1 2 3<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" 4 "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"> 5<html xmlns="http://www.w3.org/1999/xhtml"> 6 <head> 7 <meta http-equiv="Content-Type" content="text/html; charset=utf-8" /> 8 9 <title>Tutorial: Image Gradient - Boost.GIL documentation</title> 10 <link rel="stylesheet" href="../_static/pygments.css" type="text/css" /> 11 <link rel="stylesheet" href="../_static/style.css" type="text/css" /> 12 <script type="text/javascript"> 13 var DOCUMENTATION_OPTIONS = { 14 URL_ROOT: '../', 15 VERSION: '', 16 COLLAPSE_MODINDEX: false, 17 FILE_SUFFIX: '.html' 18 }; 19 </script> 20 <script type="text/javascript" src="../_static/jquery.js"></script> 21 <script type="text/javascript" src="../_static/underscore.js"></script> 22 <script type="text/javascript" src="../_static/doctools.js"></script> 23 <link rel="index" title="Index" href="../genindex.html" /> 24 <link rel="search" title="Search" href="../search.html" /> 25 <link rel="top" title="Boost.GIL documentation" href="../index.html" /> 26 <link rel="next" title="Naming Conventions" href="../naming.html" /> 27 <link rel="prev" title="Tutorial: Histogram" href="histogram.html" /> 28 </head> 29 <body> 30 <div class="header"> 31 <table border="0" cellpadding="7" cellspacing="0" width="100%" summary= 32 "header"> 33 <tr> 34 <td valign="top" width="300"> 35 <h3><a href="../index.html"><img 36 alt="C++ Boost" src="../_static/gil.png" border="0"></a></h3> 37 </td> 38 39 <td > 40 <h1 align="center"><a href="../index.html"></a></h1> 41 </td> 42 <td> 43 <div id="searchbox" style="display: none"> 44 <form class="search" action="../search.html" method="get"> 45 <input type="text" name="q" size="18" /> 46 <input type="submit" value="Search" /> 47 <input type="hidden" name="check_keywords" value="yes" /> 48 <input type="hidden" name="area" value="default" /> 49 </form> 50 </div> 51 <script type="text/javascript">$('#searchbox').show(0);</script> 52 </td> 53 </tr> 54 </table> 55 </div> 56 <hr/> 57 <div class="content"> 58 <div class="navbar" style="text-align:right;"> 59 60 61 <a class="prev" title="Tutorial: Histogram" href="histogram.html"><img src="../_static/prev.png" alt="prev"/></a> 62 <a class="next" title="Naming Conventions" href="../naming.html"><img src="../_static/next.png" alt="next"/></a> 63 64 </div> 65 66 <div class="section" id="tutorial-image-gradient"> 67<h1>Tutorial: Image Gradient</h1> 68<div class="contents local topic" id="contents"> 69<ul class="simple"> 70<li><a class="reference internal" href="#interface-and-glue-code" id="id1">Interface and Glue Code</a></li> 71<li><a class="reference internal" href="#first-implementation" id="id2">First Implementation</a></li> 72<li><a class="reference internal" href="#using-locators" id="id3">Using Locators</a></li> 73<li><a class="reference internal" href="#creating-a-generic-version-of-gil-algorithms" id="id4">Creating a Generic Version of GIL Algorithms</a></li> 74<li><a class="reference internal" href="#image-view-transformations" id="id5">Image View Transformations</a></li> 75<li><a class="reference internal" href="#d-pixel-iterators" id="id6">1D pixel iterators</a></li> 76<li><a class="reference internal" href="#stl-equivalent-algorithms" id="id7">STL Equivalent Algorithms</a></li> 77<li><a class="reference internal" href="#color-conversion" id="id8">Color Conversion</a></li> 78<li><a class="reference internal" href="#image" id="id9">Image</a></li> 79<li><a class="reference internal" href="#virtual-image-views" id="id10">Virtual Image Views</a></li> 80<li><a class="reference internal" href="#run-time-specified-images-and-image-views" id="id11">Run-Time Specified Images and Image Views</a></li> 81<li><a class="reference internal" href="#conclusion" id="id12">Conclusion</a></li> 82</ul> 83</div> 84<p>This comprehensive (and long) tutorial will walk you through an example of 85using GIL to compute the image gradients.</p> 86<p>We will start with some very simple and non-generic code and make it more 87generic as we go along. Let us start with a horizontal gradient and use the 88simplest possible approximation to a gradient - central difference.</p> 89<p>The gradient at pixel x can be approximated with the half-difference of its 90two neighboring pixels:</p> 91<div class="highlight-c++"><div class="highlight"><pre><span class="n">D</span><span class="p">[</span><span class="n">x</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span><span class="n">I</span><span class="p">[</span><span class="n">x</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">-</span> <span class="n">I</span><span class="p">[</span><span class="n">x</span><span class="o">+</span><span class="mi">1</span><span class="p">])</span> <span class="o">/</span> <span class="mi">2</span> 92</pre></div> 93</div> 94<p>For simplicity, we will also ignore the boundary cases - the pixels along the 95edges of the image for which one of the neighbors is not defined. The focus of 96this document is how to use GIL, not how to create a good gradient generation 97algorithm.</p> 98<div class="section" id="interface-and-glue-code"> 99<h2><a class="toc-backref" href="#id1">Interface and Glue Code</a></h2> 100<p>Let us first start with 8-bit unsigned grayscale image as the input and 8-bit 101signed grayscale image as the output.</p> 102<p>Here is how the interface to our algorithm looks like:</p> 103<div class="highlight-cpp"><div class="highlight"><pre><span class="cp">#include</span> <span class="cpf"><boost/gil.hpp></span><span class="cp"></span> 104<span class="k">using</span> <span class="k">namespace</span> <span class="n">boost</span><span class="o">::</span><span class="n">gil</span><span class="p">;</span> 105 106<span class="kt">void</span> <span class="nf">x_gradient</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 107<span class="p">{</span> 108 <span class="n">assert</span><span class="p">(</span><span class="n">src</span><span class="p">.</span><span class="n">dimensions</span><span class="p">()</span> <span class="o">==</span> <span class="n">dst</span><span class="p">.</span><span class="n">dimensions</span><span class="p">());</span> 109 <span class="p">...</span> <span class="c1">// compute the gradient</span> 110<span class="p">}</span> 111</pre></div> 112</div> 113<p><code class="docutils literal"><span class="pre">gray8c_view_t</span></code> is the type of the source image view - an 8-bit grayscale 114view, whose pixels are read-only (denoted by the “c”).</p> 115<p>The output is a grayscale view with a 8-bit signed (denoted by the “s”) 116integer channel type. See Appendix 1 for the complete convention GIL uses to 117name concrete types.</p> 118<p>GIL makes a distinction between an image and an image view. 119A GIL <strong>image view</strong>, is a shallow, lightweight view of a rectangular grid of 120pixels. It provides access to the pixels but does not own the pixels. 121Copy-constructing a view does not deep-copy the pixels. Image views do not 122propagate their constness to the pixels and should always be taken by a const 123reference. Whether a view is mutable or read-only (immutable) is a property of 124the view type.</p> 125<p>A GIL <cite>image</cite>, on the other hand, is a view with associated ownership. 126It is a container of pixels; its constructor/destructor allocates/deallocates 127the pixels, its copy-constructor performs deep-copy of the pixels and its 128<code class="docutils literal"><span class="pre">operator==</span></code> performs deep-compare of the pixels. Images also propagate 129their constness to their pixels - a constant reference to an image will not 130allow for modifying its pixels.</p> 131<p>Most GIL algorithms operate on image views; images are rarely 132needed. GIL’s design is very similar to that of the STL. The STL 133equivalent of GIL’s image is a container, like <code class="docutils literal"><span class="pre">std::vector</span></code>, 134whereas GIL’s image view corresponds to STL range, which is often 135represented with a pair of iterators. STL algorithms operate on 136ranges, just like GIL algorithms operate on image views.</p> 137<p>GIL’s image views can be constructed from raw data - the dimensions, 138the number of bytes per row and the pixels, which for chunky views are 139represented with one pointer. Here is how to provide the glue between 140your code and GIL:</p> 141<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">ComputeXGradientGray8</span><span class="p">(</span> 142 <span class="kt">unsigned</span> <span class="kt">char</span> <span class="k">const</span><span class="o">*</span> <span class="n">src_pixels</span><span class="p">,</span> <span class="kt">ptrdiff_t</span> <span class="n">src_row_bytes</span><span class="p">,</span> 143 <span class="kt">int</span> <span class="n">w</span><span class="p">,</span> <span class="kt">int</span> <span class="n">h</span><span class="p">,</span> 144 <span class="kt">signed</span> <span class="kt">char</span><span class="o">*</span> <span class="n">dst_pixels</span><span class="p">,</span> <span class="kt">ptrdiff_t</span> <span class="n">dst_row_bytes</span><span class="p">)</span> 145<span class="p">{</span> 146 <span class="n">gray8c_view_t</span> <span class="n">src</span> <span class="o">=</span> <span class="n">interleaved_view</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="p">(</span><span class="n">gray8_pixel_t</span> <span class="k">const</span><span class="o">*</span><span class="p">)</span><span class="n">src_pixels</span><span class="p">,</span> <span class="n">src_row_bytes</span><span class="p">);</span> 147 <span class="n">gray8s_view_t</span> <span class="n">dst</span> <span class="o">=</span> <span class="n">interleaved_view</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="p">(</span><span class="n">gray8s_pixel_t</span><span class="o">*</span><span class="p">)</span><span class="n">dst_pixels</span><span class="p">,</span> <span class="n">dst_row_bytes</span><span class="p">);</span> 148 <span class="n">x_gradient</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">dst</span><span class="p">);</span> 149<span class="p">}</span> 150</pre></div> 151</div> 152<p>This glue code is very fast and views are lightweight - in the above example 153the views have a size of 16 bytes. They consist of a pointer to the top left 154pixel and three integers - the width, height, and number of bytes per row.</p> 155</div> 156<div class="section" id="first-implementation"> 157<h2><a class="toc-backref" href="#id2">First Implementation</a></h2> 158<p>Focusing on simplicity at the expense of speed, we can compute the horizontal 159gradient like this:</p> 160<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">x_gradient</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 161<span class="p">{</span> 162 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">y</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">y</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">();</span> <span class="o">++</span><span class="n">y</span><span class="p">)</span> 163 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">x</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span> <span class="n">x</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">;</span> <span class="o">++</span><span class="n">x</span><span class="p">)</span> 164 <span class="n">dst</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span> <span class="o">=</span> <span class="p">(</span><span class="n">src</span><span class="p">(</span><span class="n">x</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="n">y</span><span class="p">)</span> <span class="o">-</span> <span class="n">src</span><span class="p">(</span><span class="n">x</span><span class="o">+</span><span class="mi">1</span><span class="p">,</span> <span class="n">y</span><span class="p">))</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 165<span class="p">}</span> 166</pre></div> 167</div> 168<p>We use image view’s <code class="docutils literal"><span class="pre">operator(x,y)</span></code> to get a reference to the pixel at a 169given location and we set it to the half-difference of its left and right 170neighbors. <code class="docutils literal"><span class="pre">operator()</span></code> returns a reference to a grayscale pixel. 