//======================================================= int EpetraExt_HypreIJMatrix::Multiply(bool TransA, const Epetra_MultiVector& X, Epetra_MultiVector& Y) const { //printf("Proc[%d], Row start: %d, Row End: %d\n", Comm().MyPID(), MyRowStart_, MyRowEnd_); bool SameVectors = false; int NumVectors = X.NumVectors(); if (NumVectors != Y.NumVectors()) return -1; // X and Y must have same number of vectors if(X.Pointers() == Y.Pointers()){ SameVectors = true; } for(int VecNum = 0; VecNum < NumVectors; VecNum++) { //Get values for current vector in multivector. double * x_values; double * y_values; EPETRA_CHK_ERR((*X(VecNum)).ExtractView(&x_values)); double *x_temp = x_local->data; double *y_temp = y_local->data; if(!SameVectors){ EPETRA_CHK_ERR((*Y(VecNum)).ExtractView(&y_values)); } else { y_values = new double[X.MyLength()]; } y_local->data = y_values; EPETRA_CHK_ERR(HYPRE_ParVectorSetConstantValues(par_y,0.0)); // Temporarily make a pointer to data in Hypre for end // Replace data in Hypre vectors with epetra values x_local->data = x_values; // Do actual computation. if(TransA) { // Use transpose of A in multiply EPETRA_CHK_ERR(HYPRE_ParCSRMatrixMatvecT(1.0, ParMatrix_, par_x, 1.0, par_y)); } else { EPETRA_CHK_ERR(HYPRE_ParCSRMatrixMatvec(1.0, ParMatrix_, par_x, 1.0, par_y)); } if(SameVectors){ int NumEntries = Y.MyLength(); std::vector<double> new_values; new_values.resize(NumEntries); std::vector<int> new_indices; new_indices.resize(NumEntries); for(int i = 0; i < NumEntries; i++){ new_values[i] = y_values[i]; new_indices[i] = i; } EPETRA_CHK_ERR((*Y(VecNum)).ReplaceMyValues(NumEntries, &new_values[0], &new_indices[0])); delete[] y_values; } x_local->data = x_temp; y_local->data = y_temp; } double flops = (double) NumVectors * (double) NumGlobalNonzeros(); UpdateFlops(flops); return 0; } //Multiply()
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; }
int32_t impl_bHYPRE_IJParCSRMatrix_ApplyAdjoint( /* in */ bHYPRE_IJParCSRMatrix self, /* in */ bHYPRE_Vector b, /* inout */ bHYPRE_Vector* x, /* out */ sidl_BaseInterface *_ex) { *_ex = 0; { /* DO-NOT-DELETE splicer.begin(bHYPRE.IJParCSRMatrix.ApplyAdjoint) */ /* Insert-Code-Here {bHYPRE.IJParCSRMatrix.ApplyAdjoint} (ApplyAdjoint method) */ /* ApplyAdjoing means to multiply by a vector, y = A'*x , where A' is the * adjoint of A (=transpose, this is a real matrix). Here, we call * the HYPRE Matvec function which performs y = a*A*x + b*y (we set * a=1 and b=0). */ int ierr=0; void * object; struct bHYPRE_IJParCSRMatrix__data * data; struct bHYPRE_IJParCSRVector__data * data_x, * data_b; bHYPRE_IJParCSRVector bHYPREP_b, bHYPREP_x; HYPRE_IJMatrix ij_A; HYPRE_IJVector ij_x, ij_b; HYPRE_ParVector xx, bb; HYPRE_ParCSRMatrix A; data = bHYPRE_IJParCSRMatrix__get_data( self ); ij_A = data -> ij_A; ierr += HYPRE_IJMatrixGetObject( ij_A, &object ); A = (HYPRE_ParCSRMatrix) object; /* A bHYPRE_Vector is just an interface, we have no knowledge of its * contents. Check whether it's something we know how to handle. * If not, die. */ bHYPREP_b = (bHYPRE_IJParCSRVector) bHYPRE_Vector__cast2(b, "bHYPRE.IJParCSRVector", _ex ); SIDL_CHECK(*_ex); hypre_assert( bHYPREP_b!=NULL ); bHYPREP_x = (bHYPRE_IJParCSRVector) bHYPRE_Vector__cast2( *x, "bHYPRE.IJParCSRVector", _ex ); SIDL_CHECK(*_ex); hypre_assert( bHYPREP_x!=NULL ); data_x = bHYPRE_IJParCSRVector__get_data( bHYPREP_x ); ij_x = data_x -> ij_b; ierr += HYPRE_IJVectorGetObject( ij_x, &object ); xx = (HYPRE_ParVector) object; data_b = bHYPRE_IJParCSRVector__get_data( bHYPREP_b ); ij_b = data_b -> ij_b; ierr += HYPRE_IJVectorGetObject( ij_b, &object ); bb = (HYPRE_ParVector) object; ierr += HYPRE_ParCSRMatrixMatvecT( 1.0, A, bb, 0.0, xx ); bHYPRE_IJParCSRVector_deleteRef( bHYPREP_b, _ex ); SIDL_CHECK(*_ex); bHYPRE_IJParCSRVector_deleteRef( bHYPREP_x, _ex ); SIDL_CHECK(*_ex); return( ierr ); hypre_babel_exception_return_error(_ex); /* DO-NOT-DELETE splicer.end(bHYPRE.IJParCSRMatrix.ApplyAdjoint) */ } }