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 defined(PETSC_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 defined(PETSC_HAVE_OPENMP)
22: #include <omp.h>
23: #endif
25: static PetscErrorCode LandauGPUMapsDestroy(void *ptr)
26: {
27: P4estVertexMaps *maps = (P4estVertexMaps *)ptr;
29: PetscFunctionBegin;
30: // free device data
31: if (maps[0].deviceType != LANDAU_CPU) {
32: #if defined(PETSC_HAVE_KOKKOS)
33: if (maps[0].deviceType == LANDAU_KOKKOS) {
34: PetscCall(LandauKokkosDestroyMatMaps(maps, maps[0].numgrids)); // implies Kokkos does
35: }
36: #endif
37: }
38: // free host data
39: for (PetscInt grid = 0; grid < maps[0].numgrids; grid++) {
40: PetscCall(PetscFree(maps[grid].c_maps));
41: PetscCall(PetscFree(maps[grid].gIdx));
42: }
43: PetscCall(PetscFree(maps));
44: PetscFunctionReturn(PETSC_SUCCESS);
45: }
46: static PetscErrorCode energy_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
47: {
48: PetscReal v2 = 0;
50: PetscFunctionBegin;
51: /* compute v^2 / 2 */
52: for (PetscInt i = 0; i < dim; ++i) v2 += x[i] * x[i];
53: /* evaluate the Maxwellian */
54: u[0] = v2 / 2;
55: PetscFunctionReturn(PETSC_SUCCESS);
56: }
58: /* needs double */
59: static PetscErrorCode gamma_m1_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
60: {
61: PetscReal *c2_0_arr = ((PetscReal *)actx);
62: double u2 = 0, c02 = (double)*c2_0_arr, xx;
64: PetscFunctionBegin;
65: /* compute u^2 / 2 */
66: for (PetscInt i = 0; i < dim; ++i) u2 += x[i] * x[i];
67: /* gamma - 1 = g_eps, for conditioning and we only take derivatives */
68: xx = u2 / c02;
69: #if defined(PETSC_USE_DEBUG)
70: u[0] = PetscSqrtReal(1. + xx);
71: #else
72: u[0] = xx / (PetscSqrtReal(1. + xx) + 1.) - 1.; // better conditioned. -1 might help condition and only used for derivative
73: #endif
74: PetscFunctionReturn(PETSC_SUCCESS);
75: }
77: /*
78: LandauFormJacobian_Internal - Evaluates Jacobian matrix.
80: Input Parameters:
81: . globX - input vector
82: . actx - optional user-defined context
83: . dim - dimension
85: Output Parameter:
86: . J0acP - Jacobian matrix filled, not created
87: */
88: static PetscErrorCode LandauFormJacobian_Internal(Vec a_X, Mat JacP, const PetscInt dim, PetscReal shift, void *a_ctx)
89: {
90: LandauCtx *ctx = (LandauCtx *)a_ctx;
91: PetscInt numCells[LANDAU_MAX_GRIDS], Nq, Nb;
92: PetscQuadrature quad;
93: PetscReal Eq_m[LANDAU_MAX_SPECIES]; // could be static data w/o quench (ex2)
94: PetscScalar *cellClosure = NULL;
95: const PetscScalar *xdata = NULL;
96: PetscDS prob;
97: PetscContainer container;
98: P4estVertexMaps *maps;
99: Mat subJ[LANDAU_MAX_GRIDS * LANDAU_MAX_BATCH_SZ];
101: PetscFunctionBegin;
104: PetscAssertPointer(ctx, 5);
105: /* check for matrix container for GPU assembly. Support CPU assembly for debugging */
106: PetscCheck(ctx->plex[0] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
107: PetscCall(PetscLogEventBegin(ctx->events[10], 0, 0, 0, 0));
108: PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids
109: PetscCall(PetscObjectQuery((PetscObject)JacP, "assembly_maps", (PetscObject *)&container));
110: if (container) {
111: PetscCheck(ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "maps but no GPU assembly");
112: PetscCall(PetscContainerGetPointer(container, (void **)&maps));
113: PetscCheck(maps, ctx->comm, PETSC_ERR_ARG_WRONG, "empty GPU matrix container");
114: for (PetscInt i = 0; i < ctx->num_grids * ctx->batch_sz; i++) subJ[i] = NULL;
115: } else {
116: PetscCheck(!ctx->gpu_assembly, ctx->comm, PETSC_ERR_ARG_WRONG, "No maps but GPU assembly");
117: for (PetscInt tid = 0; tid < ctx->batch_sz; tid++) {
118: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCreateMatrix(ctx->plex[grid], &subJ[LAND_PACK_IDX(tid, grid)]));
119: }
120: maps = NULL;
121: }
122: // 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)
123: PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad));
124: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL));
125: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
126: 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);
127: 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);
128: // get metadata for collecting dynamic data
129: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
130: PetscInt cStart, cEnd;
131: PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
132: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
133: numCells[grid] = cEnd - cStart; // grids can have different topology
134: }
135: PetscCall(PetscLogEventEnd(ctx->events[10], 0, 0, 0, 0));
136: 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 */
137: DM pack;
138: PetscCall(VecGetDM(a_X, &pack));
139: PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "pack has no DM");
140: PetscCall(PetscLogEventBegin(ctx->events[1], 0, 0, 0, 0));
141: for (PetscInt fieldA = 0; fieldA < ctx->num_species; fieldA++) {
142: Eq_m[fieldA] = ctx->Ez * ctx->t_0 * ctx->charges[fieldA] / (ctx->v_0 * ctx->masses[fieldA]); /* normalize dimensionless */
143: if (dim == 2) Eq_m[fieldA] *= 2 * PETSC_PI; /* add the 2pi term that is not in Landau */
144: }
145: if (!ctx->gpu_assembly) {
146: Vec *locXArray, *globXArray;
147: PetscScalar *cellClosure_it;
148: PetscInt cellClosure_sz = 0, nDMs, Nf[LANDAU_MAX_GRIDS];
149: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
150: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
151: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
152: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
153: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
154: }
155: /* count cellClosure size */
156: PetscCall(DMCompositeGetNumberDM(pack, &nDMs));
157: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) cellClosure_sz += Nb * Nf[grid] * numCells[grid];
158: PetscCall(PetscMalloc1(cellClosure_sz * ctx->batch_sz, &cellClosure));
159: cellClosure_it = cellClosure;
160: PetscCall(PetscMalloc(sizeof(*locXArray) * nDMs, &locXArray));
161: PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray));
162: PetscCall(DMCompositeGetLocalAccessArray(pack, a_X, nDMs, NULL, locXArray));
163: PetscCall(DMCompositeGetAccessArray(pack, a_X, nDMs, NULL, globXArray));
164: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP (once)
165: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
166: Vec locX = locXArray[LAND_PACK_IDX(b_id, grid)], globX = globXArray[LAND_PACK_IDX(b_id, grid)], locX2;
167: PetscInt cStart, cEnd, ei;
168: PetscCall(VecDuplicate(locX, &locX2));
169: PetscCall(DMGlobalToLocalBegin(ctx->plex[grid], globX, INSERT_VALUES, locX2));
170: PetscCall(DMGlobalToLocalEnd(ctx->plex[grid], globX, INSERT_VALUES, locX2));
171: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
172: for (ei = cStart; ei < cEnd; ++ei) {
173: PetscScalar *coef = NULL;
174: PetscCall(DMPlexVecGetClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef));
175: PetscCall(PetscMemcpy(cellClosure_it, coef, Nb * Nf[grid] * sizeof(*cellClosure_it))); /* change if LandauIPReal != PetscScalar */
176: PetscCall(DMPlexVecRestoreClosure(ctx->plex[grid], section[grid], locX2, ei, NULL, &coef));
177: cellClosure_it += Nb * Nf[grid];
178: }
179: PetscCall(VecDestroy(&locX2));
180: }
181: }
182: PetscCheck(cellClosure_it - cellClosure == cellClosure_sz * ctx->batch_sz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "iteration wrong %" PetscCount_FMT " != cellClosure_sz = %" PetscInt_FMT, (PetscCount)(cellClosure_it - cellClosure),
183: cellClosure_sz * ctx->batch_sz);
184: PetscCall(DMCompositeRestoreLocalAccessArray(pack, a_X, nDMs, NULL, locXArray));
185: PetscCall(DMCompositeRestoreAccessArray(pack, a_X, nDMs, NULL, globXArray));
186: PetscCall(PetscFree(locXArray));
187: PetscCall(PetscFree(globXArray));
188: xdata = NULL;
189: } else {
190: PetscMemType mtype;
191: if (ctx->jacobian_field_major_order) { // get data in batch ordering
192: PetscCall(VecScatterBegin(ctx->plex_batch, a_X, ctx->work_vec, INSERT_VALUES, SCATTER_FORWARD));
193: PetscCall(VecScatterEnd(ctx->plex_batch, a_X, ctx->work_vec, INSERT_VALUES, SCATTER_FORWARD));
194: PetscCall(VecGetArrayReadAndMemType(ctx->work_vec, &xdata, &mtype));
195: } else {
196: PetscCall(VecGetArrayReadAndMemType(a_X, &xdata, &mtype));
197: }
198: PetscCheck(mtype == PETSC_MEMTYPE_HOST || ctx->deviceType != LANDAU_CPU, ctx->comm, PETSC_ERR_ARG_WRONG, "CPU run with device data: use -mat_type aij");
199: cellClosure = NULL;
200: }
201: PetscCall(PetscLogEventEnd(ctx->events[1], 0, 0, 0, 0));
202: } else xdata = cellClosure = NULL;
204: /* do it */
205: if (ctx->deviceType == LANDAU_KOKKOS) {
206: #if defined(PETSC_HAVE_KOKKOS)
207: 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));
208: #else
209: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos");
210: #endif
211: } else { /* CPU version */
212: PetscTabulation *Tf; // used for CPU and print info. Same on all grids and all species
213: 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];
214: 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;
215: PetscReal *nu_alpha = (PetscReal *)ctx->SData_d.alpha, *nu_beta = (PetscReal *)ctx->SData_d.beta, *invMass = (PetscReal *)ctx->SData_d.invMass;
216: PetscReal(*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal(*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas;
217: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
218: PetscScalar *coo_vals = NULL;
219: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
220: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
221: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
222: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
223: }
224: /* count IPf size, etc */
225: PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids
226: const PetscReal *const BB = Tf[0]->T[0], *const DD = Tf[0]->T[1];
227: ip_offset[0] = ipf_offset[0] = elem_offset[0] = 0;
228: for (PetscInt grid = 0; grid < num_grids; grid++) {
229: PetscInt nfloc = ctx->species_offset[grid + 1] - ctx->species_offset[grid];
230: elem_offset[grid + 1] = elem_offset[grid] + numCells[grid];
231: ip_offset[grid + 1] = ip_offset[grid] + numCells[grid] * Nq;
232: ipf_offset[grid + 1] = ipf_offset[grid] + Nq * nfloc * numCells[grid];
233: }
234: IPf_sz_glb = ipf_offset[num_grids];
235: IPf_sz_tot = IPf_sz_glb * ctx->batch_sz;
236: // prep COO
237: PetscCall(PetscMalloc1(ctx->SData_d.coo_size, &coo_vals)); // allocate every time?
238: if (shift == 0.0) { /* compute dynamic data f and df and init data for Jacobian */
239: #if defined(PETSC_HAVE_THREADSAFETY)
240: double starttime, endtime;
241: starttime = MPI_Wtime();
242: #endif
243: PetscCall(PetscLogEventBegin(ctx->events[8], 0, 0, 0, 0));
244: PetscCall(PetscMalloc4(IPf_sz_tot, &ff, IPf_sz_tot, &dudx, IPf_sz_tot, &dudy, (dim == 3 ? IPf_sz_tot : 0), &dudz));
245: // F df/dx
246: for (PetscInt tid = 0; tid < ctx->batch_sz * elem_offset[num_grids]; tid++) { // for each element
247: 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
248: // find my grid:
249: PetscInt grid = 0;
250: while (b_elem_idx >= elem_offset[grid + 1]) grid++; // yuck search for grid
251: {
252: 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];
253: 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];
254: PetscScalar *coef, coef_buff[LANDAU_MAX_SPECIES * LANDAU_MAX_NQND];
255: PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem * Nq) * dim * dim]; // ingJ is static data on batch 0
256: PetscInt b, f, q;
257: if (cellClosure) {
258: coef = &cellClosure[b_id * IPf_sz_glb + ipf_offset[grid] + loc_elem * Nb * loc_Nf]; // this is const
259: } else {
260: coef = coef_buff;
261: for (f = 0; f < loc_Nf; ++f) {
262: LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][f][0];
263: for (b = 0; b < Nb; ++b) {
264: PetscInt idx = Idxs[b];
265: if (idx >= 0) {
266: coef[f * Nb + b] = xdata[idx + moffset];
267: } else {
268: idx = -idx - 1;
269: coef[f * Nb + b] = 0;
270: for (q = 0; q < maps[grid].num_face; q++) {
271: PetscInt id = maps[grid].c_maps[idx][q].gid;
272: PetscScalar scale = maps[grid].c_maps[idx][q].scale;
273: coef[f * Nb + b] += scale * xdata[id + moffset];
274: }
275: }
276: }
277: }
278: }
279: /* get f and df */
280: for (PetscInt qi = 0; qi < Nq; qi++) {
281: const PetscReal *invJ = &invJe[qi * dim * dim];
282: const PetscReal *Bq = &BB[qi * Nb];
283: const PetscReal *Dq = &DD[qi * Nb * dim];
284: PetscReal u_x[LANDAU_DIM];
285: /* get f & df */
286: for (f = 0; f < loc_Nf; ++f) {
287: const PetscInt idx = b_id * IPf_sz_glb + ipf_offset[grid] + f * loc_nip + loc_elem * Nq + qi;
288: PetscInt b, e;
289: PetscReal refSpaceDer[LANDAU_DIM];
290: ff[idx] = 0.0;
291: for (PetscInt d = 0; d < LANDAU_DIM; ++d) refSpaceDer[d] = 0.0;
292: for (b = 0; b < Nb; ++b) {
293: const PetscInt cidx = b;
294: ff[idx] += Bq[cidx] * PetscRealPart(coef[f * Nb + cidx]);
295: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] += Dq[cidx * dim + d] * PetscRealPart(coef[f * Nb + cidx]);
296: }
297: for (PetscInt d = 0; d < LANDAU_DIM; ++d) {
298: for (e = 0, u_x[d] = 0.0; e < LANDAU_DIM; ++e) u_x[d] += invJ[e * dim + d] * refSpaceDer[e];
299: }
300: dudx[idx] = u_x[0];
301: dudy[idx] = u_x[1];
302: #if LANDAU_DIM == 3
303: dudz[idx] = u_x[2];
304: #endif
305: }
306: } // q
307: } // grid
308: } // grid*batch
309: PetscCall(PetscLogEventEnd(ctx->events[8], 0, 0, 0, 0));
310: #if defined(PETSC_HAVE_THREADSAFETY)
311: endtime = MPI_Wtime();
312: if (ctx->stage) ctx->times[LANDAU_F_DF] += (endtime - starttime);
313: #endif
314: } // Jacobian setup
315: // assemble Jacobian (or mass)
316: for (PetscInt tid = 0; tid < ctx->batch_sz * elem_offset[num_grids]; tid++) { // for each element
317: const PetscInt b_Nelem = elem_offset[num_grids];
318: const PetscInt glb_elem_idx = tid % b_Nelem, b_id = tid / b_Nelem;
319: PetscInt grid = 0;
320: #if defined(PETSC_HAVE_THREADSAFETY)
321: double starttime, endtime;
322: starttime = MPI_Wtime();
323: #endif
324: while (glb_elem_idx >= elem_offset[grid + 1]) grid++;
325: {
326: const PetscInt loc_Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], loc_elem = glb_elem_idx - elem_offset[grid];
327: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset), totDim = loc_Nf * Nq, elemMatSize = totDim * totDim;
328: PetscScalar *elemMat;
329: const PetscReal *invJe = &invJ_a[(ip_offset[grid] + loc_elem * Nq) * dim * dim];
330: PetscCall(PetscMalloc1(elemMatSize, &elemMat));
331: PetscCall(PetscMemzero(elemMat, elemMatSize * sizeof(*elemMat)));
332: if (shift == 0.0) { // Jacobian
333: PetscCall(PetscLogEventBegin(ctx->events[4], 0, 0, 0, 0));
334: } else { // mass
335: PetscCall(PetscLogEventBegin(ctx->events[16], 0, 0, 0, 0));
336: }
337: for (PetscInt qj = 0; qj < Nq; ++qj) {
