Area of research:
Diploma & Master Thesis
The position is suitable for part-time employment.
CUDA and OpenCL are platforms to write fast and efficient general-purpose applications for GPUs. OpenCL is running on a variety of different hardware. CUDA is developed by NVIDIA and supports NVIDIA GPUs only. On the other hand, CUDA is a more mature technology which provides significantly better development tools and a large collection of high-performance libraries. The NVIDIA GPUDirect technology is invaluable for low latency real-time application. It allows direct communication between GPUs and other devices on the PCIe bus, but is only available on CUDA platform as well. While both platforms have similar concepts and define a very similar syntax, the communication between OpenCL and CUDA components is not directly possible.
The goal of the project to investigate how it is possible to inter-operate modules written in CUDA and OpenCL with a minimal penalty to the performance. The student is expected to evaluate different approaches, measure associated performance penalty, and apply the selected approach to OpenCLbased UFO image-processing framework.
The goal is to
Facilitate GPUDirect technology if NVIDIA GPUs are usedEnable utilization of NVIDIA developer tools with OpenCL pluginsAdd support for plugins written in CUDA
There are several possible approaches to explore:
Both CUDA and OpenCL are able to inter-operate with OpenGL to combine computing and visualization. OpenGL can be used as intermediate layer to provide inter-operation between CUDA and OpenCL.POCL (Portable Computing Language) is an opensource OpenCL platform supporting CPUs and GPUs. For NVIDIA GPUs, it serves as a bridge between an OpenCL application and CUDA framework. It should be possible to extend POCL to expose CUDA structures hidden beneath OpenCL abstraction.LLVM infrastructure may be utilized to convert between CUDA and OpenCL dialects on the fly. Particularly, AMD develops ROCm framework which includes tool to convert CUDA applications to AMD HCC. The resulting code can be, then, executed on NVIDIA with CUDA or AMD using OpenCL.