1827bd09bSSatish Balay 211cc89d2SBarry Smith /* 3827bd09bSSatish Balay Module Name: xyt 4827bd09bSSatish Balay Module Info: 5827bd09bSSatish Balay 6827bd09bSSatish Balay author: Henry M. Tufo III 7827bd09bSSatish Balay e-mail: hmt@asci.uchicago.edu 8827bd09bSSatish Balay contact: 9827bd09bSSatish Balay +--------------------------------+--------------------------------+ 10827bd09bSSatish Balay |MCS Division - Building 221 |Department of Computer Science | 11827bd09bSSatish Balay |Argonne National Laboratory |Ryerson 152 | 12827bd09bSSatish Balay |9700 S. Cass Avenue |The University of Chicago | 13827bd09bSSatish Balay |Argonne, IL 60439 |Chicago, IL 60637 | 14827bd09bSSatish Balay |(630) 252-5354/5986 ph/fx |(773) 702-6019/8487 ph/fx | 15827bd09bSSatish Balay +--------------------------------+--------------------------------+ 16827bd09bSSatish Balay 17827bd09bSSatish Balay Last Modification: 3.20.01 1811cc89d2SBarry Smith */ 19c6db04a5SJed Brown #include <../src/ksp/pc/impls/tfs/tfs.h> 20827bd09bSSatish Balay 21827bd09bSSatish Balay #define LEFT -1 22827bd09bSSatish Balay #define RIGHT 1 23827bd09bSSatish Balay #define BOTH 0 24827bd09bSSatish Balay 25827bd09bSSatish Balay typedef struct xyt_solver_info { 2652f87cdaSBarry Smith PetscInt n, m, n_global, m_global; 2752f87cdaSBarry Smith PetscInt nnz, max_nnz, msg_buf_sz; 2852f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo, nfo, *stages; 2952f87cdaSBarry Smith PetscInt *xcol_sz, *xcol_indices; 30a501084fSBarry Smith PetscScalar **xcol_vals, *x, *solve_uu, *solve_w; 3152f87cdaSBarry Smith PetscInt *ycol_sz, *ycol_indices; 32a501084fSBarry Smith PetscScalar **ycol_vals, *y; 3352f87cdaSBarry Smith PetscInt nsolves; 34a501084fSBarry Smith PetscScalar tot_solve_time; 35827bd09bSSatish Balay } xyt_info; 36827bd09bSSatish Balay 37827bd09bSSatish Balay typedef struct matvec_info { 3852f87cdaSBarry Smith PetscInt n, m, n_global, m_global; 3952f87cdaSBarry Smith PetscInt *local2global; 40ca8e9878SJed Brown PCTFS_gs_ADT PCTFS_gs_handle; 41a501084fSBarry Smith PetscErrorCode (*matvec)(struct matvec_info *, PetscScalar *, PetscScalar *); 42827bd09bSSatish Balay void *grid_data; 43827bd09bSSatish Balay } mv_info; 44827bd09bSSatish Balay 45827bd09bSSatish Balay struct xyt_CDT { 4652f87cdaSBarry Smith PetscInt id; 4752f87cdaSBarry Smith PetscInt ns; 4852f87cdaSBarry Smith PetscInt level; 49827bd09bSSatish Balay xyt_info *info; 50827bd09bSSatish Balay mv_info *mvi; 51827bd09bSSatish Balay }; 52827bd09bSSatish Balay 5352f87cdaSBarry Smith static PetscInt n_xyt = 0; 5452f87cdaSBarry Smith static PetscInt n_xyt_handles = 0; 55827bd09bSSatish Balay 56827bd09bSSatish Balay /* prototypes */ 573fdc5746SBarry Smith static PetscErrorCode do_xyt_solve(xyt_ADT xyt_handle, PetscScalar *rhs); 583fdc5746SBarry Smith static PetscErrorCode check_handle(xyt_ADT xyt_handle); 593fdc5746SBarry Smith static PetscErrorCode det_separators(xyt_ADT xyt_handle); 603fdc5746SBarry Smith static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u); 6138dcbb09SBarry Smith static PetscErrorCode xyt_generate(xyt_ADT xyt_handle); 6238dcbb09SBarry Smith static PetscErrorCode do_xyt_factor(xyt_ADT xyt_handle); 635c8f6a95SKarl Rupp static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, PetscErrorCode (*matvec)(mv_info *, PetscScalar *, PetscScalar *), void *grid_data); 64827bd09bSSatish Balay 65d71ae5a4SJacob Faibussowitsch xyt_ADT XYT_new(void) 66d71ae5a4SJacob Faibussowitsch { 67827bd09bSSatish Balay xyt_ADT xyt_handle; 68827bd09bSSatish Balay 69827bd09bSSatish Balay /* rolling count on n_xyt ... pot. problem here */ 70827bd09bSSatish Balay n_xyt_handles++; 71a501084fSBarry Smith xyt_handle = (xyt_ADT)malloc(sizeof(struct xyt_CDT)); 72827bd09bSSatish Balay xyt_handle->id = ++n_xyt; 73827bd09bSSatish Balay xyt_handle->info = NULL; 74827bd09bSSatish Balay xyt_handle->mvi = NULL; 75827bd09bSSatish Balay 76827bd09bSSatish Balay return (xyt_handle); 77827bd09bSSatish Balay } 78827bd09bSSatish Balay 7938dcbb09SBarry Smith PetscErrorCode XYT_factor(xyt_ADT xyt_handle, /* prev. allocated xyt handle */ 8052f87cdaSBarry Smith PetscInt *local2global, /* global column mapping */ 8152f87cdaSBarry Smith PetscInt n, /* local num rows */ 8252f87cdaSBarry Smith PetscInt m, /* local num cols */ 835c8f6a95SKarl Rupp PetscErrorCode (*matvec)(void *, PetscScalar *, PetscScalar *), /* b_loc=A_local.x_loc */ 841147fc2aSKarl Rupp void *grid_data) /* grid data for matvec */ 85827bd09bSSatish Balay { 86*3ba16761SJacob Faibussowitsch PetscFunctionBegin; 87*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_comm_init()); 88*3ba16761SJacob Faibussowitsch PetscCall(check_handle(xyt_handle)); 89827bd09bSSatish Balay 90827bd09bSSatish Balay /* only 2^k for now and all nodes participating */ 9163a3b9bcSJacob Faibussowitsch PetscCheck((1 << (xyt_handle->level = PCTFS_i_log2_num_nodes)) == PCTFS_num_nodes, PETSC_COMM_SELF, PETSC_ERR_PLIB, "only 2^k for now and MPI_COMM_WORLD!!! %d != %d", 1 << PCTFS_i_log2_num_nodes, PCTFS_num_nodes); 92827bd09bSSatish Balay 93827bd09bSSatish Balay /* space for X info */ 94a501084fSBarry Smith xyt_handle->info = (xyt_info *)malloc(sizeof(xyt_info)); 95827bd09bSSatish Balay 96827bd09bSSatish Balay /* set up matvec handles */ 975c8f6a95SKarl Rupp xyt_handle->mvi = set_mvi(local2global, n, m, (PetscErrorCode(*)(mv_info *, PetscScalar *, PetscScalar *))matvec, grid_data); 98827bd09bSSatish Balay 99827bd09bSSatish Balay /* matrix is assumed