Repository navigation
Back-port seals inside layers, gradient repetition and WebGPU scratch bounds to 3.1.x - #429
Merged
JimBobSquarePants merged 14 commits intoOct 28, 2026
Merged
Conversation
compose_source evaluates the blend overlap once per call instead of inside five switch cases. The fine pipeline compiles in 0.5 s instead of 1.7 s with identical arithmetic per mode. The scheduling pipelines compile in parallel when the device is probed. The fine pipeline compiles once per target format, started when the target is created, instead of four fixed variants started at the first render. Stroke scratch estimates are bounded from the outline geometry (offset edges, caps, joins) instead of the bounding-box diagonal. The PTCL seed is a per-crossing bound plus one chunk per target tile. (cherry picked from commit 8dd5241)
CMD_FILL and CMD_SOLID reference their raster interest rectangle by its info-stream offset instead of copying four words into every tile. CMD_COLOR references the packed colour in the scene stream and the draw flags in the info stream. Fill goes from 9 words to 6, solid from 5 to 2, colour from 4 to 3. The PTCL seed bound drops from 14 to 11 words per crossing. The fine shader stores each paint command's per-pixel colour and coverage and composes them at one site after the command switch. One compose site keeps the blend-mode switch from being inlined into every paint case, which cuts the cold fine compile from 1.8 s to 1.1 s. (cherry picked from commit 1536258)
The encoder accumulates a PTCL word bound per draw: one CMD_FILL (6 words) plus the paint command coarse writes for that draw tag, times the draw's tile-crossing bound. The bound flows through partitions, checkpoints, ranges, and the scene, and replaces the fixed 11 words per crossing in the scratch seed. A solid-colour scene now seeds 9 words per crossing. (cherry picked from commit 620eae3)
Gradient brushes with Repeat and Reflect now repeat before the start point. Repeat subtracts the floor of the position, and Reflect uses the magnitude of the position. The WebGPU shaders use the same rules. Flush, RenderScene, CreateScene and CopyPixelsFrom inside an open layer keep the layer whole. A retained scene renders inside a layer on both backends. The WebGPU encoder reads it in place through a scene segment table. CopyPixelsFrom inside a layer reads the source rectangle with ReadRegion and draws the image into the layer with Src. The WebGPU encoder computes upper bounds for the scratch buffers, with row bounds for tile-row chunks and blend depth regions for the blend spill. fine adds the coverage of each pixel in fixed point, so the output does not depend on the GPU scheduling order, and it equals the CPU output. An offscreen WebGPU render checks its scratch status before it returns, and it renders again with larger buffers after an overflow. The deferred corrective re-render is removed, so a scene releases its resources when it is disposed. After an overflow, fine writes the backdrop, so a deferred presentation frame keeps the previous pixels. Each backend finds the retained scenes with ordered work through DrawingCommandBatch.FindOrderedScenes. The public API does not change. Tests cover each seal inside a layer on both backends, the gradient repetition, the scratch bounds, and a 100k-line offscreen render that overflows and chunks. The WebGPU reference images that held the old output are updated: the new WebGPU output equals the CPU output.
A plain WebGPU render now submits its work and returns. It reads no scratch status when every capacity is an upper bound of the GPU use. Ordered ranges, a binning bound above the binding limit, and chunked scenes without row bounds still read the status. At 100k lines the flush call takes about 45 ms against 97 ms on the baseline, and the time per frame is the same. The bounds are now true upper bounds: - The row bounds keep one entry for each tile row, so a chunk shorter than one bin gets the bound of its own rows. - The stroke style flags hold the CPU arc count, and the shader limits each round arc to it. - A long inner join adds its own crossing bound, and the inner miter uses the shader limit. - The PTCL bound comes from the chunk allocation in coarse. - Each clip estimate grows by one pixel on each side. - The outline radius of a Miter join is sqrt(m^2 + 1) half widths, which is the clipped bevel point of both strokers. - path_count makes the y clip exact, so each line adds at most its tile count and the row index stays in the bbox. A test renders 100k lines in chunks shorter than one bin and compares them with an unchunked render. The layer tests compare their pixels with the CPU backend.
