From 6cacfad010b05cd9d79056a81c4dc7d19ca57273 Mon Sep 17 00:00:00 2001
From: Jules Blok <jules.blok@gmail.com>
Date: Thu, 30 Oct 2014 02:08:15 +0100
Subject: [PATCH] GeometryShader: Transform the projection within the geometry
 shader.

Reduces the amount of data transferred through uniforms.
The shearing transformation is reduced to a single multiplication/addition for optimization.
---
 Source/Core/VideoCommon/ConstantManager.h     |  2 +-
 Source/Core/VideoCommon/GeometryShaderGen.cpp | 12 ++++++-----
 Source/Core/VideoCommon/ShaderGenCommon.h     |  4 ++--
 .../Core/VideoCommon/VertexShaderManager.cpp  | 20 ++++---------------
 4 files changed, 14 insertions(+), 24 deletions(-)

diff --git a/Source/Core/VideoCommon/ConstantManager.h b/Source/Core/VideoCommon/ConstantManager.h
index 1c3ad55778..e31806bbd1 100644
--- a/Source/Core/VideoCommon/ConstantManager.h
+++ b/Source/Core/VideoCommon/ConstantManager.h
@@ -43,6 +43,6 @@ struct VertexShaderConstants
 	float4 normalmatrices[32];
 	float4 posttransformmatrices[64];
 	float4 depthparams;
-	float4 stereoprojection[8];
+	float4 stereooffset;
 };
 
diff --git a/Source/Core/VideoCommon/GeometryShaderGen.cpp b/Source/Core/VideoCommon/GeometryShaderGen.cpp
index f016acd6f4..6a39e2ca1d 100644
--- a/Source/Core/VideoCommon/GeometryShaderGen.cpp
+++ b/Source/Core/VideoCommon/GeometryShaderGen.cpp
@@ -70,16 +70,18 @@ static inline void GenerateGeometryShader(T& out, u32 components, API_TYPE ApiTy
 	out.Write("flat out int layer;\n");
 
 	out.Write("void main()\n{\n");
-	if (!g_ActiveConfig.backend_info.bSupportsGSInstancing)
+	if (g_ActiveConfig.backend_info.bSupportsGSInstancing)
+		out.Write("\tlayer = gl_InvocationID;\n");
+	else
 		out.Write("\tfor (layer = 0; layer < %d; ++layer) {\n", g_ActiveConfig.bStereo ? 2 : 1);
+	out.Write("\tgl_Layer = layer;\n");
+	out.Write("\tvec4 stereoproj = "I_PROJECTION"[0];\n");
+	out.Write("\tstereoproj[2] += "I_STEREOOFFSET"[layer] * stereoproj[0];\n");
 	out.Write("\tfor (int i = 0; i < gl_in.length(); ++i) {\n");
 	out.Write("\t\to = v[i];\n");
-	if (g_ActiveConfig.backend_info.bSupportsGSInstancing)
-		out.Write("\t\tlayer = gl_InvocationID;\n");
 	if (g_ActiveConfig.bStereo)
-		out.Write("\t\to.pos = float4(dot(" I_STEREOPROJECTION"[layer * 4 + 0], v[i].rawpos), dot(" I_STEREOPROJECTION"[layer * 4 + 1], v[i].rawpos), dot(" I_STEREOPROJECTION"[layer * 4 + 2], v[i].rawpos), dot(" I_STEREOPROJECTION"[layer * 4 + 3], v[i].rawpos)); \n");
+		out.Write("\t\to.pos = float4(dot(stereoproj, v[i].rawpos), dot(" I_PROJECTION"[1], v[i].rawpos), dot(" I_PROJECTION"[2], v[i].rawpos), dot(" I_PROJECTION"[3], v[i].rawpos)); \n");
 	out.Write("\t\tgl_Position = o.pos;\n");
-	out.Write("\t\tgl_Layer = layer;\n");
 	out.Write("\t\tEmitVertex();\n");
 	out.Write("\t}\n");
 	out.Write("\tEndPrimitive();\n");
diff --git a/Source/Core/VideoCommon/ShaderGenCommon.h b/Source/Core/VideoCommon/ShaderGenCommon.h
index 11a5ccafa8..5e9e24e7fa 100644
--- a/Source/Core/VideoCommon/ShaderGenCommon.h
+++ b/Source/Core/VideoCommon/ShaderGenCommon.h
@@ -239,7 +239,7 @@ private:
 #define I_NORMALMATRICES        "cnmtx"
 #define I_POSTTRANSFORMMATRICES "cpostmtx"
 #define I_DEPTHPARAMS           "cDepth" // farZ, zRange
-#define I_STEREOPROJECTION      "csproj"
+#define I_STEREOOFFSET          "csoffset"
 
 static const char s_shader_uniforms[] =
 	"\tfloat4 " I_POSNORMALMATRIX"[6];\n"
@@ -251,4 +251,4 @@ static const char s_shader_uniforms[] =
 	"\tfloat4 " I_NORMALMATRICES"[32];\n"
 	"\tfloat4 " I_POSTTRANSFORMMATRICES"[64];\n"
 	"\tfloat4 " I_DEPTHPARAMS";\n"
-	"\tfloat4 " I_STEREOPROJECTION"[8];\n";
+	"\tfloat4 " I_STEREOOFFSET";\n";
diff --git a/Source/Core/VideoCommon/VertexShaderManager.cpp b/Source/Core/VideoCommon/VertexShaderManager.cpp
index e40eb92152..5d65a49186 100644
--- a/Source/Core/VideoCommon/VertexShaderManager.cpp
+++ b/Source/Core/VideoCommon/VertexShaderManager.cpp
@@ -515,25 +515,13 @@ void VertexShaderManager::SetConstants()
 
 		if (g_ActiveConfig.bStereo && xfmem.projection.type == GX_PERSPECTIVE)
 		{
-			Matrix44 projMtx;
-			Matrix44::Set(projMtx, g_fProjectionMatrix);
-
-			Matrix44 leftShearMtx, rightShearMtx;
-			Matrix44::Shear(leftShearMtx, g_ActiveConfig.iStereoSeparation / (200.0f * g_ActiveConfig.iStereoFocalLength));
-			Matrix44::Shear(rightShearMtx, -g_ActiveConfig.iStereoSeparation / (200.0f * g_ActiveConfig.iStereoFocalLength));
-
-			Matrix44 leftProjMtx, rightProjMtx, leftCorrectedMtx, rightCorrectedMtx;
-			Matrix44::Multiply(projMtx, leftShearMtx, leftProjMtx);
-			Matrix44::Multiply(s_viewportCorrection, leftProjMtx, leftCorrectedMtx);
-			Matrix44::Multiply(projMtx, rightShearMtx, rightProjMtx);
-			Matrix44::Multiply(s_viewportCorrection, rightProjMtx, rightCorrectedMtx);
-			memcpy(constants.stereoprojection, leftCorrectedMtx.data, 4*16);
-			memcpy(constants.stereoprojection + 4, rightCorrectedMtx.data, 4*16);
+			float offset = g_ActiveConfig.iStereoSeparation / (200.0f * g_ActiveConfig.iStereoFocalLength);
+			constants.stereooffset[0] = offset;
+			constants.stereooffset[1] = -offset;
 		}
 		else
 		{
-			memcpy(constants.stereoprojection, constants.projection, 4 * 16);
-			memcpy(constants.stereoprojection + 4, constants.projection, 4 * 16);
+			constants.stereooffset[0] = constants.stereooffset[1] = 0;
 		}
 
 		dirty = true;