Bug fixed.

1. Fix the bugs in SetDevice functions and FlushToMem functions.
2. Fix the minor errors.

source/tensor/XMem.cpp

@@ -1632,7 +1632,8 @@ XMem * XMemManager::GetMem(const int devID)
             mem = GPUMems + devID;
         }
         else{
-            XPRINT1(0, stderr, "Cannot get the memory (%d). Please check your device id!", devID);
+            XPRINT1(0, stderr, "Please check your device id (%d)!", devID);
+            ShowNTErrors("Cannot get the memory!");
         }
     }
     

source/tensor/XTensor.cpp

@@ -520,9 +520,8 @@ relocate the data on the target device
 */
 void XTensor::SetDevice(int myDevId, XMem* myMem)
 {
-    if(myMem == NULL){
+    if(myMem == NULL)
         myMem = GMems.GetMem(myDevId);
-    }
     FlushToMem(myMem);
     isInGlobalMem = false;
 }
@@ -1914,6 +1913,7 @@ void XTensor::FlushToMem(XMem* targetMem)
         else if (mem != targetMem) {
             void* tmpData = targetMem->Alloc(targetMem->devID, GetDataSizeInChar());
             XMemCopy(tmpData, targetMem->devID, data, devID, GetDataSizeInChar());
+            mem->Release(data, GetDataSizeInChar(), signature);
             data = tmpData;
             mem = targetMem;
             devID = mem->devID;
@@ -1925,13 +1925,19 @@ void XTensor::FlushToMem(XMem* targetMem)
     else {
         if (devID >= 0) {
 #ifdef USE_CUDA
-            CudaGPUToCPUFlush(this);
-            mem = targetMem;
-            devID = mem->devID;
+            CudaGPUToCPUFlush(this, targetMem->devID, targetMem);
 #else
             ShowNTErrors("Recompile the code with USE_CUDA!");
 #endif
         }
+        else if (mem != targetMem) {
+            void* tmpData = targetMem->Alloc(targetMem->devID, GetDataSizeInChar());
+            XMemCopy(tmpData, targetMem->devID, data, devID, GetDataSizeInChar());
+            mem->Release(data, GetDataSizeInChar(), signature);
+            data = tmpData;
+            mem = targetMem;
+            devID = mem->devID;
+        }
     }
 }
 

source/tensor/core/utilities/FlushToMem.cu

@@ -78,6 +78,8 @@ void CudaCPUToGPUFlush(TensorList * mList, int devID, XMem * GPUMem)
 
         if(m->mem == NULL)
             delete[] (char*)m->data;
+        else
+            m->mem->Release(m->data, m->GetDataSizeInChar(), m->signature);
 
         m->dataHost = NULL;
         m->data = GPUData + p;
@@ -94,7 +96,36 @@ void CudaCPUToGPUFlush(TensorList * mList, int devID, XMem * GPUMem)
 #endif
 }
 
-/* copy the data from GPU memory to CPU memory */
+/* copy the data from GPU memory to CPU memory (memory pool) */
+void CudaGPUToCPUFlush(XTensor * tensor, int devID, XMem * CPUMem)
+{
+#ifdef USE_CUDA
+    CheckNTErrors((tensor->devID >= 0), "Cannot do cpu-flush on matrices that are already on CPU.");
+
+    /* compute the requried memory size */
+    int size = 0;
+    if (m->isSparse)
+        size = sizeof(int) + (sizeof(int) + tensor->unitSize) * tensor->unitNumNonZero;
+    else
+        size = tensor->unitSize * tensor->unitNum;
+
+    char * CPUData = CPUMem != NULL ? (char*)CPUMem->Alloc(CPUMem->devID, size):
+                                      (char*)XMemAlloc(devID, size);
+
+    /* copy from CPU memory to GPU memory */
+    cudaMemcpy(CPUData, tensor->data, size, cudaMemcpyDeviceToHost);
+
+    if (tensor->dataHost != NULL)
+        delete[](char*)tensor->dataHost;
+    tensor->dataHost = NULL;
+    tensor->mem->Release(tensor->data, tensor->GetDataSizeInChar(), tensor->signature);
+    tensor->data = CPUData;
+    tensor->devID = CPUMem != NULL ? CPUMem->devID : devID;
+    tensor->mem = CPUMem;
+#endif
+}
+
+/* copy the data from GPU memory to CPU memory ((dataHost)) and do not delete the data */
 void CudaGPUToCPUFlush(XTensor * tensor)
 {
     CheckNTErrors((sizeof(DTYPE) == tensor->unitSize), "Unsupported data type.");