OpenMW on Apple Silicon: the OpenGL core-profile port and its benchmark harness

This repository holds the working notes, tools and results of an effort to get OpenMW off Apple's legacy OpenGL 2.1 context on Apple Silicon: an OpenGL core-profile port of its renderer, and the harness used to measure it. The code changes themselves live on branches of the two upstream projects (see below); this repo is the harness, the measurements and the write-ups that justify them.

The work was done with heavy LLM assistance. Every change was verified by the harness in this repo — screenshot diffs against the unmodified renderer and frame-time statistics on fixed stations — and the survey documents record what broke, why, and how it was found, so the reasoning can be checked rather than trusted.

What is here

path	what
CORE-PROFILE-SURVEY.md	Design and findings of the port: why a core profile, what had to change and how, and the non-obvious bugs found on the way (two of them in OpenSceneGraph itself).
HARNESS.md, harness/	The benchmark and regression harness: a Lua driver mod that pins weather/time/camera at fixed stations, a runner that captures per-frame statistics and screenshots, comparison tools, and results/ with every run referenced in the write-ups.
harness/RESULTS.md	The numbers: baselines, the heavy ("modded") profile, and core versus compatibility on every station.
PATCHES.md	Map of the upstream patch series for OpenMW and for OpenMW's OpenSceneGraph fork, with what each commit does and how it is meant to be reviewed.

The code

• OpenMW: branch core-profile (three commits over upstream master) adds [Video] opengl profile = core. Compatibility stays the default; core is opt-in. On macOS it is the only way past OpenGL 2.1. This branch is the reference tree; upstream is removing fixed-function state along its own path (#9239) and the work is being offered in pieces that fit it.
• OpenSceneGraph (OpenMW's fork, branch 3.6): two bug fixes found on the way, a stale vertex-array-object pointer in VertexArrayState (OpenMW/osg#51) and a use-after-free in GLExtensions during static destruction (OpenMW/osg#50, a backport of mainline's fix), were merged upstream on 2026-09-10; core-profile-fork-v2 (three commits) adds what lets a GL2-profile OSG build run on a core context at runtime.

Result on one Apple Silicon Mac (RESULTS.md, "Core profile versus compatibility, final"): screenshots within 2 % RMSE of the compatibility renderer on every static station, zero GL errors under per-attribute checking, mean frame rate ahead of the compatibility profile on six of nine stations (+9…+38 %, level on two, behind on one) and on the heavy profile at both stations tried (+8…+11 %), with better frame-time tails at every station (1 %-low +4…+190 %).

Reproducing

1. Build OpenMW from the core-profile branch against the patched OSG (build notes in HARNESS.md; the important one: replace OSG dylibs by rename, never in place — macOS kills processes whose mapped libraries change underneath them).
2. Point harness/config/openmw.cfg at a Morrowind installation (GOG/Steam; innoextract unpacks the GOG installer on macOS).
3. harness/tools/bench.sh <label> [-s overlays/core.cfg] [-n 3] [stations…] — one launch per station, per-frame stats and a screenshot each; tools/compare.py and tools/shotdiff.sh compare two runs.

Status and what is next

The port runs end to end on macOS and is measured there; it has not been tried on Windows or Linux core contexts, on OpenCS, or with stereo rendering. It is being submitted upstream in pieces that fit OpenMW's own fixed-function removal (#9239); the patch map is PATCHES.md. What it does not change is the CPU-side draw-submission floor (roughly 6–7 ms per frame on this hardware); that would take a different rendering backend, for which the harness now provides a measured target.