FHElium

Appearance

Choose operation implementations and lowerings ​

Implementation selection chooses the Backend code that executes an operation. Lowering selection chooses the operations that represent the calculation. Use these controls to retain a whole CKKS operation, expose RNS/NTT work, or request a particular implementation for a measured workload.

Prerequisites ​

Have a correct Eager evaluator or a Compilation and a clear-message reference. Keep the input state, parameter set, and measurement scope fixed while comparing candidates. See Build and transform a Program for a manual compilation and Compile a callable for a callable interface.

1. Inspect the available implementations ​

python
from fhelium.backend import OperationBackend

backend = OperationBackend()
for implementation in backend.registry.implementations:
    print(implementation.name, [op.name for op in implementation.operation_types])

An implementation name is resolved together with an operation class. The presence of a name does not imply support for every operation, representation, device, or in-place call. OperationBackend owns the selected registry and its resource workspace; placement follows Tensor operands and the execution implementation's requirements.

2. Record a per-operation requirement ​

An NTT assignment can request the native implementation and its transform policy:

python
from fhelium import compile as fc

ntt_selection = fc.AssignNttImplementationPass(
    "native-ntt", ntt_backend="radix2_indexed"
)

For other operations, AssignImplementationsPass(selections={operation_name: implementation_name}) records the requested name on matching operations. Obtain both names from the registry inventory. overwrite=True permits replacing a prior assignment; otherwise a conflicting assignment fails. This pass records a constraint and does not establish executable coverage.

Insert assignment where the intended operations exist and before a transformation that could replace them. Inspect pass decisions and the transformed Program to confirm that the requirement reached the expected operations. The default recipe performs local lowering and fusion selection; it does not benchmark alternatives or retry an execution failure with another implementation.

3. Preserve a whole operation or expose its numerical decomposition ​

python
lowering = fc.LowerCkksToRnsNttPass(
    selections={"fhelium_ckks.rescale": "rns-drop-leading-prime"},
    preserve={"fhelium_ckks.multiply"},
)

This pass preserves whole CKKS multiplication and selects the named rescale lowering. The recipe name is an implementation identifier; the mathematical rescale divisor still follows the configured dropped depth group. The resulting mixed-level Program can contain both whole-operation implementations and RNS/NTT operations.

An assigned whole operation must be preserved when applying this lowering pass. Lowering rejects a selected operation if the rewrite would erase its recorded implementation assignment. Unhandled operations can remain in the Program for another pass or an implementation at their current level.

4. Apply a transform policy in Eager ​

python
import fhelium as fh
from fhelium.eager import Engine

engine = Engine(
    fh.Preset.slots8192_scale40_depth7_int64,
    ntt_backend="radix2_indexed",
)

This selects the Engine's NTT policy while the actual device is chosen by factories or operands. Keep key material on the operation device. For CUDA alternatives, screen compatible NTT backends, then interpret the timing evidence. The retained legacy screening runner measures its own evaluator, so confirm the result with the current execution path being deployed.

5. Validate and record the selected execution ​

Prepare materials, link with the same Backend, and inspect executable.manifest for bound implementation and resource information. Compare decoded results against the clear-message reference before accepting performance evidence. Record whether timing includes transformation, preparation, linking, first device-code compilation, input staging, and steady execution.

See Operation declaration and implementation selection for registration, RNS and NTT for numerical resources, and Generated kernels and fusion for fused execution.