2023-02-05 07:34:23 +01:00

327 lines
8.5 KiB
C

/*
* Copyright (C) 2022 NHR@FAU, University Erlangen-Nuremberg.
* All rights reserved. This file is part of nusif-solver.
* Use of this source code is governed by a MIT style
* license that can be found in the LICENSE file.
*/
#include <stddef.h>
#include <stdio.h>
#include <stdlib.h>
#include "comm.h"
// subroutines local to this module
static int sizeOfRank(int rank, int size, int N)
{
return N / size + ((N % size > rank) ? 1 : 0);
}
static void setupCommunication(Comm* c, int direction, int layer)
{
MPI_Datatype type;
size_t dblsize = sizeof(double);
int imaxLocal = c->imaxLocal;
int jmaxLocal = c->jmaxLocal;
int sizes[NDIMS];
int subSizes[NDIMS];
int starts[NDIMS];
int offset = 0;
sizes[IDIM] = imaxLocal + 2;
sizes[JDIM] = jmaxLocal + 2;
if (layer == HALO) {
offset = 1;
}
switch (direction) {
case LEFT:
subSizes[IDIM] = 1;
subSizes[JDIM] = jmaxLocal;
starts[IDIM] = 1 - offset;
starts[JDIM] = 1;
break;
case RIGHT:
subSizes[IDIM] = 1;
subSizes[JDIM] = jmaxLocal;
starts[IDIM] = imaxLocal + offset;
starts[JDIM] = 1;
break;
case BOTTOM:
subSizes[IDIM] = imaxLocal;
subSizes[JDIM] = 1;
starts[IDIM] = 1;
starts[JDIM] = 1 - offset;
break;
case TOP:
subSizes[IDIM] = imaxLocal;
subSizes[JDIM] = 1;
starts[IDIM] = 1;
starts[JDIM] = jmaxLocal + offset;
break;
}
MPI_Type_create_subarray(NDIMS,
sizes,
subSizes,
starts,
MPI_ORDER_C,
MPI_DOUBLE,
&type);
MPI_Type_commit(&type);
if (layer == HALO) {
c->rbufferTypes[direction] = type;
} else if (layer == BULK) {
c->sbufferTypes[direction] = type;
}
}
static void assembleResult(Comm* c,
double* src,
double* dst,
int imaxLocal[],
int jmaxLocal[],
int offset[],
int jmax,
int imax)
{
MPI_Request* requests;
int numRequests = 1;
if (c->rank == 0) {
numRequests = c->size + 1;
} else {
numRequests = 1;
}
requests = (MPI_Request*)malloc(numRequests * sizeof(MPI_Request));
/* all ranks send their bulk array */
MPI_Datatype bulkType;
int oldSizes[NDIMS] = { c->jmaxLocal + 2, c->imaxLocal + 2 };
int newSizes[NDIMS] = { c->jmaxLocal, c->imaxLocal };
int starts[NDIMS] = { 1, 1 };
MPI_Type_create_subarray(NDIMS,
oldSizes,
newSizes,
starts,
MPI_ORDER_C,
MPI_DOUBLE,
&bulkType);
MPI_Type_commit(&bulkType);
MPI_Isend(src, 1, bulkType, 0, 0, c->comm, &requests[0]);
/* rank 0 assembles the subdomains */
if (c->rank == 0) {
for (int i = 0; i < c->size; i++) {
MPI_Datatype domainType;
int oldSizes[NDIMS] = { jmax, imax };
int newSizes[NDIMS] = { jmaxLocal[i], imaxLocal[i] };
int starts[NDIMS] = { offset[i * NDIMS + JDIM], offset[i * NDIMS + IDIM] };
MPI_Type_create_subarray(NDIMS,
oldSizes,
newSizes,
starts,
MPI_ORDER_C,
MPI_DOUBLE,
&domainType);
MPI_Type_commit(&domainType);
MPI_Irecv(dst, 1, domainType, i, 0, c->comm, &requests[i + 1]);
}
}
MPI_Waitall(numRequests, requests, MPI_STATUSES_IGNORE);
}
static int sum(int* sizes, int position)
{
int sum = 0;
for (int i = 0; i < position; i++) {
sum += sizes[i];
}
return sum;
}
// exported subroutines
void commReduction(double* v, int op)
{
if (op == MAX) {
MPI_Allreduce(MPI_IN_PLACE, v, 1, MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD);
} else if (op == SUM) {
MPI_Allreduce(MPI_IN_PLACE, v, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
}
}
int commIsBoundary(Comm* c, int direction)
{
switch (direction) {
case LEFT:
return c->coords[IDIM] == 0;
break;
case RIGHT:
return c->coords[IDIM] == (c->dims[IDIM] - 1);
break;
case BOTTOM:
return c->coords[JDIM] == 0;
break;
case TOP:
return c->coords[JDIM] == (c->dims[JDIM] - 1);
break;
}
return 0;
}
void commExchange(Comm* c, double* grid)