338: const PetscInt jpidx_glb = ip_offset[grid] + qj + loc_elem * Nq;
339: 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
340: PetscInt d, d2, dp, d3, IPf_idx;
341: if (shift == 0.0) { // Jacobian
342: const PetscReal *const invJj = &invJe[qj * dim * dim];
343: 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];
344: const PetscReal vj[3] = {xx[jpidx_glb], yy[jpidx_glb], zz ? zz[jpidx_glb] : 0}, wj = ww[jpidx_glb];
345: // create g2 & g3
346: for (d = 0; d < LANDAU_DIM; d++) { // clear accumulation data D & K
347: gg2_temp[d] = 0;
348: for (d2 = 0; d2 < LANDAU_DIM; d2++) gg3_temp[d][d2] = 0;
349: }
350: /* inner beta reduction */
351: IPf_idx = 0;
352: 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
353: PetscInt nip_loc_r = numCells[grid_r] * Nq, Nfloc_r = Nf[grid_r];
354: for (PetscInt ei_r = 0, loc_fdf_idx = 0; ei_r < numCells[grid_r]; ++ei_r) {
355: for (PetscInt qi = 0; qi < Nq; qi++, ipidx++, loc_fdf_idx++) {
356: const PetscReal wi = ww[ipidx], x = xx[ipidx], y = yy[ipidx];
357: PetscReal temp1[3] = {0, 0, 0}, temp2 = 0;
358: #if LANDAU_DIM == 2
359: 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.;
360: LandauTensor2D(vj, x, y, Ud, Uk, mask);
361: #else
362: 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.;
363: if (ctx->use_relativistic_corrections) {
364: LandauTensor3DRelativistic(vj, x, y, z, U, mask, C_0(ctx->v_0));
365: } else {
366: LandauTensor3D(vj, x, y, z, U, mask);
367: }
368: #endif
369: for (PetscInt f = 0; f < Nfloc_r; ++f) {
370: const PetscInt idx = b_id * IPf_sz_glb + ipf_offset[grid_r] + f * nip_loc_r + ei_r * Nq + qi; // IPf_idx + f*nip_loc_r + loc_fdf_idx;
371: temp1[0] += dudx[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
372: temp1[1] += dudy[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
373: #if LANDAU_DIM == 3
374: temp1[2] += dudz[idx] * nu_beta[f + f_off] * invMass[f + f_off] * (*lambdas)[grid][grid_r];
375: #endif
376: temp2 += ff[idx] * nu_beta[f + f_off] * (*lambdas)[grid][grid_r];
377: }
378: temp1[0] *= wi;
379: temp1[1] *= wi;
380: #if LANDAU_DIM == 3
381: temp1[2] *= wi;
382: #endif
383: temp2 *= wi;
384: #if LANDAU_DIM == 2
385: for (d2 = 0; d2 < 2; d2++) {
386: for (d3 = 0; d3 < 2; ++d3) {
387: /* K = U * grad(f): g2=e: i,A */
388: gg2_temp[d2] += Uk[d2][d3] * temp1[d3];
389: /* D = -U * (I \kron (fx)): g3=f: i,j,A */
390: gg3_temp[d2][d3] += Ud[d2][d3] * temp2;
391: }
392: }
393: #else
394: for (d2 = 0; d2 < 3; ++d2) {
395: for (d3 = 0; d3 < 3; ++d3) {
396: /* K = U * grad(f): g2 = e: i,A */
397: gg2_temp[d2] += U[d2][d3] * temp1[d3];
398: /* D = -U * (I \kron (fx)): g3 = f: i,j,A */
399: gg3_temp[d2][d3] += U[d2][d3] * temp2;
400: }
401: }
402: #endif
403: } // qi
404: } // ei_r
405: IPf_idx += nip_loc_r * Nfloc_r;
406: } /* grid_r - IPs */
407: 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);
408: // add alpha and put in gg2/3
409: for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) {
410: for (d2 = 0; d2 < LANDAU_DIM; d2++) {
411: gg2[fieldA][d2] = gg2_temp[d2] * nu_alpha[fieldA + f_off];
412: for (d3 = 0; d3 < LANDAU_DIM; d3++) gg3[fieldA][d2][d3] = -gg3_temp[d2][d3] * nu_alpha[fieldA + f_off] * invMass[fieldA + f_off];
413: }
414: }
415: /* add electric field term once per IP */
416: for (PetscInt fieldA = 0, f_off = ctx->species_offset[grid]; fieldA < loc_Nf; ++fieldA) gg2[fieldA][LANDAU_DIM - 1] += Eq_m[fieldA + f_off];
417: /* Jacobian transform - g2, g3 */
418: for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) {
419: for (d = 0; d < dim; ++d) {
420: g2[fieldA][d] = 0.0;
421: for (d2 = 0; d2 < dim; ++d2) {
422: g2[fieldA][d] += invJj[d * dim + d2] * gg2[fieldA][d2];
423: g3[fieldA][d][d2] = 0.0;
424: for (d3 = 0; d3 < dim; ++d3) {
425: for (dp = 0; dp < dim; ++dp) g3[fieldA][d][d2] += invJj[d * dim + d3] * gg3[fieldA][d3][dp] * invJj[d2 * dim + dp];
426: }
427: g3[fieldA][d][d2] *= wj;
428: }
429: g2[fieldA][d] *= wj;
430: }
431: }
432: } else { // mass
433: PetscReal wj = ww[jpidx_glb];
434: /* Jacobian transform - g0 */
435: for (PetscInt fieldA = 0; fieldA < loc_Nf; ++fieldA) {
436: if (dim == 2) {
437: g0[fieldA] = wj * shift * 2. * PETSC_PI; // move this to below and remove g0
438: } else {
439: g0[fieldA] = wj * shift; // move this to below and remove g0
440: }
441: }
442: }
443: /* FE matrix construction */
444: {
445: PetscInt fieldA, d, f, d2, g;
446: const PetscReal *BJq = &BB[qj * Nb], *DIq = &DD[qj * Nb * dim];
447: /* assemble - on the diagonal (I,I) */
448: for (fieldA = 0; fieldA < loc_Nf; fieldA++) {
449: for (f = 0; f < Nb; f++) {
450: const PetscInt i = fieldA * Nb + f; /* Element matrix row */
451: for (g = 0; g < Nb; ++g) {
452: const PetscInt j = fieldA * Nb + g; /* Element matrix column */
453: const PetscInt fOff = i * totDim + j;
454: if (shift == 0.0) {
455: for (d = 0; d < dim; ++d) {
456: elemMat[fOff] += DIq[f * dim + d] * g2[fieldA][d] * BJq[g];
457: for (d2 = 0; d2 < dim; ++d2) elemMat[fOff] += DIq[f * dim + d] * g3[fieldA][d][d2] * DIq[g * dim + d2];
458: }
459: } else { // mass
460: elemMat[fOff] += BJq[f] * g0[fieldA] * BJq[g];
461: }
462: }
463: }
464: }
465: }
466: } /* qj loop */
467: if (shift == 0.0) { // Jacobian
468: PetscCall(PetscLogEventEnd(ctx->events[4], 0, 0, 0, 0));
469: } else {
470: PetscCall(PetscLogEventEnd(ctx->events[16], 0, 0, 0, 0));
471: }
472: #if defined(PETSC_HAVE_THREADSAFETY)
473: endtime = MPI_Wtime();
474: if (ctx->stage) ctx->times[LANDAU_KERNEL] += (endtime - starttime);
475: #endif
476: /* assemble matrix */
477: if (!container) {
478: PetscInt cStart;
479: PetscCall(PetscLogEventBegin(ctx->events[6], 0, 0, 0, 0));
480: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, NULL));
481: PetscCall(DMPlexMatSetClosure(ctx->plex[grid], section[grid], globsection[grid], subJ[LAND_PACK_IDX(b_id, grid)], loc_elem + cStart, elemMat, ADD_VALUES));
482: PetscCall(PetscLogEventEnd(ctx->events[6], 0, 0, 0, 0));
483: } else { // GPU like assembly for debugging
484: 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];
485: 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};
486: 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;
487: /* assemble - from the diagonal (I,I) in this format for DMPlexMatSetClosure */
488: for (fieldA = 0; fieldA < loc_Nf; fieldA++) {
489: LandauIdx *const Idxs = &maps[grid].gIdx[loc_elem][fieldA][0];
490: for (f = 0; f < Nb; f++) {
491: PetscInt idx = Idxs[f];
492: if (idx >= 0) {
493: nr = 1;
494: rows0[0] = idx;
495: row_scale[0] = 1.;
496: } else {
497: idx = -idx - 1;
498: for (q = 0, nr = 0; q < maps[grid].num_face; q++, nr++) {
499: if (maps[grid].c_maps[idx][q].gid < 0) break;
500: rows0[q] = maps[grid].c_maps[idx][q].gid;
501: row_scale[q] = maps[grid].c_maps[idx][q].scale;
502: }
503: }
504: for (g = 0; g < Nb; ++g) {
505: idx = Idxs[g];
506: if (idx >= 0) {
507: nc = 1;
508: cols0[0] = idx;
509: col_scale[0] = 1.;
510: } else {
511: idx = -idx - 1;
512: nc = maps[grid].num_face;
513: for (q = 0, nc = 0; q < maps[grid].num_face; q++, nc++) {
514: if (maps[grid].c_maps[idx][q].gid < 0) break;
515: cols0[q] = maps[grid].c_maps[idx][q].gid;
516: col_scale[q] = maps[grid].c_maps[idx][q].scale;
517: }
518: }
519: const PetscInt i = fieldA * Nb + f; /* Element matrix row */
520: const PetscInt j = fieldA * Nb + g; /* Element matrix column */
521: const PetscScalar Aij = elemMat[i * totDim + j];
522: if (coo_vals) { // mirror (i,j) in CreateStaticGPUData
523: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb;
524: 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];
525: for (PetscInt q = 0, idx2 = idx0; q < nr; q++) {
526: for (PetscInt d = 0; d < nc; d++, idx2++) coo_vals[idx2] = row_scale[q] * col_scale[d] * Aij;
527: }
528: } else {
529: for (q = 0; q < nr; q++) rows[q] = rows0[q] + moffset;
530: for (d = 0; d < nc; d++) cols[d] = cols0[d] + moffset;
531: for (q = 0; q < nr; q++) {
532: for (d = 0; d < nc; d++) vals[q * nc + d] = row_scale[q] * col_scale[d] * Aij;
533: }
534: PetscCall(MatSetValues(JacP, nr, rows, nc, cols, vals, ADD_VALUES));
535: }
536: }
537: }
538: }
539: }
540: if (loc_elem == -1) {
541: PetscCall(PetscPrintf(ctx->comm, "CPU Element matrix\n"));
542: for (PetscInt d = 0; d < totDim; ++d) {
543: for (PetscInt f = 0; f < totDim; ++f) PetscCall(PetscPrintf(ctx->comm, " %12.5e", (double)PetscRealPart(elemMat[d * totDim + f])));
544: PetscCall(PetscPrintf(ctx->comm, "\n"));
545: }
546: exit(12);
547: }
548: PetscCall(PetscFree(elemMat));
549: } /* grid */
550: } /* outer element & batch loop */
551: if (shift == 0.0) { // mass
552: PetscCall(PetscFree4(ff, dudx, dudy, dudz));
553: }
554: if (!container) { // 'CPU' assembly move nest matrix to global JacP
555: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // OpenMP
556: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
557: 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];
558: PetscInt nloc, nzl, colbuf[1024], row;
559: const PetscInt *cols;
560: const PetscScalar *vals;
561: Mat B = subJ[LAND_PACK_IDX(b_id, grid)];
562: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
563: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
564: PetscCall(MatGetSize(B, &nloc, NULL));
565: for (PetscInt i = 0; i < nloc; i++) {
566: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
567: PetscCheck(nzl <= 1024, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Row too big: %" PetscInt_FMT, nzl);
568: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = moffset + cols[j];
569: row = moffset + i;
570: PetscCall(MatSetValues(JacP, 1, &row, nzl, colbuf, vals, ADD_VALUES));
571: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
572: }
573: PetscCall(MatDestroy(&B));
574: }
575: }
576: }
577: if (coo_vals) {
578: PetscCall(MatSetValuesCOO(JacP, coo_vals, ADD_VALUES));
579: PetscCall(PetscFree(coo_vals));
580: }
581: } /* CPU version */
582: PetscCall(MatAssemblyBegin(JacP, MAT_FINAL_ASSEMBLY));
583: PetscCall(MatAssemblyEnd(JacP, MAT_FINAL_ASSEMBLY));
584: /* clean up */
585: if (cellClosure) PetscCall(PetscFree(cellClosure));
586: if (xdata) PetscCall(VecRestoreArrayReadAndMemType(a_X, &xdata));
587: PetscFunctionReturn(PETSC_SUCCESS);
588: }
590: static PetscErrorCode GeometryDMLandau(DM base, PetscInt point, PetscInt dim, const PetscReal abc[], PetscReal xyz[], void *a_ctx)
591: {
592: PetscReal r = abc[0], z = abc[1];
593: LandauCtx *ctx = (LandauCtx *)a_ctx;
595: PetscFunctionBegin;
596: if (ctx->sphere && dim == 3) { // make sphere: works for one AMR and Q2
597: PetscInt nzero = 0, idx = 0;
598: xyz[0] = r;
599: xyz[1] = z;
600: xyz[2] = abc[2];
601: for (PetscInt i = 0; i < 3; i++) {
602: if (PetscAbs(xyz[i]) < PETSC_SQRT_MACHINE_EPSILON) nzero++;
603: else idx = i;
604: }
605: if (nzero == 2) xyz[idx] *= 1.732050807568877; // sqrt(3)
606: else if (nzero == 1) {
607: for (PetscInt i = 0; i < 3; i++) xyz[i] *= 1.224744871391589; // sqrt(3/2)
608: }
609: } else {
610: xyz[0] = r;
611: xyz[1] = z;
612: if (dim == 3) xyz[2] = abc[2];
613: }
614: PetscFunctionReturn(PETSC_SUCCESS);
615: }
617: /* create DMComposite of meshes for each species group */
618: static PetscErrorCode LandauDMCreateVMeshes(MPI_Comm comm_self, const PetscInt dim, const char prefix[], LandauCtx *ctx, DM pack)
619: {
620: PetscFunctionBegin;
621: { /* p4est, quads */
622: /* Create plex mesh of Landau domain */
623: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
624: PetscReal par_radius = ctx->radius_par[grid], perp_radius = ctx->radius_perp[grid];
625: if (!ctx->sphere && !ctx->simplex) { // 2 or 3D (only 3D option)
626: PetscReal lo[] = {-perp_radius, -par_radius, -par_radius}, hi[] = {perp_radius, par_radius, par_radius};
627: DMBoundaryType periodicity[3] = {DM_BOUNDARY_NONE, dim == 2 ? DM_BOUNDARY_NONE : DM_BOUNDARY_NONE, DM_BOUNDARY_NONE};
628: if (dim == 2) lo[0] = 0;
629: else {
630: lo[1] = -perp_radius;
631: hi[1] = perp_radius; // 3D y is a perp
632: }
633: 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
634: PetscCall(DMLocalizeCoordinates(ctx->plex[grid])); /* needed for periodic */
635: if (dim == 3) PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cube"));
636: else PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "half-plane"));
637: } else if (dim == 2) {
638: size_t len;
639: PetscCall(PetscStrlen(ctx->filename, &len));
640: if (len) {
641: Vec coords;
642: PetscScalar *x;
643: PetscInt N;
644: char str[] = "-dm_landau_view_file_0";
645: str[21] += grid;
646: PetscCall(DMPlexCreateFromFile(comm_self, ctx->filename, "plexland.c", PETSC_TRUE, &ctx->plex[grid]));
647: PetscCall(DMPlexOrient(ctx->plex[grid]));
648: PetscCall(DMGetCoordinatesLocal(ctx->plex[grid], &coords));
649: PetscCall(VecGetSize(coords, &N));
650: PetscCall(VecGetArray(coords, &x));
651: /* scale by domain size */
652: for (PetscInt i = 0; i < N; i += 2) {
653: x[i + 0] *= ctx->radius_perp[grid];
654: x[i + 1] *= ctx->radius_par[grid];
655: }
656: PetscCall(VecRestoreArray(coords, &x));
657: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], ctx->filename));
658: PetscCall(PetscInfo(ctx->plex[grid], "%d) Read %s mesh file (%s)\n", (int)grid, ctx->filename, str));
659: PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, str));
660: } else { // simplex forces a sphere
661: PetscInt numCells = ctx->simplex ? 12 : 6, cell_size = ctx->simplex ? 3 : 4, j;
662: const PetscInt numVerts = 11;
663: PetscInt cellsT[][4] = {
664: {0, 1, 6, 5 },
665: {1, 2, 7, 6 },
666: {2, 3, 8, 7 },
667: {3, 4, 9, 8 },
668: {5, 6, 7, 10},
669: {10, 7, 8, 9 }
670: };
671: PetscInt cellsS[][3] = {
672: {0, 1, 6 },
673: {1, 2, 6 },
674: {6, 2, 7 },
675: {7, 2, 8 },
676: {8, 2, 3 },
677: {8, 3, 4 },
678: {0, 6, 5 },
679: {5, 6, 7 },
680: {5, 7, 10},
681: {10, 7, 9 },
682: {9, 7, 8 },
683: {9, 8, 4 }
684: };
685: const PetscInt *pcell = (const PetscInt *)(ctx->simplex ? &cellsS[0][0] : &cellsT[0][0]);
686: PetscReal coords[11][2], *flatCoords = (PetscReal *)&coords[0][0];
687: PetscReal rad = ctx->radius[grid];
688: for (j = 0; j < 5; j++) { // outside edge
689: PetscReal z, r, theta = -PETSC_PI / 2 + (j % 5) * PETSC_PI / 4;
690: r = rad * PetscCosReal(theta);
691: coords[j][0] = r;
692: z = rad * PetscSinReal(theta);
693: coords[j][1] = z;
694: }
695: coords[j][0] = 0;
696: coords[j++][1] = -rad * ctx->sphere_inner_radius_90degree[grid];
697: coords[j][0] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
698: coords[j++][1] = -rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
699: coords[j][0] = rad * ctx->sphere_inner_radius_90degree[grid];
700: coords[j++][1] = 0;
701: coords[j][0] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
702: coords[j++][1] = rad * ctx->sphere_inner_radius_45degree[grid] * 0.707106781186548;
