docs/dispatch.html

   1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN" "http://www.w3.org/TR/html4/loose.dtd">
   2 <html lang="en">
   3 <head>
   4   <meta http-equiv="content-type" content="text/html; charset=utf-8">
   5   <title>GL Dispatch</title>
   6   <link rel="stylesheet" type="text/css" href="mesa.css">
   7 </head>
   8 <body>
   9
  10 <div class="header">
  11   The Mesa 3D Graphics Library
  12 </div>
  13
  14 <iframe src="contents.html"></iframe>
  15 <div class="content">
  16
  17 <h1>GL Dispatch</h1>
  18
  19 <p>Several factors combine to make efficient dispatch of OpenGL functions
  20 fairly complicated.  This document attempts to explain some of the issues
  21 and introduce the reader to Mesa's implementation.  Readers already familiar
  22 with the issues around GL dispatch can safely skip ahead to the <a
  23 href="#overview">overview of Mesa's implementation</a>.</p>
  24
  25 <h2>1. Complexity of GL Dispatch</h2>
  26
  27 <p>Every GL application has at least one object called a GL <em>context</em>.
  28 This object, which is an implicit parameter to every GL function, stores all
  29 of the GL related state for the application.  Every texture, every buffer
  30 object, every enable, and much, much more is stored in the context.  Since
  31 an application can have more than one context, the context to be used is
  32 selected by a window-system dependent function such as
  33 <code>glXMakeContextCurrent</code>.</p>
  34
  35 <p>In environments that implement OpenGL with X-Windows using GLX, every GL
  36 function, including the pointers returned by <code>glXGetProcAddress</code>, are
  37 <em>context independent</em>.  This means that no matter what context is
  38 currently active, the same <code>glVertex3fv</code> function is used.</p>
  39
  40 <p>This creates the first bit of dispatch complexity.  An application can
  41 have two GL contexts.  One context is a direct rendering context where
  42 function calls are routed directly to a driver loaded within the
  43 application's address space.  The other context is an indirect rendering
  44 context where function calls are converted to GLX protocol and sent to a
  45 server.  The same <code>glVertex3fv</code> has to do the right thing depending
  46 on which context is current.</p>
  47
  48 <p>Highly optimized drivers or GLX protocol implementations may want to
  49 change the behavior of GL functions depending on current state.  For
  50 example, <code>glFogCoordf</code> may operate differently depending on whether
  51 or not fog is enabled.</p>
  52
  53 <p>In multi-threaded environments, it is possible for each thread to have a
  54 different GL context current.  This means that poor old <code>glVertex3fv</code>
  55 has to know which GL context is current in the thread where it is being
  56 called.</p>
  57
  58 <h2 id="overview">2. Overview of Mesa's Implementation</h2>
  59
  60 <p>Mesa uses two per-thread pointers.  The first pointer stores the address
  61 of the context current in the thread, and the second pointer stores the
  62 address of the <em>dispatch table</em> associated with that context.  The
  63 dispatch table stores pointers to functions that actually implement
  64 specific GL functions.  Each time a new context is made current in a thread,
  65 these pointers a updated.</p>
  66
  67 <p>The implementation of functions such as <code>glVertex3fv</code> becomes
  68 conceptually simple:</p>
  69
  70 <ul>
  71 <li>Fetch the current dispatch table pointer.</li>
  72 <li>Fetch the pointer to the real <code>glVertex3fv</code> function from the
  73 table.</li>
  74 <li>Call the real function.</li>
  75 </ul>
  76
  77 <p>This can be implemented in just a few lines of C code.  The file
  78 <code>src/mesa/glapi/glapitemp.h</code> contains code very similar to this.</p>
  79
  80 <blockquote>
  81 <table border="1">
  82 <tr><td><pre>
  83 void glVertex3f(GLfloat x, GLfloat y, GLfloat z)
  84 {
  85     const struct _glapi_table * const dispatch = GET_DISPATCH();
  86
  87     (*dispatch-&gt;Vertex3f)(x, y, z);
  88 }</pre></td></tr>
  89 <tr><td>Sample dispatch function</td></tr></table>
  90 </blockquote>
  91
  92 <p>The problem with this simple implementation is the large amount of
  93 overhead that it adds to every GL function call.</p>
  94
  95 <p>In a multithreaded environment, a naive implementation of
  96 <code>GET_DISPATCH</code> involves a call to <code>pthread_getspecific</code> or a
  97 similar function.  Mesa provides a wrapper function called
  98 <code>_glapi_get_dispatch</code> that is used by default.</p>
  99
 100 <h2>3. Optimizations</h2>
 101
 102 <p>A number of optimizations have been made over the years to diminish the
 103 performance hit imposed by GL dispatch.  This section describes these
 104 optimizations.  The benefits of each optimization and the situations where
 105 each can or cannot be used are listed.</p>
 106
 107 <h3>3.1. Dual dispatch table pointers</h3>
 108
 109 <p>The vast majority of OpenGL applications use the API in a single threaded
 110 manner.  That is, the application has only one thread that makes calls into
 111 the GL.  In these cases, not only do the calls to
 112 <code>pthread_getspecific</code> hurt performance, but they are completely
 113 unnecessary!  It is possible to detect this common case and avoid these
 114 calls.</p>
 115
 116 <p>Each time a new dispatch table is set, Mesa examines and records the ID
 117 of the executing thread.  If the same thread ID is always seen, Mesa knows
 118 that the application is, from OpenGL's point of view, single threaded.</p>
 119
 120 <p>As long as an application is single threaded, Mesa stores a pointer to
 121 the dispatch table in a global variable called <code>_glapi_Dispatch</code>.