{
int counts[NDIRS] = { 1, 1, 1, 1 };
MPI_Aint displs[NDIRS] = { 0, 0, 0, 0 };
MPI_Neighbor_alltoallw(grid,
counts,
displs,
c->sbufferTypes,
grid,
counts,
displs,
c->rbufferTypes,
c->comm);
}
void commShift(Comm* c, double* f, double* g)
{
MPI_Request requests[4] = { MPI_REQUEST_NULL,
MPI_REQUEST_NULL,
MPI_REQUEST_NULL,
MPI_REQUEST_NULL };
/* shift G */
/* receive ghost cells from bottom neighbor */
MPI_Irecv(g,
1,
c->rbufferTypes[BOTTOM],
c->neighbours[BOTTOM],
0,
c->comm,
&requests[0]);
/* send ghost cells to top neighbor */
MPI_Isend(g, 1, c->sbufferTypes[TOP], c->neighbours[TOP], 0, c->comm, &requests[1]);
/* shift F */
/* receive ghost cells from left neighbor */
MPI_Irecv(f, 1, c->rbufferTypes[LEFT], c->neighbours[LEFT], 1, c->comm, &requests[2]);
/* send ghost cells to right neighbor */
MPI_Isend(f,
1,
c->sbufferTypes[RIGHT],
c->neighbours[RIGHT],
1,
c->comm,
&requests[3]);
MPI_Waitall(4, requests, MPI_STATUSES_IGNORE);
}
void commCollectResult(Comm* c,
double* ug,
double* vg,
double* pg,
double* u,
double* v,
double* p,
int jmax,
int imax)
{
int offset[c->size * NDIMS];
int imaxLocal[c->size];
int jmaxLocal[c->size];
MPI_Gather(&c->imaxLocal, 1, MPI_INT, imaxLocal, 1, MPI_INT, 0, MPI_COMM_WORLD);
MPI_Gather(&c->jmaxLocal, 1, MPI_INT, jmaxLocal, 1, MPI_INT, 0, MPI_COMM_WORLD);
if (c->rank == 0) {
for (int i = 0; i < c->size; i++) {
int coords[NDIMS];
MPI_Cart_coords(c->comm, i, NDIMS, coords);
offset[i * NDIMS + IDIM] = sum(imaxLocal, coords[IDIM]);
offset[i * NDIMS + JDIM] = sum(jmaxLocal, coords[JDIM]);
printf("Rank: %d, Coords(j,i): %d %d, Size(j,i): %d %d "
"Offset(j,i): %d %d\n",
i,
coords[JDIM],
coords[IDIM],
jmaxLocal[i],
imaxLocal[i],
offset[i * NDIMS + JDIM],
offset[i * NDIMS + IDIM]);
}
}
/* collect P */
assembleResult(c, p, pg, imaxLocal, jmaxLocal, offset, jmax, imax);
/* collect U */
assembleResult(c, u, ug, imaxLocal, jmaxLocal, offset, jmax, imax);
/* collect V */
assembleResult(c, v, vg, imaxLocal, jmaxLocal, offset, jmax, imax);
}
void commPrintConfig(Comm* c)
{
fflush(stdout);
MPI_Barrier(MPI_COMM_WORLD);
if (commIsMaster(c)) {
printf("Communication setup:\n");
}
for (int i = 0; i < c->size; i++) {
if (i == c->rank) {
printf("\tRank %d of %d\n", c->rank, c->size);
printf("\tNeighbours (bottom, top, left, right): %d %d, %d, %d\n",
c->neighbours[BOTTOM],
c->neighbours[TOP],
c->neighbours[LEFT],
c->neighbours[RIGHT]);
printf("\tCoordinates (j,i) %d %d\n", c->coords[JDIM], c->coords[IDIM]);
printf("\tLocal domain size (j,i) %dx%d\n", c->jmaxLocal, c->imaxLocal);
fflush(stdout);
}
}
MPI_Barrier(MPI_COMM_WORLD);
}
void commInit(Comm* c, int jmax, int imax)
{
/* setup communication */
MPI_Comm_rank(MPI_COMM_WORLD, &(c->rank));
MPI_Comm_size(MPI_COMM_WORLD, &(c->size));
int dims[NDIMS] = { 0, 0 };
int periods[NDIMS] = { 0, 0 };
MPI_Dims_create(c->size, NDIMS, dims);
MPI_Cart_create(MPI_COMM_WORLD, NDIMS, dims, periods, 0, &c->comm);
MPI_Cart_shift(c->comm, IDIM, 1, &c->neighbours[LEFT], &c->neighbours[RIGHT]);
MPI_Cart_shift(c->comm, JDIM, 1, &c->neighbours[BOTTOM], &c->neighbours[TOP]);
MPI_Cart_get(c->comm, NDIMS, c->dims, periods, c->coords);
c->imaxLocal = sizeOfRank(c->rank, dims[IDIM], imax);
c->jmaxLocal = sizeOfRank(c->rank, dims[JDIM], jmax);
// setup buffer types for communication
setupCommunication(c, LEFT, BULK);
setupCommunication(c, LEFT, HALO);
setupCommunication(c, RIGHT, BULK);
setupCommunication(c, RIGHT, HALO);
setupCommunication(c, BOTTOM, BULK);
setupCommunication(c, BOTTOM, HALO);
setupCommunication(c, TOP, BULK);
setupCommunication(c, TOP, HALO);
}