iter19 close + r5 full sweep scoreboard

Adds iter19 campaign-close doc (mesa-panvk-bifrost r7 ship, XFB
packed-varying channel-extract fix). Also lands the r5 full
dEQP-VK sweep scoreboard (2.26M tests, 97.65% runnable pass rate)
that surfaced the holes_vert crash in the first place.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
2026-05-25 11:55:53 +02:00
parent 6019cce7d1
commit 02b212b29b
2 changed files with 176 additions and 0 deletions
+98
View File
@@ -0,0 +1,98 @@
# iter19 — XFB store channel-extract fix for packed varyings (campaign close)
Shipped: 2026-05-25
PR: https://git.reauktion.de/marfrit/marfrit-packages/pulls/96
Merge commit: 902de73a02d9
Package: `mesa-panvk-bifrost-26.0.6.r7-1-aarch64`
## What shipped
Single 21+/3- line patch to `src/panfrost/vulkan/panvk_vX_xfb_lower.c`
(`0007-panvk-bifrost-xfb-component-base-fix.patch`). Eliminates a
reproducible SIGSEGV in `vkCreateGraphicsPipeline` for any shader with
XFB-bound varyings declared at non-zero `layout (component=N)`.
## How surfaced
The r5 full dEQP-VK sweep (2,258,378 tests, 97.65% runnable pass rate)
on 2026-05-24/25 included 27 fails in `transform_feedback` plus a
SIGKILL on the chunk when `transform_feedback.simple.holes_vert` hit
`FATAL ERROR: Test program crashed`. Isolated repro produced a userspace
SIGSEGV with no kernel GPU fault (dmesg clean), pointing at a pure
libvulkan_panfrost bug.
## Root cause
`lower_xfb_output_iter17` (and identically upstream
`pan_nir_lower_xfb.c::lower_xfb_output`, which carries a `// TODO`)
computed the source-channel mask as `mask << channel_idx`. `channel_idx`
is the varying-location component (0..3) but `src` only contains
channels starting at `nir_intrinsic_component(intr)`. For a scalar
declared `layout (component=2) flat out float vegeta`, NIR emits
`store_output src=<vec1>, component=2`, and the lowering computed
`mask << 2` against a 1-component src — out-of-range; the malformed
nir_def then segfaulted during downstream NIR constant-folding in
`nir_constant_expressions.c::evaluate_*`.
The `assert(nir_intrinsic_component(intr) == 0)` precondition was
inherited from upstream Mesa as a documented `// TODO`; release builds
(`-DNDEBUG`) elided it, turning the precondition violation directly
into a SIGSEGV.
## Fix
1. Compute `src_channel = channel_idx - nir_intrinsic_component(intr)`
and use `mask << src_channel` instead.
2. Convert both elided asserts to explicit release-mode early-return
guards (closes the same elision class as the original bug).
3. Add a dispatcher-side comment explaining why `i*2+j` is the
varying-location component index.
## Verification
| Family | Result |
|----------------------------------------------|-------------------|
| `transform_feedback.simple.holes_vert` | Crash → Fail (color) |
| `transform_feedback.simple.holes_extra_draw_vert` | Crash → Fail (color) |
| `transform_feedback.simple.basic_*` | 36/36 Pass |
| `transform_feedback.simple.depth_clip_*` | 1 Pass + 4 NotSupp |
| `transform_feedback.simple.lines_or_triangles*` | 16 NotSupp |
| `transform_feedback.simple.holes_geom*` | NotSupp (no GS on G52) |
Zero new regressions on previously passing tests.
## Process
Full 8-step bugfix process followed:
- Phase 0 — characterize: `iter19/phase0_holes_vert_close.md`
- Phase 1 — source map: `iter19/phase1_holes_vert_situation.md`
- Phase 2 — root-cause + ranked options: `iter19/phase2_holes_vert_situation.md`
- Phase 3 — implement + inner-test: `iter19/phase3_holes_vert_close.md`
- Phase 4 — wider verify (focused subset)
- Phase 5 — 2nd-model review: **APPROVE WITH CHANGES (non-blocking)** → defensive-guards added
- Phase 7 — final retest: clean
- Phase 8 — ship: r7 PR → CI green → marfrit repo published → ohm `pacman -S` upgraded → smoke test confirms Crash→Fail
## Three-point ship-check verdict
1. ✓ PR #96 merged at `902de73a02d9`
2. ✓ CI runs #1380 (`mesa-panvk-bifrost-aarch64`) and #1381 (`mesa-panvk-bifrost-video-aarch64`) both `success`
3.`pacman -Q mesa-panvk-bifrost` on ohm reports `26.0.6.r7-1`; installed lib BuildID `9f8dacfc...`; holes_vert no longer SIGSEGVs
## Open follow-ups (not blocking r7 ship)
- **iter20: holes_vert color-check residual** — when `panvk_per_arch(nir_lower_xfb)` removes the original store_output post-XFB-lowering, varyings that are *both* XFB-bound *and* read by the fragment shader lose their rasterizer-side path. holes_vert's fragment shader sees `goku`/`vegeta` as undef → outputs black instead of expected blue. Naive "keep store_output" probe fixed color but blew up compile-time on `max_output_components_128` — needs a more nuanced fix.
- **iter20+: pre-existing latent crashes** unmasked by r7 — `dEQP-VK.transform_feedback.simple.max_output_components_{64,128,256}` coredump on shipped r6 baseline too (confirmed independently); they were never reached during the r5 sweep because the watchdog killed transform_feedback after holes_vert.
## Lineage
| rev | What | Date |
|-----|-----------------------------------------------|------------|
| r1 | KHR_robustness2 + nullDescriptor + nullDescriptor on Bifrost | iter8 |
| r2 | has_vk1_1/has_vk1_2 = true on Bifrost | iter9 |
| r3 | VK_EXT_transform_feedback (iter13) | iter13 |
| r4 | XFB primitive decomposition (iter17) | iter17 |
| r5 | fragmentStoresAndAtomics = true | 2026-05-23 |
| r6 | VK_EXT_legacy_dithering | 2026-05-25 |
| **r7** | **XFB packed-varying channel-extract fix** | **2026-05-25** |