[{"data":1,"prerenderedAt":403},["ShallowReactive",2],{"finding:p1-axis1-phase-g-pareto-landscape":3,"finding-runs:p1-axis1-phase-g-pareto-landscape":253,"finding-related:p1-axis1-phase-g-pareto-landscape":254},{"meta":4,"impact":35,"sections":41},{"id":5,"title":6,"subtitle":7,"eyebrow":8,"date":9,"status":10,"category":11,"polarity":12,"axes":13,"tags":15,"task_code":25,"related_runs":26,"related_findings":27},"p1-axis1-phase-g-pareto-landscape","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 \u002F +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 \u002F G.2 brush dispatch audit \u002F G.3 SH progressive \u002F G.4 multi-cam batch) を 8 scene chain で検証。\u003Cstrong>結論\u003C\u002Fstrong>: G.3 alone 30k が **universal Pareto improvement** (8 scene mean +0.107 dB at +10% wallclock、7\u002F8 scene で win、mic +0.244 \u002F Lego +0.278 \u002F chair +0.142)。G.1 stop15k は -62% wall \u002F -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 \u002F -61% wall) だが multi-scene mean -1.49 dB (ficus\u002Fdrums\u002Fmic で SH warmup 早期停止 destroy)。G.2 は 4.7× gap が architectural dispatch (Burn\u002FCubeCL batching) と判明、kernel\u002Falgorithm 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。","P1 axis 1 · Phase G omnibus · Pareto landscape · G.3 universal win","2026-05-26","stable","audit","positive",[14],1,[16,17,18,19,20,21,22,23,24],"p1-axis1","phase-g","omnibus","pareto-front","sh-progressive","early-stop","brush-comparison","multi-scene","calibration","P1 Phase G omnibus",[],[28,29,30,31,32,33,34],"p1-axis1-phase-g3-sh-progressive","p1-axis1-phase-g2-brush-dispatch-architecture","p1-axis1-phase-f1-emit-simd-falsified","p1-axis1-phase-f3-radix-gpu-prefix-falsified","p1-axis1-metal-opt-audit","p1-d-multi-scene-rechain","m4-brush-bench",{"summary":36,"rank":37,"verdict":38,"delta_psnr":39,"delta_wallclock":40},"Phase G (速度改善 loop) で 4 candidate × 8 scene の Pareto landscape 確定。\u003Cstrong>結論\u003C\u002Fstrong>: **G.3 alone 30k = universal quality improvement** (8 scene mean **33.592 dB**、Phase D 33.485 比 **+0.107 dB**、7\u002F8 scene で win、wallclock +10.2%、splats +29%)、特に **mic は stacked で -6.05 dB だったが G.3 alone なら +0.244 dB 救済**。\u003Cstrong>G.1 stop15k\u003C\u002Fstrong> は dense scene で acceptable (Lego\u002Fchair\u002Fhotdog で Phase D 比 -0.5 dB 圏内) も sparse scene で大幅劣化 (mic -5.84\u002Fficus -1.79)、8 scene mean -1.39 dB で Pareto worse、brush parity (32.86) を -0.76 dB 下回り。\u003Cstrong>G.1+G.3 stacked\u003C\u002Fstrong> は Lego-specific Pareto sweet spot (16m13s \u002F 36.254 dB \u002F +0.15 dB) だが multi-scene mean 31.998 dB で stacked \u003C G.1 alone、SH warmup の効果が early stop で truncate されて sparse scene で逆効果 (ficus -3.15、mic -6.05)。\u003Cstrong>G.2 audit\u003C\u002Fstrong> は 4.7× gap が architectural dispatch (Burn\u002FCubeCL 内部 batching vs Metal 直 per-kernel sync) と判明、Phase F 5 連続 falsification への統一的説明、kernel\u002Falgorithm 軸では覆せない構造的 finding。\u003Cstrong>axis 1 future work ROI 階層\u003C\u002Fstrong>: \u003Cstrong>algorithmic compute reduction > architectural dispatch > kernel-level micro-opt\u003C\u002Fstrong>。