Module target 

SpatialTarget — the described MESH a spatial program is placed onto.

One descriptor per chip class: how many sites, what each site holds (local memory, accumulator tiles, movement engines, NOCs), the native compute tile, which interconnect and handshake primitives exist, and which compute capabilities (tile matmul, elementwise, row-reduce, a scalar datapath) the site offers. The passes read ONLY this descriptor; every number is cited at its constructor, and the fields that a public datasheet does not give are listed in modeled so the descriptor never pretends to know them.

The scalar datapath capability EMBEDS the existing FabricTarget descriptor — the fabric lane is a consumer of the spatial IR, not a sibling of it.

Structs

CreditLatency

The latencies a bounded buffer’s credit loop is sized against, in cycles.

MeshShape

The mesh of sites: x columns × y rows.

NocRule

The NOC-direction convention a mesh imposes on multicast roles — measured, not chosen: on Wormhole the in0 senders must multicast on NOC 0 and the receivers listen on NOC 1 (all-on-one-NOC inverted the multicast rectangle under ttsim); the writer takes the complement, except the first column.

NumericsKnobs

The three numerics knobs the measurement harness varies, and the PCC the program then requires of itself. scripts/tt-matmul-measure.sh sets them as LOGOS_TT_FIDELITY / _FP32 / _L1ACC, and every ledger row records what the program PRINTED rather than what was asked for — so this parser is part of the measurement contract, not a convenience.

SiteSpec

What one site holds.

SpatialTarget

The described mesh.

TileShape

Geometry of one native compute tile — 32×32 on Tensix, a 16×16 fragment on a tensor core, 1×1 for a scalar fabric cell.

Enums

BlockApi

How a matrix engine’s sub-block API runs one compute op over an osh × osw sub-block and its ibw inner steps: the calls it makes and the operand tiles it loads.

ComputeCap

A compute capability a site offers.

DataflowPolicy

Which operand stays resident in the compute engine while the other streams.

Fidelity

Matmul math fidelity of a tile engine.

HandshakeKind

The synchronization primitive an edge may use.

Interconnect

How sites exchange data.

PackerAccumulate

Whether a matrix engine’s packer can ADD into a local-memory page instead of overwriting it.

WriteOrdering

Whether NOC writes retire in issue order (Wormhole) or need an explicit flush before a dependent semaphore multicast (Blackhole).

Functions

int_precision_max

The largest value a signed integer precision holds.

int_precision_of_bits

The signed integer precision of exactly bits bits, if the dial has one.

precision_bytes

Bytes per element of a precision.

precision_int_bits

Bit width of an integer precision (the operand truncation the systolic reference applies).

precision_is_int

Whether a precision is an integer format.

sim_grid

The compute-with-storage grid the cycle-accurate simulator exposes for a chip id (slow dispatch): wormhole_b0_80_arch.yaml → 8×9, blackhole_140_arch.yaml → 13×10 (one column of the 14×10 reserved). Read from the vendor examples’ compute_with_storage_grid_size() on each descriptor.