
29 //===----------------------------------------------------------------------===//
30 // Async runtime API.
31 //===----------------------------------------------------------------------===//
32 
33 namespace mlir {
34 namespace runtime {
35 namespace {
36 
37 // Forward declare class defined below.
38 class RefCounted;
39 
40 // -------------------------------------------------------------------------- //
41 // AsyncRuntime orchestrates all async operations and Async runtime API is built
42 // on top of the default runtime instance.
43 // -------------------------------------------------------------------------- //
44 
45 class AsyncRuntime {
46 public:
47   AsyncRuntime() : numRefCountedObjects(0) {}
48 
49   ~AsyncRuntime() {
50     assert(getNumRefCountedObjects() == 0 &&
51            "all ref counted objects must be destroyed");
52   }
53 
54   int32_t getNumRefCountedObjects() {
55     return numRefCountedObjects.load(std::memory_order_relaxed);
56   }
57 
58 private:
59   friend class RefCounted;
60 
61   // Count the total number of reference counted objects in this instance
62   // of an AsyncRuntime. For debugging purposes only.
63   void addNumRefCountedObjects() {
64     numRefCountedObjects.fetch_add(1, std::memory_order_relaxed);
65   }
66   void dropNumRefCountedObjects() {
67     numRefCountedObjects.fetch_sub(1, std::memory_order_relaxed);
68   }
69 
70   std::atomic<int32_t> numRefCountedObjects;
71 };
72 
73 // -------------------------------------------------------------------------- //
74 // A base class for all reference counted objects created by the async runtime.
75 // -------------------------------------------------------------------------- //
76 
77 class RefCounted {
78 public:
79   RefCounted(AsyncRuntime *runtime, int32_t refCount = 1)
80       : runtime(runtime), refCount(refCount) {
81     runtime->addNumRefCountedObjects();
82   }
83 
84   virtual ~RefCounted() {
85     assert(refCount.load() == 0 && "reference count must be zero");
86     runtime->dropNumRefCountedObjects();
87   }
88 
89   RefCounted(const RefCounted &) = delete;
90   RefCounted &operator=(const RefCounted &) = delete;
91 
92   void addRef(int32_t count = 1) { refCount.fetch_add(count); }
93 
94   void dropRef(int32_t count = 1) {
95     int32_t previous = refCount.fetch_sub(count);
96     assert(previous >= count && "reference count should not go below zero");
97     if (previous == count)
98       destroy();
99   }
100 
101 protected:
102   virtual void destroy() { delete this; }
103 
104 private:
105   AsyncRuntime *runtime;
106   std::atomic<int32_t> refCount;
107 };
108 
109 } // namespace
110 
111 // Returns the default per-process instance of an async runtime.
112 static AsyncRuntime *getDefaultAsyncRuntimeInstance() {
113   static auto runtime = std::make_unique<AsyncRuntime>();
114   return runtime.get();
115 }
116 
117 struct AsyncToken : public RefCounted {
118   // AsyncToken created with a reference count of 2 because it will be returned
119   // to the `async.execute` caller and also will be later on emplaced by the
120   // asynchronously executed task. If the caller immediately will drop its
121   // reference we must ensure that the token will be alive until the
122   // asynchronous operation is completed.
123   AsyncToken(AsyncRuntime *runtime) : RefCounted(runtime, /*count=*/2) {}
124 
125   // Internal state below guarded by a mutex.
126   std::mutex mu;
127   std::condition_variable cv;
128 
129   bool ready = false;
130   std::vector<std::function<void()>> awaiters;
131 };
132 
133 struct AsyncGroup : public RefCounted {
134   AsyncGroup(AsyncRuntime *runtime)
135       : RefCounted(runtime), pendingTokens(0), rank(0) {}
136 
137   std::atomic<int> pendingTokens;
138   std::atomic<int> rank;
139 
140   // Internal state below guarded by a mutex.
141   std::mutex mu;
142   std::condition_variable cv;
143 
144   std::vector<std::function<void()>> awaiters;
145 };
146 
147 } // namespace runtime
148 } // namespace mlir
149 
150 // Adds references to reference counted runtime object.
151 extern "C" void mlirAsyncRuntimeAddRef(RefCountedObjPtr ptr, int32_t count) {
152   RefCounted *refCounted = static_cast<RefCounted *>(ptr);
153   refCounted->addRef(count);
154 }
155 
156 // Drops references from reference counted runtime object.
157 extern "C" void mlirAsyncRuntimeDropRef(RefCountedObjPtr ptr, int32_t count) {
158   RefCounted *refCounted = static_cast<RefCounted *>(ptr);
159   refCounted->dropRef(count);
160 }
161 
162 // Create a new `async.token` in not-ready state.
