Rework aliased rasterization and path lowering to allow full hinting

ImageSharp.Drawing.All.sln

@@ -513,7 +513,6 @@ Global
 		SolutionGuid = {5F8B9D1F-CD8B-4CC5-8216-D531E25BD795}
 	EndGlobalSection
 	GlobalSection(SharedMSBuildProjectFiles) = preSolution
-		shared-infrastructure\src\SharedInfrastructure\SharedInfrastructure.projitems*{061582c2-658f-40ae-a978-7d74a4eb2c0a}*SharedItemsImports = 5
 		shared-infrastructure\src\SharedInfrastructure\SharedInfrastructure.projitems*{2e33181e-6e28-4662-a801-e2e7dc206029}*SharedItemsImports = 5
 		shared-infrastructure\src\SharedInfrastructure\SharedInfrastructure.projitems*{68a8cc40-6aed-4e96-b524-31b1158fdeea}*SharedItemsImports = 13
 	EndGlobalSection

src/ImageSharp.Drawing.WebGPU/GpuSceneDrawTag.cs

@@ -12,7 +12,7 @@ internal static class GpuSceneDrawTag
 {
     // These values are not a plain enum because each word also encodes the path-count, clip-count,
     // scene-word-count, and info-word-count increments consumed by the scan/reduction stages.
-    // Visible-fill tags carry five extra info words: coverage threshold plus raster interest. Must match drawtag.wgsl.
+    // Visible-fill tags carry five extra info words: coverage data plus raster interest. Must match drawtag.wgsl.
     public const uint Nop = 0U;
     public const uint FillColor = 0x188U;
     public const uint FillRecolor = 0x184U;

src/ImageSharp.Drawing.WebGPU/ImageSharp.Drawing.WebGPU.csproj

@@ -79,8 +79,6 @@
     <PackageReference Include="Silk.NET.Windowing.Sdl" Version="2.23.0" PrivateAssets="compile;build;analyzers;contentfiles;buildtransitive" />
   </ItemGroup>
 
-  <Import Project="..\..\shared-infrastructure\src\SharedInfrastructure\SharedInfrastructure.projitems" Label="Shared" />
-
   <ItemGroup>
     <ProjectReference Include="..\ImageSharp.Drawing\ImageSharp.Drawing.csproj" />
     <ProjectReference Include="..\ImageSharp.Drawing.WebGPU.ShaderGen\ImageSharp.Drawing.WebGPU.ShaderGen.csproj" OutputItemType="Analyzer" ReferenceOutputAssembly="false" />

src/ImageSharp.Drawing.WebGPU/Shaders/BboxClearComputeShader.cs

@@ -8,7 +8,7 @@ namespace SixLabors.ImageSharp.Drawing.Processing.Backends;
 
 /// <summary>
 /// GPU stage that initializes every per-path bounding box to an empty (inverted) box so
-/// that flatten can accumulate extents with atomic min/max. Wraps <c>bbox_clear.wgsl</c>.
+/// that path lowering can accumulate extents with atomic min/max. Wraps <c>bbox_clear.wgsl</c>.
 /// </summary>
 internal static unsafe class BboxClearComputeShader
 {

src/ImageSharp.Drawing.WebGPU/Shaders/FineAreaComputeShader.cs

@@ -103,7 +103,9 @@ public static unsafe bool TryCreateBindGroupLayout(
         //   6 gradients (sampled gradient ramp texture)
         //   7 image_atlas (sampled image atlas texture)
         //   8 backdrop_texture (sampled existing target contents)
-        WGPUBindGroupLayoutEntry* entries = stackalloc WGPUBindGroupLayoutEntry[9];
+        //   9 scene_data (read-only profile ranges and contour links in the packed scene stream)
+        //   10 path_tiles (read-only neighboring segment-slice indices for aliased row halos)
+        WGPUBindGroupLayoutEntry* entries = stackalloc WGPUBindGroupLayoutEntry[11];
         entries[0] = CreateUniformEntry(0, (nuint)sizeof(GpuSceneConfig));
         entries[1] = CreateStorageEntry(1, WGPUBufferBindingType.ReadOnlyStorage, 0);
         entries[2] = CreateStorageEntry(2, WGPUBufferBindingType.ReadOnlyStorage, 0);
@@ -113,10 +115,12 @@ public static unsafe bool TryCreateBindGroupLayout(
         entries[6] = CreateSampledTextureEntry(6);
         entries[7] = CreateSampledTextureEntry(7);
         entries[8] = CreateSampledTextureEntry(8);
+        entries[9] = CreateStorageEntry(9, WGPUBufferBindingType.ReadOnlyStorage, 0);
+        entries[10] = CreateStorageEntry(10, WGPUBufferBindingType.ReadOnlyStorage, 0);
 
