AsyncLoadingJob

cubecl_matmul::components::global::read

Trait AsyncLoadingJob 

Source 
pub trait AsyncLoadingJob<IP: MatrixPrecision, TL: TilingLayout>:
    CubeType
    + Copy
    + Clone {
    // Required methods
    fn execute_task<CM: CopyMechanism, G: GlobalConfig>(
        this: &mut Self,
        task_id: u32,
        tensor_reader: &GlobalIterator<Line<IP::Global>>,
        stage_memory: &mut StridedStage<IP::Stage, TL>,
        mechanism: &CM,
        config: G,
    );
    fn task_count(this: &Self) -> u32;
    fn __expand_execute_task<CM: CopyMechanism, G: GlobalConfig>(
        scope: &mut Scope,
        this: <Self as CubeType>::ExpandType,
        task_id: <u32 as CubeType>::ExpandType,
        tensor_reader: <GlobalIterator<Line<IP::Global>> as CubeType>::ExpandType,
        stage_memory: <StridedStage<IP::Stage, TL> as CubeType>::ExpandType,
        mechanism: <CM as CubeType>::ExpandType,
        config: G,
    ) -> <() as CubeType>::ExpandType;
    fn __expand_task_count(
        scope: &mut Scope,
        this: <Self as CubeType>::ExpandType,
    ) -> u32;
}
Expand description

A loading job represents a sequence of loading tasks. Each task is the smallest unit of loading work: one unit at one iteration, operating at a specific point within a read view. The job holds shared information reused across read views and iterations. By calling execute_task at strategic moments, one can hope to speed up the matmul.

Required Methods§

Source

fn execute_task<CM: CopyMechanism, G: GlobalConfig>( this: &mut Self, task_id: u32, tensor_reader: &GlobalIterator<Line<IP::Global>>, stage_memory: &mut StridedStage<IP::Stage, TL>, mechanism: &CM, config: G, )

Execute the task_idth loading task

Source

fn task_count(this: &Self) -> u32

Get the number of tasks

Source

fn __expand_execute_task<CM: CopyMechanism, G: GlobalConfig>( scope: &mut Scope, this: <Self as CubeType>::ExpandType, task_id: <u32 as CubeType>::ExpandType, tensor_reader: <GlobalIterator<Line<IP::Global>> as CubeType>::ExpandType, stage_memory: <StridedStage<IP::Stage, TL> as CubeType>::ExpandType, mechanism: <CM as CubeType>::ExpandType, config: G, ) -> <() as CubeType>::ExpandType

Source

fn __expand_task_count( scope: &mut Scope, this: <Self as CubeType>::ExpandType, ) -> u32

Dyn Compatibility§

This trait is not dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety", so this trait is not object safe.

Implementors§