renderer: add partial rendering support

Partial Rendering refers to a rendering technique where
only a portion of the scene or screen is updated, rather
than redrawing the entire output. It is commonly used as
a performance optimization strategy, focusing on redrawing
only the regions that have changed, often called dirty regions.

This introduces RenderDirtyRegion, which assists
in collecting a compact dirty region from render tasks.
Each backend can utilize this class to support efficient partial rendering.
This is implemented using a Line Sweep and Subdivision Merging O(NlogN).

The basic per-frame workflow is as follows:

1. RenderDirtyRegion::prepare() //Call this in Renderer::preRender().
2. RenderDirtyRegion::add() //Add all dirty paints for the frame before rendering.
3. RenderDirtyRegion::commit() //Generate the partial rendering region list before rendering.
4. RenderDirtyRegion::get() //Retrieve the current dirty region list and use it when drawing paints.
5. RenderDirtyRegion::clear() //Reset the state.

also removed the skip() from paint since engines need to capture the
drawable object in any case.

issue: https://github.com/thorvg/thorvg/issues/1747

src/renderer/tvgPaint.h

@@ -144,6 +144,11 @@ namespace tvg
             return (renderFlag & flag) ? true : false;
         }
 
+        void damage()
+        {
+            if (renderer) renderer->damage(bounds(renderer));
+        }
+
         void mark(RenderUpdateFlag flag)
         {
             renderFlag |= flag;

src/renderer/tvgRender.cpp

@@ -20,6 +20,7 @@
  * SOFTWARE.
  */
 
+#include <algorithm>
 #include "tvgMath.h"
 #include "tvgRender.h"
 
@@ -54,6 +55,18 @@ bool RenderMethod::viewport(const RenderRegion& vp)
 }
 
 
+void RenderMethod::damage(const RenderRegion& region)
+{
+    dirtyRegion.add(region);
+}
+
+
+bool RenderMethod::partial(bool disable)
+{
+    std::swap(dirtyRegion.disabled, disable);
+    return disable;
+}
+
 /************************************************************************/
 /* RenderPath Class Implementation                                      */
 /************************************************************************/
@@ -130,6 +143,130 @@ void RenderRegion::intersect(const RenderRegion& rhs)
     if (max.y < min.y) max.y = min.y;
 }
 
