GPT-6.2 Astra 3.58×
Harness: Codex · switches: GS_FUSED_ADAM=1 GS_MASKED_REPLAY=1 GS_PACKET_FORWARD=1 GS_DENSE_ADAM=1 GS_COMPACT_TILES=1 GS_RAW_ACTIVATIONS=1 GS_STREAM_REPLAY=1 GS_DENSE_STATS=1 GS_COMPACT_RESIZE=1 GS_DEFER_LOGGING=1 GS_ZERO_ADJOINTS=1 GS_TILED_SH=1 GS_PRIORITIZE_TILES=1 GS_FUSED_IMAGE_LOSS=1 GS_REPLAY_MOMENTS=1 GS_COMPRESSED_GRADIENTS=1 GS_HALF_PACKETS=1 GS_QUARTER_PACKETS=1 GS_REPLAY_RECIPROCAL=1 GS_PIPELINED_LOGGING=1 GS_FUSE_TILE_COUNT=1 GS_PACKED_SORT=1 GS_INPLACE_SH=1 GS_PACKED_SPLAT_VALUES=1 GS_FUSED_PROJECTION_ADJOINT=1 GS_PACKET_BATCH=512 GS_PAIRED_SPLAT_FIELDS=1 GS_TRANSIENT_ADJOINTS=1 GS_VECTOR_CONIC_ATOMICS=1
- Judged
- 3.58×
- Tier
- N1
- Self-best
- 3.54×
- Rounds
- 57
- Hours
- 10.0
- API cost
- $97.4
- Input tokens · 97.2% cached
- 117 M
- Output tokens
- 0.7 M
- Requests
- 814
§ ARaw trace
Raw work log57 sections
0.0 hRound 1 — dense Adam
- Started 2026-11-17 18:47:30 UTC on one GPU. Judgeprofile, unchanged tree and empty overrides, PID 2416, log/tmp/r1/profile.log, output/tmp/r1/profile. Finished 18:48:31 UTC. - Steps 7000–7019: raster backward 34.7% of CUDA time (per-Gaussian blending 27.6%), raster forward 26.9% (tile render 16.9%), Adam operations 22.6% (42.84 ms). First lever: use the installed PyTorch fused dense Adam. Same six groups, dense updates, float32 states, and LR schedule.GS_FUSED_ADAM=1, off by default. - Detached judge time PID 2662 (/tmp/r1/time.log), quick PID 2857 (/tmp/r1/quick.log). - End 18:56:05 UTC:/tmp/r1/time/time.json264.276 s, 1.1565×;/tmp/r1/quick/check.jsonPASS (all exact items). Promoted for full checking. - Full check/tmp/r1/check/check.jsoncompleted 19:09:26 UTC: PASS N1, 1.1464×, all exact items and all four statistical endpoints passed. Timing windows for the initial time screen: 74.820, 96.664, 92.792 s.
0.2 hRound 2 — profile the fused-Adam tree
- Started 18:56:07 UTC. Detached profile PID 3717,/tmp/r2/profile.log, output/tmp/r2/profile. - Full check of Round 1 is separate from the round, PID 4048,/tmp/r1/check.log. - Profile: per-Gaussian backward blending 31.4% of CUDA, forward blending ~19%, fused Adam 13.2%, projection 5.9%, tile duplication 5.7%. 44–50k blocks in typical backward calls, with a warp serially replaying 256 pixels including all non-contributors. Target: backward blend only. Capture the existing per-pixel acceptance decisions in forward; replay their set bits in depth order in backward. Reassign work from pixels to Gaussians inside a block for the gradient reduction. - Round 2 detachedtimePID 4628, log/tmp/r2/time.log; quick PID 5050, log/tmp/r2/quick.log. Experiments remain off the submitted tree. - A separate locked, untimed reference-vs-candidate rendering/gradient probe uses small synthetic Gaussian clouds including partial tiles, density variation, and a depth adjoint (PID 5294,/tmp/r2/numerical-probe.log). Round 3 profile is queued after results: PID 5332,/tmp/r3/profile.log, on the fastest screened candidate. - Probe: bitwise-identical RGB, inverse depth and radii in all three fixtures; maximum relative L2 over the nine backward tensors <8e-7 in two cases and 2.33e-3 in a denser case. Follow-up separates upstream nondeterministic tile reduction from replay arithmetic. - Follow-up/tmp/r2/tile-state-probe.log(PID 5698): the dense case has 1.04e-2 relative L2 between two reference repeats; using the same forward tile metadata reduces the maximum candidate-reference difference to 3.20e-6. No change to the reference's tile reduction or the image model. - End 19:13:10 UTC.GS_MASKED_REPLAY=1at extension build./tmp/r2/time/time.json: 217.443 s (61.407 / 79.868 / 76.168 s), 1.4056×. Saves another 46.833 s vs fused Adam across the run./tmp/r2/quick/check.json: PASS. Full check warranted. - Full check/tmp/r2/check/check.jsonfinished 19:24:00 UTC: PASS N1, 1.4128×. Promoted to submission.
0.5 hRound 3 — late-stage profile
- Started 19:13:16 UTC, one GPU;/tmp/r3/profile, PID 5332. Fused Adam + masked replay; profile steps 15010–15021 after densification has ended. - The separate full check of Round 2 is PID 5946,/tmp/r2/check.log. Round 3 profile finished 19:15:06: forward blend 33.15% (2.504 ms/call), fused Adam 17.37%, masked replay 13.80%, backward preprocessing 8.11%, tile duplication 6.95%. Forward still gathers RGB/depth and computes the inverse depth inside the pixel loop. Test one forward blend kernel with cached attributes and warp-owned 8x4 packets within the unchanged sorted 16x16 tile. A conservative inflated ellipse bound removes irrelevant packets; ill-conditioned conics fall back to all packets. Per-pixel alpha and early termination math and depth order are unchanged. - Detached jobs:timePID 6232 (/tmp/r3/time.log), quick PID 6525 (/tmp/r3/quick.log). - Numerical probe PID 6605 (/tmp/r3/numerical-probe.log); bitwise RGB, inverse depth, radii on all fixtures including full-resolution, unusually thin splats. The thin-splat gradient comparison has large variation including reference vs reference: under investigation. Round 4 profile queued PID 6941 (/tmp/r4/profile.log). - Follow-up with identical forward buffers,/tmp/r3/same-forward-probe.log(PID 7188): in the thin-splat stress case, relative L2 for screen positions, RGB, opacity and SH gradients is at most 3.6e-6. The subsequent ill-conditioned covariance derivative amplifies summation-order differences for positions, scales and rotations (also unstable between reference repeats). The packet change has not modified that derivative. This is not a claim of bitwise backward equivalence. - End 19:27:10 UTC.GS_PACKET_FORWARD=1at build, cumulative/tmp/r3/time/time.json182.774 s (53.159 / 66.202 / 63.413), 1.6722×. Saves a further 34.669 s over Round 2./tmp/r3/quick/check.json: PASS. -/tmp/r3/check/check.json: PASS N1, 1.6616×. Promoted to submission.
0.7 hRound 4 — densification-phase profile
- Started 19:27:12 UTC. Profile PID 6941,/tmp/r4/profile.log, steps 11011–11026 away from the actual point-set resize. - Separate full check of Round 3: PID 7388,/tmp/r3/check.log. Profile of current fastest version: dense fused Adam 22.48% of device time, backward replay 16.92%, packet render 14.91%, combined preprocess 12.58%, tile duplication 8.83%, radix sort 4.83%. Target dense Adam: one batched launch, 16-byte vectorized loads/stores even when point count has a remainder, host bias corrections, matching the reference's first-moment interpolation; all parameter and moment entries updated on every dense update. No visibility mask. - Detached jobs: time PID 7688 (/tmp/r4/time.log); quick PID 8017 (/tmp/r4/quick.log). - Initialtimeandquick: ERROR before step 1; a raw parameter in the initialized scene is not row-major, rejected by the dense kernel's contiguous-view guard. Standalone dense test (PID 8123,/tmp/r4/adam-probe.log) agrees with foreach Adam to 2.4e-7 max absolute parameter error over 100 steps, including dense zero-gradient momentum updates. Fix the layout adapter, keep the same kernel. - Fix: operate on tensors in their physical stride order, staging and copying back only if the corresponding view is not dense. Re-test PID 8558 (/tmp/r4/retest.log). End 19:41:36 UTC./tmp/r4/quick2/check.jsonPASS,/tmp/r4/time2/time.json: 166.914 s (48.199 / 60.645 / 58.070), 1.8311×, another 15.860 s saved. -/tmp/r4/check/check.jsonfinished 19:59:48 UTC: PASS N1, 1.8142×. Promoted.
1.0 hRound 5 — reprofile after dense Adam
- GPU-lock start 19:49:04 UTC; judge profile PID 8910,/tmp/r5/profile.log, on Round 4, steps 7011–7026. Separate full check of Round 4 PID 9006,/tmp/r4/check.log. - Profile: backward blend 19.4%, packet forward 15.2%, dense Adam 14.7%, preprocess 13.0%, tile duplication 10.2% plus radix sort 5.3%. TheidentifyTileRangeslaunch covers 3.76M records for a camera with at most 1.56M real tile pairs; only 43.8% can be real over this window. Test exact overlap counting before the sort. Small rectangles cache the accepted tile mask, long rectangles use a warp to count/emit pairs. Keys preserve original per-point order and depth bits; no tile is discarded unless the upstream predicate rejects it. - Detached time PID 9177 (/tmp/r5/time.log), quick PID 9409 (/tmp/r5/quick.log). -/tmp/r5/pairs-probe.log(PID 9625): sorted IDs, keys, RGB, inverse depth and radii bitwise equal in all three stress fixtures, including depth ties. One fixture compacted 1,433,570 records to exactly 386,119 valid pairs. End 19:59:15 UTC./tmp/r5/time/time.json: 153.484 s (44.217 / 56.627 / 52.641), 1.9914×. Saves another 13.430 s./tmp/r5/quick/check.json: PASS. - Full check/tmp/r5/check/check.jsonfinished 20:10:27 UTC: PASS N1, 1.9880×. Promoted.
