Actual source code: plexland.c
1: #include <../src/mat/impls/aij/seq/aij.h>
2: #include <petsc/private/dmpleximpl.h>
3: #include <petsclandau.h>
4: #include <petscts.h>
5: #include <petscdmforest.h>
6: #include <petscdmcomposite.h>
8: /* Landau collision operator */
10: /* relativistic terms */
11: #if PetscDefined(USE_REAL_SINGLE)
12: #define SPEED_OF_LIGHT 2.99792458e8F
13: #define C_0(v0) (SPEED_OF_LIGHT / v0) /* needed for relativistic tensor on all architectures */
14: #else
15: #define SPEED_OF_LIGHT 2.99792458e8
16: #define C_0(v0) (SPEED_OF_LIGHT / v0) /* needed for relativistic tensor on all architectures */
17: #endif
19: #include "land_tensors.h"
21: #if PetscDefined(HAVE_OPENMP)
22: #include <omp.h>
23: #endif
25: static PetscErrorCode LandauGPUMapsDestroy(PetscCtxRt ptr)
26: {
27: P4estVertexMaps *maps = *(P4estVertexMaps **)ptr;
29: PetscFunctionBegin;
30: // free device data
31: if (maps[0].deviceType != LANDAU_CPU) {
32: #if PetscDefined(HAVE_KOKKOS)
33: if (maps[0].deviceType == LANDAU_KOKKOS) PetscCall(LandauKokkosDestroyMatMaps(maps, maps[0].numgrids)); // implies Kokkos does
34: #endif
35: }
36: // free host data
37: for (PetscInt grid = 0; grid < maps[0].numgrids; grid++) {
38: PetscCall(PetscFree(maps[grid].c_maps));
39: PetscCall(PetscFree(maps[grid].gIdx));
40: }
41: PetscCall(PetscFree(maps));
42: PetscFunctionReturn(PETSC_SUCCESS);
43: }
44: static PetscErrorCode energy_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
45: {
46: PetscReal v2 = 0;
48: PetscFunctionBegin;
49: /* compute v^2 / 2 */
50: for (PetscInt i = 0; i < dim; ++i) v2 += x[i] * x[i];
51: /* evaluate the Maxwellian */
52: u[0] = v2 / 2;
53: PetscFunctionReturn(PETSC_SUCCESS);
54: }
56: /* needs double */
57: static PetscErrorCode gamma_m1_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
58: {
59: PetscReal *c2_0_arr = ((PetscReal *)actx);
60: double u2 = 0, c02 = (double)*c2_0_arr, xx;
62: PetscFunctionBegin;
63: /* compute u^2 / 2 */
64: for (PetscInt i = 0; i < dim; ++i) u2 += x[i] * x[i];
65: /* gamma - 1 = g_eps, for conditioning and we only take derivatives */
66: xx = u2 / c02;
67: if (PetscDefined(USE_DEBUG)) u[0] = PetscSqrtReal(1. + xx);
68: else u[0] = xx / (PetscSqrtReal(1. + xx) + 1.) - 1.; // better conditioned. -1 might help condition and only used for derivative
69: PetscFunctionReturn(PETSC_SUCCESS);
70: }
72: /* Evaluates Jacobian matrix; fills JacP, does not create it */
73: static PetscErrorCode LandauFormJacobian_Internal(Vec a_X, Mat JacP, const PetscInt dim, PetscReal shift, void *a_ctx)
74: {
75: LandauCtx *ctx = (LandauCtx *)a_ctx;
76: PetscInt numCells[LANDAU_MAX_GRIDS], Nq, Nb;
77: PetscQuadrature quad;
78: PetscReal Eq_m[LANDAU_MAX_SPECIES]; // could be static data w/o quench (ex2)
79: PetscScalar *cellClosure = NULL;
80: const PetscScalar *xdata = NULL;
81: PetscDS prob;
82: PetscContainer container;
83: P4estVertexMaps *maps;
84: Mat subJ[LANDAU_MAX_GRIDS * LANDAU_MAX_BATCH_SZ];
86: PetscFunctionBegin;
89: PetscAssertPointer(ctx, 5);
90: /* check for matrix container for GPU assembly. Support CPU assembly for debugging */
91: PetscCheck(ctx->plex[0] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
92: PetscCall(PetscLogEventBegin(ctx->events[10], 0, 0, 0, 0));
93: PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids
94: PetscCall(PetscObjectQuery((PetscObject)JacP, "assembly_maps", (PetscObject *)&container));
95: if (container) {
96: PetscCheck(ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "maps but no GPU assembly");
97: PetscCall(PetscContainerGetPointer(container, &maps));
98: PetscCheck(maps, ctx->comm, PETSC_ERR_ARG_WRONG, "empty GPU matrix container");
99: for (PetscInt i = 0; i < ctx->num_grids * ctx->batch_sz; i++) subJ[i] = NULL;
100: } else {
101: PetscCheck(!ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "No maps but GPU assembly");
102: for (PetscInt tid = 0; tid < ctx->batch_sz; tid++) {
103: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCreateMatrix(ctx->plex[grid], &subJ[LAND_PACK_IDX(tid, grid)]));
104: }
105: maps = NULL;
106: }
107: // get dynamic data (Eq is odd, for quench and Spitzer test) for CPU assembly and raw data for Jacobian GPU assembly. Get host numCells[], Nq (yuck)
108: PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad));
109: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL));
110: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
111: PetscCheck(Nq <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nq, LANDAU_MAX_NQND);
112: PetscCheck(Nb <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nb = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nb, LANDAU_MAX_NQND);
113: // get metadata for collecting dynamic data
114: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
115: PetscInt cStart, cEnd;
116: PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
117: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
118: numCells[grid] = cEnd - cStart; // grids can have different topology
119: }
120: PetscCall(PetscLogEventEnd(ctx->events[10], 0, 0, 0, 0));
121: if (shift == 0) { /* create dynamic point data: f_alpha for closure of each cell (cellClosure[nbatch,ngrids,ncells[g],f[Nb,ns[g]]]) or xdata */
122: DM pack;
123: PetscCall(VecGetDM(a_X, &pack));
124: PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "pack has no DM");
125: PetscCall(PetscLogEventBegin(ctx->events[1], 0, 0, 0, 0));
126: for (PetscInt fieldA = 0; fieldA < ctx->num_species; fieldA++) {
127: Eq_m[fieldA] = ctx->Ez * ctx->t_0 * ctx->charges[fieldA] / (ctx->v_0 * ctx->masses[fieldA]); /* normalize dimensionless */
128: if (dim == 2) Eq_m[fieldA] *= 2 * PETSC_PI; /* add the 2pi term that is not in Landau */
129: }
130: if (!ctx->gpu_assembly) {
131: Vec *locXArray, *globXArray;
132: PetscScalar *cellClosure_it;
133: PetscInt cellClosure_sz = 0, nDMs, Nf[LANDAU_MAX_GRIDS];
134: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
135: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
136: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
137: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
138: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
139: }
140: /* count cellClosure size */
141: PetscCall(DMCompositeGetNumberDM(pack, &nDMs));
142: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) cellClosure_sz += Nb * Nf[grid] * numCells[grid];
143: PetscCall(PetscMalloc1(cellClosure_sz * ctx->batch_sz, &cellClosure));
144: cellClosure_it = cellClosure;
145: PetscCall(PetscMalloc(sizeof(*locXArray) * nDMs, &locXArray));
146: PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray));
147: PetscCall(DMCompositeGetLocalAccessArray(pack, a_X, nDMs, NULL, locXArray));
148: PetscCall(DMCompositeGetAccessArray(pack, a_X, nDMs, NULL, globXArray));
149: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP (once)
150: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
151: Vec locX = locXArray[LAND_PACK_IDX(b_id, grid)], globX = globXArray[LAND_PACK_IDX(b_id, grid)], locX2;
152: PetscInt cStart, cEnd, ei;
153: PetscCall(VecDuplicate(locX, &locX2));
154: PetscCall(DMGlobalToLocalBegin(ctx->plex[grid], globX, INSERT_VALUES, locX2));
155: PetscCall(DMGlobalToLocalEnd(ctx->plex[grid], globX, INSERT_VALUES, locX2));
156: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
157: for (ei = cStart; ei < cEnd; ++ei) {
158: PetscScalar *coef = NULL;
159: PetscCall(DMPlexVecGetClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef));
160: PetscCall(PetscMemcpy(cellClosure_it, coef, Nb * Nf[grid] * sizeof(*cellClosure_it))); /* change if LandauIPReal != PetscScalar */
161: PetscCall(DMPlexVecRestoreClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef));
162: cellClosure_it += Nb * Nf[grid];
163: }
164: PetscCall(VecDestroy(&locX2));
165: }
166: }
167: PetscCheck(cellClosure_it - cellClosure == cellClosure_sz * ctx->batch_sz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "iteration wrong %" PetscCount_FMT " != cellClosure_sz = %" PetscInt_FMT, cellClosure_it - cellClosure, cellClosure_sz * ctx->batch_sz);
168: PetscCall(DMCompositeRestoreLocalAccessArray(pack, a_X, nDMs, NULL, locXArray));
169: PetscCall(DMCompositeRestoreAccessArray(pack, a_X, nDMs, NULL, globXArray));
170: PetscCall(PetscFree(locXArray));
171: PetscCall(PetscFree(globXArray));
172: xdata = NULL;
173: } else {
174: PetscMemType mtype;
175: if (ctx->jacobian_field_major_order) { // get data in batch ordering
176: PetscCall(VecScatterBegin(ctx->plex_batch, a_X, ctx->work_vec, INSERT_VALUES, SCATTER_FORWARD));
177: PetscCall(VecScatterEnd(ctx->plex_batch, a_X, ctx->work_vec, INSERT_VALUES, SCATTER_FORWARD));
178: PetscCall(VecGetArrayReadAndMemType(ctx->work_vec, &xdata, &mtype));
179: } else {
180: PetscCall(VecGetArrayReadAndMemType(a_X, &xdata, &mtype));
181: }
182: PetscCheck(mtype == PETSC_MEMTYPE_HOST || ctx->deviceType != LANDAU_CPU, ctx->comm, PETSC_ERR_ARG_WRONG, "CPU run with device data: use -mat_type aij");
183: cellClosure = NULL;
184: }
185: PetscCall(PetscLogEventEnd(ctx->events[1], 0, 0, 0, 0));
186: } else xdata = cellClosure = NULL;
188: /* do it */
189: if (ctx->deviceType == LANDAU_KOKKOS) {
190: #if PetscDefined(HAVE_KOKKOS)
191: PetscCall(LandauKokkosJacobian(ctx->plex, Nq, Nb, ctx->batch_sz, ctx->num_grids, numCells, Eq_m, cellClosure, xdata, &ctx->SData_d, shift, ctx->events, ctx->mat_offset, ctx->species_offset, subJ, JacP));
192: #else
193: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos");
194: #endif
195: } else { /* CPU version */
196: PetscTabulation *Tf; // used for CPU and print info. Same on all grids and all species
197: PetscInt ip_offset[LANDAU_MAX_GRIDS + 1], ipf_offset[LANDAU_MAX_GRIDS + 1], elem_offset[LANDAU_MAX_GRIDS + 1], IPf_sz_glb, IPf_sz_tot, num_grids = ctx->num_grids, Nf[LANDAU_MAX_GRIDS];
198: PetscReal *ff, *dudx, *dudy, *dudz, *invJ_a = (PetscReal *)ctx->SData_d.invJ, *xx = (PetscReal *)ctx->SData_d.x, *yy = (PetscReal *)ctx->SData_d.y, *zz = (PetscReal *)ctx->SData_d.z, *ww = (PetscReal *)ctx->SData_d.w;
199: PetscReal *nu_alpha = (PetscReal *)ctx->SData_d.alpha, *nu_beta = (PetscReal *)ctx->SData_d.beta, *invMass = (PetscReal *)ctx->SData_d.invMass;
200: PetscReal (*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal (*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas;
201: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
202: PetscScalar *coo_vals = NULL;
203: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
204: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
205: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
206: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
207: }
208: /* count IPf size, etc */
209: PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids
210: const PetscReal *const BB = Tf[0]->T[0], *const DD = Tf[0]->T[1];
211: ip_offset[0] = ipf_offset[0] = elem_offset[0] = 0;
212: for (PetscInt grid = 0; grid < num_grids; grid++) {
213: PetscInt nfloc = ctx->species_offset[grid + 1] - ctx->species_offset[grid];
214: elem_offset[grid + 1] = elem_offset[grid] + numCells[grid];
215: ip_offset[grid + 1] = ip_offset[grid] + numCells[grid] * Nq;
216: ipf_offset[grid + 1] = ipf_offset[grid] + Nq * nfloc * numCells[grid];
217: }
218: IPf_sz_glb = ipf_offset[num_grids];
219: IPf_sz_tot = IPf_sz_glb * ctx->batch_sz;
220: // prep COO
221: PetscCall(PetscMalloc1(ctx->SData_d.coo_size, &coo_vals)); // allocate every time?
222: if (shift == 0.0) { /* compute dynamic data f and df and init data for Jacobian */
223: #if PetscDefined(HAVE_THREADSAFETY)
224: double starttime, endtime;
225: starttime = MPI_Wtime();
226: #endif
227: PetscCall(PetscLogEventBegin(ctx->events[8], 0, 0, 0, 0));
228: PetscCall(PetscMalloc4(IPf_sz_tot, &ff, IPf_sz_tot, &dudx, IPf_sz_tot, &dudy, (dim == 3 ? IPf_sz_tot : 0), &dudz));
229: // F df/dx
230: for (PetscInt tid = 0; tid < ctx->batch_sz * elem_offset[num_grids]; tid++) { // for each element
231: const PetscInt b_Nelem = elem_offset[num_grids], b_elem_idx = tid % b_Nelem, b_id = tid / b_Nelem; // b_id == OMP thd_id in batch
232: // find my grid:
233: PetscInt grid = 0;
234: while (b_elem_idx >= elem_offset[grid + 1]) grid++; // yuck search for grid
235: {
236: const PetscInt loc_nip = numCells[grid] * Nq, loc_Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], loc_elem = b_elem_idx - elem_offset[grid];
237: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset); //b_id*b_N + ctx->mat_offset[grid];
238: PetscScalar *coef, coef_buff[LANDAU_MAX_SPECIES * LANDAU_MAX_NQND];
239: PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem * Nq) * dim * dim]; // ingJ is static data on batch 0
240: PetscInt b, f, q;
241: if (cellClosure) {
242: coef = &cellClosure[b_id * IPf_sz_glb + ipf_offset[grid] + loc_elem * Nb * loc_Nf]; // this is const
243: } else {
244: coef = coef_buff;
245: for (f = 0; f < loc_Nf; ++f) {
246: LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][f][0];
247: for (b = 0; b < Nb; ++b) {
248: PetscInt idx = Idxs[b];
249: if (idx >= 0) {
250: coef[f * Nb + b] = xdata[idx + moffset];
251: } else {
252: idx = -idx - 1;
253: coef[f * Nb + b] = 0;
254: for (q = 0; q < maps[grid].num_face; q++) {
255: PetscInt id = maps[grid].c_maps[idx][q].gid;
256: PetscScalar scale = maps[grid].c_maps[idx][q].scale;
257: coef[f * Nb + b] += scale * xdata[id + moffset];
258: }
259: }
260: }
261: }
262: }
263: /* get f and df */
264: for (PetscInt qi = 0; qi < Nq; qi++) {
265: const PetscReal *invJ = &invJe[qi * dim * dim];
266: const PetscReal *Bq = &BB[qi * Nb];
267: const PetscReal *Dq = &DD[qi * Nb * dim];
268: PetscReal u_x[LANDAU_DIM];
269: /* get f & df */
270: for (f = 0; f < loc_Nf; ++f) {
271: const PetscInt idx = b_id * IPf_sz_glb + ipf_offset[grid] + f * loc_nip + loc_elem * Nq + qi;
272: PetscInt e;
273: PetscReal refSpaceDer[LANDAU_DIM];
274: ff[idx] = 0.0;
275: for (PetscInt d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0;
276: for (PetscInt b = 0; b < Nb; ++b) {
277: const PetscInt cidx = b;
278: ff[idx] += Bq[cidx] * PetscRealPart(coef[f * Nb + cidx]);
279: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] += Dq[cidx * dim + d] * PetscRealPart(coef[f * Nb + cidx]);
280: }
281: for (PetscInt d = 0; d < LANDAU_DIM; ++d) {
282: for (e = 0, u_x[d] = 0.0; e < LANDAU_DIM; ++e) u_x[d] += invJ[e * dim + d] * refSpaceDer[e];
283: }
284: dudx[idx] = u_x[0];
285: dudy[idx] = u_x[1];
286: #if LANDAU_DIM == 3
287: dudz[idx] = u_x[2];
288: #endif
289: }
290: } // q
291: } // grid
292: } // grid*batch
293: PetscCall(PetscLogEventEnd(ctx->events[8], 0, 0, 0, 0));
294: #if PetscDefined(HAVE_THREADSAFETY)
295: endtime = MPI_Wtime();
296: if (ctx->stage) ctx->times[LANDAU_F_DF] += (endtime - starttime);
297: #endif
298: } // Jacobian setup
299: // assemble Jacobian (or mass)
300: for (PetscInt tid = 0; tid < ctx->batch_sz * elem_offset[num_grids]; tid++) { // for each element
301: const PetscInt b_Nelem = elem_offset[num_grids];
302: const PetscInt glb_elem_idx = tid % b_Nelem, b_id = tid / b_Nelem;
303: PetscInt grid = 0;
304: #if PetscDefined(HAVE_THREADSAFETY)
305: double starttime, endtime;
306: starttime = MPI_Wtime();
307: #endif
308: while (glb_elem_idx >= elem_offset[grid + 1]) grid++;
309: {
310: const PetscInt loc_Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], loc_elem = glb_elem_idx - elem_offset[grid];
311: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset), totDim = loc_Nf * Nq, elemMatSize = totDim * totDim;
312: PetscScalar *elemMat;
313: const PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem * Nq) * dim * dim];
314: PetscCall(PetscMalloc1(elemMatSize, &elemMat));
315: PetscCall(PetscMemzero(elemMat, elemMatSize * sizeof(*elemMat)));
316: if (shift == 0.0) { // Jacobian
317: PetscCall(PetscLogEventBegin(ctx->events[4], 0, 0, 0, 0));
318: } else { // mass
319: PetscCall(PetscLogEventBegin(ctx->events[16], 0, 0, 0, 0));
320: }
321: for (PetscInt qj = 0; qj < Nq; ++qj) {
322: const PetscInt jpidx_glb = ip_offset[grid] + qj + loc_elem * Nq;
323: PetscReal g0[LANDAU_MAX_SPECIES], g2[LANDAU_MAX_SPECIES][LANDAU_DIM], g3[LANDAU_MAX_SPECIES][LANDAU_DIM][LANDAU_DIM]; // could make a LANDAU_MAX_SPECIES_GRID ~ number of ions - 1
324: PetscInt d, d2, dp, d3, IPf_idx;
325: if (shift == 0.0) { // Jacobian
326: const PetscReal *const invJj = &invJe[qj * dim * dim];
327: PetscReal gg2[LANDAU_MAX_SPECIES][LANDAU_DIM], gg3[LANDAU_MAX_SPECIES][LANDAU_DIM][LANDAU_DIM], gg2_temp[LANDAU_DIM], gg3_temp[LANDAU_DIM][LANDAU_DIM];
328: const PetscReal vj[3] = {xx[jpidx_glb], yy[jpidx_glb], zz ? zz[jpidx_glb] : 0}, wj = ww[jpidx_glb];
329: // create g2 & g3
330: for (d = 0; d < LANDAU_DIM; d++) { // clear accumulation data D & K
331: gg2_temp[d] = 0;
332: for (d2 = 0; d2 < LANDAU_DIM; d2++) gg3_temp[d][d2] = 0;
333: }
334: /* inner beta reduction */
335: IPf_idx = 0;
336: for (PetscInt grid_r = 0, f_off = 0, ipidx = 0; grid_r < ctx->num_grids; grid_r++, f_off = ctx->species_offset[grid_r]) { // IPf_idx += nip_loc_r*Nfloc_r
337: PetscInt nip_loc_r = numCells[grid_r] * Nq, Nfloc_r = Nf[grid_r];
338: for (PetscInt ei_r = 0; ei_r < numCells[grid_r]; ++ei_r) {
339: for (PetscInt qi = 0; qi < Nq; qi++, ipidx++) {
