fhelium.config.ntt

Configuration and compatibility checks for CPU and CUDA NTT backends.

IndexedRadix2Policy

class View source

IndexedRadix2Policy(name: str)

Table-driven indexed radix-2 execution on CPU and CUDA.

CPU parallelizes the flattened batch/limb/butterfly space separately for each stage, with a stage barrier preserving transform dependencies. CUDA launches one kernel per stage. The same indexed schedule also provides the cross-device baseline for validating compact CUDA policies.

Attributes

Name	Type	Default/value
name	str

CompactRadix2Policy

class View source

CompactRadix2Policy(name: str, grouped_radix2_stage_count: Literal[2, 3, 4])

Grouped radix-2 execution over canonical compact twiddle rows.

grouped_radix2_stage_count counts radix-2 stages fused into a global kernel. The smem8 backend-name suffix records the current native production implementation, whose shared-memory tile and fusion depth are compiled CUDA resources rather than Python policy parameters.

Attributes

Name	Type	Default/value
name	str
grouped_radix2_stage_count	Literal[2, 3, 4]

group_width

property

group_width: int

Number of coefficients held by one grouped radix-2 tuple.

CompactFixedRadixPolicy

class View source

CompactFixedRadixPolicy(name: str, radix: Literal[4, 8, 16])

Strict execution in which every digit has exactly one radix.

Let $N=2^L$ and $\mathtt{radix}=2^b$. A compatible transform has $D=L/b$ fixed-radix digits and is rejected when $L$ is not divisible by $b$. Shared-memory tile size and fusion depth are native CUDA implementation choices: they do not affect the mathematical policy, table layout, or backend identity represented here.

Attributes

Name	Type	Default/value
name	str
radix	Literal[4, 8, 16]

radix_bits

property

radix_bits: Literal[2, 3, 4]

Return ${\log }_2(\mathtt{radix})$, the bits covered by one digit.

NttBackendPolicy

data View source

NttBackendPolicy: TypeAlias = IndexedRadix2Policy | CompactRadix2Policy | CompactFixedRadixPolicy

NTT_BACKEND_POLICIES

constant View source

NTT_BACKEND_POLICIES: Final = MappingProxyType({policy.name: policy for policy in _POLICY_SEQUENCE})

SUPPORTED_NTT_BACKENDS

constant View source

SUPPORTED_NTT_BACKENDS: Final = tuple(NTT_BACKEND_POLICIES)

DEFAULT_NTT_BACKEND

constant View source

DEFAULT_NTT_BACKEND: Final[str] = 'radix2_compact_group8_smem8'

DEFAULT_CPU_NTT_BACKEND

constant View source

DEFAULT_CPU_NTT_BACKEND: Final[str] = 'radix2_indexed'

resolve_ntt_backend_policy

function View source

def resolve_ntt_backend_policy(name: str) -> NttBackendPolicy: ...

Return the exact policy named by configuration.

Names are deliberately not case-normalized and no compatibility aliases are accepted.

validate_ntt_backend_for_log_n

function View source

def validate_ntt_backend_for_log_n(policy: NttBackendPolicy, log_ring_dimension: int) -> None: ...

Reject a named policy that cannot factor the requested ring size.

compatible_ntt_backends

function View source

def compatible_ntt_backends(log_ring_dimension: int) -> tuple[str, ...]: ...

Return canonical policy names executable for one logN.

Names retain registry order. Strict fixed-radix policies whose digit width does not divide log_ring_dimension are omitted; grouped radix-2 policies remain available for every supported ring dimension. A non-positive dimension raises ValueError.