PDF tiling patterns: colour-through-base fix, BBox clip, registry dedup

src/odr/internal/html/pdf_file.cpp

@@ -41,6 +41,18 @@ namespace {
 /// the extra digits add up across a page full of path data.
 double round2(const double v) { return std::round(v * 100.0) / 100.0; }
 
+/// Serialize a transform as an SVG `matrix(...)`. Only the translation (e, f)
+/// is rounded — it lives in page-box units where 1/100 px is plenty; the linear
+/// part (a..d) keeps full precision so small scale/skew factors aren't
+/// quantized to zero. Used for `transform`, `gradientTransform` and
+/// `patternTransform`.
+std::string svg_matrix(const util::math::Transform2D &m) {
+  std::ostringstream f;
+  f << "matrix(" << m.a << ',' << m.b << ',' << m.c << ',' << m.d << ','
+    << round2(m.e) << ',' << round2(m.f) << ')';
+  return std::move(f).str();
+}
+
 /// Convert a PDF device color to a CSS `rgb(...)` string. Non-device color
 /// spaces (Separation/ICCBased/… — stage 4.4) and the unknown space fall back
 /// to black, the PDF initial color.
@@ -222,25 +234,60 @@ std::string svg_image_fragment(const pdf::ImageElement &image,
   if (!clip_id.empty()) {
     f << "<g clip-path=\"url(#" << clip_id << ")\">";
   }
-  f << R"(<image width="1" height="1" preserveAspectRatio="none" transform="matrix()"
-    << m.a << ',' << m.b << ',' << m.c << ',' << m.d << ',' << round2(m.e)
-    << ',' << round2(m.f) << ")\"";
+  f << R"(<image width="1" height="1" preserveAspectRatio="none" transform=")"
+    << svg_matrix(m) << '"';
   f << " href=\"" << file_to_url(image.data, image.mime) << "\"/>";
   if (!clip_id.empty()) {
     f << "</g>";
   }
   return std::move(f).str();
 }
 
+/// Shared bookkeeping for the per-page `<defs>` registries below (clips,
+/// gradients, tiling patterns): a signature->id cache that deduplicates
+/// repeated definitions, a per-page monotonic id counter, and the accumulated
+/// `<defs>` markup (emitted once into the page's hidden `<svg>`). Ids are
+/// namespaced per page as `<prefix><page>_<n>`.
+class DefsRegistry {
+public:
+  explicit DefsRegistry(const std::uint32_t page) : m_page{page} {}
+
+  [[nodiscard]] std::string defs() const { return m_defs.str(); }
+
+protected:
+  /// The id for `signature`, minting `<prefix><page>_<n>` the first time it is
+  /// seen. `inserted` is true only on that first sight — when the caller still
+  /// needs to emit the definition into `m_defs`.
+  struct Entry {
+    std::string id;
+    bool inserted;
+  };
+  Entry intern(const std::string &signature, const char *prefix) {
+    const auto [it, inserted] = m_id_by_signature.try_emplace(signature);
+    if (inserted) {
+      it->second = std::string(prefix) + std::to_string(m_page) + "_" +
+                   std::to_string(++m_count);
+    }
+    return {it->second, inserted};
+  }
+
+  std::ostringstream m_defs;
+
+private:
+  std::uint32_t m_page;
+  std::uint32_t m_count{0};
+  std::unordered_map<std::string, std::string> m_id_by_signature;
+};
+
 /// Registers a page's clip regions as nested `<clipPath>` defs, deduplicating
 /// shared prefixes. PDF's current clip is the *intersection* of an ordered list
 /// of regions; SVG expresses intersection by chaining `clip-path` from one
 /// `<clipPath>` to the next, so region i's clipPath references region i-1's and
 /// the painted element references the last. Ids are namespaced per page
-/// (`c<page>_<n>`); `defs()` is emitted once in a hidden `<svg>` for the page.
-class ClipRegistry {
+/// (`c<page>_<n>`).
+class ClipRegistry : public DefsRegistry {
 public:
-  explicit ClipRegistry(std::uint32_t page) : m_page{page} {}
+  using DefsRegistry::DefsRegistry;
 