1.2 hRound 6 — post-densification profile
- Started 19:59:50 UTC. Profile PID 9863,/tmp/r6/profile.log, steps 17521–17536 of the compact-binning tree. - Full check of Round 5 PID 10486,/tmp/r5/check.log. Profile: replay 21.4%, dense Adam 17.5%, projection/SH kernels 15.9%, packet forward 14.7%, compact binning 4.6%, radix sort 3.5%. Feeding the projection kernels requires whole-array activation work: quaternion division backward 137 µs, norm backward 88 µs, expand backward 57 µs, forward norm 59 µs per step. Move exp-scale, normalized quaternion and sigmoid opacity and their corresponding adjoints into CUDA projection; external rasterizer callers retain activated-input behavior by default. - Detached time PID 10659 (/tmp/r6/time.log), quick PID 11164 (/tmp/r6/quick.log). - Probe PID 11231/tmp/r6/activation-probe.log: radii matched; RGB/depth relative L2 <8e-8, raw scale/rotation/opacity gradient relative L2 <9e-7 for normal and antialiased rendering, including sub-epsilon quaternions. - End 20:10:43 UTC./tmp/r6/time/time.json141.080 s (41.471 / 51.289 / 48.319), 2.1665×, 12.404 s faster than Round 5./tmp/r6/quick/check.jsonPASS. - Full check/tmp/r6/check/check.jsonfinished 20:17:13 UTC: PASS N1, 2.1616×. Promoted.
1.4 hRound 7 — reprofile replay and projection
- Started 20:10:45 UTC; PID 11455/tmp/r7/profile.log. Steps 9009–9024 of the raw-activation candidate. - Round 6 full check PID 11918/tmp/r6/check.log. Profile: packet raster 20.6%, masked replay 20.1%, dense Adam 18.3%, projection/preprocessing 17.7%. Try streaming one bucket per warp instead of splitting it between four warps: cap the in-flight coefficient matrix at 12 hit records/pixel, and contract the running color residual with the pixel's upstream gradient. Use an RGB-only specialization when autograd supplies no depth adjoint. The forward raster is unchanged. - Detached time PID 12406,/tmp/r7/time.log. - Detached quick PID 12678,/tmp/r7/quick.log. - Matched-forward gradient probe PID 12802,/tmp/r7/replay-probe.log, covering RGB-only and RGB+depth. End 20:26:03 UTC./tmp/r7/time/time.json: 137.356 s (40.949 / 49.892 / 46.516), 2.2252×. Further 3.724 s saved, largest in late training./tmp/r7/quick/check.jsonPASS. - Full check/tmp/r7/check/check.jsonfinished 20:33:13 UTC: PASS N1, 2.2220×. Promoted.
1.7 hRound 8 — profile training statistics
- Started 20:26:07 UTC. Profile PID 12901 (/tmp/r8/profile.log), steps 9509–9528 of the streamed-replay candidate. - Separate full check of Round 7 PID 13398/tmp/r7/check.log. Profile leaders: replay 22.52%, dense Adam 18.09%, packet forward 17.90%. Statistics path: 20 synchronizingnonzerocalls (14.111 ms host inclusive, not all removable overhead), 100 gathers, 60 scatters and 20 norms in 20 steps. Replace sparse indexing with a single dense radius-tested kernel for max radius, sum of screen-gradient lengths and observation count. Keep the screen-space autograd placeholder logically zero using a scalar-expanded leaf; no need to retain its gradient after statistics end. Default renderer still returns the index tensor. - Detached time PID 13538 (/tmp/r8/time.log), quick PID 13754 (/tmp/r8/quick.log). - Reference statistics probe PID 13872 (/tmp/r8/stats-probe.log): exact equality of all three accumulators for 3, 1,001 and 1,024,003 points. Next profile of the resize iteration queued PID 13978 (/tmp/r9/profile.log). - End 20:35:25 UTC./tmp/r8/time/time.json: 125.182 s (35.884 / 45.731 / 43.567), 2.4416×, another 12.174 s saved./tmp/r8/quick/check.jsonPASS. -/tmp/r8/check/check.jsonfinished 20:43:14 UTC: PASS N1, 2.3943×. Promoted.
1.8 hRound 9 — population resizing
- Started 20:35:27 UTC; PID 13978,/tmp/r9/profile.log, steps 13999–14002 including one scheduled densify/prune. - Separate full check of Round 8 PID 14925,/tmp/r8/check.log. In this four-step event window indexing takes 11.04 ms of device self-time (37.5%, 60 calls), with 7.31 ms vectorized gathers; it is periodic (once per 100 steps in the growth phase), not 37.5% of the entire run. Build one canonical destination mapping: retained originals, clones, then the same two repeated split-child batches, followed by pruning; gather each parameter and retained moment once. Make the same split random draw before pruning. Zero child moments and reset statistics exactly as before, without explicitly purging the allocator cache. - Detached time PID 15064 (/tmp/r9/time.log); quick PID 15459 (/tmp/r9/quick.log). - Matched-input population probe (PID 15578,/tmp/r9/resize-probe.log): exact ordered six parameters, optimizer step and both moments, statistics and RNG state after resizing for a mixed clone/split/prune case, a prune-only case, and pruning to empty. Round 10 profile queued PID 15708,/tmp/r10/profile.log. - End 20:45:42 UTC./tmp/r9/time/time.json: 123.185 s (34.717 / 44.951 / 43.517), 2.4812×. A further 1.997 s saved, concentrated in the growth window. Peak reserved memory 23.95 GiB vs 7.39 GiB./tmp/r9/quick/check.json: PASS. Candidate commitdb9e7a9, awaiting full check.
2.0 hRound 10 — reporting barriers
- Started 20:45:44 UTC. Judge profile PID 15708,/tmp/r10/profile.log, steps 17011–17026 on Round 9. Device leaders: replay 21.38%, dense Adam 20.13%, packet forward 17.63%, projection/SH forward 13.84%. Host: 16 stream synchronizations (23.931 ms inclusive), 112 scalar reads (24.293 ms inclusive, includes six CPU optimizer-state reads per step), and 32 event records/16 elapsed-time calls. These waits include real GPU work. Remove the per-step logging barrier, not the computation: without a per-step TensorBoard writer, transfer accumulated detached scalar losses only at display intervals, applying the original ordered Python double EMA. Preserve writer timing and validation reports. - Detached time PID 16494 (/tmp/r10/time.log), quick PID 17118 (/tmp/r10/quick.log). The 13.84% projection kernel above is the backward SH/covariance preprocess (the float3 means signature), not the forward kernel; forward is 4.12%. - End 20:54:18 UTC./tmp/r10/time/time.json116.698 s (32.181 / 42.944 / 41.573), 2.6191×, saving a further 6.487 s./tmp/r10/quick/check.jsonPASS. Combined rounds 9+10 full check PID 17434,/tmp/r10/check.log, on a frozen worktree at052b7a7.
2.2 hRound 11 — growth-phase projection profile
- GPU-lock start 21:00:53 UTC. Detached profile PID 17340/tmp/r11/profile.log, round 10, steps 12813–12828. - Device: streamed replay 22.54%, packet forward 19.17%, dense Adam 18.40%, backward SH/covariance projection 13.87% (581 µs/step). Target the last kernel: tile replay marks a point in the unused third conic-adjoint slot if and only if it accumulates pixel adjoints. Use that conservative support for projection backward, whose outputs are pre-zeroed, instead of evaluating the Jacobian even for occluded/no-loss points. This is independent of, and does not change, the dense optimizer. Detached time PID 17954,/tmp/r11/time.log. - Quick PID 18484 (/tmp/r11/quick.log); matched-input adjoint probe PID 18485 (/tmp/r11/projection-probe.log), including all-zero, sparse RGB-only and depth-only adjoints. - Accumulated Round 9+10 full check/tmp/r10/check/check.jsonfinished 21:09:56 UTC: PASS N1, 2.6297×. Promoted both separately committed changes and switches. - End 21:10:12 UTC./tmp/r11/time/time.json: 111.329 s (31.203 / 40.963 / 39.162), 2.7454×, saving 5.369 s over round 10./tmp/r11/quick/check.jsonPASS. Probe: positive radii for all 8,011 splats, but only 5,324 had any nonzero gradient for full RGB, 1,210 under sparse RGB and none for a zero adjoint. All nine returned gradient tensors agreed within 3e-4 relative L2; zero-gradient rows were exact (also for depth and antialiasing). Full check PID 19301 (/tmp/r11/check.log).
2.4 hRound 12 — replay packet scheduling
- Profile PID 18927 (/tmp/r12/profile.log), GPU-lock start 21:10:14 UTC. Profile of round 11, steps 15513–15528: streamed replay 21.94%, packet render 21.44%, dense Adam 20.32%, backward preprocess 9.68%. Existing backward scans eight pixel stripes per bucket and gathers Gaussian attributes before knowing whether any pixel needs them. Store eight per-bucket unions of the actual forward acceptance masks, order checkpoints by 8×4 pixel packet, and visit only nonempty packets and the attributes they actually reference. No change to compositing, thresholds or hit decisions. Time PID 19457 (/tmp/r12/time.log). - Quick PID 19857 (/tmp/r12/quick.log); numerical probe PID 19948 (/tmp/r12/packet-probe.log). Probe uses identical frame/geometry state and each implementation's own checkpoint layout. RGB, depth and radii were bitwise equal, and all eight adjoint fixtures (including AA, RGB-only, depth-only and zero) passed. - Round 11 full check/tmp/r11/check/check.jsonfinished 21:18:19 UTC: PASS N1, 2.7466×. Promoted round 11. - End 21:18:36 UTC./tmp/r12/time/time.json: 115.498 s (32.758 / 42.371 / 40.368), 2.6463×;/tmp/r12/quick/check.jsonPASS. All three timing windows regressed (4.169 s total). Extra metadata and packet reordering did not pay back: in the synthetic probe 3,235 of 3,432 packets were live. Committed the experiment asa2133ab, rejected it, and restored round 11 for further work.
2.5 hRound 13 — late-stage SH projection
- Profile on the round 11 tree, PID 20422 (/tmp/r13/profile.log), GPU-lock start 21:18:37 UTC; steps 18637–18656. - Device profile: Adam 21.39%, replay 21.18%, packet forward 20.52%, backward projection 8.94%. Test a local reassociation of the SH-to-position vector-Jacobian product: first dot each RGB coefficient with the clamped incoming adjoint, then evaluate scalar direction derivatives, instead of building a three-channel Jacobian before contracting it. Other gradients, forward SH and all parameter layouts stay unchanged. Detached time PID 20747 (/tmp/r13/time.log), quick PID 21286 (/tmp/r13/quick.log), paired-implementation untimed gradient probe PID 21288 (/tmp/r13/SH-probe.log). - End 21:25:39 UTC./tmp/r13/time/time.json111.795 s (31.366 / 41.085 / 39.344), 2.7340×;/tmp/r13/quick/check.jsonPASS, all gradient fixtures passed. No measurable speed benefit (0.466 s slower than round 11). Committed as3a0a3a7, excluded it, and restored round 11. This arithmetic simplification did not address the bottleneck.
