Performance tools

...Making Python fast...

Traditional Python performance

• Python has a number of strikes against it from a performance perspective
  • Dynamic typing
  • Intepreted
  • Extremely general object model
• There is a long history of projects that have sought to make Python fast
  • SWIG (call C/C++ from Python)
  • f2py (call Fortran from Python)
  • ctypes (call shared libraries from Python)
  • Boost.Python (call C++ from Python)

New tools for performance

• New breed of tools are emerging
  • Just in time (JIT) compilation
  • LLVM
  • Optimization techniques
  • Add static types and type inference to Python
• In this talk:
  • Cython
  • Numba
• Others
  • Theano
  • Parakeet

Cython

• Superset of Python that adds explicit static typing
• Can also call C/C++ code
• Compiled into fast C/C++ with autogenerated Python bindings
• Established, robust, stable
• Best solution for wrapping existing C/C++ codes
• Solid support for multidimensional numerical arrays
• Open source, community effort
• The GitHub project page

Numba

• Start with regular Python code
• Parses the Python code into its abstract syntax tree (AST)
• Static typing and type inference
• Uses LLVM to JIT compile directly into machine code
• New, but gaining momentum
• Solid support for multidimensional numerical arrays
• The GitHub project page
• Open source, developed primarily by Continuum Analytics

Performance benchmarks!