dart_pdf_editor_flutter_gpu 0.2.0

Opt-in Impeller flutter_gpu tile backend for dart_pdf_editor.

Changelog

0.2.0

• Require the dart-pdf 4.0.0 suite and its expanded tile-backend warm-up and transparency interfaces.

• Run requested scene warm-up for route-change benchmarks, matching the viewer's idle-prepared production path while suppressing the unmatched timing marker when the Canvas base has no GPU session.

• Preserve per-paint Normal, Multiply, and Screen state inside isolated offscreen transparency groups.

• Retain ordinary images in single-paint and mixed vector/text transparency groups, reusing the scene texture cache and the exact offscreen group pass.

• Retain outlined text in single-paint and mixed vector transparency groups, including clipped isolated groups that require the exact offscreen tile pass for group alpha and outer blending.

• Retain exact macOS CJK substitutions for Songti, Heiti, Hiragino Sans, and Hiragino Mincho, including CFF outlines inside OpenType collections. Six additional PDF.js corpus pages now stay on the GPU, with a dedicated CJK same-host scenario.

• Render every PDF blend mode exactly with retained GPU destination sampling. Advanced modes use bounded ping-pong tile attachments; provably disjoint paints share one blend pass, while overlapping paints retain painter order. Oversized tiles and scenes that combine advanced blends with offscreen transparency groups keep the conservative Canvas fallback.

• Retain isolated knockout groups made entirely from vector fills. Later sibling shapes use exact source replacement only inside their retained path, while the group's alpha and outer blend remain a single offscreen composite.

• Retain isolated knockout groups containing a base fill and one vector-soft-masked fill, including clipped vector mask bounds.

• Render overlapping Normal paints in isolated transparency groups into a shader-readable tile attachment, then apply group alpha and outer blend once when sampling it into the page pass.

• Elide zero-alpha and zero-area transparency groups before GPU auditing; their composite alpha or form BBox clip makes all nested paints exact no-ops.

• Retain non-Normal transparency groups whose padded fill bounds are provably disjoint, preserving exact outer blending without an offscreen surface.

• Treat C1 controls as non-rendering glyph placements in the exact native outline seam, matching Flutter shaping while preserving their PDF advances.

• Retain stroke-only and fill-plus-stroke outline text, including page-space alpha and zero-width hairlines, by reusing the exact path stroke pipeline. Filled glyphs keep the analytic atlas while their stroke overlays in painter order.

• Retain ordinary filled text as six-vertex glyph quads backed by a scale-independent analytic curve atlas. The shader derives device scale per tile, eliminating repeated flattened outline fans across LoDs while exact stencil rendering remains the automatic fallback for unsupported glyphs.

• Add an exact substituted-text outline seam for the retained backend. Hosts can resolve their registered TrueType/TTC bytes directly, while the opt-in native adapter probes the standard platform substitution faces and declines any unavailable, complex-shaped, or unplaceable run. Add the same-host system-font route-change benchmark to the Metal CI matrix.

• Render zero-width PDF hairlines as exact one-device-pixel contours generated for each tile LoD, while retaining the flattened source path with the scene.

• Collapse isolated transparency groups containing one clipped vector fill or stroke into retained stencil geometry, preserving group alpha and the group's page blend mode without an intermediate texture; one-element knockout groups use the same exact path because they have no sibling elements to knock out. Alpha-one, normal-blend non-knockout groups can retain multiple ordered fills and strokes because source-over is associative.

• Render platform-decoded images with deferred /SMask companion surfaces on the GPU by caching both textures and combining them in the soft-mask shader.

• Combine a single vector fill with one opaque grayscale image soft mask in a retained stencil pass. Rectangular vector-mask fills, including alpha or luminosity /BC and linearized /TR, stay entirely in retained geometry.

• Render Multiply and Screen exactly with fixed-function blending over the backend's opaque page surface.

• Warm the nonzero stencil-cover render state used by retained fills, moving its first-use Metal compilation into the existing idle pipeline phase.

• Render tiled stencil images exactly using scene-scoped, hand-built mipmaps. Mip levels participate in the shared texture budget and cache identity, and worker-reconstructed tiling cells retain their decoded images.

• Warm every tile shader with a one-pixel idle submission, shared per Impeller context and MSAA mode, then compile and submit each live page's retained scene at one-pixel scale. Both passes wait for useful pixels plus a 750 ms quiet window, so first-use driver, geometry, and upload work does not land on the first deep-zoom tile or contend with immediate navigation. Proactive warm-up defaults to desktop; mobile stays on-demand unless the host opts in, avoiding a large idle context allocation on memory-constrained devices.

• Benchmark tiling-pattern and radial-shading pipeline warm-up, scene warm-up, first GPU tiles, and Canvas parity three times on the designated macOS Metal CI lane. PRs receive a main comparison with details collapsed beneath the headline, and retain the normalized trace as a downloadable artifact.

• Accelerate exact vector tiling cells, axial gradients (including embedded outline text), and nested-circle radial gradients while retaining explicit Canvas fallback for unsafe variants.

0.1.6

• Align the experimental Flutter GPU backend with the 3.8.0 package suite. No public GPU API changes since 0.1.5.

0.1.5

• Align the experimental Flutter GPU backend with the 3.7.0 package suite.

0.1.4

• Isolate pipelines, geometry buffers, and image textures by Flutter GPU context so native windows never reuse resources owned by another view.
• Keep image-texture leases alive until in-flight command buffers complete, preventing stale blocks or unrelated fragments after structural edits.
• Report context identities and switches in the backend diagnostics.

0.1.3

• Rebuild the packaged shader bundle with Flutter 3.47's impellerc so the runtime accepts the new bundle format.

0.1.2

• Support both the Flutter 3.44 and 3.47 flutter_gpu APIs after shader loading became asynchronous and draw vertex counts moved to draw.

0.1.1

• Fall back to the Canvas renderer for deferred image soft masks that the GPU backend cannot yet reproduce exactly, preserving correct output on Impeller.

0.1.0

• Render final GPU tile textures directly and admit one per repaint, avoiding deferred slab-to-tile texture-copy bursts. Track synchronous issue time and command-buffer completion latency, failures, and in-flight submissions.
• Render ordinary non-rectangular PDF clip stacks exactly with retained stencil geometry, including nested even-odd/nonzero clips and save/restore; rectangular clips keep the cheaper scissor path.
• Report compiled clip paths and per-tile clip-mask rebuilds in backend stats.
• Add an opt-in Impeller flutter_gpu backend for retained-scene LoD tiles.
• Retain geometry and byte-budgeted image textures across tile renders and combine supported image soft masks directly in the tile shader.
• Pack command-heavy CAD geometry into backend-wide reusable 16 MiB device buffers under a strict byte budget, releasing leases only after GPU work completes and falling back instead of overshooting the ceiling.
• Fall back to the Canvas backend for unsupported pages and expose a web stub that preserves the same host API.
• Expose JSON-safe diagnostics for the latest tile route, accepted/rejected/ active sessions, runtime fallbacks, compile/replay timings, cache pressure, uploads/readbacks, and live GPU resource leases.