]> git.sesse.net Git - stockfish/blobdiff - src/thread.h
Merge remote-tracking branch 'upstream/master'
[stockfish] / src / thread.h
index 37a4a6ca2bc6215448bab8634656d9d1f8a1d7ed..43e2e1423ced3cadad508f5e1b8bc2b5bfc55978 100644 (file)
@@ -1,6 +1,6 @@
 /*
   Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
+  Copyright (C) 2004-2024 The Stockfish developers (see AUTHORS file)
 
   Stockfish is free software: you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
 #include <condition_variable>
 #include <cstddef>
 #include <cstdint>
+#include <functional>
+#include <memory>
 #include <mutex>
 #include <vector>
 
-#include "movepick.h"
+#include "numa.h"
 #include "position.h"
 #include "search.h"
 #include "thread_win32_osx.h"
-#include "types.h"
 
 namespace Stockfish {
 
-// Thread class keeps together all the thread-related stuff. We use
-// per-thread pawn and material hash tables so that once we get a
-// pointer to an entry its lifetime is unlimited and we don't have
-// to care about someone changing the entry under our feet.
-class Thread {
 
-    std::mutex              mutex;
-    std::condition_variable cv;
-    size_t                  idx;
-    bool                    exit = false, searching = true;  // Set before starting std::thread
-    NativeThread            stdThread;
+class OptionsMap;
+using Value = int;
 
+// Sometimes we don't want to actually bind the threads, but the recipient still
+// needs to think it runs on *some* NUMA node, such that it can access structures
+// that rely on NUMA node knowledge. This class encapsulates this optional process
+// such that the recipient does not need to know whether the binding happened or not.
+class OptionalThreadToNumaNodeBinder {
    public:
-    explicit Thread(size_t);
-    virtual ~Thread();
-    virtual void search();
-    void         clear();
-    void         idle_loop();
-    void         start_searching();
-    void         wait_for_search_finished();
-    size_t       id() const { return idx; }
-
-    size_t                pvIdx, pvLast;
-    std::atomic<uint64_t> nodes, tbHits, bestMoveChanges;
-    int                   selDepth, nmpMinPly;
-    Value                 bestValue, optimism[COLOR_NB];
-
-    Position              rootPos;
-    StateInfo             rootState;
-    Search::RootMoves     rootMoves;
-    Depth                 rootDepth, completedDepth;
-    Value                 rootDelta;
-    Value                 rootSimpleEval;
-    CounterMoveHistory    counterMoves;
-    ButterflyHistory      mainHistory;
-    CapturePieceToHistory captureHistory;
-    ContinuationHistory   continuationHistory[2][2];
-    PawnHistory           pawnHistory;
+    OptionalThreadToNumaNodeBinder(NumaIndex n) :
+        numaConfig(nullptr),
+        numaId(n) {}
+
+    OptionalThreadToNumaNodeBinder(const NumaConfig& cfg, NumaIndex n) :
+        numaConfig(&cfg),
+        numaId(n) {}
+
+    NumaReplicatedAccessToken operator()() const {
+        if (numaConfig != nullptr)
+            return numaConfig->bind_current_thread_to_numa_node(numaId);
+        else
+            return NumaReplicatedAccessToken(numaId);
+    }
+
+   private:
+    const NumaConfig* numaConfig;
+    NumaIndex         numaId;
 };
 
+// Abstraction of a thread. It contains a pointer to the worker and a native thread.
+// After construction, the native thread is started with idle_loop()
+// waiting for a signal to start searching.
+// When the signal is received, the thread starts searching and when
+// the search is finished, it goes back to idle_loop() waiting for a new signal.
+class Thread {
+   public:
+    Thread(Search::SharedState&,
+           std::unique_ptr<Search::ISearchManager>,
+           size_t,
+           OptionalThreadToNumaNodeBinder);
+    virtual ~Thread();
 
-// MainThread is a derived class specific for main thread
-struct MainThread: public Thread {
+    void idle_loop();
+    void start_searching();
+    void clear_worker();
+    void run_custom_job(std::function<void()> f);
 
-    using Thread::Thread;
+    void ensure_network_replicated();
 