340: const PetscReal wi = ww[ipidx], x = xx[ipidx], y = yy[ipidx];
341: PetscReal temp1[3] = {0, 0, 0}, temp2 = 0;
342: #if LANDAU_DIM == 2
343: PetscReal Ud[2][2], Uk[2][2], mask = (PetscAbs(vj[0] - x) < 100 * PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[1] - y) < 100 * PETSC_SQRT_MACHINE_EPSILON) ? 0. : 1.;
344: LandauTensor2D(vj, x, y, Ud, Uk, mask);
345: #else
346: PetscReal U[3][3], z = zz[ipidx], mask = (PetscAbs(vj[0] - x) < 100 * PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[1] - y) < 100 * PETSC_SQRT_MACHINE_EPSILON && PetscAbs(vj[2] - z) < 100 * PETSC_SQRT_MACHINE_EPSILON) ? 0. : 1.;
347: if (ctx->use_relativistic_corrections) {
348: LandauTensor3DRelativistic(vj, x, y, z, U, mask, C_0(ctx->v_0));
349: } else {
350: LandauTensor3D(vj, x, y, z, U, mask);
351: }
352: #endif
353: for (PetscInt f = 0; f < Nfloc_r; ++f) {
354: const PetscInt idx = b_id * IPf_sz_glb + ipf_offset[grid_r] + f * nip_loc_r + ei_r * Nq + qi;
356: temp1[0] += dudx[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
357: temp1[1] += dudy[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
358: #if LANDAU_DIM == 3
359: temp1[2] += dudz[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
360: #endif
361: temp2 += ff[idx] * nu_beta[f + f_off] * (*lambdas)[grid][grid_r];
362: }
363: temp1[0] *= wi;
364: temp1[1] *= wi;
365: #if LANDAU_DIM == 3
366: temp1[2] *= wi;
367: #endif
368: temp2 *= wi;
369: #if LANDAU_DIM == 2
370: for (d2 = 0; d2 < 2; d2++) {
371: for (d3 = 0; d3 < 2; ++d3) {
372: /* K = U * grad(f): g2=e: i,A */
373: gg2_temp[d2] += Uk[d2][d3] * temp1[d3];
374: /* D = -U * (I \kron (fx)): g3=f: i,j,A */
375: gg3_temp[d2][d3] += Ud[d2][d3] * temp2;
376: }
377: }
378: #else
379: for (d2 = 0; d2 < 3; ++d2) {
380: for (d3 = 0; d3 < 3; ++d3) {
381: /* K = U * grad(f): g2 = e: i,A */
382: gg2_temp[d2] += U[d2][d3] * temp1[d3];
383: /* D = -U * (I \kron (fx)): g3 = f: i,j,A */
384: gg3_temp[d2][d3] += U[d2][d3] * temp2;
385: }
386: }
387: #endif
388: } // qi
389: } // ei_r
390: IPf_idx += nip_loc_r * Nfloc_r;
391: } /* grid_r - IPs */
392: PetscCheck(IPf_idx == IPf_sz_glb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "IPf_idx != IPf_sz %" PetscInt_FMT " %" PetscInt_FMT, IPf_idx, IPf_sz_glb);
393: // add alpha and put in gg2/3
394: for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) {
395: for (d2 = 0; d2 < LANDAU_DIM; d2++) {
396: gg2[fieldA][d2] = gg2_temp[d2] * nu_alpha[fieldA + f_off];
397: for (d3 = 0; d3 < LANDAU_DIM; d3++) gg3[fieldA][d2][d3] = -gg3_temp[d2][d3] * nu_alpha[fieldA + f_off] * invMass[fieldA + f_off];
398: }
399: }
400: /* add electric field term once per IP */
401: for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) gg2[fieldA][LANDAU_DIM - 1] += Eq_m[fieldA + f_off];
402: /* Jacobian transform - g2, g3 */
403: for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) {
404: for (d = 0; d < dim; ++d) {
405: g2[fieldA][d] = 0.0;
406: for (d2 = 0; d2 < dim; ++d2) {
407: g2[fieldA][d] += invJj[d * dim + d2] * gg2[fieldA][d2];
408: g3[fieldA][d][d2] = 0.0;
409: for (d3 = 0; d3 < dim; ++d3) {
410: for (dp = 0; dp < dim; ++dp) g3[fieldA][d][d2] += invJj[d * dim + d3] * gg3[fieldA][d3][dp] * invJj[d2 * dim + dp];
411: }
412: g3[fieldA][d][d2] *= wj;
413: }
414: g2[fieldA][d] *= wj;
415: }
416: }
417: } else { // mass
418: PetscReal wj = ww[jpidx_glb];
419: /* Jacobian transform - g0 */
420: for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) {
421: if (dim == 2) {
422: g0[fieldA] = wj * shift * 2. * PETSC_PI; // move this to below and remove g0
423: } else {
424: g0[fieldA] = wj * shift; // move this to below and remove g0
425: }
426: }
427: }
428: /* FE matrix construction */
429: {
430: PetscInt fieldA, d, f, d2, g;
431: const PetscReal *BJq = &BB[qj * Nb], *DIq = &DD[qj * Nb * dim];
432: /* assemble - on the diagonal (I,I) */
433: for (fieldA = 0; fieldA < loc_Nf; fieldA++) {
434: for (f = 0; f < Nb; f++) {
435: const PetscInt i = fieldA * Nb + f; /* Element matrix row */
436: for (g = 0; g < Nb; ++g) {
437: const PetscInt j = fieldA * Nb + g; /* Element matrix column */
438: const PetscInt fOff = i * totDim + j;
439: if (shift == 0.0) {
440: for (d = 0; d < dim; ++d) {
441: elemMat[fOff] += DIq[f * dim + d] * g2[fieldA][d] * BJq[g];
442: for (d2 = 0; d2 < dim; ++d2) elemMat[fOff] += DIq[f * dim + d] * g3[fieldA][d][d2] * DIq[g * dim + d2];
443: }
444: } else { // mass
445: elemMat[fOff] += BJq[f] * g0[fieldA] * BJq[g];
446: }
447: }
448: }
449: }
450: }
451: } /* qj loop */
452: if (shift == 0.0) { // Jacobian
453: PetscCall(PetscLogEventEnd(ctx->events[4], 0, 0, 0, 0));
454: } else {
455: PetscCall(PetscLogEventEnd(ctx->events[16], 0, 0, 0, 0));
456: }
457: #if PetscDefined(HAVE_THREADSAFETY)
458: endtime = MPI_Wtime();
459: if (ctx->stage) ctx->times[LANDAU_KERNEL] += (endtime - starttime);
460: #endif
461: /* assemble matrix */
462: if (!container) {
463: PetscInt cStart;
464: PetscCall(PetscLogEventBegin(ctx->events[6], 0, 0, 0, 0));
465: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, NULL));
466: PetscCall(DMPlexMatSetClosure(ctx->plex[grid], section[grid], globsection[grid], subJ[LAND_PACK_IDX(b_id, grid)], loc_elem + cStart, elemMat, ADD_VALUES));
467: PetscCall(PetscLogEventEnd(ctx->events[6], 0, 0, 0, 0));
468: } else { // GPU like assembly for debugging
469: PetscInt fieldA, q, f, g, d, nr, nc, rows0[LANDAU_MAX_Q_FACE] = {0}, cols0[LANDAU_MAX_Q_FACE] = {0}, rows[LANDAU_MAX_Q_FACE], cols[LANDAU_MAX_Q_FACE];
470: PetscScalar vals[LANDAU_MAX_Q_FACE * LANDAU_MAX_Q_FACE] = {0}, row_scale[LANDAU_MAX_Q_FACE] = {0}, col_scale[LANDAU_MAX_Q_FACE] = {0};
471: LandauIdx *coo_elem_offsets = (LandauIdx *)ctx->SData_d.coo_elem_offsets, *coo_elem_fullNb = (LandauIdx *)ctx->SData_d.coo_elem_fullNb, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = (LandauIdx(*)[LANDAU_MAX_NQND + 1]) ctx->SData_d.coo_elem_point_offsets;
472: /* assemble - from the diagonal (I,I) in this format for DMPlexMatSetClosure */
473: for (fieldA = 0; fieldA < loc_Nf; fieldA++) {
474: LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][fieldA][0];
475: for (f = 0; f < Nb; f++) {
476: PetscInt idx = Idxs[f];
477: if (idx >= 0) {
478: nr = 1;
479: rows0[0] = idx;
480: row_scale[0] = 1.;
481: } else {
482: idx = -idx - 1;
483: for (q = 0, nr = 0; q < maps[grid].num_face; q++, nr++) {
484: if (maps[grid].c_maps[idx][q].gid < 0) break;
485: rows0[q] = maps[grid].c_maps[idx][q].gid;
486: row_scale[q] = maps[grid].c_maps[idx][q].scale;
487: }
488: }
489: for (g = 0; g < Nb; ++g) {
490: idx = Idxs[g];
491: if (idx >= 0) {
492: nc = 1;
493: cols0[0] = idx;
494: col_scale[0] = 1.;
495: } else {
496: idx = -idx - 1;
497: nc = maps[grid].num_face;
498: for (q = 0, nc = 0; q < maps[grid].num_face; q++, nc++) {
499: if (maps[grid].c_maps[idx][q].gid < 0) break;
500: cols0[q] = maps[grid].c_maps[idx][q].gid;
501: col_scale[q] = maps[grid].c_maps[idx][q].scale;
502: }
503: }
504: const PetscInt i = fieldA * Nb + f; /* Element matrix row */
505: const PetscInt j = fieldA * Nb + g; /* Element matrix column */
506: const PetscScalar Aij = elemMat[i * totDim + j];
507: if (coo_vals) { // mirror (i,j) in CreateStaticGPUData
508: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb;
509: const PetscInt idx0 = b_id * coo_elem_offsets[elem_offset[num_grids]] + coo_elem_offsets[glb_elem_idx] + fieldA * fullNb2 + fullNb * coo_elem_point_offsets[glb_elem_idx][f] + nr * coo_elem_point_offsets[glb_elem_idx][g];
510: for (PetscInt q = 0, idx2 = idx0; q < nr; q++) {
511: for (PetscInt d = 0; d < nc; d++, idx2++) coo_vals[idx2] = row_scale[q] * col_scale[d] * Aij;
512: }
513: } else {
514: for (q = 0; q < nr; q++) rows[q] = rows0[q] + moffset;
515: for (d = 0; d < nc; d++) cols[d] = cols0[d] + moffset;
516: for (q = 0; q < nr; q++) {
517: for (d = 0; d < nc; d++) vals[q * nc + d] = row_scale[q] * col_scale[d] * Aij;
518: }
519: PetscCall(MatSetValues(JacP, nr, rows, nc, cols, vals, ADD_VALUES));
520: }
521: }
522: }
523: }
524: }
525: if (loc_elem == -1) {
526: PetscCall(PetscPrintf(ctx->comm, "CPU Element matrix\n"));
527: for (PetscInt d = 0; d < totDim; ++d) {
528: for (PetscInt f = 0; f < totDim; ++f) PetscCall(PetscPrintf(ctx->comm, " %12.5e", (double)PetscRealPart(elemMat[d * totDim + f])));
529: PetscCall(PetscPrintf(ctx->comm, "\n"));
530: }
531: exit(12);
532: }
533: PetscCall(PetscFree(elemMat));
534: } /* grid */
535: } /* outer element & batch loop */
536: if (shift == 0.0) { // mass
537: PetscCall(PetscFree4(ff, dudx, dudy, dudz));
538: }
539: if (!container) { // 'CPU' assembly move nest matrix to global JacP
540: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP
541: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
542: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset); // b_id*b_N + ctx->mat_offset[grid];
543: PetscInt nloc, nzl, colbuf[1024], row;
544: const PetscInt *cols;
545: const PetscScalar *vals;
546: Mat B = subJ[LAND_PACK_IDX(b_id, grid)];
547: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
548: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
549: PetscCall(MatGetSize(B, &nloc, NULL));
550: for (PetscInt i = 0; i < nloc; i++) {
551: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
552: PetscCheck(nzl <= 1024, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT, nzl);
553: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = moffset + cols[j];
554: row = moffset + i;
555: PetscCall(MatSetValues(JacP, 1, &row, nzl, colbuf, vals, ADD_VALUES));
556: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
557: }
558: PetscCall(MatDestroy(&B));
559: }
560: }
561: }
562: if (coo_vals) {
563: PetscCall(MatSetValuesCOO(JacP, coo_vals, ADD_VALUES));
564: PetscCall(PetscFree(coo_vals));
565: }
566: } /* CPU version */
567: PetscCall(MatAssemblyBegin(JacP, MAT_FINAL_ASSEMBLY));
568: PetscCall(MatAssemblyEnd(JacP, MAT_FINAL_ASSEMBLY));
569: /* clean up */
570: PetscCall(PetscFree(cellClosure));
571: if (xdata) PetscCall(VecRestoreArrayReadAndMemType(a_X, &xdata));
572: PetscFunctionReturn(PETSC_SUCCESS);
573: }
575: /* create DMComposite of meshes for each species group */
576: static PetscErrorCode LandauDMCreateVMeshes(MPI_Comm comm_self, const PetscInt dim, const char prefix[], LandauCtx *ctx, DM pack)
577: {
578: PetscFunctionBegin;
579: /* p4est, quads */
580: /* Create plex mesh of Landau domain */
581: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
582: PetscReal par_radius = ctx->radius_par[grid], perp_radius = ctx->radius_perp[grid];
583: if (!ctx->sphere && !ctx->simplex) { // 2 or 3D (only 3D option)
584: PetscReal lo[] = {-perp_radius, -par_radius, -par_radius}, hi[] = {perp_radius, par_radius, par_radius};
585: DMBoundaryType periodicity[3] = {DM_BOUNDARY_NONE, dim == 2 ? DM_BOUNDARY_NONE : DM_BOUNDARY_NONE, DM_BOUNDARY_NONE};
586: if (dim == 2) lo[0] = 0;
587: else {
588: lo[1] = -perp_radius;
589: hi[1] = perp_radius; // 3D y is a perp
590: }
591: PetscCall(DMPlexCreateBoxMesh(comm_self, dim, PETSC_FALSE, ctx->cells0, lo, hi, periodicity, PETSC_TRUE, 0, PETSC_TRUE, &ctx->plex[grid])); // TODO: make composite and create dm[grid] here
592: PetscCall(DMLocalizeCoordinates(ctx->plex[grid])); /* needed for periodic */
593: if (dim == 3) PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cube"));
594: else PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "half-plane"));
595: } else if (dim == 2) {
596: size_t len;
597: PetscCall(PetscStrlen(ctx->filename, &len));
598: if (len) {
599: Vec coords;
600: PetscScalar *x;
601: PetscInt N;
602: char str[] = "-dm_landau_view_file_0";
603: str[21] += grid;
604: PetscCall(DMPlexCreateFromFile(comm_self, ctx->filename, "plexland.c", PETSC_TRUE, &ctx->plex[grid]));
605: PetscCall(DMPlexOrient(ctx->plex[grid]));
606: PetscCall(DMGetCoordinatesLocal(ctx->plex[grid], &coords));
607: PetscCall(VecGetSize(coords, &N));
608: PetscCall(VecGetArray(coords, &x));
609: /* scale by domain size */
610: for (PetscInt i = 0; i < N; i += 2) {
611: x[i + 0] *= ctx->radius_perp[grid];
612: x[i + 1] *= ctx->radius_par[grid];
613: }
614: PetscCall(VecRestoreArray(coords, &x));
615: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], ctx->filename));
616: PetscCall(PetscInfo(ctx->plex[grid], "%" PetscInt_FMT ") Read %s mesh file (%s)\n", grid, ctx->filename, str));
617: PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, str));
618: } else { // simplex forces a sphere
619: PetscInt numCells = ctx->simplex ? 12 : 6, cell_size = ctx->simplex ? 3 : 4, j;
620: const PetscInt numVerts = 11;
621: PetscInt cellsT[][4] = {
622: {0, 1, 6, 5 },
623: {1, 2, 7, 6 },
624: {2, 3, 8, 7 },
625: {3, 4, 9, 8 },
626: {5, 6, 7, 10},
627: {10, 7, 8, 9 }
628: };
629: PetscInt cellsS[][3] = {
630: {0, 1, 6 },
631: {1, 2, 6 },
632: {6, 2, 7 },
633: {7, 2, 8 },
634: {8, 2, 3 },
635: {8, 3, 4 },
636: {0, 6, 5 },
637: {5, 6, 7 },
638: {5, 7, 10},
639: {10, 7, 9 },
640: {9, 7, 8 },
641: {9, 8, 4 }
642: };
643: const PetscInt *pcell = (const PetscInt *)(ctx->simplex ? &cellsS[0][0] : &cellsT[0][0]);
644: PetscReal coords[11][2], *flatCoords = &coords[0][0];
645: PetscReal rad = ctx->radius[grid];
646: for (j = 0; j < 5; j++) { // outside edge
647: PetscReal z, r, theta = -PETSC_PI / 2 + (j % 5) * PETSC_PI / 4;
648: r = rad * PetscCosReal(theta);
649: coords[j][0] = r;
650: z = rad * PetscSinReal(theta);
651: coords[j][1] = z;
652: }
653: coords[j][0] = 0;
654: coords[j++][1] = -rad * ctx->sphere_inner_radius_90degree[grid];
655: coords[j][0] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
656: coords[j++][1] = -rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
657: coords[j][0] = rad * ctx->sphere_inner_radius_90degree[grid];
658: coords[j++][1] = 0;
659: coords[j][0] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
660: coords[j++][1] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
661: coords[j][0] = 0;
662: coords[j++][1] = rad * ctx->sphere_inner_radius_90degree[grid];
663: coords[j][0] = 0;
664: coords[j++][1] = 0;
665: PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 2, numCells, numVerts, cell_size, ctx->interpolate, pcell, 2, flatCoords, &ctx->plex[grid]));
666: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "semi-circle"));
667: PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make circle %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor"));
668: }
669: } else {
670: PetscCheck(dim == 3 && ctx->sphere && !ctx->simplex, ctx->comm, PETSC_ERR_ARG_WRONG, "not: dim == 3 && ctx->sphere && !ctx->simplex");
671: PetscReal rad = ctx->radius[grid], inner_rad = rad * ctx->sphere_inner_radius_90degree[grid], outer_rad = rad;
672: const PetscInt numCells = 7, cell_size = 8, numVerts = 16;
673: const PetscInt cells[][8] = {
674: {0, 3, 2, 1, 4, 5, 6, 7 },
675: {0, 4, 5, 1, 8, 9, 13, 12},
676: {1, 5, 6, 2, 9, 10, 14, 13},
677: {2, 6, 7, 3, 10, 11, 15, 14},
678: {0, 3, 7, 4, 8, 12, 15, 11},
679: {0, 1, 2, 3, 8, 11, 10, 9 },
680: {4, 7, 6, 5, 12, 13, 14, 15}
681: };
682: PetscReal coords[16 /* numVerts */][3];
683: for (PetscInt j = 0; j < 4; j++) { // inner edge, low
684: coords[j][0] = inner_rad * (j == 0 || j == 3 ? 1 : -1);
685: coords[j][1] = inner_rad * (j / 2 < 1 ? 1 : -1);
686: coords[j][2] = inner_rad * -1;
687: }
688: for (PetscInt j = 0, jj = 4; j < 4; j++, jj++) { // inner edge, hi
689: coords[jj][0] = inner_rad * (j == 0 || j == 3 ? 1 : -1);
690: coords[jj][1] = inner_rad * (j / 2 < 1 ? 1 : -1);
691: coords[jj][2] = inner_rad * 1;
692: }
693: for (PetscInt j = 0, jj = 8; j < 4; j++, jj++) { // outer edge, low
694: coords[jj][0] = outer_rad * (j == 0 || j == 3 ? 1 : -1);
695: coords[jj][1] = outer_rad * (j / 2 < 1 ? 1 : -1);
696: coords[jj][2] = outer_rad * -1;
697: }
698: for (PetscInt j = 0, jj = 12; j < 4; j++, jj++) { // outer edge, hi
699: coords[jj][0] = outer_rad * (j == 0 || j == 3 ? 1 : -1);
700: coords[jj][1] = outer_rad * (j / 2 < 1 ? 1 : -1);
701: coords[jj][2] = outer_rad * 1;
702: }
703: PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 3, numCells, numVerts, cell_size, ctx->interpolate, (const PetscInt *)cells, 3, (const PetscReal *)coords, &ctx->plex[grid]));
704: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cubed sphere"));
705: PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make cubed sphere %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor"));
706: }
707: PetscCall(DMSetOptionsPrefix(ctx->plex[grid], prefix));
708: PetscCall(DMSetFromOptions(ctx->plex[grid]));
709: } // grid loop
710: PetscCall(DMSetOptionsPrefix(pack, prefix));
711: { /* convert to p4est (or whatever), wait for discretization to create pack */
712: char convType[256];
713: PetscBool flg;
715: PetscOptionsBegin(ctx->comm, prefix, "Mesh conversion options", "DMPLEX");
716: PetscCall(PetscOptionsFList("-dm_landau_type", "Convert DMPlex to another format (p4est)", "plexland.c", DMList, DMPLEX, convType, 256, &flg));
717: PetscOptionsEnd();
718: if (flg) {
719: ctx->use_p4est = PETSC_TRUE; /* flag for Forest */
720: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
721: DM dmforest;
722: PetscBool isForest;
724: PetscCall(DMConvert(ctx->plex[grid], convType, &dmforest));
725: PetscCheck(dmforest, ctx->comm, PETSC_ERR_PLIB, "Convert failed?");