+
+void RenderDirtyRegion::subdivide(Array<RenderRegion>& targets, uint32_t idx, RenderRegion& lhs, RenderRegion& rhs)
+{
+    RenderRegion temp[5];
+    int cnt = 0;
+    temp[cnt++] = RenderRegion::intersect(lhs, rhs);
+    auto max = std::min(lhs.max.x, rhs.max.x);
+
+    auto subtract = [&](RenderRegion& lhs, RenderRegion& rhs) {
+        //top
+        if (rhs.min.y < lhs.min.y) {
+            temp[cnt++] = {{rhs.min.x, rhs.min.y}, {rhs.max.x, lhs.min.y}};
+            rhs.min.y = lhs.min.y;
+        }
+        //bottom
+        if (rhs.max.y > lhs.max.y) {
+            temp[cnt++] = {{rhs.min.x, lhs.max.y}, {rhs.max.x, rhs.max.y}};
+            rhs.max.y = lhs.max.y;
+        }
+        //left
+        if (rhs.min.x < lhs.min.x) {
+            temp[cnt++] = {{rhs.min.x, rhs.min.y}, {lhs.min.x, rhs.max.y}};
+            rhs.min.x = lhs.min.x;
+        }
+        //right
+        if (rhs.max.x > lhs.max.x) {
+            temp[cnt++] = {{lhs.max.x, rhs.min.y}, {rhs.max.x, rhs.max.y}};
+            //rhs.max.x = lhs.max.x;
+        }
+    };
+
+    subtract(temp[0], lhs);
+    subtract(temp[0], rhs);
+
+    //TODO: remove this
+    if (targets.reserved <  targets.count + cnt - 1) {
+        TVGERR("RENDERER", "reserved: %d, required: %d (+%d)\n", targets.reserved, targets.count + cnt - 1, cnt - 1);
+        abort();
+    }
+
+    /* Note: We considered using a list to avoid memory shifting,
+       but ultimately, the array outperformed the list due to better cache locality. */
+
+    //shift data
+    auto dst = &targets[idx + cnt];
+    memmove(dst, &targets[idx + 1], sizeof(RenderRegion) * (targets.count - idx - 1));
+    memcpy(&targets[idx], temp, sizeof(RenderRegion) * cnt);
+    targets.count += (cnt - 1);
+
+    //sorting by x coord again, only for the updated region
+    while (dst < targets.end() && dst->min.x < max) ++dst;
+    stable_sort(&targets[idx], dst, [](const RenderRegion& a, const RenderRegion& b) -> bool {
+        return a.min.x < b.min.x;
+    });
+}
+
+
+void RenderDirtyRegion::commit()
+{
+    if (disabled) return;
+
+    auto& targets = list[current];
+    if (targets.empty()) return;
+
+    current = !current;  //swapping buffers
+
+    auto& output = list[current];
+
+    //sorting by x coord. guarantee the stable performance: O(NlogN)
+    stable_sort(targets.begin(), targets.end(), [](const RenderRegion& a, const RenderRegion& b) -> bool {
+        return a.min.x < b.min.x;
+    });
+
+    //Optimized using sweep-line algorithm: O(NlogN)
+    for (uint32_t i = 0; i < targets.count; ++i) {
+        auto& lhs = targets[i];
+        if (lhs.invalid()) continue;
+        auto merged = false;
+
+        for (uint32_t j = i + 1; j < targets.count; ++j) {
+            auto& rhs = targets[j];
+            if (rhs.invalid()) continue;
+            if (lhs.max.x < rhs.min.x) break;   //line sweeping
+
+            //fully overlapped. drop lhs
+            if (rhs.contained(lhs)) {
+                merged = true;
+                break;
+            }
+            //fully overlapped. replace the lhs with rhs
+            if (lhs.contained(rhs)) {
+                rhs = {};
+                continue;
+            }
+            //just merge & expand on x axis
+            if (lhs.min.y == rhs.min.y && lhs.max.y == rhs.max.y) {
+                if (lhs.min.x <= rhs.max.x && rhs.min.x <= lhs.max.x) {
+                    rhs.min.x = std::min(lhs.min.x, rhs.min.x);
+                    rhs.max.x = std::max(lhs.max.x, rhs.max.x);
+                    merged = true;
+                    break;
+                }
+            }
+            //just merge & expand on y axis
+            if (lhs.min.x == rhs.min.x && lhs.max.x == rhs.max.x) {
+                if (lhs.min.y <= rhs.max.y && rhs.min.y < lhs.max.y) {
+                    rhs.min.y = std::min(lhs.min.y, rhs.min.y);
+                    rhs.max.y = std::max(lhs.max.y, rhs.max.y);
+                    merged = true;
+                    break;
+                }
+            }
+            //subdivide regions
+            if (lhs.intersected(rhs)) {
+                subdivide(targets, j, lhs, rhs);
+                merged = true;
+                break;
+            }
+        }
+        if (!merged) output.push(lhs);  //this region is complete isolated
+        lhs = {};
+    }
+}
+
 /************************************************************************/
 /* RenderTrimPath Class Implementation                                  */
 /************************************************************************/

src/renderer/tvgRender.h

@@ -50,7 +50,6 @@ static inline RenderUpdateFlag operator|(const RenderUpdateFlag a, const RenderU
     return RenderUpdateFlag(uint16_t(a) | uint16_t(b));
 }
 
-
 struct RenderSurface
 {
     union {
@@ -111,6 +110,11 @@ struct RenderRegion
         return ret;
     }
 
+    static constexpr RenderRegion add(const RenderRegion& lhs, const RenderRegion& rhs)
+    {
+        return {{std::min(lhs.min.x, rhs.min.x), std::min(lhs.min.y, rhs.min.y)}, {std::max(lhs.max.x, rhs.max.x), std::max(lhs.max.y, rhs.max.y)}};
+    }
+
     void intersect(const RenderRegion& rhs);
 
     void add(const RenderRegion& rhs)
@@ -121,6 +125,16 @@ struct RenderRegion
         if (rhs.max.y > max.y) max.y = rhs.max.y;
     }
 
