GPT-6.2 Sol 2.35×
Harness: Codex · switches: GS_STAGE_BACKWARD_PIXELS=1 GS_STAGE_FORWARD_SPLATS=1 GS_TRITON_ADAM=1 GS_ELIDE_DEPTH_ADJOINT=1 GS_FUSED_DENSIFICATION_STATS=1 GS_BATCH_PROGRESS_LOG=1 GS_COOPERATIVE_TILE_DUPLICATION=1 GS_ELIDE_UNUSED_INVDEPTH=1 GS_ELIDE_BLACK_BACKGROUND_ADJOINT=1 GS_TRITON_RENDER_ACTIVATIONS=1 GS_COMPACT_BACKWARD_PIXELS=1 GS_SKIP_EMPTY_PIXEL_BUCKETS=1 GS_GROUP_FORWARD_BUCKETS=1 GS_BOUNDED_ALPHA_RECIPROCAL=1 GS_FACTORED_SH_ADAM=1 GS_PACK_FORWARD_COLORS=1 GS_ORDER_FORWARD_TILES=1
- Judged
- 2.35×
- Tier
- N1
- Self-best
- 2.35×
- Rounds
- 48
- Hours
- 10.0
- API cost
- $22.3
- Input tokens · 97.4% cached
- 99 M
- Output tokens
- 0.7 M
- Requests
- 711
§ ARaw trace
Raw work log48 sections
0.0 hRound 1 — dense Adam fusion (started 2026-11-17 18:47:35 UTC)
- Unchanged reference, seed 0, steps 7000–7019:/tmp/r1/profile(PID 2433,/tmp/r1/profile.log). By stage, rasterizer backward 34.3% of CUDA time (per-Gaussian render backward 27.2%), rasterizer forward 26.4% (tile render 16.5%, projection 5.1%, duplicate 5.1%), and Adam approximately 25.8% (foreach addcdiv 5.4%, lerp 3.8%, addcmul 3.7%, div 3.4%, sqrt 3.3%, mul 2.9%). Densification at step 7000 also contributed gather traffic. CPU render call waits for rasterizer host readbacks; the percentage above is GPU attribution, not a timing forecast. - Enable the existing float32 fused dense Adam backend viaGS_FUSED_ADAM=1; leave group names, scheduled LRs, epsilon, moments, and exposure optimizer alone. Numerics target N1. Time then quick check in the experiment tree (PID 2647;/tmp/r1/time.log,/tmp/r1/quick.log). - Timing finished at 18:55:33 UTC:/tmp/r1/time/time.json, 265.085 s, 1.1599×, windows 74.449 / 97.183 / 93.453 s, 5.33 GiB peak allocated. Quick check/tmp/r1/quick/check.json: PASS at 18:55:56 UTC. Round ended at 18:55:56 UTC; promoted on time + exact evidence, N1 check pending. - Full check/tmp/r1/check/check.jsonat 19:11:02 UTC: PASS (N1), 1.1632×, 264.340 s. Exact items match both seed 0 short and seed 1 full training. Endpoint differences in σ (PSNR, SSIM, LPIPS, log count): −1.75, −1.36, +0.15, +0.07, all within ±3σ.
0.1 hRound 2 — cooperative backward rasterizer loads (started 2026-11-17 18:58:26 UTC)
- Post-densification profile of the fused-Adam version at steps 17001–17020, then round 1 full check. Detached runner PID 3462 (/tmp/r2/profile.log,/tmp/r1/check.log). - Late profile (/tmp/r2/profile): rasterizer backward 43.7% of CUDA time (per-Gaussian blending 34.7%, 3.031 ms/step), forward 31.3% (tile blending 18.7%, projection 6.5%, tile duplication 6.3%), fused Adam 16.6%. This is a distinct steady-state phase; round 1's 7000-step trace includes a densification iteration. - In/tmp/r2/tree, testGS_STAGE_BACKWARD_PIXELS=1: coalesce bucket checkpoints and pixel adjoint loads, then share the data within each warp instead of loading them serially in lane 0. Preserve the existing wavefront recurrence, thresholds, float32 gradients and atomics. Four independent warps per block; numerics target N1 due to atomic scheduling. - PID 4234 (/tmp/r2/runner.log), build after round 1's timing completes (/tmp/r2/build.log), time and quick after the round 1 full check releases the GPU (/tmp/r2/time.log,/tmp/r2/quick.log). - Timing completed 19:15:17 UTC:/tmp/r2/time/time.json, 1.2689×, 242.331 s (68.150 / 88.922 / 85.259 s; 5.31 GiB peak allocated), a 22.754 s improvement over round 1. Quick exact check/tmp/r2/quick/check.json: PASS at 19:15:40 UTC. - Operator probe: independent forwards produced identical RGB, depth and radii, but some multibucket gradient comparisons were not valid as standalone backward checks (up to 2.3% relative L2, consistent with the untouched forward's CUB maximum-contributor state not being controlled). Passing the same forward buffers and adjoints into each kernel resolved the mismatch: bitwise single tiles and worst relative L2 below 5e-8 across all raster gradients on eight single/multi-tile cases, partial edges, SH degrees 0/1/3 and depth+RGB adjoints (/tmp/r2/shared_forward_compare.log). No forward or gradient rule changed. Round ended 19:19:17 UTC, pending full check. - Full check/tmp/r2/check/check.jsonat 19:29:57 UTC: PASS (N1), 1.2636×, 243.334 s. Both exact gates passed; endpoint differences in σ (PSNR, SSIM, LPIPS, log count): −1.33, −1.82, −0.38, +0.43. Promoted31c75f2with the switch enabled.
0.5 hRound 3 — forward splat staging (started 2026-11-17 19:19:45 UTC)
- Densification-phase profile at steps 7001–7020, then round 2 full check, PID 5738 (/tmp/r3/profile.log,/tmp/r2/check.log). - Profile/tmp/r3/profile: forward 35.3% (tile blending 22.1%, projection 6.9%, duplication 6.8%); backward 36.4% (staged blending 26.9%, 1.909 ms/step); Adam 18.0%. Unlike round 1, the window excludes the scheduled densification step. - TryGS_STAGE_FORWARD_SPLATS=1in/tmp/r3/tree: extend the cooperative tile batch load to include the splat RGB and its inverse depth, sharing them across pixels instead of gathering and dividing repeatedly. Keep depth output, bucket checkpoints, sort, blend factors and termination criteria the same. - Detached build and timing PID 6001 (/tmp/r3/runner.log,/tmp/r3/build.log,/tmp/r3/time.log,/tmp/r3/quick.log). Timing proceeds only if the round 2 full check passes. - Time/tmp/r3/time/time.jsonat 19:33:59 UTC: 1.3171×, 233.456 s (65.751 / 85.628 / 82.077 s; 5.33 GiB peak allocated), a further 8.875 s saved. Quick exact check/tmp/r3/quick/check.json: PASS at 19:34:21 UTC. - Forward operator probe/tmp/r3/compare_forward.log: RGB, inverse depth, radii, per-pixel contributor counts, bucket-to-tile mapping, and every consumed float32 blending checkpoint bitwise identical on eight synthetic cases with partial edges and varied SH and antialiasing. Round ended 19:34:21 UTC; seed 1 validation queued. - Seed 1 check/tmp/r3/n1/check.jsonat 19:41:57 UTC: PASS (N1), including full-trajectory exact items. Differences in σ (PSNR, SSIM, LPIPS, log count): −2.18, −0.14, −2.06, +1.41. Promoted9272db4after separate time, quick, and N1 passes. - Full check of the submission/tmp/r3/check/check.jsonat 19:57:16 UTC: PASS (N1), 1.3160×, 233.653 s. Both exact gates passed; endpoint differences in σ: +0.29, +1.88, −1.55, +0.98.
0.8 hRound 4 — conservative subthreshold-alpha culling (started 2026-11-17 19:35:11 UTC)
- Steady-state profile at steps 17001–17020, then round 3 seed 1 statistical gate. PID 7254,/tmp/r4/profile.log,/tmp/r3/n1.log. - Profile/tmp/r4/profile: backward 39.5% (blending 29.2%, 2.198 ms/step), forward 31.6% (blending 17.0%, 1.282 ms/step; projection 7.5%, duplication 7.3%), Adam 19.2%. Blending still accounts for 46.2% across the two stages. - TryGS_CULL_SUBTHRESHOLD_ALPHA=1in/tmp/r4/tree: use the next power of two above each finite positive normal opacity (exponents at most 127) as an upper bound. Precompute a power cutoff for alpha below 1/256 with a guard margin, then skipexponly below that cutoff. The original 1/255 test, alpha cap, and blend arithmetic still run for all other comparisons, including inputs outside the proven range. - Build, time and quick PID 7555 (/tmp/r4/build.log,/tmp/r4/time.log,/tmp/r4/quick.log). Timing proceeds only if round 3 passes the statistical gate. - Time/tmp/r4/time/time.jsonat 19:46:01 UTC: 1.3135×, 234.098 s (66.150 / 86.216 / 81.732 s), 0.642 s slower than round 3. Quick exact check/tmp/r4/quick/check.json: PASS. Round ended at 19:46:23 UTC; no net gain, not enabled or promoted. The cutoff's additional branch/load may offset the skipped exponentials.
1.0 hRound 5 — specialized dense Adam (started 2026-11-17 19:47:07 UTC)
- Densification-phase profile at steps 10001–10020, then the combined full check of round 3. PID 8467 (/tmp/r5/profile.log,/tmp/r3/check.log). - Profile/tmp/r5/profile: backward 37.1% (blending 27.4%), forward 33.8% (blending 20.4%, projection 7.1%, duplication 6.7%), dense Adam 19.0% (1.447 ms/step). The native fused backend uses double scalar arithmetic for moment updates and maintains generic multi-tensor metadata; test a contiguous float32 kernel with host-computed bias corrections. -GS_TRITON_ADAM=1in/tmp/r5/treereplaces only the Gaussian default optimizer backend. It retains named groups, full dense moment decay and updates even for zero gradients, CPU step tensors, float32 moment dtypes, the same epsilon and step-dependent bias correction. Non-contiguous or unsupported parameters fall back to PyTorch's Adam functional. Numerics target N1. - Detached PID 8758 (/tmp/r5/runner.log,/tmp/r5/build.log,/tmp/r5/compare_adam.log,/tmp/r5/time.log,/tmp/r5/quick.log). Operator probe and timing proceed after the current full check passes. - Operator probe/tmp/r5/compare_adam.log: all six shapes with odd point counts, 128 updates including zero and absent gradients and changing position LR. CPU step and first moments bitwise, second moments within tolerance, maximum raw parameter difference 4.77e-7 against reference foreach Adam. Float32 throughout. - Time/tmp/r5/time/time.jsonat 20:01:13 UTC: 1.3900×, 221.214 s (63.009 / 81.100 / 77.104 s; 5.33 GiB peak allocated). Further 12.242 s saved replacing PyTorch fused Adam with the specialized backend. Quick exact check/tmp/r5/quick/check.json: PASS at 20:01:36 UTC, round closed. Seed 1 gate pending. - Seed 1 check/tmp/r5/n1/check.jsonat 20:08:41 UTC: PASS (N1) including full exact items. Endpoint differences in σ: −0.39, +1.45, −1.12, +0.03. Promotedf0ec3d3withGS_TRITON_ADAM=1replacing the built-in fused backend switch.
