Module provider 

CUDA kernel provider implementation

This module provides the CudaKernelProvider which manages pre-compiled PTX kernels for GPU execution of relational operations (join, dedup, groupby).

Modules

arith_kernels

Kernel function names in the arithmetic module

cache_kernels

Kernel function names in the cache module

circuit_kernels

Kernel function names in the circuit module

cnf_kernels

Kernel function names in the GPU CNF encoder module.

d4_kernels

Kernel function names in the GPU Decision-DNNF compiler module (CNF validation + circuit levelization).

dedup_kernels

Kernel function names in the dedup module

epistemic_kernels

Kernel function names in the epistemic module.

filter_kernels

Kernel function names in the filter module

groupby_kernels

Kernel function names in the groupby module

ilp_credit_kernels

Kernel function names in the ILP credit module.

ilp_exact_kernels

Kernel function names in the native bounded exact-induction module.

ilp_kernels

Kernel function names in the ILP module.

join_kernels

Kernel function names in the join module

kernel_paths

mc_eval_kernels

Kernel function names in the Monte Carlo evaluation module

mc_resident_kernels

Kernel function names in the GPU-resident Datalog/MC engine module.

mc_sample_kernels

Kernel function names in the Monte Carlo sampling module

neural_kernels

Kernel function names in the neural fast-path module.

pack_kernels

Kernel function names in the pack module (GPU-side key packing)

pir_kernels

Kernel function names in the PIR interning module.

sat_kernels

Kernel function names in the SAT module

scan_kernels

Kernel function names in the scan module

set_ops_kernels

Kernel function names in the set_ops module

sort_kernels

Kernel function names in the sort module

wcoj_kernels

Kernel function names in the GPU WCOJ module.

weights_kernels

Kernel function names in the weights module.

Structs

CudaKernelProvider

CUDA kernel provider for xlog GPU operations

FjDeltaCols

Column roles for one factorized delta step. The delta atom binds (carry, key); the stable relation binds (carry, value) in head column order — r_carry/r_value therefore double as the output placement (the novel buffer is built in full_r’s schema). The static side is always consumed key-first (layout-normalized by the caller), so it needs no indices here.

FjNode

One plan node: iterate the cover subatom (bulk EXPAND over the whole frontier), then refine every probe subatom (PROBE + compaction). Probe variables must already be bound.

FjPlan

A host-side Free Join plan over inputs (design §3). Callers hand-build the plan today; planner construction from binary joins (binary2fj) is a downstream integration surface.

FjSubAtom

One subatom: an atom (input_idx) restricted to the variables its next var_positions.len() physical columns bind/probe. Across the whole plan, each atom’s subatoms consume its columns in order and must partition them exactly (design §3).

HashTableU64

Bucketed hash table for u64 hashes.

HostLaunchMetadataTransferStats

HostTransferStats

JoinIndexV2

Cached build-side join index for v2 hash join.

PtxLoadProfile

Per-module PTX load timing (populated only when XLOG_WARMUP_PROFILE=1).

RadixSortScratch

Scratch buffers for stable radix sorting of u32 key/value pairs.

Enums

CompareOp

Comparison operators for filtering

JoinType

Join types for hash_join_v2

Constants

ARITH_MODULE

CACHE_MODULE

CIRCUIT_MODULE

CNF_MODULE

D4_MODULE

DEDUP_MODULE

DEFAULT_JOIN_MAX_OUTPUT

Default maximum output size for join operations. This prevents memory overflow when joining large tables with high cardinality matches.

EPISTEMIC_MODULE

FILTER_MODULE

FJ_DELTA_MAX_DOMAIN

Dense-domain bound for the spike bitmap (domain²/8 bytes = 512 MB at the bound; gate fixtures use domain ≤ 2^13).

GROUPBY_MODULE

ILP_CREDIT_MODULE

ILP_EXACT_MODULE

ILP_MODULE

JOIN_MODULE

Module names for loaded PTX modules

MC_EVAL_MODULE

MC_RESIDENT_MODULE

MC_SAMPLE_MODULE

NESTED_LOOP_TOTAL_THRESHOLD

Nested-loop join eligibility threshold (Cartesian product upper bound). The dispatcher routes to nested-loop iff num_left * num_right <= NESTED_LOOP_TOTAL_THRESHOLD; the provider validates the same invariant fail-closed before any allocation.

NEURAL_MODULE

PACK_MODULE

PIR_MODULE

SAT_MODULE

SCAN_MODULE

SET_OPS_MODULE

SORT_MODULE

WCOJ_MODULE

WEIGHTS_MODULE