/*--------------------------------------------------------------------------
* hypre_SStructSharedDOF_ParcsrMatRowsComm
* Given a sstruct_grid & parcsr matrix with rows corresponding to the
* sstruct_grid, determine and extract the rows that must be communicated.
* These rows are for shared dof that geometrically lie on processor
* boundaries but internally are stored on one processor.
* Algo:
* for each cellbox
* RECVs:
* i) stretch the cellbox to the variable box
* ii) in the appropriate (dof-dependent) direction, take the
* boundary and boxman_intersect to extract boxmanentries
* that contain these boundary edges.
* iii)loop over the boxmanentries and see if they belong
* on this proc or another proc
* a) if belong on another proc, these are the recvs:
* count and prepare the communication buffers and
* values.
*
* SENDs:
* i) form layer of cells that is one layer off cellbox
* (stretches in the appropriate direction)
* ii) boxman_intersect with the cellgrid boxman
* iii)loop over the boxmanentries and see if they belong
* on this proc or another proc
* a) if belong on another proc, these are the sends:
* count and prepare the communication buffers and
* values.
*
* Note: For the recv data, the dof can come from only one processor.
* For the send data, the dof can go to more than one processor
* (the same dof is on the boundary of several cells).
*--------------------------------------------------------------------------*/
HYPRE_Int
hypre_SStructSharedDOF_ParcsrMatRowsComm( hypre_SStructGrid *grid,
hypre_ParCSRMatrix *A,
HYPRE_Int *num_offprocrows_ptr,
hypre_MaxwellOffProcRow ***OffProcRows_ptr)
{
MPI_Comm A_comm= hypre_ParCSRMatrixComm(A);
MPI_Comm grid_comm= hypre_SStructGridComm(grid);
HYPRE_Int matrix_type= HYPRE_PARCSR;
HYPRE_Int nparts= hypre_SStructGridNParts(grid);
HYPRE_Int ndim = hypre_SStructGridNDim(grid);
hypre_SStructGrid *cell_ssgrid;
hypre_SStructPGrid *pgrid;
hypre_StructGrid *cellgrid;
hypre_BoxArray *cellboxes;
hypre_Box *box, *cellbox, vbox, boxman_entry_box;
hypre_Index loop_size, start;
HYPRE_Int loopi, loopj, loopk;
HYPRE_Int start_rank, end_rank, rank;
HYPRE_Int i, j, k, m, n, t, part, var, nvars;
HYPRE_SStructVariable *vartypes;
HYPRE_Int nbdry_slabs;
hypre_BoxArray *recv_slabs, *send_slabs;
hypre_Index varoffset;
hypre_BoxManager **boxmans, *cell_boxman;
hypre_BoxManEntry **boxman_entries, *entry;
HYPRE_Int nboxman_entries;
hypre_Index ishift, jshift, kshift, zero_index;
hypre_Index ilower, iupper, index;
HYPRE_Int proc, nprocs, myproc;
HYPRE_Int *SendToProcs, *RecvFromProcs;
HYPRE_Int **send_RowsNcols; /* buffer for rows & ncols */
HYPRE_Int *send_RowsNcols_alloc;
HYPRE_Int *send_ColsData_alloc;
HYPRE_Int *tot_nsendRowsNcols, *tot_sendColsData;
double **vals; /* buffer for cols & data */
HYPRE_Int *col_inds;
double *values;
hypre_MPI_Request *requests;
