[{"data":1,"prerenderedAt":357},["ShallowReactive",2],{"finding:a-7-multi-scene-batched":3,"finding-runs:a-7-multi-scene-batched":220,"finding-related:a-7-multi-scene-batched":303},{"meta":4,"impact":39,"sections":45},{"id":5,"title":6,"subtitle":7,"eyebrow":8,"date":9,"status":10,"category":11,"polarity":12,"axes":13,"tags":15,"task_code":23,"related_runs":24,"related_findings":35},"a-7-multi-scene-batched","A.7 × multi-scene — batching 効果は scene 依存 (-1.6% 〜 -18.6% で 12x の幅)","A.7 batched cmd buffer (env SPLAT_BATCHED_FORWARD=1) を chair \u002F ficus \u002F drums \u002F hotdog で 30k 実行。wallclock 改善は -1.6% (ficus) 〜 -18.6% (chair) と scene 依存で 12 倍の幅。splat 数だけでは説明不可、scene geometry の compute\u002Fcommit ratio が batching benefit を決定。PSNR drift は hotdog で -0.83 dB と高品質シーンほど SIMD atomic 非決定性の影響大。","Multi-scene finding · scene-dependent batching","2026-05-23","stable","experiment","mixed",[14],3,[16,17,18,19,20,21,22],"phase-5","a-7","icb","batched","multi-scene","scene-dependency","apple-silicon","A.7",[25,26,27,28,29,30,31,32,33,34],"chair-30k","chair-batched-30k","ficus-30k","ficus-batched-30k","drums-30k","drums-batched-30k","hotdog-30k","hotdog-batched-30k","lego-sh3-30k","lego-a7-batched-30k",[36,37,38],"a-7-icb-batching-results","a-10-variance-baseline","a-4-nerf-synthetic-scene-results",{"summary":40,"rank":41,"verdict":42,"delta_wallclock":43,"delta_psnr":44},"A.7 batched cmd buffer の wallclock 改善は scene 依存で chair -18.6% \u002F hotdog -5.4% \u002F drums -3.4% \u002F ficus -1.6% \u002F lego -6.16% の 5 シーン (12x の幅)。chair の突出は splat 数最大 (~130k) + scene geometry の compute\u002Fcommit ratio が高いことが要因と推測。一方 ficus \u002F drums は variance 範囲内、独立 effect 断定不可。卒論で「A.7 effective ≠ universal、scene 選択 + workload analysis 必須」と honest framing。","high","partial","-1.6% 〜 -18.6% (mean -7.0%)","+0.130 dB 〜 -0.828 dB (mean -0.21 dB)",[46,49,64,67,121,123,125,146,148,151,157,159,161,163,165,189,191,193,199,204,206,211,213],{"type":47,"text":48},"lead","\u003Ca href=\"\u002Ffindings\u002Fa-7-icb-batching-results\">A.7 lego 単独 finding\u003C\u002Fa> (-6.16%) の汎化性を確認するため、chair \u002F ficus \u002F drums \u002F hotdog で同じ env \u003Ccode>SPLAT_BATCHED_FORWARD=1\u003C\u002Fcode> を適用し 30k bench。結果は \u003Cstrong>scene 依存で 12 倍の幅\u003C\u002Fstrong>、A.7 batching は scene-specific な workload で大きく effect が変動することが判明。",{"type":50,"items":51},"kv",[52,55,58,61],{"key":53,"value":54},"実施日","2026-05-23 Phase B (bench chain Phase B)",{"key":56,"value":57},"config","configs\u002F2026-05-23-1000-{chair,ficus,drums,hotdog}-batched-30k.toml",{"key":59,"value":60},"binary","main HEAD (5573681 A.7 cherry-pick 適用後)",{"key":62,"value":63},"env","SPLAT_BATCHED_FORWARD=1 (trainer.forward_with_state_batched_tail + backward_combined_batched 経由)",{"type":65,"text":66},"heading","実測値 (multi-scene wallclock 