Module stage

Module stage 

Source

Structs§

ColMajorTilingOrder
Tiles laid out in column-major order.
ColMajorTilingOrderExpand
ContiguousTilingLayout
Each tile is stored contiguously in shared memory. Global memory loads may require remapping to match this layout.
FilledStage
FilledStageExpand
FilledStageFamily
NoEvent
Use when there is no event listening to do
NoEventExpand
NoTilingLayout
Dummy tiling layout that panics if it’s used. Can be used when the reader is known to be a FillReader
NumStages
Number of stages in one shared memory, i.e. buffers for double buffering
OrderedTilingOrder
Tiles are laid out in column-major order across a fixed number of rows, with all tiles from those rows placed contiguously side by side.
OrderedTilingOrderExpand
PartitionScheduler
Schedules global indices for M, N, and K axes in a partitioned matmul. Internally uses an AxisScheduler per axis.
PlaneMatmulFamily
Plane Matmul family for any precision
PlanePartitioner
Defines how to partition across planes
RowMajorTilingOrder
Tiles laid out in row-major order.
RowMajorTilingOrderExpand
StageMemoryConfig
StridedStage
Wrapper over the shared memory used for staging, abstracting its layout
StridedStageExpand
StridedStageFamily
StridedTilingLayout
Tiles follow a strided layout that often mirrors global memory layout. Not all tiles are contiguous in shared memory, but mapping is more direct.
UnitMatmulFamily
Unit Matmul family for any precision
UnitPartitioner
Defines how to partition across units

Enums§

PartitionBuffering
PartitionSchedulerScheme
Defines how partition indices are scheduled across axes.
StageEvent
Events that occur during the process of loading tiles to registers and executing inner Tile Matmuls
TilingOrderEnum
Enum for the available traits

Traits§

Stage
Stage that can be divided into tiles, with the same kind used by the tile matmul readers.
StageConfig
Configuration for the Stage matmul (SMM) level
StageEventListener
Function that is called at each StageEvent
StageFamily
Stage family for any precision
StageMatmul
Provides matrix multiplication operations at the stage level.
StageMatmulFamily
A family of StageMatmul implementations that operate with any precision.
StagePartitioner
Defines how the stage is partitioned among compute primitives (e.g., units or planes). Controls global writeback and and compute indexing.
TilingLayout
Describes how tiles are arranged in shared memory.
TilingOrder
Determines the order in which tiles are stored in shared memory, if TilingLayout is contiguous
TilingValidation