+    bool contained(const RenderRegion& rhs)
+    {
+        return (min.x <= rhs.min.x && max.x >= rhs.max.x && min.y <= rhs.min.y && max.y >= rhs.max.y);
+    }
+
+    bool intersected(const RenderRegion& rhs) const
+    {
+        return (rhs.min.x < max.x && rhs.max.x > min.x && rhs.min.y < max.y && rhs.max.y > min.y);
+    }
+
     bool operator==(const RenderRegion& rhs) const
     {
         return (min.x == rhs.min.x && min.y == rhs.min.y && max.x == rhs.max.x && max.y == rhs.max.y);
@@ -141,6 +155,56 @@ struct RenderRegion
     uint32_t h() const { return (uint32_t) sh(); }
 };
 
+struct RenderDirtyRegion
+{
+    void add(const RenderRegion& region)
+    {
+        if (disabled || region.invalid()) return;
+        ScopedLock lock(key);
+        list[current].push(region);
+    }
+
+    bool prepare(uint32_t count = 0)
+    {
+        if (disabled) return false;
+
+        count *= 120;   //FIXME: enough?
+
+        list[0].reserve(count);
+        list[1].reserve(count);
+
+        return true;
+    }
+
+    bool deactivated()
+    {
+        if (disabled) return true;
+        return false;
+    }
+
+    void clear()
+    {
+        list[0].clear();
+        list[1].clear();
+    }
+
+    const Array<RenderRegion>& get()
+    {
+        return list[current];
+    }
+
+    void commit();
+
+    bool disabled = false;
+
+private:
+    void subdivide(Array<RenderRegion>& targets, uint32_t idx, RenderRegion& lhs, RenderRegion& rhs);
+
+    Array<RenderRegion> list[2];  //double buffer swapping
+    Key key;
+    uint8_t current = 0;          //list index. 0 or 1
+};
+
 struct RenderPath
 {
     Array<PathCommand> cmds;
@@ -420,11 +484,12 @@ struct RenderEffectTritone : RenderEffect
 class RenderMethod
 {
 private:
-    uint32_t refCnt = 0;        //reference count
+    uint32_t refCnt = 0;
     Key key;
 
 protected:
     RenderRegion vport;         //viewport
+    RenderDirtyRegion dirtyRegion;
 
 public:
     //common implementation
@@ -448,11 +513,10 @@ public:
     virtual bool blend(BlendMethod method) = 0;
     virtual ColorSpace colorSpace() = 0;
     virtual const RenderSurface* mainSurface() = 0;
-
     virtual bool clear() = 0;
     virtual bool sync() = 0;
 
-    //compositions
+    //composition
     virtual RenderCompositor* target(const RenderRegion& region, ColorSpace cs, CompositionFlag flags) = 0;
     virtual bool beginComposite(RenderCompositor* cmp, MaskMethod method, uint8_t opacity) = 0;
     virtual bool endComposite(RenderCompositor* cmp) = 0;
@@ -462,6 +526,10 @@ public:
     virtual bool region(RenderEffect* effect) = 0;
     virtual bool render(RenderCompositor* cmp, const RenderEffect* effect, bool direct) = 0;
     virtual void dispose(RenderEffect* effect) = 0;
+
+    //partial rendering
+    virtual void damage(const RenderRegion& region);
+    virtual bool partial(bool disable);
 };
 
 static inline bool MASK_REGION_MERGING(MaskMethod method)

src/renderer/tvgScene.h

@@ -131,6 +131,9 @@ struct SceneImpl : Scene
         vport = renderer->viewport();
         vdirty = true;
 
+        //bounds(renderer) here hinders parallelization.
+        if (effects) renderer->damage(vport);
+
         return true;
     }
 
@@ -257,7 +260,10 @@ struct SceneImpl : Scene
     {
         auto itr = paints.begin();
         while (itr != paints.end()) {
-            PAINT((*itr))->unref();
+            auto paint = PAINT((*itr));
+            //when the paint is destroyed damage will be triggered
+            if (paint->refCnt > 1) paint->damage();
+            paint->unref();
             paints.erase(itr++);
         }
         return Result::Success;
@@ -266,6 +272,8 @@ struct SceneImpl : Scene
     Result remove(Paint* paint)
     {
         if (PAINT(paint)->parent != this) return Result::InsufficientCondition;
+        //when the paint is destroyed damage will be triggered
+        if (PAINT(paint)->refCnt > 1) PAINT(paint)->damage();
         PAINT(paint)->unref();
         paints.remove(paint);
         return Result::Success;
@@ -310,6 +318,7 @@ struct SceneImpl : Scene
             }
             delete(effects);
             effects = nullptr;
+            impl.renderer->damage(vport);
         }
         return Result::Success;
     }