Date: 2026-04-05 Branch: wgpu-device-renderer Last commits: 6c4856545 (gpu_buffer textures), f70fae350 (REPETITION variants)

Current (P11): 309/413 pass (74.8%) — 104 failures Previous (P10): 303/413 pass (73.4%) — 110 failures Net change: +6 tests passing (+1.4 percentage points)

backface (10/10), crash (2/2), scrolling (26/26), snap (3/3), split (16/16), tiles (5/5)

• filters: -1 (51→50). filter-blur-downscaled-task regressed (max diff 147). Likely atlas-layout sensitivity: the gpu_buffer fix means cs_* gradient shaders now write real gradient content into the texture cache, which may shift atlas packing and affect blur filter readback coordinates.

File: renderer/mod.rs — draw_cache_target_tasks_wgpu()

Root cause: draw_cache_target_tasks_wgpu() builds a base_textures struct for cs_* shader dispatches (linear/radial/conic gradient, blur, etc.). This struct lacked gpu_buffer_f and gpu_buffer_i bindings. The CsLinearGradient, CsRadialGradient, and CsConicGradient shaders all call sample_gradient() → fetch_from_gpu_buffer_2f() → texelFetch(sGpuBufferF, ...). Without the binding, the shader sampled from a dummy texture, producing white output.

Fix: Added gpu_buffer_f and gpu_buffer_i to base_textures:

let base_textures = TextureBindings {
    gpu_cache: ctx.gpu_cache,
    gpu_buffer_f: ctx.gpu_buffer_f,
    gpu_buffer_i: ctx.gpu_buffer_i,
    dither: ctx.dither,
    ..Default::default()
};

Impact: Fixed gradient rendering for all non-fast-linear gradients. +3 gradient tests, +3 border tests (border gradients), +1 image test.

Files: device/wgpu_device.rs, renderer/mod.rs

Root cause: The build system generates brush_image WGSL variants with WR_FEATURE_REPETITION (which enables UV tiling via fract() in brush_image.glsl). The wgpu backend had no corresponding WgpuShaderVariant entries. All batches with BatchFeatures::REPETITION were silently routed to the non-repeating BrushImage variant, causing tiled content to render once without repeating.

Fix:

• Added BrushImageRepeat and BrushImageRepeatAlpha enum variants
• Added shader_key/from_shader_key mappings with feature strings "ANTIALIASING,REPETITION,TEXTURE_2D" (alphabetically sorted)
• Changed batch_key_to_pipeline_key() to accept BatchFeatures parameter and route based on BatchFeatures::REPETITION
• Updated all 3 call sites to pass batch.features

Impact: Fixed tiling for gradients and images rendered via BrushImage.

• tiling-linear-3: max diff 254→3
• tiling-radial-3: max diff 254→4
• tiling-conic-3: max diff 223→4
• Other tiling tests also improved but still fail on dithering tolerance

Investigated the 4 gradient tests with max difference 255 (fundamentally wrong output, not dithering). None are gradient shader bugs:

State leakage bug. Passes when run in isolation. Fails when preceded by other tests in the full suite. The rendering output is pixel-identical to the reference when captured via png subcommand. The reftest runner's readback or render state is contaminated by a preceding test.

Clip mask rendering bug. The test uses a circular clip (radius=300 on a 300x300 rect, creating a full circle). The wgpu backend renders the gradient into the full square without applying the circular clip mask. Pixels at the square corners show gradient colors where the reference has white (background). This is the same class of bug affecting the 6 clip category failures.

Texture cache tile leakage. The test renders a 2048x2048 conic gradient masked by white rects to show only a 250x250 corner. A stray 50x50 yellow square appears near the bottom-center of the output, absent from the reference. This is likely a texture cache atlas tile that bleeds through due to incorrect tile bounds or UV clamping in the large gradient case.

Clip mask rendering bug. Uses a rounded-rect clip (radius=32) on a large 750x500 primitive. Same category as conic-color-wheel — the clip mask isn't fully applied at the rounded corners. Only 995 pixels are affected (the corner regions).

Gradient and conic tests with max diff 3-4, caused by float-to-unorm rounding differences between GL-rendered reference PNGs and wgpu shader output. Not fixable without increasing fuzzy tolerances or regenerating references.

clip (6), conic-color-wheel, radial-border-radius-large, and likely some compositor/compositor-surface failures involve clip masks not being correctly applied. This is a high-value fix target.

Dominated by backdrop-filter (7) and SVG filter (6) failures. Backdrop filters likely need the same two-pass approach as MixBlend.

At least 1 test fails due to render state not being properly reset between tests. May affect more tests when preceded by specific test patterns.

text (11), image (6), compositor-surface (6), compositor (4), mask (4), boxshadow (3), border (2), transforms (2), aa (1), performance (1).

1. Clip mask rendering — fixing the circular/rounded-rect clip mask application would fix ~12 tests across gradient, clip, and possibly compositor categories. Investigate how clip masks are dispatched in the wgpu backend and whether they're correctly sampled during BrushImage and gradient rendering.

2. Backdrop filter — 7 tests. Similar to MixBlend two-pass rendering.

3. State leakage — investigate what render state persists between tests causing gradient_cache_clamp to fail in the full suite. May be a depth/stencil clear issue or texture cache state.

4. SVG filters (6) — may need additional shader variants or filter graph processing.

5. Text failures (11) — shadow rendering (5), subpixel AA / dual-source blending, and other text-specific issues.

6. C4 render pass sharing — optimization plan at logical-whistling-mochi.md

• BrushImageRepeat and BrushImageRepeatAlpha enum variants
• shader_key() / from_shader_key() mappings for REPETITION variants
• instance_layout() match arm updated

• draw_cache_target_tasks_wgpu(): Added gpu_buffer_f and gpu_buffer_i to base_textures (commit 6c4856545)
• batch_key_to_pipeline_key(): Added features parameter, routes BrushImage batches based on BatchFeatures::REPETITION (commit f70fae350)
• Three call sites updated to pass batch.features