(WIP) KTX2Loader: Return DataTexture when transcoding to uncompressed format

[donmccurdy]

Work in progress. Tested and working for:

• DataTexture
• DataArrayTexture
• DataCubeTexture ❌
• CompressedTexture
• CompressedArrayTexture
• CompressedCubeTexture

Related:

• KTX2Loader: Return DataTexture for uncompressed formats #29922
• KTX2Loader: Fix .minFilter default for untranscoded compressed textures #29904 (comment)
• (WIP) KTX2Loader: Add support for u8, f16, and f32 array and cube textures #26642

[github-actions]

📦 Bundle size

Full ESM build, minified and gzipped.

	Before	After	Diff
WebGL	339.13 79	339.14 79	+16 B +2 B
WebGPU	478.72 132.72	478.73 132.72	+16 B +1 B
WebGPU Nodes	478.18 132.6	478.2 132.61	+16 B +4 B

🌳 Bundle size after tree-shaking

Minimal build including a renderer, camera, empty scene, and dependencies.

	Before	After	Diff
WebGL	464.59 111.96	464.59 111.96	+0 B +0 B
WebGPU	546.88 148.18	546.88 148.18	+0 B +0 B
WebGPU Nodes	502.76 137.89	502.76 137.89	+0 B +0 B

[donmccurdy]

@Mugen87 I'm running into the old problems from #26642 (comment) here, I'm just not sure how to properly set up mipmaps for DataTexture, DataArrayTexture, and DataCubeTexture classes. I've left Data3DTexture out of scope for this PR. I see what looks like a TODO about it here:

three.js/src/renderers/webgl/WebGLTextures.js

Lines 1301 to 1307 in de9b967

	if ( useTexStorage && allocateMemory ) {
	// TODO: Uniformly handle mipmap definitions
	// Normal textures and compressed cube textures define base level + mips with their mipmap array
	// Uncompressed cube textures use their mipmap array only for mips (no base level)
	if ( mipmaps.length > 0 ) levels ++;

I also feel worried about this pattern...

texture = new DataCubeTexture( mipmaps[ 0 ].data, format, type );
texture.mipmaps = mipmaps.map( ( mip ) => new DataCubeTexture( mip.data, mip.width, mip.height ) );

... since the other existing pattern is to initialize a CubeTexture with an array of 6 DataTextures, but here we have one DataCubeTexture per mip, not per face, and it's pretty inconsistent. Is this something we could work out for the Data*Texture classes, independently of KTX2Loader?

[Mugen87]

Is this something we could work out for the Data*Texture classes, independently of KTX2Loader?

Yes, it think this TODO should eventually be resolved. Ideally, the image and mipmaps property behaves the same for all texture types.

The policy should be: If no mipmaps are present, the engine evaluates the image property. If mipmaps are defined, the first entry in the mipmaps array is the base level, the subsequent entries represent the mips. Cube textures should be changed so they honor this policy.