{
 "metadata": {
  "name": ""
 },
 "nbformat": 3,
 "nbformat_minor": 0,
 "worksheets": [
  {
   "cells": [
    {
     "cell_type": "code",
     "collapsed": false,
     "input": [
      "%load_ext load_style\n",
      "%load_style talk.css"
     ],
     "language": "python",
     "metadata": {
      "slideshow": {
       "slide_type": "-"
      }
     },
     "outputs": []
    },
    {
     "cell_type": "heading",
     "level": 1,
     "metadata": {
      "slideshow": {
       "slide_type": "slide"
      }
     },
     "source": [
      "<h1 class=\"bigtitle\">Performance tools</h1>\n",
      "\n",
      "<h2>...Making Python fast...</h2>"
     ]
    },
    {
     "cell_type": "heading",
     "level": 1,
     "metadata": {
      "slideshow": {
       "slide_type": "slide"
      }
     },
     "source": [
      "Traditional Python performance"
     ]
    },
    {
     "cell_type": "markdown",
     "metadata": {},
     "source": [
      "* Python has a number of strikes against it from a performance perspective\n",
      "  - Dynamic typing\n",
      "  - Intepreted\n",
      "  - Extremely general object model\n",
      "* There is a long history of projects that have sought to make Python fast\n",
      "  - SWIG (call C/C++ from Python)\n",
      "  - f2py (call Fortran from Python)\n",
      "  - ctypes (call shared libraries from Python)\n",
      "  - Boost.Python (call C++ from Python)\n",
      "\n",
      "  "
     ]
    },
    {
     "cell_type": "heading",
     "level": 1,
     "metadata": {
      "slideshow": {
       "slide_type": "slide"
      }
     },
     "source": [
      "New tools for performance"
     ]
    },
    {
     "cell_type": "markdown",
     "metadata": {},
     "source": [
      "* New breed of tools are emerging\n",
      "  - Just in time (JIT) compilation\n",
      "  - LLVM\n",
      "  - Optimization techniques\n",
      "  - Add static types and type inference to Python\n",
      "* In this talk:\n",
      "  - Cython\n",
      "  - Numba\n",
      "* Others\n",
      "  - [Theano](http://deeplearning.net/software/theano/)\n",
      "  - [Parakeet](http://iskandr.github.io/parakeet/)\n",
      "  "
     ]
    },
    {
     "cell_type": "heading",
     "level": 1,
     "metadata": {
      "slideshow": {
       "slide_type": "slide"
      }
     },
     "source": [
      "Cython"
     ]
    },
    {
     "cell_type": "markdown",
     "metadata": {},
     "source": [
      "* Superset of Python that adds explicit static typing\n",
      "* Can also call C/C++ code\n",
      "* Compiled into fast C/C++ with autogenerated Python bindings\n",
      "* Established, robust, stable\n",
      "* Best solution for wrapping existing C/C++ codes\n",
      "* Solid support for multidimensional numerical arrays\n",
      "* Open source, community effort\n",
      "* The [GitHub project page](https://github.com/cython/cython)"
     ]
    },
    {
     "cell_type": "heading",
     "level": 1,
     "metadata": {
      "slideshow": {
       "slide_type": "slide"
      }
     },
     "source": [
      "Numba"
     ]
    },
    {
     "cell_type": "markdown",
     "metadata": {},
     "source": [
      "* Start with regular Python code\n",
      "* Parses the Python code into its abstract syntax tree (AST)\n",
      "* Static typing and type inference\n",
      "* Uses LLVM to JIT compile directly into machine code\n",
      "* New, but gaining momentum\n",
      "* Solid support for multidimensional numerical arrays\n",
      "* The [GitHub project page](https://github.com/numba/numba)\n",
      "* Open source, developed primarily by [Continuum Analytics](http://www.continuum.io/)"
     ]
    },
    {
     "cell_type": "markdown",
     "metadata": {
      "slideshow": {
       "slide_type": "slide"
      }
     },
     "source": [
      "<h1 class=\"bigtitle\">Performance benchmarks!</h1>"
     ]
    }
   ],
   "metadata": {}
  }
 ]
}