Expand description
Bench plans (plan.v1): everything the runner needs, self-contained —
candidate configs, PTX paths, launch geometry, and the argument layout
matching cuda-oxide’s PTX parameter lowering (each slice becomes a
ptr, len pair of .param slots; scalars are single slots).
A kernel’s [bench] section in kernel.toml declares the workload;
sizes and grid shapes are arithmetic expressions over the kernel’s
dimensions plus the elements variable.