{"id":"3bf9b924-6bb0-40b6-a68a-3bdc8a52422f","arxiv_id":"2506.17770","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"New wave-intrinsic algorithms (List Merge, Box Sampling, Mask Sampling) coordinate texel production across shader lanes, achieving perfect bilinear and bicubic filtering under magnification with at most one texel evaluation per pixel for sufficiently large magnification.","lead":"This paper presents new GPU algorithms that let nearby pixels share texture-decoding work using wave-level communication, so magnified textures can often be filtered with zero error using at most one texture sample per pixel. It matters for real-time rendering with expensive compressed textures, where stochastic sampling previously caused noise or flicker.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1.59/2.35 zero-error thresholds rest on an unproved worst-case assumption about 8x4 waves; phase offsets, anisotropic Jacobians, and perspective gradients can inflate the unique-texel union beyond 32 lanes, so the stated guarantee needs scoping or a proof.","rationale":"The paper is honest, well-scoped in its experimental claims, and the core idea - distributing unique texel production across wave lanes to enable exact filtering under magnification - is sound and well supported by the measurements. The reader's weakest-assumption analysis correctly identifies spatial coherence of the 8x4 wave as a load-bearing premise. However, the more precise load-bearing condition is the combinatorial worst case: the union of all 2x2 bilinear footprints over a wave must fit in 32 texels. That condition is sensitive not only to active-lane coherence but also to the sub-texel phase of the pixel grid, the full Jacobian of the screen-to-texel mapping, and the variation of that Jacobian across the tile. The paper provides Figure 4 for one empirical scene but no formal derivation or exhaustive scan, so the introduction's 'guarantee' at 1.59x for all orientations is stronger than what is actually shown. This does not invalidate the method or its practical value; even if the exact threshold rises for some phase/anisotropy configurations, the existential claim 'for sufficiently large magnification' likely still holds. But the paper should either prove the worst-case bound for the stated thresholds or explicitly qualify the guarantee to locally uniform, fully active waves. A short exhaustive enumeration would settle the point cheaply and would materially strengthen the paper. Until that is done, a cautious acceptance should be conditional on scoping the guarantee.","tokens_in":22380,"tokens_out":21711,"duration_ms":222787,"concrete_test":"Write an offline brute-force checker for an 8x4 pixel tile: enumerate sub-texel phase offsets on a fine grid (e.g., 0.01 texel steps); for each phase, evaluate a family of 2x2 Jacobians including the uniform case whose largest singular value is 1/1.59, plus anisotropic cases with the same determinant but condition numbers 2, 4, and 8, and a coarse set of perspective-like Jacobians that vary across the tile. Compute the union of integer texel coordinates of the 32 bilinear 2x2 footprints, count the unique texels, and check whether Mask Sampling's 16x16 mask would overflow. Repeat with active-lane masks containing random holes to simulate silhouette coverage.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that with 32-lane 8x4 waves, zero-error bilinear filtering is guaranteed at magnification factors above about 1.59 (Mask Sampling/List Merge) or 2.35 (Box Sampling). The only support for these thresholds is Figure 4, an empirical plot from a single scene: a quad viewed head-on, uniformly scaled, and rotated only in the image plane. The number of unique texels required by a wave is not a function of magnification alone; it also depends on (a) the sub-texel phase offset of the 8x4 pixel grid, (b) the full 2x2 screen-to-texel Jacobian, including anisotropic scales and perspective shear, and (c) the active-lane mask at silhouettes. None of these are bounded in the statement of the guarantee, and the term 'magnification factor' is never formally defined. In anisotropic cases, a scalar magnification bound does not bound the larger singular value of the Jacobian, so the union of 2x2 footprints across the wave can exceed 32 lanes even when the texture is nominally magnified. The paper acknowledges fallbacks at edges (Section S2) and perspective-induced AABB growth for Box Sampling, but the abstract and Section 1 present zero-error as essentially unconditional for sufficiently large magnification. The load-bearing condition is therefore that the union of bilinear footprints over an 8x4 wave fits into 32 texels; this is asserted for the quad scene but not proved or shown to be worst-case over phases, Jacobians, and active-lane patterns.