to be of full rank */ 100827bd09bSSatish Balay /* LATER we can reset to indicate rank def. */ 101827bd09bSSatish Balay xyt_handle->ns = 0; 102827bd09bSSatish Balay 103827bd09bSSatish Balay /* determine separators and generate firing order - NB xyt info set here */ 104*3ba16761SJacob Faibussowitsch PetscCall(det_separators(xyt_handle)); 105827bd09bSSatish Balay 106*3ba16761SJacob Faibussowitsch PetscCall(do_xyt_factor(xyt_handle)); 107*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 108827bd09bSSatish Balay } 109827bd09bSSatish Balay 110d71ae5a4SJacob Faibussowitsch PetscErrorCode XYT_solve(xyt_ADT xyt_handle, PetscScalar *x, PetscScalar *b) 111d71ae5a4SJacob Faibussowitsch { 112*3ba16761SJacob Faibussowitsch PetscFunctionBegin; 113*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_comm_init()); 114*3ba16761SJacob Faibussowitsch PetscCall(check_handle(xyt_handle)); 115827bd09bSSatish Balay 116827bd09bSSatish Balay /* need to copy b into x? */ 117*3ba16761SJacob Faibussowitsch if (b) PetscCall(PCTFS_rvec_copy(x, b, xyt_handle->mvi->n)); 118*3ba16761SJacob Faibussowitsch PetscCall(do_xyt_solve(xyt_handle, x)); 119*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 120827bd09bSSatish Balay } 121827bd09bSSatish Balay 122d71ae5a4SJacob Faibussowitsch PetscErrorCode XYT_free(xyt_ADT xyt_handle) 123d71ae5a4SJacob Faibussowitsch { 124*3ba16761SJacob Faibussowitsch PetscFunctionBegin; 125*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_comm_init()); 126*3ba16761SJacob Faibussowitsch PetscCall(check_handle(xyt_handle)); 127827bd09bSSatish Balay n_xyt_handles--; 128827bd09bSSatish Balay 129a501084fSBarry Smith free(xyt_handle->info->nsep); 130a501084fSBarry Smith free(xyt_handle->info->lnsep); 131a501084fSBarry Smith free(xyt_handle->info->fo); 132a501084fSBarry Smith free(xyt_handle->info->stages); 133a501084fSBarry Smith free(xyt_handle->info->solve_uu); 134a501084fSBarry Smith free(xyt_handle->info->solve_w); 135a501084fSBarry Smith free(xyt_handle->info->x); 136a501084fSBarry Smith free(xyt_handle->info->xcol_vals); 137a501084fSBarry Smith free(xyt_handle->info->xcol_sz); 138a501084fSBarry Smith free(xyt_handle->info->xcol_indices); 139a501084fSBarry Smith free(xyt_handle->info->y); 140a501084fSBarry Smith free(xyt_handle->info->ycol_vals); 141a501084fSBarry Smith free(xyt_handle->info->ycol_sz); 142a501084fSBarry Smith free(xyt_handle->info->ycol_indices); 143a501084fSBarry Smith free(xyt_handle->info); 144a501084fSBarry Smith free(xyt_handle->mvi->local2global); 145*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_gs_free(xyt_handle->mvi->PCTFS_gs_handle)); 146a501084fSBarry Smith free(xyt_handle->mvi); 147a501084fSBarry Smith free(xyt_handle); 148827bd09bSSatish Balay 149827bd09bSSatish Balay /* if the check fails we nuke */ 150a501084fSBarry Smith /* if NULL pointer passed to free we nuke */ 151827bd09bSSatish Balay /* if the calls to free fail that's not my problem */ 152*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 153827bd09bSSatish Balay } 154827bd09bSSatish Balay 15511cc89d2SBarry Smith /* This function is currently not used */ 156d71ae5a4SJacob Faibussowitsch PetscErrorCode XYT_stats(xyt_ADT xyt_handle) 157d71ae5a4SJacob Faibussowitsch { 15852f87cdaSBarry Smith PetscInt op[] = {NON_UNIFORM, GL_MIN, GL_MAX, GL_ADD, GL_MIN, GL_MAX, GL_ADD, GL_MIN, GL_MAX, GL_ADD}; 15952f87cdaSBarry Smith PetscInt fop[] = {NON_UNIFORM, GL_MIN, GL_MAX, GL_ADD}; 16052f87cdaSBarry Smith PetscInt vals[9], work[9]; 161a501084fSBarry Smith PetscScalar fvals[3], fwork[3]; 162827bd09bSSatish Balay 16363a3b9bcSJacob Faibussowitsch PetscFunctionBegin; 16463a3b9bcSJacob Faibussowitsch PetscCall(PCTFS_comm_init()); 16563a3b9bcSJacob Faibussowitsch PetscCall(check_handle(xyt_handle)); 166827bd09bSSatish Balay 167827bd09bSSatish Balay /* if factorization not done there are no stats */ 1687d0a6c19SBarry Smith if (!xyt_handle->info || !xyt_handle->mvi) { 16963a3b9bcSJacob Faibussowitsch if (!PCTFS_my_id) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "XYT_stats() :: no stats available!\n")); 170*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 171827bd09bSSatish Balay } 172827bd09bSSatish Balay 173827bd09bSSatish Balay vals[0] = vals[1] = vals[2] = xyt_handle->info->nnz; 174827bd09bSSatish Balay vals[3] = vals[4] = vals[5] = xyt_handle->mvi->n; 175827bd09bSSatish Balay vals[6] = vals[7] = vals[8] = xyt_handle->info->msg_buf_sz; 17663a3b9bcSJacob Faibussowitsch PetscCall(PCTFS_giop(vals, work, PETSC_STATIC_ARRAY_LENGTH(op) - 1, op)); 177827bd09bSSatish Balay 1782fa5cd67SKarl Rupp fvals[0] = fvals[1] = fvals[2] = xyt_handle->info->tot_solve_time / xyt_handle->info->nsolves++; 17963a3b9bcSJacob Faibussowitsch PetscCall(PCTFS_grop(fvals, fwork, PETSC_STATIC_ARRAY_LENGTH(fop) - 1, fop)); 180827bd09bSSatish Balay 181b1c944f5SJed Brown if (!PCTFS_my_id) { 18263a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: min xyt_nnz=%" PetscInt_FMT "\n", PCTFS_my_id, vals[0])); 18363a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: max xyt_nnz=%" PetscInt_FMT "\n", PCTFS_my_id, vals[1])); 18463a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: avg xyt_nnz=%g\n", PCTFS_my_id, (double)(1.0 * vals[2] / PCTFS_num_nodes))); 18563a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: tot xyt_nnz=%" PetscInt_FMT "\n", PCTFS_my_id, vals[2])); 18663a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: xyt C(2d) =%g\n", PCTFS_my_id, (double)(vals[2] / (PetscPowReal(1.0 * vals[5], 1.5))))); 18763a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: xyt C(3d) =%g\n", PCTFS_my_id, (double)(vals[2] / (PetscPowReal(1.0 * vals[5], 1.6667))))); 18863a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: min xyt_n =%" PetscInt_FMT "\n", PCTFS_my_id, vals[3])); 18963a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: max xyt_n =%" PetscInt_FMT "\n", PCTFS_my_id, vals[4])); 19063a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: avg xyt_n =%g\n", PCTFS_my_id, (double)(1.0 * vals[5] / PCTFS_num_nodes))); 19163a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: tot xyt_n =%" PetscInt_FMT "\n", PCTFS_my_id, vals[5])); 19263a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: min xyt_buf=%" PetscInt_FMT "\n", PCTFS_my_id, vals[6])); 19363a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: max xyt_buf=%" PetscInt_FMT "\n", PCTFS_my_id, vals[7])); 19463a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: avg xyt_buf=%g\n", PCTFS_my_id, (double)(1.0 * vals[8] / PCTFS_num_nodes))); 19563a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: min xyt_slv=%g\n", PCTFS_my_id, (double)PetscRealPart(fvals[0]))); 19663a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: max xyt_slv=%g\n", PCTFS_my_id, (double)PetscRealPart(fvals[1]))); 19763a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d :: avg xyt_slv=%g\n", PCTFS_my_id, (double)PetscRealPart(fvals[2] / PCTFS_num_nodes))); 198827bd09bSSatish Balay } 199*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 200827bd09bSSatish Balay } 201827bd09bSSatish Balay 20211cc89d2SBarry Smith /* 203827bd09bSSatish Balay 204827bd09bSSatish Balay Description: get A_local, local portion of global coarse matrix which 205827bd09bSSatish Balay is a row dist. nxm matrix w/ n<m. 206827bd09bSSatish Balay o my_ml holds address of ML struct associated w/A_local and coarse grid 207827bd09bSSatish Balay o local2global holds global number of column i (i=0,...,m-1) 208827bd09bSSatish Balay o local2global holds global number of row i (i=0,...,n-1) 209827bd09bSSatish Balay o mylocmatvec performs A_local . vec_local (note that gs is performed using 210ca8e9878SJed Brown PCTFS_gs_init/gop). 211827bd09bSSatish Balay 212827bd09bSSatish Balay mylocmatvec = my_ml->Amat[grid_tag].matvec->external; 213827bd09bSSatish Balay mylocmatvec (void :: void *data, double *in, double *out) 21411cc89d2SBarry Smith */ 215d71ae5a4SJacob Faibussowitsch static PetscErrorCode do_xyt_factor(xyt_ADT xyt_handle) 216d71ae5a4SJacob Faibussowitsch { 2177b1ae94cSBarry Smith return xyt_generate(xyt_handle); 218827bd09bSSatish Balay } 219827bd09bSSatish Balay 220d71ae5a4SJacob Faibussowitsch static PetscErrorCode xyt_generate(xyt_ADT xyt_handle) 221d71ae5a4SJacob Faibussowitsch { 22252f87cdaSBarry Smith PetscInt i, j, k, idx; 22352f87cdaSBarry Smith PetscInt dim, col; 224a501084fSBarry Smith PetscScalar *u, *uu, *v, *z, *w, alpha, alpha_w; 22552f87cdaSBarry Smith PetscInt *segs; 22652f87cdaSBarry Smith PetscInt op[] = {GL_ADD, 0}; 22752f87cdaSBarry Smith PetscInt off, len; 228a501084fSBarry Smith PetscScalar *x_ptr, *y_ptr; 22952f87cdaSBarry Smith PetscInt *iptr, flag; 23052f87cdaSBarry Smith PetscInt start = 0, end, work; 23152f87cdaSBarry Smith PetscInt op2[] = {GL_MIN, 0}; 232ca8e9878SJed Brown PCTFS_gs_ADT PCTFS_gs_handle; 23352f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo; 23452f87cdaSBarry Smith PetscInt a_n = xyt_handle->mvi->n; 23552f87cdaSBarry Smith PetscInt a_m = xyt_handle->mvi->m; 23652f87cdaSBarry Smith PetscInt *a_local2global = xyt_handle->mvi->local2global; 23752f87cdaSBarry Smith PetscInt level; 23852f87cdaSBarry Smith PetscInt n, m; 23952f87cdaSBarry Smith PetscInt *xcol_sz, *xcol_indices, *stages; 240a501084fSBarry Smith PetscScalar **xcol_vals, *x; 24152f87cdaSBarry Smith PetscInt *ycol_sz, *ycol_indices; 242a501084fSBarry Smith PetscScalar **ycol_vals, *y; 24352f87cdaSBarry Smith PetscInt n_global; 24452f87cdaSBarry Smith PetscInt xt_nnz = 0, xt_max_nnz = 0; 24552f87cdaSBarry Smith PetscInt yt_nnz = 0, yt_max_nnz = 0; 2464a0de3f6SBarry Smith PetscBLASInt i1 = 1, dlen; 247a501084fSBarry Smith PetscScalar dm1 = -1.0; 248827bd09bSSatish Balay 249827bd09bSSatish Balay n = xyt_handle->mvi->n; 250827bd09bSSatish Balay nsep = xyt_handle->info->nsep; 251827bd09bSSatish Balay lnsep = xyt_handle->info->lnsep; 252827bd09bSSatish Balay fo = xyt_handle->info->fo; 253827bd09bSSatish Balay end = lnsep[0]; 254827bd09bSSatish Balay level = xyt_handle->level; 255ca8e9878SJed Brown PCTFS_gs_handle = xyt_handle->mvi->PCTFS_gs_handle; 256827bd09bSSatish Balay 257827bd09bSSatish Balay /* is there a null space? */ 258827bd09bSSatish Balay /* LATER add in ability to detect null space by checking alpha */ 2592fa5cd67SKarl Rupp for (i = 0, j = 0; i <= level; i++) j += nsep[i]; 260827bd09bSSatish Balay 261827bd09bSSatish Balay m = j - xyt_handle->ns; 26248a46eb9SPierre Jolivet if (m != j) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "xyt_generate() :: null space exists %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", m, j, xyt_handle->ns)); 263827bd09bSSatish Balay 26463a3b9bcSJacob Faibussowitsch PetscCall(PetscInfo(0, "xyt_generate() :: X(%" PetscInt_FMT ",%" PetscInt_FMT ")\n", n, m)); 265827bd09bSSatish Balay 266827bd09bSSatish Balay /* get and initialize storage for x local */ 267827bd09bSSatish Balay /* note that x local is nxm and stored by columns */ 