Provides transparent scalability for using multiple GPUs with JCuda
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
src
.gitignore
LICENSE
README.md

README.md

TransScale

Provides transparent scalability for using multiple GPUs with JCuda. TransScale facilitates agile CUDA development in Java using JCuda. With TransScale, it is now easy to load modules on single or multiple CUDA-enabled GPUs and harness their processing power.

TransScale also provides wrappers for 2D CUDA primitive arrays of structs with arbitrary number of fields (such as a 2D array of char's representing an array of colors in ARGB space). These 2D allocations can be automatically pitched if possible and the pitch values are converted to a number that can be treated as a regular "width" similar to any 2D allocated array. The wrapped pointers are extensions of CUdeviceptr which allows transparent manipulation using JCuda.

Note that the project is still very immature and more features and samples may be added in the future. The sample package of the source code contains the example of a vector addition solution using CUDA.

Acknowledgement

I would like to thank Marco Hutter for creating JCuda and his support. JCuda's homepage is located here