…tation The documentation said that the pen expands in local geometry space and that the outline is transformed after. Both backends stroke the transformed path in device space, with the pen width multiplied by the square root of the absolute determinant of the transform. The documentation now describes this behavior.
The documentation gives only the observable contract of the property: the transform that applies to vector output before rasterization.
The CopyPixelsFromSelf case composites green 128 at 50% over white. The exact green value is 191.5, which is a rounding tie. The WebGPU backend on macOS gives 191, and the reference image has 192. The difference is 0.0184% in 96 pixels. The reference comparison of this case now uses 0.0185%, which is the measured difference plus one unit. The other cases keep 0.0003%.
Aliased rasterization on the CPU and WebGPU backends: - A pixel centre exactly on a vertical edge belongs to the shape left of the edge. A centre exactly on a horizontal edge belongs to the shape above the edge. A centre on a sloped edge stays covered. A 1 pixel pen line on whole-pixel coordinates now sets one row or one column. - Correct the sign of the aliased start cover for edges that are clipped left of the band. - Find thin horizontal features between two row centres also in a row that has no row-centre crossing. - Do not light the midpoint pixel of an interval that continues outside the left or right edge of the band. - CPU: keep a horizontal edge that lies on the bottom of the interest, and treat a vertical edge exactly on the left of the interest as a line, as for contained geometry. - WebGPU: give the row walk the crossing of a vertical edge on a tile edge, and keep a vertical line on the left edge of the path bbox as a halo-only segment. Keep the real end of a vertical segment that is stretched along the tile's left edge, so the row walks of the tile and of the tile on its left do not read crossings below that end. Compare y_edge in rounded 24.8. - Inflate stroke bounds by Pen.GetOutlineRadius on both backends. - Remove the clipped-left path flag from the WebGPU pipeline.
A thin horizontal feature that touches or crosses the edge of a 16-row band on the CPU, or of a tile row on WebGPU, was not set. The column pass of each band started from a zero winding and saw only the crossings inside the band. - Each column walk now starts from the winding just above the centre of the first row, and walks down and up from there. An edge exactly through the column centre on that row counts only when the start point is right of it. - The walk reads the column crossings within half a pixel above and below the band or tile. The CPU replays the lines of the neighbouring bands. An aliased WebGPU fill command holds the path tiles above and below, and a horizontal line exactly on a tile row boundary is kept as a halo-only segment. - Each band or tile sets only the midpoints in its own rows. - Both backends keep 16 crossings for each column and skip a column that has more. - Both backends sort coincident row crossings with the same key: vertical edge, then downward direction, then profile. Before, the direction bit had opposite meanings on the two backends, so two crossings at one position could form one interval on the CPU and two on WebGPU.
The row pass and the column pass test each pixel centre against the same edge, from two separate interpolations in 24.8 fixed point. Both used truncating division. One pass could put a centre outside the edge while the other put the edge exactly on the centre. Then neither pass set the pixel, and a thin sloped sliver got a one pixel hole. The CPU and WebGPU both had this fault. - Both backends now round each crossing down, and move an inexact value that lands on a pixel centre up by one unit. Only an exact hit then equals a centre, so the two passes always put a centre on the same side of an edge. - Two pixels of the aliased ellipse reference images change. Their centres are 0.003 px outside the flattened ellipse.
WebGPU renders a large scene in tile-row chunks when a tile buffer is larger than the binding limit. The aliased column walk reads the path tiles above and below a tile to find thin features within half a pixel of a tile row boundary. In a chunk those tiles did not exist at the first and last chunk rows, so a thin feature across a chunk edge lost its pixel. - path_row_alloc keeps one tile row on each side of the chunk. Coarse writes commands for the chunk rows only, and it allocates the segments of a tile outside the chunk for the fill tile next to it. - Fine skips a neighbouring tile that has no segment allocation, for example a tile where coarse skipped the draw inside a clip. - The path rows, tiles and segments of a chunk are sized for the chunk rows and the two extra rows, and they scale back to the scene budget with the same rows, so the budget does not grow from flush to flush.
