From fa856fbf68831267fde969feaa905fb5a4fd1d1d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Sat, 23 May 2026 12:14:37 +0200
Subject: [PATCH] Pass the culling unit tests.

---
 GPU/Common/DrawEngineCommon.cpp | 35 +++++++++++++++++++--------------
 GPU/Common/DrawEngineCommon.h   |  5 -----
 GPU/GPUState.h                  |  6 ++++++
 3 files changed, 26 insertions(+), 20 deletions(-)

diff --git a/GPU/Common/DrawEngineCommon.cpp b/GPU/Common/DrawEngineCommon.cpp
index 7813332170..2911bf166d 100644
--- a/GPU/Common/DrawEngineCommon.cpp
+++ b/GPU/Common/DrawEngineCommon.cpp
@@ -252,13 +252,25 @@ bool DrawEngineCommon::TestBoundingBox(const void *vdata, const void *inds, int
 		}
 	}
 
+	// Unclear why the top/left is off by a pixel.
+	const int left = gstate.getOffsetX() + std::max(gstate.getRegionX1(), gstate.getScissorX1()) - 1;
+	const int top = gstate.getOffsetY() + std::max(gstate.getRegionY1(), gstate.getScissorY1()) - 1;
+	const int right = gstate.getOffsetX() + std::min(gstate.getRegionX2(), gstate.getScissorX2()) + 1;
+	const int bottom = gstate.getOffsetY() + std::min(gstate.getRegionY2(), gstate.getScissorY2()) + 1;
+
+	// This is strange, it seems if the draw box is at all outside the 4096x4096 coordinate space, all checks pass.
+	// It seems very odd that the hardware would have checks for this.
+	if (right >= 4096 || bottom >= 4096 || left < 1.0f || top < 1.0f) {
+		return true;
+	}
+
 	// TODO: How accurate should we be?
 	int insideCount[6] = {0};
 	for (int i = 0; i < vertexCount; i++) {
 		// Complete the transform to see if the vertex should be ignored. Not sure if we need to go to these lengths...
 		const float *objpos = verts + i * 3;
 
-		float projpos[4]{};
+		float projpos[4];
 		Vec3ByMatrix44(projpos, objpos, gstate_c.worldviewproj);
 
 		if (projpos[2] >= -projpos[3]) {
@@ -268,28 +280,26 @@ bool DrawEngineCommon::TestBoundingBox(const void *vdata, const void *inds, int
 			insideCount[5]++;
 		}
 
-		float invW = 1.0f / projpos[3];
+		const float invW = 1.0f / projpos[3];
 		const float screenpos[3] = {
 			projpos[0] * gstate.getViewportXScale() * invW + gstate.getViewportXCenter(),
 			projpos[1] * gstate.getViewportYScale() * invW + gstate.getViewportYCenter(),
 			projpos[2] * gstate.getViewportZScale() * invW + gstate.getViewportZCenter(),
 		};
 
-		const float drawX = screenpos[0] - gstate.getOffsetX();
-		const float drawY = screenpos[1] - gstate.getOffsetY();
+		const float drawX = screenpos[0];
+		const float drawY = screenpos[1];
 
-		const int regionX2 = gstate.getRegionX2() + 1;
-		const int regionY2 = gstate.getRegionY2() + 1;
-		if (drawX >= 0.0f) {
+		if (drawX >= left) {
 			insideCount[0]++;
 		}
-		if (drawX <= regionX2) {
+		if (drawX <= right) {
 			insideCount[1]++;
 		}
-		if (drawY >= 0.0f) {
+		if (drawY >= top) {
 			insideCount[2]++;
 		}
-		if (drawY <= regionY2) {
+		if (drawY <= bottom) {
 			insideCount[3]++;
 		}
 	}
@@ -328,11 +338,6 @@ bool DrawEngineCommon::TestBoundingBoxFast(const void *vdata, int vertexCount, c
 		return true;
 	}
 
-	// Also let's just bail if offsetOutsideEdge_ is set, instead of handling the cases.
-	// NOTE: This is written to in UpdatePlanes so can't check it before.
-	if (offsetOutsideEdge_)
-		return true;
-
 	// Simple, most common case.
 	int stride = dec->VertexSize();
 	int offset = dec->posoff;
diff --git a/GPU/Common/DrawEngineCommon.h b/GPU/Common/DrawEngineCommon.h
index 5228bc4559..ef7d5039fb 100644
--- a/GPU/Common/DrawEngineCommon.h
+++ b/GPU/Common/DrawEngineCommon.h
@@ -350,11 +350,6 @@ protected:
 	// Hardware tessellation
 	TessellationDataTransfer *tessDataTransfer = nullptr;
 
-	// Culling
-	Vec2f minOffset_{};
-	Vec2f maxOffset_{};
-	bool offsetOutsideEdge_ = false;
-
 	GPUCommon *gpuCommon_ = nullptr;
 
 	// Software depth raster
diff --git a/GPU/GPUState.h b/GPU/GPUState.h
index a8b517a585..5f9f19b788 100644
--- a/GPU/GPUState.h
+++ b/GPU/GPUState.h
@@ -375,8 +375,14 @@ struct GPUgstate {
 	int getScissorY1() const { return (scissor1 >> 10) & 0x3FF; }
 	int getScissorX2() const { return scissor2 & 0x3FF; }
 	int getScissorY2() const { return (scissor2 >> 10) & 0x3FF; }
+
+	// When drawing, these registers do weird stuff. Must be set to 0 for normal behavior.
 	int getRegionRateX() const { return 0x100 + (region1 & 0x3FF); }
 	int getRegionRateY() const { return 0x100 + ((region1 >> 10) & 0x3FF); }
+
+	// However when testing for bbox culling, these do act like X1/X2.
+	int getRegionX1() const { return (region1 & 0x3FF); }
+	int getRegionY1() const { return (region1 >> 10) & 0x3FF; }
 	int getRegionX2() const { return (region2 & 0x3FF); }
 	int getRegionY2() const { return (region2 >> 10) & 0x3FF; }