Add Log10 display transform mode and WebGL-safe NaN handling

[eeholmes]

Background

For many of my datasets the data are highly skewed and strictly positive. It is helpful to allow data transforms for display. This was fairly easy to implement but surfaced an issue with fragile NaN generation in the current code.

Two related issues came up in visualization:

1. Some environmental variables are much easier to interpret on a logarithmic scale than a linear scale.
2. Shader code that generates NaN via expressions like sqrt(-1.0) can be driver-dependent and unreliable (black globe while the histogram showed values)

Shader safety

Replaced sqrt(-1.0)-style NaN generation with a helper based on uintBitsToFloat(0x7FC00000u).

Example code sketch

function applyDisplayTransformToData(data: Float32Array, mode: "linear" | "log10") {
  if (mode === "linear") return data;
  for (let i = 0; i < data.length; i++) {
    const v = data[i];
    data[i] = Number.isFinite(v) && v > 0 ? Math.log10(v) : Number.NaN;
  }
  return data;
}

float makeNaN() {
  return uintBitsToFloat(0x7FC00000u);
}

How this was solved in my fork

In my fork, I:

• added linear and log10 transform modes to shared types/store state,
• added a transform control in the colormap UI,
• applied the transform across all grid renderers before computing final bounds/histograms,
• normalized HEALPix missing/fill handling so transforms do not mis-handle sentinel values,
• replaced shader NaN generation with a quiet-NaN helper.

Related PR

• Fork PR: https://github.com/eeholmes/gridlook/pull/23

Untransformed

Transformed

[Karinon]

Hi Eli, thank you for creating this issue.

Replaced sqrt(-1.0)-style NaN generation with a helper based on uintBitsToFloat(0x7FC00000u).

can you provide me with one example dataset where this matters and tell me something about the device/browser (os, graphics card, browser)? I would like to reproduce it myself.

added linear and log10 transform modes to shared types/store state

While I see the issue here, I am not really happy with this solution as you have transformed all the values from linear to log10, but the unit in your example still remains meter. In my eyes the UI is no longer self-explanatory after that. As of now I don't have a better solution up my sleeve, but I will think about it.

[eeholmes]

Here is an example. CHLA variable is skewed
https://storage.googleapis.com/nmfs_odp_nwfsc/CB/fish-pace-datasets/chla-z/zarr

[eeholmes-uw]

@Karinon I changed the UI to clearly show that the data are transformed and removed units. I use the log10 transform feature all the time. Perhaps the data I look at is more commonly highly skewed and positive.