Reuse 5 slots instead of 4

[stockfish] / src / tt.cpp

diff --git a/src/tt.cpp b/src/tt.cpp
index c1a8b1b950b24edbc514fdb43baaae2dc6909288..5ea6a808acf86b2ad256f8715bf5c59772e015b1 100644 (file)
--- a/src/tt.cpp
+++ b/src/tt.cpp
@@ -29,6 +29,12 @@
 #include "movegen.h"
 #include "tt.h"
 
+#if defined(_MSC_VER)
+#include <xmmintrin.h>
+#endif
+
+// The main transposition table
+TranspositionTable TT;
 
 ////
 //// Functions
@@ -56,16 +62,16 @@ void TranspositionTable::set_size(unsigned mbSize) {
 
   unsigned newSize = 1024;
 
-  // We store a cluster of 4 TTEntry for each position and newSize is
-  // the maximum number of storable positions
-  while ((2 * newSize) * 4 * (sizeof(TTEntry)) <= (mbSize << 20))
+  // We store a cluster of ClusterSize number of TTEntry for each position
+  // and newSize is the maximum number of storable positions.
+  while ((2 * newSize) * sizeof(TTCluster) <= (mbSize << 20))
       newSize *= 2;
 
   if (newSize != size)
   {
       size = newSize;
       delete [] entries;
-      entries = new TTEntry[size * 4];
+      entries = new TTCluster[size];
       if (!entries)
       {
           std::cerr << "Failed to allocate " << mbSize
@@ -84,7 +90,17 @@ void TranspositionTable::set_size(unsigned mbSize) {
 
 void TranspositionTable::clear() {
 
-  memset(entries, 0, size * 4 * sizeof(TTEntry));
+  memset(entries, 0, size * sizeof(TTCluster));
+}
+
+
+/// TranspositionTable::first_entry returns a pointer to the first
+/// entry of a cluster given a position. The low 32 bits of the key
+/// are used to get the index in the table.
+
+inline TTEntry* TranspositionTable::first_entry(const Key posKey) const {
+
+  return entries[uint32_t(posKey) & (size - 1)].data;
 }
 
 
@@ -101,11 +117,12 @@ void TranspositionTable::clear() {
 void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d, Move m) {
 
   TTEntry *tte, *replace;
+  uint32_t posKey32 = posKey >> 32; // Use the high 32 bits as key
 
   tte = replace = first_entry(posKey);
-  for (int i = 0; i < 4; i++, tte++)
+  for (int i = 0; i < ClusterSize; i++, tte++)
   {
-      if (!tte->key() || tte->key() == posKey) // empty or overwrite old
+      if (!tte->key() || tte->key() == posKey32) // empty or overwrite old
       {
           // Do not overwrite when new type is VALUE_TYPE_EVAL
           if (tte->key() && t == VALUE_TYPE_EVAL)
@@ -114,7 +131,7 @@ void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d,
           if (m == MOVE_NONE)
               m = tte->move();
 
-          *tte = TTEntry(posKey, v, t, d, m, generation);
+          *tte = TTEntry(posKey32, v, t, d, m, generation);
           return;
       }
       else if (i == 0)  // replace would be a no-op in this common case
@@ -127,7 +144,7 @@ void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d,
       if (c1 + c2 + c3 > 0)
           replace = tte;
   }
-  *replace = TTEntry(posKey, v, t, d, m, generation);
+  *replace = TTEntry(posKey32, v, t, d, m, generation);
   writes++;
 }
 
@@ -138,24 +155,39 @@ void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d,
 
 TTEntry* TranspositionTable::retrieve(const Key posKey) const {
 
-  TTEntry *tte = first_entry(posKey);
+  uint32_t posKey32 = posKey >> 32;
+  TTEntry* tte = first_entry(posKey);
 
-  for (int i = 0; i < 4; i++, tte++)
-      if (tte->key() == posKey)
+  for (int i = 0; i < ClusterSize; i++, tte++)
+      if (tte->key() == posKey32)
           return tte;
 
   return NULL;
 }
 
 
-/// TranspositionTable::first_entry returns a pointer to the first
-/// entry of a cluster given a position.
-
-inline TTEntry* TranspositionTable::first_entry(const Key posKey) const {
-
-  return entries + (int(posKey & (size - 1)) << 2);
+/// TranspositionTable::prefetch looks up the current position in the
+/// transposition table and load it in L1/L2 cache. This is a non
+/// blocking function and do not stalls the CPU waiting for data
+/// to be loaded from RAM, that can be very slow. When we will
+/// subsequently call retrieve() the TT data will be already
+/// quickly accessible in L1/L2 CPU cache.
+
+void TranspositionTable::prefetch(const Key posKey) const {
+
+#if defined(_MSC_VER)
+   char* addr = (char*)first_entry(posKey);
+  _mm_prefetch(addr, _MM_HINT_T0);
+  _mm_prefetch(addr+64, _MM_HINT_T0);
+#else
+  // We need to force an asm volatile here because gcc builtin
+  // is optimized away by Intel compiler.
+  char* addr = (char*)first_entry(posKey);
+  asm volatile("prefetcht0 %0" :: "m" (addr));
+#endif
 }
 
+
 /// TranspositionTable::new_search() is called at the beginning of every new
 /// search. It increments the "generation" variable, which is used to
 /// distinguish transposition table entries from previous searches from
@@ -207,7 +239,7 @@ void TranspositionTable::extract_pv(const Position& pos, Move pv[]) {
        tte && tte->move() != MOVE_NONE && !stop;
        tte = retrieve(p.get_key()), ply++)
   {
-      if (!move_is_legal(p, tte->move(), p.pinned_pieces(p.side_to_move())))
+      if (!move_is_legal(p, tte->move()))
           break;
       pv[ply] = tte->move();
       p.do_move(pv[ply], st[ply]);
@@ -224,6 +256,6 @@ void TranspositionTable::extract_pv(const Position& pos, Move pv[]) {
 
 int TranspositionTable::full() const {
 
-  double N = double(size) * 4.0;
+  double N = double(size) * ClusterSize;
   return int(1000 * (1 - exp(writes * log(1.0 - 1.0/N))));
 }