Expand description
The analytical model behind --backend model (S6).
It estimates relative cost within one kernel’s space from occupancy
and wave count — nothing else. Its output is labelled estimated on
every surface, and it ships only with its measured Spearman rank
correlation against real hardware attached (docs/LIMITATIONS.md): the model
is gated on measured quality, not on plausibility.
Structs§
- Device
Params - Per-SM limits by compute capability. Only the parts this project has measured on are listed; an unknown cc is an error, never a guess.
- Estimate
- One candidate’s estimate.
costis a unitless relative score within a kernel’s space — smaller is predicted faster. It is NOT a time.
Enums§
Constants§
Functions§
- device
- estimate
- Estimate one candidate. Model: blocks-per-SM limited by threads, smem and the block cap; cost = waves / occupancy — a candidate that needs more waves of less-occupied SMs is predicted slower.
- smem_
bytes - Shared-memory bytes per block for a candidate: the
[model]smem_bytesexpression in kernel.toml, over the kernel’s dimensions. - spearman
- Spearman rank correlation between two paired samples (average ranks for ties). Returns None below 3 pairs — a correlation of two points is noise dressed up as a number.