726: PetscCall(DMSetOptionsPrefix(dmforest, prefix));
727: PetscCall(DMIsForest(dmforest, &isForest));
728: PetscCheck(isForest, ctx->comm, PETSC_ERR_PLIB, "Converted to non Forest?");
729: PetscCall(DMDestroy(&ctx->plex[grid]));
730: ctx->plex[grid] = dmforest; // Forest for adaptivity
731: }
732: } else ctx->use_p4est = PETSC_FALSE; /* flag for Forest */
733: }
734: PetscCall(DMSetDimension(pack, dim));
735: PetscCall(PetscObjectSetName((PetscObject)pack, "Mesh"));
736: PetscCall(DMSetApplicationContext(pack, ctx));
737: PetscFunctionReturn(PETSC_SUCCESS);
738: }
740: static PetscErrorCode SetupDS(DM pack, PetscInt dim, PetscInt grid, const char prefix[], LandauCtx *ctx)
741: {
742: PetscInt ii, i0;
743: char buf[256];
744: PetscSection section;
746: PetscFunctionBegin;
747: for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
748: if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "e"));
749: else PetscCall(PetscSNPrintf(buf, sizeof(buf), "i%" PetscInt_FMT, ii));
750: /* Setup Discretization - FEM */
751: PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, prefix, PETSC_DECIDE, &ctx->fe[ii]));
752: PetscCall(PetscObjectSetName((PetscObject)ctx->fe[ii], buf));
753: PetscCall(DMSetField(ctx->plex[grid], i0, NULL, (PetscObject)ctx->fe[ii]));
754: }
755: PetscCall(DMCreateDS(ctx->plex[grid]));
756: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion));
757: for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
758: if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "se"));
759: else PetscCall(PetscSNPrintf(buf, sizeof(buf), "si%" PetscInt_FMT, ii));
760: PetscCall(PetscSectionSetComponentName(section, i0, 0, buf));
761: }
762: PetscFunctionReturn(PETSC_SUCCESS);
763: }
765: /* Define a Maxwellian function for testing out the operator. */
767: /* Using cartesian velocity space coordinates, the particle */
768: /* density, [1/m^3], is defined according to */
770: /* $$ n=\int_{R^3} dv^3 \left(\frac{m}{2\pi T}\right)^{3/2}\exp [- mv^2/(2T)] $$ */
772: /* Using some constant, c, we normalize the velocity vector into a */
773: /* dimensionless variable according to v=c*x. Thus the density, $n$, becomes */
775: /* $$ n=\int_{R^3} dx^3 \left(\frac{mc^2}{2\pi T}\right)^{3/2}\exp [- mc^2/(2T)*x^2] $$ */
777: /* Defining $\theta=2T/mc^2$, we thus find that the probability density */
778: /* for finding the particle within the interval in a box dx^3 around x is */
780: /* f(x;\theta)=\left(\frac{1}{\pi\theta}\right)^{3/2} \exp [ -x^2/\theta ] */
782: typedef struct {
783: PetscReal v_0;
784: PetscReal kT_m;
785: PetscReal n;
786: PetscReal shift;
787: } MaxwellianCtx;
789: static PetscErrorCode maxwellian(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
790: {
791: MaxwellianCtx *mctx = (MaxwellianCtx *)actx;
792: PetscReal v2 = 0, theta = 2 * mctx->kT_m / (mctx->v_0 * mctx->v_0), shift; /* theta = 2kT/mc^2 */
794: PetscFunctionBegin;
795: /* compute the exponents, v^2 */
796: for (PetscInt i = 0; i < dim; ++i) v2 += x[i] * x[i];
797: /* evaluate the Maxwellian */
798: if (mctx->shift < 0) shift = -mctx->shift;
799: else {
800: u[0] = mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta));
801: shift = mctx->shift;
802: }
803: if (shift != 0.) {
804: v2 = 0;
805: for (PetscInt i = 0; i < dim - 1; ++i) v2 += x[i] * x[i];
806: v2 += (x[dim - 1] - shift) * (x[dim - 1] - shift);
807: /* evaluate the shifted Maxwellian */
808: u[0] += mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta));
809: }
810: PetscFunctionReturn(PETSC_SUCCESS);
811: }
813: /*@
814: DMPlexLandauAddMaxwellians - Add a Maxwellian distribution to a state
816: Collective
818: Input Parameters:
819: + dm - The mesh (local)
820: . time - Current time
821: . temps - Temperatures of each species (global)
822: . ns - Number density of each species (global)
823: . grid - index into current grid - just used for offset into `temp` and `ns`
824: . b_id - batch index
825: . n_batch - number of batches
826: - actx - Landau context
828: Output Parameter:
829: . X - The state (local to this grid)
831: Level: beginner
833: .seealso: `DMPlexLandauCreateVelocitySpace()`
834: @*/
835: PetscErrorCode DMPlexLandauAddMaxwellians(DM dm, Vec X, PetscReal time, PetscReal temps[], PetscReal ns[], PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx)
836: {
837: LandauCtx *ctx = (LandauCtx *)actx;
838: PetscErrorCode (*initu[LANDAU_MAX_SPECIES])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *);
839: PetscInt dim;
840: MaxwellianCtx *mctxs[LANDAU_MAX_SPECIES], data[LANDAU_MAX_SPECIES];
842: PetscFunctionBegin;
843: PetscCall(DMGetDimension(dm, &dim));
844: if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx));
845: for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
846: mctxs[i0] = &data[i0];
847: data[i0].v_0 = ctx->v_0; // v_0 same for all grids
848: data[i0].kT_m = ctx->k * temps[ii] / ctx->masses[ii]; /* kT/m */
849: data[i0].n = ns[ii];
850: initu[i0] = maxwellian;
851: data[i0].shift = 0;
852: }
853: data[0].shift = ctx->electronShift;
854: /* need to make ADD_ALL_VALUES work - TODO */
855: PetscCall(DMProjectFunction(dm, time, initu, (void **)mctxs, INSERT_ALL_VALUES, X));
856: PetscFunctionReturn(PETSC_SUCCESS);
857: }
859: /* Adds Maxwellians to X for the given grid and batch using temperatures and densities from actx */
860: static PetscErrorCode LandauSetInitialCondition(DM dm, Vec X, PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx)
861: {
862: LandauCtx *ctx = (LandauCtx *)actx;
864: PetscFunctionBegin;
865: if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx));
866: PetscCall(VecZeroEntries(X));
867: PetscCall(DMPlexLandauAddMaxwellians(dm, X, 0.0, ctx->thermal_temps, ctx->n, grid, b_id, n_batch, ctx));
868: PetscFunctionReturn(PETSC_SUCCESS);
869: }
871: // adapt a level once. Forest in/out
872: #if PetscDefined(USE_INFO)
873: static const char *s_refine_names[] = {"RE", "Z1", "Origin", "Z2", "Uniform"};
874: #endif
875: static PetscErrorCode adaptToleranceFEM(PetscFE fem, Vec sol, PetscInt type, PetscInt grid, LandauCtx *ctx, DM *newForest)
876: {
877: DM forest, plex, adaptedDM = NULL;
878: PetscDS prob;
879: PetscBool isForest;
880: PetscQuadrature quad;
881: PetscInt Nq, Nb, *Nb2, cStart, cEnd, c, dim, qj, k;
882: DMLabel adaptLabel = NULL;
884: PetscFunctionBegin;
885: forest = ctx->plex[grid];
886: PetscCall(DMCreateDS(forest));
887: PetscCall(DMGetDS(forest, &prob));
888: PetscCall(DMGetDimension(forest, &dim));
889: PetscCall(DMIsForest(forest, &isForest));
890: PetscCheck(isForest, ctx->comm, PETSC_ERR_ARG_WRONG, "! Forest");
891: PetscCall(DMConvert(forest, DMPLEX, &plex));
892: PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd));
893: PetscCall(DMLabelCreate(PETSC_COMM_SELF, "adapt", &adaptLabel));
894: PetscCall(PetscFEGetQuadrature(fem, &quad));
895: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL));
896: PetscCheck(Nq <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nq, LANDAU_MAX_NQND);
897: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
898: PetscCall(PetscDSGetDimensions(prob, &Nb2));
899: PetscCheck(Nb2[0] == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, " Nb = %" PetscInt_FMT " != Nb (%" PetscInt_FMT ")", Nb, Nb2[0]);
900: PetscCheck(Nb <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nb = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nb, LANDAU_MAX_NQND);
901: PetscCall(PetscInfo(sol, "%" PetscInt_FMT ") Refine phase: %s\n", grid, s_refine_names[type]));
902: if (type == 4) {
903: for (c = cStart; c < cEnd; c++) PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE));
904: } else if (type == 2) {
905: PetscInt rCellIdx[8], nr = 0, nrmax = (dim == 3) ? 8 : 2;
906: PetscReal minRad = PETSC_INFINITY, r;
907: for (c = cStart; c < cEnd; c++) {
908: PetscReal tt, v0[LANDAU_MAX_NQND * 3], J[LANDAU_MAX_NQND * 9], invJ[LANDAU_MAX_NQND * 9], detJ[LANDAU_MAX_NQND];
909: PetscCall(DMPlexComputeCellGeometryFEM(plex, c, quad, v0, J, invJ, detJ));
910: (void)J;
911: (void)invJ;
912: for (qj = 0; qj < Nq; ++qj) {
913: tt = PetscSqr(v0[dim * qj + 0]) + PetscSqr(v0[dim * qj + 1]) + PetscSqr((dim == 3) ? v0[dim * qj + 2] : 0);
914: r = PetscSqrtReal(tt);
915: if (r < minRad - PETSC_SQRT_MACHINE_EPSILON * 10.) {
916: minRad = r;
917: nr = 0;
918: rCellIdx[nr++] = c;
919: PetscCall(PetscInfo(sol, "\t\t%" PetscInt_FMT ") Found first inner r=%e, cell %" PetscInt_FMT ", qp %" PetscInt_FMT "/%" PetscInt_FMT "\n", grid, (double)r, c, qj + 1, Nq));
920: } else if ((r - minRad) < PETSC_SQRT_MACHINE_EPSILON * 100. && nr < nrmax) {
921: for (k = 0; k < nr; k++)
922: if (c == rCellIdx[k]) break;
923: if (k == nr) {
924: rCellIdx[nr++] = c;
925: PetscCall(PetscInfo(sol, "\t\t\t%" PetscInt_FMT ") Found another inner r=%e, cell %" PetscInt_FMT ", qp %" PetscInt_FMT "/%" PetscInt_FMT ", d=%e\n", grid, (double)r, c, qj + 1, Nq, (double)(r - minRad)));
926: }
927: }
928: }
929: }
930: for (k = 0; k < nr; k++) PetscCall(DMLabelSetValue(adaptLabel, rCellIdx[k], DM_ADAPT_REFINE));
931: PetscCall(PetscInfo(sol, "\t\t\t%" PetscInt_FMT ") Refined %" PetscInt_FMT " origin cells %" PetscInt_FMT ",%" PetscInt_FMT " r=%g\n", grid, nr, rCellIdx[0], rCellIdx[1], (double)minRad));
932: } else if (type == 0 || type == 1 || type == 3) { /* refine along r=0 axis */
933: PetscScalar *coef = NULL;
934: Vec coords;
935: PetscInt csize, Nv, d, nz, nrefined = 0;
936: DM cdm;
937: PetscSection cs;
938: PetscCall(DMGetCoordinatesLocal(forest, &coords));
939: PetscCall(DMGetCoordinateDM(forest, &cdm));
940: PetscCall(DMGetLocalSection(cdm, &cs));
941: for (c = cStart; c < cEnd; c++) {
942: PetscInt doit = 0, outside = 0;
943: PetscCall(DMPlexVecGetClosure(cdm, cs, coords, c, &csize, &coef));
944: Nv = csize / dim;
945: for (nz = d = 0; d < Nv; d++) {
946: PetscReal z = PetscRealPart(coef[d * dim + (dim - 1)]), x = PetscSqr(PetscRealPart(coef[d * dim + 0])) + ((dim == 3) ? PetscSqr(PetscRealPart(coef[d * dim + 1])) : 0);
947: x = PetscSqrtReal(x);
948: if (type == 0) {
949: if (ctx->re_radius > PETSC_SQRT_MACHINE_EPSILON && (z < -PETSC_MACHINE_EPSILON * 10. || z > ctx->re_radius + PETSC_MACHINE_EPSILON * 10.)) outside++; /* first pass don't refine bottom */
950: } else if (type == 1 && (z > ctx->vperp0_radius1 || z < -ctx->vperp0_radius1)) {
951: outside++; /* don't refine outside electron refine radius */
952: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type]));
953: } else if (type == 3 && (z > ctx->vperp0_radius2 || z < -ctx->vperp0_radius2)) {
954: outside++; /* refine r=0 cells on refinement front */
955: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type]));
956: }
957: if (x < PETSC_MACHINE_EPSILON * 10. && (type != 0 || ctx->re_radius > PETSC_SQRT_MACHINE_EPSILON)) nz++;
958: }
959: PetscCall(DMPlexVecRestoreClosure(cdm, cs, coords, c, &csize, &coef));
960: if (doit || (outside < Nv && nz)) {
961: PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE));
962: nrefined++;
963: }
964: }
965: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") Refined %" PetscInt_FMT " cells\n", grid, nrefined));
966: }
967: PetscCall(DMDestroy(&plex));
968: PetscCall(DMAdaptLabel(forest, adaptLabel, &adaptedDM));
969: PetscCall(DMLabelDestroy(&adaptLabel));
970: *newForest = adaptedDM;
971: if (adaptedDM) {
972: if (isForest) PetscCall(DMForestSetAdaptivityForest(adaptedDM, NULL)); // ????
973: PetscCall(DMConvert(adaptedDM, DMPLEX, &plex));
974: PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd));
975: PetscCall(PetscInfo(sol, "\t\t\t\t%" PetscInt_FMT ") %" PetscInt_FMT " cells, %" PetscInt_FMT " total quadrature points\n", grid, cEnd - cStart, Nq * (cEnd - cStart)));
976: PetscCall(DMDestroy(&plex));
977: } else *newForest = NULL;
978: PetscFunctionReturn(PETSC_SUCCESS);
979: }
981: // forest goes in (ctx->plex[grid]), plex comes out
982: static PetscErrorCode adapt(PetscInt grid, LandauCtx *ctx, Vec *uu)
983: {
984: PetscFunctionBegin;
985: PetscInt type, limits[5] = {(grid == 0) ? ctx->numRERefine : 0, (grid == 0) ? ctx->nZRefine1 : 0, ctx->numAMRRefine[grid], (grid == 0) ? ctx->nZRefine2 : 0, ctx->postAMRRefine[grid]};
986: for (type = 0; type < 5; type++) {
987: for (PetscInt adaptIter = 0; adaptIter < limits[type]; adaptIter++) {
988: DM newForest = NULL;
989: PetscCall(adaptToleranceFEM(ctx->fe[0], *uu, type, grid, ctx, &newForest));
990: if (newForest) {
991: PetscCall(DMDestroy(&ctx->plex[grid]));
992: PetscCall(VecDestroy(uu));
993: PetscCall(DMCreateGlobalVector(newForest, uu));
994: PetscCall(LandauSetInitialCondition(newForest, *uu, grid, 0, 1, ctx));
995: ctx->plex[grid] = newForest;
996: } else {
997: PetscCall(PetscInfo(*uu, "No refinement\n"));
998: }
999: }
1000: }
1001: PetscCall(PetscObjectSetName((PetscObject)*uu, "uAMR"));
1002: PetscFunctionReturn(PETSC_SUCCESS);
1003: }
1005: // make log(Lambdas) from NRL Plasma formulary
1006: static PetscErrorCode makeLambdas(LandauCtx *ctx)
1007: {
1008: PetscFunctionBegin;
1009: for (PetscInt gridi = 0; gridi < ctx->num_grids; gridi++) {
1010: PetscInt iii = ctx->species_offset[gridi];
1011: PetscReal Ti_ev = (ctx->thermal_temps[iii] / 1.1604525e7) * 1000; // convert (back) to eV
1012: PetscReal ni = ctx->n[iii] * ctx->n_0;
1013: for (PetscInt gridj = gridi; gridj < ctx->num_grids; gridj++) {
1014: PetscInt jjj = ctx->species_offset[gridj];
1015: PetscReal Zj = ctx->charges[jjj] / 1.6022e-19;
1016: if (gridi == 0) {
1017: if (gridj == 0) { // lam_ee
1018: ctx->lambdas[gridi][gridj] = 23.5 - PetscLogReal(PetscSqrtReal(ni) * PetscPowReal(Ti_ev, -1.25)) - PetscSqrtReal(1e-5 + PetscSqr(PetscLogReal(Ti_ev) - 2) / 16);
1019: } else { // lam_ei == lam_ie
1020: if (10 * Zj * Zj > Ti_ev) {
1021: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 23 - PetscLogReal(PetscSqrtReal(ni) * Zj * PetscPowReal(Ti_ev, -1.5));
1022: } else {
1023: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 24 - PetscLogReal(PetscSqrtReal(ni) / Ti_ev);
1024: }
1025: }
1026: } else { // lam_ii'
1027: PetscReal mui = ctx->masses[iii] / 1.6720e-27, Zi = ctx->charges[iii] / 1.6022e-19;
1028: PetscReal Tj_ev = (ctx->thermal_temps[jjj] / 1.1604525e7) * 1000; // convert (back) to eV
1029: PetscReal muj = ctx->masses[jjj] / 1.6720e-27;
1030: PetscReal nj = ctx->n[jjj] * ctx->n_0;
1031: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 23 - PetscLogReal(Zi * Zj * (mui + muj) / (mui * Tj_ev + muj * Ti_ev) * PetscSqrtReal(ni * Zi * Zi / Ti_ev + nj * Zj * Zj / Tj_ev));
1032: }
1033: }
1034: }
1035: //PetscReal v0 = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */
1036: PetscFunctionReturn(PETSC_SUCCESS);
1037: }
1039: static PetscErrorCode ProcessOptions(LandauCtx *ctx, const char prefix[])
1040: {
1041: PetscBool flg, fileflg;
1042: PetscInt ii, nt, nm, nc, num_species_grid[LANDAU_MAX_GRIDS], non_dim_grid;
1043: PetscReal lnLam = 10;
1044: DM dummy;
1046: PetscFunctionBegin;
1047: PetscCall(DMCreate(ctx->comm, &dummy));
1048: /* get options - initialize context */
1049: ctx->verbose = 1; // should be 0 for silent compliance
1050: ctx->batch_sz = 1;
1051: ctx->batch_view_idx = 0;
1052: ctx->interpolate = PETSC_TRUE;
1053: ctx->gpu_assembly = PETSC_TRUE;
1054: ctx->norm_state = 0;
1055: ctx->electronShift = 0;
1056: ctx->M = NULL;
1057: ctx->J = NULL;
1058: /* geometry and grids */
1059: ctx->sphere = PETSC_FALSE;
1060: ctx->map_sphere = PETSC_TRUE;
1061: ctx->use_p4est = PETSC_FALSE;
1062: ctx->simplex = PETSC_FALSE;
1063: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1064: ctx->radius[grid] = 5.; /* thermal radius (velocity) */
1065: ctx->radius_perp[grid] = 5.; /* thermal radius (velocity) */
1066: ctx->radius_par[grid] = 5.; /* thermal radius (velocity) */
1067: ctx->numAMRRefine[grid] = 0;
1068: ctx->postAMRRefine[grid] = 0;
1069: ctx->species_offset[grid + 1] = 1; // one species default
1070: num_species_grid[grid] = 0;
1071: ctx->plex[grid] = NULL; /* cache as expensive to Convert */
1072: }
1073: ctx->species_offset[0] = 0;
1074: ctx->re_radius = 0.;
1075: ctx->vperp0_radius1 = 0;
1076: ctx->vperp0_radius2 = 0;
1077: ctx->nZRefine1 = 0;
1078: ctx->nZRefine2 = 0;
1079: ctx->numRERefine = 0;
1080: num_species_grid[0] = 1; // one species default
1081: /* species - [0] electrons, [1] one ion species eg, duetarium, [2] heavy impurity ion, ... */
1082: ctx->charges[0] = -1; /* electron charge (MKS) */
1083: ctx->masses[0] = 1 / 1835.469965278441013; /* temporary value in proton mass */
1084: ctx->n[0] = 1;
1085: ctx->v_0 = 1; /* thermal velocity, we could start with a scale != 1 */
1086: ctx->thermal_temps[0] = 1;
1087: /* constants, etc. */
1088: ctx->epsilon0 = 8.8542e-12; /* permittivity of free space (MKS) F/m */
1089: ctx->k = 1.38064852e-23; /* Boltzmann constant (MKS) J/K */
1090: ctx->n_0 = 1.e20; /* typical plasma n, but could set it to 1 */
1091: ctx->Ez = 0;
1092: for (PetscInt grid = 0; grid < LANDAU_NUM_TIMERS; grid++) ctx->times[grid] = 0;
1093: for (PetscInt ii = 0; ii < LANDAU_DIM; ii++) ctx->cells0[ii] = 2;
1094: if (LANDAU_DIM == 2) ctx->cells0[0] = 1;
1095: ctx->use_matrix_mass = PETSC_FALSE;
1096: ctx->use_relativistic_corrections = PETSC_FALSE;
1097: ctx->use_energy_tensor_trick = PETSC_FALSE; /* Use Eero's trick for energy conservation v --> grad(v^2/2) */