26852f87cdaSBarry Smith xcol_sz = (PetscInt *)malloc(m * sizeof(PetscInt)); 26952f87cdaSBarry Smith xcol_indices = (PetscInt *)malloc((2 * m + 1) * sizeof(PetscInt)); 270a501084fSBarry Smith xcol_vals = (PetscScalar **)malloc(m * sizeof(PetscScalar *)); 271db4deed7SKarl Rupp for (i = j = 0; i < m; i++, j += 2) { 272827bd09bSSatish Balay xcol_indices[j] = xcol_indices[j + 1] = xcol_sz[i] = -1; 273827bd09bSSatish Balay xcol_vals[i] = NULL; 274827bd09bSSatish Balay } 275827bd09bSSatish Balay xcol_indices[j] = -1; 276827bd09bSSatish Balay 277827bd09bSSatish Balay /* get and initialize storage for y local */ 278827bd09bSSatish Balay /* note that y local is nxm and stored by columns */ 27952f87cdaSBarry Smith ycol_sz = (PetscInt *)malloc(m * sizeof(PetscInt)); 28052f87cdaSBarry Smith ycol_indices = (PetscInt *)malloc((2 * m + 1) * sizeof(PetscInt)); 281a501084fSBarry Smith ycol_vals = (PetscScalar **)malloc(m * sizeof(PetscScalar *)); 282db4deed7SKarl Rupp for (i = j = 0; i < m; i++, j += 2) { 283827bd09bSSatish Balay ycol_indices[j] = ycol_indices[j + 1] = ycol_sz[i] = -1; 284827bd09bSSatish Balay ycol_vals[i] = NULL; 285827bd09bSSatish Balay } 286827bd09bSSatish Balay ycol_indices[j] = -1; 287827bd09bSSatish Balay 288827bd09bSSatish Balay /* size of separators for each sub-hc working from bottom of tree to top */ 289827bd09bSSatish Balay /* this looks like nsep[]=segments */ 29052f87cdaSBarry Smith stages = (PetscInt *)malloc((level + 1) * sizeof(PetscInt)); 29152f87cdaSBarry Smith segs = (PetscInt *)malloc((level + 1) * sizeof(PetscInt)); 292*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_ivec_zero(stages, level + 1)); 293ca8e9878SJed Brown PCTFS_ivec_copy(segs, nsep, level + 1); 2942fa5cd67SKarl Rupp for (i = 0; i < level; i++) segs[i + 1] += segs[i]; 295827bd09bSSatish Balay stages[0] = segs[0]; 296827bd09bSSatish Balay 297827bd09bSSatish Balay /* temporary vectors */ 298a501084fSBarry Smith u = (PetscScalar *)malloc(n * sizeof(PetscScalar)); 299a501084fSBarry Smith z = (PetscScalar *)malloc(n * sizeof(PetscScalar)); 300a501084fSBarry Smith v = (PetscScalar *)malloc(a_m * sizeof(PetscScalar)); 301a501084fSBarry Smith uu = (PetscScalar *)malloc(m * sizeof(PetscScalar)); 302a501084fSBarry Smith w = (PetscScalar *)malloc(m * sizeof(PetscScalar)); 303827bd09bSSatish Balay 304827bd09bSSatish Balay /* extra nnz due to replication of vertices across separators */ 3052fa5cd67SKarl Rupp for (i = 1, j = 0; i <= level; i++) j += nsep[i]; 306827bd09bSSatish Balay 307827bd09bSSatish Balay /* storage for sparse x values */ 308827bd09bSSatish Balay n_global = xyt_handle->info->n_global; 30985ec1a3cSBarry Smith xt_max_nnz = yt_max_nnz = (PetscInt)(2.5 * PetscPowReal(1.0 * n_global, 1.6667) + j * n / 2) / PCTFS_num_nodes; 310a501084fSBarry Smith x = (PetscScalar *)malloc(xt_max_nnz * sizeof(PetscScalar)); 311a501084fSBarry Smith y = (PetscScalar *)malloc(yt_max_nnz * sizeof(PetscScalar)); 312827bd09bSSatish Balay 313827bd09bSSatish Balay /* LATER - can embed next sep to fire in gs */ 314827bd09bSSatish Balay /* time to make the donuts - generate X factor */ 315db4deed7SKarl Rupp for (dim = i = j = 0; i < m; i++) { 316827bd09bSSatish Balay /* time to move to the next level? */ 317d4af0d30SBarry Smith while (i == segs[dim]) { 31808401ef6SPierre Jolivet PetscCheck(dim != level, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim about to exceed level"); 319827bd09bSSatish Balay stages[dim++] = i; 320827bd09bSSatish Balay end += lnsep[dim]; 321827bd09bSSatish Balay } 322827bd09bSSatish Balay stages[dim] = i; 323827bd09bSSatish Balay 324827bd09bSSatish Balay /* which column are we firing? */ 325827bd09bSSatish Balay /* i.e. set v_l */ 326827bd09bSSatish Balay /* use new seps and do global min across hc to determine which one to fire */ 327827bd09bSSatish Balay (start < end) ? (col = fo[start]) : (col = INT_MAX); 328*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_giop_hc(&col, &work, 1, op2, dim)); 329827bd09bSSatish Balay 330827bd09bSSatish Balay /* shouldn't need this */ 331db4deed7SKarl Rupp if (col == INT_MAX) { 3329566063dSJacob Faibussowitsch PetscCall(PetscInfo(0, "hey ... col==INT_MAX??\n")); 333827bd09bSSatish Balay continue; 334827bd09bSSatish Balay } 335827bd09bSSatish Balay 336827bd09bSSatish Balay /* do I own it? I should */ 337*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(v, a_m)); 3382fa5cd67SKarl Rupp if (col == fo[start]) { 339827bd09bSSatish Balay start++; 340ca8e9878SJed Brown idx = PCTFS_ivec_linear_search(col, a_local2global, a_n); 3412fa5cd67SKarl Rupp if (idx != -1) { 3422fa5cd67SKarl Rupp v[idx] = 1.0; 3432fa5cd67SKarl Rupp j++; 344546078acSJacob Faibussowitsch } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_PLIB, "NOT FOUND!"); 345db4deed7SKarl Rupp } else { 346ca8e9878SJed Brown idx = PCTFS_ivec_linear_search(col, a_local2global, a_m); 3472fa5cd67SKarl Rupp if (idx != -1) v[idx] = 1.0; 348827bd09bSSatish Balay } 349827bd09bSSatish Balay 350827bd09bSSatish Balay /* perform u = A.v_l */ 351*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(u, n)); 352*3ba16761SJacob Faibussowitsch PetscCall(do_matvec(xyt_handle->mvi, v, u)); 353827bd09bSSatish Balay 354827bd09bSSatish Balay /* uu = X^T.u_l (local portion) */ 355827bd09bSSatish Balay /* technically only need to zero out first i entries */ 356827bd09bSSatish Balay /* later turn this into an XYT_solve call ? */ 357*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(uu, m)); 