From 0c0c254ccf656ca34771dea937a0e44a36acff1f Mon Sep 17 00:00:00 2001 From: Meta Vens Date: Sat, 3 Oct 2026 13:40:42 +0200 Subject: [PATCH] Research: prior art for measuring culling efficiency against an oracle Eyes pass for cultmath-tapes:question:wells-ceiling-oracle. Co-Authored-By: Claude Opus 5.5 --- .../culling-efficiency-oracle-prior-art.md | 305 ++++++++++++++++++ 1 file changed, 305 insertions(+) create mode 100644 docs/research/culling-efficiency-oracle-prior-art.md diff --git a/docs/research/culling-efficiency-oracle-prior-art.md b/docs/research/culling-efficiency-oracle-prior-art.md new file mode 100644 index 00000000..cf0d8077 --- /dev/null +++ b/docs/research/culling-efficiency-oracle-prior-art.md @@ -0,0 +1,305 @@ +# Culling Efficiency Against an Oracle: Prior Art Survey + +This is the facts file for Eureka question `cultmath-tapes:question:wells-ceiling-oracle`, +which interprets ruling `cultmath-tapes:ruling:operator-wells-contract-vs-ceiling`. That +ruling holds scenario (c) to "at least 90% of the probe-free ceiling (the cost of the +dense cells alone)". The question asks what "ceiling" means: + +- (a) the pre-pass's own mask, under which a pre-pass that skips nothing scores 0.98; +- (b) an oracle mask, the cells where at least one ray samples nonzero density. + +Under (b), Imagination measured 0.38 at warp 0 and 0.19 at warp D. + +This survey covers how published work measures culling or skipping efficiency against +ground truth, and what efficiency figures conservative bounds reach. Eyes pass, +2026-10-03. It gives facts and evidence pointers only, with no recommendation. + +Verification marks: + +- **[fetched]** The claim was read in the cited source's full text during this pass. + PDFs were converted with `pdftotext -layout`. +- **[search]** The claim came from a search-result summary, not a full read. +- **[not found]** The source was read, and the named kind of measurement is absent from it. + +Reference kinds used below: + +- **oracle**: an exact or ground-truth answer computed independently of the method, + such as the exact visible set or the set of truly non-empty cells. +- **self**: the method's own result, or a baseline that does no culling. +- **peer**: another approximate method. + +## 1. Visibility culling: conservative versus exact visible set + +### 1.1 Survey vocabulary + +- **[fetched]** Cohen-Or, Chrysanthou, Silva, Durand, *A Survey of Visibility for + Walkthrough Applications* (IEEE TVCG 2003). Section 2 lists "Tightness of + approximation" as a review criterion. It notes that few papers report the ratio + between the size of the potentially visible set and the size of the visible set, + and asks later work to report it. Most surveyed techniques are conservative, which + means they overestimate the visible set. + https://www.cs.tau.ac.il/~dcor/online_papers/papers/visibility-survey-ieee.pdf +- **[fetched]** Bittner, *Hierarchical Techniques for Visibility Computations* (PhD + thesis, CTU Prague, 2002), chapter 2, equations 2.1 to 2.4. It defines a quality + measure against the exact PVS `S_E`: + - relative overestimation `Q_o = |S_A \ S_E| / |S_E|`; + - relative underestimation `Q_u = |S_E \ S_A| / |S_E|`; + - the expected quality, which is the expectation of these over inputs. + + The same text appears in Bittner and Wonka, *Visibility in Computer Graphics* + (section 4.3). + https://dcgi.fel.cvut.cz/home/bittner/publications/diss_noimages.pdf + http://www.casa.ucl.ac.uk/mike-michigan-april1/mike's%20stuff/attach/bittner_wonka.pdf +- **[fetched]** Nirenstein, Blake, Gain, *Exact From-Region Visibility Culling* (EGWR + 2002), introduction and Table 1. The paper classifies algorithms as exact, + conservative, aggressive or approximate. Conservative methods "incur a false + visibility error". The paper states: + - exact visibility "provides a benchmark for comparing other approaches"; + - the percentage