row major xtensor assignment is so slow

[argman]

using tensor_type_row = xt::xtensor<float, 4, xt::layout_type::row_major>;
using tensor_type_col = xt::xtensor<float, 4, xt::layout_type::column_major>;
auto tmp_a = tensor_type_col({200, 512, 512, 5});
tensor_type_col tmp_b;
tmp_b = xt::view(tmp_a, xt::all(), xt::all(), xt::all(), xt::range(1, 2));

this demo code cost about 400ms if i use row_major, 100ms of column major

how can i speed up this ?

its even slow than using for loop with openmp enabled.

[spectre-ns]

I'm pretty sure this is slow because the column major xtensor will be strided in memory between elements. This will always be slow. I haven't run the assignment loop in this case but it could very well be using the stepper assignment method which won't be improved with openmp.

[tdegeus]

Could we quantify the performance by writing an explicit assignment in a loop (making use of the strides)?