trying to call cblas without python function call overhead

numba_dot

{
 "metadata": {
  "name": ""
 },
 "nbformat": 3,
 "nbformat_minor": 0,
 "worksheets": [
  {
   "cells": [
    {
     "cell_type": "code",
     "collapsed": false,
     "input": [
      "# using ubuntu 12.10\n",
      "# shut up debug symbols: https://groups.google.com/a/continuum.io/forum/#!topic/numba-users/wwC80xyjuu4\n",
      "# http://arjones6.blogspot.ch/2013/05/passing-multidimensional-numpy-arrays.html"
     ],
     "language": "python",
     "metadata": {},
     "outputs": [],
     "prompt_number": 25
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": [
      "from cffi import FFI\n",
      "ffi = FFI()\n",
      "ffi.cdef(\"\"\"     // some declarations from the man page\n",
      "    double cblas_ddot (const int N, const double * x, const int incx, const double * y, const int incy);\n",
      "\"\"\")\n",
      "C = ffi.verify(\"\"\"   // passed to the real C compiler\n",
      "#include <gsl/gsl_cblas.h>\n",
      "\"\"\", libraries=['gslcblas'])   # or a list of libraries to link with"
     ],
     "language": "python",
     "metadata": {},
     "outputs": [],
     "prompt_number": 18
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": [
      "import numpy as np\n",
      "import ctypes\n",
      "from numba import double\n",
      "\n",
      "length = 33\n",
      "x = np.arange(length, dtype=np.float64)\n",
      "y = x.copy()\n",
      "\n",
      "x_p = ffi.cast(\"double *\", x.ctypes.data)\n",
      "y_p = ffi.cast(\"double *\", y.ctypes.data)\n",
      "\n",
      "assert C.cblas_ddot(length,x_p, 1, y_p, 1) == np.dot(x,y)\n",
      "print 'np.dot'\n",
      "%timeit np.dot(x,y)\n",
      "print 'cblas ddot'\n",
      "%timeit C.cblas_ddot(length,x_p, 1, y_p, 1)"
     ],
     "language": "python",
     "metadata": {},
     "outputs": [
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "np.dot\n",
        "1000000 loops, best of 3: 642 ns per loop"
       ]
      },
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "\n",
        "cblas ddot\n",
        "1000000 loops, best of 3: 364 ns per loop"
       ]
      },
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "\n"
       ]
      }
     ],
     "prompt_number": 27
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": [
      "from numba import autojit\n",
      "from numba.decorators import jit, autojit\n",
      "X = np.arange(300*300).reshape(300,300)\n",
      "\n",
      "\n",
      "def sum2d(arr):\n",
      "    M, N = arr.shape\n",
      "    result = 0.0\n",
      "    for i in range(M):\n",
      "        for j in range(N):\n",
      "            result += arr[i,j]\n",
      "    return result\n",
      "\n",
      "sum2d_numba = autojit(sum2d)\n",
      "\n",
      "print 'sum'\n",
      "%timeit sum2d(X)\n",
      "print 'sum + numba'\n",
      "%timeit sum2d_numba(X)"
     ],
     "language": "python",
     "metadata": {},
     "outputs": [
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "sum\n",
        "10 loops, best of 3: 42.2 ms per loop"
       ]
      },
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "\n",
        "sum + numba\n"
       ]
      },
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "10000 loops, best of 3: 110 \u00b5s per loop\n"
       ]
      }
     ],
     "prompt_number": 28
    },
    {
     "cell_type": "code",
     "collapsed": false,
     "input": [
      "def sumdot(x, y, n):\n",
      "    result = 0.0\n",
      "    for i in xrange(n):\n",
      "        result += np.dot(x,y)\n",
      "    return result\n",
      "\n",
      "print 'np.dot'\n",
      "%timeit sumdot(x, y, 1000000)\n",
      "sumdot_numba = autojit(sumdot)\n",
      "print 'np.dot + numba'\n",
      "%timeit sumdot_numba(x,y, 1000000)\n",
      "\n",
      "\n",
      "print 'ddot (cblas/cffi)'\n",
      "from numba import autojit\n",
      "\n",
      "x_p = ffi.cast(\"double *\", x.ctypes.data)\n",
      "y_p = ffi.cast(\"double *\", y.ctypes.data)\n",
      "\n",
      "def sumdot_cffi(x_p, y_p, n):\n",
      "\n",
      "    length = len(x)\n",
      "    result = 0.0\n",
      "    for i in xrange(n):\n",
      "        result += C.cblas_ddot(length, x_p, 1, y_p, 1)\n",
      "    return result\n",
      "\n",
      "assert sumdot_cffi(x_p, y_p, 1000000) == sumdot(x, y, 1000000)\n",
      "\n",
      "%timeit sumdot_cffi(x_p, y_p, 1000000)\n",
      "sumdot_cffi_numba = autojit(sumdot_cffi)\n",
      "print 'ddot (cblas/cffi) + numba'\n",
      "%timeit sumdot_cffi(x_p, y_p, 1000000)"
     ],
     "language": "python",
     "metadata": {},
     "outputs": [
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "np.dot\n",
        "1 loops, best of 3: 902 ms per loop"
       ]
      },
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "\n",
        "np.dot + numba\n"
       ]
      },
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "1 loops, best of 3: 528 ms per loop\n",
        "ddot (cblas/cffi)\n",
        "1 loops, best of 3: 347 ms per loop"
       ]
      },
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "\n",
        "ddot (cblas/cffi) + numba\n",
        "1 loops, best of 3: 347 ms per loop"
       ]
      },
      {
       "output_type": "stream",
       "stream": "stdout",
       "text": [
        "\n"
       ]
      }
     ],
     "prompt_number": 26
    }
   ],
   "metadata": {}
  }
 ]
}