改善)",{"type":68,"columns":69,"align":77,"rows":80,"caption":120},"table",[70,71,72,73,74,75,76],"scene","baseline wall","batched wall","Δ wall %","baseline PSNR","batched PSNR","Δ PSNR",[78,79,79,79,79,79,79],"left","right",[81,89,97,105,113],[82,83,84,85,86,87,88],"**chair**","22m54s (1374s)","**18m39s (1119s)**","**-18.6%**","22.883","22.780","-0.103",[90,91,92,93,94,95,96],"lego","23m13s (1393s)","21m47s (1307s)","**-6.16%**","24.879","24.577","-0.302",[98,99,100,101,102,103,104],"**hotdog**","23m52s (1432s)","22m35s (1355s)","**-5.4%**","30.290","**29.462**","**-0.828**",[106,107,108,109,110,111,112],"drums","21m27s (1287s)","20m43s (1243s)","-3.4%","17.773","17.903","+0.130",[114,83,115,116,117,118,119],"ficus","22m33s (1353s)","-1.6%","13.959","14.054","+0.095","wallclock 改善は chair で突出 (-18.6%)、ficus でほぼゼロ (-1.6%)。PSNR drift も hotdog で -0.83 dB、他 scene は ±0.13 dB 程度。",{"type":65,"text":122},"scene 依存性の analysis",{"type":65,"level":14,"text":124},"(1) wallclock 改善 vs splat 数の弱相関",{"type":68,"columns":126,"align":129,"rows":130,"caption":145},[70,127,128],"final splats (baseline)","wallclock Δ %",[78,79,79],[131,134,138,141,143],[132,133,85],"chair","~130k",[135,136,137],"hotdog","~80k","-5.4%",[90,139,140],"83,734","-6.16%",[106,142,109],"~65k",[114,144,116],"~50k","splat 数が多いほど batching 効果が大きい trend は見えるが、hotdog (80k) と lego (84k) で大きな差。単純な correlation でない。",{"type":65,"level":14,"text":147},"(2) 仮説: compute\u002Fcommit ratio + scene geometry",{"type":149,"text":150},"paragraph","\u003Cstrong>batching benefit は \"commit overhead 削減量\" だが、その効果が iter time に占める割合は scene による\u003C\u002Fstrong>:",{"type":152,"items":153},"list",[154,155,156],"chair (-18.6%): 椅子の \u003Cstrong>thin geometry (脚 \u002F 背)\u003C\u002Fstrong> は per-tile 多数 splat 通過 → per-tile compute は重いが各 dispatch は短く済む → commit overhead 占有率 高 → batching 効果 大","hotdog (-5.4%): \u003Cstrong>平面 + 滑らかな形状\u003C\u002Fstrong> は per-pixel compositing が単純、iter time は GPU compute bound → commit overhead 占有率 低 → batching 効果 小","ficus (-1.6%): \u003Cstrong>細い枝の繰り返し\u003C\u002Fstrong>、refine の grad-driven split が高頻度 → per-iter で host CPU 側 refine 判定が dominant → batching 対象外 (refine は変更不可)",{"type":65,"level":14,"text":158},"(3) PSNR drift は高品質シーンで顕在化",{"type":149,"text":160},"\u003Cstrong>hotdog -0.828 dB\u003C\u002Fstrong> は variance σ ±0.32 dB の 2.6x で有意。低品質シーン (ficus 13.9 dB) では drift +0.095 dB と variance 内。