2.7 hRound 14 — SH adjoint memory layout
- Profile on the round 11 tree, PID 21489 (/tmp/r14/profile.log), GPU-lock start 21:26:04 UTC; steps 14211–14226. - Replay 22.82%, dense Adam 20.47%, packet forward 19.08%, backward projection 9.59%, fills 4.34%. The SH gradient is a dense 45-float row for each of ~2.02M points (>360 MB), initially zeroed then populated by scattered per-point stores. Stage 32 rows per warp with coalesced float4 transfers, apply the original SH derivative, and write each dense gradient row once, including explicit zeros for non-contributors and inactive degrees. Time PID 21660 (/tmp/r14/time.log). - Initial/tmp/r14/time/time.json: 107.756 s (30.208 / 39.538 / 38.010), 2.8364×; quick PID 21994/tmp/r14/quick/check.jsonPASS. Local gradient probe PID 21997 (/tmp/r14/SH-tiling-probe.log): all nine outputs agreed across SH degrees 0–3, AA off/on, sparse/full/zero RGB and pure depth, with an 8,011-point partial final warp. - Added scalar fallback for otherwise contiguous external SH views whose storage offset is not 16-byte aligned. Revised full timing PID 22426 (/tmp/r14/time-aligned.log); unaligned degree sweep PID 22946 (/tmp/r14/SH-unaligned-probe.log). Full check on immutable revised tree PID 22947 (/tmp/r14/check.log). - End 21:37:27 UTC. Revised/tmp/r14/time-aligned/time.json: 107.881 s (30.224 / 39.645 / 38.012), 2.8331×, 3.448 s gained over round 11. All 32 deliberately unaligned input/degree/adjoint cases passed. Candidate commit721baa0; its full check runs separately.
2.8 hRound 15 — raster-stage profile
- Judge profile queued PID 23159 (/tmp/r15/profile.log), steps 9631–9650 of round 14. Untimed tile-work inspection queued after the profile, PID 23178 (/tmp/r15/tile-work.jsonand/tmp/r15/tile-work.log), to count sorted pairs and actual pixel-bucket occupancy for six growth-phase views of the scene. - Round 14 full check/tmp/r14/check/check.jsonfinished 21:41:58 UTC: PASS N1, 2.8327×. Promoted. Round 15 profile GPU-lock start 21:42:01 UTC. - Profile: replay 24.22%, packet forward 22.31%, Adam 19.51%. Two synchronous count reads per rasterization: 20 first reads averaged 1,739.6 µs host-inclusive, 20 second reads 194.9 µs. Eliminate the second read by bounding the allocated number of 32-entry tile buckets by min(R, floor((R+31T)/32)), where R is pair count and T is tile count. Replay still checks the GPU-resident exact count. Six inspected camera views needed 34,526–44,644 actual buckets and only 1,033–1,069 extra reserved entries; these views are immediately after an opacity reset. - Time PID 23822 (/tmp/r15/time.log); quick PID 24416 (/tmp/r15/quick.log), matched-buffer probe PID 24418 (/tmp/r15/capacity-probe.log). End 21:53:14 UTC./tmp/r15/time/time.json: 107.801 s (30.129 / 39.613 / 38.058), 2.8353×. Difference of 0.080 s from round 14 not resolved, last window slightly worse. Quick PASS; bitwise RGB/depth/radii and gradient probes passed. Excludedfa07d22; the eliminated host wait mostly overlapped useful GPU work.
3.1 hRound 16 — forward tile load balance
- Profile PID 24590 (/tmp/r16/profile.log), steps 16771–16790 of round 14 on the mature population. Per-tile counts on six corresponding views queued after profile, PID 24684 (/tmp/r16/tile-work.json). - GPU-lock start 21:53:26 UTC. Replay 23.37%, dense Adam 21.78%, packet rendering 19.64%. Tile candidate counts had medians 393–742 and maxima 3,517–7,561; the longest tenth of tiles accounted for 25–38% of the sum of the last processed indices. Test independent tile scheduling: within each frame, bucket tile IDs by log₂(list length), render long lists first, and reuse the now-dead count buffer for this permutation. Within-tile depth order and all pixel arithmetic remain unchanged. Time PID 24876 (/tmp/r16/time.log), quick PID 25100 (/tmp/r16/quick.log), numerical comparison PID 25103 (/tmp/r16/order-probe.log). - End 22:04:32 UTC./tmp/r16/time/time.json106.978 s (30.272 / 39.119 / 37.586), 2.8571×. Small effect concentrated after growth; repeated judgetimePID 25655/tmp/r16/repeat/time.json: 106.509 s (30.018 / 39.015 / 37.476), 2.8696×. Both late windows improved on both runs versus round 14./tmp/r16/quick/check.jsonPASS and independent eight-case probe found identical RGB/depth/radii and matching adjoints. Candidatef6032abis undergoing full check, PID 25925 (/tmp/r16/check.log).
3.3 hRound 17 — image loss
- Judge profile on round 16 PID 25831 (/tmp/r17/profile.log), steps 8341–8364. - GPU-lock start 22:04:58 UTC. Replay 23.53%, Adam 21.35%, packet forward 18.77%. SSIM forward/backward kernels 2.26% / 1.53% plus surrounding image pointwise operations, reductions and copies: in 24 steps, 96 means, 48 abs, 24 signs, 144 contiguous calls and 120 zeros_like. Test a single image objective: compute the same eleven-tap separable window for all five moments together, reduce full-image L1 and SSIM per tile and globally, and convolve all three SSIM adjoints together with the L1 sign term. No sampling. Ground-truth values are stored contiguously once. Time PID 26252 (/tmp/r17/time.log), quick PID 26606 (/tmp/r17/quick.log), scalar/gradient comparison PID 26609 (/tmp/r17/image-loss-probe.log). - Round 16 full check/tmp/r16/check/check.jsonfinished 22:10:09 UTC: PASS N1, 2.8646×. Promoted. - First/tmp/r17/time/time.json101.313 s (28.092 / 37.359 / 35.863), 3.0168×, and/tmp/r17/quick/check.jsonPASS. Not promoted: the local mixed VJP agreed at full resolution (relative L2 2.33e-7 at λ=0.2), but missed the probe tolerance on a flat patch with pure SSIM (6.23e-3). Diagnosis PID 27147 (/tmp/r17/loss-diagnosis.log): supplying the same derivative maps made backward exactly equal. Float32 arithmetic comparison identified the issue: reference rounds the target mean's square before subtracting it from E[y²], whereas the new forward used a fused subtract. Match the reference rounding for variances/covariance. Use out-of-place addition for optional depth loss since the two returned scalars share storage. - Revised time PID 27244 (/tmp/r17/time-matched.log), exact check PID 27803 (/tmp/r17/quick-matched.log), unchanged local tolerance suite PID 27805 (/tmp/r17/loss-matched-probe.log). - End 22:17:52 UTC./tmp/r17/time-matched/time.json101.640 s (28.139 / 37.501 / 36.000), 3.0071×. Improved all windows vs round 16; 5.338 s faster than its first screen./tmp/r17/quick-matched/check.jsonPASS. The same unmodified tolerance suite now passes all 15 image/weight combinations. Flat pure-SSIM loss error fell to zero and mixed gradient relative L2 to 0.00137; full-resolution λ=0.2 relative L2 1.66e-7. Commit809e230, full-check PID 28106 (/tmp/r17/check.log).
3.5 hRound 18 — replay work distribution
- Judge profile PID 28011 (/tmp/r18/profile.log), steps 16517–16536 on round 17. Untimed hit-mask inspection PID 28124 (/tmp/r18/replay-work.json) sampling every 29th bucket across six corresponding mature views, to count per-pixel and per-splat lane occupancy. - GPU-lock start 22:18:19 UTC. Replay 23.81%, Adam 23.25%, packet forward 19.00%. Every replayed pair used the same per-Gaussian opacity, conic and frame scale in its center/conic adjoint. Test accumulating weighted first and second moments of the pixel displacement and applying those factors once at the end of the bucket. Per-pixel ordered alpha/transmittance and color/opacity terms are unchanged. Time PID 28294 (/tmp/r18/time.log). - Quick PID 28704 (/tmp/r18/quick.log), matched-forward gradient probe PID 28734 (/tmp/r18/moment-probe.log). The sampled splat lanes used 22.9–30.4% of the possible lane-iterations in their pixel accumulation loop; the corresponding pixel-owned loop was 46.8–60.3% occupied (these counts ignore zero image gradients). The 1-in-29 bucket samples contained 1.58–2.11M forward-accepted pairs per view. - Round 17 full check/tmp/r17/check/check.jsonfinished 22:25:20 UTC: PASS N1, 3.0122×. Promoted. - End 22:27:08 UTC./tmp/r18/time/time.json: 99.230 s (27.287 / 36.609 / 35.334), 3.0801×. All eight same-forward gradient fixtures and/tmp/r18/quick/check.jsonpassed. Commita05e10e; full-check PID 29618 (/tmp/r18/check.log), immutable round-18 tree. Completed 22:34:27 UTC:/tmp/r18/check/check.jsonPASS N1, 3.0742×. Promoted.
3.8 hRound 19 — batched splat-adjoint reductions
- Profile GPU-lock start 22:29:20 UTC, PID 29526 (/tmp/r19/profile.log), steps 14731–14746 on round 18: packet forward 22.64%, dense Adam 22.20%, streamed replay 21.96%. The replay work counts suggest significant divergence in the per-Gaussian part of the backward. Test a warp matrix product for the batched adjoints, leaving the per-pixel depth-ordered replay alone. -GS_MATRIX_REPLAY: FP64 8×8×4 warp products aggregate the color adjoint and tile-local moments. Exact replay acceptance masks select coefficients; after the reduction each splat translates the moments to its own center. Time PID 30182 (/tmp/r19/time.log), quick check and same-forward numerical probe queued under PID 30777 (/tmp/r19/quick.log,/tmp/r19/matrix-probe.log). - End 22:40:28 UTC./tmp/r19/time/time.json: 103.319 s (28.071 / 38.251 / 36.997), 2.9582×, 4.089 s slower than round 18./tmp/r19/quick/check.jsonand all eight same-forward gradient fixtures passed. Even though the MMA removes the splat-owned scalar reduction, building and walking largely sparse matrices was slower in all windows. Committed as442e6c9and excluded.