163 extern "C" AsyncToken *mlirAsyncRuntimeCreateToken() {
164   AsyncToken *token = new AsyncToken(getDefaultAsyncRuntimeInstance());
165   return token;
166 }
167 
168 // Create a new `async.group` in empty state.
169 extern "C" AsyncGroup *mlirAsyncRuntimeCreateGroup() {
170   AsyncGroup *group = new AsyncGroup(getDefaultAsyncRuntimeInstance());
171   return group;
172 }
173 
174 extern "C" int64_t mlirAsyncRuntimeAddTokenToGroup(AsyncToken *token,
175                                                    AsyncGroup *group) {
176   std::unique_lock<std::mutex> lockToken(token->mu);
177   std::unique_lock<std::mutex> lockGroup(group->mu);
178 
179   // Get the rank of the token inside the group before we drop the reference.
180   int rank = group->rank.fetch_add(1);
181   group->pendingTokens.fetch_add(1);
182 
183   auto onTokenReady = [group]() {
184     // Run all group awaiters if it was the last token in the group.
185     if (group->pendingTokens.fetch_sub(1) == 1) {
186       group->cv.notify_all();
187       for (auto &awaiter : group->awaiters)
188         awaiter();
189     }
190   };
191 
192   if (token->ready) {
193     // Update group pending tokens immediately and maybe run awaiters.
194     onTokenReady();
195 
196   } else {
197     // Update group pending tokens when token will become ready. Because this
198     // will happen asynchronously we must ensure that `group` is alive until
199     // then, and re-ackquire the lock.
200     group->addRef();
201 
202     token->awaiters.push_back([group, onTokenReady]() {
203       // Make sure that `dropRef` does not destroy the mutex owned by the lock.
204       {
205         std::unique_lock<std::mutex> lockGroup(group->mu);
206         onTokenReady();
207       }
208       group->dropRef();
209     });
210   }
211 
212   return rank;
213 }
214 
215 // Switches `async.token` to ready state and runs all awaiters.
216 extern "C" void mlirAsyncRuntimeEmplaceToken(AsyncToken *token) {
217   // Make sure that `dropRef` does not destroy the mutex owned by the lock.
218   {
219     std::unique_lock<std::mutex> lock(token->mu);
220     token->ready = true;
221     token->cv.notify_all();
222     for (auto &awaiter : token->awaiters)
223       awaiter();
224   }
225 
226   // Async tokens created with a ref count `2` to keep token alive until the
227   // async task completes. Drop this reference explicitly when token emplaced.
228   token->dropRef();
229 }
230 
231 extern "C" void mlirAsyncRuntimeAwaitToken(AsyncToken *token) {
232   std::unique_lock<std::mutex> lock(token->mu);
233   if (!token->ready)
234     token->cv.wait(lock, [token] { return token->ready; });
235 }
236 
237 extern "C" void mlirAsyncRuntimeAwaitAllInGroup(AsyncGroup *group) {
238   std::unique_lock<std::mutex> lock(group->mu);
239   if (group->pendingTokens != 0)
240     group->cv.wait(lock, [group] { return group->pendingTokens == 0; });
241 }
242 
243 extern "C" void mlirAsyncRuntimeExecute(CoroHandle handle, CoroResume resume) {
244   (*resume)(handle);
245 }
246 
247 extern "C" void mlirAsyncRuntimeAwaitTokenAndExecute(AsyncToken *token,
248                                                      CoroHandle handle,
249                                                      CoroResume resume) {
250   std::unique_lock<std::mutex> lock(token->mu);
251   auto execute = [handle, resume]() { (*resume)(handle); };
252   if (token->ready)
253     execute();
254   else
255     token->awaiters.push_back([execute]() { execute(); });
256 }
257 
258 extern "C" void mlirAsyncRuntimeAwaitAllInGroupAndExecute(AsyncGroup *group,
259                                                           CoroHandle handle,
260                                                           CoroResume resume) {
261   std::unique_lock<std::mutex> lock(group->mu);
262   auto execute = [handle, resume]() { (*resume)(handle); };
263   if (group->pendingTokens == 0)
264     execute();
265   else
266     group->awaiters.push_back([execute]() { execute(); });
267 }
268 
269 //===----------------------------------------------------------------------===//
270 // Small async runtime support library for testing.
271 //===----------------------------------------------------------------------===//
272 
273 extern "C" void mlirAsyncRuntimePrintCurrentThreadId() {
274   static thread_local std::thread::id thisId = std::this_thread::get_id();
275   std::cout << "Current thread id: " << thisId << std::endl;
276 }
277 
278 #endif // MLIR_ASYNCRUNTIME_DEFINE_FUNCTIONS
279