         WGPUBindGroupLayoutDescriptor descriptor = new()
         {
-            entryCount = 9,
+            entryCount = 11,
             entries = entries
         };
 

src/ImageSharp.Drawing.WebGPU/Shaders/PathCountComputeShader.cs

@@ -7,7 +7,7 @@
 namespace SixLabors.ImageSharp.Drawing.Processing.Backends;
 
 /// <summary>
-/// GPU stage that counts, per tile, how many flattened line segments cross it: one thread per
+/// GPU stage that counts, per tile, how many final line segments cross it: one thread per
 /// line walks the line's tile crossings, bumps tile segment counts and backdrops, and emits one
 /// SegmentCount record per crossing for path-tiling. Wraps <c>path_count.wgsl</c>.
 /// </summary>
@@ -29,7 +29,7 @@ internal static unsafe class PathCountComputeShader
     /// ceil(<paramref name="lineCount"/> / 256). At runtime the stage is normally driven by
     /// the indirect count written by path-count-setup using the same divisor.
     /// </summary>
-    /// <param name="lineCount">The number of flattened lines.</param>
+    /// <param name="lineCount">The number of final lines.</param>
     /// <returns>The X dispatch dimension in workgroups.</returns>
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static uint GetDispatchX(uint lineCount)
@@ -52,7 +52,7 @@ public static bool TryCreateBindGroupLayout(
         // Bindings match path_count.wgsl:
         //   0 config uniform
         //   1 bump allocators (read-write; lines read, seg_counts bump-allocated)
-        //   2 lines (read-only LineSoup from flatten)
+        //   2 lines (read-only LineSoup from path lowering)
         //   3 paths (read-only Path records from path_row_alloc)
         //   4 rows (read-only PathRow spans finalized by tile_alloc)
         //   5 tile (read-write; backdrop and segment counts updated atomically)

src/ImageSharp.Drawing.WebGPU/Shaders/PathCountSetupComputeShader.cs

@@ -8,7 +8,7 @@ namespace SixLabors.ImageSharp.Drawing.Processing.Backends;
 
 /// <summary>
 /// GPU stage that computes the indirect dispatch size for the per-line stages
-/// (path-row-span and path-count) by dividing the flattened line count by the workgroup
+/// (path-row-span and path-count) by dividing the final line count by the workgroup
 /// size, or zero workgroups after an upstream allocation failure. Wraps
 /// <c>path_count_setup.wgsl</c>.
 /// </summary>

src/ImageSharp.Drawing.WebGPU/Shaders/FlattenComputeShader.cs → src/ImageSharp.Drawing.WebGPU/Shaders/PathLoweringComputeShader.cs

@@ -7,16 +7,16 @@
 namespace SixLabors.ImageSharp.Drawing.Processing.Backends;
 
 /// <summary>
-/// GPU stage that flattens encoded path segments into a device-space line soup, expands
-/// strokes via the CPU PolygonStroker port, and accumulates per-path bounding boxes for the
-/// downstream stages. Wraps <c>flatten.wgsl</c>.
+/// GPU stage that lowers encoded path segments into device-space lines and expands strokes
+/// with the CPU PolygonStroker port. The stage also accumulates per-path bounding boxes for
+/// downstream stages. It wraps <c>path_lowering.wgsl</c>.
 /// </summary>
-internal static unsafe class FlattenComputeShader
+internal static unsafe class PathLoweringComputeShader
 {
     /// <summary>
-    /// Gets the generated WGSL source bytes for the flatten stage.
+    /// Gets the generated WGSL source bytes for the path-lowering stage.
     /// </summary>
-    public static ReadOnlySpan<byte> ShaderCode => GeneratedWgslShaderSources.FlattenCode;
+    public static ReadOnlySpan<byte> ShaderCode => GeneratedWgslShaderSources.PathLoweringCode;
 