358827bd09bSSatish Balay y_ptr = y; 359827bd09bSSatish Balay iptr = ycol_indices; 360db4deed7SKarl Rupp for (k = 0; k < i; k++) { 361827bd09bSSatish Balay off = *iptr++; 362827bd09bSSatish Balay len = *iptr++; 3639566063dSJacob Faibussowitsch PetscCall(PetscBLASIntCast(len, &dlen)); 364792fecdfSBarry Smith PetscCallBLAS("BLASdot", uu[k] = BLASdot_(&dlen, u + off, &i1, y_ptr, &i1)); 365827bd09bSSatish Balay y_ptr += len; 366827bd09bSSatish Balay } 367827bd09bSSatish Balay 368827bd09bSSatish Balay /* uu = X^T.u_l (comm portion) */ 3699566063dSJacob Faibussowitsch PetscCall(PCTFS_ssgl_radd(uu, w, dim, stages)); 370827bd09bSSatish Balay 371827bd09bSSatish Balay /* z = X.uu */ 372*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(z, n)); 373827bd09bSSatish Balay x_ptr = x; 374827bd09bSSatish Balay iptr = xcol_indices; 375db4deed7SKarl Rupp for (k = 0; k < i; k++) { 376827bd09bSSatish Balay off = *iptr++; 377827bd09bSSatish Balay len = *iptr++; 3789566063dSJacob Faibussowitsch PetscCall(PetscBLASIntCast(len, &dlen)); 379792fecdfSBarry Smith PetscCallBLAS("BLASaxpy", BLASaxpy_(&dlen, &uu[k], x_ptr, &i1, z + off, &i1)); 380827bd09bSSatish Balay x_ptr += len; 381827bd09bSSatish Balay } 382827bd09bSSatish Balay 383827bd09bSSatish Balay /* compute v_l = v_l - z */ 384*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(v + a_n, a_m - a_n)); 3859566063dSJacob Faibussowitsch PetscCall(PetscBLASIntCast(n, &dlen)); 386792fecdfSBarry Smith PetscCallBLAS("BLASaxpy", BLASaxpy_(&dlen, &dm1, z, &i1, v, &i1)); 387827bd09bSSatish Balay 388827bd09bSSatish Balay /* compute u_l = A.v_l */ 389*3ba16761SJacob Faibussowitsch if (a_n != a_m) PetscCall(PCTFS_gs_gop_hc(PCTFS_gs_handle, v, "+\0", dim)); 390*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(u, n)); 391*3ba16761SJacob Faibussowitsch PetscCall(do_matvec(xyt_handle->mvi, v, u)); 392827bd09bSSatish Balay 393827bd09bSSatish Balay /* compute sqrt(alpha) = sqrt(u_l^T.u_l) - local portion */ 3949566063dSJacob Faibussowitsch PetscCall(PetscBLASIntCast(n, &dlen)); 395792fecdfSBarry Smith PetscCallBLAS("BLASdot", alpha = BLASdot_(&dlen, u, &i1, u, &i1)); 396827bd09bSSatish Balay /* compute sqrt(alpha) = sqrt(u_l^T.u_l) - comm portion */ 397*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_grop_hc(&alpha, &alpha_w, 1, op, dim)); 398827bd09bSSatish Balay 3998f1a2a5eSBarry Smith alpha = (PetscScalar)PetscSqrtReal((PetscReal)alpha); 400827bd09bSSatish Balay 401827bd09bSSatish Balay /* check for small alpha */ 402827bd09bSSatish Balay /* LATER use this to detect and determine null space */ 40363a3b9bcSJacob Faibussowitsch PetscCheck(PetscAbsScalar(alpha) >= 1.0e-14, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bad alpha! %g", (double)PetscAbsScalar(alpha)); 404827bd09bSSatish Balay 405827bd09bSSatish Balay /* compute v_l = v_l/sqrt(alpha) */ 406*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_scale(v, 1.0 / alpha, n)); 407*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_scale(u, 1.0 / alpha, n)); 408827bd09bSSatish Balay 409827bd09bSSatish Balay /* add newly generated column, v_l, to X */ 410827bd09bSSatish Balay flag = 1; 411827bd09bSSatish Balay off = len = 0; 412db4deed7SKarl Rupp for (k = 0; k < n; k++) { 413db4deed7SKarl Rupp if (v[k] != 0.0) { 414827bd09bSSatish Balay len = k; 4159371c9d4SSatish Balay if (flag) { 4169371c9d4SSatish Balay off = k; 4179371c9d4SSatish Balay flag = 0; 4189371c9d4SSatish Balay } 419827bd09bSSatish Balay } 420827bd09bSSatish Balay } 421827bd09bSSatish Balay 422827bd09bSSatish Balay len -= (off - 1); 423827bd09bSSatish Balay 424db4deed7SKarl Rupp if (len > 0) { 425db4deed7SKarl Rupp if ((xt_nnz + len) > xt_max_nnz) { 4269566063dSJacob Faibussowitsch PetscCall(PetscInfo(0, "increasing space for X by 2x!\n")); 427827bd09bSSatish Balay xt_max_nnz *= 2; 428a501084fSBarry Smith x_ptr = (PetscScalar *)malloc(xt_max_nnz * sizeof(PetscScalar)); 429*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_copy(x_ptr, x, xt_nnz)); 430a501084fSBarry Smith free(x); 431827bd09bSSatish Balay x = x_ptr; 432827bd09bSSatish Balay x_ptr += xt_nnz; 433827bd09bSSatish Balay } 434827bd09bSSatish Balay xt_nnz += len; 435*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_copy(x_ptr, v + off, len)); 436827bd09bSSatish Balay 437827bd09bSSatish Balay xcol_indices[2 * i] = off; 438827bd09bSSatish Balay xcol_sz[i] = xcol_indices[2 * i + 1] = len; 439827bd09bSSatish Balay xcol_vals[i] = x_ptr; 440db4deed7SKarl Rupp } else { 441827bd09bSSatish Balay xcol_indices[2 * i] = 0; 442827bd09bSSatish Balay xcol_sz[i] = xcol_indices[2 * i + 1] = 0; 443827bd09bSSatish Balay xcol_vals[i] = x_ptr; 444827bd09bSSatish Balay } 445827bd09bSSatish Balay 446827bd09bSSatish Balay /* add newly generated column, u_l, to Y */ 447827bd09bSSatish Balay flag = 1; 448827bd09bSSatish Balay off = len = 0; 449db4deed7SKarl Rupp for (k = 0; k < n; k++) { 450db4deed7SKarl Rupp if (u[k] != 0.0) { 451827bd09bSSatish Balay len = k; 4529371c9d4SSatish Balay if (flag) { 4539371c9d4SSatish Balay off = k; 4549371c9d4SSatish Balay flag = 0; 4559371c9d4SSatish Balay } 456827bd09bSSatish Balay } 457827bd09bSSatish Balay } 458827bd09bSSatish Balay 459827bd09bSSatish Balay len -= (off - 1); 460827bd09bSSatish Balay 461db4deed7SKarl Rupp if (len > 0) { 462db4deed7SKarl Rupp if ((yt_nnz + len) > yt_max_nnz) { 4639566063dSJacob Faibussowitsch PetscCall(PetscInfo(0, "increasing space for Y by 2x!