OpenGL may be implemented by any combination of hardware and software. At the high-end, hardware may implement virtually all of OpenGL while at the low-end, OpenGL may be implemented entirely in software. In between are combination software/hardware implementations. More money buys more hardware and better performance.

Intro-level workstation hardware and the recent PC 3-D hardware typically implement point, line, and polygon rasterization in hardware but implement floating point transformations, lighting, and clipping in software. This is a good strategy since the bottleneck in 3-D rendering is usually rasterization and modern CPU's have sufficient floating point performance to handle the transformation stage.

OpenGL developers must remember that their application may be used on a wide variety of OpenGL implementations. Therefore one should consider using all possible optimizations, even those which have little return on the development system, since other systems may benefit greatly.

From this point of view it may seem wise to develop your application on a low-end system. There is a pitfall however; some operations which are cheep in software may be expensive in hardware. The moral is: test your application on a variety of systems to be sure the performance is dependable.

Multiprocessing

One should consider multiprocessing in these situations. By assigning rendering and computation to different threads they may be executed in parallel on multiprocessor computers.

For many applications, supporting multiprocessing is just a matter of partitioning the render and compute operations into separate threads which share common data structures and coordinate with synchronization primitives.

SGI's Performer is an example of a high level toolkit designed for this purpose.

Image quality vs. performance

Complexity may refer to the geometric or rendering attributes of a database. Here are a few examples.

• During interactive rotation (i.e. mouse button held down) render a reduced-polygon model. When drawing a static image draw the full polygon model.
• During animation, disable dithering, smooth shading, and/or texturing. Enable them for the static image.
• If texturing is required, use GL_NEAREST sampling and glHint( GL_PERSPECTIVE_CORRECTION_HINT, GL_FASTEST ).
• During animation, disable antialiasing. Enable antialiasing for the static image.
• Use coarser NURBS/evaluator tesselation during animation. Use glPolygonMode( GL_FRONT_AND_BACK, GL_LINE ) to inspect tesselation granularity and reduce if possible.

Level of detail management and culling

Objects which are entirely outside of the field of view may be culled. This type of high level cull testing can be done efficiently with bounding boxes or spheres and have a major impact on performance. Again, toolkits such as Inventor and Performer have this feature.

Basically, one wants data structures which can be traversed quickly and passed to the graphics library in an efficient manner. For example, suppose we need to render a triangle strip. The data structure which stores the list of vertices may be implemented with a linked list or an array. Clearly the array can be traversed more quickly than a linked list. The way in which a vertex is stored in the data structure is also significant. High performance hardware can process vertexes specified by a pointer more quickly than those specified by three separate parameters.

An Example

	struct city {
		float latitute, longitude;	/* city location */
		char *name;			/* city's name */
		int large_flag;  		/* 0 = small, 1 = large */
	};

Our first attempt at rendering this information may be:

	void draw_cities( int n, struct city citylist[] )
	{
	   int i;
	   for (i=0; i < n; i++) {
	      if (citylist[i].large_flag) {
		 glPointSize( 4.0 );
	      }
	      else {
	         glPointSize( 2.0 );
 	      }
	      glBegin( GL_POINTS );
	      glVertex2f( citylist[i].longitude, citylist[i].latitude );
	      glEnd();
              glRasterPos2f( citylist[i].longitude, citylist[i].latitude );
	      glCallLists( strlen(citylist[i].name),
			   GL_BYTE, 
			   citylist[i].name );
	   }
	}

• glPointSize is called for every loop iteration.
• only one point is drawn between glBegin and glEnd
• the vertices aren't being specified in the most efficient manner

	void draw_cities( int n, struct city citylist[] )
	{
	   int i;
	   /* draw small dots first */
	   glPointSize( 2.0 );
	   glBegin( GL_POINTS );
	   for (i=0; i < n ;i++) {
	      if (citylist[i].large_flag==0) {
	         glVertex2f( citylist[i].longitude, citylist[i].latitude );
	      }
	   }
	   glEnd();
	   /* draw large dots second */
	   glPointSize( 4.0 );
	   glBegin( GL_POINTS );
	   for (i=0; i < n ;i++) {
	      if (citylist[i].large_flag==1) {
	         glVertex2f( citylist[i].longitude, citylist[i].latitude );
	      }
	   }
	   glEnd();
	   /* draw city labels third */
	   for (i=0; i < n ;i++) {
              glRasterPos2f( citylist[i].longitude, citylist[i].latitude );
	      glCallLists( strlen(citylist[i].name),
			   GL_BYTE, 
			   citylist[i].name );
	   }
	}

