From 01a1438dcecd302def73b6181313027dc8782df2 Mon Sep 17 00:00:00 2001
From: Henrik Rydgard <hrydgard@gmail.com>
Date: Mon, 8 Sep 2014 23:10:23 +0200
Subject: [PATCH] Improve depth buffering in D3D by computing a depth-flipped
 proj matrix but only when needed.

---
 GPU/Directx9/ShaderManagerDX9.cpp         | 24 +++++++++++------------
 GPU/Directx9/StateMappingDX9.cpp          | 10 +++++++---
 GPU/Directx9/VertexShaderGeneratorDX9.cpp |  2 +-
 GPU/GPUState.cpp                          |  7 ++++++-
 GPU/GPUState.h                            |  1 +
 5 files changed, 27 insertions(+), 17 deletions(-)

diff --git a/GPU/Directx9/ShaderManagerDX9.cpp b/GPU/Directx9/ShaderManagerDX9.cpp
index ea087efbc1..23ec98e073 100644
--- a/GPU/Directx9/ShaderManagerDX9.cpp
+++ b/GPU/Directx9/ShaderManagerDX9.cpp
@@ -255,19 +255,13 @@ void LinkedShaderDX9::SetMatrix(D3DXHANDLE uniform, const float* pMatrix) {
 	m_vs->constant->SetMatrix(pD3Ddevice, uniform, pDxMat);
 }
 
-// Depth in ogl is between -1;1 we need between 0;1
-// Pretty sure this is wrong, our Z buffer is screwed up anyhow..
-void ConvertProjMatrixToD3D(Matrix4x4 & in) {
-	/*
-	in.zz *= 0.5f;
-	in.wz += 1.f;
-	*/
+// Depth in ogl is between -1;1 we need between 0;1 and optionally reverse it
+void ConvertProjMatrixToD3D(Matrix4x4 & in, bool invert) {
 	Matrix4x4 s;
 	Matrix4x4 t;
-	s.setScaling(Vec3(1, 1, 0.5f));
+	s.setScaling(Vec3(1, 1, invert ? -0.5 : 0.5f));
 	t.setTranslation(Vec3(0, 0, 0.5f));
-	in = in * s;
-	in = in * t;
+	in = in * s * t;
 }
 
 void LinkedShaderDX9::use() {
@@ -297,8 +291,14 @@ void LinkedShaderDX9::updateUniforms() {
 			flippedMatrix[0] = -flippedMatrix[0];
 			flippedMatrix[12] = -flippedMatrix[12];
 		}
+		bool invert = false;
+		if (gstate_c.vpDepth < 0) {
+			invert = true;
+			//flippedMatrix[9] = -flippedMatrix[9];
+			//flippedMatrix[14] = -flippedMatrix[14];
+		}
 		// Convert matrices !
-		ConvertProjMatrixToD3D(flippedMatrix);
+		ConvertProjMatrixToD3D(flippedMatrix, invert);
 
 		SetMatrix(u_proj, flippedMatrix.getReadPtr());
 	}
@@ -308,7 +308,7 @@ void LinkedShaderDX9::updateUniforms() {
 		proj_through.setOrtho(0.0f, gstate_c.curRTWidth, gstate_c.curRTHeight, 0, 0, 1);
 
 		// Convert matrices !
-		ConvertProjMatrixToD3D(proj_through);
+		ConvertProjMatrixToD3D(proj_through, false);
 
 		SetMatrix(u_proj_through, proj_through.getReadPtr());
 	}
diff --git a/GPU/Directx9/StateMappingDX9.cpp b/GPU/Directx9/StateMappingDX9.cpp
index aae6261624..4f72e746ce 100644
--- a/GPU/Directx9/StateMappingDX9.cpp
+++ b/GPU/Directx9/StateMappingDX9.cpp
@@ -381,7 +381,6 @@ void TransformDrawEngineDX9::ApplyDrawState(int prim) {
 		float vpY0 = vpYb - offsetY + vpYa;   // Need to account for sign of Y
 		gstate_c.vpWidth = vpXa * 2.0f;
 		gstate_c.vpHeight = -vpYa * 2.0f;
-
 		float vpWidth = fabsf(gstate_c.vpWidth);
 		float vpHeight = fabsf(gstate_c.vpHeight);
 
@@ -399,8 +398,13 @@ void TransformDrawEngineDX9::ApplyDrawState(int prim) {
 
 		float zScale = getFloat24(gstate.viewportz1) / 65535.0f;
 		float zOff = getFloat24(gstate.viewportz2) / 65535.0f;
-		float depthRangeMin = zOff - zScale;
-		float depthRangeMax = zOff + zScale;
+
+		float depthRangeMin = zOff - fabsf(zScale);
+		float depthRangeMax = zOff + fabsf(zScale);
+
+		gstate_c.vpDepth = zScale * 2;
+
+		// D3D does not like viewports outside the screen. Let's clamp for now.
 
 		dxstate.viewport.set(vpX0 + renderX, vpY0 + renderY, vpWidth, vpHeight, depthRangeMin, depthRangeMax);
 	}
diff --git a/GPU/Directx9/VertexShaderGeneratorDX9.cpp b/GPU/Directx9/VertexShaderGeneratorDX9.cpp
index dca22cd963..cc4d189854 100644
--- a/GPU/Directx9/VertexShaderGeneratorDX9.cpp
+++ b/GPU/Directx9/VertexShaderGeneratorDX9.cpp
@@ -593,7 +593,7 @@ void GenerateVertexShaderDX9(int prim, char *buffer, bool useHWTransform) {
 
 	}
 
-	WRITE(p, "Out.gl_Position.z = (Out.gl_Position.z + Out.gl_Position.w) * 0.5f;");
+	// WRITE(p, "Out.gl_Position.z = (Out.gl_Position.z + Out.gl_Position.w) * 0.5f;");
 	WRITE(p, "	return Out;             ");
 	WRITE(p, "}\n");
 }
diff --git a/GPU/GPUState.cpp b/GPU/GPUState.cpp
index 9ed6c211cb..8a6422bea7 100644
--- a/GPU/GPUState.cpp
+++ b/GPU/GPUState.cpp
@@ -266,7 +266,7 @@ struct GPUStateCache_v0
 };
 
 void GPUStateCache::DoState(PointerWrap &p) {
-	auto s = p.Section("GPUStateCache", 0, 3);
+	auto s = p.Section("GPUStateCache", 0, 4);
 	if (!s) {
 		// Old state, this was not versioned.
 		GPUStateCache_v0 old;
@@ -328,6 +328,11 @@ void GPUStateCache::DoState(PointerWrap &p) {
 
 	p.Do(vpWidth);
 	p.Do(vpHeight);
+	if (s >= 4) {
+		p.Do(vpDepth);
+	} else {
+		vpDepth = 1.0f;  // any positive value should be fine
+	}
 
 	p.Do(curRTWidth);
 	p.Do(curRTHeight);
diff --git a/GPU/GPUState.h b/GPU/GPUState.h
index cc7b0e5424..178fe11d07 100644
--- a/GPU/GPUState.h
+++ b/GPU/GPUState.h
@@ -475,6 +475,7 @@ struct GPUStateCache
 
 	float vpWidth;
 	float vpHeight;
+	float vpDepth;
 
 	u32 curRTWidth;
 	u32 curRTHeight;