reduce memory footprint for XMem

source/tensor/XMem.cpp

@@ -155,8 +155,13 @@ void XMem::Initialize(int myDevID, MEMPOOL_MODE myMode, MTYPE myBlockSize, int m
 
     bufSize = myBufSize;
 
+#ifdef SMALL_DEVICE
+    if (myMode == FREE_ON_THE_FLY)
+        SetIndex(50000);
+#else
     if (myMode == FREE_ON_THE_FLY)
         SetIndex(MILLION);
+#endif
 
     signature++;
 }
@@ -659,6 +664,9 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
             cur.next = &next;
             cur.size = needed;
             
+            if(next.next != NULL)
+                next.next->pre = &next;
+            
             next.state = 1;
             next.size = remaining;
             next.blockID = cur.blockID;
@@ -683,7 +691,7 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
         }
         /* if there is still no available memory piece, we have to obtain a new block of memory. */
         else{
-            MTYPE used = 0;
+            /*MTYPE used = 0;
             MTYPE total = 0;
             MTYPE free = 0;
             for(int i = 0; i < blockNum; i++){
@@ -706,11 +714,11 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
                         node = next;
                     }
                     else{
-                        if(node->size >= mySize){
+                        if(node->head.state == 1 && node->size >= mySize){
                             fprintf(stderr, "hit!!!!!!!!!!!\n");
                         }
 
-                        fprintf(stderr, "%d %lld %lld %lld\n", node->head.blockID, free, node->size, mySize);
+                        //fprintf(stderr, "%d %lld %lld %lld\n", node->head.blockID, free, node->size, mySize);
 
                         free += node->size;
                         node = node->next;
@@ -718,19 +726,6 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
                 }
             }
 
-            for(int i = 0; i < bufNodeCount; i++){
-                for(int j = i + 1; j < bufNodeCount; j++){
-                    MTYPE gap1 = (MTYPE)bufNodes[i].p - (MTYPE)bufNodes[j].p - bufNodes[j].head.size;
-                    MTYPE gap2 = (MTYPE)bufNodes[j].p - (MTYPE)bufNodes[i].p - bufNodes[i].head.size;
-                    if(gap1 > 0 && gap1 < 128){
-                        int nnn = 0;
-                    }
-                    if(gap2 > 0 && gap2 < 128){
-                        int nnn = 0;
-                    }
-                }
-            }
-
             MTYPE headSize = 0;
             MTYPE headSizeUsed = 0;
             for(int i = 0, j = 0; i < blockNum; i++){
@@ -753,7 +748,7 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
             delete[] bufNodes;
 
             fprintf(stderr, "%lld %lld\n", headSize, headSizeUsed);
-            fprintf(stderr, "mem: %lld %lld %lld %lld\n", used, total, free, mySize);
+            fprintf(stderr, "mem: %lld %lld %lld %lld\n", used, total, free, mySize);*/
 
             int bi;
             for(bi = 0; bi < blockNum; bi++){
@@ -1086,14 +1081,6 @@ void XMem::ReleaseStandard(int myDevID, void * p, MTYPE size)
             hit->head.size += pre->size;
             RemoveAllocIndexNode(pre->indexNode);
 
-            if(pre == blocks[head->blockID].head){
-                int nnn = 0;
-            }
-
-            if(head == blocks[head->blockID].head){
-                int nnn = 0;
-            }
-
             if(pre == blocks[head->blockID].head)
                 blocks[head->blockID].head = head;
         }
@@ -1182,7 +1169,7 @@ void XMem::RebuildIndex()
                 
                 if(head->state == 1){
                     newNode->size = (char*)p + head->size -
-                                    ( head->state == 1 ? (char*)GetPitchedAddress((char*)p, MY_PITCH) : (char*)head->indexNode->pReal);
+                                    (head->state == 1 ? (char*)GetPitchedAddress((char*)p, MY_PITCH) : (char*)head->indexNode->pReal);
                 }
                 else
                     newNode->size = node->size;
@@ -1203,7 +1190,6 @@ void XMem::RebuildIndex()
                 
                 if(newLast != NULL)
                     newLast->next = newHeader;
-                newHeader->pre = newLast;
                 
                 if(head->state == 1){
                     newNode->pReal = NULL;

source/tensor/test/TXMem.cpp

@@ -68,6 +68,7 @@ bool TestXMemCase1()
                 int j = rand() % caseNum;
 
                 //fprintf(stderr, "%d %d %d\n", testxmemid, j, ok);
+                //fprintf(stderr, "iter %d %d %d\n", iter, i, j);
 
                 if (p[j] == NULL) {
                     p[j] = (int*)mem.AllocStandard(mem.devID, size[j] * sizeof(int));