You are here:

Fragment Shader Complexity

The vast majority of applications are fragment bound on current Tegra devices. Optimizing fragment shaders is best done in a directed way, based on total contribution to the scene. PerfHUD ES can help you to identify the key shaders (those contributing significantly to render-time), to focus this process. All fragment shaders with a significant contribution to the scene should be reviewed carefully and optimized where possible.

Go to Frame Profiler.
Check all 3 bucket definition check boxes.
Go to Frame Debugger -> Shader Viewer.
Click Export to Disk File.
Select the By performance contribution and all program radio buttons.
Select a destination folder and click Save.

You should now have a folder containing a set of fragment shader source files, along with shader_readme.txt which shows how much of the render time is due to each shader.

Check each significant fragment shader (usually only the top 5 or so buckets contribute >90% of perf), and apply standard Tegra optimizations:

You can perform much of the analysis (and even experimentation using PerfHUD's dynamic shader re-compilation system) from within PHES. However, it is difficult to identify true unique shaders. Multiple instances of a single vertex shader could be used with a single fragment shader, which would result in that fragment shader spanning many buckets.

Use low precision wherever possible. In particular color and normalized vector data can be made low with zero visual impact.
Remove as much work as possible from fragment shaders. Burn as much computation as possible into uniform/vertex shader/texture lookups.
Performance can usually be improved by replacing computations that depend only on a few inputs, including build-in functions calls such as pow(x, y), with a pre-computed texture look up.
Only normalize vectors when absolutely necessary.
Prefer clamp(val, 0.0, 1.0) to max(val, 0.0) and/or min(val, 1.0).
Supply optimally-formatted data to shaders via textures whenever possible.
LOD-biasing, in texture lookups, may incur a performance penalty on Tegra hardware.
Conditional logic should be used only when absolutely necessary.