1098: ctx->SData_d.w = NULL;
1099: ctx->SData_d.x = NULL;
1100: ctx->SData_d.y = NULL;
1101: ctx->SData_d.z = NULL;
1102: ctx->SData_d.invJ = NULL;
1103: ctx->jacobian_field_major_order = PETSC_FALSE;
1104: ctx->SData_d.coo_elem_offsets = NULL;
1105: ctx->SData_d.coo_elem_point_offsets = NULL;
1106: ctx->SData_d.coo_elem_fullNb = NULL;
1107: ctx->SData_d.coo_size = 0;
1108: PetscOptionsBegin(ctx->comm, prefix, "Options for Fokker-Plank-Landau collision operator", "none");
1109: #if PetscDefined(HAVE_KOKKOS)
1110: ctx->deviceType = LANDAU_KOKKOS;
1111: PetscCall(PetscStrncpy(ctx->filename, "kokkos", sizeof(ctx->filename)));
1112: #else
1113: ctx->deviceType = LANDAU_CPU;
1114: PetscCall(PetscStrncpy(ctx->filename, "cpu", sizeof(ctx->filename)));
1115: #endif
1116: PetscCall(PetscOptionsString("-dm_landau_device_type", "Use kernels on 'cpu' 'kokkos'", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), NULL));
1117: PetscCall(PetscStrcmp("cpu", ctx->filename, &flg));
1118: if (flg) {
1119: ctx->deviceType = LANDAU_CPU;
1120: } else {
1121: PetscCall(PetscStrcmp("kokkos", ctx->filename, &flg));
1122: PetscCheck(flg, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_device_type %s", ctx->filename);
1123: ctx->deviceType = LANDAU_KOKKOS;
1124: }
1125: ctx->filename[0] = '\0';
1126: PetscCall(PetscOptionsString("-dm_landau_filename", "file to read mesh from", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), &fileflg));
1127: PetscCall(PetscOptionsReal("-dm_landau_electron_shift", "Shift in thermal velocity of electrons", "none", ctx->electronShift, &ctx->electronShift, NULL));
1128: PetscCall(PetscOptionsInt("-dm_landau_verbose", "Level of verbosity output", "plexland.c", ctx->verbose, &ctx->verbose, NULL));
1129: PetscCall(PetscOptionsInt("-dm_landau_batch_size", "Number of 'vertices' to batch", "ex2.c", ctx->batch_sz, &ctx->batch_sz, NULL));
1130: PetscCheck(LANDAU_MAX_BATCH_SZ >= ctx->batch_sz, ctx->comm, PETSC_ERR_ARG_WRONG, "LANDAU_MAX_BATCH_SZ %d < ctx->batch_sz %" PetscInt_FMT, LANDAU_MAX_BATCH_SZ, ctx->batch_sz);
1131: PetscCall(PetscOptionsInt("-dm_landau_batch_view_idx", "Index of batch for diagnostics like plotting", "ex2.c", ctx->batch_view_idx, &ctx->batch_view_idx, NULL));
1132: PetscCheck(ctx->batch_view_idx < ctx->batch_sz, ctx->comm, PETSC_ERR_ARG_WRONG, "-ctx->batch_view_idx %" PetscInt_FMT " > ctx->batch_sz %" PetscInt_FMT, ctx->batch_view_idx, ctx->batch_sz);
1133: PetscCall(PetscOptionsReal("-dm_landau_Ez", "Initial parallel electric field in unites of Conner-Hastie critical field", "plexland.c", ctx->Ez, &ctx->Ez, NULL));
1134: PetscCall(PetscOptionsReal("-dm_landau_n_0", "Normalization constant for number density", "plexland.c", ctx->n_0, &ctx->n_0, NULL));
1135: PetscCall(PetscOptionsBool("-dm_landau_use_mataxpy_mass", "Use fast but slightly fragile MATAXPY to add mass term", "plexland.c", ctx->use_matrix_mass, &ctx->use_matrix_mass, NULL));
1136: PetscCall(PetscOptionsBool("-dm_landau_use_relativistic_corrections", "Use relativistic corrections", "plexland.c", ctx->use_relativistic_corrections, &ctx->use_relativistic_corrections, NULL));
1137: PetscCall(PetscOptionsBool("-dm_landau_simplex", "Use simplex elements", "plexland.c", ctx->simplex, &ctx->simplex, NULL));
1138: PetscCall(PetscOptionsBool("-dm_landau_sphere", "use sphere/semi-circle domain instead of rectangle", "plexland.c", ctx->sphere, &ctx->sphere, NULL));
1139: PetscCall(PetscOptionsBool("-dm_landau_map_sphere", "Map to sphere/semi-circle domain instead of rectangle", "plexland.c", ctx->map_sphere, &ctx->map_sphere, NULL));
1140: if (LANDAU_DIM == 2 && ctx->use_relativistic_corrections) ctx->use_relativistic_corrections = PETSC_FALSE; // should warn
1141: PetscCall(PetscOptionsBool("-dm_landau_use_energy_tensor_trick", "Use Eero's trick of using grad(v^2/2) instead of v as args to Landau tensor to conserve energy with relativistic corrections and Q1 elements", "plexland.c", ctx->use_energy_tensor_trick,
1142: &ctx->use_energy_tensor_trick, NULL));
1144: /* get num species with temperature, set defaults */
1145: for (ii = 1; ii < LANDAU_MAX_SPECIES; ii++) {
1146: ctx->thermal_temps[ii] = 1;
1147: ctx->charges[ii] = 1;
1148: ctx->masses[ii] = 1;
1149: ctx->n[ii] = 1;
1150: }
1151: nt = LANDAU_MAX_SPECIES;
1152: PetscCall(PetscOptionsRealArray("-dm_landau_thermal_temps", "Temperature of each species [e,i_0,i_1,...] in keV (must be set to set number of species)", "plexland.c", ctx->thermal_temps, &nt, &flg));
1153: PetscCheck(flg, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_thermal_temps ,t1,t2,.. must be provided to set the number of species");
1154: PetscCall(PetscInfo(dummy, "num_species set to number of thermal temps provided (%" PetscInt_FMT ")\n", nt));
1155: ctx->num_species = nt;
1156: for (ii = 0; ii < ctx->num_species; ii++) ctx->thermal_temps[ii] *= 1.1604525e7; /* convert to Kelvin */
1157: nm = LANDAU_MAX_SPECIES - 1;
1158: PetscCall(PetscOptionsRealArray("-dm_landau_ion_masses", "Mass of each species in units of proton mass [i_0=2,i_1=40...]", "plexland.c", &ctx->masses[1], &nm, &flg));
1159: PetscCheck(!flg || nm == ctx->num_species - 1, ctx->comm, PETSC_ERR_ARG_WRONG, "num ion masses %" PetscInt_FMT " != num species %" PetscInt_FMT, nm, ctx->num_species - 1);
1160: nm = LANDAU_MAX_SPECIES;
1161: PetscCall(PetscOptionsRealArray("-dm_landau_n", "Number density of each species = n_s * n_0", "plexland.c", ctx->n, &nm, &flg));
1162: PetscCheck(!flg || nm == ctx->num_species, ctx->comm, PETSC_ERR_ARG_WRONG, "wrong num n: %" PetscInt_FMT " != num species %" PetscInt_FMT, nm, ctx->num_species);
1163: for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] *= 1.6720e-27; /* scale by proton mass kg */
1164: ctx->masses[0] = 9.10938356e-31; /* electron mass kg (should be about right already) */
1165: nc = LANDAU_MAX_SPECIES - 1;
1166: PetscCall(PetscOptionsRealArray("-dm_landau_ion_charges", "Charge of each species in units of proton charge [i_0=2,i_1=18,...]", "plexland.c", &ctx->charges[1], &nc, &flg));
1167: if (flg) PetscCheck(nc == ctx->num_species - 1, ctx->comm, PETSC_ERR_ARG_WRONG, "num charges %" PetscInt_FMT " != num species %" PetscInt_FMT, nc, ctx->num_species - 1);
1168: for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->charges[ii] *= 1.6022e-19; /* electron/proton charge (MKS) */
1169: /* geometry and grids */
1170: nt = LANDAU_MAX_GRIDS;
1171: PetscCall(PetscOptionsIntArray("-dm_landau_num_species_grid", "Number of species on each grid: [ 1, ....] or [S, 0 ....] for single grid", "plexland.c", num_species_grid, &nt, &flg));
1172: if (flg) {
1173: ctx->num_grids = nt;
1174: for (ii = nt = 0; ii < ctx->num_grids; ii++) nt += num_species_grid[ii];
1175: PetscCheck(ctx->num_species == nt, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_num_species_grid: sum %" PetscInt_FMT " != num_species = %" PetscInt_FMT ". %" PetscInt_FMT " grids (check that number of grids <= LANDAU_MAX_GRIDS = %d)", nt, ctx->num_species,
1176: ctx->num_grids, LANDAU_MAX_GRIDS);
1177: } else {
1178: if (ctx->num_species > LANDAU_MAX_GRIDS) {
1179: num_species_grid[0] = 1;
1180: num_species_grid[1] = ctx->num_species - 1;
1181: ctx->num_grids = 2;
1182: } else {
1183: ctx->num_grids = ctx->num_species;
1184: for (ii = 0; ii < ctx->num_grids; ii++) num_species_grid[ii] = 1;
1185: }
1186: }
1187: for (ctx->species_offset[0] = ii = 0; ii < ctx->num_grids; ii++) ctx->species_offset[ii + 1] = ctx->species_offset[ii] + num_species_grid[ii];
1188: PetscCheck(ctx->species_offset[ctx->num_grids] == ctx->num_species, ctx->comm, PETSC_ERR_ARG_WRONG, "ctx->species_offset[ctx->num_grids] %" PetscInt_FMT " != ctx->num_species = %" PetscInt_FMT " ???????????", ctx->species_offset[ctx->num_grids],
1189: ctx->num_species);
1190: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1191: PetscInt iii = ctx->species_offset[grid]; // normalize with first (arbitrary) species on grid
1192: ctx->thermal_speed[grid] = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */
1193: }
1194: // get lambdas here because we need them for t_0 etc
1195: PetscCall(PetscOptionsReal("-dm_landau_ln_lambda", "Universal cross section parameter. Default uses NRL formulas", "plexland.c", lnLam, &lnLam, &flg));
1196: if (flg) {
1197: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1198: for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) ctx->lambdas[gridj][grid] = lnLam; /* cross section ratio large - small angle collisions */
1199: }
1200: } else {
1201: PetscCall(makeLambdas(ctx));
1202: }
1203: non_dim_grid = 0;
1204: PetscCall(PetscOptionsInt("-dm_landau_normalization_grid", "Index of grid to use for setting v_0, m_0, t_0. (Not recommended)", "plexland.c", non_dim_grid, &non_dim_grid, &flg));
1205: if (non_dim_grid != 0) PetscCall(PetscInfo(dummy, "Normalization grid set to %" PetscInt_FMT ", but non-default not well verified\n", non_dim_grid));
1206: PetscCheck(non_dim_grid >= 0 && non_dim_grid < ctx->num_species, ctx->comm, PETSC_ERR_ARG_WRONG, "Normalization grid wrong: %" PetscInt_FMT, non_dim_grid);
1207: ctx->v_0 = ctx->thermal_speed[non_dim_grid]; /* arbitrary units for non dimensionalization: global mean velocity in 1D of electrons */
1208: ctx->m_0 = ctx->masses[non_dim_grid]; /* arbitrary reference mass, electrons */
1209: ctx->t_0 = 8 * PETSC_PI * PetscSqr(ctx->epsilon0 * ctx->m_0 / PetscSqr(ctx->charges[non_dim_grid])) / ctx->lambdas[non_dim_grid][non_dim_grid] / ctx->n_0 * PetscPowReal(ctx->v_0, 3); /* note, this t_0 makes nu[non_dim_grid,non_dim_grid]=1 */
1210: /* domain */
1211: nt = LANDAU_MAX_GRIDS;
1212: PetscCall(PetscOptionsRealArray("-dm_landau_domain_radius", "Phase space size in units of thermal velocity of grid", "plexland.c", ctx->radius, &nt, &flg));
1213: if (flg) {
1214: PetscCheck(nt >= ctx->num_grids, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_domain_radius: given %" PetscInt_FMT " radius != number grids %" PetscInt_FMT, nt, ctx->num_grids);
1215: while (nt--) ctx->radius_par[nt] = ctx->radius_perp[nt] = ctx->radius[nt];
1216: } else {
1217: nt = LANDAU_MAX_GRIDS;
1218: PetscCall(PetscOptionsRealArray("-dm_landau_domain_max_par", "Parallel velocity domain size in units of thermal velocity of grid", "plexland.c", ctx->radius_par, &nt, &flg));
1219: if (flg) PetscCheck(nt >= ctx->num_grids, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_domain_max_par: given %" PetscInt_FMT " radius != number grids %" PetscInt_FMT, nt, ctx->num_grids);
1220: PetscCall(PetscOptionsRealArray("-dm_landau_domain_max_perp", "Perpendicular velocity domain size in units of thermal velocity of grid", "plexland.c", ctx->radius_perp, &nt, &flg));
1221: if (flg) PetscCheck(nt >= ctx->num_grids, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_domain_max_perp: given %" PetscInt_FMT " radius != number grids %" PetscInt_FMT, nt, ctx->num_grids);
1222: }
1223: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1224: if (flg && ctx->radius[grid] <= 0) { /* negative is ratio of c - need to set par and perp with this -- todo */
1225: if (ctx->radius[grid] == 0) ctx->radius[grid] = 0.75;
1226: else ctx->radius[grid] = -ctx->radius[grid];
1227: ctx->radius[grid] = ctx->radius[grid] * SPEED_OF_LIGHT / ctx->v_0; // use any species on grid to normalize (v_0 same for all on grid)
1228: PetscCall(PetscInfo(dummy, "Change domain radius to %g for grid %" PetscInt_FMT "\n", (double)ctx->radius[grid], grid));
1229: }
1230: ctx->radius[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1231: ctx->radius_perp[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1232: ctx->radius_par[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1233: }
1234: /* amr parameters */
1235: if (!fileflg) {
1236: nt = LANDAU_MAX_GRIDS;
1237: PetscCall(PetscOptionsIntArray("-dm_landau_amr_levels_max", "Number of AMR levels of refinement around origin, after (RE) refinements along z", "plexland.c", ctx->numAMRRefine, &nt, &flg));
1238: PetscCheck(!flg || nt >= ctx->num_grids, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_amr_levels_max: given %" PetscInt_FMT " != number grids %" PetscInt_FMT, nt, ctx->num_grids);
1239: nt = LANDAU_MAX_GRIDS;
1240: PetscCall(PetscOptionsIntArray("-dm_landau_amr_post_refine", "Number of levels to uniformly refine after AMR", "plexland.c", ctx->postAMRRefine, &nt, &flg));
1241: for (ii = 1; ii < ctx->num_grids; ii++) ctx->postAMRRefine[ii] = ctx->postAMRRefine[0]; // all grids the same now
1242: PetscCall(PetscOptionsInt("-dm_landau_amr_re_levels", "Number of levels to refine along v_perp=0, z>0", "plexland.c", ctx->numRERefine, &ctx->numRERefine, &flg));
1243: PetscCall(PetscOptionsInt("-dm_landau_amr_z_refine_pre", "Number of levels to refine along v_perp=0 before origin refine", "plexland.c", ctx->nZRefine1, &ctx->nZRefine1, &flg));
1244: PetscCall(PetscOptionsInt("-dm_landau_amr_z_refine_post", "Number of levels to refine along v_perp=0 after origin refine", "plexland.c", ctx->nZRefine2, &ctx->nZRefine2, &flg));
1245: PetscCall(PetscOptionsReal("-dm_landau_re_radius", "velocity range to refine on positive (z>0) r=0 axis for runaways", "plexland.c", ctx->re_radius, &ctx->re_radius, &flg));
1246: PetscCall(PetscOptionsReal("-dm_landau_z_radius_pre", "velocity range to refine r=0 axis (for electrons)", "plexland.c", ctx->vperp0_radius1, &ctx->vperp0_radius1, &flg));
1247: PetscCall(PetscOptionsReal("-dm_landau_z_radius_post", "velocity range to refine r=0 axis (for electrons) after origin AMR", "plexland.c", ctx->vperp0_radius2, &ctx->vperp0_radius2, &flg));
1248: /* spherical domain */
1249: if (ctx->sphere || ctx->simplex) {
1250: ctx->sphere_uniform_normal = PETSC_FALSE;
1251: PetscCall(PetscOptionsBool("-dm_landau_sphere_uniform_normal", "Scaling of circle radius to get uniform particles per cell with Maxwellians (not used)", "plexland.c", ctx->sphere_uniform_normal, &ctx->sphere_uniform_normal, NULL));
1252: if (!ctx->sphere_uniform_normal) { // true
1253: nt = LANDAU_MAX_GRIDS;
1254: PetscCall(PetscOptionsRealArray("-dm_landau_sphere_inner_radius_90degree_scale", "Scaling of radius for inner circle on 90 degree grid", "plexland.c", ctx->sphere_inner_radius_90degree, &nt, &flg));
1255: if (flg && nt < ctx->num_grids) {
1256: for (PetscInt grid = nt; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = ctx->sphere_inner_radius_90degree[0];
1257: } else if (!flg || nt == 0) {
1258: if (ctx->sphere && !ctx->simplex && LANDAU_DIM == 3) {
1259: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.35; // optimized for R=6, Q4, AMR=0, 0 refinement
1260: } else {
1261: if (LANDAU_DIM == 2) {
1262: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.4; // optimized for R=5, Q4, AMR=0
1263: } else {
1264: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.577 * 0.40;
1265: }
1266: }
1267: }
1268: nt = LANDAU_MAX_GRIDS;
1269: PetscCall(PetscOptionsRealArray("-dm_landau_sphere_inner_radius_45degree_scale", "Scaling of radius for inner circle on 45 degree grid", "plexland.c", ctx->sphere_inner_radius_45degree, &nt, &flg));
1270: if (flg && nt < ctx->num_grids) {
1271: for (PetscInt grid = nt; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = ctx->sphere_inner_radius_45degree[0];
1272: } else if (!flg || nt == 0) {
1273: if (LANDAU_DIM == 2) {
1274: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = 0.45; // optimized for R=5, Q4, AMR=0
1275: } else {
1276: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = 0.4; // 3D sphere
1277: }
1278: }
1279: if (ctx->sphere) PetscCall(PetscInfo(ctx->plex[0], "sphere : , 45 degree scaling = %g; 90 degree scaling = %g\n", (double)ctx->sphere_inner_radius_45degree[0], (double)ctx->sphere_inner_radius_90degree[0]));
1280: } else {
1281: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1282: switch (ctx->numAMRRefine[grid]) {
1283: case 0:
1284: case 1:
1285: case 2:
1286: case 3:
1287: default:
1288: if (LANDAU_DIM == 2) {
1289: ctx->sphere_inner_radius_90degree[grid] = 0.40;
1290: ctx->sphere_inner_radius_45degree[grid] = 0.45;
1291: } else {
1292: ctx->sphere_inner_radius_45degree[grid] = 0.25;
1293: }
1294: }
1295: }
1296: }
1297: } else {
1298: nt = LANDAU_DIM;
1299: PetscCall(PetscOptionsIntArray("-dm_landau_num_cells", "Number of cells in each dimension of base grid", "plexland.c", ctx->cells0, &nt, &flg));
1300: }
1301: }
1302: /* processing options */
1303: PetscCall(PetscOptionsBool("-dm_landau_gpu_assembly", "Assemble Jacobian on GPU", "plexland.c", ctx->gpu_assembly, &ctx->gpu_assembly, NULL));
1304: PetscCall(PetscOptionsBool("-dm_landau_jacobian_field_major_order", "Reorder Jacobian for GPU assembly with field major, or block diagonal, ordering (DEPRECATED)", "plexland.c", ctx->jacobian_field_major_order, &ctx->jacobian_field_major_order, NULL));
1305: if (ctx->jacobian_field_major_order) PetscCheck(ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order requires -dm_landau_gpu_assembly");
1306: PetscCheck(!ctx->jacobian_field_major_order, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED");
1307: PetscOptionsEnd();
1309: for (ii = ctx->num_species; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] = ctx->thermal_temps[ii] = ctx->charges[ii] = 0;
