Struct DmaContext 

pub struct DmaContext<const DMA: Dma> { /* private fields */ }

Context for a specific DMA engine.

Implementations

impl<const DMA: Dma> DmaContext<DMA>

pub fn hbm_cluster_shuffle<D: Scalar, Chip: M, Element: M>( &mut self, _tensor: &HbmTensor<D, Chip, Element>, _shuffle_pattern: &[usize], ) -> HbmTensor<D, Chip, Element>

Perform cluster shuffle operation using DMA commands (HBM <-> HBM transfer). This operation redistributes data across clusters according to the shuffle pattern.

Arguments

• tensor - Input tensor with cluster dimension
• shuffle_pattern - Array mapping source cluster to destination cluster

Example

For Cluster=2 with shuffle_pattern=[1,0]:

• Data from Cluster 1 goes to Cluster 0
• Data from Cluster 0 goes to Cluster 1

pub fn hbm_chip_shuffle<D: Scalar, Chip: M, Element: M>( &mut self, _tensor: &HbmTensor<D, Chip, Element>, _shuffle_pattern: &[usize], ) -> HbmTensor<D, Chip, Element>

Perform chip shuffle using PCIe/Tensor DMA commands (HBM <-> HBM transfer across chips). This operation redistributes data across chips according to the shuffle pattern.

Arguments

• tensor - Input tensor with chip dimension
• shuffle_pattern - Array mapping source chip to destination chip

Example

For Chip=4 with shuffle_pattern=[1,2,3,0]:

• Data from Chip 1 goes to Chip 0
• Data from Chip 2 goes to Chip 1
• Data from Chip 3 goes to Chip 2
• Data from Chip 0 goes to Chip 3

impl DmaContext<{ Dma::Tensor }>

pub fn dm_cluster_shuffle<D: Scalar, const CLUSTER_DIM: usize, Chip: M, Cluster: M, Slice: M, Element: M>( &mut self, tensor: DmTensorView<'_, D, Chip, Cluster, Slice, Element>, shuffle_pattern: &[usize], ) -> DmTensor<D, Chip, Cluster, Slice, Element>

Perform cluster shuffle operation using DMA commands (DM <-> DM transfer). This operation redistributes data across clusters according to the shuffle pattern.

Arguments

• tensor - Input tensor with cluster dimension
• shuffle_pattern - Array mapping source cluster to destination cluster

Example

For Cluster=2 with shuffle_pattern=[1,0]:

• Data from Cluster 1 goes to Cluster 0
• Data from Cluster 0 goes to Cluster 1

pub fn dm_chip_shuffle<D: Scalar, const CHIP_DIM: usize, Chip: M, Cluster: M, Slice: M, Element: M>( &mut self, tensor: DmTensorView<'_, D, Chip, Cluster, Slice, Element>, shuffle_pattern: &[usize; CHIP_DIM], ) -> DmTensor<D, Chip, Cluster, Slice, Element>

Perform chip shuffle using Tensor DMA commands (DM <-> DM transfer across chips). This operation redistributes data across chips according to the shuffle pattern.

Arguments

• tensor - Input tensor with chip dimension
• shuffle_pattern - Array mapping source chip to destination chip

Example

For Chip=4 with shuffle_pattern=[1,2,3,0]:

• Data from Chip 1 goes to Chip 0
• Data from Chip 2 goes to Chip 1
• Data from Chip 3 goes to Chip 2
• Data from Chip 0 goes to Chip 3

Trait Implementations

impl<const DMA: Dma> Debug for DmaContext<DMA>

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

Formats the value using the given formatter.