GL Dispatch

Several factors combine to make efficient dispatch of OpenGL functions fairly complicated. This document attempts to explain some of the issues and introduce the reader to Mesa's implementation. Readers already familiar with the issues around GL dispatch can safely skip ahead to the overview of Mesa's implementation.

1. Complexity of GL Dispatch

Every GL application has at least one object called a GL context. This object, which is an implicit parameter to every GL function, stores all of the GL related state for the application. Every texture, every buffer object, every enable, and much, much more is stored in the context. Since an application can have more than one context, the context to be used is selected by a window-system dependent function such as glXMakeContextCurrent.

In environments that implement OpenGL with X-Windows using GLX, every GL function, including the pointers returned by glXGetProcAddress, are context independent. This means that no matter what context is currently active, the same glVertex3fv function is used.

This creates the first bit of dispatch complexity. An application can have two GL contexts. One context is a direct rendering context where function calls are routed directly to a driver loaded within the application's address space. The other context is an indirect rendering context where function calls are converted to GLX protocol and sent to a server. The same glVertex3fv has to do the right thing depending on which context is current.

Highly optimized drivers or GLX protocol implementations may want to change the behavior of GL functions depending on current state. For example, glFogCoordf may operate differently depending on whether or not fog is enabled.

In multi-threaded environments, it is possible for each thread to have a different GL context current. This means that poor old glVertex3fv has to know which GL context is current in the thread where it is being called.

2. Overview of Mesa's Implementation

Mesa uses two per-thread pointers. The first pointer stores the address of the context current in the thread, and the second pointer stores the address of the dispatch table associated with that context. The dispatch table stores pointers to functions that actually implement specific GL functions. Each time a new context is made current in a thread, these pointers a updated.

The implementation of functions such as glVertex3fv becomes conceptually simple:

Fetch the current dispatch table pointer.
Fetch the pointer to the real glVertex3fv function from the table.
Call the real function.

This can be implemented in just a few lines of C code. The file src/mesa/glapi/glapitemp.h contains code very similar to this.

void glVertex3f(GLfloat x, GLfloat y, GLfloat z)
{
    const struct _glapi_table * const dispatch = GET_DISPATCH();

    (*dispatch->Vertex3f)(x, y, z);
}

Sample dispatch function

The problem with this simple implementation is the large amount of overhead that it adds to every GL function call.

In a multithreaded environment, a naive implementation of GET_DISPATCH involves a call to pthread_getspecific or a similar function. Mesa provides a wrapper function called _glapi_get_dispatch that is used by default.

3. Optimizations

A number of optimizations have been made over the years to diminish the performance hit imposed by GL dispatch. This section describes these optimizations. The benefits of each optimization and the situations where each can or cannot be used are listed.

3.1. Dual dispatch table pointers

The vast majority of OpenGL applications use the API in a single threaded manner. That is, the application has only one thread that makes calls into the GL. In these cases, not only do the calls to pthread_getspecific hurt performance, but they are completely unnecessary! It is possible to detect this common case and avoid these calls.

Each time a new dispatch table is set, Mesa examines and records the ID of the executing thread. If the same thread ID is always seen, Mesa knows that the application is, from OpenGL's point of view, single threaded.

As long as an application is single threaded, Mesa stores a pointer to the dispatch table in a global variable called _glapi_Dispatch. The pointer is also stored in a per-thread location via pthread_setspecific. When Mesa detects that an application has become multithreaded, NULL is stored in _glapi_Dispatch.

Using this simple mechanism the dispatch functions can detect the multithreaded case by comparing _glapi_Dispatch to NULL. The resulting implementation of GET_DISPATCH is slightly more complex, but it avoids the expensive pthread_getspecific call in the common case.

#define GET_DISPATCH() \
    (_glapi_Dispatch != NULL) \
        ? _glapi_Dispatch : pthread_getspecific(&_glapi_Dispatch_key)

Improved GET_DISPATCH Implementation

3.2. ELF TLS

Starting with the 2.4.20 Linux kernel, each thread is allocated an area of per-thread, global storage. Variables can be put in this area using some extensions to GCC. By storing the dispatch table pointer in this area, the expensive call to pthread_getspecific and the test of _glapi_Dispatch can be avoided.

