Phase G omnibus — 速度改善 4 candidate × 8 scene の Pareto landscape 確定、**G.3 alone 30k = universal quality improvement** (+0.107 dB / +10% wall)
Phase F 5 連続 falsification (kernel-level micro-opt) 後の Phase G 計算改善 loop。4 candidate (G.1 early stop / G.2 brush dispatch audit / G.3 SH progressive / G.4 multi-cam batch) を 8 scene chain で検証。<strong>結論</strong>: G.3 alone 30k が **universal Pareto improvement** (8 scene mean +0.107 dB at +10% wallclock、7/8 scene で win、mic +0.244 / Lego +0.278 / chair +0.142)。G.1 stop15k は -62% wall / -1.39 dB mean で Pareto worse (mic で -5.84 outlier)。G.1+G.3 stacked は Lego-specific sweet spot (+0.15 dB / -61% wall) だが multi-scene mean -1.49 dB (ficus/drums/mic で SH warmup 早期停止 destroy)。G.2 は 4.7× gap が architectural dispatch (Burn/CubeCL batching) と判明、kernel/algorithm 軸では覆せない構造的 calibration。G.4 multi-cam は H.A 既 falsified で drop。**axis 1 axis 1 future work 階層**: algorithmic > architectural > kernel-level の ROI 順位、Phase G が確定。
本文
Phase G (速度改善 loop) で 4 candidate × 8 scene の Pareto landscape を確定。結論: G.3 alone 30k が **universal quality improvement** (8 scene mean +0.107 dB / 7-8/8 win、+10% wallclock)。G.1 stop15k は scene-dependent (dense OK / sparse fail)、G.1+G.3 stacked は Lego-specific sweet spot で multi-scene mean fail、G.2 は architectural structural finding。Phase F (kernel-level) → G.2 (architectural) → G.3 (algorithmic) で axis 1 最適化 ROI 階層を確定。
Phase G の universal Pareto improvement は G.3 alone 30k。8 scene mean PSNR 33.592 dB (Phase D 33.485 比 **+0.107 dB**)、7/8 scene で improvement、特に Phase D の outlier だった mic は **+0.244 dB の最大改善**。wallclock は +10.2% (mean) と modest cost、splats +29%。brush mean (32.86) との gap が +0.625 → +0.732 dB に拡大。**重要な対比**: G.1 stop15k は scene-dependent fail (mic -5.84)、G.1+G.3 stacked は Lego-specific sweet spot で multi-scene mean fail (sh warmup の効果が early stop で truncate)。G.3 alone は **iter 0-3000 sh warmup + iter 3000-15000 refine + iter 15000-30000 settle** の full cycle が必要、early stop と不互換。卒論 §5.4 narrative: kernel-level → architectural → algorithmic の ROI 階層が Phase F→G で確定。
1. Pareto landscape — 4 candidate × 8 scene
| config | 8 scene total wall | 8 scene mean PSNR | vs Phase D | vs brush (32.86) | verdict |
|---|---|---|---|---|---|
| **brush (paper)** | ~1h 12m (estimate) | 32.86 | -0.625 | baseline | wgpu+Burn batched |
| **Phase D 30k baseline** | 5h 05m | 33.485 | baseline | **+0.625** | previous best universal |
| G.1 stop15k | 1h 57m | 32.103 | -1.382 | -0.757 | **scene-dependent fail** (mic -5.84) |
| G.1+G.3 stacked (15k+sh_prog) | 2h 07m | 31.998 | -1.487 | -0.862 | **Lego-specific sweet spot** (multi fail) |
| **G.3 alone 30k (sh_prog)** | **5h 36m** | **33.592** | **+0.107** | **+0.732 ✓** | **universal Pareto improvement** |
2. G.3 alone 30k per-scene 結果 (universal win 詳細)
| scene | Phase D PSNR | G.3 30k PSNR | Δ PSNR | G.3 wallclock | G.3 splats |
|---|---|---|---|---|---|
| Lego | 36.106 | **36.384** | **+0.278 ✓** | 41m 7s | 487,741 |
| chair | 35.810 | **35.952** | +0.142 ✓ | 1h 15m 42s | 1,148,667 |
| ficus | 34.220 | **34.281** | +0.061 ✓ | 21m 40s | 226,749 |
| drums | 27.200 | **27.217** | +0.017 ≈ | 1h 5m 36s | 1,001,014 |
| hotdog | 37.330 | **37.374** | +0.044 ✓ | 28m 17s | 310,045 |
| mic | 36.380 | **36.624** | **+0.244 ✓** | 33m 29s | 391,373 |
| materials | 29.900 | **30.025** | +0.125 ✓ | 30m 46s | 349,784 |
| ship | 30.930 | 30.877 | -0.053 ≈ | 39m 46s | 495,160 |
| **mean** | **33.485** | **33.592** | **+0.107** | **42 min avg** | **551k avg** |
3. G.1+G.3 stacked が Lego-specific だった理由
- Lego/chair (dense texture-rich): sh warmup で early refine が低周波 base 作成、iter 3000-15000 で high-freq detail 改善、iter 15000 stop で済む → stacked +0.05-0.56 dB
