ADS-Tile: Adaptive Tiling Methods
- ADS-Tile is an umbrella term for adaptive, tile-based methods that partition computation, imagery, and dataflow at granular levels for optimized scheduling.
- These methods span applications from transformer accelerators and CUDA kernels to dynamic tiling in video analytics and texture synthesis.
- Empirical results demonstrate significant gains in memory access reduction, throughput improvement, and efficient detector scheduling across various implementations.
Searching arXiv for the cited papers and nearest matches to “ADS-Tile”. arXiv search: (Du et al., 13 Mar 2026) ADS-Tile is not presented in the cited arXiv literature as a single standardized method name. In current usage it is better understood as an informal umbrella for adaptive or analysis-driven tile-based methods whose common operation is to partition computation, imagery, geometry, or dataflow into tiles and then optimize selection, scheduling, packing, or fusion at tile granularity. In the supplied corpus, the nearest realizations span tile-based adaptive stationary for transformer accelerators, editable CUDA tile SDPA kernels, sliding tile attention for video diffusion, adaptive image/video tiling, tile-aware Gaussian Splatting acceleration, controllable texture tiling, and adaptive Delaunay sampling that uses a tetrahedral scaffold rather than regular image tiles (Li et al., 25 Mar 2025, Khan, 2 Mar 2026, Zhang et al., 6 Feb 2025, Nguyen et al., 2023, Kittivorawong et al., 25 May 2026, Jo et al., 21 Apr 2026, Huang et al., 22 Jun 2026, Takikawa et al., 5 May 2026).
1. Terminological status and scope
The most precise characterization of ADS-Tile in the present literature is negative: it is not a stable canonical title shared by a single arXiv method. Several nearby papers explicitly state that their methods are not named ADS-Tile, even when they occupy the same design space. AdaGScale is described as “a related tile-acceleration alternative” for 3D Gaussian Splatting rather than as “ADS-Tile” or a rebranding of any prior tile accelerator (Jo et al., 21 Apr 2026). “Tile-based Adaptive Stationary (TAS)” is likewise described as conceptually close to adaptive tile-level dataflow scheduling, but “it is not called ‘ADS-Tile’ in the paper” (Li et al., 25 Mar 2025). “Adaptive Delaunay Sampling (ADS)” is a different use of the initials, centered on a tetrahedral sampling scaffold for occupancy functions rather than on regular image or compute tiles (Takikawa et al., 5 May 2026).
A second source of ambiguity is that “tiling” itself denotes different operations across subfields. In CAD code generation, “Design-Specification Tiling” is introduced as a coverage objective over multi-granular design components, not as a spatial image partition (Du et al., 13 Mar 2026). In texture synthesis, tiling means controlled repetition of a reference pattern according to frequency, orientation, and scale (Huang et al., 22 Jun 2026). In video analytics, tiling denotes fine-grained spatiotemporal pruning and packing of relevant regions (Kittivorawong et al., 25 May 2026). In accelerator and kernel work, it denotes schedule-level decomposition of matrices, attention blocks, or dataflow regions (Khan, 2 Mar 2026, Shen et al., 15 Dec 2025).
This suggests that ADS-Tile is best treated as a cross-domain family resemblance term: methods in this family share tile granularity and adaptive control, but not a single formal problem statement, optimization target, or implementation substrate.
2. Computational tiling in kernels and accelerator mapping
One major ADS-Tile-like lineage is computational tiling: tiles are units of arithmetic scheduling, data reuse, and memory movement. TAS is a representative example. It targets transformer linear projections and “selects the input or weight stationary in a tile granularity, based on the input sequence length,” with the stated goal of reducing external memory access in a shape-aware manner (Li et al., 25 Mar 2025). Its adaptive rule distinguishes regimes where input-stationary or weight-stationary reuse is preferable, then combines that choice with output-stationary partial-sum retention. In this sense, tile granularity is the locus of dataflow adaptation, not merely a blocking convenience.
TiledAttention occupies a different point in the same design space. It is a “CUDA Tile SDPA Kernel for PyTorch,” implemented in cuTile Python / TileIR and exposed as a PyTorch-callable function (Khan, 2 Mar 2026). Its defining properties are tiled execution over sequence length, streaming tiles, online softmax accumulation, no materialization of the full attention matrix, and schedule-level editability from Python. The emphasis here is not only throughput but a research surface in which tile shapes, staging depth, and shared-memory layout are first-class experimental parameters. That makes it a close computational analogue of an ADS-Tile methodology: the tile schedule itself is the object of design.