これは SIMD atomic 非決定性が \u003Cstrong>absolute PSNR 値が高いシーンほど visible\u003C\u002Fstrong> という性質 (relative ratio で同じ drift でも高品質シーンで「目立つ」)。",{"type":65,"text":162},"有意性判定 (variance baseline 適用)",{"type":149,"text":164},"\u003Ca href=\"\u002Ffindings\u002Fa-10-variance-baseline\">A.10 variance baseline\u003C\u002Fa> から PSNR σ ±0.32 dB \u002F wall σ ±2.4% を有意性 noise floor とすると:",{"type":68,"columns":166,"align":171,"rows":172,"caption":188},[70,167,168,169,170],"wall Δ","wall 有意?","PSNR Δ","PSNR 有意?",[78,79,78,79,78],[173,177,180,183,186],[132,174,175,88,176],"-18.6%","**有意** (7.8 σ)","× 内 (0.3 σ)",[90,140,178,96,179],"**有意** (2.6 σ)","× 内 (0.9 σ)",[135,137,181,182,178],"**有意** (2.3 σ)","-0.828",[106,109,184,112,185],"× 内 (1.4 σ)","× 内 (0.4 σ)",[114,116,187,119,176],"× 内 (0.7 σ)","5 scene 中 3 scene (chair \u002F lego \u002F hotdog) で wallclock 有意改善、drums \u002F ficus は variance 内。PSNR drift は hotdog のみ有意悪化。",{"type":65,"text":190},"卒論への含意",{"type":149,"text":192},"Chapter 第 3 軸 (Apple Silicon 固有最適化) で A.7 batching を「\u003Cstrong>scene-specific positive finding\u003C\u002Fstrong>」として記述:",{"type":152,"items":194},[195,196,197,198],"適用シーンを選べば wallclock -5〜-19% 改善 (chair \u002F lego \u002F hotdog で有意)","適用シーンによっては PSNR drift -0.83 dB (hotdog) — atomic 非決定性が高品質シーンで顕在化","scene-independent な \"汎用最適化\" としては不採用が妥当","scene workload analysis (compute\u002Fcommit ratio profiling) を deployment 前に行うべき",{"type":200,"label":201,"variant":202,"text":203},"callout","Lesson","info","「\u003Cstrong>single-scene bench の数字を universal claim に使うのは危険\u003C\u002Fstrong>」が今回最大の方法論的 lesson。A.7 lego -6.16% で commit して、他 4 シーン bench せずに「+6% positive」と書けば misleading だった。卒論 evaluation には multi-scene bench を必須にすべき。",{"type":65,"text":205},"残作業 (defer)",{"type":152,"items":207},[208,209,210],"scene geometry の compute\u002Fcommit ratio を Instruments Metal Trace で profile (仮説の rigorous 検証)","backward atomic を non-atomic reduction に書き換え、hotdog の PSNR drift を解消する A.7 part 2","Tanks & Temples real-world scene で A.7 適用 (現状 NeRF Synthetic のみ、real-world は別 dynamics の可能性)",{"type":65,"text":212},"関連",{"type":152,"items":214},[215,216,217,218,219],"A.7 lego 単独 finding: \u003Ccode>a-7-icb-batching-results\u003C\u002Fcode>","A.7 implementation plan: \u003Ccode>a-7-icb-batching-plan\u003C\u002Fcode>","A.10 variance baseline (有意性 noise floor): \u003Ccode>a-10-variance-baseline\u003C\u002Fcode>","A.4 NeRF Synthetic multi-scene 8 シーン: \u003Ccode>a-4-nerf-synthetic-scene-results\u003C\u002Fcode>","A.5 final ablation 表: \u003Ccode>final-ablation-table\u003C\u002Fcode>",[221,230,237,244,251,260,270,278,286,294],{"id":26,"title":26,"subtitle":222,"date":9,"workspace":223,"tags":224,"verdict":42,"psnr":226,"psnr_unit":-1,"wallclock":227,"splats":228,"summary_url":229,"detail_path":229},"A.7 × multi-scene: chair 30k with SPLAT_BATCHED_FORWARD=1","splat",[225,26],"auto-bench",22.77984046936035,"18m 39s",39451,"\u002Fruns\u002Fchair-batched-30k\u002F",{"id":30,"title":30,"subtitle":231,"date":9,"workspace":223,"tags":232,"verdict":42,"psnr":233,"psnr_unit":-1,"wallclock":234,"splats":235,"summary_url":236,"detail_path":236},"A.7 × multi-scene: drums 30k with SPLAT_BATCHED_FORWARD=1",[225,30],17.90349578857422,"20m 