Java and GPU … are we nearly there yet?
summary
Argues that combining Code Reflection with the Panama FFM lets Java generate GPU kernels, pack C99-style layouts as Panama segments, and auto-infer task graphs so vendors can build NDRange-style runtimes on top. Demonstrates a hard-cascade face detector that processes ~2 million scaled windows, 130 million half-features and 271 million area ops; Java parallel stream runs at ~522 ms, sequential ~3.5 s, GPU back end runs the image in ~53 ms — a working, order-of-magnitude speedup.
AI-written summary · about
chapters (6)
- 0:00Start
- 4:46Why? Performance!
- 11:10Java GPGPU Programming Models
- 27:55Consider multiple kernels with data dependencies
- 31:13Panama FFM API MemorySegments as GPGPU Parameters
- 38:43Demo: Viola Jones Detection Performance
related talks
40mJava GPGPU Enablement: Are We There Yet?
51mJava on the GPU Where are we now?
49mMassive Parallelism with GPUs in Java
50mUnder the HAT: Empowering GPU Acceleration for Java
44mWrite Once, Accelerate Everywhere: GPU-Ready Java with TornadoVM
43mHow to make your electrons do more in Java?
40mHow to make your electrons do more in Java
46mTornadoVM: Supercharge your Java with GPU acceleration