3.9 hRound 20 — full-size late training profile
- GPU-lock start 22:41:16 UTC, judge profile PID 30976 (/tmp/r20/profile.log), steps 18713–18728 of the submitted round-18 version. - Adam 23.62%, streamed replay 21.39%, packet forward 20.81%. SASS of the packet inner loop shows eight range-reduction instructions around the Gaussian approximate exponent instruction, plus four separatevalue * alphaproducts. Test the hardware exponential for the EWA (fall back to accurate exp within 1e-8 of the alpha cutoff), consistently in forward and backward; reusealpha * transmittanceacross the four channels. Time PID 31229 (/tmp/r20/time.log). - End 22:48:01 UTC./tmp/r20/time/time.json: 102.085 s (28.408 / 37.603 / 36.075), 2.9940×, 2.855 s slower, all windows regressed. Quick/probe PID 31802:/tmp/r20/quick/check.jsonPASS, six forward fixtures had identical last-hit indices and maximum RGB/depth differences below 3.6e-7, and all eight gradient fixtures passed. A shorter common exponential path with a data-dependent guard did not translate into better full-run time. Committedc9de52c, excluded.
4.0 hRound 21 — dense-memory traffic
- GPU-lock start 22:48:01 UTC, profile PID 32042 (/tmp/r21/profile.log), steps 15793–15808 on round 18. Device attribute queries report support for both CUDA virtual memory management and generic lossless memory compression. - Dense Adam 22.50%, packet forward 22.15%, streamed replay 21.97%, tiled SH 6.86%. Dense Adam streams all parameters and two moment arrays every step; test CUDA compressible physical pages for allocations by the training thread, behind a PyTorch caching memory pool. The optimizer and all GPU arithmetic are untouched, and unsupported allocations fall back to ordinary device memory. Time PID 32232 (/tmp/r21/time.log). - End 22:53:55 UTC./tmp/r21/time/time.json: 99.160 s (27.540 / 36.512 / 35.109), 3.0823×, only 0.070 s faster than round 18 with a slower first window, not resolved from noise. The private pool increased reserved memory from 24.42 to 42.74 GiB. Quick/probe PID 32604:/tmp/r21/quick/check.jsonPASS; the driver confirmed compressible backing at all three tested allocation sizes, with correct data across CUDA streams. Committed37a1393; not promoted.
4.1 hRound 22 — derivative storage
- GPU-lock start 22:54:02 UTC. Judge profile of round 18, steps 13731–13746, then an untimed sampling of the actual SH parameters, gradients, and moments in round 21, PID 33064 (/tmp/r22/profile.log,/tmp/r22/adam-storage.json). Inspect actual compression properties and zero 32-byte sectors; no timing inferred from the sampling. - Profile: stream replay 22.88%, Adam 22.42%, packet forward 20.28%, tiled SH adjoint 6.85%. Initial sampler missed an explicit optimizer-module import; corrected retry PID 33270,/tmp/r22/adam-storage-retry.json. Across four views of 2.0M points only 3.6% of sampled parameter sectors and 4.8–4.9% of moment sectors were zero, versus 62.8–72.8% of gradient sectors. Driver inspection showed that in round 21 only the gradients were not compressed, owing to autograd's thread. Test compressing precisely the rasterizer's backward allocations instead of the entire training thread. Time PID 33395 (/tmp/r22/time.log). -/tmp/r22/time/time.json: 98.757 s (27.832 / 36.177 / 34.749), 3.0949×. Quick/diagnostic PID 33725,/tmp/r22/quick/check.jsonPASS. Actual SH gradients had driver compression type 1 on four sampled steps, parameters and moments remained at type 0. First timing saved only 0.473 s, so repeated on immutable source (PID 34227,/tmp/r22/time-repeat.log). - End 23:02:12 UTC. Repeat/tmp/r22/time-repeat/time.json: 98.781 s (27.093 / 36.791 / 34.897), 3.0941×. Total gain reproduced, although individual windows vary. Commit747c79c; full-check PID 34487 (/tmp/r22/check.log). Round 18 remains submitted pending the verdict.
4.3 hRound 23 — spatial packet granularity
- Judge profile PID 34393 (/tmp/r23/profile.log), steps 16117–16132 of the round-22 tree. Sampled half-warp versus full-warp packet candidate counts on the same model configuration queued separately, PID 34506 (/tmp/r23/packet-work.json). - GPU-lock start 23:02:39 UTC; replay 22.63%, Adam 22.17%, packet forward 20.45%. Replace each 8×4 pixel warp's shared candidate mask by one mask for each 4×4 half-warp; the two halves traverse eligible splats in their original order. Tile work counts across six mature views suggest 4.6–6.4% fewer warp candidate iterations (geometric bound, not a speed measurement). Time PID 34659 (/tmp/r23/time.log). - Quick and exact-forward/active-checkpoint probe PID 35197 (/tmp/r23/quick.log,/tmp/r23/half-packet-probe.log). Separate full check of round 22 finished 23:08:57 UTC:/tmp/r22/check/check.jsonPASS N1, 3.1070×. Promoted. - End 23:10:58 UTC./tmp/r23/time/time.json: 96.553 s (26.900 / 35.523 / 34.130), 3.1655×, 2.204 s faster than round 22's first screen./tmp/r23/quick/check.jsonPASS. Six normal and thin-splat fixtures were bitwise identical for RGB, inverse depth, all radii, per-pixel last contributor, and the hit masks and five prefix fields for every live pixel/bucket. Commit31fb571; full-check PID 35927 (/tmp/r23/check.log).
4.4 hRound 24 — coefficient exchange in replay
- Judge profile PID 35835 (/tmp/r24/profile.log), steps 14371–14386 of the round-23 tree; inspecting shared-memory exchange between the pixel-owned and splat-owned halves of the replay. - GPU-lock start 23:11:34 UTC. Stream replay 23.22%, dense Adam 22.63%. The sparse coefficient lookup needs a per-pair rank/popcount and multiple exchanges in stripes with over 12 hits per pixel. Test a bank-swizzled, directly addressed shared-memory table; amortize mask transposition across the entire stripe while retaining sparse scalar reductions. Round-23 full check/tmp/r23/check/check.jsonfinished 23:16:35 UTC: PASS N1, 3.1597×. Promoted. - Packet forward 18.36%. Detached time PID 36408 (/tmp/r24/time.log), quick and same-forward gradient probe PID 36744 (/tmp/r24/quick.log,/tmp/r24/direct-probe.log). Compiler resource report: still 56 registers per thread and no local memory, but shared storage increases from 21,376 to 40,960 bytes per block. - End 23:23:03 UTC./tmp/r24/time/time.json: 98.374 s (26.859 / 36.376 / 35.139), 3.1069×, 1.821 s slower than round 23./tmp/r24/quick/check.jsonand the eight same-forward gradient fixtures passed. Eliminating rank calculations did not pay for the substantially larger per-warp storage and its occupancy cost. Commit853b55d, excluded.
4.6 hRound 25 — packet candidate bounds
- GPU-lock start 23:23:07 UTC, profile of round 23 at steps 11873–11888, then untimed sampled candidate counts for multiple smaller packet shapes and a supplementary conic-axis bound. PID 37237 (/tmp/r25/profile.log,/tmp/r25/fine-work.json). - Replay 22.83%, Adam 22.46%, packet forward 19.24%. Further subdivision to 2×4 or 4×2 avoided only about 3% of sampled candidate iterations; a separating-axis bound on the existing 4×4 packets avoided about 11–14%. Use one conic row as the separating normal: any supported displacement obeys |row·displacement| ≤ sqrt(diagonal × level). Expand the interval for pixel extent and float rounding, retain the prior box for ill-conditioned/out-of-range inputs or unhelpful shapes. Time PID 37485 (/tmp/r25/time.log). Exact production mask tested on 1,048,576 generated conics including near-singular, high-offset and opacity/ellipse cutoff cases (/tmp/r25/axis-geometry.log): zero misses among 71,290,197 accepted pixel pairs. - End 23:37:03 UTC./tmp/r25/time/time.json: 97.046 s (27.069 / 35.654 / 34.323), 3.1495×, 0.493 s slower than round 23 with three small regressions. Quick/fixture PID 38179:/tmp/r25/quick/check.jsonPASS; six end-to-end render fixtures produced identical RGB, depth, hit masks and all live checkpoints. Computing the tighter bound inside serial tile batches offset its reduction in candidate visits. Commit57f5a37, excluded.
4.9 hRound 26 — gather and bin layout
- GPU-lock start 23:39:51 UTC, profile PID 38388 (/tmp/r26/profile.log), steps 17431–17446 of the verified round-23 tree. - Dense Adam 23.01%, replay 22.62%, packet forward 18.77%. Both raster passes gather the same projected centers, conics, RGB and depths indirectly through the exact tile-depth permutation. Test gathering once into three contiguous float4 arrays per sorted record, including the unchanged packet mask, point ID and inverse depth; both tile kernels consume these arrays. Time PID 38548 (/tmp/r26/time.log). - End 23:45:58 UTC./tmp/r26/time/time.json: 98.018 s (27.335 / 36.083 / 34.600), 3.1182×, 1.465 s slower than round 23. Quick/fixture PID 38790,/tmp/r26/quick/check.jsonPASS; six exact render/checkpoint fixtures and eight gradient comparisons passed. The gather saves repeated indirect loads but materializes 48 additional bytes for every sorted pair, including unused tail pairs. All three timing windows regressed. Commit2a01452, excluded.
5.0 hRound 27 — inexpensive packet subdivision
- GPU-lock start 23:45:59 UTC, profile PID 39357 (/tmp/r27/profile.log), steps 10131–10146 on the round-23 tree. Revisit the simpler axis-aligned alternative quantified by round 25's packet-work diagnostic. - Replay 23.98%, Adam 21.62%, packet forward 19.15%. Four independent 2×4 pixel packets per warp keep the original axis-aligned bounds and depth-sorted splat order and use a 32-bit membership mask. The six-view diagnostic estimated 2.5–4.3% fewer warp candidate iterations over round 23. Time PID 39528 (/tmp/r27/time.log). - End 23:51:09 UTC./tmp/r27/time/time.json: 95.949 s (26.735 / 35.313 / 33.901), 3.1855×, 0.604 s faster than round 23; all windows improved. Quick/fixture PID 39895,/tmp/r27/quick/check.jsonPASS, six fixtures bitwise identical across all rendered outputs and live checkpoint fields. Candidate commit23b87b4; full-check PID 40452 (/tmp/r27/check.log).
