Module limits 

Device limits and the dispatch geometry derived from them.

Everything in here except GpuLimits::from_client is a pure function of GpuLimits. That is deliberate: it means the behaviour on a device with half the shared memory, a quarter of the units per cube or a smaller plane can be asserted in a unit test on a machine that has none of those properties.

Structs

GpuLimits

Every device limit that dispatch geometry and staging decisions depend on.

Functions

checked_cube_count

Build a static cube count, checked against the device’s per-dimension limit.

fits_binding

Check a single allocation against the device’s per-binding size limit.

fits_shared_memory

Check a kernel’s shared-memory footprint against the device budget.

grid_2d

Split a flat cube count into a 2D grid within the device’s x/y limits.

grid_2d_limited

Split a flat cube count into a 2D grid bounded by max_dim per dimension.

plane_partitions

How many whole planes a wg_size-wide workgroup divides into.

plane_uniform

Whether a wg_size-wide workgroup is guaranteed to be exactly one plane.

resident_workgroups

How many workgroups of a given shared-memory footprint stay resident.

resolve_workgroup_size

Make a preferred workgroup width legal on the target device.