WebGL2-only: drop WebGL1 path, adopt VAOs (per-flush attribute setup → bindVertexArray) #1509

Activity

1. obiot

   added

   Improvement

   on Jun 21, 2026
2. obiot commented on Jun 28, 2026

   

   obiot

   on Jun 28, 2026

   MemberAuthor

   More actions

   Folded in from #1391 (closed as duplicate)

   #1391 ("Drop WebGL1 support") tracked the same WebGL2-only baseline. Merging its additional analysis here so nothing is lost.

   Extra cleanup item (extends scope A)

   • Remove the preferWebGL1 application setting (and any docs/types referencing it).
   • Simplify IndexBuffer to always use Uint32Array (WebGL2 guarantees OES_element_index_uint), and drop the WebGL1-specific bufferData overloads (WebGL2 accepts the offset+length form).

   Shader modernization — GLSL #version 300 es (not yet in this ticket's scope)

   Going WebGL2-only also unlocks modern shaders:

   • attribute/varying → in/out; texture2D() → texture(); gl_FragColor → an explicit out variable.
   • Sampler array indexing (texture(uSamplers[idx], uv)) — eliminates the if/else sampler chain in the multi-texture batching fragment shader (a real per-fragment win + simpler codegen).
   • Per-variable precision qualifiers instead of the single global one (src/video/webgl/utils/precision.js).

   Other WebGL2-only wins now unblocked (extends scope C)

   Beyond the texStorage2D / MSAA / instancing / transform-feedback already listed:

   • UBOs (Uniform Buffer Objects) — share projection matrix / global tint / time uniforms across shaders without per-draw re-upload.
   • MRT (Multiple Render Targets) — multi-texture output in one pass (G-buffers / deferred effects).
   • 3D textures / texture arrays — e.g. a single texture array for tilemap rendering instead of multiple atlases.
   • NPOT textures with no isPowerOfTwo / nextPowerOfTwo guards (the repeat-guard removal is already in scope A; this also drops the POT checks in src/video/texture/cache.js).

   Additional references (from #1391)

   • Context creation also lives in src/video/rendertarget/canvasrendertarget.js (besides webgl_renderer.js).
   • POT checks: src/video/webgl/batchers/material_batcher.js, src/video/texture/cache.js.
   • Shader precision: src/video/webgl/utils/precision.js.
3. obiot commented on Jul 29, 2026

   

   obiot

   on Jul 29, 2026

   MemberAuthor

   More actions

   Acceptance criteria status (PR #1553)

   Four of five are met. Recording the fifth honestly rather than closing quietly over it.

   • ✅ WebGL renderer creates only a webgl2 context; missing → Canvas2D, with the Camera3d should fail loudly when used with the Canvas renderer #1479 loud signal preserved for an explicit renderer: video.WEBGL
   • ✅ All WebGLVersion > 1 branches removed; single code path
   • ✅ Attributes specified once per layout, not per flush — verified by GL spy in tests/webgl_vao_call_counts.spec.js: 19 vertexAttribPointer + 19 enableVertexAttribArray at startup (4+5+3+3+4 across the five batchers), 0 per frame thereafter, and 0 across 100 alternating batcher switches
   • ✅ VAOs recreated after context loss/restore — tests/webgl_vao_recreation.spec.js, driving a real WEBGL_lose_context cycle
   • ⚠️ "No visual regression … FPS/draw-time on the night-city stress scene improves measurably" — visual regression: verified (7 examples captured before/after; four pixel-identical, the three animated ones below their measured same-code capture noise). Frame-time: not measurable.

   Measured on night-city

   Batcher switches / frame	2.88
   Indexed flushes / frame	43.1
   GL calls eliminated / frame	~310
   Measured cost per eliminated call	0.018 µs (desktop Chrome command buffer)
   Time saved / frame	~5.6 µs

   For comparison, a 2D scene (platformer) saves ~42 calls/frame, so the stress scene does show the ~7× larger saving this ticket predicted — the mechanism is confirmed. But ~5.6 µs against a 16.6 ms budget is roughly 0.03%, which is not observable as FPS. Measured headless (SwiftShader), so absolute timings on real hardware will differ; the call counts are exact and hardware-independent.

   One correction to the ticket's premise

   The motivation says the batcher re-specifies attributes "on every flush". That was true only for the mesh family: base Batcher.flush() bound its own glVertexBuffer and re-applied the pointers, because vertexAttribPointer captures whichever ARRAY_BUFFER is bound when it is called — so switching from the shared buffer to its own invalidated the pointers set at useShader() time. QuadBatcher.flush() overrode the base and did not re-specify, since it uploads into the already-bound shared renderer.vertexBuffer.

   That asymmetry was correct behaviour for a no-VAO design, not redundancy — and it is exactly what a VAO removes, since the VAO records buffer and pointer together, so restoring it needs no re-pointing after a buffer switch. It does mean the per-flush cost this ticket describes was a 3D-mesh cost, not a 2D-sprite one, which is why sprite scenes save far less than the framing implies.

   Suggestion

   Amend criterion 5 to "GL call count per frame drops from O(flushes × attribs) to O(layouts), verified by GL spy" — which is met and machine-checkable — and drop the FPS clause, or re-run night-city on real hardware before closing if a frame-time figure is wanted.
4. obiot

   added a commit that references this issue on Jul 29, 2026

   WebGL 2 only renderer + Vertex Array Objects (#1509) — 20.0.0 (#1553)

   ...

   a6284cc
5. obiot

   closed this as completedin #1553on Jul 29, 2026
6. obiot

   added a commit that references this issue on Jul 31, 2026

   test: scope the WebGL tripwire to CI so a GL-less machine still runs

   ...

   303fd14