X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Ftt.cpp;h=0396b287cc5734b9517569e1124f91ddf0fcde73;hp=1e98aa24901549c888a3704576b5215619f6a6bf;hb=f4140ecc0c78d3d89f4e2459105e3ce3a1ab3ce1;hpb=ae4967744686ae6e95ca0cdb9e0d788de05d4e03

diff --git a/src/tt.cpp b/src/tt.cpp
index 1e98aa24..0396b287 100644
--- a/src/tt.cpp
+++ b/src/tt.cpp
@@ -29,6 +29,15 @@
 #include "movegen.h"
 #include "tt.h"
 
+#if defined(_MSC_VER)
+#include <xmmintrin.h>
+#endif
+
+// This is the number of TTEntry slots for each position
+static const int ClusterSize = 4;
+
+// The main transposition table
+TranspositionTable TT;
 
 ////
 //// Functions
@@ -56,16 +65,16 @@ 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) * ClusterSize * (sizeof(TTEntry)) <= (mbSize << 20))
       newSize *= 2;
 
   if (newSize != size)
   {
       size = newSize;
       delete [] entries;
-      entries = new TTEntry[size * 4];
+      entries = new TTEntry[size * ClusterSize];
       if (!entries)
       {
           std::cerr << "Failed to allocate " << mbSize
@@ -84,7 +93,17 @@ void TranspositionTable::set_size(unsigned mbSize) {
 
 void TranspositionTable::clear() {
 
-  memset(entries, 0, size * 4 * sizeof(TTEntry));
+  memset(entries, 0, size * ClusterSize * sizeof(TTEntry));
+}
+
+
+/// 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)) * ClusterSize);
 }
 
 
@@ -101,11 +120,12 @@ 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 +134,7 @@ 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 +147,7 @@ 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 +158,37 @@ 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.
+/// 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.
 
-inline TTEntry* TranspositionTable::first_entry(const Key posKey) const {
+void TranspositionTable::prefetch(const Key posKey) const {
 
-  return entries + (int(posKey & (size - 1)) << 2);
+#if defined(_MSC_VER)
+  _mm_prefetch((char*)first_entry(posKey), _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
@@ -224,6 +257,6 @@ 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))));
 }