     /// <summary>
     /// Gets the WGSL entry point used by this shader.
@@ -35,7 +35,7 @@ public static uint GetDispatchX(uint pathTagCount)
         => (pathTagCount + 255U) / 256U;
 
     /// <summary>
-    /// Creates the bind-group layout required by the flatten stage.
+    /// Creates the bind-group layout required by the path-lowering stage.
     /// </summary>
     /// <param name="api">The WebGPU API facade.</param>
     /// <param name="device">The device that owns the staged-scene pipelines.</param>
@@ -48,7 +48,7 @@ public static bool TryCreateBindGroupLayout(
         out WGPUBindGroupLayoutImpl* layout,
         out string? error)
     {
-        // Bindings match flatten.wgsl:
+        // Bindings match path_lowering.wgsl:
         //   0 config uniform
         //   1 scene (read-only path tags, points, styles and transforms)
         //   2 tag_monoids (read-only pathtag prefix sums from the scan stages)
@@ -72,7 +72,7 @@ public static bool TryCreateBindGroupLayout(
         layout = api.DeviceCreateBindGroupLayout(device, in descriptor);
         if (layout is null)
         {
-            error = "Failed to create the flatten bind-group layout.";
+            error = "Failed to create the path-lowering bind-group layout.";
             return false;
         }
 

src/ImageSharp.Drawing.WebGPU/Shaders/PathRowAllocComputeShader.cs

@@ -56,17 +56,19 @@ public static bool TryCreateBindGroupLayout(
         //   3 bump allocators (read-write; path_rows bump-allocated, failure bit set on overflow)
         //   4 paths (read-write; Path record written per draw object)
         //   5 rows (read-write; PathRow records initialized to empty spans)
-        WGPUBindGroupLayoutEntry* entries = stackalloc WGPUBindGroupLayoutEntry[6];
+        //   6 path_bboxes (read-only path bounds before clipping to the drawing bounds)
+        WGPUBindGroupLayoutEntry* entries = stackalloc WGPUBindGroupLayoutEntry[7];
         entries[0] = SceneShaderBindingLayoutHelper.CreateUniformEntry(0, (nuint)sizeof(GpuSceneConfig));
         entries[1] = SceneShaderBindingLayoutHelper.CreateStorageEntry(1, WGPUBufferBindingType.ReadOnlyStorage);
         entries[2] = SceneShaderBindingLayoutHelper.CreateStorageEntry(2, WGPUBufferBindingType.ReadOnlyStorage);
         entries[3] = SceneShaderBindingLayoutHelper.CreateStorageEntry(3, WGPUBufferBindingType.Storage, (nuint)sizeof(GpuSceneBumpAllocators));
         entries[4] = SceneShaderBindingLayoutHelper.CreateStorageEntry(4, WGPUBufferBindingType.Storage);
         entries[5] = SceneShaderBindingLayoutHelper.CreateStorageEntry(5, WGPUBufferBindingType.Storage);
+        entries[6] = SceneShaderBindingLayoutHelper.CreateStorageEntry(6, WGPUBufferBindingType.ReadOnlyStorage);
 
         WGPUBindGroupLayoutDescriptor descriptor = new()
         {
-            entryCount = 6,
+            entryCount = 7,
             entries = entries
         };
 

src/ImageSharp.Drawing.WebGPU/Shaders/PathRowSpanComputeShader.cs

@@ -6,7 +6,7 @@
 namespace SixLabors.ImageSharp.Drawing.Processing.Backends;
 
