triton.experimental.gluon.language.amd.cdna5.tdm.async_load_fused

triton.experimental.gluon.language.amd.cdna5.tdm.async_load_fused(members: List[Tuple[tensor_descriptor, shared_memory_descriptor, ttgl.constexpr | int]], cache_modifier='', _semantic=None) → None

Emit one explicit fused TDM load for 2-4 descriptor/destination pairs.

This can perform better than several consecutive separate TDM loads, especially for more than two loads. Under the hood, different warps load different descriptor/destination pairs according to each member’s warp_used_hint; collectively all participating warps load all pairs at the block level.

Each member is (desc, dest, warp_used_hint). The descriptors must already encode their tile offsets, predicates, and bounds; use update_tensor_descriptor() before calling this helper when needed. All members share one cache modifier, matching the fused IR operation.

Parameters:

• members – 2-4 (desc, dest, warp_used_hint) tuples. Hints must be legal, pairwise-disjoint bitmasks.

• cache_modifier (str, optional) – Cache behavior shared by all members.