1.2 hRound 6 — elide an absent depth adjoint (started 2026-11-17 20:02:10 UTC)
- Steady-state profile at 17001–17020 then round 5 statistical gate. PID 10155 (/tmp/r6/profile.log,/tmp/r5/n1.log). - Profile/tmp/r6/profile: backward 42.4% (blending 31.3%, 2.199 ms/step); forward 34.0% (blending 18.2%, projection 8.0%, duplication 7.8%); specialized Adam 13.1%, 0.918 ms/step vs 1.445 ms at the same phase with PyTorch fused Adam. -GS_ELIDE_DEPTH_ADJOINT=1in/tmp/r6/tree: stop materializing unused depth output gradients as zeros, and dispatch a CUDA backward specialization without depth shuffles, memory traffic, derivatives or atomics when that adjoint is absent. RGB, forward depth and the derivative path for a real depth loss remain available. Same math for finite inputs, N1 target due to atomic ordering. - Build, same-forward-buffer gradient comparison and screening PID 10524 (/tmp/r6/build.log,/tmp/r6/compare_depth.log,/tmp/r6/time.log,/tmp/r6/quick.log). GPU tests wait for round 5 to pass N1. - Operator comparison/tmp/r6/compare_depth.log: same forward buffers, absent depth adjoint via autograd versus an explicit zero tensor, all nine CUDA gradients relative L2 below 4e-8 on eight cases and exact on single tiles. Kernel register count 64→56, shared memory 28,160→23,584 B per block for staged backward. - Time/tmp/r6/time/time.jsonat 20:12:29 UTC: 1.4283×, 215.277 s (61.702 / 78.622 / 74.953 s; 5.32 GiB allocated), further 5.937 s saved. Quick exact check/tmp/r6/quick/check.json: PASS at 20:12:52 UTC; round closed. N1 validation pending. - Seed 1 check/tmp/r6/n1/check.jsonat 20:17:42 UTC: PASS (N1) including full exact items. Endpoint differences in σ: −1.93, +0.15, −0.68, +0.70. Promoteded60178. - Full check of the submission/tmp/r6/check/check.jsonat 20:27:01 UTC: PASS (N1), 1.4279×, 215.342 s. Both exact gates passed and endpoint differences in σ were −1.27, −1.76, +0.69, +0.63.
1.4 hRound 7 — fuse visibility and densification bookkeeping (started 2026-11-17 20:17:43 UTC)
- Round 6 statistical gate then profile at steps 7001–7020 if it passes; PID 11528 (/tmp/r6/n1.log,/tmp/r7/profile.log). - Full check of the round 6 submission queued after the profile, PID 12041 (/tmp/r6/check.log). - Profile/tmp/r7/profileat 7001–7020: backward blending 28.3% CUDA, forward blending 21.4%, Adam 12.8%.aten::nonzeroused for a compact visibility list includes a blocking host wait (19.1% inclusive CPU, mostly waiting for the queued render); the list then drives 5 gathers and 3 scatters per step in this phase. - TryGS_FUSED_DENSIFICATION_STATS=1in/tmp/r7/tree: a single predicated float32 kernel updates max radii, gradient norm sum, and visibility count for the sameradii > 0points, avoiding the compact list and CPU readback. Keep optional visibility computation for other renderer callers and fall back to the original indexing path on unsupported layouts. - Build/probe/time/quick job PID 12221 (/tmp/r7/runner.log,/tmp/r7/build.log,/tmp/r7/compare_stats.log,/tmp/r7/time.log,/tmp/r7/quick.log). GPU tests wait for the round 6 full check to pass. - Operator probe/tmp/r7/compare_stats.log: all three stats bitwise identical after 20 accumulated updates for eight shapes (0–262,147 points, with both visible and invisible radii and tail cases). End-to-end tier target is N1 because removing synchronization can affect atomic ordering. - Time/tmp/r7/time/time.jsonat 20:30:38 UTC: 1.4980×, 205.267 s (58.148 / 74.959 / 72.159 s; 5.30 GiB peak allocated), further 10.010 s saved. Quick exact check/tmp/r7/quick/check.json: PASS at 20:31:00 UTC; round ended. Seed 1 check pending. - Seed 1 check/tmp/r7/n1/check.jsonat 20:37:41 UTC: PASS (N1) including full exact items. Endpoint differences in σ: −1.77, −1.44, −1.34, +0.52. Promoted5d316e3.
1.7 hRound 8 — eliminate the bucket-count host readback (started 2026-11-17 20:31:38 UTC)
- Steady-state profile at steps 17001–17020, then round 7 N1 gate. PID 13691 (/tmp/r8/profile.log,/tmp/r7/n1.log). - Profile/tmp/r8/profile: backward blending 29.4% CUDA, forward blending 18.9%, dense Adam 13.5%; no compact visibility list. Runtime-to-GPU memcpy correlations show two blocking rasterizer DtoH counts per step, with median inclusive CPU times 427 µs (point/tile-pair count) and 865 µs (bucket count). This includes queued GPU work, not recoverable wall time by itself. -GS_BOUND_BUCKET_ALLOCATION=1in/tmp/r8/tree: size the sample buffer usingfloor((R + 31 * tile_count) / 32)from the already-known pair countR. This upper-bounds the sum of per-tile ceilings even with discarded pairs. The actual count remains on-device and guards backward warps before any sample access. Sort and all forwards/checkpoint math are unchanged. - Build, forward probe, time and quick PID 14086 (/tmp/r8/build.log,/tmp/r8/compare_bucket.log,/tmp/r8/time.log,/tmp/r8/quick.log). GPU tests wait for round 7's N1 gate. - Forward probe/tmp/r8/compare_bucket.log: bitwise RGB, radii, depth, contributor counts and all active checkpoints across eight cases. - Time/tmp/r8/time/time.jsonat 20:41:20 UTC: 1.4978×, 205.286 s (58.682 / 74.675 / 71.929 s), effectively unchanged from 205.267 s. Peak allocated memory increased to 5.91 GiB. Quick exact check/tmp/r8/quick/check.json: PASS at 20:41:42 UTC; round ended. No promotion: the host wait mostly covered GPU work still required, and extra capacity increased memory.
1.9 hRound 9 — batch progress scalar readbacks (started 2026-11-17 20:42:40 UTC)
- Profile at steps 4001–4020 on the submitted round 7 stack, PID 15118 (/tmp/r9/profile.log). - Profile/tmp/r9/profile: backward blending 28.2% of CUDA, forward blending 22.3%, Adam 12.5%. Progress loss has one stream synchronization per step (42.2 ms inclusive CPU wait across 20 steps), separate from the CPU-only Adam step counters. -GS_BATCH_PROGRESS_LOG=1in/tmp/r9/tree: when no TensorBoard writer uses per-step metrics, batch detached loss scalars at the existing progress update interval. Replay the identical EMA recurrence in the same order and skip unconsumed CUDA event timing. TensorBoard mode and the training graph remain unchanged. Detached timing/quick PID 15324 (/tmp/r9/runner.log,/tmp/r9/time.log,/tmp/r9/quick.log). No submission change yet. - Time/tmp/r9/time/time.jsonat 20:49:09 UTC: 1.5348×, 200.335 s (54.113 / 76.925 / 69.297 s; 5.31 GiB peak allocated), 4.932 s below round 7 with mixed window gains. Quick exact check/tmp/r9/quick/check.json: PASS. Round ended at 20:49:43 UTC. Experiment committed asdb25b93(N1 target); full check before promotion. - Full check/tmp/r9/check/check.jsonat 20:59:52 UTC: PASS (N1), 1.5858×, 193.893 s (53.017 / 71.744 / 69.132 s). Both exact gates passed. Endpoint differences in σ (PSNR, SSIM, LPIPS, log count): −2.75, −1.70, +0.08, −0.27. Promoteddb25b93and enabled its switch.
2.1 hRound 10 — warp-cooperative duplication of wide splats (started 2026-11-17 20:50:04 UTC)
- Detached runner PID 16324 (/tmp/r10/profile-check-runner.log) profiles/tmp/r9/treeat steps 17001–17020 (/tmp/r10/profile,/tmp/r10/profile.log), then runs round 9's combined full check (/tmp/r9/check,/tmp/r9/check.log). The submission remains on round 7 pending validation. - Profile/tmp/r10/profile: rasterizer backward 40.8% CUDA (blending 29.4%), forward 35.0% (blending 18.9%, preprocess 8.3%, tile duplication 8.0% or 0.542 ms/step), dense Adam 13.3%. Tile duplication's per-Gaussian traversal stalls a warp behind the largest rectangle; test more parallel work without changing the rectangle or opacity test. -GS_COOPERATIVE_TILE_DUPLICATION=1in/tmp/r10/tree: a warp cooperates on bounding rectangles of at least 32 tiles; smaller rectangles retain the per-lane path. Each Gaussian keeps its original prefix-sum span. Wide splats place the same accepted key/ID pairs in that span, with culled slots becoming the same sentinels, so stable sorting still orders by the same keys in the same order. Build PID 16800 (/tmp/r10/build.log) and detached probe/time/quick PID 17108 (/tmp/r10/runner.log,/tmp/r10/compare_binning.log,/tmp/r10/time.log,/tmp/r10/quick.log). Timing requires round 9's full check to pass. - Operator probe/tmp/r10/compare_binning.log: bitwise sorted keys and IDs, RGB, radii, depth, contributors and active checkpoints in ten cases, including tested wide path execution, partial edge tiles and tail Gaussian lanes. Probe finished before the concurrent N1 evaluator's verdict was written, so the initial runner stopped before timing; restarted time + quick after PASS (PID 17675,/tmp/r10/time-quick-runner.log). - Time/tmp/r10/time/time.jsonat 21:04:04 UTC: 1.6637×, 184.816 s (51.574 / 68.118 / 65.124 s; 5.30 GiB peak allocated). All windows faster than round 9 screening, 15.519 s saved in total, or 9.077 s against the round 9 full check. Quick exact check/tmp/r10/quick/check.json: PASS at 21:04:23 UTC; round ended. Committed0269509(N1 target); pending combined full check before promotion. - Full check/tmp/r10/check/check.jsonat 21:14:02 UTC: PASS (N1), 1.6736×, 183.724 s (50.731 / 67.840 / 65.154 s). Both exact gates passed. Endpoint differences in σ: +0.64 PSNR, −1.34 SSIM, +0.68 LPIPS, +1.43 log count. Promoted0269509and enabled its switch.
2.3 hRound 11 — omit unconsumed inverse-depth blending (started 2026-11-17 21:05:42 UTC)
- Detached runner PID 18130 (/tmp/r11/profile-check-runner.log) profiles the faster round 10 stack at steps 11001–11020 (/tmp/r11/profile,/tmp/r11/profile.log), then runs round 10's full check (/tmp/r10/check,/tmp/r10/check.log). Round 9 remains submitted pending N1. - Profile/tmp/r11/profile: backward blending 30.5% of CUDA, forward blending 22.7% (1.481 ms/step), Adam 13.3%, projection 8.8%, tile duplication after round 10 2.1% (0.136 ms/step). Target forward blending's extra inverse-depth channel; the RGB and transmittance inputs to the loss don't depend on it. -GS_ELIDE_UNUSED_INVDEPTH=1in/tmp/r11/tree: training explicitly requests RGB-only output for cameras without reliable depth. Skip the otherwise unused inverse-depth blend and checkpoint storage/copy; leave RGB, transmittance, ordering, geometry, and default depth output for viewer/evaluation/reliable-depth callers unchanged. Build PID 18482 (/tmp/r11/build.log); staged/unstaged probes, time and quick PID 18797 (/tmp/r11/runner.log,/tmp/r11/compare_invdepth.log,/tmp/r11/with_depth.log,/tmp/r11/time.log,/tmp/r11/quick.log) after round 10's full check passes. - Operator probes: bitwise RGB, radii, contributors and active RGB checkpoints in ten cases with both forward kernels. The compact sample buffer drops exactly one inverse-depth plane per bucket; shared-forward gradient comparisons have max relative L2 3.80865e-07. Depth+RGB backward still runs on the default path (/tmp/r11/with_depth.log). - Time/tmp/r11/time/time.jsonat 21:17:25 UTC: 1.6967×, 181.227 s (50.044 / 66.761 / 64.422 s; 5.24 GiB peak allocated), 3.589 s below round 10 screening with gains in all three windows. Quick exact check/tmp/r11/quick/check.json: PASS at 21:17:43 UTC; round ended. Committeddf6edf7(N1 target); pending seed 1 validation. - Seed 1 check/tmp/r11/n1/check.jsonat 21:24:10 UTC: PASS (N1), including full exact items. Endpoint differences in σ: −2.89 PSNR, −1.75 SSIM, −0.63 LPIPS, +0.08 log count. The relatively small gain and PSNR margin warrant stronger validation before enabling this in the submission. Round 10 remains submitted while round 12 screens.
