PetscErrorCode VecStashCreate_Private(MPI_Comm comm,PetscInt bs,VecStash *stash)
{
PetscErrorCode ierr;
PetscInt max,*opt,nopt;
PetscBool flg;
PetscFunctionBegin;
/* Require 2 tags, get the second using PetscCommGetNewTag() */
stash->comm = comm;
ierr = PetscCommGetNewTag(stash->comm,&stash->tag1);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(stash->comm,&stash->tag2);CHKERRQ(ierr);
ierr = MPI_Comm_size(stash->comm,&stash->size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(stash->comm,&stash->rank);CHKERRQ(ierr);
nopt = stash->size;
ierr = PetscMalloc1(nopt,&opt);CHKERRQ(ierr);
ierr = PetscOptionsGetIntArray(NULL,"-vecstash_initial_size",opt,&nopt,&flg);CHKERRQ(ierr);
if (flg) {
if (nopt == 1) max = opt[0];
else if (nopt == stash->size) max = opt[stash->rank];
else if (stash->rank < nopt) max = opt[stash->rank];
else max = 0; /* use default */
stash->umax = max;
} else {
stash->umax = 0;
}
ierr = PetscFree(opt);CHKERRQ(ierr);
if (bs <= 0) bs = 1;
stash->bs = bs;
stash->nmax = 0;
stash->oldnmax = 0;
stash->n = 0;
stash->reallocs = -1;
stash->idx = 0;
stash->array = 0;
stash->send_waits = 0;
stash->recv_waits = 0;
stash->send_status = 0;
stash->nsends = 0;
stash->nrecvs = 0;
stash->svalues = 0;
stash->rvalues = 0;
stash->rmax = 0;
stash->nprocs = 0;
stash->nprocessed = 0;
stash->donotstash = PETSC_FALSE;
stash->ignorenegidx = PETSC_FALSE;
PetscFunctionReturn(0);
}
PetscErrorCode MatStashCreate_Private(MPI_Comm comm,PetscInt bs,MatStash *stash)
{
PetscErrorCode ierr;
PetscInt max,*opt,nopt,i;
PetscTruth flg;
PetscFunctionBegin;
/* Require 2 tags,get the second using PetscCommGetNewTag() */
stash->comm = comm;
ierr = PetscCommGetNewTag(stash->comm,&stash->tag1);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(stash->comm,&stash->tag2);CHKERRQ(ierr);
ierr = MPI_Comm_size(stash->comm,&stash->size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(stash->comm,&stash->rank);CHKERRQ(ierr);
ierr = PetscMalloc(2*stash->size*sizeof(PetscMPIInt),&stash->flg_v);CHKERRQ(ierr);
for (i=0; i<2*stash->size; i++) stash->flg_v[i] = -1;
nopt = stash->size;
ierr = PetscMalloc(nopt*sizeof(PetscInt),&opt);CHKERRQ(ierr);
ierr = PetscOptionsGetIntArray(PETSC_NULL,"-matstash_initial_size",opt,&nopt,&flg);CHKERRQ(ierr);
if (flg) {
if (nopt == 1) max = opt[0];
else if (nopt == stash->size) max = opt[stash->rank];
else if (stash->rank < nopt) max = opt[stash->rank];
else max = 0; /* Use default */
stash->umax = max;
} else {
stash->umax = 0;
}
ierr = PetscFree(opt);CHKERRQ(ierr);
if (bs <= 0) bs = 1;
stash->bs = bs;
stash->nmax = 0;
stash->oldnmax = 0;
stash->n = 0;
stash->reallocs = -1;
stash->space_head = 0;
stash->space = 0;
stash->send_waits = 0;
stash->recv_waits = 0;
stash->send_status = 0;
stash->nsends = 0;
stash->nrecvs = 0;
stash->svalues = 0;
stash->rvalues = 0;
stash->rindices = 0;
stash->nprocessed = 0;
PetscFunctionReturn(0);
}
/*@C
PetscObjectGetNewTag - Gets a unique new tag from a PETSc object. All
processors that share the object MUST call this routine EXACTLY the same
number of times. This tag should only be used with the current objects
communicator; do NOT use it with any other MPI communicator.
Collective on PetscObject
Input Parameter:
. obj - the PETSc object; this must be cast with a (PetscObject), for example,
PetscObjectGetNewTag((PetscObject)mat,&tag);
Output Parameter:
. tag - the new tag
Level: developer
Concepts: tag^getting
Concepts: message tag^getting
Concepts: MPI message tag^getting
.seealso: PetscCommGetNewTag()
@*/
PetscErrorCode PetscObjectGetNewTag(PetscObject obj,PetscMPIInt *tag)
{
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscCommGetNewTag(obj->comm,tag);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode PetscCommBuildTwoSided_Ibarrier(MPI_Comm comm,PetscMPIInt count,MPI_Datatype dtype,PetscInt nto,const PetscMPIInt *toranks,const void *todata,PetscInt *nfrom,PetscMPIInt **fromranks,void *fromdata)
{
PetscErrorCode ierr;
PetscMPIInt nrecvs,tag,unitbytes,done;
PetscInt i;
char *tdata;
MPI_Request *sendreqs,barrier;
PetscSegBuffer segrank,segdata;
PetscFunctionBegin;
ierr = PetscCommGetNewTag(comm,&tag);CHKERRQ(ierr);
ierr = MPI_Type_size(dtype,&unitbytes);CHKERRQ(ierr);
tdata = (char*)todata;
ierr = PetscMalloc(nto*sizeof(MPI_Request),&sendreqs);CHKERRQ(ierr);
for (i=0; i<nto; i++) {
ierr = MPI_Issend((void*)(tdata+count*unitbytes*i),count,dtype,toranks[i],tag,comm,sendreqs+i);CHKERRQ(ierr);
}
ierr = PetscSegBufferCreate(sizeof(PetscMPIInt),4,&segrank);CHKERRQ(ierr);
ierr = PetscSegBufferCreate(unitbytes,4*count,&segdata);CHKERRQ(ierr);
nrecvs = 0;
barrier = MPI_REQUEST_NULL;
for (done=0; !done; ) {
PetscMPIInt flag;
MPI_Status status;
ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag,comm,&flag,&status);CHKERRQ(ierr);
if (flag) { /* incoming message */
PetscMPIInt *recvrank;
void *buf;
ierr = PetscSegBufferGet(&segrank,1,&recvrank);CHKERRQ(ierr);
ierr = PetscSegBufferGet(&segdata,count,&buf);CHKERRQ(ierr);
*recvrank = status.MPI_SOURCE;
ierr = MPI_Recv(buf,count,dtype,status.MPI_SOURCE,tag,comm,MPI_STATUS_IGNORE);CHKERRQ(ierr);
nrecvs++;
}
if (barrier == MPI_REQUEST_NULL) {
PetscMPIInt sent,nsends;
ierr = PetscMPIIntCast(nto,&nsends);CHKERRQ(ierr);
ierr = MPI_Testall(nsends,sendreqs,&sent,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
if (sent) {
ierr = MPI_Ibarrier(comm,&barrier);CHKERRQ(ierr);
ierr = PetscFree(sendreqs);CHKERRQ(ierr);
}
} else {
ierr = MPI_Test(&barrier,&done,MPI_STATUS_IGNORE);CHKERRQ(ierr);
}
}
*nfrom = nrecvs;
ierr = PetscSegBufferExtractAlloc(&segrank,fromranks);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segrank);CHKERRQ(ierr);
ierr = PetscSegBufferExtractAlloc(&segdata,fromdata);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segdata);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@C
PetscGatherMessageLengths2 - Computes info about messages that a MPI-node will receive,
including (from-id,length) pairs for each message. Same functionality as PetscGatherMessageLengths()
except it takes TWO ilenths and output TWO olengths.
Collective on MPI_Comm
Input Parameters:
+ comm - Communicator
. nsends - number of messages that are to be sent.
. nrecvs - number of messages being received
- ilengths1, ilengths2 - array of integers of length sizeof(comm)
a non zero ilengths[i] represent a message to i of length ilengths[i]
Output Parameters:
+ onodes - list of node-ids from which messages are expected
- olengths1, olengths2 - corresponding message lengths
Level: developer
Concepts: mpi utility
Notes:
With this info, the correct MPI_Irecv() can be posted with the correct
from-id, with a buffer with the right amount of memory required.
The calling function deallocates the memory in onodes and olengths
To determine nrecevs, one can use PetscGatherNumberOfMessages()
.seealso: PetscGatherMessageLengths() and PetscGatherNumberOfMessages()
@*/
PetscErrorCode PetscGatherMessageLengths2(MPI_Comm comm,PetscMPIInt nsends,PetscMPIInt nrecvs,const PetscMPIInt ilengths1[],const PetscMPIInt ilengths2[],PetscMPIInt **onodes,PetscMPIInt **olengths1,PetscMPIInt **olengths2)
{
PetscErrorCode ierr;
PetscMPIInt size,tag,i,j,*buf_s = NULL,*buf_r = NULL,*buf_j = NULL;
MPI_Request *s_waits = NULL,*r_waits = NULL;
MPI_Status *w_status = NULL;
PetscFunctionBegin;
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(comm,&tag);CHKERRQ(ierr);
/* cannot use PetscMalloc5() because r_waits and s_waits must be contiguous for the call to MPI_Waitall() */
ierr = PetscMalloc4(nrecvs+nsends,&r_waits,2*nrecvs,&buf_r,2*nsends,&buf_s,nrecvs+nsends,&w_status);CHKERRQ(ierr);
s_waits = r_waits + nrecvs;
/* Post the Irecv to get the message length-info */
ierr = PetscMalloc1(nrecvs+1,olengths1);CHKERRQ(ierr);
ierr = PetscMalloc1(nrecvs+1,olengths2);CHKERRQ(ierr);
for (i=0; i<nrecvs; i++) {
buf_j = buf_r + (2*i);
ierr = MPI_Irecv(buf_j,2,MPI_INT,MPI_ANY_SOURCE,tag,comm,r_waits+i);CHKERRQ(ierr);
}
/* Post the Isends with the message length-info */
for (i=0,j=0; i<size; ++i) {
if (ilengths1[i]) {
buf_j = buf_s + (2*j);
buf_j[0] = *(ilengths1+i);
buf_j[1] = *(ilengths2+i);
ierr = MPI_Isend(buf_j,2,MPI_INT,i,tag,comm,s_waits+j);CHKERRQ(ierr);
j++;
}
}
if (j != nsends) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"j %d not equal to expected number of sends %d\n",j,nsends);
/* Post waits on sends and receivs */
if (nrecvs+nsends) {ierr = MPI_Waitall(nrecvs+nsends,r_waits,w_status);CHKERRQ(ierr);}
/* Pack up the received data */
ierr = PetscMalloc1(nrecvs+1,onodes);CHKERRQ(ierr);
for (i=0; i<nrecvs; ++i) {
(*onodes)[i] = w_status[i].MPI_SOURCE;
buf_j = buf_r + (2*i);
(*olengths1)[i] = buf_j[0];
(*olengths2)[i] = buf_j[1];
}
ierr = PetscFree4(r_waits,buf_r,buf_s,w_status);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode PetscCommBuildTwoSidedFReq_Reference(MPI_Comm comm,PetscMPIInt count,MPI_Datatype dtype,PetscMPIInt nto,const PetscMPIInt *toranks,const void *todata,
PetscMPIInt *nfrom,PetscMPIInt **fromranks,void *fromdata,PetscMPIInt ntags,MPI_Request **toreqs,MPI_Request **fromreqs,
PetscErrorCode (*send)(MPI_Comm,const PetscMPIInt[],PetscMPIInt,PetscMPIInt,void*,MPI_Request[],void*),
PetscErrorCode (*recv)(MPI_Comm,const PetscMPIInt[],PetscMPIInt,void*,MPI_Request[],void*),void *ctx)
{
PetscErrorCode ierr;
PetscMPIInt i,*tag;
MPI_Aint lb,unitbytes;
MPI_Request *sendreq,*recvreq;
PetscFunctionBegin;
ierr = PetscMalloc1(ntags,&tag);CHKERRQ(ierr);
if (ntags > 0) {
ierr = PetscCommDuplicate(comm,&comm,&tag[0]);CHKERRQ(ierr);
}
for (i=1; i<ntags; i++) {
ierr = PetscCommGetNewTag(comm,&tag[i]);CHKERRQ(ierr);
}
/* Perform complete initial rendezvous */
ierr = PetscCommBuildTwoSided(comm,count,dtype,nto,toranks,todata,nfrom,fromranks,fromdata);CHKERRQ(ierr);
ierr = PetscMalloc1(nto*ntags,&sendreq);CHKERRQ(ierr);
ierr = PetscMalloc1(*nfrom*ntags,&recvreq);CHKERRQ(ierr);
ierr = MPI_Type_get_extent(dtype,&lb,&unitbytes);CHKERRQ(ierr);
if (lb != 0) SETERRQ1(comm,PETSC_ERR_SUP,"Datatype with nonzero lower bound %ld\n",(long)lb);
for (i=0; i<nto; i++) {
PetscMPIInt k;
for (k=0; k<ntags; k++) sendreq[i*ntags+k] = MPI_REQUEST_NULL;
ierr = (*send)(comm,tag,i,toranks[i],((char*)todata)+count*unitbytes*i,sendreq+i*ntags,ctx);CHKERRQ(ierr);
}
for (i=0; i<*nfrom; i++) {
void *header = (*(char**)fromdata) + count*unitbytes*i;
PetscMPIInt k;
for (k=0; k<ntags; k++) recvreq[i*ntags+k] = MPI_REQUEST_NULL;
ierr = (*recv)(comm,tag,(*fromranks)[i],header,recvreq+i*ntags,ctx);CHKERRQ(ierr);
}
ierr = PetscFree(tag);CHKERRQ(ierr);
ierr = PetscCommDestroy(&comm);CHKERRQ(ierr);
*toreqs = sendreq;
*fromreqs = recvreq;
PetscFunctionReturn(0);
}
PetscErrorCode DMPlexVTKWritePartition_ASCII(DM dm, FILE *fp)
{
MPI_Comm comm;
PetscInt numCells = 0, cellHeight;
PetscInt numLabelCells, cMax, cStart, cEnd, c;
PetscMPIInt numProcs, rank, proc, tag;
PetscBool hasLabel;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(comm, &tag);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = DMPlexGetVTKCellHeight(dm, &cellHeight);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, cellHeight, &cStart, &cEnd);CHKERRQ(ierr);
ierr = DMPlexGetHybridBounds(dm, &cMax, NULL, NULL, NULL);CHKERRQ(ierr);
if (cMax >= 0) cEnd = PetscMin(cEnd, cMax);
ierr = DMPlexGetStratumSize(dm, "vtk", 1, &numLabelCells);CHKERRQ(ierr);
hasLabel = numLabelCells > 0 ? PETSC_TRUE : PETSC_FALSE;
for (c = cStart; c < cEnd; ++c) {
if (hasLabel) {
PetscInt value;
ierr = DMPlexGetLabelValue(dm, "vtk", c, &value);CHKERRQ(ierr);
if (value != 1) continue;
}
++numCells;
}
if (!rank) {
for (c = 0; c < numCells; ++c) {ierr = PetscFPrintf(comm, fp, "%d\n", rank);CHKERRQ(ierr);}
for (proc = 1; proc < numProcs; ++proc) {
MPI_Status status;
ierr = MPI_Recv(&numCells, 1, MPIU_INT, proc, tag, comm, &status);CHKERRQ(ierr);
for (c = 0; c < numCells; ++c) {ierr = PetscFPrintf(comm, fp, "%d\n", proc);CHKERRQ(ierr);}
}
} else {
ierr = MPI_Send(&numCells, 1, MPIU_INT, 0, tag, comm);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
/*@C
PetscGatherMessageLengths - Computes info about messages that a MPI-node will receive,
including (from-id,length) pairs for each message.