5.1 hRound 28 — compositing adjoint division
- Profile PID 40317 (/tmp/r28/profile.log) at steps 16689–16704 on round 27, separating the compositing recurrence from its sparse reduction. - GPU-lock start 23:51:23 UTC. Dense Adam 23.35%, replay 22.72%, packet forward 17.98%. Each accepted pair divides its scalar adjoint by1-alpha, a positive normal divisor because upstream clamps alpha to 0.99. Test CUDA's native float division specifically at that site; forward computations and all thresholds remain unchanged. Time PID 40620 (/tmp/r28/time.log). - Quick and normal/saturated gradient fixtures PID 41150 (/tmp/r28/quick.log,/tmp/r28/reciprocal-probe.log). Round 27's full check/tmp/r27/check/check.jsonfinished 23:56:52 UTC: PASS N1, 3.1852×. Promoted. - End 23:58:54 UTC./tmp/r28/time/time.json: 95.435 s (26.380 / 35.268 / 33.787), 3.2026×, 0.514 s faster than round 27, again improving all windows./tmp/r28/quick/check.jsonPASS and all 16 matched-forward gradient fixtures passed, including saturated opacities with pure depth or sparse/all-zero color adjoints. Commit1d9f26c; full check PID 41847 (/tmp/r28/check.log).
5.2 hRound 29 — progress-report synchronization
- Profile PID 41752 (/tmp/r29/profile.log), 24 steps starting at 19973 on round 28. Investigate host gaps rather than launching another geometry rewrite. In round 28's official trace, the two largest gaps (472 and 389 µs) immediately followed exactly 40 bytes of D2H fromaten::to/_to_copy/copy_and preceded Gaussian Adam; this matches the ten-float logging read. - GPU-lock start November 17 23:59:17 UTC. The new trace reproduced both gaps (459 and 399 µs); Adam was 23.56%, replay 21.49%, packet forward 18.61%. Submit the ten-scalar copy into pinned memory just before backward, record a stream event, then consume it at the original logging point. The Python EMA and progress report retain their original order and iteration. Time PID 42013 (/tmp/r29/time.log), November 18 UTC. - Round 28's separate full check/tmp/r28/check/check.jsonfinished November 18 00:04:55 UTC: PASS N1, 3.1654× (96.556 s). This standalone full timing did not reproduce its initial 0.514 s gain over round 27; do not promote round 28 alone on that basis. Confirmation time PID 42875 (/tmp/r28/time-repeat.log). - End November 18 00:06:52 UTC./tmp/r29/time/time.json: 94.197 s (26.335 / 34.625 / 33.237), 3.2447×, 1.238 s faster than round 28's first screen. Quick/exact-EMA fixture PID 42634:/tmp/r29/quick/check.jsonPASS, every loss and ordered CPU EMA matched exactly at three resume boundaries in float32 and float64. Commit18dccc2; full-check PID 43200 (/tmp/r29/check.log).
5.4 hRound 30 — dense Adam memory scheduling
- Official profile PID 43181 (/tmp/r30/profile.log), steps 18413–18432 of round 29. Inspect the interior vector path separately from the small per-group tail; the kernel uses 46 registers with no local memory. - Round-28 confirmation/tmp/r28/time-repeat/time.json, completed 00:08:40 UTC: 95.680 s (26.642 / 35.203 / 33.835), 3.1944×. Two standalone screens favor the reciprocal, one full timing did not; the combined round-29 check is still the promotion decision. - Round 29 full check/tmp/r29/check/check.jsonfinished 00:12:44 UTC: PASS N1, 3.2417×, 94.285 s. Promoted rounds 28 and 29 together. Round 30's profile took the GPU lock at 00:12:44 UTC. - Adam 23.57%, replay 21.82%, packet forward 18.65%, tiled SH 6.94%. The two 40-byte progress copies now target pinned memory, followed by 7/8 µs gaps before backward kernels, rather than 400+ µs gaps before Adam. For Adam, prove alignment and bounds once per complete group block, issue two independent sets of vector loads before consuming either, and send partial/misaligned blocks down a coalesced scalar path. All entries still perform the same update. Time PID 43868 (/tmp/r30/time.log). - End 00:18:39 UTC./tmp/r30/time/time.json: 95.475 s (26.922 / 34.913 / 33.640), 3.2013×. All three windows regressed relative to round 29. Quick/probe PID 44219 (/tmp/r30/quick.log,/tmp/r30/Adam-probe.log):/tmp/r30/quick/check.jsonPASS, but the synthetic exact Adam comparison reported a mismatch. Commitf0924b5, excluded; the submitted optimizer is unchanged.
5.5 hRound 31 — mature-stage profile
- GPU-lock start 00:19:30 UTC November 18. Official profile of the verified round-29 tree at steps 16681–16698, PID 44810 (/tmp/r31/profile.log). - Adam 23.81%, stream replay 22.58%, packet render 16.75%, tiled SH backward 7.04%. Test the forward-to-backward checkpoint layout: interleave three color prefix channels and transmittance in one aligned 16-byte vector per pixel/bucket; retain depth and hit masks and every floating-point field. Timing PID 45027 (/tmp/r31/time.log). The follow-up round-30 probe, PID 44938 (/tmp/r30/tail-diagnostic.log), localized its first mismatch to the scalar tail block, with maximum absolute parameter difference 5.96e-8. - Quick check and same-forward/active-checkpoint probes PID 45293 (/tmp/r31/quick.log,/tmp/r31/prefix-probe.log,/tmp/r31/gradient-probe.log). - First timing:/tmp/r31/time/time.json93.985 s (26.020 / 34.636 / 33.328), 3.2520×./tmp/r31/quick/check.jsonPASS, six fixtures bitwise equal on RGB/depth/radii, hits and all live checkpoint fields, and all eight reverse fixtures passed. Commit80f7c8e. Only the first timing window clearly improved; repeat PID 45972 (/tmp/r31/time-repeat.log) before drawing a conclusion. - End 00:28:36 UTC./tmp/r31/time-repeat/time.json: 94.906 s (26.766 / 34.798 / 33.342), 3.2205×. Neither mature timing window beat round 29 on either screen. Excluded; fewer instructions did not resolve an end-to-end advantage.
5.7 hRound 32 — projection and tile enumeration
- Official profile of round 29 at steps 12981–12998, then an untimed footprint histogram over six training views; PID 46238 (/tmp/r32/profile.log,/tmp/r32/footprints.json), GPU-lock start 00:28:56 UTC. - Replay 23.78%, Adam 22.03%, packet render 18.66%; forward projection 4.83% and the tile-counting pass 3.76%. Rectangles of at most 32 tiles cover 97.6–99.1% of visible points in the sample; only 8.9–14.9% of all warps contain a larger rectangle. Compute the normal rectangle's exact overlap mask and popcount in projection while its conic and bounds are in registers. Large rectangles retain the cooperative count pass; every emitted pair still uses the original overlap test and depth key. Time PID 46459 (/tmp/r32/time.log); quick and exact binning/forward probes PID 46682 (/tmp/r32/quick.log,/tmp/r32/binning-probe.log,/tmp/r32/prefix-probe.log). - End 00:35:14 UTC./tmp/r32/time/time.json: 93.696 s (26.240 / 34.400 / 33.056), 3.2621×;/tmp/r32/quick/check.jsonPASS. All depth keys, sorted IDs, radii, RGB and inverse depth matched on three binning fixtures including ties/long thin splats. Six additional fixtures matched every live checkpoint. Commit7c0e666; full check PID 47356 (/tmp/r32/check.log).
5.8 hRound 33 — exact depth ordering
- Profile of the screened round-32 tree at mature steps 17471–17486, PID 47354 (/tmp/r33/profile.log). Assess the cost of sorting composite 64-bit tile/depth keys and the number of ties a one-word coarse key would need to resolve exactly. - GPU-lock start 00:37:44 UTC. Adam 22.69%, replay 22.04%, packet forward 21.31%; 96 CUB onesweep passes for 16 views, 4.36% of CUDA time. Emit monotone 32-bit tile/coarse-depth keys, stable radix sort them, and stably reorder ties up to 32 records against their full unsigned depth bits in the range-identification kernel. Long ties trigger re-emission and the original full-width stable sort, flagged in the existing bucket-count readback. Sampled key collisions PID 47474 (/tmp/r33/key-runs.json); timing PID 47702 (/tmp/r33/time.log). - In six actual views, the longest coarse ties ranged from 13–24 records, no full-sort fallback required; 22.1–30.9% of records belonged to a tie. Quick and exact depth-sort/forward probes PID 47956 (/tmp/r33/quick.log,/tmp/r33/depth-key-probe.log,/tmp/r33/prefix-probe.log). The controlled depth inputs exercise short reversed ties, full ties and long-run fallback explicitly. - Round 32 full check/tmp/r32/check/check.jsonfinished 00:44:09 UTC: PASS N1, 3.2446× (94.200 s). Endpoint differences were −0.69σ PSNR, +0.92σ SSIM, −1.23σ LPIPS and +0.45σ log-count. Promoted the identical tree/configuration. - End 00:46:17 UTC./tmp/r33/time/time.json: 91.597 s (25.445 / 33.729 / 32.423), 3.3368×; 2.099 s faster than round 32 with savings in all windows./tmp/r33/quick/check.jsonPASS. The five depth-order fixtures reproduced IDs, ranges, radii, RGB and depth bitwise, including 2,403 corrected reversed neighbors and a 2,101-record fallback run. All six ordinary/thin fixtures matched their live checkpoints exactly. Commita956de1; full-check PID 48859 (/tmp/r33/check.log).
6.0 hRound 34 — SH tile storage
- Judge profile of round 33 at mature steps 18793–18810, PID 48857 (/tmp/r34/profile.log). Static resource accounting on the preceding tree found 47,104 bytes of shared memory in the SH adjoint; inspect the new profile before selecting its tile staging path. - GPU-lock start 00:46:30 UTC. Adam 23.52%, replay 22.24%, packet forward 20.56%, tiled SH 7.05%. Radix sorting now takes four passes per view; repair/range detection is 0.68%. In the SH adjoint, consume all inputs of a degree band before overwriting them with coefficient gradients. Zero only absent/unused entries as the tile is staged, permitting a single shared tile for inputs and outputs. Time PID 49134 (/tmp/r34/time.log); quick and all-degree SH gradient probes PID 49331 (/tmp/r34/quick.log,/tmp/r34/SH-*.log). - Round 33 full check/tmp/r33/check/check.jsonfinished 00:51:57 UTC: PASS N1, 3.3201× (92.057 s). Per-endpoint deviations were +0.54σ PSNR, +1.44σ SSIM, −2.17σ LPIPS and +0.51σ log-count. Promoted the checked code and switches. - End 00:54:03 UTC./tmp/r34/time/time.json: 90.309 s (25.052 / 33.293 / 31.964), 3.3844×, 1.288 s faster than round 33./tmp/r34/quick/check.jsonPASS and all 64 aligned/unaligned all-degree adjoint fixtures passed the numerical comparison. Shared memory falls to 24,064 bytes, no spills. An additional one-tile exact probe matched through degree 2 but found a 1.61e-8 relative-L2 difference in the degree-3 coefficient gradient after rescheduling the band. Classify as N1, not bitwise. Commitf627bed, full-check PID 50340 (/tmp/r34/check.log).