Phase G が axis 1 「Apple Silicon native 最適化の ROI 階層」を構造的に確定、卒論 §5.4 narrative の集大成。","high","g3-universal-pareto-confirmed","**G.3 alone 30k = +0.107 dB universal mean** (7\u002F8 scene improve)、Lego +0.278 \u002F mic +0.244 \u002F chair +0.142 \u002F materials +0.125","**G.3 alone 30k = +10.2%** (mean、scene 依存 +5-50%)、G.1 stop15k = -62%、G.1+G.3 stacked = -58%",[42,45,50,53,101,103,175,177,185,187,222,224,233,235,242,244],{"type":43,"text":44},"lead","Phase G (速度改善 loop) で 4 candidate × 8 scene の Pareto landscape を確定。\u003Cstrong>結論\u003C\u002Fstrong>: G.3 alone 30k が **universal quality improvement** (8 scene mean +0.107 dB \u002F 7-8\u002F8 win、+10% wallclock)。G.1 stop15k は scene-dependent (dense OK \u002F 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。",{"type":46,"label":47,"variant":48,"text":49},"callout","Headline (G.3 alone 30k = universal Pareto improvement on quality axis)","success","\u003Cstrong>Phase G の universal Pareto improvement は G.3 alone 30k\u003C\u002Fstrong>。8 scene mean PSNR 33.592 dB (Phase D 33.485 比 **+0.107 dB**)、7\u002F8 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%。\u003Cstrong>brush mean (32.86) との gap が +0.625 → +0.732 dB に拡大\u003C\u002Fstrong>。**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。",{"type":51,"text":52},"heading","1. Pareto landscape — 4 candidate × 8 scene",{"type":54,"columns":55,"align":62,"rows":65,"caption":100},"table",[56,57,58,59,60,61],"config","8 scene total wall","8 scene mean PSNR","vs Phase D","vs brush (32.86)","verdict",[63,64,64,64,64,63],"left","right",[66,73,79,86,93],[67,68,69,70,71,72],"**brush (paper)**","~1h 12m (estimate)","32.86","-0.625","baseline","wgpu+Burn batched",[74,75,76,71,77,78],"**Phase D 30k baseline**","5h 05m","33.485","**+0.625**","previous best universal",[80,81,82,83,84,85],"G.1 stop15k","1h 57m","32.103","-1.382","-0.757","**scene-dependent fail** (mic -5.84)",[87,88,89,90,91,92],"G.1+G.3 stacked (15k+sh_prog)","2h 07m","31.998","-1.487","-0.862","**Lego-specific sweet spot** (multi fail)",[94,95,96,97,98,99],"**G.3 alone 30k (sh_prog)**","**5h 36m**","**33.592**","**+0.107**","**+0.732 ✓**","**universal Pareto improvement**","\u003Cstrong>G.3 alone 30k が universal quality improvement\u003C\u002Fstrong>: +0.107 dB at +10% wallclock。G.1 \u002F stacked は scene-dependent で multi-scene mean fail。brush mean との gap が +0.625 → +0.732 dB と更に拡大。",{"type":51,"text":102},"2. G.3 alone 30k per-scene 結果 (universal win 詳細)",{"type":54,"columns":104,"align":111,"rows":112,"caption":174},[105,106,107,108,109,110],"scene","Phase D PSNR","G.3 30k PSNR","Δ PSNR","G.3 wallclock","G.3 splats",[63,64,64,64,64,64],[113,120,127,134,141,148,155,162,169],[114,115,116,117,118,119],"Lego","36.106","**36.384**","**+0.278 ✓**","41m 7s","487,741",[121,122,123,124,125,126],"chair","35.810","**35.952**","+0.142 ✓","1h 15m 42s","1,148,667",[128,129,130,131,132,133],"ficus","34.220","**34.281**","+0.061 ✓","21m 40s","226,749",[135,136,137,138,139,140],"drums","27.200","**27.217**","+0.017 ≈","1h 5m 36s","1,001,014",[142,143,144,145,146,147],"hotdog","37.330","**37.374**","+0.044 ✓","28m 17s","310,045",[149,150,151,152,153,154],"mic","36.380","**36.624**","**+0.244 ✓**","33m 29s","391,373",[156,157,158,159,160,161],"materials","29.900","**30.025**","+0.125 ✓","30m 46s","349,784",[163,164,165,166,167,168],"ship","30.930","30.877","-0.053 ≈","39m 46s","495,160",[170,171,96,97,172,173],"**mean**","**33.485**","**42 min avg**","**551k avg**","\u003Cstrong>7\u002F8 scene で improvement\u003C\u002Fstrong>、ship のみ -0.053 dB (noise 圏)。