2.5 hRound 12 — skip the zero black-background adjoint (started 2026-11-17 21:18:42 UTC)
- Detached runner PID 20020 (/tmp/r12/profile-n1-runner.log) profiles the faster round 11 stack at steps 17001–17020 (/tmp/r12/profile,/tmp/r12/profile.log), then runs round 11's seed 1 gate (/tmp/r11/n1,/tmp/r11/n1.log). Round 10 remains submitted pending N1. - Profile/tmp/r12/profile: backward blending 31.3% CUDA (1.974 ms/step), forward RGB-only blending 19.7% (1.241 ms/step), dense Adam 14.3%, projection 8.9%. Disassembly of the backward inner loop shows both an alpha reciprocal and a separate background-transmittance division per contributing splat/pixel pair. -GS_ELIDE_BLACK_BACKGROUND_ADJOINT=1in/tmp/r12/tree: stage the same pixel values and vote once per warp; when the RGB background is zero and the tile adjoints/transmittances are finite, skip only the zero background derivative term. Nonzero backgrounds and nonfinite pixel state use the same original term; RGB and depth gradient expressions are unchanged. Build PID 20399 (/tmp/r12/build.log); matched-buffer backward probes and judge time/quick PID 20692 (/tmp/r12/runner.log,/tmp/r12/compare_background.log,/tmp/r12/time.log,/tmp/r12/quick.log), waiting for round 11's N1 gate. - Operator probes/tmp/r12/compare_background.log: 32 comparisons; max relative L2 3.1e-08 with the same forward state, covering black/nonblack, missing/present depth, and a nonfinite-adjoint fallback. Exceptional-value masks also agree. - Time/tmp/r12/time/time.jsonat 21:27:31 UTC: 1.7412×, 176.592 s (48.777 / 65.187 / 62.628 s; 5.23 GiB peak allocated), 4.635 s below round 11 screening with gains in each window. Quick exact check/tmp/r12/quick/check.json: PASS at 21:27:48 UTC; round ended. Committed4e87e4a(N1 target). Validate the whole stack in one full check before promotion. - Combined full check/tmp/r12/check/check.jsonat 21:37:47 UTC: PASS (N1), 1.7113×, 179.679 s (51.111 / 65.777 / 62.790 s). Both exact gates passed. Endpoint differences in σ: −2.10 PSNR, +0.29 SSIM, −1.37 LPIPS, +0.73 log count. Promoted bothdf6edf7and4e87e4aand enabled their switches.
2.7 hRound 13 — checkpoint per-pixel contribution masks (started 2026-11-17 21:28:35 UTC)
- Detached runner PID 21580 (/tmp/r13/profile-check-runner.log) profiles the faster round 12 stack at steps 17001–17020 (/tmp/r13/profile,/tmp/r13/profile.log), then runs round 12's combined full check (/tmp/r12/check,/tmp/r12/check.log). Round 10 remains submitted pending N1. - Profile/tmp/r13/profile: backward blending 29.0% CUDA (1.783 ms/step), forward blending 20.5% (1.258 ms/step), dense Adam 14.7%, projection 9.2%. The backward recomputes which splats passed each pixel's forward tests before using their weights. -GS_CHECKPOINT_CONTRIBUTOR_MASKS=1in/tmp/r13/tree: stash the forward decisions as one 32-bit mask per pixel/bucket and feed them to the staged backward kernel. Backward reconstructs the same weights in the same order and skips inactive pairs before computing their Gaussian exponent. Additional storage: one plane per bucket. Build PID 21961 (/tmp/r13/build-runner.log,/tmp/r13/build.log) starts after round 12 leaves timing. Probe/time/quick PID 22075 (/tmp/r13/runner.log,/tmp/r13/compare_masks.log,/tmp/r13/time.log,/tmp/r13/quick.log) requires round 12's full check to pass. - Operator probes/tmp/r13/compare_masks.log: 20 comparisons; max relative L2 6.17468e-07 with shared forward state. Exact outputs, contributor counts, sorted keys and active checkpoints with or without inverse depth; every last-contributor bit is set. - Time/tmp/r13/time/time.jsonat 21:40:59 UTC: 1.7539×, 175.314 s (47.315 / 65.733 / 62.267 s; 5.29 GiB peak allocated). Only 1.278 s below round 12 screening and the middle window is slower; not resolved against timing variation. Quick exact check/tmp/r13/quick/check.json: PASS at 21:41:16 UTC. Round ended; committed the experiment asd590a7c(N1 target), no promotion or full gate.
3.0 hRound 14 — compile the Gaussian activation graph (started 2026-11-17 21:48:16 UTC)
- Detached profile PID 23594 (/tmp/r14/profile.log) profiles the round 12 submission at steps 11001–11020 (/tmp/r14/profile). - Profile/tmp/r14/profile: backward blending 28.4% CUDA, forward blending 23.0%, Adam 13.9%, projection 9.2%. Correlating sigmoid, exp and quaternion normalization autograd sequence IDs to launches attributes 420 kernels over 20 steps and 8.830 of 124.197 ms (7.1%) of CUDA kernel time to these activations/adjoints. -GS_COMPILE_RENDER_ACTIVATIONS=1in/tmp/r14/treegroups the existing scale exp, quaternion normalization and opacity sigmoid into one dynamically shapedtorch.compileregion, including their backward. Raw float32 Gaussian parameters, activation definitions, optimizer, and density changes are unmodified. Detached build/probe/time/quick PID 24296 (/tmp/r14/build.log,/tmp/r14/compare_activations.log,/tmp/r14/time.log,/tmp/r14/quick.log). - First attempt: operator probe passed on ordinary tensors, but/tmp/r14/time/time.jsonreturned ERROR after step 1300. Dynamo guarded allnn.Parametershapes as static despite dynamic compilation and hit its recompilation limit after densification. No speed figure for this attempt. - Pass differentiable tensor views instead of Parameter objects to the same compiled function. Probe the real GaussianModel method with twelve replacement Parameter shapes and log recompilation diagnostics; then retry the unchanged judge recipe. Detached PID 26564 (/tmp/r14/dynamic-runner.log,/tmp/r14/compare_dynamic.log,/tmp/r14/time-dynamic.log,/tmp/r14/quick-dynamic.log). - The replacement-parameter probe found one rotation gradient component outside elementwise tolerance, with 3.34e-6 absolute difference. To distinguish cancellation from incorrect arithmetic, compare both paths to the analytic FP64 normalization derivative, using per-row upstream gradient norm divided by the clamped quaternion norm as the error scale. Retain the original checks on all activation outputs and other adjoints; no judge change. Probe/time/quick PID 26716 (/tmp/r14/scaled-runner.log,/tmp/r14/compare_scaled.log). - Twelve replacement Parameter shapes pass the probe with no logged recompilations. Largest scaled FP64 rotation errors at the largest shape: 3.55e-7 eager, 4.47e-7 compiled. - Time/tmp/r14/time-dynamic/time.jsonat 22:02:20 UTC: 1.3314×, 230.938 s (62.523 / 85.527 / 82.888 s; 5.20 GiB peak allocated). Slower in all windows than round 12. Quick exact check/tmp/r14/quick-dynamic/check.json: PASS at 22:02:42 UTC. Round ended; no promotion. Experiment committed as147474a(N1 target).
3.3 hRound 15 — fused Triton rasterizer activations (started 2026-11-17 22:03:33 UTC)
- Detached runner PID 27715 (/tmp/r15/profile-runner.log) profiles the round 12 submission at steps 11001–11020 (/tmp/r15/profile,/tmp/r15/profile.log), then profiles the slow compiled candidate over the same steps (/tmp/r14/profile-compiled,/tmp/r14/profile-compiled.log) to identify the regression. - Current stack profile: backward blending 28.3%, forward blending 23.0%, Adam 13.9%, projection 9.2%, eager activations/adjoints approximately 0.44 ms per step. Compiled diagnostic: forward + backward 2.43 ms per step, including about 2.38 ms in its two norm-reduction kernels. The slowdown is inside the generated GPU operations rather than residual recompilation. -GS_TRITON_RENDER_ACTIVATIONS=1in/tmp/r15/tree: batch many Gaussians per block, with one fused activation kernel and one fused adjoint kernel, and a fixed four-channel quaternion reduction. Use the eager exp/sigmoid/normalize formulas and float32, with explicit rounded division and no FMA contraction. Unsupported input layouts or custom activations use the existing getters. Build/probe/time/quick PID 28297 (/tmp/r15/runner.log,/tmp/r15/build.log,/tmp/r15/compare_activations.log,/tmp/r15/time.log,/tmp/r15/quick.log). - Operator probes/tmp/r15/compare_activations.log: verified the fused path fires across twelve replacement sizes including tail tiles and zero/clamped quaternion norms. Scale/opacity outputs and adjoints pass; worst scaled FP64 rotation-adjoint error 4.18e-7 (eager 3.55e-7). - Time/tmp/r15/time/time.jsonat 22:11:01 UTC: 1.8463×, 166.541 s (45.642 / 62.096 / 58.804 s; 5.22 GiB peak allocated), further 10.051 s saved vs. round 12's screening with gains in every window. Quick exact check/tmp/r15/quick/check.json: PASS at 22:11:19 UTC. Round ended; experiment committed asd159843with combined full check pending. - Full check/tmp/r15/check/check.jsonat 22:20:48 UTC: PASS (N1), 1.8222×, 168.739 s (47.539 / 62.223 / 58.976 s). Both exact gates passed. Endpoint differences in σ: −0.18 PSNR, +0.15 SSIM, −0.26 LPIPS, −0.27 log count. Promotedd159843and its switch to the submission.
3.4 hRound 16 — compact the backward pixel wavefront (started 2026-11-17 22:12:12 UTC)
- Detached PID 29366 (/tmp/r16/profile-check-runner.log) profiles the faster round 15 stack at steps 17001–17020 (/tmp/r16/profile,/tmp/r16/profile.log), then runs round 15's combined full check (/tmp/r15/check,/tmp/r15/check.log). Round 12 remains submitted pending N1. - Profile/tmp/r16/profile: backward blending 30.9% CUDA (1.777 ms/step), forward blending 21.5% (1.234 ms/step), Adam 15.7%, projection 9.8%. The backward loop still traverses the entire tile even when pixels terminated before this bucket or are outside a partial tile. -GS_COMPACT_BACKWARD_PIXELS=1in/tmp/r16/tree: warp ballots compact only the valid pixels whose last contributor is beyond the bucket's first index. Traverse them in original pixel order with the same per-splat recurrence, skip no contributing pixels, and leave forward untouched. Build (after round 15 timing), same-state probes, time and quick PID 29799 (/tmp/r16/runner.log,/tmp/r16/build.log,/tmp/r16/compare_pixels.log,/tmp/r16/time.log,/tmp/r16/quick.log). Requires round 15's N1 pass before GPU probes and screening. - Shared-forward-state probes: 27 cases including partial tiles, black/nonblack backgrounds, and RGB-only or inverse-depth gradients. Maximum relative L2 8.11e-7; special-value masks agree. - Time/tmp/r16/time/time.jsonat 22:23:49 UTC: 1.9100×, 160.983 s (45.125 / 59.368 / 56.489 s), a further 5.558 s below round 15 screening. Quick exact check/tmp/r16/quick/check.json: PASS at 22:24:09 UTC; round ended. Committedb48790f; submission stays on round 15 pending the full check. - Full check/tmp/r16/check/check.jsonat 22:32:01 UTC: PASS (N1), 1.9084×, 161.122 s (43.677 / 60.898 / 56.547 s). Both exact gates passed. Endpoint differences in σ: −0.66 PSNR, −0.42 SSIM, −0.53 LPIPS, +1.20 log count. Promotedb48790fand enabled compaction.
