- // This is needed because it counteracts the precision issues we get
- // because we sample the input texture with normalized coordinates
- // (especially when the repeat count along the axis is not a power of
- // two); we very rapidly end up in narrowly missing a texel center,
- // which causes precision loss to propagate throughout the FFT.
+ // Because of the memory layout (see below) and because we use offsets,
+ // the support texture values for many consecutive values will be
+ // the same. Thus, we can store a smaller texture (giving a small
+ // performance boost) and just sample it with NEAREST. Also, this
+ // counteracts any precision issues we might get from linear
+ // interpolation.