Collective on MPI_Comm
Input Parameters:
+ comm - Communicator
. nsends - number of messages that are to be sent.
. nrecvs - number of messages being received
- ilengths - an array of integers of length sizeof(comm)
a non zero ilengths[i] represent a message to i of length ilengths[i]
Output Parameters:
+ onodes - list of node-ids from which messages are expected
- olengths - corresponding message lengths
Level: developer
Concepts: mpi utility
Notes:
With this info, the correct MPI_Irecv() can be posted with the correct
from-id, with a buffer with the right amount of memory required.
The calling function deallocates the memory in onodes and olengths
To determine nrecevs, one can use PetscGatherNumberOfMessages()
.seealso: PetscGatherNumberOfMessages()
@*/
PetscErrorCode PETSC_DLLEXPORT PetscGatherMessageLengths(MPI_Comm comm,PetscMPIInt nsends,PetscMPIInt nrecvs,const PetscMPIInt ilengths[],PetscMPIInt **onodes,PetscMPIInt **olengths)
{
PetscErrorCode ierr;
PetscMPIInt size,tag,i,j;
MPI_Request *s_waits = PETSC_NULL,*r_waits = PETSC_NULL;
MPI_Status *w_status = PETSC_NULL;
PetscFunctionBegin;
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(comm,&tag);CHKERRQ(ierr);
/* cannot use PetscMalloc3() here because in the call to MPI_Waitall() they MUST be contiguous */
ierr = PetscMalloc2(nrecvs+nsends,MPI_Request,&r_waits,nrecvs+nsends,MPI_Status,&w_status);CHKERRQ(ierr);
s_waits = r_waits+nrecvs;
/* Post the Irecv to get the message length-info */
ierr = PetscMalloc(nrecvs*sizeof(PetscMPIInt),olengths);CHKERRQ(ierr);
for (i=0; i<nrecvs; i++) {
ierr = MPI_Irecv((*olengths)+i,1,MPI_INT,MPI_ANY_SOURCE,tag,comm,r_waits+i);CHKERRQ(ierr);
}
/* Post the Isends with the message length-info */
for (i=0,j=0; i<size; ++i) {
if (ilengths[i]) {
ierr = MPI_Isend((void*)(ilengths+i),1,MPI_INT,i,tag,comm,s_waits+j);CHKERRQ(ierr);
j++;
}
}
/* Post waits on sends and receivs */
if (nrecvs+nsends) {ierr = MPI_Waitall(nrecvs+nsends,r_waits,w_status);CHKERRQ(ierr);}
/* Pack up the received data */
ierr = PetscMalloc(nrecvs*sizeof(PetscMPIInt),onodes);CHKERRQ(ierr);
for (i=0; i<nrecvs; ++i) {
(*onodes)[i] = w_status[i].MPI_SOURCE;
}
ierr = PetscFree2(r_waits,w_status);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode PetscCommBuildTwoSided_Allreduce(MPI_Comm comm,PetscMPIInt count,MPI_Datatype dtype,PetscInt nto,const PetscMPIInt *toranks,const void *todata,PetscInt *nfrom,PetscMPIInt **fromranks,void *fromdata)
{
PetscErrorCode ierr;
PetscMPIInt size,*iflags,nrecvs,tag,unitbytes,*franks;
PetscInt i;
char *tdata,*fdata;
MPI_Request *reqs,*sendreqs;
MPI_Status *statuses;
PetscFunctionBegin;
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = PetscMalloc(size*sizeof(*iflags),&iflags);CHKERRQ(ierr);
ierr = PetscMemzero(iflags,size*sizeof(*iflags));CHKERRQ(ierr);
for (i=0; i<nto; i++) iflags[toranks[i]] = 1;
ierr = PetscGatherNumberOfMessages(comm,iflags,NULL,&nrecvs);CHKERRQ(ierr);
ierr = PetscFree(iflags);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(comm,&tag);CHKERRQ(ierr);
ierr = MPI_Type_size(dtype,&unitbytes);CHKERRQ(ierr);
ierr = PetscMalloc(nrecvs*count*unitbytes,&fdata);CHKERRQ(ierr);
tdata = (char*)todata;
ierr = PetscMalloc2(nto+nrecvs,MPI_Request,&reqs,nto+nrecvs,MPI_Status,&statuses);CHKERRQ(ierr);
sendreqs = reqs + nrecvs;
for (i=0; i<nrecvs; i++) {
ierr = MPI_Irecv((void*)(fdata+count*unitbytes*i),count,dtype,MPI_ANY_SOURCE,tag,comm,reqs+i);CHKERRQ(ierr);
}
for (i=0; i<nto; i++) {
ierr = MPI_Isend((void*)(tdata+count*unitbytes*i),count,dtype,toranks[i],tag,comm,sendreqs+i);CHKERRQ(ierr);
}
ierr = MPI_Waitall(nto+nrecvs,reqs,statuses);CHKERRQ(ierr);
ierr = PetscMalloc(nrecvs*sizeof(PetscMPIInt),&franks);CHKERRQ(ierr);
for (i=0; i<nrecvs; i++) franks[i] = statuses[i].MPI_SOURCE;
ierr = PetscFree2(reqs,statuses);CHKERRQ(ierr);
*nfrom = nrecvs;
*fromranks = franks;
*(void**)fromdata = fdata;
PetscFunctionReturn(0);
}
/*
MatStashCreate_Private - Creates a stash,currently used for all the parallel
matrix implementations. The stash is where elements of a matrix destined
to be stored on other processors are kept until matrix assembly is done.
This is a simple minded stash. Simply adds entries to end of stash.
Input Parameters:
comm - communicator, required for scatters.
bs - stash block size. used when stashing blocks of values
Output Parameters:
stash - the newly created stash
*/
PetscErrorCode MatStashCreate_Private(MPI_Comm comm,PetscInt bs,MatStash *stash)
{
PetscErrorCode ierr;
PetscInt max,*opt,nopt,i;
PetscBool flg;
PetscFunctionBegin;
/* Require 2 tags,get the second using PetscCommGetNewTag() */
stash->comm = comm;
ierr = PetscCommGetNewTag(stash->comm,&stash->tag1);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(stash->comm,&stash->tag2);CHKERRQ(ierr);
ierr = MPI_Comm_size(stash->comm,&stash->size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(stash->comm,&stash->rank);CHKERRQ(ierr);
ierr = PetscMalloc1(2*stash->size,&stash->flg_v);CHKERRQ(ierr);
for (i=0; i<2*stash->size; i++) stash->flg_v[i] = -1;
nopt = stash->size;
ierr = PetscMalloc1(nopt,&opt);CHKERRQ(ierr);
ierr = PetscOptionsGetIntArray(NULL,NULL,"-matstash_initial_size",opt,&nopt,&flg);CHKERRQ(ierr);
if (flg) {
if (nopt == 1) max = opt[0];
else if (nopt == stash->size) max = opt[stash->rank];
else if (stash->rank < nopt) max = opt[stash->rank];
else max = 0; /* Use default */
stash->umax = max;
} else {
stash->umax = 0;
}
ierr = PetscFree(opt);CHKERRQ(ierr);
if (bs <= 0) bs = 1;
stash->bs = bs;
stash->nmax = 0;
stash->oldnmax = 0;
stash->n = 0;
stash->reallocs = -1;
stash->space_head = 0;
stash->space = 0;
stash->send_waits = 0;
stash->recv_waits = 0;
stash->send_status = 0;
stash->nsends = 0;
stash->nrecvs = 0;
stash->svalues = 0;
stash->rvalues = 0;
stash->rindices = 0;
stash->nprocessed = 0;
stash->reproduce = PETSC_FALSE;
stash->blocktype = MPI_DATATYPE_NULL;
ierr = PetscOptionsGetBool(NULL,NULL,"-matstash_reproduce",&stash->reproduce,NULL);CHKERRQ(ierr);
#if !defined(PETSC_HAVE_MPIUNI)
ierr = PetscOptionsGetBool(NULL,NULL,"-matstash_legacy",&flg,NULL);CHKERRQ(ierr);
if (!flg) {
stash->ScatterBegin = MatStashScatterBegin_BTS;
stash->ScatterGetMesg = MatStashScatterGetMesg_BTS;
stash->ScatterEnd = MatStashScatterEnd_BTS;
stash->ScatterDestroy = MatStashScatterDestroy_BTS;
} else {
#endif
stash->ScatterBegin = MatStashScatterBegin_Ref;
stash->ScatterGetMesg = MatStashScatterGetMesg_Ref;
stash->ScatterEnd = MatStashScatterEnd_Ref;
stash->ScatterDestroy = NULL;
#if !defined(PETSC_HAVE_MPIUNI)
}
#endif
PetscFunctionReturn(0);
}
PetscErrorCode DMPlexVTKWriteCells_ASCII(DM dm, FILE *fp, PetscInt *totalCells)
{
MPI_Comm comm;
DMLabel label;
IS globalVertexNumbers = NULL;
const PetscInt *gvertex;
PetscInt dim;
PetscInt numCorners = 0, totCorners = 0, maxCorners, *corners;
PetscInt numCells = 0, totCells = 0, maxCells, cellHeight;
PetscInt numLabelCells, maxLabelCells, cMax, cStart, cEnd, c, vStart, vEnd, v;
PetscMPIInt numProcs, rank, proc, tag;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(comm, &tag);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr);
ierr = DMPlexGetVTKCellHeight(dm, &cellHeight);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, cellHeight, &cStart, &cEnd);CHKERRQ(ierr);
ierr = DMPlexGetDepthStratum(dm, 0, &vStart, &vEnd);CHKERRQ(ierr);
ierr = DMPlexGetHybridBounds(dm, &cMax, NULL, NULL, NULL);CHKERRQ(ierr);
if (cMax >= 0) cEnd = PetscMin(cEnd, cMax);
ierr = DMPlexGetLabel(dm, "vtk", &label);CHKERRQ(ierr);
ierr = DMPlexGetStratumSize(dm, "vtk", 1, &numLabelCells);CHKERRQ(ierr);
ierr = MPI_Allreduce(&numLabelCells, &maxLabelCells, 1, MPIU_INT, MPI_MAX, comm);CHKERRQ(ierr);
if (!maxLabelCells) label = NULL;
for (c = cStart; c < cEnd; ++c) {
PetscInt *closure = NULL;
PetscInt closureSize, value;
if (label) {
ierr = DMLabelGetValue(label, c, &value);CHKERRQ(ierr);
if (value != 1) continue;
}
ierr = DMPlexGetTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
for (v = 0; v < closureSize*2; v += 2) {
if ((closure[v] >= vStart) && (closure[v] < vEnd)) ++numCorners;
}
ierr = DMPlexRestoreTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
++numCells;
}
maxCells = numCells;
ierr = MPI_Reduce(&numCells, &totCells, 1, MPIU_INT, MPI_SUM, 0, comm);CHKERRQ(ierr);
ierr = MPI_Reduce(&numCells, &maxCells, 1, MPIU_INT, MPI_MAX, 0, comm);CHKERRQ(ierr);
ierr = MPI_Reduce(&numCorners, &totCorners, 1, MPIU_INT, MPI_SUM, 0, comm);CHKERRQ(ierr);
ierr = MPI_Reduce(&numCorners, &maxCorners, 1, MPIU_INT, MPI_MAX, 0, comm);CHKERRQ(ierr);
ierr = DMPlexGetVertexNumbering(dm, &globalVertexNumbers);CHKERRQ(ierr);
ierr = ISGetIndices(globalVertexNumbers, &gvertex);CHKERRQ(ierr);
ierr = PetscMalloc1(maxCells, &corners);CHKERRQ(ierr);
ierr = PetscFPrintf(comm, fp, "CELLS %d %d\n", totCells, totCorners+totCells);CHKERRQ(ierr);
if (!rank) {
PetscInt *remoteVertices;
int *vertices;
ierr = PetscMalloc1(maxCorners, &vertices);CHKERRQ(ierr);
for (c = cStart, numCells = 0; c < cEnd; ++c) {
PetscInt *closure = NULL;
PetscInt closureSize, value, nC = 0;
if (label) {
ierr = DMLabelGetValue(label, c, &value);CHKERRQ(ierr);
if (value != 1) continue;
}
ierr = DMPlexGetTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
for (v = 0; v < closureSize*2; v += 2) {
if ((closure[v] >= vStart) && (closure[v] < vEnd)) {
const PetscInt gv = gvertex[closure[v] - vStart];
vertices[nC++] = gv < 0 ? -(gv+1) : gv;
}
}
ierr = DMPlexRestoreTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
corners[numCells++] = nC;
ierr = PetscFPrintf(comm, fp, "%d ", nC);CHKERRQ(ierr);
ierr = DMPlexInvertCell(dim, nC, vertices);CHKERRQ(ierr);
for (v = 0; v < nC; ++v) {
ierr = PetscFPrintf(comm, fp, " %d", vertices[v]);CHKERRQ(ierr);
}
ierr = PetscFPrintf(comm, fp, "\n");CHKERRQ(ierr);
}
if (numProcs > 1) {ierr = PetscMalloc1(maxCorners+maxCells, &remoteVertices);CHKERRQ(ierr);}
for (proc = 1; proc < numProcs; ++proc) {
MPI_Status status;
ierr = MPI_Recv(&numCorners, 1, MPIU_INT, proc, tag, comm, &status);CHKERRQ(ierr);
ierr = MPI_Recv(remoteVertices, numCorners, MPIU_INT, proc, tag, comm, &status);CHKERRQ(ierr);
for (c = 0; c < numCorners;) {
PetscInt nC = remoteVertices[c++];
for (v = 0; v < nC; ++v, ++c) {
vertices[v] = remoteVertices[c];
}
ierr = DMPlexInvertCell(dim, nC, vertices);CHKERRQ(ierr);
ierr = PetscFPrintf(comm, fp, "%d ", nC);CHKERRQ(ierr);
for (v = 0; v < nC; ++v) {
ierr = PetscFPrintf(comm, fp, " %d", vertices[v]);CHKERRQ(ierr);
}