6.1 hRound 35 — replay stripe work distribution
- GPU-lock start 00:54:18 UTC. Profile of round 34, steps 15643–15660, then untimed work accounting on six production views for alternative pixel stripes, PID 50338 (/tmp/r35/profile.log,/tmp/r35/stripe-work.json). - Adam 24.25%, replay 23.41%, packet forward 19.11%; the new SH adjoint is 5.63%. The actual acceptance masks predict that lending alternating pixel members to a splat's idle warp partner reduces peak splat-side loop iterations by 22.8% without more pixel-side work. Test loans only on stripes with at least four members, and return the partner's accumulated moments once per bucket. More dispersed stripes would reduce splat iterations but increase pixel-side maxima by 20–35%; compact 8×4 stripes instead reduce pixel maxima by 11.8% at a 3.6% splat-side cost. Time PID 50924 (/tmp/r35/time.log); quick and reverse probes PID 51115 (/tmp/r35/quick.log,/tmp/r35/gradient-probe.log,/tmp/r35/single-tile.log). - Round-34 deterministic follow-up PID 50468 (/tmp/r34/SH-rounding.log): across all degrees, AA settings and adjoint types only four returned tensors differed bitwise; worst relative L2 was 3.36e-8. - Round 34 full check/tmp/r34/check/check.json: PASS N1, 3.3780× (90.480 s). Endpoints −0.53σ PSNR, −0.86σ SSIM, −1.24σ LPIPS, +0.83σ log-count; promoted. - End 01:02:32 UTC./tmp/r35/time/time.json: 90.901 s (25.318 / 33.416 / 32.167), 3.3624×, 0.592 s slower than round 34, with regressions in all windows./tmp/r35/quick/check.jsonPASS, all numerical fixtures passed, worst relative L2 1.99e-7 on the deterministic one-tile cases. RGB-only replay rose from 56 to 72 registers per thread (depth: 80); reduced loop imbalance did not pay for the extra routing and persistent partials. Commitceae9e9, excluded.
6.3 hRound 36 — per-Gaussian raster attributes
- GPU-lock start 01:03:31 UTC. Profile of the verified round-34 tree at steps 13741–13758, PID 51899 (/tmp/r36/profile.log); examine color and inverse-depth loads in the two rendering kernels. - Adam 23.30%, replay 23.29%, packet forward 20.53%, projection 6.14%. Store internally evaluated RGB and reciprocal depth as one aligned float4 during projection and gather it directly in blend and replay. This is one record per Gaussian, not the extra sorted-pair staging pass rejected in round 26. Precomputed RGB retains its ordinary layout and separate depths. Timing PID 52080 (/tmp/r36/time.log); quick plus exact internally shaded/precomputed RGB, depth, checkpoint and derivative probes PID 52271 (/tmp/r36/quick.log,/tmp/r36/prefix-*.log,/tmp/r36/gradient-probe.log). - End 01:10:20 UTC./tmp/r36/time/time.json: 89.875 s (24.875 / 33.239 / 31.761), 3.4008×, 0.434 s faster than round 34 with faster times in all windows./tmp/r36/quick/check.jsonPASS; all 12 SH/precomputed-color forward fixtures matched live state bitwise, and all 32 deterministic derivative comparisons matched bitwise. Commitbe14afd; full-check PID 52956 (/tmp/r36/check.log).
6.4 hRound 37 — independent forward work
- GPU-lock start 01:10:39 UTC, round-36 profile at steps 16817–16834, PID 52954 (/tmp/r37/profile.log). Adam 24.74%, replay 22.97%, packet render 18.67%, forward projection including shading 6.57%, tiled SH adjoint 5.80%. The binning pass needs projected geometry, not SH colors; test overlapping bounded-grid shading with the binning work and its two existing host count reads. - The worker waits for projected radii and depths, writes the same per-Gaussian color/clamp data, and signals an event before blend. Precomputed colors, very small inputs and debug mode stay on their synchronous path. Time PID 53394 (/tmp/r37/time.log); quick and exact attribute/forward/derivative checks PID 53639 (/tmp/r37/quick.log,/tmp/r37/shading-*.log,/tmp/r37/gradient-probe.log). - Round 36 full check/tmp/r36/check/check.json: PASS N1, 3.3977× (89.955 s), endpoint deviations −2.32σ PSNR, −1.67σ SSIM, −0.01σ LPIPS and +0.10σ log-count. Promoted the checked tree and switches. - End 01:17:52 UTC./tmp/r37/time/time.json: 91.055 s (26.031 / 33.168 / 31.857), 3.3567×, 1.180 s slower than round 36 with an especially large first-window regression./tmp/r37/quick/check.jsonPASS, but the exact forward probe found a small nonzero image difference (below 4e-6 on the first fixture); follow-up PID 54350 (/tmp/r37/float-check.log). Separate shading requires an extra visibility/input traversal and stream dependencies; not a wall-time win. Commit2bac72b, excluded.
6.5 hRound 38 — compact replay pixel stripes
- Official profile of the verified round-36 tree at steps 17323–17338, PID 54348 (/tmp/r38/profile.log). Revisit the pixel-side imbalance measured with actual acceptance masks during round 35, without extra shared totals or any new checkpoint metadata. - GPU-lock start 01:18:48 UTC. Adam 24.55%, replay 22.46%, packet forward 19.59%, tiled SH 5.77%. Replace replay's two 16-pixel scan lines with four adjacent 8-pixel segments; each warp still covers the identical tile in eight stripes. Only the inter-pixel summation order and address mapping change. Time PID 54524 (/tmp/r38/time.log), quick and derivative probes PID 54526 (/tmp/r38/quick.log,/tmp/r38/gradient-probe.log,/tmp/r38/single-tile.log). The rejected round-37 follow-up matched raw visible shading values, but the thin-splat stress fixture exposed larger downstream image differences (0.0021 RGB / 0.0074 inverse depth); it remains excluded. - End 01:24:33 UTC./tmp/r38/time/time.json: 90.163 s (25.608 / 32.942 / 31.614), 3.3899×. Later windows improved but the first regressed 0.733 s, leaving the full run 0.288 s slower than round 36./tmp/r38/quick/check.jsonPASS; all 64 numerical cases passed (one-tile max relative L2 6.00e-8). Commitf833557, excluded rather than selecting a configuration on two partial windows.
6.6 hRound 39 — forward candidate-mask construction
- Official round-36 profile, steps 14841–14858, then independent all-group finite-density sampling of unchanged Adam writes (profiling only); PID 55397 (/tmp/r39/profile.log,/tmp/r39/dense-writes.json). Inspect whether masks can be shared across the eight pixel warps instead of balloted independently. - GPU-lock start 01:25:06 UTC. Adam 23.63%, replay 23.35%, packet forward 19.42%, tiled SH 5.72%. Transpose each 32-Gaussian × 32-packet membership matrix once in its loader warp. Replace the four ballots in each consuming eight-pixel packet group with a direct shared candidate-mask lookup. No change to masks, splat order, checkpoints or pixel arithmetic is intended. Time PID 55652 (/tmp/r39/time.log); quick and bitwise live-state probe PID 55654 (/tmp/r39/quick.log,/tmp/r39/mask-probe.log). - The dense-write sample found only 6.0% of 32-byte SH parameter sectors unchanged over a step (5.7% first moment, 5.1% second). Other parameter groups were at 0–0.1% unchanged sectors. This does not motivate conditional stores in this memory-bound optimizer; no sparse or lazy update was introduced. - First time/tmp/r39/time/time.json: 89.778 s (25.067 / 33.015 / 31.696), 3.4044×./tmp/r39/quick/check.jsonPASS; all six forward fixtures matched including hits and live checkpoints. Only 0.097 s faster than the R36 screen, so a repeat is running (PID 56514,/tmp/r39/time-repeat.log). - End 01:34:36 UTC. Repeat/tmp/r39/time-repeat/time.json: 90.706 s (25.700 / 33.139 / 31.867), 3.3696×; 0.831 s slower than the verified configuration's screen. Excluded commit8ec4b4d; the small first gain did not repeat.
6.8 hRound 40 — projection adjoints
- Official profile of the verified R36 tree, steps 11680–11699, PID 56753 (/tmp/r40/profile.log). - GPU-lock start 01:34:37 UTC. Dense Adam 23.78%, replay 22.93%, packet forward 20.31%; the two per-point projection-adjoint passes take 2.02% + 3.32%.GS_FUSED_PROJECTION_ADJOINTmerges these passes, sharing the live predicate, mean and six covariance partials in a thread. It still returns every public gradient, and separates the inline-SH variant from the M=15 tiled path. Time PID 56934 (/tmp/r40/time.log); quick and single-/multi-tile all-degree probes PID 57137 (/tmp/r40/quick.log,/tmp/r40/gradient-probe.log,/tmp/r40/gradient-multi.log). - End 01:41:29 UTC./tmp/r40/time/time.json: 88.826 s (24.909 / 32.604 / 31.313), 3.4409×, 1.049 s faster than R36./tmp/r40/quick/check.jsonPASS. All 32 deterministic AA/degree/adjoint combinations matched in every gradient bit; all 32 multi-tile combinations also passed. The common fused specialization uses 60 registers, no stack/local memory, versus 54 and 64 registers for the individual kernels. Commita1b048a; full check queued, PID 57862 (/tmp/r40/check.log).
6.9 hRound 41 — repeated pixel work in replay
- Profile of screened R40 at steps 18357–18372, PID 57861 (/tmp/r41/profile.log). - GPU-lock start 01:41:30 UTC. Dense Adam 25.15%, replay 22.19%, packet forward 19.41%, forward projection 6.60%. The same RGB gradient, final-image dot product and background term are fetched/recomputed for every live pixel/bucket. Precompute one float4 per pixel with RGB gradient and the bucket-independent residual. The RGB-only replay then needs one vector load; depth-adjoint rendering retains its original path. Algebraically identical dot products reassociate, so check gradient error as well as the exact gate. Timing PID 58129 (/tmp/r41/time.log); quick and gradient probes PID 58331 (/tmp/r41/quick.log,/tmp/r41/gradient-*.log). - R40 full check/tmp/r40/check/check.jsonfinished 01:46:48 UTC: PASS N1, 3.4331×, 89.029 s (24.861 / 32.721 / 31.447). All exact items and four model endpoints passed. Promoted as42e04f5and verified its content and switches against the checked snapshot. - End 01:55:07 UTC./tmp/r41/time/time.json: 88.810 s (24.799 / 32.688 / 31.322), 3.4415×./tmp/r41/quick/check.jsonPASS. All fixtures passed; maximum deterministic relative L2 gradient error was 1.77e-5. Excluded commit3a66823: only 0.016 s faster overall, both mature windows slower than R40, with a new rounding difference and an extra scratch allocation.