of polygons culled is "a very crude and imperfect metric", because + it ignores the type, distribution and magnitude of visibility error. + + Reported results (Table 2) are culling fractions of the whole scene: 99.45% for the + town and 99.12% for the forest. No conservative method is measured against the + exact set in this paper **[not found]**. A search summary attributed a 95.5% figure + for extended projection to this paper, but that figure is not in the full text. + https://people.cs.uct.ac.za/~jgain/wp-content/papercite-data/pdf/nirenstein2002.pdf + +### 1.2 Measured conservative-versus-exact ratios + +- **[fetched]** Bittner, Wonka, Wimmer, *Fast Exact From-Region Visibility in Urban + Scenes* (EGSR 2005), section 6. The exact 2.5D algorithm is used as the oracle, and + the paper simulates conservative methods against it: + - **Section 6.2, one penumbra wedge per occluder.** The conservative PVS was more + than 47% larger than the exact PVS for 2 m view cells and more than 76% larger + for 4 m view cells. As exact/conservative ratios these are ≤ 0.68 and ≤ 0.57. + - **Section 6.3, object-space discretization** (Wonka et al.'s method, a 1400×1400 + grid, Vienna). The PVS overestimate averaged 17%, a ratio of about 0.85. + - **Section 6.3, Table 4, line-space discretization** (Leyvand et al.). PVS sizes by + grid resolution, with the exact PVS last: + + | Scene | 512 | 1024 | 2048 | 16384 | exact | + |---------|--------|--------|--------|-------|-------| + | Vienna | 633 | 535 | 489 | 455 | 235 | + | Atlanta | 50,744 | 25,751 | 10,082 | 2,656 | 1,152 | + + The ratios exact/conservative below are computed from that table, not quoted from + the paper: + - Vienna: 0.37, 0.44, 0.48, 0.52; + - Atlanta: 0.023, 0.045, 0.11, 0.43. + + The paper says the overestimation comes from conservative shrinking and from gaps + "between formerly connected occluders". + + https://peterwonka.net/Publications/pdfs/2005.EGSR.Bittner.FastExactFromRegionVisibility.pdf + +### 1.3 Run-time occlusion culling against an ideal renderer + +- **[fetched]** Mattausch, Bittner, Wimmer, *CHC++: Coherent Hierarchical Culling + Revisited* (Eurographics 2008), section 7 (Figure 7) and the conclusion. The + reference algorithm "OPT" renders only the visible nodes of the hierarchy and + issues no queries. It was implemented by recording visibility with an exact + stop-and-wait pass, so it is an oracle at node granularity that is independent of + the method under test. CHC++ is reported to be "for most cases within a few percent + of the 'ideal' method". + + Caveat: occlusion queries are exact per-node tests, not conservative range bounds. + The gap to OPT measures query overhead and latency, not bound looseness. + + The paper also separates out the "optimal" algorithm defined by Guthe et al. + (NOHC-OPT), and shows CHC++ "clearly below" that optimum. That earlier optimum was a + cost-model optimum, not an oracle. + https://dcgi.fel.cvut.cz/home/bittner/publications/chc++.pdf +- **[fetched]** Zhang et al., *Visibility Culling using Hierarchical Occlusion Maps* + (SIGGRAPH 1997). It reports "% culling" of the total model, and frame time + normalized to rendering without HOM. Both are **self** references. No comparison + with the exact visible set **[not found]**. + https://www.cs.princeton.edu/courses/archive/spr01/cs598b/papers/zhang97.pdf +- **[search]** Wonka et al., *Guided Visibility Sampling* (SIGGRAPH 2006). This is + an aggressive, sampling-based PVS method. Its Table 2 reports error as false pixels + on a 1000×1000 screen, together with PVS size. Section 4.5 compares it to Bittner's + exact algorithm on one scene, because exact solvers do not run on the larger test + scenes. The text was extracted in this pass, but not read in full. + https://peterwonka.net/Publications/pdfs/2006.SG.Wonka.GuidedVisibilitySampling.final.pdf + +## 2. Empty-space skipping in volume rendering + +- **[fetched]** Li, Mueller, Kaufman, *Empty Space Skipping and Occlusion Clipping for + Texture-based Volume Rendering* (IEEE Vis 2003): + - **Section 5, Table 1.