1310: if (ctx->verbose != 0) {
1311: PetscReal pmassunit = PetscRealConstant(1.6720e-27);
1313: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "masses: e=%10.3e; ions in proton mass units: %10.3e %10.3e ...\n", (double)ctx->masses[0], (double)(ctx->masses[1] / pmassunit), (double)(ctx->num_species > 2 ? ctx->masses[2] / pmassunit : 0)));
1314: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "charges: e=%10.3e; charges in elementary units: %10.3e %10.3e\n", (double)ctx->charges[0], (double)(-ctx->charges[1] / ctx->charges[0]), (double)(ctx->num_species > 2 ? -ctx->charges[2] / ctx->charges[0] : 0)));
1315: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "n: e: %10.3e i: %10.3e %10.3e\n", (double)ctx->n[0], (double)ctx->n[1], (double)(ctx->num_species > 2 ? ctx->n[2] : 0)));
1316: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "thermal T (K): e=%10.3e i=%10.3e %10.3e. Normalization grid %" PetscInt_FMT ": v_0=%10.3e (%10.3ec) n_0=%10.3e t_0=%10.3e %" PetscInt_FMT " batched, view batch %" PetscInt_FMT "\n", (double)ctx->thermal_temps[0],
1317: (double)ctx->thermal_temps[1], (double)((ctx->num_species > 2) ? ctx->thermal_temps[2] : 0), non_dim_grid, (double)ctx->v_0, (double)(ctx->v_0 / SPEED_OF_LIGHT), (double)ctx->n_0, (double)ctx->t_0, ctx->batch_sz, ctx->batch_view_idx));
1318: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "Domain radius (AMR levels) grid %d: par=%10.3e perp=%10.3e (%" PetscInt_FMT ") ", 0, (double)ctx->radius_par[0], (double)ctx->radius_perp[0], ctx->numAMRRefine[0]));
1319: for (ii = 1; ii < ctx->num_grids; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, ", %" PetscInt_FMT ": par=%10.3e perp=%10.3e (%" PetscInt_FMT ") ", ii, (double)ctx->radius_par[ii], (double)ctx->radius_perp[ii], ctx->numAMRRefine[ii]));
1320: if (ctx->use_relativistic_corrections) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nUse relativistic corrections\n"));
1321: else PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1322: }
1323: PetscCall(DMDestroy(&dummy));
1324: {
1325: PetscMPIInt rank;
1326: PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank));
1327: ctx->stage = 0;
1328: PetscCall(PetscLogEventRegister("Landau Create", DM_CLASSID, &ctx->events[13])); /* 13 */
1329: PetscCall(PetscLogEventRegister(" GPU ass. setup", DM_CLASSID, &ctx->events[2])); /* 2 */
1330: PetscCall(PetscLogEventRegister(" Build matrix", DM_CLASSID, &ctx->events[12])); /* 12 */
1331: PetscCall(PetscLogEventRegister(" Assembly maps", DM_CLASSID, &ctx->events[15])); /* 15 */
1332: PetscCall(PetscLogEventRegister("Landau Mass mat", DM_CLASSID, &ctx->events[14])); /* 14 */
1333: PetscCall(PetscLogEventRegister("Landau Operator", DM_CLASSID, &ctx->events[11])); /* 11 */
1334: PetscCall(PetscLogEventRegister("Landau Jacobian", DM_CLASSID, &ctx->events[0])); /* 0 */
1335: PetscCall(PetscLogEventRegister("Landau Mass", DM_CLASSID, &ctx->events[9])); /* 9 */
1336: PetscCall(PetscLogEventRegister(" Preamble", DM_CLASSID, &ctx->events[10])); /* 10 */
1337: PetscCall(PetscLogEventRegister(" static IP Data", DM_CLASSID, &ctx->events[7])); /* 7 */
1338: PetscCall(PetscLogEventRegister(" dynamic IP-Jac", DM_CLASSID, &ctx->events[1])); /* 1 */
1339: PetscCall(PetscLogEventRegister(" Kernel-init", DM_CLASSID, &ctx->events[3])); /* 3 */
1340: PetscCall(PetscLogEventRegister(" Jac-f-df (GPU)", DM_CLASSID, &ctx->events[8])); /* 8 */
1341: PetscCall(PetscLogEventRegister(" J Kernel (GPU)", DM_CLASSID, &ctx->events[4])); /* 4 */
1342: PetscCall(PetscLogEventRegister(" M Kernel (GPU)", DM_CLASSID, &ctx->events[16])); /* 16 */
1343: PetscCall(PetscLogEventRegister(" Copy to CPU", DM_CLASSID, &ctx->events[5])); /* 5 */
1344: PetscCall(PetscLogEventRegister(" CPU assemble", DM_CLASSID, &ctx->events[6])); /* 6 */
1346: if (rank) { /* turn off output stuff for duplicate runs - do we need to add the prefix to all this? */
1347: PetscCall(PetscOptionsClearValue(NULL, "-snes_converged_reason"));
1348: PetscCall(PetscOptionsClearValue(NULL, "-ksp_converged_reason"));
1349: PetscCall(PetscOptionsClearValue(NULL, "-snes_monitor"));
1350: PetscCall(PetscOptionsClearValue(NULL, "-ksp_monitor"));
1351: PetscCall(PetscOptionsClearValue(NULL, "-ts_monitor"));
1352: PetscCall(PetscOptionsClearValue(NULL, "-ts_view"));
1353: PetscCall(PetscOptionsClearValue(NULL, "-ts_adapt_monitor"));
1354: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_dm_view"));
1355: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_vec_view"));
1356: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_dm_view"));
1357: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_view"));
1358: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_jacobian_view"));
1359: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mat_view"));
1360: PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_converged_reason"));
1361: PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_monitor"));
1362: PetscCall(PetscOptionsClearValue(NULL, "-"));
1363: PetscCall(PetscOptionsClearValue(NULL, "-info"));
1364: }
1365: }
1366: PetscFunctionReturn(PETSC_SUCCESS);
1367: }
1369: /* Build c_maps and gIdx from PetscSection constraint data and cMat, replacing the probing strategy */
1370: static PetscErrorCode LandauBuildConstraintMaps_PetscSection(DM dm, PetscInt Nf_grid, PetscSection section, PetscSection globsection, P4estVertexMaps *maps, pointInterpolationP4est (*pointMaps)[LANDAU_MAX_Q_FACE], PetscInt MAP_BF_SIZE, LandauIdx *coo_elem_fullNb, LandauIdx *coo_elem_offsets, PetscInt glb_elem_idx_start)
1371: {
1372: PetscDS ds;
1373: PetscSection aSec, cSec;
1374: IS aIS, clpermIS = NULL;
1375: Mat cMat;
1376: const PetscInt *anchors = NULL, *clperm_arr = NULL;
1377: PetscInt cStart, cEnd, aStart, aEnd, sStart, sEnd;
1378: const PetscInt **fieldPerms[LANDAU_MAX_SPECIES];
1379: PetscInt fieldFoffs[LANDAU_MAX_SPECIES]; /* running offset per field in natural order */
1380: PetscInt fullNb[LANDAU_MAX_SPECIES]; /* unconstrained DOF count per field for this element */
1381: PetscInt foffs[LANDAU_MAX_SPECIES + 1]; /* cumulative field offsets in natural closure order */
1382: PetscInt clTotDof; /* total closure DOFs (same for all cells, from DS) */
1384: PetscFunctionBegin;
1385: PetscCheck(Nf_grid <= LANDAU_MAX_SPECIES, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Nf_grid %" PetscInt_FMT " > LANDAU_MAX_SPECIES %d", Nf_grid, LANDAU_MAX_SPECIES);
1386: PetscCall(DMGetDS(dm, &ds));
1387: /* per-field offsets and total DOF count are stored in the DS; no need to recompute per cell */
1388: for (PetscInt f = 0; f < Nf_grid; f++) PetscCall(PetscDSGetFieldOffset(ds, f, &foffs[f]));
1389: PetscCall(PetscDSGetTotalDimension(ds, &clTotDof));
1390: foffs[Nf_grid] = clTotDof;
1391: PetscCall(DMGetDefaultConstraints(dm, &cSec, &cMat, NULL));
1392: PetscCall(DMPlexGetAnchors(dm, &aSec, &aIS));
1393: PetscCall(DMPlexGetHeightStratum(dm, 0, &cStart, &cEnd));
1394: if (aSec) {
1395: PetscCall(PetscSectionGetChart(aSec, &aStart, &aEnd));
1396: PetscCall(ISGetIndices(aIS, &anchors));
1397: } else {
1398: aStart = aEnd = 0;
1399: anchors = NULL;
1400: }
1401: PetscCall(PetscSectionGetChart(section, &sStart, &sEnd));
1402: /* Query closure inverse permutation once; identical for all height-0 cells. */
1403: if (cStart < cEnd) {
1404: PetscSection clSec = NULL;
1405: PetscCall(PetscSectionGetClosureIndex(section, (PetscObject)dm, &clSec, NULL));
1406: if (clSec) { /* closure permutation exists */
1407: PetscInt depth0;
1408: PetscCall(DMPlexGetPointDepth(dm, cStart, &depth0));
1409: PetscCall(PetscSectionGetClosureInversePermutation(section, (PetscObject)dm, depth0, clTotDof, &clpermIS));
1410: PetscCall(ISGetIndices(clpermIS, &clperm_arr));
1411: }
1412: }
1413: for (PetscInt ej = cStart, eidx = 0; ej < cEnd; ++ej, ++eidx) {
1414: PetscInt glb_elem_idx = glb_elem_idx_start + eidx;
1415: PetscInt *closure = NULL;
1416: PetscInt closureSize;
1418: if (coo_elem_offsets) coo_elem_offsets[glb_elem_idx + 1] = coo_elem_offsets[glb_elem_idx];
1419: PetscCall(DMPlexGetTransitiveClosure(dm, ej, PETSC_TRUE, &closureSize, &closure)); /* original closure */
1420: for (PetscInt f = 0; f < Nf_grid; f++) PetscCall(PetscSectionGetFieldPointSyms(section, f, closureSize, closure, &fieldPerms[f], NULL)); /* orientation perms; flips affect sign only and are not needed for index assignment */
1421: PetscCall(PetscArrayzero(fieldFoffs, LANDAU_MAX_SPECIES));
1422: PetscCall(PetscArrayzero(fullNb, LANDAU_MAX_SPECIES));
1423: for (PetscInt ci = 0; ci < closureSize; ci++) { /* fill gIdx / c_maps */
1424: PetscInt p = closure[2 * ci];
1425: if (p < sStart || p >= sEnd) continue;
1426: for (PetscInt f = 0; f < Nf_grid; f++) {
1427: const PetscInt *fcdofs = NULL;
1428: const PetscInt *perm = (fieldPerms[f] && fieldPerms[f][ci]) ? fieldPerms[f][ci] : NULL;
1429: PetscInt fdof = 0, cfdof = 0;
1430: PetscInt pGlobOff = 0;
1431: PetscInt globFieldInPoint = 0; /* unconstrained DOF offset for field f at this point */
1432: PetscInt cind = 0; /* index into fcdofs[] */
1434: PetscCall(PetscSectionGetFieldDof(section, p, f, &fdof));
1435: if (!fdof) continue;
1436: PetscCall(PetscSectionGetFieldConstraintDof(section, p, f, &cfdof));
1437: if (cfdof) PetscCall(PetscSectionGetFieldConstraintIndices(section, p, f, &fcdofs));
1438: PetscCall(PetscSectionGetOffset(globsection, p, &pGlobOff));
1439: for (PetscInt g = 0; g < f; g++) { /* unconstrained DOFs from earlier fields */
1440: PetscInt gfdof = 0, gcfdof = 0;
1441: PetscCall(PetscSectionGetFieldDof(section, p, g, &gfdof));
1442: PetscCall(PetscSectionGetFieldConstraintDof(section, p, g, &gcfdof));
1443: globFieldInPoint += gfdof - gcfdof;
1444: }
1445: for (PetscInt b = 0; b < fdof; b++) {
1446: PetscInt preind = foffs[f] + fieldFoffs[f] + (perm ? perm[b] : b); /* natural order position */
1447: PetscInt q = clperm_arr ? clperm_arr[preind] : preind; /* permuted position */
1448: PetscBool isConstrained = (cfdof > 0 && cind < cfdof && b == fcdofs[cind]) ? PETSC_TRUE : PETSC_FALSE;
1450: q -= foffs[f]; /* subtract field base offset to get q within [0, Nb) */
1451: if (isConstrained) { /* constrained DOF */
1452: PetscInt cOff, row, bDof = 0, bOff2 = 0;
1453: PetscInt nNonzero = 0;
1454: PetscInt trivGid = -1;
1455: PetscScalar trivVal = 0;
1457: PetscCheck(cSec, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Constrained DOF found but constraint section is NULL");
1458: PetscCall(PetscSectionGetFieldOffset(cSec, p, f, &cOff));
1459: row = cOff + cind;
1460: if (aSec && p >= aStart && p < aEnd) {
1461: PetscCall(PetscSectionGetDof(aSec, p, &bDof));
1462: PetscCall(PetscSectionGetOffset(aSec, p, &bOff2));
1463: }
1464: PetscCheck(!bDof || anchors, PETSC_COMM_SELF, PETSC_ERR_PLIB, "constrained point has anchors but anchor array is NULL");
1465: for (PetscInt ai = 0; ai < bDof; ai++) { /* scan cMat row for non-zeros */
1466: PetscInt aLocOff, aGlobOff, aDof = 0;
1467: PetscInt aGlobFieldInPoint = 0;
1468: PetscInt a = anchors[bOff2 + ai];
1470: if (a >= sStart && a < sEnd) PetscCall(PetscSectionGetFieldDof(section, a, f, &aDof));
1471: if (!aDof) continue;
1472: PetscCall(PetscSectionGetFieldOffset(section, a, f, &aLocOff));
1473: PetscCall(PetscSectionGetOffset(globsection, a, &aGlobOff));
1474: for (PetscInt g = 0; g < f; g++) {
1475: PetscInt agfdof = 0, agcfdof = 0;
1476: PetscCall(PetscSectionGetFieldDof(section, a, g, &agfdof));
1477: PetscCall(PetscSectionGetFieldConstraintDof(section, a, g, &agcfdof));
1478: aGlobFieldInPoint += agfdof - agcfdof;
1479: }
1480: for (PetscInt e = 0; e < aDof; e++) {
1481: PetscScalar val;
1482: PetscInt col = aLocOff + e;
1483: PetscCall(MatGetValues(cMat, 1, &row, 1, &col, &val));
1484: if (PetscAbs(PetscRealPart(val)) > PETSC_MACHINE_EPSILON) {
1485: nNonzero++;
1486: trivVal = val;
1487: trivGid = aGlobOff + aGlobFieldInPoint + e;
1488: }
1489: }
1490: }
1491: PetscCheck(nNonzero > 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "constrained DOF (cell %" PetscInt_FMT " field %" PetscInt_FMT ") has no anchor non-zeros in cMat", eidx, f);
1492: if (nNonzero == 1 && PetscAbs(PetscRealPart(trivVal) - 1.0) < 10 * PETSC_MACHINE_EPSILON) { /* trivial constraint */
1493: maps->gIdx[eidx][f][q] = trivGid;
1494: fullNb[f]++;
1495: } else { /* non-trivial constraint: build pointMap */
1496: PetscInt jj = 0;
1498: maps->gIdx[eidx][f][q] = -(maps->num_reduced + 1);
1499: for (PetscInt ai = 0; ai < bDof && jj < maps->num_face; ai++) {
1500: PetscInt aLocOff, aGlobOff, aDof = 0;
1501: PetscInt aGlobFieldInPoint = 0;
1502: PetscInt a = anchors[bOff2 + ai];
1504: if (a >= sStart && a < sEnd) PetscCall(PetscSectionGetFieldDof(section, a, f, &aDof));
1505: if (!aDof) continue;
1506: PetscCall(PetscSectionGetFieldOffset(section, a, f, &aLocOff));
1507: PetscCall(PetscSectionGetOffset(globsection, a, &aGlobOff));
1508: for (PetscInt g = 0; g < f; g++) {
1509: PetscInt agfdof = 0, agcfdof = 0;
1510: PetscCall(PetscSectionGetFieldDof(section, a, g, &agfdof));
1511: PetscCall(PetscSectionGetFieldConstraintDof(section, a, g, &agcfdof));
1512: aGlobFieldInPoint += agfdof - agcfdof;
1513: }
1514: for (PetscInt e = 0; e < aDof && jj < maps->num_face; e++) {
1515: PetscScalar val;
1516: PetscReal rval;
1517: PetscInt col = aLocOff + e;
1519: PetscCall(MatGetValues(cMat, 1, &row, 1, &col, &val));
1520: rval = PetscRealPart(val);
1521: if (PetscAbs(rval) <= PETSC_MACHINE_EPSILON) rval = 0.0;
1522: pointMaps[maps->num_reduced][jj].scale = rval;
1523: pointMaps[maps->num_reduced][jj].gid = (rval == 0.0) ? -1 : aGlobOff + aGlobFieldInPoint + e;
1524: if (rval != 0.0) fullNb[f]++;
1525: jj++;
1526: }
1527: }
1528: while (jj < maps->num_face) {
1529: pointMaps[maps->num_reduced][jj].scale = 0.0;
1530: pointMaps[maps->num_reduced][jj].gid = -1;
1531: jj++;
1532: }
1533: maps->num_reduced++;
1534: PetscCheck(maps->num_reduced < MAP_BF_SIZE, PETSC_COMM_SELF, PETSC_ERR_PLIB, "maps->num_reduced %" PetscInt_FMT " >= MAP_BF_SIZE %" PetscInt_FMT, maps->num_reduced, MAP_BF_SIZE);
1535: }
1536: cind++;
1537: } else { /* unconstrained DOF */
1538: PetscCheck(pGlobOff >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Landau per-grid plex must be on PETSC_COMM_SELF; got off-process global offset");
1539: maps->gIdx[eidx][f][q] = pGlobOff + globFieldInPoint + b - cind;
1540: fullNb[f]++;
1541: }
1542: }
1543: fieldFoffs[f] += fdof;
1544: }
1545: }
1546: for (PetscInt f = 0; f < Nf_grid; f++) PetscCall(PetscSectionRestoreFieldPointSyms(section, f, closureSize, closure, &fieldPerms[f], NULL));
1547: PetscCall(DMPlexRestoreTransitiveClosure(dm, ej, PETSC_TRUE, &closureSize, &closure));
1548: if (coo_elem_offsets) { /* COO offsets */
1549: for (PetscInt f = 0; f < Nf_grid; f++) {
1550: coo_elem_offsets[glb_elem_idx + 1] += fullNb[f] * fullNb[f];
1551: if (f == 0) coo_elem_fullNb[glb_elem_idx] = fullNb[f];
1552: else PetscCheck(coo_elem_fullNb[glb_elem_idx] == fullNb[f], PETSC_COMM_SELF, PETSC_ERR_PLIB, "full element size change with species %" PetscInt_FMT " %" PetscInt_FMT, coo_elem_fullNb[glb_elem_idx], fullNb[f]);
1553: }
1554: }
1555: } /* cell */
1556: if (clpermIS) {
1557: PetscCall(ISRestoreIndices(clpermIS, &clperm_arr));
1558: PetscCall(ISDestroy(&clpermIS));
1559: }
1560: if (aSec) PetscCall(ISRestoreIndices(aIS, &anchors));
1561: PetscFunctionReturn(PETSC_SUCCESS);
1562: }
1564: static PetscErrorCode CreateStaticData(PetscInt dim, IS grid_batch_is_inv[], const char prefix[], LandauCtx *ctx)
1565: {
1566: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
1567: PetscQuadrature quad;
1568: const PetscReal *quadWeights;
1569: PetscReal invMass[LANDAU_MAX_SPECIES], nu_alpha[LANDAU_MAX_SPECIES], nu_beta[LANDAU_MAX_SPECIES];
1570: PetscInt numCells[LANDAU_MAX_GRIDS], Nq, Nb, Nf[LANDAU_MAX_GRIDS], ncellsTot = 0, MAP_BF_SIZE = 64 * LANDAU_DIM * LANDAU_DIM * LANDAU_MAX_Q_FACE * LANDAU_MAX_SPECIES;
1571: PetscTabulation *Tf;
1572: PetscDS prob;
1574: PetscFunctionBegin;
1575: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
1576: PetscCheck(Nb <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nb = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nb, LANDAU_MAX_NQND);
1577: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1578: for (PetscInt ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++) {
1579: invMass[ii] = ctx->m_0 / ctx->masses[ii];
1580: nu_alpha[ii] = PetscSqr(ctx->charges[ii] / ctx->m_0) * ctx->m_0 / ctx->masses[ii];
1581: nu_beta[ii] = PetscSqr(ctx->charges[ii] / ctx->epsilon0) / (8 * PETSC_PI) * ctx->t_0 * ctx->n_0 / PetscPowReal(ctx->v_0, 3);
1582: }
1583: }
1584: if (ctx->verbose == 4) {
1585: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "nu_alpha: "));
1586: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1587: PetscInt iii = ctx->species_offset[grid];
1588: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_alpha[ii]));
1589: }
1590: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_beta: "));
1591: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1592: PetscInt iii = ctx->species_offset[grid];
1593: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_beta[ii]));
1594: }
1595: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_alpha[i]*nu_beta[j]*lambda[i][j]:\n"));
