Script.atomic.shared_cas

Script.atomic.shared_cas(dst, compare, values, *, sem='relaxed', scope='cta', output=None)

Element-wise compare-and-swap on shared memory.

Per element: old = dst[i]; if (old == compare[i]) dst[i] = values[i], atomically. The returned output (if bound) holds old, which the caller typically inspects to decide whether the swap succeeded.

Parameters:

• dst (SharedTensor) – Destination tile in shared memory.

• compare (RegisterTensor) – Expected-old-value tile; same shape and dtype as dst.

• values (RegisterTensor) – Tile of replacement values; same shape and dtype as dst.

• sem (str) – See shared_add().

• scope (str) – See shared_add().

• output (RegisterTensor | None) – See shared_add().

Returns:

Pre-CAS value at each element when output is consumed; None otherwise. Note that, unlike the arithmetic ops, CAS has no red.* form, so an unused output still costs a register allocation at the PTX level.

Return type:

RegisterTensor or None

Notes

• Thread group: Can be executed by any sized thread group.

• Hardware: Requires compute capability 7.0+ (sm_70).

• PTX: atom.{sem}.{scope}.shared.cas.s32.