**最大改善は mic +0.244 dB** (Phase D 36.38 → G.3 36.62、stacked では mic -6.05 だったので G.3 alone の universal win 性が rescue 効果として顕著)。",{"type":51,"text":176},"3. G.1+G.3 stacked が Lego-specific だった理由",{"type":178,"ordered":179,"items":180},"list",true,[181,182,183,184],"\u003Cstrong>Lego\u002Fchair (dense texture-rich)\u003C\u002Fstrong>: 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","\u003Cstrong>ficus\u002Fdrums\u002Fmic (sparse smooth)\u003C\u002Fstrong>: 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","\u003Cstrong>G.3 alone 30k\u003C\u002Fstrong>: 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","\u003Cstrong>機構の本質\u003C\u002Fstrong>: SH progressive は \u003Cstrong>warmup の利益を fully realize するには 30k 必要\u003C\u002Fstrong>、early stop と不互換。stacked variant は Lego 1 シーンのみ Pareto sweet spot",{"type":51,"text":186},"4. axis 1 最適化 ROI 階層 (Phase F + G 統合 calibration)",{"type":54,"columns":188,"align":194,"rows":196,"caption":221},[189,190,191,192,193],"family","examples","expected","actual 8 scene","ROI ranking",[63,63,64,64,195],"center",[197,203,209,215],[198,199,200,201,202],"kernel-level micro-opt (Phase F)","emit_simd \u002F f16 fwd \u002F radix GPU prefix \u002F refine GPU \u002F target cache","audit -0.5-1.0% × 5","5 連続 falsification (-2% 〜 +4.7% regression)","**LOW**",[204,205,206,207,208],"architectural dispatch (Phase G.2)","Burn\u002FCubeCL batching vs Metal 直 per-kernel sync","4.7× per-iter gap origin","structural finding only (移植 6-10 週)","**HIGH but cost**",[210,211,212,213,214],"algorithmic compute reduction (Phase G.3)","SH progressive growth, full 30k iters","advisor ≤1% hedge","**+0.107 dB universal at +10% wallclock**","**HIGH ✓**",[216,217,218,219,220],"scene-dependent config (Phase G.1)","early stop @ 15k","-50% wall hint","scene-dependent (-1.39 dB mean、mic -5.84)","**MEDIUM**","\u003Cstrong>Apple Silicon native Metal 最適化の ROI 階層\u003C\u002Fstrong>: algorithmic > architectural > kernel-level の順、Phase F+G 8 個 candidate で実証。\u003Cstrong>G.3 alone 30k\u003C\u002Fstrong> が axis 1 universal improvement の唯一の confirmed candidate、他全 candidate は scene-dependent or falsified。",{"type":51,"text":223},"5. 卒論 §5.4 narrative 統合",{"type":178,"items":225},[226,227,228,229,230,231,232],"\u003Cstrong>Phase F 5 連続 falsification\u003C\u002Fstrong> (kernel-level micro-opt): emit_pairs SIMD \u002F f16 forward \u002F radix GPU prefix scan \u002F refine GPU 化 \u002F target cache async — 全部 audit theoretical prediction が overestimate、empirical で regression or noise 圏","\u003Cstrong>G.2 structural finding\u003C\u002Fstrong>: 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」","\u003Cstrong>G.3 universal win\u003C\u002Fstrong>: 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**","\u003Cstrong>G.1 scene-dependent\u003C\u002Fstrong>: early stop は dense scene で acceptable、sparse scene で fail (mic -5.84) — universal config として不適","\u003Cstrong>G.1+G.3 stacked\u003C\u002Fstrong>: 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","\u003Cstrong>方法論 calibration\u003C\u002Fstrong>: 5k smoke の cascading effect は refine.stop_iter=1500 artifact、30k validate なしに smoke A\u002FB 結論を信用すべきでない。