For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

get_waves_per_eu

def get_waves_per_eu(depth: Int) -> Int

Return the recommended rocdl.waves_per_eu hint for an AMD MHA kernel.

AMD GCN/CDNA schedulers use this hint to decide how many wavefronts to co-issue per execution unit. Shallow heads (depth 64 or 128) benefit from two waves to hide memory latency, while deeper heads use one wave to conserve register file capacity.

Args:

depth (Int): Attention head depth (key/value dimension per head).

Returns:

Int: 2 for depth 64 or 128, 1 otherwise.