Sliding Tile Attention (STA) extends the same principle to 3D video diffusion attention. Instead of token-wise sliding windows, STA “operates tile-by-tile with a novel hardware-aware sliding window design,” exploiting the empirical observation that attention in pretrained video diffusion models is concentrated within localized spatial-temporal windows (Zhang et al., 6 Feb 2025). The crucial systems idea is block regularity: dense and empty blocks are preserved, while mixed blocks are eliminated. This converts sparsity from a theoretical FLOP reduction into a practical GPU speedup.
“Design in Tiles (DiT)” generalizes tile-level reasoning beyond attention to GEMM deployment on tile-based many-PE accelerators (Shen et al., 15 Dec 2025). Its schedule abstraction jointly exposes tiling and mapping, data layout, and dataflow, then lowers these decisions through an IR-based toolchain into executable kernels. A plausible implication is that, in accelerator contexts, ADS-Tile-like methods are less about one particular operator than about making tile placement and tile communication explicit design variables.
3. Spatial tiling for video, rendering, and texture control
A second lineage treats tiles as spatial support regions for sensing or rendering. Dynamic Tiling is the most direct image-analytics example. It is “a model-agnostic, adaptive, and scalable approach for small object detection” that starts from non-overlapping tiles, then uses dynamic overlapping rates and a tile minimizer to recover fragmented objects near boundaries while reducing unnecessary forward passes (Nguyen et al., 2023). The adaptive trigger is geometric: detections near tile edges or corners cause local re-tiling, rather than a global fixed-overlap crop schedule.
Tetris pushes this further in track materialization for stationary video (Kittivorawong et al., 25 May 2026). It decomposes frames into a tile-based polyomino data model, runs a classifier to identify relevant tiles, groups them into polyominoes, prunes redundant polyominoes with an ILP under a user-specified accuracy constraint, and then packs survivors into detector-sized canvases. Here the tile is simultaneously a spatial support primitive, an optimization variable, and a packing unit for detector batching.
AdaGScale shows the same idea in rendering rather than vision analytics (Jo et al., 21 Apr 2026). It targets the Gaussian-tile pair bottleneck in 3D Gaussian Splatting by estimating the peripheral contribution of each projected Gaussian and adaptively shrinking its support only for tile intersection testing. The original support is retained during color accumulation, so the method prunes low-importance Gaussian-tile pairs without changing the radiance accumulation rule for retained pairs. In this setting, the tile is a rasterization bin whose count controls intersection, sorting, and splatting cost.
Texture tiling provides another, non-accelerator meaning. “Controllable Texture Tiling with Transformed RoPE-Enhanced Diffusion Models” treats tiling as repeated placement of a reference pattern under user-defined affine control (Huang et al., 22 Jun 2026). Instead of pixel-space warping, the method transforms coordinates inside attention via a Coordinate-Transformed Rotary Embedding, while a Disjoint Attention Mask prevents semantic leakage from corrupting the reference stream. The tile is therefore a repeated appearance unit rather than a computation block.
4. Optimization objectives and algorithmic formulations
Across these literatures, ADS-Tile-like methods are distinguished not merely by partitioning into tiles, but by coupling tiles to explicit optimization objectives. In the CAD-code-generation setting, “Design-Specification Tiling” defines “knowledge sufficiency” as the objective for in-context exemplar selection and quantifies it through a surrogate tiling ratio that measures how much of a query specification is covered by selected exemplars (Du et al., 13 Mar 2026). The abstract further states that maximizing this objective constitutes submodular maximization and admits a polynomial-time greedy algorithm with a -approximation guarantee. Here “tiling” is best understood as compositional coverage.
Tetris is more operational. Its classifier produces candidate relevant tiles, but the decisive optimization step is an ILP that prunes redundant polyominoes under a user-specified accuracy constraint and then a packer that minimizes detector calls by assembling the survivors into canvases (Kittivorawong et al., 25 May 2026). The objective is not semantic coverage but minimizing processed tile area and detector invocations while satisfying learned maximum-gap constraints.
Dynamic Tiling uses a more heuristic control logic (Nguyen et al., 2023). Boundary-adjacent detections define where dynamic patches should be generated, and a tile minimizer then aggregates them into fewer detector inputs. The optimization is therefore implicit and pipeline-driven rather than stated as a formal objective function.
In accelerator work, DiT formalizes tiling, mapping, layout, and dataflow as a structured deployment space for GEMM (Shen et al., 15 Dec 2025), while STA formalizes local attention at tile granularity so that all queries in a tile share the same neighborhood of key tiles (Zhang et al., 6 Feb 2025). This suggests a recurring ADS-Tile pattern: a coarse partition first exposes tile-level structure, then a downstream optimizer chooses which tiles to keep, how often to revisit them, or how to map them onto hardware.