ierr = PetscFPrintf(comm, fp, "\n");CHKERRQ(ierr);
}
}
if (numProcs > 1) {ierr = PetscFree(remoteVertices);CHKERRQ(ierr);}
ierr = PetscFree(vertices);CHKERRQ(ierr);
} else {
PetscInt *localVertices, numSend = numCells+numCorners, k = 0;
ierr = PetscMalloc1(numSend, &localVertices);CHKERRQ(ierr);
for (c = cStart, numCells = 0; c < cEnd; ++c) {
PetscInt *closure = NULL;
PetscInt closureSize, value, nC = 0;
if (label) {
ierr = DMLabelGetValue(label, c, &value);CHKERRQ(ierr);
if (value != 1) continue;
}
ierr = DMPlexGetTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
for (v = 0; v < closureSize*2; v += 2) {
if ((closure[v] >= vStart) && (closure[v] < vEnd)) {
const PetscInt gv = gvertex[closure[v] - vStart];
closure[nC++] = gv < 0 ? -(gv+1) : gv;
}
}
corners[numCells++] = nC;
localVertices[k++] = nC;
for (v = 0; v < nC; ++v, ++k) {
localVertices[k] = closure[v];
}
ierr = DMPlexRestoreTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
}
if (k != numSend) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB, "Invalid number of vertices to send %d should be %d", k, numSend);
ierr = MPI_Send(&numSend, 1, MPIU_INT, 0, tag, comm);CHKERRQ(ierr);
ierr = MPI_Send(localVertices, numSend, MPIU_INT, 0, tag, comm);CHKERRQ(ierr);
ierr = PetscFree(localVertices);CHKERRQ(ierr);
}
ierr = ISRestoreIndices(globalVertexNumbers, &gvertex);CHKERRQ(ierr);
ierr = PetscFPrintf(comm, fp, "CELL_TYPES %d\n", totCells);CHKERRQ(ierr);
if (!rank) {
PetscInt cellType;
for (c = 0; c < numCells; ++c) {
ierr = DMPlexVTKGetCellType(dm, dim, corners[c], &cellType);CHKERRQ(ierr);
ierr = PetscFPrintf(comm, fp, "%d\n", cellType);CHKERRQ(ierr);
}
for (proc = 1; proc < numProcs; ++proc) {
MPI_Status status;
ierr = MPI_Recv(&numCells, 1, MPIU_INT, proc, tag, comm, &status);CHKERRQ(ierr);
ierr = MPI_Recv(corners, numCells, MPIU_INT, proc, tag, comm, &status);CHKERRQ(ierr);
for (c = 0; c < numCells; ++c) {
ierr = DMPlexVTKGetCellType(dm, dim, corners[c], &cellType);CHKERRQ(ierr);
ierr = PetscFPrintf(comm, fp, "%d\n", cellType);CHKERRQ(ierr);
}
}
} else {
ierr = MPI_Send(&numCells, 1, MPIU_INT, 0, tag, comm);CHKERRQ(ierr);
ierr = MPI_Send(corners, numCells, MPIU_INT, 0, tag, comm);CHKERRQ(ierr);
}
ierr = PetscFree(corners);CHKERRQ(ierr);
*totalCells = totCells;
PetscFunctionReturn(0);
}
PetscErrorCode DMPlexVTKWriteSection_ASCII(DM dm, PetscSection section, PetscSection globalSection, Vec v, FILE *fp, PetscInt enforceDof, PetscInt precision, PetscReal scale)
{
MPI_Comm comm;
const MPI_Datatype mpiType = MPIU_SCALAR;
PetscScalar *array;
PetscInt numDof = 0, maxDof;
PetscInt numLabelCells, cellHeight, cMax, cStart, cEnd, numLabelVertices, vMax, vStart, vEnd, pStart, pEnd, p;
PetscMPIInt numProcs, rank, proc, tag;
PetscBool hasLabel;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
PetscValidHeaderSpecific(dm,DM_CLASSID,1);
PetscValidHeaderSpecific(v,VEC_CLASSID,4);
if (precision < 0) precision = 6;
ierr = PetscCommGetNewTag(comm, &tag);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = PetscSectionGetChart(section, &pStart, &pEnd);CHKERRQ(ierr);
/* VTK only wants the values at cells or vertices */
ierr = DMPlexGetVTKCellHeight(dm, &cellHeight);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, cellHeight, &cStart, &cEnd);CHKERRQ(ierr);
ierr = DMPlexGetDepthStratum(dm, 0, &vStart, &vEnd);CHKERRQ(ierr);
ierr = DMPlexGetHybridBounds(dm, &cMax, NULL, NULL, &vMax);CHKERRQ(ierr);
if (cMax >= 0) cEnd = PetscMin(cEnd, cMax);
if (vMax >= 0) vEnd = PetscMin(vEnd, vMax);
pStart = PetscMax(PetscMin(cStart, vStart), pStart);
pEnd = PetscMin(PetscMax(cEnd, vEnd), pEnd);
ierr = DMPlexGetStratumSize(dm, "vtk", 1, &numLabelCells);CHKERRQ(ierr);
ierr = DMPlexGetStratumSize(dm, "vtk", 2, &numLabelVertices);CHKERRQ(ierr);
hasLabel = numLabelCells > 0 || numLabelVertices > 0 ? PETSC_TRUE : PETSC_FALSE;
for (p = pStart; p < pEnd; ++p) {
/* Reject points not either cells or vertices */
if (((p < cStart) || (p >= cEnd)) && ((p < vStart) || (p >= vEnd))) continue;
if (hasLabel) {
PetscInt value;
if (((p >= cStart) && (p < cEnd) && numLabelCells) ||
((p >= vStart) && (p < vEnd) && numLabelVertices)) {
ierr = DMPlexGetLabelValue(dm, "vtk", p, &value);CHKERRQ(ierr);
if (value != 1) continue;
}
}
ierr = PetscSectionGetDof(section, p, &numDof);CHKERRQ(ierr);
if (numDof) break;
}
ierr = MPI_Allreduce(&numDof, &maxDof, 1, MPIU_INT, MPI_MAX, comm);CHKERRQ(ierr);
enforceDof = PetscMax(enforceDof, maxDof);
ierr = VecGetArray(v, &array);CHKERRQ(ierr);
if (!rank) {
char formatString[8];
ierr = PetscSNPrintf(formatString, 8, "%%.%de", precision);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
/* Here we lose a way to filter points by keeping them out of the Numbering */
PetscInt dof, off, goff, d;
/* Reject points not either cells or vertices */
if (((p < cStart) || (p >= cEnd)) && ((p < vStart) || (p >= vEnd))) continue;
if (hasLabel) {
PetscInt value;
if (((p >= cStart) && (p < cEnd) && numLabelCells) ||
((p >= vStart) && (p < vEnd) && numLabelVertices)) {
ierr = DMPlexGetLabelValue(dm, "vtk", p, &value);CHKERRQ(ierr);
if (value != 1) continue;
}
}
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(globalSection, p, &goff);CHKERRQ(ierr);
if (dof && goff >= 0) {
for (d = 0; d < dof; d++) {
if (d > 0) {
ierr = PetscFPrintf(comm, fp, " ");CHKERRQ(ierr);
}
ierr = PetscFPrintf(comm, fp, formatString, PetscRealPart(array[off+d])*scale);CHKERRQ(ierr);
}
for (d = dof; d < enforceDof; d++) {
ierr = PetscFPrintf(comm, fp, " 0.0");CHKERRQ(ierr);
}
ierr = PetscFPrintf(comm, fp, "\n");CHKERRQ(ierr);
}
}
for (proc = 1; proc < numProcs; ++proc) {
PetscScalar *remoteValues;
PetscInt size = 0, d;
MPI_Status status;
ierr = MPI_Recv(&size, 1, MPIU_INT, proc, tag, comm, &status);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &remoteValues);CHKERRQ(ierr);
ierr = MPI_Recv(remoteValues, size, mpiType, proc, tag, comm, &status);CHKERRQ(ierr);
for (p = 0; p < size/maxDof; ++p) {
for (d = 0; d < maxDof; ++d) {
if (d > 0) {
ierr = PetscFPrintf(comm, fp, " ");CHKERRQ(ierr);
}
ierr = PetscFPrintf(comm, fp, formatString, PetscRealPart(remoteValues[p*maxDof+d])*scale);CHKERRQ(ierr);
}
for (d = maxDof; d < enforceDof; ++d) {
ierr = PetscFPrintf(comm, fp, " 0.0");CHKERRQ(ierr);
}
ierr = PetscFPrintf(comm, fp, "\n");CHKERRQ(ierr);
}
ierr = PetscFree(remoteValues);CHKERRQ(ierr);
}
} else {
PetscScalar *localValues;
PetscInt size, k = 0;
ierr = PetscSectionGetStorageSize(section, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &localValues);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt dof, off, goff, d;
/* Reject points not either cells or vertices */
if (((p < cStart) || (p >= cEnd)) && ((p < vStart) || (p >= vEnd))) continue;
if (hasLabel) {
PetscInt value;
if (((p >= cStart) && (p < cEnd) && numLabelCells) ||
((p >= vStart) && (p < vEnd) && numLabelVertices)) {
ierr = DMPlexGetLabelValue(dm, "vtk", p, &value);CHKERRQ(ierr);
if (value != 1) continue;
}
}
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(globalSection, p, &goff);CHKERRQ(ierr);
if (goff >= 0) {
for (d = 0; d < dof; ++d) {
localValues[k++] = array[off+d];
}
}
}
ierr = MPI_Send(&k, 1, MPIU_INT, 0, tag, comm);CHKERRQ(ierr);
ierr = MPI_Send(localValues, k, mpiType, 0, tag, comm);CHKERRQ(ierr);
ierr = PetscFree(localValues);CHKERRQ(ierr);
}
ierr = VecRestoreArray(v, &array);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C)
{
PetscErrorCode ierr;
Mat Cmpi;
Mat_PtAPMPI *ptap;
PetscFreeSpaceList free_space=NULL,current_space=NULL;
Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c;
Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data;
Mat_SeqAIJ *p_loc,*p_oth;
PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ;
PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz;
PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row;
PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n;
PetscBT lnkbt;
MPI_Comm comm;
PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0;
PetscInt **buf_rj,**buf_ri,**buf_ri_k;
PetscInt len,proc,*dnz,*onz,*owners;
PetscInt nzi,*pti,*ptj;
PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci;
MPI_Request *swaits,*rwaits;
MPI_Status *sstatus,rstatus;
Mat_Merge_SeqsToMPI *merge;
PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0;
PetscReal afill=1.0,afill_tmp;
PetscInt rmax;
#if defined(PTAP_PROFILE)
PetscLogDouble t0,t1,t2,t3,t4;
#endif
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t0);CHKERRQ(ierr);
#endif
/* check if matrix local sizes are compatible */
if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) {
SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Arow (%D, %D) != Prow (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend);
}
if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) {
SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Acol (%D, %D) != Prow (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend);
}
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
/* create struct Mat_PtAPMPI and attached it to C later */
ierr = PetscNew(&ptap);CHKERRQ(ierr);
ierr = PetscNew(&merge);CHKERRQ(ierr);
ptap->merge = merge;
ptap->reuse = MAT_INITIAL_MATRIX;
/* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */
ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr);
/* get P_loc by taking all local rows of P */
ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr);
p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data;
p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data;
pi_loc = p_loc->i; pj_loc = p_loc->j;
pi_oth = p_oth->i; pj_oth = p_oth->j;
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t1);CHKERRQ(ierr);
#endif
/* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */
/*-------------------------------------------------------------------*/
ierr = PetscMalloc1((am+1),&api);CHKERRQ(ierr);
api[0] = 0;
/* create and initialize a linked list */
ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr);
/* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */
ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr);
current_space = free_space;
for (i=0; i<am; i++) {
/* diagonal portion of A */
nzi = adi[i+1] - adi[i];
aj = ad->j + adi[i];
for (j=0; j<nzi; j++) {
row = aj[j];
pnz = pi_loc[row+1] - pi_loc[row];
Jptr = pj_loc + pi_loc[row];
/* add non-zero cols of P into the sorted linked list lnk */
ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr);
}
/* off-diagonal portion of A */
nzi = aoi[i+1] - aoi[i];
aj = ao->j + aoi[i];
for (j=0; j<nzi; j++) {