   /// The clipPath id to reference on a path painted under `clip`, registering
   /// any not-yet-seen regions. Empty when `clip` is empty (unclipped).
@@ -253,11 +300,9 @@ class ClipRegistry {
       signature += region.even_odd ? 'E' : 'N';
       signature += d;
       signature += ';';
-      const auto [it, inserted] = m_id_by_signature.try_emplace(signature);
+      const auto [id, inserted] = intern(signature, "c");
       if (inserted) {
-        it->second =
-            "c" + std::to_string(m_page) + "_" + std::to_string(++m_count);
-        m_defs << "<clipPath id=\"" << it->second << '"';
+        m_defs << "<clipPath id=\"" << id << '"';
         if (!parent.empty()) {
           m_defs << " clip-path=\"url(#" << parent << ")\"";
         }
@@ -267,18 +312,10 @@ class ClipRegistry {
         }
         m_defs << "/></clipPath>";
       }
-      parent = it->second;
+      parent = id;
     }
     return parent;
   }
-
-  [[nodiscard]] std::string defs() const { return m_defs.str(); }
-
-private:
-  std::uint32_t m_page;
-  std::uint32_t m_count{0};
-  std::unordered_map<std::string, std::string> m_id_by_signature;
-  std::ostringstream m_defs;
 };
 
 /// Registers a page's shadings (axial/radial) as `<linearGradient>`/
@@ -293,9 +330,9 @@ class ClipRegistry {
 /// shading is over-painted beyond its interval instead of being masked to it;
 /// `Shading::background` and `Shading::bbox` are likewise not yet honoured.
 /// Honouring them needs the fill clipped to the gradient band/annulus.
-class GradientRegistry {
+class GradientRegistry : public DefsRegistry {
 public:
-  explicit GradientRegistry(const std::uint32_t page) : m_page{page} {}
+  using DefsRegistry::DefsRegistry;
 
   /// The gradient id to reference via `fill="url(#id)"` for `shading` placed by
   /// `m` (shading space -> page box). Empty for an unrepresentable shading.
@@ -308,12 +345,10 @@ class GradientRegistry {
     sig << shading.type << ':' << static_cast<const void *>(&shading) << ':'
         << m.a << ',' << m.b << ',' << m.c << ',' << m.d << ',' << m.e << ','
         << m.f;
-    const auto [it, inserted] = m_id_by_signature.try_emplace(sig.str());
+    const auto [id, inserted] = intern(sig.str(), "g");
     if (!inserted) {
-      return it->second;
+      return id;
     }
-    it->second = "g" + std::to_string(m_page) + "_" + std::to_string(++m_count);
-    const std::string &id = it->second;
 
     const std::array<double, 6> &c = shading.coords;
     if (shading.type == 2) {
@@ -327,27 +362,15 @@ class GradientRegistry {
              << "\" cy=\"" << c[4] << "\" r=\"" << c[5] << "\" fx=\"" << c[0]
              << "\" fy=\"" << c[1] << "\" fr=\"" << c[2] << '"';
     }
-    // Only the translation (e, f) is rounded — it lives in page-box units where
-    // 1/100 px is plenty; the linear part (a..d) keeps full precision so small
-    // scale/skew factors aren't quantized to zero.
-    m_defs << " gradientUnits=\"userSpaceOnUse\" gradientTransform=\"matrix("
-           << m.a << ',' << m.b << ',' << m.c << ',' << m.d << ','
-           << round2(m.e) << ',' << round2(m.f) << ")\">";
+    m_defs << " gradientUnits=\"userSpaceOnUse\" gradientTransform=\""
+           << svg_matrix(m) << "\">";
     for (const pdf::GradientStop &stop : shading.stops) {
       m_defs << "<stop offset=\"" << round2(stop.offset) << "\" stop-color=\""
              << rgb_to_css(stop.rgb) << "\"/>";
     }
     m_defs << (shading.type == 2 ? "</linearGradient>" : "</radialGradient>");
     return id;
   }
-
-  [[nodiscard]] std::string defs() const { return m_defs.str(); }
-
-private:
-  std::uint32_t m_page{};
-  std::uint32_t m_count{0};
-  std::unordered_map<std::string, std::string> m_id_by_signature;
-  std::ostringstream m_defs;
 };
 