43s",64575,"\u002Fruns\u002Fdrums-batched-30k\u002F",{"id":28,"title":28,"subtitle":238,"date":9,"workspace":223,"tags":239,"verdict":42,"psnr":240,"psnr_unit":-1,"wallclock":241,"splats":242,"summary_url":243,"detail_path":243},"A.7 × multi-scene: ficus 30k with SPLAT_BATCHED_FORWARD=1",[225,28],14.05406379699707,"22m 32s",61599,"\u002Fruns\u002Fficus-batched-30k\u002F",{"id":32,"title":32,"subtitle":245,"date":9,"workspace":223,"tags":246,"verdict":42,"psnr":247,"psnr_unit":-1,"wallclock":248,"splats":249,"summary_url":250,"detail_path":250},"A.7 × multi-scene: hotdog 30k with SPLAT_BATCHED_FORWARD=1",[225,32],29.461942672729492,"22m 35s",80841,"\u002Fruns\u002Fhotdog-batched-30k\u002F",{"id":34,"title":34,"subtitle":252,"date":9,"workspace":223,"tags":253,"verdict":42,"psnr":256,"psnr_unit":-1,"wallclock":257,"splats":258,"summary_url":259,"detail_path":259},"A.7 batched cmd buffer — extract+rasterize.forward と backward chain を 1 cmd buffer 集約",[17,254,255,16],"icb-batching","lego-30k",24.5767765045166,"21m 47s",82855,"\u002Fruns\u002Flego-a7-batched-30k\u002F",{"id":25,"title":25,"subtitle":261,"date":262,"workspace":223,"tags":263,"verdict":42,"psnr":266,"psnr_unit":-1,"wallclock":267,"splats":268,"summary_url":269,"detail_path":269},"A.4 NeRF Synthetic 他シーン展開 — chair 30k baseline (sh=3)","2026-05-22",[264,25,265,16],"scene-ablation","scene-chair",22.88273811340332,"19m 51s",38928,"\u002Fruns\u002Fchair-30k\u002F",{"id":29,"title":29,"subtitle":271,"date":262,"workspace":223,"tags":272,"verdict":42,"psnr":274,"psnr_unit":-1,"wallclock":275,"splats":276,"summary_url":277,"detail_path":277},"A.4 NeRF Synthetic 他シーン展開 — drums 30k baseline (sh=3)",[264,29,273,16],"scene-drums",17.772979736328125,"21m 23s",64515,"\u002Fruns\u002Fdrums-30k\u002F",{"id":27,"title":27,"subtitle":279,"date":262,"workspace":223,"tags":280,"verdict":42,"psnr":282,"psnr_unit":-1,"wallclock":283,"splats":284,"summary_url":285,"detail_path":285},"A.4 NeRF Synthetic 他シーン展開 — ficus 30k baseline (sh=3)",[264,27,281,16],"scene-ficus",13.958805084228516,"22m 48s",60938,"\u002Fruns\u002Fficus-30k\u002F",{"id":31,"title":31,"subtitle":287,"date":262,"workspace":223,"tags":288,"verdict":42,"psnr":290,"psnr_unit":-1,"wallclock":291,"splats":292,"summary_url":293,"detail_path":293},"A.4 NeRF Synthetic 他シーン展開 — hotdog 30k baseline (sh=3)",[264,31,289,16],"scene-hotdog",30.290376663208008,"23m 46s",82154,"\u002Fruns\u002Fhotdog-30k\u002F",{"id":33,"title":33,"subtitle":295,"date":262,"workspace":223,"tags":296,"verdict":42,"psnr":299,"psnr_unit":-1,"wallclock":300,"splats":301,"summary_url":302,"detail_path":302},"A.8 SH degree ablation — sh_degree=3 (DC only, no view-dependence)",[297,255,298,16],"sh-ablation","sh-3",24.87872886657715,"23m 