1596: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1597: PetscInt iii = ctx->species_offset[grid];
1598: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) {
1599: for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) {
1600: PetscInt jjj = ctx->species_offset[gridj];
1601: for (PetscInt jj = jjj; jj < ctx->species_offset[gridj + 1]; jj++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %14.9e", (double)(nu_alpha[ii] * nu_beta[jj] * ctx->lambdas[grid][gridj])));
1602: }
1603: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1604: }
1605: }
1606: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "lambda[i][j]:\n"));
1607: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1608: PetscInt iii = ctx->species_offset[grid];
1609: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) {
1610: for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) {
1611: PetscInt jjj = ctx->species_offset[gridj];
1612: for (PetscInt jj = jjj; jj < ctx->species_offset[gridj + 1]; jj++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %14.9e", (double)ctx->lambdas[grid][gridj]));
1613: }
1614: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1615: }
1616: }
1617: }
1618: PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids
1619: PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids
1620: /* DS, Tab and quad is same on all grids */
1621: PetscCheck(ctx->plex[0], ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
1622: PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad));
1623: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, &quadWeights));
1624: PetscCheck(Nq <= LANDAU_MAX_NQND, ctx->comm, PETSC_ERR_ARG_WRONG, "Order too high. Nq = %" PetscInt_FMT " > LANDAU_MAX_NQND (%d)", Nq, LANDAU_MAX_NQND);
1625: /* setup each grid */
1626: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1627: PetscInt cStart, cEnd;
1628: PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
1629: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1630: numCells[grid] = cEnd - cStart; // grids can have different topology
1631: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
1632: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
1633: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
1634: ncellsTot += numCells[grid];
1635: }
1636: /* create GPU assembly data */
1637: if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
1638: PetscContainer container;
1639: pointInterpolationP4est(*pointMaps)[LANDAU_MAX_Q_FACE];
1640: P4estVertexMaps *maps;
1641: const PetscInt *plex_batch = NULL;
1642: LandauIdx *coo_elem_offsets = NULL, *coo_elem_fullNb = NULL, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = NULL;
1643: PetscCall(PetscLogEventBegin(ctx->events[2], 0, 0, 0, 0));
1644: PetscCall(PetscMalloc(sizeof(*maps) * ctx->num_grids, &maps));
1645: PetscCall(PetscMalloc(sizeof(*pointMaps) * MAP_BF_SIZE, &pointMaps));
1647: { // setup COO assembly -- put COO metadata directly in ctx->SData_d
1648: PetscCall(PetscMalloc3(ncellsTot + 1, &coo_elem_offsets, ncellsTot, &coo_elem_fullNb, ncellsTot, &coo_elem_point_offsets)); // array of integer pointers
1649: coo_elem_offsets[0] = 0; // finish later
1650: PetscCall(PetscInfo(ctx->plex[0], "COO initialization, %" PetscInt_FMT " cells\n", ncellsTot));
1651: ctx->SData_d.coo_n_cellsTot = ncellsTot;
1652: ctx->SData_d.coo_elem_offsets = (void *)coo_elem_offsets;
1653: ctx->SData_d.coo_elem_fullNb = (void *)coo_elem_fullNb;
1654: ctx->SData_d.coo_elem_point_offsets = (void *)coo_elem_point_offsets;
1655: }
1657: ctx->SData_d.coo_max_fullnb = 0;
1658: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1659: if (grid_batch_is_inv[grid]) PetscCall(ISGetIndices(grid_batch_is_inv[grid], &plex_batch));
1660: PetscCheck(!plex_batch, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED");
1661: // make maps
1662: maps[grid].d_self = NULL;
1663: maps[grid].num_elements = numCells[grid];
1664: maps[grid].num_face = (PetscInt)(pow(Nq, 1. / ((double)dim)) + .001); // Q
1665: maps[grid].num_face = (PetscInt)(pow(maps[grid].num_face, (double)(dim - 1)) + .001); // Q^2
1666: maps[grid].num_reduced = 0;
1667: maps[grid].deviceType = ctx->deviceType;
1668: maps[grid].numgrids = ctx->num_grids;
1669: PetscCall(PetscMalloc(maps[grid].num_elements * sizeof(*maps[grid].gIdx), &maps[grid].gIdx));
1670: PetscCall(LandauBuildConstraintMaps_PetscSection(ctx->plex[grid], Nf[grid], section[grid], globsection[grid], &maps[grid], pointMaps, MAP_BF_SIZE, coo_elem_fullNb, coo_elem_offsets, glb_elem_idx));
1671: for (PetscInt ej = 0; ej < numCells[grid]; ej++) {
1672: if (coo_elem_fullNb[glb_elem_idx + ej] > ctx->SData_d.coo_max_fullnb) ctx->SData_d.coo_max_fullnb = coo_elem_fullNb[glb_elem_idx + ej];
1673: }
1674: glb_elem_idx += numCells[grid];
1675: // allocate and copy point data maps[grid].gIdx[eidx][field][q]
1676: PetscCall(PetscMalloc(maps[grid].num_reduced * sizeof(*maps[grid].c_maps), &maps[grid].c_maps));
1677: for (PetscInt ej = 0; ej < maps[grid].num_reduced; ++ej) {
1678: for (PetscInt q = 0; q < maps[grid].num_face; ++q) {
1679: maps[grid].c_maps[ej][q].scale = pointMaps[ej][q].scale;
1680: maps[grid].c_maps[ej][q].gid = pointMaps[ej][q].gid;
1681: }
1682: }
1683: #if PetscDefined(HAVE_KOKKOS)
1684: if (ctx->deviceType == LANDAU_KOKKOS) PetscCall(LandauKokkosCreateMatMaps(maps, pointMaps, Nf, grid)); // implies Kokkos does
1685: #endif
1686: if (plex_batch) {
1687: PetscCall(ISRestoreIndices(grid_batch_is_inv[grid], &plex_batch));
1688: PetscCall(ISDestroy(&grid_batch_is_inv[grid])); // we are done with this
1689: }
1690: } /* grids */
1691: // finish COO
1692: { // setup COO assembly
1693: PetscInt *oor, *ooc;
1694: ctx->SData_d.coo_size = (PetscCount)coo_elem_offsets[ncellsTot] * ctx->batch_sz;
1695: PetscCall(PetscMalloc2(ctx->SData_d.coo_size, &oor, ctx->SData_d.coo_size, &ooc));
1696: for (PetscCount i = 0; i < ctx->SData_d.coo_size; i++) oor[i] = ooc[i] = -1;
1697: // get
1698: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1699: for (PetscInt ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) {
1700: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx];
1701: const LandauIdx *const Idxs = &maps[grid].gIdx[ej][0][0]; // just use field-0 maps, They should be the same but this is just for COO storage
1702: coo_elem_point_offsets[glb_elem_idx][0] = 0;
1703: for (PetscInt f = 0, cnt2 = 0; f < Nb; f++) {
1704: PetscInt idx = Idxs[f];
1705: coo_elem_point_offsets[glb_elem_idx][f + 1] = coo_elem_point_offsets[glb_elem_idx][f]; // start at last
1706: if (idx >= 0) {
1707: cnt2++;
1708: coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc
1709: } else {
1710: idx = -idx - 1;
1711: for (PetscInt q = 0; q < maps[grid].num_face; q++) {
1712: if (maps[grid].c_maps[idx][q].gid >= 0) { // skip zero-scale (gid=-1) entries; do not break - they may be non-contiguous in 3D AMR
1713: cnt2++;
1714: coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc
1715: }
1716: }
1717: }
1718: PetscCheck(cnt2 <= fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "wrong count %" PetscInt_FMT " < %" PetscInt_FMT, fullNb, cnt2);
1719: }
1720: PetscCheck(coo_elem_point_offsets[glb_elem_idx][Nb] == fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "coo_elem_point_offsets size %" PetscInt_FMT " != fullNb=%" PetscInt_FMT, coo_elem_point_offsets[glb_elem_idx][Nb], fullNb);
1721: }
1722: }
1723: // set
1724: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
1725: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1726: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
1727: for (PetscInt ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) {
1728: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb;
1729: // set (i,j)
1730: for (PetscInt fieldA = 0; fieldA < Nf[grid]; fieldA++) {
1731: const LandauIdx *const Idxs = &maps[grid].gIdx[ej][fieldA][0];
1732: PetscInt rows[LANDAU_MAX_Q_FACE], cols[LANDAU_MAX_Q_FACE];
1733: for (PetscInt f = 0; f < Nb; ++f) {
1734: const PetscInt nr = coo_elem_point_offsets[glb_elem_idx][f + 1] - coo_elem_point_offsets[glb_elem_idx][f];
1735: if (nr == 1) rows[0] = Idxs[f];
1736: else {
1737: const PetscInt idx = -Idxs[f] - 1;
1738: for (PetscInt q = 0, ri = 0; q < maps[grid].num_face; q++)
1739: if (maps[grid].c_maps[idx][q].gid >= 0) rows[ri++] = maps[grid].c_maps[idx][q].gid;
1740: }
1741: for (PetscInt g = 0; g < Nb; ++g) {
1742: const PetscInt nc = coo_elem_point_offsets[glb_elem_idx][g + 1] - coo_elem_point_offsets[glb_elem_idx][g];
1743: if (nc == 1) cols[0] = Idxs[g];
1744: else {
1745: const PetscInt idx = -Idxs[g] - 1;
1746: for (PetscInt q = 0, ci = 0; q < maps[grid].num_face; q++)
1747: if (maps[grid].c_maps[idx][q].gid >= 0) cols[ci++] = maps[grid].c_maps[idx][q].gid;
1748: }
1749: const PetscInt idx0 = b_id * coo_elem_offsets[ncellsTot] + coo_elem_offsets[glb_elem_idx] + fieldA * fullNb2 + fullNb * coo_elem_point_offsets[glb_elem_idx][f] + nr * coo_elem_point_offsets[glb_elem_idx][g];
1750: for (PetscInt q = 0, idx = idx0; q < nr; q++) {
1751: for (PetscInt d = 0; d < nc; d++, idx++) {
1752: oor[idx] = rows[q] + moffset;
1753: ooc[idx] = cols[d] + moffset;
1754: }
1755: }
1756: }
1757: }
1758: }
1759: } // cell
1760: } // grid
1761: } // batch
1762: PetscCall(MatSetPreallocationCOO(ctx->J, ctx->SData_d.coo_size, oor, ooc));
1763: PetscCall(PetscFree2(oor, ooc));
1764: }
1765: PetscCall(PetscFree(pointMaps));
1766: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
1767: PetscCall(PetscContainerSetPointer(container, (void *)maps));
1768: PetscCall(PetscContainerSetCtxDestroy(container, LandauGPUMapsDestroy));
1769: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "assembly_maps", (PetscObject)container));
1770: PetscCall(PetscContainerDestroy(&container));
1771: PetscCall(PetscLogEventEnd(ctx->events[2], 0, 0, 0, 0));
1772: } // end GPU assembly
1773: { /* create static point data, Jacobian called first, only one vertex copy */
1774: PetscReal *invJe, *ww, *xx, *yy, *zz = NULL, *invJ_a;
1775: PetscInt outer_ipidx, outer_ej, grid, nip_glb = 0;
1776: PetscFE fe;
1777: PetscCall(PetscLogEventBegin(ctx->events[7], 0, 0, 0, 0));
1778: PetscCall(PetscInfo(ctx->plex[0], "Initialize static data\n"));
1779: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) nip_glb += Nq * numCells[grid];
1780: /* collect f data, first time is for Jacobian, but make mass now */
1781: if (ctx->verbose != 0) {
1782: PetscInt ncells = 0, N;
1783: MatInfo info;
1784: PetscCall(MatGetInfo(ctx->J, MAT_LOCAL, &info));
1785: PetscCall(MatGetSize(ctx->J, &N, NULL));
1786: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ncells += numCells[grid];
1787: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%d) %s %" PetscInt_FMT " IPs, %" PetscInt_FMT " cells total, Nb=%" PetscInt_FMT ", Nq=%" PetscInt_FMT ", dim=%" PetscInt_FMT ", Tab: Nb=%" PetscInt_FMT " Nf=%" PetscInt_FMT " Np=%" PetscInt_FMT " cdim=%" PetscInt_FMT " N=%" PetscInt_FMT " nnz= %" PetscInt_FMT "\n", 0, "FormLandau", nip_glb, ncells, Nb, Nq, dim, Nb,
1788: ctx->num_species, Nb, dim, N, (PetscInt)info.nz_used));
1789: }
1790: PetscCall(PetscMalloc4(nip_glb, &ww, nip_glb, &xx, nip_glb, &yy, nip_glb * dim * dim, &invJ_a));
1791: if (dim == 3) PetscCall(PetscMalloc1(nip_glb, &zz));
1792: if (ctx->use_energy_tensor_trick) {
1793: PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, prefix, PETSC_DECIDE, &fe));
1794: PetscCall(PetscObjectSetName((PetscObject)fe, "energy"));
1795: }
1796: /* init each grids static data - no batch */
1797: for (grid = 0, outer_ipidx = 0, outer_ej = 0; grid < ctx->num_grids; grid++) { // OpenMP (once)
1798: Vec v2_2 = NULL; // projected function: v^2/2 for non-relativistic, gamma... for relativistic
1799: PetscSection e_section;
1800: DM dmEnergy;
1801: PetscInt cStart, cEnd, ej;
1803: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1804: // prep energy trick, get v^2 / 2 vector
1805: if (ctx->use_energy_tensor_trick) {
1806: PetscErrorCode (*energyf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {ctx->use_relativistic_corrections ? gamma_m1_f : energy_f};
1807: Vec glob_v2;
1808: PetscReal *c2_0[1], data[1] = {PetscSqr(C_0(ctx->v_0))};
1810: PetscCall(DMClone(ctx->plex[grid], &dmEnergy));
1811: PetscCall(PetscObjectSetName((PetscObject)dmEnergy, "energy"));
1812: PetscCall(DMSetField(dmEnergy, 0, NULL, (PetscObject)fe));
1813: PetscCall(DMCreateDS(dmEnergy));
1814: PetscCall(DMGetLocalSection(dmEnergy, &e_section));
1815: PetscCall(DMGetGlobalVector(dmEnergy, &glob_v2));
1816: PetscCall(PetscObjectSetName((PetscObject)glob_v2, "trick"));
1817: c2_0[0] = &data[0];
1818: PetscCall(DMProjectFunction(dmEnergy, 0., energyf, (void **)c2_0, INSERT_ALL_VALUES, glob_v2));
1819: PetscCall(DMGetLocalVector(dmEnergy, &v2_2));
1820: PetscCall(VecZeroEntries(v2_2)); /* zero BCs so don't set */
1821: PetscCall(DMGlobalToLocalBegin(dmEnergy, glob_v2, INSERT_VALUES, v2_2));
1822: PetscCall(DMGlobalToLocalEnd(dmEnergy, glob_v2, INSERT_VALUES, v2_2));
1823: PetscCall(DMViewFromOptions(dmEnergy, NULL, "-energy_dm_view"));
1824: PetscCall(VecViewFromOptions(glob_v2, NULL, "-energy_vec_view"));
1825: PetscCall(DMRestoreGlobalVector(dmEnergy, &glob_v2));
1826: }
1827: /* append part of the IP data for each grid */
1828: for (ej = 0; ej < numCells[grid]; ++ej, ++outer_ej) {
1829: PetscScalar *coefs = NULL;
1830: PetscReal vj[LANDAU_MAX_NQND * LANDAU_DIM], detJj[LANDAU_MAX_NQND], Jdummy[LANDAU_MAX_NQND * LANDAU_DIM * LANDAU_DIM], c0 = C_0(ctx->v_0), c02 = PetscSqr(c0);
1831: invJe = invJ_a + outer_ej * Nq * dim * dim;
1832: PetscCall(DMPlexComputeCellGeometryFEM(ctx->plex[grid], ej + cStart, quad, vj, Jdummy, invJe, detJj));
1833: if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecGetClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs));
1834: /* create static point data */
1835: for (PetscInt qj = 0; qj < Nq; qj++, outer_ipidx++) {
1836: const PetscInt gidx = outer_ipidx;
1837: const PetscReal *invJ = &invJe[qj * dim * dim];
1838: ww[gidx] = detJj[qj] * quadWeights[qj];
1839: if (dim == 2) ww[gidx] *= vj[qj * dim + 0]; /* cylindrical coordinate, w/o 2pi */
1840: // get xx, yy, zz
1841: if (ctx->use_energy_tensor_trick) {
1842: double refSpaceDer[3], eGradPhi[3];
1843: const PetscReal *const DD = Tf[0]->T[1];
1844: const PetscReal *Dq = &DD[qj * Nb * dim];
1845: for (PetscInt d = 0; d < 3; ++d) refSpaceDer[d] = eGradPhi[d] = 0.0;
1846: for (PetscInt b = 0; b < Nb; ++b) {
1847: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b * dim + d] * PetscRealPart(coefs[b]);
1848: }
1849: xx[gidx] = 1e10;
1850: if (ctx->use_relativistic_corrections) {
1851: double dg2_c2 = 0;
1852: //for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] *= c02;
1853: for (PetscInt d = 0; d < dim; ++d) dg2_c2 += PetscSqr(refSpaceDer[d]);
1854: dg2_c2 *= (double)c02;
1855: if (dg2_c2 >= .999) {
1856: xx[gidx] = vj[qj * dim + 0]; /* coordinate */
1857: yy[gidx] = vj[qj * dim + 1];
1858: if (dim == 3) zz[gidx] = vj[qj * dim + 2];
1859: PetscCall(PetscPrintf(ctx->comm, "Error: %12.5e %" PetscInt_FMT ".%" PetscInt_FMT ") dg2/c02 = %12.5e x= %12.5e %12.5e %12.5e\n", (double)PetscSqrtReal(xx[gidx] * xx[gidx] + yy[gidx] * yy[gidx] + zz[gidx] * zz[gidx]), ej, qj, dg2_c2, (double)xx[gidx], (double)yy[gidx], (double)zz[gidx]));
1860: } else {
1861: PetscReal fact = c02 / PetscSqrtReal(1. - dg2_c2);
1862: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] *= fact;
1863: // could test with other point u' that (grad - grad') * U (refSpaceDer, refSpaceDer') == 0
1864: }
1865: }
1866: if (xx[gidx] == 1e10) {
1867: for (PetscInt d = 0; d < dim; ++d) {
1868: for (PetscInt e = 0; e < dim; ++e) eGradPhi[d] += invJ[e * dim + d] * refSpaceDer[e];
1869: }
1870: xx[gidx] = eGradPhi[0];
1871: yy[gidx] = eGradPhi[1];
1872: if (dim == 3) zz[gidx] = eGradPhi[2];
1873: }
1874: } else {
1875: xx[gidx] = vj[qj * dim + 0]; /* coordinate */
1876: yy[gidx] = vj[qj * dim + 1];
1877: if (dim == 3) zz[gidx] = vj[qj * dim + 2];
1878: }
1879: } /* q */
1880: if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecRestoreClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs));
1881: } /* ej */
1882: if (ctx->use_energy_tensor_trick) {
1883: PetscCall(DMRestoreLocalVector(dmEnergy, &v2_2));
1884: PetscCall(DMDestroy(&dmEnergy));
1885: }
1886: } /* grid */
1887: if (ctx->use_energy_tensor_trick) PetscCall(PetscFEDestroy(&fe));
1888: /* cache static data */
1889: if (ctx->deviceType == LANDAU_KOKKOS) {
1890: #if PetscDefined(HAVE_KOKKOS)
1891: PetscCall(LandauKokkosStaticDataSet(ctx->plex[0], Nq, Nb, ctx->batch_sz, ctx->num_grids, numCells, ctx->species_offset, ctx->mat_offset, nu_alpha, nu_beta, invMass, (PetscReal *)ctx->lambdas, invJ_a, xx, yy, zz, ww, &ctx->SData_d));
1892: /* free */
1893: PetscCall(PetscFree4(ww, xx, yy, invJ_a));
1894: if (dim == 3) PetscCall(PetscFree(zz));
1895: #else
1896: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type kokkos not built");
1897: #endif
1898: } else { /* CPU version, just copy in, only use part */
1899: PetscReal *nu_alpha_p = (PetscReal *)ctx->SData_d.alpha, *nu_beta_p = (PetscReal *)ctx->SData_d.beta, *invMass_p = (PetscReal *)ctx->SData_d.invMass, *lambdas_p = NULL; // why set these ?
