SSE_TrilinearInteroplation.cpp

// Scalar Trilinear Interoplation Lambda using SSE. 
auto trilerp_v = [&](int i0, int j0, int k0, float z, float y, float x) -> auto
{

	vec4 aa(this->getdata(i0, j0, k0), this->getdata(i0, j0 + 1, k0), this->getdata(i0 + 1, j0, k0), this->getdata(i0 + 1, j0 + 1, k0));
	vec4 bb(this->getdata(i0, j0, k0 + 1), this->getdata(i0, j0 + 1, k0 + 1), this->getdata(i0 + 1, j0, k0 + 1), this->getdata(i0 + 1, j0 + 1, k0 + 1));
	vec4 coeff_z(z); vec4 ccoeff_z(1.0f - z);
	__m128 L_a = _mm_fmadd_ps(ccoeff_z.sa, aa.sa, _mm_mul_ps(coeff_z.sa, bb.sa));

	__m128 cc = _mm_shuffle_ps(L_a, L_a, _MM_SHUFFLE(1, 0, 1, 0));  // (z0,z1,|z0,z1)
	__m128 dd = _mm_shuffle_ps(L_a, L_a, _MM_SHUFFLE(2, 1, 2, 1));  // (z1,z2,|z1,z2)
	vec4 coeff_y(y); vec4 ccoeff_y(1.0f - y);
	__m128 L_b = _mm_fmadd_ps(ccoeff_y.sa, cc, _mm_mul_ps(coeff_y.sa, dd));

	__m128 ee = _mm_shuffle_ps(L_b, L_b, _MM_SHUFFLE(0, 0, 0, 0)); // (y0,|y1,y0,y0)
	__m128 ff = _mm_shuffle_ps(L_b, L_b, _MM_SHUFFLE(1, 1, 1, 1)); // (y1,|y1,y0,y1)
	vec4 coeff_x(x); vec4 ccoeff_x(1.0f - x);
	__m128 L_C = _mm_fmadd_ps(ccoeff_x.sa, ee, _mm_mul_ps(coeff_x.sa, ff));

	return vec4(L_C).x; 

};