 /// Serialize an `sh` shading flood to an SVG `<rect>` covering the page box,
@@ -376,12 +399,14 @@ std::string svg_shading_fragment(const std::string &gradient_id,
 /// every `/XStep`/`/YStep`, and `patternTransform` (pattern space -> page box)
 /// places the lattice. An uncoloured pattern (`/PaintType 2`) ignores its
 /// content's own colours and paints in the path's fill colour, so the cache key
-/// folds that colour in. Ids are namespaced per page (`pat<page>_<n>`). Only
-/// paths and images inside the tile are rendered (nested text/shadings/patterns
-/// are skipped — rare). Returns "" for an unrepresentable pattern.
-class PatternRegistry {
+/// folds that colour in. Each cell is clipped to its `/BBox` so marks outside
+/// the cell (or in the gap when a step exceeds the BBox) don't leak into the
+/// tile. Ids are namespaced per page (`pat<page>_<n>`). Only paths and images
+/// inside the tile are rendered (nested text/shadings/patterns are skipped —
+/// rare). Returns "" for an unrepresentable pattern.
+class PatternRegistry : public DefsRegistry {
 public:
-  explicit PatternRegistry(const std::uint32_t page) : m_page{page} {}
+  using DefsRegistry::DefsRegistry;
 
   std::string register_pattern(const pdf::Pattern &pattern,
                                const util::math::Transform2D &m,
@@ -398,12 +423,10 @@ class PatternRegistry {
     if (uncoloured) {
       sig << ':' << device_color_to_css(fill_color);
     }
-    const auto [it, inserted] = m_id_by_signature.try_emplace(sig.str());
+    const auto [id, inserted] = intern(sig.str(), "pat");
     if (!inserted) {
-      return it->second;
+      return id;
     }
-    it->second =
-        "pat" + std::to_string(m_page) + "_" + std::to_string(++m_count);
 
     // Tile content is laid out in pattern space (identity page transform); the
     // y-flip and placement live in `patternTransform`.
@@ -423,24 +446,29 @@ class PatternRegistry {
       }
     }
 
-    m_defs << "<pattern id=\"" << it->second
+    m_defs << "<pattern id=\"" << id
            << "\" patternUnits=\"userSpaceOnUse\" x=\""
            << round2(pattern.bbox[0]) << "\" y=\"" << round2(pattern.bbox[1])
            << "\" width=\"" << round2(std::abs(pattern.x_step))
            << "\" height=\"" << round2(std::abs(pattern.y_step))
-           << "\" patternTransform=\"matrix(" << m.a << ',' << m.b << ',' << m.c
-           << ',' << m.d << ',' << round2(m.e) << ',' << round2(m.f) << ")\">"
-           << std::move(tile).str() << "</pattern>";
-    return it->second;
+           << "\" patternTransform=\"" << svg_matrix(m) << "\">";
+    // Clip each cell to its `/BBox` (ISO 32000-1 8.7.3.1). An overlapping
+    // lattice (a step smaller than the BBox) can't be expressed as a single SVG
+    // `<pattern>` and is not reproduced.
+    const double bbox_w = pattern.bbox[2] - pattern.bbox[0];
+    const double bbox_h = pattern.bbox[3] - pattern.bbox[1];
+    if (bbox_w > 0 && bbox_h > 0) {
+      m_defs << "<clipPath id=\"" << id << "c\"><rect x=\""
+             << round2(pattern.bbox[0]) << "\" y=\"" << round2(pattern.bbox[1])
+             << "\" width=\"" << round2(bbox_w) << "\" height=\""
+             << round2(bbox_h) << "\"/></clipPath><g clip-path=\"url(#" << id
+             << "c)\">" << std::move(tile).str() << "</g>";
+    } else {
+      m_defs << std::move(tile).str();
+    }
+    m_defs << "</pattern>";
+    return id;
   }
-
-  [[nodiscard]] std::string defs() const { return m_defs.str(); }
-
-private:
-  std::uint32_t m_page{};
-  std::uint32_t m_count{0};
-  std::unordered_map<std::string, std::string> m_id_by_signature;
-  std::ostringstream m_defs;
 };
 
