Minimum Image convention in BoxDim.h (Difference GPU/CPU-implementation in z-direction)

[akataniel]

Hello HOOMD developers,

I use the function
HOSTDEVICE Scalar3 minImage(const Scalar3& v) const {}
of BoxDim.h in Line 317 quite often in my personal modifications and was asking myself why for GPU-computation we calculate
Scalar img = slow::rint(w.z * m_Linv.z);
in z-direction in Line 325 but for CPU-computation this is missing. For x- and y-direction it is implement with e.g.
int i = int(w.y * m_Linv.y + Scalar(0.5));
in Line 364.
Unluckily I am not that deep into the domain decomposition but still curious if there maybe is a mistake in the code but since the domain can also be decomposed in z-direction.

Kind regards and thank you in advance
Daniel

[joaander]

Yes, of course HOOMD-blue correctly wraps vectors in the z direction on both the CPU and GPU. Many unit tests and all validation tests would fail if this were not the case.

Read the full code carefully and you will see that it performs the same operation (for shifts of up to one box image). The comment explains why the code paths are different:

hoomd-blue/hoomd/BoxDim.h

Lines 343 to 358 in ec0b84c

	// on the cpu, branches are faster than calling rint
	if (m_periodic.z)
	{
	if (w.z >= m_hi.z)
	{
	w.z -= L.z;
	w.y -= L.z * m_yz;
	w.x -= L.z * m_xz;
	}
	else if (w.z < m_lo.z)
	{
	w.z += L.z;
	w.y += L.z * m_yz;
	w.x += L.z * m_xz;
	}
	}