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 struct

SplitCounterSync

struct SplitCounterSync[pipeline_stages: Int, block_rows: Int, warp_rows: Int, reads_per_warp_block: Int]

Split counter synchronization strategy.

Uses separate producer and consumer counters per tile to reduce atomic contention. Producers only write to producer counters, consumers only write to consumer counters.

Phase progression:

• Producer phase advances by reads_per_warp_block (waits for N consumers)
• Consumer phase advances by writes_per_warp_block (waits for 1 producer)
• This asymmetry reflects the 1-producer-to-N-consumers relationship

Parameters

• ​pipeline_stages (Int): Number of pipeline stages in the ring buffer.
• ​block_rows (Int): Total number of rows in the work block.
• ​warp_rows (Int): Number of rows processed by a single warp.
• ​reads_per_warp_block (Int): Number of consumer read operations per warp block.

Fields

• ​producer_counters (SplitCounterSync[pipeline_stages, block_rows, warp_rows, reads_per_warp_block].ProducerCounterArray):
• ​consumer_counters (SplitCounterSync[pipeline_stages, block_rows, warp_rows, reads_per_warp_block].ConsumerCounterArray):

Implemented traits

AnyType, Copyable, Deinitable, ImplicitlyCopyable, Movable, RegisterPassable, SyncStrategy, TrivialRegisterPassable

comptime members

block_warps

comptime block_warps = (block_rows // warp_rows)

ConsumerCounterArray

comptime ConsumerCounterArray = SMemArray[Int32, SplitCounterSync[pipeline_stages, block_rows, warp_rows, reads_per_warp_block].total_tiles]

ProducerCounterArray

comptime ProducerCounterArray = SMemArray[Int32, SplitCounterSync[pipeline_stages, block_rows, warp_rows, reads_per_warp_block].total_tiles]

total_tiles

comptime total_tiles = (SplitCounterSync[pipeline_stages, block_rows, warp_rows, reads_per_warp_block].block_warps * pipeline_stages)

writes_per_warp_block

comptime writes_per_warp_block = 1

Methods

__init__

def __init__() -> Self

Initialize with internally allocated producer and consumer counters.

get_staged_idx

def get_staged_idx(self, tile_idx: Int, stage: Int) -> Int

Returns:

Int

wait_producer_acquire

def wait_producer_acquire(self, tile_idx: Int, stage: Int, phase: Int32)

Producer waits on consumer counter.

Args:

• ​tile_idx (Int): Index of the tile within a stage (0 to block_warps-1).
• ​stage (Int): Pipeline stage (0 to pipeline_stages-1).
• ​phase (Int32): Counter threshold to wait for before acquiring the tile.

signal_producer_release

def signal_producer_release(mut self, tile_idx: Int, stage: Int)

Producer increments producer counter.

Args:

• ​tile_idx (Int): Index of the tile within a stage (0 to block_warps-1).
• ​stage (Int): Pipeline stage (0 to pipeline_stages-1).

wait_consumer_acquire

def wait_consumer_acquire(self, tile_idx: Int, stage: Int, phase: Int32)

Consumer waits on producer counter.

Args:

• ​tile_idx (Int): Index of the tile within a stage (0 to block_warps-1).
• ​stage (Int): Pipeline stage (0 to pipeline_stages-1).
• ​phase (Int32): Counter threshold to wait for before acquiring the tile.

signal_consumer_release

def signal_consumer_release(mut self, tile_idx: Int, stage: Int)

Consumer increments consumer counter by 1.

Args:

• ​tile_idx (Int): Index of the tile within a stage (0 to block_warps-1).
• ​stage (Int): Pipeline stage (0 to pipeline_stages-1).

get_producer_phase_increment

def get_producer_phase_increment(self) -> Int32

Producer phase advances by reads_per_warp_block.

Returns:

Int32

get_consumer_phase_increment

def get_consumer_phase_increment(self) -> Int32

Consumer phase advances by writes_per_warp_block.

Returns:

Int32