 122 The pointer is also stored in a per-thread location via
 123 <code>pthread_setspecific</code>.  When Mesa detects that an application has
 124 become multithreaded, <code>NULL</code> is stored in <code>_glapi_Dispatch</code>.</p>
 125
 126 <p>Using this simple mechanism the dispatch functions can detect the
 127 multithreaded case by comparing <code>_glapi_Dispatch</code> to <code>NULL</code>.
 128 The resulting implementation of <code>GET_DISPATCH</code> is slightly more
 129 complex, but it avoids the expensive <code>pthread_getspecific</code> call in
 130 the common case.</p>
 131
 132 <blockquote>
 133 <table border="1">
 134 <tr><td><pre>
 135 #define GET_DISPATCH() \
 136     (_glapi_Dispatch != NULL) \
 137         ? _glapi_Dispatch : pthread_getspecific(&amp;_glapi_Dispatch_key)
 138 </pre></td></tr>
 139 <tr><td>Improved <code>GET_DISPATCH</code> Implementation</td></tr></table>
 140 </blockquote>
 141
 142 <h3>3.2. ELF TLS</h3>
 143
 144 <p>Starting with the 2.4.20 Linux kernel, each thread is allocated an area
 145 of per-thread, global storage.  Variables can be put in this area using some
 146 extensions to GCC.  By storing the dispatch table pointer in this area, the
 147 expensive call to <code>pthread_getspecific</code> and the test of
 148 <code>_glapi_Dispatch</code> can be avoided.</p>
 149
 150 <p>The dispatch table pointer is stored in a new variable called
 151 <code>_glapi_tls_Dispatch</code>.  A new variable name is used so that a single
 152 libGL can implement both interfaces.  This allows the libGL to operate with
 153 direct rendering drivers that use either interface.  Once the pointer is
 154 properly declared, <code>GET_DISPACH</code> becomes a simple variable
 155 reference.</p>
 156
 157 <blockquote>
 158 <table border="1">
 159 <tr><td><pre>
 160 extern __thread struct _glapi_table *_glapi_tls_Dispatch
 161     __attribute__((tls_model("initial-exec")));
 162
 163 #define GET_DISPATCH() _glapi_tls_Dispatch
 164 </pre></td></tr>
 165 <tr><td>TLS <code>GET_DISPATCH</code> Implementation</td></tr></table>
 166 </blockquote>
 167
 168 <p>Use of this path is controlled by the preprocessor define
 169 <code>USE_ELF_TLS</code>.  Any platform capable of using ELF TLS should use this
 170 as the default dispatch method.</p>
 171
 172 <h3>3.3. Assembly Language Dispatch Stubs</h3>
 173
 174 <p>Many platforms has difficulty properly optimizing the tail-call in the
 175 dispatch stubs.  Platforms like x86 that pass parameters on the stack seem
 176 to have even more difficulty optimizing these routines.  All of the dispatch
 177 routines are very short, and it is trivial to create optimal assembly
 178 language versions.  The amount of optimization provided by using assembly
 179 stubs varies from platform to platform and application to application.
