Optimize floatVectorToFloatArray for performance improvements

[weegeekps]

This changes the floatVectorToFloatArray function used heavily during loads to use a faster method for copying the memory from the emscripten context to JS context. This was impacting us during loads of large scenes in CesiumJS, and has resulted in about a full second faster load time with the Microsoft campus dataset.

Our PR in our fork of this repo has some additional instructions and info on how we tested this out: CesiumGS#3

Key changes

• The function now uses slice to copy data from the WASM heap, which is more efficient and avoids mapping over the buffer by default.
• Moved from using a map to a for-loop. For small datasets, this doesn't matter, but when simultaneously loading multiple datasets, each with >4 million Gaussians, the additional overhead of map quickly becomes costly.
• The enhancementFunc parameter is now optional and only applied if defined, reducing unnecessary processing for the default case.

[weegeekps]

Hi, @drumath2237. This performance issue bubbled up for us with CesiumJS. It would be great to know when you're planning on performing the next release for spz-loader. We are going live with our next release of CesiumJS on the 1st of April. This issue isn't a critical issue for us, but we are rolling out some 3DGS improvements in April that would benefit from this fix.

To make it clear, my intent is not to rush you but just to understand so we can plan appropriately.

Thank you!

[drumath2237]

@weegeekps

Thank you for your contribution!
I will review this tomorrow or later, as soon as I possible.

[drumath2237]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors the floatVectorToFloatArray function to improve performance when handling large data volumes. The changes involve using HEAPF32.slice for more direct memory copying from the WASM heap and employing a for loop for conditional data processing, which is more efficient than the previous .map implementation. The enhancementFunc is also now optional, avoiding unnecessary work. After a thorough review, I have not identified any issues or further improvement opportunities in the submitted changes.

[drumath2237]

LGTM!

When I tested it locally with racoonfamily.spz, the decoding speed improved by about 100–300 ms. Thank you so much!

[drumath2237]

@weegeekps

I have just published version 0.3.1 of spz-loader/core.
Please check it out.

https://github.com/drumath2237/spz-loader/releases/tag/v0.3.1