MPI.cl File Reference

MPI syncing point. More...

#include "resources/Scripts/types/types.h"

Include dependency graph for MPI.cl:

Functions
__kernel void	mask_planes (__global unsigned int *mpi_mask, const __global int *imove, const __global vec *r_in, unsigned int mpi_rank, vec mpi_planes_orig, vec mpi_planes_dist, usize N)
	Compute the particles sending mask.
__kernel void	copy (__global vec *mpi_r, __global vec *mpi_u, __global vec *mpi_dudt, __global float *mpi_rho, __global float *mpi_drhodt, __global float *mpi_m, const __global vec *r_in, const __global vec *u_in, const __global vec *dudt_in, const __global float *rho_in, const __global float *drhodt_in, const __global float *m, usize N)
	Copy the data.
__kernel void	restore (__global vec *r_in, __global vec *u_in, __global vec *dudt_in, __global float *rho_in, __global float *drhodt_in, __global float *m, __global int *imove, const __global unsigned int *mpi_mask, const __global vec *mpi_r, const __global vec *mpi_u, const __global vec *mpi_dudt, const __global float *mpi_rho, const __global float *mpi_drhodt, const __global float *mpi_m, unsigned int mpi_rank, usize nbuffer, usize N)
	Add the particles received from other processes.
__kernel void	sort (const __global unsigned int *mpi_mask_in, __global unsigned int *mpi_mask, const __global usize *id_sorted, usize N)
	Sort the mask.
__kernel void	drop_planes (__global int *imove, __global vec *r, unsigned int mpi_rank, vec mpi_planes_orig, vec mpi_planes_dist, vec domain_max, usize N)
	Drop particles belonging to different processes.

Detailed Description

MPI syncing point.

Function Documentation

copy()

__kernel void copy	(	__global vec *	mpi_r,
		__global vec *	mpi_u,
		__global vec *	mpi_dudt,
		__global float *	mpi_rho,
		__global float *	mpi_drhodt,
		__global float *	mpi_m,
		const __global vec *	r_in,
		const __global vec *	u_in,
		const __global vec *	dudt_in,
		const __global float *	rho_in,
		const __global float *	drhodt_in,
		const __global float *	m,
		usize	N )

Copy the data.

To make the operation as asynchronous as possible, we are copying all the available data, regardless it is useful or not. That is a bit computationally less efficient, but allows to start transfering data ASAP, while we still operate to neglect the particles

It is intended that the fields copied here are the same subsequently exchanged later with the mpi-sync tool

Parameters

mpi_r	Position \( \mathbf{r} \) MPI copy
mpi_u	Velocity \( \mathbf{u} \) MPI copy
mpi_dudt	Velocity rate of change \( \frac{d \mathbf{u}}{d t} \) MPI copy
mpi_rho	Density \( \rho \) MPI copy
mpi_drhodt	Density rate of change \( \frac{d \rho}{d t} \) MPI copy
mpi_m	Mass \( m \) MPI copy
r_in	Position \( \mathbf{r} \)
u_in	Velocity \( \mathbf{u} \)
dudt_in	Velocity rate of change \( \frac{d \mathbf{u}}{d t} \)
rho_in	Density \( \rho \)
drhodt_in	Density rate of change \( \frac{d \rho}{d t} \)
m	Mass \( m \)
N	Number of particles

drop_planes()

__kernel void drop_planes	(	__global int *	imove,
		__global vec *	r,
		unsigned int	mpi_rank,
		vec	mpi_planes_orig,
		vec	mpi_planes_dist,
		vec	domain_max,
		usize	N )

Drop particles belonging to different processes.

Transfer the particles belonging to different processes, i.e. outside the current process bounding planes, to "the buffer". It should be always bear in mind that the buffer particles are not made available until next time step

Parameters

imove	Moving flags imove > 0 for regular fluid particles imove = 0 for sensors imove < 0 for boundary elements/particles
r	Position \( \mathbf{r} \)
mpi_mask	Output processes mask
mpi_rank	MPI process index
mpi_planes_orig	Center of the first interface (between procs) plane
mpi_planes_dist	Distance between interface planes
domain_max	Top-left-frontal corner of the computational domain
N	Number of particles

mask_planes()

__kernel void mask_planes	(	__global unsigned int *	mpi_mask,
		const __global int *	imove,
		const __global vec *	r_in,
		unsigned int	mpi_rank,
		vec	mpi_planes_orig,
		vec	mpi_planes_dist,
		usize	N )

Compute the particles sending mask.

The sending mask is just an array where it is stored the process each particle will be sent for, just in case it shall be sent at all

This method assumes equally spaced interface planes

No special treatment is required at the bounding processes, since setting the mask with an invalid process is just ignoring it. However, such a situation would indicate problems. Thus, it is strongly recommended to use MPI sync with the computational domain module, in such a way mpi_planes_orig matchs the minimum computational domain point

Parameters

mpi_mask	Output processes mask
imove	Moving flags imove > 0 for regular fluid particles imove = 0 for sensors imove < 0 for boundary elements/particles
r_in	Position \( \mathbf{r} \)
mpi_rank	MPI process index
mpi_planes_orig	Center of the first interface (between procs) plane
mpi_planes_dist	Distance between interface planes
N	Number of particles

restore()

__kernel void restore	(	__global vec *	r_in,
		__global vec *	u_in,
		__global vec *	dudt_in,
		__global float *	rho_in,
		__global float *	drhodt_in,
		__global float *	m,
		__global int *	imove,
		const __global unsigned int *	mpi_mask,
		const __global vec *	mpi_r,
		const __global vec *	mpi_u,
		const __global vec *	mpi_dudt,
		const __global float *	mpi_rho,
		const __global float *	mpi_drhodt,
		const __global float *	mpi_m,
		unsigned int	mpi_rank,
		usize	nbuffer,
		usize	N )

Add the particles received from other processes.

The restored particles has always imove=0 flag

Parameters

imove	Moving flags imove > 0 for regular fluid particles imove = 0 for sensors imove < 0 for boundary elements/particles
r_in	Position \( \mathbf{r} \)
u_in	Velocity \( \mathbf{u} \)
dudt_in	Velocity rate of change \( \frac{d \mathbf{u}}{d t} \)
rho_in	Density \( \rho \)
drhodt_in	Density rate of change \( \frac{d \rho}{d t} \)
m	Mass \( m \)
mpi_mask	Incoming processes mask
mpi_r	Position \( \mathbf{r} \) MPI copy
mpi_u	Velocity \( \mathbf{u} \) MPI copy
mpi_dudt	Velocity rate of change \( \frac{d \mathbf{u}}{d t} \) MPI copy
mpi_rho	Density \( \rho \) MPI copy
mpi_drhodt	Density rate of change \( \frac{d \rho}{d t} \) MPI copy
mpi_m	Mass \( m \) MPI copy
mpi_rank	MPI process index
nbuffer	Number of buffer particles
N	Number of particles

sort()

__kernel void sort	(	const __global unsigned int *	mpi_mask_in,
		__global unsigned int *	mpi_mask,
		const __global usize *	id_sorted,
		usize	N )

Sort the mask.

The processes mask can be used for debugging purposes. To this end we should sort it before.

Parameters

mpi_mask_in	Unsorted processes mask
mpi_mask	Sorted processes mask
id_sorted	Permutations list from the unsorted space to the sorted one.
N	Number of particles