Reikna is a library containing various GPU algorithms built on top of PyCUDA and PyOpenCL. The main design goals are:

• separation of computation cores (matrix multiplication, random numbers generation etc) from simple transformations on their input and output values (scaling, typecast etc);
• separation of the preparation and execution stage, maximizing the performance of the execution stage at the expense of the preparation stage (in other words, aiming at large simulations)
• partial abstraction from CUDA/OpenCL

For more information proceed to the project documentation page. If you have a general question that does not qualify as an issue, you can ask it at the discussion forum.