PetscErrorCode MatSolve_SeqSBSTRM_4_inplace(Mat A,Vec bb,Vec xx)
{
Mat_SeqSBAIJ *a =(Mat_SeqSBAIJ*)A->data;
IS isrow=a->row;
PetscInt mbs =a->mbs,*ai=a->i,*aj=a->j,bs=A->rmap->bs,bs2=a->bs2;
PetscErrorCode ierr;
const PetscInt *r;
PetscInt nz,*vj,k,idx;
PetscScalar *x,*b,x0,x1,x2,x3,*t,*tp;
Mat_SeqSBSTRM *sbstrm = (Mat_SeqSBSTRM*)A->spptr;
MatScalar *as =sbstrm->as,*diag;
PetscScalar tp0, tp1, tp2, tp3;
const MatScalar *v0, *v1, *v2, *v3;
PetscInt slen;
PetscFunctionBegin;
slen = 4*(ai[mbs]-ai[0]);
v0 = as + 16*ai[0];
v1 = v0 + slen;
v2 = v1 + slen;
v3 = v2 + slen;
ierr = VecGetArray(bb,&b);CHKERRQ(ierr);
ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
t = a->solve_work;
ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr);
/* solve U^T * D * y = b by forward substitution */
tp = t;
for (k=0; k<mbs; k++) { /* t <- perm(b) */
idx = 4*r[k];
tp[0] = b[idx];
tp[1] = b[idx+1];
tp[2] = b[idx+2];
tp[3] = b[idx+3];
tp += 4;
}
for (k=0; k<mbs; k++) {
vj = aj + ai[k];
tp = t + k*4;
x0 = tp[0]; x1=tp[1]; x2=tp[2]; x3=tp[3];
nz = ai[k+1] - ai[k];
tp = t + (*vj)*4;
while (nz--) {
tp[0] += v0[0]*x0 + v1[0]*x1 + v2[0]*x2 + v3[0]*x3;
tp[1] += v0[1]*x0 + v1[1]*x1 + v2[1]*x2 + v3[1]*x3;
tp[2] += v0[2]*x0 + v1[2]*x1 + v2[2]*x2 + v3[2]*x3;
tp[3] += v0[3]*x0 + v1[3]*x1 + v2[3]*x2 + v3[3]*x3;
vj++; tp = t + (*vj)*4;
v0 += 4; v1 += 4; v2 += 4; v3 += 4;
}
/* xk = inv(Dk)*(Dk*xk) */
diag = as+k*16; /* ptr to inv(Dk) */
tp = t + k*4;
tp[0] = diag[0]*x0 + diag[4]*x1 + diag[8]*x2 + diag[12]*x3;
tp[1] = diag[1]*x0 + diag[5]*x1 + diag[9]*x2 + diag[13]*x3;
tp[2] = diag[2]*x0 + diag[6]*x1 + diag[10]*x2+ diag[14]*x3;
tp[3] = diag[3]*x0 + diag[7]*x1 + diag[11]*x2+ diag[15]*x3;
}
/* solve U*x = y by back substitution */
for (k=mbs-1; k>=0; k--) {
vj = aj + ai[k+1];
tp = t + k*4;
x0 = tp[0]; x1=tp[1]; x2=tp[2]; x3=tp[3]; /* xk */
nz = ai[k+1] - ai[k];
tp = t + (*vj)*4;
while (nz--) {
/* xk += U(k,* */
v0 -= 4; v1 -= 4; v2 -= 4; v3 -= 4;
vj--; tp = t + (*vj)*4;
tp0 = tp[0]; tp1 = tp[1]; tp2 = tp[2]; tp3 = tp[3];
x0 += v0[3]*tp3 + v0[2]*tp2 + v0[1]*tp1 + v0[0]*tp0;
x1 += v1[3]*tp3 + v1[2]*tp2 + v1[1]*tp1 + v1[0]*tp0;
x2 += v2[3]*tp3 + v2[2]*tp2 + v2[1]*tp1 + v2[0]*tp0;
x3 += v3[3]*tp3 + v3[2]*tp2 + v3[1]*tp1 + v3[0]*tp0;
}
tp = t + k*4;
tp[0] = x0; tp[1]=x1; tp[2]=x2; tp[3]=x3;
idx = 4*r[k];
x[idx] = x0;
x[idx+1] = x1;
x[idx+2] = x2;
x[idx+3] = x3;
}
ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr);
ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr);
ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
ierr = PetscLogFlops(4.0*bs2*a->nz - (bs+2.0*bs2)*mbs);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*
This function is used before applying a
symmetric reordering to matrix A that is
in SBAIJ format.
The permutation is assumed to be symmetric, i.e.,
P = P^T (= inv(P)),
so the permuted matrix P*A*inv(P)=P*A*P^T is ensured to be symmetric.
