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

sched_barrier_pairs

def sched_barrier_pairs[pairs: Int, valu_cnt: Int, group: Int]()

Emits pairs schedule groups of shape [1 MFMA, valu_cnt VALU].

Each call expands to two schedule_group_barrier invocations under sync_id=group: one declaring "1 MFMA in this group" and one declaring "valu_cnt VALUs in this group". pairs > 1 recurses to emit additional pairs in the same sync_id, all of which LLVM orders relative to each other.

Note valu_cnt is the count of VALU instructions in each group, not a VALU-to-MFMA ratio; LLVM derives the interleave from the group declaration. Mojo's underlying schedule_group_barrier(mask, size, sync_id) docstring labels size as a "repeat count" — that is misleading; it is the instruction count per group.

Parameters:

pairs (Int): Number of (MFMA, VALU) groups to emit, all sharing sync_id=group.
valu_cnt (Int): VALU instructions per group.
group (Int): IGroupLP sync_id. Reuse within one cluster's pair sequence; pick distinct ids for distinct clusters so their constraints stay independent. 1-10 typical for MHA main+epilogue.