We can still do better, however. If we redesign the data structures used to represent the city information we can improve the efficiency of drawing the city points. For example:

	struct city_list {
		int num_cities;		/* how many cities in the list */
		float *position;	/* pointer to lat/lon coordinates */
		char **name;		/* pointer to city names */
		float size;		/* size of city points */
	};

	/* indicates if server can do GL_EXT_vertex_array: */
	GLboolean varray_available;

	void draw_cities( struct city_list *list )
	{
	    int i;
	    GLboolean use_begin_end;

	    /* draw the points */
	    glPointSize( list->size );

	#ifdef GL_EXT_vertex_array
	    if (varray_available) {
		glVertexPointerEXT( 2, GL_FLOAT, 0, list->num_cities, list->position );
		glDrawArraysEXT( GL_POINTS, 0, list->num_cities );
		use_begin_end = GL_FALSE;
	    }
	    else
	#else
	    {
	    	use_begin_end = GL_TRUE;
	    }
	#endif

	    if (use_begin_end) {
		glBegin(GL_POINTS);
		for (i=0; i < list->num_cities; i++) {
		    glVertex2fv( &position[i*2] );
		}
		glEnd();
	    }

	    /* draw city labels */
	    for (i=0; i < list->num_cities ;i++) {
		glRasterPos2fv( list->position[i*2] );
		glCallLists( strlen(list->name[i]),
				GL_BYTE, list->name[i] );
	    }
	}

In the following sections the techniques for maximizing performance, as seen above, are explained.

After each of the following techniques look for a bracketed list of symbols which relates the significance of the optimization to your OpenGL system:

• H - beneficial for high-end hardware
• L - beneficial for low-end hardware
• S - beneficial for software implementations
• all - probably beneficial for all implementations

Use connected primitives

Connected primitives such as GL_LINES, GL_LINE_LOOP, GL_TRIANGLE_STRIP, GL_TRIANGLE_FAN, and GL_QUAD_STRIP require fewer vertices to describe an object than individual line, triangle, or polygon primitives. This reduces data transfer and transformation workload. [all]

Use the vertex array extension

On some architectures function calls are somewhat expensive so replacing many glVertex/glColor/glNormal calls with the vertex array mechanism may be very beneficial. [all]

Store vertex data in consecutive memory locations

When maximum performance is needed on high-end systems it's good to store vertex data in contiguous memory to maximize through put of data from host memory to graphics subsystem. [H,L]

Use the vector versions of glVertex, glColor, glNormal and glTexCoord

The glVertex, glColor, etc. functions which take a pointer to their arguments such as glVertex3fv(v) may be much faster than those which take individual arguments such as glVertex3f(x,y,z) on systems with DMA-driven graphics hardware. [H,L]

Reduce quantity of primitives

Be careful not to render primitives which are over-tesselated. Experiment with the GLU primitives, for example, to determine the best compromise of image quality vs. tesselation level. Textured objects in particular may still be rendered effectively with low geometric complexity. [all]

Display lists

Use display lists to encapsulate frequently drawn objects. Display list data may be stored in the graphics subsystem rather than host memory thereby eliminating host-to-graphics data movement. Display lists are also very beneficial when rendering remotely. [all]

Don't specify unneeded per-vertex information

If lighting is disabled don't call glNormal. If texturing is disabled don't call glTexCoord, etc.

Minimize code between glBegin/glEnd

For maximum performance on high-end systems it's extremely important to send vertex data to the graphics system as fast as possible. Avoid extraneous code between glBegin/glEnd.

Example:

	glBegin( GL_TRIANGLE_STRIP );
	for (i=0; i < n; i++) {
		if (lighting) {
			glNormal3fv( norm[i] );
		}
		glVertex3fv( vert[i] );
	}
	glEnd();
	

This is a very bad construct. The following is much better:

	if (lighting) {
		glBegin( GL_TRIANGLE_STRIP );
		for (i=0; i < n ;i++) {
			glNormal3fv( norm[i] );
			glVertex3fv( vert[i] );
		}
		glEnd();
	}
	else {
		glBegin( GL_TRIANGLE_STRIP );
		for (i=0; i < n ;i++) {
			glVertex3fv( vert[i] );
		}
		glEnd();
	}
	

Also consider manually unrolling important rendering loops to maximize the function call rate.