171A grayscale pixel is convertible to its channel type (<code class="docutils literal"><span class="pre">unsigned</span> <span class="pre">char</span></code> for 172<code class="docutils literal"><span class="pre">src</span></code>) and it can be copy-constructed from a channel. 173(This is only true for grayscale pixels).</p> 174<p>While the above code is easy to read, it is not very fast, because the binary 175<code class="docutils literal"><span class="pre">operator()</span></code> computes the location of the pixel in a 2D grid, which involves 176addition and multiplication. Here is a faster version of the above:</p> 177<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">x_gradient</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 178<span class="p">{</span> 179 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">y</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">y</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">();</span> <span class="o">++</span><span class="n">y</span><span class="p">)</span> 180 <span class="p">{</span> 181 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">src_it</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 182 <span class="n">gray8s_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">dst_it</span> <span class="o">=</span> <span class="n">dst</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 183 184 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">x</span><span class="o">=</span><span class="mi">1</span><span class="p">;</span> <span class="n">x</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">;</span> <span class="o">++</span><span class="n">x</span><span class="p">)</span> 185 <span class="n">dst_it</span><span class="p">[</span><span class="n">x</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span><span class="n">src_it</span><span class="p">[</span><span class="n">x</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">-</span> <span class="n">src_it</span><span class="p">[</span><span class="n">x</span><span class="o">+</span><span class="mi">1</span><span class="p">])</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 186 <span class="p">}</span> 187<span class="p">}</span> 188</pre></div> 189</div> 190<p>We use pixel iterators initialized at the beginning of each row. GIL’s 191iterators are Random Access Traversal iterators. If you are not 192familiar with random access iterators, think of them as if they were 193pointers. In fact, in the above example the two iterator types are raw 194C pointers and their <code class="docutils literal"><span class="pre">operator[]</span></code> is a fast pointer indexing 195operator.</p> 196<p>The code to compute gradient in the vertical direction is very 197similar:</p> 198<p>Instead of looping over the rows, we loop over each column and create a 199<code class="docutils literal"><span class="pre">y_iterator</span></code>, an iterator moving vertically. In this case a simple pointer 200cannot be used because the distance between two adjacent pixels equals the 201number of bytes in each row of the image. GIL uses here a special step 202iterator class whose size is 8 bytes - it contains a raw C pointer and a step. 203Its <code class="docutils literal"><span class="pre">operator[]</span></code> multiplies the index by its step.</p> 204<p>The above version of <code class="docutils literal"><span class="pre">y_gradient</span></code>, however, is much slower (easily an order 205of magnitude slower) than <code class="docutils literal"><span class="pre">x_gradient</span></code> because of the memory access pattern; 206traversing an image vertically results in lots of cache misses. A much more 207efficient and cache-friendly version will iterate over the columns in the inner 208loop:</p> 209<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">y_gradient</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 210<span class="p">{</span> 211 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">y</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span> <span class="n">y</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">;</span> <span class="o">++</span><span class="n">y</span><span class="p">)</span> 212 <span class="p">{</span> 213 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">src1_it</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="o">-</span><span class="mi">1</span><span class="p">);</span> 214 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">src2_it</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="o">+</span><span class="mi">1</span><span class="p">);</span> 215 <span class="n">gray8s_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">dst_it</span> <span class="o">=</span> <span class="n">dst</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 216 217 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">x</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">x</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">();</span> <span class="o">++</span><span class="n">x</span><span class="p">)</span> 218 <span class="p">{</span> 219 <span class="o">*</span><span class="n">dst_it</span> <span class="o">=</span> <span class="p">((</span><span class="o">*</span><span class="n">src1_it</span><span class="p">)</span> <span class="o">-</span> <span class="p">(</span><span class="o">*</span><span class="n">src2_it</span><span class="p">))</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 220 <span class="o">++</span><span class="n">dst_it</span><span class="p">;</span> 221 <span class="o">++</span><span class="n">src1_it</span><span class="p">;</span> 222 <span class="o">++</span><span class="n">src2_it</span><span class="p">;</span> 223 <span class="p">}</span> 224 <span class="p">}</span> 225<span class="p">}</span> 226</pre></div> 227</div> 228<p>This sample code also shows an alternative way of using pixel iterators - 229instead of <code class="docutils literal"><span class="pre">operator[]</span></code> one could use increments and dereferences.</p> 230</div> 231<div class="section" id="using-locators"> 232<h2><a class="toc-backref" href="#id3">Using Locators</a></h2> 233<p>Unfortunately this cache-friendly version requires the extra hassle of 234maintaining two separate iterators in the source view. For every pixel, we 235want to access its neighbors above and below it. Such relative access can be 236done with GIL locators:</p> 237<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">y_gradient</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 238<span class="p">{</span> 239 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">xy_locator</span> <span class="n">src_loc</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">xy_at</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span><span class="mi">1</span><span class="p">);</span> 240 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">y</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span> <span class="n">y</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">;</span> <span class="o">++</span><span class="n">y</span><span class="p">)</span> 241 <span class="p">{</span> 242 <span class="n">gray8s_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">dst_it</span> <span class="o">=</span> <span class="n">dst</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 243 244 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">x</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">x</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">();</span> <span class="o">++</span><span class="n">x</span><span class="p">)</span> 245 <span class="p">{</span> 246 <span class="p">(</span><span class="o">*</span><span class="n">dst_it</span><span class="p">)</span> <span class="o">=</span> <span class="p">(</span><span class="n">src_loc</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span><span class="o">-</span><span class="mi">1</span><span class="p">)</span> <span class="o">-</span> <span class="n">src_loc</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span><span class="mi">1</span><span class="p">))</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 247 <span class="o">++</span><span class="n">dst_it</span><span class="p">;</span> 248 <span class="o">++</span><span class="n">src_loc</span><span class="p">.</span><span class="n">x</span><span class="p">();</span> <span class="c1">// each dimension can be advanced separately</span> 249 <span class="p">}</span> 250 <span class="n">src_loc</span><span class="o">+=</span><span class="n">point</span><span class="o"><</span><span class="n">std</span><span class="o">::</span><span class="kt">ptrdiff_t</span><span class="o">></span><span class="p">(</span><span class="o">-</span><span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">(),</span> <span class="mi">1</span><span class="p">);</span> <span class="c1">// carriage return</span> 251 <span class="p">}</span> 252<span class="p">}</span> 253</pre></div> 254</div> 255<p>The first line creates a locator pointing to the first pixel of the 256second row of the source view. A GIL pixel locator is very similar to 257an iterator, except that it can move both horizontally and 258vertically. <code class="docutils literal"><span class="pre">src_loc.x()</span></code> and <code class="docutils literal"><span class="pre">src_loc.y()</span></code> return references to a 259horizontal and a vertical iterator respectively, which can be used to 260move the locator along the desired dimension, as shown 261above. Additionally, the locator can be advanced in both dimensions 262simultaneously using its <code class="docutils literal"><span class="pre">operator+=</span></code> and <code class="docutils literal"><span class="pre">operator-=</span></code>. Similar to 263image views, locators provide binary <code class="docutils literal"><span class="pre">operator()</span></code> which returns a 264reference to a pixel with a relative offset to the current locator 265position. For example, <code class="docutils literal"><span class="pre">src_loc(0,1)</span></code> returns a reference to the 266neighbor below the current pixel. Locators are very lightweight 267objects - in the above example the locator has a size of 8 bytes - it 268consists of a raw pointer to the current pixel and an int indicating 269the number of bytes from one row to the next (which is the step when 270moving vertically). The call to <code class="docutils literal"><span class="pre">++src_loc.x()</span></code> corresponds to a 271single C pointer increment. However, the example above performs more 272computations than necessary. The code <code class="docutils literal"><span class="pre">src_loc(0,1)</span></code> has to compute 273the offset of the pixel in two dimensions, which is slow. Notice 274though that the offset of the two neighbors is the same, regardless of 275the pixel location. To improve the performance, GIL can cache and 276reuse this offset:</p> 277<div class="highlight-c++"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">y_gradient</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 278<span class="p">{</span> 279 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">xy_locator</span> <span class="n">src_loc</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">xy_at</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span><span class="mi">1</span><span class="p">);</span> 280 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">xy_locator</span><span class="o">::</span><span class="n">cached_location_t</span> <span class="n">above</span> <span class="o">=</span> <span class="n">src_loc</span><span class="p">.</span><span class="n">cache_location</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span><span class="o">-</span><span class="mi">1</span><span class="p">);</span> 281 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">xy_locator</span><span class="o">::</span><span class="n">cached_location_t</span> <span class="n">below</span> <span class="o">=</span> <span class="n">src_loc</span><span class="p">.