To _roll(To to, From from, long shift, long axis,
types::array<long, N> const &shape, utils::int_<M>)
{
long n = shape[N - M];
long offset = std::accumulate(shape.begin() + N - M + 1, shape.end(), 1L,
std::multiplies<long>());
if (axis == N - M) {
for (long i = 0; i < shift; ++i)
to = _roll(to, from + (n - shift + i) * offset, shift, axis, shape,
utils::int_<M - 1>());
for (long i = shift; i < n; ++i)
to = _roll(to, from + (i - shift) * offset, shift, axis, shape,
utils::int_<M - 1>());
} else
for (From end = from + n * offset; from != end; from += offset)
to = _roll(to, from, shift, axis, shape, utils::int_<M - 1>());
return to;
}
types::ndarray<T,N> roll(types::ndarray<T,N> const& expr, long shift, long axis)
{
auto&& expr_shape = expr.shape();
while(shift<0) shift+=expr_shape[axis];
types::ndarray<T,N> out(expr_shape, __builtin__::None);
_roll(out.fbegin(), expr.fbegin(), shift, axis, expr_shape, utils::int_<N>());
return out;
}