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces collaborative texture filtering (CTF), a family of wave-intrinsic algorithms for magnified texture filtering. The three proposed methods—List Merge, Box Sampling, and Mask Sampling—coordinate the lanes of an 8x4 GPU wave so that each lane produces a unique texel and then gathers the 2x2 (or 4x4) texels needed for exact bilinear or bicubic filtering. When the union of required texels fits in the active lanes, filtering is exact with at most one texel evaluation per lane; otherwise, the system falls back to novel stochastic methods that the authors show improve on One-tap STF and Wave Communication STF. The paper reports perfect bilinear filtering for magnification factors above 1.59 (Mask Sampling/List Merge) or 2.35 (Box Sampling) on a head-on rotated quad, evaluates PSNR/FLIP/ColorVideoVDP quality and RTX 5090 runtimes, and demonstrates extensions to Catmull-Rom filters and to distributing expensive decompression work across lanes.","tokens_in":22679,"tokens_out":8506,"duration_ms":92020,"significance":"If the threshold claims hold, this is a significant practical advance: it would give exact magnified filtering for expensive texture formats at no more than one texel evaluation per pixel, eliminating the noise and flicker of stochastic texture filtering while retaining its cost advantage. The core algorithmic mechanism is constructive and sound under a clearly stated condition: when the wave-wide unique-texel set fits in the active lanes, every required texel is produced and gathered. The supplemental pseudocode is concrete and reproducible, the evaluation is careful and transparent about metrics and hardware, and the fallback methods are a useful contribution in their own right. The main weakness is that the generality of the 1.59/2.35 thresholds is asserted more strongly than the evidence supports.","major_comments":[{"comment":"The abstract and Section 1 state that zero-error bilinear filtering is guaranteed for magnification factors of at least 1.59 (Mask Sampling/List Merge) or 2.35 (Box Sampling) with 32-lane waves, but Figure 4 only measures a single scene: a quad viewed head-on, uniformly scaled, and rotated in the image plane. The number of unique texels required by an 8x4 wave is a function of the full screen-to-texel Jacobian (including perspective shear and anisotropy), the sub-texel phase offset of the wave, and the active-lane mask at silhouettes. None of these are bounded in the statement of the guarantee, and 'magnification factor' is never formally defined. Please define the term precisely (e.g., as a bound on the singular values of the texel-to-screen Jacobian) and either prove a worst-case upper bound on the unique-texel count for 8x4 waves under that condition, or explicitly restrict the guarantee to locally isotropic, full-wave magnification. As written, the headline claim is stronger than the evidence in Figure 4 supports, and the fallback paths described in Section 3.2 and Section S2 show that the method itself is conditional by design.","section":"Section 1 and Section 4.1, Fig. 4"},{"comment":"The statement 'We have verified experimentally that for a screen-aligned quad with magnification factor 1.0, the bilinear filter requires no more than 54 unique texels ... Hence, ≤2 texel evaluations are sufficient to always succeed' uses the word 'always' on the basis of a single configuration family (a rotated quad viewed head-on). This is load-bearing for the paper's claim that simple extensions achieve perfect filtering at magnification 1.0. To support this, the authors should either provide a proof over phases and Jacobians or report a broader parameter sweep that includes perspective, anisotropic scaling, and partial waves. If such a sweep is not possible within scope, the sentence should be rephrased as an experimental observation for the tested configurations.","section":"Section 5, Discussion and Limitations"}],"minor_comments":[{"comment":"In the pseudocode for both Box Sampling (line 22) and Mask Sampling (line 41), the condition 'if (curLaneIdx <= activeTexelsNeeded)' should likely be 'curLaneIdx < activeTexelsNeeded'. When activeTexelsNeeded is less than 32, the current condition selects one extra lane and may invoke the bijective mapping with an index outside the set of set bits; the prose in Section 3.3 correctly says 'If i < n, where i is the current lane's index'.","section":"Supplement S1.1 and S1.2"},{"comment":"There is a typo: 'If the arean of the global bounding box' should read 'If the area of the global bounding box'.","section":"Section 3.2, fourth paragraph"},{"comment":"The term 'magnification factor' appears in Figure 4, Figures 6 and 7, and throughout Section 4.1 without a formal definition. Adding one sentence defining it would substantially improve precision, especially given Major Comment 1.","section":"Section 4.1, Fig. 4 caption and surrounding text"},{"comment":"The phrase 'occuring' in 'no more than 54 unique texels (occuring at, e.g., 30◦ rotation)' should be 'occurring'.