3.6 hRound 17 — vectorize the dense Adam memory path (started 2026-11-17 22:24:32 UTC)
- Detached PID 31506 (/tmp/r17/profile-check-runner.log) profiles the faster round 16 candidate at steps 11001–11020 (/tmp/r17/profile,/tmp/r17/profile.log), then runs its full check (/tmp/r16/check,/tmp/r16/check.log). - Profile/tmp/r17/profile: backward blending 27.0% CUDA (1.540 ms/step), forward blending 25.7%, dense Adam 15.3% (0.872 ms/step), projection 10.1%. Triton PTX showed scalar loads and stores despite four adjacent elements per lane: a per-element tail predicate prevents instruction vectorization in every block. -GS_VECTORIZED_DENSE_ADAM=1in/tmp/r17/tree: use an unmasked block path whenever all lanes have valid elements, retaining the same bounds mask for the tail. Both paths call the same update body; no change to the state, step counters, or float32 arithmetic. Detached probe, PTX inspection, judge time and quick PID 31936 (/tmp/r17/runner.log,/tmp/r17/compare_adam.log,/tmp/r17/time.log,/tmp/r17/quick.log) wait for round 16's full PASS. - Offline SM90 compilation confirms that each full interior block uses 128-bit loads/stores (8 load and 6 store instructions versus 32 and 24 scalar instructions). The tail still uses guarded accesses. - Runtime probe/tmp/r17/compare_adam.log: bitwise parameters and moments after 128 steps across seven sizes covering tails and block boundaries and one misaligned case; actual launches also generate 128-bit accesses. - Time/tmp/r17/time/time.jsonat 22:36:56 UTC: 1.9004×, 161.800 s (44.759 / 60.675 / 56.366 s). This is 0.817 s slower than round 16 screening, no resolved improvement. Quick exact check/tmp/r17/quick/check.json: PASS at 22:37:15 UTC; round ended. Committed267be88in the experiment worktree, not promoted.
3.8 hRound 18 — contract the backward color residual (started 2026-11-17 22:37:33 UTC)
- Detached profile PID 33760 (/tmp/r18/profile.log) measures the round 16 submission at steps 7001–7020 (/tmp/r18/profile) to cover the densification phase as well as the earlier post-densification profiles. - Profile took the GPU lock at 22:39:31 UTC after a cache miss rebuilding the actual submission's extensions. Backward blending is 27.4% CUDA (1.375 ms/step), forward blending 25.2%, Adam 15.2%, projection 9.7%. -GS_CONTRACT_BACKWARD_COLOR_ADJOINT=1in/tmp/r18/tree: in staged RGB-only backward, use linearity of the dot product with the constant per-pixel RGB adjoint to carry a single scalar residual between warp lanes. Preserve per-channel RGB gradients, transmittance and contributor tests, and use the original path for depth. This reassociates float32 arithmetic, so target N1. - Detached build/probe/time/quick/conditional N1 PID 34505 (/tmp/r18/runner.log,/tmp/r18/build.log,/tmp/r18/compare_adjoint.log,/tmp/r18/time.log,/tmp/r18/quick.log,/tmp/r18/n1.log). N1 runs only after a quick PASS and a screening time below 158.5 s. - Operator probe: 18 shared-forward-state cases with black and nonblack backgrounds, varying SH and antialiasing, partial tiles, and RGB-only gradients; worst relative L2 2.19e-6. - Time/tmp/r18/time/time.jsonat 22:46:11 UTC: 1.8962×, 162.158 s (49.714 / 57.627 / 54.817 s). Later windows improved, but early training regressed and the total is 1.175 s slower than round 16 screening. Quick exact check/tmp/r18/quick/check.json: PASS at 22:46:29 UTC; round ended. N1 not run (no resolved gain); committed1397f34in the experiment worktree and left it disabled.
4.0 hRound 19 — streamline splat falloff evaluation (started 2026-11-17 22:46:42 UTC)
- Detached profile PID 35513 (/tmp/r19/profile.log) measures the submitted round 16 stack after densification at steps 15001–15020 (/tmp/r19/profile). - Profile: forward blending 26.9% CUDA (1.602 ms/step), backward blending 26.6% (1.587 ms/step), Adam 15.2%, projection 9.9%. The baseline disassembly has a multi-instruction range-reduction sequence around each exponent, even though floor/cap and contributor checks use the exponent only over a narrow active range. -GS_FAST_SPLAT_FALLOFF=1in/tmp/r19/tree: specialize both directions to the same float32__expfintrinsic for the splat falloff. Default math is unchanged; no change to tests, thresholds, weights, ordering, or precision elsewhere. N1 target. Detached build/probes/time/quick/conditional N1 PID 35803 (/tmp/r19/runner.log,/tmp/r19/build.log,/tmp/r19/compare_falloff.log,/tmp/r19/time.log,/tmp/r19/quick.log,/tmp/r19/n1.log). - Initial probe stopped on a wide-splat multi-tile stress case with gradient relative L2 2.00e-4 despite unchanged last contributor counts and RGB relative L2 below 7e-8. Guard the alpha-floor neighborhood with the original exponent (within 1e-5 relative of 1/255) and keep the original alpha comparison. Guard build/probes PID 36607 (/tmp/r19/build-guard.log,/tmp/r19/compare_guard.log). - Repeated original-exponent controls on that large stress shape (/tmp/r19/cutoff-control.log,/tmp/r19/repeat-{0,1}.log) also show up to 2.43e-4 relative gradient spread with identical forward results, with or without backward pixel compaction. The original operator probe cannot distinguish a change there. Keep the independent wide-shape diagnostic, but constrain the high-scale operator stand-in to one tile to isolate the altered arithmetic; same error thresholds and no judge changes. Detached guarded probe/time/quick/conditional N1 PID 37300 (/tmp/r19/runner-controlled.log,/tmp/r19/compare_controlled.log,/tmp/r19/time.log,/tmp/r19/quick.log,/tmp/r19/n1.log). - Guarded variant: 18 controlled cases with exact radii and last contributor counts; worst relative output/adjoint L2 1.65e-6. Time/tmp/r19/time/time.jsonat 23:00:02 UTC: 1.8911×, 162.598 s (45.250 / 60.600 / 56.748 s). Quick/tmp/r19/quick/check.json: PASS at 23:00:21 UTC. Slower than round 16; not promoted. Committed7ba90c3in the experiment worktree. - Screen the original unguarded arithmetic using the same controlled probes. Its extension is already built. Detached PID 37973 uses/tmp/r19/unguarded-treeand/tmp/r19/switches.json; logs/tmp/r19/{runner-unguarded,compare_unguarded,time-unguarded,quick-unguarded,n1-unguarded}.log. - Unguarded variant: 18 controlled cases, worst relative output/adjoint L2 1.65e-6, exact radii and last contributor counts. Time/tmp/r19/time-unguarded/time.jsonat 23:05:02 UTC: 1.9509×, 157.613 s (43.812 / 58.414 / 55.387 s); all three windows improve against round 16 screening, 3.370 s saved. Quick/tmp/r19/quick-unguarded/check.json: PASS at 23:05:20 UTC; round ended. Committed168cbe5on top of round 16. Seed 1 validation is running. - Seed 1/tmp/r19/n1-unguarded/check.jsonat 23:08:37 UTC: INCORRECT. PSNR −3.25σ exceeds the 3σ bound; SSIM −1.39σ, LPIPS −0.21σ, log count +0.43σ. Despite exactness and operator evidence, this candidate is rejected and not promoted or retried.
4.3 hRound 20 — projection arithmetic (profile queued 2026-11-17 23:05:45 UTC)
- Detached profile PID 38787 (/tmp/r20/profile-runner.log,/tmp/r20/profile.log) waits for round 19 validation and profiles the resulting best stack at steps 11001–11020 (/tmp/r20/profile). - Profile took the lock at 23:08:42 UTC on round 16: backward blending 27.0% CUDA, forward blending 25.4%, Adam 15.3%, forward projection 10.1% (0.576 ms/step). Projection disassembly shows multiple reciprocal/refinement sequences for quotients with the same depth denominator. -GS_SHARE_PROJECTION_RECIPROCAL=1in/tmp/r20/tree: form the reciprocal and squared reciprocal once in both EWA projection directions, reuse them for the Jacobian quotients, and leave the covariance/shading/extent rules unchanged. Float32 multiplication in place of repeated division is N1. Detached build/probe/time/quick/conditional N1 PID 39313 (/tmp/r20/{runner,build,compare_projection,time,quick,n1}.log). - Operator probes: 18 controlled cases, exact radii and last contributor counts, worst relative output/adjoint L2 1.55e-6. Time/tmp/r20/time/time.jsonat 23:16:27 UTC: 1.8905×, 162.645 s (45.352 / 60.894 / 56.399 s). Quick/tmp/r20/quick/check.json: PASS at 23:16:45 UTC; round ended. No resolved gain, no N1 run or promotion. Committedbcba6f4in the experiment worktree.
4.5 hRound 21 — tune staged backward warp grouping (started 2026-11-17 23:17:14 UTC)
- Detached profile PID 40436 (/tmp/r21/profile.log) measures the current best stack at steps 10001–10020 (/tmp/r21/profile). - Profile: forward blending 27.2% CUDA (1.513 ms/step), backward blending 26.3% (1.463 ms/step), Adam 15.0%, projection 9.8%. The staged compact backward kernel uses 64 registers and 25,632 bytes of shared storage per four-warp CTA; all work and synchronization are warp-local. Test smaller CTA grouping for load balance while leaving per-warp capacity comparable. -GS_TWO_WARP_STAGED_BACKWARD=1in/tmp/r21/tree: two independent staged bucket warps per block instead of four. No changes to the warp's pixel scan, arithmetic, memory order within a warp, or atomic destinations. End-to-end target N1 due to atomic ordering. Detached build/probe/time/quick/conditional N1 PID 40662 (/tmp/r21/{runner,build,compare_warps,time,quick,n1}.log). - Shared-forward-state probe: 18 cases, max relative gradient L2 8.53e-7. Time/tmp/r21/time/time.jsonat 23:23:54 UTC: 1.9076×, 161.186 s (44.416 / 60.308 / 56.462 s). Quick/tmp/r21/quick/check.json: PASS at 23:24:13 UTC; round ended. Comparable to round 16 within noise, no promotion or N1 run. Committedfbdce73in the experiment worktree.
