fix: force half float in texture if it's exact and OES_texture_float_linear is not available

[slak44]

Context

The origin of the problem is this bug in OHIF, a consumer of vtk: OHIF/Viewers#4286
For certain mobile Android devices, instead of rendering a texture, the canvas was solid gray/black. It happens regardless of OS version/browser, but eventually I narrowed it down to devices that use Mali GPUs (which is likely because they don't support OES_texture_float_linear).

Since it was working everywhere previously and I had a good version of OHIF, I bisected until this first bad commit, which among other things, bumps the version of vtk.js from 29.3.0 to 29.7.0. After checking all the vtk versions in that range, I found that 29.5.0 was the last good one, and 29.5.1 was the first bad one with gray/black canvases. This issue has been present on master ever since.

Luckily (or unluckily) this was a patch release that only included 2 commits. Of them, the relevant one is this.

The commit changed the texture format to sometimes use *32F versions, and added this warning:

The problem with using full floats is in webgl2 they aren't texture filterable without OES_texture_float_linear, while half floats are always filterable even without the extension. Using full floats triggers the warning/black textures.

I assume all of these are the same problem and would be fixed by this change:

• [Bug] Handling of clamped arrays? #3030
• [Bug] OES_texture_float_linear cornerstonejs/cornerstone3D#1152
• [Bug] OHIF not working on some Android devices OHIF/Viewers#4286

Results

With this change, canvases are no longer black in OHIF and render the images again everywhere.

Changes

This change forces half floats in cases where:

1. Full floats would have been used instead
2. The extension for texture filtering is not supported in the gl context
3. The data can be stored accurately in half floats

I don't know if there's any way to solve this if the data doesn't fit in halfs, maybe with another option similar to preferSizeOverAccuracy.

PR and Code Checklist

• semantic-release commit messages
• Run npm run reformat to have correctly formatted code

Testing

I only reproduced this on Androids with Mali GPUs. I would assume it happens on any device that supports webgl2 but not OES_texture_float_linear.
If you have such a device, you can see the problem here (a gray square is rendered, instead of an image of a brain CT): https://viewer.ohif.org/viewer?StudyInstanceUIDs=2.16.840.1.114362.1.11972228.22789312658.616067305.306.2

For testing the changes, the easiest way I know is to clone, compile and run OHIF, and configure resolutions in its package.json to use these changes in vtk:

  "resolutions": {
    "@kitware/vtk.js": "link:../vtk-js/dist/esm",

Then open this: http://localhost:3000/viewer?StudyInstanceUIDs=2.16.840.1.114362.1.11972228.22789312658.616067305.306.2

With the changes enabled you should see the brain CT, without the changes you should not.

[sankhesh]

LGTM

[github-actions]

🎉 This PR is included in version 32.1.1 🎉

The release is available on:

• npm package (@latest dist-tag)
• npm package (@latest dist-tag)
• GitHub release

Your semantic-release bot 📦🚀