sw_engine: add support for partial rendering

This implements RenderDirtyRegion.

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

src/renderer/sw_engine/tvgSwRenderer.cpp

@@ -40,7 +40,7 @@ struct SwTask : Task
 {
     SwSurface* surface = nullptr;
     SwMpool* mpool = nullptr;
-    RenderRegion bbox;                          //Rendering Region
+    RenderRegion bbox[2] = {{}, {}};  //Rendering Region 0:current, 1:prevous
     Matrix transform;
     Array<RenderData> clips;
     RenderUpdateFlag flags = RenderUpdateFlag::None;
@@ -52,7 +52,7 @@ struct SwTask : Task
     {
         //Can we skip the synchronization?
         done();
-        return bbox;
+        return bbox[0];
     }
 
     virtual void dispose() = 0;
@@ -92,7 +92,7 @@ struct SwShapeTask : SwTask
     bool clip(SwRle* target) override
     {
         if (shape.strokeRle) return rleClip(target, shape.strokeRle);
-        if (shape.fastTrack) return rleClip(target, &bbox);
+        if (shape.fastTrack) return rleClip(target, &bbox[0]);
         if (shape.rle) return rleClip(target, shape.rle);
         return false;
     }
@@ -101,7 +101,7 @@ struct SwShapeTask : SwTask
     {
         //Invisible
         if (opacity == 0 && !clipper) {
-            bbox.reset();
+            bbox[0].reset();
             return;
         }
 
@@ -115,7 +115,7 @@ struct SwShapeTask : SwTask
             updateFill = (MULTIPLY(rshape->color.a, opacity) || rshape->fill);
             if (updateShape) shapeReset(&shape);
             if (updateFill || clipper) {
-                if (shapePrepare(&shape, rshape, transform, bbox, renderBox, mpool, tid, clips.count > 0 ? true : false)) {
+                if (shapePrepare(&shape, rshape, transform, bbox[0], renderBox, mpool, tid, clips.count > 0 ? true : false)) {
                     if (!shapeGenRle(&shape, rshape, antialiasing(strokeWidth))) goto err;
                 } else {
                     updateFill = false;
@@ -135,7 +135,7 @@ struct SwShapeTask : SwTask
         if (updateShape || flags & RenderUpdateFlag::Stroke) {
             if (strokeWidth > 0.0f) {
                 shapeResetStroke(&shape, rshape, transform);
-                if (!shapeGenStrokeRle(&shape, rshape, transform, bbox, renderBox, mpool, tid)) goto err;
+                if (!shapeGenStrokeRle(&shape, rshape, transform,  bbox[0], renderBox, mpool, tid)) goto err;
                 if (auto fill = rshape->strokeFill()) {
                     auto ctable = (flags & RenderUpdateFlag::GradientStroke) ? true : false;
                     if (ctable) shapeResetStrokeFill(&shape);
@@ -157,12 +157,11 @@ struct SwShapeTask : SwTask
             if (!clipShapeRle && !clipStrokeRle) goto err;
         }
 
-        bbox = renderBox; //sync
+        bbox[0] = renderBox; //sync
-
         return;
 
     err:
-        bbox.reset();
+        bbox[0].reset();
         shapeReset(&shape);
         rleReset(shape.strokeRle);
         shapeDelOutline(&shape, mpool, tid);
@@ -188,7 +187,7 @@ struct SwImageTask : SwTask
 
     void run(unsigned tid) override
     {
-        auto clipBox = bbox;
+        auto clipBox = bbox[0];
 
