Metal texture read() vs sample() performance

Question

In trying to optimize my Metal rendering performance in the fragment shader, I've noticed that there are several ways to read from a texture, depending on the scenario. I was wondering if the texture read approach has performance implications.

Here are three different alternatives I've tried:

1. Using tex2D.read() without a sampler (uses uint2 coordinates).
2. Using tex2D.sample() with pixel coordinates (uses float2 coordinates)
3. Using tex2D.sample() with normalized coordinates (uses float2 coordinates).

Should I expect performance differences in iOS between these options?

Answer 1

They are generally similar. One might require more work than another to convert between uint2 / float2 for the coordinate, but this effect should be minor.

Read has undefined behavior for reading outside the image, so that is something to consider.