Phase G.3 30k chain で artifact 検出","\u003Cstrong>axis 1 future work 階層\u003C\u002Fstrong>: 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",{"type":51,"text":234},"6. 結論と recommended deployment",{"type":178,"ordered":179,"items":236},[237,238,239,240,241],"\u003Cstrong>Phase D recipe + G.3 SH progressive (init=0, max=3, unlock_interval=1000)\u003C\u002Fstrong> を新 universal default として推奨。8 scene mean +0.107 dB \u002F brush 比 +0.732 dB \u002F wallclock +10%","\u003Cstrong>G.1 stop15k\u003C\u002Fstrong> は scene-dependent debug mode として保持、Lego 単独実験等で wallclock 節約用途","\u003Cstrong>G.1+G.3 stacked\u003C\u002Fstrong> は Lego-specific optimization、卒論 §6 future work に「scene-dependent SH schedule + early stop tuning」として記述","\u003Cstrong>G.2 architectural backport\u003C\u002Fstrong> (async readback) は future work、+1-3% hedged ROI で 1-2 週 cost、prototype-gated 投入","\u003Cstrong>G.4 multi-cam batch\u003C\u002Fstrong> は H.A 既 falsified で drop",{"type":51,"text":243},"7. 関連",{"type":178,"items":245},[246,247,248,249,250,251,252],"Phase G.3 implementation + smoke + 30k validation: \u003Ccode>p1-axis1-phase-g3-sh-progressive\u003C\u002Fcode>","Phase G.2 architectural audit: \u003Ccode>p1-axis1-phase-g2-brush-dispatch-architecture\u003C\u002Fcode>","Phase F 5 連続 falsification: \u003Ccode>p1-axis1-phase-f1-emit-simd-falsified\u003C\u002Fcode>、\u003Ccode>p1-axis1-phase-f3-radix-gpu-prefix-falsified\u003C\u002Fcode>、\u003Ccode>p1-e-refine-gpu-smoke\u003C\u002Fcode>、\u003Ccode>p1-axis1-target-cache\u003C\u002Fcode>","Phase D baseline (P1.M5 完遂): \u003Ccode>p1-d-multi-scene-rechain\u003C\u002Fcode>","brush comparison reference: \u003Ccode>m4-brush-bench\u003C\u002Fcode>","axis 1 audit baseline: \u003Ccode>p1-axis1-metal-opt-audit\u003C\u002Fcode>","卒論統合候補: \u003Ccode>chapter-5-4-negative-findings\u003C\u002Fcode> (Phase F+G の ROI 階層化 paragraph)",[],[255,276,300,317,334,356,385],{"id":32,"title":256,"date":257,"status":10,"polarity":12,"category":258,"axes":259,"tags":260,"task_code":269,"related_runs":270,"delta_psnr":271,"delta_wallclock":272,"rank":37,"verdict":273,"impact_summary":274,"detail_path":275},"P1 axis 1 Metal 最適化候補 audit — 5 候補 + 既実装 gate flip 機会、Tier 1 -1.0% wallclock 即時 actionable","2026-05-25","design",[14],[261,262,263,264,265,266,267,268],"p1-profile","axis-1","metal-optimization","kernel-audit","tbdr","simd-reduction","apple-silicon","gate-flip","P1 axis 1 Metal kernel audit",[],"N\u002FA (audit)","estimated -1.5 〜 -2.5% (Tier 1+2)","design-complete-actionable","5 kernel (clean baseline share 合計 55.1%) を Explore subagent で構造的 audit、Apple Silicon 特化最適化候補を kernel 単位で 2-4 個ずつ抽出。