         //Convert colorspace if it's not aligned.
         rasterConvertCS(source, surface->cs);
@@ -204,9 +203,11 @@ struct SwImageTask : SwTask
         if ((flags & (RenderUpdateFlag::Image | RenderUpdateFlag::Transform | RenderUpdateFlag::Color)) && (opacity > 0)) {
             imageReset(&image);
             if (!image.data || image.w == 0 || image.h == 0) goto end;
-            if (!imagePrepare(&image, transform, clipBox,  bbox, mpool, tid)) goto end;
+
+            if (!imagePrepare(&image, transform, clipBox, bbox[0], mpool, tid)) goto end;
+
             if (clips.count > 0) {
-                if (!imageGenRle(&image, bbox, false)) goto end;
+                if (!imageGenRle(&image, bbox[0], false)) goto end;
                 if (image.rle) {
                     //Clear current task memorypool here if the clippers would use the same memory pool
                     imageDelOutline(&image, mpool, tid);
@@ -220,7 +221,7 @@ struct SwImageTask : SwTask
         }
         goto end;
     err:
-        bbox.reset();
+        bbox[0].reset();
         rleReset(image.rle);
     end:
         imageDelOutline(&image, mpool, tid);
@@ -266,7 +267,10 @@ SwRenderer::~SwRenderer()
 
 bool SwRenderer::clear()
 {
-    if (surface) return rasterClear(surface, 0, 0, surface->w, surface->h);
+    if (surface) {
+        fulldraw = true;
+        return rasterClear(surface, 0, 0, surface->w, surface->h);
+    }
     return false;
 }
 
@@ -334,7 +338,28 @@ bool SwRenderer::postUpdate()
 
 bool SwRenderer::preRender()
 {
-    return surface != nullptr;
+    if (!surface) return false;
+    if (fulldraw || !dirtyRegion.prepare(tasks.count)) return true;
+
+    //TODO: optimize to remove this iteration.
+    //collect the old and new dirtry regions
+    ARRAY_FOREACH(p, tasks) {
+        auto task = *p;
+        task->done();
+        auto& cur = task->bbox[0];
+        auto& prv = task->bbox[1];
+        //quick generous merge if two regions are close enough.
+        if (abs(cur.min.y - prv.min.y) < 5 && abs(cur.max.y - prv.max.y) < 5 && abs(cur.min.x - prv.min.x) < 5 && abs(cur.max.x - prv.max.x) < 5) {
+            dirtyRegion.add(RenderRegion::add(task->bbox[0], task->bbox[1]));
+        } else {
+            dirtyRegion.add(task->bbox[0]);
+            dirtyRegion.add(task->bbox[1]);
+        }
+    }
+
+    dirtyRegion.commit();
+
+    return true;
 }
 
 
@@ -359,9 +384,15 @@ bool SwRenderer::postRender()
 
     ARRAY_FOREACH(p, tasks) {
         if ((*p)->disposed) delete(*p);
-        else (*p)->pushed = false;
+        else {
+            (*p)->bbox[1] = (*p)->bbox[0];
+            (*p)->pushed = false;
+        }
     }
     tasks.clear();
+    dirtyRegion.clear();
+
+    fulldraw = false;
 
     return true;
 }
@@ -375,7 +406,18 @@ bool SwRenderer::renderImage(RenderData data)
     if (task->opacity == 0) return true;
 
     //full scene or partial rendering
-    return rasterImage(surface, &task->image, task->transform, task->bbox, task->opacity);
+    if (fulldraw || task->pushed || dirtyRegion.deactivated()) {
+        rasterImage(surface, &task->image, task->transform, task->bbox[0], task->opacity);
+    } else {
+         ARRAY_FOREACH(p, dirtyRegion.get()) {
+            if (task->bbox[0].min.x >= p->max.x) break;  //dirtyRegion is sorted in x order
+            if (task->bbox[0].intersected(*p)) {
+                auto bbox = RenderRegion::intersect(task->bbox[0], *p);
+                rasterImage(surface, &task->image, task->transform, bbox, task->opacity);
+            }
+        }
+    }
+    return true;
 }
 
 
@@ -411,12 +453,27 @@ bool SwRenderer::renderShape(RenderData data)
         }
     };
 