</span><span class="n">cache_location</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="mi">1</span><span class="p">);</span> 282 283 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">y</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span> <span class="n">y</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">;</span> <span class="o">++</span><span class="n">y</span><span class="p">)</span> 284 <span class="p">{</span> 285 <span class="n">gray8s_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">dst_it</span> <span class="o">=</span> <span class="n">dst</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 286 287 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">x</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">x</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">();</span> <span class="o">++</span><span class="n">x</span><span class="p">)</span> 288 <span class="p">{</span> 289 <span class="p">(</span><span class="o">*</span><span class="n">dst_it</span><span class="p">)</span> <span class="o">=</span> <span class="p">(</span><span class="n">src_loc</span><span class="p">[</span><span class="n">above</span><span class="p">]</span> <span class="o">-</span> <span class="n">src_loc</span><span class="p">[</span><span class="n">below</span><span class="p">])</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 290 <span class="o">++</span><span class="n">dst_it</span><span class="p">;</span> 291 <span class="o">++</span><span class="n">src_loc</span><span class="p">.</span><span class="n">x</span><span class="p">();</span> 292 <span class="p">}</span> 293 <span class="n">src_loc</span><span class="o">+=</span><span class="n">point</span><span class="o"><</span><span class="n">std</span><span class="o">::</span><span class="kt">ptrdiff_t</span><span class="o">></span><span class="p">(</span><span class="o">-</span><span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">(),</span> <span class="mi">1</span><span class="p">);</span> 294 <span class="p">}</span> 295<span class="p">}</span> 296</pre></div> 297</div> 298<p>In this example <code class="docutils literal"><span class="pre">src_loc[above]</span></code> corresponds to a fast pointer indexing 299operation and the code is efficient.</p> 300</div> 301<div class="section" id="creating-a-generic-version-of-gil-algorithms"> 302<h2><a class="toc-backref" href="#id4">Creating a Generic Version of GIL Algorithms</a></h2> 303<p>Let us make our <code class="docutils literal"><span class="pre">x_gradient</span></code> more generic. It should work with any image 304views, as long as they have the same number of channels. The gradient 305operation is to be computed for each channel independently.</p> 306<p>Here is how the new interface looks like:</p> 307<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 308<span class="kt">void</span> <span class="n">x_gradient</span><span class="p">(</span><span class="k">const</span> <span class="n">SrcView</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 309<span class="p">{</span> 310 <span class="n">gil_function_requires</span><span class="o"><</span><span class="n">ImageViewConcept</span><span class="o"><</span><span class="n">SrcView</span><span class="o">></span> <span class="o">></span><span class="p">();</span> 311 <span class="n">gil_function_requires</span><span class="o"><</span><span class="n">MutableImageViewConcept</span><span class="o"><</span><span class="n">DstView</span><span class="o">></span> <span class="o">></span><span class="p">();</span> 312 <span class="n">gil_function_requires</span> 313 <span class="o"><</span> 314 <span class="n">ColorSpacesCompatibleConcept</span> 315 <span class="o"><</span> 316 <span class="k">typename</span> <span class="n">color_space_type</span><span class="o"><</span><span class="n">SrcView</span><span class="o">>::</span><span class="n">type</span><span class="p">,</span> 317 <span class="k">typename</span> <span class="n">color_space_type</span><span class="o"><</span><span class="n">DstView</span><span class="o">>::</span><span class="n">type</span> 318 <span class="o">></span> 319 <span class="o">></span><span class="p">();</span> 320 321 <span class="p">...</span> <span class="c1">// compute the gradient</span> 322<span class="p">}</span> 323</pre></div> 324</div> 325<p>The new algorithm now takes the types of the input and output image 326views as template parameters. That allows using both built-in GIL 327image views, as well as any user-defined image view classes. The 328first three lines are optional; they use <code class="docutils literal"><span class="pre">boost::concept_check</span></code> to 329ensure that the two arguments are valid GIL image views, that the 330second one is mutable and that their color spaces are compatible 331(i.e. have the same set of channels).</p> 332<p>GIL does not require using its own built-in constructs. You are free 333to use your own channels, color spaces, iterators, locators, views and 334images. However, to work with the rest of GIL they have to satisfy a 335set of requirements; in other words, they have to e model the 336corresponding GIL _concept_. GIL’s concepts are defined in the user 337guide.</p> 338<p>One of the biggest drawbacks of using templates and generic 339programming in C++ is that compile errors can be very difficult to 340comprehend. This is a side-effect of the lack of early type 341checking - a generic argument may not satisfy the requirements of a 342function, but the incompatibility may be triggered deep into a nested 343call, in code unfamiliar and hardly related to the problem. GIL uses 344<code class="docutils literal"><span class="pre">boost::concept_check</span></code> to mitigate this problem. The above three 345lines of code check whether the template parameters are valid models 346of their corresponding concepts. If a model is incorrect, the compile 347error will be inside <code class="docutils literal"><span class="pre">gil_function_requires</span></code>, which is much closer 348to the problem and easier to track. Furthermore, such checks get 349compiled out and have zero performance overhead. The disadvantage of 350using concept checks is the sometimes severe impact they have on 351compile time. This is why GIL performs concept checks only in debug 352mode, and only if <code class="docutils literal"><span class="pre">BOOST_GIL_USE_CONCEPT_CHECK</span></code> is defined (off by 353default).</p> 354<p>The body of the generic function is very similar to that of the 355concrete one. The biggest difference is that we need to loop over the 356channels of the pixel and compute the gradient for each channel:</p> 357<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 358<span class="kt">void</span> <span class="n">x_gradient</span><span class="p">(</span><span class="k">const</span> <span class="n">SrcView</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 359<span class="p">{</span> 360 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">y</span><span class="o">=</span><span class="mi">0</span><span class="p">;</span> <span class="n">y</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">();</span> <span class="o">++</span><span class="n">y</span><span class="p">)</span> 361 <span class="p">{</span> 362 <span class="k">typename</span> <span class="n">SrcView</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">src_it</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 363 <span class="k">typename</span> <span class="n">DstView</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">dst_it</span> <span class="o">=</span> <span class="n">dst</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 364 365 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">x</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span> <span class="n">x</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">;</span> <span class="o">++</span><span class="n">x</span><span class="p">)</span> 366 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">c</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">c</span> <span class="o"><</span> <span class="n">num_channels</span><span class="o"><</span><span class="n">SrcView</span><span class="o">>::</span><span class="n">value</span><span class="p">;</span> <span class="o">++</span><span class="n">c</span><span class="p">)</span> 367 <span class="n">dst_it</span><span class="p">[</span><span class="n">x</span><span class="p">][</span><span class="n">c</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span><span class="n">src_it</span><span class="p">[</span><span class="n">x</span><span class="o">-</span><span class="mi">1</span><span class="p">][</span><span class="n">c</span><span class="p">]</span><span class="o">-</span> <span class="n">src_it</span><span class="p">[</span><span class="n">x</span><span class="o">+</span><span class="mi">1</span><span class="p">][</span><span class="n">c</span><span class="p">])</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 368 <span class="p">}</span> 369<span class="p">}</span> 370</pre></div> 371</div> 372<p>Having an explicit loop for each channel could be a performance problem. 373GIL allows us to abstract out such per-channel operations:</p> 374<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">Out</span><span class="o">></span> 375<span class="k">struct</span> <span class="n">halfdiff_cast_channels</span> 376<span class="p">{</span> 377 <span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">T</span><span class="o">></span> <span class="n">Out</span> <span class="k">operator</span><span class="p">()(</span><span class="n">T</span> <span class="k">const</span><span class="o">&</span> <span class="n">in1</span><span class="p">,</span> <span class="n">T</span> <span class="k">const</span><span class="o">&</span> <span class="n">in2</span><span class="p">)</span> <span class="k">const</span> 378 <span class="p">{</span> 379 <span class="k">return</span> <span class="n">Out</span><span class="p">((</span><span class="n">in1</span> <span class="o">-</span> <span class="n">in2</span><span class="p">)</span> <span class="o">/</span> <span class="mi">2</span><span class="p">);</span> 380 <span class="p">}</span> 381<span class="p">};</span> 382 383<span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 384<span class="kt">void</span> <span class="n">x_gradient</span><span class="p">(</span><span class="k">const</span> <span class="n">SrcView</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 385<span class="p">{</span> 386 <span class="k">typedef</span> <span class="k">typename</span> <span class="n">channel_type</span><span class="o"><</span><span class="n">DstView</span><span class="o">>::</span><span class="n">type</span> <span class="n">dst_channel_t</span><span class="p">;</span> 387 388 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">y</span><span class="o">=</span><span class="mi">0</span><span class="p">;</span> <span class="n">y</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">();</span> <span class="o">++</span><span class="n">y</span><span class="p">)</span> 389 <span class="p">{</span> 390 <span class="k">typename</span> <span class="n">SrcView</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">src_it</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 391 <span class="k">typename</span> <span class="n">DstView</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">dst_it</span> <span class="o">=</span> <span class="n">dst</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 392 393 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">x</span><span class="o">=</span><span class="mi">1</span><span class="p">;</span> <span class="n">x</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">()</span> <span class="o">-</span> <span class="mi">1</span><span class="p">;</span> <span class="o">++</span><span class="n">x</span><span class="p">)</span> 394 <span class="p">{</span> 395 <span class="n">static_transform</span><span class="p">(</span><span class="n">src_it</span><span class="p">[</span><span class="n">x</span><span class="o">-</span><span class="mi">1</span><span class="p">],</span> <span class="n">src_it</span><span class="p">[</span><span class="n">x</span><span class="o">+</span><span class="mi">1</span><span class="p">],</span> <span class="n">dst_it</span><span class="p">[</span><span class="n">x</span><span class="p">],</span> 396 <span class="n">halfdiff_cast_channels</span><span class="o"><</span><span class="n">dst_channel_t</span><span class="o">></span><span class="p">());</span> 397 <span class="p">}</span> 398 <span class="p">}</span> 399<span class="p">}</span> 400</pre></div> 401</div> 402<p>The <code class="docutils literal"><span class="pre">static_transform</span></code> is an example of a channel-level GIL algorithm. 