\n")); 464827bd09bSSatish Balay yt_max_nnz *= 2; 465a501084fSBarry Smith y_ptr = (PetscScalar *)malloc(yt_max_nnz * sizeof(PetscScalar)); 466*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_copy(y_ptr, y, yt_nnz)); 467a501084fSBarry Smith free(y); 468827bd09bSSatish Balay y = y_ptr; 469827bd09bSSatish Balay y_ptr += yt_nnz; 470827bd09bSSatish Balay } 471827bd09bSSatish Balay yt_nnz += len; 472*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_copy(y_ptr, u + off, len)); 473827bd09bSSatish Balay 474827bd09bSSatish Balay ycol_indices[2 * i] = off; 475827bd09bSSatish Balay ycol_sz[i] = ycol_indices[2 * i + 1] = len; 476827bd09bSSatish Balay ycol_vals[i] = y_ptr; 477db4deed7SKarl Rupp } else { 478827bd09bSSatish Balay ycol_indices[2 * i] = 0; 479827bd09bSSatish Balay ycol_sz[i] = ycol_indices[2 * i + 1] = 0; 480827bd09bSSatish Balay ycol_vals[i] = y_ptr; 481827bd09bSSatish Balay } 482827bd09bSSatish Balay } 483827bd09bSSatish Balay 484827bd09bSSatish Balay /* close off stages for execution phase */ 485db4deed7SKarl Rupp while (dim != level) { 486827bd09bSSatish Balay stages[dim++] = i; 48763a3b9bcSJacob Faibussowitsch PetscCall(PetscInfo(0, "disconnected!!! dim(%" PetscInt_FMT ")!=level(%" PetscInt_FMT ")\n", dim, level)); 488827bd09bSSatish Balay } 489827bd09bSSatish Balay stages[dim] = i; 490827bd09bSSatish Balay 491827bd09bSSatish Balay xyt_handle->info->n = xyt_handle->mvi->n; 492827bd09bSSatish Balay xyt_handle->info->m = m; 493827bd09bSSatish Balay xyt_handle->info->nnz = xt_nnz + yt_nnz; 494827bd09bSSatish Balay xyt_handle->info->max_nnz = xt_max_nnz + yt_max_nnz; 495827bd09bSSatish Balay xyt_handle->info->msg_buf_sz = stages[level] - stages[0]; 496a501084fSBarry Smith xyt_handle->info->solve_uu = (PetscScalar *)malloc(m * sizeof(PetscScalar)); 497a501084fSBarry Smith xyt_handle->info->solve_w = (PetscScalar *)malloc(m * sizeof(PetscScalar)); 498827bd09bSSatish Balay xyt_handle->info->x = x; 499827bd09bSSatish Balay xyt_handle->info->xcol_vals = xcol_vals; 500827bd09bSSatish Balay xyt_handle->info->xcol_sz = xcol_sz; 501827bd09bSSatish Balay xyt_handle->info->xcol_indices = xcol_indices; 502827bd09bSSatish Balay xyt_handle->info->stages = stages; 503827bd09bSSatish Balay xyt_handle->info->y = y; 504827bd09bSSatish Balay xyt_handle->info->ycol_vals = ycol_vals; 505827bd09bSSatish Balay xyt_handle->info->ycol_sz = ycol_sz; 506827bd09bSSatish Balay xyt_handle->info->ycol_indices = ycol_indices; 507827bd09bSSatish Balay 508a501084fSBarry Smith free(segs); 509a501084fSBarry Smith free(u); 510a501084fSBarry Smith free(v); 511a501084fSBarry Smith free(uu); 512a501084fSBarry Smith free(z); 513a501084fSBarry Smith free(w); 514827bd09bSSatish Balay 515*3ba16761SJacob Faibussowitsch return PETSC_SUCCESS; 516827bd09bSSatish Balay } 517827bd09bSSatish Balay 518d71ae5a4SJacob Faibussowitsch static PetscErrorCode do_xyt_solve(xyt_ADT xyt_handle, PetscScalar *uc) 519d71ae5a4SJacob Faibussowitsch { 52052f87cdaSBarry Smith PetscInt off, len, *iptr; 52152f87cdaSBarry Smith PetscInt level = xyt_handle->level; 52252f87cdaSBarry Smith PetscInt n = xyt_handle->info->n; 52352f87cdaSBarry Smith PetscInt m = xyt_handle->info->m; 52452f87cdaSBarry Smith PetscInt *stages = xyt_handle->info->stages; 52552f87cdaSBarry Smith PetscInt *xcol_indices = xyt_handle->info->xcol_indices; 52652f87cdaSBarry Smith PetscInt *ycol_indices = xyt_handle->info->ycol_indices; 527a501084fSBarry Smith PetscScalar *x_ptr, *y_ptr, *uu_ptr; 528a501084fSBarry Smith PetscScalar *solve_uu = xyt_handle->info->solve_uu; 529a501084fSBarry Smith PetscScalar *solve_w = xyt_handle->info->solve_w; 530a501084fSBarry Smith PetscScalar *x = xyt_handle->info->x; 531a501084fSBarry Smith PetscScalar *y = xyt_handle->info->y; 5324a0de3f6SBarry Smith PetscBLASInt i1 = 1, dlen; 533827bd09bSSatish Balay 5340924e98cSBarry Smith PetscFunctionBegin; 535827bd09bSSatish Balay uu_ptr = solve_uu; 536*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(uu_ptr, m)); 537827bd09bSSatish Balay 538827bd09bSSatish Balay /* x = X.Y^T.b */ 539827bd09bSSatish Balay /* uu = Y^T.b */ 540947b95d8SBarry Smith for (y_ptr = y, iptr = ycol_indices; *iptr != -1; y_ptr += len) { 541c5df96a5SBarry Smith off = *iptr++; 542c5df96a5SBarry Smith len = *iptr++; 5439566063dSJacob Faibussowitsch PetscCall(PetscBLASIntCast(len, &dlen)); 544792fecdfSBarry Smith PetscCallBLAS("BLASdot", *uu_ptr++ = BLASdot_(&dlen, uc + off, &i1, y_ptr, &i1)); 545827bd09bSSatish Balay } 546827bd09bSSatish Balay 547d5b43468SJose E. Roman /* communication of beta */ 548827bd09bSSatish Balay uu_ptr = solve_uu; 5499566063dSJacob Faibussowitsch if (level) PetscCall(PCTFS_ssgl_radd(uu_ptr, solve_w, level, stages)); 550*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(uc, n)); 551827bd09bSSatish Balay 552827bd09bSSatish Balay /* x = X.uu */ 553db4deed7SKarl Rupp for (x_ptr = x, iptr = xcol_indices; *iptr != -1; x_ptr += len) { 554c5df96a5SBarry Smith off = *iptr++; 555c5df96a5SBarry Smith len = *iptr++; 5569566063dSJacob Faibussowitsch PetscCall(PetscBLASIntCast(len, &dlen)); 557792fecdfSBarry Smith PetscCallBLAS("BLASaxpy", BLASaxpy_(&dlen, uu_ptr++, x_ptr, &i1, uc + off, &i1)); 558827bd09bSSatish Balay } 559*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 560827bd09bSSatish Balay } 561827bd09bSSatish Balay 562d71ae5a4SJacob Faibussowitsch static PetscErrorCode check_handle(xyt_ADT xyt_handle) 563d71ae5a4SJacob Faibussowitsch { 56452f87cdaSBarry Smith PetscInt vals[2], work[2], op[] = {NON_UNIFORM, GL_MIN, GL_MAX}; 565827bd09bSSatish Balay 5660924e98cSBarry Smith PetscFunctionBegin; 5673274405dSPierre Jolivet PetscCheck(xyt_handle, PETSC_COMM_SELF, PETSC_ERR_PLIB, "check_handle() :: bad handle :: NULL %p", (void *)xyt_handle); 568827bd09bSSatish Balay 569827bd09bSSatish Balay vals[0] = vals[1] = xyt_handle->id; 570*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_giop(vals, work, PETSC_STATIC_ARRAY_LENGTH(op) - 1, op)); 57163a3b9bcSJacob Faibussowitsch PetscCheck(!