3.2 Transformation

Lighting

• Avoid using positional lights, i.e. light positions should be of the form (x,y,z,0) [L,S]
• Avoid using spotlights. [all]
• Avoid using two-sided lighting. [all]
• Avoid using negative material and light color coefficients [S]
• Avoid using the local viewer lighting model. [L,S]
• Avoid frequent changes to the GL_SHININESS material parameter. [L,S]
• Some OpenGL implementations are optimized for the case of a single light source.
• Consider pre-lighting complex objects before rendering, ala radiosity. You can get the effect of lighting by specifying vertex colors instead of vertex normals. [S]

Two sided lighting

If you want both the front and back of polygons shaded the same try using two light sources instead of two-sided lighting. Position the two light sources on opposite sides of your object. That way, a polygon will always be lit correctly whether it's back or front facing. [L,S]

Disable normal vector normalization when not needed

glEnable/Disable(GL_NORMALIZE) controls whether normal vectors are scaled to unit length before lighting. If you do not use glScale you may be able to disable normalization without ill effects. Normalization is disabled by default. [L,S]

Use connected primitives

Connected primitives such as GL_LINES, GL_LINE_LOOP, GL_TRIANGLE_STRIP, GL_TRIANGLE_FAN, and GL_QUAD_STRIP decrease traversal and transformation load.

glRect usage

If you have to draw many rectangles consider using glBegin(GL_QUADS) ... glEnd() instead. [all]

3.3 Rasterization

Disable smooth shading when not needed

Smooth shading is enabled by default. Flat shading doesn't require interpolation of the four color components and is usually faster than smooth shading in software implementations. Hardware may perform flat and smooth-shaded rendering at the same rate though there's at least one case in which smooth shading is faster than flat shading (E&S Freedom). [S]

Disable depth testing when not needed

Background objects, for example, can be drawn without depth testing if they're drawn first. Foreground objects can be drawn without depth testing if they're drawn last. [L,S]

Disable dithering when not needed

This is easy to forget when developing on a high-end machine. Disabling dithering can make a big difference in software implementations of OpenGL on lower-end machines with 8 or 12-bit color buffers. Dithering is enabled by default. [S]

Use back-face culling whenever possible.

If you're drawing closed polyhedra or other objects for which back facing polygons aren't visible there's probably no point in drawing those polygons. [all]

The GL_SGI_cull_vertex extension

SGI's Cosmo GL supports a new culling extension which looks at vertex normals to try to improve the speed of culling.

Avoid extra fragment operations

Stenciling, blending, stippling, alpha testing and logic ops can all take extra time during rasterization. Be sure to disable the operations which aren't needed. [all]

Reduce the window size or screen resolution

A simple way to reduce rasterization time is to reduce the number of pixels drawn. If a smaller window or reduced display resolution are acceptable it's an easy way to improve rasterization speed. [L,S]

3.4 Texturing

Use efficient image formats

The GL_UNSIGNED_BYTE component format is typically the fastest for specifying texture images. Experiment with the internal texture formats offered by the GL_EXT_texture extension. Some formats are faster than others on some systems (16-bit texels on the Reality Engine, for example). [all]

Encapsulate texture maps in texture objects or display lists

This is especially important if you use several texture maps. By putting textures into display lists or texture objects the graphics system can manage their storage and minimize data movement between the client and graphics subsystem. [all]

Use smaller texture maps

Smaller images can be moved from host to texture memory faster than large images. More small texture can be stored simultaneously in texture memory, reducing texture memory swapping. [all]

Use simpler sampling functions

Experiment with the minification and magnification texture filters to determine which performs best while giving acceptable results. Generally, GL_NEAREST is fastest and GL_LINEAR is second fastest. [all]

Use the same sampling function for minification and magnification

If both the minification and magnification filters are GL_NEAREST or GL_LINEAR then there's no reason OpenGL has to compute the lambda value which determines whether to use minification or magnification sampling for each fragment. Avoiding the lambda calculation can be a good performace improvement.

Use a simpler texture environment function

Some texture environment modes may be faster than others. For example, the GL_DECAL or GL_REPLACE_EXT functions for 3 component textures is a simple assignment of texel samples to fragments while GL_MODULATE is a linear interpolation between texel samples and incoming fragments. [S,L]

Combine small textures

If you are using several small textures consider tiling them together as a larger texture and modify your texture coordinates to address the subtexture you want. This technique can eliminate texture bindings.

Use glHint(GL_PERSPECTIVE_CORRECTION_HINT, GL_FASTEST)

This hint can improve the speed of texturing when perspective- correct texture coordinate interpolation isn't needed, such as when using a glOrtho() projection.