4.7 hRound 22 — skip empty pixel buckets (started 2026-11-17 23:25:39 UTC)
- Detached profile PID 41701 (/tmp/r22/profile.log) measures the submitted round 16 stack at steps 15001–15020 (/tmp/r22/profile). - Profile: forward blending 26.9% CUDA (1.599 ms/step), backward blending 26.6% (1.583 ms/step), Adam 15.2%, projection 9.9%. The current backward compaction only removes pixels after their last contributor, not empty buckets between contributing ones. -GS_SKIP_EMPTY_PIXEL_BUCKETS=1in/tmp/r22/tree: retain a checkpoint when a pixel accepts its first splat in the bucket (before modifying transmittance or color). Mark interior empty buckets with zero transmittance, which cannot be a valid checkpoint under the existing termination threshold. Compact those pixels out of the staged backward scan as well. No extra buffer plane, depth/SH changes, or threshold changes; target N1. Detached build PID 42028 (/tmp/r22/build.log), probe/time/quick/conditional N1 PID 42323 (/tmp/r22/{runner,compare_empty,time,quick,n1}.log). - Operator probes: 18 cases, bitwise RGB, depth, radii, sorted pairs, last contributors, and all retained checkpoints. All 62,012 empty sentinels (of 941,012 active pixel buckets in the probe) match unchanged transmittance across the reference checkpoint pair. Worst relative L2 2.58e-7 with shared-forward-state gradients. - Time/tmp/r22/time/time.jsonat 23:42:14 UTC: 1.9523×, 157.500 s (44.799 / 57.780 / 54.921 s), 3.483 s below round 16 screening, with gains in all three windows. Quick/tmp/r22/quick/check.json: PASS at 23:42:33 UTC; round ended. Committed9acf0cc, N1 validation running. - Seed 1/tmp/r22/n1/check.jsonat 23:45:51 UTC: PASS (N1) with the full-trajectory exact items. Endpoint differences in σ: −2.18 PSNR, −1.61 SSIM, −0.80 LPIPS, +0.54 log count. Promoted9acf0ccand its switch after independent time, quick, and N1 passes; combined full check queued. - Combined full check/tmp/r22/check/check.jsonat 23:53:12 UTC: PASS (N1), 1.8992×, 161.904 s (48.886 / 57.975 / 55.043 s). Both exact gates pass. Endpoint differences in σ: −0.24 PSNR, −0.61 SSIM, −0.17 LPIPS, +0.15 log count. The later windows remain faster than round 16's comparable windows, but the first was 4.087 s slower than in round 22 screening, so the additional whole-run gain is not yet resolved in the combined check. Keep the validated round 22 stack pending follow-up evidence.
4.9 hRound 23 — group forward checkpoint loops (queued 2026-11-17 23:42:59 UTC)
- Detached PID 43232 (/tmp/r23/profile-check-runner.log) waits for round 22's N1 verdict, then profiles the resulting best stack at steps 7001–7020 (/tmp/r23/profile,/tmp/r23/profile.log). If round 22 passes, runs its combined full check (/tmp/r22/check,/tmp/r22/check.log). - Profile took the lock at 23:45:55 UTC on round 22. Forward blending 26.0% CUDA (1.273 ms/step), backward blending 25.6% (1.252 ms/step), Adam 15.5%, projection 9.9%. Checkpoint boundary detection still executes in every forward pair iteration. -GS_GROUP_FORWARD_BUCKETS=1in/tmp/r23/tree: iterate each fetched batch in 32-splat buckets, advancing the same checkpoint state once per bucket rather than testingj % 32per splat. Keep the same pair math, order, end-of-tile voting, and sparse checkpoint format. Build/probe/time/quick/conditional N1 PID 44010 (/tmp/r23/{runner,build,compare_grouped,time,quick,n1}.log); build waits for round 22 timing, probes wait for the full PASS. - Probe: 18 cases, bitwise RGB, depth, radii, sorted pairs, last contributors, sparse sentinels and every consumed checkpoint. Shared-state adjoints have worst relative L2 5.46e-7. Time/tmp/r23/time/time.jsonat 23:56:03 UTC: 2.0297×, 151.489 s (43.566 / 55.793 / 52.130 s), 6.011 s below round 22 screening, gains in all windows. Quick/tmp/r23/quick/check.json: PASS at 23:56:22 UTC; round ended. Committed58e777a; N1 validation running. - Seed 1/tmp/r23/n1/check.jsonat 23:59:32 UTC: PASS (N1) with full-trajectory exact items. Endpoint differences in σ: +0.55 PSNR, +1.40 SSIM, −2.57 LPIPS, +0.80 log count. Promoted58e777aand its switch after separate time, quick and N1 PASS; combined full check queued. - Combined full check/tmp/r23/check/check.jsonat 00:06:57 UTC: PASS (N1), 2.0710×, 148.470 s (41.440 / 55.012 / 52.018 s). Both exact gates passed. Endpoint differences in σ: −0.88 PSNR, −0.76 SSIM, +0.51 LPIPS, −0.22 log count. Keep round 23 submitted.
5.2 hRound 24 — conservative splat row cutoff (queued 2026-11-17 23:56:44 UTC)
- Detached PID 45845 (/tmp/r24/profile-check-runner.log) waits for round 23's N1 result, then profiles the resulting best stack at steps 11001–11020 (/tmp/r24/profile,/tmp/r24/profile.log). If round 23 passes, runs its combined full check (/tmp/r23/check,/tmp/r23/check.log). - Profile/tmp/r24/profile: backward blending 26.37% of CUDA (1.389 ms/step), forward blending 22.12% (1.165 ms/step), Adam 16.59%, forward projection 10.91%, radix sort 5.13%. -GS_CULL_SPLAT_ROWS=1in/tmp/r24/tree: calculate a conservative row-distance limit per splat from its positive-definite conic, with guards on the conic, dimensions, coordinates and opacity. Skip a pair in either blend direction only when the guarded limit proves its alpha is below the existing threshold; other pairs retain the original tests and arithmetic. Build, shared-state probes, judge time, quick and conditional N1 in detached PID 46604 (/tmp/r24/{runner,build,compare_rows,time,quick,n1}.log). The GPU probes and training wait for round 23's full PASS. The experiment is not submitted. - Operator probes: 18 cases, bitwise RGB, depth, radii, sorted pairs, last contributors, sparse sentinels and consumed checkpoints. Shared-state gradients differ by at most 1.98e-7 relative L2. Timing started at 00:07:05 UTC. - Time/tmp/r24/time/time.jsonat 00:09:48 UTC: 1.9861×, 154.815 s (41.813 / 60.640 / 52.362 s). Slower than both round 23 measurements, chiefly in the middle window. Quick exact/tmp/r24/quick/check.json: PASS at 00:10:06 UTC; round ended. No N1 run or promotion; committed29d4250in its experiment worktree.
5.4 hRound 25 — bounded alpha reciprocal (queued 2026-11-18 00:10:42 UTC)
- Detached PID 48295 (/tmp/r25/profile.log) profiles validated round 23 at steps 20001–20020 (/tmp/r25/profile). - Late profile: backward blending 25.66% of CUDA (1.314 ms/step), forward blending 20.69%, Adam 17.42%, forward projection 11.50%. Backward disassembly shows the reciprocal's normal-range refinement behind a per-pair range check and slow call. -GS_BOUNDED_ALPHA_RECIPROCAL=1in/tmp/r25/tree: the alpha cap and floor restrict1 - alphato positive normals near [0.01, 1]. Use the same reciprocal estimate and two rounded FMAs directly on this interval; keep the nonzero-background division unchanged. Detached PID 48598 runs an exhaustive denominator equality probe, build, shared-buffer backward probes, judge time, quick and conditional N1 (/tmp/r25/{runner,probe-build,probe-reciprocal,build,compare_backward,time,quick,n1}.log). - Standalone CUDA probe covers all 56,305,664 FP32 denominators in the interval, with zero bit mismatches against division. Disassembly confirms removal of the range branch and slow call for the RGB/depth reciprocal. Shared-state probes: 18 cases with black/RGB-only and nonblack/RGB+depth adjoints, partial tiles and varied SH degrees; worst relative L2 7.96e-7. Timing started at 00:16:01 UTC. - Time/tmp/r25/time/time.jsonat 00:18:37 UTC: 2.0921×, 146.975 s (41.262 / 54.299 / 51.415 s), 4.514 s below round 23 screening and 1.495 s below its combined check, with gains in all three windows. Quick exact/tmp/r25/quick/check.json: PASS at 00:19:03 UTC; round ended. - Seed 1/tmp/r25/n1/check.jsonat 00:22:10 UTC: PASS (N1) including full exact items. Endpoint differences in σ: −1.41 PSNR, −1.45 SSIM, −0.68 LPIPS, +0.93 log count. Committed and promotedb49c8f5plus its switch after independent time/quick/N1; combined check queued. - Combined full check/tmp/r25/check/check.jsonat 00:31:57 UTC: PASS (N1), 2.1202×, 145.024 s (40.144 / 53.704 / 51.177 s), 3.446 s faster than round 23's combined check and faster in each window. Both exact gates passed. Endpoint differences in σ: +0.19 PSNR, +0.32 SSIM, +0.82 LPIPS, +0.76 log count. Keep round 25 submitted.
5.6 hRound 26 — chunk backward checkpoint staging (queued 2026-11-18 00:23:13 UTC)
- Detached PID 50011 (/tmp/r26/profile-check-runner.log) profiles the round 25 stack at steps 5501–5520 (/tmp/r26/profile,/tmp/r26/profile.log), then runs round 25's combined full check (/tmp/r25/check,/tmp/r25/check.log). - Early profile: backward blending 25.05% of CUDA (1.058 ms/step), forward blending 21.70%, dense Adam 17.28%, forward projection 10.99%. The staged wavefront tests the next 32-pixel chunk boundary on every loop iteration. -GS_GROUP_BACKWARD_PIXELS=1in/tmp/r26/tree: stage checkpoint/residual batches in an outer loop and run the existing pair processing over each chunk. Preserve pixel ordering, lane handoff, tail steps, warp synchronization and gradients. Build/probe/time/quick/conditional N1 detached PID 50631 (/tmp/r26/{runner,build,compare_backward,time,quick,n1}.log); build waits for round 25 timing to finish and probes wait for its full PASS. - Shared-state probes: 18 cases, max relative L2 1.58e-7. Time/tmp/r26/time/time.jsonat 00:34:42 UTC: 2.1006×, 146.376 s (41.016 / 54.603 / 50.757 s), 0.599 s faster than round 25 screening but 1.352 s slower than its combined check, with mixed window changes. Quick exact/tmp/r26/quick/check.json: PASS at 00:35:01 UTC; round ended. No resolved gain, N1 run or promotion. Committed636f544in the experiment worktree.
5.8 hRound 27 — keep projection covariance local (queued 2026-11-18 00:35:36 UTC)
- Detached PID 52456 (/tmp/r27/profile.log) profiles the validated round 25 submission at steps 9001–9020 (/tmp/r27/profile). - Profile: forward blending 23.12% of CUDA, backward blending 22.78%, Adam 17.50%, forward projection 11.96% (0.597 ms/step). Disassembly of projection shows immediate global reloads of covariance coefficients stored by the same thread, plus stores before rejecting offscreen rectangles. -GS_LOCAL_PROJECTION_COVARIANCE=1in/tmp/r27/tree: compute the same covariance in local scalars, use it directly for projection, and store it for backward only after the existing visibility tests. Precomputed covariance and the default kernel are unchanged. Build, bitwise forward/covariance probes, time, quick and conditional N1 detached PID 52705 (/tmp/r27/{runner,build,compare_covariance,time,quick,n1}.log). - Probes: 20 forward cases, bitwise RGB/depth/radii and visible covariances, including explicitly offscreen rows; 18 backward cases, max relative L2 6.32e-7. Time/tmp/r27/time/time.jsonat 00:42:31 UTC: 2.0981×, 146.555 s (41.485 / 53.746 / 51.325 s), 1.531 s slower than round 25's combined check. Quick/tmp/r27/quick/check.json: PASS at 00:42:48 UTC; round ended. No resolved gain, N1 run or promotion. Committed59f6f46in the experiment worktree.
