greedy_layout

rstsr_core::prelude_dev

Function greedy_layout 

Source 
pub fn greedy_layout<D>(
    layout: &Layout<D>,
    keep_shape: bool,
) -> (Layout<D>, Vec<isize>)
where
    D: DimDevAPI,
Expand description

This function will return a f-prefer layout that make minimal memory accessing efforts (pointers will not frequently back-and-forth).

Note that this function should only be used for iteration.

§Parameter keep_shape

Keep size of output layout when input layout is boardcasted. This option should be false if TensorIterOrder::K and true if TensorIterOrder::G.

For example of layout shape [5, 1, 2, 1, 3, 6] and stride [1000, 10, 10, 40, 0, 100],

  • false: shape [2, 6, 5, 1, 1, 1] and stride [10, 100, 1000, 0, 0, 0]; meaning that broadcasted shapes are eliminated and moved to last axes.
  • true: shape [3, 1, 1, 2, 6, 5] and stride [0, 10, 40, 10, 100, 1000]; meaning that broadcasted shapes are iterated with most priority.

§Returns

  • layout: The output layout of greedy iteration.
  • index: Transpose index from input layout to output layout.