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 in Mesa</title>
6  <link rel="stylesheet" type="text/css" href="mesa.css">
7</head>
8<body>
9<h1>GL Dispatch in Mesa</h1>
10
11<p>Several factors combine to make efficient dispatch of OpenGL functions
12fairly complicated.  This document attempts to explain some of the issues
13and introduce the reader to Mesa's implementation.  Readers already familiar
14with the issues around GL dispatch can safely skip ahead to the <a
15href="#overview">overview of Mesa's implementation</a>.</p>
16
17<h2>1. Complexity of GL Dispatch</h2>
18
19<p>Every GL application has at least one object called a GL <em>context</em>.
20This object, which is an implicit parameter to ever GL function, stores all
21of the GL related state for the application.  Every texture, every buffer
22object, every enable, and much, much more is stored in the context.  Since
23an application can have more than one context, the context to be used is
24selected by a window-system dependent function such as
25<tt>glXMakeContextCurrent</tt>.</p>
26
27<p>In environments that implement OpenGL with X-Windows using GLX, every GL
28function, including the pointers returned by <tt>glXGetProcAddress</tt>, are
29<em>context independent</em>.  This means that no matter what context is
30currently active, the same <tt>glVertex3fv</tt> function is used.</p>
31
32<p>This creates the first bit of dispatch complexity.  An application can
33have two GL contexts.  One context is a direct rendering context where
34function calls are routed directly to a driver loaded within the
35application's address space.  The other context is an indirect rendering
36context where function calls are converted to GLX protocol and sent to a
37server.  The same <tt>glVertex3fv</tt> has to do the right thing depending
38on which context is current.</p>
39
40<p>Highly optimized drivers or GLX protocol implementations may want to
41change the behavior of GL functions depending on current state.  For
42example, <tt>glFogCoordf</tt> may operate differently depending on whether
43or not fog is enabled.</p>
44
45<p>In multi-threaded environments, it is possible for each thread to have a
46differnt GL context current.  This means that poor old <tt>glVertex3fv</tt>
47has to know which GL context is current in the thread where it is being
48called.</p>
49
50<h2 id="overview">2. Overview of Mesa's Implementation</h2>
51
52<p>Mesa uses two per-thread pointers.  The first pointer stores the address
53of the context current in the thread, and the second pointer stores the
54address of the <em>dispatch table</em> associated with that context.  The
55dispatch table stores pointers to functions that actually implement
56specific GL functions.  Each time a new context is made current in a thread,
57these pointers a updated.</p>
58
59<p>The implementation of functions such as <tt>glVertex3fv</tt> becomes
60conceptually simple:</p>
61
62<ul>
63<li>Fetch the current dispatch table pointer.</li>
64<li>Fetch the pointer to the real <tt>glVertex3fv</tt> function from the
65table.</li>
66<li>Call the real function.</li>
67</ul>
68
69<p>This can be implemented in just a few lines of C code.  The file
70<tt>src/mesa/glapi/glapitemp.h</tt> contains code very similar to this.</p>
71
72<blockquote>
73<table border="1">
74<tr><td><pre>
75void glVertex3f(GLfloat x, GLfloat y, GLfloat z)
76{
77    const struct _glapi_table * const dispatch = GET_DISPATCH();
78
79    (*dispatch-&gt;Vertex3f)(x, y, z);
80}</pre></td></tr>
81<tr><td>Sample dispatch function</td></tr></table>
82</blockquote>
83
84<p>The problem with this simple implementation is the large amount of
85overhead that it adds to every GL function call.</p>
86
87<p>In a multithreaded environment, a naive implementation of
88<tt>GET_DISPATCH</tt> involves a call to <tt>pthread_getspecific</tt> or a
89similar function.  Mesa provides a wrapper function called
90<tt>_glapi_get_dispatch</tt> that is used by default.</p>
91
92<h2>3. Optimizations</h2>
93
94<p>A number of optimizations have been made over the years to diminish the
95performance hit imposed by GL dispatch.  This section describes these
96optimizations.  The benefits of each optimization and the situations where
97each can or cannot be used are listed.</p>
98
99<h3>3.1. Dual dispatch table pointers</h3>
100
101<p>The vast majority of OpenGL applications use the API in a single threaded
102manner.  That is, the application has only one thread that makes calls into
103the GL.  In these cases, not only do the calls to
104<tt>pthread_getspecific</tt> hurt performance, but they are completely
105unnecessary!  It is possible to detect this common case and avoid these
106calls.</p>
107
108<p>Each time a new dispatch table is set, Mesa examines and records the ID
109of the executing thread.  If the same thread ID is always seen, Mesa knows
110that the application is, from OpenGL's point of view, single threaded.</p>
111
112<p>As long as an application is single threaded, Mesa stores a pointer to
113the dispatch table in a global variable called <tt>_glapi_Dispatch</tt>.
