From efe5bc6a3cb7f2069646883f29a16d3f345f74f8 Mon Sep 17 00:00:00 2001
From: Peter Boyle <paboyle@ph.ed.ac.uk>
Date: Sat, 9 May 2020 22:27:56 -0400
Subject: [PATCH] Split allocator cache into two pools of different sizes

---
 Grid/allocator/AlignedAllocator.cc | 69 +++++++++++++++---------------
 Grid/allocator/AlignedAllocator.h  | 17 ++++----
 2 files changed, 44 insertions(+), 42 deletions(-)

diff --git a/Grid/allocator/AlignedAllocator.cc b/Grid/allocator/AlignedAllocator.cc
index d53c4dc2..77646410 100644
--- a/Grid/allocator/AlignedAllocator.cc
+++ b/Grid/allocator/AlignedAllocator.cc
@@ -6,21 +6,19 @@ NAMESPACE_BEGIN(Grid);
 MemoryStats *MemoryProfiler::stats = nullptr;
 bool         MemoryProfiler::debug = false;
 
-#ifdef GRID_NVCC
-#define SMALL_LIMIT (0)
-#else
-#define SMALL_LIMIT (4096)
-#endif
-
-#ifdef POINTER_CACHE
-int PointerCache::victim;
-
+int PointerCache::Victim;
+int PointerCache::VictimSmall;
 PointerCache::PointerCacheEntry PointerCache::Entries[PointerCache::Ncache];
+PointerCache::PointerCacheEntry PointerCache::EntriesSmall[PointerCache::NcacheSmall];
 
-void *PointerCache::Insert(void *ptr,size_t bytes) {
-
-  if (bytes < SMALL_LIMIT ) return ptr;
-
+void *PointerCache::Insert(void *ptr,size_t bytes) 
+{
+  if (bytes < GRID_ALLOC_SMALL_LIMIT ) 
+    return Insert(ptr,bytes,EntriesSmall,NcacheSmall,VictimSmall);
+  return Insert(ptr,bytes,Entries,Ncache,Victim);  
+}
+void *PointerCache::Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim) 
+{
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
@@ -28,8 +26,8 @@ void *PointerCache::Insert(void *ptr,size_t bytes) {
   void * ret = NULL;
   int v = -1;
 
-  for(int e=0;e<Ncache;e++) {
-    if ( Entries[e].valid==0 ) {
+  for(int e=0;e<ncache;e++) {
+    if ( entries[e].valid==0 ) {
       v=e; 
       break;
     }
@@ -37,40 +35,43 @@ void *PointerCache::Insert(void *ptr,size_t bytes) {
 
   if ( v==-1 ) {
     v=victim;
-    victim = (victim+1)%Ncache;
+    victim = (victim+1)%ncache;
   }
 
-  if ( Entries[v].valid ) {
-    ret = Entries[v].address;
-    Entries[v].valid = 0;
-    Entries[v].address = NULL;
-    Entries[v].bytes = 0;
+  if ( entries[v].valid ) {
+    ret = entries[v].address;
+    entries[v].valid = 0;
+    entries[v].address = NULL;
+    entries[v].bytes = 0;
   }
 
-  Entries[v].address=ptr;
-  Entries[v].bytes  =bytes;
-  Entries[v].valid  =1;
+  entries[v].address=ptr;
+  entries[v].bytes  =bytes;
+  entries[v].valid  =1;
 
   return ret;
 }
 
-void *PointerCache::Lookup(size_t bytes) {
-
-  if (bytes < SMALL_LIMIT ) return NULL;
-
+void *PointerCache::Lookup(size_t bytes)
+{
+  if (bytes < GRID_ALLOC_SMALL_LIMIT ) 
+    return Lookup(bytes,EntriesSmall,NcacheSmall);
+  return Lookup(bytes,Entries,Ncache);
+}
+void *PointerCache::Lookup(size_t bytes,PointerCacheEntry *entries,int ncache) 
+{
 #ifdef GRID_OMP
   assert(omp_in_parallel()==0);
 #endif 
-
-  for(int e=0;e<Ncache;e++){
-    if ( Entries[e].valid && ( Entries[e].bytes == bytes ) ) {
-      Entries[e].valid = 0;
-      return Entries[e].address;
+  for(int e=0;e<ncache;e++){
+    if ( entries[e].valid && ( entries[e].bytes == bytes ) ) {
+      entries[e].valid = 0;
+      return entries[e].address;
     }
   }
   return NULL;
 }
-#endif
+
 
 void check_huge_pages(void *Buf,uint64_t BYTES)
 {
diff --git a/Grid/allocator/AlignedAllocator.h b/Grid/allocator/AlignedAllocator.h
index 8c189be8..d6e2e073 100644
--- a/Grid/allocator/AlignedAllocator.h
+++ b/Grid/allocator/AlignedAllocator.h
@@ -42,21 +42,19 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
 
 #define POINTER_CACHE
 #define GRID_ALLOC_ALIGN (2*1024*1024)
+#define GRID_ALLOC_SMALL_LIMIT (4096)
 
 NAMESPACE_BEGIN(Grid);
 
 // Move control to configure.ac and Config.h?
-#ifdef POINTER_CACHE
+
 class PointerCache {
 private:
 /*Pinning pages is costly*/
 /*Could maintain separate large and small allocation caches*/
-#ifdef GRID_NVCC 
-  static const int Ncache=128;
-#else
+/* Could make these configurable, perhaps up to a max size*/
+  static const int NcacheSmall=128; 
   static const int Ncache=8;
-#endif
-  static int victim;
 
   typedef struct { 
     void *address;
@@ -65,14 +63,17 @@ private:
   } PointerCacheEntry;
     
   static PointerCacheEntry Entries[Ncache];
+  static int Victim;
+  static PointerCacheEntry EntriesSmall[NcacheSmall];
+  static int VictimSmall;
 
 public:
 
   static void *Insert(void *ptr,size_t bytes) ;
+  static void *Insert(void *ptr,size_t bytes,PointerCacheEntry *entries,int ncache,int &victim) ;
   static void *Lookup(size_t bytes) ;
-
+  static void *Lookup(size_t bytes,PointerCacheEntry *entries,int ncache) ;
 };
-#endif  
 
 std::string sizeString(size_t bytes);