5.9 hRound 28 — render bookkeeping profile (queued 2026-11-18 00:43:53 UTC)
- Detached profile PID 53662 (/tmp/r28/profile.log) measures the validated round 25 submission at steps 4501–4520 (/tmp/r28/profile), lock held at 00:43:53 UTC. - Profile: backward blending 27.99% CUDA, forward blending 17.75%, Adam 17.29%, forward projection 10.11%; fill kernels 3.54% across the step. Trace and source identify a dense zero initialization and add-by-zero for the screen-space placeholder, which is not used in the rasterizer forward. Its adjoint is only consumed by densification. -GS_COMPACT_VIEWSPACE_POINTS=1in/tmp/r28/tree: expose the same FP32 zero values and shape as an expanded scalar leaf; keep its dense adjoint through densification and stop storing it afterward. The rasterizer still computes all internal projection derivatives. Default/external render calls are unchanged. CPU placeholder probe passed; detached time/quick/conditional N1 PID 53922 (/tmp/r28/{runner,time,quick,n1}.log). - Time/tmp/r28/time/time.jsonat 00:50:00 UTC: 2.1002×, 146.408 s (41.791 / 53.527 / 51.090 s), 1.384 s slower than round 25's combined check with mixed window changes. Quick exact/tmp/r28/quick/check.json: PASS at 00:50:18 UTC; round ended. No resolved gain, N1 run or promotion. Committed57b310cin the experiment worktree.
6.1 hRound 29 — binning profile (queued 2026-11-18 00:51:29 UTC)
- Detached profile PID 54801 (/tmp/r29/profile.log) measures the validated round 25 stack at steps 10501–10520 (/tmp/r29/profile). - Profile: backward blending 25.65% of CUDA, forward blending 22.13%, dense Adam 16.67%, projection 11.34%, radix scatter 5.17% across six passes per iteration. -GS_COMPACT_TILE_DEPTH_KEYS=1in/tmp/r29/tree: subtract the existing positive near-depth encoding from each float bit pattern and use 28 depth bits only if every projected Gaussian fits. The overflow flag is combined with the existing pair-count readback. Stable sorting and the full key fallback preserve raw depth order and equal-depth ties; compact keys remove one pass for 12-bit tile grids. Build, sorted-output and overflow probes, time, quick and conditional N1 detached PID 55092 (/tmp/r29/{runner,build,compare_sort,time,quick,n1}.log). - First build stopped before GPU tests:adam.cuincludedauxiliary.hwithout a fixed-width integer declaration. Added the direct<cstdint>include and restarted the same probe/time/quick/N1 sequence in detached PID 55490 (/tmp/r29/{runner-fixed,build-fixed}.log). - Probes: 18 compact cases plus two overflow fallbacks, covering both scalar and cooperative duplication. Ordered Gaussian IDs and rendered RGB/depth/radii matched bitwise; compact keys matched their lossless encoding and fallback keys were unchanged. Time/tmp/r29/time/time.jsonat 01:00:50 UTC: 2.1093×, 145.775 s (41.133 / 53.892 / 50.751 s), 0.751 s slower than round 25's combined check with mixed window changes. Quick exact/tmp/r29/quick/check.json: PASS at 01:01:08 UTC; round ended. No resolved gain, N1 run or promotion. Committed69e032ain the experiment worktree.
6.2 hRound 30 — densification-step profile (queued 2026-11-18 01:01:40 UTC)
- Detached profile PID 56471 (/tmp/r30/profile.log) measures steps 10899–10902 on the validated round 25 stack, including the scheduled densification at 10900 (/tmp/r30/profile). - Densification-window profile: indexing 33.17% CUDA (11.774 ms inclusive, 60 calls), backward blending 16.28%, forward blending 10.77%, Adam 7.96%, and concatenations 4.81%. Seven cache-release calls use 1.26 ms inclusive CPU; the larger opportunity is avoiding intermediate full-model copies. -GS_BATCH_DENSIFICATION_UPDATES=1in/tmp/r30/tree: for positive split thresholds, generate clones and split children in their existing order before one append. The appended clones have zero padded split gradients and cannot be selected; fold original split-parent removal into the unchanged final prune. Preserve random draws, zero-initialized child Adam state, parameter replacement and stats reset. Nonpositive thresholds retain the default path. Detached PID 56716 builds, compares full model/optimizer/RNG state, times, checks quick and conditionally checks N1 (/tmp/r30/{runner,build,compare_densification,time,quick,n1}.log). - Operator probe: five cases with clone/split/prune mixes, no children, screen pruning and the nonpositive-threshold fallback. Parameters, Adam tensors, densification stats and CUDA RNG state matched. Time/tmp/r30/time/time.jsonat 01:09:29 UTC: 2.1010×, 146.351 s (41.396 / 53.510 / 51.445 s), 1.327 s slower than round 25's combined check with mixed windows. Quick exact/tmp/r30/quick/check.json: PASS at 01:09:47 UTC; round ended. No resolved gain, N1 run or promotion. Committed66d40e9in the experiment worktree.
6.4 hRound 31 — backward writeback profile (queued 2026-11-18 01:10:46 UTC)
- Detached profile PID 57695 (/tmp/r31/profile.log) measures validated round 25 at steps 18501–18520 (/tmp/r31/profile). - Late profile: backward blending 24.54% CUDA (1.240 ms/step), forward blending 21.25%, dense Adam 17.78%, projection 11.28%. Each surviving splat lane currently writes the three conic derivatives and one opacity derivative as separate float32 atomic additions. -GS_PACK_CONIC_OPACITY_ADJOINTS=1in/tmp/r31/tree: reuse the unused fourth conic-gradient slot for opacity and issue a vector float4 atomic on Hopper; reconstruct the original opacity-gradient tensor in the already-required projection adjoint, including antialiasing. Other architectures use scalar atomics; reductions remain float32. Build, same-buffer probes, time, quick and conditional N1 detached PID 57961 (/tmp/r31/{runner,build,compare_writeback,time,quick,n1}.log). - Shared-state probes: 18 cases, maximum relative L2 5.12e-7. Time/tmp/r31/time/time.jsonat 01:18:23 UTC: 2.0670×, 148.760 s (43.178 / 54.275 / 51.307 s), slower than both round 25 measurements, with the largest regression in the first window. Quick exact/tmp/r31/quick/check.json: PASS at 01:18:41 UTC; round ended. No N1 run or promotion; committed1b3cef7in the experiment worktree.
6.6 hRound 32 — photometric loss profile (queued 2026-11-18 01:20:34 UTC)
- Detached profile PID 58937 (/tmp/r32/profile.log) measures round 25 at steps 6501–6520 (/tmp/r32/profile). - Profile: backward blending 28.2% of CUDA self time (1.211 ms/step), forward blending 18.7% (0.804 ms), Adam 16.9%, backward SH/covariance adjoint 10.7%. The SSIM forward and backward kernels take 79.8 and 54.2 µs/step. Attribution to the loss found 10.6 µs/step of forward fills, 6.9 µs/step of its image copies, 12.1 µs/step of mean-adjoint materializations, and separate L1 pointwise/gradient kernels. -GS_FUSED_PHOTO_TERMS=1in/tmp/r32/tree: pack ground truth once, write absolute residuals directly from the SSIM tile loads, and retain the original PyTorch means and scalar weights. The joint adjoint applies each scalar mean gradient before the unchanged SSIM convolution and adds the original signed L1 gradient per pixel. Fully overwritten output maps use uninitialized allocations. Detached build PID 59312 (/tmp/r32/build.log), operator probe/time/quick/conditional N1 runner PID 59627 (/tmp/r32/{runner,compare_photo,time,quick,n1}.log). - Extension build passed. The initial runner stopped before taking the GPU lock becausebuildprints JSON; corrected the path parser and restarted the detached runner as PID 59805 (/tmp/r32/runner-resume.log). - Operator probe: both FP32 scalar terms and all image gradients matched bitwise across 21 shape/pattern cases, including 979×546, interleaved targets, partial tiles, equal images and four loss weights. Time/tmp/r32/time/time.jsonat 01:34:12 UTC: 2.1020×, 146.282 s (40.919 / 54.606 / 50.757 s), 0.693 s below round 25 screening but 1.258 s above its full check. Quick exact/tmp/r32/quick/check.json: PASS at 01:34:30 UTC; round ended. No resolved gain, N1 run or promotion. Committed3b28d2fin the experiment worktree.
6.8 hRound 33 — late-stage projection and gradient profile (queued 2026-11-18 01:35:13 UTC)
- Detached profile PID 60786 (/tmp/r33/profile.log) measures the validated round 25 stack at steps 17001–17020 (/tmp/r33/profile). - Profile: backward blending 26.0% CUDA (1.317 ms/step), forward blending 19.8% (1.000 ms), dense Adam 17.8% (0.898 ms, of which SH-rest alone takes 0.648 ms), backward SH/covariance adjoint 11.2%. The[P,15,3]SH-rest adjoint is the largest backward fill, 79.4 µs/step, before the shading adjoint rewrites and Adam reads it. The forward and backwardpreprocessCUDAsymbols are distinguished by their first data argument (float const*andfloat3 const*, respectively). -GS_FACTORED_SH_ADAM=1in/tmp/r33/tree: return the original float32 SH basis and clamped RGB adjoint without their outer product; expand it inside dense Adam, preserving FP32 multiplication, both moment recurrences and the update for every coefficient, including zero-gradient rows. Compare the originating Parameter by identity so topology replacement discards this adjoint exactly as before. Dense gradients remain the public/default rasterizer path. Initial unswitched build PID 61241 stopped in favor of the judge's configured build. Detached configured build/probe/time/quick/conditional N1 PID 61664 (/tmp/r33/{build-config,compare_sh_adjoint,time,quick,n1,runner}.log). - Probes: bitwise rasterizer adjoints in 12 same-buffer cases (all four SH degrees, partial tiles, RGB/depth, visible/cull mix) and bitwise parameters and Adam state over 24 steps at each of five point counts. The probe observed 6,126,644 actual parameter changes on zero-input-gradient steps, confirming that existing moments still update them. - Time/tmp/r33/time/time.jsonat 01:50:39 UTC: 2.2578×, 136.185 s (40.469 / 49.081 / 46.635 s), 8.839 s below the round 25 full check. Quick/tmp/r33/quick/check.jsonPASS at 01:50:57 UTC. Seed 1/tmp/r33/n1/check.jsonPASS (N1) at 01:53:53 UTC, including full-trajectory exact items. Committed and promoted1f00f5dwith the new switch at 01:55 UTC. Combined full check queued separately. - Combined check/tmp/r33/check/check.jsonat 02:05:09 UTC: 2.2501× PASS (N1), 136.653 s (39.694 / 50.170 / 46.789 s). Both exact gates pass; endpoint biases are +0.76σ PSNR, +1.82σ SSIM, −2.51σ LPIPS and −0.12σ log count. Standalone seed 1's four biases were within 0.84σ.
7.2 hRound 34 — profile the factored SH stack (queued 2026-11-18 01:55:40 UTC)
- Detached PID 63257 (/tmp/r34/profile-check-runner.log) profiles the submitted round 33 stack at steps 16001–16020 (/tmp/r34/profile,/tmp/r34/profile.log), then runs the round 33 combined full check (/tmp/r33/check,/tmp/r33/check.log). - Profile took the lock at 01:57:56 UTC. Backward blending 28.3% of CUDA self time (1.321 ms/step), forward blending 22.8% (1.067 ms/step), Adam 18.1%, factored SH/covariance adjoint 5.43%, forward projection 3.66%. The staged forward kernel uses 42 registers/thread and 9,260 bytes of shared memory, and currently waits to gather each batch before blending. -GS_PREFETCH_FORWARD_SPLATS=1in/tmp/r34/tree: double-buffer the existing batch slots and asynchronously gather the next RGB/conic/coordinate batch during blending. Preserves the original CTA geometry, contributor reduction, FP32 math and depth fallback. Initial detached build PID 64655 (/tmp/r34/build.log) reported a CUDA/C++ captured-type template error; fixed the template dispatch. Configured build/probe/time/quick/conditional N1 retry PID 65218 (/tmp/r34/{build-retry,runner-retry,export,compare,time,quick,n1}.log); initial runner PID 65149 exited without GPU work. - Nine probes: bitwise RGB/depth/radii, contributor counts and consumed checkpoints, including 5–7 fetched batches, early termination, partial or empty tiles, and depth fallback. Shared-state gradient relative L2 ≤2.56e-9. Time/tmp/r34/time/time.jsonat 02:11:06 UTC: 2.2748×, 135.171 s (38.969 / 49.521 / 46.681 s). Quick exact/tmp/r34/quick/check.json: PASS at 02:11:24 UTC; round ended. Only 1.014 s faster than round 33 screening, with mixed later windows and no resolved gain. No N1 or promotion; committedb978cb2in the experiment worktree.