(vals[0] != vals[1]) && !(xyt_handle->id <= 0), PETSC_COMM_SELF, PETSC_ERR_PLIB, "check_handle() :: bad handle :: id mismatch min/max %" PetscInt_FMT "/%" PetscInt_FMT " %" PetscInt_FMT, vals[0], vals[1], xyt_handle->id); 572*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 573827bd09bSSatish Balay } 574827bd09bSSatish Balay 575d71ae5a4SJacob Faibussowitsch static PetscErrorCode det_separators(xyt_ADT xyt_handle) 576d71ae5a4SJacob Faibussowitsch { 57752f87cdaSBarry Smith PetscInt i, ct, id; 57852f87cdaSBarry Smith PetscInt mask, edge, *iptr; 57952f87cdaSBarry Smith PetscInt *dir, *used; 58052f87cdaSBarry Smith PetscInt sum[4], w[4]; 581a501084fSBarry Smith PetscScalar rsum[4], rw[4]; 58252f87cdaSBarry Smith PetscInt op[] = {GL_ADD, 0}; 583a501084fSBarry Smith PetscScalar *lhs, *rhs; 58452f87cdaSBarry Smith PetscInt *nsep, *lnsep, *fo, nfo = 0; 585ca8e9878SJed Brown PCTFS_gs_ADT PCTFS_gs_handle = xyt_handle->mvi->PCTFS_gs_handle; 58652f87cdaSBarry Smith PetscInt *local2global = xyt_handle->mvi->local2global; 58752f87cdaSBarry Smith PetscInt n = xyt_handle->mvi->n; 58852f87cdaSBarry Smith PetscInt m = xyt_handle->mvi->m; 58952f87cdaSBarry Smith PetscInt level = xyt_handle->level; 590ab824b78SBarry Smith PetscInt shared = 0; 591827bd09bSSatish Balay 5920924e98cSBarry Smith PetscFunctionBegin; 59352f87cdaSBarry Smith dir = (PetscInt *)malloc(sizeof(PetscInt) * (level + 1)); 59452f87cdaSBarry Smith nsep = (PetscInt *)malloc(sizeof(PetscInt) * (level + 1)); 59552f87cdaSBarry Smith lnsep = (PetscInt *)malloc(sizeof(PetscInt) * (level + 1)); 59652f87cdaSBarry Smith fo = (PetscInt *)malloc(sizeof(PetscInt) * (n + 1)); 59752f87cdaSBarry Smith used = (PetscInt *)malloc(sizeof(PetscInt) * n); 598827bd09bSSatish Balay 599*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_ivec_zero(dir, level + 1)); 600*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_ivec_zero(nsep, level + 1)); 601*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_ivec_zero(lnsep, level + 1)); 602*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_ivec_set(fo, -1, n + 1)); 603*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_ivec_zero(used, n)); 604827bd09bSSatish Balay 6058cda6cd7SBarry Smith lhs = (PetscScalar *)malloc(sizeof(PetscScalar) * m); 6068cda6cd7SBarry Smith rhs = (PetscScalar *)malloc(sizeof(PetscScalar) * m); 607827bd09bSSatish Balay 608827bd09bSSatish Balay /* determine the # of unique dof */ 609*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(lhs, m)); 610*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_set(lhs, 1.0, n)); 611*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_gs_gop_hc(PCTFS_gs_handle, lhs, "+\0", level)); 6129566063dSJacob Faibussowitsch PetscCall(PetscInfo(0, "done first PCTFS_gs_gop_hc\n")); 613*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_rvec_zero(rsum, 2)); 614947b95d8SBarry Smith for (i = 0; i < n; i++) { 6159371c9d4SSatish Balay if (lhs[i] != 0.0) { 6169371c9d4SSatish Balay rsum[0] += 1.0 / lhs[i]; 6179371c9d4SSatish Balay rsum[1] += lhs[i]; 6189371c9d4SSatish Balay } 6192fa5cd67SKarl Rupp if (lhs[i] != 1.0) shared = 1; 620827bd09bSSatish Balay } 621827bd09bSSatish Balay 622*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_grop_hc(rsum, rw, 2, op, level)); 623827bd09bSSatish Balay rsum[0] += 0.1; 624827bd09bSSatish Balay rsum[1] += 0.1; 625827bd09bSSatish Balay 62652f87cdaSBarry Smith xyt_handle->info->n_global = xyt_handle->info->m_global = (PetscInt)rsum[0]; 62752f87cdaSBarry Smith xyt_handle->mvi->n_global = xyt_handle->mvi->m_global = (PetscInt)rsum[0]; 628827bd09bSSatish Balay 629827bd09bSSatish Balay /* determine separator sets top down */ 6300fdf79fbSJacob Faibussowitsch PetscCheck(!shared, PETSC_COMM_SELF, PETSC_ERR_PLIB, "shared dof separator determination not ready ... see hmt!!!"); 631827bd09bSSatish Balay /* solution is to do as in the symmetric shared case but then */ 632827bd09bSSatish Balay /* pick the sub-hc with the most free dofs and do a mat-vec */ 633827bd09bSSatish Balay /* and pick up the responses on the other sub-hc from the */ 634827bd09bSSatish Balay /* initial separator set obtained from the symm. shared case */ 6354e619c20SJed Brown /* [dead code deleted since it is unlikely to be completed] */ 636db4deed7SKarl Rupp for (iptr = fo + n, id = PCTFS_my_id, mask = PCTFS_num_nodes >> 1, edge = level; edge > 0; edge--, mask >>= 1) { 637827bd09bSSatish Balay /* set rsh of hc, fire, and collect lhs responses */ 638*3ba16761SJacob Faibussowitsch PetscCall((id < mask) ? PCTFS_rvec_zero(lhs, m) : PCTFS_rvec_set(lhs, 1.0, m)); 639*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_gs_gop_hc(PCTFS_gs_handle, lhs, "+\0", edge)); 640827bd09bSSatish Balay 641827bd09bSSatish Balay /* set lsh of hc, fire, and collect rhs responses */ 642*3ba16761SJacob Faibussowitsch PetscCall((id < mask) ? PCTFS_rvec_set(rhs, 1.0, m) : PCTFS_rvec_zero(rhs, m)); 643*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_gs_gop_hc(PCTFS_gs_handle, rhs, "+\0", edge)); 