For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo module
tile_scheduler_splitk
Tile scheduler for split-K matmul kernels on SM100 structured hardware.
Wraps the B200 tile scheduler to partition the K dimension into multiple splits, coordinating per-split workspace reductions and lock-based synchronization across CTAs that contribute to the same output tile.
comptime values
SPLITK_COUNTER_SLOTS
comptime SPLITK_COUNTER_SLOTS = 4
Counter slots per kernel. Each device context, which launches on its own stream, takes a slot of its own, so launches on different streams never share counters; contexts past the last slot use per-call counters.
SPLITK_COUNTER_STRIDE
comptime SPLITK_COUNTER_STRIDE = 32
Words between tile counters: one 128-byte line each, so the tiles' atomics do not share a line.
SPLITK_SLOT_WORDS
comptime SPLITK_SLOT_WORDS = 8192
Words in one counter slot.
SPLITK_TILE_COUNTERS
comptime SPLITK_TILE_COUNTERS = 256
Tile counters per slot. A parallel split-K grid has every CTA resident, at most one per SM, so it has fewer tiles than this.
SplitKTileCounters
comptime SplitKTileCounters = Pointer[UInt32, MutUntrackedOrigin, address_space=AddressSpace.GLOBAL]
Pointer to a kernel's tile counters (see splitk_tile_counters).
Structs
-
SchedulerWorkIteratorSplitK: Work iterator for Scheduler warp (split-K) using next-style iteration. -
TileScheduler: Tile scheduler that partitions the K dimension into multiple splits. -
WaitAndAdvanceContextSplitK: Context for waiting on CLC barrier and advancing work iterator (Split-K). -
WorkInfo: Describes a unit of split-K work for a single output tile. -
WorkIteratorSplitK: Per-warp work iterator for split-K using next-style iteration.
Functions
-
get_num_tiles: Computes the number of output tiles aligned up to the cluster shape. -
get_required_locks_buffer_size_bytes: Computes the byte size of the lock buffer needed for split-K reduction. -
splitk_k_tiles_per_split: Returns the K tiles per split, rounded up to a whole K group. -
splitk_tile_counters: Returns the calling kernel's tile counters for the parallel split-K reduction.