hypre_MPI_Status *status;
HYPRE_Int **rbuffer_RowsNcols;
double **rbuffer_ColsData;
HYPRE_Int num_sends, num_recvs;
hypre_MaxwellOffProcRow **OffProcRows;
HYPRE_Int *starts;
HYPRE_Int ierr= 0;
hypre_MPI_Comm_rank(A_comm, &myproc);
hypre_MPI_Comm_size(grid_comm, &nprocs);
start_rank= hypre_ParCSRMatrixFirstRowIndex(A);
end_rank = hypre_ParCSRMatrixLastRowIndex(A);
hypre_SetIndex(ishift, 1, 0, 0);
hypre_SetIndex(jshift, 0, 1, 0);
hypre_SetIndex(kshift, 0, 0, 1);
hypre_SetIndex(zero_index, 0, 0, 0);
/* need a cellgrid boxman to determine the send boxes -> only the cell dofs
are unique so a boxman intersect can be used to get the edges that
must be sent. */
HYPRE_SStructGridCreate(grid_comm, ndim, nparts, &cell_ssgrid);
vartypes= hypre_CTAlloc(HYPRE_SStructVariable, 1);
vartypes[0]= HYPRE_SSTRUCT_VARIABLE_CELL;
for (i= 0; i< nparts; i++)
{
pgrid= hypre_SStructGridPGrid(grid, i);
cellgrid= hypre_SStructPGridCellSGrid(pgrid);
cellboxes= hypre_StructGridBoxes(cellgrid);
hypre_ForBoxI(j, cellboxes)
{
box= hypre_BoxArrayBox(cellboxes, j);
HYPRE_SStructGridSetExtents(cell_ssgrid, i,
hypre_BoxIMin(box), hypre_BoxIMax(box));
}
HYPRE_SStructGridSetVariables(cell_ssgrid, i, 1, vartypes);
}
int
main(int argc, char *argv[])
{
GRID *g;
DOF *u_h;
MAT *A, *A0, *B;
MAP *map;
INT i;
size_t nnz, mem, mem_peak;
VEC *x, *y0, *y1, *y2;
double t0, t1, dnz, dnz1, mflops, mop;
char *fn = "../test/cube.dat";
FLOAT mem_max = 300;
INT refine = 0;
phgOptionsRegisterFilename("-mesh_file", "Mesh file", (char **)&fn);
phgOptionsRegisterInt("-loop_count", "Loop count", &loop_count);
phgOptionsRegisterInt("-refine", "Refinement level", &refine);
phgOptionsRegisterFloat("-mem_max", "Maximum memory", &mem_max);
phgInit(&argc, &argv);
g = phgNewGrid(-1);
if (!phgImport(g, fn, FALSE))
phgError(1, "can't read file \"%s\".\n", fn);
phgRefineAllElements(g, refine);
u_h = phgDofNew(g, DOF_DEFAULT, 1, "u_h", DofNoAction);
while (TRUE) {
phgPrintf("\n");
if (phgBalanceGrid(g, 1.2, 1, NULL, 0.))
phgPrintf("Repartition mesh, %d submeshes, load imbalance: %lg\n",
g->nprocs, (double)g->lif);
map = phgMapCreate(u_h, NULL);
A = phgMapCreateMat(map, map);
A->handle_bdry_eqns = TRUE;
build_matrix(A, u_h);
phgMatAssemble(A);
/* Note: A is unsymmetric (A' != A) if boundary entries not removed */
phgMatRemoveBoundaryEntries(A);
#if 0
/* test block matrix operation */
A0 = phgMatCreateBlockMatrix(g->comm, 1, 1, &A, NULL);
#else
A0 = A;
#endif
phgPrintf("%d DOF, %d elems, %d submeshes, matrix size: %d, LIF: %lg\n",
DofGetDataCountGlobal(u_h), g->nleaf_global,
g->nprocs, A->rmap->nglobal, (double)g->lif);
/* test PHG mat-vec multiply */
x = phgMapCreateVec(A->cmap, 1);
y1 = phgMapCreateVec(A->rmap, 1);
phgVecRandomize(x, 123);
phgMatVec(MAT_OP_N, 1.0, A0, x, 0.0, &y1);
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