An aliased fill sets the midpoint pixel of a thin feature between two pixel centres. When such a feature crossed the edge of the image, a clip or a region, both backends dropped the geometry outside the edge, so the pixel inside was lost. The output inside the edge must equal the output with no edge. - CPU: an aliased item rasterizes one more pixel on each side of its interest. The first row band can then be band -1. The band info carries the interest, and the coverage row handler clips every span to it. This also stops an antialiased stroke on a region canvas from writing outside the region, so one reference image changes. - WebGPU: binning keeps one more pixel around the interest and the clips of an aliased fill. Where a path touches a target edge, path_count and path_tiling also walk the tile row or column just outside it, and store those pieces in the edge tile as halo-only segments. Fine reads their real crossing positions. The encoder bounds count the extra pixel and the edge row. - path_row_alloc gives a draw with no tile rows or columns an empty bbox, and path_count writes a record that path_tiling skips for every reserved crossing slot that no tile keeps.
FillPath_AliasedThinBarsAcrossChunkEdges_MatchUnchunkedRender uses the 100k-line scene of FillManyLines_ChunksSmallerThanOneBin_MatchUnchunkedRender, which draws the lines antialiased. It drew them aliased, and on the macOS CI runners that render did not finish within the 5 second readback timeout. The lines are now antialiased, and only the bars across the chunk edges are aliased. The test still fails without the chunk halo fix.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Prerequisites
Description
This PR back-ports the changes of #428 to
release/3.1.x. It corrects three faults inDrawingCanvas. It also changes how the WebGPU backend sets the size of its scratch buffers, and it corrects faults in aliased fills on the two backends.Calls inside a layer
A
Flush,RenderScene,CreateSceneorCopyPixelsFromcall betweenSaveLayerandRestorebroke the layer:Applycall lost its earlier content.Now these calls keep the layer whole on the two backends:
Flushdoes nothing while a layer is open.RenderScenedraws the scene into the open layer.CreateScenecloses the open layer at the end of the scene. Then the canvas opens the layer again.CopyPixelsFromreads the source rectangle and draws it into the open layer.Scenes in
CreateSceneCreateScenelost the scenes thatRenderScenerecorded before it. Now the new scene contains these scenes, in their recorded order.Gradient repetition
With
RepeatorReflect, a position before the start point got the color of the first stop. Now the two modes repeat the gradient in the two directions. The CPU backend and the WebGPU shaders use the same rule.WebGPU scratch buffers
The encoder calculates an upper bound for each scratch buffer. If all the bounds hold, a render submits its work and returns. It does not read back the scratch status.
A scene with ordered work, for example
Apply, still reads the status. If an offscreen render overflows a buffer, the backend renders the scene again with larger buffers.The fine stage adds pixel coverage as fixed-point values. Thus, the order of the GPU work does not change the output.
Aliased fills
Aliased fills now use one rule for a pixel centre exactly on an edge. The CPU backend and the WebGPU shaders use the same rule for all shapes, text included:
Thus, a 1 pixel pen line on whole-pixel coordinates sets one row or one column.
These faults are also corrected on the two backends:
Aliased output changes where a pixel centre is exactly on a vertical or horizontal edge. This includes aliased text.
WebGPU commits from
mainThis PR also contains three earlier WebGPU commits from
main:Other changes
LineJoin.Miterstroke now include the clipped bevel point.DrawingOptions.Transformdocumentation no longer says that strokes expand in local space.The public API does not change. Some WebGPU reference images change, and the new images match the CPU reference images. The two aliased ellipse reference images change at 2 pixels. The centres of these pixels are 0.003 px outside the flattened ellipse. The region canvas reference image changes at 8 pixels outside the region.
Tests on Windows: net8.0 and net10.0 have 2843 passed, 0 failed and 9 skipped each.