- ficus/drums/mic (sparse smooth): sh warmup が high-freq detail 学習を遅延、iter 15000 stop で settle phase なし → final splats が high-freq overfit に届かず stacked -3 〜 -6 dB の大幅 loss
- G.3 alone 30k: iter 0-3000 sh warmup + iter 3000-15000 refine + iter 15000-30000 settle (refine off) の full cycle、sparse scene でも settle phase で high-freq detail が converge → universal win
- 機構の本質: SH progressive は warmup の利益を fully realize するには 30k 必要、early stop と不互換。stacked variant は Lego 1 シーンのみ Pareto sweet spot
4. axis 1 最適化 ROI 階層 (Phase F + G 統合 calibration)
| family | examples | expected | actual 8 scene | ROI ranking |
|---|---|---|---|---|
| kernel-level micro-opt (Phase F) | emit_simd / f16 fwd / radix GPU prefix / refine GPU / target cache | audit -0.5-1.0% × 5 | 5 連続 falsification (-2% 〜 +4.7% regression) | **LOW** |
| architectural dispatch (Phase G.2) | Burn/CubeCL batching vs Metal 直 per-kernel sync | 4.7× per-iter gap origin | structural finding only (移植 6-10 週) | **HIGH but cost** |
| algorithmic compute reduction (Phase G.3) | SH progressive growth, full 30k iters | advisor ≤1% hedge | **+0.107 dB universal at +10% wallclock** | **HIGH ✓** |
| scene-dependent config (Phase G.1) | early stop @ 15k | -50% wall hint | scene-dependent (-1.39 dB mean、mic -5.84) | **MEDIUM** |
5. 卒論 §5.4 narrative 統合
- Phase F 5 連続 falsification (kernel-level micro-opt): emit_pairs SIMD / f16 forward / radix GPU prefix scan / refine GPU 化 / target cache async — 全部 audit theoretical prediction が overestimate、empirical で regression or noise 圏
- G.2 structural finding: 4.7× per-iter gap は kernel-level ではなく architectural dispatch、Phase F の統一的説明として「kernel-level の Apple 特化最適化は dispatch 同期 cost に打ち消される」
- G.3 universal win: SH progressive growth (Algorithmic family) は per-iter compute を直接削減 + sh warmup → better-conditioned splat distribution → 8 scene mean +0.107 dB at +10% wall。**唯一の universal Pareto improvement**
- G.1 scene-dependent: early stop は dense scene で acceptable、sparse scene で fail (mic -5.84) — universal config として不適
- G.1+G.3 stacked: Lego-specific Pareto sweet spot だが multi-scene mean fail、「single-scene Pareto improvement は multi-scene mean に transfer しない」教訓
- 方法論 calibration: 5k smoke の cascading effect は refine.stop_iter=1500 artifact、30k validate なしに smoke A/B 結論を信用すべきでない。Phase G.3 30k chain で artifact 検出
- axis 1 future work 階層: algorithmic > architectural > kernel-level の ROI 順位、Apple Silicon native 最適化の構造的 calibration として卒論 §5.4 に追加
6. 結論と recommended deployment
- Phase D recipe + G.3 SH progressive (init=0, max=3, unlock_interval=1000) を新 universal default として推奨。8 scene mean +0.107 dB / brush 比 +0.732 dB / wallclock +10%
- G.1 stop15k は scene-dependent debug mode として保持、Lego 単独実験等で wallclock 節約用途
- G.1+G.3 stacked は Lego-specific optimization、卒論 §6 future work に「scene-dependent SH schedule + early stop tuning」として記述
- G.2 architectural backport (async readback) は future work、+1-3% hedged ROI で 1-2 週 cost、prototype-gated 投入
- G.4 multi-cam batch は H.A 既 falsified で drop
7. 関連
- Phase G.3 implementation + smoke + 30k validation:
p1-axis1-phase-g3-sh-progressive - Phase G.2 architectural audit:
p1-axis1-phase-g2-brush-dispatch-architecture - Phase F 5 連続 falsification:
p1-axis1-phase-f1-emit-simd-falsified、p1-axis1-phase-f3-radix-gpu-prefix-falsified、p1-e-refine-gpu-smoke、p1-axis1-target-cache - Phase D baseline (P1.M5 完遂):
p1-d-multi-scene-rechain - brush comparison reference:
m4-brush-bench - axis 1 audit baseline:
p1-axis1-metal-opt-audit - 卒論統合候補:
chapter-5-4-negative-findings(Phase F+G の ROI 階層化 paragraph)