5. Empirical behavior and reported gains
The cited literature reports tile-level gains in very different metrics—external memory access, detector calls, Gaussian-tile pair count, SDPA throughput, and end-to-end latency—so direct comparison is not meaningful. What is comparable is that tile granularity is the primary lever in each case.
| Method | Domain | Reported effect |
|---|---|---|
| TAS (Li et al., 25 Mar 2025) | Transformer linear projections | reduces EMA by more than 97% |
| TiledAttention (Khan, 2 Mar 2026) | SDPA research kernel | mean 28.15x over PyTorch SDPA math; mean 14.36x over eager attention; mean 0.632x of fused PyTorch SDPA |
| STA (Zhang et al., 6 Feb 2025) | Video diffusion attention | 2.8–17x over FA2; 1.6–10x over FA3; 58.79% MFU; 945s to 685s without quality degradation; 268s with 0.09% VBench drop after finetuning |
| AdaGScale (Jo et al., 21 Apr 2026) | 3D Gaussian Splatting | 13.8x geometric mean speedup over original 3D-GS with about 0.5 dB PSNR degradation on city-scale scenes |
| Tetris (Kittivorawong et al., 25 May 2026) | Track materialization | within a 5% tracking accuracy loss; up to 17.4x higher throughput than prior systems; up to 68.8x higher than the reference pipeline |
| DiT (Shen et al., 15 Dec 2025) | GEMM on tile-based many-PE accelerators | 1.2–2.0x speedup across diverse matrix shapes |
| Dynamic Tiling (Nguyen et al., 2023) | Small object detection | best reported configuration reaches 45.5 mAP@0.5 and 31.2 [email protected] at 0.301 s |
A notable commonality is that the strongest results are paired with execution-aware control mechanisms rather than tile partitioning alone. TiledAttention couples tiling to online softmax and streaming (Khan, 2 Mar 2026). STA couples local windows to block-regular kernel execution (Zhang et al., 6 Feb 2025). Tetris couples tile classification to ILP pruning and packing (Kittivorawong et al., 25 May 2026). AdaGScale couples tile intersection to peripheral-contribution estimation while leaving accumulation unchanged (Jo et al., 21 Apr 2026). This suggests that ADS-Tile-like performance gains arise when tile selection is inseparable from downstream scheduling or reconstruction.
6. Limitations, non-equivalences, and adjacent tile literatures
The first limitation is terminological: the cited papers do not establish ADS-Tile as a single method. AdaGScale is explicitly described as “not the same method” as an ADS-Tile-style accelerator (Jo et al., 21 Apr 2026). TAS is “not called ‘ADS-Tile’ in the paper,” even though it is a strong conceptual match for adaptive tile-level dataflow (Li et al., 25 Mar 2025). Adaptive Delaunay Sampling uses “ADS” in a different sense entirely and replaces grid tiles with a progressively refined Delaunay tetrahedralization (Takikawa et al., 5 May 2026). In the CAD example, the abstract introduces “Design-Specification Tiling,” but the accompanying detailed note states that the supplied paper content did not establish either “ADS-Tile” or “DST” terminology in the available body text (Du et al., 13 Mar 2026).
A second limitation is domain specificity. Dynamic Tiling and Tetris assume stationary or structured vision workloads and optimize detector scheduling rather than arithmetic kernels (Nguyen et al., 2023, Kittivorawong et al., 25 May 2026). STA and TiledAttention target GPU attention kernels rather than scene partitioning (Zhang et al., 6 Feb 2025, Khan, 2 Mar 2026). Texture tiling is about controlled repetition in diffusion models, not sparse execution (Huang et al., 22 Jun 2026). Consequently, transferring one ADS-Tile-like method across domains usually requires changing both the tile semantics and the optimization criterion.
Finally, “tile” has an entirely separate meaning in theoretical and molecular self-assembly. There, the core questions involve tile concentrations, universality, seed complexity, and tile-set synthesis rather than adaptive scheduling or spatial pruning. Representative examples include concentration optimization in DNA tile self-assembly (Chen et al., 2012), intrinsic universality of the abstract Tile Assembly Model (Doty et al., 2011), and NP-hard search for minimal tile sets in patterned self-assembly (Göös et al., 2014). In the supplied literature, this is a distinct lineage and should not be conflated with computational or vision-oriented ADS-Tile usage unless the intended context is explicitly aTAM, 2HAM, or DNA self-assembly.
