Over the last years, we have worked in the area of automatic and transparent acceleration of legacy sequential CPU applications. We are proud to announce that we have now open-sourced HTrOP, the prototypical implementation of a compilation and runtime environment for multi-accelerator platforms that was devised and evaluated within this research.
HTrOP is able to automatically identify computational-intensive application parts, parallelize the sequential code, generate corresponding OpenCL kernel and host code and transparently offload the hotspots to different OpenCL-enabled computing devices.
A journal article (open access) on this topic has just been published at ACM Transactions on Architecture and Code Optimization (TACO).
Further information can be found on the project's webpage.