-    void search() override;
-    void check_time();
+    // Thread has been slightly altered to allow running custom jobs, so
+    // this name is no longer correct. However, this class (and ThreadPool)
+    // require further work to make them properly generic while maintaining
+    // appropriate specificity regarding search, from the point of view of an
+    // outside user, so renaming of this function is left for whenever that happens.
+    void   wait_for_search_finished();
+    size_t id() const { return idx; }
 
-    double           previousTimeReduction;
-    Value            bestPreviousScore;
-    Value            bestPreviousAverageScore;
-    Value            iterValue[4];
-    int              callsCnt;
-    bool             stopOnPonderhit;
-    std::atomic_bool ponder;
+    std::unique_ptr<Search::Worker> worker;
+    std::function<void()>           jobFunc;
+
+   private:
+    std::mutex                mutex;
+    std::condition_variable   cv;
+    size_t                    idx, nthreads;
+    bool                      exit = false, searching = true;  // Set before starting std::thread
+    NativeThread              stdThread;
+    NumaReplicatedAccessToken numaAccessToken;
 };
 
 
 // ThreadPool struct handles all the threads-related stuff like init, starting,
 // parking and, most importantly, launching a thread. All the access to threads
 // is done through this class.
-struct ThreadPool {
+class ThreadPool {
+   public:
+    ThreadPool() {}
+
+    ~ThreadPool() {
+        // destroy any existing thread(s)
+        if (threads.size() > 0)
+        {
+            main_thread()->wait_for_search_finished();
+
+            threads.clear();
+        }
+    }
+
+    ThreadPool(const ThreadPool&) = delete;
+    ThreadPool(ThreadPool&&)      = delete;
 
-    void start_thinking(Position&, StateListPtr&, const Search::LimitsType&, bool = false);
-    void clear();
-    void set(size_t);
+    ThreadPool& operator=(const ThreadPool&) = delete;
+    ThreadPool& operator=(ThreadPool&&)      = delete;
 
-    MainThread* main() const { return static_cast<MainThread*>(threads.front()); }
-    uint64_t    nodes_searched() const { return accumulate(&Thread::nodes); }
-    uint64_t    tb_hits() const { return accumulate(&Thread::tbHits); }
-    Thread*     get_best_thread() const;
-    void        start_searching();
-    void        wait_for_search_finished() const;
+    void   start_thinking(const OptionsMap&, Position&, StateListPtr&, Search::LimitsType);
+    void   run_on_thread(size_t threadId, std::function<void()> f);
+    void   wait_on_thread(size_t threadId);
+    size_t num_threads() const;
+    void   clear();
+    void   set(const NumaConfig& numaConfig,
+               Search::SharedState,
+               const Search::SearchManager::UpdateContext&);
 
-    std::atomic_bool stop, increaseDepth;
+    Search::SearchManager* main_manager();
+    Thread*                main_thread() const { return threads.front().get(); }
+    uint64_t               nodes_searched() const;
+    uint64_t               tb_hits() const;
+    Thread*                get_best_thread() const;
+    void                   start_searching();
+    void                   wait_for_search_finished() const;
+
+    std::vector<size_t> get_bound_thread_count_by_numa_node() const;
+
+    void ensure_network_replicated();
+
+    std::atomic_bool stop, abortedSearch, increaseDepth;
 
     auto cbegin() const noexcept { return threads.cbegin(); }
     auto begin() noexcept { return threads.begin(); }
@@ -119,20 +160,19 @@ struct ThreadPool {
     auto empty() const noexcept { return threads.empty(); }
 
    private:
-    StateListPtr         setupStates;
-    std::vector<Thread*> threads;
+    StateListPtr                         setupStates;
+    std::vector<std::unique_ptr<Thread>> threads;
+    std::vector<NumaIndex>               boundThreadToNumaNode;
 
-    uint64_t accumulate(std::atomic<uint64_t> Thread::*member) const {
+    uint64_t accumulate(std::atomic<uint64_t> Search::Worker::*member) const {
 
         uint64_t sum = 0;
-        for (Thread* th : threads)
-            sum += (th->*member).load(std::memory_order_relaxed);
+        for (auto&& th : threads)
+            sum += (th->worker.get()->*member).load(std::memory_order_relaxed);
         return sum;
     }
 };
 
-extern ThreadPool Threads;
-
 }  // namespace Stockfish
 
 #endif  // #ifndef THREAD_H_INCLUDED