114The pointer is also stored in a per-thread location via
115<tt>pthread_setspecific</tt>.  When Mesa detects that an application has
116become multithreaded, <tt>NULL</tt> is stored in <tt>_glapi_Dispatch</tt>.</p>
117
118<p>Using this simple mechanism the dispatch functions can detect the
119multithreaded case by comparing <tt>_glapi_Dispatch</tt> to <tt>NULL</tt>.
120The resulting implementation of <tt>GET_DISPATCH</tt> is slightly more
121complex, but it avoids the expensive <tt>pthread_getspecific</tt> call in
122the common case.</p>
123
124<blockquote>
125<table border="1">
126<tr><td><pre>
127#define GET_DISPATCH() \
128    (_glapi_Dispatch != NULL) \
129        ? _glapi_Dispatch : pthread_getspecific(&_glapi_Dispatch_key)
130</pre></td></tr>
131<tr><td>Improved <tt>GET_DISPATCH</tt> Implementation</td></tr></table>
132</blockquote>
133
134<h3>3.2. ELF TLS</h3>
135
136<p>Starting with the 2.4.20 Linux kernel, each thread is allocated an area
137of per-thread, global storage.  Variables can be put in this area using some
138extensions to GCC.  By storing the dispatch table pointer in this area, the
139expensive call to <tt>pthread_getspecific</tt> and the test of
140<tt>_glapi_Dispatch</tt> can be avoided.</p>
141
142<p>The dispatch table pointer is stored in a new variable called
143<tt>_glapi_tls_Dispatch</tt>.  A new variable name is used so that a single
144libGL can implement both interfaces.  This allows the libGL to operate with
145direct rendering drivers that use either interface.  Once the pointer is
146properly declared, <tt>GET_DISPACH</tt> becomes a simple variable
147reference.</p>
148
149<blockquote>
150<table border="1">
151<tr><td><pre>
152extern __thread struct _glapi_table *_glapi_tls_Dispatch
153    __attribute__((tls_model("initial-exec")));
154
155#define GET_DISPATCH() _glapi_tls_Dispatch
156</pre></td></tr>
157<tr><td>TLS <tt>GET_DISPATCH</tt> Implementation</td></tr></table>
158</blockquote>
159
160<p>Use of this path is controlled by the preprocessor define
161<tt>GLX_USE_TLS</tt>.  Any platform capable of using TLS should use this as
162the default dispatch method.</p>
163
164<h3>3.3. Assembly Language Dispatch Stubs</h3>
165
166<p>Many platforms has difficulty properly optimizing the tail-call in the
167dispatch stubs.  Platforms like x86 that pass parameters on the stack seem
168to have even more difficulty optimizing these routines.  All of the dispatch
169routines are very short, and it is trivial to create optimal assembly
170language versions.  The amount of optimization provided by using assembly
171stubs varies from platform to platform and application to application.
