Spatially Tiled Integral Histograms and Fixed-Point Lookup-Table Entropy from 512³ to Real 1024³ (1.07B Voxels) and 512⁴ Scale
High-throughput sliding-window tensor processing, quantized neural-network and large-language-model activation kernels, and volumetric directional entropy evaluation on modern GPUs are fundamentally constrained by multiple hardware bottlenecks: special-function-unit latency on transcendental operations, neighborhood scaling per thread, and catastrophic video-memory explosion when storing volumetric and higher-dimensional prefix-sum and integral-histogram tables. This work introduces a strictly division-free, square-root-free, and logarithm-free CUDA lookup-table and spatially tiled prefix-sum architecture, validated on NVIDIA Ampere and Turing GPUs and extended to extreme multidimensional tensor scales and higher-dimensional tables. The central contribution is the replacement of runtime floating-point divisions and transcendental logarithms with a precomputed fixed-point integer lookup table resident in the GPU cache hierarchy. Multi-architecture benchmarks reveal critical hardware principles for quantized GPU and large-language-model kernels: first, on special-function-unit-constrained GPUs, a cache-resident global lookup table outperforms native transcendental math and matches special-function-unit throughput on other architectures; second, allocating large static per-block shared-memory lookup tables throttles multiprocessor warp occupancy, making cache-resident global lookup tables faster than static shared-memory alternatives on both tested GPU generations. For GPU vector binning, replacing transcendental trigonometric and Euclidean-norm operations with integer hemisphere tests, quadratic thresholds, and cross-product sector comparisons achieves exact bin agreement with the reference method and a substantial speedup, eliminating transcendental operations from the inner loop. To overcome the memory wall that prevents scaling to billion-voxel volumes and higher-dimensional tables, a spatial tiling scheme with bounded-precision integer prefix accumulation drastically reduces prefix-table memory footprint and enables billions of sliding windows to be completed within viable memory limits. Building the prefix table once for multiple simultaneous multi-scale radii consistently outperforms separable sliding-sum baselines, and higher-dimensional corner-based prefix tables achieve even larger speedups. Finally, a local Fannes–Audenaert continuity bound is proved and verified with zero violations, and unwrapping-free detection of severe phase aliasing is demonstrated.
Authors
- Andres Sebaatian Pirolo (ORCID: https://orcid.org/0009-0004-3899-1222)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24257906
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- preprint