row = aj[j];
pnz = pi_oth[row+1] - pi_oth[row];
Jptr = pj_oth + pi_oth[row];
ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr);
}
apnz = lnk[0];
api[i+1] = api[i] + apnz;
if (ap_rmax < apnz) ap_rmax = apnz;
/* if free space is not available, double the total space in the list */
if (current_space->local_remaining<apnz) {
ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr);
nspacedouble++;
}
/* Copy data into free space, then initialize lnk */
ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr);
current_space->array += apnz;
current_space->local_used += apnz;
current_space->local_remaining -= apnz;
}
/* Allocate space for apj, initialize apj, and */
/* destroy list of free space and other temporary array(s) */
ierr = PetscMalloc1((api[am]+1),&apj);CHKERRQ(ierr);
ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr);
afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1);
if (afill_tmp > afill) afill = afill_tmp;
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2);CHKERRQ(ierr);
#endif
/* determine symbolic Co=(p->B)^T*AP - send to others */
/*----------------------------------------------------*/
ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr);
/* then, compute symbolic Co = (p->B)^T*AP */
pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors
>= (num of nonzero rows of C_seq) - pn */
ierr = PetscMalloc1((pon+1),&coi);CHKERRQ(ierr);
coi[0] = 0;
/* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */
nnz = fill*(poti[pon] + api[am]);
ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr);
current_space = free_space;
for (i=0; i<pon; i++) {
pnz = poti[i+1] - poti[i];
ptJ = potj + poti[i];
for (j=0; j<pnz; j++) {
row = ptJ[j]; /* row of AP == col of Pot */
apnz = api[row+1] - api[row];
Jptr = apj + api[row];
/* add non-zero cols of AP into the sorted linked list lnk */
ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr);
}
nnz = lnk[0];
/* If free space is not available, double the total space in the list */
if (current_space->local_remaining<nnz) {
ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr);
nspacedouble++;
}
/* Copy data into free space, and zero out denserows */
ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr);
current_space->array += nnz;
current_space->local_used += nnz;
current_space->local_remaining -= nnz;
coi[i+1] = coi[i] + nnz;
}
ierr = PetscMalloc1((coi[pon]+1),&coj);CHKERRQ(ierr);
ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr);
afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1);
if (afill_tmp > afill) afill = afill_tmp;
ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr);
/* send j-array (coj) of Co to other processors */
/*----------------------------------------------*/
/* determine row ownership */
ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr);
merge->rowmap->n = pn;
merge->rowmap->bs = 1;
ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr);
owners = merge->rowmap->range;
/* determine the number of messages to send, their lengths */
ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr);
ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr);
ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr);
len_s = merge->len_s;
merge->nsend = 0;
ierr = PetscMalloc1((size+2),&owners_co);CHKERRQ(ierr);
proc = 0;
for (i=0; i<pon; i++) {
while (prmap[i] >= owners[proc+1]) proc++;
len_si[proc]++; /* num of rows in Co to be sent to [proc] */
len_s[proc] += coi[i+1] - coi[i];
}
len = 0; /* max length of buf_si[] */
owners_co[0] = 0;
for (proc=0; proc<size; proc++) {
owners_co[proc+1] = owners_co[proc] + len_si[proc];
if (len_si[proc]) {
merge->nsend++;
len_si[proc] = 2*(len_si[proc] + 1);
len += len_si[proc];
}
}
/* determine the number and length of messages to receive for coi and coj */
ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr);
ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr);
/* post the Irecv and Isend of coj */
ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr);
ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr);
ierr = PetscMalloc1((merge->nsend+1),&swaits);CHKERRQ(ierr);
for (proc=0, k=0; proc<size; proc++) {
if (!len_s[proc]) continue;
i = owners_co[proc];
ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr);
k++;
}
/* receives and sends of coj are complete */
for (i=0; i<merge->nrecv; i++) {
ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr);
}
ierr = PetscFree(rwaits);CHKERRQ(ierr);
if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);}
/* send and recv coi */
/*-------------------*/
ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr);
ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr);
ierr = PetscMalloc1((len+1),&buf_s);CHKERRQ(ierr);
buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
for (proc=0,k=0; proc<size; proc++) {
if (!len_s[proc]) continue;
/* form outgoing message for i-structure:
buf_si[0]: nrows to be sent
[1:nrows]: row index (global)
[nrows+1:2*nrows+1]: i-structure index
*/
/*-------------------------------------------*/
nrows = len_si[proc]/2 - 1;
buf_si_i = buf_si + nrows+1;
buf_si[0] = nrows;
buf_si_i[0] = 0;
nrows = 0;
for (i=owners_co[proc]; i<owners_co[proc+1]; i++) {
nzi = coi[i+1] - coi[i];
buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */
buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */
nrows++;
}
ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr);
k++;
buf_si += len_si[proc];
}
i = merge->nrecv;
while (i--) {
ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr);
}
ierr = PetscFree(rwaits);CHKERRQ(ierr);
if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);}
ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr);
ierr = PetscFree(len_ri);CHKERRQ(ierr);
ierr = PetscFree(swaits);CHKERRQ(ierr);
ierr = PetscFree(buf_s);CHKERRQ(ierr);
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t3);CHKERRQ(ierr);
#endif
/* compute the local portion of C (mpi mat) */
/*------------------------------------------*/
ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr);
/* allocate pti array and free space for accumulating nonzero column info */
ierr = PetscMalloc1((pn+1),&pti);CHKERRQ(ierr);
pti[0] = 0;
/* set initial free space to be fill*(nnz(P) + nnz(AP)) */
nnz = fill*(pi_loc[pm] + api[am]);
ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr);
current_space = free_space;
ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr);
for (k=0; k<merge->nrecv; k++) {
buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
nrows = *buf_ri_k[k];
nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
}
ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr);
rmax = 0;
for (i=0; i<pn; i++) {
/* add pdt[i,:]*AP into lnk */
pnz = pdti[i+1] - pdti[i];
ptJ = pdtj + pdti[i];
for (j=0; j<pnz; j++) {
row = ptJ[j]; /* row of AP == col of Pt */
apnz = api[row+1] - api[row];
Jptr = apj + api[row];
/* add non-zero cols of AP into the sorted linked list lnk */
ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr);
}
/* add received col data into lnk */
for (k=0; k<merge->nrecv; k++) { /* k-th received message */
if (i == *nextrow[k]) { /* i-th row */
nzi = *(nextci[k]+1) - *nextci[k];
Jptr = buf_rj[k] + *nextci[k];
ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr);
nextrow[k]++; nextci[k]++;
}
}
nnz = lnk[0];
/* if free space is not available, make more free space */
if (current_space->local_remaining<nnz) {
ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr);
nspacedouble++;
}
/* copy data into free space, then initialize lnk */
ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr);
ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr);
current_space->array += nnz;
current_space->local_used += nnz;
current_space->local_remaining -= nnz;
pti[i+1] = pti[i] + nnz;
if (nnz > rmax) rmax = nnz;
}
ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr);
ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr);
ierr = PetscMalloc1((pti[pn]+1),&ptj);CHKERRQ(ierr);
ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr);
afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1);
if (afill_tmp > afill) afill = afill_tmp;
ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr);
/* create symbolic parallel matrix Cmpi */
/*--------------------------------------*/
ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr);
ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,P->cmap->bs);CHKERRQ(ierr);
ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr);
ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr);
merge->bi = pti; /* Cseq->i */
merge->bj = ptj; /* Cseq->j */
merge->coi = coi; /* Co->i */
merge->coj = coj; /* Co->j */
merge->buf_ri = buf_ri;
merge->buf_rj = buf_rj;
merge->owners_co = owners_co;
merge->destroy = Cmpi->ops->destroy;
merge->duplicate = Cmpi->ops->duplicate;
/* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */
Cmpi->assembled = PETSC_FALSE;
Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP;
Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP;
/* attach the supporting struct to Cmpi for reuse */
c = (Mat_MPIAIJ*)Cmpi->data;
c->ptap = ptap;
ptap->api = api;
ptap->apj = apj;
ptap->rmax = ap_rmax;
*C = Cmpi;
/* flag 'scalable' determines which implementations to be used:
0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa;
1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */
/* set default scalable */
ptap->scalable = PETSC_TRUE;
ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr);
if (!ptap->scalable) { /* Do dense axpy */
ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr);
} else {
ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr);
}
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t4);CHKERRQ(ierr);
if (rank==1) PetscPrintf(MPI_COMM_SELF," [%d] PtAPSymbolic %g/P + %g/AP + %g/comm + %g/PtAP = %g\n",rank,t1-t0,t2-t1,t3-t2,t4-t3,t4-t0);CHKERRQ(ierr);
#endif
#if defined(PETSC_USE_INFO)
if (pti[pn] != 0) {
ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr);
ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr);
} else {
ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr);
}
#endif
PetscFunctionReturn(0);
}
PetscErrorCode heavyEdgeMatchAgg(IS perm,Mat a_Gmat,PetscInt verbose,PetscCoarsenData **a_locals_llist)
{
PetscErrorCode ierr;
PetscBool isMPI;
MPI_Comm wcomm = ((PetscObject)a_Gmat)->comm;
PetscInt sub_it,kk,n,ix,*idx,*ii,iter,Iend,my0;
PetscMPIInt rank,size;
const PetscInt nloc = a_Gmat->rmap->n,n_iter=6; /* need to figure out how to stop this */
