GPU Expressions

[TimSylvester]

We would like to eventually move as much of the style property evaluation as is practical to the shaders. This sets up some foundational aspects of that and allows basic zoom-based expressions in line layers to be evaluated in shaders.

The legacy code used uniforms for these same zoom-based expressions, updating them on each frame. Unfortunately, the entire UBO still needs to be updated if any of the properties are CPU-evaluated, so we don't really gain any efficiency from the binding or data transfer. We skip some of the actual property evaluation, but that barely shows up in profiling.

To be really beneficial, we will need to get to the point where we can use shaders for per-attribute expressions, so that the much larger vertex attribute arrays do not need to be updated constantly. That will involve making the feature data available to the shaders with some pre-processing.

Currently only applies to non-custom line layers.

[github-actions]

Bloaty Results (iOS) 🐋

Compared to main

    FILE SIZE        VM SIZE    
 --------------  -------------- 
  +1.2%  +210Ki  +1.2%  +208Ki    TOTAL

Full report: https://maplibre-native.s3.eu-central-1.amazonaws.com/bloaty-results-ios/pr-2202-compared-to-main.txt

[github-actions]

Bloaty Results 🐋

Compared to main

    FILE SIZE        VM SIZE    
 --------------  -------------- 
  +0.3%  +435Ki  +0.2% +68.0Ki    TOTAL

Full report: https://maplibre-native.s3.eu-central-1.amazonaws.com/bloaty-results/pr-2202-compared-to-main.txt

Compared to d387090 (legacy)

    FILE SIZE        VM SIZE    
 --------------  -------------- 
   +21% +24.2Mi  +409% +24.4Mi    TOTAL

Full report: https://maplibre-native.s3.eu-central-1.amazonaws.com/bloaty-results/pr-2202-compared-to-legacy.txt

[github-actions]

Benchmark Results ⚡

Benchmark                                                     Time             CPU      Time Old      Time New       CPU Old       CPU New
------------------------------------------------------------------------------------------------------------------------------------------
OVERALL_GEOMEAN                                            +0.0049         +0.0053             0             0             0             0

Full report: https://maplibre-native.s3.eu-central-1.amazonaws.com/benchmark-results/pr-2202-compared-to-main.txt

[sjg-wdw]

Looks good to me.

[louwers]

The pre-commit hook tried to format the code, but it could not push because the fork is owned by an organization. You can run pre-commit locally with:

pre-commit run --all-files

Could you update the PR description to include a description of the changes made to the iOS Benchmark app?

The PR title could be a bit more descriptive. This is not the first PR on GPU Expressions, and I am assuming it won't be the last. If you hover over a piece of code or use git blame it's nice to get a short description with a summary to know as part of what effort the code in question was added.

A few more details about the implementation and the reason behind them in the description would also be appreciated. Each PR should ideally have a self-contained description of the changes.

Other than that, code looks good. Thanks!

[TimSylvester]

@louwers I have pre-commit set up to run on commits, and running with --all-files changes a bunch of files that I haven't interacted with, e.g., bin/offline.cpp. It would be better to to apply all those changes to main first, rather than including them here, I think.

[louwers]

@TimSylvester Sure. I'll make a PR.

Edit: I don't see any changes to bin/offline.cpp. #2245

[TimSylvester]

Updated benchmark, pretty much the same.

CPU:

Average frame encoding time: 2.78 ms
Average frame rendering time: 6.55 ms

Average frame encoding time: 2.77 ms
Average frame rendering time: 6.47 ms

Average frame encoding time: 2.79 ms
Average frame rendering time: 6.57 ms

GPU:

Average frame encoding time: 2.76 ms
Average frame rendering time: 6.70 ms

Average frame encoding time: 2.79 ms
Average frame rendering time: 6.77 ms

Average frame encoding time: 2.76 ms
Average frame rendering time: 6.76 ms

So, a small hit in render time with no real benefit to encoding time. The CPU work we're avoiding is not very costly, and we can't avoid updating and binding the evaluated property UBO unless every single property could be done on the GPU.

done