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

emit_minimal_barrier_block

emit_minimal_barrier_block(block: MMABlockSpec, wrap_waits: Bool, global_before_frag: Bool = False) -> List[OpDesc]

Emit one block in the "minimal-barrier + cross-stage rotation" shape — for schedules that override build_explicit_blocks.

Layout (per block):

Sync-group A: [sched_barrier] entry_wait entry_wait_lgkm [sched_barrier]. Fences emitted iff wrap_waits=True and either entry-wait field is present.
Load section: frags + globals, in order controlled by global_before_frag (False = ping-pong default; True = load-before-frag for kernels like 4-wave inline that benefit).
Sync-group B: [sched_barrier] pre_sync [barrier post_barrier_lgkm] [sched_barrier]. Fences emitted iff wrap_waits=True and either pre_sync or pre_mma_barrier is present.
Final mma.

Reads wait values, frag/global ops, barrier flags from block — typically populated by _construct_mma_blocks and patched by derive_waits_from_blocks. Schedules consume the derived structure and emit it in their preferred order, without the conditional template branching of MMABlockSpec.expand.

Bypasses pre_mma_set_prio / post_mma_* / fused_mma / drain flags — those don't apply under the minimal-barriers pattern. Schedules with different needs should write their own emitter.

Args:

block (MMABlockSpec): The block spec to emit.
wrap_waits (Bool): Wrap each contiguous wait/barrier group with schedule_barrier on both sides.
global_before_frag (Bool): Load globals before frags inside the block.

Returns:

List[OpDesc]: Ordered op list for the block, ready to be appended to a per-block emission override.