phgMatVec(MAT_OP_N, 1.0, A0, x, 0.0, &y1);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
y0 = phgVecCopy(y1, NULL);
nnz = A->nnz_d + A->nnz_o;
#if USE_MPI
dnz1 = nnz;
MPI_Reduce(&dnz1, &dnz, 1, MPI_DOUBLE, MPI_SUM, 0, g->comm);
#else
dnz = nnz;
#endif
mop = loop_count * (dnz + dnz - A->rmap->nlocal) * 1e-6;
phgPrintf("\n");
t1 -= t0;
phgPrintf(" PHG: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF)\n",
t1, dnz, mop / (t1 == 0 ? 1. : t1), mflops);
/* test trans(A)*x */
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
phgMatVec(MAT_OP_T, 1.0, A0, x, 0.0, &y1);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
t1 -= t0;
phgPrintf(" A'*x: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1 == 0 ? 1. : t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
/* time A * trans(A) */
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
B = phgMatMat(MAT_OP_N, MAT_OP_N, 1.0, A, A, 0.0, NULL);
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
nnz = B->nnz_d + B->nnz_o;
#if USE_MPI
dnz1 = nnz;
MPI_Reduce(&dnz1, &dnz, 1, MPI_DOUBLE, MPI_SUM, 0, g->comm);
#else
dnz = nnz;
#endif
/* compare B*x <--> A*A*x */
y2 = phgMatVec(MAT_OP_N, 1.0, B, x, 0.0, NULL);
phgMatVec(MAT_OP_N, 1.0, A0, y0, 0.0, &y1);
phgMatDestroy(&B);
t1 -= t0;
phgPrintf(" A*A: time %0.4lf, nnz %0.16lg, %0.2lfMF, err: %le\n",
t1, dnz, mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y1, 1.0, &y2), 0, NULL));
#if USE_PETSC
{
Mat ma, mb;
MatInfo info;
Vec va, vb, vc;
PetscScalar *vec;
ma = phgPetscCreateMatAIJ(A);
MatGetVecs(ma, PETSC_NULL, &va);
VecDuplicate(va, &vb);
VecGetArray(va, &vec);
memcpy(vec, x->data, x->map->nlocal * sizeof(*vec));
VecRestoreArray(va, &vec);
MatMult(ma, va, vb);
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
MatMult(ma, va, vb);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
VecGetArray(vb, &vec);
memcpy(y1->data, vec, x->map->nlocal * sizeof(*vec));
VecRestoreArray(vb, &vec);
MatGetInfo(ma, MAT_GLOBAL_SUM, &info);
/*phgPrintf(" --------------------------------------------"
"-------------------------\n");*/
phgPrintf("\n");
t1 -= t0;
dnz = info.nz_used;
phgPrintf(" PETSc: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1==0 ? 1.:t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
MatMultTranspose(ma, va, vb);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
VecGetArray(vb, &vec);
memcpy(y1->data, vec, x->map->nlocal * sizeof(*vec));
VecRestoreArray(vb, &vec);
t1 -= t0;
phgPrintf(" A'*x: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1==0 ? 1.:t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
MatMatMult(ma, ma, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &mb);
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
t1 -= t0;
MatGetInfo(mb, MAT_GLOBAL_SUM, &info);
dnz = info.nz_used;
VecDuplicate(va, &vc);