1900: ctx->SData_d.w = (void *)ww;
1901: ctx->SData_d.x = (void *)xx;
1902: ctx->SData_d.y = (void *)yy;
1903: ctx->SData_d.z = (void *)zz;
1904: ctx->SData_d.invJ = (void *)invJ_a;
1905: PetscCall(PetscMalloc4(ctx->num_species, &nu_alpha_p, ctx->num_species, &nu_beta_p, ctx->num_species, &invMass_p, LANDAU_MAX_GRIDS * LANDAU_MAX_GRIDS, &lambdas_p));
1906: for (PetscInt ii = 0; ii < ctx->num_species; ii++) {
1907: nu_alpha_p[ii] = nu_alpha[ii];
1908: nu_beta_p[ii] = nu_beta[ii];
1909: invMass_p[ii] = invMass[ii];
1910: }
1911: ctx->SData_d.alpha = (void *)nu_alpha_p;
1912: ctx->SData_d.beta = (void *)nu_beta_p;
1913: ctx->SData_d.invMass = (void *)invMass_p;
1914: ctx->SData_d.lambdas = (void *)lambdas_p;
1915: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1916: PetscReal (*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal (*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas;
1917: for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) (*lambdas)[grid][gridj] = ctx->lambdas[grid][gridj];
1918: }
1919: }
1920: PetscCall(PetscLogEventEnd(ctx->events[7], 0, 0, 0, 0));
1921: } // initialize
1922: PetscFunctionReturn(PETSC_SUCCESS);
1923: }
1925: /* < v, u > */
1926: static void g0_1(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[])
1927: {
1928: g0[0] = 1.;
1929: }
1931: /* < v, u > */
1932: static void g0_fake(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[])
1933: {
1934: static double ttt = 1e-12;
1935: g0[0] = ttt++;
1936: }
1938: /* < v, u > */
1939: static void g0_r(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, PetscReal u_tShift, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar g0[])
1940: {
1941: g0[0] = 2. * PETSC_PI * x[0];
1942: }
1944: /* Creates ctx->J sparsity pattern without real data; supports field-major ordering */
1945: static PetscErrorCode LandauCreateJacobianMatrix(MPI_Comm comm, Vec X, IS grid_batch_is_inv[LANDAU_MAX_GRIDS], LandauCtx *ctx)
1946: {
1947: PetscInt *idxs = NULL;
1948: Mat subM[LANDAU_MAX_GRIDS];
1950: PetscFunctionBegin;
1951: if (!ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
1952: PetscFunctionReturn(PETSC_SUCCESS);
1953: }
1954: // get the RCM for this grid to separate out species into blocks -- create 'idxs' & 'ctx->batch_is' -- not used
1955: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) PetscCall(PetscMalloc1(ctx->mat_offset[ctx->num_grids] * ctx->batch_sz, &idxs));
1956: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1957: const PetscInt *values, n = ctx->mat_offset[grid + 1] - ctx->mat_offset[grid];
1958: Mat gMat;
1959: DM massDM;
1960: PetscDS prob;
1961: Vec tvec;
1962: // get "mass" matrix for reordering
1963: PetscCall(DMClone(ctx->plex[grid], &massDM));
1964: PetscCall(DMCopyFields(ctx->plex[grid], PETSC_DETERMINE, PETSC_DETERMINE, massDM));
1965: PetscCall(DMCreateDS(massDM));
1966: PetscCall(DMGetDS(massDM, &prob));
1967: for (PetscInt ix = 0, ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++, ix++) PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_fake, NULL, NULL, NULL));
1968: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only")); // this trick is need to both sparsify the matrix and avoid runtime error
1969: PetscCall(DMCreateMatrix(massDM, &gMat));
1970: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
1971: PetscCall(MatSetOption(gMat, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
1972: PetscCall(MatSetOption(gMat, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
1973: PetscCall(DMCreateLocalVector(ctx->plex[grid], &tvec));
1974: PetscCall(DMPlexSNESComputeJacobianFEM(massDM, tvec, gMat, gMat, ctx));
1975: PetscCall(MatViewFromOptions(gMat, NULL, "-dm_landau_reorder_mat_view"));
1976: PetscCall(DMDestroy(&massDM));
1977: PetscCall(VecDestroy(&tvec));
1978: subM[grid] = gMat;
1979: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
1980: MatOrderingType rtype = MATORDERINGRCM;
1981: IS isrow, isicol;
1982: PetscCall(MatGetOrdering(gMat, rtype, &isrow, &isicol));
1983: PetscCall(ISInvertPermutation(isrow, PETSC_DECIDE, &grid_batch_is_inv[grid]));
1984: PetscCall(ISGetIndices(isrow, &values));
1985: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
1986: #if !defined(LANDAU_SPECIES_MAJOR)
1987: PetscInt N = ctx->mat_offset[ctx->num_grids], n0 = ctx->mat_offset[grid] + b_id * N;
1988: for (PetscInt ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0;
1989: #else
1990: PetscInt n0 = ctx->mat_offset[grid] * ctx->batch_sz + b_id * n;
1991: for (PetscInt ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0;
1992: #endif
1993: }
1994: PetscCall(ISRestoreIndices(isrow, &values));
1995: PetscCall(ISDestroy(&isrow));
1996: PetscCall(ISDestroy(&isicol));
1997: }
1998: }
1999: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) PetscCall(ISCreateGeneral(comm, ctx->mat_offset[ctx->num_grids] * ctx->batch_sz, idxs, PETSC_OWN_POINTER, &ctx->batch_is));
2000: // get a block matrix
2001: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2002: Mat B = subM[grid];
2003: PetscInt nloc, nzl, *colbuf, row, COL_BF_SIZE = 1024;
2004: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2005: PetscCall(MatGetSize(B, &nloc, NULL));
2006: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2007: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2008: const PetscInt *cols;
2009: const PetscScalar *vals;
2010: for (PetscInt i = 0; i < nloc; i++) {
2011: PetscCall(MatGetRow(B, i, &nzl, NULL, NULL));
2012: if (nzl > COL_BF_SIZE) {
2013: PetscCall(PetscFree(colbuf));
2014: PetscCall(PetscInfo(ctx->plex[grid], "Realloc buffer %" PetscInt_FMT " to %" PetscInt_FMT " (row size %" PetscInt_FMT ") \n", COL_BF_SIZE, 2 * COL_BF_SIZE, nzl));
2015: COL_BF_SIZE = nzl;
2016: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2017: }
2018: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
2019: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset;
2020: row = i + moffset;
2021: PetscCall(MatSetValues(ctx->J, 1, &row, nzl, colbuf, vals, INSERT_VALUES));
2022: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
2023: }
2024: }
2025: PetscCall(PetscFree(colbuf));
2026: }
2027: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid]));
2028: PetscCall(MatAssemblyBegin(ctx->J, MAT_FINAL_ASSEMBLY));
2029: PetscCall(MatAssemblyEnd(ctx->J, MAT_FINAL_ASSEMBLY));
2031: // debug
2032: PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_mat_view"));
2033: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
2034: Mat mat_block_order;
2035: PetscCall(MatCreateSubMatrix(ctx->J, ctx->batch_is, ctx->batch_is, MAT_INITIAL_MATRIX, &mat_block_order)); // use MatPermute
2036: PetscCall(MatViewFromOptions(mat_block_order, NULL, "-dm_landau_mat_view"));
2037: PetscCall(MatDestroy(&mat_block_order));
2038: PetscCall(VecScatterCreate(X, ctx->batch_is, X, NULL, &ctx->plex_batch));
2039: PetscCall(VecDuplicate(X, &ctx->work_vec));
2040: }
2041: PetscFunctionReturn(PETSC_SUCCESS);
2042: }
2044: static void LandauSphereMapping(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar f[])
2045: {
2046: PetscReal u_max = 0, u_norm = 0, scale, square_inner_radius = PetscRealPart(constants[0]), square_radius = PetscRealPart(constants[1]);
2048: for (PetscInt d = 0; d < dim; ++d) {
2049: PetscReal val = PetscAbsReal(PetscRealPart(u[d]));
2050: if (val > u_max) u_max = val;
2051: u_norm += PetscRealPart(u[d]) * PetscRealPart(u[d]);
2052: }
2053: u_norm = PetscSqrtReal(u_norm);
2055: if (u_max < square_inner_radius) {
2056: for (PetscInt d = 0; d < dim; ++d) f[d] = u[d];
2057: return;
2058: }
2060: /* Map rays linearly from inner cube (|u|=square_inner_radius) to outer cube (|u|=square_radius),
2061: projecting outer-cube points onto the sphere of radius square_radius*sqrt(3). */
2062: if (u_max > square_radius + 1e-5) (void)PetscPrintf(PETSC_COMM_SELF, "Error: Point outside outer radius: u_max %g > %g\n", (double)u_max, (double)square_radius);
2063: /* if (PetscAbsReal(u_max - square_inner_radius) < 1e-5 || PetscAbsReal(u_max - square_radius) < 1e-5) {
2064: (void)PetscPrintf(PETSC_COMM_SELF, "Warning: Point near corner of inner and outer cube: u_max %g, inner %g, outer %g\n", (double)u_max, (double)square_inner_radius, (double)square_radius);
2065: } */
2066: {
2067: PetscReal u_0_norm = u_norm * square_inner_radius / u_max;
2068: PetscReal R_max = square_radius * PetscSqrtReal((PetscReal)dim);
2069: PetscReal t = (u_max - square_inner_radius) / (square_radius - square_inner_radius);
2070: PetscReal rho_prime = (1.0 - t) * u_0_norm + t * R_max;
2071: scale = rho_prime / u_norm;
2072: }
2073: for (PetscInt d = 0; d < dim; ++d) f[d] = u[d] * scale;
2074: }
2076: static PetscErrorCode LandauSphereMesh(DM dm, PetscReal inner, PetscReal radius)
2077: {
2078: DM cdm;
2079: PetscDS cds;
2080: PetscScalar consts[2];
2082: PetscFunctionBegin;
2083: consts[0] = inner;
2084: consts[1] = radius;
2085: PetscCall(DMGetCoordinateDM(dm, &cdm));
2086: PetscCall(DMGetDS(cdm, &cds));
2087: PetscCall(PetscDSSetConstants(cds, 2, consts));
2088: PetscCall(DMPlexRemapGeometry(dm, 0.0, LandauSphereMapping));
2089: PetscFunctionReturn(PETSC_SUCCESS);
2090: }
2092: PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat);
2094: /*@C
2095: DMPlexLandauCreateVelocitySpace - Create a `DMPLEX` velocity space mesh
2097: Collective
2099: Input Parameters:
2100: + comm - The MPI communicator
2101: . dim - velocity space dimension (2 for axisymmetric, 3 for full 3X + 3V solver)
2102: - prefix - prefix for options (not tested)
2104: Output Parameters:
2105: + pack - The `DM` object representing the mesh
2106: . X - A vector (user destroys)
2107: - J - Optional matrix (object destroys)
2109: Level: beginner
2111: .seealso: `DMPlexCreate()`, `DMPlexLandauDestroyVelocitySpace()`
2112: @*/
2113: PetscErrorCode DMPlexLandauCreateVelocitySpace(MPI_Comm comm, PetscInt dim, const char prefix[], Vec *X, Mat *J, DM *pack)
2114: {
2115: LandauCtx *ctx;
2116: Vec Xsub[LANDAU_MAX_GRIDS];
2117: IS grid_batch_is_inv[LANDAU_MAX_GRIDS];
2119: PetscFunctionBegin;
2120: PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Only 2D and 3D supported");
2121: PetscCheck(LANDAU_DIM == dim, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " != LANDAU_DIM %d", dim, LANDAU_DIM);
2122: PetscCall(PetscNew(&ctx));
2123: ctx->comm = comm; /* used for diagnostics and global errors */
2124: /* process options */
2125: PetscCall(ProcessOptions(ctx, prefix));
2126: if (dim == 2) ctx->use_relativistic_corrections = PETSC_FALSE;
2127: /* Create Mesh */
2128: PetscCall(DMCompositeCreate(PETSC_COMM_SELF, pack));
2129: PetscCall(PetscLogEventBegin(ctx->events[13], 0, 0, 0, 0));
2130: PetscCall(PetscLogEventBegin(ctx->events[15], 0, 0, 0, 0));
2131: PetscCall(LandauDMCreateVMeshes(PETSC_COMM_SELF, dim, prefix, ctx, *pack)); // creates grids (Forest of AMR)
2132: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2133: /* create FEM */
2134: PetscCall(SetupDS(ctx->plex[grid], dim, grid, prefix, ctx));
2135: /* set initial state */
2136: PetscCall(DMCreateGlobalVector(ctx->plex[grid], &Xsub[grid]));
2137: PetscCall(PetscObjectSetName((PetscObject)Xsub[grid], "u_orig"));
2138: /* initial static refinement, no solve */
2139: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, 0, 1, ctx));
2140: /* forest refinement - forest goes in (if forest), plex comes out */
2141: if (ctx->use_p4est) {
2142: DM plex;
2143: PetscCall(adapt(grid, ctx, &Xsub[grid])); // forest goes in, plex comes out
2144: // convert to plex, all done with this level
2145: PetscCall(DMConvert(ctx->plex[grid], DMPLEX, &plex));
2146: PetscCall(DMDestroy(&ctx->plex[grid]));
2147: ctx->plex[grid] = plex;
2148: } else if (ctx->sphere && dim == 3) {
2149: if (ctx->map_sphere) PetscCall(LandauSphereMesh(ctx->plex[grid], ctx->radius[grid] * ctx->sphere_inner_radius_90degree[grid], ctx->radius[grid]));
2150: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, 0, 1, ctx));
2151: }
2152: if (grid == 0) {
2153: PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, "-dm_landau_amr_dm_view"));
2154: PetscCall(VecSetOptionsPrefix(Xsub[grid], prefix));
2155: PetscCall(VecViewFromOptions(Xsub[grid], NULL, "-dm_landau_amr_vec_view"));
2156: }
2157: #if !defined(LANDAU_SPECIES_MAJOR)
2158: PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2159: #else
2160: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
2161: PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2162: }
2163: #endif
2164: PetscCall(DMSetApplicationContext(ctx->plex[grid], ctx));
2165: }
2166: #if !defined(LANDAU_SPECIES_MAJOR)
2167: // stack the batched DMs, could do it all here!!! b_id=0
2168: for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) {
2169: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2170: }
2171: #endif
2172: // create ctx->mat_offset
2173: ctx->mat_offset[0] = 0;
2174: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2175: PetscInt n;
2176: PetscCall(VecGetLocalSize(Xsub[grid], &n));
2177: ctx->mat_offset[grid + 1] = ctx->mat_offset[grid] + n;
2178: }
2179: // creat DM & Jac
2180: PetscCall(DMSetApplicationContext(*pack, ctx));
2181: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only"));
2182: PetscCall(DMCreateMatrix(*pack, &ctx->J));
2183: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
2184: PetscCall(MatSetOption(ctx->J, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
2185: PetscCall(MatSetOption(ctx->J, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
2186: PetscCall(PetscObjectSetName((PetscObject)ctx->J, "Jac"));
2187: // construct initial conditions in X
2188: PetscCall(DMCreateGlobalVector(*pack, X));
2189: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2190: PetscInt n;
2191: PetscCall(VecGetLocalSize(Xsub[grid], &n));
2192: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2193: PetscScalar const *values;
2194: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2195: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, b_id, ctx->batch_sz, ctx));
2196: PetscCall(VecGetArrayRead(Xsub[grid], &values)); // Drop whole grid in Plex ordering
2197: for (PetscInt i = 0, idx = moffset; i < n; i++, idx++) PetscCall(VecSetValue(*X, idx, values[i], INSERT_VALUES));
2198: PetscCall(VecRestoreArrayRead(Xsub[grid], &values));
2199: }
2200: }
2201: // cleanup
2202: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(VecDestroy(&Xsub[grid]));
2203: /* check for correct matrix type */
2204: if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
2205: PetscBool flg;
2206: if (ctx->deviceType == LANDAU_KOKKOS) {
2207: PetscCall(PetscObjectTypeCompareAny((PetscObject)ctx->J, &flg, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, MATAIJKOKKOS, ""));
2208: #if PetscDefined(HAVE_KOKKOS)
2209: PetscCheck(flg, ctx->comm, PETSC_ERR_ARG_WRONG, "must use '-dm_mat_type aijkokkos -dm_vec_type kokkos' for GPU assembly and Kokkos or use '-dm_landau_device_type cpu'");
2210: #else
2211: PetscCheck(flg, ctx->comm, PETSC_ERR_ARG_WRONG, "must configure with '--download-kokkos-kernels' for GPU assembly and Kokkos or use '-dm_landau_device_type cpu'");
2212: #endif
2213: }
2214: }
2215: PetscCall(PetscLogEventEnd(ctx->events[15], 0, 0, 0, 0));
2217: // create field major ordering
2218: ctx->work_vec = NULL;
2219: ctx->plex_batch = NULL;
2220: ctx->batch_is = NULL;
2221: for (PetscInt i = 0; i < LANDAU_MAX_GRIDS; i++) grid_batch_is_inv[i] = NULL;
2222: PetscCall(PetscLogEventBegin(ctx->events[12], 0, 0, 0, 0));
2223: PetscCall(LandauCreateJacobianMatrix(comm, *X, grid_batch_is_inv, ctx));
2224: PetscCall(PetscLogEventEnd(ctx->events[12], 0, 0, 0, 0));
2226: // create AMR GPU assembly maps and static GPU data
2227: PetscCall(CreateStaticData(dim, grid_batch_is_inv, prefix, ctx));
2229: PetscCall(PetscLogEventEnd(ctx->events[13], 0, 0, 0, 0));
2231: // create mass matrix
2232: PetscCall(DMPlexLandauCreateMassMatrix(*pack, NULL));
2234: if (J) *J = ctx->J;
2236: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
2237: PetscContainer container;
2238: // cache ctx for KSP with batch/field major Jacobian ordering -ksp_type gmres/etc -dm_landau_jacobian_field_major_order
2239: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2240: PetscCall(PetscContainerSetPointer(container, (void *)ctx));
2241: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "LandauCtx", (PetscObject)container));
2242: PetscCall(PetscContainerDestroy(&container));
2243: // batch solvers need to map -- can batch solvers work
2244: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2245: PetscCall(PetscContainerSetPointer(container, (void *)ctx->plex_batch));
2246: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "plex_batch_is", (PetscObject)container));
2247: PetscCall(PetscContainerDestroy(&container));
2248: }
2249: // for batch solvers
2250: {
2251: PetscContainer container;
2252: PetscInt *pNf;
2253: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2254: PetscCall(PetscMalloc1(sizeof(*pNf), &pNf));
2255: *pNf = ctx->batch_sz;
2256: PetscCall(PetscContainerSetPointer(container, (void *)pNf));
2257: PetscCall(PetscContainerSetCtxDestroy(container, PetscCtxDestroyDefault));
2258: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "batch size", (PetscObject)container));
2259: PetscCall(PetscContainerDestroy(&container));
2260: }
2261: PetscFunctionReturn(PETSC_SUCCESS);
2262: }
2264: /*@C
2265: DMPlexLandauAccess - Access to the distribution function with user callback
2267: Collective
2269: Input Parameters:
2270: + pack - the `DMCOMPOSITE`
2271: . func - call back function
2272: - user_ctx - application context
2274: Input/Output Parameter:
2275: . X - Vector to data to
2277: Level: advanced
2279: .seealso: `DMPlexLandauCreateVelocitySpace()`
2280: @*/
2281: PetscErrorCode DMPlexLandauAccess(DM pack, Vec X, PetscErrorCode (*func)(DM, Vec, PetscInt, PetscInt, PetscInt, void *), void *user_ctx)
2282: {
2283: LandauCtx *ctx;
2285: PetscFunctionBegin;
2286: PetscCall(DMGetApplicationContext(pack, &ctx)); // uses ctx->num_grids; ctx->plex[grid]; ctx->batch_sz; ctx->mat_offset
2287: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2288: PetscInt dim, n;
2289: PetscCall(DMGetDimension(pack, &dim));
2290: for (PetscInt sp = ctx->species_offset[grid], i0 = 0; sp < ctx->species_offset[grid + 1]; sp++, i0++) {
2291: Vec vec;
2292: PetscInt vf[1] = {i0};
2293: IS vis;
2294: DM vdm;
2295: PetscCall(DMCreateSubDM(ctx->plex[grid], 1, vf, &vis, &vdm));
2296: PetscCall(DMSetApplicationContext(vdm, ctx)); // the user might want this
2297: PetscCall(DMCreateGlobalVector(vdm, &vec));
2298: PetscCall(VecGetSize(vec, &n));
2299: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2300: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2301: PetscCall(VecZeroEntries(vec));
2302: /* Add your data with 'dm' for species 'sp' to 'vec' */
2303: PetscCall(func(vdm, vec, i0, grid, b_id, user_ctx));
2304: /* add to global */
2305: PetscScalar const *values;
2306: const PetscInt *offsets;
2307: PetscCall(VecGetArrayRead(vec, &values));
2308: PetscCall(ISGetIndices(vis, &offsets));
2309: for (PetscInt i = 0; i < n; i++) PetscCall(VecSetValue(X, moffset + offsets[i], values[i], ADD_VALUES));
2310: PetscCall(VecRestoreArrayRead(vec, &values));
2311: PetscCall(ISRestoreIndices(vis, &offsets));
2312: } // batch
2313: PetscCall(VecDestroy(&vec));
2314: PetscCall(ISDestroy(&vis));
2315: PetscCall(DMDestroy(&vdm));
2316: }
2317: } // grid
2318: PetscFunctionReturn(PETSC_SUCCESS);
2319: }
2321: /*@
2322: DMPlexLandauDestroyVelocitySpace - Destroy a `DMPLEX` velocity space mesh
2324: Collective
2326: Input/Output Parameters:
2327: . dm - the `DM` to destroy
2329: Level: beginner
2331: .seealso: `DMPlexLandauCreateVelocitySpace()`
2332: @*/
2333: PetscErrorCode DMPlexLandauDestroyVelocitySpace(DM *dm)
2334: {
2335: LandauCtx *ctx;
2337: PetscFunctionBegin;
2338: PetscCall(DMGetApplicationContext(*dm, &ctx));
2339: PetscCall(MatDestroy(&ctx->M));
2340: PetscCall(MatDestroy(&ctx->J));
2341: for (PetscInt ii = 0; ii < ctx->num_species; ii++) PetscCall(PetscFEDestroy(&ctx->fe[ii]));
2342: PetscCall(ISDestroy(&ctx->batch_is));
2343: PetscCall(VecDestroy(&ctx->work_vec));
2344: PetscCall(VecScatterDestroy(&ctx->plex_batch));
2345: if (ctx->deviceType == LANDAU_KOKKOS) {
2346: #if PetscDefined(HAVE_KOKKOS)
2347: PetscCall(LandauKokkosStaticDataClear(&ctx->SData_d));
2348: #else
2349: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos");
2350: #endif
2351: } else {
2352: if (ctx->SData_d.x) { /* in a CPU run */
2353: PetscReal *invJ = (PetscReal *)ctx->SData_d.invJ, *xx = (PetscReal *)ctx->SData_d.x, *yy = (PetscReal *)ctx->SData_d.y, *zz = (PetscReal *)ctx->SData_d.z, *ww = (PetscReal *)ctx->SData_d.w;
2354: LandauIdx *coo_elem_offsets = (LandauIdx *)ctx->SData_d.coo_elem_offsets, *coo_elem_fullNb = (LandauIdx *)ctx->SData_d.coo_elem_fullNb, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = (LandauIdx(*)[LANDAU_MAX_NQND + 1]) ctx->SData_d.coo_elem_point_offsets;
2355: PetscCall(PetscFree4(ww, xx, yy, invJ));
2356: PetscCall(PetscFree(zz));
2357: if (coo_elem_offsets) PetscCall(PetscFree3(coo_elem_offsets, coo_elem_fullNb, coo_elem_point_offsets)); // could be NULL
2358: PetscCall(PetscFree4(ctx->SData_d.alpha, ctx->SData_d.beta, ctx->SData_d.invMass, ctx->SData_d.lambdas));
2359: }
2360: }
2362: if (ctx->times[LANDAU_MATRIX_TOTAL] > 0) { // OMP timings
2363: PetscCall(PetscPrintf(ctx->comm, "TSStep N 1.0 %10.3e\n", ctx->times[LANDAU_EX2_TSSOLVE]));
2364: PetscCall(PetscPrintf(ctx->comm, "2: Solve: %10.3e with %" PetscInt_FMT " threads\n", ctx->times[LANDAU_EX2_TSSOLVE] - ctx->times[LANDAU_MATRIX_TOTAL], ctx->batch_sz));
2365: PetscCall(PetscPrintf(ctx->comm, "3: Landau: %10.3e\n", ctx->times[LANDAU_MATRIX_TOTAL]));
2366: PetscCall(PetscPrintf(ctx->comm, "Landau Jacobian %" PetscInt_FMT " 1.0 %10.3e\n", (PetscInt)ctx->times[LANDAU_JACOBIAN_COUNT], ctx->times[LANDAU_JACOBIAN]));
2367: PetscCall(PetscPrintf(ctx->comm, "Landau Operator N 1.0 %10.3e\n", ctx->times[LANDAU_OPERATOR]));
2368: PetscCall(PetscPrintf(ctx->comm, "Landau Mass N 1.0 %10.3e\n", ctx->times[LANDAU_MASS]));
2369: PetscCall(PetscPrintf(ctx->comm, " Jac-f-df (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_F_DF]));
2370: PetscCall(PetscPrintf(ctx->comm, " Kernel (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_KERNEL]));
2371: PetscCall(PetscPrintf(ctx->comm, "MatLUFactorNum X 1.0 %10.3e\n", ctx->times[KSP_FACTOR]));
2372: PetscCall(PetscPrintf(ctx->comm, "MatSolve X 1.0 %10.3e\n", ctx->times[KSP_SOLVE]));
2373: }
2374: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMDestroy(&ctx->plex[grid]));
2375: PetscCall(PetscFree(ctx));
2376: PetscCall(DMDestroy(dm));
2377: PetscFunctionReturn(PETSC_SUCCESS);
2378: }
2380: /* < v, ru > */
2381: static void f0_s_den(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0)
2382: {
2383: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2384: f0[0] = u[ii];
2385: }
2387: /* < v, ru > */
2388: static void f0_s_mom(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0)
2389: {
2390: PetscInt ii = (PetscInt)PetscRealPart(constants[0]), jj = (PetscInt)PetscRealPart(constants[1]);
2391: f0[0] = x[jj] * u[ii]; /* x momentum */
2392: }
2394: static void f0_s_v2(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0)
2395: {
2396: PetscInt i, ii = (PetscInt)PetscRealPart(constants[0]);
2397: double tmp1 = 0.;
2398: for (i = 0; i < dim; ++i) tmp1 += x[i] * x[i];
2399: f0[0] = tmp1 * u[ii];
2400: }
2402: static PetscErrorCode gamma_n_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *u, void *actx)
2403: {
2404: const PetscReal *c2_0_arr = ((PetscReal *)actx);
2405: const PetscReal c02 = c2_0_arr[0];
2407: PetscFunctionBegin;
2408: for (PetscInt s = 0; s < Nf; s++) {
2409: PetscReal tmp1 = 0.;
2410: for (PetscInt i = 0; i < dim; ++i) tmp1 += x[i] * x[i];
2411: if (PetscDefined(USE_DEBUG)) u[s] = PetscSqrtReal(1. + tmp1 / c02); // u[0] = PetscSqrtReal(1. + xx);
2412: else {
2413: PetscReal xx = tmp1 / c02;
2414: u[s] = xx / (PetscSqrtReal(1. + xx) + 1.); // better conditioned = xx/(PetscSqrtReal(1. + xx) + 1.)
