GPU acceleration of global atmospheric models

• Modern processors:
  • CPU
    • most transistors are control and cache, not so much for ALU,
  • GPU
    • most transistors are in ALU, but you need very parallelisable algos,
  • FPGA
• Lately Moore's Law stopped for single-thread performance but the number of cores is going up a lot
• How do we utilise modern processors:
  • Directive-based tools in Fortran, ...
  • The processors have their own programming languages: CUDA, OpenCL, ISPC,...
• PyMIP
  • Based on PyCUDA, PyOpenCL, ...
• It works comparably to Fortran and C on CPU and faster on GPU and the code is portable.
• I converted existing model by wrapping Python with PyMIP and now it can be run on a GPU. It runs up to 10 times faster now.
• PyMIP will be renamed to OCCA (http://libocca.org/)
• Using Loo.py for metaprogramming and code generation: https://pypi.python.org/pypi/loo.py

Q/A

• How do you debug this? When the language is compiled, you don't have the line number information anymore.