/* compare B*x <--> A*A*x */
MatMult(ma, vb, vc);
MatMult(mb, va, vb);
VecGetArray(vb, &vec);
memcpy(y1->data, vec, x->map->nlocal * sizeof(*vec));
VecRestoreArray(vb, &vec);
VecGetArray(vc, &vec);
memcpy(y2->data, vec, x->map->nlocal * sizeof(*vec));
VecRestoreArray(vc, &vec);
phgPrintf(" A*A: time %0.4lf, nnz %0.16lg, %0.2lfMF, err: %le\n",
t1, dnz, mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y1, 1.0, &y2), 0, NULL));
phgPetscMatDestroy(&mb);
phgPetscMatDestroy(&ma);
phgPetscVecDestroy(&va);
phgPetscVecDestroy(&vb);
phgPetscVecDestroy(&vc);
}
#endif /* USE_PETSC */
#if USE_HYPRE
{
HYPRE_IJMatrix ma;
HYPRE_IJVector va, vb, vc;
HYPRE_ParCSRMatrix par_ma;
hypre_ParCSRMatrix *par_mb;
HYPRE_ParVector par_va, par_vb, par_vc;
HYPRE_Int offset, *ni, start, end;
assert(sizeof(INT)==sizeof(int) && sizeof(FLOAT)==sizeof(double));
setup_hypre_mat(A, &ma);
ni = phgAlloc(2 * A->rmap->nlocal * sizeof(*ni));
offset = A->cmap->partition[A->cmap->rank];
for (i = 0; i < A->rmap->nlocal; i++)
ni[i] = i + offset;
HYPRE_IJVectorCreate(g->comm, offset, offset + A->rmap->nlocal - 1,
&va);
HYPRE_IJVectorCreate(g->comm, offset, offset + A->rmap->nlocal - 1,
&vb);
HYPRE_IJVectorCreate(g->comm, offset, offset + A->rmap->nlocal - 1,
&vc);
HYPRE_IJVectorSetObjectType(va, HYPRE_PARCSR);
HYPRE_IJVectorSetObjectType(vb, HYPRE_PARCSR);
HYPRE_IJVectorSetObjectType(vc, HYPRE_PARCSR);
HYPRE_IJVectorSetMaxOffProcElmts(va, 0);
HYPRE_IJVectorSetMaxOffProcElmts(vb, 0);
HYPRE_IJVectorSetMaxOffProcElmts(vc, 0);
HYPRE_IJVectorInitialize(va);
HYPRE_IJVectorInitialize(vb);
HYPRE_IJVectorInitialize(vc);
HYPRE_IJMatrixGetObject(ma, (void **)(void *)&par_ma);
HYPRE_IJVectorGetObject(va, (void **)(void *)&par_va);
HYPRE_IJVectorGetObject(vb, (void **)(void *)&par_vb);
HYPRE_IJVectorGetObject(vc, (void **)(void *)&par_vc);
HYPRE_IJVectorSetValues(va, A->cmap->nlocal, ni, (double *)x->data);
HYPRE_IJVectorAssemble(va);
HYPRE_IJVectorAssemble(vb);
HYPRE_IJVectorAssemble(vc);
HYPRE_IJMatrixGetRowCounts(ma, A->cmap->nlocal,
ni, ni + A->rmap->nlocal);
for (i = 0, nnz = 0; i < A->rmap->nlocal; i++)
nnz += ni[A->rmap->nlocal + i];
#if USE_MPI
dnz1 = nnz;
MPI_Reduce(&dnz1, &dnz, 1, MPI_DOUBLE, MPI_SUM, 0, g->comm);
#else
dnz = nnz;
#endif
HYPRE_ParCSRMatrixMatvec(1.0, par_ma, par_va, 0.0, par_vb);
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
HYPRE_ParCSRMatrixMatvec(1.0, par_ma, par_va, 0.0, par_vb);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
HYPRE_IJVectorGetValues(vb, A->rmap->nlocal, ni, (double*)y1->data);
/*phgPrintf(" --------------------------------------------"
"-------------------------\n");*/
phgPrintf("\n");
t1 -= t0;
phgPrintf(" HYPRE: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1==0 ? 1.:t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
for (i = 0; i < loop_count; i++) {
