Module layout 

The order a tensor’s dimensions occupy memory in. The order a tensor’s dimensions occupy memory in.

A tensor’s strides say which dimension is innermost, which next, and so on outward — its dimension order. Contiguous NCHW is the identity order; a convolution that computes channels-last and hands back a permuted view is [0, 2, 3, 1]. A kernel that iterates a tensor in its own dimension order reads it linearly; one that iterates in any other order reads it strided.

Two questions are asked here. dim_order asks for the order of a tensor that is dense — that fills its buffer with no gaps — which is what a kernel needs before it may treat the buffer as a flat run of elements. nested_dim_order asks only that the dimensions nest, tolerating the gap a pitched or tile-aligned allocation leaves under a dimension, which is enough to decide what order to iterate in.

Functions

dim_order

The dimension order of a tensor that is dense in memory, or None if it is not dense.

is_contiguous_order

Whether a dimension order is the contiguous one, [0, 1, .., rank - 1].

nested_dim_order

The dimension order of a tensor whose dimensions nest without overlapping, or None if they do not.

Type Aliases

DimOrder

The order a tensor’s dimensions appear in memory, outermost first.