vec4 color32_to_vec4(color32 color)
{
	__m128 colorf = _mm_cvtpu8_ps(_mm_set_pi32(0.0f, color));
	colorf = _mm_shuffle_ps(colorf, colorf, 0x1B);
	return _mm_mul_ps(colorf, _mm_set1_ps(1.0f / 255.0f));
}

color32 vec4_to_color32(__m128 m)
{
	m = _mm_shuffle_ps(m, m, 0x1B);
	return (_m_to_int(_mm_cvtps_pi8(_mm_mul_ps(m, _mm_set1_ps(127.0f)))) * 2) | _m_to_int(_mm_cvtps_pi8(m));
}