reduce memory footprint of XMem freeotf

source/sample/transformer/T2TTrainer.cpp

@@ -167,6 +167,22 @@ void T2TTrainer::Train(const char * fn, T2TModel * model)
             /* get probabilities */
             float prob = GetProb(&output, &gold, NULL);
 
+            MTYPE totalUsed = 0;
+            MTYPE totalSize = 0;
+            
+            for (int i = 0; i <= mem->curBlockID; i++) {
+                totalSize += mem->blocks[i].size;
+                totalUsed += mem->blocks[i].used;
+                if (totalSize <= 0) {
+                    int nnn = 0;
+                }
+                if (totalUsed <= 0) {
+                    int nnn = 0;
+                }
+            }
+
+            //fprintf(stderr, "%d(%ld,%ld,%f)\n", mem->curBlockID, totalUsed, totalSize, (float)totalUsed/totalSize);
+            
             loss += -prob;
             wordCount += wc;
             wordCountTotal += wc;

source/tensor/XMem.cpp

@@ -532,7 +532,7 @@ release a piece of memory
 */
 void XMem::Release(void * p, MTYPE size, MTYPE code)
 {
-    if(code == signature)
+    //if(code == signature)
         Release(devID, p, size);
 }
 
@@ -667,6 +667,7 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
         hit->size = mySize;
         hit->head.state = 2;
         hit->pReal = beg;
+        blocks[hit->head.blockID].used += mySize;
         
         RemoveFreeIndexNode(hit);
         AddAllocIndexNode(hit);
@@ -731,6 +732,7 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
                 header.blockID = bi;
                     
                 block->head = &header;
+                block->used = 0;
                     
                 result = AllocStandard(myDevID, mySize, myIsRebuiltIndex);
                 break;
@@ -917,6 +919,8 @@ void XMem::ReleaseStandard(int myDevID, void * p, MTYPE size)
     
     RemoveAllocIndexNode(hit);
     AddFreeIndexNode(hit);
+
+    blocks[hit->head.blockID].used -= hit->head.size;
 }
 
 /* rebuild index to merge small fragments of memory and free the block with no use */
@@ -934,6 +938,7 @@ void XMem::RebuildIndex()
         CheckNTErrors(head->size <= block->size, "Illegal memory block!");
         
         block->head = NULL;
+        block->used = 0;
 
         /* if the block is not used, we delete it */
         if(head->state == 1 && head->size == block->size){
@@ -976,7 +981,8 @@ void XMem::RebuildIndex()
                 /* make a new index node */
                 MPieceNode * newNode = memIndex2 + nodeNumUsed2++;
                 newNode->p = p;
-                newNode->size = (char*)p + node->size - (char*)GetPitchedAddress((char*)p, MY_PITCH);
+                newNode->size = (char*)p + head->size - 
+                                ( head->state == 1 ? (char*)GetPitchedAddress((char*)p, MY_PITCH) : head->indexNode->pReal);
                 newNode->pre = NULL;
                 newNode->next = NULL;
                 
@@ -1004,6 +1010,7 @@ void XMem::RebuildIndex()
                     newNode->pReal = head->indexNode->pReal;
                     MPieceNode * entry = memIndex2 + indexEntryNum + FindIndexEntry(newNode->size);
                     AddAllocIndexNode(newNode, entry);
+                    block->used += head->size;
                 }
                 
                 if(newLast == NULL)

source/tensor/XTensor.cpp

@@ -162,6 +162,7 @@ XTensor::XTensor(const XTensor &reference)
         devID = reference.devID;
         mem = reference.mem;
         data = reference.data;
+        signature = reference.signature;
         
         /* what we really want to do is "reference.data = NULL;"
            As "reference" is constant, we cannot reset reference.data
@@ -299,6 +300,7 @@ XTensor& XTensor::operator= (const XTensor& tensor)
         newTensor->SetTMP();
         newTensor->data = data;
         newTensor->dataHost = dataHost;
+        newTensor->signature = tensor.signature;
         
         XLink::Replace(this, newTensor);
         XLink::ClearOutgoing(this);