403Other such algorithms are <code class="docutils literal"><span class="pre">static_generate</span></code>, <code class="docutils literal"><span class="pre">static_fill</span></code> and 404<code class="docutils literal"><span class="pre">static_for_each</span></code>. They are the channel-level equivalents of STL 405<code class="docutils literal"><span class="pre">generate</span></code>, <code class="docutils literal"><span class="pre">transform</span></code>, <code class="docutils literal"><span class="pre">fill</span></code> and <code class="docutils literal"><span class="pre">for_each</span></code> respectively. 406GIL channel algorithms use static recursion to unroll the loops; they never 407loop over the channels explicitly.</p> 408<p>Note that sometimes modern compilers (at least Visual Studio 8) already unroll 409channel-level loops, such as the one above. However, another advantage of 410using GIL’s channel-level algorithms is that they pair the channels 411semantically, not based on their order in memory. For example, the above 412example will properly match an RGB source with a BGR destination.</p> 413<p>Here is how we can use our generic version with images of different types:</p> 414<div class="highlight-cpp"><div class="highlight"><pre><span class="c1">// Calling with 16-bit grayscale data</span> 415<span class="kt">void</span> <span class="nf">XGradientGray16_Gray32</span><span class="p">(</span> 416 <span class="kt">unsigned</span> <span class="kt">short</span> <span class="k">const</span><span class="o">*</span> <span class="n">src_pixels</span><span class="p">,</span> <span class="kt">ptrdiff_t</span> <span class="n">src_row_bytes</span><span class="p">,</span> 417 <span class="kt">int</span> <span class="n">w</span><span class="p">,</span> <span class="kt">int</span> <span class="n">h</span><span class="p">,</span> 418 <span class="kt">signed</span> <span class="kt">int</span><span class="o">*</span> <span class="n">dst_pixels</span><span class="p">,</span> <span class="kt">ptrdiff_t</span> <span class="n">dst_row_bytes</span><span class="p">)</span> 419<span class="p">{</span> 420 <span class="n">gray16c_view_t</span> <span class="n">src</span><span class="o">=</span><span class="n">interleaved_view</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="p">(</span><span class="n">gray16_pixel_t</span> <span class="k">const</span><span class="o">*</span><span class="p">)</span><span class="n">src_pixels</span><span class="p">,</span> <span class="n">src_row_bytes</span><span class="p">);</span> 421 <span class="n">gray32s_view_t</span> <span class="n">dst</span><span class="o">=</span><span class="n">interleaved_view</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="p">(</span><span class="n">gray32s_pixel_t</span><span class="o">*</span><span class="p">)</span><span class="n">dst_pixels</span><span class="p">,</span> <span class="n">dst_row_bytes</span><span class="p">);</span> 422 <span class="n">x_gradient</span><span class="p">(</span><span class="n">src</span><span class="p">,</span><span class="n">dst</span><span class="p">);</span> 423<span class="p">}</span> 424 425<span class="c1">// Calling with 8-bit RGB data into 16-bit BGR</span> 426<span class="kt">void</span> <span class="nf">XGradientRGB8_BGR16</span><span class="p">(</span> 427 <span class="kt">unsigned</span> <span class="kt">char</span> <span class="k">const</span><span class="o">*</span> <span class="n">src_pixels</span><span class="p">,</span> <span class="kt">ptrdiff_t</span> <span class="n">src_row_bytes</span><span class="p">,</span> 428 <span class="kt">int</span> <span class="n">w</span><span class="p">,</span> <span class="kt">int</span> <span class="n">h</span><span class="p">,</span> 429 <span class="kt">signed</span> <span class="kt">short</span><span class="o">*</span> <span class="n">dst_pixels</span><span class="p">,</span> <span class="kt">ptrdiff_t</span> <span class="n">dst_row_bytes</span><span class="p">)</span> 430<span class="p">{</span> 431 <span class="n">rgb8c_view_t</span> <span class="n">src</span> <span class="o">=</span> <span class="n">interleaved_view</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="p">(</span><span class="n">rgb8_pixel_t</span> <span class="k">const</span><span class="o">*</span><span class="p">)</span><span class="n">src_pixels</span><span class="p">,</span> <span class="n">src_row_bytes</span><span class="p">);</span> 432 <span class="n">bgr16s_view_t</span> <span class="n">dst</span> <span class="o">=</span> <span class="n">interleaved_view</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="p">(</span><span class="n">bgr16s_pixel_t</span><span class="o">*</span><span class="p">)</span><span class="n">dst_pixels</span><span class="p">,</span> <span class="n">dst_row_bytes</span><span class="p">);</span> 433 <span class="n">x_gradient</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">dst</span><span class="p">);</span> 434<span class="p">}</span> 435 436<span class="c1">// Either or both the source and the destination could be planar - the gradient code does not change</span> 437<span class="kt">void</span> <span class="nf">XGradientPlanarRGB8_RGB32</span><span class="p">(</span> 438 <span class="kt">unsigned</span> <span class="kt">short</span> <span class="k">const</span><span class="o">*</span> <span class="n">src_r</span><span class="p">,</span> <span class="kt">unsigned</span> <span class="kt">short</span> <span class="k">const</span><span class="o">*</span> <span class="n">src_g</span><span class="p">,</span> <span class="kt">unsigned</span> <span class="kt">short</span> <span class="k">const</span><span class="o">*</span> <span class="n">src_b</span><span class="p">,</span> 439 <span class="kt">ptrdiff_t</span> <span class="n">src_row_bytes</span><span class="p">,</span> <span class="kt">int</span> <span class="n">w</span><span class="p">,</span> <span class="kt">int</span> <span class="n">h</span><span class="p">,</span> 440 <span class="kt">signed</span> <span class="kt">int</span><span class="o">*</span> <span class="n">dst_pixels</span><span class="p">,</span> <span class="kt">ptrdiff_t</span> <span class="n">dst_row_bytes</span><span class="p">)</span> 441<span class="p">{</span> 442 <span class="n">rgb16c_planar_view_t</span> <span class="n">src</span> <span class="o">=</span> <span class="n">planar_rgb_view</span> <span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="n">src_r</span><span class="p">,</span> <span class="n">src_g</span><span class="p">,</span> <span class="n">src_b</span><span class="p">,</span> <span class="n">src_row_bytes</span><span class="p">);</span> 443 <span class="n">rgb32s_view_t</span> <span class="n">dst</span> <span class="o">=</span> <span class="n">interleaved_view</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">,(</span><span class="n">rgb32s_pixel_t</span><span class="o">*</span><span class="p">)</span><span class="n">dst_pixels</span><span class="p">,</span> <span class="n">dst_row_bytes</span><span class="p">);</span> 444 <span class="n">x_gradient</span><span class="p">(</span><span class="n">src</span><span class="p">,</span><span class="n">dst</span><span class="p">);</span> 445<span class="p">}</span> 446</pre></div> 447</div> 448<p>As these examples illustrate, both the source and the destination can be 449interleaved or planar, of any channel depth (assuming the destination channel 450is assignable to the source), and of any compatible color spaces.</p> 451<p>GIL 2.1 can also natively represent images whose channels are not 452byte-aligned, such as 6-bit RGB222 image or a 1-bit Gray1 image. 453GIL algorithms apply to these images natively. See the design guide or sample 454files for more on using such images.</p> 455</div> 456<div class="section" id="image-view-transformations"> 457<h2><a class="toc-backref" href="#id5">Image View Transformations</a></h2> 458<p>One way to compute the y-gradient is to rotate the image by 90 degrees, 459compute the x-gradient and rotate the result back. 460Here is how to do this in GIL:</p> 461<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 462<span class="kt">void</span> <span class="n">y_gradient</span><span class="p">(</span><span class="k">const</span> <span class="n">SrcView</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 463<span class="p">{</span> 464 <span class="n">x_gradient</span><span class="p">(</span><span class="n">rotated90ccw_view</span><span class="p">(</span><span class="n">src</span><span class="p">),</span> <span class="n">rotated90ccw_view</span><span class="p">(</span><span class="n">dst</span><span class="p">));</span> 465<span class="p">}</span> 466</pre></div> 467</div> 468<p><code class="docutils literal"><span class="pre">rotated90ccw_view</span></code> takes an image view and returns an image view 469representing 90-degrees counter-clockwise rotation of its input. It is 470an example of a GIL view transformation function. GIL provides a 471variety of transformation functions that can perform any axis-aligned 472rotation, transpose the view, flip it vertically or horizontally, 473extract a rectangular subimage, perform color conversion, subsample 474view, etc. The view transformation functions are fast and shallow - 475they don’t copy the pixels, they just change the “coordinate system” 476of accessing the pixels. <code class="docutils literal"><span class="pre">rotated90cw_view</span></code>, for example, returns a 477view whose horizontal iterators are the vertical iterators of the 478original view. The above code to compute <code class="docutils literal"><span class="pre">y_gradient</span></code> is slow 479because of the memory access pattern; using <code class="docutils literal"><span class="pre">rotated90cw_view</span></code> does 480not make it any slower.</p> 481<p>Another example: suppose we want to compute the gradient of the N-th 482channel of a color image. Here is how to do that:</p> 483<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 484<span class="kt">void</span> <span class="n">nth_channel_x_gradient</span><span class="p">(</span><span class="k">const</span> <span class="n">SrcView</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="kt">int</span> <span class="n">n</span><span class="p">,</span> <span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 485<span class="p">{</span> 486 <span class="n">x_gradient</span><span class="p">(</span><span class="n">nth_channel_view</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">n</span><span class="p">),</span> <span class="n">dst</span><span class="p">);</span> 487<span class="p">}</span> 488</pre></div> 489</div> 490<p><code class="docutils literal"><span class="pre">nth_channel_view</span></code> is a view transformation function that takes any 491view and returns a single-channel (grayscale) view of its N-th 492channel. For interleaved RGB view, for example, the returned view is 493a step view - a view whose horizontal iterator skips over two channels 494when incremented. If applied on a planar RGB view, the returned type 495is a simple grayscale view whose horizontal iterator is a C pointer. 496Image view transformation functions can be piped together. For 497example, to compute the y gradient of the second channel of the even 498pixels in the view, use:</p> 499<div class="highlight-cpp"><div class="highlight"><pre><span class="n">y_gradient</span><span class="p">(</span><span class="n">subsampled_view</span><span class="p">(</span><span class="n">nth_channel_view</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span> <span class="mi">2</span><span class="p">,</span><span class="mi">2</span><span class="p">),</span> <span class="n">dst</span><span class="p">);</span> 500</pre></div> 501</div> 502<p>GIL can sometimes simplify piped views. For example, two nested 503subsampled views (views that skip over pixels in X and in Y) can be 504represented as a single subsampled view whose step is the product of 505the steps of the two views.