fhelium.backend.triton._ntt

Compose generated NTT endpoints and synchronized native middle stages.

A schedule fuses component expressions into transform loads and stores. Each interior stage group retains its native synchronization and scratch storage. Successive transforms materialize the preceding transform's live result while its pointwise consumers can become the next transform's input expressions.

CompactNttExecution

class View source

CompactNttExecution(output_roots: tuple[ValueExpr, ...], outputs: tuple[ValueExpr, ...], transforms: tuple[_Transform, ...], dtype: torch.dtype, device: torch.device)

Prepared compact radix-2 execution for one input layout.

Preparation records table operand positions and emits kernels. Execution accepts current Tensor operands, allocates result/scratch storage and launches those kernels; no input or evaluation-key Tensor is retained by the plan.

Attributes

Name	Type	Default/value
output_roots	tuple[ValueExpr, ...]
outputs	tuple[ValueExpr, ...]
transforms	tuple[_Transform, ...]
dtype	torch.dtype
device	torch.device
execution_source	str	field(init=False, repr=False, compare=False)

build

method

def build(graph: BoundGraph, values: tuple[torch.Tensor, ...], tables: tuple[torch.Tensor, ...]) -> CompactNttExecution: ...

execute

method

def execute(values: tuple[torch.Tensor, ...], tables: tuple[torch.Tensor, ...]) -> tuple[torch.Tensor, ...]: ...

Execute the prepared compact schedule using fresh result and scratch storage.