 /// Deduplicates CSS declarations into atomic, single-property classes. PDF text

src/odr/internal/pdf/AGENTS.md

@@ -590,9 +590,12 @@ stage exists to avoid.
   every `/XStep`/`/YStep`, with `patternTransform` (the pattern `/Matrix`)
   placing the lattice; a `/PatternType 1` fill references it as `fill="url(#…)"`.
   Coloured (`/PaintType 1`) cells carry their own colours; uncoloured
-  (`/PaintType 2`) cells are painted in the current fill colour. Only paths and
-  images inside a tile are rendered (nested text/shadings/patterns are skipped —
-  rare).
+  (`/PaintType 2`) cells are painted in the current fill colour (resolved
+  through the Pattern colour space's base, so `[/Pattern /DeviceRGB]` keeps its
+  tint). Each cell is clipped to its `/BBox`; an overlapping lattice (a step
+  smaller than the BBox) can't be expressed as one SVG `<pattern>` and is not
+  reproduced. Only paths and images inside a tile are rendered (nested
+  text/shadings/patterns are skipped — rare).
 - **SVG residue** — where no 1:1 primitive exists; all at generation time, never
   rasterization: mesh/function shadings (types 1, 4–7) → tessellate into small
   flat polygons (pdf.js's approach); color spaces

src/odr/internal/pdf/pdf_color.cpp

@@ -148,6 +148,13 @@ ColorSpaceDef::to_rgb(const std::vector<double> &c) const {
     return alternate->to_rgb(tint->eval(c));
   }
   case ColorSpaceKind::pattern:
+    // An uncoloured pattern (`/PaintType 2`) carries its colour in the Pattern
+    // space's underlying base (e.g. `[/Pattern /DeviceRGB]`); convert through
+    // it. Without a base there is no device colour to convert.
+    if (base != nullptr) {
+      return base->to_rgb(c);
+    }
+    return {0, 0, 0};
   case ColorSpaceKind::unknown:
     return {0, 0, 0};
   }

src/odr/internal/pdf/pdf_document_parser.cpp

@@ -860,7 +860,7 @@ Pattern *parse_pattern(State &state, const ObjectReference &reference,
     if (object.has_stream) {
       pattern->content = parser.read_decoded_stream(object);
     }
-    if (dictionary.has_key("Resources")) {
+    if (dictionary.has_value("Resources")) {
       pattern->resources = parse_resources(state, dictionary["Resources"]);
     }
   }

test/src/internal/pdf/pdf_page_extractor.cpp

@@ -961,6 +961,37 @@ TEST(PdfPageExtractor, scn_uncoloured_tiling_pattern_carries_colour) {
   EXPECT_EQ(p.fill_pattern->paint_type, 2);
 }
 
+// An uncoloured pattern selected through a *named* Pattern colour space with an
+// underlying base (`[/Pattern /DeviceRGB]`) resolves its leading components
+// through that base — `1 0 0` is red, not black (the base would be ignored if
+// the Pattern space's `to_rgb` dropped it).
+TEST(PdfPageExtractor, scn_uncoloured_tiling_pattern_colour_through_base) {
+  Pattern pattern;
+  pattern.type = Pattern::Type::tiling;
+  pattern.paint_type = 2;
+
+  std::vector<Object> array{Object(Name{"Pattern"}), Object(Name{"DeviceRGB"})};
+  ColorSpaceContext ctx;
+  ctx.resolve = [](const Object &o) { return o; };
+  ctx.load_stream = [](const Object &) { return std::string{}; };
+  ctx.named = nullptr;
+
+  Resources res;
+  res.color_space["CS1"] =
+      parse_color_space(Object(Array(std::move(array))), ctx);
+  res.pattern["P2"] = &pattern;
+
+  const auto page =
+      extract_page("/CS1 cs 1 0 0 /P2 scn 0 0 10 10 re f", res, Logger::null());
+  ASSERT_EQ(page.size(), 1);
+  const PathElement &p = std::get<PathElement>(page[0]);
+  ASSERT_NE(p.fill_pattern, nullptr);
+  EXPECT_EQ(p.fill_color.space, ColorSpace::device_rgb);
+  EXPECT_DOUBLE_EQ(p.fill_color.rgb[0], 1.0);
+  EXPECT_DOUBLE_EQ(p.fill_color.rgb[1], 0.0);
+  EXPECT_DOUBLE_EQ(p.fill_color.rgb[2], 0.0);
+}
+
 // The `sh` operator floods the current clip with a named `/Shading`, emitting a
 // `ShadingElement` placed by the CTM.
 TEST(PdfPageExtractor, sh_emits_shading_element) {