The unit testing of test_lstm_op and test_gru_op failed due to the Eigen in debug mode.

[qingqing01]

In the debug mode, the unit testing of test_lstm_op and test_gru_op failed, error:

python: /home/dangqingqing/.third_party/eigen3/src/extern_eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h:704: static bool Eigen::internal::OuterReducer<Self, Op, Eigen::GpuDevice>::run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) [with Device = Eigen::GpuDevice; OutputType = double; Self = Eigen::TensorEvaluator<const Eigen::TensorReductionOp<Eigen::internal::SumReducer<double>, const Eigen::array<int, 1ul>, const Eigen::TensorMap<Eigen::Tensor<const double, 2, 1, long int>, 0, Eigen::MakePointer>, Eigen::MakePointer>, Eigen::GpuDevice>; Op = Eigen::internal::SumReducer<double>; typename Self::Index = long int]: Assertion `false && "Should only be called to reduce doubles or floats on a gpu device"' failed.

Thread 1 "python" received signal SIGABRT, Aborted.

The Eigen sum function failed with double type on GPU.

[qingqing01]

The elementwise_mul_op have the same problem. But from link: https://bitbucket.org/eigen/eigen/pull-requests/216/enable-efficient-tensor-reduction-for/diff

It seems the Eigen has enabled efficient Tensor reduction for doubles on the GPU. And there is no error in release mode.