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"This is a strong HPG paper with a sound conditional algorithmic core and a careful evaluation. My recommended revision is scoping: either prove the worst-case bound behind the 1.59/2.35 thresholds or qualify them explicitly as empirical results for locally isotropic magnification. I do not see grounds for rejection; the algorithms produce exact results whenever the stated lane-budget condition holds, and the fallback prevents incorrect rendering otherwise. The paper would be acceptable after the authors tighten the claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first practical way I've seen to get exact bilinear (and bicubic) texture filtering under magnification for formats that can't go through the fixed-function texture units, at roughly one texel decode per pixel. The wave-level work allocation is the real contribution, and it is genuinely new relative to WPAM25.\n\nThe paper gives three collaborative algorithms — List Merge, Box Sampling, Mask Sampling — plus a fallback estimator and edge remapping. The logic is constructive and clear: each lane reports its filter footprint, the wave builds the union, and if that union fits in 32 lanes, each lane produces one unique texel and gathers via wave intrinsics. When it doesn't fit, the fallback still beats prior STF methods. The evaluation is honest: multiple metrics, denoised and non-denoised, real NTC integration, and the performance numbers on the RTX 5090 are plausible. The Catmull-Rom handling with negative weights using a single tap is a nice bonus. No circularity; the comparison against ground-truth bilinear is external.\n\nThe soft spots are about the scope of the headline guarantee, not the core mechanism. The 1.59 / 2.35 thresholds come from one empirical scene (an 8x4 wave on a uniformly scaled, head-on rotated quad). The paper states these as if they are general guarantees, but the actual condition is simply that the union of needed texels over the wave fits in the active lanes. That union depends on sub-texel phase, anisotropic Jacobians, perspective shear, and active-lane masks. The paper does acknowledge fallback at edges and perspective growth for Box Sampling, but it never bounds the worst case for arbitrary content. That's a real gap, though it is a scoping issue rather than a flaw in the algorithm: when the set fits, the result is exact; when it doesn't, the fallback degrades gracefully. The authors could fix this by either proving a worst-case bound for 8x4 pixel waves or explicitly stating the guarantee as conditional on the unique-texel count.\n\nThe spatial-coherence assumption (lanes are an 8x4 tile in screen space) is also worth flagging — if a future GPU schedules lanes in scattered patterns, the whole scheme loses its footing. That's a hardware assumption, not a failure of the current implementation.\n\nWho is this for: anyone working on real-time rendering with neural or custom-compressed textures, and anyone building on wave intrinsics for shader-side filtering. It deserves a serious referee; the algorithms are reproducible enough that the core claims can be verified and the caveats can be tightened.\n\nRecommendation: send it to review. It is a solid HPG-quality contribution, with the main caveat that the abstract's 'guarantee' language should be scoped to the empirical setting or backed by a worst-case analysis.","headline":"A genuinely useful set of wave-communication tricks that make exact filtered texture lookups practical for expensive decompression formats; the headline magnification thresholds are empirically motivated rather than proven worst-case, but the core idea holds up.","tokens_in":23194,"tokens_out":2288,"would_cite":true,"duration_ms":22622,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that by distributing unique texel evaluations across the lanes of a GPU wave, exact zero-error texture filtering is possible at one texel evaluation per pixel whenever magnification is sufficiently large.","keywords":["stochastic texture filtering","wave intrinsics","texture magnification","bilinear filtering","bicubic filtering","texture compression","GPU shading","exact filtering"],"falsifier":"Render a magnified textured quad with an 8x4 pixel wave layout at magnification factor at least 1.59 and any