172However, by using the assembly stubs, many platforms can use an additional
173space optimization (see <a href="#fixedsize">below</a>).</p>
174
175<p>The biggest hurdle to creating assembly stubs is handling the various
176ways that the dispatch table pointer can be accessed.  There are four
177different methods that can be used:</p>
178
179<ol>
180<li>Using <tt>_glapi_Dispatch</tt> directly in builds for non-multithreaded
181environments.</li>
182<li>Using <tt>_glapi_Dispatch</tt> and <tt>_glapi_get_dispatch</tt> in
183multithreaded environments.</li>
184<li>Using <tt>_glapi_Dispatch</tt> and <tt>pthread_getspecific</tt> in
185multithreaded environments.</li>
186<li>Using <tt>_glapi_tls_Dispatch</tt> directly in TLS enabled
187multithreaded environments.</li>
188</ol>
189
190<p>People wishing to implement assembly stubs for new platforms should focus
191on #4 if the new platform supports TLS.  Otherwise, implement #2 followed by
192#3.  Environments that do not support multithreading are uncommon and not
193terribly relevant.</p>
194
195<p>Selection of the dispatch table pointer access method is controlled by a
196few preprocessor defines.</p>
197
198<ul>
199<li>If <tt>GLX_USE_TLS</tt> is defined, method #4 is used.</li>
200<li>If <tt>HAVE_PTHREAD</tt> is defined, method #3 is used.</li>
201<li>If <tt>WIN32_THREADS</tt> is defined, method #2 is used.</li>
202<li>If none of the preceeding are defined, method #1 is used.</li>
203</ul>
204
205<p>Two different techniques are used to handle the various different cases.
206On x86 and SPARC, a macro called <tt>GL_STUB</tt> is used.  In the preamble
207of the assembly source file different implementations of the macro are
208selected based on the defined preprocessor variables.  The assmebly code
209then consists of a series of invocations of the macros such as:
210
211<blockquote>
212<table border="1">
213<tr><td><pre>
214GL_STUB(Color3fv, _gloffset_Color3fv)
215</pre></td></tr>
216<tr><td>SPARC Assembly Implementation of <tt>glColor3fv</tt></td></tr></table>
217</blockquote>
218
219<p>The benefit of this technique is that changes to the calling pattern
220(i.e., addition of a new dispatch table pointer access method) require fewer
221changed lines in the assembly code.</p>
222
223<p>However, this technique can only be used on platforms where the function
224implementation does not change based on the parameters passed to the
225function.  For example, since x86 passes all parameters on the stack, no
226additional code is needed to save and restore function parameters around a
227call to <tt>pthread_getspecific</tt>.  Since x86-64 passes parameters in
228registers, varying amounts of code needs to be inserted around the call to
229<tt>pthread_getspecific</tt> to save and restore the GL function's
230parameters.</p>
231
232<p>The other technique, used by platforms like x86-64 that cannot use the
233first technique, is to insert <tt>#ifdef</tt> within the assembly
234implementation of each function.  This makes the assembly file considerably
235larger (e.g., 29,332 lines for <tt>glapi_x86-64.S</tt> versus 1,155 lines for
236<tt>glapi_x86.S</tt>) and causes simple changes to the function
237implementation to generate many lines of diffs.  Since the assmebly files
238are typically generated by scripts (see <a href="#autogen">below</a>), this
239isn't a significant problem.</p>
240
241<p>Once a new assembly file is created, it must be inserted in the build
242system.  There are two steps to this.  The file must first be added to
243<tt>src/mesa/sources</tt>.  That gets the file built and linked.  The second
244step is to add the correct <tt>#ifdef</tt> magic to
245<tt>src/mesa/glapi/glapi_dispatch.c</tt> to prevent the C version of the
246dispatch functions from being built.</p>
247
248<h3 id="fixedsize">3.4. Fixed-Length Dispatch Stubs</h3>
249
250<p>To implement <tt>glXGetProcAddress</tt>, Mesa stores a table that
251associates function names with pointers to those functions.  This table is
252stored in <tt>src/mesa/glapi/glprocs.h</tt>.  For different reasons on
253different platforms, storing all of those pointers is inefficient.  On most
254platforms, including all known platforms that support TLS, we can avoid this
255added overhead.</p>
256
257<p>If the assembly stubs are all the same size, the pointer need not be
258stored for every function.  The location of the function can instead be
259calculated by multiplying the size of the dispatch stub by the offset of the
260function in the table.  This value is then added to the address of the first
261dispatch stub.</p>
262
263<p>This path is activated by adding the correct <tt>#ifdef</tt> magic to
264<tt>src/mesa/glapi/glapi.c</tt> just before <tt>glprocs.h</tt> is
265included.</p>
266
267<h2 id="autogen">4. Automatic Generation of Dispatch Stubs</h2>
268
269</body>
270</html>
271