- a wrong assumption! This code needs rework! -- Hong
The function is modified from sro.f of YSMP. The description from YSMP:
C THE NONZERO ENTRIES OF THE MATRIX M ARE ASSUMED TO BE STORED
C SYMMETRICALLY IN (IA,JA,A) FORMAT (I.E., NOT BOTH M(I,J) AND M(J,I)
C ARE STORED IF I NE J).
C
C SRO DOES NOT REARRANGE THE ORDER OF THE ROWS, BUT DOES MOVE
C NONZEROES FROM ONE ROW TO ANOTHER TO ENSURE THAT IF M(I,J) WILL BE
C IN THE UPPER TRIANGLE OF M WITH RESPECT TO THE NEW ORDERING, THEN
C M(I,J) IS STORED IN ROW I (AND THUS M(J,I) IS NOT STORED); WHEREAS
C IF M(I,J) WILL BE IN THE STRICT LOWER TRIANGLE OF M, THEN M(J,I) IS
C STORED IN ROW J (AND THUS M(I,J) IS NOT STORED).
-- output: new index set (inew, jnew) for A and a map a2anew that maps
values a to anew, such that all
nonzero A_(perm(i),iperm(k)) will be stored in the upper triangle.
Note: matrix A is not permuted by this function!
*/
PetscErrorCode MatReorderingSeqSBAIJ(Mat A,IS perm)
{
Mat_SeqSBAIJ *a=(Mat_SeqSBAIJ*)A->data;
const PetscInt mbs=a->mbs;
PetscFunctionBegin;
if (!mbs) PetscFunctionReturn(0);
SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Matrix reordering is not supported for sbaij matrix. Use aij format");
#if 0
PetscErrorCode ierr;
const PetscInt *rip,*riip;
PetscInt *ai,*aj,*r;
PetscInt *nzr,nz,jmin,jmax,j,k,ajk,i;
IS iperm; /* inverse of perm */
ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr);
ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr);
ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr);
for (i=0; i<mbs; i++) {
if (rip[i] != riip[i]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Non-symmetric permutation, use symmetric permutation for symmetric matrices");
}
ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr);
ierr = ISDestroy(&iperm);CHKERRQ(ierr);
if (!a->inew) {
ierr = PetscMalloc2(mbs+1,&ai, 2*a->i[mbs],&aj);CHKERRQ(ierr);
} else {
ai = a->inew; aj = a->jnew;
}
ierr = PetscMemcpy(ai,a->i,(mbs+1)*sizeof(PetscInt));CHKERRQ(ierr);
ierr = PetscMemcpy(aj,a->j,(a->i[mbs])*sizeof(PetscInt));CHKERRQ(ierr);
/*
Phase 1: Find row index r in which to store each nonzero.
Initialize count of nonzeros to be stored in each row (nzr).
At the end of this phase, a nonzero a(*,*)=a(r(),aj())
s.t. a(perm(r),perm(aj)) will fall into upper triangle part.
*/
ierr = PetscMalloc1(mbs,&nzr);CHKERRQ(ierr);
ierr = PetscMalloc1(ai[mbs],&r);CHKERRQ(ierr);
for (i=0; i<mbs; i++) nzr[i] = 0;
for (i=0; i<ai[mbs]; i++) r[i] = 0;
/* for each nonzero element */
for (i=0; i<mbs; i++) {
nz = ai[i+1] - ai[i];
j = ai[i];
/* printf("nz = %d, j=%d\n",nz,j); */
while (nz--) {
/* --- find row (=r[j]) and column (=aj[j]) in which to store a[j] ...*/
k = aj[j]; /* col. index */
/* printf("nz = %d, k=%d\n", nz,k); */
/* for entry that will be permuted into lower triangle, swap row and col. index */
if (rip[k] < rip[i]) aj[j] = i;
else k = i;
r[j] = k; j++;
nzr[k]++; /* increment count of nonzeros in that row */
}
}
/* Phase 2: Find new ai and permutation to apply to (aj,a).
Determine pointers (r) to delimit rows in permuted (aj,a).
Note: r is different from r used in phase 1.
At the end of this phase, (aj[j],a[j]) will be stored in
(aj[r(j)],a[r(j)]).