703: coords[j][0] = 0;
704: coords[j++][1] = rad * ctx->sphere_inner_radius_90degree[grid];
705: coords[j][0] = 0;
706: coords[j++][1] = 0;
707: PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 2, numCells, numVerts, cell_size, ctx->interpolate, pcell, 2, flatCoords, &ctx->plex[grid]));
708: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "semi-circle"));
709: PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make circle %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor"));
710: }
711: } else {
712: PetscCheck(dim == 3 && ctx->sphere && !ctx->simplex, ctx->comm, PETSC_ERR_ARG_WRONG, "not: dim == 3 && ctx->sphere && !ctx->simplex");
713: PetscReal rad = ctx->radius[grid] / 1.732050807568877, inner_rad = rad * ctx->sphere_inner_radius_45degree[grid], outer_rad = rad;
714: const PetscInt numCells = 7, cell_size = 8, numVerts = 16;
715: const PetscInt cells[][8] = {
716: {0, 3, 2, 1, 4, 5, 6, 7 },
717: {0, 4, 5, 1, 8, 9, 13, 12},
718: {1, 5, 6, 2, 9, 10, 14, 13},
719: {2, 6, 7, 3, 10, 11, 15, 14},
720: {0, 3, 7, 4, 8, 12, 15, 11},
721: {0, 1, 2, 3, 8, 11, 10, 9 },
722: {4, 7, 6, 5, 12, 13, 14, 15}
723: };
724: PetscReal coords[16 /* numVerts */][3];
725: for (PetscInt j = 0; j < 4; j++) { // inner edge, low
726: coords[j][0] = inner_rad * (j == 0 || j == 3 ? 1 : -1);
727: coords[j][1] = inner_rad * (j / 2 < 1 ? 1 : -1);
728: coords[j][2] = inner_rad * -1;
729: }
730: for (PetscInt j = 0, jj = 4; j < 4; j++, jj++) { // inner edge, hi
731: coords[jj][0] = inner_rad * (j == 0 || j == 3 ? 1 : -1);
732: coords[jj][1] = inner_rad * (j / 2 < 1 ? 1 : -1);
733: coords[jj][2] = inner_rad * 1;
734: }
735: for (PetscInt j = 0, jj = 8; j < 4; j++, jj++) { // outer edge, low
736: coords[jj][0] = outer_rad * (j == 0 || j == 3 ? 1 : -1);
737: coords[jj][1] = outer_rad * (j / 2 < 1 ? 1 : -1);
738: coords[jj][2] = outer_rad * -1;
739: }
740: for (PetscInt j = 0, jj = 12; j < 4; j++, jj++) { // outer edge, hi
741: coords[jj][0] = outer_rad * (j == 0 || j == 3 ? 1 : -1);
742: coords[jj][1] = outer_rad * (j / 2 < 1 ? 1 : -1);
743: coords[jj][2] = outer_rad * 1;
744: }
745: PetscCall(DMPlexCreateFromCellListPetsc(comm_self, 3, numCells, numVerts, cell_size, ctx->interpolate, (const PetscInt *)cells, 3, (const PetscReal *)coords, &ctx->plex[grid]));
746: PetscCall(PetscObjectSetName((PetscObject)ctx->plex[grid], "cubed sphere"));
747: PetscCall(PetscInfo(ctx->plex[grid], "\t%" PetscInt_FMT ") Make cubed sphere %s mesh\n", grid, ctx->simplex ? "simplex" : "tensor"));
748: }
749: PetscCall(DMSetFromOptions(ctx->plex[grid]));
750: } // grid loop
751: PetscCall(PetscObjectSetOptionsPrefix((PetscObject)pack, prefix));
752: { /* convert to p4est (or whatever), wait for discretization to create pack */
753: char convType[256];
754: PetscBool flg;
756: PetscOptionsBegin(ctx->comm, prefix, "Mesh conversion options", "DMPLEX");
757: PetscCall(PetscOptionsFList("-dm_landau_type", "Convert DMPlex to another format (p4est)", "plexland.c", DMList, DMPLEX, convType, 256, &flg));
758: PetscOptionsEnd();
759: if (flg) {
760: ctx->use_p4est = PETSC_TRUE; /* flag for Forest */
761: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
762: DM dmforest;
763: PetscCall(DMConvert(ctx->plex[grid], convType, &dmforest));
764: if (dmforest) {
765: PetscBool isForest;
766: PetscCall(PetscObjectSetOptionsPrefix((PetscObject)dmforest, prefix));
767: PetscCall(DMIsForest(dmforest, &isForest));
768: if (isForest) {
769: if (ctx->sphere) PetscCall(DMForestSetBaseCoordinateMapping(dmforest, GeometryDMLandau, ctx));
770: PetscCall(DMDestroy(&ctx->plex[grid]));
771: ctx->plex[grid] = dmforest; // Forest for adaptivity
772: } else SETERRQ(ctx->comm, PETSC_ERR_PLIB, "Converted to non Forest?");
773: } else SETERRQ(ctx->comm, PETSC_ERR_PLIB, "Convert failed?");
774: }
775: } else ctx->use_p4est = PETSC_FALSE; /* flag for Forest */
776: }
777: } /* non-file */
778: PetscCall(DMSetDimension(pack, dim));
779: PetscCall(PetscObjectSetName((PetscObject)pack, "Mesh"));
780: PetscCall(DMSetApplicationContext(pack, ctx));
781: PetscFunctionReturn(PETSC_SUCCESS);
782: }
784: static PetscErrorCode SetupDS(DM pack, PetscInt dim, PetscInt grid, LandauCtx *ctx)
785: {
786: PetscInt ii, i0;
787: char buf[256];
788: PetscSection section;
790: PetscFunctionBegin;
791: for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
792: if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "e"));
793: else PetscCall(PetscSNPrintf(buf, sizeof(buf), "i%" PetscInt_FMT, ii));
794: /* Setup Discretization - FEM */
795: PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, NULL, PETSC_DECIDE, &ctx->fe[ii]));
796: PetscCall(PetscObjectSetName((PetscObject)ctx->fe[ii], buf));
797: PetscCall(DMSetField(ctx->plex[grid], i0, NULL, (PetscObject)ctx->fe[ii]));
798: }
799: PetscCall(DMCreateDS(ctx->plex[grid]));
800: PetscCall(DMGetSection(ctx->plex[grid], §ion));
801: for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
802: if (ii == 0) PetscCall(PetscSNPrintf(buf, sizeof(buf), "se"));
803: else PetscCall(PetscSNPrintf(buf, sizeof(buf), "si%" PetscInt_FMT, ii));
804: PetscCall(PetscSectionSetComponentName(section, i0, 0, buf));
805: }
806: PetscFunctionReturn(PETSC_SUCCESS);
807: }
809: /* Define a Maxwellian function for testing out the operator. */
811: /* Using cartesian velocity space coordinates, the particle */
812: /* density, [1/m^3], is defined according to */
814: /* $$ n=\int_{R^3} dv^3 \left(\frac{m}{2\pi T}\right)^{3/2}\exp [- mv^2/(2T)] $$ */
816: /* Using some constant, c, we normalize the velocity vector into a */
817: /* dimensionless variable according to v=c*x. Thus the density, $n$, becomes */
819: /* $$ n=\int_{R^3} dx^3 \left(\frac{mc^2}{2\pi T}\right)^{3/2}\exp [- mc^2/(2T)*x^2] $$ */
821: /* Defining $\theta=2T/mc^2$, we thus find that the probability density */
822: /* for finding the particle within the interval in a box dx^3 around x is */
824: /* f(x;\theta)=\left(\frac{1}{\pi\theta}\right)^{3/2} \exp [ -x^2/\theta ] */
826: typedef struct {
827: PetscReal v_0;
828: PetscReal kT_m;
829: PetscReal n;
830: PetscReal shift;
831: } MaxwellianCtx;
833: static PetscErrorCode maxwellian(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf_dummy, PetscScalar *u, void *actx)
834: {
835: MaxwellianCtx *mctx = (MaxwellianCtx *)actx;
836: PetscInt i;
837: PetscReal v2 = 0, theta = 2 * mctx->kT_m / (mctx->v_0 * mctx->v_0), shift; /* theta = 2kT/mc^2 */
839: PetscFunctionBegin;
840: /* compute the exponents, v^2 */
841: for (i = 0; i < dim; ++i) v2 += x[i] * x[i];
842: /* evaluate the Maxwellian */
843: if (mctx->shift < 0) shift = -mctx->shift;
844: else {
845: u[0] = mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta));
846: shift = mctx->shift;
847: }
848: if (shift != 0.) {
849: v2 = 0;
850: for (i = 0; i < dim - 1; ++i) v2 += x[i] * x[i];
851: v2 += (x[dim - 1] - shift) * (x[dim - 1] - shift);
852: /* evaluate the shifted Maxwellian */
853: u[0] += mctx->n * PetscPowReal(PETSC_PI * theta, -1.5) * (PetscExpReal(-v2 / theta));
854: }
855: PetscFunctionReturn(PETSC_SUCCESS);
856: }
858: /*@
859: DMPlexLandauAddMaxwellians - Add a Maxwellian distribution to a state
861: Collective
863: Input Parameters:
864: + dm - The mesh (local)
865: . time - Current time
866: . temps - Temperatures of each species (global)
867: . ns - Number density of each species (global)
868: . grid - index into current grid - just used for offset into `temp` and `ns`
869: . b_id - batch index
870: . n_batch - number of batches
871: - actx - Landau context
873: Output Parameter:
874: . X - The state (local to this grid)
876: Level: beginner
878: .seealso: `DMPlexLandauCreateVelocitySpace()`
879: @*/
880: PetscErrorCode DMPlexLandauAddMaxwellians(DM dm, Vec X, PetscReal time, PetscReal temps[], PetscReal ns[], PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx)
881: {
882: LandauCtx *ctx = (LandauCtx *)actx;
883: PetscErrorCode (*initu[LANDAU_MAX_SPECIES])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *);
884: PetscInt dim;
885: MaxwellianCtx *mctxs[LANDAU_MAX_SPECIES], data[LANDAU_MAX_SPECIES];
887: PetscFunctionBegin;
888: PetscCall(DMGetDimension(dm, &dim));
889: if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx));
890: for (PetscInt ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
891: mctxs[i0] = &data[i0];
892: data[i0].v_0 = ctx->v_0; // v_0 same for all grids
893: data[i0].kT_m = ctx->k * temps[ii] / ctx->masses[ii]; /* kT/m */
894: data[i0].n = ns[ii];
895: initu[i0] = maxwellian;
896: data[i0].shift = 0;
897: }
898: data[0].shift = ctx->electronShift;
899: /* need to make ADD_ALL_VALUES work - TODO */
900: PetscCall(DMProjectFunction(dm, time, initu, (void **)mctxs, INSERT_ALL_VALUES, X));
901: PetscFunctionReturn(PETSC_SUCCESS);
902: }
904: /*
905: LandauSetInitialCondition - Adds Maxwellians with context
907: Collective
909: Input Parameters:
910: . dm - The mesh
911: - grid - index into current grid - just used for offset into temp and ns
912: . b_id - batch index
913: - n_batch - number of batches
914: + actx - Landau context with T and n
916: Output Parameter:
917: . X - The state
919: Level: beginner
921: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauAddMaxwellians()`
922: */
923: static PetscErrorCode LandauSetInitialCondition(DM dm, Vec X, PetscInt grid, PetscInt b_id, PetscInt n_batch, void *actx)
924: {
925: LandauCtx *ctx = (LandauCtx *)actx;
927: PetscFunctionBegin;
928: if (!ctx) PetscCall(DMGetApplicationContext(dm, &ctx));
929: PetscCall(VecZeroEntries(X));
930: PetscCall(DMPlexLandauAddMaxwellians(dm, X, 0.0, ctx->thermal_temps, ctx->n, grid, b_id, n_batch, ctx));
931: PetscFunctionReturn(PETSC_SUCCESS);
932: }
934: // adapt a level once. Forest in/out
935: #if defined(PETSC_USE_INFO)
936: static const char *s_refine_names[] = {"RE", "Z1", "Origin", "Z2", "Uniform"};
937: #endif
938: static PetscErrorCode adaptToleranceFEM(PetscFE fem, Vec sol, PetscInt type, PetscInt grid, LandauCtx *ctx, DM *newForest)
939: {
940: DM forest, plex, adaptedDM = NULL;
941: PetscDS prob;
942: PetscBool isForest;
943: PetscQuadrature quad;
944: PetscInt Nq, Nb, *Nb2, cStart, cEnd, c, dim, qj, k;
945: DMLabel adaptLabel = NULL;
947: PetscFunctionBegin;
948: forest = ctx->plex[grid];
949: PetscCall(DMCreateDS(forest));
950: PetscCall(DMGetDS(forest, &prob));
951: PetscCall(DMGetDimension(forest, &dim));
952: PetscCall(DMIsForest(forest, &isForest));
953: PetscCheck(isForest, ctx->comm, PETSC_ERR_ARG_WRONG, "! Forest");
954: PetscCall(DMConvert(forest, DMPLEX, &plex));
955: PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd));
956: PetscCall(DMLabelCreate(PETSC_COMM_SELF, "adapt", &adaptLabel));
957: PetscCall(PetscFEGetQuadrature(fem, &quad));
958: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, NULL));
959: 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);
960: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
961: PetscCall(PetscDSGetDimensions(prob, &Nb2));
962: PetscCheck(Nb2[0] == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, " Nb = %" PetscInt_FMT " != Nb (%d)", Nb, (int)Nb2[0]);
963: 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);
964: PetscCall(PetscInfo(sol, "%" PetscInt_FMT ") Refine phase: %s\n", grid, s_refine_names[type]));
965: if (type == 4) {
966: for (c = cStart; c < cEnd; c++) PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE));
967: } else if (type == 2) {
968: PetscInt rCellIdx[8], nr = 0, nrmax = (dim == 3) ? 8 : 2;
969: PetscReal minRad = PETSC_INFINITY, r;
970: for (c = cStart; c < cEnd; c++) {
971: PetscReal tt, v0[LANDAU_MAX_NQND * 3], J[LANDAU_MAX_NQND * 9], invJ[LANDAU_MAX_NQND * 9], detJ[LANDAU_MAX_NQND];
972: PetscCall(DMPlexComputeCellGeometryFEM(plex, c, quad, v0, J, invJ, detJ));
973: (void)J;
974: (void)invJ;
975: for (qj = 0; qj < Nq; ++qj) {
976: tt = PetscSqr(v0[dim * qj + 0]) + PetscSqr(v0[dim * qj + 1]) + PetscSqr((dim == 3) ? v0[dim * qj + 2] : 0);
977: r = PetscSqrtReal(tt);
978: if (r < minRad - PETSC_SQRT_MACHINE_EPSILON * 10.) {
979: minRad = r;
980: nr = 0;
981: rCellIdx[nr++] = c;
982: 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));
983: } else if ((r - minRad) < PETSC_SQRT_MACHINE_EPSILON * 100. && nr < nrmax) {
984: for (k = 0; k < nr; k++)
985: if (c == rCellIdx[k]) break;
986: if (k == nr) {
987: rCellIdx[nr++] = c;
988: 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)));
989: }
990: }
991: }
992: }
993: for (k = 0; k < nr; k++) PetscCall(DMLabelSetValue(adaptLabel, rCellIdx[k], DM_ADAPT_REFINE));
994: 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));
995: } else if (type == 0 || type == 1 || type == 3) { /* refine along r=0 axis */
996: PetscScalar *coef = NULL;
997: Vec coords;
998: PetscInt csize, Nv, d, nz, nrefined = 0;
999: DM cdm;
1000: PetscSection cs;
1001: PetscCall(DMGetCoordinatesLocal(forest, &coords));
1002: PetscCall(DMGetCoordinateDM(forest, &cdm));
1003: PetscCall(DMGetLocalSection(cdm, &cs));
1004: for (c = cStart; c < cEnd; c++) {
1005: PetscInt doit = 0, outside = 0;
1006: PetscCall(DMPlexVecGetClosure(cdm, cs, coords, c, &csize, &coef));
1007: Nv = csize / dim;
1008: for (nz = d = 0; d < Nv; d++) {
1009: PetscReal z = PetscRealPart(coef[d * dim + (dim - 1)]), x = PetscSqr(PetscRealPart(coef[d * dim + 0])) + ((dim == 3) ? PetscSqr(PetscRealPart(coef[d * dim + 1])) : 0);
1010: x = PetscSqrtReal(x);
1011: if (type == 0) {
1012: 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 */
1013: } else if (type == 1 && (z > ctx->vperp0_radius1 || z < -ctx->vperp0_radius1)) {
1014: outside++; /* don't refine outside electron refine radius */
1015: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type]));
1016: } else if (type == 3 && (z > ctx->vperp0_radius2 || z < -ctx->vperp0_radius2)) {
1017: outside++; /* refine r=0 cells on refinement front */
1018: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") (debug) found %s cells\n", grid, s_refine_names[type]));
1019: }
1020: if (x < PETSC_MACHINE_EPSILON * 10. && (type != 0 || ctx->re_radius > PETSC_SQRT_MACHINE_EPSILON)) nz++;
1021: }
1022: PetscCall(DMPlexVecRestoreClosure(cdm, cs, coords, c, &csize, &coef));
1023: if (doit || (outside < Nv && nz)) {
1024: PetscCall(DMLabelSetValue(adaptLabel, c, DM_ADAPT_REFINE));
1025: nrefined++;
1026: }
1027: }
1028: PetscCall(PetscInfo(sol, "\t%" PetscInt_FMT ") Refined %" PetscInt_FMT " cells\n", grid, nrefined));
1029: }
1030: PetscCall(DMDestroy(&plex));
1031: PetscCall(DMAdaptLabel(forest, adaptLabel, &adaptedDM));
1032: PetscCall(DMLabelDestroy(&adaptLabel));
1033: *newForest = adaptedDM;
1034: if (adaptedDM) {
1035: if (isForest) {
1036: PetscCall(DMForestSetAdaptivityForest(adaptedDM, NULL)); // ????
