The more computations and processing your shader code needs to do, the more it will impact the performance of your game. For example, supporting color per material is nice to make a shader more flexible, but if you always leave that color set to white then useless computations are performed for each vertex or pixel rendered on screen.
The frequency of computations will also impact the performance of your game. Usually there are many more pixels rendered (and subsequently more pixel shader executions) than there are vertices (vertex shader executions), and more vertices than objects being rendered. Where possible, move computations out of the pixel shader code into the the vertex shader code, or move them out of shaders completely and set the values in a script.
Surface Shaders are great for writing shaders that interact with lighting. However, their default options are tuned to cover a broad number of general cases. Tweak these for specific situations to make shaders run faster or at least be smaller:
approxviewdirective for shaders that use view direction (i.e. Specular) makes the view direction normalized per vertex instead of per pixel. This is approximate, but often good enough.
halfasviewfor Specular shader types is even faster. The half-vector (halfway between lighting direction and view vector) is computed and normalized per vertex, and the lighting function receives the half-vector as a parameter instead of the view vector.
noforwardaddmakes a shader fully support one-directional light in Forward rendering only. The rest of the lights can still have an effect as per-vertex lights or spherical harmonics. This is great to make your shader smaller and make sure it always renders in one pass, even with multiple lights present.
noambientdisables ambient lighting and spherical harmonics lights on a shader. This can make performance slightly faster.
When writing shaders in Cg/HLSL, there are three basic number types:
fixed (see Data Types and Precision).
For good performance, always use the lowest precision that is possible. This is especially important on mobile platforms like iOS and Android. Good rules of thumb are:
halfprecision. Increase only if necessary.
In practice, exactly which number type you should use for depends on the platform and the GPU. Generally speaking:
float/half/fixedend up being exactly the same underneath. This can make testing difficult, as it’s harder to see if half/fixed precision is really enough, so always test your shaders on the target device for accurate results.
halfprecision support. This is usually faster, and uses less power to do calculations.
Fixedprecision is generally only useful for older mobile GPUs. Most modern GPUs (the ones that can run OpenGL ES 3 or Metal) internally treat
halfprecision exactly the same.
See Data Types and Precision for more details.
The fixed-function AlphaTest - or its programmable equivalent,
clip() - has different performance characteristics on different platforms:
On some platforms (mostly mobile GPUs found in iOS and Android devices), using ColorMask to leave out some channels (e.g.
ColorMask RGB) can be resource-intensive, so only use it if really necessary.