LineMaterial: Fix world-unit extrusion with orthographic cameras. - #34540
LineMaterial: Fix world-unit extrusion with orthographic cameras.#34540diarmidmackenzie wants to merge 9 commits into
Conversation
📦 Bundle sizeFull ESM build, minified and gzipped.
🌳 Bundle size after tree-shakingMinimal build including a renderer, camera, empty scene, and dependencies.
|
|
I've also patched the TSL version of the line material. The unit tests is problematic because of the intermediate renderer creation. I've removed it for now since the test failed and throwed context creation errors. |
|
Thanks for looking at this. Are you happy to merge without tests, or do we need to come up with another plan for testing? |
…variant Replaces the hand-written override with the exact line from mrdoob/three.js#34540 (milestone r187), and drops everything that existed only to support a local variant of it. Byte-identical to upstream is the point, not the line count. It makes this visibly the same fix rather than a lookalike, and it makes removal free: resolveVertexShader() now recognises three states -- already fixed, stock, or neither -- so when the bundled three ships r187 it finds the fixed line present and returns the shader untouched. Deleting this block at that point changes no pixel and needs no coordination with a super-three bump. Dropped along the way: - The bespoke D4 degenerate-cross guard. It fires exactly when the segment runs along the view axis, which is exactly when the extruded quad has zero screen area -- measured, it cannot change a rendered pixel either way. All it bought was a defined zero-area quad instead of NaN vertex positions, at the cost of diverging from upstream on a case that cannot render. If that NaN is worth closing it belongs in the three.js PR, not in a vendored copy. - check-shader-target.mjs and its `npm run dist` gate. Its premise was that the throw fired at first LineMaterial construction, so a broken bump would survive a build and a smoke load. Hoisting the resolution to module scope already fixed that -- it now throws when the bundle loads, on any page -- so the script was a second copy of the target strings guarding a hole that was already closed, and it needed its own anti-drift check to stay honest. This departs from the plan's D2; restoring it is one file. Re-measured after the change: orthographic 400/400/400/400/650 against 400/400/400/400/650 predicted from geometry, and all five perspective fixtures still byte-identical. Asserted against the running material built from this branch's bundle: the compiled shader carries the fixed line and no longer carries the stock one. Net: 258 insertions -> 86, over one fewer file. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013TR1kzRUyHfCWmFXYgpi15
|
Sorry, but I'm afraid the PR is not ready to merge. If you add an orthographic camera to Dev: 
PR: 
Open both screenshots in separate tabs and switch back and forth. You will see that your change cuts off parts of the segment at the right. The root cause is a mismatch in your version. You have updated the vertex shader but the fragment shader still seems to work with a perspective ray. |
🖼️ E2E screenshot tests✅ All examples render correctly again (run). |
|
Thanks for spotting that problem. Fragment shader now also updated in the PR. I've added a camera picker (perspective / orthographic) to I've also created a demo here showing original, my 1st fix, and the latest fix, for a couple of different lines, at a variable offset from an orthograhic camera. |
2 participants
World-unit lines rendered with an orthographic camera become narrower away from the image centre and thin strokes can disappear. The vertex shader builds its extrusion basis from the camera-space segment midpoint, which assumes perspective viewing rays.
Use camera-space -z for the orthographic extrusion basis, selected using the existing projection-matrix classification. The perspective expression is preserved, and the same material can switch between camera types.
This moves the extrusion fix described in diarmidmackenzie/aframe-components#56 into LineMaterial itself.
Adds a QUnit rendering regression covering centred and off-axis horizontal, vertical and diagonal segments, an asymmetric orthographic frustum, and camera switching. The fixture uses an uninterrupted dash to isolate vertex extrusion from the solid-line fragment endcap calculation.
Validation:
npm run lintpasses.git diff --checkpass.The rendering test was run through a temporary Puppeteer harness loading
UnitTestsAddons.html?filter=LineMaterial; the standard runner splits its testPage argument at every=and drops the filter value. The full unit suite was not run.