</p> 506</div> 507<div class="section" id="d-pixel-iterators"> 508<h2><a class="toc-backref" href="#id6">1D pixel iterators</a></h2> 509<p>Let’s go back to <code class="docutils literal"><span class="pre">x_gradient</span></code> one more time. Many image view 510algorithms apply the same operation for each pixel and GIL provides an 511abstraction to handle them. However, our algorithm has an unusual 512access pattern, as it skips the first and the last column. It would be 513nice and instructional to see how we can rewrite it in canonical 514form. The way to do that in GIL is to write a version that works for 515every pixel, but apply it only on the subimage that excludes the first 516and last column:</p> 517<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">x_gradient_unguarded</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 518<span class="p">{</span> 519 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">y</span><span class="o">=</span><span class="mi">0</span><span class="p">;</span> <span class="n">y</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">();</span> <span class="o">++</span><span class="n">y</span><span class="p">)</span> 520 <span class="p">{</span> 521 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">src_it</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 522 <span class="n">gray8s_view_t</span><span class="o">::</span><span class="n">x_iterator</span> <span class="n">dst_it</span> <span class="o">=</span> <span class="n">dst</span><span class="p">.</span><span class="n">row_begin</span><span class="p">(</span><span class="n">y</span><span class="p">);</span> 523 524 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">x</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">x</span> <span class="o"><</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">();</span> <span class="o">++</span><span class="n">x</span><span class="p">)</span> 525 <span class="n">dst_it</span><span class="p">[</span><span class="n">x</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span><span class="n">src_it</span><span class="p">[</span><span class="n">x</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">-</span> <span class="n">src_it</span><span class="p">[</span><span class="n">x</span><span class="o">+</span><span class="mi">1</span><span class="p">])</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 526 <span class="p">}</span> 527<span class="p">}</span> 528 529<span class="kt">void</span> <span class="nf">x_gradient</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 530<span class="p">{</span> 531 <span class="n">assert</span><span class="p">(</span><span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">()</span><span class="o">>=</span><span class="mi">2</span><span class="p">);</span> 532 <span class="n">x_gradient_unguarded</span><span class="p">(</span><span class="n">subimage_view</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">()</span><span class="o">-</span><span class="mi">2</span><span class="p">,</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">()),</span> 533 <span class="n">subimage_view</span><span class="p">(</span><span class="n">dst</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="n">src</span><span class="p">.</span><span class="n">width</span><span class="p">()</span><span class="o">-</span><span class="mi">2</span><span class="p">,</span> <span class="n">src</span><span class="p">.</span><span class="n">height</span><span class="p">()));</span> 534<span class="p">}</span> 535</pre></div> 536</div> 537<p><code class="docutils literal"><span class="pre">subimage_view</span></code> is another example of a GIL view transformation 538function. It takes a source view and a rectangular region (in this 539case, defined as x_min,y_min,width,height) and returns a view 540operating on that region of the source view. The above implementation 541has no measurable performance degradation from the version that 542operates on the original views.</p> 543<p>Now that <code class="docutils literal"><span class="pre">x_gradient_unguarded</span></code> operates on every pixel, we can 544rewrite it more compactly:</p> 545<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">x_gradient_unguarded</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 546<span class="p">{</span> 547 <span class="n">gray8c_view_t</span><span class="o">::</span><span class="n">iterator</span> <span class="n">src_it</span> <span class="o">=</span> <span class="n">src</span><span class="p">.</span><span class="n">begin</span><span class="p">();</span> 548 <span class="k">for</span> <span class="p">(</span><span class="n">gray8s_view_t</span><span class="o">::</span><span class="n">iterator</span> <span class="n">dst_it</span> <span class="o">=</span> <span class="n">dst</span><span class="p">.</span><span class="n">begin</span><span class="p">();</span> <span class="n">dst_it</span><span class="o">!=</span><span class="n">dst</span><span class="p">.</span><span class="n">end</span><span class="p">();</span> <span class="o">++</span><span class="n">dst_it</span><span class="p">,</span> <span class="o">++</span><span class="n">src_it</span><span class="p">)</span> 549 <span class="o">*</span><span class="n">dst_it</span> <span class="o">=</span> <span class="p">(</span><span class="n">src_it</span><span class="p">.</span><span class="n">x</span><span class="p">()[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">-</span> <span class="n">src_it</span><span class="p">.</span><span class="n">x</span><span class="p">()[</span><span class="mi">1</span><span class="p">])</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 550<span class="p">}</span> 551</pre></div> 552</div> 553<p>GIL image views provide <code class="docutils literal"><span class="pre">begin()</span></code> and <code class="docutils literal"><span class="pre">end()</span></code> methods that return 554one dimensional pixel iterators which iterate over each pixel in the 555view, left to right and top to bottom. They do a proper “carriage 556return” - they skip any unused bytes at the end of a row. As such, 557they are slightly suboptimal, because they need to keep track of their 558current position with respect to the end of the row. Their increment 559operator performs one extra check (are we at the end of the row?), a 560check that is avoided if two nested loops are used instead. These 561iterators have a method <code class="docutils literal"><span class="pre">x()</span></code> which returns the more lightweight 562horizontal iterator that we used previously. Horizontal iterators have 563no notion of the end of rows. In this case, the horizontal iterators 564are raw C pointers. In our example, we must use the horizontal 565iterators to access the two neighbors properly, since they could 566reside outside the image view.</p> 567</div> 568<div class="section" id="stl-equivalent-algorithms"> 569<h2><a class="toc-backref" href="#id7">STL Equivalent Algorithms</a></h2> 570<p>GIL provides STL equivalents of many algorithms. For example, 571<code class="docutils literal"><span class="pre">std::transform</span></code> is an STL algorithm that sets each element in a 572destination range the result of a generic function taking the 573corresponding element of the source range. In our example, we want to 574assign to each destination pixel the value of the half-difference of 575the horizontal neighbors of the corresponding source pixel. If we 576abstract that operation in a function object, we can use GIL’s 577<code class="docutils literal"><span class="pre">transform_pixel_positions</span></code> to do that:</p> 578<div class="highlight-cpp"><div class="highlight"><pre><span class="k">struct</span> <span class="n">half_x_difference</span> 579<span class="p">{</span> 580 <span class="kt">int</span> <span class="k">operator</span><span class="p">()(</span><span class="k">const</span> <span class="n">gray8c_loc_t</span><span class="o">&</span> <span class="n">src_loc</span><span class="p">)</span> <span class="k">const</span> 581 <span class="p">{</span> 582 <span class="k">return</span> <span class="p">(</span><span class="n">src_loc</span><span class="p">.</span><span class="n">x</span><span class="p">()[</span><span class="o">-</span><span class="mi">1</span><span class="p">]</span> <span class="o">-</span> <span class="n">src_loc</span><span class="p">.</span><span class="n">x</span><span class="p">()[</span><span class="mi">1</span><span class="p">])</span> <span class="o">/</span> <span class="mi">2</span><span class="p">;</span> 583 <span class="p">}</span> 584<span class="p">};</span> 585 586<span class="kt">void</span> <span class="nf">x_gradient_unguarded</span><span class="p">(</span><span class="n">gray8c_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 587<span class="p">{</span> 588 <span class="n">transform_pixel_positions</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">dst</span><span class="p">,</span> <span class="n">half_x_difference</span><span class="p">());</span> 589<span class="p">}</span> 590</pre></div> 591</div> 592<p>GIL provides the algorithms <code class="docutils literal"><span class="pre">for_each_pixel</span></code> and 593<code class="docutils literal"><span class="pre">transform_pixels</span></code> which are image view equivalents of STL 594<code class="docutils literal"><span class="pre">std::for_each</span></code> and <code class="docutils literal"><span class="pre">std::transform</span></code>. It also provides 595<code class="docutils literal"><span class="pre">for_each_pixel_position</span></code> and <code class="docutils literal"><span class="pre">transform_pixel_positions</span></code>, which 596instead of references to pixels, pass to the generic function pixel 597locators. This allows for more powerful functions that can use the 598pixel neighbors through the passed locators. GIL algorithms iterate 599through the pixels using the more efficient two nested loops (as 600opposed to the single loop using 1-D iterators)</p> 601</div> 602<div class="section" id="color-conversion"> 603<h2><a class="toc-backref" href="#id8">Color Conversion</a></h2> 604<p>Instead of computing the gradient of each color plane of an image, we 605often want to compute the gradient of the luminosity. In other words, 606we want to convert the color image to grayscale and compute the 607gradient of the result. Here how to compute the luminosity gradient of 608a 32-bit float RGB image:</p> 609<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">x_gradient_rgb_luminosity</span><span class="p">(</span><span class="n">rgb32fc_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 610<span class="p">{</span> 611 <span class="n">x_gradient</span><span class="p">(</span><span class="n">color_converted_view</span><span class="o"><</span><span class="n">gray8_pixel_t</span><span class="o">></span><span class="p">(</span><span class="n">src</span><span class="p">),</span> <span class="n">dst</span><span class="p">);</span> 612<span class="p">}</span> 613</pre></div> 614</div> 615<p><code class="docutils literal"><span class="pre">color_converted_view</span></code> is a GIL view transformation function that 616takes any image view and returns a view in a target color space and 617channel depth (specified as template parameters). In our example, it 618constructs an 8-bit integer grayscale view over 32-bit float RGB 619pixels. Like all other view transformation functions, 620<code class="docutils literal"><span class="pre">color_converted_view</span></code> is very fast and shallow. It doesn’t copy the 621data or perform any color conversion. Instead it returns a view that 622performs color conversion every time its pixels are accessed.