PetscInt *lid_cprowID,*lid_gid;
PetscBool *lid_matched;
Mat_SeqAIJ *matA, *matB=0;
Mat_MPIAIJ *mpimat=0;
PetscScalar one=1.;
PetscCoarsenData *agg_llists = PETSC_NULL,*deleted_list = PETSC_NULL;
Mat cMat,tMat,P;
MatScalar *ap;
PetscMPIInt tag1,tag2;
PetscFunctionBegin;
ierr = MPI_Comm_rank(wcomm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(wcomm, &size);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(a_Gmat, &my0, &Iend);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(wcomm, &tag1);CHKERRQ(ierr);
ierr = PetscCommGetNewTag(wcomm, &tag2);CHKERRQ(ierr);
ierr = PetscMalloc(nloc*sizeof(PetscInt), &lid_gid);CHKERRQ(ierr); /* explicit array needed */
ierr = PetscMalloc(nloc*sizeof(PetscInt), &lid_cprowID);CHKERRQ(ierr);
ierr = PetscMalloc(nloc*sizeof(PetscBool), &lid_matched);CHKERRQ(ierr);
ierr = PetscCDCreate(nloc, &agg_llists);CHKERRQ(ierr);
/* ierr = PetscCDSetChuckSize(agg_llists, nloc+1);CHKERRQ(ierr); */
*a_locals_llist = agg_llists;
ierr = PetscCDCreate(size, &deleted_list);CHKERRQ(ierr);
ierr = PetscCDSetChuckSize(deleted_list, 100);CHKERRQ(ierr);
/* setup 'lid_gid' for scatters and add self to all lists */
for (kk=0;kk<nloc;kk++) {
lid_gid[kk] = kk + my0;
ierr = PetscCDAppendID(agg_llists, kk, my0+kk);CHKERRQ(ierr);
}
/* make a copy of the graph, this gets destroyed in iterates */
ierr = MatDuplicate(a_Gmat,MAT_COPY_VALUES,&cMat);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)a_Gmat, MATMPIAIJ, &isMPI);CHKERRQ(ierr);
iter = 0;
while(iter++ < n_iter) {
PetscScalar *cpcol_gid,*cpcol_max_ew,*cpcol_max_pe,*lid_max_ew;
PetscBool *cpcol_matched;
PetscMPIInt *cpcol_pe,proc;
Vec locMaxEdge,locMaxPE,ghostMaxEdge,ghostMaxPE;
PetscInt nEdges,n_nz_row,jj;
Edge *Edges;
PetscInt gid;
const PetscInt *perm_ix, n_sub_its = 120;
/* get submatrices of cMat */
if (isMPI) {
mpimat = (Mat_MPIAIJ*)cMat->data;
matA = (Mat_SeqAIJ*)mpimat->A->data;
matB = (Mat_SeqAIJ*)mpimat->B->data;
/* force compressed storage of B */
matB->compressedrow.check = PETSC_TRUE;
ierr = MatCheckCompressedRow(mpimat->B,&matB->compressedrow,matB->i,cMat->rmap->n,-1.0);CHKERRQ(ierr);
assert(matB->compressedrow.use);
} else {
matA = (Mat_SeqAIJ*)cMat->data;
}
assert(matA && !matA->compressedrow.use);
assert(matB==0 || matB->compressedrow.use);
/* set max edge on nodes */
ierr = MatGetVecs(cMat, &locMaxEdge, 0);CHKERRQ(ierr);
ierr = MatGetVecs(cMat, &locMaxPE, 0);CHKERRQ(ierr);
/* get 'cpcol_pe' & 'cpcol_gid' & init. 'cpcol_matched' using 'mpimat->lvec' */
if (mpimat) {
Vec vec;
PetscScalar vval;
ierr = MatGetVecs(cMat, &vec, 0);CHKERRQ(ierr);
/* cpcol_pe */
vval = (PetscScalar)(rank);
for (kk=0,gid=my0;kk<nloc;kk++,gid++) {
ierr = VecSetValues(vec, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr); /* set with GID */
}
ierr = VecAssemblyBegin(vec);CHKERRQ(ierr);
ierr = VecAssemblyEnd(vec);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,vec,mpimat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,vec,mpimat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(mpimat->lvec, &cpcol_gid);CHKERRQ(ierr); /* get proc ID in 'cpcol_gid' */
ierr = VecGetLocalSize(mpimat->lvec, &n);CHKERRQ(ierr);
ierr = PetscMalloc(n*sizeof(PetscInt), &cpcol_pe);CHKERRQ(ierr);
for (kk=0;kk<n;kk++) cpcol_pe[kk] = (PetscMPIInt)PetscRealPart(cpcol_gid[kk]);
ierr = VecRestoreArray(mpimat->lvec, &cpcol_gid);CHKERRQ(ierr);
/* cpcol_gid */
for (kk=0,gid=my0;kk<nloc;kk++,gid++) {
vval = (PetscScalar)(gid);
ierr = VecSetValues(vec, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr); /* set with GID */
}
ierr = VecAssemblyBegin(vec);CHKERRQ(ierr);
ierr = VecAssemblyEnd(vec);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,vec,mpimat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,vec,mpimat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecDestroy(&vec);CHKERRQ(ierr);
ierr = VecGetArray(mpimat->lvec, &cpcol_gid);CHKERRQ(ierr); /* get proc ID in 'cpcol_gid' */
/* cpcol_matched */
ierr = VecGetLocalSize(mpimat->lvec, &n);CHKERRQ(ierr);
ierr = PetscMalloc(n*sizeof(PetscBool), &cpcol_matched);CHKERRQ(ierr);
for (kk=0;kk<n;kk++) cpcol_matched[kk] = PETSC_FALSE;
}
/* need an inverse map - locals */
for (kk=0;kk<nloc;kk++) lid_cprowID[kk] = -1;
/* set index into compressed row 'lid_cprowID' */
if (matB) {
ii = matB->compressedrow.i;
for (ix=0; ix<matB->compressedrow.nrows; ix++) {
lid_cprowID[matB->compressedrow.rindex[ix]] = ix;
}
}
/* get removed IS, use '' */
/* if (iter==1) { */
/* PetscInt *lid_rem,idx; */
/* ierr = PetscMalloc(nloc*sizeof(PetscInt), &lid_rem);CHKERRQ(ierr); */
/* for (kk=idx=0;kk<nloc;kk++){ */
/* PetscInt nn,lid=kk; */
/* ii = matA->i; nn = ii[lid+1] - ii[lid]; */
/* if ((ix=lid_cprowID[lid]) != -1) { /\* if I have any ghost neighbors *\/ */
/* ii = matB->compressedrow.i; */
/* nn += ii[ix+1] - ii[ix]; */
/* } */
/* if (nn < 2) { */
/* lid_rem[idx++] = kk + my0; */
/* } */
/* } */
/* ierr = PetscCDSetRemovedIS(agg_llists, wcomm, idx, lid_rem);CHKERRQ(ierr); */
/* ierr = PetscFree(lid_rem);CHKERRQ(ierr); */
/* } */
/* compute 'locMaxEdge' & 'locMaxPE', and create list of edges, count edges' */
for (nEdges=0,kk=0,gid=my0;kk<nloc;kk++,gid++){
PetscReal max_e = 0., tt;
PetscScalar vval;
PetscInt lid = kk;
PetscMPIInt max_pe=rank,pe;
ii = matA->i; n = ii[lid+1] - ii[lid]; idx = matA->j + ii[lid];
ap = matA->a + ii[lid];
for (jj=0; jj<n; jj++) {
PetscInt lidj = idx[jj];
if (lidj != lid && PetscRealPart(ap[jj]) > max_e) max_e = PetscRealPart(ap[jj]);
if (lidj > lid) nEdges++;
}
if ((ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n ; jj++) {
if ((tt=PetscRealPart(ap[jj])) > max_e) max_e = tt;
nEdges++;
if ((pe=cpcol_pe[idx[jj]]) > max_pe) max_pe = pe;
}
}
vval = max_e;
ierr = VecSetValues(locMaxEdge, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr);
vval = (PetscScalar)max_pe;
ierr = VecSetValues(locMaxPE, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = VecAssemblyBegin(locMaxEdge);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxEdge);CHKERRQ(ierr);
ierr = VecAssemblyBegin(locMaxPE);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxPE);CHKERRQ(ierr);
/* get 'cpcol_max_ew' & 'cpcol_max_pe' */
if (mpimat) {
ierr = VecDuplicate(mpimat->lvec, &ghostMaxEdge);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,locMaxEdge,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxEdge,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecDuplicate(mpimat->lvec, &ghostMaxPE);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,locMaxPE,ghostMaxPE,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxPE,ghostMaxPE,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxPE, &cpcol_max_pe);CHKERRQ(ierr);
}
/* setup sorted list of edges */
ierr = PetscMalloc(nEdges*sizeof(Edge), &Edges);CHKERRQ(ierr);
ierr = ISGetIndices(perm, &perm_ix);CHKERRQ(ierr);
for (nEdges=n_nz_row=kk=0;kk<nloc;kk++){
PetscInt nn, lid = perm_ix[kk];
ii = matA->i; nn = n = ii[lid+1] - ii[lid]; idx = matA->j + ii[lid];
ap = matA->a + ii[lid];
for (jj=0; jj<n; jj++) {
PetscInt lidj = idx[jj]; assert(PetscRealPart(ap[jj])>0.);
if (lidj > lid) {
Edges[nEdges].lid0 = lid;
Edges[nEdges].gid1 = lidj + my0;
Edges[nEdges].cpid1 = -1;
Edges[nEdges].weight = PetscRealPart(ap[jj]);
nEdges++;
}
}
if ((ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
nn += n;
for (jj=0 ; jj<n ; jj++) {
assert(PetscRealPart(ap[jj])>0.);
Edges[nEdges].lid0 = lid;
Edges[nEdges].gid1 = (PetscInt)PetscRealPart(cpcol_gid[idx[jj]]);
Edges[nEdges].cpid1 = idx[jj];
Edges[nEdges].weight = PetscRealPart(ap[jj]);
nEdges++;
}
}
if (nn > 1) n_nz_row++;
else if (iter == 1){
/* should select this because it is technically in the MIS but lets not */
ierr = PetscCDRemoveAll(agg_llists, lid);CHKERRQ(ierr);
}
}
ierr = ISRestoreIndices(perm,&perm_ix);CHKERRQ(ierr);
qsort(Edges, nEdges, sizeof(Edge), gamg_hem_compare);
/* projection matrix */
ierr = MatCreateAIJ(wcomm, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE, 1, 0, 1, 0, &P);CHKERRQ(ierr);
/* clear matched flags */
for (kk=0;kk<nloc;kk++) lid_matched[kk] = PETSC_FALSE;
/* process - communicate - process */
for (sub_it=0;sub_it<n_sub_its;sub_it++){
PetscInt nactive_edges;
ierr = VecGetArray(locMaxEdge, &lid_max_ew);CHKERRQ(ierr);
for (kk=nactive_edges=0;kk<nEdges;kk++){
/* HEM */
const Edge *e = &Edges[kk];
const PetscInt lid0=e->lid0,gid1=e->gid1,cpid1=e->cpid1,gid0=lid0+my0,lid1=gid1-my0;
PetscBool isOK = PETSC_TRUE;
/* skip if either (local) vertex is done already */
if (lid_matched[lid0] || (gid1>=my0 && gid1<Iend && lid_matched[gid1-my0])) {
continue;
}
/* skip if ghost vertex is done */
if (cpid1 != -1 && cpcol_matched[cpid1]) {
continue;
}
nactive_edges++;
/* skip if I have a bigger edge someplace (lid_max_ew gets updated) */
if (PetscRealPart(lid_max_ew[lid0]) > e->weight + 1.e-12) {
continue;
}
if (cpid1 == -1) {
if (PetscRealPart(lid_max_ew[lid1]) > e->weight + 1.e-12) {
continue;
}
} else {
/* see if edge might get matched on other proc */
PetscReal g_max_e = PetscRealPart(cpcol_max_ew[cpid1]);
if (g_max_e > e->weight + 1.e-12) {
continue;
} else if (e->weight > g_max_e - 1.e-12 && (PetscMPIInt)PetscRealPart(cpcol_max_pe[cpid1]) > rank) {
/* check for max_e == to this edge and larger processor that will deal with this */
continue;
}
}
/* check ghost for v0 */
if (isOK){
PetscReal max_e,ew;
if ((ix=lid_cprowID[lid0]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n && isOK; jj++) {
PetscInt lidj = idx[jj];
if (cpcol_matched[lidj]) continue;
ew = PetscRealPart(ap[jj]); max_e = PetscRealPart(cpcol_max_ew[lidj]);
/* check for max_e == to this edge and larger processor that will deal with this */
if (ew > max_e - 1.e-12 && ew > PetscRealPart(lid_max_ew[lid0]) - 1.e-12 && (PetscMPIInt)PetscRealPart(cpcol_max_pe[lidj]) > rank){
isOK = PETSC_FALSE;
}
}
}
/* for v1 */
if (cpid1 == -1 && isOK){
if ((ix=lid_cprowID[lid1]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n && isOK ; jj++) {
PetscInt lidj = idx[jj];
if (cpcol_matched[lidj]) continue;
ew = PetscRealPart(ap[jj]); max_e = PetscRealPart(cpcol_max_ew[lidj]);
/* check for max_e == to this edge and larger processor that will deal with this */
if (ew > max_e - 1.e-12 && ew > PetscRealPart(lid_max_ew[lid1]) - 1.e-12 && (PetscMPIInt)PetscRealPart(cpcol_max_pe[lidj]) > rank) {
isOK = PETSC_FALSE;
}
}
}
}
}
/* do it */
if (isOK){
if (cpid1 == -1) {
lid_matched[lid1] = PETSC_TRUE; /* keep track of what we've done this round */
ierr = PetscCDAppendRemove(agg_llists, lid0, lid1);CHKERRQ(ierr);
