/** Copy lattice data with layout described by fieldtype.
* @param r_buffer data destination
* @param s_buffer data source
* @param type field layout type
*/
MDINLINE void halo_dtcopy(void *r_buffer, void *s_buffer, int count, Fieldtype type) {
int i, j;
HALO_TRACE(fprintf(stderr, "%d: halo_dtcopy r_buffer=%p s_buffer=%p blocks=%d stride=%d skip=%d\n",this_node,r_buffer,s_buffer,type->vblocks,type->vstride,type->vskip));
if (type->subtype) {
halo_copy_vector(r_buffer, s_buffer, count, type, type->vflag);
} else {
for (i=0; i<count; i++, s_buffer+=type->extent, r_buffer+=type->extent) {
if (!type->count) {
memcpy(r_buffer,s_buffer,type->extent);
} else {
for (j=0; j<type->count; j++) {
memcpy(r_buffer+type->disps[j],s_buffer+type->disps[j],type->lengths[j]);
}
}
}
}
}
/** Frees datastrutures associated with a halo communicator
* @param hc halo communicator to be released
*/
void release_halo_communication(HaloCommunicator *hc) {
int n;
for (n=0; n<hc->num; n++) {
HALO_TRACE(fprintf(stderr,"%d: freeing %p\n",this_node,&(hc->halo_info[n].datatype)));
MPI_Type_free(&(hc->halo_info[n].datatype));
}
free(hc->halo_info);
}
MDINLINE void halo_copy_vector(void *r_buffer, void *s_buffer, int count, Fieldtype type, int vflag) {
int i, j;
void *dest, *src;
int vblocks = type->vblocks;
int vstride = type->vstride;
int vskip = type->vskip;
int extent = type->extent;
HALO_TRACE(fprintf(stderr, "%d: halo_copy_vector %p %p vblocks=%d vstride=%d vskip=%d extent=%d subtype_extent=%d\n",this_node,r_buffer,s_buffer,vblocks,vstride,vskip,extent,type->subtype->extent));
if (vflag){
vskip *= type->subtype->extent;
}
for (i=0; i<count; i++, s_buffer+=extent, r_buffer+=extent) {
for (j=0, dest=r_buffer, src=s_buffer; j<vblocks; j++, dest+=vskip, src+=vskip) {
halo_dtcopy(dest,src,vstride,type->subtype);
}
}
}
/** Copy lattice data with layout described by fieldtype.
* @param r_buffer data destination
* @param s_buffer data source
* @param type field layout type
*/
MDINLINE void halo_dtcopy(void *r_buffer, void *s_buffer, Fieldtype type) {
int i, j, k;
void *dest, *src;
int vblocks = type->vblocks;
int vstride = type->vstride;
int vskip = type->vskip;
int count = type->count;
int *lens = type->lengths;
int *disps = type->disps;
int extent = type->extent;
HALO_TRACE(fprintf(stderr, "%d: halo comm local copy r_buffer=%p s_buffer=%p\n",this_node,r_buffer,s_buffer));
for (i=0; i<vblocks; i++, r_buffer+=vskip*extent, s_buffer+=vskip*extent) {
for (j=0, dest=r_buffer, src=s_buffer; j<vstride; j++, dest+=extent, src+=extent) {
for (k=0; k<count; k++) {
memcpy(dest+disps[k],src+disps[k],lens[k]);
}
}
}
}
/** Perform communication according to the parallelization scheme
* described by the halo communicator
* @param hc halo communicator describing the parallelization scheme
*/
void halo_communication(HaloCommunicator *hc, void *base) {
int n, comm_type, s_node, r_node;
void *s_buffer, *r_buffer ;
Fieldtype fieldtype;
MPI_Datatype datatype;
MPI_Request request;
MPI_Status status;
HALO_TRACE(fprintf(stderr, "%d: halo_comm base=%p num=%d\n", this_node, base, hc->num)) ;
for (n = 0; n < hc->num; n++) {