 180 However, by using the assembly stubs, many platforms can use an additional
 181 space optimization (see <a href="#fixedsize">below</a>).</p>
 182
 183 <p>The biggest hurdle to creating assembly stubs is handling the various
 184 ways that the dispatch table pointer can be accessed.  There are four
 185 different methods that can be used:</p>
 186
 187 <ol>
 188 <li>Using <code>_glapi_Dispatch</code> directly in builds for non-multithreaded
 189 environments.</li>
 190 <li>Using <code>_glapi_Dispatch</code> and <code>_glapi_get_dispatch</code> in
 191 multithreaded environments.</li>
 192 <li>Using <code>_glapi_Dispatch</code> and <code>pthread_getspecific</code> in
 193 multithreaded environments.</li>
 194 <li>Using <code>_glapi_tls_Dispatch</code> directly in TLS enabled
 195 multithreaded environments.</li>
 196 </ol>
 197
 198 <p>People wishing to implement assembly stubs for new platforms should focus
 199 on #4 if the new platform supports TLS.  Otherwise, implement #2 followed by
 200 #3.  Environments that do not support multithreading are uncommon and not
 201 terribly relevant.</p>
 202
 203 <p>Selection of the dispatch table pointer access method is controlled by a
 204 few preprocessor defines.</p>
 205
 206 <ul>
 207 <li>If <code>USE_ELF_TLS</code> is defined, method #3 is used.</li>
 208 <li>If <code>HAVE_PTHREAD</code> is defined, method #2 is used.</li>
 209 <li>If none of the preceding are defined, method #1 is used.</li>
 210 </ul>
 211
 212 <p>Two different techniques are used to handle the various different cases.
 213 On x86 and SPARC, a macro called <code>GL_STUB</code> is used.  In the preamble
 214 of the assembly source file different implementations of the macro are
 215 selected based on the defined preprocessor variables.  The assembly code
 216 then consists of a series of invocations of the macros such as:
 217
 218 <blockquote>
 219 <table border="1">
 220 <tr><td><pre>
 221 GL_STUB(Color3fv, _gloffset_Color3fv)
 222 </pre></td></tr>
 223 <tr><td>SPARC Assembly Implementation of <code>glColor3fv</code></td></tr></table>
 224 </blockquote>
 225
 226 <p>The benefit of this technique is that changes to the calling pattern
 227 (i.e., addition of a new dispatch table pointer access method) require fewer
 228 changed lines in the assembly code.</p>
 229
 230 <p>However, this technique can only be used on platforms where the function
 231 implementation does not change based on the parameters passed to the
 232 function.  For example, since x86 passes all parameters on the stack, no
 233 additional code is needed to save and restore function parameters around a
 234 call to <code>pthread_getspecific</code>.  Since x86-64 passes parameters in
 235 registers, varying amounts of code needs to be inserted around the call to
 236 <code>pthread_getspecific</code> to save and restore the GL function's
 237 parameters.</p>
 238
 239 <p>The other technique, used by platforms like x86-64 that cannot use the
 240 first technique, is to insert <code>#ifdef</code> within the assembly
 241 implementation of each function.  This makes the assembly file considerably
 242 larger (e.g., 29,332 lines for <code>glapi_x86-64.S</code> versus 1,155 lines for
 243 <code>glapi_x86.S</code>) and causes simple changes to the function
 244 implementation to generate many lines of diffs.  Since the assembly files
 245 are typically generated by scripts (see <a href="#autogen">below</a>), this
 246 isn't a significant problem.</p>
 247
 248 <p>Once a new assembly file is created, it must be inserted in the build
 249 system.  There are two steps to this.  The file must first be added to
 250 <code>src/mesa/sources</code>.  That gets the file built and linked.  The second
 251 step is to add the correct <code>#ifdef</code> magic to
 252 <code>src/mesa/glapi/glapi_dispatch.c</code> to prevent the C version of the
 253 dispatch functions from being built.</p>
 254
 255 <h3 id="fixedsize">3.4. Fixed-Length Dispatch Stubs</h3>
 256
 257 <p>To implement <code>glXGetProcAddress</code>, Mesa stores a table that
 258 associates function names with pointers to those functions.  This table is
 259 stored in <code>src/mesa/glapi/glprocs.h</code>.  For different reasons on
 260 different platforms, storing all of those pointers is inefficient.  On most
 261 platforms, including all known platforms that support TLS, we can avoid this
 262 added overhead.</p>
 263
 264 <p>If the assembly stubs are all the same size, the pointer need not be
 265 stored for every function.  The location of the function can instead be
 266 calculated by multiplying the size of the dispatch stub by the offset of the
 267 function in the table.  This value is then added to the address of the first
 268 dispatch stub.</p>
 269
 270 <p>This path is activated by adding the correct <code>#ifdef</code> magic to
 271 <code>src/mesa/glapi/glapi.c</code> just before <code>glprocs.h</code> is
 272 included.</p>
 273
 274 <h2 id="autogen">4. Automatic Generation of Dispatch Stubs</h2>
 275
 276 </div>
 277 </body>
 278 </html>