1037: }
1038: PetscCall(DMConvert(adaptedDM, DMPLEX, &plex));
1039: PetscCall(DMPlexGetHeightStratum(plex, 0, &cStart, &cEnd));
1040: PetscCall(PetscInfo(sol, "\t\t\t\t%" PetscInt_FMT ") %" PetscInt_FMT " cells, %" PetscInt_FMT " total quadrature points\n", grid, cEnd - cStart, Nq * (cEnd - cStart)));
1041: PetscCall(DMDestroy(&plex));
1042: } else *newForest = NULL;
1043: PetscFunctionReturn(PETSC_SUCCESS);
1044: }
1046: // forest goes in (ctx->plex[grid]), plex comes out
1047: static PetscErrorCode adapt(PetscInt grid, LandauCtx *ctx, Vec *uu)
1048: {
1049: PetscInt adaptIter;
1051: PetscFunctionBegin;
1052: 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]};
1053: for (type = 0; type < 5; type++) {
1054: for (adaptIter = 0; adaptIter < limits[type]; adaptIter++) {
1055: DM newForest = NULL;
1056: PetscCall(adaptToleranceFEM(ctx->fe[0], *uu, type, grid, ctx, &newForest));
1057: if (newForest) {
1058: PetscCall(DMDestroy(&ctx->plex[grid]));
1059: PetscCall(VecDestroy(uu));
1060: PetscCall(DMCreateGlobalVector(newForest, uu));
1061: PetscCall(LandauSetInitialCondition(newForest, *uu, grid, 0, 1, ctx));
1062: ctx->plex[grid] = newForest;
1063: } else {
1064: PetscCall(PetscInfo(*uu, "No refinement\n"));
1065: }
1066: }
1067: }
1068: PetscCall(PetscObjectSetName((PetscObject)*uu, "uAMR"));
1069: PetscFunctionReturn(PETSC_SUCCESS);
1070: }
1072: // make log(Lambdas) from NRL Plasma formulary
1073: static PetscErrorCode makeLambdas(LandauCtx *ctx)
1074: {
1075: PetscFunctionBegin;
1076: for (PetscInt gridi = 0; gridi < ctx->num_grids; gridi++) {
1077: PetscInt iii = ctx->species_offset[gridi];
1078: PetscReal Ti_ev = (ctx->thermal_temps[iii] / 1.1604525e7) * 1000; // convert (back) to eV
1079: PetscReal ni = ctx->n[iii] * ctx->n_0;
1080: for (PetscInt gridj = gridi; gridj < ctx->num_grids; gridj++) {
1081: PetscInt jjj = ctx->species_offset[gridj];
1082: PetscReal Zj = ctx->charges[jjj] / 1.6022e-19;
1083: if (gridi == 0) {
1084: if (gridj == 0) { // lam_ee
1085: ctx->lambdas[gridi][gridj] = 23.5 - PetscLogReal(PetscSqrtReal(ni) * PetscPowReal(Ti_ev, -1.25)) - PetscSqrtReal(1e-5 + PetscSqr(PetscLogReal(Ti_ev) - 2) / 16);
1086: } else { // lam_ei == lam_ie
1087: if (10 * Zj * Zj > Ti_ev) {
1088: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 23 - PetscLogReal(PetscSqrtReal(ni) * Zj * PetscPowReal(Ti_ev, -1.5));
1089: } else {
1090: ctx->lambdas[gridi][gridj] = ctx->lambdas[gridj][gridi] = 24 - PetscLogReal(PetscSqrtReal(ni) / Ti_ev);
1091: }
1092: }
1093: } else { // lam_ii'
1094: PetscReal mui = ctx->masses[iii] / 1.6720e-27, Zi = ctx->charges[iii] / 1.6022e-19;
1095: PetscReal Tj_ev = (ctx->thermal_temps[jjj] / 1.1604525e7) * 1000; // convert (back) to eV
1096: PetscReal muj = ctx->masses[jjj] / 1.6720e-27;
1097: PetscReal nj = ctx->n[jjj] * ctx->n_0;
1098: 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));
1099: }
1100: }
1101: }
1102: //PetscReal v0 = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */
1103: PetscFunctionReturn(PETSC_SUCCESS);
1104: }
1106: static PetscErrorCode ProcessOptions(LandauCtx *ctx, const char prefix[])
1107: {
1108: PetscBool flg, fileflg;
1109: PetscInt ii, nt, nm, nc, num_species_grid[LANDAU_MAX_GRIDS], non_dim_grid;
1110: PetscReal lnLam = 10;
1111: DM dummy;
1113: PetscFunctionBegin;
1114: PetscCall(DMCreate(ctx->comm, &dummy));
1115: /* get options - initialize context */
1116: ctx->verbose = 1; // should be 0 for silent compliance
1117: ctx->batch_sz = 1;
1118: ctx->batch_view_idx = 0;
1119: ctx->interpolate = PETSC_TRUE;
1120: ctx->gpu_assembly = PETSC_TRUE;
1121: ctx->norm_state = 0;
1122: ctx->electronShift = 0;
1123: ctx->M = NULL;
1124: ctx->J = NULL;
1125: /* geometry and grids */
1126: ctx->sphere = PETSC_FALSE;
1127: ctx->use_p4est = PETSC_FALSE;
1128: ctx->simplex = PETSC_FALSE;
1129: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1130: ctx->radius[grid] = 5.; /* thermal radius (velocity) */
1131: ctx->radius_perp[grid] = 5.; /* thermal radius (velocity) */
1132: ctx->radius_par[grid] = 5.; /* thermal radius (velocity) */
1133: ctx->numAMRRefine[grid] = 0;
1134: ctx->postAMRRefine[grid] = 0;
1135: ctx->species_offset[grid + 1] = 1; // one species default
1136: num_species_grid[grid] = 0;
1137: ctx->plex[grid] = NULL; /* cache as expensive to Convert */
1138: }
1139: ctx->species_offset[0] = 0;
1140: ctx->re_radius = 0.;
1141: ctx->vperp0_radius1 = 0;
1142: ctx->vperp0_radius2 = 0;
1143: ctx->nZRefine1 = 0;
1144: ctx->nZRefine2 = 0;
1145: ctx->numRERefine = 0;
1146: num_species_grid[0] = 1; // one species default
1147: /* species - [0] electrons, [1] one ion species eg, duetarium, [2] heavy impurity ion, ... */
1148: ctx->charges[0] = -1; /* electron charge (MKS) */
1149: ctx->masses[0] = 1 / 1835.469965278441013; /* temporary value in proton mass */
1150: ctx->n[0] = 1;
1151: ctx->v_0 = 1; /* thermal velocity, we could start with a scale != 1 */
1152: ctx->thermal_temps[0] = 1;
1153: /* constants, etc. */
1154: ctx->epsilon0 = 8.8542e-12; /* permittivity of free space (MKS) F/m */
1155: ctx->k = 1.38064852e-23; /* Boltzmann constant (MKS) J/K */
1156: ctx->n_0 = 1.e20; /* typical plasma n, but could set it to 1 */
1157: ctx->Ez = 0;
1158: for (PetscInt grid = 0; grid < LANDAU_NUM_TIMERS; grid++) ctx->times[grid] = 0;
1159: for (PetscInt ii = 0; ii < LANDAU_DIM; ii++) ctx->cells0[ii] = 2;
1160: if (LANDAU_DIM == 2) ctx->cells0[0] = 1;
1161: ctx->use_matrix_mass = PETSC_FALSE;
1162: ctx->use_relativistic_corrections = PETSC_FALSE;
1163: ctx->use_energy_tensor_trick = PETSC_FALSE; /* Use Eero's trick for energy conservation v --> grad(v^2/2) */
1164: ctx->SData_d.w = NULL;
1165: ctx->SData_d.x = NULL;
1166: ctx->SData_d.y = NULL;
1167: ctx->SData_d.z = NULL;
1168: ctx->SData_d.invJ = NULL;
1169: ctx->jacobian_field_major_order = PETSC_FALSE;
1170: ctx->SData_d.coo_elem_offsets = NULL;
1171: ctx->SData_d.coo_elem_point_offsets = NULL;
1172: ctx->SData_d.coo_elem_fullNb = NULL;
1173: ctx->SData_d.coo_size = 0;
1174: PetscOptionsBegin(ctx->comm, prefix, "Options for Fokker-Plank-Landau collision operator", "none");
1175: #if defined(PETSC_HAVE_KOKKOS)
1176: ctx->deviceType = LANDAU_KOKKOS;
1177: PetscCall(PetscStrncpy(ctx->filename, "kokkos", sizeof(ctx->filename)));
1178: #else
1179: ctx->deviceType = LANDAU_CPU;
1180: PetscCall(PetscStrncpy(ctx->filename, "cpu", sizeof(ctx->filename)));
1181: #endif
1182: PetscCall(PetscOptionsString("-dm_landau_device_type", "Use kernels on 'cpu' 'kokkos'", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), NULL));
1183: PetscCall(PetscStrcmp("cpu", ctx->filename, &flg));
1184: if (flg) {
1185: ctx->deviceType = LANDAU_CPU;
1186: } else {
1187: PetscCall(PetscStrcmp("kokkos", ctx->filename, &flg));
1188: if (flg) ctx->deviceType = LANDAU_KOKKOS;
1189: else SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_device_type %s", ctx->filename);
1190: }
1191: ctx->filename[0] = '\0';
1192: PetscCall(PetscOptionsString("-dm_landau_filename", "file to read mesh from", "plexland.c", ctx->filename, ctx->filename, sizeof(ctx->filename), &fileflg));
1193: PetscCall(PetscOptionsReal("-dm_landau_electron_shift", "Shift in thermal velocity of electrons", "none", ctx->electronShift, &ctx->electronShift, NULL));
1194: PetscCall(PetscOptionsInt("-dm_landau_verbose", "Level of verbosity output", "plexland.c", ctx->verbose, &ctx->verbose, NULL));
1195: PetscCall(PetscOptionsInt("-dm_landau_batch_size", "Number of 'vertices' to batch", "ex2.c", ctx->batch_sz, &ctx->batch_sz, NULL));
1196: PetscCheck(LANDAU_MAX_BATCH_SZ >= ctx->batch_sz, ctx->comm, PETSC_ERR_ARG_WRONG, "LANDAU_MAX_BATCH_SZ %" PetscInt_FMT " < ctx->batch_sz %" PetscInt_FMT, (PetscInt)LANDAU_MAX_BATCH_SZ, ctx->batch_sz);
1197: 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));
1198: 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);
1199: PetscCall(PetscOptionsReal("-dm_landau_Ez", "Initial parallel electric field in unites of Conner-Hastie critical field", "plexland.c", ctx->Ez, &ctx->Ez, NULL));
1200: PetscCall(PetscOptionsReal("-dm_landau_n_0", "Normalization constant for number density", "plexland.c", ctx->n_0, &ctx->n_0, NULL));
1201: 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));
1202: PetscCall(PetscOptionsBool("-dm_landau_use_relativistic_corrections", "Use relativistic corrections", "plexland.c", ctx->use_relativistic_corrections, &ctx->use_relativistic_corrections, NULL));
1203: PetscCall(PetscOptionsBool("-dm_landau_simplex", "Use simplex elements", "plexland.c", ctx->simplex, &ctx->simplex, NULL));
1204: PetscCall(PetscOptionsBool("-dm_landau_sphere", "use sphere/semi-circle domain instead of rectangle", "plexland.c", ctx->sphere, &ctx->sphere, NULL));
1205: if (LANDAU_DIM == 2 && ctx->use_relativistic_corrections) ctx->use_relativistic_corrections = PETSC_FALSE; // should warn
1206: 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,
1207: &ctx->use_energy_tensor_trick, NULL));
1209: /* get num species with temperature, set defaults */
1210: for (ii = 1; ii < LANDAU_MAX_SPECIES; ii++) {
1211: ctx->thermal_temps[ii] = 1;
1212: ctx->charges[ii] = 1;
1213: ctx->masses[ii] = 1;
1214: ctx->n[ii] = 1;
1215: }
1216: nt = LANDAU_MAX_SPECIES;
1217: 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));
1218: if (flg) {
1219: PetscCall(PetscInfo(dummy, "num_species set to number of thermal temps provided (%" PetscInt_FMT ")\n", nt));
1220: ctx->num_species = nt;
1221: } else SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_thermal_temps ,t1,t2,.. must be provided to set the number of species");
1222: for (ii = 0; ii < ctx->num_species; ii++) ctx->thermal_temps[ii] *= 1.1604525e7; /* convert to Kelvin */
1223: nm = LANDAU_MAX_SPECIES - 1;
1224: 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));
1225: 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);
1226: nm = LANDAU_MAX_SPECIES;
1227: PetscCall(PetscOptionsRealArray("-dm_landau_n", "Number density of each species = n_s * n_0", "plexland.c", ctx->n, &nm, &flg));
1228: 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);
1229: for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] *= 1.6720e-27; /* scale by proton mass kg */
1230: ctx->masses[0] = 9.10938356e-31; /* electron mass kg (should be about right already) */
1231: nc = LANDAU_MAX_SPECIES - 1;
1232: 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));
1233: 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);
1234: for (ii = 0; ii < LANDAU_MAX_SPECIES; ii++) ctx->charges[ii] *= 1.6022e-19; /* electron/proton charge (MKS) */
1235: /* geometry and grids */
1236: nt = LANDAU_MAX_GRIDS;
1237: 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));
1238: if (flg) {
1239: ctx->num_grids = nt;
1240: for (ii = nt = 0; ii < ctx->num_grids; ii++) nt += num_species_grid[ii];
1241: 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,
1242: ctx->num_grids, LANDAU_MAX_GRIDS);
1243: } else {
1244: if (ctx->num_species > LANDAU_MAX_GRIDS) {
1245: num_species_grid[0] = 1;
1246: num_species_grid[1] = ctx->num_species - 1;
1247: ctx->num_grids = 2;
1248: } else {
1249: ctx->num_grids = ctx->num_species;
1250: for (ii = 0; ii < ctx->num_grids; ii++) num_species_grid[ii] = 1;
1251: }
1252: }
1253: 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];
1254: 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],
1255: ctx->num_species);
1256: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1257: PetscInt iii = ctx->species_offset[grid]; // normalize with first (arbitrary) species on grid
1258: ctx->thermal_speed[grid] = PetscSqrtReal(ctx->k * ctx->thermal_temps[iii] / ctx->masses[iii]); /* arbitrary units for non-dimensionalization: plasma formulary def */
1259: }
1260: // get lambdas here because we need them for t_0 etc
1261: PetscCall(PetscOptionsReal("-dm_landau_ln_lambda", "Universal cross section parameter. Default uses NRL formulas", "plexland.c", lnLam, &lnLam, &flg));
1262: if (flg) {
1263: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1264: for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) ctx->lambdas[gridj][grid] = lnLam; /* cross section ratio large - small angle collisions */
1265: }
1266: } else {
1267: PetscCall(makeLambdas(ctx));
1268: }
1269: non_dim_grid = 0;
1270: 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));
1271: if (non_dim_grid != 0) PetscCall(PetscInfo(dummy, "Normalization grid set to %" PetscInt_FMT ", but non-default not well verified\n", non_dim_grid));
1272: 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);
1273: ctx->v_0 = ctx->thermal_speed[non_dim_grid]; /* arbitrary units for non dimensionalization: global mean velocity in 1D of electrons */
1274: ctx->m_0 = ctx->masses[non_dim_grid]; /* arbitrary reference mass, electrons */
1275: 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 */
1276: /* domain */
1277: nt = LANDAU_MAX_GRIDS;
1278: PetscCall(PetscOptionsRealArray("-dm_landau_domain_radius", "Phase space size in units of thermal velocity of grid", "plexland.c", ctx->radius, &nt, &flg));
1279: if (flg) {
1280: 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);
1281: while (nt--) ctx->radius_par[nt] = ctx->radius_perp[nt] = ctx->radius[nt];
1282: } else {
1283: nt = LANDAU_MAX_GRIDS;
1284: 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));
1285: 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);
1286: 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));
1287: 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);
1288: }
1289: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1290: if (flg && ctx->radius[grid] <= 0) { /* negative is ratio of c - need to set par and perp with this -- todo */
1291: if (ctx->radius[grid] == 0) ctx->radius[grid] = 0.75;
1292: else ctx->radius[grid] = -ctx->radius[grid];
1293: 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)
1294: PetscCall(PetscInfo(dummy, "Change domain radius to %g for grid %" PetscInt_FMT "\n", (double)ctx->radius[grid], grid));
1295: }
1296: ctx->radius[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1297: ctx->radius_perp[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1298: ctx->radius_par[grid] *= ctx->thermal_speed[grid] / ctx->v_0; // scale domain by thermal radius relative to v_0
1299: }
1300: /* amr parameters */
1301: if (!fileflg) {
1302: nt = LANDAU_MAX_GRIDS;
1303: 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));
1304: 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);
1305: nt = LANDAU_MAX_GRIDS;
1306: PetscCall(PetscOptionsIntArray("-dm_landau_amr_post_refine", "Number of levels to uniformly refine after AMR", "plexland.c", ctx->postAMRRefine, &nt, &flg));
1307: for (ii = 1; ii < ctx->num_grids; ii++) ctx->postAMRRefine[ii] = ctx->postAMRRefine[0]; // all grids the same now
1308: 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));
1309: 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));
1310: 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));
1311: 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));
1312: 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));
1313: 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));
1314: /* spherical domain */
1315: if (ctx->sphere || ctx->simplex) {
1316: ctx->sphere_uniform_normal = PETSC_FALSE;
1317: 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));
1318: if (!ctx->sphere_uniform_normal) { // true
1319: nt = LANDAU_MAX_GRIDS;
1320: 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));
1321: if (flg && nt < ctx->num_grids) {
1322: for (PetscInt grid = nt; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = ctx->sphere_inner_radius_90degree[0];
1323: } else if (!flg || nt == 0) {
1324: if (LANDAU_DIM == 2) {
1325: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.4; // optimized for R=5, Q4, AMR=0
1326: } else {
1327: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_90degree[grid] = 0.577 * 0.40;
1328: }
1329: }
1330: nt = LANDAU_MAX_GRIDS;
1331: 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));
1332: if (flg && nt < ctx->num_grids) {
1333: for (PetscInt grid = nt; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = ctx->sphere_inner_radius_45degree[0];
1334: } else if (!flg || nt == 0) {
1335: if (LANDAU_DIM == 2) {
1336: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = 0.45; // optimized for R=5, Q4, AMR=0
1337: } else {
1338: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ctx->sphere_inner_radius_45degree[grid] = 0.4; // 3D sphere
1339: }
1340: }
1341: 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]));
1342: } else {
1343: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1344: switch (ctx->numAMRRefine[grid]) {
1345: case 0:
1346: case 1:
1347: case 2:
1348: case 3:
1349: default:
1350: if (LANDAU_DIM == 2) {
1351: ctx->sphere_inner_radius_90degree[grid] = 0.40;
1352: ctx->sphere_inner_radius_45degree[grid] = 0.45;
1353: } else {
1354: ctx->sphere_inner_radius_45degree[grid] = 0.25;
1355: }
1356: }
1357: }
1358: }
1359: } else {
1360: nt = LANDAU_DIM;
1361: PetscCall(PetscOptionsIntArray("-dm_landau_num_cells", "Number of cells in each dimension of base grid", "plexland.c", ctx->cells0, &nt, &flg));
1362: }
1363: }
1364: /* processing options */
1365: PetscCall(PetscOptionsBool("-dm_landau_gpu_assembly", "Assemble Jacobian on GPU", "plexland.c", ctx->gpu_assembly, &ctx->gpu_assembly, NULL));
1366: 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));
1367: 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");
1368: PetscCheck(!ctx->jacobian_field_major_order, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED");
1369: PetscOptionsEnd();
1371: for (ii = ctx->num_species; ii < LANDAU_MAX_SPECIES; ii++) ctx->masses[ii] = ctx->thermal_temps[ii] = ctx->charges[ii] = 0;
1372: if (ctx->verbose != 0) {
