Module kernels 

Runtime-dispatched compute kernels. Runtime-dispatched dot-product and packed-bit Hamming kernels.

Callers validate dimensions once at the engine boundary. Every hot kernel therefore uses debug_assert! for the equal-length and row-shape contracts instead of repeating release-mode validation in the scan loop. Call crate::kernels::initialize during engine initialization to apply the optional ZE_KERNEL=scalar|neon|avx2 override and surface unsupported requests as a typed crate::kernels::KernelInitError. If no explicit initialization occurs, the first kernel call safely caches the best runtime-detected arm.

The scalar implementation is the behavioral oracle. crate::kernels::KernelVariant exposes every runtime-supported table to property tests, fuzzing, and the future Task 27 frontier harness without moving selection to build time.

Modules

postings

Posting-block decode kernels: bit-unpack and delta prefix sum. Posting-block decode kernels: bit-unpack and delta prefix sum.

Structs

BaselineKernelConfig

Baseline micro-shape selected before the Task 27 B1 campaign.

Bit4Row

One bounds-validated packed four-bit row borrowed from a mapped region.

Bit4Rows4

Four equally-sized packed rows whose addresses were validated before scoring.

KernelFeatures

Runtime CPU capabilities relevant to current and reserved kernel tiers.

KernelVariant

One concrete runtime dispatch table.

KnobSpace

Enumerable tuning declarations consumed directly by Task 27-H.

Enums

GatherShapeError

Rejected shape at the safe four-row gather boundary.

InstructionTier

Instruction tiers enumerated for the Task 27 B1 frontier campaign.

KernelArm

User-selectable runtime kernel families.

KernelInitError

Typed failures produced while applying a runtime kernel override.

Constants

BASELINE_KERNEL_CONFIG

Shipped baseline configuration; B1 may replace every value with evidence.

KERNEL_KNOB_SPACE

Registered Task 27-H kernel tuning space.

MAX_DOT_I8_DIMENSION

Largest supported dimension for an i8 dot product.

Functions

detected_features

Returns cached runtime feature detection, including reserved B1 features.

dot_bit4

Scores a signed-byte query against one packed four-bit row.

dot_bit4_batch

Scores one signed-byte query against contiguous packed four-bit rows.

dot_f16

Computes an f32 dot product from equal-length IEEE f16 bit vectors.

dot_f32

Computes the dot product of equal-length f32 vectors.

dot_i8

Computes the dot product of equal-length signed-byte vectors.

dot_i8_batch

Scores one i8 query against contiguous row-major vectors.

hamming_u1

Computes Hamming distance over equal-length packed-byte vectors.

hamming_u1_batch

Scores one packed-bit query against contiguous row-major byte vectors.

initialize

Applies runtime feature detection and the optional ZE_KERNEL override.

is_arm_supported

Reports whether a user-selectable arm is executable on this CPU.

prefetch_bit4_row_group

Prefetches one already-validated four-row gather group.

prefetch_bit4_rows

Prefetches caller-selected packed rows into L1 for a later gather group.

score_bit4_ptrs

Scores four arbitrary validated packed four-bit rows through runtime dispatch.

selected_arm

Returns the cached arm, selecting the best safe runtime arm if necessary.