HYPRE_ParCSRMatrixMatvecT(1.0, par_ma, par_va, 0.0, par_vb);
}
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
HYPRE_IJVectorGetValues(vb, A->rmap->nlocal, ni, (double*)y1->data);
t1 -= t0;
phgPrintf(" A'*x: time %0.4lf, nnz %0.16lg, %0.2lfMF (%0.2lfMF), "
"err: %le\n", t1, dnz, mop / (t1==0 ? 1.:t1), mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y0, 1.0, &y1), 0, NULL));
phgPerfGetMflops(g, NULL, NULL); /* reset flops counter */
t0 = phgGetTime(NULL);
/* Note: 'HYPRE_ParCSRMatrix' is currently typedef'ed to
* 'hypre_ParCSRMatrix *' */
par_mb = hypre_ParMatmul((hypre_ParCSRMatrix *)par_ma,
(hypre_ParCSRMatrix *)par_ma);
t1 = phgGetTime(NULL);
mflops = phgPerfGetMflops(g, NULL, NULL);
start = hypre_ParCSRMatrixFirstRowIndex(par_mb);
end = hypre_ParCSRMatrixLastRowIndex(par_mb) + 1;
for (i = start, nnz = 0; i < end; i++) {
HYPRE_Int ncols;
hypre_ParCSRMatrixGetRow(par_mb, i, &ncols, NULL, NULL);
hypre_ParCSRMatrixRestoreRow(par_mb, i, &ncols, NULL, NULL);
nnz += ncols;
}
#if USE_MPI
dnz1 = nnz;
MPI_Reduce(&dnz1, &dnz, 1, MPI_DOUBLE, MPI_SUM, 0, g->comm);
#else
dnz = nnz;
#endif
/* compare B*x <--> A*A*x */
HYPRE_ParCSRMatrixMatvec(1.0, par_ma, par_vb, 0.0, par_vc);
HYPRE_ParCSRMatrixMatvec(1.0, (void *)par_mb, par_va, 0.0, par_vb);
HYPRE_IJVectorGetValues(vb, A->rmap->nlocal, ni, (double*)y1->data);
HYPRE_IJVectorGetValues(vc, A->rmap->nlocal, ni, (double*)y2->data);
hypre_ParCSRMatrixDestroy((par_mb));
t1 -= t0;
phgPrintf(" A*A: time %0.4lf, nnz %0.16lg, %0.2lfMF, err: %le\n",
t1, dnz, mflops,
(double)phgVecNorm2(phgVecAXPBY(-1.0, y1, 1.0, &y2), 0, NULL));
phgFree(ni);
HYPRE_IJMatrixDestroy(ma);
HYPRE_IJVectorDestroy(va);
HYPRE_IJVectorDestroy(vb);
HYPRE_IJVectorDestroy(vc);
}
#endif /* USE_HYPRE */
if (A0 != A)
phgMatDestroy(&A0);
#if 0
if (A->rmap->nglobal > 1000) {
VEC *v = phgMapCreateVec(A->rmap, 3);
for (i = 0; i < v->map->nlocal; i++) {
v->data[i + 0 * v->map->nlocal] = 1 * (i + v->map->partition[g->rank]);
v->data[i + 1 * v->map->nlocal] = 2 * (i + v->map->partition[g->rank]);
v->data[i + 2 * v->map->nlocal] = 3 * (i + v->map->partition[g->rank]);
}
phgMatDumpMATLAB(A, "A", "A.m");
phgVecDumpMATLAB(v, "v", "v.m");
phgFinalize();
exit(0);
}
#endif
phgMatDestroy(&A);
phgVecDestroy(&x);
phgVecDestroy(&y0);
phgVecDestroy(&y1);
phgVecDestroy(&y2);
phgMapDestroy(&map);
mem = phgMemoryUsage(g, &mem_peak);
dnz = mem / (1024.0 * 1024.0);
dnz1 = mem_peak / (1024.0 * 1024.0);
/*phgPrintf(" --------------------------------------------"
"-------------------------\n");*/
phgPrintf("\n");
phgPrintf(" Memory: current %0.4lgMB, peak %0.4lgMB\n", dnz, dnz1);
#if 0
{
static int loop_count = 0;
if (++loop_count == 4)
break;
}
#endif
if (mem_peak > 1024 * (size_t)1024 * mem_max)
break;
phgRefineAllElements(g, 1);
}
phgDofFree(&u_h);
phgFreeGrid(&g);
phgFinalize();
return 0;
}
