シェーダーにおける条件

Sometimes, you want the same shader to do different things under different circumstances. For example, you might want to configure different settings for different materials, define functionality for different hardware, or dynamically change the behavior of shaders at runtime. You might also want to avoid executing computationally expensive code when it’s not needed, such as texture reads, vertex inputs, interpolators, or loops.

条件を使用することで、GPU が特定の条件下でのみ実行する動作を定義することができ ます。

条件のタイプ

シェーダーでの条件の使用には、以下アプローチがあります。

• 静的分岐: シェーダーコンパイラーがコンパイル時に条件付きコードを評価します。
• 動的分岐: GPU がランタイムに条件付きコードを評価します。
• シェーダーバリアント: Unity は、静的分岐を使用して、シェーダーソースコードを複数のシェーダープログラムにコンパイルします。その上で Unity は、ランタイムで条件に一致するシェーダープログラムを使用します。

どちらのタイプの条件をどのような場合に使用するか

シェーダーにおける条件には、“全てのケースに適したアプローチ” はありません。個々のプロジェクトで、使用しているシェーダーを踏まえて、それぞれのアプローチのメリットとデメリットを検討する必要があります。

Bear in mind the following information:

• If you know the conditions at compile time (for example, if you know that you are building for given hardware), then static branching is likely the best choice. Static branching code is simple to write and maintain, and it does not negatively impact build times, file sizes, or runtime performance. However, you cannot use it to execute code for different conditions at runtime.
• If you need to execute code for different conditions at runtime, you should consider the following options:
  • Shader variants do not incur any GPU performance penalty. However, a large number of shader variants in a project can lead to significant problems: increased build times, file sizes, runtime memory usage, and loading times. Shader variants also introduce additional code complexity when manually preloading (“prewarming”) shaders.
  • Dynamic branching incurs a GPU performance penalty, which might be small or large depending on the shader code and the hardware. However, it allows you to use conditionals at runtime without increasing the number of shader variants in your project.
  • It’s possible to author your shader code so that it avoids branching or variants, and instead uses mathematical operations that return 0 or 1 to execute conditional code. This can lead to complex code that is difficult to maintain. Depending on the circumstances, it might result in only a very small performance improvement over dynamic branching, or no advantage at all.

In general, the best approach is to profile the performance of your application and carefully consider your decisions case-by-case. For example, if you can afford the slightly increased GPU cost, it might be best to use dynamic branching to achieve “good enough” GPU performance and reduce the risk of introducing more variants. However, if GPU performance is the primary concern for this shader and you have already accounted for the cost of additional variants, you might choose to use variants.