class CVector3D
{
    public:
        float X, Y, Z;

    public:
        CVector3D() : X(0.0f), Y(0.0f), Z(0.0f) {}
        CVector3D(float x, float y, float z) : X(x), Y(y), Z(z) {}

        int operator!() const;

        float& operator[](int index) { return *((&X)+index); }
        const float& operator[](int index) const { return *((&X)+index); }

        // vector equality (testing float equality, so please be careful if necessary)
        bool operator==(const CVector3D &vector) const
        {
            return (X == vector.X && Y == vector.Y && Z == vector.Z);
        }

        bool operator!=(const CVector3D& vector) const
        {
            return !operator==(vector);
        }

        CVector3D operator+(const CVector3D& vector) const
        {
            return CVector3D(X + vector.X, Y + vector.Y, Z + vector.Z);
        }

        CVector3D& operator+=(const CVector3D& vector)
        {
            X += vector.X;
            Y += vector.Y;
            Z += vector.Z;
            return *this;
        }

        CVector3D operator-(const CVector3D& vector) const
        {
            return CVector3D(X - vector.X, Y - vector.Y, Z - vector.Z);
        }

        CVector3D& operator-=(const CVector3D& vector)
        {
            X -= vector.X;
            Y -= vector.Y;
            Z -= vector.Z;
            return *this;
        }

        CVector3D operator*(float value) const
        {
            return CVector3D(X * value, Y * value, Z * value);
        }

        CVector3D& operator*=(float value)
        {
            X *= value;
            Y *= value;
            Z *= value;
            return *this;
        }

        CVector3D operator-() const
        {
            return CVector3D(-X, -Y, -Z);
        }

    public:
        float Dot (const CVector3D &vector) const
        {
            return ( X * vector.X +
                     Y * vector.Y +
                     Z * vector.Z );
        }

        CVector3D Cross (const CVector3D &vector) const
        {
            CVector3D Temp;
            Temp.X = (Y * vector.Z) - (Z * vector.Y);
            Temp.Y = (Z * vector.X) - (X * vector.Z);
            Temp.Z = (X * vector.Y) - (Y * vector.X);
            return Temp;
        }

        float Length () const;
        float LengthSquared () const;
        void Normalize ();
        CVector3D Normalized () const;

        // Returns 3 element array of floats, e.g. for glVertex3fv
        const float* GetFloatArray() const { return &X; }
};

#include <stdint.h>
#include <algorithm>

using RGBColor = CVector3D;
using u8 = uint8_t;

template <typename T>
inline T clamp(T value, T min, T max)
{
    if (value <= min) return min;
    else if (value >= max) return max;
    else return value;
}

// SColor4ub: structure for packed RGBA colors
struct SColor4ub
{
    u8 R;
    u8 G;
    u8 B;
    u8 A;

    SColor4ub() { }
    SColor4ub(u8 _r, u8 _g, u8 _b, u8 _a) : R(_r), G(_g), B(_b), A(_a) { }
};

SColor4ub fallback_ConvertRGBColorTo4ub(const RGBColor& src)
{
    SColor4ub result;
    result.R = clamp(static_cast<int>(src.X * 255), 0, 255);
    result.G = clamp(static_cast<int>(src.Y * 255), 0, 255);
    result.B = clamp(static_cast<int>(src.Z * 255), 0, 255);
    result.A = 255;
    return result;
}

#include <emmintrin.h>

SColor4ub sse_ConvertRGBColorTo4ub(const RGBColor& src)
{
    const __m128 zero = _mm_setzero_ps();
    const __m128 _255 = _mm_set_ss(255.0f);
    __m128 r = _mm_load_ss(&src.X);
    __m128 g = _mm_load_ss(&src.Y);
    __m128 b = _mm_load_ss(&src.Z);

    // C = min(255, 255*max(C, 0)) ( == clamp(255*C, 0, 255) )
    r = _mm_max_ss(r, zero);
    g = _mm_max_ss(g, zero);
    b = _mm_max_ss(b, zero);

    r = _mm_mul_ss(r, _255);
    g = _mm_mul_ss(g, _255);
    b = _mm_mul_ss(b, _255);

    r = _mm_min_ss(r, _255);
    g = _mm_min_ss(g, _255);
    b = _mm_min_ss(b, _255);

    // convert to integer and combine channels using bit logic
    int ri = _mm_cvtss_si32(r);
    int gi = _mm_cvtss_si32(g);
    int bi = _mm_cvtss_si32(b);

    return SColor4ub(ri, gi, bi, 0xFF);
}

#include <iostream>

int main()
{
    SColor4ub a;
    SColor4ub b;

    for (int i = 0; i < 500000000; ++i)
    {
        RGBColor col(0.4f, 0.5f, i / 50.0f);
        a = sse_ConvertRGBColorTo4ub(col);
        b.B += a.B;
    }

    if (b.B > 50)
        return 0;
    return 1;
}