Struct StreamTensor 

pub struct StreamTensor<'l, const T: Tu, P: Position, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> { /* private fields */ }

Tensor streamed through the pipeline.

Implementations

impl<'l, const T: Tu, P: Position, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> StreamTensor<'l, T, P, D, Chip, Cluster, Slice, Time, Packet>

pub type Mapping = Pair<Chip, Pair<Cluster, Pair<Slice, Pair<Time, Packet>>>>

Mapping type alias.

pub fn new(ctx: &'l mut TuContext<T>, inner: Tensor<D, Self::Mapping>) -> Self

Creates a new stream tensor.

impl<'l, const T: Tu, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> StreamTensor<'l, T, PositionBegin, D, Chip, Cluster, Slice, Time, Packet>

pub fn fetch<D2: Scalar, Time2: M, Packet2: M>( self, ) -> FetchTensor<'l, T, D2, Chip, Cluster, Slice, Time2, Packet2>

where D: FetchCast<D2>,

Performs fetch operation to create a fetched tensor.

impl<'l, const T: Tu, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> StreamTensor<'l, T, PositionFetch, D, Chip, Cluster, Slice, Time, Packet>

pub fn switch<Slice2: M, Time2: M>( self, config: SwitchConfig, ) -> SwitchTensor<'l, T, D, Chip, Cluster, Slice2, Time2, Packet>

Performs switching operation to create a switched tensor.

Applies switching network routing only. The packet passes through unchanged — no padding, no reshaping. Use SwitchTensor::collect afterwards to normalize the packet to flit-sized chunks.

pub fn collect<Time2: M, Packet2: M>( self, ) -> CollectTensor<'l, T, D, Chip, Cluster, Slice, Time2, Packet2>

Skips the switching network and goes directly to collect.

Slice and Time are preserved from fetch; only the packet is normalized to flit-sized chunks.

impl<'l, const T: Tu, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> StreamTensor<'l, T, PositionSwitch, D, Chip, Cluster, Slice, Time, Packet>

pub fn collect<Time2: M, Packet2: M>( self, ) -> CollectTensor<'l, T, D, Chip, Cluster, Slice, Time2, Packet2>

Normalizes packet to exactly 32 bytes (one flit).

Pads to flit-aligned boundary, then splits: inner 32 bytes become Packet2, outer flit portion is absorbed into Time2. For packets already ≤ 32 bytes, only padding is added.

impl<'l, const T: Tu, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> StreamTensor<'l, T, PositionCollect, D, Chip, Cluster, Slice, Time, Packet>

pub fn to_trf<Row: M, Element: M>( self, address: TrfAddress, ) -> TrfTensor<D, Chip, Cluster, Slice, Row, Element>

Stores to the tensor register file.

pub fn to_vrf<Element2: M>( self, address: Address, ) -> VrfTensor<D, Chip, Cluster, Slice, Element2>

where D: VeScalar,

Stores to the vector register file.

pub fn vector_init( self, ) -> VectorInitTensor<'l, T, D, Chip, Cluster, Slice, Time, Packet>

where D: VeScalar,

Initializes Vector Engine processing for this tensor.

pub fn align<OutTime: M, OutPacket: M, Row: M, TrfElement: M>( self, trf_tensor: &TrfTensor<D, Chip, Cluster, Slice, Row, TrfElement>, ) -> AlignedPair<'l, T, D, Chip, Cluster, Slice, Row, OutTime, OutPacket>

where Chip: M, Cluster: M, Slice: M,

Aligns LHS stream (via Feed Buffer) and RHS TRF (via TRF Sequencer) to computation shape.

pub fn transpose<OutTime: M, OutPacket: M>( self, ) -> TransposeTensor<'l, T, D, Chip, Cluster, Slice, OutTime, OutPacket>

Performs transpose operation.

impl<'l, const T: Tu, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> StreamTensor<'l, T, PositionContraction, D, Chip, Cluster, Slice, Time, Packet>

pub fn vector_init( self, ) -> VectorInitTensor<'l, T, D, Chip, Cluster, Slice, Time, Packet>

where D: VeScalar,

Initializes Vector Engine processing from contraction output.

impl<'l, const T: Tu, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> StreamTensor<'l, T, PositionVectorFinal, D, Chip, Cluster, Slice, Time, Packet>

pub fn to_vrf<Element: M>( self, address: Address, ) -> VrfTensor<D, Chip, Cluster, Slice, Element>

where D: VeScalar,

Stores to the vector register file after VE pipeline.

pub fn transpose<Time2: M, Packet2: M>( self, ) -> TransposeTensor<'l, T, D, Chip, Cluster, Slice, Time2, Packet2>

Performs transpose operation (transitions to Transpose engine).

impl<'l, const T: Tu, P: Position, D: Scalar, Chip: M, Cluster: M, Slice: M, Time: M, Packet: M> StreamTensor<'l, T, P, D, Chip, Cluster, Slice, Time, Packet>

pub fn commit<Element: M>( self, address: Address, ) -> DmTensor<D, Chip, Cluster, Slice, Element>

Commits to the data memory.

pub fn commit_view<Element: M>( self, dst: DmTensorViewMut<'l, D, Chip, Cluster, Slice, Element>, )

Commits to mutable tensor view in the data memory.

Trait Implementations

impl<'l, const T: Tu, P: Debug + Position, D: Debug + Scalar, Chip: Debug + M, Cluster: Debug + M, Slice: Debug + M, Time: Debug + M, Packet: Debug + M> Debug for StreamTensor<'l, T, P, D, Chip, Cluster, Slice, Time, Packet>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.