Struct AcceleratedMatmul

pub struct AcceleratedMatmul;

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

Trait Implementations

impl<MP: MatmulPrecision> TileMatmul<MP> for AcceleratedMatmul

type Config = AcceleratedConfig

The configuration type associated with this Matmul.

type Lhs = Matrix<<MP as MatmulPrecision>::ES>

Contains Lhs data for computation

type Rhs = Matrix<<MP as MatmulPrecision>::ES>

Contains Rhs data for computation

type Accumulator = Matrix<<MP as MatmulPrecision>::EA>

Contains and accumulates results of the Tile Matmul execution

fn execute( lhs: &Self::Lhs, rhs: &Self::Rhs, out: &mut Self::Accumulator, _config: Self::Config, )

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

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

Create the container for Lhs

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

Create the container for Rhs

fn fill_lhs(tile: &Tile<MP::ES>, lhs: &mut Self::Lhs, config: Self::Config)

Fill the container of Lhs with tile data

fn fill_rhs(tile: &Tile<MP::ES>, rhs: &mut Self::Rhs, config: Self::Config)

Fill the container of Rhs with tile data

fn fill_accumulator( tile: &Tile<MP::EA>, acc: &mut Self::Accumulator, config: Self::Config, )

Fill the accumulator with data

fn write_results( out: &Self::Accumulator, slice: &mut SliceMut<Line<MP::EO>>, config: Self::Config, )

Write the content of the output container to the given slice

fn allocate_accumulator(config: Self::Config) -> Self::Accumulator

Allocate the container to receive the execution output.

fn zero_accumulator(acc: &mut Self::Accumulator, _config: Self::Config)

Fill the accumulator with zeros.

fn __expand_execute( scope: &mut Scope, lhs: <Self::Lhs as CubeType>::ExpandType, rhs: <Self::Rhs as CubeType>::ExpandType, out: <Self::Accumulator as CubeType>::ExpandType, _config: Self::Config, ) -> <() as CubeType>::ExpandType

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

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

fn __expand_fill_lhs( scope: &mut Scope, tile: <Tile<MP::ES> as CubeType>::ExpandType, lhs: <Self::Lhs as CubeType>::ExpandType, config: Self::Config, ) -> <() as CubeType>::ExpandType

fn __expand_fill_rhs( scope: &mut Scope, tile: <Tile<MP::ES> as CubeType>::ExpandType, rhs: <Self::Rhs as CubeType>::ExpandType, config: Self::Config, ) -> <() as CubeType>::ExpandType

fn __expand_fill_accumulator( scope: &mut Scope, tile: <Tile<MP::EA> as CubeType>::ExpandType, acc: <Self::Accumulator as CubeType>::ExpandType, config: Self::Config, ) -> <() as CubeType>::ExpandType

fn __expand_write_results( scope: &mut Scope, out: <Self::Accumulator as CubeType>::ExpandType, slice: <SliceMut<Line<MP::EO>> as CubeType>::ExpandType, config: Self::Config, ) -> <() as CubeType>::ExpandType

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

fn __expand_zero_accumulator( scope: &mut Scope, acc: <Self::Accumulator as CubeType>::ExpandType, _config: Self::Config, ) -> <() as CubeType>::ExpandType

impl TileMatmulFamily for AcceleratedMatmul

type Matmul<MP: MatmulPrecision> = AcceleratedMatmul

The specific TileMatmul implementation associated with this family.

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<MP: MatmulPrecision, R: Runtime>( client: &ComputeClient<R::Server, R::Channel>, 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.