fhelium.engine.rns.parameters

Device-resident parameter tables for RNS operations.

RnsRowParameters

class View source

RnsRowParameters(prime_ids: tuple[int, ...], parameter_row_start: int, native_parameters: torch.Tensor, montgomery_reduction_parameters: tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor], montgomery_r2: torch.Tensor, scaled_montgomery_r2: torch.Tensor, twice_modulus: torch.Tensor, moduli: tuple[int, ...], mixed_radix_normalizers: torch.Tensor | None = None, basis_extension_coefficients: torch.Tensor | None = None, mixed_radix_propagation_coefficients: torch.Tensor | None = None)

Device tables for one exact contiguous prime_ids interval.

Every tensor is integral on one execution device. One-dimensional tables have shape [limb] in prime_ids order. montgomery_r2[j] is $R^{2}mod{q}_i$, scaled_montgomery_r2[j] is ${\mathrm{\Delta }}_0{R}^{2}mod{q}_i$, and twice_modulus[j] is $2q_i$ for i=prime_ids[j]. The four reduction vectors are respectively the low and high split words of $q_i$ and $-q_i^{-1}modR$.

Mixed-radix tables exist only for multi-row source digits. Normalizers have shape [digit - 1]; propagation coefficients have shape [digit - 1, digit]; basis-extension coefficients have shape [digit - 1, destination_limb] in canonical level-zero QP destination order. Their entries include the Montgomery factors required by their native consumers.

parameter_row_start identifies the first row in the engine's canonical level-zero QP order. native_parameters is the cached zero-copy [parameter, limb] view consumed by native RNS kernels.

Attributes

Name	Type	Default/value
prime_ids	tuple[int, ...]
parameter_row_start	int
native_parameters	torch.Tensor
montgomery_reduction_parameters	tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]
montgomery_r2	torch.Tensor
scaled_montgomery_r2	torch.Tensor
twice_modulus	torch.Tensor
moduli	tuple[int, ...]
mixed_radix_normalizers	torch.Tensor | None	None
basis_extension_coefficients	torch.Tensor | None	None
mixed_radix_propagation_coefficients	torch.Tensor | None	None

RnsParameterStore

class View source

RnsParameterStore(*, rns_layout: RnsLayout, montgomery_parameters: MontgomeryParameters, device: torch.device, torch_dtype: torch.dtype, rns_basis_level_count: int, level_row_starts: list[int], basis_row_stops: tuple[int, int], montgomery_reduction_parameter_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor], native_parameter_tensor: torch.Tensor, montgomery_r2: torch.Tensor, scaled_montgomery_r2: torch.Tensor, twice_modulus: torch.Tensor, moduli: list[int])

Build engine-owned parameter views for RNS basis extension.

A process owns one device and one dense canonical [Q | P] prime order. Level $\ell$ selects the contiguous interval beginning at Q prime id level; a Q basis ends before P and a QP basis includes the fixed P suffix. Views preserve this exact order and do not allocate or mutate the source tables. The store contains no device fanout or communication policy.

row_parameters

method

def row_parameters(key) -> RnsRowParameters: ...

basis_parameters

method

def basis_parameters(level: int, *, include_p: bool=False) -> RnsRowParameters: ...

twice_modulus_for_basis

method

def twice_modulus_for_basis(level: int, *, include_p: bool=False) -> torch.Tensor: ...

moduli_for_basis

method

def moduli_for_basis(level: int, *, include_p: bool=False) -> list[int]: ...