*/
for (i=0; i<mbs; i++) {
ai[i+1] = ai[i] + nzr[i];
nzr[i] = ai[i+1];
}
/* determine where each (aj[j], a[j]) is stored in new (aj,a)
for each nonzero element (in reverse order) */
jmin = ai[0]; jmax = ai[mbs];
nz = jmax - jmin;
j = jmax-1;
while (nz--) {
i = r[j]; /* row value */
if (aj[j] == i) r[j] = ai[i]; /* put diagonal nonzero at beginning of row */
else { /* put off-diagonal nonzero in last unused location in row */
nzr[i]--; r[j] = nzr[i];
}
j--;
}
a->a2anew = aj + ai[mbs];
ierr = PetscMemcpy(a->a2anew,r,ai[mbs]*sizeof(PetscInt));CHKERRQ(ierr);
/* Phase 3: permute (aj,a) to upper triangular form (wrt new ordering) */
for (j=jmin; j<jmax; j++) {
while (r[j] != j) {
k = r[j]; r[j] = r[k]; r[k] = k;
ajk = aj[k]; aj[k] = aj[j]; aj[j] = ajk;
/* ak = aa[k]; aa[k] = aa[j]; aa[j] = ak; */
}
}
ierr= ISRestoreIndices(perm,&rip);CHKERRQ(ierr);
a->inew = ai;
a->jnew = aj;
ierr = ISDestroy(&a->row);CHKERRQ(ierr);
ierr = ISDestroy(&a->icol);CHKERRQ(ierr);
ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
ierr = ISDestroy(&a->row);CHKERRQ(ierr);
a->row = perm;
ierr = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
ierr = ISDestroy(&a->icol);CHKERRQ(ierr);
a->icol = perm;
ierr = PetscFree(nzr);CHKERRQ(ierr);
ierr = PetscFree(r);CHKERRQ(ierr);
PetscFunctionReturn(0);
#endif
}
static PetscErrorCode MatIncreaseOverlap_MPISBAIJ_Once(Mat C,PetscInt is_max,IS is[])
{
Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
PetscErrorCode ierr;
PetscMPIInt size,rank,tag1,tag2,*len_s,nrqr,nrqs,*id_r1,*len_r1,flag,len,*iwork;
const PetscInt *idx_i;
PetscInt idx,isz,col,*n,*data1,**data1_start,*data2,*data2_i,*data,*data_i,
Mbs,i,j,k,*odata1,*odata2,
proc_id,**odata2_ptr,*ctable=0,*btable,len_max,len_est;
PetscInt proc_end=0,len_unused,nodata2;
PetscInt ois_max; /* max no of is[] in each of processor */
char *t_p;
MPI_Comm comm;
MPI_Request *s_waits1,*s_waits2,r_req;
MPI_Status *s_status,r_status;
PetscBT *table; /* mark indices of this processor's is[] */
PetscBT table_i;
PetscBT otable; /* mark indices of other processors' is[] */
PetscInt bs=C->rmap->bs,Bn = c->B->cmap->n,Bnbs = Bn/bs,*Bowners;
IS garray_local,garray_gl;
PetscFunctionBegin;
comm = ((PetscObject)C)->comm;
size = c->size;
rank = c->rank;
Mbs = c->Mbs;
ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr);
ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr);
/* create tables used in
step 1: table[i] - mark c->garray of proc [i]
step 3: table[i] - mark indices of is[i] when whose=MINE
table[0] - mark incideces of is[] when whose=OTHER */
len = PetscMax(is_max, size);CHKERRQ(ierr);
ierr = PetscMalloc2(len,PetscBT,&table,(Mbs/PETSC_BITS_PER_BYTE+1)*len,char,&t_p);CHKERRQ(ierr);
for (i=0; i<len; i++) {
table[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i;
}
ierr = MPI_Allreduce(&is_max,&ois_max,1,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr);
/* 1. Send this processor's is[] to other processors */
/*---------------------------------------------------*/
/* allocate spaces */
ierr = PetscMalloc(is_max*sizeof(PetscInt),&n);CHKERRQ(ierr);
len = 0;
for (i=0; i<is_max; i++) {
ierr = ISGetLocalSize(is[i],&n[i]);CHKERRQ(ierr);
len += n[i];
}
if (!len) {
is_max = 0;
} else {
len += 1 + is_max; /* max length of data1 for one processor */
}
ierr = PetscMalloc((size*len+1)*sizeof(PetscInt),&data1);CHKERRQ(ierr);
ierr = PetscMalloc(size*sizeof(PetscInt*),&data1_start);CHKERRQ(ierr);
for (i=0; i<size; i++) data1_start[i] = data1 + i*len;
ierr = PetscMalloc4(size,PetscInt,&len_s,size,PetscInt,&btable,size,PetscMPIInt,&iwork,size+1,PetscInt,&Bowners);CHKERRQ(ierr);
/* gather c->garray from all processors */
ierr = ISCreateGeneral(comm,Bnbs,c->garray,PETSC_COPY_VALUES,&garray_local);CHKERRQ(ierr);
