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

BlockReducer

struct BlockReducer[BLOCK_SIZE: Int]

Reduces a scalar across BLOCK_SIZE threads in a block. Broadcasts the result to every thread.

Parameters

• ​BLOCK_SIZE (Int): Number of threads in the launching block.

Implemented traits

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

Methods

__init__

def __init__() -> Self

Default-initializes a BlockReducer.

sum

def sum[dtype: DType](self, val: Scalar[dtype]) -> Scalar[dtype]

Returns the sum of val across the block.

Parameters:

• ​dtype (DType): The scalar dtype.

Args:

• ​val (Scalar[dtype]): The per-thread value.

Returns:

Scalar[dtype]: The block-wide sum, broadcast to every thread.

max

def max[dtype: DType](self, val: Scalar[dtype]) -> Scalar[dtype]

Returns the maximum of val across the block.

Parameters:

• ​dtype (DType): The scalar dtype.

Args:

• ​val (Scalar[dtype]): The per-thread value.

Returns:

Scalar[dtype]: The block-wide maximum, broadcast to every thread.

min

def min[dtype: DType](self, val: Scalar[dtype]) -> Scalar[dtype]

Returns the minimum of val across the block.

Parameters:

• ​dtype (DType): The scalar dtype.

Args:

• ​val (Scalar[dtype]): The per-thread value.

Returns:

Scalar[dtype]: The block-wide minimum, broadcast to every thread.

generic

def generic[S: ReduceOp](self, mut state: S)

Block-wide all-reduce over state.join; on return every thread holds the combined value.

Dispatches on state size (a perf choice — either path is byte-safe since ReduceOp is TrivialRegisterPassable):

• Small (<= _WARP_SHUFFLE_MAX_WORDS uint32 words): a register-only within-warp warp-shuffle butterfly, then a sparse one-entry-per-warp shmem store + a second shuffle over warp 0 (mirrors legacy welford_block_all_reduce) — BLOCK_SIZE/WARP_SIZE stores + one barrier, ~30x less shmem traffic than the tree.
• Large: a log2(BLOCK_SIZE)-step block-wide shmem tree. Past a few words the per-word shuffle + register pressure makes the tree comparable or better.

Parameters:

• ​S (ReduceOp): The monoid type being combined.

Args:

• ​state (S): The per-thread state; on return, holds the block-wide combined value on every thread.