HALO_TRACE(fprintf(stderr, "%d: halo_comm round %d\n", this_node, n)) ;
comm_type = hc->halo_info[n].type ;
s_buffer = (char *)base + hc->halo_info[n].s_offset;
r_buffer = (char *)base + hc->halo_info[n].r_offset;
switch (comm_type) {
case HALO_LOCL:
fieldtype = hc->halo_info[n].fieldtype;
halo_dtcopy(r_buffer,s_buffer,1,fieldtype);
break ;
case HALO_SENDRECV:
datatype = hc->halo_info[n].datatype;
s_node = hc->halo_info[n].source_node ;
r_node = hc->halo_info[n].dest_node ;
HALO_TRACE(fprintf(stderr,"%d: halo_comm sendrecv %d to %d (%d) (%p)\n",this_node,s_node,r_node,REQ_HALO_SPREAD,&datatype));
MPI_Sendrecv(s_buffer, 1, datatype, r_node, REQ_HALO_SPREAD,
r_buffer, 1, datatype, s_node, REQ_HALO_SPREAD,
MPI_COMM_WORLD, &status);
break ;
case HALO_SEND:
datatype = hc->halo_info[n].datatype;
fieldtype = hc->halo_info[n].fieldtype;
s_node = hc->halo_info[n].source_node ;
r_node = hc->halo_info[n].dest_node ;
HALO_TRACE(fprintf(stderr,"%d: halo_comm send to %d.\n",this_node,r_node));
MPI_Isend(s_buffer, 1, datatype, r_node, REQ_HALO_SPREAD, MPI_COMM_WORLD, &request);
halo_dtset(r_buffer,0,fieldtype);
MPI_Wait(&request,&status);
break;
case HALO_RECV:
datatype = hc->halo_info[n].datatype;
s_node = hc->halo_info[n].source_node ;
r_node = hc->halo_info[n].dest_node ;
HALO_TRACE(fprintf(stderr,"%d: halo_comm recv from %d.\n",this_node,s_node));
MPI_Irecv(r_buffer, 1, datatype, s_node, REQ_HALO_SPREAD, MPI_COMM_WORLD, &request);
MPI_Wait(&request,&status);
break;
case HALO_OPEN:
fieldtype = hc->halo_info[n].fieldtype;
HALO_TRACE(fprintf(stderr,"%d: halo_comm open boundaries\n",this_node));
/* \todo this does not work for the n_i - <n_i> */
halo_dtset(r_buffer,0,fieldtype);
break;
}
}
}
/** Preparation of the halo parallelization scheme. Sets up the
* necessary datastructures for \ref halo_communication
* @param hc halo communicator beeing created (Input/Output)
* @param lattice lattice the communcation is created for (Input)
* @param fieldtype field layout of the lattice data (Input)
* @param datatype MPI datatype for the lattice data (Input)
*/
void prepare_halo_communication(HaloCommunicator *hc, Lattice *lattice, Fieldtype fieldtype, MPI_Datatype datatype) {
int k, n, dir, lr, cnt, num = 0 ;
int *grid = lattice->grid ;
int *period = lattice->halo_grid ;
for (n=0; n<hc->num; n++) {
MPI_Type_free(&(hc->halo_info[n].datatype));
}
num = 2*3; /* two communications in each space direction */
hc->num = num ;
hc->halo_info = realloc(hc->halo_info,num*sizeof(HaloInfo)) ;
int extent = fieldtype->extent;
cnt = 0 ;
for (dir=0; dir<3; dir++) {
for (lr=0; lr<2; lr++) {
HaloInfo *hinfo = &(hc->halo_info[cnt]) ;
int nblocks = 1 ;
for (k=dir+1;k<3;k++) {
nblocks *= period[k] ;
}
int stride = 1 ;
for (k=0;k<dir;k++) {
stride *= period[k] ;
}
int skip = 1 ;
for (k=0;k<dir+1 && k<2;k++) {
skip *= period[k] ;
}
if (lr==0) {
/* send to left, recv from right */
hinfo->s_offset = extent * stride * 1;
hinfo->r_offset = extent * stride * (grid[dir]+1);
} else {
/* send to right, recv from left */
hinfo->s_offset = extent * stride * grid[dir];
hinfo->r_offset = extent * stride * 0;
}
hinfo->source_node = node_neighbors[2*dir+1-lr];