} else if (sub_it != n_sub_its-1) {
/* add gid1 to list of ghost deleted by me -- I need their children */
proc = cpcol_pe[cpid1];
cpcol_matched[cpid1] = PETSC_TRUE; /* messing with VecGetArray array -- needed??? */
ierr = PetscCDAppendID(deleted_list, proc, cpid1);CHKERRQ(ierr); /* cache to send messages */
ierr = PetscCDAppendID(deleted_list, proc, lid0);CHKERRQ(ierr);
} else {
continue;
}
lid_matched[lid0] = PETSC_TRUE; /* keep track of what we've done this round */
/* set projection */
ierr = MatSetValues(P,1,&gid0,1,&gid0,&one,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(P,1,&gid1,1,&gid0,&one,INSERT_VALUES);CHKERRQ(ierr);
} /* matched */
} /* edge loop */
/* deal with deleted ghost on first pass */
if (size>1 && sub_it != n_sub_its-1){
PetscCDPos pos; PetscBool ise = PETSC_FALSE;
PetscInt nSend1, **sbuffs1,nSend2;
#define REQ_BF_SIZE 100
MPI_Request *sreqs2[REQ_BF_SIZE],*rreqs2[REQ_BF_SIZE];
MPI_Status status;
/* send request */
for (proc=0,nSend1=0;proc<size;proc++){
ierr = PetscCDEmptyAt(deleted_list,proc,&ise);CHKERRQ(ierr);
if (!ise) nSend1++;
}
ierr = PetscMalloc(nSend1*sizeof(PetscInt*), &sbuffs1);CHKERRQ(ierr);
/* ierr = PetscMalloc4(nSend1, PetscInt*, sbuffs1, nSend1, PetscInt*, rbuffs1, nSend1, MPI_Request*, sreqs1, nSend1, MPI_Request*, rreqs1);CHKERRQ(ierr); */
/* PetscFree4(sbuffs1,rbuffs1,sreqs1,rreqs1); */
for (proc=0,nSend1=0;proc<size;proc++){
/* count ghosts */
ierr = PetscCDSizeAt(deleted_list,proc,&n);CHKERRQ(ierr);
if (n>0){
#define CHUNCK_SIZE 100
PetscInt *sbuff,*pt;
MPI_Request *request;
assert(n%2==0);
n /= 2;
ierr = PetscMalloc((2 + 2*n + n*CHUNCK_SIZE)*sizeof(PetscInt) + 2*sizeof(MPI_Request), &sbuff);CHKERRQ(ierr);
/* PetscMalloc4(2+2*n,PetscInt,sbuffs1[nSend1],n*CHUNCK_SIZE,PetscInt,rbuffs1[nSend1],1,MPI_Request,rreqs2[nSend1],1,MPI_Request,sreqs2[nSend1]); */
/* save requests */
sbuffs1[nSend1] = sbuff;
request = (MPI_Request*)sbuff;
sbuff = pt = (PetscInt*)(request+1);
*pt++ = n; *pt++ = rank;
ierr = PetscCDGetHeadPos(deleted_list,proc,&pos);CHKERRQ(ierr);
while(pos){
PetscInt lid0, cpid, gid;
ierr = PetscLLNGetID(pos, &cpid);CHKERRQ(ierr);
gid = (PetscInt)PetscRealPart(cpcol_gid[cpid]);
ierr = PetscCDGetNextPos(deleted_list,proc,&pos);CHKERRQ(ierr);
ierr = PetscLLNGetID(pos, &lid0);CHKERRQ(ierr);
ierr = PetscCDGetNextPos(deleted_list,proc,&pos);CHKERRQ(ierr);
*pt++ = gid; *pt++ = lid0;
}
/* send request tag1 [n, proc, n*[gid1,lid0] ] */
ierr = MPI_Isend(sbuff, 2*n+2, MPIU_INT, proc, tag1, wcomm, request);CHKERRQ(ierr);
/* post recieve */
request = (MPI_Request*)pt;
rreqs2[nSend1] = request; /* cache recv request */
pt = (PetscInt*)(request+1);
ierr = MPI_Irecv(pt, n*CHUNCK_SIZE, MPIU_INT, proc, tag2, wcomm, request);CHKERRQ(ierr);
/* clear list */
ierr = PetscCDRemoveAll(deleted_list, proc);CHKERRQ(ierr);
nSend1++;
}
}
/* recieve requests, send response, clear lists */
kk = nactive_edges;
ierr = MPI_Allreduce(&kk,&nactive_edges,1,MPIU_INT,MPI_SUM,wcomm);CHKERRQ(ierr); /* not correct syncronization and global */
nSend2 = 0;
while(1){
#define BF_SZ 10000
PetscMPIInt flag,count;
PetscInt rbuff[BF_SZ],*pt,*pt2,*pt3,count2,*sbuff,count3;
MPI_Request *request;
ierr = MPI_Iprobe(MPI_ANY_SOURCE, tag1, wcomm, &flag, &status);CHKERRQ(ierr);
if (!flag) break;
ierr = MPI_Get_count(&status, MPIU_INT, &count);CHKERRQ(ierr);
if (count > BF_SZ) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"buffer too small for recieve: %d",count);
proc = status.MPI_SOURCE;
/* recieve request tag1 [n, proc, n*[gid1,lid0] ] */
ierr = MPI_Recv(rbuff, count, MPIU_INT, proc, tag1, wcomm, &status);CHKERRQ(ierr);
/* count sends */
pt = rbuff; count3 = count2 = 0;
n = *pt++; kk = *pt++; assert(kk==proc);
while(n--){
PetscInt gid1=*pt++, lid1=gid1-my0; kk=*pt++; assert(lid1>=0 && lid1<nloc);
if (lid_matched[lid1]){
PetscPrintf(PETSC_COMM_SELF,"\t *** [%d]%s %d) ERROR recieved deleted gid %d, deleted by (lid) %d from proc %d\n",rank,__FUNCT__,sub_it,gid1,kk);
PetscSleep(1);
}
assert(!lid_matched[lid1]);
lid_matched[lid1] = PETSC_TRUE; /* keep track of what we've done this round */
ierr = PetscCDSizeAt(agg_llists, lid1, &kk);CHKERRQ(ierr);
count2 += kk + 2;
count3++; /* number of verts requested (n) */
}
assert(pt-rbuff==count);
if (count2 > count3*CHUNCK_SIZE) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Irecv will be too small: %d",count2);
/* send tag2 *[lid0, n, n*[gid] ] */
ierr = PetscMalloc(count2*sizeof(PetscInt) + sizeof(MPI_Request), &sbuff);CHKERRQ(ierr);
request = (MPI_Request*)sbuff;
sreqs2[nSend2++] = request; /* cache request */
if (nSend2==REQ_BF_SIZE) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"buffer too small for requests: %d",nSend2);
pt2 = sbuff = (PetscInt*)(request+1);
pt = rbuff;
n = *pt++; kk = *pt++; assert(kk==proc);
while(n--){
/* read [n, proc, n*[gid1,lid0] */
PetscInt gid1=*pt++, lid1=gid1-my0, lid0=*pt++; assert(lid1>=0 && lid1<nloc);
/* write [lid0, n, n*[gid] ] */
*pt2++ = lid0;
pt3 = pt2++; /* save pointer for later */
/* for (pos=PetscCDGetHeadPos(agg_llists,lid1) ; pos ; pos=PetscCDGetNextPos(agg_llists,lid1,pos)){ */
ierr = PetscCDGetHeadPos(agg_llists,lid1,&pos);CHKERRQ(ierr);
while(pos){
PetscInt gid;
ierr = PetscLLNGetID(pos, &gid);CHKERRQ(ierr);
ierr = PetscCDGetNextPos(agg_llists,lid1,&pos);CHKERRQ(ierr);
*pt2++ = gid;
}
*pt3 = (pt2-pt3)-1;
/* clear list */
ierr = PetscCDRemoveAll(agg_llists, lid1);CHKERRQ(ierr);
}
assert(pt2-sbuff==count2); assert(pt-rbuff==count);
/* send requested data tag2 *[lid0, n, n*[gid1] ] */
ierr = MPI_Isend(sbuff, count2, MPIU_INT, proc, tag2, wcomm, request);CHKERRQ(ierr);
}
/* recieve tag2 *[lid0, n, n*[gid] ] */
for (kk=0;kk<nSend1;kk++){
PetscMPIInt count;
MPI_Request *request;
PetscInt *pt, *pt2;
request = rreqs2[kk]; /* no need to free -- buffer is in 'sbuffs1' */
ierr = MPI_Wait(request, &status);CHKERRQ(ierr);
ierr = MPI_Get_count(&status, MPIU_INT, &count);CHKERRQ(ierr);
pt = pt2 = (PetscInt*)(request+1);
while(pt-pt2 < count){
PetscInt lid0 = *pt++, n = *pt++; assert(lid0>=0 && lid0<nloc);
while(n--){
PetscInt gid1 = *pt++;
ierr = PetscCDAppendID(agg_llists, lid0, gid1);CHKERRQ(ierr);
}
}
assert(pt-pt2==count);
}
/* wait for tag1 isends */
while(nSend1--){
MPI_Request *request;
request = (MPI_Request*)sbuffs1[nSend1];
ierr = MPI_Wait(request, &status);CHKERRQ(ierr);
ierr = PetscFree(request);CHKERRQ(ierr);
}
ierr = PetscFree(sbuffs1);CHKERRQ(ierr);
/* wait for tag2 isends */
while(nSend2--){
MPI_Request *request = sreqs2[nSend2];
ierr = MPI_Wait(request, &status);CHKERRQ(ierr);
ierr = PetscFree(request);CHKERRQ(ierr);
}
ierr = VecRestoreArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecRestoreArray(ghostMaxPE, &cpcol_max_pe);CHKERRQ(ierr);
/* get 'cpcol_matched' - use locMaxPE, ghostMaxEdge, cpcol_max_ew */
for (kk=0,gid=my0;kk<nloc;kk++,gid++) {
PetscScalar vval = lid_matched[kk] ? 1.0 : 0.0;
ierr = VecSetValues(locMaxPE, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr); /* set with GID */
}
ierr = VecAssemblyBegin(locMaxPE);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxPE);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,locMaxPE,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxPE,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecGetLocalSize(mpimat->lvec, &n);CHKERRQ(ierr);
for (kk=0;kk<n;kk++) {
cpcol_matched[kk] = (PetscBool)(PetscRealPart(cpcol_max_ew[kk]) != 0.0);
}
ierr = VecRestoreArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
} /* size > 1 */
/* compute 'locMaxEdge' */
ierr = VecRestoreArray(locMaxEdge, &lid_max_ew);CHKERRQ(ierr);
for (kk=0,gid=my0;kk<nloc;kk++,gid++){
PetscReal max_e = 0.,tt;
PetscScalar vval;
PetscInt lid = kk;
if (lid_matched[lid]) vval = 0.;
else {
ii = matA->i; n = ii[lid+1] - ii[lid]; idx = matA->j + ii[lid];
ap = matA->a + ii[lid];
for (jj=0; jj<n; jj++) {
PetscInt lidj = idx[jj];
if (lid_matched[lidj]) continue; /* this is new - can change local max */
if (lidj != lid && PetscRealPart(ap[jj]) > max_e) max_e = PetscRealPart(ap[jj]);
}
if (lid_cprowID && (ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n ; jj++) {
PetscInt lidj = idx[jj];
if (cpcol_matched[lidj]) continue;
if ((tt=PetscRealPart(ap[jj])) > max_e) max_e = tt;
}
}
}
vval = (PetscScalar)max_e;
ierr = VecSetValues(locMaxEdge, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr); /* set with GID */
}
ierr = VecAssemblyBegin(locMaxEdge);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxEdge);CHKERRQ(ierr);
if (size>1 && sub_it != n_sub_its-1){
/* compute 'cpcol_max_ew' */
ierr = VecScatterBegin(mpimat->Mvctx,locMaxEdge,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxEdge,ghostMaxEdge,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecGetArray(locMaxEdge, &lid_max_ew);CHKERRQ(ierr);
/* compute 'cpcol_max_pe' */
for (kk=0,gid=my0;kk<nloc;kk++,gid++){
PetscInt lid = kk;
PetscReal ew,v1_max_e,v0_max_e=PetscRealPart(lid_max_ew[lid]);
PetscScalar vval;
PetscMPIInt max_pe=rank,pe;
if (lid_matched[lid]) vval = (PetscScalar)rank;
else if ((ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i; n = ii[ix+1] - ii[ix];
ap = matB->a + ii[ix];
idx = matB->j + ii[ix];
for (jj=0 ; jj<n ; jj++) {
PetscInt lidj = idx[jj];
if (cpcol_matched[lidj]) continue;
ew = PetscRealPart(ap[jj]); v1_max_e = PetscRealPart(cpcol_max_ew[lidj]);
/* get max pe that has a max_e == to this edge w */
if ((pe=cpcol_pe[idx[jj]]) > max_pe && ew > v1_max_e - 1.e-12 && ew > v0_max_e - 1.e-12) max_pe = pe;
assert(ew < v0_max_e + 1.e-12 && ew < v1_max_e + 1.e-12);
}
vval = (PetscScalar)max_pe;
}
ierr = VecSetValues(locMaxPE, 1, &gid, &vval, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = VecAssemblyBegin(locMaxPE);CHKERRQ(ierr);
ierr = VecAssemblyEnd(locMaxPE);CHKERRQ(ierr);
ierr = VecScatterBegin(mpimat->Mvctx,locMaxPE,ghostMaxPE,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(mpimat->Mvctx,locMaxPE,ghostMaxPE,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecGetArray(ghostMaxPE, &cpcol_max_pe);CHKERRQ(ierr);
ierr = VecRestoreArray(locMaxEdge, &lid_max_ew);CHKERRQ(ierr);
} /* deal with deleted ghost */
if (verbose>2) PetscPrintf(wcomm,"\t[%d]%s %d.%d: %d active edges.\n",
rank,__FUNCT__,iter,sub_it,nactive_edges);
if (!nactive_edges) break;
} /* sub_it loop */
/* clean up iteration */
ierr = PetscFree(Edges);CHKERRQ(ierr);
if (mpimat){
ierr = VecRestoreArray(ghostMaxEdge, &cpcol_max_ew);CHKERRQ(ierr);
ierr = VecDestroy(&ghostMaxEdge);CHKERRQ(ierr);
