#include <cassert>
#include <cmath>
#include <cstring>
+#include <xmmintrin.h>
+#include "movegen.h"
#include "tt.h"
+// The main transposition table
+TranspositionTable TT;
////
//// Functions
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
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;
}
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)
+ // Do not overwrite when new type is VALUE_TYPE_EV_LO
+ if (tte->key() && t == VALUE_TYPE_EV_LO)
return;
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
if (c1 + c2 + c3 > 0)
replace = tte;
}
- *replace = TTEntry(posKey, v, t, d, m, generation);
+ *replace = TTEntry(posKey32, v, t, d, m, generation);
writes++;
}
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 {
+
+#if defined(__INTEL_COMPILER) || defined(__ICL)
+ // This hack prevents prefetches to be optimized away by the
+ // Intel compiler. Both MSVC and gcc seems not affected.
+ __asm__ ("");
+#endif
- return entries + (int(posKey & (size - 1)) << 2);
+ char const* addr = (char*)first_entry(posKey);
+ _mm_prefetch(addr, _MM_HINT_T2);
+ _mm_prefetch(addr+64, _MM_HINT_T2); // 64 bytes ahead
}
+
/// 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
}
+/// TranspositionTable::extract_pv() extends a PV by adding moves from the
+/// transposition table at the end. This should ensure that the PV is almost
+/// always at least two plies long, which is important, because otherwise we
+/// will often get single-move PVs when the search stops while failing high,
+/// and a single-move PV means that we don't have a ponder move.
+
+void TranspositionTable::extract_pv(const Position& pos, Move pv[], int pvSize) {
+
+ const TTEntry* tte;
+ StateInfo st;
+ Position p(pos);
+ int ply = 0;
+
+ // Update position to the end of current PV
+ while (pv[ply] != MOVE_NONE)
+ p.do_move(pv[ply++], st);
+
+ // Try to add moves from TT while possible
+ while ( (tte = retrieve(p.get_key())) != NULL
+ && tte->move() != MOVE_NONE
+ && move_is_legal(p, tte->move())
+ && (!p.is_draw() || ply < 2)
+ && ply < pvSize)
+ {
+ pv[ply] = tte->move();
+ p.do_move(pv[ply++], st);
+ }
+ pv[ply] = MOVE_NONE;
+}
+
+
/// TranspositionTable::full() returns the permill of all transposition table
/// entries which have received at least one write during the current search.
/// It is used to display the "info hashfull ..." information in UCI.
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))));
}