FHElium

fhelium.backend.rns.tables ​

RNS quotient tables and depth-indexed rescaling data.

centered_basis_extension_tables ​

function View source

python
def centered_basis_extension_tables(source: RnsContext, target: RnsContext, source_prime_ids: tuple[int, ...], target_prime_ids: tuple[int, ...]) -> dict[str, torch.Tensor]: ...

Prepare Garner and centered-extension operands for two residue bases.

Source and destination use the same word format and device. For source product Q, subtraction of Q mod p recenters the upper CRT half in each target row p. All row parameters and half digits are ordinary tensors; execution does not retain either preparation context.

rescale_inverse_matrix ​

function View source

python
def rescale_inverse_matrix(rns_context: RnsContext) -> torch.Tensor: ...

trailing_quotient_inverses ​

function View source

python
def trailing_quotient_inverses(moduli: tuple[int, ...], drop_count: int, radix: int, *, dtype: torch.dtype, device: torch.device) -> torch.Tensor: ...

Build pj−1Rmodqi for successive trailing-row removal.

Row j removes the j-th prime from the end. Only columns before that prime participate; unused columns are zero.

moddown_inverse_tables ​

function View source

python
def moddown_inverse_tables(rns_context: RnsContext) -> tuple[torch.Tensor, ...]: ...

RescaleTables ​

class View source

python
RescaleTables(rns_context: RnsContext, dropped_prime_inverses_montgomery: torch.Tensor)

Own pairwise and prefix-product inverses used by RNS rescaling.

Attributes ​

NameTypeDefault/value
rns_contextRnsContext
dropped_prime_inverses_montgomerytorch.Tensor

device ​

property

python
device: torch.device

active_views ​

method

python
def active_views(row_count: int, *, include_p: bool) -> tuple[torch.Tensor, torch.Tensor, int]: ...

Return preselected parameters, inverses, and dropped modulus.

prefix_inverse ​

method

python
def prefix_inverse(row_count: int, drop_count: int, *, include_p: bool) -> torch.Tensor: ...

Return M−1Rmodqi for each surviving prime.

M is the product of the leading drop_count active primes. The resource retains the vector for reuse with the same prime interval.

tensor_operands ​

method

python
def tensor_operands(row_count: int, drop_count: int, *, include_p: bool, ntt_context: NttContext | None=None) -> tuple[dict[str, torch.Tensor], dict[str, object]]: ...

Expose quotient tables and scalar rounding offsets for one group.

MODDOWN_RESCALE_TABLE_NAMES ​

constant View source

python
MODDOWN_RESCALE_TABLE_NAMES = ('joint_remaining_parameters', 'joint_dropped_parameters', 'joint_ordered_parameters', 'joint_p_times_base', 'joint_inverses', 'joint_product_inverse', 'joint_dropped_offset', 'joint_remaining_offset', 'joint_drop_positions', 'joint_remaining_positions')

moddown_rescale_tables ​

function View source

python
def moddown_rescale_tables(context: RnsContext, q_prime_ids: tuple[int, ...], p_prime_ids: tuple[int, ...], drop_prime_ids: tuple[int, ...], *, moddown_rounding: str='nearest', rescale_rounding: str='nearest') -> dict[str, torch.Tensor]: ...

Prepare the exact quotient of v+Pa+HP+PHD by PD.

Q and P identify actual ordered input rows; D identifies the complete group consumed by rescaling. The remaining Q rows retain their input order. All returned values are ordinary numerical Tensors on this RNS context's device.