 /// <summary>
-/// GPU stage that derives sparse per-row tile column spans from the flattened line stream:
+/// GPU stage that derives sparse per-row tile column spans from the final line stream:
 /// one thread per line replays the tile-crossing traversal, atomically growing each touched
 /// row's span and accumulating per-row winding backdrops for tile-alloc. Wraps
 /// <c>path_row_span.wgsl</c>.
@@ -40,7 +40,7 @@ public static bool TryCreateBindGroupLayout(
         // Bindings match path_row_span.wgsl:
         //   0 config uniform
         //   1 bump allocators (read-write because the buffer is atomic; this stage only reads the lines counter)
-        //   2 lines (read-only LineSoup from flatten)
+        //   2 lines (read-only LineSoup from path lowering)
         //   3 paths (read-only Path records from path_row_alloc)
         //   4 rows (read-write; spans, backdrops and flags accumulated atomically)
         WGPUBindGroupLayoutEntry* entries = stackalloc WGPUBindGroupLayoutEntry[5];

src/ImageSharp.Drawing.WebGPU/Shaders/PathTilingComputeShader.cs

@@ -40,7 +40,7 @@ public static bool TryCreateBindGroupLayout(
         // Bindings match path_tiling.wgsl:
         //   0 bump allocators (read-write because the buffer is atomic; this stage only reads the seg_counts total)
         //   1 seg_counts (read-only SegmentCount records from path_count)
-        //   2 lines (read-only LineSoup from flatten)
+        //   2 lines (read-only LineSoup from path lowering)
         //   3 paths (read-only Path records)
         //   4 rows (read-only sparse PathRow records)
         //   5 tiles (read-only; segment_count_or_ix holds the inverted segment base index from coarse)

src/ImageSharp.Drawing.WebGPU/Shaders/PathtagScanComputeShader.cs

@@ -7,7 +7,7 @@ namespace SixLabors.ImageSharp.Drawing.Processing.Backends;
 
 /// <summary>
 /// GPU stage that finishes the pathtag scan, writing an exclusive prefix TagMonoid per 4-byte
-/// tag word for flatten to locate path data in the scene stream. Two generated variants share
+/// tag word for path lowering to locate path data in the scene stream. Two generated variants share
 /// this wrapper: the small variant scans the pathtag-reduce partials directly, while the large
 /// variant reads per-workgroup prefixes precomputed by pathtag-scan1. Wraps
 /// <c>pathtag_scan.wgsl</c>.

src/ImageSharp.Drawing.WebGPU/Shaders/WgslSource/Shared/bbox.wgsl

@@ -3,30 +3,28 @@
 
 // Per-path device-space bounding box record.
 //
-// Cleared by bbox_clear.wgsl, atomically widened by flatten.wgsl (via its
+// Cleared by bbox_clear.wgsl, atomically widened by path_lowering.wgsl (via its
 // field-compatible AtomicPathBbox view of the same buffer), then consumed by
 // binning.wgsl, clip_reduce.wgsl, clip_leaf.wgsl, and draw_leaf.wgsl.
 // Ported from Vello's shader/shared/bbox.wgsl (linebender/vello) with
-// ImageSharp additions (coverage_threshold, interest).
+// ImageSharp additions (coverage_data, interest).
 
 // The annotated bounding box for a transformed path.
 // Coordinates are integer pixels (for the convenience of atomic update)
 // but will probably become fixed-point fractions for rectangles.
 //
 // `draw_flags` is propagated to the draw info stream and carries fill-rule, aliasing, and blend state.
-// Field layout must mirror AtomicPathBbox in flatten.wgsl.
+// Field layout must mirror AtomicPathBbox in path_lowering.wgsl.
 struct PathBbox {
     x0: i32, // min bounds; i32 max while empty (see bbox_clear)
     y0: i32,
     x1: i32, // max bounds; x0 > x1 denotes an empty box
     y1: i32,
     draw_flags: u32,
     trans_ix: u32, // index of the path's transform in the scene stream
-    // Per-fill coverage parameter, propagated from the style record to the fine pass.
-    // For aliased fills this is the quantization cutoff in [0,1]; for antialiased fills it
-    // carries the perceptual coverage boost for text (zero when disabled). The draw-flags
-    // aliased bit selects the interpretation; the two uses are mutually exclusive.
-    coverage_threshold: f32,
+    // Per-fill antialiased coverage adjustment propagated from the style record to the fine pass.
+    // Aliased fills use exact centre sampling and leave this value at zero.
+    coverage_data: f32,
     _padding: u32,
     interest: vec4<f32>, // raster interest rect (x0, y0, x1, y1) in pixels
 }