</p> 623<p>In the generic version of this algorithm we might like to convert the 624color space to grayscale, but keep the channel depth the same. We do 625that by constructing the type of a GIL grayscale pixel with the same 626channel as the source, and color convert to that pixel type:</p> 627<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 628<span class="kt">void</span> <span class="n">x_luminosity_gradient</span><span class="p">(</span><span class="n">SrcView</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">DstView</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 629<span class="p">{</span> 630 <span class="k">using</span> <span class="n">gray_pixel_t</span> <span class="o">=</span> <span class="n">pixel</span><span class="o"><</span><span class="k">typename</span> <span class="n">channel_type</span><span class="o"><</span><span class="n">SrcView</span><span class="o">>::</span><span class="n">type</span><span class="p">,</span> <span class="n">gray_layout_t</span><span class="o">></span><span class="p">;</span> 631 <span class="n">x_gradient</span><span class="p">(</span><span class="n">color_converted_view</span><span class="o"><</span><span class="n">gray_pixel_t</span><span class="o">></span><span class="p">(</span><span class="n">src</span><span class="p">),</span> <span class="n">dst</span><span class="p">);</span> 632<span class="p">}</span> 633</pre></div> 634</div> 635<p>When the destination color space and channel type happens to be the 636same as the source one, color conversion is unnecessary. GIL detects 637this case and avoids calling the color conversion code at all - 638i.e. <code class="docutils literal"><span class="pre">color_converted_view</span></code> returns back the source view unchanged.</p> 639</div> 640<div class="section" id="image"> 641<h2><a class="toc-backref" href="#id9">Image</a></h2> 642<p>The above example has a performance problem - <code class="docutils literal"><span class="pre">x_gradient</span></code> 643dereferences most source pixels twice, which will cause the above code 644to perform color conversion twice. Sometimes it may be more efficient 645to copy the color converted image into a temporary buffer and use it 646to compute the gradient - that way color conversion is invoked once 647per pixel. Using our non-generic version we can do it like this:</p> 648<div class="highlight-cpp"><div class="highlight"><pre><span class="kt">void</span> <span class="nf">x_luminosity_gradient</span><span class="p">(</span><span class="n">rgb32fc_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">gray8s_view_t</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 649<span class="p">{</span> 650 <span class="n">gray8_image_t</span> <span class="n">ccv_image</span><span class="p">(</span><span class="n">src</span><span class="p">.</span><span class="n">dimensions</span><span class="p">());</span> 651 <span class="n">copy_pixels</span><span class="p">(</span><span class="n">color_converted_view</span><span class="o"><</span><span class="n">gray8_pixel_t</span><span class="o">></span><span class="p">(</span><span class="n">src</span><span class="p">),</span> <span class="n">view</span><span class="p">(</span><span class="n">ccv_image</span><span class="p">));</span> 652 653 <span class="n">x_gradient</span><span class="p">(</span><span class="n">const_view</span><span class="p">(</span><span class="n">ccv_image</span><span class="p">),</span> <span class="n">dst</span><span class="p">);</span> 654<span class="p">}</span> 655</pre></div> 656</div> 657<p>First we construct an 8-bit grayscale image with the same dimensions 658as our source. Then we copy a color-converted view of the source into 659the temporary image. Finally we use a read-only view of the temporary 660image in our <code class="docutils literal"><span class="pre">x_gradient</span> <span class="pre">algorithm</span></code>. As the example shows, GIL 661provides global functions <code class="docutils literal"><span class="pre">view</span></code> and <code class="docutils literal"><span class="pre">const_view</span></code> that take an 662image and return a mutable or an immutable view of its pixels.</p> 663<p>Creating a generic version of the above is a bit trickier:</p> 664<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 665<span class="kt">void</span> <span class="n">x_luminosity_gradient</span><span class="p">(</span><span class="k">const</span> <span class="n">SrcView</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 666<span class="p">{</span> 667 <span class="k">using</span> <span class="n">d_channel_t</span> <span class="o">=</span> <span class="k">typename</span> <span class="n">channel_type</span><span class="o"><</span><span class="n">DstView</span><span class="o">>::</span><span class="n">type</span><span class="p">;</span> 668 <span class="k">using</span> <span class="n">channel_t</span> <span class="o">=</span> <span class="k">typename</span> <span class="n">channel_convert_to_unsigned</span><span class="o"><</span><span class="n">d_channel_t</span><span class="o">>::</span><span class="n">type</span><span class="p">;</span> 669 <span class="k">using</span> <span class="n">gray_pixel_t</span> <span class="o">=</span> <span class="n">pixel</span><span class="o"><</span><span class="n">channel_t</span><span class="p">,</span> <span class="n">gray_layout_t</span><span class="o">></span><span class="p">;</span> 670 <span class="k">using</span> <span class="n">gray_image_t</span> <span class="o">=</span> <span class="n">image</span><span class="o"><</span><span class="n">gray_pixel_t</span><span class="p">,</span> <span class="nb">false</span><span class="o">></span><span class="p">;</span> 671 672 <span class="n">gray_image_t</span> <span class="nf">ccv_image</span><span class="p">(</span><span class="n">src</span><span class="p">.</span><span class="n">dimensions</span><span class="p">());</span> 673 <span class="n">copy_pixels</span><span class="p">(</span><span class="n">color_converted_view</span><span class="o"><</span><span class="n">gray_pixel_t</span><span class="o">></span><span class="p">(</span><span class="n">src</span><span class="p">),</span> <span class="n">view</span><span class="p">(</span><span class="n">ccv_image</span><span class="p">));</span> 674 <span class="n">x_gradient</span><span class="p">(</span><span class="n">const_view</span><span class="p">(</span><span class="n">ccv_image</span><span class="p">),</span> <span class="n">dst</span><span class="p">);</span> 675<span class="p">}</span> 676</pre></div> 677</div> 678<p>First we use the <code class="docutils literal"><span class="pre">channel_type</span></code> metafunction to get the channel type 679of the destination view. A metafunction is a function operating on 680types. In GIL metafunctions are class templates (declared with 681<code class="docutils literal"><span class="pre">struct</span></code> type specifier) which take their parameters as template 682parameters and return their result in a nested typedef called 683<code class="docutils literal"><span class="pre">type</span></code>. In this case, <code class="docutils literal"><span class="pre">channel_type</span></code> is a unary metafunction which 684in this example is called with the type of an image view and returns 685the type of the channel associated with that image view.</p> 686<p>GIL constructs that have an associated pixel type, such as pixels, 687pixel iterators, locators, views and images, all model 688<code class="docutils literal"><span class="pre">PixelBasedConcept</span></code>, which means that they provide a set of 689metafunctions to query the pixel properties, such as <code class="docutils literal"><span class="pre">channel_type</span></code>, 690<code class="docutils literal"><span class="pre">color_space_type</span></code>, <code class="docutils literal"><span class="pre">channel_mapping_type</span></code>, and <code class="docutils literal"><span class="pre">num_channels</span></code>.</p> 691<p>After we get the channel type of the destination view, we use another 692metafunction to remove its sign (if it is a signed integral type) and 693then use it to generate the type of a grayscale pixel. From the pixel 694type we create the image type. GIL’s image class is specialized over 695the pixel type and a boolean indicating whether the image should be 696planar or interleaved. Single-channel (grayscale) images in GIL must 697always be interleaved. There are multiple ways of constructing types 698in GIL. Instead of instantiating the classes directly we could have 699used type factory metafunctions. The following code is equivalent:</p> 700<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 701<span class="kt">void</span> <span class="n">x_luminosity_gradient</span><span class="p">(</span><span class="n">SrcView</span> <span class="k">const</span><span class="o">&</span> <span class="n">src</span><span class="p">,</span> <span class="n">DstView</span> <span class="k">const</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 702<span class="p">{</span> 703 <span class="k">typedef</span> <span class="k">typename</span> <span class="n">channel_type</span><span class="o"><</span><span class="n">DstView</span><span class="o">>::</span><span class="n">type</span> <span class="n">d_channel_t</span><span class="p">;</span> 704 <span class="k">typedef</span> <span class="k">typename</span> <span class="n">channel_convert_to_unsigned</span><span class="o"><</span><span class="n">d_channel_t</span><span class="o">>::</span><span class="n">type</span> <span class="n">channel_t</span><span class="p">;</span> 705 <span class="k">typedef</span> <span class="k">typename</span> <span class="n">image_type</span><span class="o"><</span><span class="n">channel_t</span><span class="p">,</span> <span class="n">gray_layout_t</span><span class="o">>::</span><span class="n">type</span> <span class="n">gray_image_t</span><span class="p">;</span> 706 <span class="k">typedef</span> <span class="k">typename</span> <span class="n">gray_image_t</span><span class="o">::</span><span class="n">value_type</span> <span class="n">gray_pixel_t</span><span class="p">;</span> 707 708 <span class="n">gray_image_t</span> <span class="nf">ccv_image</span><span class="p">(</span><span class="n">src</span><span class="p">.</span><span class="n">dimensions</span><span class="p">());</span> 709 <span class="n">copy_and_convert_pixels</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">view</span><span class="p">(</span><span class="n">ccv_image</span><span class="p">));</span> 710 <span class="n">x_gradient</span><span class="p">(</span><span class="n">const_view</span><span class="p">(</span><span class="n">ccv_image</span><span class="p">),</span> <span class="n">dst</span><span class="p">);</span> 711<span class="p">}</span> 712</pre></div> 713</div> 714<p>GIL provides a set of metafunctions that generate GIL types - 715<code class="docutils literal"><span class="pre">image_type</span></code> is one such meta-function that constructs the type of 716an image from a given channel type, color layout, and 717planar/interleaved option (the default is interleaved). There are also 718similar meta-functions to construct the types of pixel references, 719iterators, locators and image views. GIL also has metafunctions 720<code class="docutils literal"><span class="pre">derived_pixel_reference_type</span></code>, <code class="docutils literal"><span class="pre">derived_iterator_type</span></code>, 721<code class="docutils literal"><span class="pre">derived_view_type</span></code> and <code class="docutils literal"><span class="pre">derived_image_type</span></code> that construct the 722type of a GIL construct from a given source one by changing one or 723more properties of the type and keeping the rest.</p> 724<p>From the image type we can use the nested typedef <code class="docutils literal"><span class="pre">value_type</span></code> to 725obtain the type of a pixel. GIL images, image views and locators have 726nested typedefs <code class="docutils literal"><span class="pre">value_type</span></code> and <code class="docutils literal"><span class="pre">reference</span></code> to obtain the type of 727the pixel and a reference to the pixel. If you have a pixel iterator, 728you can get these types from its <code class="docutils literal"><span class="pre">iterator_traits</span></code>. Note also the 729algorithm <code class="docutils literal"><span class="pre">copy_and_convert_pixels</span></code>, which is an abbreviated version 730of <code class="docutils literal"><span class="pre">copy_pixels</span></code> with a color converted source view.