\u003Cstrong>最大の発見\u003C\u002Fstrong>: **emit_pairs_simd PSO は既に実装済**、\u003Ccode>use_simd_emit: Cell::new(false)\u003C\u002Fcode> で gate off、comment に「30k validation 後 default true 化予定」(tile_bin.rs:86-87)、**Phase D 30k 完遂で即 flip 可能** (-0.7-1.0% wallclock 即時、zero risk)。同様の即 actionable 機会: f16 forward kernel \u003Ccode>render_splats_f16\u003C\u002Fcode> は env \u003Ccode>SPLAT_F16_FORWARD=1\u003C\u002Fcode> gate (現在 disabled、A\u002FB test 必須 PSNR risk MED-HIGH)。\u003Cstrong>Tier 2 (Phase E scope)\u003C\u002Fstrong>: radix_sort GPU prefix sum (-0.54-0.82% wallclock、CPU-GPU 16-pass sync 除去)、backward_raster imageblock+TBDR (-0.67-1.07% wallclock、tile-local 累積で atomic 大幅削減)。\u003Cstrong>累計 -1.5-2.5% wallclock 改善余地確定、卒論 §6 future work 候補と pilot 実装目標\u003C\u002Fstrong>。backward SIMD reduction は既に default 有効 (rasterize.rs:642、2.43× win 享受中で確認済)、SSIM は eval-only で training 直接寄与なしのため Tier 3。","\u002Ffindings\u002Fp1-axis1-metal-opt-audit\u002F",{"id":30,"title":277,"date":257,"status":10,"polarity":278,"category":11,"axes":279,"tags":280,"task_code":289,"related_runs":290,"delta_psnr":294,"delta_wallclock":295,"rank":296,"verdict":297,"impact_summary":298,"detail_path":299},"Phase F.1 emit_pairs_simd + f16 forward gate flip — audit Tier 1 仮説 falsified、現規模で net regression \u002F no improvement","negative",[14],[16,281,282,283,284,285,286,287,288],"phase-f","emit-simd","f16-forward","tier-1","falsified","negative-finding","ab-test","lego-5k","P1 Phase F.1 \u002F F.2",[291,292,293],"lego-phase-f1-emit-simd-5k","lego-phase-f1-baseline-5k","lego-phase-f2-f16-fwd-5k","±0.13 dB (両者とも許容範囲、atomic\u002Ffp 順序由来)","+4.7% (emit_simd net regression) \u002F +2.5% (f16 fwd noise 圏内)","medium","audit-falsified-tier-1","audit (p1-axis1-metal-opt-audit) で Tier 1「即 actionable gate flip、-0.7-1.0% wallclock、zero risk」と分類した 2 候補を Lego 5k smoke A\u002FB で実証検証。\u003Cstrong>emit_pairs_simd は total wallclock +4.7% の net regression\u003C\u002Fstrong> (112.11s → 117.38s、~10 kernel 平均なので noise floor 小、real regression 確定)、ただし per-kernel emit_pairs 単体は +8.5% で baseline 2 sample 変動 (4.814 \u002F 5.129、6.5%) と近い hedge 必要。\u003Cstrong>f16 forward は ~+2.5% wallclock\u003C\u002Fstrong> (114.97s)、run-to-run variance 圏内で improvement \u002F regression いずれも明確に検出できず。\u003Cstrong>PSNR は両者で許容範囲\u003C\u002Fstrong> (emit_simd -0.132 dB、f16 +0.075 dB、atomic order \u002F fp 順序由来想定)。**audit の予測 calibration data**: Tier 1 SIMD-reduction 系の効果は theory より小さく overhead が打ち消し、Tier 2 別 mechanism (CPU-GPU sync 除去) は別途検証必要、Tier 2 同 family (backward TBDR) は falsification 拡大適用で skip 判断強化。卒論 narrative 価値: 「audit theoretical predictions vs empirical measurements」の方法論 paragraph を §5.4 negative findings 章 (chapter-5-4-negative-findings.md) に追加候補。","\u002Ffindings\u002Fp1-axis1-phase-f1-emit-simd-falsified\u002F",{"id":31,"title":301,"date":257,"status":10,"polarity":278,"category":11,"axes":302,"tags":303,"task_code":308,"related_runs":309,"delta_psnr":312,"delta_wallclock":313,"rank":37,"verdict":314,"impact_summary":315,"detail_path":316},"Phase F.3 radix GPU prefix scan — bit-exact 実装完成だが Metal implicit fences で +7.4% wallclock \u002F +35-41% per-call regression、audit Tier 2 仮説 falsified",[14],[16,281,304,305,306,285,286,307,265,287,288],"radix-sort","gpu-prefix-scan","tier-2","metal-fences","P1 Phase F.3",[310,311],"lego-phase-f3-baseline-5k","lego-phase-f3-gpu-scan-5k","+0.03 dB (parity、bit-exact 経路、session noise 内)","+7.4% (+8.58s @ 5k iter、sanity rerun +4.8% でも regression 確定)","audit-falsified-tier-2","audit (p1-axis1-metal-opt-audit) で Tier 2 「radix_sort GPU prefix sum、-0.5-0.8% wallclock、LOW PSNR risk」と分類した候補を empirical 検証。