Fr0zkaandClaude Opus 5.5
6e52fb48b2
Tagged the passage bucket: per-tile prepareSweptTubeBase dominates (0.47-0.92 s summed per view), smooth-min second. Segment lower-bound pruning (-17% pairs) and deferred fallback floor work stayed hash-exact but moved near-ready <3%. Still +17..+26% vs L109. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
59 lines
4.1 KiB
Markdown
59 lines
4.1 KiB
Markdown
# Linux118 — passage tile cost
|
||
|
||
## Result
|
||
|
||
The L115 output is exact: owner4 hashes match at W1/W8/W12 under Clang and GCC. Walk audits passed, and the fast tier passed with 259 checks plus 5 diagnostic-only baseline comparisons.
|
||
|
||
The near-ready target was not met. The final candidate is 17.2–25.7% slower than L109 across the eight W8/W12 cells; whole-view readiness is 9.4–13.9% slower. Both designs leave readiness close to L115.
|
||
|
||
## Field tags
|
||
|
||
One W1 probe per stratum, summed over LOD0–2 passage-feature tiles. `SmoothMinDistance` time is estimated from 1/128 sampled calls. Point time contains `CachedColumn`; cache read time is inside `CachedColumn`, so those columns overlap.
|
||
|
||
| Stratum | Feature tiles | Prepare ms | Smooth-min calls / ms est. | Point / CachedColumn ms est. | Shared reads ms est. (initial / locked) |
|
||
| --- | ---: | ---: | ---: | ---: | ---: |
|
||
| S0 | 158 | 467 | 7.14M / 137 | 404 / 226 | 24 / 4 |
|
||
| S1 | 195 | 919 | 15.60M / 289 | 0 / 0 | 0 / 0 |
|
||
| S2 | 187 | 848 | 13.93M / 249 | 0 / 0 | 0 / 0 |
|
||
| S3 | 182 | 567 | 2.62M / 49 | 0 / 0 | 0 / 0 |
|
||
|
||
The Heightfield2D/cache tags occur in S0; the other three strata do not use that path in these views.
|
||
|
||
## Designs and passage counters
|
||
|
||
Design 1 added a conservative segment-to-sample lower bound in `prepareSweptTubeBase`. It skips distance, floor, and shell work when a segment cannot improve the sample. Hashes stayed exact. It pruned 17–18% of candidate pairs, but the readiness matrix remained near L115.
|
||
|
||
Design 2 deferred floor-profile interpolation in exact fallbacks until a segment could improve the nearest route or floor result. It also stayed hash-exact, with only a small timing change. Restricting fallbacks to prepared feature spans changed hashes, so that restriction was removed.
|
||
|
||
W8 passage-feature tile totals, averaged over two profiles per stratum. L118 pairs show evaluated plus bound-pruned counts; their sum equals L115's candidate-pair total.
|
||
|
||
| Stratum | Sweep pairs: L109 → L115; L118 evaluated + pruned | Fallbacks: L109 → L115 → L118 | Visits: L109 → L115 → L118 |
|
||
| --- | ---: | ---: | ---: |
|
||
| S0 | 3.36M → 7.56M; 6.25M + 1.31M | 0.241M → 0.413M → 0.413M | 1.19M → 3.41M → 3.41M |
|
||
| S1 | 6.19M → 16.22M; 13.25M + 2.97M | 0.369M → 0.719M → 0.719M | 1.91M → 7.60M → 7.59M |
|
||
| S2 | 6.15M → 15.52M; 12.69M + 2.83M | 0.315M → 0.634M → 0.634M | 1.66M → 6.88M → 6.87M |
|
||
| S3 | 4.09M → 9.63M; 7.98M + 1.64M | 0.074M → 0.131M → 0.131M | 0.48M → 1.41M → 1.41M |
|
||
|
||
## Readiness
|
||
|
||
Five interleaved measured runs per version, stratum, and worker count, after one warm-up per version. Values are median seconds: near LOD0–2 / whole view.
|
||
|
||
| Stratum | Workers | Near L109 → L115 → L118 (Δ vs L109) | Whole L109 → L115 → L118 (Δ vs L109) |
|
||
| --- | ---: | ---: | ---: |
|
||
| S0 | 8 | 0.771 → 0.941 → 0.939 (+21.8%) | 1.731 → 1.968 → 1.938 (+12.0%) |
|
||
| S0 | 12 | 0.781 → 0.927 → 0.932 (+19.4%) | 1.716 → 1.878 → 1.897 (+10.6%) |
|
||
| S1 | 8 | 0.964 → 1.235 → 1.212 (+25.7%) | 1.886 → 2.122 → 2.088 (+10.7%) |
|
||
| S1 | 12 | 0.908 → 1.166 → 1.136 (+25.1%) | 1.754 → 2.022 → 1.997 (+13.8%) |
|
||
| S2 | 8 | 0.991 → 1.257 → 1.241 (+25.2%) | 1.863 → 2.136 → 2.121 (+13.8%) |
|
||
| S2 | 12 | 0.929 → 1.172 → 1.143 (+23.1%) | 1.736 → 1.989 → 1.976 (+13.8%) |
|
||
| S3 | 8 | 0.830 → 0.981 → 0.997 (+20.1%) | 1.726 → 1.918 → 1.912 (+10.8%) |
|
||
| S3 | 12 | 0.780 → 0.933 → 0.914 (+17.2%) | 1.625 → 1.795 → 1.778 (+9.4%) |
|
||
|
||
The timing gate recorded one-minute load 0.65 and 99.2% CPU idle. Each headless run used CPUs 0–11, the 6 GiB address-space limit, and a 1,740 s timeout. Design 1 timings are preserved in `Saved/Linux/l118/timing-design1/`; final timings are in `Saved/Linux/l118/timing/`.
|
||
|
||
## Unreal and workspace
|
||
|
||
The post-timing Linux `VFHostEditor` build passed (27/27 actions, 71.13 s) with `-NoUBA -MaxParallelActions=4`; it compiled `VFCore_Core_src_VoxelForgeCore_cpp.cpp`. Unreal ran without `prlimit`. Three existing RealtimeMesh deprecation warnings remain.
|
||
|
||
Work stayed on `wip/l111-halfpipe-spline` at `c2b8d39`. No commit, push, or stash was made. Logs and summaries are under `Saved/Linux/l118/`.
|