From 2fd24277c1cac52cd21884bdbed01312441b3b9f Mon Sep 17 00:00:00 2001
From: refractionpcsx2 <refraction@gmail.com>
Date: Thu, 16 Nov 2023 23:16:21 +0000
Subject: [PATCH] GS: Pre-round/truncate STQ values based on hardware tests

---
 pcsx2/GS/GSState.cpp | 54 +++++++++++++++++++++++++++++++++++++++++++-
 pcsx2/GS/GSState.h   |  1 +
 2 files changed, 54 insertions(+), 1 deletion(-)

diff --git a/pcsx2/GS/GSState.cpp b/pcsx2/GS/GSState.cpp
index bf8ebb9288..87b2917c95 100644
--- a/pcsx2/GS/GSState.cpp
+++ b/pcsx2/GS/GSState.cpp
@@ -27,6 +27,7 @@
 #include <cfloat>
 #include <fstream>
 #include <iomanip>
+#include <bit>
 
 int GSState::s_n = 0;
 int GSState::s_last_transfer_draw_n = 0;
@@ -503,7 +504,7 @@ void GSState::DumpVertices(const std::string& filename)
 			file << uv_U << DEL << uv_V;
 		}
 		else
-			file << v.ST.S << DEL << v.ST.T << DEL << v.RGBAQ.Q;
+			file << v.ST.S << "(" << std::bit_cast<u32>(v.ST.S) << ")" << DEL << v.ST.T << "(" << std::bit_cast<u32>(v.ST.T) << ")" << DEL << v.RGBAQ.Q << "(" << std::bit_cast<u32>(v.RGBAQ.Q) << ")";
 
 		file << std::endl;
 	}
@@ -1593,6 +1594,13 @@ inline bool GSState::TestDrawChanged()
 	return false;
 }
 
+u32 GSState::CalcMask(int exp, int max_exp)
+{
+	const int amount = 9 + (max_exp - exp);
+
+	return (1 << std::min(amount, 23)) - 1;
+}
+
 void GSState::FlushPrim()
 {
 	if (m_index.tail > 0)
@@ -1668,6 +1676,50 @@ void GSState::FlushPrim()
 
 		m_vt.Update(m_vertex.buff, m_index.buff, m_vertex.tail, m_index.tail, GSUtil::GetPrimClass(PRIM->PRIM));
 
+		// Texel coordinate rounding
+		// Helps Manhunt (lights shining through objects).
+		// Can help with some alignment issues when upscaling too, and is for both Software and Hardware renderers.
+		// Sometimes hardware doesn't get affected, likely due to the difference in how GPU's handle textures (Persona minimap).
+		if (m_env.PRIM.TME && (GSUtil::GetPrimClass(PRIM->PRIM) == GS_PRIM_CLASS::GS_SPRITE_CLASS || m_vt.m_eq.z))
+		{
+			if (!m_env.PRIM.FST) // STQ's
+			{
+				const bool is_sprite = GSUtil::GetPrimClass(PRIM->PRIM) == GS_PRIM_CLASS::GS_SPRITE_CLASS;
+				// ST's have the lowest 9 bits (or greater depending on exponent difference) rounding down (from hardware tests).
+				for (int i = m_index.tail - 1; i >= 0; i--)
+				{
+					GSVertex* v = &m_vertex.buff[m_index.buff[i]];
+
+					// Only Q on the second vertex is valid
+					if (!(i & 1) && is_sprite)
+						v->RGBAQ.Q = m_vertex.buff[m_index.buff[i + 1]].RGBAQ.Q;
+
+					int T = std::bit_cast<int>(v->ST.T);
+					int Q = std::bit_cast<int>(v->RGBAQ.Q);
+					int S = std::bit_cast<int>(v->ST.S);
+					const int expS = (S >> 23) & 0xff;
+					const int expT = (T >> 23) & 0xff;
+					const int expQ = (Q >> 23) & 0xff;
+					int max_exp = std::max(expS, expQ);
+
+					u32 mask = CalcMask(expS, max_exp);
+					S &= ~mask;
+					v->ST.S = std::bit_cast<float>(S);
+					max_exp = std::max(expT, expQ);
+					mask = CalcMask(expT, max_exp);
+					T &= ~mask;
+					v->ST.T = std::bit_cast<float>(T);
+					Q &= ~0xff;
+
+					if (!is_sprite || (i & 1))
+						v->RGBAQ.Q = std::bit_cast<float>(Q);
+
+					m_vt.m_min.t.x = std::min(m_vt.m_min.t.x, (v->ST.S / v->RGBAQ.Q) * (1 << m_context->TEX0.TW));
+					m_vt.m_min.t.y = std::min(m_vt.m_min.t.y, (v->ST.T / v->RGBAQ.Q) * (1 << m_context->TEX0.TH));
+				}
+			}
+		}
+
 		// Skip draw if Z test is enabled, but set to fail all pixels.
 		const bool skip_draw = (m_context->TEST.ZTE && m_context->TEST.ZTST == ZTST_NEVER);
 
diff --git a/pcsx2/GS/GSState.h b/pcsx2/GS/GSState.h
index 4a6f214c55..0e49400269 100644
--- a/pcsx2/GS/GSState.h
+++ b/pcsx2/GS/GSState.h
@@ -383,6 +383,7 @@ public:
 	virtual void UpdateSettings(const Pcsx2Config::GSOptions& old_config);
 
 	void Flush(GSFlushReason reason);
+	u32 CalcMask(int exp, int max_exp);
 	void FlushPrim();
 	bool TestDrawChanged();
 	void FlushWrite();