NPU Bench Swarm: instrument, models, and complete raw dataset

Artifact for a research article by Jos Timanta Tarigan on in-situ latency benchmarking of neural game mechanics on a mobile neural processing unit (under review; the full citation will be added to this record upon acceptance). A 2D swarm-shooter game built as a measurement instrument: enemy swarms are driven by a batched policy-network inference every 60 Hz tick (batch size = swarm size, 8–80), and gesture-cast abilities exercise a sporadic input-path CNN. Measurements cover CPU (XNNPACK), GPU-delegate, and NPU (NNAPI) backends of TensorFlow Lite / LiteRT 1.4.2 on a Poco X6 Pro (MediaTek Dimensity 8300-Ultra, APU 780 NPU, Android 16). Contents: full native C++/GLES3 source and Gradle build (npubench-swarm-code.zip, MIT license); 16 trained TFLite models (npubench-swarm-models.zip, CC BY 4.0); installable APK (npubench-swarm-v0.1.apk, MIT); complete raw dataset of 194 CSV files with 159,735 per-tick records, 161 cold-start records, and 1,306 gesture strokes (npubench-swarm-runs.zip, CC BY 4.0); analysis pipeline and generated figures (npubench-swarm-analysis.zip); project documentation (npubench-swarm-docs.zip). See DATA_README.md for the folder layout, CSV column dictionaries, and reproduction instructions; SHA256SUMS.txt covers every file. Development repository: https://github.com/jttarigan/npubench-swarm