** Empty-space skipping is measured as a frame-rate speedup + over no skipping, 2.8× on average. This is a **self** reference. + - **Section 5, Table 2.** It reports the percentage of non-empty voxels skipped + through occlusion, 48% to 64% with a 56.2% average. The denominator is all + non-empty voxels, not the truly occluded voxels. The authors write that "there + are still many occluded voxels that have not been skipped" and call the number + "smaller than we have expected". The truly occluded set is not computed + **[not found]**. + - **Section 5.** Overhead is reported separately. The opacity map and clipping cost + about 10% of the empty-space-skipping render time, measured with a mode that pays + the overhead but skips nothing ("ESS+M"). + + https://www3.cs.stonybrook.edu/~mueller/papers/OcclusionClipping.pdf +- **[fetched]** Zellmann, *Comparing Hierarchical Data Structures for Sparse Volume + Rendering with Empty Space Skipping* (arXiv 1912.09596, 2019). It reports frame + rates and build times, and compares only against naive ray marching and simple + macrocell grids, as the paper itself states. No step count or sample count against + an ideal traversal **[not found]**. + https://arxiv.org/abs/1912.09596 +- **[search]** Hadwiger et al., *SparseLeap* (IEEE TVCG 24(1), 2018). It builds + per-pixel lists of ray segments by rasterizing occupancy geometry. The full text + was not retrieved (the repository returned HTML), so it is unknown whether it + reports skip efficiency against the true empty space. + https://vccvisualization.org/research/sparseleap/ +- **[search]** Deakin and Knackstedt, *Efficient ray casting of volumetric images using + distance maps for empty space skipping* (Computational Visual Media 6(1), 2020). The + full text was not retrieved. + +## 3. Interval and affine arithmetic for implicits, noise and SDF marching + +- **[fetched]** Keeter, *Massively Parallel Rendering of Complex Closed-Form Implicit + Surfaces* (SIGGRAPH 2020): + - **Section 3, Interval arithmetic.** Ambiguous regions "either contain the shape's + boundary or are a false positive (due to interval arithmetic's conservative + behavior)". The false-positive share is not measured **[not found]**. + - **Section 2.1, Figure 5.** Reported efficiency is against **self** and peer + baselines: a brute-force per-pixel interpreter (about 58 ms/MP) and a compiled + kernel (about 3 ms/MP). + - **Section 3.2, Figure 6.** Tape length is reduced 17× in 64² tiles and 216× in 8² + subtiles, from 6056 clauses to means of 356 and 28. + - **Section 5, Figures 9 and 10.** Heatmaps of work per pixel show extra work at + model edges, where interval evaluation "can't skip entire regions". The text says + many smooth blends lead "to less useful interval evaluation results". + - **Section 6.** It names reduced affine arithmetic as a way "to more tightly bound + intervals". + + https://www.mattkeeter.com/research/mpr/keeter_mpr20.pdf +- **[fetched]** Keeter, *Gradients are the new intervals* (blog, 2025-05-14). It + compares interval evaluation with gradient-based pseudo-intervals by tape length + (120.2 against 121.5 clauses) and by render time. These are peer references. There + are no tile-classification counts and no ground truth **[not found]**. + https://www.mattkeeter.com/blog/2025-05-14-gradients/ +- **[fetched]** Heidrich, Slusallek, Seidel, *Sampling Procedural Shaders Using Affine + Arithmetic* (ACM TOG 17(3), 1998): + - **What "error" measures.** Error is the width of the computed range per pixel. It + is not compared with the true range from dense sampling **[not found]**. The + tightness against the true variation is argued qualitatively, from shader cross + sections. + - **Section 3.2.