Animated textures

If you want to use an animated texture, perhaps live video textures, don't use glTexImage2D to repeatedly change the texture. Use glTexSubImage2D or glTexCopyTexSubImage2D. These functions are standard in OpenGL 1.1 and available as extensions to 1.0.

3.5 Clearing

Use glClear carefully [all]

Clear all relevant color buffers with one glClear.

Wrong:

  glClear( GL_COLOR_BUFFER_BIT );
  if (stenciling) {
	glClear( GL_STENCIL_BUFFER_BIT );
  }

Right:

  if (stenciling) {
	glClear( GL_COLOR_BUFFER_BIT | GL_STENCIL_BUFFER_BIT );
  }
  else {
	glClear( GL_COLOR_BUFFER_BIT );
  }

Disable dithering

Disable dithering before clearing the color buffer. Visually, the difference between dithered and undithered clears is usually negligable.

Use scissoring to clear a smaller area

If you don't need to clear the whole buffer use glScissor() to restrict clearing to a smaller area. [L].

Don't clear the color buffer at all

If the scene you're drawing opaquely covers the entire window there is no reason to clear the color buffer.

Eliminate depth buffer clearing

If the scene you're drawing covers the entire window there is a trick which let's you omit the depth buffer clear. The idea is to only use half the depth buffer range for each frame and alternate between using GL_LESS and GL_GREATER as the depth test function.

Example:

   int EvenFlag;

   /* Call this once during initialization and whenever the window
    * is resized.
    */
   void init_depth_buffer( void )
   {
	glClearDepth( 1.0 );
	glClear( GL_DEPTH_BUFFER_BIT );
	glDepthRange( 0.0, 0.5 );
	glDepthFunc( GL_LESS );
	EvenFlag = 1;
   }

   /* Your drawing function */
   void display_func( void )
   {
	if (EvenFlag) {
	    glDepthFunc( GL_LESS );
	    glDepthRange( 0.0, 0.5 );
	}
	else {
	    glDepthFunc( GL_GREATER );
	    glDepthRange( 1.0, 0.5 );
	}
	EvenFlag = !EvenFlag;

	/* draw your scene */
   }

Avoid glClearDepth( d ) where d!=1.0

Some software implementations may have optimized paths for clearing the depth buffer to 1.0. [S]

3.6 Miscellaneous

Avoid "round-trip" calls

Calls such as glGetFloatv, glGetIntegerv, glIsEnabled, glGetError, glGetString require a slow, round trip transaction between the application and renderer. Especially avoid them in your main rendering code.

Note that software implementations of OpenGL may actually perform these operations faster than hardware systems. If you're developing on a low-end system be aware of this fact. [H,L]

Avoid glPushAttrib

If only a few pieces of state need to be saved and restored it's often faster to maintain the information in the client program. glPushAttrib( GL_ALL_ATTRIB_BITS ) in particular can be very expensive on hardware systems. This call may be faster in software implementations than in hardware. [H,L]

Check for GL errors during development

During development call glGetError inside your rendering/event loop to catch errors. GL errors raised during rendering can slow down rendering speed. Remove the glGetError call for production code since it's a "round trip" command and can cause delays. [all]

Use glColorMaterial instead of glMaterial

If you need to change a material property on a per vertex basis, glColorMaterial may be faster than glMaterial. [all]

glDrawPixels

• glDrawPixels often performs best with GL_UNSIGNED_BYTE color components [all]
• Disable all unnecessary raster operations before calling glDrawPixels. [all]
• Use the GL_EXT_abgr extension to specify color components in alpha, blue, green, red order on systems which were designed for IRIS GL. [H,L].

Avoid using viewports which are larger than the window

Software implementations may have to do additional clipping in this situation. [S]

Alpha planes

Don't allocate alpha planes in the color buffer if you don't need them. Specifically, they are not needed for transparency effects. Systems without hardware alpha planes may have to resort to a slow software implementation. [L,S]

Accumulation, stencil, overlay planes

Do not allocate accumulation, stencil or overlay planes if they are not needed. [all]

Be aware of the depth buffer's depth

Your OpenGL may support several different sizes of depth buffers- 16 and 24-bit for example. Shallower depth buffers may be faster than deep buffers both for software and hardware implementations. However, the precision of of a 16-bit depth buffer may not be sufficient for some applications. [L,S]

Transparency may be implemented with stippling instead of blending

If you need simple transparent objects consider using polygon stippling instead of alpha blending. The later is typically faster and may actually look better in some situations. [L,S]

Group state changes together

Try to mimimize the number of GL state changes in your code. When GL state is changed, internal state may have to be recomputed, introducing delays. [all]

Avoid using glPolygonMode

If you need to draw many polygon outlines or vertex points use glBegin with GL_POINTS, GL_LINES, GL_LINE_LOOP or GL_LINE_STRIP instead as it can be much faster. [all]

3.7 Window System Integration

Minimize calls to the make current call

The glXMakeCurrent call, for example, can be expensive on hardware systems because the context switch may involve moving a large amount of data in and out of the hardware.

