Introduce coordination between searching threads (#2204)

this patch improves threading performance by introducing some coordination between threads.

The observation is that threading is an area where a lot of Elo can potentially be gained:
https://github.com/glinscott/fishtest/wiki/UsefulData#elo-from-threading
At STC, 8 threads gain roughly 320 Elo, vs sequential at the same time,
however, loses 66 Elo against a single thread with 8x more time.
This 66 Elo should be partially recoverable with improved threading.

To improve threading, this patch introduces some LMR at nodes that are already being searched by other threads.
This requires some coordination between threads, avoiding however synchronisation.
To do so, threads leave a trail of breadcrumbs to mark the nodes they are searching.
These breadcrumbs are stored in a small hash table, which is only probed at low plies (currently ply < 8).

A couple of variants of this patch passed both STC and LTC threaded tests.
I picked the simpler, more robust version.
I expect that further tests can find further improvements.

STC (5+0.05 @ 8 threads):
LLR: 2.95 (-2.94,2.94) [0.50,4.50]
Total: 26209 W: 5359 L: 5079 D: 15771
http://tests.stockfishchess.org/tests/view/5d0a9b030ebc5925cf0a8e6f

LTC (20+0.2 @ 8 threads):
LLR: 2.96 (-2.94,2.94) [0.00,3.50]
Total: 34832 W: 5650 L: 5382 D: 23800
http://tests.stockfishchess.org/tests/view/5d0c67a20ebc5925cf0aafa7

other passed/tested variants:
http://tests.stockfishchess.org/tests/view/5d0a9b030ebc5925cf0a8e6f
http://tests.stockfishchess.org/tests/view/5d0c67ca0ebc5925cf0aafa9
http://tests.stockfishchess.org/tests/view/5d0c67810ebc5925cf0aafa3
http://tests.stockfishchess.org/tests/view/5d0958ca0ebc5925cf0a74c6

For the sequential code there is no change in bench, and an earlier version of this patch passed a non-regression test.
STC (10+0.1 @ 1 thread)
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 10471 W: 2364 L: 2220 D: 5887
http://tests.stockfishchess.org/tests/view/5d087ee20ebc5925cf0a6381

passed the additional non-regression tests at 2 and 4 threads 20+0.2 TC. The code was rebased on master prior to testing.

2 threads:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 218863 W: 40927 L: 41153 D: 136783
http://tests.stockfishchess.org/tests/view/5d18c6c30ebc5925cf0b9566

4threads:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 16839 W: 3017 L: 2889 D: 10933
http://tests.stockfishchess.org/tests/view/5d18c6ea0ebc5925cf0b9568

No functional change.

diff --git a/src/search.cpp b/src/search.cpp
index 37fc2ec48a8bfa99b7e4a125f68f540fe1ebaba0..e008bec1e45d8a063cca7d05dd803411d7cb38c6 100644 (file)
--- a/src/search.cpp
+++ b/src/search.cpp
@@ -102,6 +102,48 @@ namespace {
     Move best = MOVE_NONE;
   };
 
     Move best = MOVE_NONE;
   };
 

+  // Breadcrumbs are used to mark nodes as being searched by a given thread.
+  struct Breadcrumb {
+    std::atomic<Thread*> thread;
+    std::atomic<Key> key;
+  };
+  std::array<Breadcrumb, 1024> breadcrumbs;
+
+  // ThreadHolding keeps track of which thread left breadcrumbs at the given node for potential reductions.
+  // A free node will be marked upon entering the moves loop, and unmarked upon leaving that loop, by the ctor/dtor of this struct.
+  struct ThreadHolding {
+    explicit ThreadHolding(Thread* thisThread, Key posKey, int ply) {
+       location = ply < 8 ? &breadcrumbs[posKey & (breadcrumbs.size() - 1)] : nullptr;
+       otherThread = false;
+       owning = false;
+       if (location)
+       {
+          // see if another already marked this location, if not, mark it ourselves.
+          Thread* tmp = (*location).thread.load(std::memory_order_relaxed);
+          if (tmp == nullptr)
+          {
+              (*location).thread.store(thisThread, std::memory_order_relaxed);
+              (*location).key.store(posKey, std::memory_order_relaxed);
+              owning = true;
+          }
+          else if (   tmp != thisThread
+                   && (*location).key.load(std::memory_order_relaxed) == posKey)
+              otherThread = true;
+       }
+    }
+
+    ~ThreadHolding() {
+       if (owning) // free the marked location.
+           (*location).thread.store(nullptr, std::memory_order_relaxed);
+    }
+
+    bool marked() { return otherThread; }
+
+    private:
+    Breadcrumb* location;
+    bool otherThread, owning;
+  };
+

   template <NodeType NT>
   Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, bool cutNode);
 
   template <NodeType NT>
   Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, bool cutNode);
 


@@ -847,6 +889,9 @@ moves_loop: // When in check, search starts from here
     moveCountPruning = false;
     ttCapture = ttMove && pos.capture_or_promotion(ttMove);
 
     moveCountPruning = false;
     ttCapture = ttMove && pos.capture_or_promotion(ttMove);
 

+    // Mark this node as being searched.
+    ThreadHolding th(thisThread, posKey, ss->ply);
+

     // Step 12. Loop through all pseudo-legal moves until no moves remain
     // or a beta cutoff occurs.
     while ((move = mp.next_move(moveCountPruning)) != MOVE_NONE)
     // Step 12. Loop through all pseudo-legal moves until no moves remain
     // or a beta cutoff occurs.
     while ((move = mp.next_move(moveCountPruning)) != MOVE_NONE)


@@ -1019,6 +1064,10 @@ moves_loop: // When in check, search starts from here
       {
           Depth r = reduction(improving, depth, moveCount);
 
       {
           Depth r = reduction(improving, depth, moveCount);
 

+          // Reduction if other threads are searching this position.
+         if (th.marked())
+              r += ONE_PLY;
+

           // Decrease reduction if position is or has been on the PV
           if (ttPv)
               r -= 2 * ONE_PLY;
           // Decrease reduction if position is or has been on the PV
           if (ttPv)
               r -= 2 * ONE_PLY;