ierr = VecRestoreArray(ghostMaxPE, &cpcol_max_pe);CHKERRQ(ierr);
ierr = VecDestroy(&ghostMaxPE);CHKERRQ(ierr);
ierr = PetscFree(cpcol_pe);CHKERRQ(ierr);
ierr = PetscFree(cpcol_matched);CHKERRQ(ierr);
}
ierr = VecDestroy(&locMaxEdge);CHKERRQ(ierr);
ierr = VecDestroy(&locMaxPE);CHKERRQ(ierr);
if (mpimat){
ierr = VecRestoreArray(mpimat->lvec, &cpcol_gid);CHKERRQ(ierr);
}
/* create next G if needed */
if (iter == n_iter) { /* hard wired test - need to look at full surrounded nodes or something */
ierr = MatDestroy(&P);CHKERRQ(ierr);
ierr = MatDestroy(&cMat);CHKERRQ(ierr);
break;
} else {
Vec diag;
/* add identity for unmatched vertices so they stay alive */
for (kk=0,gid=my0;kk<nloc;kk++,gid++){
if (!lid_matched[kk]) {
gid = kk+my0;
ierr = MatGetRow(cMat,gid,&n,0,0);CHKERRQ(ierr);
if (n>1){
ierr = MatSetValues(P,1,&gid,1,&gid,&one,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = MatRestoreRow(cMat,gid,&n,0,0);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(P,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(P,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
/* project to make new graph with colapsed edges */
ierr = MatPtAP(cMat,P,MAT_INITIAL_MATRIX,1.0,&tMat);CHKERRQ(ierr);
ierr = MatDestroy(&P);CHKERRQ(ierr);
ierr = MatDestroy(&cMat);CHKERRQ(ierr);
cMat = tMat;
ierr = MatGetVecs(cMat, &diag, 0);CHKERRQ(ierr);
ierr = MatGetDiagonal(cMat, diag);CHKERRQ(ierr); /* effectively PCJACOBI */
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = VecSqrtAbs(diag);CHKERRQ(ierr);
ierr = MatDiagonalScale(cMat, diag, diag);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
}
} /* coarsen iterator */
/* make fake matrix */
if (size>1){
Mat mat;
PetscCDPos pos;
PetscInt gid, NN, MM, jj = 0, mxsz = 0;
for (kk=0;kk<nloc;kk++){
ierr = PetscCDSizeAt(agg_llists, kk, &jj);CHKERRQ(ierr);
if (jj > mxsz) mxsz = jj;
}
ierr = MatGetSize(a_Gmat, &MM, &NN);CHKERRQ(ierr);
if (mxsz > MM-nloc) mxsz = MM-nloc;
ierr = MatCreateAIJ(wcomm, nloc, nloc,PETSC_DETERMINE, PETSC_DETERMINE,0, 0, mxsz, 0, &mat);CHKERRQ(ierr);
/* */
for (kk=0,gid=my0;kk<nloc;kk++,gid++){
/* for (pos=PetscCDGetHeadPos(agg_llists,kk) ; pos ; pos=PetscCDGetNextPos(agg_llists,kk,pos)){ */
ierr = PetscCDGetHeadPos(agg_llists,kk,&pos);CHKERRQ(ierr);
while(pos){
PetscInt gid1;
ierr = PetscLLNGetID(pos, &gid1);CHKERRQ(ierr);
ierr = PetscCDGetNextPos(agg_llists,kk,&pos);CHKERRQ(ierr);
if (gid1 < my0 || gid1 >= my0+nloc) {
ierr = MatSetValues(mat,1,&gid,1,&gid1,&one,ADD_VALUES);CHKERRQ(ierr);
}
}
}
ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscCDSetMat(agg_llists, mat);CHKERRQ(ierr);
}
ierr = PetscFree(lid_cprowID);CHKERRQ(ierr);
ierr = PetscFree(lid_gid);CHKERRQ(ierr);
ierr = PetscFree(lid_matched);CHKERRQ(ierr);
ierr = PetscCDDestroy(deleted_list);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode VecAssemblyBegin_MPI_BTS(Vec X)
{
Vec_MPI *x = (Vec_MPI*)X->data;
PetscErrorCode ierr;
MPI_Comm comm;
PetscInt i,j,jb,bs;
PetscFunctionBegin;
if (X->stash.donotstash) PetscFunctionReturn(0);
ierr = PetscObjectGetComm((PetscObject)X,&comm);CHKERRQ(ierr);
ierr = VecGetBlockSize(X,&bs);CHKERRQ(ierr);
#if defined(PETSC_USE_DEBUG)
{
InsertMode addv;
ierr = MPIU_Allreduce((PetscEnum*)&X->stash.insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,comm);CHKERRQ(ierr);
if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(comm,PETSC_ERR_ARG_NOTSAMETYPE,"Some processors inserted values while others added");
}
#endif
X->bstash.insertmode = X->stash.insertmode; /* Block stash implicitly tracks InsertMode of scalar stash */
ierr = VecStashSortCompress_Private(&X->stash);CHKERRQ(ierr);
ierr = VecStashSortCompress_Private(&X->bstash);CHKERRQ(ierr);
if (!x->sendranks) {
PetscMPIInt nowners,bnowners,*owners,*bowners;
PetscInt ntmp;
ierr = VecStashGetOwnerList_Private(&X->stash,X->map,&nowners,&owners);CHKERRQ(ierr);
ierr = VecStashGetOwnerList_Private(&X->bstash,X->map,&bnowners,&bowners);CHKERRQ(ierr);
ierr = PetscMergeMPIIntArray(nowners,owners,bnowners,bowners,&ntmp,&x->sendranks);CHKERRQ(ierr);
x->nsendranks = ntmp;
ierr = PetscFree(owners);CHKERRQ(ierr);
ierr = PetscFree(bowners);CHKERRQ(ierr);
ierr = PetscMalloc1(x->nsendranks,&x->sendhdr);CHKERRQ(ierr);
ierr = PetscCalloc1(x->nsendranks,&x->sendptrs);CHKERRQ(ierr);
}
for (i=0,j=0,jb=0; i<x->nsendranks; i++) {
PetscMPIInt rank = x->sendranks[i];
x->sendhdr[i].insertmode = X->stash.insertmode;
/* Initialize pointers for non-empty stashes the first time around. Subsequent assemblies with
* VEC_SUBSET_OFF_PROC_ENTRIES will leave the old pointers (dangling because the stash has been collected) when
* there is nothing new to send, so that size-zero messages get sent instead. */
x->sendhdr[i].count = 0;
if (X->stash.n) {
x->sendptrs[i].ints = &X->stash.idx[j];
x->sendptrs[i].scalars = &X->stash.array[j];
for ( ; j<X->stash.n && X->stash.idx[j] < X->map->range[rank+1]; j++) x->sendhdr[i].count++;
}
x->sendhdr[i].bcount = 0;
if (X->bstash.n) {
x->sendptrs[i].intb = &X->bstash.idx[jb];
x->sendptrs[i].scalarb = &X->bstash.array[jb*bs];
for ( ; jb<X->bstash.n && X->bstash.idx[jb]*bs < X->map->range[rank+1]; jb++) x->sendhdr[i].bcount++;
}
}
if (!x->segrecvint) {ierr = PetscSegBufferCreate(sizeof(PetscInt),1000,&x->segrecvint);CHKERRQ(ierr);}
if (!x->segrecvscalar) {ierr = PetscSegBufferCreate(sizeof(PetscScalar),1000,&x->segrecvscalar);CHKERRQ(ierr);}
if (!x->segrecvframe) {ierr = PetscSegBufferCreate(sizeof(VecAssemblyFrame),50,&x->segrecvframe);CHKERRQ(ierr);}
if (x->recvhdr) { /* VEC_SUBSET_OFF_PROC_ENTRIES and this is not the first assembly */
PetscMPIInt tag[4];
if (!x->assembly_subset) SETERRQ(comm,PETSC_ERR_PLIB,"Attempt to reuse rendezvous when not VEC_SUBSET_OFF_PROC_ENTRIES");
for (i=0; i<4; i++) {ierr = PetscCommGetNewTag(comm,&tag[i]);CHKERRQ(ierr);}
for (i=0; i<x->nsendranks; i++) {
ierr = VecAssemblySend_MPI_Private(comm,tag,i,x->sendranks[i],x->sendhdr+i,x->sendreqs+4*i,X);CHKERRQ(ierr);
}
for (i=0; i<x->nrecvranks; i++) {
ierr = VecAssemblyRecv_MPI_Private(comm,tag,x->recvranks[i],x->recvhdr+i,x->recvreqs+4*i,X);CHKERRQ(ierr);
}
x->use_status = PETSC_TRUE;
} else { /* First time */
ierr = PetscCommBuildTwoSidedFReq(comm,3,MPIU_INT,x->nsendranks,x->sendranks,(PetscInt*)x->sendhdr,&x->nrecvranks,&x->recvranks,&x->recvhdr,4,&x->sendreqs,&x->recvreqs,VecAssemblySend_MPI_Private,VecAssemblyRecv_MPI_Private,X);CHKERRQ(ierr);
x->use_status = PETSC_FALSE;
}
{
PetscInt nstash,reallocs;
ierr = VecStashGetInfo_Private(&X->stash,&nstash,&reallocs);CHKERRQ(ierr);
ierr = PetscInfo2(X,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr);
ierr = VecStashGetInfo_Private(&X->bstash,&nstash,&reallocs);CHKERRQ(ierr);
ierr = PetscInfo2(X,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
static PetscErrorCode PetscCommBuildTwoSidedFReq_Ibarrier(MPI_Comm comm,PetscMPIInt count,MPI_Datatype dtype,PetscMPIInt nto,const PetscMPIInt *toranks,const void *todata,
PetscMPIInt *nfrom,PetscMPIInt **fromranks,void *fromdata,PetscMPIInt ntags,MPI_Request **toreqs,MPI_Request **fromreqs,
PetscErrorCode (*send)(MPI_Comm,const PetscMPIInt[],PetscMPIInt,PetscMPIInt,void*,MPI_Request[],void*),
PetscErrorCode (*recv)(MPI_Comm,const PetscMPIInt[],PetscMPIInt,void*,MPI_Request[],void*),void *ctx)
{
PetscErrorCode ierr;
PetscMPIInt nrecvs,tag,*tags,done,i;
MPI_Aint lb,unitbytes;
char *tdata;
MPI_Request *sendreqs,*usendreqs,*req,barrier;
PetscSegBuffer segrank,segdata,segreq;
PetscBool barrier_started;
PetscFunctionBegin;
ierr = PetscCommDuplicate(comm,&comm,&tag);CHKERRQ(ierr);
ierr = PetscMalloc1(ntags,&tags);CHKERRQ(ierr);
for (i=0; i<ntags; i++) {
ierr = PetscCommGetNewTag(comm,&tags[i]);CHKERRQ(ierr);
}
ierr = MPI_Type_get_extent(dtype,&lb,&unitbytes);CHKERRQ(ierr);
if (lb != 0) SETERRQ1(comm,PETSC_ERR_SUP,"Datatype with nonzero lower bound %ld\n",(long)lb);
tdata = (char*)todata;
ierr = PetscMalloc1(nto,&sendreqs);CHKERRQ(ierr);
ierr = PetscMalloc1(nto*ntags,&usendreqs);CHKERRQ(ierr);
/* Post synchronous sends */
for (i=0; i<nto; i++) {
ierr = MPI_Issend((void*)(tdata+count*unitbytes*i),count,dtype,toranks[i],tag,comm,sendreqs+i);CHKERRQ(ierr);
}
/* Post actual payloads. These are typically larger messages. Hopefully sending these later does not slow down the
* synchronous messages above. */
for (i=0; i<nto; i++) {
PetscMPIInt k;
for (k=0; k<ntags; k++) usendreqs[i*ntags+k] = MPI_REQUEST_NULL;
ierr = (*send)(comm,tags,i,toranks[i],tdata+count*unitbytes*i,usendreqs+i*ntags,ctx);CHKERRQ(ierr);
}
ierr = PetscSegBufferCreate(sizeof(PetscMPIInt),4,&segrank);CHKERRQ(ierr);
ierr = PetscSegBufferCreate(unitbytes,4*count,&segdata);CHKERRQ(ierr);
ierr = PetscSegBufferCreate(sizeof(MPI_Request),4,&segreq);CHKERRQ(ierr);
nrecvs = 0;
barrier = MPI_REQUEST_NULL;
/* MPICH-3.2 sometimes does not create a request in some "optimized" cases. This is arguably a standard violation,
* but we need to work around it. */
barrier_started = PETSC_FALSE;
for (done=0; !done; ) {
PetscMPIInt flag;
MPI_Status status;
ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag,comm,&flag,&status);CHKERRQ(ierr);
if (flag) { /* incoming message */
PetscMPIInt *recvrank,k;
void *buf;
ierr = PetscSegBufferGet(segrank,1,&recvrank);CHKERRQ(ierr);
ierr = PetscSegBufferGet(segdata,count,&buf);CHKERRQ(ierr);
*recvrank = status.MPI_SOURCE;
ierr = MPI_Recv(buf,count,dtype,status.MPI_SOURCE,tag,comm,MPI_STATUS_IGNORE);CHKERRQ(ierr);
ierr = PetscSegBufferGet(segreq,ntags,&req);CHKERRQ(ierr);
for (k=0; k<ntags; k++) req[k] = MPI_REQUEST_NULL;
ierr = (*recv)(comm,tags,status.MPI_SOURCE,buf,req,ctx);CHKERRQ(ierr);
nrecvs++;
}
if (!barrier_started) {
PetscMPIInt sent,nsends;
ierr = PetscMPIIntCast(nto,&nsends);CHKERRQ(ierr);
ierr = MPI_Testall(nsends,sendreqs,&sent,MPI_STATUSES_IGNORE);CHKERRQ(ierr);
if (sent) {
#if defined(PETSC_HAVE_MPI_IBARRIER)
ierr = MPI_Ibarrier(comm,&barrier);CHKERRQ(ierr);
#elif defined(PETSC_HAVE_MPIX_IBARRIER)
ierr = MPIX_Ibarrier(comm,&barrier);CHKERRQ(ierr);
#endif
barrier_started = PETSC_TRUE;
}
} else {
ierr = MPI_Test(&barrier,&done,MPI_STATUS_IGNORE);CHKERRQ(ierr);
}
}
*nfrom = nrecvs;
ierr = PetscSegBufferExtractAlloc(segrank,fromranks);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segrank);CHKERRQ(ierr);
ierr = PetscSegBufferExtractAlloc(segdata,fromdata);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segdata);CHKERRQ(ierr);
*toreqs = usendreqs;
ierr = PetscSegBufferExtractAlloc(segreq,fromreqs);CHKERRQ(ierr);
ierr = PetscSegBufferDestroy(&segreq);CHKERRQ(ierr);
