replace the cuda thread allocation code

source/tensor/core/movement/Spread.cu

@@ -252,11 +252,6 @@ void _CudaSpreadForGather(XTensor * source, XTensor * collection, int dim,
         }
     }
     else{
-        GDevs.GetCudaThread2D(source->devID, blockNum * indexSize, stride, MAX_INT, cudaGrids, cudaBlocks);
-
-        dim3 blocks(cudaGrids[0], cudaGrids[1]);
-        dim3 threads(cudaBlocks[0], cudaBlocks[1]);
-
         XMem * mem = source->mem;
         int * si = mem != NULL ? 
                   (int*)mem->AllocBuf(mem->devID, sizeof(int) * indexSize) : 
@@ -271,6 +266,11 @@ void _CudaSpreadForGather(XTensor * source, XTensor * collection, int dim,
         DTYPE * s = (DTYPE*)source->data;
         DTYPE * c = (DTYPE*)collection->data;
 
+        GDevs.GetCudaThread2D(source->devID, blockNum * indexSize, stride, MAX_INT, cudaGrids, cudaBlocks);
+
+        dim3 blocks(cudaGrids[0], cudaGrids[1]);
+        dim3 threads(cudaBlocks[0], cudaBlocks[1]);
+
         KernelSpreadForGatherFuzed<<<blocks, threads >>>(s, c, blockNum, blockSizeSrc, blockSizeColl, stride, indexSize, si, ci);
 
         if (mem != NULL) {