Struct AcceleratedMatmul 

pub struct AcceleratedMatmul<Acc: TileKind> { /* private fields */ }

Uses one plane to perform a small matmul using accelerated instructions.

Trait Implementations

impl<L: Numeric, R: Numeric, A: Numeric, AccTile: TileKind> TileMatmul<L, R, A> for AcceleratedMatmul<AccTile>

where CmmaStageReader<AccTile>: CmmaFragmentReader<TileKind = AccTile>,

type Config = AcceleratedConfig

The configuration type associated with this Matmul.

type LhsFragment = Matrix<L>

Contains Lhs data for computation

type RhsFragment = Matrix<R>

Contains Rhs data for computation

type AccFragment = Matrix<A>

Contains and accumulates results of the Tile Matmul execution

type LhsTile = Strided

Tile for the lhs data

type RhsTile = Strided

Tile for the rhs data

type AccTile = AccTile

Tile for the accumulator data

type OutTile = Strided

Tile for the output data

fn execute( lhs: &Self::LhsFragment, rhs: &Self::RhsFragment, out: &mut Self::AccFragment, _config: Self::Config, )

Executes the matrix multiplication of Lhs and Rhs, adding the result to the accumulator

fn allocate_lhs(config: Self::Config) -> Self::LhsFragment

Create the container for Lhs

fn allocate_rhs(config: Self::Config) -> Self::RhsFragment

Create the container for Rhs

fn load_lhs<E: Numeric>( tile: &StridedTile<E>, lhs: &mut Self::LhsFragment, _config: Self::Config, )

Load the container of Lhs from tile data

fn load_rhs<E: Numeric>( tile: &StridedTile<E>, rhs: &mut Self::RhsFragment, _config: Self::Config, )

Load the container of Rhs from tile data

fn load_acc<E: Numeric>( tile: &AccTile::Tile<E>, acc: &mut Self::AccFragment, config: Self::Config, )

Load the container of Acc from tile data

fn write_results<E: Numeric>( tile: &mut StridedTile<E, ReadWrite>, out: &Self::AccFragment, _config: Self::Config, )

Write the content of the output container to the given slice

fn allocate_acc(config: Self::Config) -> Self::AccFragment

Allocate the container to receive the execution output.

fn __expand_execute( scope: &mut Scope, lhs: <Self::LhsFragment as CubeType>::ExpandType, rhs: <Self::RhsFragment as CubeType>::ExpandType, out: <Self::AccFragment as CubeType>::ExpandType, _config: Self::Config, ) -> <() as CubeType>::ExpandType

fn __expand_allocate_lhs( scope: &mut Scope, config: Self::Config, ) -> <Self::LhsFragment as CubeType>::ExpandType

fn __expand_allocate_rhs( scope: &mut Scope, config: Self::Config, ) -> <Self::RhsFragment as CubeType>::ExpandType

fn __expand_load_lhs<E: Numeric>( scope: &mut Scope, tile: <StridedTile<E> as CubeType>::ExpandType, lhs: <Self::LhsFragment as CubeType>::ExpandType, _config: Self::Config, ) -> <() as CubeType>::ExpandType

fn __expand_load_rhs<E: Numeric>( scope: &mut Scope, tile: <StridedTile<E> as CubeType>::ExpandType, rhs: <Self::RhsFragment as CubeType>::ExpandType, _config: Self::Config, ) -> <() as CubeType>::ExpandType

fn __expand_load_acc<E: Numeric>( scope: &mut Scope, tile: <AccTile::Tile<E> as CubeType>::ExpandType, acc: <Self::AccFragment as CubeType>::ExpandType, config: Self::Config, ) -> <() as CubeType>::ExpandType

fn __expand_write_results<E: Numeric>( scope: &mut Scope, tile: <StridedTile<E, ReadWrite> as CubeType>::ExpandType, out: <Self::AccFragment as CubeType>::ExpandType, _config: Self::Config, ) -> <() as CubeType>::ExpandType

fn __expand_allocate_acc( scope: &mut Scope, config: Self::Config, ) -> <Self::AccFragment as CubeType>::ExpandType

impl<Tile: TileKind> TileMatmulFamily for AcceleratedMatmul<Tile>

where CmmaStageReader<Tile>: CmmaFragmentReader<TileKind = Tile>,

type Matmul<L: Numeric, R: Numeric, A: Numeric> = AcceleratedMatmul<Tile>

The specific TileMatmul implementation associated with this family.

type LhsTile = Strided

Tile kind for Lhs

type RhsTile = Strided

Tile kind for Rhs

type AccTile = Tile

Tile kind for Acc

type OutTile = Strided

Tile kind for Out

type Config = AcceleratedConfig

The configuration type associated with this matmul family.

fn requires_accelerator() -> bool

Returns whether this tile matmul requires specialized hardware accelerators (e.g., tensor cores).

fn computation_resources() -> Result<ComputeResources, InvalidConfigError>

Returns the compute resources required to run this tile matmul.

fn setup<Lhs: Numeric, Rhs: Numeric, Acc: Numeric, R: Runtime>( client: &ComputeClient<R::Server>, problem: &MatmulProblem, selection: &MatmulSelection, matmul_line_sizes: &MatmulLineSizes, ) -> Result<Self::Config, MatmulSetupError>

Constructs the configuration based on the matmul problem, selection, and line sizes.

fn filter_line_sizes( available_line_sizes: AvailableLineSizes, ) -> AvailableLineSizes

Filters out line sizes that are incompatible with this matmul family.