CUDA Kernel Reference

The CUDA surface of xlog lives in kernels/, crates/xlog-cuda/src/, and crates/xlog-cuda/kernels/. Detailed per-function reference pages are generated with Doxygen from those sources during the docs build; the table below summarizes the kernel modules and what each provides.

Kernel Modules

File	Kernels	Purpose
`arith.cu`	`arith_add_`, `arith_sub_`, `arith_mul_`, `arith_div_`, `arith_mod_`, `arith_abs_`, `arith_neg_*`	Arithmetic operations for `is` expressions
`join.cu`	`hash_join_bucket_count_v2`, `hash_join_scatter_v2`, `hash_join_probe_v2`, `hash_join_semi`, `hash_join_anti`, `compute_composite_hash`	Hash joins (v2 with bucketed layout) + composite hashing
`pack.cu`	`pack_keys`, `pack_and_hash_keys`, `hash_packed_keys`, `gather_packed_rows`, `compare_packed_keys`	Key packing/hashing + packed-row utilities
`dedup.cu`	`mark_unique_*`, `compact_rows`	Sort-based deduplication
`filter.cu`	`filter_compare_`, `compact__by_mask`, `mask_{and,or,not}`	Filtering and stream compaction
`sort.cu`	`radix_histogram`, `radix_scatter_`, `init_indices`, `apply_permutation_`, `gather_keys_*`	Stable radix sort + permutation apply
`groupby.cu`	`detect_group_boundaries`, `extract_group_keys`, `groupby_`, `groupby_logsumexp_`	Sorted aggregation
`scan.cu`	`exclusive_scan_mask`, `count_mask`, `multiblock_scan_*`	Prefix sum operations
`set_ops.cu`	`concat_{u32,bytes}`, `sorted_diff_mark`	Union/difference operations
`circuit.cu`	`xgcf_forward_level`, `xgcf_backward_level_*`	XGCF circuit eval + reverse-mode gradients (probabilistic)
`sat.cu`	`sat_cdcl_solve`, `sat_check_model`, `sat_proof_check`, `sat_assert_`, `sat_xgcf_cnf_`, `sat_emit_not_phi`	GPU CDCL verifier + equivalence query construction helpers
`mc_sample.cu`	`mc_sample_bernoulli`	Bernoulli sampling (Monte Carlo inference)

See the architecture overview for how these kernels are dispatched by the runtime, and GPU Execution for the execution model.

​Kernel Modules

Kernel Modules