From 510229b68b1c00110beb747f0f10f431a35e2286 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Tue, 22 Oct 2019 23:46:26 +0200
Subject: [PATCH] SoftGPU: Detect through-mode rectangles from triangle strips

---
 GPU/Math3D.h                   |  4 ++++
 GPU/Software/Clipper.cpp       |  2 +-
 GPU/Software/TransformUnit.cpp | 38 ++++++++++++++++++++++++++++++++--
 3 files changed, 41 insertions(+), 3 deletions(-)

diff --git a/GPU/Math3D.h b/GPU/Math3D.h
index 292c63be05..ada1a6931e 100644
--- a/GPU/Math3D.h
+++ b/GPU/Math3D.h
@@ -625,6 +625,10 @@ public:
 		*this = *this / f;
 	}
 
+	bool operator ==(const Vec4 &other) const {
+		return x == other.x && y == other.y && z == other.z && w == other.w;
+	}
+
 	T Length2() const
 	{
 		return x*x + y*y + z*z + w*w;
diff --git a/GPU/Software/Clipper.cpp b/GPU/Software/Clipper.cpp
index ca04f7f97b..746896c431 100644
--- a/GPU/Software/Clipper.cpp
+++ b/GPU/Software/Clipper.cpp
@@ -198,7 +198,7 @@ void ProcessRect(const VertexData& v0, const VertexData& v1)
 
 		// Color and depth values of second vertex are used for the whole rectangle
 		buf[0].color0 = buf[1].color0 = buf[2].color0 = buf[3].color0;
-		buf[0].color1 = buf[1].color1 = buf[2].color1 = buf[3].color1;
+		buf[0].color1 = buf[1].color1 = buf[2].color1 = buf[3].color1;  // is color1 ever used in through mode?
 		buf[0].clippos.w = buf[1].clippos.w = buf[2].clippos.w = buf[3].clippos.w = 1.0f;
 		buf[0].fogdepth = buf[1].fogdepth = buf[2].fogdepth = buf[3].fogdepth = 1.0f;
 
diff --git a/GPU/Software/TransformUnit.cpp b/GPU/Software/TransformUnit.cpp
index 3b0603fd79..2848ce600f 100644
--- a/GPU/Software/TransformUnit.cpp
+++ b/GPU/Software/TransformUnit.cpp
@@ -319,8 +319,7 @@ void TransformUnit::SubmitPrimitive(void* vertices, void* indices, GEPrimitiveTy
 
 	VertexReader vreader(buf, vtxfmt, vertex_type);
 
-	const int max_vtcs_per_prim = 3;
-	static VertexData data[max_vtcs_per_prim];
+	static VertexData data[4];  // Normally max verts per prim is 3, but we temporarily need 4 to detect rectangles from strips.
 	// This is the index of the next vert in data (or higher, may need modulus.)
 	static int data_index = 0;
 
@@ -439,6 +438,41 @@ void TransformUnit::SubmitPrimitive(void* vertices, void* indices, GEPrimitiveTy
 			// Don't draw a triangle when loading the first two vertices.
 			int skip_count = data_index >= 2 ? 0 : 2 - data_index;
 
+			// If index count == 4, check if we can convert to a rectangle.
+			// This is for Darkstalkers (and should speed up many 2D games).
+			if (vertex_count == 4 && gstate.isModeThrough()) {
+				for (int vtx = 0; vtx < 4; ++vtx) {
+					if (indices) {
+						vreader.Goto(ConvertIndex(vtx) - index_lower_bound);
+					}
+					else {
+						vreader.Goto(vtx);
+					}
+					data[vtx] = ReadVertex(vreader);
+				}
+
+				// OK, now let's look at data to detect rectangles. There are a few possibilities
+				// but we focus on Darkstalkers for now.
+				if (data[0].screenpos.x == data[1].screenpos.x &&
+					data[0].screenpos.y == data[2].screenpos.y &&
+					data[2].screenpos.x == data[3].screenpos.x &&
+					data[1].screenpos.y == data[3].screenpos.y &&
+					data[1].screenpos.y > data[0].screenpos.y &&  // Avoid rotation handling
+					data[2].screenpos.x > data[0].screenpos.x &&
+					data[0].texturecoords.x == data[1].texturecoords.x &&
+					data[0].texturecoords.y == data[2].texturecoords.y &&
+					data[2].texturecoords.x == data[3].texturecoords.x &&
+					data[1].texturecoords.y == data[3].texturecoords.y &&
+					data[1].texturecoords.y >  data[0].texturecoords.y &&
+					data[2].texturecoords.x >  data[0].texturecoords.x &&
+					data[0].color0 == data[1].color0 &&
+					data[1].color0 == data[2].color0 &&
+					data[2].color0 == data[3].color0) {
+					// It's a rectangle!
+					Clipper::ProcessRect(data[0], data[3]);
+				}
+			}
+
 			for (int vtx = 0; vtx < vertex_count; ++vtx) {
 				if (indices) {
 					vreader.Goto(ConvertIndex(vtx) - index_lower_bound);