#2 add lecture 2

Author: martinjrobins

12_optimisation_1/lectures/lectures_01_cython.ipynb

@@ -0,0 +1,193 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "slideshow": {
+     "slide_type": "slide"
+    }
+   },
+   "source": [
+    "# Optimisation - Cython\n",
+    "## Martin Robinson\n",
+    "## Oct 2019"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "slideshow": {
+     "slide_type": "slide"
+    }
+   },
+   "source": [
+    "# Using lower-level languages\n",
+    "\n",
+    "- Interpreted languages are fundamentially speed-limited when they only consider *type* at run-time.\n",
+    "- e.g. consider what happens with the types of the variables in the following function\n",
+    "```python\n",
+    "def norm(arg_list, p):\n",
+    "    sum = 0               # sum is an int here\n",
+    "    for x in arg_list:    # type of x depends on input container\n",
+    "        sum += abs(x)**p  # type of rhs depends on both x and p, sum could *change* type here\n",
+    "    return sum**(1.0/p)   # return value is probably float due to 1.0\n",
+    "```\n",
+    "- how much memory to allocate for sum? does this memory need to be re-allocated during the loop? are conversion routines between types required during the loop?"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "slideshow": {
+     "slide_type": "slide"
+    }
+   },
+   "source": [
+    "- compare to equivilant C++ code\n",
+    "```cpp\n",
+    "float norm(std::vector<float>& arg_list, float p) {\n",
+    "    float sum = 0.0f;\n",
+    "    for (size_t i = 0; i < arg_list.size(); ++i) {\n",
+    "        sum += std::pow(std::abs(arg_list), p);\n",
+    "    }\n",
+    "    return std::pow(sum, 1.0f/p);\n",
+    "}\n",
+    "```\n",
+    "- compiler can pre-allocate the stack size because the sizes of local variables known, no type conversions neccessary\n",
+    "- compiler can generate efficient machine code because the programmer has provided more information (i.e. types)\n",
+    "    "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# \"Compiling\" Python code\n",
+    "\n",
+    "- All python implementations (CPython, PyPy, IronPython) compile to *bytecode*, which is then either interpreted at run-time, or perhaps further compiled to native machine code\n",
+    "- Implementations that compile to native machine code usually implement something close to normal python, but with restrictions or additions that alter the nature of the language. These include:\n",
+    "    - Cython (Python-to-C)\n",
+    "    - Nuitka (Python-to-C++)\n",
+    "    - Numba (Python-to-LLVM IR)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# \"Wrapping\" C and C++ for use in Python\n",
+    "\n",
+    "- the compilers in the previous slide implement an altered version of python, yet another language to learn!\n",
+    "- If your already comfortable with C, C++ or Fortran, why not use this directly and write a *wrapper* to call from Python?\n",
+    "- "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Cython"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Your first Cython program"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "```python\n",
+    "def norm(a, p):\n",
+    "    s = 0\n",
+    "    x_max = a.shape[0]\n",
+    "    y_max = a.shape[1]\n",
+    "    for i in range(x_max):\n",
+    "        for j in range(y_max):\n",
+    "            s += abs(a[i, j])**p\n",
+    "    return s**(1.0/p)\n",
+    "\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Manual compilation"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Examining the generated code"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Adding types"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# memoryviews"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Tuning indexing further"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "```python\n",
+    "from libc.math cimport abs\n",
+    "cimport cython\n",
+    "\n",
+    "@cython.boundscheck(False)  # Deactivate bounds checking\n",
+    "@cython.wraparound(False)   # Deactivate negative indexing.\n",
+    "@cython.cdivision(True)     # Deactivate normal python division checking\n",
+    "cdef double norm(double [:, :] a, int p):\n",
+    "    cdef double s = 0\n",
+    "    cdef Py_ssize_t x_max = a.shape[0]\n",
+    "    cdef Py_ssize_t y_max = a.shape[1]\n",
+    "    for i in range(x_max):\n",
+    "        for j in range(y_max):\n",
+    "            s += abs(a[i, j])**p\n",
+    "    return s**(1.0/p)\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Packaging Cython programs"
+   ]
+  }
+ ],
+ "metadata": {
+  "language_info": {
+   "name": "python",
+   "pygments_lexer": "ipython3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}