gluon_authoring

Pin the tuned Triton champion at its best config, dump its .ttgir, and recover layouts with ttgir_bridge.py, which uses the compiler's own parser with round-trip proofs. Then verify layout equivalence, measure the transcription debt with parity_gate.py, and re-inject the software pipeline before optimizing further.

A faithful transcription lands below the comparator because gluon_to_ttgir omits Triton's scheduling and pipelining passes, losing cross-iteration overlap. Measure the debt with plain@ns=1, then re-inject the pipeline passes using gluon_swp.py, which wraps the compiler in-process without modifying installed files.

cdna4.async_copy imports on gfx942 but does not lower there except at the 32-bit per-thread direct-to-LDS width that generation supports. On CDNA3, authored overlap for dot-free loops uses synchronous staging with allocate_shared_memory plus the warp_pipeline_stage hint instead.

Gluon pays only when the performance residual is layout-shaped, because LDS swizzle and padding choice plus LDS deduplication are the two things plain Triton has no syntax for. If the bottleneck is not layout-related, tuning the plain Triton kernel's config is the cheaper direction.

Variants differing only by a constexpr layout constant share a Triton cache entry, so the second silently runs the first's binary and all arms tie. Give each arm its own kernel object and TRITON_CACHE_DIR, expose a fingerprint() hook, and run ab_bench.py with --permute to check whether numbers follow the code or the position.