</p> 731</div> 732<div class="section" id="virtual-image-views"> 733<h2><a class="toc-backref" href="#id10">Virtual Image Views</a></h2> 734<p>So far we have been dealing with images that have pixels stored in 735memory. GIL allows you to create an image view of an arbitrary image, 736including a synthetic function. To demonstrate this, let us create a 737view of the Mandelbrot set. First, we need to create a function 738object that computes the value of the Mandelbrot set at a given 739location (x,y) in the image:</p> 740<div class="highlight-cpp"><div class="highlight"><pre><span class="c1">// models PixelDereferenceAdaptorConcept</span> 741<span class="k">struct</span> <span class="n">mandelbrot_fn</span> 742<span class="p">{</span> 743 <span class="k">typedef</span> <span class="n">point</span><span class="o"><</span><span class="kt">ptrdiff_t</span><span class="o">></span> <span class="n">point_t</span><span class="p">;</span> 744 745 <span class="k">typedef</span> <span class="n">mandelbrot_fn</span> <span class="n">const_t</span><span class="p">;</span> 746 <span class="k">typedef</span> <span class="n">gray8_pixel_t</span> <span class="n">value_type</span><span class="p">;</span> 747 <span class="k">typedef</span> <span class="n">value_type</span> <span class="n">reference</span><span class="p">;</span> 748 <span class="k">typedef</span> <span class="n">value_type</span> <span class="n">const_reference</span><span class="p">;</span> 749 <span class="k">typedef</span> <span class="n">point_t</span> <span class="n">argument_type</span><span class="p">;</span> 750 <span class="k">typedef</span> <span class="n">reference</span> <span class="n">result_type</span><span class="p">;</span> 751 <span class="k">static</span> <span class="kt">bool</span> <span class="k">constexpr</span> <span class="n">is_mutable</span> <span class="o">=</span> <span class="nb">false</span><span class="p">;</span> 752 753 <span class="n">mandelbrot_fn</span><span class="p">()</span> <span class="p">{}</span> 754 <span class="n">mandelbrot_fn</span><span class="p">(</span><span class="k">const</span> <span class="n">point_t</span><span class="o">&</span> <span class="n">sz</span><span class="p">)</span> <span class="o">:</span> <span class="n">_img_size</span><span class="p">(</span><span class="n">sz</span><span class="p">)</span> <span class="p">{}</span> 755 756 <span class="n">result_type</span> <span class="k">operator</span><span class="p">()(</span><span class="k">const</span> <span class="n">point_t</span><span class="o">&</span> <span class="n">p</span><span class="p">)</span> <span class="k">const</span> 757 <span class="p">{</span> 758 <span class="c1">// normalize the coords to (-2..1, -1.5..1.5)</span> 759 <span class="kt">double</span> <span class="n">t</span><span class="o">=</span><span class="n">get_num_iter</span><span class="p">(</span><span class="n">point</span><span class="o"><</span><span class="kt">double</span><span class="o">></span><span class="p">(</span><span class="n">p</span><span class="p">.</span><span class="n">x</span><span class="o">/</span><span class="p">(</span><span class="kt">double</span><span class="p">)</span><span class="n">_img_size</span><span class="p">.</span><span class="n">x</span><span class="o">*</span><span class="mi">3</span><span class="o">-</span><span class="mi">2</span><span class="p">,</span> <span class="n">p</span><span class="p">.</span><span class="n">y</span><span class="o">/</span><span class="p">(</span><span class="kt">double</span><span class="p">)</span><span class="n">_img_size</span><span class="p">.</span><span class="n">y</span><span class="o">*</span><span class="mi">3</span><span class="o">-</span><span class="mf">1.5f</span><span class="p">));</span> 760 <span class="k">return</span> <span class="nf">value_type</span><span class="p">((</span><span class="n">bits8</span><span class="p">)(</span><span class="n">pow</span><span class="p">(</span><span class="n">t</span><span class="p">,</span><span class="mf">0.2</span><span class="p">)</span><span class="o">*</span><span class="mi">255</span><span class="p">));</span> <span class="c1">// raise to power suitable for viewing</span> 761 <span class="p">}</span> 762<span class="k">private</span><span class="o">:</span> 763 <span class="n">point_t</span> <span class="n">_img_size</span><span class="p">;</span> 764 765 <span class="kt">double</span> <span class="nf">get_num_iter</span><span class="p">(</span><span class="k">const</span> <span class="n">point</span><span class="o"><</span><span class="kt">double</span><span class="o">>&</span> <span class="n">p</span><span class="p">)</span> <span class="k">const</span> 766 <span class="p">{</span> 767 <span class="n">point</span><span class="o"><</span><span class="kt">double</span><span class="o">></span> <span class="n">Z</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span><span class="mi">0</span><span class="p">);</span> 768 <span class="k">for</span> <span class="p">(</span><span class="kt">int</span> <span class="n">i</span><span class="o">=</span><span class="mi">0</span><span class="p">;</span> <span class="n">i</span><span class="o"><</span><span class="mi">100</span><span class="p">;</span> <span class="o">++</span><span class="n">i</span><span class="p">)</span> <span class="c1">// 100 iterations</span> 769 <span class="p">{</span> 770 <span class="n">Z</span> <span class="o">=</span> <span class="n">point</span><span class="o"><</span><span class="kt">double</span><span class="o">></span><span class="p">(</span><span class="n">Z</span><span class="p">.</span><span class="n">x</span><span class="o">*</span><span class="n">Z</span><span class="p">.</span><span class="n">x</span> <span class="o">-</span> <span class="n">Z</span><span class="p">.</span><span class="n">y</span><span class="o">*</span><span class="n">Z</span><span class="p">.</span><span class="n">y</span> <span class="o">+</span> <span class="n">p</span><span class="p">.</span><span class="n">x</span><span class="p">,</span> <span class="mi">2</span><span class="o">*</span><span class="n">Z</span><span class="p">.</span><span class="n">x</span><span class="o">*</span><span class="n">Z</span><span class="p">.</span><span class="n">y</span> <span class="o">+</span> <span class="n">p</span><span class="p">.</span><span class="n">y</span><span class="p">);</span> 771 <span class="k">if</span> <span class="p">(</span><span class="n">Z</span><span class="p">.</span><span class="n">x</span><span class="o">*</span><span class="n">Z</span><span class="p">.</span><span class="n">x</span> <span class="o">+</span> <span class="n">Z</span><span class="p">.</span><span class="n">y</span><span class="o">*</span><span class="n">Z</span><span class="p">.</span><span class="n">y</span> <span class="o">></span> <span class="mi">4</span><span class="p">)</span> 772 <span class="k">return</span> <span class="n">i</span><span class="o">/</span><span class="p">(</span><span class="kt">double</span><span class="p">)</span><span class="mi">100</span><span class="p">;</span> 773 <span class="p">}</span> 774 <span class="k">return</span> <span class="mi">0</span><span class="p">;</span> 775 <span class="p">}</span> 776<span class="p">};</span> 777</pre></div> 778</div> 779<p>We can now use GIL’s <code class="docutils literal"><span class="pre">virtual_2d_locator</span></code> with this function object 780to construct a Mandelbrot view of size 200x200 pixels:</p> 781<div class="highlight-cpp"><div class="highlight"><pre><span class="k">typedef</span> <span class="n">mandelbrot_fn</span><span class="o">::</span><span class="n">point_t</span> <span class="n">point_t</span><span class="p">;</span> 782<span class="k">typedef</span> <span class="n">virtual_2d_locator</span><span class="o"><</span><span class="n">mandelbrot_fn</span><span class="p">,</span><span class="nb">false</span><span class="o">></span> <span class="n">locator_t</span><span class="p">;</span> 783<span class="k">typedef</span> <span class="n">image_view</span><span class="o"><</span><span class="n">locator_t</span><span class="o">></span> <span class="n">my_virt_view_t</span><span class="p">;</span> 784 785<span class="n">point_t</span> <span class="nf">dims</span><span class="p">(</span><span class="mi">200</span><span class="p">,</span><span class="mi">200</span><span class="p">);</span> 786 787<span class="c1">// Construct a Mandelbrot view with a locator, taking top-left corner (0,0) and step (1,1)</span> 788<span class="n">my_virt_view_t</span> <span class="nf">mandel</span><span class="p">(</span><span class="n">dims</span><span class="p">,</span> <span class="n">locator_t</span><span class="p">(</span><span class="n">point_t</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span><span class="mi">0</span><span class="p">),</span> <span class="n">point_t</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span><span class="mi">1</span><span class="p">),</span> <span class="n">mandelbrot_fn</span><span class="p">(</span><span class="n">dims</span><span class="p">)));</span> 789</pre></div> 790</div> 791<p>We can treat the synthetic view just like a real one. For example, 792let’s invoke our <code class="docutils literal"><span class="pre">x_gradient</span></code> algorithm to compute the gradient of 793the 90-degree rotated view of the Mandelbrot set and save the original 794and the result:</p> 795<div class="highlight-cpp"><div class="highlight"><pre><span class="n">gray8s_image_t</span> <span class="nf">img</span><span class="p">(</span><span class="n">dims</span><span class="p">);</span> 796<span class="n">x_gradient</span><span class="p">(</span><span class="n">rotated90cw_view</span><span class="p">(</span><span class="n">mandel</span><span class="p">),</span> <span class="n">view</span><span class="p">(</span><span class="n">img</span><span class="p">));</span> 797 798<span class="c1">// Save the Mandelbrot set and its 90-degree rotated gradient (jpeg cannot save signed char; must convert to unsigned char)</span> 799<span class="n">jpeg_write_view</span><span class="p">(</span><span class="s">"mandel.jpg"</span><span class="p">,</span><span class="n">mandel</span><span class="p">);</span> 800<span class="n">jpeg_write_view</span><span class="p">(</span><span class="s">"mandel_grad.jpg"</span><span class="p">,</span><span class="n">color_converted_view</span><span class="o"><</span><span class="n">gray8_pixel_t</span><span class="o">></span><span class="p">(</span><span class="n">const_view</span><span class="p">(</span><span class="n">img</span><span class="p">)));</span> 801</pre></div> 802</div> 803<p>Here is what the two files look like:</p> 804<img alt="../_images/mandel.jpg" src="../_images/mandel.jpg" /> 805</div> 806<div class="section" id="run-time-specified-images-and-image-views"> 807<h2><a class="toc-backref" href="#id11">Run-Time Specified Images and Image Views</a></h2> 808<p>So far we have created a generic function that computes the image 809gradient of an image view template specialization. Sometimes, 810however, the properties of an image view, such as its color space and 811channel depth, may not be available at compile time. GIL’s 812<code class="docutils literal"><span class="pre">dynamic_image</span></code> extension allows for working with GIL constructs 813that are specified at run time, also called _variants_. GIL provides 814models of a run-time instantiated image, <code class="docutils literal"><span class="pre">any_image</span></code>, and a run-time 815instantiated image view, <code class="docutils literal"><span class="pre">any_image_view</span></code>. The mechanisms are in 816place to create other variants, such as <code class="docutils literal"><span class="pre">any_pixel</span></code>, 817<code class="docutils literal"><span class="pre">any_pixel_iterator</span></code>, etc. Most of GIL’s algorithms and all of the 818view transformation functions also work with run-time instantiated 819image views and binary algorithms, such as <code class="docutils literal"><span class="pre">copy_pixels</span></code> can have 820either or both arguments be variants.