ierr = ISAllGather(garray_local, &garray_gl);CHKERRQ(ierr);
ierr = ISDestroy(&garray_local);CHKERRQ(ierr);
ierr = MPI_Allgather(&Bnbs,1,MPIU_INT,Bowners+1,1,MPIU_INT,comm);CHKERRQ(ierr);
Bowners[0] = 0;
for (i=0; i<size; i++) Bowners[i+1] += Bowners[i];
if (is_max){
/* hash table ctable which maps c->row to proc_id) */
ierr = PetscMalloc(Mbs*sizeof(PetscInt),&ctable);CHKERRQ(ierr);
for (proc_id=0,j=0; proc_id<size; proc_id++) {
for (; j<C->rmap->range[proc_id+1]/bs; j++) {
ctable[j] = proc_id;
}
}
/* hash tables marking c->garray */
ierr = ISGetIndices(garray_gl,&idx_i);CHKERRQ(ierr);
for (i=0; i<size; i++){
table_i = table[i];
ierr = PetscBTMemzero(Mbs,table_i);CHKERRQ(ierr);
for (j = Bowners[i]; j<Bowners[i+1]; j++){ /* go through B cols of proc[i]*/
ierr = PetscBTSet(table_i,idx_i[j]);CHKERRQ(ierr);
}
}
ierr = ISRestoreIndices(garray_gl,&idx_i);CHKERRQ(ierr);
} /* if (is_max) */
ierr = ISDestroy(&garray_gl);CHKERRQ(ierr);
/* evaluate communication - mesg to who, length, and buffer space */
for (i=0; i<size; i++) len_s[i] = 0;
/* header of data1 */
for (proc_id=0; proc_id<size; proc_id++){
iwork[proc_id] = 0;
*data1_start[proc_id] = is_max;
data1_start[proc_id]++;
for (j=0; j<is_max; j++) {
if (proc_id == rank){
*data1_start[proc_id] = n[j];
} else {
*data1_start[proc_id] = 0;
}
data1_start[proc_id]++;
}
}
for (i=0; i<is_max; i++) {
ierr = ISGetIndices(is[i],&idx_i);CHKERRQ(ierr);
for (j=0; j<n[i]; j++){
idx = idx_i[j];
*data1_start[rank] = idx; data1_start[rank]++; /* for local proccessing */
proc_end = ctable[idx];
for (proc_id=0; proc_id<=proc_end; proc_id++){ /* for others to process */
if (proc_id == rank ) continue; /* done before this loop */
if (proc_id < proc_end && !PetscBTLookup(table[proc_id],idx))
continue; /* no need for sending idx to [proc_id] */
*data1_start[proc_id] = idx; data1_start[proc_id]++;
len_s[proc_id]++;
}
}
/* update header data */
for (proc_id=0; proc_id<size; proc_id++){
if (proc_id== rank) continue;
*(data1 + proc_id*len + 1 + i) = len_s[proc_id] - iwork[proc_id];
iwork[proc_id] = len_s[proc_id] ;
}
ierr = ISRestoreIndices(is[i],&idx_i);CHKERRQ(ierr);
}
nrqs = 0; nrqr = 0;
for (i=0; i<size; i++){
data1_start[i] = data1 + i*len;
if (len_s[i]){
nrqs++;
len_s[i] += 1 + is_max; /* add no. of header msg */
}
}
for (i=0; i<is_max; i++) {
ierr = ISDestroy(&is[i]);CHKERRQ(ierr);
}
ierr = PetscFree(n);CHKERRQ(ierr);
ierr = PetscFree(ctable);CHKERRQ(ierr);
/* Determine the number of messages to expect, their lengths, from from-ids */
ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&nrqr);CHKERRQ(ierr);
ierr = PetscGatherMessageLengths(comm,nrqs,nrqr,len_s,&id_r1,&len_r1);CHKERRQ(ierr);
/* Now post the sends */
ierr = PetscMalloc2(size,MPI_Request,&s_waits1,size,MPI_Request,&s_waits2);CHKERRQ(ierr);
k = 0;
for (proc_id=0; proc_id<size; proc_id++){ /* send data1 to processor [proc_id] */
if (len_s[proc_id]){
ierr = MPI_Isend(data1_start[proc_id],len_s[proc_id],MPIU_INT,proc_id,tag1,comm,s_waits1+k);CHKERRQ(ierr);
k++;
}
}
/* 2. Receive other's is[] and process. Then send back */
/*-----------------------------------------------------*/
len = 0;
for (i=0; i<nrqr; i++){
if (len_r1[i] > len)len = len_r1[i];
}
ierr = PetscFree(len_r1);CHKERRQ(ierr);
ierr = PetscFree(id_r1);CHKERRQ(ierr);
for (proc_id=0; proc_id<size; proc_id++)
len_s[proc_id] = iwork[proc_id] = 0;
ierr = PetscMalloc((len+1)*sizeof(PetscInt),&odata1);CHKERRQ(ierr);
ierr = PetscMalloc(size*sizeof(PetscInt**),&odata2_ptr);CHKERRQ(ierr);
ierr = PetscBTCreate(Mbs,&otable);CHKERRQ(ierr);
len_max = ois_max*(Mbs+1); /* max space storing all is[] for each receive */