-    if (task->rshape->strokeFirst()) {
+    //full scene or partial rendering
-        stroke(task, surface, task->bbox);
+    if (fulldraw || task->pushed || dirtyRegion.deactivated()) {
-        fill(task, surface, task->shape.bbox);
+        if (task->rshape->strokeFirst()) {
+            stroke(task, surface, task->bbox[0]);
+            fill(task, surface, task->shape.bbox);
+        } else {
+            fill(task, surface, task->shape.bbox);
+            stroke(task, surface, task->bbox[0]);
+        }
     } else {
-        fill(task, surface, task->shape.bbox);
+        //TODO: skip the stroke bbox if they are invalid.
-        stroke(task, surface, task->bbox);
+         ARRAY_FOREACH(p, dirtyRegion.get()) {
+            if (task->bbox[0].min.x >= p->max.x) break;   //dirtyRegion is sorted in x order
+            if (task->rshape->strokeFirst()) {
+                if (task->bbox[0].intersected(*p)) stroke(task, surface, RenderRegion::intersect(task->bbox[0], *p));
+                if (task->shape.bbox.intersected(*p)) fill(task, surface, RenderRegion::intersect(task->shape.bbox, *p));
+            } else {
+                if (task->shape.bbox.intersected(*p)) fill(task, surface, RenderRegion::intersect(task->shape.bbox, *p));
+                if (task->bbox[0].intersected(*p)) stroke(task, surface, RenderRegion::intersect(task->bbox[0], *p));
+            }
+        }
     }
 
     return true;
@@ -477,9 +534,9 @@ bool SwRenderer::blend(BlendMethod method)
 }
 
 
-void SwRenderer::damage(TVG_UNUSED const RenderRegion& region)
+void SwRenderer::damage(const RenderRegion& region)
 {
-    //TODO:
+    dirtyRegion.add(region);
 }
 
 
@@ -686,6 +743,9 @@ void SwRenderer::dispose(RenderData data)
     task->done();
     task->dispose();
 
+    //should be updated for the region; the current paint is removed
+    dirtyRegion.add(task->bbox[0]);
+
     if (task->pushed) task->disposed = true;
     else delete(task);
 }
@@ -706,7 +766,7 @@ void* SwRenderer::prepareCommon(SwTask* task, const Matrix& transform, const Arr
 
     task->surface = surface;
     task->mpool = mpool;
-    task->bbox = RenderRegion::intersect(vport, {{0, 0}, {int32_t(surface->w), int32_t(surface->h)}});
+    task->bbox[0] = RenderRegion::intersect(vport, {{0, 0}, {int32_t(surface->w), int32_t(surface->h)}});
     task->transform = transform;
     task->clips = clips;
     task->opacity = opacity;

src/renderer/sw_engine/tvgSwRenderer.h

@@ -71,12 +71,14 @@ public:
     static bool term();
 
 private:
+    RenderDirtyRegion    dirtyRegion;
     SwSurface*           surface = nullptr;           //active surface
     Array<SwTask*>       tasks;                       //async task list
     Array<SwSurface*>    compositors;                 //render targets cache list
     SwMpool*             mpool;                       //private memory pool
     RenderRegion         vport;                       //viewport
     bool                 sharedMpool;                 //memory-pool behavior policy
+    bool                 fulldraw = true;             //buffer is cleared (need to redraw full screen)
 
     SwRenderer();
     ~SwRenderer();

src/renderer/tvgRender.cpp

@@ -20,6 +20,7 @@
  * SOFTWARE.
  */
 
+#include <algorithm>
 #include "tvgMath.h"
 #include "tvgRender.h"
 
@@ -116,6 +117,134 @@ 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 (skip || disabled) return;
+
+    auto& targets = list[current];
+    if (targets.empty()) return;
+
+    if (targets.count > THRESHOLD) {
+        skip = true;
+        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

@@ -106,6 +106,11 @@ struct RenderRegion
         return {{std::max(lhs.min.x, rhs.min.x), std::max(lhs.min.y, rhs.min.y)}, {std::min(lhs.max.x, rhs.max.x), std::min(lhs.max.y, rhs.max.y)}};
     }
 
+    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)
@@ -116,6 +121,16 @@ 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);
@@ -477,7 +492,7 @@ struct RenderEffectTritone : RenderEffect
 class RenderMethod
 {
 private:
-    uint32_t refCnt = 0;        //reference count
+    uint32_t refCnt = 0;
     Key key;
 
 public: