Rasterizer: Pre-divide vertex attributes by W

Execute the division-by-W for perspective-correct interpolation of
values in the clipper, moving them out of the rasterization inner loop.

src/video_core/clipper.cpp

@@ -91,10 +91,17 @@ static void InitScreenCoordinates(OutputVertex& vtx)
     viewport.zscale     = float24::FromRawFloat24(registers.viewport_depth_range);
     viewport.offset_z   = float24::FromRawFloat24(registers.viewport_depth_far_plane);
 
+    float24 inv_w = float24::FromFloat32(1.f) / vtx.pos.w;
+    vtx.color *= inv_w;
+    vtx.tc0 *= inv_w;
+    vtx.tc1 *= inv_w;
+    vtx.tc2 *= inv_w;
+    vtx.pos.w = inv_w;
+
     // TODO: Not sure why the viewport width needs to be divided by 2 but the viewport height does not
-    vtx.screenpos[0] = (vtx.pos.x / vtx.pos.w + float24::FromFloat32(1.0)) * viewport.halfsize_x + viewport.offset_x;
-    vtx.screenpos[1] = (vtx.pos.y / vtx.pos.w + float24::FromFloat32(1.0)) * viewport.halfsize_y + viewport.offset_y;
-    vtx.screenpos[2] = viewport.offset_z - vtx.pos.z / vtx.pos.w * viewport.zscale;
+    vtx.screenpos[0] = (vtx.pos.x * inv_w + float24::FromFloat32(1.0)) * viewport.halfsize_x + viewport.offset_x;
+    vtx.screenpos[1] = (vtx.pos.y * inv_w + float24::FromFloat32(1.0)) * viewport.halfsize_y + viewport.offset_y;
+    vtx.screenpos[2] = viewport.offset_z - vtx.pos.z * inv_w * viewport.zscale;
 }
 
 void ProcessTriangle(OutputVertex &v0, OutputVertex &v1, OutputVertex &v2) {

src/video_core/pica.h

@@ -757,6 +757,26 @@ struct float24 {
         return float24::FromFloat32(ToFloat32() - flt.ToFloat32());
     }
 
+    float24& operator *= (const float24& flt) {
+        value *= flt.ToFloat32();
+        return *this;
+    }
+
+    float24& operator /= (const float24& flt) {
+        value /= flt.ToFloat32();
+        return *this;
+    }
+
+    float24& operator += (const float24& flt) {
+        value += flt.ToFloat32();
+        return *this;
+    }
+
+    float24& operator -= (const float24& flt) {
+        value -= flt.ToFloat32();
+        return *this;
+    }
+
     float24 operator - () const {
         return float24::FromFloat32(-ToFloat32());
     }

src/video_core/rasterizer.cpp

@@ -106,10 +106,7 @@ void ProcessTriangle(const VertexShader::OutputVertex& v0,
     int bias1 = IsRightSideOrFlatBottomEdge(vtxpos[1].xy(), vtxpos[2].xy(), vtxpos[0].xy()) ? -1 : 0;
     int bias2 = IsRightSideOrFlatBottomEdge(vtxpos[2].xy(), vtxpos[0].xy(), vtxpos[1].xy()) ? -1 : 0;
 
-    const Math::Vec3<float24> w_inverse = Math::MakeVec(
-            float24::FromFloat32(1.0f) / v0.pos.w,
-            float24::FromFloat32(1.0f) / v1.pos.w,
-            float24::FromFloat32(1.0f) / v2.pos.w);
+    auto w_inverse = Math::MakeVec(v0.pos.w, v1.pos.w, v2.pos.w);
 
     auto textures = registers.GetTextures();
     auto tev_stages = registers.GetTevStages();
@@ -158,7 +155,7 @@ void ProcessTriangle(const VertexShader::OutputVertex& v0,
             //
             // The generalization to three vertices is straightforward in baricentric coordinates.
             auto GetInterpolatedAttribute = [&](float24 attr0, float24 attr1, float24 attr2) {
-                auto attr_over_w = Math::MakeVec(attr0, attr1, attr2) * w_inverse;
+                auto attr_over_w = Math::MakeVec(attr0, attr1, attr2);
                 float24 interpolated_attr_over_w = Math::Dot(attr_over_w, baricentric_coordinates);
                 return interpolated_attr_over_w * interpolated_w_inverse;
             };