fix the bug of incorrect pointer to the head of memory pieces

source/tensor/XMem.cpp

@@ -685,13 +685,16 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
             int bi;
             for(bi = 0; bi < blockNum; bi++){
                 XMemBlock * block = blocks + bi;
-                if(block->mem == NULL){
+                if (block->mem != NULL && (block->head != NULL || block->size < mySize + 2 * MY_PITCH))
+                    continue;
+                
+                if (block->mem == NULL) {
                     block->size = MAX(maxBlockSize, mySize + 2 * MY_PITCH);
                     if (myDevID < 0) {
                         block->mem = new char[block->size];
                         memset(block->mem, 0, block->size);
                     }
-                    else{
+                    else {
 #ifdef USE_CUDA
                         int devIDBackup = -1;
                         cudaGetDevice(&devIDBackup);
@@ -706,6 +709,7 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
                         ShowNTErrors("Please specify USE_CUDA for compiling this program.");
 #endif
                     }
+                }
 
                 curBlockID = MAX(curBlockID, bi);
                     
@@ -731,7 +735,6 @@ void * XMem::AllocStandard(int myDevID, MTYPE mySize, bool myIsRebuiltIndex)
                 result = AllocStandard(myDevID, mySize, myIsRebuiltIndex);
                 break;
             }
-            }
             CheckNTErrors(bi < blockNum, "No enough memory is available!");
         }
     }