From 73566ff7a990cdfe8d8f023997b57942dc785fc4 Mon Sep 17 00:00:00 2001
From: Subv <subv2112@gmail.com>
Date: Sun, 11 Jun 2017 11:55:35 -0500
Subject: [PATCH] SwRasterizer: Flip the vertex quaternions before clipping (if
 necessary).

---
 src/common/vector_math.h                   |  2 +-
 src/video_core/swrasterizer/clipper.cpp    | 11 ++++++++++
 src/video_core/swrasterizer/rasterizer.cpp | 24 ++++------------------
 3 files changed, 16 insertions(+), 21 deletions(-)

diff --git a/src/common/vector_math.h b/src/common/vector_math.h
index c7a461a1e..d0fe0e405 100644
--- a/src/common/vector_math.h
+++ b/src/common/vector_math.h
@@ -462,7 +462,7 @@ public:
         z -= other.z;
         w -= other.w;
     }
-    template <typename Q = T, class = typename std::enable_if<std::is_signed<Q>::value>::type>
+    template <typename Q = T>
     Vec4<decltype(-T{})> operator-() const {
         return MakeVec(-x, -y, -z, -w);
     }
diff --git a/src/video_core/swrasterizer/clipper.cpp b/src/video_core/swrasterizer/clipper.cpp
index 6fb923756..7537689b7 100644
--- a/src/video_core/swrasterizer/clipper.cpp
+++ b/src/video_core/swrasterizer/clipper.cpp
@@ -95,6 +95,17 @@ void ProcessTriangle(const OutputVertex& v0, const OutputVertex& v1, const Outpu
     static const size_t MAX_VERTICES = 9;
     static_vector<Vertex, MAX_VERTICES> buffer_a = {v0, v1, v2};
     static_vector<Vertex, MAX_VERTICES> buffer_b;
+
+    auto FlipQuaternionIfOpposite = [](auto& a, const auto& b) {
+        if (Math::Dot(a, b) < float24::Zero())
+            a = -a;
+    };
+
+    // Flip the quaternions if they are opposite to prevent interpolating them over the wrong
+    // direction.
+    FlipQuaternionIfOpposite(buffer_a[1].quat, buffer_a[0].quat);
+    FlipQuaternionIfOpposite(buffer_a[2].quat, buffer_a[0].quat);
+
     auto* output_list = &buffer_a;
     auto* input_list = &buffer_b;
 
diff --git a/src/video_core/swrasterizer/rasterizer.cpp b/src/video_core/swrasterizer/rasterizer.cpp
index 2c804b6e7..76f793c86 100644
--- a/src/video_core/swrasterizer/rasterizer.cpp
+++ b/src/video_core/swrasterizer/rasterizer.cpp
@@ -362,13 +362,6 @@ std::tuple<Math::Vec4<u8>, Math::Vec4<u8>> ComputeFragmentsColors(const Math::Qu
     };
 }
 
-static bool AreQuaternionsOpposite(Math::Vec4<Pica::float24> qa, Math::Vec4<Pica::float24> qb) {
-    Math::Vec4f a{ qa.x.ToFloat32(), qa.y.ToFloat32(), qa.z.ToFloat32(), qa.w.ToFloat32() };
-    Math::Vec4f b{ qb.x.ToFloat32(), qb.y.ToFloat32(), qb.z.ToFloat32(), qb.w.ToFloat32() };
-
-    return (Math::Dot(a, b) < 0.f);
-}
-
 MICROPROFILE_DEFINE(GPU_Rasterization, "GPU", "Rasterization", MP_RGB(50, 50, 240));
 
 /**
@@ -462,15 +455,6 @@ static void ProcessTriangleInternal(const Vertex& v0, const Vertex& v1, const Ve
     int bias2 =
         IsRightSideOrFlatBottomEdge(vtxpos[2].xy(), vtxpos[0].xy(), vtxpos[1].xy()) ? -1 : 0;
 
-    // Flip the quaternions if they are opposite to prevent interpolating them over the wrong direction.
-    auto v1_quat = v1.quat;
-    auto v2_quat = v2.quat;
-
-    if (AreQuaternionsOpposite(v0.quat, v1.quat))
-        v1_quat = v1_quat * float24::FromFloat32(-1.0f);
-    if (AreQuaternionsOpposite(v0.quat, v2.quat))
-        v2_quat = v2_quat * float24::FromFloat32(-1.0f);
-
     auto w_inverse = Math::MakeVec(v0.pos.w, v1.pos.w, v2.pos.w);
 
     auto textures = regs.texturing.GetTextures();
@@ -571,11 +555,11 @@ static void ProcessTriangleInternal(const Vertex& v0, const Vertex& v1, const Ve
 
             Math::Quaternion<float> normquat{
                 {
-                    GetInterpolatedAttribute(v0.quat.x, v1_quat.x, v2_quat.x).ToFloat32(),
-                    GetInterpolatedAttribute(v0.quat.y, v1_quat.y, v2_quat.y).ToFloat32(),
-                    GetInterpolatedAttribute(v0.quat.z, v1_quat.z, v2_quat.z).ToFloat32()
+                    GetInterpolatedAttribute(v0.quat.x, v1.quat.x, v2.quat.x).ToFloat32(),
+                    GetInterpolatedAttribute(v0.quat.y, v1.quat.y, v2.quat.y).ToFloat32(),
+                    GetInterpolatedAttribute(v0.quat.z, v1.quat.z, v2.quat.z).ToFloat32()
                 },
-                GetInterpolatedAttribute(v0.quat.w, v1_quat.w, v2_quat.w).ToFloat32(),
+                GetInterpolatedAttribute(v0.quat.w, v1.quat.w, v2.quat.w).ToFloat32(),
             };
 
             Math::Vec3<float> fragment_position{