len_est = 2*len_max; /* estimated space of storing is[] for all receiving messages */
ierr = PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);CHKERRQ(ierr);
nodata2 = 0; /* nodata2+1: num of PetscMalloc(,&odata2_ptr[]) called */
odata2_ptr[nodata2] = odata2;
len_unused = len_est; /* unused space in the array odata2_ptr[nodata2]-- needs to be >= len_max */
k = 0;
while (k < nrqr){
/* Receive messages */
ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status);CHKERRQ(ierr);
if (flag){
ierr = MPI_Get_count(&r_status,MPIU_INT,&len);CHKERRQ(ierr);
proc_id = r_status.MPI_SOURCE;
ierr = MPI_Irecv(odata1,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);CHKERRQ(ierr);
ierr = MPI_Wait(&r_req,&r_status);CHKERRQ(ierr);
/* Process messages */
/* make sure there is enough unused space in odata2 array */
if (len_unused < len_max){ /* allocate more space for odata2 */
ierr = PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);CHKERRQ(ierr);
odata2_ptr[++nodata2] = odata2;
len_unused = len_est;
}
ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,odata1,OTHER,odata2,&otable);CHKERRQ(ierr);
len = 1 + odata2[0];
for (i=0; i<odata2[0]; i++){
len += odata2[1 + i];
}
/* Send messages back */
ierr = MPI_Isend(odata2,len,MPIU_INT,proc_id,tag2,comm,s_waits2+k);CHKERRQ(ierr);
k++;
odata2 += len;
len_unused -= len;
len_s[proc_id] = len; /* num of messages sending back to [proc_id] by this proc */
}
}
ierr = PetscFree(odata1);CHKERRQ(ierr);
ierr = PetscBTDestroy(&otable);CHKERRQ(ierr);
/* 3. Do local work on this processor's is[] */
/*-------------------------------------------*/
/* make sure there is enough unused space in odata2(=data) array */
len_max = is_max*(Mbs+1); /* max space storing all is[] for this processor */
if (len_unused < len_max){ /* allocate more space for odata2 */
ierr = PetscMalloc((len_est+1)*sizeof(PetscInt),&odata2);CHKERRQ(ierr);
odata2_ptr[++nodata2] = odata2;
}
data = odata2;
ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,data1_start[rank],MINE,data,table);CHKERRQ(ierr);
ierr = PetscFree(data1_start);CHKERRQ(ierr);
/* 4. Receive work done on other processors, then merge */
/*------------------------------------------------------*/
/* get max number of messages that this processor expects to recv */
ierr = MPI_Allreduce(len_s,iwork,size,MPI_INT,MPI_MAX,comm);CHKERRQ(ierr);
ierr = PetscMalloc((iwork[rank]+1)*sizeof(PetscInt),&data2);CHKERRQ(ierr);
ierr = PetscFree4(len_s,btable,iwork,Bowners);CHKERRQ(ierr);
k = 0;
while (k < nrqs){
/* Receive messages */
ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status);CHKERRQ(ierr);
if (flag){
ierr = MPI_Get_count(&r_status,MPIU_INT,&len);CHKERRQ(ierr);
proc_id = r_status.MPI_SOURCE;
ierr = MPI_Irecv(data2,len,MPIU_INT,proc_id,r_status.MPI_TAG,comm,&r_req);CHKERRQ(ierr);
ierr = MPI_Wait(&r_req,&r_status);CHKERRQ(ierr);
if (len > 1+is_max){ /* Add data2 into data */
data2_i = data2 + 1 + is_max;
for (i=0; i<is_max; i++){
table_i = table[i];
data_i = data + 1 + is_max + Mbs*i;
isz = data[1+i];
for (j=0; j<data2[1+i]; j++){
col = data2_i[j];
if (!PetscBTLookupSet(table_i,col)) {data_i[isz++] = col;}
}
data[1+i] = isz;
if (i < is_max - 1) data2_i += data2[1+i];
}
}
k++;
}
}
ierr = PetscFree(data2);CHKERRQ(ierr);
ierr = PetscFree2(table,t_p);CHKERRQ(ierr);
/* phase 1 sends are complete */
ierr = PetscMalloc(size*sizeof(MPI_Status),&s_status);CHKERRQ(ierr);
if (nrqs) {ierr = MPI_Waitall(nrqs,s_waits1,s_status);CHKERRQ(ierr);}
ierr = PetscFree(data1);CHKERRQ(ierr);
/* phase 2 sends are complete */
if (nrqr){ierr = MPI_Waitall(nrqr,s_waits2,s_status);CHKERRQ(ierr);}
ierr = PetscFree2(s_waits1,s_waits2);CHKERRQ(ierr);
ierr = PetscFree(s_status);CHKERRQ(ierr);
/* 5. Create new is[] */
