# Mid-to-low expansion benchmark

Run with the `1016306` benchmark entry point and `64199d4` ownership resolver.
All workloads use 1472 tiles, F143 projection/MLP weights with scale -4, and the
normal planning defaults. MLP and attention shapes are SigLIP defaults (729
tokens, 1152 width, 4304 MLP hidden width; 16 attention heads). ViT is the full
single-layer benchmark, including patch projection and MAP, with 378x378 input.

Each workload runs on a separate set of eight physical CPU cores; mid search
can use those eight, while expansion is serial within a workload. Workloads
run concurrently, without CPU affinity overlap. They still share memory/cache
resources. No device execution occurs. Exact commands and raw JSON accompany
this file. `run.py` launches the suite and `summarize.py` reports distributions.

The timed interval starts with an already selected MidProgram and includes
normal tile expansion, low simplification and analytical cycle costing. A
separate timer measures `lower_to_tiles`, which builds per-tile work lists.
Footprint screening, tile mapping, allocation, physical byte-transfer
preparation, scheduling, row encoding, linking, and low-program destruction
are excluded. `planning_ms` is reported separately and is not an expansion
measurement. Each retained candidate is expanded once, not scheduled.

The Repeat MLP has one top-level mid operation because its body is structured;
tile/kernel/transfer counts include that body once, not three unrolled copies.
The finalist count can exceed the requested shortlist because the planner
retains additional memory/layout alternatives.

The user stopped the exhaustive ViT sweeps before completion. Both processes
were terminated; there are no completed ViT timing distributions. The MLP and
attention JSON reports are complete. A short perf sample of the running B2
expansion is retained. Its self-time samples point primarily to byte_spans,
block-major indexing, and radix sorting. Call-chain unwinding was incomplete,
so inclusive caller percentages should not be treated as reliable attribution.