2415: }
2416: }
2417: PetscFunctionReturn(PETSC_SUCCESS);
2418: }
2420: /* < v, ru > */
2421: static void f0_s_rden(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0)
2422: {
2423: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2424: f0[0] = 2. * PETSC_PI * x[0] * u[ii];
2425: }
2427: /* < v, ru > */
2428: static void f0_s_rmom(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0)
2429: {
2430: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2431: f0[0] = 2. * PETSC_PI * x[0] * x[1] * u[ii];
2432: }
2434: static void f0_s_rv2(PetscInt dim, PetscInt Nf, PetscInt NfAux, const PetscInt uOff[], const PetscInt uOff_x[], const PetscScalar u[], const PetscScalar u_t[], const PetscScalar u_x[], const PetscInt aOff[], const PetscInt aOff_x[], const PetscScalar a[], const PetscScalar a_t[], const PetscScalar a_x[], PetscReal t, const PetscReal x[], PetscInt numConstants, const PetscScalar constants[], PetscScalar *f0)
2435: {
2436: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2437: f0[0] = 2. * PETSC_PI * x[0] * (x[0] * x[0] + x[1] * x[1]) * u[ii];
2438: }
2440: /*@
2441: DMPlexLandauPrintNorms - collects moments and prints them
2443: Collective
2445: Input Parameters:
2446: + X - the state
2447: - stepi - current step to print
2449: Level: beginner
2451: .seealso: `DMPlexLandauCreateVelocitySpace()`
2452: @*/
2453: PetscErrorCode DMPlexLandauPrintNorms(Vec X, PetscInt stepi)
2454: {
2455: LandauCtx *ctx;
2456: PetscDS prob;
2457: DM pack;
2458: PetscInt cStart, cEnd, dim, ii, i0, nDMs;
2459: PetscScalar xmomentumtot = 0, ymomentumtot = 0, zmomentumtot = 0, energytot = 0, densitytot = 0, tt[LANDAU_MAX_SPECIES];
2460: PetscScalar xmomentum[LANDAU_MAX_SPECIES], ymomentum[LANDAU_MAX_SPECIES], zmomentum[LANDAU_MAX_SPECIES], energy[LANDAU_MAX_SPECIES], density[LANDAU_MAX_SPECIES];
2461: Vec *globXArray;
2463: PetscFunctionBegin;
2464: PetscCall(VecGetDM(X, &pack));
2465: PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Vector has no DM");
2466: PetscCall(DMGetDimension(pack, &dim));
2467: PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " not in [2,3]", dim);
2468: PetscCall(DMGetApplicationContext(pack, &ctx));
2469: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2470: /* print momentum and energy */
2471: PetscCall(DMCompositeGetNumberDM(pack, &nDMs));
2472: PetscCheck(nDMs == ctx->num_grids * ctx->batch_sz, PETSC_COMM_WORLD, PETSC_ERR_PLIB, "#DM wrong %" PetscInt_FMT " %" PetscInt_FMT, nDMs, ctx->num_grids * ctx->batch_sz);
2473: PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray));
2474: PetscCall(DMCompositeGetAccessArray(pack, X, nDMs, NULL, globXArray));
2475: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2476: Vec Xloc = globXArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)];
2477: PetscCall(DMGetDS(ctx->plex[grid], &prob));
2478: for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
2479: PetscScalar user[2] = {(PetscScalar)i0, ctx->charges[ii]};
2480: PetscCall(PetscDSSetConstants(prob, 2, user));
2481: if (dim == 2) { /* 2/3X + 3V (cylindrical coordinates) */
2482: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rden));
2483: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2484: density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii];
2485: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rmom));
2486: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2487: zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2488: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rv2));
2489: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2490: energy[ii] = tt[0] * 0.5 * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii];
2491: zmomentumtot += zmomentum[ii];
2492: energytot += energy[ii];
2493: densitytot += density[ii];
2494: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%3" PetscInt_FMT ") species-%" PetscInt_FMT ": charge density= %20.13e z-momentum= %20.13e energy= %20.13e", stepi, ii, (double)PetscRealPart(density[ii]), (double)PetscRealPart(zmomentum[ii]), (double)PetscRealPart(energy[ii])));
2495: } else { /* 2/3Xloc + 3V */
2496: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_den));
2497: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2498: density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii];
2499: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_mom));
2500: user[1] = 0;
2501: PetscCall(PetscDSSetConstants(prob, 2, user));
2502: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2503: xmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2504: user[1] = 1;
2505: PetscCall(PetscDSSetConstants(prob, 2, user));
2506: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2507: ymomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2508: user[1] = 2;
2509: PetscCall(PetscDSSetConstants(prob, 2, user));
2510: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2511: zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2512: if (ctx->use_relativistic_corrections) {
2513: /* gamma * M * f */
2514: if (ii == 0 && grid == 0) { // do all at once
2515: Vec Mf, globGamma, *globMfArray, *globGammaArray;
2516: PetscErrorCode (*gammaf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {gamma_n_f};
2517: PetscReal *c2_0[1], data[1];
2519: PetscCall(VecDuplicate(X, &globGamma));
2520: PetscCall(VecDuplicate(X, &Mf));
2521: PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globMfArray));
2522: PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globGammaArray));
2523: /* M * f */
2524: PetscCall(MatMult(ctx->M, X, Mf));
2525: /* gamma */
2526: PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2527: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { // yes a grid loop in a grid loop to print nice, need to fix for batching
2528: Vec v1 = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)];
2529: data[0] = PetscSqr(C_0(ctx->v_0));
2530: c2_0[0] = &data[0];
2531: PetscCall(DMProjectFunction(ctx->plex[grid], 0., gammaf, (void **)c2_0, INSERT_ALL_VALUES, v1));
2532: }
2533: PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2534: /* gamma * Mf */
2535: PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2536: PetscCall(DMCompositeGetAccessArray(pack, Mf, nDMs, NULL, globMfArray));
2537: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { // yes a grid loop in a grid loop to print nice
2538: PetscInt Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], N, bs;
2539: Vec Mfsub = globMfArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], Gsub = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], v1, v2;
2540: // get each component
2541: PetscCall(VecGetSize(Mfsub, &N));
2542: PetscCall(VecCreate(ctx->comm, &v1));
2543: PetscCall(VecSetSizes(v1, PETSC_DECIDE, N / Nf));
2544: PetscCall(VecCreate(ctx->comm, &v2));
2545: PetscCall(VecSetSizes(v2, PETSC_DECIDE, N / Nf));
2546: PetscCall(VecSetFromOptions(v1)); // ???
2547: PetscCall(VecSetFromOptions(v2));
2548: // get each component
2549: PetscCall(VecGetBlockSize(Gsub, &bs));
2550: PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT " in Gsub", bs, Nf);
2551: PetscCall(VecGetBlockSize(Mfsub, &bs));
2552: PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT, bs, Nf);
2553: for (PetscInt i = 0, ix = ctx->species_offset[grid]; i < Nf; i++, ix++) {
2554: PetscScalar val;
2555: PetscCall(VecStrideGather(Gsub, i, v1, INSERT_VALUES)); // this is not right -- TODO
2556: PetscCall(VecStrideGather(Mfsub, i, v2, INSERT_VALUES));
2557: PetscCall(VecDot(v1, v2, &val));
2558: energy[ix] = PetscRealPart(val) * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ix];
2559: }
2560: PetscCall(VecDestroy(&v1));
2561: PetscCall(VecDestroy(&v2));
2562: } /* grids */
2563: PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2564: PetscCall(DMCompositeRestoreAccessArray(pack, Mf, nDMs, NULL, globMfArray));
2565: PetscCall(PetscFree(globGammaArray));
2566: PetscCall(PetscFree(globMfArray));
2567: PetscCall(VecDestroy(&globGamma));
2568: PetscCall(VecDestroy(&Mf));
2569: }
2570: } else {
2571: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_v2));
2572: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2573: energy[ii] = 0.5 * tt[0] * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii];
2574: }
2575: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%3" PetscInt_FMT ") species %" PetscInt_FMT ": density=%20.13e, x-momentum=%20.13e, y-momentum=%20.13e, z-momentum=%20.13e, energy=%21.13e", stepi, ii, (double)PetscRealPart(density[ii]), (double)PetscRealPart(xmomentum[ii]), (double)PetscRealPart(ymomentum[ii]), (double)PetscRealPart(zmomentum[ii]), (double)PetscRealPart(energy[ii])));
2576: xmomentumtot += xmomentum[ii];
2577: ymomentumtot += ymomentum[ii];
2578: zmomentumtot += zmomentum[ii];
2579: energytot += energy[ii];
2580: densitytot += density[ii];
2581: }
2582: if (ctx->num_species > 1) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
2583: }
2584: }
2585: PetscCall(DMCompositeRestoreAccessArray(pack, X, nDMs, NULL, globXArray));
2586: PetscCall(PetscFree(globXArray));
2587: /* totals */
2588: PetscCall(DMPlexGetHeightStratum(ctx->plex[0], 0, &cStart, &cEnd));
2589: if (ctx->num_species > 1) {
2590: if (dim == 2) {
2591: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\t%3" PetscInt_FMT ") Total: charge density=%21.13e, momentum=%21.13e, energy=%21.13e (m_i[0]/m_e = %g, %" PetscInt_FMT " cells on electron grid)", stepi, (double)PetscRealPart(densitytot), (double)PetscRealPart(zmomentumtot), (double)PetscRealPart(energytot),
2592: (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart));
2593: } else {
2594: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\t%3" PetscInt_FMT ") Total: charge density=%21.13e, x-momentum=%21.13e, y-momentum=%21.13e, z-momentum=%21.13e, energy=%21.13e (m_i[0]/m_e = %g, %" PetscInt_FMT " cells)", stepi, (double)PetscRealPart(densitytot), (double)PetscRealPart(xmomentumtot), (double)PetscRealPart(ymomentumtot), (double)PetscRealPart(zmomentumtot), (double)PetscRealPart(energytot),
2595: (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart));
2596: }
2597: } else PetscCall(PetscPrintf(PETSC_COMM_WORLD, " -- %" PetscInt_FMT " cells", cEnd - cStart));
2598: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
2599: PetscFunctionReturn(PETSC_SUCCESS);
2600: }
2602: /*@
2603: DMPlexLandauCreateMassMatrix - Create mass matrix for Landau in Plex space (not field major order of Jacobian)
2604: - puts mass matrix into ctx->M
2606: Collective
2608: Input Parameter:
2609: . pack - the `DM` object. Puts matrix in Landau context M field
2611: Output Parameter:
2612: . Amat - The mass matrix (optional), mass matrix is added to the `DM` context
2614: Level: beginner
2616: .seealso: `DMPlexLandauCreateVelocitySpace()`
2617: @*/
2618: PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat)
2619: {
2620: DM mass_pack, massDM[LANDAU_MAX_GRIDS];
2621: PetscDS prob;
2622: PetscInt ii, dim, N1 = 1, N2;
2623: LandauCtx *ctx;
2624: Mat packM, subM[LANDAU_MAX_GRIDS];
2626: PetscFunctionBegin;
2628: if (Amat) PetscAssertPointer(Amat, 2);
2629: PetscCall(DMGetApplicationContext(pack, &ctx));
2630: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2631: PetscCall(PetscLogEventBegin(ctx->events[14], 0, 0, 0, 0));
2632: PetscCall(DMGetDimension(pack, &dim));
2633: PetscCall(DMCompositeCreate(PetscObjectComm((PetscObject)pack), &mass_pack));
2634: /* create pack mass matrix */
2635: for (PetscInt grid = 0, ix = 0; grid < ctx->num_grids; grid++) {
2636: PetscCall(DMClone(ctx->plex[grid], &massDM[grid]));
2637: PetscCall(DMCopyFields(ctx->plex[grid], PETSC_DETERMINE, PETSC_DETERMINE, massDM[grid]));
2638: PetscCall(DMCreateDS(massDM[grid]));
2639: PetscCall(DMGetDS(massDM[grid], &prob));
2640: for (ix = 0, ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++, ix++) {
2641: if (dim == 3) PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_1, NULL, NULL, NULL));
2642: else PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_r, NULL, NULL, NULL));
2643: }
2644: #if !defined(LANDAU_SPECIES_MAJOR)
2645: PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2646: #else
2647: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
2648: PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2649: }
2650: #endif
2651: PetscCall(DMCreateMatrix(massDM[grid], &subM[grid]));
2652: }
2653: #if !defined(LANDAU_SPECIES_MAJOR)
2654: // stack the batched DMs
2655: for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) {
2656: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2657: }
2658: #endif
2659: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only"));
2660: PetscCall(DMCreateMatrix(mass_pack, &packM));
2661: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
2662: PetscCall(MatSetOption(packM, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
2663: PetscCall(MatSetOption(packM, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
2664: PetscCall(DMDestroy(&mass_pack));
2665: /* make mass matrix for each block */
2666: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2667: Vec locX;
2668: DM plex = massDM[grid];
2669: PetscCall(DMGetLocalVector(plex, &locX));
2670: /* Mass matrix is independent of the input, so no need to fill locX */
2671: PetscCall(DMPlexSNESComputeJacobianFEM(plex, locX, subM[grid], subM[grid], ctx));
2672: PetscCall(DMRestoreLocalVector(plex, &locX));
2673: PetscCall(DMDestroy(&massDM[grid]));
2674: }
2675: PetscCall(MatGetSize(ctx->J, &N1, NULL));
2676: PetscCall(MatGetSize(packM, &N2, NULL));
2677: PetscCheck(N1 == N2, PetscObjectComm((PetscObject)pack), PETSC_ERR_PLIB, "Incorrect matrix sizes: |Jacobian| = %" PetscInt_FMT ", |Mass|=%" PetscInt_FMT, N1, N2);
2678: /* assemble block diagonals */
2679: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2680: Mat B = subM[grid];
2681: PetscInt nloc, nzl, *colbuf, COL_BF_SIZE = 1024, row;
2682: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2683: PetscCall(MatGetSize(B, &nloc, NULL));
2684: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2685: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2686: const PetscInt *cols;
2687: const PetscScalar *vals;
2688: for (PetscInt i = 0; i < nloc; i++) {
2689: PetscCall(MatGetRow(B, i, &nzl, NULL, NULL));
2690: if (nzl > COL_BF_SIZE) {
2691: PetscCall(PetscFree(colbuf));
2692: PetscCall(PetscInfo(pack, "Realloc buffer %" PetscInt_FMT " to %" PetscInt_FMT " (row size %" PetscInt_FMT ") \n", COL_BF_SIZE, 2 * COL_BF_SIZE, nzl));
2693: COL_BF_SIZE = nzl;
2694: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2695: }
2696: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
2697: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset;
2698: row = i + moffset;
2699: PetscCall(MatSetValues(packM, 1, &row, nzl, colbuf, vals, INSERT_VALUES));
2700: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
2701: }
2702: }
2703: PetscCall(PetscFree(colbuf));
2704: }
2705: // cleanup
2706: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid]));
2707: PetscCall(MatAssemblyBegin(packM, MAT_FINAL_ASSEMBLY));
2708: PetscCall(MatAssemblyEnd(packM, MAT_FINAL_ASSEMBLY));
2709: PetscCall(PetscObjectSetName((PetscObject)packM, "mass"));
2710: PetscCall(MatViewFromOptions(packM, NULL, "-dm_landau_mass_view"));
2711: ctx->M = packM;
2712: if (Amat) *Amat = packM;
2713: PetscCall(PetscLogEventEnd(ctx->events[14], 0, 0, 0, 0));
2714: PetscFunctionReturn(PETSC_SUCCESS);
2715: }
2717: /*@
2718: DMPlexLandauIFunction - `TS` residual calculation, confusingly this computes the Jacobian w/o mass
2720: Collective
2722: Input Parameters:
2723: + ts - The time stepping context
2724: . time_dummy - current time (not used)
2725: . X - Current state
2726: . X_t - Time derivative of current state
2727: - actx - Landau context
2729: Output Parameter:
2730: . F - The residual
2732: Level: beginner
2734: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIJacobian()`
2735: @*/
2736: PetscErrorCode DMPlexLandauIFunction(TS ts, PetscReal time_dummy, Vec X, Vec X_t, Vec F, void *actx)
2737: {
2738: LandauCtx *ctx = (LandauCtx *)actx;
2739: PetscInt dim;
2740: DM pack;
2741: #if PetscDefined(HAVE_THREADSAFETY)
2742: double starttime, endtime;
2743: #endif
2744: PetscObjectState state;
2746: PetscFunctionBegin;
2747: PetscCall(TSGetDM(ts, &pack));
2748: PetscCall(DMGetApplicationContext(pack, &ctx));
2749: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2750: if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage));
2751: PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0));
2752: PetscCall(PetscLogEventBegin(ctx->events[0], 0, 0, 0, 0));
2753: #if PetscDefined(HAVE_THREADSAFETY)
2754: starttime = MPI_Wtime();
2755: #endif
2756: PetscCall(DMGetDimension(pack, &dim));
2757: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2758: if (state != ctx->norm_state) {
2759: PetscCall(MatZeroEntries(ctx->J));
2760: PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, 0.0, (void *)ctx));
2761: PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_jacobian_view"));
2762: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2763: ctx->norm_state = state;
2764: } else {
2765: PetscCall(PetscInfo(ts, "WARNING Skip forming Jacobian, has not changed %" PetscInt64_FMT "\n", state));
2766: }
2767: /* mat vec for op */
2768: PetscCall(MatMult(ctx->J, X, F)); /* C*f */
2769: /* add time term */
2770: if (X_t) PetscCall(MatMultAdd(ctx->M, X_t, F, F));
2771: #if PetscDefined(HAVE_THREADSAFETY)
2772: if (ctx->stage) {
2773: endtime = MPI_Wtime();
2774: ctx->times[LANDAU_OPERATOR] += (endtime - starttime);
2775: ctx->times[LANDAU_JACOBIAN] += (endtime - starttime);
2776: ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime);
2777: ctx->times[LANDAU_JACOBIAN_COUNT] += 1;
2778: }
2779: #endif
2780: PetscCall(PetscLogEventEnd(ctx->events[0], 0, 0, 0, 0));
2781: PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0));
2782: if (ctx->stage) PetscCall(PetscLogStagePop());
2783: PetscFunctionReturn(PETSC_SUCCESS);
2784: }
2786: /*@
2787: DMPlexLandauIJacobian - `TS` Jacobian construction, confusingly this adds mass
2789: Collective
2791: Input Parameters:
2792: + ts - The time stepping context
2793: . time_dummy - current time (not used)
2794: . X - Current state
2795: . U_tdummy - Time derivative of current state (not used)
2796: . shift - shift for du/dt term
2797: - actx - Landau context
2799: Output Parameters:
2800: + Amat - Jacobian
2801: - Pmat - same as Amat
2803: Level: beginner
2805: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIFunction()`
2806: @*/
2807: PetscErrorCode DMPlexLandauIJacobian(TS ts, PetscReal time_dummy, Vec X, Vec U_tdummy, PetscReal shift, Mat Amat, Mat Pmat, void *actx)
2808: {
2809: LandauCtx *ctx = NULL;
2810: PetscInt dim;
2811: DM pack;
2812: #if PetscDefined(HAVE_THREADSAFETY)
2813: double starttime, endtime;
2814: #endif
2815: PetscObjectState state;
2817: PetscFunctionBegin;
2818: PetscCall(TSGetDM(ts, &pack));
2819: PetscCall(DMGetApplicationContext(pack, &ctx));
2820: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2821: PetscCheck(Amat == Pmat && Amat == ctx->J, ctx->comm, PETSC_ERR_PLIB, "Amat!=Pmat || Amat!=ctx->J");
2822: PetscCall(DMGetDimension(pack, &dim));
2823: /* get collision Jacobian into A */
2824: if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage));
2825: PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0));
2826: PetscCall(PetscLogEventBegin(ctx->events[9], 0, 0, 0, 0));
2827: #if PetscDefined(HAVE_THREADSAFETY)
2828: starttime = MPI_Wtime();
2829: #endif
2830: PetscCheck(shift != 0.0, ctx->comm, PETSC_ERR_PLIB, "zero shift");
2831: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2832: PetscCheck(state == ctx->norm_state, ctx->comm, PETSC_ERR_PLIB, "wrong state, %" PetscInt64_FMT " %" PetscInt64_FMT, ctx->norm_state, state);
2833: if (!ctx->use_matrix_mass) {
2834: PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, shift, (void *)ctx));
2835: } else { /* add mass */
2836: PetscCall(MatAXPY(Pmat, shift, ctx->M, SAME_NONZERO_PATTERN));
2837: }
2838: #if PetscDefined(HAVE_THREADSAFETY)
2839: if (ctx->stage) {
2840: endtime = MPI_Wtime();
2841: ctx->times[LANDAU_OPERATOR] += (endtime - starttime);
2842: ctx->times[LANDAU_MASS] += (endtime - starttime);
2843: ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime);
2844: }
2845: #endif
2846: PetscCall(PetscLogEventEnd(ctx->events[9], 0, 0, 0, 0));
2847: PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0));
2848: if (ctx->stage) PetscCall(PetscLogStagePop());
2849: PetscFunctionReturn(PETSC_SUCCESS);
2850: }