/*--------------------*/
for (i=0; i<is_max; i++) {
data_i = data + 1 + is_max + Mbs*i;
ierr = ISCreateGeneral(PETSC_COMM_SELF,data[1+i],data_i,PETSC_COPY_VALUES,is+i);CHKERRQ(ierr);
}
for (k=0; k<=nodata2; k++){
ierr = PetscFree(odata2_ptr[k]);CHKERRQ(ierr);
}
ierr = PetscFree(odata2_ptr);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatIncreaseOverlap_MPISBAIJ(Mat C,PetscInt is_max,IS is[],PetscInt ov)
{
PetscErrorCode ierr;
PetscInt i,N=C->cmap->N, bs=C->rmap->bs,M=C->rmap->N,Mbs=M/bs,*nidx,isz,iov;
IS *is_new,*is_row;
Mat *submats;
Mat_MPISBAIJ *c=(Mat_MPISBAIJ*)C->data;
Mat_SeqSBAIJ *asub_i;
PetscBT table;
PetscInt *ai,brow,nz,nis,l,nmax,nstages_local,nstages,max_no,pos;
const PetscInt *idx;
PetscBool flg,*allcolumns,*allrows;
PetscFunctionBegin;
ierr = PetscMalloc(is_max*sizeof(IS),&is_new);CHKERRQ(ierr);
/* Convert the indices into block format */
ierr = ISCompressIndicesGeneral(N,C->rmap->n,bs,is_max,is,is_new);CHKERRQ(ierr);
if (ov < 0){ SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");}
/* ----- previous non-scalable implementation ----- */
flg=PETSC_FALSE;
ierr = PetscOptionsHasName(PETSC_NULL, "-IncreaseOverlap_old", &flg);CHKERRQ(ierr);
if (flg){ /* previous non-scalable implementation */
printf("use previous non-scalable implementation...\n");
for (i=0; i<ov; ++i) {
ierr = MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);CHKERRQ(ierr);
}
} else { /* implementation using modified BAIJ routines */
ierr = PetscMalloc((Mbs+1)*sizeof(PetscInt),&nidx);CHKERRQ(ierr);
ierr = PetscBTCreate(Mbs,&table);CHKERRQ(ierr); /* for column search */
ierr = PetscMalloc2(is_max+1,PetscBool,&allcolumns,is_max+1,PetscBool,&allrows);CHKERRQ(ierr);
/* Create is_row */
ierr = PetscMalloc(is_max*sizeof(IS **),&is_row);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,Mbs,0,1,&is_row[0]);CHKERRQ(ierr);
allrows[0] = PETSC_TRUE;
for (i=1; i<is_max; i++) {
is_row[i] = is_row[0]; /* reuse is_row[0] */
allrows[i] = PETSC_TRUE;
}
/* Allocate memory to hold all the submatrices - Modified from MatGetSubMatrices_MPIBAIJ() */
ierr = PetscMalloc((is_max+1)*sizeof(Mat),&submats);CHKERRQ(ierr);
/* Check for special case: each processor gets entire matrix columns */
for (i=0; i<is_max; i++) {
ierr = ISIdentity(is_new[i],&flg);CHKERRQ(ierr);
ierr = ISGetLocalSize(is_new[i],&isz);CHKERRQ(ierr);
if (flg && isz == Mbs){
allcolumns[i] = PETSC_TRUE;
} else {
allcolumns[i] = PETSC_FALSE;
}
}
/* Determine the number of stages through which submatrices are done */
nmax = 20*1000000 / (c->Nbs * sizeof(PetscInt));
if (!nmax) nmax = 1;
nstages_local = is_max/nmax + ((is_max % nmax)?1:0);
/* Make sure every processor loops through the nstages */
ierr = MPI_Allreduce(&nstages_local,&nstages,1,MPIU_INT,MPI_MAX,((PetscObject)C)->comm);CHKERRQ(ierr);
for (iov=0; iov<ov; ++iov) {
/* 1) Get submats for column search */
for (i=0,pos=0; i<nstages; i++) {
if (pos+nmax <= is_max) max_no = nmax;
else if (pos == is_max) max_no = 0;
else max_no = is_max-pos;
c->ijonly = PETSC_TRUE;
ierr = MatGetSubMatrices_MPIBAIJ_local(C,max_no,is_row+pos,is_new+pos,MAT_INITIAL_MATRIX,allrows+pos,allcolumns+pos,submats+pos);CHKERRQ(ierr);
pos += max_no;
}
/* 2) Row search */
ierr = MatIncreaseOverlap_MPIBAIJ_Once(C,is_max,is_new);CHKERRQ(ierr);
/* 3) Column search */
for (i=0; i<is_max; i++){
asub_i = (Mat_SeqSBAIJ*)submats[i]->data;
ai=asub_i->i;;
/* put is_new obtained from MatIncreaseOverlap_MPIBAIJ() to table */
ierr = PetscBTMemzero(Mbs,table);CHKERRQ(ierr);
ierr = ISGetIndices(is_new[i],&idx);CHKERRQ(ierr);
ierr = ISGetLocalSize(is_new[i],&nis);CHKERRQ(ierr);
for (l=0; l<nis; l++) {
ierr = PetscBTSet(table,idx[l]);CHKERRQ(ierr);
nidx[l] = idx[l];
}
