Gap 27: CPU-GPU equivalence verified on an AMD Radeon 8060S (Vulkan, Linux)

First non-Apple GPU verification: 16 tests / 135 cases pass, 124 at
max 1 LSB, 10 bit-identical, one at max 2 LSB (0.0002% of pixels), worst
mean 0.0003 LSB. Recorded in gpu-pipeline.md, hardware-parity.md and
gaps.md. Also as_chunks in spots.rs for the newer clippy lint.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

crates/pipeline/src/spots.rs

@@ -247,8 +247,8 @@ fn fill_box(src: &mut [[f32; 3]], tgt: &[[f32; 3]], alpha: &[f32], bw: usize, bh
     let flat: Vec<f32> = tgt.iter().flatten().copied().collect();
     let out = lightcraft_inpaint::complete(bw, bh, 3, &flat, &hole, &lightcraft_inpaint::CompleteParams::default())
         .unwrap_or_else(|| lightcraft_inpaint::membrane_fill(bw, bh, 3, &flat, &hole));
-    for (d, v) in src.iter_mut().zip(out.chunks_exact(3)) {
-        *d = [v[0], v[1], v[2]];
+    for (d, v) in src.iter_mut().zip(out.as_chunks::<3>().0) {
+        *d = *v;
     }
 }
 

docs/gaps.md

@@ -1,6 +1,6 @@
 # Gaps: where LightCraft falls short of Lightroom
 
-> **Last reviewed:** 2026-10-11 · **Last updated:** 2026-10-11 · **Change:** minor (CR3 entry names the EOS R7 C-RAW fix and the refused CRX variants; CI and Windows / Linux entries note the 0.6.0 riscv64 build; camera colour: tone curve from black-and-white camera JPEGs; Compare and Reference get sharp zoom windows; gap 4 shrinks; gap 14: the HDR switch is read from crs XMP; gaps 2 and 10: EOS R5 Mark II CR3 verified) · **Target:** Adobe Lightroom Classic 15.5 / Lightroom 9.6
+> **Last reviewed:** 2026-10-11 · **Last updated:** 2026-10-12 · **Change:** minor (gap 27: GPU equivalence verified on an AMD Radeon on Linux) · **Target:** Adobe Lightroom Classic 15.5 / Lightroom 9.6
 
 Every known shortfall, one entry each, ranked by user impact (how many Lightroom users hit it, and
 how hard it blocks them), then by effort. **Agents pick work from the top** (the first entry nobody is
@@ -41,7 +41,7 @@ Hours are Opus 5.5 agent wall-clock hours (see [target-app-parity.md](target-app
 | 24 | **Brush live preview and masking feel** (#517, #518); Feather / Edge mask refinement (15.5) | issues | masking feels slower than Lightroom | 8–12 | [ui-parity](ui-parity.md) |
 | 25 | **File writes**: lossy DNG, Render to DNG; AVIF import (PSD export landed with Edit in PhotoCraft, #953) | file-format-parity.md | interchange edge cases | 21–37 | [file-format-parity](file-format-parity.md) |
 | 26 | **Performance at scale**: 85k-photo library opens in 1.7–4.7 s; 100k grid test pending; colour NR at half resolution | ROADMAP log, gpu-pipeline.md | large-catalog users | 15–30 | [target-app-parity](target-app-parity.md#by-dimension) |
-| 27 | **Windows / Linux runtime coverage**: Windows installer UI unverified on Windows (#79); GPU path proven on Apple only; the Linux riscv64 tarball (new in 0.6.0, #844) is cross-compiled and never run beyond a QEMU smoke test | ROADMAP, issues, `release.yml` | non-Mac users find bugs first | 15–30 | [hardware-parity](hardware-parity.md) |
+| 27 | **Windows / Linux runtime coverage**: Windows installer UI unverified on Windows (#79); GPU path proven on Apple and on one AMD GPU on Linux (Radeon 8060S, Vulkan); NVIDIA, Intel and DX12 unverified; the Linux riscv64 tarball (new in 0.6.0, #844) is cross-compiled and never run beyond a QEMU smoke test | ROADMAP, issues, `release.yml` | non-Mac users find bugs first | 13–27 | [hardware-parity](hardware-parity.md) |
 | 28 | **Small Classic shortcuts and keys**: rating `[` `]`, flag cycling, filter-bar keys, F1 help, panel keys (Tab, F5–F8) | [ui-parity](ui-parity.md#shortcuts) | Classic muscle memory | 4–8 | [ui-parity](ui-parity.md) |
 
 Beta-blocking gaps (B) total **~230–430 h**; with the stability and platform work they need around them,
@@ -72,6 +72,7 @@ Beta-blocking gaps (B) total **~230–430 h**; with the stability and platform w
 
 | Date | Change | Summary |
 |---|---|---|
+| 2026-10-12 | minor | Gap 27: the CPU ↔ GPU equivalence suite passes on an AMD Radeon 8060S (Vulkan, Linux), the first non-Apple GPU verified; 15–30 h → 13–27 h; gap not closed |
 | 2026-10-11 | minor | Reviewed against main after 0.6.0: CR3 entry names the CRX variants `crx.rs` still refuses and the R7 / R6 Mark II C-RAW state (decoding, not verified, so the measured numbers don't move); CI and Windows / Linux entries note the riscv64 release build (#844); R7 fix added to Closed recently. Panel docking (UI.9, #869) closes no listed gap |
 | 2026-10-11 | minor | Camera colour entry: raws with a black-and-white camera JPEG take its tone curve (#577); gap not closed |
 | 2026-10-11 | minor | Gap 4: Compare and Reference views get zoom windows per pane (1:1 was a 2,560 px render magnified); overlays and soft proofing remain, 8–15 h → 4–8 h |

docs/gpu-pipeline.md

@@ -232,6 +232,12 @@ from outside the window: the vignette geometry arrives as the `VIG_VIEW` paramet
 noise-reduction size come from the whole frame, and a window that spots or Auto Mask strokes grew is
 cut back after the readback. The tests skip (pass with a note) when no adapter exists.
 