rotation, with the fallback disabled, and count for every wave the number of distinct integer texel coordinates lying in the bilinear footprints of its pixels. Mask Sampling and List Merge predict this count is never above 32; a single wave with 33 or more needed texels, or a pixel whose four taps are not all present after the gather, would refute the zero-error claim. For Box Sampling the corresponding critical magnification is 2.35.","tokens_in":22185,"feed_emoji":"🖼️","tokens_out":8003,"duration_ms":84039,"temperature":0.7,"pith_summary":"The paper claims that when a texture is magnified, exact texture filtering does not require each pixel to decode all of its own texels. Instead, the lanes of a GPU wave can collaboratively determine every texel the wave needs, have each lane produce one unique texel, and then share those values so every pixel filters exactly. The paper presents three collaboration schemes, Box Sampling, Mask Sampling, and List Merge, that guarantee perfect bilinear filtering on 32-lane 8x4 waves once magnification reaches about 2.35 for Box Sampling and 1.59 for Mask Sampling and List Merge, at any orientation. This matters because modern compressed and neural texture formats cannot use the GPU's built-in texture filtering hardware, leaving stochastic filtering as the only cheap option, which introduces noise and flicker under magnification. The paper also contributes fallback methods that maintain higher quality than prior stochastic approaches when the wave cannot produce every needed texel.","feed_headline":"One texel per pixel is enough for exact texture filtering","feed_subtitle":"Wave-wide collaboration erases stochastic noise under magnification, even for neural and compressed texture formats.","key_machinery":"The machinery is GPU wave intrinsics, the register-level communication instructions that let shader lanes executing together as one wave exchange values without memory traffic, combined with three schedules for deciding which lane produces which texel. Box Sampling computes a wave-wide axis-aligned bounding box of all needed texel coordinates using WaveActiveMin and WaveActiveMax, then maps lane indices bijectively into that box. Mask Sampling encodes the wave's required texels in a 16x16 bitmask combined across lanes with WaveActiveBitOr, and uses a rank-and-select mapping that sends lane i to the i-th set bit and reverses the mapping when gathering taps. List Merge builds the exact duplicate-free list of needed texels hierarchically in shared memory and serves as the upper bound on how often perfect filtering is possible. These mechanisms convert redundant per-pixel texel decompression into one cooperative pass per wave.","core_discovery":"The central discovery is that exact texture filtering under magnification does not require each pixel to own its filter taps. The wave treats the union of all pixels' required texels as a shared job list: lanes first communicate which texel coordinates they need, then each lane produces one not-yet-produced texel from that list, then every lane gathers its filter taps from neighboring lanes via wave intrinsics and applies the true filter weights. When the number of unique texels fits within the 32 lanes, the result is identical to full bilinear or bicubic filtering, with zero error and at most one texel evaluation per pixel. For bilinear filtering, this success is guaranteed for magnification factors of about 1.59 and above with Mask Sampling and List Merge, and about 2.35 and above with Box Sampling, under the assumed 8x4 pixel wave layout. The same approach also handles filters with negative weights, such as Catmull-Rom, with at most one texel per lane, in contrast to the two taps that positivization requires in one-tap stochastic filtering.","pith_inferences":["If wave size grew beyond 32 lanes, the same counting argument would push the perfect-bilinear magnification threshold below 1.59 and toward 1.0 for wide enough waves; the paper states the methods generalize to any wave size but does not measure this scaling.","The bitmask rank-and-select scheme only requires each pixel's needed values to form a small enumerable set, so the same cooperative pattern could be reused for other per-pixel gathers, such as sharing material samples or light-list queries among neighboring pixels.","For decompressors whose per-texel cost is very high, the fallback regime may also beat per-pixel deterministic filtering even below the zero-error thresholds; the paper's expensive-decompressor measurement only demonstrates the high-magnification regime.","A hybrid that combines several waves through shared memory, which the authors mention as future work, could extend exact filtering to lower magnification