** Noise over an affine form that spans several lattice cells falls + back to interval arithmetic, and a span of more than two cells returns the full + noise range [0, 1]. For turbulence (blue marble), 81% of pixels lie within 5% + error and 59% within 1%. The eroded shader, with high-frequency noise, gets 77% + and 53%. + - **Section 4.** Against AA, interval arithmetic gave errors "up to 50 percent where + affine arithmetic produces errors below 1/256" (wood shader, Figures 6 and 7). + + https://vccimaging.org/Publications/Heidrich1998SPS/Heidrich1998SPS.pdf +- **[fetched]** Gamito and Maddock, *Ray casting implicit fractal surfaces with reduced + affine arithmetic* (The Visual Computer 23(3), 2007). Section 5, Tables 1 to 3, use + a sphere hypertextured with three octaves of noise. The accuracy proxy is the + average number of interval evaluations per ray, a **peer** comparison with no oracle + count. The rows below are reconstructed from a mangled text layout, and the + reconstruction agrees with the paper's prose. + + | Noise | IA evals/ray | IA time | Standard AA evals/ray | Reduced AA evals/ray | Reduced AA + interval optimisation, evals/ray | Its time | + |------------------------|--------------|-----------|-----------------------|----------------------|-----------------------------------------------|-----------| + | Perlin gradient | 78.40 | 8m27.5s | 34.60 | 34.60 | 20.81 | 3m43.2s | + | Lewis sparse convolution | 48.44 | 26m31.0s | 23.25 | 23.25 | 13.46 | 6m29.3s | + | Worley cellular | 45.70 | 25m54.7s | 32.56 | 43.78 | 21.29 | 14m25.9s | + + The text cites "the excessive conservativeness of IA estimates", and says IA + compensates by being fast. + https://eprints.whiterose.ac.uk/id/eprint/3762/1/gamito_newraycastvisual.pdf +- **[fetched]** Knoll, Hijazi, Kensler, Schott, Hansen, Hagen, *Fast Ray Tracing of + Arbitrary Implicit Surfaces with Interval and Affine Arithmetic* (CGF 28(1), 2009). + Section 6.1.1 and Tables 1 and 2 report frames per second for IA against reduced + AA: + - RAA is "generally 1.5 to 2× faster", and "3 to 4 times faster" for functions + "with high bound overestimation"; + - IA is faster for superquadrics. + + These are **peer** references, with no oracle count of wasted subdivisions + **[not found]**. + https://sci.utah.edu/~knolla/cgrtia.pdf +- **[fetched]** Sharp and Jacobson, *Spelunking the Deep: Guaranteed Queries on General + Neural Implicit Surfaces via Range Analysis* (SIGGRAPH 2022), section 3.4 and Table 1. + Tightness is measured as "the typical size of a region which can be bounded away + from the zero": + + | Method | 1D length | 3D volume | Ray-cast time (relative to the fastest) | + |---------------|-----------|-----------------|-----------------------------------------| + | interval | 0.011 | < 0.001 × 10⁻³ | 34.3× | + | affine (full) | 0.821 | 3.499 × 10⁻³ | 8.4× | + | affine (fixed)| 0.306 | 0.267 × 10⁻³ | 1.0× | + + The prose says interval arithmetic "yields extremely pessimistic bounds". These are + **peer** references, with no oracle region size **[not found]**. + https://arxiv.org/abs/2202.02444 +- **[fetched]** Mitchell, *Robust Ray Intersection with Interval Arithmetic* (Graphics + Interface 1990). It reports root-finding timings for its method and for peer + root finders. No pruning-against-oracle measure **[not found]**. + http://graphicsinterface.org/wp-content/uploads/gi1990-8.pdf +- Duff, *Interval Arithmetic and Recursive Subdivision for Implicit Functions and + Constructive Solid Geometry* (SIGGRAPH 1992), and Snyder, *Interval Analysis for + Computer Graphics* (SIGGRAPH 1992), were not retrieved in this pass. + https://doi.org/10.1145/142920.134027 + +## 4. Conservative rasterization and tile classification + +- **[fetched]** Hasselgren, Akenine-Möller, Ohlsson, *Conservative Rasterization* + (GPU Gems 2, chapter 42): + - **Section 42.1.