</p> 821<p>Lets make our <code class="docutils literal"><span class="pre">x_luminosity_gradient</span></code> algorithm take a variant image 822view. For simplicity, let’s assume that only the source view can be a 823variant. (As an example of using multiple variants, see GIL’s image 824view algorithm overloads taking multiple variants.)</p> 825<p>First, we need to make a function object that contains the templated 826destination view and has an application operator taking a templated 827source view:</p> 828<div class="highlight-cpp"><div class="highlight"><pre><span class="cp">#include</span> <span class="cpf"><boost/gil/extension/dynamic_image/dynamic_image_all.hpp></span><span class="cp"></span> 829 830<span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 831<span class="k">struct</span> <span class="n">x_gradient_obj</span> 832<span class="p">{</span> 833 <span class="k">typedef</span> <span class="kt">void</span> <span class="n">result_type</span><span class="p">;</span> <span class="c1">// required typedef</span> 834 835 <span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">_dst</span><span class="p">;</span> 836 <span class="n">x_gradient_obj</span><span class="p">(</span><span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> <span class="o">:</span> <span class="n">_dst</span><span class="p">(</span><span class="n">dst</span><span class="p">)</span> <span class="p">{}</span> 837 838 <span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcView</span><span class="o">></span> 839 <span class="kt">void</span> <span class="k">operator</span><span class="p">()(</span><span class="k">const</span> <span class="n">SrcView</span><span class="o">&</span> <span class="n">src</span><span class="p">)</span> <span class="k">const</span> <span class="p">{</span> <span class="n">x_luminosity_gradient</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">_dst</span><span class="p">);</span> <span class="p">}</span> 840<span class="p">};</span> 841</pre></div> 842</div> 843<p>The second step is to provide an overload of <code class="docutils literal"><span class="pre">x_luminosity_gradient</span></code> that 844takes image view variant and calls GIL’s <code class="docutils literal"><span class="pre">apply_operation</span></code> passing it the 845function object:</p> 846<div class="highlight-cpp"><div class="highlight"><pre><span class="k">template</span> <span class="o"><</span><span class="k">typename</span> <span class="n">SrcViews</span><span class="p">,</span> <span class="k">typename</span> <span class="n">DstView</span><span class="o">></span> 847<span class="kt">void</span> <span class="n">x_luminosity_gradient</span><span class="p">(</span><span class="k">const</span> <span class="n">any_image_view</span><span class="o"><</span><span class="n">SrcViews</span><span class="o">>&</span> <span class="n">src</span><span class="p">,</span> <span class="k">const</span> <span class="n">DstView</span><span class="o">&</span> <span class="n">dst</span><span class="p">)</span> 848<span class="p">{</span> 849 <span class="n">apply_operation</span><span class="p">(</span><span class="n">src</span><span class="p">,</span> <span class="n">x_gradient_obj</span><span class="o"><</span><span class="n">DstView</span><span class="o">></span><span class="p">(</span><span class="n">dst</span><span class="p">));</span> 850<span class="p">}</span> 851</pre></div> 852</div> 853<p><code class="docutils literal"><span class="pre">any_image_view<SrcViews></span></code> is the image view variant. It is 854templated over <code class="docutils literal"><span class="pre">SrcViews</span></code>, an enumeration of all possible view types 855the variant can take. <code class="docutils literal"><span class="pre">src</span></code> contains inside an index of the 856currently instantiated type, as well as a block of memory containing 857the instance. <code class="docutils literal"><span class="pre">apply_operation</span></code> goes through a switch statement 858over the index, each case of which casts the memory to the correct 859view type and invokes the function object with it. Invoking an 860algorithm on a variant has the overhead of one switch 861statement. Algorithms that perform an operation for each pixel in an 862image view have practically no performance degradation when used with 863a variant.</p> 864<p>Here is how we can construct a variant and invoke the algorithm:</p> 865<div class="highlight-cpp"><div class="highlight"><pre><span class="cp">#include</span> <span class="cpf"><boost/mp11.hpp></span><span class="cp"></span> 866<span class="cp">#include</span> <span class="cpf"><boost/gil/extension/io/jpeg/old.hpp></span><span class="cp"></span> 867 868<span class="k">typedef</span> <span class="n">mp11</span><span class="o">::</span><span class="n">mp_list</span><span class="o"><</span><span class="n">gray8_image_t</span><span class="p">,</span> <span class="n">gray16_image_t</span><span class="p">,</span> <span class="n">rgb8_image_t</span><span class="p">,</span> <span class="n">rgb16_image_t</span><span class="o">></span> <span class="n">my_img_types</span><span class="p">;</span> 869<span class="n">any_image</span><span class="o"><</span><span class="n">my_img_types</span><span class="o">></span> <span class="n">runtime_image</span><span class="p">;</span> 870<span class="n">jpeg_read_image</span><span class="p">(</span><span class="s">"input.jpg"</span><span class="p">,</span> <span class="n">runtime_image</span><span class="p">);</span> 871 872<span class="n">gray8s_image_t</span> <span class="nf">gradient</span><span class="p">(</span><span class="n">runtime_image</span><span class="p">.</span><span class="n">dimensions</span><span class="p">());</span> 873<span class="n">x_luminosity_gradient</span><span class="p">(</span><span class="n">const_view</span><span class="p">(</span><span class="n">runtime_image</span><span class="p">),</span> <span class="n">view</span><span class="p">(</span><span class="n">gradient</span><span class="p">));</span> 874<span class="n">jpeg_write_view</span><span class="p">(</span><span class="s">"x_gradient.jpg"</span><span class="p">,</span> <span class="n">color_converted_view</span><span class="o"><</span><span class="n">gray8_pixel_t</span><span class="o">></span><span class="p">(</span><span class="n">const_view</span><span class="p">(</span><span class="n">gradient</span><span class="p">)));</span> 875</pre></div> 876</div> 877<p>In this example, we create an image variant that could be 8-bit or 87816-bit RGB or grayscale image. We then use GIL’s I/O extension to load 879the image from file in its native color space and channel depth. If 880none of the allowed image types matches the image on disk, an 881exception will be thrown. We then construct a 8 bit signed 882(i.e. <code class="docutils literal"><span class="pre">char</span></code>) image to store the gradient and invoke <code class="docutils literal"><span class="pre">x_gradient</span></code> 883on it. Finally we save the result into another file. We save the view 884converted to 8-bit unsigned, because JPEG I/O does not support signed 885char.</p> 886<p>Note how free functions and methods such as <code class="docutils literal"><span class="pre">jpeg_read_image</span></code>, 887<code class="docutils literal"><span class="pre">dimensions</span></code>, <code class="docutils literal"><span class="pre">view</span></code> and <code class="docutils literal"><span class="pre">const_view</span></code> work on both templated and 888variant types. For templated images <code class="docutils literal"><span class="pre">view(img)</span></code> returns a templated 889view, whereas for image variants it returns a view variant. For 890example, the return type of <code class="docutils literal"><span class="pre">view(runtime_image)</span></code> is 891<code class="docutils literal"><span class="pre">any_image_view<Views></span></code> where <code class="docutils literal"><span class="pre">Views</span></code> enumerates four views 892corresponding to the four image types. <code class="docutils literal"><span class="pre">const_view(runtime_image)</span></code> 893returns a <code class="docutils literal"><span class="pre">any_image_view</span></code> of the four read-only view types, etc.</p> 894<p>A warning about using variants: instantiating an algorithm with a 895variant effectively instantiates it with every possible type the 896variant can take. For binary algorithms, the algorithm is 897instantiated with every possible combination of the two input types! 898This can take a toll on both the compile time and the executable size.</p> 899</div> 900<div class="section" id="conclusion"> 901<h2><a class="toc-backref" href="#id12">Conclusion</a></h2> 902<p>This tutorial provides a glimpse at the challenges associated with 903writing generic and efficient image processing algorithms in GIL. We 904have taken a simple algorithm and shown how to make it work with image 905representations that vary in bit depth, color space, ordering of the 906channels, and planar/interleaved structure. We have demonstrated that 907the algorithm can work with fully abstracted virtual images, and even 908images whose type is specified at run time. The associated video 909presentation also demonstrates that even for complex scenarios the 910generated assembly is comparable to that of a C version of the 911algorithm, hand-written for the specific image types.</p> 912<p>Yet, even for such a simple algorithm, we are far from making a fully 913generic and optimized code. In particular, the presented algorithms 914work on homogeneous images, i.e. images whose pixels have channels 915that are all of the same type. There are examples of images, such as a 916packed 565 RGB format, which contain channels of different 917types. While GIL provides concepts and algorithms operating on 918heterogeneous pixels, we leave the task of extending x_gradient as an 919exercise for the reader. Second, after computing the value of the 920gradient we are simply casting it to the destination channel 921type. This may not always be the desired operation. For example, if 922the source channel is a float with range [0..1] and the destination is 923unsigned char, casting the half-difference to unsigned char will 924result in either 0 or 1. Instead, what we might want to do is scale 925the result into the range of the destination channel. GIL’s 926channel-level algorithms might be useful in such cases. For example, 927p channel_convert converts between channels by linearly scaling the 928source channel value into the range of the destination channel.</p> 929<p>There is a lot to be done in improving the performance as 930well. Channel-level operations, such as the half-difference, could be 931abstracted out into atomic channel-level algorithms and performance 932overloads could be provided for concrete channel 933types. Processor-specific operations could be used, for example, to 934perform the operation over an entire row of pixels simultaneously, or 935the data could be pre-fetched. All of these optimizations can be 936realized as performance specializations of the generic 937algorithm. Finally, compilers, while getting better over time, are 938still failing to fully optimize generic code in some cases, such as 939failing to inline some functions or put some variables into 940registers. If performance is an issue, it might be worth trying your 941code with different compilers.</p> 942</div> 943</div> 944 945 946 <div class="navbar" style="text-align:right;"> 947 948 949 <a class="prev" title="Tutorial: Histogram" href="histogram.html"><img src="../_static/prev.png" alt="prev"/></a> 950 <a class="next" title="Naming Conventions" href="../naming.html"><img src="../_static/next.png" alt="next"/></a> 951 952 </div> 953 </div> 954 <div class="footer" role="contentinfo"> 955 Last updated on 2020-08-11 15:08:48. 956 Created using <a href="http://sphinx-doc.org/">Sphinx</a> 1.5.6. 957 </div> 958 </body> 959</html>