isz = nis;
/* add column entries to table */
for (brow=0; brow<Mbs; brow++){
nz = ai[brow+1] - ai[brow];
if (nz) {
if (!PetscBTLookupSet(table,brow)) nidx[isz++] = brow;
}
}
ierr = ISRestoreIndices(is_new[i],&idx);CHKERRQ(ierr);
ierr = ISDestroy(&is_new[i]);CHKERRQ(ierr);
/* create updated is_new */
ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,is_new+i);CHKERRQ(ierr);
}
/* Free tmp spaces */
for (i=0; i<is_max; i++){
ierr = MatDestroy(&submats[i]);CHKERRQ(ierr);
}
}
ierr = PetscFree2(allcolumns,allrows);CHKERRQ(ierr);
ierr = PetscBTDestroy(&table);CHKERRQ(ierr);
ierr = PetscFree(submats);CHKERRQ(ierr);
ierr = ISDestroy(&is_row[0]);CHKERRQ(ierr);
ierr = PetscFree(is_row);CHKERRQ(ierr);
ierr = PetscFree(nidx);CHKERRQ(ierr);
}
for (i=0; i<is_max; i++) {ierr = ISDestroy(&is[i]);CHKERRQ(ierr);}
ierr = ISExpandIndicesGeneral(N,N,bs,is_max,is_new,is);CHKERRQ(ierr);
for (i=0; i<is_max; i++) {ierr = ISDestroy(&is_new[i]);CHKERRQ(ierr);}
ierr = PetscFree(is_new);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PETSC_EXTERN PetscErrorCode AOCreate_Basic(AO ao)
{
AO_Basic *aobasic;
PetscMPIInt size,rank,count,*lens,*disp;
PetscInt napp,*allpetsc,*allapp,ip,ia,N,i,*petsc=NULL,start;
PetscErrorCode ierr;
IS isapp=ao->isapp,ispetsc=ao->ispetsc;
MPI_Comm comm;
const PetscInt *myapp,*mypetsc=NULL;
PetscFunctionBegin;
/* create special struct aobasic */
ierr = PetscNewLog(ao, AO_Basic, &aobasic);CHKERRQ(ierr);
ao->data = (void*) aobasic;
ierr = PetscMemcpy(ao->ops,&AOOps_Basic,sizeof(struct _AOOps));CHKERRQ(ierr);
ierr = PetscObjectChangeTypeName((PetscObject)ao,AOBASIC);CHKERRQ(ierr);
ierr = ISGetLocalSize(isapp,&napp);CHKERRQ(ierr);
ierr = ISGetIndices(isapp,&myapp);CHKERRQ(ierr);
ierr = PetscMPIIntCast(napp,&count);CHKERRQ(ierr);
/* transmit all lengths to all processors */
ierr = PetscObjectGetComm((PetscObject)isapp,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = PetscMalloc2(size,PetscMPIInt, &lens,size,PetscMPIInt,&disp);CHKERRQ(ierr);
ierr = MPI_Allgather(&count, 1, MPI_INT, lens, 1, MPI_INT, comm);CHKERRQ(ierr);
N = 0;
for (i = 0; i < size; i++) {
ierr = PetscMPIIntCast(N,disp+i);CHKERRQ(ierr); /* = sum(lens[j]), j< i */
N += lens[i];
}
ao->N = N;
ao->n = N;
/* If mypetsc is 0 then use "natural" numbering */
if (napp) {
if (!ispetsc) {
start = disp[rank];
ierr = PetscMalloc((napp+1) * sizeof(PetscInt), &petsc);CHKERRQ(ierr);
for (i=0; i<napp; i++) petsc[i] = start + i;
} else {
ierr = ISGetIndices(ispetsc,&mypetsc);CHKERRQ(ierr);
petsc = (PetscInt*)mypetsc;
}
}
/* get all indices on all processors */
ierr = PetscMalloc2(N,PetscInt,&allpetsc,N,PetscInt,&allapp);CHKERRQ(ierr);
ierr = MPI_Allgatherv(petsc, count, MPIU_INT, allpetsc, lens, disp, MPIU_INT, comm);CHKERRQ(ierr);
ierr = MPI_Allgatherv((void*)myapp, count, MPIU_INT, allapp, lens, disp, MPIU_INT, comm);CHKERRQ(ierr);
ierr = PetscFree2(lens,disp);CHKERRQ(ierr);
#if defined(PETSC_USE_DEBUG)
{
PetscInt *sorted;
ierr = PetscMalloc(N*sizeof(PetscInt),&sorted);CHKERRQ(ierr);
ierr = PetscMemcpy(sorted,allpetsc,N*sizeof(PetscInt));CHKERRQ(ierr);
ierr = PetscSortInt(N,sorted);CHKERRQ(ierr);
for (i=0; i<N; i++) {
if (sorted[i] != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"PETSc ordering requires a permutation of numbers 0 to N-1\n it is missing %D has %D",i,sorted[i]);
}
ierr = PetscMemcpy(sorted,allapp,N*sizeof(PetscInt));CHKERRQ(ierr);
ierr = PetscSortInt(N,sorted);CHKERRQ(ierr);
for (i=0; i<N; i++) {
if (sorted[i] != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Application ordering requires a permutation of numbers 0 to N-1\n it is missing %D has %D",i,sorted[i]);
}
ierr = PetscFree(sorted);CHKERRQ(ierr);
}
#endif
/* generate a list of application and PETSc node numbers */
