CPU Tuning: Basics

CPU Tuning: Basics

The first stage of the rendering pipeline is traversal of the data and sending the current rendering data to the rest of the pipeline. In theory, the entire rendering database (scene graph) must be traversed in some fashion for each frame because both scene content and viewer position can be dynamic.

To get the best possible CPU performance, follow these two overall guidelines:

Compile your application for optimum speed.
Compile all object files with at least -O2. Note that the compiler option for debugging, g, turns off all optimization. If you must run the debugger on optimized code, you can use g3 with -O2 with limited success. If you're not compiling with -xansi (the default) or -ansi, you may need to include float for faster floating-point operations.

On certain platforms, other compile-time options, such as -mips3 or -mips4, are available.
Use a simple data structure and a fast traversal method.
The CPU tuning strategy focuses on developing fast database traversal for drawing with a simple, easily accessed data structure. The fastest rendering is achieved with an inner loop that traverses a completely flattened (non-hierarchical) database. Most applications cannot achieve this level of simplicity for a variety of reasons. For example, some databases occupy too much memory when completely flattened. Note also that you run a greater risk of cache misses if you flatten the data.

When an application is CPU limited, the entire graphics pipeline may be sitting idle for periods of time. The following sections describe techniques for structuring application code so that the CPU doesn't become the bottleneck.

Immediate Mode Drawing Versus Display Lists

In deciding whether to use immediate mode drawing or display lists, consider the following advantages and disadvantages of display lists. In the end, the characteristics of your application determine which approach to take.

Display lists have the following advantages:

You don't have to optimize traversal of the data yourself; in fact, display list traversal is typically well-tuned and runs more efficiently than user programs.
Display lists manage their own data storage. This is particularly useful for algorithmically generated objects.
Display lists are significantly better for remote graphics over a network. The display list can be cached on the remote CPU so that the data for the display list does not have to be re-sent every frame. Furthermore, the remote CPU handles much of the responsibility for traversal.
Display lists are preferable for direct rendering if they contain enough primitives (a total of about ten) because display lists are stored efficiently. If the lists are short, the setup performance cost is not offset by the more efficient storage or saving in CPU time.
For information on display lists on Indigo2 IMPACT systems, see "Using Display Lists Effectively".

Display lists do have drawbacks that may affect some applications:

The most troublesome drawback of display lists is data expansion. To achieve fast, simple traversal on all systems, all data is copied directly into the display list. Therefore, the display list contains an entire copy of all your data plus additional overhead for each command.
If vertices are shared in structures more complex than the OpenGL primitives (line strip, triangle strip, triangle fan, quad strip), they are stored more than once.
If the database becomes sufficiently large, paging eventually hinders performance. Therefore, when contemplating the use of OpenGL display lists for really large databases, consider the amount of main memory.
Compiling display lists may take some time.

For moderate amounts of static data, there is seldom a good reason not to use display lists. Note the reason that the following section on immediate mode drawing is much longer than the section on display lists: Display lists are optimized by the system, and therefore require less work. That alone may be a good reason to prefer display lists under the appropriate circumstances.

Immediate Mode Drawing Versus Display Lists