8.2. Shared Layout

Shared layout defines how tilus gets the offset of a tensor element in the shared memory. We can specify the layout of a shared tensor using the layout argument in shared_tensor(). If we do not provide the layout argument, tilus will try to infer the layout based on the usage of the shared tensor.

The difference between the shared layout and the global layout is that we require the shared layout have a constant shape, while the global layout supports symbolic shape. Another difference is that the offset computation in shared layout can involve the invariant variables that are invariant in the lifetime (i.e., the time after the shared tensor allocation and before the shared tensor free), while the global layout only supports grid-invariant variables.

We support the following functions to create a global layout:

tilus.ir.layout.shared_row_major(*shape)	Create a shared layout with row-major order.
tilus.ir.layout.shared_column_major(*shape)	Create a shared layout with column-major order.
tilus.ir.layout.shared_compose(lhs, rhs, *others)	Compose multiple shared layouts together.
tilus.ir.SharedLayout.create(shape, size, ...)	Create a shared layout.

Please refer to the SharedLayout class for internals of the shared layout.