src/ImageSharp.Drawing.WebGPU/Shaders/WgslSource/Shared/bump.wgsl

@@ -7,7 +7,7 @@
 // compare the result against the capacity in Config; on overflow they set
 // their STAGE_* bit in `failed` and the C# side reads the block back, grows
 // the buffers, and retries the render. Imported by most compute stages
-// (binning, flatten, tile/row allocation, counting, coarse, chunk_reset,
+// (binning, path lowering, tile/row allocation, counting, coarse, chunk_reset,
 // prepare, and the indirect-dispatch setup shaders).
 //
 // Ported from Vello's shader/shared/bump.wgsl (linebender/vello), where the
@@ -16,7 +16,7 @@
 // Bitflags for each stage that can fail allocation.
 const STAGE_BINNING: u32 = 0x1u;
 const STAGE_TILE_ALLOC: u32 = 0x2u;
-const STAGE_FLATTEN: u32 = 0x4u;
+const STAGE_PATH_LOWERING: u32 = 0x4u;
 const STAGE_PATH_COUNT: u32 = 0x8u;
 const STAGE_COARSE: u32 = 0x10u;
 
@@ -33,7 +33,7 @@ struct BumpAllocators {
     seg_counts: atomic<u32>, // SegmentCount records
     segments: atomic<u32>, // Segment records
     blend_spill: atomic<u32>, // blend stack slots spilled past BLEND_STACK_SPLIT
-    lines: atomic<u32>, // LineSoup records from flattening
+    lines: atomic<u32>, // final LineSoup records from path lowering
 }
 
 // Workgroup counts for dispatchWorkgroupsIndirect, written by the setup

src/ImageSharp.Drawing.WebGPU/Shaders/WgslSource/Shared/config.wgsl

@@ -55,9 +55,16 @@ struct Config {
     segments_size: u32,
     blend_size: u32,
     ptcl_size: u32,
-    // Scene-wide coverage cutoff consumed by the aliased fine pass; per-fill
-    // thresholds (CmdFill.coverage_threshold) override it where set.
-    fine_coverage_threshold: f32,
+
+    // Word offset of the segment-profile tag table, or zero when the table is absent. Each encoded
+    // point pair has one slot, which lets a fill segment find its X and Y profile identifiers.
+    profile_slots_base: u32,
+    // Word offset of the profile record table, or zero when the table is absent. The table starts
+    // with the X and Y record counts. Each record then stores two axis limits and one contour link.
+    profile_records_base: u32,
+    profile_pad0: u32,
+    profile_pad1: u32,
+    profile_pad2: u32,
 }
 
