Project Proposal: GPU support
me at alcidesfonseca.com
Wed Aug 29 01:02:54 UTC 2012
I am also interesting in participating on this project. In my masters thesis I developed ÆminiumGPU a GPU-backed Map-Reduce library for Java (and Æminium, a JVM-based language).
My approach differs from the ones proposed in which it is a Source-to-Source compiler that takes Java source code, looks up for the lambda functions of higher-order functions (map, reduce and a couple others) and translates them to OpenCL functions. At runtime these functions are integrated in kernel templates (one for map, other for reduce, …) and executed using the JavaCL library.
Another of the novel features is that it automatically decides if a given operation will be executed on the CPU or on the CPU. We have both a regression solution as well as Machine Learning classification systems.
This is something that might be interesting to include in the Java solution, even if it is bytecode-based.
— Alcides Fonseca
University of Coimbra
 https://github.com/aeminium/AeminiumGPU and https://github.com/aeminium/AeminiumGPUCompiler
More information about the discuss