Module global

Solves full reductions by loading blocks into shared memory. Handles memory movement, bound checks, plane specialization.

Re-exports

pub use quantization::*;

Modules

args

global_memory

load

Loaders read from global memory and write to stage memory

multi_stage

quantization

single_stage

Structs

LoadSpecializationConfig

Configuration for how each input tensor (Lhs and Rhs) is loaded, specifying the plane roles responsible for loading them.

MaxLoaderPlanes

Maximal number of planes each loader can handle to divide its workload evenly

PlaneRoleConfig

Contains the number of plane in each role and the rule to distinguish planes based on their plane id

PlaneWriter

Writes tiles from out shared memory to output global memory using a plane for each tile

SpecializedLoadingSides

Aggregates loading sides for both main flow and load only roles

Specializer

Specialization information in cube functions

UnitWriter

Writes tiles from out shared memory to output global memory using a unit for each tile

ZeroAccumulatorLoader

Accumulator loader that zeros the accumulator

ZeroAccumulatorLoaderExpand

Enums

LoadingSides

Specifies which input(s) a plane role participates in loading.

RoleRule

Rule to distinguish a plane’s role based on its plane id

RoleRuleConfig

Comptime version of RoleRule

SpecializationTensorConfig

Determines which types of planes are responsible for loading a tensor.

SpecializerKind

Comptime information of specializer

Traits

AccumulatorLoader

Loads an accumulator with pre-defined data

CopyMechanism

Allows to copy a slice of data from global to shared memory asynchronously

GlobalConfig

Configuration for the global matmul level.

GlobalMatmul

Provides matrix multiplication operations at the global level.

GlobalMatmulFamily

A family of matmuls working with any precision.

GlobalWriter

Responsible of writing the accumulated stage matmul output to global memory