SSE Fixes

CColor <-> CVector4f implicit casting

include/CColor.hpp

@@ -1,7 +1,8 @@
 #ifndef CCOLOR_HPP
 #define CCOLOR_HPP
 
-#include "MathLib.hpp"
+#include "Math.hpp"
+#include "TVectorUnion.hpp"
 #include <iostream>
 
 #if BYTE_ORDER == __ORDER_LITTLE_ENDIAN__
@@ -25,6 +26,8 @@ typedef union
 typedef uint8_t Comp8;
 typedef uint32_t Comp32;
 
+class CVector4f;
+
 class alignas(16) CColor
 {
 public:
@@ -54,6 +57,9 @@ public:
     CColor(Comp32 rgba) { fromRGBA32(rgba); }
     CColor(const Comp8* rgba) { fromRGBA8(rgba[0], rgba[1], rgba[2], rgba[3]); }
 
+    CColor(const CVector4f& other);
+    CColor& operator=(const CVector4f& other);
+
 #if ZE_ATHENA_TYPES
     inline void readRGBA(Athena::io::IStreamReader& reader)
     {
@@ -195,12 +201,16 @@ public:
     }
     inline float magSquared() const
     {
-#if __SSE4_1__
-        TVectorUnion result;
-        result.mVec128 = _mm_dp_ps(mVec128, mVec128, 0xF1);
-        return result.v[0];
-#elif __SSE__
+#if __SSE__
         TVectorUnion result;
+#if __SSE4_1__ || __SSE4_2__
+        if (cpuFeatures().SSE41 || cpuFeatures().SSE42)
+        {
+            result.mVec128 = _mm_dp_ps(mVec128, mVec128, 0xF1);
+            return result.v[0];
+        }
+#endif
+
         result.mVec128 = _mm_mul_ps(mVec128, mVec128);
         return result.v[0] + result.v[1] + result.v[2] + result.v[3];
 #else

include/CVector4f.hpp

@@ -13,6 +13,7 @@
 
 namespace Zeus
 {
+class CColor;
 class alignas(16) CVector4f
 {
     public:
@@ -35,6 +36,7 @@ class alignas(16) CVector4f
 
     CVector4f(float xyzw) {splat(xyzw);}
     CVector4f(float x, float y, float z, float w) {v[0] = x; v[1] = y; v[2] = z; v[3] = w;}
+    CVector4f(const CColor& other);
 #if ZE_ATHENA_TYPES
     CVector4f(Athena::io::IStreamReader& input)
     {
@@ -53,6 +55,7 @@ class alignas(16) CVector4f
         w = 1.0f;
     }
 
+    CVector4f& operator=(const CColor& other);
     inline bool operator ==(const CVector4f& rhs) const
     {
 #if __SSE__

src/CColor.cpp

@@ -1,4 +1,5 @@
 #include "CColor.hpp"
+#include "CVector4f.hpp"
 
 namespace Zeus
 {
@@ -27,6 +28,18 @@ float hueToRgb(float p, float q, float t)
     return p;
 }
 
+CColor::CColor(const CVector4f& other)
+{ r = other.x; g = other.y; b = other.z; a = other.w; }
+
+CColor& CColor::operator=(const CVector4f& other)
+{
+    r = other.x;
+    g = other.y;
+    b = other.z;
+    a = other.w;
+
+    return *this;
+}
 void CColor::fromHSV(float h, float s, float v, float _a)
 {
     int i = int(h * 6);

src/CVector4f.cpp

@@ -1,2 +1,20 @@
 #include "CVector4f.hpp"
+#include "CColor.hpp"
 
+namespace Zeus
+{
+CVector4f::CVector4f(const Zeus::CColor& other)
+    : x(other.r), y(other.g), z(other.b), w(other.a)
+{
+}
+
+CVector4f& CVector4f::operator=(const CColor& other)
+{
+    x = other.r;
+    y = other.g;
+    z = other.b;
+    w = other.a;
+
+    return *this;
+}
+}