Enum ArgSource 

pub enum ArgSource {
    TensorAddr(TensorId),
    ItemStart,
    ItemCount,
    ItemY0,
    ItemX0,
    ItemCountX,
    McastStart {
        operand: Operand,
        axis: Axis,
    },
    McastEnd {
        operand: Operand,
        axis: Axis,
    },
    McastNumDests {
        operand: Operand,
    },
    McastSender {
        operand: Operand,
        axis: Axis,
    },
    SemHandle(SemId),
    CyclesAddr,
    CyclesSlot(u8),
    TensorDim {
        tensor: TensorId,
        axis: Axis,
    },
}

Variants

TensorAddr(TensorId)

DRAM base address of a tensor.

ItemStart

First work item (linear id) of the site — items are consecutive in row-major block order.

ItemCount

Number of work items of the site.

ItemY0

Strided item enumeration (a wrapped blocked plan): first block row / column and the count per row.

ItemX0

ItemCountX

McastStart

Multicast rectangle of the RECEIVERS of the site’s group for an operand (near corner / far corner).

Fields

operand: Operand

axis: Axis

McastEnd

Fields

operand: Operand

axis: Axis

McastNumDests

Fields

operand: Operand

McastSender

Physical coordinate of the sender of the site’s group for an operand.

Fields

operand: Operand

axis: Axis

SemHandle(SemId)

The program-wide semaphore handle (get_semaphore(id) in the kernel).

CyclesAddr

Base of the device-cycle scratch, and this site’s word in it. Stamped by the kernel that writes the output, whose lifetime spans the whole program on its core.

CyclesSlot(u8)

TensorDim

A tensor’s tile-grid extent (rows_t / cols_t), handed to a data-movement kernel at run time: the ONLY way shape enters a kernel. Every tile stride and loop count is printed as a block-extent literal times or over one of these, so the address arithmetic has exactly the vendor’s form — constant-folding the shape restructures it and slows the NOC issue rate (+844 cycles at 640³ on ttsim), while a runtime value the compiler cannot see is 1 costs a multiply per tile (+300 cycles). A movement kernel is thereby shape-generic: the same binary serves every shape of the same blocking.

Fields

tensor: TensorId

axis: Axis

Trait Implementations

impl Clone for ArgSource

fn clone(&self) -> ArgSource

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ArgSource

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

Formats the value using the given formatter.

impl PartialEq for ArgSource

fn eq(&self, other: &ArgSource) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Eq for ArgSource

impl StructuralPartialEq for ArgSource