7.5 hRound 35 — backward blending profile (queued 2026-11-18 02:13:39 UTC)
- Detached profile PID 66368 (/tmp/r35/profile.log) measures the round 33 submission at steps 8501–8520 (/tmp/r35/profile). - Profile: backward blending 27.5% CUDA self time (1.200 ms/step), forward blending 23.6% (1.030 ms), Adam 18.3%. The backward inner loop applies a splat's constant opacity, conic coefficients and NDC scales to every accepted pixel's screen-space geometry adjoint. -GS_FACTOR_SPLAT_GEOMETRY_ADJOINT=1in/tmp/r35/tree: accumulate the weighted first/second offset moments in the existing FP32 per-splat registers, and apply their constant chain rule once before the original atomic adds. RGB, opacity, ordering and transmittance remain unchanged. Keep the original gradient path for depth, nonblack background, nonfinite staged pixel gradients, or unsupported splat values. Numerics target N1. Detached build/probe/time/quick/conditional N1 PID 66686 (/tmp/r35/{runner,build,compare,time,quick,n1}.log); reuses round 34's saved same-forward-state probe cases. - Nine same-state cases: bitwise forward items; gradient relative L2 ≤2.39e-7. Time/tmp/r35/time/time.jsonat 02:21:46 UTC: 2.2740×, 135.218 s (39.538 / 49.229 / 46.451 s). Quick exact/tmp/r35/quick/check.json: PASS at 02:22:04 UTC; round ended. Gain against round 33 screening is 0.967 s with mixed window differences, below resolution. No N1 run or promotion; committedf815d73in the experiment worktree.
7.6 hRound 36 — late backward writeback profile (queued 2026-11-18 02:23:15 UTC)
- Detached profile PID 67762 (/tmp/r36/profile.log) measures the round 33 submission at steps 20501–20520 (/tmp/r36/profile). - Profile: backward blending 27.4% of CUDA self time (1.235 ms/step), forward blending 22.5% (1.012 ms), Adam 18.6%, radix sort passes 6.1%. The backward blend epilogue currently issues 9 FP32 atomics for every valid RGB splat, even for a bucket lane with no accepted pixels or a splat known to have exactly one tile pair. -GS_ELIDE_REDUNDANT_SPLAT_ATOMICS=1in/tmp/r36/tree: leave already-zero buffers alone for lanes with no accepted pixels; substitute a store with the same FP32 flush-to-zero addition to positive zero only when the existing bounding tile count guarantees exclusive ownership. Keep atomics for all other writers and leave dense Adam unchanged. Detached atom/store bit probe, configured build, same-buffer probe, time, quick and conditional N1 PID 68128 (/tmp/r36/{runner,compile-probe,compare_atomic,build,compare,time,quick,n1}.log). - The atom/store probe matched all 3,072 selected FP32 bit patterns, including both signed zeros, subnormals, infinities and NaNs. Nine same-state cases were forward-bitwise; gradient relative L2 ≤2.56e-9. Time/tmp/r36/time/time.jsonat 02:32:15 UTC: 2.2289×, 137.954 s (41.873 / 49.222 / 46.859 s), slower than the submitted round 33 stack. Quick/tmp/r36/quick/check.jsonPASS at 02:32:34 UTC; round ended. No N1 run or promotion; committedffcfcd4in the experiment worktree.
7.8 hRound 37 — profile staged rasterizer blending (queued 2026-11-18 02:33:57 UTC)
- Detached profile PID 69274 (/tmp/r37/profile.log) measures the round 33 submission at steps 11001–11020 (/tmp/r37/profile). - Profile: backward blending 27.9% CUDA self time (1.310 ms/step), forward blending 24.2% (1.137 ms/step), dense Adam 17.5%, radix sort passes 5.8%. Re-screen the staged blending choices together: the round 34 prefetch and round 35 factored geometry adjoint modify separate kernels and each had a small unresolved gain alone. -/tmp/r37/treecherry-picksb978cb2andf815d73separately as91ea443and7822450; both independent switches are enabled in/tmp/r37/switches.json, no other implementation change. Detached configured build/same-state probe/time/quick/conditional N1 PID 69568 (/tmp/r37/{runner,build,compare,time,quick,n1}.log). - Nine same-state cases passed (exact forward, gradient relative L2 ≤2.39e-7). Time/tmp/r37/time/time.jsonat 02:40:22 UTC: 2.2288×, 137.958 s (41.485 / 49.575 / 46.898 s). Quick exact/tmp/r37/quick/check.json: PASS at 02:40:40 UTC; round ended. The combination regressed versus round 33; no N1 or promotion.
7.9 hRound 38 — early backward workspace profile (queued 2026-11-18 02:41:40 UTC)
- Detached profile PID 70651 (/tmp/r38/profile.log) measures the round 33 submission at steps 901–920 (/tmp/r38/profile). - Profile: backward blending 39.6% CUDA (0.656 ms/step), forward blending 20.8% (0.345 ms), dense Adam incl. initialization 11.7%, sort passes 6.1%. CUDA launch ancestry attributes nine fills per step (16.0 µs/step) directly to the rasterizer's backward output/scratch preparation. -GS_FUSE_BACKWARD_BUFFER_CLEAR=1in/tmp/r38/tree: allocate the same separate tensors, then initialize all required FP32 gradient regions with a single vectorized clear launch, skipping absent and fully written factored SH buffers. No view aliasing, changed gradient values, or optimizer change. Detached configured build/shared-state probes/time/quick/conditional N1 PID 70899 (/tmp/r38/{runner,build,compare,compare_sh,time,quick,n1}.log). The submission remains the fully checked round 33 tree. - Same-state blending and SH adjoint probes passed. Time/tmp/r38/time/time.jsonat 02:48:38 UTC: 2.2404×, 137.244 s; quick exact/tmp/r38/quick/check.json: PASS at 02:48:56 UTC; round ended. Experiment committed as16cd01d(N1 intended); no resolved benefit, N1 run, or promotion.
8.2 hRound 39 — compact tile binning (GPU lock 2026-11-18 02:51:58 UTC)
- Profile/tmp/r39/profile/timing_profile.profile.txt(steps 12501–12520) of round 33: backward blending 29.85% of CUDA self time, forward blending 22.73%, dense Adam 17.19%, radix sort 5.90%, tile duplication 2.94%. -GS_COMPACT_SMALL_TILE_RECTS=1in/tmp/r39/treecaches the original overlap decisions for rectangles of at most 16 tiles during projection. Use actual accepted counts in the same scan and emit only accepted pairs; wide rectangles keep the original duplication. Guard the all-sentinel tile range start. Configured build PID 72571 (/tmp/r39/build.log). - Detached scalar/cooperative sorted-pair and forward-state probes, time, quick, conditional N1: PID 73011 (/tmp/r39/{runner,export-0,export-1,compare-0,compare-1,time,quick,n1}.log). - Eight operator cases per duplication path: identical ordered valid keys/IDs, RGB/depth, radii, contributor counts and consumed checkpoints, including mixed rectangles, zero compacted pairs and an all-sentinel wide list. Time/tmp/r39/time/time.jsonat 03:04:46 UTC: 2.2740×, 135.215 s (39.528 / 49.738 / 45.949 s). Quick exact/tmp/r39/quick/check.json: PASS at 03:05:05 UTC; round ended. The 0.970 s screening improvement has mixed windows and is unresolved. No N1 or promotion; experiment committed as1001ff9.
8.3 hRound 40 — late-training profile (queued 2026-11-18 03:05:48 UTC)
- Detached profile PID 73790 (/tmp/r40/profile.log) measures the submitted round 33 stack at steps 24001–24020 (/tmp/r40/profile). - GPU lock at 03:05:49 UTC. Profile: backward blending 27.77% of CUDA (1.211 ms/step), forward blending 20.30% (0.886 ms/step), dense Adam 19.12% (0.834 ms/step), radix sort 6.42%. The factored Adam specialization uses explicit FP32 square root and division; disassembly confirmed per-value square-root slow branches, including for zero second moments. -GS_ADAM_REFINED_SQRT=1in/tmp/r40/treereplaces only the root's control flow: exponent-scale small normal moments, use reciprocal-square-root plus rounded FMA refinement, and select the existing signed-zero/infinity cases without a branch. Division, moments, parameter layout and dense steps are unchanged. Detached exhaustive square-root and optimizer-state probes, judge time, quick and conditional N1 PID 74213 (/tmp/r40/{runner,build,compare,time,quick,n1}.log). - Exhaustive SM90 probe matchedsqrt_rnon all 4,294,967,296 FP32 encodings (NaN classifications equal; payloads not compared). Four dense and factored cases, 28 steps each, matched parameters and Adam state bitwise. Time/tmp/r40/time/time.jsonat 03:15:24 UTC: 2.2688×, 135.528 s (39.319 / 49.790 / 46.418 s). Quick exact/tmp/r40/quick/check.json: PASS at 03:15:43 UTC; round ended. A 0.657 s gain with mixed windows is unresolved. No N1 or promotion; experiment committed asbca4e9b.
8.5 hRound 41 — early backward-pipeline profile (queued 2026-11-18 03:16:29 UTC)
- Detached profile PID 75356 (/tmp/r41/profile.log) measures the submitted round 33 stack at steps 1501–1520 (/tmp/r41/profile). - GPU lock at 03:16:30 UTC. Backward blending 32.86% of CUDA (0.752 ms/step), forward blending 19.35% (0.443 ms/step), dense Adam 9.75%, densification indexing 6.00% self, radix sort 5.45%. The compact backward loop drains all 31 possible stages even for partial buckets with fewer live contributors. -GS_TRIM_BACKWARD_WAVEFRONT=1in/tmp/r41/treecomputes the last possible phase by reducing the final 32 compacted pixels' recorded contributor counts. Earlier pixels cannot determine a later phase. Existing contribution guards, FP32 arithmetic, per-lane operation order and atomic writes are unchanged. Detached build, saved-forward-state and SH adjoint probes, judge time, quick and conditional N1 PID 75611 (/tmp/r41/{runner,build,compare,compare-sh,time,quick,n1}.log). - Nine same-state cases passed (exact forward, gradient relative L2 ≤3.66e-9); the factored SH consistency probes also passed. Time/tmp/r41/time/time.jsonat 03:23:17 UTC: 2.2434×, 137.058 s (40.293 / 50.099 / 46.666 s), slower than round 33. Quick exact/tmp/r41/quick/check.json: PASS at 03:23:35 UTC; round ended. No N1 or promotion; experiment committed ased59596.
8.6 hRound 42 — binning profile (queued 2026-11-18 03:23:55 UTC)
- Detached profile PID 76769 (/tmp/r42/profile.log) measures the submitted round 33 stack at steps 10501–10520 (/tmp/r42/profile). - GPU lock at 03:23:56 UTC. Profile: backward blending 24.88% of CUDA (1.306 ms/step), forward blending 20.96% (1.100 ms/step), dense Adam 14.49%, scheduled densification indexing 10.08% self, radix scatter 5.04% (six passes, 0.265 ms/step). -GS_SEGMENT_TILE_SORT=1in/tmp/r42/treereserves per-tile slots with warp-aggregated histograms, scans offsets, then uses CUB segmented key sorting. Keys include the original depth bits and Gaussian ID, so equal-depth order is the original stable duplication order. Invalid pairs stay out of all segments; range unpack also produces bucket counts. Configured build PID 77079 (/tmp/r42/build.log); detached scalar/cooperative probes and judge time/quick/conditional N1 PID 77393 (/tmp/r42/{runner,export-0,export-1,compare-0,compare-1,time,quick,n1}.log). - Twenty-four probes reproduced ordered valid keys/IDs, ranges, RGB/depth/radii, contributor counts and checkpoints bitwise, including depth ties, one-pair, all-sentinel, long and production-size segments. Time/tmp/r42/time/time.jsonat 03:31:20 UTC: 2.2040×, 139.508 s (41.418 / 50.813 / 47.277 s), slower in all windows. Quick exact/tmp/r42/quick/check.json: PASS at 03:31:38 UTC; round ended. No N1 or promotion; experiment committed asb178beb.