hinfo->dest_node = node_neighbors[2*dir+lr];
halo_create_field_vector(nblocks, stride, skip, fieldtype, &hinfo->fieldtype);
MPI_Type_vector(nblocks, stride, skip, datatype, &hinfo->datatype);
MPI_Type_commit(&hinfo->datatype);
#ifdef PARTIAL_PERIODIC
if ( !PERIODIC(dir) && (boundary[2*dir+lr] != 0 || boundary[2*dir+1-lr] != 0) ) {
if (node_grid[dir] == 1) {
hinfo->type = HALO_OPEN;
}
else if (lr == 0) {
if (boundary[2*dir+lr] == 1) {
hinfo->type = HALO_RECV;
} else {
hinfo->type = HALO_SEND;
}
}
else {
if (boundary[2*dir+lr] == -1) {
hinfo->type = HALO_RECV;
} else {
hinfo->type = HALO_SEND;
}
}
} else
#endif
{
if (node_grid[dir] == 1) {
hc->halo_info[cnt].type = HALO_LOCL;
}
else {
hc->halo_info[cnt].type = HALO_SENDRECV;
}
}
HALO_TRACE(fprintf(stderr,"%d: prepare_halo_communication dir=%d lr=%d s_offset=%ld r_offset=%ld s_node=%d d_node=%d type=%d\n",this_node,dir,lr,hinfo->s_offset,hinfo->r_offset,hinfo->source_node,hinfo->dest_node,hinfo->type)) ;
cnt++;
}
}
}
/** Perform communication according to the parallelization scheme
* described by the halo communicator
* @param hc halo communicator describing the parallelization scheme
*/
void halo_communication(HaloCommunicator *hc) {
int n, comm_type, s_node, r_node;
void *s_buffer, *r_buffer ;
Fieldtype fieldtype;
MPI_Datatype datatype;
MPI_Request request;
MPI_Status status;
HALO_TRACE(fprintf(stderr, "%d: halo_comm %p (num=%d)\n", this_node, hc, hc->num)) ;
for (n = 0; n < hc->num; n++) {
HALO_TRACE(fprintf(stderr, "%d: halo_comm round %d\n", this_node, n)) ;
comm_type = hc->halo_info[n].type ;
s_buffer = hc->halo_info[n].send_buffer ;
r_buffer = hc->halo_info[n].recv_buffer ;
switch (comm_type) {
case HALO_LOCL:
fieldtype = hc->halo_info[n].fieldtype;
halo_dtcopy(r_buffer,s_buffer,fieldtype);
break ;
case HALO_SENDRECV:
datatype = hc->halo_info[n].datatype;
s_node = hc->halo_info[n].source_node ;
r_node = hc->halo_info[n].dest_node ;
HALO_TRACE(fprintf(stderr,"%d: halo_comm sendrecv %d to %d (%d) (%p)\n",this_node,s_node,r_node,REQ_HALO_SPREAD,&datatype));
MPI_Sendrecv(s_buffer, 1, datatype, r_node, REQ_HALO_SPREAD,
r_buffer, 1, datatype, s_node, REQ_HALO_SPREAD,
MPI_COMM_WORLD, &status);
break ;
case HALO_SEND:
datatype = hc->halo_info[n].datatype;
fieldtype = hc->halo_info[n].fieldtype;
s_node = hc->halo_info[n].source_node ;
r_node = hc->halo_info[n].dest_node ;
HALO_TRACE(fprintf(stderr,"%d: halo_comm send to %d.\n",this_node,r_node));
MPI_Isend(s_buffer, 1, datatype, r_node, REQ_HALO_SPREAD, MPI_COMM_WORLD, &request);
halo_dtset(r_buffer,0,fieldtype);
MPI_Wait(&request,&status);
break;
case HALO_RECV:
datatype = hc->halo_info[n].datatype;
s_node = hc->halo_info[n].source_node ;
r_node = hc->halo_info[n].dest_node ;
HALO_TRACE(fprintf(stderr,"%d: halo_comm recv from %d.\n",this_node,s_node));
MPI_Irecv(r_buffer, 1, datatype, s_node, REQ_HALO_SPREAD, MPI_COMM_WORLD, &request);
MPI_Wait(&request,&status);
break;
case HALO_OPEN:
fieldtype = hc->halo_info[n].fieldtype;
HALO_TRACE(fprintf(stderr,"%d: halo_comm open boundaries\n",this_node));
halo_dtset(r_buffer,0,fieldtype);
break;
}
}
HALO_TRACE(fprintf(stderr, "%d: halo_comm %p finished\n", this_node, hc));
}
