fhelium.engine.ntt.plans

Host-side construction of NTT schedules and twiddle tables.

CompactPowerOfTwoRadixNttPlan

class View source

CompactPowerOfTwoRadixNttPlan(ckks_config: CkksConfig, policy: CompactFixedRadixPolicy, *, device: str | int | torch.device | None=None)

Radix-specific outer twists and fixed cyclic-root powers.

A radix-R digit evaluates one twisted cyclic R-point NTT. Radix-4, radix-8, and radix-16 use dedicated butterflies. Every digit has the exact policy radix, and incompatible ring dimensions are rejected. This does not alias grouped radix-2 execution.

Outer tables have integral shape [prime, N - 1] and root-power tables have shape [prime, radix] on device in standard representation. Prime rows follow ckks_config.moduli exactly; the four tables are separately allocated and do not alias.

CompactRadix2NttPlan

class View source

CompactRadix2NttPlan(ckks_config: CkksConfig, *, device: str | int | torch.device | None=None)

Canonical compact forward and inverse standard-residue twiddle rows.

Compact CUDA kernels compute butterfly indices in-kernel and consume one integral [prime, coefficient] twiddle table per transform direction. Prime rows follow ckks_config.moduli exactly; final extent is $N$ and both tables use ckks_config.torch_dtype on device. Construction is functional and the forward and inverse tables do not alias.

IndexedRadix2NttPlan

class View source

IndexedRadix2NttPlan(ckks_config: CkksConfig, *, device: str | int | torch.device | None=None)

Radix-2 schedules used by the indexed backend.

Forward and inverse indices have shape [2, logN, N/2]. Expanded twiddles contain only the nontrivial odd lane and have shape [prime, logN, N/2]. No all-one even-lane tensor is allocated. Schedules use torch.int32; standard-residue twiddles use the config's integral dtype. All tensors are on device and prime rows follow ckks_config.moduli exactly. These immutable plan tables do not alias.