docs/dispatch.rst

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