 // Geometry of tiles and bins

src/ImageSharp.Drawing.WebGPU/Shaders/WgslSource/Shared/drawtag.wgsl

@@ -4,7 +4,7 @@
 // Draw tag stream decoding: the draw monoid, tag values, and draw-flag bits.
 //
 // Imported by binning.wgsl, clip_leaf.wgsl, coarse.wgsl, draw_leaf.wgsl,
-// draw_reduce.wgsl, fine.wgsl, flatten.wgsl, and path_row_alloc.wgsl.
+// draw_reduce.wgsl, fine.wgsl, path_lowering.wgsl, and path_row_alloc.wgsl.
 // Ported from Vello's shader/shared/drawtag.wgsl (linebender/vello) with
 // ImageSharp additions (recolor, elliptic and path gradient tags, the
 // aliased-coverage bit, and the blend mode/alpha draw-flag fields).
@@ -28,7 +28,7 @@ struct DrawMonoid {
 // Each draw object has a 32-bit draw tag, which is a bit-packed
 // version of the draw monoid: bit 0 = clip count, bits 2..4 = scene words,
 // bits 6..9 = info words (see map_draw_tag).
-// Visible-fill draw tags carry five extra info words: coverage threshold plus raster interest.
+// Visible-fill draw tags carry five extra info words: coverage data plus raster interest.
 const DRAWTAG_NOP = 0u;
 const DRAWTAG_FILL_COLOR = 0x188u;
 const DRAWTAG_FILL_RECOLOR = 0x184u;
@@ -45,9 +45,8 @@ const DRAWTAG_END_CLIP = 0x21u;
 // draw object stream but is used after the draw objects have been reduced on the GPU.
 // 0 represents a non-zero fill. 1 represents an even-odd fill.
 const DRAW_INFO_FLAGS_FILL_RULE_BIT = 1u;
-// Per-fill coverage rule. When set, the fill is rasterized aliased (coverage quantized against
-// config.fine_coverage_threshold) instead of using analytic area coverage. Carried in a free
-// high bit of the draw-flags word alongside the fill-rule bit.
+// Per-fill coverage rule. When set, the fill is rasterized using exact centre sampling instead
+// of analytic area coverage. Carried in a free high bit alongside the fill-rule bit.
 const DRAW_INFO_FLAGS_ALIASED_BIT = 0x40000000u;
 // Blend state packed into the draw-flags word by the C# encoder
 // (WebGPUSceneEncoder.PackStyleDrawFlags): bits 1..13 hold the (mix << 8) | compose
@@ -59,10 +58,8 @@ const DRAW_FLAGS_BLEND_ALPHA_MASK = 0x3fffc000u;
 
 // Flag bits carried in the high bits of the clip blend word, set by the C# encoder
 // (WebGPUSceneEncoder.AppendClipBeginData). DIFFERENCE marks an ImageSharp Difference
-// clip so fine inverts the mask; HARD marks a hard-edge (aliased) clip mask. Declared
-// here because every consumer (draw_leaf, coarse, fine) imports this module.
+// clip so fine inverts the mask. Declared here because every consumer imports this module.
 const CLIP_DIFFERENCE_MASK_BIT = 0x80000000u;
-const CLIP_HARD_MASK_BIT = 0x40000000u;
 // Marks an ISOLATED group (a layer). Isolated groups seed with transparent content and
 // composite back with blend_mix_compose, matching the CPU backend's clean layer targets.
 // Non-isolated groups (canvas clips) seed with a copy of the current tile content and pop

src/ImageSharp.Drawing.WebGPU/Shaders/WgslSource/Shared/pathtag.wgsl

@@ -5,9 +5,9 @@
 //
 // The scene encodes each path as a byte stream of tags (one per segment,
 // plus transform/style/path markers). The pathtag_reduce/scan stages compute
-// a prefix sum of TagMonoid over that stream so flatten.wgsl can locate each
+// a prefix sum of TagMonoid over that stream so path_lowering.wgsl can locate each
 // segment's point data, transform, and style in O(1). Imported by
-// flatten.wgsl, pathtag_reduce.wgsl, pathtag_reduce2.wgsl, pathtag_scan.wgsl,
+// path_lowering.wgsl, pathtag_reduce.wgsl, pathtag_reduce2.wgsl, pathtag_scan.wgsl,
 // and pathtag_scan1.wgsl.
 //
 // Ported from Vello's shader/shared/pathtag.wgsl (linebender/vello). Tag
@@ -43,7 +43,7 @@ const STYLE_SIZE_IN_WORDS: u32 = 10u;
 // Flag bits of the first word of a Style record, packed by the C# encoder.
 // STYLE marks a stroke (vs fill) and FILL marks the even-odd fill rule
 // (clear = non-zero); the cap and join fields drive the GPU stroker in
-// flatten.wgsl.
+// path_lowering.wgsl.
 const STYLE_FLAGS_STYLE: u32 = 0x80000000u;
 const STYLE_FLAGS_FILL: u32 = 0x40000000u;