| # Optimization ledger — one row per variant, never delete a row | |
| The point: a promising optimization abandoned over a *fixable bug* is the most common way to leave 2x | |
| on the table. So the "why dropped" column distinguishes **slower** from **merely broken**. Anything | |
| marked `broken` is a candidate to revisit, not a dead end. | |
| Keep the last-known-good runnable at all times. | |
| | # | variant | correct? | metric | vs best | status | why dropped | | |
| |---|---|---|---|---|---|---| | |
| | 0 | reference (spec) | yes | — | 1.00x | baseline | — | | |
| | 1 | naive port | | | | | | | |
| | 2 | + shared-mem tiling | | | | | | | |
| | 3 | + async copy / TMA | | | | | | | |
| | 4 | + warp specialization | | | | | | | |
| | 5 | shape-specialized dispatch | | | | | | | |
| status ∈ {best, kept, slower, **broken**, untested} | |
| ## Time attribution — where the remaining time goes | |
| Update after each profile. Attack the largest row; never stop because a percentage looks good. | |
| | component | us | % of total | bounded by | idea | | |
| |---|---|---|---|---| | |
| | | | | | | | |