Make Frozen Time rendering 4 times faster by grouping draw segments in batches of 100 (old algorithm processed 32000 draw segs per sprite!)

src/swrenderer/scene/r_translucent_pass.cpp

@@ -165,6 +165,7 @@ namespace swrenderer
 	{
 		CollectPortals();
 		Thread->SpriteList->Sort();
+		Thread->DrawSegments->BuildSegmentGroups();
 
 		Clip3DFloors *clip3d = Thread->Clip3DFloors.get();
 		if (clip3d->height_top == nullptr)

src/swrenderer/segments/r_drawsegment.cpp

@@ -40,6 +40,7 @@
 #include "swrenderer/things/r_visiblesprite.h"
 #include "swrenderer/scene/r_light.h"
 #include "swrenderer/viewport/r_viewport.h"
+#include "swrenderer/r_renderthread.h"
 
 namespace swrenderer
 {
@@ -83,4 +84,83 @@ namespace swrenderer
 	{
 		InterestingSegments.Push(segment);
 	}
+
+	void DrawSegmentList::BuildSegmentGroups()
+	{
+		SegmentGroups.Clear();
+
+		unsigned int groupSize = 100;
+		for (unsigned int index = BeginIndex(); index < EndIndex(); index += groupSize)
+		{
+			auto ds = Segment(index);
+
+			DrawSegmentGroup group;
+			group.BeginIndex = index;
+			group.EndIndex = MIN(index + groupSize, EndIndex());
+			group.x1 = ds->x1;
+			group.x2 = ds->x2;
+			group.neardepth = MIN(ds->sz1, ds->sz2);
+			group.fardepth = MAX(ds->sz1, ds->sz2);
+
+			for (unsigned int groupIndex = group.BeginIndex + 1; groupIndex < group.EndIndex; groupIndex++)
+			{
+				ds = Segment(groupIndex);
+				group.x1 = MIN(group.x1, ds->x1);
+				group.x2 = MAX(group.x2, ds->x2);
+				group.neardepth = MIN(group.neardepth, ds->sz1);
+				group.neardepth = MIN(group.neardepth, ds->sz2);
+				group.fardepth = MAX(ds->sz1, group.fardepth);
+				group.fardepth = MAX(ds->sz2, group.fardepth);
+			}
+
+			for (int x = group.x1; x < group.x2; x++)
+			{
+				cliptop[x] = 0;
+				clipbottom[x] = viewheight;
+			}
+
+			for (unsigned int groupIndex = group.BeginIndex; groupIndex < group.EndIndex; groupIndex++)
+			{
+				ds = Segment(groupIndex);
+
+				// kg3D - no clipping on fake segs
+				if (ds->fake) continue;
+
+				if (ds->silhouette & SIL_BOTTOM)
+				{
+					short *clip1 = clipbottom + ds->x1;
+					const short *clip2 = ds->sprbottomclip;
+					int i = ds->x2 - ds->x1;
+					do
+					{
+						if (*clip1 > *clip2)
+							*clip1 = *clip2;
+						clip1++;
+						clip2++;
+					} while (--i);
+				}
+
+				if (ds->silhouette & SIL_TOP)
+				{
+					short *clip1 = cliptop + ds->x1;
+					const short *clip2 = ds->sprtopclip;
+					int i = ds->x2 - ds->x1;
+					do
+					{
+						if (*clip1 < *clip2)
+							*clip1 = *clip2;
+						clip1++;
+						clip2++;
+					} while (--i);
+				}
+			}
+
+			group.sprtopclip = Thread->FrameMemory->AllocMemory<short>(group.x2 - group.x1);
+			group.sprbottomclip = Thread->FrameMemory->AllocMemory<short>(group.x2 - group.x1);
+			memcpy(group.sprtopclip, cliptop + group.x1, (group.x2 - group.x1) * sizeof(short));
+			memcpy(group.sprbottomclip, clipbottom + group.x1, (group.x2 - group.x1) * sizeof(short));
+
+			SegmentGroups.Push(group);
+		}
+	}
 }

src/swrenderer/segments/r_drawsegment.h

@@ -47,11 +47,23 @@ namespace swrenderer
 		int CurrentPortalUniq; // [ZZ] to identify the portal that this drawseg is in. used for sprite clipping.
 	};
 