ierr = PetscFree(sendreqs);CHKERRQ(ierr);
ierr = PetscFree(tags);CHKERRQ(ierr);
ierr = PetscCommDestroy(&comm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
void PETSC_STDCALL petsccommgetnewtag_(MPI_Fint * comm,PetscMPIInt *tag, int *__ierr ){
*__ierr = PetscCommGetNewTag(
MPI_Comm_f2c( *(comm) ),tag);
}
PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C)
{
PetscErrorCode ierr;
Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c=(Mat_MPIAIJ*)C->data;
Mat_SeqAIJ *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data;
Mat_SeqAIJ *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data;
Mat_SeqAIJ *p_loc,*p_oth;
Mat_PtAPMPI *ptap;
Mat_Merge_SeqsToMPI *merge;
PetscInt *adi=ad->i,*aoi=ao->i,*adj,*aoj,*apJ,nextp;
PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj;
PetscInt i,j,k,anz,pnz,apnz,nextap,row,*cj;
MatScalar *ada,*aoa,*apa,*pa,*ca,*pa_loc,*pa_oth,valtmp;
PetscInt am =A->rmap->n,cm=C->rmap->n,pon=(p->B)->cmap->n;
MPI_Comm comm;
PetscMPIInt size,rank,taga,*len_s;
PetscInt *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci;
PetscInt **buf_ri,**buf_rj;
PetscInt cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */
MPI_Request *s_waits,*r_waits;
MPI_Status *status;
MatScalar **abuf_r,*ba_i,*pA,*coa,*ba;
PetscInt *api,*apj,*coi,*coj;
PetscInt *poJ=po->j,*pdJ=pd->j,pcstart=P->cmap->rstart,pcend=P->cmap->rend;
PetscBool scalable;
#if defined(PTAP_PROFILE)
PetscLogDouble t0,t1,t2,t3,t4,et2_AP=0.0,et2_PtAP=0.0,t2_0,t2_1,t2_2;
#endif
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)C,&comm);CHKERRQ(ierr);
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t0);CHKERRQ(ierr);
#endif
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ptap = c->ptap;
if (!ptap) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_INCOMP,"MatPtAP() has not been called to create matrix C yet, cannot use MAT_REUSE_MATRIX");
merge = ptap->merge;
apa = ptap->apa;
scalable = ptap->scalable;
/* 1) get P_oth = ptap->P_oth and P_loc = ptap->P_loc */
/*--------------------------------------------------*/
if (ptap->reuse == MAT_INITIAL_MATRIX) {
/* P_oth and P_loc are obtained in MatPtASymbolic(), skip calling MatGetBrowsOfAoCols() and MatMPIAIJGetLocalMat() */
ptap->reuse = MAT_REUSE_MATRIX;
} else { /* update numerical values of P_oth and P_loc */
ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_REUSE_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr);
ierr = MatMPIAIJGetLocalMat(P,MAT_REUSE_MATRIX,&ptap->P_loc);CHKERRQ(ierr);
}
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t1);CHKERRQ(ierr);
#endif
/* 2) compute numeric C_seq = P_loc^T*A_loc*P - dominating part */
/*--------------------------------------------------------------*/
/* get data from symbolic products */
p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data;
p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data;
pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a;
pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a;
coi = merge->coi; coj = merge->coj;
ierr = PetscCalloc1(coi[pon]+1,&coa);CHKERRQ(ierr);
bi = merge->bi; bj = merge->bj;
owners = merge->rowmap->range;
ierr = PetscCalloc1(bi[cm]+1,&ba);CHKERRQ(ierr); /* ba: Cseq->a */
api = ptap->api; apj = ptap->apj;
if (!scalable) { /* Do dense axpy on apa (length of pN, stores A[i,:]*P) - nonscalable, but fast */
ierr = PetscInfo(C,"Using non-scalable dense axpy\n");CHKERRQ(ierr);
/*-----------------------------------------------------------------------------------------------------*/
for (i=0; i<am; i++) {
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2_0);CHKERRQ(ierr);
#endif
/* 2-a) form i-th sparse row of A_loc*P = Ad*P_loc + Ao*P_oth */
/*------------------------------------------------------------*/
apJ = apj + api[i];
/* diagonal portion of A */
anz = adi[i+1] - adi[i];
adj = ad->j + adi[i];
ada = ad->a + adi[i];
for (j=0; j<anz; j++) {
row = adj[j];
pnz = pi_loc[row+1] - pi_loc[row];
pj = pj_loc + pi_loc[row];
pa = pa_loc + pi_loc[row];
/* perform dense axpy */
valtmp = ada[j];
for (k=0; k<pnz; k++) {
apa[pj[k]] += valtmp*pa[k];
}
ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr);
}
/* off-diagonal portion of A */
anz = aoi[i+1] - aoi[i];
aoj = ao->j + aoi[i];
aoa = ao->a + aoi[i];
for (j=0; j<anz; j++) {
row = aoj[j];
pnz = pi_oth[row+1] - pi_oth[row];
pj = pj_oth + pi_oth[row];
pa = pa_oth + pi_oth[row];
/* perform dense axpy */
valtmp = aoa[j];
for (k=0; k<pnz; k++) {
apa[pj[k]] += valtmp*pa[k];
}
ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr);
}
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2_1);CHKERRQ(ierr);
et2_AP += t2_1 - t2_0;
#endif
/* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */
/*--------------------------------------------------------------*/
apnz = api[i+1] - api[i];
/* put the value into Co=(p->B)^T*AP (off-diagonal part, send to others) */
pnz = po->i[i+1] - po->i[i];
poJ = po->j + po->i[i];
pA = po->a + po->i[i];
for (j=0; j<pnz; j++) {
row = poJ[j];
cnz = coi[row+1] - coi[row];
cj = coj + coi[row];
ca = coa + coi[row];
/* perform dense axpy */
valtmp = pA[j];
for (k=0; k<cnz; k++) {
ca[k] += valtmp*apa[cj[k]];
}
ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr);
}
/* put the value into Cd (diagonal part) */
pnz = pd->i[i+1] - pd->i[i];
pdJ = pd->j + pd->i[i];
pA = pd->a + pd->i[i];
for (j=0; j<pnz; j++) {
row = pdJ[j];
cnz = bi[row+1] - bi[row];
cj = bj + bi[row];
ca = ba + bi[row];
/* perform dense axpy */
valtmp = pA[j];
for (k=0; k<cnz; k++) {
ca[k] += valtmp*apa[cj[k]];
}
ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr);
}
/* zero the current row of A*P */
for (k=0; k<apnz; k++) apa[apJ[k]] = 0.0;
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2_2);CHKERRQ(ierr);
et2_PtAP += t2_2 - t2_1;
#endif
}
} else { /* Do sparse axpy on apa (length of ap_rmax, stores A[i,:]*P) - scalable, but slower */
ierr = PetscInfo(C,"Using scalable sparse axpy\n");CHKERRQ(ierr);
/*-----------------------------------------------------------------------------------------*/
pA=pa_loc;
for (i=0; i<am; i++) {
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2_0);CHKERRQ(ierr);
#endif
/* form i-th sparse row of A*P */
apnz = api[i+1] - api[i];
apJ = apj + api[i];
/* diagonal portion of A */
anz = adi[i+1] - adi[i];
adj = ad->j + adi[i];
ada = ad->a + adi[i];
for (j=0; j<anz; j++) {
row = adj[j];
pnz = pi_loc[row+1] - pi_loc[row];
pj = pj_loc + pi_loc[row];
pa = pa_loc + pi_loc[row];
valtmp = ada[j];
nextp = 0;
for (k=0; nextp<pnz; k++) {
if (apJ[k] == pj[nextp]) { /* col of AP == col of P */
apa[k] += valtmp*pa[nextp++];
}
}
ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr);
}
/* off-diagonal portion of A */
anz = aoi[i+1] - aoi[i];
aoj = ao->j + aoi[i];
aoa = ao->a + aoi[i];
for (j=0; j<anz; j++) {
row = aoj[j];
pnz = pi_oth[row+1] - pi_oth[row];
pj = pj_oth + pi_oth[row];
pa = pa_oth + pi_oth[row];
valtmp = aoa[j];
nextp = 0;
for (k=0; nextp<pnz; k++) {
if (apJ[k] == pj[nextp]) { /* col of AP == col of P */
apa[k] += valtmp*pa[nextp++];
}
}
ierr = PetscLogFlops(2.0*pnz);CHKERRQ(ierr);
}
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2_1);CHKERRQ(ierr);
et2_AP += t2_1 - t2_0;
#endif
/* 2-b) Compute Cseq = P_loc[i,:]^T*AP[i,:] using outer product */
/*--------------------------------------------------------------*/
pnz = pi_loc[i+1] - pi_loc[i];
pJ = pj_loc + pi_loc[i];
for (j=0; j<pnz; j++) {
nextap = 0;
row = pJ[j]; /* global index */
if (row < pcstart || row >=pcend) { /* put the value into Co */
row = *poJ;
cj = coj + coi[row];
ca = coa + coi[row]; poJ++;
} else { /* put the value into Cd */
row = *pdJ;
cj = bj + bi[row];
ca = ba + bi[row]; pdJ++;
}
valtmp = pA[j];
for (k=0; nextap<apnz; k++) {
if (cj[k]==apJ[nextap]) ca[k] += valtmp*apa[nextap++];
}
ierr = PetscLogFlops(2.0*apnz);CHKERRQ(ierr);
}
pA += pnz;
/* zero the current row info for A*P */
ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr);
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2_2);CHKERRQ(ierr);
et2_PtAP += t2_2 - t2_1;
#endif
}
}
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2);CHKERRQ(ierr);
#endif
/* 3) send and recv matrix values coa */
/*------------------------------------*/
buf_ri = merge->buf_ri;
buf_rj = merge->buf_rj;
len_s = merge->len_s;
ierr = PetscCommGetNewTag(comm,&taga);CHKERRQ(ierr);
ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr);
ierr = PetscMalloc2(merge->nsend+1,&s_waits,size,&status);CHKERRQ(ierr);
for (proc=0,k=0; proc<size; proc++) {
if (!len_s[proc]) continue;
i = merge->owners_co[proc];
ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr);
k++;
}
if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);}
if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);}
ierr = PetscFree2(s_waits,status);CHKERRQ(ierr);
ierr = PetscFree(r_waits);CHKERRQ(ierr);
ierr = PetscFree(coa);CHKERRQ(ierr);
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t3);CHKERRQ(ierr);
#endif
/* 4) insert local Cseq and received values into Cmpi */
/*------------------------------------------------------*/
ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr);
for (k=0; k<merge->nrecv; k++) {
buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
nrows = *(buf_ri_k[k]);
nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */
nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
}
for (i=0; i<cm; i++) {
row = owners[rank] + i; /* global row index of C_seq */
bj_i = bj + bi[i]; /* col indices of the i-th row of C */
ba_i = ba + bi[i];
bnz = bi[i+1] - bi[i];
/* add received vals into ba */
for (k=0; k<merge->nrecv; k++) { /* k-th received message */
/* i-th row */
if (i == *nextrow[k]) {
cnz = *(nextci[k]+1) - *nextci[k];
cj = buf_rj[k] + *(nextci[k]);
ca = abuf_r[k] + *(nextci[k]);
nextcj = 0;
for (j=0; nextcj<cnz; j++) {
if (bj_i[j] == cj[nextcj]) { /* bcol == ccol */
ba_i[j] += ca[nextcj++];
}
}
nextrow[k]++; nextci[k]++;
ierr = PetscLogFlops(2.0*cnz);CHKERRQ(ierr);
}
}
ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscFree(ba);CHKERRQ(ierr);
ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr);
ierr = PetscFree(abuf_r);CHKERRQ(ierr);
ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr);
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t4);CHKERRQ(ierr);
if (rank==1) PetscPrintf(MPI_COMM_SELF," [%d] PtAPNum %g/P + %g/PtAP( %g + %g ) + %g/comm + %g/Cloc = %g\n\n",rank,t1-t0,t2-t1,et2_AP,et2_PtAP,t3-t2,t4-t3,t4-t0);CHKERRQ(ierr);
#endif
PetscFunctionReturn(0);
}