1373: 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] / 1.6720e-27), (double)(ctx->num_species > 2 ? ctx->masses[2] / 1.6720e-27 : 0)));
1374: 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)));
1375: 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)));
1376: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "thermal T (K): e=%10.3e i=%10.3e %10.3e. Normalization grid %d: 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],
1377: (double)ctx->thermal_temps[1], (double)((ctx->num_species > 2) ? ctx->thermal_temps[2] : 0), (int)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));
1378: 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]));
1379: 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]));
1380: if (ctx->use_relativistic_corrections) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nUse relativistic corrections\n"));
1381: else PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1382: }
1383: PetscCall(DMDestroy(&dummy));
1384: {
1385: PetscMPIInt rank;
1386: PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank));
1387: ctx->stage = 0;
1388: PetscCall(PetscLogEventRegister("Landau Create", DM_CLASSID, &ctx->events[13])); /* 13 */
1389: PetscCall(PetscLogEventRegister(" GPU ass. setup", DM_CLASSID, &ctx->events[2])); /* 2 */
1390: PetscCall(PetscLogEventRegister(" Build matrix", DM_CLASSID, &ctx->events[12])); /* 12 */
1391: PetscCall(PetscLogEventRegister(" Assembly maps", DM_CLASSID, &ctx->events[15])); /* 15 */
1392: PetscCall(PetscLogEventRegister("Landau Mass mat", DM_CLASSID, &ctx->events[14])); /* 14 */
1393: PetscCall(PetscLogEventRegister("Landau Operator", DM_CLASSID, &ctx->events[11])); /* 11 */
1394: PetscCall(PetscLogEventRegister("Landau Jacobian", DM_CLASSID, &ctx->events[0])); /* 0 */
1395: PetscCall(PetscLogEventRegister("Landau Mass", DM_CLASSID, &ctx->events[9])); /* 9 */
1396: PetscCall(PetscLogEventRegister(" Preamble", DM_CLASSID, &ctx->events[10])); /* 10 */
1397: PetscCall(PetscLogEventRegister(" static IP Data", DM_CLASSID, &ctx->events[7])); /* 7 */
1398: PetscCall(PetscLogEventRegister(" dynamic IP-Jac", DM_CLASSID, &ctx->events[1])); /* 1 */
1399: PetscCall(PetscLogEventRegister(" Kernel-init", DM_CLASSID, &ctx->events[3])); /* 3 */
1400: PetscCall(PetscLogEventRegister(" Jac-f-df (GPU)", DM_CLASSID, &ctx->events[8])); /* 8 */
1401: PetscCall(PetscLogEventRegister(" J Kernel (GPU)", DM_CLASSID, &ctx->events[4])); /* 4 */
1402: PetscCall(PetscLogEventRegister(" M Kernel (GPU)", DM_CLASSID, &ctx->events[16])); /* 16 */
1403: PetscCall(PetscLogEventRegister(" Copy to CPU", DM_CLASSID, &ctx->events[5])); /* 5 */
1404: PetscCall(PetscLogEventRegister(" CPU assemble", DM_CLASSID, &ctx->events[6])); /* 6 */
1406: if (rank) { /* turn off output stuff for duplicate runs - do we need to add the prefix to all this? */
1407: PetscCall(PetscOptionsClearValue(NULL, "-snes_converged_reason"));
1408: PetscCall(PetscOptionsClearValue(NULL, "-ksp_converged_reason"));
1409: PetscCall(PetscOptionsClearValue(NULL, "-snes_monitor"));
1410: PetscCall(PetscOptionsClearValue(NULL, "-ksp_monitor"));
1411: PetscCall(PetscOptionsClearValue(NULL, "-ts_monitor"));
1412: PetscCall(PetscOptionsClearValue(NULL, "-ts_view"));
1413: PetscCall(PetscOptionsClearValue(NULL, "-ts_adapt_monitor"));
1414: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_dm_view"));
1415: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_amr_vec_view"));
1416: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_dm_view"));
1417: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mass_view"));
1418: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_jacobian_view"));
1419: PetscCall(PetscOptionsClearValue(NULL, "-dm_landau_mat_view"));
1420: PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_converged_reason"));
1421: PetscCall(PetscOptionsClearValue(NULL, "-pc_bjkokkos_ksp_monitor"));
1422: PetscCall(PetscOptionsClearValue(NULL, "-"));
1423: PetscCall(PetscOptionsClearValue(NULL, "-info"));
1424: }
1425: }
1426: PetscFunctionReturn(PETSC_SUCCESS);
1427: }
1429: static PetscErrorCode CreateStaticData(PetscInt dim, IS grid_batch_is_inv[], LandauCtx *ctx)
1430: {
1431: PetscSection section[LANDAU_MAX_GRIDS], globsection[LANDAU_MAX_GRIDS];
1432: PetscQuadrature quad;
1433: const PetscReal *quadWeights;
1434: PetscReal invMass[LANDAU_MAX_SPECIES], nu_alpha[LANDAU_MAX_SPECIES], nu_beta[LANDAU_MAX_SPECIES];
1435: 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;
1436: PetscTabulation *Tf;
1437: PetscDS prob;
1439: PetscFunctionBegin;
1440: PetscCall(PetscFEGetDimension(ctx->fe[0], &Nb));
1441: 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);
1442: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1443: for (PetscInt ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++) {
1444: invMass[ii] = ctx->m_0 / ctx->masses[ii];
1445: nu_alpha[ii] = PetscSqr(ctx->charges[ii] / ctx->m_0) * ctx->m_0 / ctx->masses[ii];
1446: nu_beta[ii] = PetscSqr(ctx->charges[ii] / ctx->epsilon0) / (8 * PETSC_PI) * ctx->t_0 * ctx->n_0 / PetscPowReal(ctx->v_0, 3);
1447: }
1448: }
1449: if (ctx->verbose == 4) {
1450: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "nu_alpha: "));
1451: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1452: PetscInt iii = ctx->species_offset[grid];
1453: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_alpha[ii]));
1454: }
1455: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_beta: "));
1456: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1457: PetscInt iii = ctx->species_offset[grid];
1458: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %e", (double)nu_beta[ii]));
1459: }
1460: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\nnu_alpha[i]*nu_beta[j]*lambda[i][j]:\n"));
1461: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1462: PetscInt iii = ctx->species_offset[grid];
1463: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) {
1464: for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) {
1465: PetscInt jjj = ctx->species_offset[gridj];
1466: 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])));
1467: }
1468: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1469: }
1470: }
1471: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "lambda[i][j]:\n"));
1472: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1473: PetscInt iii = ctx->species_offset[grid];
1474: for (PetscInt ii = iii; ii < ctx->species_offset[grid + 1]; ii++) {
1475: for (PetscInt gridj = 0; gridj < ctx->num_grids; gridj++) {
1476: PetscInt jjj = ctx->species_offset[gridj];
1477: for (PetscInt jj = jjj; jj < ctx->species_offset[gridj + 1]; jj++) PetscCall(PetscPrintf(PETSC_COMM_WORLD, " %14.9e", (double)ctx->lambdas[grid][gridj]));
1478: }
1479: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
1480: }
1481: }
1482: }
1483: PetscCall(DMGetDS(ctx->plex[0], &prob)); // same DS for all grids
1484: PetscCall(PetscDSGetTabulation(prob, &Tf)); // Bf, &Df same for all grids
1485: /* DS, Tab and quad is same on all grids */
1486: PetscCheck(ctx->plex[0], ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
1487: PetscCall(PetscFEGetQuadrature(ctx->fe[0], &quad));
1488: PetscCall(PetscQuadratureGetData(quad, NULL, NULL, &Nq, NULL, &quadWeights));
1489: 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);
1490: /* setup each grid */
1491: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1492: PetscInt cStart, cEnd;
1493: PetscCheck(ctx->plex[grid] != NULL, ctx->comm, PETSC_ERR_ARG_WRONG, "Plex not created");
1494: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1495: numCells[grid] = cEnd - cStart; // grids can have different topology
1496: PetscCall(DMGetLocalSection(ctx->plex[grid], §ion[grid]));
1497: PetscCall(DMGetGlobalSection(ctx->plex[grid], &globsection[grid]));
1498: PetscCall(PetscSectionGetNumFields(section[grid], &Nf[grid]));
1499: ncellsTot += numCells[grid];
1500: }
1501: /* create GPU assembly data */
1502: if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
1503: PetscContainer container;
1504: PetscScalar *elemMatrix, *elMat;
1505: pointInterpolationP4est(*pointMaps)[LANDAU_MAX_Q_FACE];
1506: P4estVertexMaps *maps;
1507: const PetscInt *plex_batch = NULL, elMatSz = Nb * Nb * ctx->num_species * ctx->num_species;
1508: LandauIdx *coo_elem_offsets = NULL, *coo_elem_fullNb = NULL, (*coo_elem_point_offsets)[LANDAU_MAX_NQND + 1] = NULL;
1509: /* create GPU assembly data */
1510: PetscCall(PetscInfo(ctx->plex[0], "Make GPU maps %d\n", 1));
1511: PetscCall(PetscLogEventBegin(ctx->events[2], 0, 0, 0, 0));
1512: PetscCall(PetscMalloc(sizeof(*maps) * ctx->num_grids, &maps));
1513: PetscCall(PetscMalloc(sizeof(*pointMaps) * MAP_BF_SIZE, &pointMaps));
1514: PetscCall(PetscMalloc(sizeof(*elemMatrix) * elMatSz, &elemMatrix));
1516: { // setup COO assembly -- put COO metadata directly in ctx->SData_d
1517: PetscCall(PetscMalloc3(ncellsTot + 1, &coo_elem_offsets, ncellsTot, &coo_elem_fullNb, ncellsTot, &coo_elem_point_offsets)); // array of integer pointers
1518: coo_elem_offsets[0] = 0; // finish later
1519: PetscCall(PetscInfo(ctx->plex[0], "COO initialization, %" PetscInt_FMT " cells\n", ncellsTot));
1520: ctx->SData_d.coo_n_cellsTot = ncellsTot;
1521: ctx->SData_d.coo_elem_offsets = (void *)coo_elem_offsets;
1522: ctx->SData_d.coo_elem_fullNb = (void *)coo_elem_fullNb;
1523: ctx->SData_d.coo_elem_point_offsets = (void *)coo_elem_point_offsets;
1524: }
1526: ctx->SData_d.coo_max_fullnb = 0;
1527: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1528: PetscInt cStart, cEnd, Nfloc = Nf[grid], totDim = Nfloc * Nb;
1529: if (grid_batch_is_inv[grid]) PetscCall(ISGetIndices(grid_batch_is_inv[grid], &plex_batch));
1530: PetscCheck(!plex_batch, ctx->comm, PETSC_ERR_ARG_WRONG, "-dm_landau_jacobian_field_major_order DEPRECATED");
1531: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1532: // make maps
1533: maps[grid].d_self = NULL;
1534: maps[grid].num_elements = numCells[grid];
1535: maps[grid].num_face = (PetscInt)(pow(Nq, 1. / ((double)dim)) + .001); // Q
1536: maps[grid].num_face = (PetscInt)(pow(maps[grid].num_face, (double)(dim - 1)) + .001); // Q^2
1537: maps[grid].num_reduced = 0;
1538: maps[grid].deviceType = ctx->deviceType;
1539: maps[grid].numgrids = ctx->num_grids;
1540: // count reduced and get
1541: PetscCall(PetscMalloc(maps[grid].num_elements * sizeof(*maps[grid].gIdx), &maps[grid].gIdx));
1542: for (PetscInt ej = cStart, eidx = 0; ej < cEnd; ++ej, ++eidx, glb_elem_idx++) {
1543: if (coo_elem_offsets) coo_elem_offsets[glb_elem_idx + 1] = coo_elem_offsets[glb_elem_idx]; // start with last one, then add
1544: for (PetscInt fieldA = 0; fieldA < Nf[grid]; fieldA++) {
1545: PetscInt fullNb = 0;
1546: for (PetscInt q = 0; q < Nb; ++q) {
1547: PetscInt numindices, *indices;
1548: PetscScalar *valuesOrig = elMat = elemMatrix;
1549: PetscCall(PetscArrayzero(elMat, totDim * totDim));
1550: elMat[(fieldA * Nb + q) * totDim + fieldA * Nb + q] = 1;
1551: PetscCall(DMPlexGetClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, (PetscScalar **)&elMat));
1552: if (ctx->simplex) {
1553: PetscCheck(numindices == Nb, ctx->comm, PETSC_ERR_ARG_WRONG, "numindices != Nb numindices=%d Nb=%d", (int)numindices, (int)Nb);
1554: for (PetscInt q = 0; q < numindices; ++q) { maps[grid].gIdx[eidx][fieldA][q] = (LandauIdx)indices[q]; }
1555: fullNb++;
1556: } else {
1557: for (PetscInt f = 0; f < numindices; ++f) { // look for a non-zero on the diagonal (is this too complicated for simplices?)
1558: if (PetscAbs(PetscRealPart(elMat[f * numindices + f])) > PETSC_MACHINE_EPSILON) {
1559: // found it
1560: if (PetscAbs(PetscRealPart(elMat[f * numindices + f] - 1.)) < PETSC_MACHINE_EPSILON) { // normal vertex 1.0
1561: if (plex_batch) {
1562: maps[grid].gIdx[eidx][fieldA][q] = (LandauIdx)plex_batch[indices[f]];
1563: } else {
1564: maps[grid].gIdx[eidx][fieldA][q] = (LandauIdx)indices[f];
1565: }
1566: fullNb++;
1567: } else { //found a constraint
1568: PetscInt jj = 0;
1569: PetscReal sum = 0;
1570: const PetscInt ff = f;
1571: maps[grid].gIdx[eidx][fieldA][q] = -maps[grid].num_reduced - 1; // store (-)index: id = -(idx+1): idx = -id - 1
1572: PetscCheck(!ctx->simplex, ctx->comm, PETSC_ERR_ARG_WRONG, "No constraints with simplex");
1573: do { // constraints are continuous in Plex - exploit that here
1574: PetscInt ii; // get 'scale'
1575: for (ii = 0, pointMaps[maps[grid].num_reduced][jj].scale = 0; ii < maps[grid].num_face; ii++) { // sum row of outer product to recover vector value
1576: if (ff + ii < numindices) { // 3D has Q and Q^2 interps so might run off end. We could test that elMat[f*numindices + ff + ii] > 0, and break if not
1577: pointMaps[maps[grid].num_reduced][jj].scale += PetscRealPart(elMat[f * numindices + ff + ii]);
1578: }
1579: }
1580: sum += pointMaps[maps[grid].num_reduced][jj].scale; // diagnostic
1581: // get 'gid'
1582: if (pointMaps[maps[grid].num_reduced][jj].scale == 0) pointMaps[maps[grid].num_reduced][jj].gid = -1; // 3D has Q and Q^2 interps
1583: else {
1584: if (plex_batch) {
1585: pointMaps[maps[grid].num_reduced][jj].gid = plex_batch[indices[f]];
1586: } else {
1587: pointMaps[maps[grid].num_reduced][jj].gid = indices[f];
1588: }
1589: fullNb++;
1590: }
1591: } while (++jj < maps[grid].num_face && ++f < numindices); // jj is incremented if we hit the end
1592: while (jj < maps[grid].num_face) {
1593: pointMaps[maps[grid].num_reduced][jj].scale = 0;
1594: pointMaps[maps[grid].num_reduced][jj].gid = -1;
1595: jj++;
1596: }
1597: if (PetscAbs(sum - 1.0) > 10 * PETSC_MACHINE_EPSILON) { // debug
1598: PetscInt d, f;
1599: PetscReal tmp = 0;
1600: PetscCall(PetscPrintf(PETSC_COMM_SELF, "\t\t%d.%d.%d) ERROR total I = %22.16e (LANDAU_MAX_Q_FACE=%d, #face=%d)\n", (int)eidx, (int)q, (int)fieldA, (double)sum, LANDAU_MAX_Q_FACE, (int)maps[grid].num_face));
1601: for (d = 0, tmp = 0; d < numindices; ++d) {
1602: if (tmp != 0 && PetscAbs(tmp - 1.0) > 10 * PETSC_MACHINE_EPSILON) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "%3d) %3" PetscInt_FMT ": ", (int)d, indices[d]));
1603: for (f = 0; f < numindices; ++f) tmp += PetscRealPart(elMat[d * numindices + f]);
1604: if (tmp != 0) PetscCall(PetscPrintf(ctx->comm, " | %22.16e\n", (double)tmp));
1605: }
1606: }
1607: maps[grid].num_reduced++;
1608: PetscCheck(maps[grid].num_reduced < MAP_BF_SIZE, PETSC_COMM_SELF, PETSC_ERR_PLIB, "maps[grid].num_reduced %d > %" PetscInt_FMT, (int)maps[grid].num_reduced, MAP_BF_SIZE);
1609: }
1610: break;
1611: }
1612: }
1613: } // !simplex
1614: // cleanup
1615: PetscCall(DMPlexRestoreClosureIndices(ctx->plex[grid], section[grid], globsection[grid], ej, PETSC_TRUE, &numindices, &indices, NULL, (PetscScalar **)&elMat));
1616: if (elMat != valuesOrig) PetscCall(DMRestoreWorkArray(ctx->plex[grid], numindices * numindices, MPIU_SCALAR, &elMat));
1617: }
1618: { // setup COO assembly
1619: coo_elem_offsets[glb_elem_idx + 1] += fullNb * fullNb; // one species block, adds a block for each species, on this element in this grid
1620: if (fieldA == 0) { // cache full Nb for this element, on this grid per species
1621: coo_elem_fullNb[glb_elem_idx] = fullNb;
1622: if (fullNb > ctx->SData_d.coo_max_fullnb) ctx->SData_d.coo_max_fullnb = fullNb;
1623: } else PetscCheck(coo_elem_fullNb[glb_elem_idx] == fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "full element size change with species %d %d", (int)coo_elem_fullNb[glb_elem_idx], (int)fullNb);
1624: }
1625: } // field
1626: } // cell
1627: // allocate and copy point data maps[grid].gIdx[eidx][field][q]
1628: PetscCall(PetscMalloc(maps[grid].num_reduced * sizeof(*maps[grid].c_maps), &maps[grid].c_maps));
1629: for (PetscInt ej = 0; ej < maps[grid].num_reduced; ++ej) {
1630: for (PetscInt q = 0; q < maps[grid].num_face; ++q) {
1631: maps[grid].c_maps[ej][q].scale = pointMaps[ej][q].scale;
1632: maps[grid].c_maps[ej][q].gid = pointMaps[ej][q].gid;
1633: }
1634: }
1635: #if defined(PETSC_HAVE_KOKKOS)
1636: if (ctx->deviceType == LANDAU_KOKKOS) {
1637: PetscCall(LandauKokkosCreateMatMaps(maps, pointMaps, Nf, grid)); // implies Kokkos does
1638: }
1639: #endif
1640: if (plex_batch) {
1641: PetscCall(ISRestoreIndices(grid_batch_is_inv[grid], &plex_batch));
1642: PetscCall(ISDestroy(&grid_batch_is_inv[grid])); // we are done with this
1643: }
1644: } /* grids */
1645: // finish COO
1646: { // setup COO assembly
1647: PetscInt *oor, *ooc;
1648: ctx->SData_d.coo_size = coo_elem_offsets[ncellsTot] * ctx->batch_sz;
1649: PetscCall(PetscMalloc2(ctx->SData_d.coo_size, &oor, ctx->SData_d.coo_size, &ooc));
1650: for (PetscInt i = 0; i < ctx->SData_d.coo_size; i++) oor[i] = ooc[i] = -1;
1651: // get
1652: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1653: for (PetscInt ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) {
1654: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx];
1655: 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
1656: coo_elem_point_offsets[glb_elem_idx][0] = 0;
1657: for (PetscInt f = 0, cnt2 = 0; f < Nb; f++) {
1658: PetscInt idx = Idxs[f];
1659: coo_elem_point_offsets[glb_elem_idx][f + 1] = coo_elem_point_offsets[glb_elem_idx][f]; // start at last
1660: if (idx >= 0) {
1661: cnt2++;
1662: coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc
1663: } else {
1664: idx = -idx - 1;
1665: for (PetscInt q = 0; q < maps[grid].num_face; q++) {
1666: if (maps[grid].c_maps[idx][q].gid < 0) break;
1667: cnt2++;
1668: coo_elem_point_offsets[glb_elem_idx][f + 1]++; // inc
1669: }
1670: }
1671: PetscCheck(cnt2 <= fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "wrong count %d < %d", (int)fullNb, (int)cnt2);
1672: }
1673: PetscCheck(coo_elem_point_offsets[glb_elem_idx][Nb] == fullNb, PETSC_COMM_SELF, PETSC_ERR_PLIB, "coo_elem_point_offsets size %d != fullNb=%d", (int)coo_elem_point_offsets[glb_elem_idx][Nb], (int)fullNb);
1674: }
1675: }
1676: // set
