GL skybox/environment renders blown-out white (Newport Loft HDRI) — Vulkan renders it correctly #797

Activity

1. coderabbitai commented on Aug 14, 2026

   

   coderabbitaiBot

   on Aug 14, 2026 – with coderabbitai

   Contributor

   More actions

   🔗 Related PRs

   #404 - fix(renderer): repair reflection-probe bake (was all-black) + add vis… [merged]
   #465 - Fix/flaky fog visual evidence test [merged]
   #674 - fix(renderer): GTAO black-sky artifact — exact texelFetch HZB depth copy + ULP-tolerant far-plane test [merged]
   #750 - feat(renderer): offline reference path tracer for ground-truth GI/BRDF validation (#709) [merged]
   #794 - Vulkan Phase 8: editor-frame parity, MCP diagnostic loop, command-ordered buffers (#691) [open]

   ---

   🧪 Issue enrichment is currently in open beta.

   To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

   issue_enrichment:
     auto_enrich:
       enabled: false

   💬 Have feedback or questions? Drop into our discord!
2. coderabbitai

   added

   bugSomething isn't workingSomething isn't working

   robustnessProduction hardening / shipping robustnessProduction hardening / shipping robustness

   on Aug 14, 2026
3. drsnuggles8

   added 2 commits that reference this issue on Aug 14, 2026

   fix(rhi): Renderer2D rendered vertically mirrored under Vulkan — rout…

   ...

   8e30f45

   fix(gl): framebuffer creation clobbered texture unit 0 — every GL HDR…

   ...

   60cf190
4. drsnuggles8 commented on Aug 15, 2026

   

   drsnuggles8

   on Aug 15, 2026

   OwnerAuthor

   More actions

   Root-caused and fixed (60cf190b on the #691 Phase 8 branch)

   The one-line cause: OpenGLFramebuffer::Invalidate called Utils::BindTexture(id) — which is glBindTextureUnit(0, id) — on every freshly created color and depth attachment. The attachment setup is fully DSA, so the bind did nothing useful; it only replaced whatever was published on texture unit 0.

   Why that whitened every HDRI sky: ConvertEquirectangularToCubemap publishes the loaded HDR pano on unit 0 (TEX_DIFFUSE), then RenderToCubemap creates its capture framebuffer — whose depth attachment landed on unit 0 — so all six face bakes sampled the depth clear value (constant 1.0) instead of the pano. White cubemap → white skybox + white-tinted IBL in all 13 HDRI scenes, on GL only. Vulkan resolves bindings at draw from published state, so it was correct all along — exactly as the sweep suggested.

   The evidence ladder (all via a new env-gated OLO_ENV_BAKE_DUMP diagnostic, kept in-tree):

   1. CPU source mean of newport_loft.hdr: (0.335, 0.236, 0.183) ✓
   2. Uploaded-GL-texture readback: identical to source ✓ (upload exonerated)
   3. Baked faces: uniform wrong values ✗
   4. A uv-visualizing shader variant: per-face uv means correct ✓ (geometry/varyings exonerated — only the sample wrong)
   5. A unit-0 occupancy probe at draw time: unit 0 held GL id 46 (the capture FB's depth attachment) where the pano (id 43) had been published — caught red-handed.

   Fix: delete both Utils::BindTexture calls and the orphaned helper; hoist the cube-mesh build ahead of the publish so no future resource-creation path can slot into the same gap.

   Verified: GL bake per-face means are now byte-identical to Vulkan's; PBRReference renders at a 1.18/255 mean pixel difference between backends (was ~174/255). Full suite green; the RHI boundary ratchet stays at its shrunken baseline; no goldens involved HDRI environments, so none needed rebaking.

   Recommend closing once #794 merges (the fix rides that branch).
5. drsnuggles8

   closed this as completedin 701ef35on Aug 15, 2026