TornadoVM experiments

This repository contains some experiments made with TornadoVM with GPU support.

• Hampel.java Implementation of the Hampel filter
• ReductionAvgFloats.java Average reduction of a float array
• KMeans.java Implementation of the baseline KMeans algorithm

How to run

To build your code use one of the commands inside the build script, for example:

./nvidia_docker_run.sh javac.py Hampel.java

To run:

./nvidia_docker_run.sh tornado Hampel

The directory opencl_generated/ contains some OpenCL code generated by TornadoVM. The file generated_sum.c is analyzed and commented to understand how TornadoVM handles reductions, which are not a trivial task to parallelize on a GPU. To see the generated code pass the flag tornado --printKernel when running the code (OpenCL kernel is generated at runtime, not at compilation time).

Performance

The directory perf/ contains some plots of the performances obtained with TornadoVM, compared with the sequential execution by the CPU. For example with Hampel filter:

Requirements

Docker image of TornadoVM with GPU support, https://github.com/beehive-lab/docker-tornado

References

• An Introduction to the OpenCL Programming Model
• Efficient Reductions with CUDA