Java GPGPU Enablement: Are We There Yet?
summary
HAT exposes an ND-range kernel layer plus a higher compute model so a backend can minimize buffer copies; it uses Babylon to lift Java code into a transformable model and Panama/JExtract for FFI and off‑heap MemorySegments. Demos show first-call compile latency (~2005 ms) but big wins: a healing‑brush search drops from ~3 s to 41 ms, Viola‑Jones from 1.7 s to 82 ms, and Game of Life rises from ~2 fps to ~157 fps. Backends exist for CUDA/OpenCL/PTX and a Java reference; remaining work is reducing stray copies and easing native builds.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 3:57heterogeneous accelerator toolkit
- 6:59heterogeneous toolkit
- 8:32kernel
- 11:15Babylon
- 18:10Arrays
- 22:37Healing Brush
- 25:18Face Detector
- 27:39Game of Life
- 29:19A Bad Example
- 37:20Summary
related talks
50mUnder the HAT: Empowering GPU Acceleration for Java
47mJava and GPU … are we nearly there yet?
51mJava on the GPU Where are we now?
46mHeterogeneous Accelerator Toolkit (HAT) Update
43mReflecting on HAT: A Project Babylon Case Study
49mMassive Parallelism with GPUs in Java
43mHow to make your electrons do more in Java?
44mWrite Once, Accelerate Everywhere: GPU-Ready Java with TornadoVM