Module Ir

Affine legality queries over access index vectors (gh-ocannl-494, manifesto §6 waypoints 1-2).

A process-global census of the allocation classes whose accumulation across a schedule search exhausted a 12 GB card (gh-ocannl-550).

Assignment computations over tensor nodes.

OCANNL's C-family code generator, shared by the cc, CUDA, HIP and Metal backends: renders lowered routines as C, CUDA, HIP or MSL source through a backend's C_syntax_config, and reports per-compile censuses of the renderings it chose.

Where a compile's host wall time goes (gh-ocannl-1314).

Analytic cost model, the extraction half (gh-ocannl-491 task 1): per-kernel compulsory memory footprints, arithmetic-op counts, arithmetic intensity, and the roofline lower-bound time under advisory envelope constants (Backend_intf.hardware_limits's peak_flops / peak_memory_bandwidth).

CUDA's tensor-core tables as pure functions of storage precisions, the numerics policy and the compute capability: the one arm variant the backend's mma_syntax and mma_fragment_syntax hooks dispatch on, its arch floors, markers and tiles, the capability descriptor (Backend_intf.mma_capability) autotune seeds from, and the nvrtc arch-target policy that reads the floors back out of rendered source.

The licensed floating-point algebra over raw lowered code (gh-ocannl-998): a member of the Rewrites tier, after Online_softmax.

HIP's rocWMMA tables as pure functions of storage precisions and the numerics policy: the bf16-residency predicate, the one combination table both Hip_backend mma hooks consult, the arm table it publishes (Backend_intf.codegen_capabilities.mma_arm), and the capability descriptor (Backend_intf.mma_capability) autotune seeds from.

A weakly-owned association from a tensor node to the host buffer holding its initialization data (ndarray-backed literals and persistence-loaded nodes). After gh-ocannl-333 the data is no longer stored on Tnode.t; instead it is registered here at construction and uploaded into a backend context, on demand, at link time.

Symbols and projections used to lower tensor indexing.

Interval (min/max) lattice shared between the per-tensor bounds summaries stored on Tnode.t and the interval_of analysis over Low_level.scalar_t. See docs/proposals/interval-analysis-scalar-t.md. This is deliberately the single module defining the bounds record and its operations (binding constraint 9 of the proposal): Tnode stores plain t values; Low_level extends them with a symbol environment and source tracking.

A bounded, process-local cache of compiled kernel code, keyed by exact source identity (gh-ocannl-1314).

Outlining a large kernel (gh-ocannl-1603).

The nest view of raw lowered code shared by the recognizers of the algebraic-rewrite tier (Rewrites): Online_softmax and Prefix_scan read a routine as its top-level statements, each either a perfect serial nest with one Set at the bottom or opaque to them, plus a census of who writes and reads which node.

gh-ocannl-1526: one lowering of a routine, reused by the many compiles of an autotuning search.

What a failed Metal compile or pipeline creation reports, read off the text the metal binding raises it with (Failure): the classes the backend types, and the cause it types a compiler crash as.

Metal's simdgroup_matrix tables as pure functions of storage precisions and the numerics policy: the one combination resolver both Metal_backend mma hooks consult, the arm table it publishes (Backend_intf.codegen_capabilities.mma_arm), and the capability descriptor (Backend_intf.mma_capability) autotune seeds from.

Numerics policy (gh-ocannl-478's "option 3" knob): compute-precision decisions that change results, so they are chosen by the user — via the global config or set_policy — never by the optimizer. Storage precisions live on tensor nodes (Tnode.t.storage_prec); this record governs how computations over those storages are carried out. It must be identical across sibling autotune candidates: candidate schedules compete on speed, never on numerics (the bitwise-parity discipline of the tensorized twins depends on it).

The online-softmax attention rewrite (gh-ocannl-483): the first member of the algebraic-rewrite tier over lowered code (Rewrites), a pattern-directed substitution that CHANGES the computation where the schedule transforms only rearrange it.

The prefix-scan rewrite (gh-ocannl-1758): a member of the algebraic-rewrite tier over lowered code (Rewrites), gated by the config key prefix_scan (default off, on in the approximate profile, pinned off in reproducible).

The C-tile geometry of the register-tiled Tile_mma rendering (gh-ocannl-619).

Test-only fault-injection points for resource-owning seams.

The algebraic-rewrite tier over raw lowered code: pattern-directed substitutions that CHANGE the computation, where the schedule transforms only rearrange a fixed one. Assignments.lower applies the tier between to_low_level and Low_level.optimize, ahead of the analyses, so the traced store and the placements see the rewritten routine and a rewrite may remove a node's definition outright.

The partial-schedule decision space (gh-ocannl-514 phases 1-2): the representation branch-and-bound schedule search operates on, per docs/proposals/gh-ocannl-514.md.

A unit of scheduled work: what a backend stream executes for one routine launch or one transfer. The context_lifetime payload exists only to keep the referenced contexts (and with them their buffers) reachable for as long as the task can still run.

The vector register file the C backends' explicit-SIMD renderings target: its register width and its register count, as ONE value (gh-ocannl-1399).

The Winograd F(2x2, 3x3) convolution rewrite (gh-ocannl-505): a member of the algebraic-rewrite tier over lowered code (Rewrites), gated by the config key winograd_conv (default off, on in the approximate profile, pinned off in reproducible).