WebGL best practices

Anything that requires syncing the CPU and GPU sides is potentially very slow, so if possible you should try to avoid doing that in your main rendering loops. This includes the following WebGL calls: getError(), readPixels(), and finish(). WebGL getter calls such as getParameter() and getUniformLocation() should be considered slow too, so try to cache their results in a JavaScript variable.

Fewer, larger draw operations will improve performance. If you have 1000 sprites to paint, try to do it as a single drawArrays() or drawElements() call. You can draw degenerate (flat) triangles if you need to draw discontinuous objects as a single drawArrays() call.

Fewer state changes will also improve performance. In particular, if you can pack multiple images into a single texture and select them by using the appropriate texture coordinates, that can help you do fewer texture binding changes, which improves performance.

• In some rare cases, packing greyscale textures which belong together into the color channels of a single texture might help.

Smaller textures perform better than larger ones. For this reason, mipmapping can be a performance win.

Simpler shaders perform better than complex ones. In particular, if you can remove an if statement from a shader, that will make it run faster. Division and math functions like log() should be considered expensive too.

• However, nowadays even mobile devices possess powerful GPUs that are capable of running even relatively complex shader programs. Moreover, because shaders are compiled, the eventual machine code that actually runs on the hardware may be highly optimized. What may seem like an expensive function call may in fact compile into only few (or even a single) machine instructions. This is particularly true for GLSL functions that typically operate on vectors, such as normalize(), dot() and mix(). The best advice in that regard is to use the built-in functions, rather than try to implement, for example, one's own version of a dot-product or linear interpolation, which may in fact compile to larger and less optimized machine code. Finally, it is important to keep in mind that GPUs are constructed to do complex mathematical calculations in hardware, and therefore, may support math functions, such as sin(), cos() and other, through dedicated machine instructions.

Do as much as you can in the vertex shader, rather than in the fragment shader. Because, per rendering pass, fragment shaders run many more times than vertex shaders, any calculation that can be done on the vertices and then just interpolated among fragments is a performance boon (this interpolation is done "automagically" for you, through the fixed functionality rasterization phase of the OpenGL pipeline). For example, a simple animation of a textured surface can be achieved through a time-dependent transformation of texture coordinates (simplest case is to add a uniform vector to the texture coordinates attribute vector). If visually acceptable, one can transform the texture coordinates in the vertex shader rather than in the fragment shader, to get better performance.

Always have vertex attrib 0 array enabled. If you draw with vertex attrib 0 array disabled, you will force the browser to do complicated emulation when running on desktop OpenGL (e.g. on Mac OSX). This is because in desktop OpenGL, nothing gets drawn if vertex attrib 0 is not array-enabled. You can use bindAttribLocation() to force a vertex attribute to use location 0, and use enableVertexAttribArray() to make it array-enabled.