Repository navigation
allightmapblur_gl_test compares the dispatch with the fragment passes in single precision, past AMD's 8-bit export rounding - #554
Conversation
…nt passes' single precision, which AMD's 8-bit targets do not round as the store does On the RX 6800 (driver 26.8.1) the texel for texel case failed at every blur size and convention: about 4,200 of 94,357 pixels, 4.4%, where the dispatch's occlusion was exactly one 8-bit step above the fragment passes'. The driver hands a fragment shader's colour to an RGBA8 target at half precision, rounded toward zero, before rounding it to 8 bits: the light map held exactly that of the passes' single precision at all 84,337 drawn texels. The dispatch's rg8 store rounds the single precision itself to the nearest 8 bits. The two blurs agree; only the targets' rounding of them differed. The exact case's second pass now draws into a single-precision copy of the light map, seeded with its texels so the sky keeps them, and the dispatch's store must hold the 8 bits nearest it, or at a tie either of the two. One texel lay at 165.50002 steps, which the store rounded down. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Summary
Merge Risk: ⚪ Minimal · up to
|
| Check name | Status | Explanation | Resolution |
|---|---|---|---|
| Docstring Coverage |  |
Docstring coverage is 40.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 5 functions across 2 files. | Write docstrings for the functions missing them to satisfy the coverage threshold. |
-
Check name Status Explanation Title check The title clearly identifies the test and the main change: comparing single-precision fragment passes with the dispatch while accounting for AMD 8-bit export rounding. It is somewhat long but remains … Description check The description gives clear problem, cause, change, and testing details. It is mostly complete and directly supports the pull request objectives, although it does not include the template's Related Is… Linked Issues check Check skipped because no linked issues were found for this pull request. Out of Scope Changes check Check skipped because no linked issues were found for this pull request. Full details: Title check
-
Explanation
The title clearly identifies the test and the main change: comparing single-precision fragment passes with the dispatch while accounting for AMD 8-bit export rounding. It is somewhat long but remains specific and relevant.
Full details: Description check
-
Explanation
The description gives clear problem, cause, change, and testing details. It is mostly complete and directly supports the pull request objectives, although it does not include the template's Related Issues section, completed checklist, or Additional Notes section.
-
- Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.
A rabbit checks the light-map glow
Float targets catch the colors’ flow
The fragment pass writes true
The test checks each channel, too
Then hops away through pixels bright
Comment @coderabbitai help to get the list of available commands.
Problem
allightmapblur_gl_test's "texel for texel through a single-precision scratch" case fails on an AMD RX 6800 (Windows, driver 26.8.1). It fails at every blur size and both depth conventions: about 4,200 of 94,357 pixels (4.4%) have the dispatch's occlusion exactly one 8-bit step above the fragment passes'.Cause
The driver doesn't round a fragment shader's colour straight to an RGBA8 target. It first hands it over at half precision, rounded toward zero, which drops a value just above a rounding midpoint below it. The dispatch's
rg8store rounds the single-precision value to the nearest 8 bits.I checked both with a temporary diagnostic:
unorm8(fp16_rtz(f))of the passes' single-precision value at all 84,337 drawn texels.The two blurs agree; only the targets' rounding of them differed. The viewer is unaffected beyond that rounding.
Change
blurFragment's second pass draws into a new single-precision copy of the light map,mFloatLight. It is seeded with the light map's texels so the sky keeps them.The comparison stays as sensitive to a difference in the blur's arithmetic as before: one shows wherever it moves the rounding. The other cases and
alcompute_gl_benchkeep the 8-bit targets.Testing
On the RX 6800:
allightmapblur_gl_testpasses, three runs in random order.alrender_gl_testspass.🤖 Generated with Claude Code