8.8 hRound 43 — early forward blending profile (queued 2026-11-18 03:32:38 UTC)
- Detached profile PID 78218 (/tmp/r43/profile.log) measures the submitted round 33 stack at steps 4201–4220 (/tmp/r43/profile). - GPU lock at 03:32:39 UTC. Backward blending 26.89% of CUDA (1.002 ms/step), forward blending 17.78% (0.663 ms/step), dense Adam 14.14%, densification indexing 9.56% self, radix scatter 4.99%. The staged forward path loads cached RGB from three separate shared-memory arrays after the alpha test. -GS_PACK_FORWARD_COLORS=1in/tmp/r43/treestages RGB and optional inverse depth into one aligned vector per splat, allowing a single 128-bit shared load with independent FP32 blend expressions unchanged. Configured build PID 78510 (/tmp/r43/build.log). - Detached same-state and SH adjoint probes, judge time, quick and conditional N1 PID 78827 (/tmp/r43/{runner,compare,compare-sh,time,quick,n1}.log). - SASS confirms the fusedLDS.128. Nine same-state cases preserve the forward bits and have gradient relative L2 ≤2.56e-9; factored SH adjoints pass their independent probes. Time/tmp/r43/time/time.jsonat 03:39:08 UTC: 2.2997×, 133.707 s (38.239 / 48.939 / 46.530 s), 2.478 s faster than round 33 screening and improved in each window. Quick exact/tmp/r43/quick/check.json: PASS at 03:39:26 UTC; round ended. - Seed 1/tmp/r43/n1/check.jsonat 03:42:28 UTC: PASS (N1) including full exact items and endpoint differences −0.61σ PSNR, +0.54σ SSIM, −1.19σ LPIPS, +0.12σ log count. Committed and promoted7a64c89with its switch after time, quick and N1 passed. Detached combined check PID 80103 (/tmp/r43/check.log,/tmp/r43/check) runs the identical tree and configuration from its worktree. - Combined full check/tmp/r43/check/check.jsonat 03:48:23 UTC: 2.2702× PASS (N1), 135.440 s (39.296 / 49.572 / 46.571 s). Both exact gates pass, and endpoint differences are −0.11σ PSNR, −0.20σ SSIM, −1.12σ LPIPS, +1.03σ log count. Keep7a64c89submitted.
9.0 hRound 44 — packed forward stack profile (queued 2026-11-18 03:44:49 UTC)
- Detached PID 80349 waits for round 43's combined PASS and then profiles the submitted stack at steps 15501–15520 (/tmp/r44/profile,/tmp/r44/profile.log). - GPU lock at 03:50:23 UTC. Profile: backward blending 27.73% of CUDA (1.343 ms/step), forward blending 24.66% (1.194 ms/step), dense Adam 17.49%, radix scatter 5.92%. Forward and backward checkpoint code writes/reads three distinct RGB planes for every occupied pixel bucket. -GS_PACK_CHECKPOINT_RG=1in/tmp/r44/treeinterleaves checkpoint red/green within their existing two planes for vector loads and stores. Blue, transmittance and optional depth retain their layouts and byte counts. Each FP32 value is unmodified, including zero transmittance markers. Configured build PID 82031 (/tmp/r44/build-config.log) supersedes build PID 81821 to include the header's fixed-width types directly. - Detached staged and unstaged forward/backward probes, judge time, quick and conditional N1 PID 82365 (/tmp/r44/{runner,compare-staged,export-legacy,compare-legacy,time,quick,n1}.log). - Eighteen cases preserve render, contributor and checkpoint values bitwise and gradient relative L2 ≤2.80e-8. Time/tmp/r44/time/time.jsonat 03:58:32 UTC: 2.2818×, 134.754 s (38.897 / 49.327 / 46.530 s), between round 43's screening and combined times. Quick exact/tmp/r44/quick/check.json: PASS at 03:58:51 UTC; round ended. Unresolved gain; no N1 or promotion. Experiment committed ase0a4015.
9.2 hRound 45 — late dense Adam profile (queued 2026-11-18 03:59:10 UTC)
- Detached profile PID 83219 (/tmp/r45/profile.log) measures the submitted round 43 stack at steps 20001–20020 (/tmp/r45/profile). - GPU lock at 03:59:11 UTC. Profile: backward blending 27.35% of CUDA (1.254 ms/step), forward blending 23.04% (1.057 ms/step), dense Adam 18.25% (0.837 ms/step), radix scatter 6.04%. Factored SH-rest Adam alone takes 0.592 ms per step, launching 45,052 CTAs with 64 registers/thread at the existing 2048 elements / 8 warps. -/tmp/r45/treeaddsGS_SH_ADAM_LAUNCHto try1024x8(less work per lane) or2048x4(more work per lane) only on the factored SH group. The default and other groups remain unchanged. The exact FP32 calculations and dense update are identical. The local judge has nosweepsubcommand; screen both with separatetimecalls. - Detached configured builds, bitwise state probes, both timing/quick checks and conditional N1 PID 83566 (/tmp/r45/{runner,build-*,compare,time-*,quick-*,n1-*}.log). - Both layouts matched parameters and FP32 optimizer state bitwise for 28 steps at each of five point counts.1024x8:/tmp/r45/time-1024x8/time.json, 2.3004×, 133.666 s (38.561 / 48.911 / 46.195 s), at 04:08:46 UTC; quick PASS.2048x4:/tmp/r45/time-2048x4/time.json, 2.1924×, 140.247 s (39.630 / 51.671 / 48.946 s), at 04:11:35 UTC; quick PASS. Round ended 04:11:59 UTC. The former ties round 43 screening, the latter regresses. No N1 or promotion; experiment committed asc2a6704.
9.4 hRound 46 — forward gather profile (queued 2026-11-18 04:12:17 UTC)
- Detached profile PID 85833 (/tmp/r46/profile.log) measures the submitted round 43 stack at steps 7501–7520 (/tmp/r46/profile). - GPU lock at 04:12:18 UTC. Backward blending 24.81% of CUDA (1.192 ms/step), forward blending 19.87% (0.955 ms/step), dense Adam 14.80%, densification indexing 10.20% self, radix scatter 4.91%. The packed-shared forward path still gathers coordinates, conics and colors from separate global-memory arrays for each tile batch. -GS_PACK_FORWARD_SPLATS=1in/tmp/r46/treewrites a packed forward descriptor during projection and fetches its aligned vector fields during staging. The original arrays remain available to binning and backward, and the renderer still uses each original FP32 value. The mode requires staged packed RGB. Configured build PID 86102 (/tmp/r46/build.log). - Detached SH/precomputed color, partial-tile and RGB/depth probes, judge time/quick and conditional N1 PID 86404 (/tmp/r46/{runner,compare,export-explicit,compare-explicit,time,quick,n1}.log). - SH cases pass. First explicit-color comparison stopped before judge runs: the raw probe supplied both SH and precomputed RGB, an input combination rejected by the public rasterizer that leaves the SH clamp mask uninitialized. Use RGB alone in the corrected probe; no candidate or judge change. - Corrected probe and judge sequence is detached PID 86744 (/tmp/r46/{runner-valid,export-explicit-valid,compare-explicit-valid,time,quick,n1}.log). - Nine SH and nine valid precomputed-RGB cases pass with exact forward state and gradient relative L2 ≤1.29e-8. Time/tmp/r46/time/time.jsonat 04:20:35 UTC: 2.2716×, 135.361 s (38.542 / 49.960 / 46.859 s), slower than round 43 screening. Quick exact/tmp/r46/quick/check.json: PASS at 04:20:52 UTC; round ended. No resolved benefit, N1 or promotion. Experiment committed as7912995.
9.6 hRound 47 — forward CTA scheduling profile (queued 2026-11-18 04:21:13 UTC)
- Detached profile PID 87403 (/tmp/r47/profile.log) measures the submitted round 43 stack at steps 16501–16520 (/tmp/r47/profile). - GPU lock at 04:21:14 UTC. Profile: backward blending 28.39% of CUDA (1.311 ms/step), forward blending 22.43% (1.036 ms/step), dense Adam 18.35%, radix scatter 5.96%. Each forward CTA follows a tile's distinct depth-list length; row-major launch order does not prioritize longer work. -GS_ORDER_FORWARD_TILES=1in/tmp/r47/treebuilds a coarse descending priority list by logarithmic bucket-count class in the bucket-count kernel, then maps render CTAs through that permutation. Each logical tile still has the same threads, depth list, arithmetic, reduction and checkpoint positions. Grids over 4096 tiles use original scheduling. Configured build PID 87691 (/tmp/r47/build.log). - Detached ordered-pair, rendering and gradient probes, judge time/quick and conditional N1 PID 87952 (/tmp/r47/{runner,export,compare,compare-grad,time,quick,n1}.log). - Eleven forward cases preserve ordered pairs, ranges, RGB/depth/radii, contributor counts and checkpoints bitwise, including production-size images and >4096-tile fallback. Nine shared-buffer adjoint cases have relative L2 ≤1.29e-8. Time/tmp/r47/time/time.jsonat 04:28:11 UTC: 2.3520×, 130.731 s (38.268 / 47.349 / 45.114 s), 2.976 s faster than round 43's screening and faster in all windows versus its combined check. Quick exact/tmp/r47/quick/check.json: PASS at 04:28:30 UTC; round ended. - Seed 1/tmp/r47/n1/check.jsonat 04:31:21 UTC: PASS (N1) including full exact items; endpoint differences −2.16σ PSNR, −1.18σ SSIM, −0.57σ LPIPS, +0.06σ log count. Committed and promotedbc7700bwith its switch. Detached combined check of the submitted tree PID 89320 (/tmp/r47/check.log,/tmp/r47/check). - Combined check on the actual submission/tmp/r47/check/check.jsonat 04:39:27 UTC: 2.3410× PASS (N1), 131.344 s (37.277 / 49.021 / 45.045 s), 4.096 s faster than the prior combined check and better in every window. Both exact gates passed; endpoint differences −1.90σ PSNR, −0.45σ SSIM, −1.53σ LPIPS, +0.53σ log count. Keepbc7700bsubmitted.
9.8 hRound 48 — ordered-tile stack profile (queued 2026-11-18 04:33:13 UTC)
- Detached PID 89652 waits for round 47's combined PASS and profiles the submitted stack at steps 18501–18520 (/tmp/r48/profile,/tmp/r48/profile.log). - GPU lock at 04:39:29 UTC. Profile: backward blending 27.39% of CUDA (1.220 ms/step), forward blending 21.51% (0.958 ms/step), dense Adam 18.79% (0.837 ms/step), radix scatter 6.33%. Factored SH Adam still emits a hardware divide refinement and slow-path check per coefficient. -GS_FAST_SH_ADAM_DIV=1in/tmp/r48/treeuses native FP32 approximate division only for the final factored SH update quotient with bounded positive epsilon and bias correction. Moments, square root, raw parameters, dense stepping, schedules and fallback are unchanged. Configured build PID 91086 (/tmp/r48/build.log). - Detached quotient/error and optimizer-state probes, judge time/quick and conditional N1 PID 91425 (/tmp/r48/{runner,compare,time,quick,n1}.log).