Struct SwizzleSpanMatmul

pub struct SwizzleSpanMatmul<const W: u32> {}

Iterates on global matmuls following the swizzle algorithm

The swizzle algorithm processes W elements per row in a top-down pass, then shifts to the next W columns in a bottom-up pass. This zigzag (top-down, bottom-up) repeats, covering the matrix span by span.

Trait Implementations

impl<const W: u32> CubeType for SwizzleSpanMatmul<W>

type ExpandType = SwizzleSpanMatmulExpand<W>

fn init(scope: &mut Scope, expand: Self::ExpandType) -> Self::ExpandType

Wrapper around the init method, necessary to type inference.

impl<const W: u32> SpanMatmul for SwizzleSpanMatmul<W>

fn execute<MP: MatmulPrecision, GMM: GlobalMatmul<MP>>( lhs: VirtualTensor<MP::EI>, rhs: VirtualTensor<MP::EI>, out: VirtualTensor<MP::EO, ReadWrite>, span: Span, acc: GMM::Accumulator, k_range: (u32, u32), quantization: CubeOption<Quantization<MP>>, config: GMM::Config, )

fn __expand_execute<MP: MatmulPrecision, GMM: GlobalMatmul<MP>>( context: &mut Scope, lhs: <VirtualTensor<MP::EI> as CubeType>::ExpandType, rhs: <VirtualTensor<MP::EI> as CubeType>::ExpandType, out: <VirtualTensor<MP::EO, ReadWrite> as CubeType>::ExpandType, span: <Span as CubeType>::ExpandType, acc: <GMM::Accumulator as CubeType>::ExpandType, k_range: <(u32, u32) as CubeType>::ExpandType, quantization: <CubeOption<Quantization<MP>> as CubeType>::ExpandType, config: GMM::Config, ) -> <() as CubeType>::ExpandType