Reduce Apple OpenGL vertex state churn

[LisaScheers]

Description

Reduces redundant OpenGL state translation on Apple GPUs by retaining the current vertex/index buffers across shader transitions and caching the vertex attribute pointer formats already valid for the bound buffer.

Only newly required formats are submitted. Color/emissive pointer remapping is preserved, cache state is invalidated on buffer replacement or explicit unbind, and non-Apple drivers retain the existing behavior.

A native Apple Silicon profile of the closely related Firestorm viewer showed repeated time in glVertexAttribPointerARB_Exec, OpenGL object-name hash operations, and dispatch/state updates. The official develop branch has the same unconditional shader-transition unbind and pointer resubmission path.

Related Issues

• Please link to a relevant GitHub issue for additional context.

Issue Link: None; opened as a draft for upstream performance evaluation.

---

Checklist

• I have provided a clear title and detailed description for this pull request.
• If useful, I have included media such as screenshots and video to show off my changes.
• The PR is linked to a relevant issue with sufficient context.
• I have tested the changes locally and verified they work as intended.
• All new and existing tests pass.
• Code follows the project's style guidelines.
• Documentation has been updated if needed.
• Any dependent changes have been merged and published in downstream modules.
• I have reviewed the contributing guidelines.

---

Additional Notes

Validation performed:

• clean git diff --check
• exact stable patch-ID match with the Firestorm implementation
• reviewed cache invalidation and every color/emissive attribute combination
• verified the non-Apple path remains unchanged

A full official viewer build was not available because this host lacks the autobuild dependency bundle and full Xcode toolchain. This remains a draft pending native build and benchmark coverage.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[LisaScheers]

recheck

[Copilot]

Pull request overview

This PR reduces redundant OpenGL state work on Apple GPUs by avoiding unconditional vertex/index buffer unbinds on shader transitions and by caching which vertex attribute pointer formats have already been submitted for the currently bound array buffer.

Changes:

• Skip LLVertexBuffer::unbind() during LLGLSLShader::bind() on Apple to retain buffer bindings across shader transitions.
• Add an Apple-only cache in LLVertexBuffer::setupVertexBuffer() to only submit newly needed glVertexAttribPointer* formats, with invalidation on buffer change or explicit unbind.
• Preserve existing color/emissive attribute pointer remapping while avoiding redundant resubmission on Apple.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
indra/llrender/llvertexbuffer.cpp	Adds Apple-only vertex attribute pointer caching and invalidation to reduce redundant glVertexAttribPointer* calls.
indra/llrender/llglslshader.cpp	Avoids unbinding VBO/IBO on shader transitions on Apple to reduce driver translation churn.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Geenz]

I'll take a closer look at this this week.

[Geenz]

@LisaScheers Can you share Tracy runs of this change?

[LisaScheers]

@Geenz Yes,
here are the 60-second Tracy 0.11.1 captures:
https://drive.google.com/drive/folders/1gd4ycQWzA2srAPiib4uCbV4Q-K9vkNDh?usp=drive_link

Both are ReleaseOS builds using identical dependencies and settings on the same Apple Silicon Mac.

Metric	Baseline	PR	Change
Foreground Render mean	12.07 ms	10.84 ms	−10.2%
Foreground Render median	11.25 ms	10.44 ms	−7.2%
Foreground Render p95	16.13 ms	14.55 ms	−9.8%
Foreground Render p99	23.94 ms	21.13 ms	−11.7%
LLGLSLShader::bind() inclusive mean	0.712 µs	0.617 µs	−13.3%
LLGLSLShader::bind() self mean	0.214 µs	0.179 µs	−16.4%

The PR processed 10.5% more shader-bind calls and approximately 3–7% more batch calls per frame, so it was not obviously benefiting from a lighter draw workload.
One caveat: the baseline was briefly treated as unfocused during this run, triggering the viewer’s 40 ms background sleep. Therefore, the raw FPS difference is not valid. The render figures above exclude those background frames. I’m collecting additional interleaved runs with foreground state controlled.