Reuse 5 slots instead of 4

But this time with the guarantee of an always aligned
access so that prefetching is not adversely impacted.

On Joona PC
1+0, 64Mb hash:

Orig - Mod: 174 - 237 - 359

Instead after 1000 games at 1+0 with 128MB hash size
we are at + 1 ELO (just 4 games of difference).

Signed-off-by: Marco Costalba <mcostalba@gmail.com>

diff --git a/src/tt.cpp b/src/tt.cpp
index ffb8b0124568e58101a913af300b60bf2dc195ce..5ea6a808acf86b2ad256f8715bf5c59772e015b1 100644 (file)
--- a/src/tt.cpp
+++ b/src/tt.cpp
@@ -33,9 +33,6 @@
 #include <xmmintrin.h>
 #endif
 
 #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;
 
 // The main transposition table
 TranspositionTable TT;
 


@@ -67,14 +64,14 @@ void TranspositionTable::set_size(unsigned mbSize) {
 
   // We store a cluster of ClusterSize number of TTEntry for each position
   // and newSize is the maximum number of storable positions.
 
   // 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))
+  while ((2 * newSize) * sizeof(TTCluster) <= (mbSize << 20))

       newSize *= 2;
 
   if (newSize != size)
   {
       size = newSize;
       delete [] entries;
       newSize *= 2;
 
   if (newSize != size)
   {
       size = newSize;
       delete [] entries;

-      entries = new TTEntry[size * ClusterSize];
+      entries = new TTCluster[size];

       if (!entries)
       {
           std::cerr << "Failed to allocate " << mbSize
       if (!entries)
       {
           std::cerr << "Failed to allocate " << mbSize


@@ -93,7 +90,7 @@ void TranspositionTable::set_size(unsigned mbSize) {
 
 void TranspositionTable::clear() {
 
 
 void TranspositionTable::clear() {
 

-  memset(entries, 0, size * ClusterSize * sizeof(TTEntry));
+  memset(entries, 0, size * sizeof(TTCluster));

 }
 
 
 }
 
 


@@ -103,7 +100,7 @@ void TranspositionTable::clear() {
 
 inline TTEntry* TranspositionTable::first_entry(const Key posKey) const {
 
 
 inline TTEntry* TranspositionTable::first_entry(const Key posKey) const {
 

-  return entries + ((uint32_t(posKey) & (size - 1)) * ClusterSize);
+  return entries[uint32_t(posKey) & (size - 1)].data;

 }
 
 
 }
 
 

diff --git a/src/tt.h b/src/tt.h
index e838b0902fb9c140c8cd7bea1c551a00d4878213..2c989dacdcfdd399c9e010b114e19da1979af06a 100644 (file)
--- a/src/tt.h
+++ b/src/tt.h
@@ -70,10 +70,23 @@ private:
   uint32_t data;
   int16_t value_;
   int16_t depth_;
   uint32_t data;
   int16_t value_;
   int16_t depth_;

-  uint32_t pad_to_16_bytes;

 };
 
 };
 

-/// The transposition table class.  This is basically just a huge array
+
+/// This is the number of TTEntry slots for each position
+const int ClusterSize = 5;
+
+/// Each group of ClusterSize number of TTEntry form a TTCluster
+/// that is indexed by a single position key. Cluster is padded
+/// to a cache line size so to guarantee always aligned accesses.
+
+struct TTCluster {
+  TTEntry data[ClusterSize];
+  char cache_line_padding[64 - sizeof(TTEntry[ClusterSize])];
+};
+
+
+/// The transposition table class. This is basically just a huge array

 /// containing TTEntry objects, and a few methods for writing new entries
 /// and reading new ones.
 
 /// containing TTEntry objects, and a few methods for writing new entries
 /// and reading new ones.
 


@@ -95,14 +108,14 @@ public:
 private:
   inline TTEntry* first_entry(const Key posKey) const;
 
 private:
   inline TTEntry* first_entry(const Key posKey) const;
 

-  // Be sure 'writes' is at least one cacheline away
+  // Be sure 'writes' is at least one cache line away

   // from read only variables.
   unsigned char pad_before[64 - sizeof(unsigned)];
   unsigned writes; // heavy SMP read/write access here
   unsigned char pad_after[64];
 
   unsigned size;
   // from read only variables.
   unsigned char pad_before[64 - sizeof(unsigned)];
   unsigned writes; // heavy SMP read/write access here
   unsigned char pad_after[64];
 
   unsigned size;

-  TTEntry* entries;
+  TTCluster* entries;

   uint8_t generation;
 };
 
   uint8_t generation;
 };