class FoutBitP {
    public:
        FourBitP(float lX, float lY, float uX, float uY)
            : lx_(lx), ly_(ly), ux_(ux), uy_(uy) {
        }

        __device__ Match evaluate(Patch patch) const {
#define TEX_COORDS(A, B) float A ## B = patch.origin.B + patch.size.B * A ## B ## _
            TEX_COORDS(l, x);
            TEX_COORDS(l, y);
            TEX_COORDS(u, x);
            TEX_COORDS(u, y);
#undef TEX_COORDS
#define MIDDLE(A) float m ## A = (l ## A + u ## A) / 2
            MIDDLE(l);
            MIDDLE(u);
#undef MIDDLE
            // Exactly 8 texture fetches are require for a 4BitP evalutaion:
#define FETCH(X, Y) float s ## X ## Y = tex2D(image, X ## x, Y ## y)
            FETCH(l, u); FETCH(m, u); FETCH(u, u);
            FETCH(l, m);              FETCH(u, m);
            FETCH(l, l); FETCH(m, l); FETCH(u, l);
#undef FETCH
            float left   = sml - sll - smu + slu,
                  right  = sul - sml - suu + smu,
                  top    = sum - slm - suu + slu,
                  bottom = sul - sll - sum + slm;
            return (top < bottom) << 1 + (left < right);
        }
    private:
        float lx_, ly_, ux_, uy_;
};