7.1 hRound 42 — tile scheduling granularity
- Official profile of submitted R40, steps 19219–19238, PID 59443 (/tmp/r42/profile.log). - GPU-lock start 01:55:08 UTC. Adam 25.09%, replay 22.80%, packet forward 18.89%. Existing measured tile lists have 90th percentiles of 1,599–2,244 splats across six late views, versus a 256-splat load batch. Parameterize only this batch length: 512 amortizes barriers and preloads more attributes; 128 uses less shared memory and can stop loading an opaque tile sooner. Keep the 16×16 CTA, 32-splat buckets, depth ordering, hit masks, pixel math, and final reduction intact. Both official timings PID 59629 (/tmp/r42/time-{512,128}.log), quick and exact forward fixtures PID 59861 (/tmp/r42/quick-{512,128}.log,/tmp/r42/checkpoints-{512,128}.log). - 512-splat screen/tmp/r42/time-512/time.json: 88.346 s (24.659 / 32.478 / 31.209), 3.4596×, faster in all three windows. Quick PASS and all six exact checkpoint fixtures passed. While 128 builds, an untimed attribute-sector diagnostic for future layout work uses the 512 tree (PID 60900,/tmp/r43/attribute-sectors.json,/tmp/r43/sector-probe.log). - End 02:06:05 UTC. 128-splat screen/tmp/r42/time-128/time.json: 90.758 s (25.299 / 33.418 / 32.041), 3.3677×, slower in every window despite a quick PASS and exact forward fixtures. Commit7398496; only the 512 variant proceeds to full check (PID 61433,/tmp/r42/check.log).
7.4 hRound 43 — sector locality of splat attributes
- Profile of screened R42-512 at steps 12173–12192, PID 61432 (/tmp/r43/profile.log). Six growth-phase camera fixtures at the same location in training showed 2.26–2.57 sorted instances per visible point and a predicted 0.5004–0.5007 ratio of distinct 32-byte sectors per warp for paired vs separate 16-byte conic/RGB-depth records. This is a traffic model, not timing. - GPU-lock start 02:06:06 UTC. Replay 26.05%, dense Adam 22.33%, packet forward 19.97%, projection 6.10%. Place the conic/opacity float4 and the internally shaded RGB/inverse-depth float4 in one aligned 32-byte record per Gaussian, not per emitted pair. Binning and both blend passes read the same conic field; precomputed color still reads its separate external buffer. No additional pass or duplicated field. Time PID 62020 (/tmp/r43/time.log); quick, exact precomputed/internal forward and all-degree gradient probes PID 62234 (/tmp/r43/quick.log,/tmp/r43/checkpoints-*.log,/tmp/r43/gradient-probe.log). - R42 full check/tmp/r42/check/check.jsonfinished 02:10:47 UTC: PASS N1, 3.4433×, 88.764 s (24.878 / 32.633 / 31.253). Promoted the 512 configuration. - End 02:17:26 UTC./tmp/r43/time/time.json: 88.092 s (24.905 / 32.318 / 30.869), 3.4696×. Quick PASS, all six precomputed and six internally shaded forward fixtures bitwise, and all 32 deterministic gradient fixtures bitwise. Gains concentrate in the two mature windows (0.160 and 0.340 s). Commit1847646; full check queued, PID 62963 (/tmp/r43/check.log).
7.5 hRound 44 — off-screen covariance work
- Official profile of R43 at steps 10517–10536, then an untimed count of points provably offscreen before constructing their world covariance, PID 62962 (/tmp/r44/profile.log,/tmp/r44/project-bound.json,/tmp/r44/bound.log). - GPU-lock start 02:17:27 UTC. Replay 24.20%, Adam 23.89%, packet forward 18.79%, forward projection 6.27%. The conservative norm bound rejects 34–59% of points passing the near plane, but *only* 0.07–1.80% of live warps become empty. Instead of adding another visibility test, keep computed world covariance in six registers through the existing exact radius/rectangle test. Store it for backward only for a retained rectangle. Precomputed matrices remain unchanged. Time PID 63298 (/tmp/r44/time.log), quick and exact raw/activated/precomputed geometry/gradient probes PID 63531 (/tmp/r44/quick.log,/tmp/r44/cov-*.log,/tmp/r44/gradient-probe.log). - R43 full check/tmp/r43/check/check.jsonfinished 02:22:51 UTC: PASS N1, 3.4703×, 88.073 s (24.775 / 32.284 / 31.014). All four endpoints were within 1.34σ. Promoted. - End 02:25:22 UTC./tmp/r44/time/time.json: 88.086 s (25.006 / 32.262 / 30.818), 3.4698×; quick PASS, all 18 forward fixtures exact (including live covariance bits in computed branches), all 32 deterministic adjoint fixtures exact. Commit282a8a. Excluded: 0.006 s below the R43 screen but slightly slower than its full timing; not resolved from noise.
7.7 hRound 45 — reusable packet bounds
- Official profile of verified R43 at steps 17171–17188, PID 64642 (/tmp/r45/profile.log). - GPU-lock start 02:25:23 UTC. Adam 24.83%, replay 22.33%, packet forward 20.29%, projection 6.44%. Compute the existing conservative floating box once per Gaussian in projection and encode the inclusive 2×4 packet-coordinate intervals in the spare half of a widened center vector. The raster loader uses integer mask intersections instead of redoing log/square-root and testing 12 intervals for every tile instance. Unrepresentable grids or nonfinite bounds take the ordinary formula. Offline interval comparisons matched all 1.8 million tested axis/packet pairs. Time PID 64806 (/tmp/r45/time.log); quick, thin/offscreen forward and all-degree gradient probes PID 65026 (/tmp/r45/quick.log,/tmp/r45/packet-*.log,/tmp/r45/gradient-probe.log). - End 02:32:07 UTC./tmp/r45/time/time.json: 88.353 s (24.895 / 32.388 / 31.071), 3.4593×, 0.261 s slower than R43. Quick PASS, but the exact raw offscreen fixture failed its zero-image-error test after matching counts, radii and last contributions. The probe chain stopped before further gradient tests. Excluded commit2ba3b36; an untimed absolute-error probe is queued (PID 65695,/tmp/r45/rounding.log).
7.8 hRound 46 — materialization of raster adjoints
- Official profile of verified R43 at steps 15269–15288, PID 65692 (/tmp/r46/profile.log). - GPU-lock start 02:32:07 UTC. Adam 24.57%, replay 23.21%, forward blend 19.46%, fused projection adjoint 4.42%, fills 2.45%. Autograd can tell whether it needs the six-component covariance output. When it does not, consume the six derivatives locally in the fused projection and do not clear/store their dense result; do not allocate the private inverse-depth accumulator in image-only backward. Explicit extension users still receive the complete covariance derivative by default. Time PID 65907 (/tmp/r46/time.log); quick/full-vs-omitted tensor fixtures PID 66111 (/tmp/r46/quick.log,/tmp/r46/gradient-*.log); high-level autograd fixtures PID 66378 (/tmp/r46/autograd.log). R45's follow-up found maxima of 5.67e-7 in RGB and 1.79e-6 in inverse depth on the ordinary geometry, plus differing last contributors on the thin fixture. - End 02:38:55 UTC./tmp/r46/time/time.json: 87.367 s (24.892 / 31.926 / 30.550), 3.4984×, 0.725 s faster than R43./tmp/r46/quick/check.jsonPASS. Both native-output modes matched in 32 deterministic AA/degree/adjoint cases apiece, and all four high-level covariance-requires-grad/depth-gradient combinations matched bitwise. Commite4a8567, full check queued PID 66910 (/tmp/r46/check.log).
7.9 hRound 47 — conic atomics
- Official profile of screened R46 at steps 12617–12634, PID 66909 (/tmp/r47/profile.log). - GPU-lock start 02:38:57 UTC. Replay 24.52%, Adam 22.95%, forward blend 21.43%. Replace the three scalar conic-component adds and touch exchange with one SM90 float4 atomic on the already aligned conic row. Its private third component now counts touching buckets; only its existing nonzero predicate is consumed. Pre-SM90 retains the scalar path. Time PID 67167 (/tmp/r47/time.log); quick, same-tile/multi-tile gradients and SASS inspection PID 67371 (/tmp/r47/quick.log,/tmp/r47/gradient-*.log,/tmp/r47/atomics.txt). - R46 full check/tmp/r46/check/check.jsonfinished 02:43:54 UTC: PASS N1, 3.4951×, 87.448 s (24.894 / 31.847 / 30.706). All endpoints within 1.33σ of reference. Promoted and verified tree/config equivalence with the screened version. - End 02:47:52 UTC./tmp/r47/time/time.json: 86.565 s (24.498 / 31.705 / 30.363), 3.5308×, 0.802 s faster than R46 and all three windows improved. Quick PASS, 32 deterministic gradient cases bitwise and 32 multi-tile cases within tolerance. SASS contains nativeREDG.E.ADD.F32x4, rather than four scalar operations. Commited4b1b1; full check queued PID 68446 (/tmp/r47/check.log).
8.0 hRound 48 — color/opacity atomics
- Official profile of screened R47 at steps 13739–13754, PID 68445 (/tmp/r48/profile.log). - GPU-lock start 02:47:53 UTC. Adam 24.46%, replay 23.27%, forward blend 19.73%. Put incoming RGB and opacity adjoints in a shared 16-byte row. Use a native four-component atomic instead of four scalar adds, consume the color directly at stride four in SH, and apply the opacity derivative into its public contiguous output. Direct extension calls continue to return contiguous RGB by default; autograd can use a strided view for an unrequested output. Time PID 68857 (/tmp/r48/time.log); quick plus contiguous/view, multi-tile and public-autograd probes PID 69217 (/tmp/r48/quick.log,/tmp/r48/gradient-*.log,/tmp/r48/autograd.log). - R47 full check/tmp/r47/check/check.jsonfinished 02:52:53 UTC: PASS N1, 3.5233×, 86.748 s (24.258 / 31.898 / 30.592). All four endpoints within 0.47σ. Promoted. - End 03:00:41 UTC./tmp/r48/time/time.json: 87.075 s (24.338 / 32.055 / 30.683), 3.5101×. Quick PASS, contiguous and strided internal modes exact on all 32 deterministic fixtures each, all multi-tile and public-gradient fixtures passed. Excluded commit2819954: although the first window improved, the new widened gradient storage and conversion regressed both mature windows and total time by 0.510 s.
