Make sure 'SrcVectorDim' is not same as 'DstVectorDim'

include/ck/tensor_operation/gpu/device/device_permute.hpp

@@ -109,6 +109,7 @@ struct DevicePermute
     static_assert(3 <= NumDim, "Only accept at least 3D dimension tensor");
     static_assert((NumDim - 2) <= SrcVectorDim && SrcVectorDim < NumDim);
     static_assert((NumDim - 2) <= DstVectorDim && DstVectorDim < NumDim);
+    static_assert(SrcVectorDim != DstVectorDim);
 
     template <index_t N = NumDim>
     static auto ConvertArrayToTuple(const std::array<index_t, NumDim>& array)

include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp

@@ -116,6 +116,7 @@ struct GridwisePermute
                   SrcVectorDim < InGridDesc::GetNumOfDimension());
     static_assert((OutGridDesc::GetNumOfDimension() - 2) <= DstVectorDim &&
                   DstVectorDim < OutGridDesc::GetNumOfDimension());
+    static_assert(SrcVectorDim != DstVectorDim);
 
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};