\u003Cstrong>bit-exact 実装は完成\u003C\u002Fstrong> (16-thread single-threadgroup kernel、Apple SIMD prefix exclusive sum + per-digit serial scan、100k random \u002F 500k packed keys \u002F edge cases 6 種で CPU stable sort と byte-for-byte 一致)、しかし 5k Lego smoke で \u003Cstrong>wallclock 115.83s → 124.41s (+7.4%)、ts_fwd_radix_sort 4.768 → 6.733 ms\u002Fcall (+41%)\u003C\u002Fstrong> の net regression。sanity re-run (118.73s \u002F 6.402ms) で再現確認、run-to-run 変動の上。\u003Cstrong>PSNR は parity\u003C\u002Fstrong> (31.604 → 31.635 dB、bit-exact 経路で 0 drift 期待、観測 +0.03 dB は session noise)。\u003Cstrong>Likely mechanism\u003C\u002Fstrong>: StorageModeShared buffer での back-to-back compute encoder 間で Metal が implicit fence を挿入 (hist→scan の buf_hist、scan→scatter の buf_offsets で read-after-write hazard)、TBDR pipeline stall。旧 CPU 経路は buf_hist→buf_offsets 変換を host で実行するため GPU 内 memory dependency が無く、\u003Cstrong>「除去した wait_until_completed」は実は CPU prefix scan と overlap していた active work\u003C\u002Fstrong> だった。教訓: 「CPU 介在を on-GPU に置換」族の audit 予測は overlap の存在を見落とすため systematically overestimate、Tier 2 同 family (backward TBDR、tile-local accumulator) の skip 判断強化。kernel + tests は env \u003Ccode>SPLAT_RADIX_GPU_SCAN=1\u003C\u002Fcode> で opt-in (future workload hedge)。","\u002Ffindings\u002Fp1-axis1-phase-f3-radix-gpu-prefix-falsified\u002F",{"id":29,"title":318,"date":257,"status":10,"polarity":319,"category":258,"axes":320,"tags":322,"task_code":328,"related_runs":329,"delta_psnr":271,"delta_wallclock":330,"rank":37,"verdict":331,"impact_summary":332,"detail_path":333},"Phase G.2 brush 4.7× per-iter 速度差の真因 — command buffer batching、Phase F 全 kernel-level 改善試行への統一的構造説明","mixed",[14,321],2,[16,17,22,323,324,325,326,327],"dispatch-architecture","command-buffer-batching","structural-finding","burn-cubecl","metal-direct","P1 axis 1 Phase G.2",[],"0% (audit) \u002F estimated +3-5% if async readback backport (要 prototype 検証)","structural-explanation","Explore subagent (Sonnet very thorough) で brush 18ms\u002Fiter vs splat-rs 84ms\u002Fiter (4.67×) の真因を architectural 差で構造特定。\u003Cstrong>brush は Burn\u002FCubeCL backend の \u003Ccode>launch_unchecked()\u003C\u002Fcode> async dispatch + 内部 command buffer batching\u003C\u002Fstrong> で per-iter 5-7 explicit awaits → ~5 actual GPU flushes。一方 \u003Cstrong>splat-rs は per kernel 毎に \u003Ccode>cmd.wait_until_completed()\u003C\u002Fcode> で 10-50 GPU flushes\u003C\u002Fstrong>。計算: 17ms GPU compute + ~25ms wait overhead (~2.5ms\u002Fwait × ~10) = ~42ms、実測 84ms とは ~2× ずれあるが (subagent quantification の不確実性)、order-of-magnitude は一致。