+	struct DrawSegmentGroup
+	{
+		short x1, x2;
+		float neardepth, fardepth;
+		short *sprtopclip;
+		short *sprbottomclip;
+		unsigned int BeginIndex;
+		unsigned int EndIndex;
+	};
+
 	class DrawSegmentList
 	{
 	public:
 		DrawSegmentList(RenderThread *thread);
 
+		TArray<DrawSegmentGroup> SegmentGroups;
+
 		unsigned int BeginIndex() const { return StartIndices.Last(); }
 		unsigned int EndIndex() const { return Segments.Size(); }
 		DrawSegment *Segment(unsigned int index) const { return Segments[Segments.Size() - 1 - index]; }
@@ -66,6 +78,8 @@ namespace swrenderer
 		void Push(DrawSegment *segment);
 		void PushInteresting(DrawSegment *segment);
 
+		void BuildSegmentGroups();
+
 		RenderThread *Thread = nullptr;
 
 	private:
@@ -74,5 +88,9 @@ namespace swrenderer
 
 		TArray<DrawSegment *> InterestingSegments; // drawsegs that have something drawn on them
 		TArray<unsigned int> StartInterestingIndices;
+
+		// For building segment groups
+		short cliptop[MAXWIDTH];
+		short clipbottom[MAXWIDTH];
 	};
 }

src/swrenderer/things/r_visiblesprite.cpp

@@ -285,7 +285,45 @@ namespace swrenderer
 		// The first drawseg that is closer than the sprite is the clip seg.
 
 		DrawSegmentList *segmentlist = thread->DrawSegments.get();
-		for (unsigned int index = segmentlist->BeginIndex(); index != segmentlist->EndIndex(); index++)
+		for (unsigned int groupIndex = 0; groupIndex < segmentlist->SegmentGroups.Size(); groupIndex++)
+		{
+			auto &group = segmentlist->SegmentGroups[groupIndex];
+			if (group.x1 >= x2 || group.x2 <= x1)
+				continue;
+
+			if (group.fardepth < spr->depth) 
+			{
+				r1 = MAX<int>(group.x1, x1);
+				r2 = MIN<int>(group.x2, x2);
+
+				// Clip bottom
+				clip1 = clipbot + r1;
+				clip2 = group.sprbottomclip + r1 - group.x1;
+				i = r2 - r1;
+				do
+				{
+					if (*clip1 > *clip2)
+						*clip1 = *clip2;
+					clip1++;
+					clip2++;
+				} while (--i);
+
+				// Clip top
+				clip1 = cliptop + r1;
+				clip2 = group.sprtopclip + r1 - group.x1;
+				i = r2 - r1;
+				do
+				{
+					if (*clip1 < *clip2)
+						*clip1 = *clip2;
+					clip1++;
+					clip2++;
+				} while (--i);
+			}
+			else
+			{
+				//for (unsigned int index = segmentlist->BeginIndex(); index != segmentlist->EndIndex(); index++)
+				for (unsigned int index = group.BeginIndex; index != group.EndIndex; index++)
 				{
 					DrawSegment *ds = segmentlist->Segment(index);
 
@@ -321,7 +359,6 @@ namespace swrenderer
 						}
 					}
 
-
 					// Check if sprite is in front of draw seg:
 					if ((!spr->IsWallSprite() && neardepth > spr->depth) || ((spr->IsWallSprite() || fardepth > spr->depth) &&
 						(spr->gpos.Y - ds->curline->v1->fY()) * (ds->curline->v2->fX() - ds->curline->v1->fX()) -
@@ -373,6 +410,8 @@ namespace swrenderer
 						} while (--i);
 					}
 				}
+			}
+		}
 
 		// all clipping has been performed, so draw the sprite