+**AMD Radeon 8060S (RADV STRIX_HALO, Vulkan 1.4, Mesa 26.2.4, Linux 7.2), 2026-10-12**, the first
+non-Apple GPU verified: all 16 tests pass, 135 compared cases. 124 differ by max 1 LSB, 10 are
+bit-identical, and one (the point colour range overlay) has max 2 LSB on 0.0002% of pixels. Worst
+mean 0.0003 LSB. That is the same picture as on Metal, inside the bounds
+(`cargo test -p lightcraft-gpu --release --test equivalence -- --nocapture`).
+
 Remaining differences come from f32 vs f64 coordinate math, fast-math transcendental functions on
 Metal and running-sum order in the box filters — all far below one 8-bit step.
 

docs/hardware-parity.md

@@ -1,6 +1,6 @@
 # Hardware parity
 
-> **Last reviewed:** 2026-10-10 · **Last updated:** 2026-10-10 · **Change:** major (created from the code, gpu-pipeline.md and the parity checklist) · **Target:** Adobe Lightroom Classic 15.5 / Lightroom 9.6
+> **Last reviewed:** 2026-10-10 · **Last updated:** 2026-10-12 · **Change:** minor (first non-Apple GPU verification: AMD Radeon 8060S on Linux) · **Target:** Adobe Lightroom Classic 15.5 / Lightroom 9.6
 
 What Lightroom does with the machine's hardware, per platform, and what LightCraft does. Sources:
 [gpu-pipeline.md](gpu-pipeline.md), [display-profiles.md](display-profiles.md), [denoise.md](denoise.md),
@@ -8,7 +8,7 @@ What Lightroom does with the machine's hardware, per platform, and what LightCra
 
 | Feature | Target | LightCraft: macOS | Windows | Linux / BSD | Web | Status |
 |---|---|---|---|---|---|---|
-| GPU for display and editing | Metal / DX12, "Use GPU for display / image processing / export" | wgpu Metal: geometry, resize, blur, WB, NR, guided filter, dehaze, masks, finish; CPU fallback per render | DX12 only (Vulkan opt-in, #136) | Vulkan | CPU only (no WebGPU yet) | 🟡 defringe, spot removal, heuristic masks, long brushes and high-bit exports stay on CPU |
+| GPU for display and editing | Metal / DX12, "Use GPU for display / image processing / export" | wgpu Metal: geometry, resize, blur, WB, NR, guided filter, dehaze, masks, finish; CPU fallback per render | DX12 only (Vulkan opt-in, #136) | Vulkan (CPU ↔ GPU equivalence verified on an AMD Radeon 8060S, RADV) | CPU only (no WebGPU yet) | 🟡 defringe, spot removal, heuristic masks, long brushes and high-bit exports stay on CPU |
 | GPU export | ✅ | ✅ (6000 × 4000 in ~0.3 s vs ~1.2 s CPU) | ✅ | ✅ | — | ✅ (black-export fix on an Intel iGPU, #78) |
 | AI on GPU / neural engine | Metal / DirectML / Core ML, model per feature | AI denoise on GPU or CPU (setting); SAM 3 on Metal via candle; faces on CPU | CPU / GPU denoise | CPU / GPU denoise | — | 🟡 no Core ML / Neural Engine / DirectML |
 | HDR (EDR) display | ✅ (macOS EDR, Windows HDR) | ⬜ | ⬜ | ⬜ | ⬜ | ⬜ (LR-VIEW-HDR-DISPLAY) |
@@ -28,7 +28,9 @@ What Lightroom does with the machine's hardware, per platform, and what LightCra
 **~40% ready** (estimated): the GPU editing path is real and fast (measured: 24 MP exposure drag
 3.9 ms on GPU vs 22–42 ms CPU; NR drag 18 ms vs 301–471 ms), but HDR display, pen pressure,
 tethering, printing, automatic monitor profiles and neural-engine inference are missing. GPU
-correctness is proven on Apple GPUs; other vendors rely on user reports.
+correctness is proven on Apple GPUs and on one AMD GPU on Linux (Radeon 8060S, Vulkan / RADV: the
+equivalence suite passes with the same max 1–2 LSB as on Metal, [gpu-pipeline.md](gpu-pipeline.md#correctness-cpu-oracle-and-equivalence-tests)).
+NVIDIA, Intel and Windows DX12 still rely on user reports.
 
 ## Remaining (Opus 5.5 agent hours)
 
@@ -47,4 +49,5 @@ correctness is proven on Apple GPUs; other vendors rely on user reports.
 
 | Date | Change | Summary |
 |---|---|---|
+| 2026-10-12 | minor | CPU ↔ GPU equivalence suite run on an AMD Radeon 8060S (Vulkan, RADV, Linux): 16 tests / 135 cases pass, max 2 LSB, mean ≤ 0.0003; first GPU verified outside Apple |
 | 2026-10-10 | major | Created |