13s",83734,"\u002Fruns\u002Flego-sh3-30k\u002F",[304,323,341],{"id":38,"title":305,"date":306,"status":10,"polarity":12,"category":11,"axes":307,"tags":309,"task_code":314,"related_runs":315,"delta_psnr":319,"delta_wallclock":320,"rank":41,"verdict":42,"impact_summary":321,"detail_path":322},"A.4 NeRF Synthetic 他シーン展開 — 8 シーン complete 30k 結果","2026-05-24",[308],1,[16,310,20,311,21,312,313],"nerf-synthetic","psnr","evaluation","8-scenes","A.4",[33,25,27,29,31,316,317,318],"mic-30k","materials-30k","ship-30k","-5.93 dB (8 シーン平均 18.95 vs lego 24.879、std ±6.0)","21-29 min (シーン非依存的、materials のみ +5 min)","8 シーン complete (lego + 7 新規) 30k 完遂。シーン依存性が PSNR で 17.6 dB の幅 (materials 12.71 〜 hotdog 30.29)、mean 18.95 ± 6.0 dB。本実装の brush SoTA 比 gap は scene-dependent で -7.4 dB (hotdog) 〜 -22.3 dB (ficus 含む)。共通要因仮説: SfM init.ply の sparsity (細い枝 \u002F マイク \u002F 反射 PBR で薄い) + refine grad_threshold の lego\u002Fhotdog tuning over-fit。卒論 evaluation で「lego baseline + multi-scene mean ± std」併記必須。","\u002Ffindings\u002Fa-4-nerf-synthetic-scene-results\u002F",{"id":37,"title":324,"date":9,"status":10,"polarity":325,"category":11,"axes":326,"tags":327,"task_code":332,"related_runs":333,"delta_psnr":337,"delta_wallclock":338,"rank":41,"verdict":10,"impact_summary":339,"detail_path":340},"A.10 variance baseline — σ ±0.32 dB \u002F range 0.885 dB を実測","negative",[14],[16,328,329,330,331,22],"variance","gpu-non-determinism","kahan","atomic","A.10",[33,334,335,336],"lego-variance-trial1-30k","lego-variance-trial2-30k","lego-variance-trial3-30k","σ ±0.32 dB \u002F range 0.885 dB","σ ±2.4% \u002F range 5.2%","M-3.x lego sh3 30k の PSNR variance は σ ±0.32 dB \u002F range 0.885 dB (4 run estimate)、wallclock variance は σ ±2.4% \u002F range 5.2%。原因は SIMD backward kernel の atomic_fetch_add 順序非決定性で、A.10 Kahan で消えない (compensator も bit-identical のところ)。卒論 finding として「Apple Silicon の variance band は数値精度の問題でなく GPU scheduler 由来」と確定。","\u002Ffindings\u002Fa-10-variance-baseline\u002F",{"id":36,"title":342,"date":9,"status":10,"polarity":343,"category":11,"axes":344,"tags":345,"task_code":23,"related_runs":350,"delta_psnr":351,"delta_wallclock":352,"rank":353,"verdict":354,"impact_summary":355,"detail_path":356},"A.7 batched cmd buffer — wallclock -6.2% 改善 + PSNR drift -0.30 dB","positive",[14],[16,18,346,347,348,22,349],"command-buffer","batching","metal","results",[33,34],"-0.302 dB (24.577 vs 24.879)","-6.16% (1307.26s vs 1393s = -85.74s)","mid","accepted","scope B 限定版 (forward 末尾の extract_offsets + rasterize.forward を 1 cmd buffer、backward chain の rasterize.backward + project_backwards を 1 cmd buffer) を env SPLAT_BATCHED_FORWARD=1 で活性化。30k bench で wallclock -6.2% \u002F PSNR drift -0.30 dB、Mildly positive。期待 -6〜-12% の下限、Apple Silicon の commit overhead が予想より小さい示唆。","\u002Ffindings\u002Fa-7-icb-batching-results\u002F",1782449788619]