644827bd09bSSatish Balay 645827bd09bSSatish Balay /* count number of dofs I own that have signal and not in sep set */ 646*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_ivec_zero(sum, 4)); 647*3ba16761SJacob Faibussowitsch for (ct = i = 0; i < n; i++) { 648db4deed7SKarl Rupp if (!used[i]) { 649827bd09bSSatish Balay /* number of unmarked dofs on node */ 650827bd09bSSatish Balay ct++; 651827bd09bSSatish Balay /* number of dofs to be marked on lhs hc */ 6522fa5cd67SKarl Rupp if ((id < mask) && (lhs[i] != 0.0)) sum[0]++; 653827bd09bSSatish Balay /* number of dofs to be marked on rhs hc */ 6542fa5cd67SKarl Rupp if ((id >= mask) && (rhs[i] != 0.0)) sum[1]++; 655827bd09bSSatish Balay } 656827bd09bSSatish Balay } 657827bd09bSSatish Balay 658827bd09bSSatish Balay /* for the non-symmetric case we need separators of width 2 */ 659827bd09bSSatish Balay /* so take both sides */ 660827bd09bSSatish Balay (id < mask) ? (sum[2] = ct) : (sum[3] = ct); 661*3ba16761SJacob Faibussowitsch PetscCall(PCTFS_giop_hc(sum, w, 4, op, edge)); 662827bd09bSSatish Balay 663827bd09bSSatish Balay ct = 0; 664db4deed7SKarl Rupp if (id < mask) { 665827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 666db4deed7SKarl Rupp for (i = 0; i < n; i++) { 667db4deed7SKarl Rupp if ((!used[i]) && (lhs[i] != 0.0)) { 6689371c9d4SSatish Balay ct++; 6699371c9d4SSatish Balay nfo++; 670827bd09bSSatish Balay *--iptr = local2global[i]; 671827bd09bSSatish Balay used[i] = edge; 672827bd09bSSatish Balay } 673827bd09bSSatish Balay } 674827bd09bSSatish Balay /* LSH hc summation of ct should be sum[0] */ 675db4deed7SKarl Rupp } else { 676827bd09bSSatish Balay /* mark dofs I own that have signal and not in sep set */ 677db4deed7SKarl Rupp for (i = 0; i < n; i++) { 678db4deed7SKarl Rupp if ((!used[i]) && (rhs[i] != 0.0)) { 6799371c9d4SSatish Balay ct++; 6809371c9d4SSatish Balay nfo++; 681827bd09bSSatish Balay *--iptr = local2global[i]; 682827bd09bSSatish Balay used[i] = edge; 683827bd09bSSatish Balay } 684827bd09bSSatish Balay } 685827bd09bSSatish Balay /* RSH hc summation of ct should be sum[1] */ 686827bd09bSSatish Balay } 687827bd09bSSatish Balay 688*3ba16761SJacob Faibussowitsch if (ct > 1) PetscCall(PCTFS_ivec_sort(iptr, ct)); 689827bd09bSSatish Balay lnsep[edge] = ct; 690827bd09bSSatish Balay nsep[edge] = sum[0] + sum[1]; 691827bd09bSSatish Balay dir[edge] = BOTH; 692827bd09bSSatish Balay 693827bd09bSSatish Balay /* LATER or we can recur on these to order seps at this level */ 694827bd09bSSatish Balay /* do we need full set of separators for this? */ 695827bd09bSSatish Balay 696827bd09bSSatish Balay /* fold rhs hc into lower */ 6972fa5cd67SKarl Rupp if (id >= mask) id -= mask; 698827bd09bSSatish Balay } 699827bd09bSSatish Balay 700827bd09bSSatish Balay /* level 0 is on processor case - so mark the remainder */ 701db4deed7SKarl Rupp for (ct = i = 0; i < n; i++) { 702db4deed7SKarl Rupp if (!used[i]) { 7039371c9d4SSatish Balay ct++; 7049371c9d4SSatish Balay nfo++; 705827bd09bSSatish Balay *--iptr = local2global[i]; 706827bd09bSSatish Balay used[i] = edge; 707827bd09bSSatish Balay } 708827bd09bSSatish Balay } 709*3ba16761SJacob Faibussowitsch if (ct > 1) PetscCall(PCTFS_ivec_sort(iptr, ct)); 710827bd09bSSatish Balay lnsep[edge] = ct; 711827bd09bSSatish Balay nsep[edge] = ct; 712827bd09bSSatish Balay dir[edge] = BOTH; 713827bd09bSSatish Balay 714827bd09bSSatish Balay xyt_handle->info->nsep = nsep; 715827bd09bSSatish Balay xyt_handle->info->lnsep = lnsep; 716827bd09bSSatish Balay xyt_handle->info->fo = fo; 717827bd09bSSatish Balay xyt_handle->info->nfo = nfo; 718827bd09bSSatish Balay 719a501084fSBarry Smith free(dir); 720a501084fSBarry Smith free(lhs); 721a501084fSBarry Smith free(rhs); 722a501084fSBarry Smith free(used); 723*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 724827bd09bSSatish Balay } 725827bd09bSSatish Balay 726d71ae5a4SJacob Faibussowitsch static mv_info *set_mvi(PetscInt *local2global, PetscInt n, PetscInt m, PetscErrorCode (*matvec)(mv_info *, PetscScalar *, PetscScalar *), void *grid_data) 727d71ae5a4SJacob Faibussowitsch { 728827bd09bSSatish Balay mv_info *mvi; 729827bd09bSSatish Balay 730a501084fSBarry Smith mvi = (mv_info *)malloc(sizeof(mv_info)); 731827bd09bSSatish Balay mvi->n = n; 732827bd09bSSatish Balay mvi->m = m; 733827bd09bSSatish Balay mvi->n_global = -1; 734827bd09bSSatish Balay mvi->m_global = -1; 73552f87cdaSBarry Smith mvi->local2global = (PetscInt *)malloc((m + 1) * sizeof(PetscInt)); 7362fa5cd67SKarl Rupp 737ca8e9878SJed Brown PCTFS_ivec_copy(mvi->local2global, local2global, m); 738827bd09bSSatish Balay mvi->local2global[m] = INT_MAX; 7395c8f6a95SKarl Rupp mvi->matvec = matvec; 740827bd09bSSatish Balay mvi->grid_data = grid_data; 741827bd09bSSatish Balay 742827bd09bSSatish Balay /* set xyt communication handle to perform restricted matvec */ 743ca8e9878SJed Brown mvi->PCTFS_gs_handle = PCTFS_gs_init(local2global, m, PCTFS_num_nodes); 744827bd09bSSatish Balay 745827bd09bSSatish Balay return (mvi); 746827bd09bSSatish Balay } 747827bd09bSSatish Balay 748d71ae5a4SJacob Faibussowitsch static PetscErrorCode do_matvec(mv_info *A, PetscScalar *v, PetscScalar *u) 749d71ae5a4SJacob Faibussowitsch { 7500924e98cSBarry Smith PetscFunctionBegin; 751*3ba16761SJacob Faibussowitsch PetscCall(A->matvec((mv_info *)A->grid_data, v, u)); 752*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 753827bd09bSSatish Balay } 754