The dispatch table pointer is stored in a new variable called _glapi_tls_Dispatch. A new variable name is used so that a single libGL can implement both interfaces. This allows the libGL to operate with direct rendering drivers that use either interface. Once the pointer is properly declared, GET_DISPACH becomes a simple variable reference.

extern __thread struct _glapi_table *_glapi_tls_Dispatch
    __attribute__((tls_model("initial-exec")));

#define GET_DISPATCH() _glapi_tls_Dispatch

TLS GET_DISPATCH Implementation

Use of this path is controlled by the preprocessor define USE_ELF_TLS. Any platform capable of using ELF TLS should use this as the default dispatch method.

3.3. Assembly Language Dispatch Stubs

Many platforms has difficulty properly optimizing the tail-call in the dispatch stubs. Platforms like x86 that pass parameters on the stack seem to have even more difficulty optimizing these routines. All of the dispatch routines are very short, and it is trivial to create optimal assembly language versions. The amount of optimization provided by using assembly stubs varies from platform to platform and application to application. However, by using the assembly stubs, many platforms can use an additional space optimization (see below).

The biggest hurdle to creating assembly stubs is handling the various ways that the dispatch table pointer can be accessed. There are four different methods that can be used:

Using _glapi_Dispatch directly in builds for non-multithreaded environments.
Using _glapi_Dispatch and _glapi_get_dispatch in multithreaded environments.
Using _glapi_Dispatch and pthread_getspecific in multithreaded environments.
Using _glapi_tls_Dispatch directly in TLS enabled multithreaded environments.

People wishing to implement assembly stubs for new platforms should focus on #4 if the new platform supports TLS. Otherwise, implement #2 followed by #3. Environments that do not support multithreading are uncommon and not terribly relevant.

Selection of the dispatch table pointer access method is controlled by a few preprocessor defines.

If USE_ELF_TLS is defined, method #3 is used.
If HAVE_PTHREAD is defined, method #2 is used.
If none of the preceding are defined, method #1 is used.

Two different techniques are used to handle the various different cases. On x86 and SPARC, a macro called GL_STUB is used. In the preamble of the assembly source file different implementations of the macro are selected based on the defined preprocessor variables. The assembly code then consists of a series of invocations of the macros such as:

GL_STUB(Color3fv, _gloffset_Color3fv)
SPARC Assembly Implementation of glColor3fv

The benefit of this technique is that changes to the calling pattern (i.e., addition of a new dispatch table pointer access method) require fewer changed lines in the assembly code.

However, this technique can only be used on platforms where the function implementation does not change based on the parameters passed to the function. For example, since x86 passes all parameters on the stack, no additional code is needed to save and restore function parameters around a call to pthread_getspecific. Since x86-64 passes parameters in registers, varying amounts of code needs to be inserted around the call to pthread_getspecific to save and restore the GL function's parameters.

The other technique, used by platforms like x86-64 that cannot use the first technique, is to insert #ifdef within the assembly implementation of each function. This makes the assembly file considerably larger (e.g., 29,332 lines for glapi_x86-64.S versus 1,155 lines for glapi_x86.S) and causes simple changes to the function implementation to generate many lines of diffs. Since the assembly files are typically generated by scripts (see below), this isn't a significant problem.

Once a new assembly file is created, it must be inserted in the build system. There are two steps to this. The file must first be added to src/mesa/sources. That gets the file built and linked. The second step is to add the correct #ifdef magic to src/mesa/glapi/glapi_dispatch.c to prevent the C version of the dispatch functions from being built.

3.4. Fixed-Length Dispatch Stubs

To implement glXGetProcAddress, Mesa stores a table that associates function names with pointers to those functions. This table is stored in src/mesa/glapi/glprocs.h. For different reasons on different platforms, storing all of those pointers is inefficient. On most platforms, including all known platforms that support TLS, we can avoid this added overhead.

If the assembly stubs are all the same size, the pointer need not be stored for every function. The location of the function can instead be calculated by multiplying the size of the dispatch stub by the offset of the function in the table. This value is then added to the address of the first dispatch stub.

This path is activated by adding the correct #ifdef magic to src/mesa/glapi/glapi.c just before glprocs.h is included.