1677: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
1678: for (PetscInt grid = 0, glb_elem_idx = 0; grid < ctx->num_grids; grid++) {
1679: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
1680: for (PetscInt ej = 0; ej < numCells[grid]; ++ej, glb_elem_idx++) {
1681: const PetscInt fullNb = coo_elem_fullNb[glb_elem_idx], fullNb2 = fullNb * fullNb;
1682: // set (i,j)
1683: for (PetscInt fieldA = 0; fieldA < Nf[grid]; fieldA++) {
1684: const LandauIdx *const Idxs = &maps[grid].gIdx[ej][fieldA][0];
1685: PetscInt rows[LANDAU_MAX_Q_FACE], cols[LANDAU_MAX_Q_FACE];
1686: for (PetscInt f = 0; f < Nb; ++f) {
1687: const PetscInt nr = coo_elem_point_offsets[glb_elem_idx][f + 1] - coo_elem_point_offsets[glb_elem_idx][f];
1688: if (nr == 1) rows[0] = Idxs[f];
1689: else {
1690: const PetscInt idx = -Idxs[f] - 1;
1691: for (PetscInt q = 0; q < nr; q++) rows[q] = maps[grid].c_maps[idx][q].gid;
1692: }
1693: for (PetscInt g = 0; g < Nb; ++g) {
1694: const PetscInt nc = coo_elem_point_offsets[glb_elem_idx][g + 1] - coo_elem_point_offsets[glb_elem_idx][g];
1695: if (nc == 1) cols[0] = Idxs[g];
1696: else {
1697: const PetscInt idx = -Idxs[g] - 1;
1698: for (PetscInt q = 0; q < nc; q++) cols[q] = maps[grid].c_maps[idx][q].gid;
1699: }
1700: 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];
1701: for (PetscInt q = 0, idx = idx0; q < nr; q++) {
1702: for (PetscInt d = 0; d < nc; d++, idx++) {
1703: oor[idx] = rows[q] + moffset;
1704: ooc[idx] = cols[d] + moffset;
1705: }
1706: }
1707: }
1708: }
1709: }
1710: } // cell
1711: } // grid
1712: } // batch
1713: PetscCall(MatSetPreallocationCOO(ctx->J, ctx->SData_d.coo_size, oor, ooc));
1714: PetscCall(PetscFree2(oor, ooc));
1715: }
1716: PetscCall(PetscFree(pointMaps));
1717: PetscCall(PetscFree(elemMatrix));
1718: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
1719: PetscCall(PetscContainerSetPointer(container, (void *)maps));
1720: PetscCall(PetscContainerSetUserDestroy(container, LandauGPUMapsDestroy));
1721: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "assembly_maps", (PetscObject)container));
1722: PetscCall(PetscContainerDestroy(&container));
1723: PetscCall(PetscLogEventEnd(ctx->events[2], 0, 0, 0, 0));
1724: } // end GPU assembly
1725: { /* create static point data, Jacobian called first, only one vertex copy */
1726: PetscReal *invJe, *ww, *xx, *yy, *zz = NULL, *invJ_a;
1727: PetscInt outer_ipidx, outer_ej, grid, nip_glb = 0;
1728: PetscFE fe;
1729: PetscCall(PetscLogEventBegin(ctx->events[7], 0, 0, 0, 0));
1730: PetscCall(PetscInfo(ctx->plex[0], "Initialize static data\n"));
1731: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) nip_glb += Nq * numCells[grid];
1732: /* collect f data, first time is for Jacobian, but make mass now */
1733: if (ctx->verbose != 0) {
1734: PetscInt ncells = 0, N;
1735: MatInfo info;
1736: PetscCall(MatGetInfo(ctx->J, MAT_LOCAL, &info));
1737: PetscCall(MatGetSize(ctx->J, &N, NULL));
1738: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) ncells += numCells[grid];
1739: 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,
1740: ctx->num_species, Nb, dim, N, (PetscInt)info.nz_used));
1741: }
1742: PetscCall(PetscMalloc4(nip_glb, &ww, nip_glb, &xx, nip_glb, &yy, nip_glb * dim * dim, &invJ_a));
1743: if (dim == 3) PetscCall(PetscMalloc1(nip_glb, &zz));
1744: if (ctx->use_energy_tensor_trick) {
1745: PetscCall(PetscFECreateDefault(PETSC_COMM_SELF, dim, 1, ctx->simplex, NULL, PETSC_DECIDE, &fe));
1746: PetscCall(PetscObjectSetName((PetscObject)fe, "energy"));
1747: }
1748: /* init each grids static data - no batch */
1749: for (grid = 0, outer_ipidx = 0, outer_ej = 0; grid < ctx->num_grids; grid++) { // OpenMP (once)
1750: Vec v2_2 = NULL; // projected function: v^2/2 for non-relativistic, gamma... for relativistic
1751: PetscSection e_section;
1752: DM dmEnergy;
1753: PetscInt cStart, cEnd, ej;
1755: PetscCall(DMPlexGetHeightStratum(ctx->plex[grid], 0, &cStart, &cEnd));
1756: // prep energy trick, get v^2 / 2 vector
1757: if (ctx->use_energy_tensor_trick) {
1758: PetscErrorCode (*energyf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {ctx->use_relativistic_corrections ? gamma_m1_f : energy_f};
1759: Vec glob_v2;
1760: PetscReal *c2_0[1], data[1] = {PetscSqr(C_0(ctx->v_0))};
1762: PetscCall(DMClone(ctx->plex[grid], &dmEnergy));
1763: PetscCall(PetscObjectSetName((PetscObject)dmEnergy, "energy"));
1764: PetscCall(DMSetField(dmEnergy, 0, NULL, (PetscObject)fe));
1765: PetscCall(DMCreateDS(dmEnergy));
1766: PetscCall(DMGetSection(dmEnergy, &e_section));
1767: PetscCall(DMGetGlobalVector(dmEnergy, &glob_v2));
1768: PetscCall(PetscObjectSetName((PetscObject)glob_v2, "trick"));
1769: c2_0[0] = &data[0];
1770: PetscCall(DMProjectFunction(dmEnergy, 0., energyf, (void **)c2_0, INSERT_ALL_VALUES, glob_v2));
1771: PetscCall(DMGetLocalVector(dmEnergy, &v2_2));
1772: PetscCall(VecZeroEntries(v2_2)); /* zero BCs so don't set */
1773: PetscCall(DMGlobalToLocalBegin(dmEnergy, glob_v2, INSERT_VALUES, v2_2));
1774: PetscCall(DMGlobalToLocalEnd(dmEnergy, glob_v2, INSERT_VALUES, v2_2));
1775: PetscCall(DMViewFromOptions(dmEnergy, NULL, "-energy_dm_view"));
1776: PetscCall(VecViewFromOptions(glob_v2, NULL, "-energy_vec_view"));
1777: PetscCall(DMRestoreGlobalVector(dmEnergy, &glob_v2));
1778: }
1779: /* append part of the IP data for each grid */
1780: for (ej = 0; ej < numCells[grid]; ++ej, ++outer_ej) {
1781: PetscScalar *coefs = NULL;
1782: 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);
1783: invJe = invJ_a + outer_ej * Nq * dim * dim;
1784: PetscCall(DMPlexComputeCellGeometryFEM(ctx->plex[grid], ej + cStart, quad, vj, Jdummy, invJe, detJj));
1785: if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecGetClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs));
1786: /* create static point data */
1787: for (PetscInt qj = 0; qj < Nq; qj++, outer_ipidx++) {
1788: const PetscInt gidx = outer_ipidx;
1789: const PetscReal *invJ = &invJe[qj * dim * dim];
1790: ww[gidx] = detJj[qj] * quadWeights[qj];
1791: if (dim == 2) ww[gidx] *= vj[qj * dim + 0]; /* cylindrical coordinate, w/o 2pi */
1792: // get xx, yy, zz
1793: if (ctx->use_energy_tensor_trick) {
1794: double refSpaceDer[3], eGradPhi[3];
1795: const PetscReal *const DD = Tf[0]->T[1];
1796: const PetscReal *Dq = &DD[qj * Nb * dim];
1797: for (PetscInt d = 0; d < 3; ++d) refSpaceDer[d] = eGradPhi[d] = 0.0;
1798: for (PetscInt b = 0; b < Nb; ++b) {
1799: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] += Dq[b * dim + d] * PetscRealPart(coefs[b]);
1800: }
1801: xx[gidx] = 1e10;
1802: if (ctx->use_relativistic_corrections) {
1803: double dg2_c2 = 0;
1804: //for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] *= c02;
1805: for (PetscInt d = 0; d < dim; ++d) dg2_c2 += PetscSqr(refSpaceDer[d]);
1806: dg2_c2 *= (double)c02;
1807: if (dg2_c2 >= .999) {
1808: xx[gidx] = vj[qj * dim + 0]; /* coordinate */
1809: yy[gidx] = vj[qj * dim + 1];
1810: if (dim == 3) zz[gidx] = vj[qj * dim + 2];
1811: 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]));
1812: } else {
1813: PetscReal fact = c02 / PetscSqrtReal(1. - dg2_c2);
1814: for (PetscInt d = 0; d < dim; ++d) refSpaceDer[d] *= fact;
1815: // could test with other point u' that (grad - grad') * U (refSpaceDer, refSpaceDer') == 0
1816: }
1817: }
1818: if (xx[gidx] == 1e10) {
1819: for (PetscInt d = 0; d < dim; ++d) {
1820: for (PetscInt e = 0; e < dim; ++e) eGradPhi[d] += invJ[e * dim + d] * refSpaceDer[e];
1821: }
1822: xx[gidx] = eGradPhi[0];
1823: yy[gidx] = eGradPhi[1];
1824: if (dim == 3) zz[gidx] = eGradPhi[2];
1825: }
1826: } else {
1827: xx[gidx] = vj[qj * dim + 0]; /* coordinate */
1828: yy[gidx] = vj[qj * dim + 1];
1829: if (dim == 3) zz[gidx] = vj[qj * dim + 2];
1830: }
1831: } /* q */
1832: if (ctx->use_energy_tensor_trick) PetscCall(DMPlexVecRestoreClosure(dmEnergy, e_section, v2_2, ej + cStart, NULL, &coefs));
1833: } /* ej */
1834: if (ctx->use_energy_tensor_trick) {
1835: PetscCall(DMRestoreLocalVector(dmEnergy, &v2_2));
1836: PetscCall(DMDestroy(&dmEnergy));
1837: }
1838: } /* grid */
1839: if (ctx->use_energy_tensor_trick) PetscCall(PetscFEDestroy(&fe));
1840: /* cache static data */
1841: if (ctx->deviceType == LANDAU_KOKKOS) {
1842: #if defined(PETSC_HAVE_KOKKOS)
1843: 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));
1844: /* free */
1845: PetscCall(PetscFree4(ww, xx, yy, invJ_a));
1846: if (dim == 3) PetscCall(PetscFree(zz));
1847: #else
1848: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type kokkos not built");
1849: #endif
1850: } else { /* CPU version, just copy in, only use part */
1851: 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 ?
1852: ctx->SData_d.w = (void *)ww;
1853: ctx->SData_d.x = (void *)xx;
1854: ctx->SData_d.y = (void *)yy;
1855: ctx->SData_d.z = (void *)zz;
1856: ctx->SData_d.invJ = (void *)invJ_a;
1857: 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));
1858: for (PetscInt ii = 0; ii < ctx->num_species; ii++) {
1859: nu_alpha_p[ii] = nu_alpha[ii];
1860: nu_beta_p[ii] = nu_beta[ii];
1861: invMass_p[ii] = invMass[ii];
1862: }
1863: ctx->SData_d.alpha = (void *)nu_alpha_p;
1864: ctx->SData_d.beta = (void *)nu_beta_p;
1865: ctx->SData_d.invMass = (void *)invMass_p;
1866: ctx->SData_d.lambdas = (void *)lambdas_p;
1867: for (PetscInt grid = 0; grid < LANDAU_MAX_GRIDS; grid++) {
1868: PetscReal(*lambdas)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS] = (PetscReal(*)[LANDAU_MAX_GRIDS][LANDAU_MAX_GRIDS])ctx->SData_d.lambdas;
1869: for (PetscInt gridj = 0; gridj < LANDAU_MAX_GRIDS; gridj++) { (*lambdas)[grid][gridj] = ctx->lambdas[grid][gridj]; }
1870: }
1871: }
1872: PetscCall(PetscLogEventEnd(ctx->events[7], 0, 0, 0, 0));
1873: } // initialize
1874: PetscFunctionReturn(PETSC_SUCCESS);
1875: }
1877: /* < v, u > */
1878: 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[])
1879: {
1880: g0[0] = 1.;
1881: }
1883: /* < v, u > */
1884: 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[])
1885: {
1886: static double ttt = 1e-12;
1887: g0[0] = ttt++;
1888: }
1890: /* < v, u > */
1891: 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[])
1892: {
1893: g0[0] = 2. * PETSC_PI * x[0];
1894: }
1896: static PetscErrorCode MatrixNfDestroy(void *ptr)
1897: {
1898: PetscInt *nf = (PetscInt *)ptr;
1900: PetscFunctionBegin;
1901: PetscCall(PetscFree(nf));
1902: PetscFunctionReturn(PETSC_SUCCESS);
1903: }
1905: /*
1906: LandauCreateJacobianMatrix - creates ctx->J with without real data. Hard to keep sparse.
1907: - Like DMPlexLandauCreateMassMatrix. Should remove one and combine
1908: - has old support for field major ordering
1909: */
1910: static PetscErrorCode LandauCreateJacobianMatrix(MPI_Comm comm, Vec X, IS grid_batch_is_inv[LANDAU_MAX_GRIDS], LandauCtx *ctx)
1911: {
1912: PetscInt *idxs = NULL;
1913: Mat subM[LANDAU_MAX_GRIDS];
1915: PetscFunctionBegin;
1916: if (!ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
1917: PetscFunctionReturn(PETSC_SUCCESS);
1918: }
1919: // get the RCM for this grid to separate out species into blocks -- create 'idxs' & 'ctx->batch_is' -- not used
1920: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) PetscCall(PetscMalloc1(ctx->mat_offset[ctx->num_grids] * ctx->batch_sz, &idxs));
1921: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1922: const PetscInt *values, n = ctx->mat_offset[grid + 1] - ctx->mat_offset[grid];
1923: Mat gMat;
1924: DM massDM;
1925: PetscDS prob;
1926: Vec tvec;
1927: // get "mass" matrix for reordering
1928: PetscCall(DMClone(ctx->plex[grid], &massDM));
1929: PetscCall(DMCopyFields(ctx->plex[grid], PETSC_DETERMINE, PETSC_DETERMINE, massDM));
1930: PetscCall(DMCreateDS(massDM));
1931: PetscCall(DMGetDS(massDM, &prob));
1932: 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));
1933: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only")); // this trick is need to both sparsify the matrix and avoid runtime error
1934: PetscCall(DMCreateMatrix(massDM, &gMat));
1935: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
1936: PetscCall(MatSetOption(gMat, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
1937: PetscCall(MatSetOption(gMat, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
1938: PetscCall(DMCreateLocalVector(ctx->plex[grid], &tvec));
1939: PetscCall(DMPlexSNESComputeJacobianFEM(massDM, tvec, gMat, gMat, ctx));
1940: PetscCall(MatViewFromOptions(gMat, NULL, "-dm_landau_reorder_mat_view"));
1941: PetscCall(DMDestroy(&massDM));
1942: PetscCall(VecDestroy(&tvec));
1943: subM[grid] = gMat;
1944: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
1945: MatOrderingType rtype = MATORDERINGRCM;
1946: IS isrow, isicol;
1947: PetscCall(MatGetOrdering(gMat, rtype, &isrow, &isicol));
1948: PetscCall(ISInvertPermutation(isrow, PETSC_DECIDE, &grid_batch_is_inv[grid]));
1949: PetscCall(ISGetIndices(isrow, &values));
1950: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
1951: #if !defined(LANDAU_SPECIES_MAJOR)
1952: PetscInt N = ctx->mat_offset[ctx->num_grids], n0 = ctx->mat_offset[grid] + b_id * N;
1953: for (PetscInt ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0;
1954: #else
1955: PetscInt n0 = ctx->mat_offset[grid] * ctx->batch_sz + b_id * n;
1956: for (PetscInt ii = 0; ii < n; ++ii) idxs[n0 + ii] = values[ii] + n0;
1957: #endif
1958: }
1959: PetscCall(ISRestoreIndices(isrow, &values));
1960: PetscCall(ISDestroy(&isrow));
1961: PetscCall(ISDestroy(&isicol));
1962: }
1963: }
1964: 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));
1965: // get a block matrix
1966: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
1967: Mat B = subM[grid];
1968: PetscInt nloc, nzl, *colbuf, row, COL_BF_SIZE = 1024;
1969: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
1970: PetscCall(MatGetSize(B, &nloc, NULL));
1971: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
1972: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
1973: const PetscInt *cols;
1974: const PetscScalar *vals;
1975: for (PetscInt i = 0; i < nloc; i++) {
1976: PetscCall(MatGetRow(B, i, &nzl, NULL, NULL));
1977: if (nzl > COL_BF_SIZE) {
1978: PetscCall(PetscFree(colbuf));
1979: 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));
1980: COL_BF_SIZE = nzl;
1981: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
1982: }
1983: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
1984: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset;
1985: row = i + moffset;
1986: PetscCall(MatSetValues(ctx->J, 1, &row, nzl, colbuf, vals, INSERT_VALUES));
1987: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
1988: }
1989: }
1990: PetscCall(PetscFree(colbuf));
1991: }
1992: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid]));
1993: PetscCall(MatAssemblyBegin(ctx->J, MAT_FINAL_ASSEMBLY));
1994: PetscCall(MatAssemblyEnd(ctx->J, MAT_FINAL_ASSEMBLY));
1996: // debug
1997: PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_mat_view"));
1998: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
1999: Mat mat_block_order;
2000: PetscCall(MatCreateSubMatrix(ctx->J, ctx->batch_is, ctx->batch_is, MAT_INITIAL_MATRIX, &mat_block_order)); // use MatPermute
2001: PetscCall(MatViewFromOptions(mat_block_order, NULL, "-dm_landau_mat_view"));
2002: PetscCall(MatDestroy(&mat_block_order));
2003: PetscCall(VecScatterCreate(X, ctx->batch_is, X, NULL, &ctx->plex_batch));
2004: PetscCall(VecDuplicate(X, &ctx->work_vec));
2005: }
2006: PetscFunctionReturn(PETSC_SUCCESS);
2007: }
2009: PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat);
2010: /*@C
2011: DMPlexLandauCreateVelocitySpace - Create a `DMPLEX` velocity space mesh
2013: Collective
2015: Input Parameters:
2016: + comm - The MPI communicator
2017: . dim - velocity space dimension (2 for axisymmetric, 3 for full 3X + 3V solver)
2018: - prefix - prefix for options (not tested)
2020: Output Parameters:
2021: + pack - The `DM` object representing the mesh
2022: . X - A vector (user destroys)
2023: - J - Optional matrix (object destroys)
2025: Level: beginner
2027: .seealso: `DMPlexCreate()`, `DMPlexLandauDestroyVelocitySpace()`
2028: @*/
2029: PetscErrorCode DMPlexLandauCreateVelocitySpace(MPI_Comm comm, PetscInt dim, const char prefix[], Vec *X, Mat *J, DM *pack)
2030: {
2031: LandauCtx *ctx;
2032: Vec Xsub[LANDAU_MAX_GRIDS];
2033: IS grid_batch_is_inv[LANDAU_MAX_GRIDS];
2035: PetscFunctionBegin;
2036: PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Only 2D and 3D supported");
2037: PetscCheck(LANDAU_DIM == dim, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " != LANDAU_DIM %d", dim, LANDAU_DIM);
2038: PetscCall(PetscNew(&ctx));
2039: ctx->comm = comm; /* used for diagnostics and global errors */
2040: /* process options */
2041: PetscCall(ProcessOptions(ctx, prefix));
2042: if (dim == 2) ctx->use_relativistic_corrections = PETSC_FALSE;
2043: /* Create Mesh */
2044: PetscCall(DMCompositeCreate(PETSC_COMM_SELF, pack));
2045: PetscCall(PetscLogEventBegin(ctx->events[13], 0, 0, 0, 0));
2046: PetscCall(PetscLogEventBegin(ctx->events[15], 0, 0, 0, 0));
2047: PetscCall(LandauDMCreateVMeshes(PETSC_COMM_SELF, dim, prefix, ctx, *pack)); // creates grids (Forest of AMR)
2048: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2049: /* create FEM */
2050: PetscCall(SetupDS(ctx->plex[grid], dim, grid, ctx));
2051: /* set initial state */
2052: PetscCall(DMCreateGlobalVector(ctx->plex[grid], &Xsub[grid]));
2053: PetscCall(PetscObjectSetName((PetscObject)Xsub[grid], "u_orig"));
2054: /* initial static refinement, no solve */
2055: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, 0, 1, ctx));
2056: /* forest refinement - forest goes in (if forest), plex comes out */
2057: if (ctx->use_p4est) {
2058: DM plex;
2059: PetscCall(adapt(grid, ctx, &Xsub[grid])); // forest goes in, plex comes out
2060: if (grid == 0) {
2061: PetscCall(DMViewFromOptions(ctx->plex[grid], NULL, "-dm_landau_amr_dm_view")); // need to differentiate - todo
