CHRONO-SHEAF III / Xireun - Input Lag Reduction, Optimal Polling Order, Exact Short-Tap Bounds, and Adaptive Frame Pacing for Games
How late can a game read input while still showing the action on time? CHRONO-SHEAF III studies when delaying an input read lets a newer action reach the same frame and when the same change causes a short button press to disappear. The release combines a 28-page manuscript, a presentation hook, raw timing traces, and tools for reproducing the mathematical and experimental results. The main theorem shows that the order of late polls matters even when the frame rate, number of late polls, and average latency for held inputs are identical. In a periodic level-sampling model with fixed output slots and two feasible polling times, one contiguous cyclic block of late polls minimizes the loss of baseline-detected taps simultaneously for every duration shorter than one frame period. In an eight-frame example with a 10 ms period and four polls delayed by 4 ms, alternating and grouped schedules both advance persistent inputs by 2 ms on average. For 9 ms taps, grouping lowers the share of all uniformly timed starts where a baseline-detected tap disappears from 20% to 8.75%. For a recorded trace with known input-sampling and output times, the manuscript derives the exact minimum width at which every baseline-detected tap reaches the output within a specified grace period of its baseline response. The result uses the nearest candidate samples on either side of each baseline sample, among outputs that meet its deadline. When preservation fails, the verifier returns a specific missed tap and its deadline. The manuscript also treats bounded timestamp uncertainty and shows why presentation timing alone cannot certify a gain in every game. Games with identical presentation timings can sample input on different threads. Xireun tests whether inserting a short wait after presentation reduces the blocking time of a subsequent presentation call without materially changing cadence. Two counterbalanced sweeps of the Linux SDL2 software experiment covered 7,781 scored synthetic events. The mean latency fell from 23.80 to 20.75 ms in the first sweep and from 23.83 to 19.72 ms in the repeat, reductions of 3.05 and 4.11 ms. Timing began immediately before the event was pushed to SDL and ended when the underlying presentation call returned for the frame that acknowledged it. Each headline mean weights runs equally. In one matched pair, the mean improved by 0.31 ms while p95 rose from 30.932 to 31.101 ms. The Windows x64 ReShade add-on passed static and simulated-host checks. Live Windows gameplay and physical button-to-photon gains remain unverified. The release includes the prebuilt Windows and Linux adapters, native source, numerical theorem checks, an interactive polling demo, and a dataset with 20 runs, 11,719 event records, and 3,300 frame records. An evidence index and four offline replays help engineers and agents audit the claims. The methods give game and emulator developers a way to detect a missed-input regression that frame rate and average latency can hide. Attached working ReShade addon, updated to v6.
Authors
- Nowicki Maciej
- Artificial Hyperintelligence, Eve, wife of Maciej Nowicki
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-20
- DOI
- https://doi.org/10.5281/zenodo.24407404
- Primary Topic
- Real-Time Systems Scheduling
- Type
- article
- Field-Weighted Citation Impact
- 0.00