# MLP layout sweep

59 completed; 49 hardware passes.

Calibration statistics exclude adaptive cross-products; differences ≤24 cycles are ties.

The third estimate replaces expanded exchange cost with the final placed schedule.
It is not the earlier provisional score used for finalist selection.

Runtime starts at the renderer’s default cut: the latest tile’s initial profile entry.

| Plan | Renderer cycles | Compact | Expanded | With final exchange |
|---|---:|---:|---:|---:|
| down-22 | 186,630 | 310,377 | 248,066 | 228,114 |
| up-15 | 187,842 | 311,951 | 247,784 | 226,302 |
| cross-cropped-0-0 | 187,986 | 311,951 | 247,784 | 226,409 |
| historical | 188,196 | 310,377 | 248,066 | 229,671 |
| up-11 | 188,364 | 310,833 | 250,218 | 231,960 |
| up-19 | 189,822 | 310,377 | 249,740 | 231,345 |
| down-05 | 189,930 | 314,766 | 247,303 | 226,807 |
| default | 190,578 | 303,439 | 275,103 | 238,638 |
| down-08 | 191,220 | 303,439 | 275,103 | 239,277 |
| down-13 | 191,274 | 315,602 | 245,771 | 227,843 |
| cross-cropped-2-2 | 193,290 | 315,222 | 250,364 | 235,935 |
| cross-cropped-0-2 | 193,998 | 316,340 | 247,674 | 229,225 |
| cross-cropped-2-0 | 194,052 | 310,833 | 250,218 | 238,093 |
| up-02 | 195,216 | 310,377 | 247,983 | 237,775 |
| up-01 | 196,482 | 331,895 | 266,344 | 239,986 |

## Calibration

| Estimate | Median predicted/measured | Rank correlation | Selection regret |
|---|---:|---:|---:|
| compact_cycles | 1.649 | 0.885 | 3,948 |
| expanded_cycles | 1.329 | 0.848 | 14,124 |
| refined_cycles | 1.222 | 0.931 | 1,212 |
| compact_exchange | 2.378 | 0.775 | 11,974 |
| expanded_exchange | 1.355 | 0.735 | 1,410 |

## Best constraints

```
0:4x92x4:4x1:48:interleaved:normal:complete:direct
2:4x24x15:15x1:48:standard:normal:complete:direct
```

## Rejected or failed builds

- down-00: 1; see `down-00/build.log`.
- down-09: 1; see `down-09/build.log`.
- down-10: 1; see `down-10/build.log`.
- down-15: 1; see `down-15/build.log`.
- down-19: 1; see `down-19/build.log`.
- down-20: 1; see `down-20/build.log`.
- down-21: 1; see `down-21/build.log`.
- up-09: 1; see `up-09/build.log`.
- up-17: 1; see `up-17/build.log`.
- up-18: 1; see `up-18/build.log`.

## Instruction-count checks

Offline diagnostics only; these formulas were not used to select or cost the sweep.
Each row counts distinct exported kernel/metadata keys, not independent hardware runs.

| Kernel | Keys | Median absolute error (%) | Maximum error (cycles) |
|---|---:|---:|---:|
| GELU (logical size; padding may add waves) | 49 | 0.000 | 1674 |
| Interleaved F16 GEMM | 179 | 0.301 | 690 |
| F16 reduction | 100 | 0.000 | 0 |