2062: PetscCall(VecViewFromOptions(Xsub[grid], NULL, "-dm_landau_amr_vec_view"));
2063: }
2064: // convert to plex, all done with this level
2065: PetscCall(DMConvert(ctx->plex[grid], DMPLEX, &plex));
2066: PetscCall(DMDestroy(&ctx->plex[grid]));
2067: ctx->plex[grid] = plex;
2068: }
2069: #if !defined(LANDAU_SPECIES_MAJOR)
2070: PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2071: #else
2072: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
2073: PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2074: }
2075: #endif
2076: PetscCall(DMSetApplicationContext(ctx->plex[grid], ctx));
2077: }
2078: #if !defined(LANDAU_SPECIES_MAJOR)
2079: // stack the batched DMs, could do it all here!!! b_id=0
2080: for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) {
2081: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(*pack, ctx->plex[grid]));
2082: }
2083: #endif
2084: // create ctx->mat_offset
2085: ctx->mat_offset[0] = 0;
2086: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2087: PetscInt n;
2088: PetscCall(VecGetLocalSize(Xsub[grid], &n));
2089: ctx->mat_offset[grid + 1] = ctx->mat_offset[grid] + n;
2090: }
2091: // creat DM & Jac
2092: PetscCall(DMSetApplicationContext(*pack, ctx));
2093: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only"));
2094: PetscCall(DMCreateMatrix(*pack, &ctx->J));
2095: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
2096: PetscCall(MatSetOption(ctx->J, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
2097: PetscCall(MatSetOption(ctx->J, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
2098: PetscCall(PetscObjectSetName((PetscObject)ctx->J, "Jac"));
2099: // construct initial conditions in X
2100: PetscCall(DMCreateGlobalVector(*pack, X));
2101: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2102: PetscInt n;
2103: PetscCall(VecGetLocalSize(Xsub[grid], &n));
2104: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2105: PetscScalar const *values;
2106: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2107: PetscCall(LandauSetInitialCondition(ctx->plex[grid], Xsub[grid], grid, b_id, ctx->batch_sz, ctx));
2108: PetscCall(VecGetArrayRead(Xsub[grid], &values)); // Drop whole grid in Plex ordering
2109: for (PetscInt i = 0, idx = moffset; i < n; i++, idx++) PetscCall(VecSetValue(*X, idx, values[i], INSERT_VALUES));
2110: PetscCall(VecRestoreArrayRead(Xsub[grid], &values));
2111: }
2112: }
2113: // cleanup
2114: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(VecDestroy(&Xsub[grid]));
2115: /* check for correct matrix type */
2116: if (ctx->gpu_assembly) { /* we need GPU object with GPU assembly */
2117: PetscBool flg;
2118: if (ctx->deviceType == LANDAU_KOKKOS) {
2119: PetscCall(PetscObjectTypeCompareAny((PetscObject)ctx->J, &flg, MATSEQAIJKOKKOS, MATMPIAIJKOKKOS, MATAIJKOKKOS, ""));
2120: #if defined(PETSC_HAVE_KOKKOS)
2121: 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'");
2122: #else
2123: 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'");
2124: #endif
2125: }
2126: }
2127: PetscCall(PetscLogEventEnd(ctx->events[15], 0, 0, 0, 0));
2129: // create field major ordering
2130: ctx->work_vec = NULL;
2131: ctx->plex_batch = NULL;
2132: ctx->batch_is = NULL;
2133: for (PetscInt i = 0; i < LANDAU_MAX_GRIDS; i++) grid_batch_is_inv[i] = NULL;
2134: PetscCall(PetscLogEventBegin(ctx->events[12], 0, 0, 0, 0));
2135: PetscCall(LandauCreateJacobianMatrix(comm, *X, grid_batch_is_inv, ctx));
2136: PetscCall(PetscLogEventEnd(ctx->events[12], 0, 0, 0, 0));
2138: // create AMR GPU assembly maps and static GPU data
2139: PetscCall(CreateStaticData(dim, grid_batch_is_inv, ctx));
2141: PetscCall(PetscLogEventEnd(ctx->events[13], 0, 0, 0, 0));
2143: // create mass matrix
2144: PetscCall(DMPlexLandauCreateMassMatrix(*pack, NULL));
2146: if (J) *J = ctx->J;
2148: if (ctx->gpu_assembly && ctx->jacobian_field_major_order) {
2149: PetscContainer container;
2150: // cache ctx for KSP with batch/field major Jacobian ordering -ksp_type gmres/etc -dm_landau_jacobian_field_major_order
2151: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2152: PetscCall(PetscContainerSetPointer(container, (void *)ctx));
2153: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "LandauCtx", (PetscObject)container));
2154: PetscCall(PetscContainerDestroy(&container));
2155: // batch solvers need to map -- can batch solvers work
2156: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2157: PetscCall(PetscContainerSetPointer(container, (void *)ctx->plex_batch));
2158: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "plex_batch_is", (PetscObject)container));
2159: PetscCall(PetscContainerDestroy(&container));
2160: }
2161: // for batch solvers
2162: {
2163: PetscContainer container;
2164: PetscInt *pNf;
2165: PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container));
2166: PetscCall(PetscMalloc1(sizeof(*pNf), &pNf));
2167: *pNf = ctx->batch_sz;
2168: PetscCall(PetscContainerSetPointer(container, (void *)pNf));
2169: PetscCall(PetscContainerSetUserDestroy(container, MatrixNfDestroy));
2170: PetscCall(PetscObjectCompose((PetscObject)ctx->J, "batch size", (PetscObject)container));
2171: PetscCall(PetscContainerDestroy(&container));
2172: }
2173: PetscFunctionReturn(PETSC_SUCCESS);
2174: }
2176: /*@C
2177: DMPlexLandauAccess - Access to the distribution function with user callback
2179: Collective
2181: Input Parameters:
2182: + pack - the `DMCOMPOSITE`
2183: . func - call back function
2184: - user_ctx - user context
2186: Input/Output Parameter:
2187: . X - Vector to data to
2189: Level: advanced
2191: .seealso: `DMPlexLandauCreateVelocitySpace()`
2192: @*/
2193: PetscErrorCode DMPlexLandauAccess(DM pack, Vec X, PetscErrorCode (*func)(DM, Vec, PetscInt, PetscInt, PetscInt, void *), void *user_ctx)
2194: {
2195: LandauCtx *ctx;
2197: PetscFunctionBegin;
2198: PetscCall(DMGetApplicationContext(pack, &ctx)); // uses ctx->num_grids; ctx->plex[grid]; ctx->batch_sz; ctx->mat_offset
2199: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2200: PetscInt dim, n;
2201: PetscCall(DMGetDimension(pack, &dim));
2202: for (PetscInt sp = ctx->species_offset[grid], i0 = 0; sp < ctx->species_offset[grid + 1]; sp++, i0++) {
2203: Vec vec;
2204: PetscInt vf[1] = {i0};
2205: IS vis;
2206: DM vdm;
2207: PetscCall(DMCreateSubDM(ctx->plex[grid], 1, vf, &vis, &vdm));
2208: PetscCall(DMSetApplicationContext(vdm, ctx)); // the user might want this
2209: PetscCall(DMCreateGlobalVector(vdm, &vec));
2210: PetscCall(VecGetSize(vec, &n));
2211: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2212: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2213: PetscCall(VecZeroEntries(vec));
2214: /* Add your data with 'dm' for species 'sp' to 'vec' */
2215: PetscCall(func(vdm, vec, i0, grid, b_id, user_ctx));
2216: /* add to global */
2217: PetscScalar const *values;
2218: const PetscInt *offsets;
2219: PetscCall(VecGetArrayRead(vec, &values));
2220: PetscCall(ISGetIndices(vis, &offsets));
2221: for (PetscInt i = 0; i < n; i++) PetscCall(VecSetValue(X, moffset + offsets[i], values[i], ADD_VALUES));
2222: PetscCall(VecRestoreArrayRead(vec, &values));
2223: PetscCall(ISRestoreIndices(vis, &offsets));
2224: } // batch
2225: PetscCall(VecDestroy(&vec));
2226: PetscCall(ISDestroy(&vis));
2227: PetscCall(DMDestroy(&vdm));
2228: }
2229: } // grid
2230: PetscFunctionReturn(PETSC_SUCCESS);
2231: }
2233: /*@
2234: DMPlexLandauDestroyVelocitySpace - Destroy a `DMPLEX` velocity space mesh
2236: Collective
2238: Input/Output Parameters:
2239: . dm - the `DM` to destroy
2241: Level: beginner
2243: .seealso: `DMPlexLandauCreateVelocitySpace()`
2244: @*/
2245: PetscErrorCode DMPlexLandauDestroyVelocitySpace(DM *dm)
2246: {
2247: LandauCtx *ctx;
2249: PetscFunctionBegin;
2250: PetscCall(DMGetApplicationContext(*dm, &ctx));
2251: PetscCall(MatDestroy(&ctx->M));
2252: PetscCall(MatDestroy(&ctx->J));
2253: for (PetscInt ii = 0; ii < ctx->num_species; ii++) PetscCall(PetscFEDestroy(&ctx->fe[ii]));
2254: PetscCall(ISDestroy(&ctx->batch_is));
2255: PetscCall(VecDestroy(&ctx->work_vec));
2256: PetscCall(VecScatterDestroy(&ctx->plex_batch));
2257: if (ctx->deviceType == LANDAU_KOKKOS) {
2258: #if defined(PETSC_HAVE_KOKKOS)
2259: PetscCall(LandauKokkosStaticDataClear(&ctx->SData_d));
2260: #else
2261: SETERRQ(ctx->comm, PETSC_ERR_ARG_WRONG, "-landau_device_type %s not built", "kokkos");
2262: #endif
2263: } else {
2264: if (ctx->SData_d.x) { /* in a CPU run */
2265: 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;
2266: 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;
2267: PetscCall(PetscFree4(ww, xx, yy, invJ));
2268: if (zz) PetscCall(PetscFree(zz));
2269: if (coo_elem_offsets) {
2270: PetscCall(PetscFree3(coo_elem_offsets, coo_elem_fullNb, coo_elem_point_offsets)); // could be NULL
2271: }
2272: PetscCall(PetscFree4(ctx->SData_d.alpha, ctx->SData_d.beta, ctx->SData_d.invMass, ctx->SData_d.lambdas));
2273: }
2274: }
2276: if (ctx->times[LANDAU_MATRIX_TOTAL] > 0) { // OMP timings
2277: PetscCall(PetscPrintf(ctx->comm, "TSStep N 1.0 %10.3e\n", ctx->times[LANDAU_EX2_TSSOLVE]));
2278: 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));
2279: PetscCall(PetscPrintf(ctx->comm, "3: Landau: %10.3e\n", ctx->times[LANDAU_MATRIX_TOTAL]));
2280: PetscCall(PetscPrintf(ctx->comm, "Landau Jacobian %" PetscInt_FMT " 1.0 %10.3e\n", (PetscInt)ctx->times[LANDAU_JACOBIAN_COUNT], ctx->times[LANDAU_JACOBIAN]));
2281: PetscCall(PetscPrintf(ctx->comm, "Landau Operator N 1.0 %10.3e\n", ctx->times[LANDAU_OPERATOR]));
2282: PetscCall(PetscPrintf(ctx->comm, "Landau Mass N 1.0 %10.3e\n", ctx->times[LANDAU_MASS]));
2283: PetscCall(PetscPrintf(ctx->comm, " Jac-f-df (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_F_DF]));
2284: PetscCall(PetscPrintf(ctx->comm, " Kernel (GPU) N 1.0 %10.3e\n", ctx->times[LANDAU_KERNEL]));
2285: PetscCall(PetscPrintf(ctx->comm, "MatLUFactorNum X 1.0 %10.3e\n", ctx->times[KSP_FACTOR]));
2286: PetscCall(PetscPrintf(ctx->comm, "MatSolve X 1.0 %10.3e\n", ctx->times[KSP_SOLVE]));
2287: }
2288: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMDestroy(&ctx->plex[grid]));
2289: PetscCall(PetscFree(ctx));
2290: PetscCall(DMDestroy(dm));
2291: PetscFunctionReturn(PETSC_SUCCESS);
2292: }
2294: /* < v, ru > */
2295: 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)
2296: {
2297: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2298: f0[0] = u[ii];
2299: }
2301: /* < v, ru > */
2302: 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)
2303: {
2304: PetscInt ii = (PetscInt)PetscRealPart(constants[0]), jj = (PetscInt)PetscRealPart(constants[1]);
2305: f0[0] = x[jj] * u[ii]; /* x momentum */
2306: }
2308: 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)
2309: {
2310: PetscInt i, ii = (PetscInt)PetscRealPart(constants[0]);
2311: double tmp1 = 0.;
2312: for (i = 0; i < dim; ++i) tmp1 += x[i] * x[i];
2313: f0[0] = tmp1 * u[ii];
2314: }
2316: static PetscErrorCode gamma_n_f(PetscInt dim, PetscReal time, const PetscReal x[], PetscInt Nf, PetscScalar *u, void *actx)
2317: {
2318: const PetscReal *c2_0_arr = ((PetscReal *)actx);
2319: const PetscReal c02 = c2_0_arr[0];
2321: PetscFunctionBegin;
2322: for (PetscInt s = 0; s < Nf; s++) {
2323: PetscReal tmp1 = 0.;
2324: for (PetscInt i = 0; i < dim; ++i) tmp1 += x[i] * x[i];
2325: #if defined(PETSC_USE_DEBUG)
2326: u[s] = PetscSqrtReal(1. + tmp1 / c02); // u[0] = PetscSqrtReal(1. + xx);
2327: #else
2328: {
2329: PetscReal xx = tmp1 / c02;
2330: u[s] = xx / (PetscSqrtReal(1. + xx) + 1.); // better conditioned = xx/(PetscSqrtReal(1. + xx) + 1.)
2331: }
2332: #endif
2333: }
2334: PetscFunctionReturn(PETSC_SUCCESS);
2335: }
2337: /* < v, ru > */
2338: 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)
2339: {
2340: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2341: f0[0] = 2. * PETSC_PI * x[0] * u[ii];
2342: }
2344: /* < v, ru > */
2345: 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)
2346: {
2347: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2348: f0[0] = 2. * PETSC_PI * x[0] * x[1] * u[ii];
2349: }
2351: 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)
2352: {
2353: PetscInt ii = (PetscInt)PetscRealPart(constants[0]);
2354: f0[0] = 2. * PETSC_PI * x[0] * (x[0] * x[0] + x[1] * x[1]) * u[ii];
2355: }
2357: /*@
2358: DMPlexLandauPrintNorms - collects moments and prints them
2360: Collective
2362: Input Parameters:
2363: + X - the state
2364: - stepi - current step to print
2366: Level: beginner
2368: .seealso: `DMPlexLandauCreateVelocitySpace()`
2369: @*/
2370: PetscErrorCode DMPlexLandauPrintNorms(Vec X, PetscInt stepi)
2371: {
2372: LandauCtx *ctx;
2373: PetscDS prob;
2374: DM pack;
2375: PetscInt cStart, cEnd, dim, ii, i0, nDMs;
2376: PetscScalar xmomentumtot = 0, ymomentumtot = 0, zmomentumtot = 0, energytot = 0, densitytot = 0, tt[LANDAU_MAX_SPECIES];
2377: PetscScalar xmomentum[LANDAU_MAX_SPECIES], ymomentum[LANDAU_MAX_SPECIES], zmomentum[LANDAU_MAX_SPECIES], energy[LANDAU_MAX_SPECIES], density[LANDAU_MAX_SPECIES];
2378: Vec *globXArray;
2380: PetscFunctionBegin;
2381: PetscCall(VecGetDM(X, &pack));
2382: PetscCheck(pack, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Vector has no DM");
2383: PetscCall(DMGetDimension(pack, &dim));
2384: PetscCheck(dim == 2 || dim == 3, PETSC_COMM_SELF, PETSC_ERR_PLIB, "dim %" PetscInt_FMT " not in [2,3]", dim);
2385: PetscCall(DMGetApplicationContext(pack, &ctx));
2386: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2387: /* print momentum and energy */
2388: PetscCall(DMCompositeGetNumberDM(pack, &nDMs));
2389: 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);
2390: PetscCall(PetscMalloc(sizeof(*globXArray) * nDMs, &globXArray));
2391: PetscCall(DMCompositeGetAccessArray(pack, X, nDMs, NULL, globXArray));
2392: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2393: Vec Xloc = globXArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)];
2394: PetscCall(DMGetDS(ctx->plex[grid], &prob));
2395: for (ii = ctx->species_offset[grid], i0 = 0; ii < ctx->species_offset[grid + 1]; ii++, i0++) {
2396: PetscScalar user[2] = {(PetscScalar)i0, (PetscScalar)ctx->charges[ii]};
2397: PetscCall(PetscDSSetConstants(prob, 2, user));
2398: if (dim == 2) { /* 2/3X + 3V (cylindrical coordinates) */
2399: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rden));
2400: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2401: density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii];
2402: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rmom));
2403: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2404: zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2405: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_rv2));
2406: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2407: energy[ii] = tt[0] * 0.5 * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii];
2408: zmomentumtot += zmomentum[ii];
2409: energytot += energy[ii];
2410: densitytot += density[ii];
2411: 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])));
2412: } else { /* 2/3Xloc + 3V */
2413: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_den));
2414: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2415: density[ii] = tt[0] * ctx->n_0 * ctx->charges[ii];
2416: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_mom));
2417: user[1] = 0;
2418: PetscCall(PetscDSSetConstants(prob, 2, user));
2419: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2420: xmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2421: user[1] = 1;
2422: PetscCall(PetscDSSetConstants(prob, 2, user));
2423: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2424: ymomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2425: user[1] = 2;
2426: PetscCall(PetscDSSetConstants(prob, 2, user));
2427: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2428: zmomentum[ii] = tt[0] * ctx->n_0 * ctx->v_0 * ctx->masses[ii];
2429: if (ctx->use_relativistic_corrections) {
2430: /* gamma * M * f */
2431: if (ii == 0 && grid == 0) { // do all at once
2432: Vec Mf, globGamma, *globMfArray, *globGammaArray;
2433: PetscErrorCode (*gammaf[1])(PetscInt, PetscReal, const PetscReal[], PetscInt, PetscScalar[], void *) = {gamma_n_f};
2434: PetscReal *c2_0[1], data[1];
2436: PetscCall(VecDuplicate(X, &globGamma));
2437: PetscCall(VecDuplicate(X, &Mf));
2438: PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globMfArray));
2439: PetscCall(PetscMalloc(sizeof(*globMfArray) * nDMs, &globGammaArray));
2440: /* M * f */
2441: PetscCall(MatMult(ctx->M, X, Mf));
2442: /* gamma */
2443: PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2444: 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
2445: Vec v1 = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)];
2446: data[0] = PetscSqr(C_0(ctx->v_0));
2447: c2_0[0] = &data[0];
2448: PetscCall(DMProjectFunction(ctx->plex[grid], 0., gammaf, (void **)c2_0, INSERT_ALL_VALUES, v1));
2449: }
2450: PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2451: /* gamma * Mf */
2452: PetscCall(DMCompositeGetAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2453: PetscCall(DMCompositeGetAccessArray(pack, Mf, nDMs, NULL, globMfArray));
2454: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) { // yes a grid loop in a grid loop to print nice
2455: PetscInt Nf = ctx->species_offset[grid + 1] - ctx->species_offset[grid], N, bs;
2456: Vec Mfsub = globMfArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], Gsub = globGammaArray[LAND_PACK_IDX(ctx->batch_view_idx, grid)], v1, v2;
2457: // get each component
2458: PetscCall(VecGetSize(Mfsub, &N));
2459: PetscCall(VecCreate(ctx->comm, &v1));
2460: PetscCall(VecSetSizes(v1, PETSC_DECIDE, N / Nf));
2461: PetscCall(VecCreate(ctx->comm, &v2));
2462: PetscCall(VecSetSizes(v2, PETSC_DECIDE, N / Nf));
2463: PetscCall(VecSetFromOptions(v1)); // ???