8.2 hRound 49 — early growth phase
- Official profile of verified R47 at steps 2453–2480, PID 70043 (/tmp/r49/profile.log). - GPU-lock start 03:00:42 UTC. In this early phase replay leads at 31.66%, Adam 18.16%, packet forward 17.15%. Measured hit-pattern diagnostic PID 70175 (/tmp/r49/chunks.json): relative to the existing 12-entry replay chunk, capacity 16 reduced expected chunk rounds by 16.5% and splat loop iterations by 13.2%; capacity 24 reduced them by 31.2% and 26.6%. Sweep both while retaining full FP32 storage, order within each pixel, and all hit coefficients. Timing PID 70344 (/tmp/r49/time-{16,24}.log); quick and synthetic gradient cases PID 70548 (/tmp/r49/quick-{16,24}.log,/tmp/r49/gradient-*.log). Supplemental mature hit histogram PID 71776 (/tmp/r49/mature-chunks.json). - End 03:20:14 UTC. Capacity 16/tmp/r49/time-16/time.json: 86.519 s (24.120 / 31.848 / 30.551), 3.5327×. Capacity 24/tmp/r49/time-24/time.json: 88.425 s (24.544 / 32.688 / 31.194), 3.4565×. Both quick PASS; both matched all single-tile gradients bitwise and passed all multi-tile fixtures. Late sampled groups had lower occupancy and fewer potential chunk savings. Commit24c484eexcluded: 16 saved only 0.046 s overall with worse mature windows; 24 was uniformly slower.
8.7 hRound 50 — dense raster gradient initialization
- Official mature profile of verified R47 at steps 16273–16288, PID 72144 (/tmp/r50/profile.log). - GPU-lock start 03:20:15 UTC. Adam 25.77%, replay 23.82%, packet forward 17.90%, fills 2.09%. Eightaten::zeroscalls per step account for 824 µs over the profile. Use one CUDA byte-clear on an arena sized for exactly the requested auxiliary outputs and give each its own aligned contiguous float32 view. Keep SH's separate tiled initialization. Time PID 72348 (/tmp/r50/time.log); quick, full/optional gradient-output and autograd probes PID 72663 (/tmp/r50/quick.log,/tmp/r50/gradient-*.log,/tmp/r50/autograd.log). - End 03:37:49 UTC./tmp/r50/time/time.json: 86.595 s (24.348 / 31.910 / 30.337), 3.5296×, 5.08 GiB peak allocated and 41.19 GiB reserved. Quick PASS; 64 deterministic gradient cases and all four public autograd cases exact. Excluded commite2c8559: 0.030 s slower than R47's screen, not a resolved difference.
9.0 hRound 51 — native RGB gradient pairs
- Official growth-phase profile of verified R47 at steps 14451–14468, PID 73317 (/tmp/r51/profile.log). - GPU-lock start 03:37:50 UTC. Kernel shares: Adam 24.45%, replay 23.16%, forward 19.98%. Use the native aligned pair *within* the existing contiguous three-float RGB row: red/green for even points, green/blue for odd, one scalar add for the remaining channel. No extra scratch or change to any output representation. Detached time, quick and gradient fixture sequence PID 73494 (/tmp/r51/jobs.log,/tmp/r51/time.log,/tmp/r51/quick.log,/tmp/r51/gradient-*.log). - End 03:49:53 UTC./tmp/r51/time/time.json: 86.505 s (24.410 / 31.734 / 30.361), 3.5332×. Quick PASS, 32 deterministic and 32 multi-tile gradient cases passed. Excluded107ebbebecause the 0.060 s difference is below timing resolution and neither mature window clearly improved.
9.1 hRound 52 — host raster setup
- Official profile of verified R47 at steps 7701–7724, PID 74309 (/tmp/r52/profile.log). - GPU-lock start 03:49:53 UTC. Replay 24.16%, Adam 23.73%, packet forward 18.85%; 48 full-frame device copies took 318.6 µs (0.46%) and the host made 96 synchronous-copy API calls including the count reads. CUB occupancy/device-attribute queries accounted for only 88 µs of host time. Test writing RGB and inverse depth to their public and private replay destinations directly at the end of packet shading, preserving independent buffers. Detached time/quick/forward and public gradient sequence PID 74481 (/tmp/r52/jobs.log,/tmp/r52/time.log,/tmp/r52/quick.log,/tmp/r52/frame.log,/tmp/r52/autograd.log). - End 03:56:17 UTC./tmp/r52/time/time.json: 86.695 s (24.571 / 31.749 / 30.375), 3.5255×. Quick PASS, all six forward and all four actual-state public gradient fixtures exact. Direct dual writes did not improve any window over R47's screen. Excluded4a1babe.
9.2 hRound 53 — dense Adam square root
- Official mature profile of verified R47 at steps 19625–19644, PID 75322 (/tmp/r53/profile.log). - GPU-lock start 03:56:18 UTC. Dense Adam 24.93%, replay 22.17%, packet forward 19.88%. Test non-FTZ FP32 native approximate square root of the second moment. Keep the dense update, moment interpolation, bias corrections, epsilon and division in the same order; no skipped parameters. Sequence PID 75481 (/tmp/r53/time.log,/tmp/r53/quick.log,/tmp/r53/adam-probe.log); probe includes zero/tiny variances, zero gradients and both alignment cases. - End 04:02:36 UTC./tmp/r53/time/time.json: 87.291 s (24.767 / 31.962 / 30.563), 3.5014×. Quick PASS; both moments matched bitwise and maximum relative parameter L2 was 4.17e-8 on the stress probe. All three timing windows regressed. Excluded939cff6; reducing square-root precision is not useful here.
9.3 hRound 54 — occupancy of replay coefficient storage
- Official profile of verified R47 at steps 8781–8800, then six real hit-mask samples measuring whether ragged lists fit the current per-warp capacity, PID 76295 (/tmp/r54/profile.log,/tmp/r54/ragged.json,/tmp/r54/ragged.log). - GPU-lock start 04:02:36 UTC. Adam 22.24%, replay 21.71%, forward blend 16.68%. Across the six views 48–67% of stripes with a list longer than 12 could still fit *all* hits in 384 entries, avoiding 12.5–14.4% of rounds. For just these RGB stripes, use a warp prefix sum to pack variable-length lists in the existing coefficient table and carry offsets in the unused gradient W field. No additional shared storage. Time/quick/gradient sequence PID 76547 (/tmp/r54/time.log,/tmp/r54/quick.log,/tmp/r54/gradient-*.log). - End 04:21:16 UTC./tmp/r54/time/time.json: 88.202 s (25.127 / 32.233 / 30.842), 3.4653×. Quick PASS and all deterministic and multi-tile numerical fixtures passed. Dynamic list indexing and the prefix sum lost time across all windows despite fewer theoretical rounds; excludedfd48325.
9.6 hRound 55 — requested raster outputs
- Official profile of verified R47 around step 16353, PID 77633 (/tmp/r55/profile.log). - GPU-lock start 04:21:17 UTC. Dense Adam 24.93%, replay 23.15%, packet rendering 19.91%. For training views with no depth supervision, request just the RGB, radii and screen outputs from the renderer. A specialized loop skips the unused depth channel, depth checkpoints and final depth copy; the public default and all depth-supervised renders retain the original depth calculation. Time, quick and actual-autograd numerical sequence PID 77817 (/tmp/r55/time.log,/tmp/r55/quick.log,/tmp/r55/depth.log). -/tmp/r55/time/time.json: 86.390 s (24.261 / 31.735 / 30.393), 3.5379×. Quick PASS. All 32 requested/not-requested depth autograd fixtures had exact requested forward outputs, exact single-tile gradients, and maximum multi-tile gradient relative L2 of 5.84e-7. Peak allocated dropped from 5.02 to 4.96 GiB but reserved rose from 42.72 to 62.44 GiB. The small improvement was confined to the first window, so repeat before promotion (PID 78707,/tmp/r55/repeat.log). - End 04:34:43 UTC./tmp/r55/repeat/time.json: 86.832 s (24.519 / 31.816 / 30.497), 3.5199×. Neither mature window improved on either run and the initial all-run gain did not reproduce. Excluded01eb62f.
9.8 hRound 56 — image mask processing
- Official growth-phase profile of verified R47 at steps 7183–7204, PID 78891 (/tmp/r56/profile.log). - GPU-lock start 04:34:44 UTC. Replay 23.53%, Adam 21.13%, packet forward 16.87%. The unconditional alpha multiplication and its adjoint each took about 6.8 µs/step. Skip both for implicit all-one RGB masks; keep the original mask and check its identity, version and requires-grad status so changing it resumes regular multiplication. Preserve RGBA and train/test exposure masks. Initial launch PID 79051 failed before taking the GPU lock because the snapshot was missing (/tmp/r56/time.log); corrected time, quick and mask-fixture sequence PID 79091 (/tmp/r56/time-valid.log,/tmp/r56/quick.log,/tmp/r56/mask.log). - End 04:42:16 UTC./tmp/r56/time-valid/time.json: 86.917 s (24.818 / 31.678 / 30.421), 3.5165×. Quick PASS and seven RGB/RGBA/exposure/mutation/gradient fixtures bitwise. No end-to-end benefit compared with 86.565 s on R47; excluded49231eb.
9.9 hRound 57 — replay checkpoint writes
- Profile and six real-frame prefix-usage counts at steps 16743+, PID 79930 (/tmp/r57/profile.log,/tmp/r57/coverage.json,/tmp/r57/coverage.log). - GPU-lock start 04:42:16 UTC. Adam 25.99%, replay 23.17%, packet forward 17.92%. In the six sampled mature views, 15.75–29.89% of live pixel-buckets had zero hit masks; their five float prefix fields are never read by masked replay. Save the prefixes in registers during a bucket and only emit them when that pixel accepted a Gaussian. Keep every live hit mask, all final outputs and all consumed checkpoints. Time/quick/forward and actual backward fixtures PID 80239 (/tmp/r57/time.log,/tmp/r57/quick.log,/tmp/r57/prefix.log,/tmp/r57/autograd.log).