** It defines overestimated rasterization as every pixel whose cell + intersects the polygon, and underestimated rasterization as every pixel whose + cell lies wholly inside it. + - **Section 42.5.** Overdraw is discussed qualitatively: it depends on triangle + acuity and resolution. Figure 42-7 compares the two proposed algorithms only. No + count of excess fragments against the exact overlap set **[not found]**. + + https://developer.nvidia.com/gpugems/gpugems2/part-v-image-oriented-computing/chapter-42-conservative-rasterization +- **[search]** Antochi, Juurlink, Vassiliadis, Liuha, *Scene management models and + overlap tests for tile-based rendering* (DSD 2004), and the companion *Efficient + tile-aware bounding-box overlap test*. In a search summary, the bounding-box + triangle-to-tile test "can generate up to 30% false positives" on common workloads, + measured against exact overlap. The full text could not be retrieved (the TU Delft + host refused the connection, and CiteSeerX returned HTML). + https://citeseerx.ist.psu.edu/document?doi=5c78f05f185b9513bf930588dd2a4ae334090620 + +## 5. Summary of reference kinds and reported figures + +| Source | Reference | Figure reported | +|--------|-----------|-----------------| +| Bittner et al. 2005, §6.2 | oracle (exact PVS) | conservative PVS +47% / +76% (ratio ≤ 0.68 / ≤ 0.57) | +| Bittner et al. 2005, §6.3 | oracle | object discretization +17% (≈ 0.85); line-space 0.023–0.52 by scene and resolution | +| Bittner thesis, ch. 2 | oracle | defines `Q_o`, `Q_u` relative to `|S_E|` | +| Cohen-Or et al. 2003 | — | notes the PVS/visible-set ratio is rarely reported | +| Nirenstein et al. 2002 | oracle (its own exact) | percent culled; calls that metric crude | +| CHC++ 2008 | oracle (OPT, node level) | within a few percent of OPT (exact queries, not bounds) | +| HOM 1997 | self | % of model culled, normalized frame time | +| Li, Mueller, Kaufman 2003 | self / all non-empty | 2.8× speedup; 56% of non-empty voxels skipped; overhead ≈ 10% reported separately | +| Zellmann 2019 | self | fps against no skipping | +| Keeter 2020 | self (brute force, compiled) | 17× / 216× tape reduction; false-positive share not measured | +| Heidrich et al. 1998 | none (range width) | IA error up to 50% where AA < 1/256 | +| Gamito, Maddock 2007 | peer | IA 2.1–3.8× more evaluations per ray than reduced AA + optimisation on noise | +| Knoll et al. 2009 | peer | RAA 1.5–4× faster than IA | +| Sharp, Jacobson 2022 | peer | IA provable region about 75× smaller in 1D than full AA | +| GPU Gems 2, ch. 42 | peer | qualitative overdraw | +| Antochi et al. 2004 [search] | oracle (exact overlap) | bounding box up to 30% false positives | + +On the interval side (section 3), none of the sources read in this pass reports a +pruning or culling fraction against an oracle mask, such as the truly empty cells or +the truly non-ambiguous tiles. They report only against self or peer baselines. The +oracle-referenced ratios found are all on the visibility side (sections 1.2 and 1.3), +and the one from rasterization (section 4) is search-only. No published figure was +found for interval-arithmetic culling of noise or warped-noise fields as a fraction of +an oracle. + +## 6. Local context (from the eureka mind, not re-measured) + +- **Question** `cultmath-tapes:question:wells-ceiling-oracle` (admitted 2026-10-03, + Imagination). At 3ac7250a it gives: + - warp 0: oracle efficiency 0.38, mask ceiling 0.96, cull fraction 0.95; + - warp D = 60: oracle efficiency 0.19, mask ceiling 0.99, cull fraction 0.84; + - a pre-pass that skips nothing: 0.14 against the oracle. +- **Finding** `cultmath-tapes:finding:cut-interval-ops.s2.wells-ceiling-contract-vacuous` + (Soul, confirmed by mutation). The mask-referenced (c) check reports 0.98 with the + wells skip disabled. Locations are in + `packages/cultmath/tests/CultMath.Tests/NoiseBoundTests.cs`: lines 1257–1260, 1437 + and 1461.