2464: PetscCall(VecSetFromOptions(v2));
2465: // get each component
2466: PetscCall(VecGetBlockSize(Gsub, &bs));
2467: PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT " in Gsub", bs, Nf);
2468: PetscCall(VecGetBlockSize(Mfsub, &bs));
2469: PetscCheck(bs == Nf, PETSC_COMM_SELF, PETSC_ERR_PLIB, "bs %" PetscInt_FMT " != num_species %" PetscInt_FMT, bs, Nf);
2470: for (PetscInt i = 0, ix = ctx->species_offset[grid]; i < Nf; i++, ix++) {
2471: PetscScalar val;
2472: PetscCall(VecStrideGather(Gsub, i, v1, INSERT_VALUES)); // this is not right -- TODO
2473: PetscCall(VecStrideGather(Mfsub, i, v2, INSERT_VALUES));
2474: PetscCall(VecDot(v1, v2, &val));
2475: energy[ix] = PetscRealPart(val) * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ix];
2476: }
2477: PetscCall(VecDestroy(&v1));
2478: PetscCall(VecDestroy(&v2));
2479: } /* grids */
2480: PetscCall(DMCompositeRestoreAccessArray(pack, globGamma, nDMs, NULL, globGammaArray));
2481: PetscCall(DMCompositeRestoreAccessArray(pack, Mf, nDMs, NULL, globMfArray));
2482: PetscCall(PetscFree(globGammaArray));
2483: PetscCall(PetscFree(globMfArray));
2484: PetscCall(VecDestroy(&globGamma));
2485: PetscCall(VecDestroy(&Mf));
2486: }
2487: } else {
2488: PetscCall(PetscDSSetObjective(prob, 0, &f0_s_v2));
2489: PetscCall(DMPlexComputeIntegralFEM(ctx->plex[grid], Xloc, tt, ctx));
2490: energy[ii] = 0.5 * tt[0] * ctx->n_0 * ctx->v_0 * ctx->v_0 * ctx->masses[ii];
2491: }
2492: 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])));
2493: xmomentumtot += xmomentum[ii];
2494: ymomentumtot += ymomentum[ii];
2495: zmomentumtot += zmomentum[ii];
2496: energytot += energy[ii];
2497: densitytot += density[ii];
2498: }
2499: if (ctx->num_species > 1) PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
2500: }
2501: }
2502: PetscCall(DMCompositeRestoreAccessArray(pack, X, nDMs, NULL, globXArray));
2503: PetscCall(PetscFree(globXArray));
2504: /* totals */
2505: PetscCall(DMPlexGetHeightStratum(ctx->plex[0], 0, &cStart, &cEnd));
2506: if (ctx->num_species > 1) {
2507: if (dim == 2) {
2508: 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),
2509: (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart));
2510: } else {
2511: 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),
2512: (double)(ctx->masses[1] / ctx->masses[0]), cEnd - cStart));
2513: }
2514: } else PetscCall(PetscPrintf(PETSC_COMM_WORLD, " -- %" PetscInt_FMT " cells", cEnd - cStart));
2515: PetscCall(PetscPrintf(PETSC_COMM_WORLD, "\n"));
2516: PetscFunctionReturn(PETSC_SUCCESS);
2517: }
2519: /*@
2520: DMPlexLandauCreateMassMatrix - Create mass matrix for Landau in Plex space (not field major order of Jacobian)
2521: - puts mass matrix into ctx->M
2523: Collective
2525: Input Parameter:
2526: . pack - the `DM` object. Puts matrix in Landau context M field
2528: Output Parameter:
2529: . Amat - The mass matrix (optional), mass matrix is added to the `DM` context
2531: Level: beginner
2533: .seealso: `DMPlexLandauCreateVelocitySpace()`
2534: @*/
2535: PetscErrorCode DMPlexLandauCreateMassMatrix(DM pack, Mat *Amat)
2536: {
2537: DM mass_pack, massDM[LANDAU_MAX_GRIDS];
2538: PetscDS prob;
2539: PetscInt ii, dim, N1 = 1, N2;
2540: LandauCtx *ctx;
2541: Mat packM, subM[LANDAU_MAX_GRIDS];
2543: PetscFunctionBegin;
2545: if (Amat) PetscAssertPointer(Amat, 2);
2546: PetscCall(DMGetApplicationContext(pack, &ctx));
2547: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2548: PetscCall(PetscLogEventBegin(ctx->events[14], 0, 0, 0, 0));
2549: PetscCall(DMGetDimension(pack, &dim));
2550: PetscCall(DMCompositeCreate(PetscObjectComm((PetscObject)pack), &mass_pack));
2551: /* create pack mass matrix */
2552: for (PetscInt grid = 0, ix = 0; grid < ctx->num_grids; grid++) {
2553: PetscCall(DMClone(ctx->plex[grid], &massDM[grid]));
2554: PetscCall(DMCopyFields(ctx->plex[grid], PETSC_DETERMINE, PETSC_DETERMINE, massDM[grid]));
2555: PetscCall(DMCreateDS(massDM[grid]));
2556: PetscCall(DMGetDS(massDM[grid], &prob));
2557: for (ix = 0, ii = ctx->species_offset[grid]; ii < ctx->species_offset[grid + 1]; ii++, ix++) {
2558: if (dim == 3) PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_1, NULL, NULL, NULL));
2559: else PetscCall(PetscDSSetJacobian(prob, ix, ix, g0_r, NULL, NULL, NULL));
2560: }
2561: #if !defined(LANDAU_SPECIES_MAJOR)
2562: PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2563: #else
2564: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) { // add batch size DMs for this species grid
2565: PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2566: }
2567: #endif
2568: PetscCall(DMCreateMatrix(massDM[grid], &subM[grid]));
2569: }
2570: #if !defined(LANDAU_SPECIES_MAJOR)
2571: // stack the batched DMs
2572: for (PetscInt b_id = 1; b_id < ctx->batch_sz; b_id++) {
2573: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(DMCompositeAddDM(mass_pack, massDM[grid]));
2574: }
2575: #endif
2576: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only"));
2577: PetscCall(DMCreateMatrix(mass_pack, &packM));
2578: PetscCall(PetscOptionsInsertString(NULL, "-dm_preallocate_only false"));
2579: PetscCall(MatSetOption(packM, MAT_STRUCTURALLY_SYMMETRIC, PETSC_TRUE));
2580: PetscCall(MatSetOption(packM, MAT_IGNORE_ZERO_ENTRIES, PETSC_TRUE));
2581: PetscCall(DMDestroy(&mass_pack));
2582: /* make mass matrix for each block */
2583: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2584: Vec locX;
2585: DM plex = massDM[grid];
2586: PetscCall(DMGetLocalVector(plex, &locX));
2587: /* Mass matrix is independent of the input, so no need to fill locX */
2588: PetscCall(DMPlexSNESComputeJacobianFEM(plex, locX, subM[grid], subM[grid], ctx));
2589: PetscCall(DMRestoreLocalVector(plex, &locX));
2590: PetscCall(DMDestroy(&massDM[grid]));
2591: }
2592: PetscCall(MatGetSize(ctx->J, &N1, NULL));
2593: PetscCall(MatGetSize(packM, &N2, NULL));
2594: PetscCheck(N1 == N2, PetscObjectComm((PetscObject)pack), PETSC_ERR_PLIB, "Incorrect matrix sizes: |Jacobian| = %" PetscInt_FMT ", |Mass|=%" PetscInt_FMT, N1, N2);
2595: /* assemble block diagonals */
2596: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) {
2597: Mat B = subM[grid];
2598: PetscInt nloc, nzl, *colbuf, COL_BF_SIZE = 1024, row;
2599: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2600: PetscCall(MatGetSize(B, &nloc, NULL));
2601: for (PetscInt b_id = 0; b_id < ctx->batch_sz; b_id++) {
2602: const PetscInt moffset = LAND_MOFFSET(b_id, grid, ctx->batch_sz, ctx->num_grids, ctx->mat_offset);
2603: const PetscInt *cols;
2604: const PetscScalar *vals;
2605: for (PetscInt i = 0; i < nloc; i++) {
2606: PetscCall(MatGetRow(B, i, &nzl, NULL, NULL));
2607: if (nzl > COL_BF_SIZE) {
2608: PetscCall(PetscFree(colbuf));
2609: PetscCall(PetscInfo(pack, "Realloc buffer %" PetscInt_FMT " to %" PetscInt_FMT " (row size %" PetscInt_FMT ") \n", COL_BF_SIZE, 2 * COL_BF_SIZE, nzl));
2610: COL_BF_SIZE = nzl;
2611: PetscCall(PetscMalloc(sizeof(*colbuf) * COL_BF_SIZE, &colbuf));
2612: }
2613: PetscCall(MatGetRow(B, i, &nzl, &cols, &vals));
2614: for (PetscInt j = 0; j < nzl; j++) colbuf[j] = cols[j] + moffset;
2615: row = i + moffset;
2616: PetscCall(MatSetValues(packM, 1, &row, nzl, colbuf, vals, INSERT_VALUES));
2617: PetscCall(MatRestoreRow(B, i, &nzl, &cols, &vals));
2618: }
2619: }
2620: PetscCall(PetscFree(colbuf));
2621: }
2622: // cleanup
2623: for (PetscInt grid = 0; grid < ctx->num_grids; grid++) PetscCall(MatDestroy(&subM[grid]));
2624: PetscCall(MatAssemblyBegin(packM, MAT_FINAL_ASSEMBLY));
2625: PetscCall(MatAssemblyEnd(packM, MAT_FINAL_ASSEMBLY));
2626: PetscCall(PetscObjectSetName((PetscObject)packM, "mass"));
2627: PetscCall(MatViewFromOptions(packM, NULL, "-dm_landau_mass_view"));
2628: ctx->M = packM;
2629: if (Amat) *Amat = packM;
2630: PetscCall(PetscLogEventEnd(ctx->events[14], 0, 0, 0, 0));
2631: PetscFunctionReturn(PETSC_SUCCESS);
2632: }
2634: /*@
2635: DMPlexLandauIFunction - `TS` residual calculation, confusingly this computes the Jacobian w/o mass
2637: Collective
2639: Input Parameters:
2640: + ts - The time stepping context
2641: . time_dummy - current time (not used)
2642: . X - Current state
2643: . X_t - Time derivative of current state
2644: - actx - Landau context
2646: Output Parameter:
2647: . F - The residual
2649: Level: beginner
2651: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIJacobian()`
2652: @*/
2653: PetscErrorCode DMPlexLandauIFunction(TS ts, PetscReal time_dummy, Vec X, Vec X_t, Vec F, void *actx)
2654: {
2655: LandauCtx *ctx = (LandauCtx *)actx;
2656: PetscInt dim;
2657: DM pack;
2658: #if defined(PETSC_HAVE_THREADSAFETY)
2659: double starttime, endtime;
2660: #endif
2661: PetscObjectState state;
2663: PetscFunctionBegin;
2664: PetscCall(TSGetDM(ts, &pack));
2665: PetscCall(DMGetApplicationContext(pack, &ctx));
2666: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2667: if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage));
2668: PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0));
2669: PetscCall(PetscLogEventBegin(ctx->events[0], 0, 0, 0, 0));
2670: #if defined(PETSC_HAVE_THREADSAFETY)
2671: starttime = MPI_Wtime();
2672: #endif
2673: PetscCall(DMGetDimension(pack, &dim));
2674: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2675: if (state != ctx->norm_state) {
2676: PetscCall(MatZeroEntries(ctx->J));
2677: PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, 0.0, (void *)ctx));
2678: PetscCall(MatViewFromOptions(ctx->J, NULL, "-dm_landau_jacobian_view"));
2679: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2680: ctx->norm_state = state;
2681: } else {
2682: PetscCall(PetscInfo(ts, "WARNING Skip forming Jacobian, has not changed %" PetscInt64_FMT "\n", state));
2683: }
2684: /* mat vec for op */
2685: PetscCall(MatMult(ctx->J, X, F)); /* C*f */
2686: /* add time term */
2687: if (X_t) PetscCall(MatMultAdd(ctx->M, X_t, F, F));
2688: #if defined(PETSC_HAVE_THREADSAFETY)
2689: if (ctx->stage) {
2690: endtime = MPI_Wtime();
2691: ctx->times[LANDAU_OPERATOR] += (endtime - starttime);
2692: ctx->times[LANDAU_JACOBIAN] += (endtime - starttime);
2693: ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime);
2694: ctx->times[LANDAU_JACOBIAN_COUNT] += 1;
2695: }
2696: #endif
2697: PetscCall(PetscLogEventEnd(ctx->events[0], 0, 0, 0, 0));
2698: PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0));
2699: if (ctx->stage) PetscCall(PetscLogStagePop());
2700: PetscFunctionReturn(PETSC_SUCCESS);
2701: }
2703: /*@
2704: DMPlexLandauIJacobian - `TS` Jacobian construction, confusingly this adds mass
2706: Collective
2708: Input Parameters:
2709: + ts - The time stepping context
2710: . time_dummy - current time (not used)
2711: . X - Current state
2712: . U_tdummy - Time derivative of current state (not used)
2713: . shift - shift for du/dt term
2714: - actx - Landau context
2716: Output Parameters:
2717: + Amat - Jacobian
2718: - Pmat - same as Amat
2720: Level: beginner
2722: .seealso: `DMPlexLandauCreateVelocitySpace()`, `DMPlexLandauIFunction()`
2723: @*/
2724: PetscErrorCode DMPlexLandauIJacobian(TS ts, PetscReal time_dummy, Vec X, Vec U_tdummy, PetscReal shift, Mat Amat, Mat Pmat, void *actx)
2725: {
2726: LandauCtx *ctx = NULL;
2727: PetscInt dim;
2728: DM pack;
2729: #if defined(PETSC_HAVE_THREADSAFETY)
2730: double starttime, endtime;
2731: #endif
2732: PetscObjectState state;
2734: PetscFunctionBegin;
2735: PetscCall(TSGetDM(ts, &pack));
2736: PetscCall(DMGetApplicationContext(pack, &ctx));
2737: PetscCheck(ctx, PETSC_COMM_SELF, PETSC_ERR_PLIB, "no context");
2738: PetscCheck(Amat == Pmat && Amat == ctx->J, ctx->comm, PETSC_ERR_PLIB, "Amat!=Pmat || Amat!=ctx->J");
2739: PetscCall(DMGetDimension(pack, &dim));
2740: /* get collision Jacobian into A */
2741: if (ctx->stage) PetscCall(PetscLogStagePush(ctx->stage));
2742: PetscCall(PetscLogEventBegin(ctx->events[11], 0, 0, 0, 0));
2743: PetscCall(PetscLogEventBegin(ctx->events[9], 0, 0, 0, 0));
2744: #if defined(PETSC_HAVE_THREADSAFETY)
2745: starttime = MPI_Wtime();
2746: #endif
2747: PetscCheck(shift != 0.0, ctx->comm, PETSC_ERR_PLIB, "zero shift");
2748: PetscCall(PetscObjectStateGet((PetscObject)ctx->J, &state));
2749: PetscCheck(state == ctx->norm_state, ctx->comm, PETSC_ERR_PLIB, "wrong state, %" PetscInt64_FMT " %" PetscInt64_FMT, ctx->norm_state, state);
2750: if (!ctx->use_matrix_mass) {
2751: PetscCall(LandauFormJacobian_Internal(X, ctx->J, dim, shift, (void *)ctx));
2752: } else { /* add mass */
2753: PetscCall(MatAXPY(Pmat, shift, ctx->M, SAME_NONZERO_PATTERN));
2754: }
2755: #if defined(PETSC_HAVE_THREADSAFETY)
2756: if (ctx->stage) {
2757: endtime = MPI_Wtime();
2758: ctx->times[LANDAU_OPERATOR] += (endtime - starttime);
2759: ctx->times[LANDAU_MASS] += (endtime - starttime);
2760: ctx->times[LANDAU_MATRIX_TOTAL] += (endtime - starttime);
2761: }
2762: #endif
2763: PetscCall(PetscLogEventEnd(ctx->events[9], 0, 0, 0, 0));
2764: PetscCall(PetscLogEventEnd(ctx->events[11], 0, 0, 0, 0));
2765: if (ctx->stage) PetscCall(PetscLogStagePop());
2766: PetscFunctionReturn(PETSC_SUCCESS);
2767: }