Bug fixed for Gather functions.

source/tensor/core/movement/Gather.cu

@@ -138,9 +138,8 @@ void _CudaGather(const XTensor * s, XTensor * t, XTensor * srcIndex)
         XMemCopy(sIndex, devID, srcIndex, -1, sizeof(int) * indexSize);
     }
     else {
-        int * sIndexCPU = new int[sizeof(int) * indexSize];
-        XMemCopy(sIndexCPU, -1, srcIndex, srcIndex->devID, sizeof(int) * indexSize);
-        int * sIndexData = (int*)sIndexCPU->data;
+        int * sIndexData = new int[sizeof(int) * indexSize];
+        XMemCopy(sIndexData, -1, srcIndex, srcIndex->devID, sizeof(int) * indexSize);
         for (int i = 0; i < indexSize; i++) {
             int srcIndexValue = sIndexData[i] * stride;
             CheckNTErrors(srcIndexValue < s->unitNum, "Wrong index!");
@@ -148,7 +147,7 @@ void _CudaGather(const XTensor * s, XTensor * t, XTensor * srcIndex)
 
         sIndex = (int *)srcIndex->data;
 
-        delete[] sIndexCPU;
+        delete[] sIndexData;
     }
 
     KernelGather<<<blocks, threads >>>(sData, tData, sIndex, indexSize, stride);