\u003Cstrong>これは Phase F 5 連続 falsification への統一的構造説明\u003C\u002Fstrong>: kernel-level micro-opt (SIMD reduction \u002F f16 accumulator \u002F radix GPU prefix \u002F TBDR imageblock 等) が効かなかったのは bottleneck が \u003Cstrong>per-kernel compute ではなく dispatch synchronization architecture\u003C\u002Fstrong> だったから。**主仮説 ranking** (subagent assessment): (1) command buffer batching (50% of gap、移植 VERY HIGH cost 6-10 週)、(2) async readback (15%、MEDIUM cost 1-2 週、+3-5% expected)、(3) kernel fusion (5-10%、EXTREME cost)。**ただし subagent quantification は overestimate 傾向あり** (Phase F.3 で「removed wait was overlapping with CPU work, not idle」発覚と矛盾、wait は free な場合もある)。卒論 narrative としては \u003Cstrong>structural explanation\u003C\u002Fstrong> として極めて価値高い、§5.4 negative findings 章で「Phase F 全 kernel-level 改善試行は architectural mismatch だった」統一的 paragraph 候補。","\u002Ffindings\u002Fp1-axis1-phase-g2-brush-dispatch-architecture\u002F",{"id":28,"title":335,"date":257,"status":10,"polarity":12,"category":258,"axes":336,"tags":337,"task_code":345,"related_runs":346,"delta_psnr":351,"delta_wallclock":352,"rank":37,"verdict":353,"impact_summary":354,"detail_path":355},"Phase G.3 SH-progressive — 5k smoke -14% は artifact、30k full は **quality improvement + 0.28 dB** に reframe、stacked + G.1 で **Pareto sweet spot** (Lego -61% wallclock + 0.15 dB)",[14],[16,17,20,338,19,288,339,340,341,342,343,344],"compute-reduction","lego-30k","stacked-config","implementation","unit-tests","bit-exact","smoke-artifact","P1 Phase G.3",[347,348,349,292,350],"lego-phase-g3-sh-progressive-5k","lego-phase-g3-sh-progressive-30k","lego-phase-g1g3-stacked-15k","lego-brushcompat-opacdecay-30k","+0.15 dB stacked (vs Phase D)、+0.28 dB 30k single (vs Phase D)、-0.12 dB 5k smoke (許容)","**-61% stacked** (vs Phase D)、-1.9% 30k single、-13.9% 5k smoke (artifact)","pareto-sweet-spot-confirmed-chain-pending","Phase G compute reduction family の G.3 (SH-progressive growth) を実装 + bit-exact unit tests (11 件、cargo test 43 件 全 pass) + **3 layer の Lego 結果検証**。\u003Cstrong>(1) 5k smoke\u003C\u002Fstrong>: wallclock -13.9% \u002F splats -22% \u002F PSNR -0.12 dB、cascading splat reduction を観測。\u003Cstrong>(2) 30k full validation\u003C\u002Fstrong>: wallclock **-1.9%** (5k から大幅縮小)、splats **+30%** (5k から逆転)、PSNR **+0.28 dB** (quality improvement!)。5k smoke の cascading 効果は refine.stop_iter=1500 による artifact、30k では sh unlock 完了 (iter 3000) 後に refine が iter 15000 まで full SH で継続 → splats baseline より grow。\u003Cstrong>(3) G.1+G.3 stacked (max_steps=15000 + sh_progressive)\u003C\u002Fstrong>: Lego **16m13s \u002F 36.254 dB \u002F 428k splats** = Phase D 比 **-61% wallclock + 0.15 dB PSNR** で \u003Cstrong>Pareto sweet spot\u003C\u002Fstrong> 確定。\u003Cstrong>Key reframe\u003C\u002Fstrong>: G.3 は「speed win」ではなく「**quality improvement at no speed cost**」、stacked variant で G.1 speed と SH warmup quality gain を統合。\u003Cstrong>Implementation\u003C\u002Fstrong>: \u003Ccode>[trainer.sh_progressive]\u003C\u002Fcode> section (default disabled、全 backward compat)、\u003Ccode>CameraGpu\u003C\u002Fcode> struct を \u003Ccode>sh_degree\u003C\u002Fcode> (buffer layout) と \u003Ccode>active_sh_degree\u003C\u002Fcode> (per-iter eval) に分離。