ierr = PetscMalloc2(N,PetscInt, &aobasic->app,N,PetscInt,&aobasic->petsc);CHKERRQ(ierr);
ierr = PetscLogObjectMemory(ao,2*N*sizeof(PetscInt));CHKERRQ(ierr);
ierr = PetscMemzero(aobasic->app, N*sizeof(PetscInt));CHKERRQ(ierr);
ierr = PetscMemzero(aobasic->petsc, N*sizeof(PetscInt));CHKERRQ(ierr);
for (i = 0; i < N; i++) {
ip = allpetsc[i];
ia = allapp[i];
/* check there are no duplicates */
if (aobasic->app[ip]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Duplicate in PETSc ordering at position %d. Already mapped to %d, not %d.", i, aobasic->app[ip]-1, ia);
aobasic->app[ip] = ia + 1;
if (aobasic->petsc[ia]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Duplicate in Application ordering at position %d. Already mapped to %d, not %d.", i, aobasic->petsc[ia]-1, ip);
aobasic->petsc[ia] = ip + 1;
}
if (napp && !mypetsc) {
ierr = PetscFree(petsc);CHKERRQ(ierr);
}
ierr = PetscFree2(allpetsc,allapp);CHKERRQ(ierr);
/* shift indices down by one */
for (i = 0; i < N; i++) {
aobasic->app[i]--;
aobasic->petsc[i]--;
}
ierr = ISRestoreIndices(isapp,&myapp);CHKERRQ(ierr);
if (napp) {
if (ispetsc) {
ierr = ISRestoreIndices(ispetsc,&mypetsc);CHKERRQ(ierr);
} else {
ierr = PetscFree(petsc);CHKERRQ(ierr);
}
}
PetscFunctionReturn(0);
}
static PetscErrorCode permutematrix(Mat Ain, Mat Bin, Mat *Aout, Mat *Bout, int **permIndices)
{
PetscErrorCode ierr;
MatPartitioning part;
IS isn, is, iscols;
PetscInt *nlocal,localCols,m,n;
PetscMPIInt size, rank;
MPI_Comm comm;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Ain,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MatGetSize(Ain,&m,&n);CHKERRQ(ierr);
ierr = MatPartitioningCreate(comm,&part);CHKERRQ(ierr);
ierr = MatPartitioningSetAdjacency(part,Ain);CHKERRQ(ierr);
ierr = MatPartitioningSetFromOptions(part);CHKERRQ(ierr);
/* get new processor owner number of each vertex */
ierr = MatPartitioningApply(part,&is);CHKERRQ(ierr);
/* get new global number of each old global number */
ierr = ISPartitioningToNumbering(is,&isn);CHKERRQ(ierr);
ierr = PetscMalloc(size*sizeof(int),&nlocal);CHKERRQ(ierr);
/* get number of new vertices for each processor */
ierr = ISPartitioningCount(is,size,nlocal);CHKERRQ(ierr);
ierr = ISDestroy(&is);CHKERRQ(ierr);
/* get old global number of each new global number */
ierr = ISInvertPermutation(isn,nlocal[rank],&is);CHKERRQ(ierr);
ierr = ISDestroy(&isn);CHKERRQ(ierr);
ierr = MatPartitioningDestroy(&part);CHKERRQ(ierr);
ierr = ISSort(is);CHKERRQ(ierr);
/* If matrix is square, the permutation is applied to rows and columns;
otherwise it is only applied to rows. */
if (m == n) {
iscols = is;
localCols = nlocal[rank];
} else {
PetscInt lowj, highj;
ierr = MatGetOwnershipRangeColumn(Ain,&lowj,&highj);CHKERRQ(ierr);
localCols = highj-lowj;
ierr = ISCreateStride(comm,localCols, lowj, 1, &iscols);CHKERRQ(ierr);
}
/* copy permutation */
if (permIndices) {
const PetscInt *indices;
PetscInt i;
*permIndices = malloc(sizeof(int)*(nlocal[rank]+localCols));
ierr = ISGetIndices(is, &indices);CHKERRQ(ierr);
for (i=0; i<nlocal[rank]; i++) (*permIndices)[i] = indices[i];
ierr = ISRestoreIndices(is, &indices);CHKERRQ(ierr);
ierr = ISGetIndices(iscols, &indices);CHKERRQ(ierr);
for (i=0; i<localCols; i++) (*permIndices)[i+nlocal[rank]] = indices[i];
ierr = ISRestoreIndices(iscols, &indices);CHKERRQ(ierr);
}
ierr = PetscFree(nlocal);CHKERRQ(ierr);
ierr = MatGetSubMatrix(Ain,is,iscols,MAT_INITIAL_MATRIX,Aout);CHKERRQ(ierr);
if (Bin && Bout) {
ierr = MatGetSubMatrix(Bin,is,iscols,MAT_INITIAL_MATRIX,Bout);CHKERRQ(ierr);
}
ierr = ISDestroy(&is);CHKERRQ(ierr);
if (m != n) {
ierr = ISDestroy(&iscols);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}