or wider filters without waiting for hardware with larger waves."],"forward_implications":["For neural texture decompression at magnification above the thresholds, collaborative filtering can produce zero-error results with about one texel evaluation per pixel; the paper's neural decompression test showed a path around 93 microseconds per megapixel versus 516 for per-pixel 2x2 decompression, making exact filtering practical for expensive formats.","When a wave needs only a few unique texels, the decompression work for a single texel can itself be distributed across lanes: the paper demonstrates a 1.64x speedup in a mixed scene and 2.5x when fully magnified for DCT-based decompression by splitting color channels across lanes.","Filters with negative weights, such as Catmull-Rom, can be filtered exactly with at most one texel per lane under sufficient magnification, removing the two-tap cost that positivization imposes on one-tap stochastic filtering.","Below the perfect-filtering thresholds, the new fallback estimators give higher image quality than both one-tap stochastic texture filtering and wave-communication stochastic texture filtering; the extended fallback uses otherwise idle lanes to produce additional unique texels, so quality continues to improve as magnification frees more lanes.","The same collaborative structure extends to larger discrete filters: perfect bicubic filtering is achievable with one texel per lane at higher magnification, or with two texels per lane at magnification factors close to the bilinear ones."],"supporting_citations":[{"why":"Defines one-tap stochastic texture filtering and filtering-after-shading, the baseline the new methods must beat and the source of the positivization technique for negative filter weights.","marker":"[PWSF24]"},{"why":"Introduces wave-communication stochastic texture filtering with sample reuse across nearby pixels, the direct predecessor this work extends into deterministic collaboration.","marker":"[WPAM25]"},{"why":"Documents the HLSL Shader Model 6.0 wave intrinsics, including WaveActiveMin, WaveActiveMax, WaveActiveBitOr, and WaveReadLaneAt, that all three algorithms use for register-level sharing.","marker":"[Mic21]"},{"why":"Supplies the hierarchical parallel-sum idea that List Merge adapts into a duplicate-eliminating list merge.","marker":"[Hoo11]"},{"why":"Defines random-access neural texture compression, the expensive decompressor used to demonstrate that collaborative filtering yields zero error at a fraction of the full bilinear cost.","marker":"[VSW*23]"},{"why":"ColorVideoVDP is the primary perceptual metric used to compare filtered output against ground truth in the paper's evaluation.","marker":"[MHA*24]"}],"fun_headline_variants":["Wave lanes share texel work for exact filtering with one eval each","Lane-wide collaboration makes zero-error filtering one texel per pixel","Textures: wave communication slashes filtering to one texel per pixel","Exact magnification filtering via wave-shared texels, one per lane"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The zero-error guarantee assumes the 32 lanes of a wave are arranged as a spatially coherent 8x4 tile of pixels, so the union of the texels they need is small enough to fit in 32 slots; if lanes were scattered over random screen positions or heavily divergent, the texel demand would overflow the wave and the method would fall back to approximate filtering.","fun_headline_variants_meta":{"raw":{"variants":["Wave lanes share texel work for exact filtering with one eval each","Lane-wide collaboration makes zero-error filtering one texel per pixel","Textures: wave communication slashes filtering to one texel per pixel","Exact magnification filtering via wave-shared texels, one per lane"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1193,"prompt_tokens":928,"completion_tokens":265,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":188}},"tokens_in":544,"tokens_out":265,"duration_ms":3285,"temperature":1.0,"reasoning_tokens":188,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:01:01.680445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a magnified textured quad with an 8x4 pixel wave layout at magnification factor at least 1.59 and any rotation, with the fallback disabled, and count for every wave the number of distinct integer texel coordinates lying in the bilinear footprints of its pixels. Mask Sampling and List Merge predict this count is never above 32; a single wave with 33 or more needed texels, or a pixel whose four taps are not all present after the gather, would refute the zero-error claim. For Box Sampling the corresponding critical magnification is 2.35.","supporting_citations":[],"review_version":1}