Visual / pixel format performance

Some X visuals or pixel formats may be faster than others. On PCs for example, 24-bit color buffers may be slower to read/write than 12 or 8-bit buffers. There is often a tradeoff between performance and quality of frame buffer configurations. 12-bit color may not look as nice as 24-bit color. A 16-bit depth buffer won't have the precision of a 24-bit depth buffer.

The GLX_EXT_visual_rating extension can help you select visuals based on performance or quality. GLX 1.2's visual caveat attribute can tell you if a visual has a performance penalty associated with it.

It may be worthwhile to experiment with different visuals to determine if there's any advantage of one over another.

Avoid mixing OpenGL rendering with native rendering

OpenGL allows both itself and the native window system to render into the same window. For this to be done correctly synchronization is needed. The GLX glXWaitX and glXWaitGL functions serve this purpose.

Synchronization hurts performance. Therefore, if you need to render with both OpenGL and native window system calls try to group the rendering calls to minimize synchronization.

For example, if you're drawing a 3-D scene with OpenGL and displaying text with X, draw all the 3-D elements first, call glXWaitGL to synchronize, then call all the X drawing functions.

Don't redraw more than necessary

Be sure that you're not redrawing your scene unnecissarily. For example, expose/repaint events may come in batches describing separate regions of the window which must be redrawn. Since one usually redraws the whole window image with OpenGL you only need to respond to one expose/repaint event. In the case of X, look at the count field of the XExposeEvent structure. Only redraw when it is zero.

Also, when responding to mouse motion events you should skip extra motion events in the input queue. Otherwise, if you try to process every motion event and redraw your scene there will be a noticable delay between mouse input and screen updates.

It can be a good idea to put a print statement in your redraw and event loop function so you know exactly what messages are causing your scene to be redrawn, and when.

SwapBuffer calls and graphics pipe blocking

On systems with 3-D graphics hardware the SwapBuffers call is synchronized to the monitor's vertical retrace. Input to the OpenGL command queue may be blocked until the buffer swap has completed. Therefore, don't put more OpenGL calls immediately after SwapBuffers. Instead, put application computation instructions which can overlap with the buffer swap delay.

3.8 Mesa-specific

Double buffering

The X driver supports two back color buffer implementations: Pixmaps and XImages. The MESA_BACK_BUFFER environment variable controls which is used. Which of the two that's faster depends on the nature of your rendering. Experiment.

X Visuals

As described above, some X visuals can be rendered into more quickly than others. The MESA_RGB_VISUAL environment variable can be used to determine the quickest visual by experimentation.

Depth buffers

Mesa may use a 16 or 32-bit depth buffer as specified in the src/config.h configuration file. 16-bit depth buffers are faster but may not offer the precision needed for all applications.

Flat-shaded primitives

If one is drawing a number of flat-shaded primitives all of the same color the glColor command should be put before the glBegin call.

Don't do this:

	glBegin(...);
	glColor(...);
	glVertex(...);
	...
	glEnd();
	

Do this:

	glColor(...);
	glBegin(...);
	glVertex(...);
	...
	glEnd();
	

glColor*() commands

The glColor[34]ub[v] are the fastest versions of the glColor command.

Avoid double precision valued functions

Mesa does all internal floating point computations in single precision floating point. API functions which take double precision floating point values must convert them to single precision. This can be expensive in the case of glVertex, glNormal, etc.

4. Evaluation and Tuning

Performance evaluation is a large subject and only the basics are covered here. For more information see "OpenGL on Silicon Graphics Systems".

• CPU subsystem - application code which drives the graphics subsystem
• Geometry subsystem - transformation of vertices, lighting, and clipping
• Rasterization subsystem - drawing filled polygons, line segments and per-pixel processing

4.1.1 CPU subsystem

4.1.2 Geometry subsystem

4.1.3 Rasterization subsystem

After bottlenecks have been identified the techniques outlined in section 3 can be applied. The process of identifying and reducing bottlenecks should be repeated until no further improvements can be made or your minimum performance threshold has been met.