WIP: Typed array encoding in supplyDefaults with Caching #5308
Conversation
Works for initial render, but typed array is overwritten when supplyDefaults is run again.
|
|
||
| // Normalize shape to a list | ||
| if(Number.isInteger(v.shape)) { | ||
| coerced.shape = [v.shape]; |
There was a problem hiding this comment.
Could we possibly support numeric strings here as well?
| if(Number.isInteger(v.shape)) { | ||
| coerced.shape = [v.shape]; | ||
| } else { | ||
| coerced.shape = v.shape; |
There was a problem hiding this comment.
It might be a good idea to drop this array support for shape attribute; while it cannot be a typedArray itself.
In that case for n dimensional arrays one could pass an string containing comma (or X) separated integers.
There was a problem hiding this comment.
At the moment, I am using the shape to build a multi-dimensional array when the length is more than 1. Only the inner most layer of the multi-dimensional array is made up of typed arrays, but I pretty sure it'll be a lot better than nothing.
This reverts commit 5079fc7
The fact that elements aren't regular numbers, and aren't compatible with the Math module functions would make this challenging.
|
Instead of targeting master, could you possibly open a PR to merge your work into #5230 branch. That would help us test and dev this. |
|
Base should be updated now. Thanks! |
|
💃 |
Overview
This PR builds off of #5230, with a candidate solution to the problem outlined in #5230 (comment).
Architecture Details
The approach is to perform the typed array decoding in coerce (before the supplyDefaults logic in the layout and traces). This has the advantage that none of the supply defaults or calc logic needs to be aware of the typed array specification objects.
For performance, the ArrayBuffer instances resulting from the decoding process are cached by trace uid and property name. To keep cache bounded, only one ArrayBuffer is cached per trace per property. Playing with the isosurface example in #5230, building the typed array with this caching approach is about 10x faster than when performing the base64 decoding.
Implementation Details
This PR handles decoding N-dimensional arrays into a nested collection of
Arrays. The inner most layer contains TypedArrays that are backed by the original decoded ArrayBuffer, so no copying is performed. So far, all the 2D traces I've played with have worked fine with this nesting arrangement.More testing is needed, but hopefully nothing in supplyDefaults or calc or rendering will need to change in any of the traces (apart from potential bug fixes for TypedArray handling).