\u003Cstrong>Calibration data point\u003C\u002Fstrong>: Phase F 5 連続 falsification + G.3 5k smoke artifact = audit \u002F smoke overestimate 6 例目、「smoke は production scale を representative しない」が新教訓。8 scene chain validation pending。\u003Cstrong>卒論 narrative\u003C\u002Fstrong>: Phase F (kernel-level fail) → G.2 (architectural insight) → G.3 (algorithmic reframe: speed → quality + stacked Pareto) の 3 family 比較で「Apple Silicon native 最適化は \u003Cstrong>algorithmic compute reduction + early stop の組み合わせが Pareto-optimal\u003C\u002Fstrong>」という構造的 calibration。","\u002Ffindings\u002Fp1-axis1-phase-g3-sh-progressive\u002F",{"id":33,"title":357,"date":257,"status":10,"polarity":12,"category":358,"axes":359,"tags":361,"task_code":371,"related_runs":372,"delta_psnr":380,"delta_wallclock":381,"rank":37,"verdict":382,"impact_summary":383,"detail_path":384},"P1.D multi-scene Phase D re-chain final — 8 scene mean 33.49 dB、brush mean 32.86 を +0.63 dB 上回り、universal win-win-win 実証","experiment",[14,321,360],3,[362,363,364,23,365,366,367,368,369,370],"p1","phase-d","milestone-m5","brush-parity","brush-超え","premultiplied","opacity-decay","universal-win-win-win","rechain-final","P1.D multi-scene re-chain (M5 final)",[350,373,374,375,376,377,378,379],"chair-brushcompat-opacdecay-30k","ficus-brushcompat-opacdecay-30k","drums-brushcompat-opacdecay-30k","hotdog-brushcompat-opacdecay-30k","mic-brushcompat-opacdecay-30k","materials-brushcompat-opacdecay-30k","ship-brushcompat-opacdecay-30k","8 scene mean +0.63 dB vs brush paper (33.49 vs 32.86)","-61% total chain (13h+ → 5h 5m)","accepted-m5-complete","Phase D opacity_decay (rate=0.004 brush default) を 7 scene × 30k full chain bench、Lego val Phase D 30k と合わせて 8 scene 集計。**全 scene で baseline brushcompat 30k 比 PSNR + splats + wallclock すべて改善 (universal win-win-win)**: PSNR +0.18〜+1.42 dB \u002F splats -57〜-78% \u002F wallclock -39〜-69%。8 scene mean 33.49 dB vs brush paper 8 scene mean 32.86 dB = **+0.63 dB 上回り**、本実装が brush の multi-scene mean を decisive に超えた。brush 超え 3 scene (Lego val +4.07 \u002F drums +1.05 \u002F mic +1.02)、4 scene が brush 比 ±0.7 dB 圏内 (chair -0.02 \u002F hotdog -0.39 \u002F ship -0.01 \u002F materials -0.10)、最遠 scene でも ficus -0.65 で接近。全体 wallclock baseline chain (13h+) → Phase D re-chain 5h 5m (-61%)、mean splats 1.4M → 428k (-69%) で brush 282k に肉薄。P1.M5 完全達成 (Lego val > 36 dB ✅ + multi-scene mean > 32 dB ✅)、卒論 central evaluation table の final 数字確定、universal claim 完全実証。","\u002Ffindings\u002Fp1-d-multi-scene-rechain\u002F",{"id":34,"title":386,"date":387,"status":10,"polarity":319,"category":358,"axes":388,"tags":389,"task_code":395,"related_runs":396,"delta_psnr":398,"delta_wallclock":399,"rank":37,"verdict":400,"impact_summary":401,"detail_path":402},"M4 Max 上 brush (wgpu→Metal) の 30k baseline — wgpu 抽象は自作より速かった","2026-05-23",[321],[390,391,392,71,393,394],"phase-2","brush","wgpu","m4-max","abstraction-cost","A.3",[397],"lego-sh3-30k","+11.13 dB (brush 比優位)","−65.6% (brush の方が速い)","investigative","wgpu 抽象は自作 native より遅いはず、という想定が逆。同一 M4 Max 上で brush (wgpu) が 9m08s \u002F 37.40 dB、自作 (Metal 直) が 26m32s \u002F 26.27 dB。第 2 軸 (抽象コスト定量化) の主張を再 framing する必要が確定。","\u002Ffindings\u002Fm4-brush-bench\u002F",1782449788632]